md/raid6: move raid6 data processing to raid6_pq.ko
[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/seq_file.h>
51 #include "md.h"
52 #include "raid5.h"
53 #include "bitmap.h"
54
55 /*
56  * Stripe cache
57  */
58
59 #define NR_STRIPES              256
60 #define STRIPE_SIZE             PAGE_SIZE
61 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
62 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
63 #define IO_THRESHOLD            1
64 #define BYPASS_THRESHOLD        1
65 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
66 #define HASH_MASK               (NR_HASH - 1)
67
68 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
69
70 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
71  * order without overlap.  There may be several bio's per stripe+device, and
72  * a bio could span several devices.
73  * When walking this list for a particular stripe+device, we must never proceed
74  * beyond a bio that extends past this device, as the next bio might no longer
75  * be valid.
76  * This macro is used to determine the 'next' bio in the list, given the sector
77  * of the current stripe+device
78  */
79 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
80 /*
81  * The following can be used to debug the driver
82  */
83 #define RAID5_PARANOIA  1
84 #if RAID5_PARANOIA && defined(CONFIG_SMP)
85 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
86 #else
87 # define CHECK_DEVLOCK()
88 #endif
89
90 #ifdef DEBUG
91 #define inline
92 #define __inline__
93 #endif
94
95 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
96
97 /*
98  * We maintain a biased count of active stripes in the bottom 16 bits of
99  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
100  */
101 static inline int raid5_bi_phys_segments(struct bio *bio)
102 {
103         return bio->bi_phys_segments & 0xffff;
104 }
105
106 static inline int raid5_bi_hw_segments(struct bio *bio)
107 {
108         return (bio->bi_phys_segments >> 16) & 0xffff;
109 }
110
111 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
112 {
113         --bio->bi_phys_segments;
114         return raid5_bi_phys_segments(bio);
115 }
116
117 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
118 {
119         unsigned short val = raid5_bi_hw_segments(bio);
120
121         --val;
122         bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
123         return val;
124 }
125
126 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
127 {
128         bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16);
129 }
130
131 /* Find first data disk in a raid6 stripe */
132 static inline int raid6_d0(struct stripe_head *sh)
133 {
134         if (sh->ddf_layout)
135                 /* ddf always start from first device */
136                 return 0;
137         /* md starts just after Q block */
138         if (sh->qd_idx == sh->disks - 1)
139                 return 0;
140         else
141                 return sh->qd_idx + 1;
142 }
143 static inline int raid6_next_disk(int disk, int raid_disks)
144 {
145         disk++;
146         return (disk < raid_disks) ? disk : 0;
147 }
148
149 /* When walking through the disks in a raid5, starting at raid6_d0,
150  * We need to map each disk to a 'slot', where the data disks are slot
151  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
152  * is raid_disks-1.  This help does that mapping.
153  */
154 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
155                              int *count, int syndrome_disks)
156 {
157         int slot;
158
159         if (idx == sh->pd_idx)
160                 return syndrome_disks;
161         if (idx == sh->qd_idx)
162                 return syndrome_disks + 1;
163         slot = (*count)++;
164         return slot;
165 }
166
167 static void return_io(struct bio *return_bi)
168 {
169         struct bio *bi = return_bi;
170         while (bi) {
171
172                 return_bi = bi->bi_next;
173                 bi->bi_next = NULL;
174                 bi->bi_size = 0;
175                 bio_endio(bi, 0);
176                 bi = return_bi;
177         }
178 }
179
180 static void print_raid5_conf (raid5_conf_t *conf);
181
182 static int stripe_operations_active(struct stripe_head *sh)
183 {
184         return sh->check_state || sh->reconstruct_state ||
185                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
186                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
187 }
188
189 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
190 {
191         if (atomic_dec_and_test(&sh->count)) {
192                 BUG_ON(!list_empty(&sh->lru));
193                 BUG_ON(atomic_read(&conf->active_stripes)==0);
194                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
195                         if (test_bit(STRIPE_DELAYED, &sh->state)) {
196                                 list_add_tail(&sh->lru, &conf->delayed_list);
197                                 blk_plug_device(conf->mddev->queue);
198                         } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
199                                    sh->bm_seq - conf->seq_write > 0) {
200                                 list_add_tail(&sh->lru, &conf->bitmap_list);
201                                 blk_plug_device(conf->mddev->queue);
202                         } else {
203                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
204                                 list_add_tail(&sh->lru, &conf->handle_list);
205                         }
206                         md_wakeup_thread(conf->mddev->thread);
207                 } else {
208                         BUG_ON(stripe_operations_active(sh));
209                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
210                                 atomic_dec(&conf->preread_active_stripes);
211                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
212                                         md_wakeup_thread(conf->mddev->thread);
213                         }
214                         atomic_dec(&conf->active_stripes);
215                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
216                                 list_add_tail(&sh->lru, &conf->inactive_list);
217                                 wake_up(&conf->wait_for_stripe);
218                                 if (conf->retry_read_aligned)
219                                         md_wakeup_thread(conf->mddev->thread);
220                         }
221                 }
222         }
223 }
224
225 static void release_stripe(struct stripe_head *sh)
226 {
227         raid5_conf_t *conf = sh->raid_conf;
228         unsigned long flags;
229
230         spin_lock_irqsave(&conf->device_lock, flags);
231         __release_stripe(conf, sh);
232         spin_unlock_irqrestore(&conf->device_lock, flags);
233 }
234
235 static inline void remove_hash(struct stripe_head *sh)
236 {
237         pr_debug("remove_hash(), stripe %llu\n",
238                 (unsigned long long)sh->sector);
239
240         hlist_del_init(&sh->hash);
241 }
242
243 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
244 {
245         struct hlist_head *hp = stripe_hash(conf, sh->sector);
246
247         pr_debug("insert_hash(), stripe %llu\n",
248                 (unsigned long long)sh->sector);
249
250         CHECK_DEVLOCK();
251         hlist_add_head(&sh->hash, hp);
252 }
253
254
255 /* find an idle stripe, make sure it is unhashed, and return it. */
256 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
257 {
258         struct stripe_head *sh = NULL;
259         struct list_head *first;
260
261         CHECK_DEVLOCK();
262         if (list_empty(&conf->inactive_list))
263                 goto out;
264         first = conf->inactive_list.next;
265         sh = list_entry(first, struct stripe_head, lru);
266         list_del_init(first);
267         remove_hash(sh);
268         atomic_inc(&conf->active_stripes);
269 out:
270         return sh;
271 }
272
273 static void shrink_buffers(struct stripe_head *sh, int num)
274 {
275         struct page *p;
276         int i;
277
278         for (i=0; i<num ; i++) {
279                 p = sh->dev[i].page;
280                 if (!p)
281                         continue;
282                 sh->dev[i].page = NULL;
283                 put_page(p);
284         }
285 }
286
287 static int grow_buffers(struct stripe_head *sh, int num)
288 {
289         int i;
290
291         for (i=0; i<num; i++) {
292                 struct page *page;
293
294                 if (!(page = alloc_page(GFP_KERNEL))) {
295                         return 1;
296                 }
297                 sh->dev[i].page = page;
298         }
299         return 0;
300 }
301
302 static void raid5_build_block(struct stripe_head *sh, int i);
303 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
304                             struct stripe_head *sh);
305
306 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
307 {
308         raid5_conf_t *conf = sh->raid_conf;
309         int i;
310
311         BUG_ON(atomic_read(&sh->count) != 0);
312         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
313         BUG_ON(stripe_operations_active(sh));
314
315         CHECK_DEVLOCK();
316         pr_debug("init_stripe called, stripe %llu\n",
317                 (unsigned long long)sh->sector);
318
319         remove_hash(sh);
320
321         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
322         sh->sector = sector;
323         stripe_set_idx(sector, conf, previous, sh);
324         sh->state = 0;
325
326
327         for (i = sh->disks; i--; ) {
328                 struct r5dev *dev = &sh->dev[i];
329
330                 if (dev->toread || dev->read || dev->towrite || dev->written ||
331                     test_bit(R5_LOCKED, &dev->flags)) {
332                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
333                                (unsigned long long)sh->sector, i, dev->toread,
334                                dev->read, dev->towrite, dev->written,
335                                test_bit(R5_LOCKED, &dev->flags));
336                         BUG();
337                 }
338                 dev->flags = 0;
339                 raid5_build_block(sh, i);
340         }
341         insert_hash(conf, sh);
342 }
343
344 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
345 {
346         struct stripe_head *sh;
347         struct hlist_node *hn;
348
349         CHECK_DEVLOCK();
350         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
351         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
352                 if (sh->sector == sector && sh->disks == disks)
353                         return sh;
354         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
355         return NULL;
356 }
357
358 static void unplug_slaves(mddev_t *mddev);
359 static void raid5_unplug_device(struct request_queue *q);
360
361 static struct stripe_head *
362 get_active_stripe(raid5_conf_t *conf, sector_t sector,
363                   int previous, int noblock)
364 {
365         struct stripe_head *sh;
366         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
367
368         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
369
370         spin_lock_irq(&conf->device_lock);
371
372         do {
373                 wait_event_lock_irq(conf->wait_for_stripe,
374                                     conf->quiesce == 0,
375                                     conf->device_lock, /* nothing */);
376                 sh = __find_stripe(conf, sector, disks);
377                 if (!sh) {
378                         if (!conf->inactive_blocked)
379                                 sh = get_free_stripe(conf);
380                         if (noblock && sh == NULL)
381                                 break;
382                         if (!sh) {
383                                 conf->inactive_blocked = 1;
384                                 wait_event_lock_irq(conf->wait_for_stripe,
385                                                     !list_empty(&conf->inactive_list) &&
386                                                     (atomic_read(&conf->active_stripes)
387                                                      < (conf->max_nr_stripes *3/4)
388                                                      || !conf->inactive_blocked),
389                                                     conf->device_lock,
390                                                     raid5_unplug_device(conf->mddev->queue)
391                                         );
392                                 conf->inactive_blocked = 0;
393                         } else
394                                 init_stripe(sh, sector, previous);
395                 } else {
396                         if (atomic_read(&sh->count)) {
397                           BUG_ON(!list_empty(&sh->lru));
398                         } else {
399                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
400                                         atomic_inc(&conf->active_stripes);
401                                 if (list_empty(&sh->lru) &&
402                                     !test_bit(STRIPE_EXPANDING, &sh->state))
403                                         BUG();
404                                 list_del_init(&sh->lru);
405                         }
406                 }
407         } while (sh == NULL);
408
409         if (sh)
410                 atomic_inc(&sh->count);
411
412         spin_unlock_irq(&conf->device_lock);
413         return sh;
414 }
415
416 static void
417 raid5_end_read_request(struct bio *bi, int error);
418 static void
419 raid5_end_write_request(struct bio *bi, int error);
420
421 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
422 {
423         raid5_conf_t *conf = sh->raid_conf;
424         int i, disks = sh->disks;
425
426         might_sleep();
427
428         for (i = disks; i--; ) {
429                 int rw;
430                 struct bio *bi;
431                 mdk_rdev_t *rdev;
432                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
433                         rw = WRITE;
434                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
435                         rw = READ;
436                 else
437                         continue;
438
439                 bi = &sh->dev[i].req;
440
441                 bi->bi_rw = rw;
442                 if (rw == WRITE)
443                         bi->bi_end_io = raid5_end_write_request;
444                 else
445                         bi->bi_end_io = raid5_end_read_request;
446
447                 rcu_read_lock();
448                 rdev = rcu_dereference(conf->disks[i].rdev);
449                 if (rdev && test_bit(Faulty, &rdev->flags))
450                         rdev = NULL;
451                 if (rdev)
452                         atomic_inc(&rdev->nr_pending);
453                 rcu_read_unlock();
454
455                 if (rdev) {
456                         if (s->syncing || s->expanding || s->expanded)
457                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
458
459                         set_bit(STRIPE_IO_STARTED, &sh->state);
460
461                         bi->bi_bdev = rdev->bdev;
462                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
463                                 __func__, (unsigned long long)sh->sector,
464                                 bi->bi_rw, i);
465                         atomic_inc(&sh->count);
466                         bi->bi_sector = sh->sector + rdev->data_offset;
467                         bi->bi_flags = 1 << BIO_UPTODATE;
468                         bi->bi_vcnt = 1;
469                         bi->bi_max_vecs = 1;
470                         bi->bi_idx = 0;
471                         bi->bi_io_vec = &sh->dev[i].vec;
472                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
473                         bi->bi_io_vec[0].bv_offset = 0;
474                         bi->bi_size = STRIPE_SIZE;
475                         bi->bi_next = NULL;
476                         if (rw == WRITE &&
477                             test_bit(R5_ReWrite, &sh->dev[i].flags))
478                                 atomic_add(STRIPE_SECTORS,
479                                         &rdev->corrected_errors);
480                         generic_make_request(bi);
481                 } else {
482                         if (rw == WRITE)
483                                 set_bit(STRIPE_DEGRADED, &sh->state);
484                         pr_debug("skip op %ld on disc %d for sector %llu\n",
485                                 bi->bi_rw, i, (unsigned long long)sh->sector);
486                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
487                         set_bit(STRIPE_HANDLE, &sh->state);
488                 }
489         }
490 }
491
492 static struct dma_async_tx_descriptor *
493 async_copy_data(int frombio, struct bio *bio, struct page *page,
494         sector_t sector, struct dma_async_tx_descriptor *tx)
495 {
496         struct bio_vec *bvl;
497         struct page *bio_page;
498         int i;
499         int page_offset;
500
501         if (bio->bi_sector >= sector)
502                 page_offset = (signed)(bio->bi_sector - sector) * 512;
503         else
504                 page_offset = (signed)(sector - bio->bi_sector) * -512;
505         bio_for_each_segment(bvl, bio, i) {
506                 int len = bio_iovec_idx(bio, i)->bv_len;
507                 int clen;
508                 int b_offset = 0;
509
510                 if (page_offset < 0) {
511                         b_offset = -page_offset;
512                         page_offset += b_offset;
513                         len -= b_offset;
514                 }
515
516                 if (len > 0 && page_offset + len > STRIPE_SIZE)
517                         clen = STRIPE_SIZE - page_offset;
518                 else
519                         clen = len;
520
521                 if (clen > 0) {
522                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
523                         bio_page = bio_iovec_idx(bio, i)->bv_page;
524                         if (frombio)
525                                 tx = async_memcpy(page, bio_page, page_offset,
526                                         b_offset, clen,
527                                         ASYNC_TX_DEP_ACK,
528                                         tx, NULL, NULL);
529                         else
530                                 tx = async_memcpy(bio_page, page, b_offset,
531                                         page_offset, clen,
532                                         ASYNC_TX_DEP_ACK,
533                                         tx, NULL, NULL);
534                 }
535                 if (clen < len) /* hit end of page */
536                         break;
537                 page_offset +=  len;
538         }
539
540         return tx;
541 }
542
543 static void ops_complete_biofill(void *stripe_head_ref)
544 {
545         struct stripe_head *sh = stripe_head_ref;
546         struct bio *return_bi = NULL;
547         raid5_conf_t *conf = sh->raid_conf;
548         int i;
549
550         pr_debug("%s: stripe %llu\n", __func__,
551                 (unsigned long long)sh->sector);
552
553         /* clear completed biofills */
554         spin_lock_irq(&conf->device_lock);
555         for (i = sh->disks; i--; ) {
556                 struct r5dev *dev = &sh->dev[i];
557
558                 /* acknowledge completion of a biofill operation */
559                 /* and check if we need to reply to a read request,
560                  * new R5_Wantfill requests are held off until
561                  * !STRIPE_BIOFILL_RUN
562                  */
563                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
564                         struct bio *rbi, *rbi2;
565
566                         BUG_ON(!dev->read);
567                         rbi = dev->read;
568                         dev->read = NULL;
569                         while (rbi && rbi->bi_sector <
570                                 dev->sector + STRIPE_SECTORS) {
571                                 rbi2 = r5_next_bio(rbi, dev->sector);
572                                 if (!raid5_dec_bi_phys_segments(rbi)) {
573                                         rbi->bi_next = return_bi;
574                                         return_bi = rbi;
575                                 }
576                                 rbi = rbi2;
577                         }
578                 }
579         }
580         spin_unlock_irq(&conf->device_lock);
581         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
582
583         return_io(return_bi);
584
585         set_bit(STRIPE_HANDLE, &sh->state);
586         release_stripe(sh);
587 }
588
589 static void ops_run_biofill(struct stripe_head *sh)
590 {
591         struct dma_async_tx_descriptor *tx = NULL;
592         raid5_conf_t *conf = sh->raid_conf;
593         int i;
594
595         pr_debug("%s: stripe %llu\n", __func__,
596                 (unsigned long long)sh->sector);
597
598         for (i = sh->disks; i--; ) {
599                 struct r5dev *dev = &sh->dev[i];
600                 if (test_bit(R5_Wantfill, &dev->flags)) {
601                         struct bio *rbi;
602                         spin_lock_irq(&conf->device_lock);
603                         dev->read = rbi = dev->toread;
604                         dev->toread = NULL;
605                         spin_unlock_irq(&conf->device_lock);
606                         while (rbi && rbi->bi_sector <
607                                 dev->sector + STRIPE_SECTORS) {
608                                 tx = async_copy_data(0, rbi, dev->page,
609                                         dev->sector, tx);
610                                 rbi = r5_next_bio(rbi, dev->sector);
611                         }
612                 }
613         }
614
615         atomic_inc(&sh->count);
616         async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
617                 ops_complete_biofill, sh);
618 }
619
620 static void ops_complete_compute5(void *stripe_head_ref)
621 {
622         struct stripe_head *sh = stripe_head_ref;
623         int target = sh->ops.target;
624         struct r5dev *tgt = &sh->dev[target];
625
626         pr_debug("%s: stripe %llu\n", __func__,
627                 (unsigned long long)sh->sector);
628
629         set_bit(R5_UPTODATE, &tgt->flags);
630         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
631         clear_bit(R5_Wantcompute, &tgt->flags);
632         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
633         if (sh->check_state == check_state_compute_run)
634                 sh->check_state = check_state_compute_result;
635         set_bit(STRIPE_HANDLE, &sh->state);
636         release_stripe(sh);
637 }
638
639 static struct dma_async_tx_descriptor *ops_run_compute5(struct stripe_head *sh)
640 {
641         /* kernel stack size limits the total number of disks */
642         int disks = sh->disks;
643         struct page *xor_srcs[disks];
644         int target = sh->ops.target;
645         struct r5dev *tgt = &sh->dev[target];
646         struct page *xor_dest = tgt->page;
647         int count = 0;
648         struct dma_async_tx_descriptor *tx;
649         int i;
650
651         pr_debug("%s: stripe %llu block: %d\n",
652                 __func__, (unsigned long long)sh->sector, target);
653         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
654
655         for (i = disks; i--; )
656                 if (i != target)
657                         xor_srcs[count++] = sh->dev[i].page;
658
659         atomic_inc(&sh->count);
660
661         if (unlikely(count == 1))
662                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
663                         0, NULL, ops_complete_compute5, sh);
664         else
665                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
666                         ASYNC_TX_XOR_ZERO_DST, NULL,
667                         ops_complete_compute5, sh);
668
669         return tx;
670 }
671
672 static void ops_complete_prexor(void *stripe_head_ref)
673 {
674         struct stripe_head *sh = stripe_head_ref;
675
676         pr_debug("%s: stripe %llu\n", __func__,
677                 (unsigned long long)sh->sector);
678 }
679
680 static struct dma_async_tx_descriptor *
681 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
682 {
683         /* kernel stack size limits the total number of disks */
684         int disks = sh->disks;
685         struct page *xor_srcs[disks];
686         int count = 0, pd_idx = sh->pd_idx, i;
687
688         /* existing parity data subtracted */
689         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
690
691         pr_debug("%s: stripe %llu\n", __func__,
692                 (unsigned long long)sh->sector);
693
694         for (i = disks; i--; ) {
695                 struct r5dev *dev = &sh->dev[i];
696                 /* Only process blocks that are known to be uptodate */
697                 if (test_bit(R5_Wantdrain, &dev->flags))
698                         xor_srcs[count++] = dev->page;
699         }
700
701         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
702                 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
703                 ops_complete_prexor, sh);
704
705         return tx;
706 }
707
708 static struct dma_async_tx_descriptor *
709 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
710 {
711         int disks = sh->disks;
712         int i;
713
714         pr_debug("%s: stripe %llu\n", __func__,
715                 (unsigned long long)sh->sector);
716
717         for (i = disks; i--; ) {
718                 struct r5dev *dev = &sh->dev[i];
719                 struct bio *chosen;
720
721                 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
722                         struct bio *wbi;
723
724                         spin_lock(&sh->lock);
725                         chosen = dev->towrite;
726                         dev->towrite = NULL;
727                         BUG_ON(dev->written);
728                         wbi = dev->written = chosen;
729                         spin_unlock(&sh->lock);
730
731                         while (wbi && wbi->bi_sector <
732                                 dev->sector + STRIPE_SECTORS) {
733                                 tx = async_copy_data(1, wbi, dev->page,
734                                         dev->sector, tx);
735                                 wbi = r5_next_bio(wbi, dev->sector);
736                         }
737                 }
738         }
739
740         return tx;
741 }
742
743 static void ops_complete_postxor(void *stripe_head_ref)
744 {
745         struct stripe_head *sh = stripe_head_ref;
746         int disks = sh->disks, i, pd_idx = sh->pd_idx;
747
748         pr_debug("%s: stripe %llu\n", __func__,
749                 (unsigned long long)sh->sector);
750
751         for (i = disks; i--; ) {
752                 struct r5dev *dev = &sh->dev[i];
753                 if (dev->written || i == pd_idx)
754                         set_bit(R5_UPTODATE, &dev->flags);
755         }
756
757         if (sh->reconstruct_state == reconstruct_state_drain_run)
758                 sh->reconstruct_state = reconstruct_state_drain_result;
759         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
760                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
761         else {
762                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
763                 sh->reconstruct_state = reconstruct_state_result;
764         }
765
766         set_bit(STRIPE_HANDLE, &sh->state);
767         release_stripe(sh);
768 }
769
770 static void
771 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
772 {
773         /* kernel stack size limits the total number of disks */
774         int disks = sh->disks;
775         struct page *xor_srcs[disks];
776
777         int count = 0, pd_idx = sh->pd_idx, i;
778         struct page *xor_dest;
779         int prexor = 0;
780         unsigned long flags;
781
782         pr_debug("%s: stripe %llu\n", __func__,
783                 (unsigned long long)sh->sector);
784
785         /* check if prexor is active which means only process blocks
786          * that are part of a read-modify-write (written)
787          */
788         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
789                 prexor = 1;
790                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
791                 for (i = disks; i--; ) {
792                         struct r5dev *dev = &sh->dev[i];
793                         if (dev->written)
794                                 xor_srcs[count++] = dev->page;
795                 }
796         } else {
797                 xor_dest = sh->dev[pd_idx].page;
798                 for (i = disks; i--; ) {
799                         struct r5dev *dev = &sh->dev[i];
800                         if (i != pd_idx)
801                                 xor_srcs[count++] = dev->page;
802                 }
803         }
804
805         /* 1/ if we prexor'd then the dest is reused as a source
806          * 2/ if we did not prexor then we are redoing the parity
807          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
808          * for the synchronous xor case
809          */
810         flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
811                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
812
813         atomic_inc(&sh->count);
814
815         if (unlikely(count == 1)) {
816                 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
817                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
818                         flags, tx, ops_complete_postxor, sh);
819         } else
820                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
821                         flags, tx, ops_complete_postxor, sh);
822 }
823
824 static void ops_complete_check(void *stripe_head_ref)
825 {
826         struct stripe_head *sh = stripe_head_ref;
827
828         pr_debug("%s: stripe %llu\n", __func__,
829                 (unsigned long long)sh->sector);
830
831         sh->check_state = check_state_check_result;
832         set_bit(STRIPE_HANDLE, &sh->state);
833         release_stripe(sh);
834 }
835
836 static void ops_run_check(struct stripe_head *sh)
837 {
838         /* kernel stack size limits the total number of disks */
839         int disks = sh->disks;
840         struct page *xor_srcs[disks];
841         struct dma_async_tx_descriptor *tx;
842
843         int count = 0, pd_idx = sh->pd_idx, i;
844         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
845
846         pr_debug("%s: stripe %llu\n", __func__,
847                 (unsigned long long)sh->sector);
848
849         for (i = disks; i--; ) {
850                 struct r5dev *dev = &sh->dev[i];
851                 if (i != pd_idx)
852                         xor_srcs[count++] = dev->page;
853         }
854
855         tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
856                 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
857
858         atomic_inc(&sh->count);
859         tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
860                 ops_complete_check, sh);
861 }
862
863 static void raid5_run_ops(struct stripe_head *sh, unsigned long ops_request)
864 {
865         int overlap_clear = 0, i, disks = sh->disks;
866         struct dma_async_tx_descriptor *tx = NULL;
867
868         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
869                 ops_run_biofill(sh);
870                 overlap_clear++;
871         }
872
873         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
874                 tx = ops_run_compute5(sh);
875                 /* terminate the chain if postxor is not set to be run */
876                 if (tx && !test_bit(STRIPE_OP_POSTXOR, &ops_request))
877                         async_tx_ack(tx);
878         }
879
880         if (test_bit(STRIPE_OP_PREXOR, &ops_request))
881                 tx = ops_run_prexor(sh, tx);
882
883         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
884                 tx = ops_run_biodrain(sh, tx);
885                 overlap_clear++;
886         }
887
888         if (test_bit(STRIPE_OP_POSTXOR, &ops_request))
889                 ops_run_postxor(sh, tx);
890
891         if (test_bit(STRIPE_OP_CHECK, &ops_request))
892                 ops_run_check(sh);
893
894         if (overlap_clear)
895                 for (i = disks; i--; ) {
896                         struct r5dev *dev = &sh->dev[i];
897                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
898                                 wake_up(&sh->raid_conf->wait_for_overlap);
899                 }
900 }
901
902 static int grow_one_stripe(raid5_conf_t *conf)
903 {
904         struct stripe_head *sh;
905         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
906         if (!sh)
907                 return 0;
908         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
909         sh->raid_conf = conf;
910         spin_lock_init(&sh->lock);
911
912         if (grow_buffers(sh, conf->raid_disks)) {
913                 shrink_buffers(sh, conf->raid_disks);
914                 kmem_cache_free(conf->slab_cache, sh);
915                 return 0;
916         }
917         sh->disks = conf->raid_disks;
918         /* we just created an active stripe so... */
919         atomic_set(&sh->count, 1);
920         atomic_inc(&conf->active_stripes);
921         INIT_LIST_HEAD(&sh->lru);
922         release_stripe(sh);
923         return 1;
924 }
925
926 static int grow_stripes(raid5_conf_t *conf, int num)
927 {
928         struct kmem_cache *sc;
929         int devs = conf->raid_disks;
930
931         sprintf(conf->cache_name[0],
932                 "raid%d-%s", conf->level, mdname(conf->mddev));
933         sprintf(conf->cache_name[1],
934                 "raid%d-%s-alt", conf->level, mdname(conf->mddev));
935         conf->active_name = 0;
936         sc = kmem_cache_create(conf->cache_name[conf->active_name],
937                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
938                                0, 0, NULL);
939         if (!sc)
940                 return 1;
941         conf->slab_cache = sc;
942         conf->pool_size = devs;
943         while (num--)
944                 if (!grow_one_stripe(conf))
945                         return 1;
946         return 0;
947 }
948
949 #ifdef CONFIG_MD_RAID5_RESHAPE
950 static int resize_stripes(raid5_conf_t *conf, int newsize)
951 {
952         /* Make all the stripes able to hold 'newsize' devices.
953          * New slots in each stripe get 'page' set to a new page.
954          *
955          * This happens in stages:
956          * 1/ create a new kmem_cache and allocate the required number of
957          *    stripe_heads.
958          * 2/ gather all the old stripe_heads and tranfer the pages across
959          *    to the new stripe_heads.  This will have the side effect of
960          *    freezing the array as once all stripe_heads have been collected,
961          *    no IO will be possible.  Old stripe heads are freed once their
962          *    pages have been transferred over, and the old kmem_cache is
963          *    freed when all stripes are done.
964          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
965          *    we simple return a failre status - no need to clean anything up.
966          * 4/ allocate new pages for the new slots in the new stripe_heads.
967          *    If this fails, we don't bother trying the shrink the
968          *    stripe_heads down again, we just leave them as they are.
969          *    As each stripe_head is processed the new one is released into
970          *    active service.
971          *
972          * Once step2 is started, we cannot afford to wait for a write,
973          * so we use GFP_NOIO allocations.
974          */
975         struct stripe_head *osh, *nsh;
976         LIST_HEAD(newstripes);
977         struct disk_info *ndisks;
978         int err;
979         struct kmem_cache *sc;
980         int i;
981
982         if (newsize <= conf->pool_size)
983                 return 0; /* never bother to shrink */
984
985         err = md_allow_write(conf->mddev);
986         if (err)
987                 return err;
988
989         /* Step 1 */
990         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
991                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
992                                0, 0, NULL);
993         if (!sc)
994                 return -ENOMEM;
995
996         for (i = conf->max_nr_stripes; i; i--) {
997                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
998                 if (!nsh)
999                         break;
1000
1001                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1002
1003                 nsh->raid_conf = conf;
1004                 spin_lock_init(&nsh->lock);
1005
1006                 list_add(&nsh->lru, &newstripes);
1007         }
1008         if (i) {
1009                 /* didn't get enough, give up */
1010                 while (!list_empty(&newstripes)) {
1011                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1012                         list_del(&nsh->lru);
1013                         kmem_cache_free(sc, nsh);
1014                 }
1015                 kmem_cache_destroy(sc);
1016                 return -ENOMEM;
1017         }
1018         /* Step 2 - Must use GFP_NOIO now.
1019          * OK, we have enough stripes, start collecting inactive
1020          * stripes and copying them over
1021          */
1022         list_for_each_entry(nsh, &newstripes, lru) {
1023                 spin_lock_irq(&conf->device_lock);
1024                 wait_event_lock_irq(conf->wait_for_stripe,
1025                                     !list_empty(&conf->inactive_list),
1026                                     conf->device_lock,
1027                                     unplug_slaves(conf->mddev)
1028                         );
1029                 osh = get_free_stripe(conf);
1030                 spin_unlock_irq(&conf->device_lock);
1031                 atomic_set(&nsh->count, 1);
1032                 for(i=0; i<conf->pool_size; i++)
1033                         nsh->dev[i].page = osh->dev[i].page;
1034                 for( ; i<newsize; i++)
1035                         nsh->dev[i].page = NULL;
1036                 kmem_cache_free(conf->slab_cache, osh);
1037         }
1038         kmem_cache_destroy(conf->slab_cache);
1039
1040         /* Step 3.
1041          * At this point, we are holding all the stripes so the array
1042          * is completely stalled, so now is a good time to resize
1043          * conf->disks.
1044          */
1045         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1046         if (ndisks) {
1047                 for (i=0; i<conf->raid_disks; i++)
1048                         ndisks[i] = conf->disks[i];
1049                 kfree(conf->disks);
1050                 conf->disks = ndisks;
1051         } else
1052                 err = -ENOMEM;
1053
1054         /* Step 4, return new stripes to service */
1055         while(!list_empty(&newstripes)) {
1056                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1057                 list_del_init(&nsh->lru);
1058                 for (i=conf->raid_disks; i < newsize; i++)
1059                         if (nsh->dev[i].page == NULL) {
1060                                 struct page *p = alloc_page(GFP_NOIO);
1061                                 nsh->dev[i].page = p;
1062                                 if (!p)
1063                                         err = -ENOMEM;
1064                         }
1065                 release_stripe(nsh);
1066         }
1067         /* critical section pass, GFP_NOIO no longer needed */
1068
1069         conf->slab_cache = sc;
1070         conf->active_name = 1-conf->active_name;
1071         conf->pool_size = newsize;
1072         return err;
1073 }
1074 #endif
1075
1076 static int drop_one_stripe(raid5_conf_t *conf)
1077 {
1078         struct stripe_head *sh;
1079
1080         spin_lock_irq(&conf->device_lock);
1081         sh = get_free_stripe(conf);
1082         spin_unlock_irq(&conf->device_lock);
1083         if (!sh)
1084                 return 0;
1085         BUG_ON(atomic_read(&sh->count));
1086         shrink_buffers(sh, conf->pool_size);
1087         kmem_cache_free(conf->slab_cache, sh);
1088         atomic_dec(&conf->active_stripes);
1089         return 1;
1090 }
1091
1092 static void shrink_stripes(raid5_conf_t *conf)
1093 {
1094         while (drop_one_stripe(conf))
1095                 ;
1096
1097         if (conf->slab_cache)
1098                 kmem_cache_destroy(conf->slab_cache);
1099         conf->slab_cache = NULL;
1100 }
1101
1102 static void raid5_end_read_request(struct bio * bi, int error)
1103 {
1104         struct stripe_head *sh = bi->bi_private;
1105         raid5_conf_t *conf = sh->raid_conf;
1106         int disks = sh->disks, i;
1107         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1108         char b[BDEVNAME_SIZE];
1109         mdk_rdev_t *rdev;
1110
1111
1112         for (i=0 ; i<disks; i++)
1113                 if (bi == &sh->dev[i].req)
1114                         break;
1115
1116         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1117                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1118                 uptodate);
1119         if (i == disks) {
1120                 BUG();
1121                 return;
1122         }
1123
1124         if (uptodate) {
1125                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1126                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1127                         rdev = conf->disks[i].rdev;
1128                         printk_rl(KERN_INFO "raid5:%s: read error corrected"
1129                                   " (%lu sectors at %llu on %s)\n",
1130                                   mdname(conf->mddev), STRIPE_SECTORS,
1131                                   (unsigned long long)(sh->sector
1132                                                        + rdev->data_offset),
1133                                   bdevname(rdev->bdev, b));
1134                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1135                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1136                 }
1137                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1138                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1139         } else {
1140                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1141                 int retry = 0;
1142                 rdev = conf->disks[i].rdev;
1143
1144                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1145                 atomic_inc(&rdev->read_errors);
1146                 if (conf->mddev->degraded)
1147                         printk_rl(KERN_WARNING
1148                                   "raid5:%s: read error not correctable "
1149                                   "(sector %llu on %s).\n",
1150                                   mdname(conf->mddev),
1151                                   (unsigned long long)(sh->sector
1152                                                        + rdev->data_offset),
1153                                   bdn);
1154                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1155                         /* Oh, no!!! */
1156                         printk_rl(KERN_WARNING
1157                                   "raid5:%s: read error NOT corrected!! "
1158                                   "(sector %llu on %s).\n",
1159                                   mdname(conf->mddev),
1160                                   (unsigned long long)(sh->sector
1161                                                        + rdev->data_offset),
1162                                   bdn);
1163                 else if (atomic_read(&rdev->read_errors)
1164                          > conf->max_nr_stripes)
1165                         printk(KERN_WARNING
1166                                "raid5:%s: Too many read errors, failing device %s.\n",
1167                                mdname(conf->mddev), bdn);
1168                 else
1169                         retry = 1;
1170                 if (retry)
1171                         set_bit(R5_ReadError, &sh->dev[i].flags);
1172                 else {
1173                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1174                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1175                         md_error(conf->mddev, rdev);
1176                 }
1177         }
1178         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1179         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1180         set_bit(STRIPE_HANDLE, &sh->state);
1181         release_stripe(sh);
1182 }
1183
1184 static void raid5_end_write_request(struct bio *bi, int error)
1185 {
1186         struct stripe_head *sh = bi->bi_private;
1187         raid5_conf_t *conf = sh->raid_conf;
1188         int disks = sh->disks, i;
1189         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1190
1191         for (i=0 ; i<disks; i++)
1192                 if (bi == &sh->dev[i].req)
1193                         break;
1194
1195         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1196                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1197                 uptodate);
1198         if (i == disks) {
1199                 BUG();
1200                 return;
1201         }
1202
1203         if (!uptodate)
1204                 md_error(conf->mddev, conf->disks[i].rdev);
1205
1206         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1207         
1208         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1209         set_bit(STRIPE_HANDLE, &sh->state);
1210         release_stripe(sh);
1211 }
1212
1213
1214 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1215         
1216 static void raid5_build_block(struct stripe_head *sh, int i)
1217 {
1218         struct r5dev *dev = &sh->dev[i];
1219
1220         bio_init(&dev->req);
1221         dev->req.bi_io_vec = &dev->vec;
1222         dev->req.bi_vcnt++;
1223         dev->req.bi_max_vecs++;
1224         dev->vec.bv_page = dev->page;
1225         dev->vec.bv_len = STRIPE_SIZE;
1226         dev->vec.bv_offset = 0;
1227
1228         dev->req.bi_sector = sh->sector;
1229         dev->req.bi_private = sh;
1230
1231         dev->flags = 0;
1232         dev->sector = compute_blocknr(sh, i);
1233 }
1234
1235 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1236 {
1237         char b[BDEVNAME_SIZE];
1238         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1239         pr_debug("raid5: error called\n");
1240
1241         if (!test_bit(Faulty, &rdev->flags)) {
1242                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1243                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1244                         unsigned long flags;
1245                         spin_lock_irqsave(&conf->device_lock, flags);
1246                         mddev->degraded++;
1247                         spin_unlock_irqrestore(&conf->device_lock, flags);
1248                         /*
1249                          * if recovery was running, make sure it aborts.
1250                          */
1251                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1252                 }
1253                 set_bit(Faulty, &rdev->flags);
1254                 printk(KERN_ALERT
1255                        "raid5: Disk failure on %s, disabling device.\n"
1256                        "raid5: Operation continuing on %d devices.\n",
1257                        bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1258         }
1259 }
1260
1261 /*
1262  * Input: a 'big' sector number,
1263  * Output: index of the data and parity disk, and the sector # in them.
1264  */
1265 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1266                                      int previous, int *dd_idx,
1267                                      struct stripe_head *sh)
1268 {
1269         long stripe;
1270         unsigned long chunk_number;
1271         unsigned int chunk_offset;
1272         int pd_idx, qd_idx;
1273         int ddf_layout = 0;
1274         sector_t new_sector;
1275         int sectors_per_chunk = conf->chunk_size >> 9;
1276         int raid_disks = previous ? conf->previous_raid_disks
1277                                   : conf->raid_disks;
1278         int data_disks = raid_disks - conf->max_degraded;
1279
1280         /* First compute the information on this sector */
1281
1282         /*
1283          * Compute the chunk number and the sector offset inside the chunk
1284          */
1285         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1286         chunk_number = r_sector;
1287         BUG_ON(r_sector != chunk_number);
1288
1289         /*
1290          * Compute the stripe number
1291          */
1292         stripe = chunk_number / data_disks;
1293
1294         /*
1295          * Compute the data disk and parity disk indexes inside the stripe
1296          */
1297         *dd_idx = chunk_number % data_disks;
1298
1299         /*
1300          * Select the parity disk based on the user selected algorithm.
1301          */
1302         pd_idx = qd_idx = ~0;
1303         switch(conf->level) {
1304         case 4:
1305                 pd_idx = data_disks;
1306                 break;
1307         case 5:
1308                 switch (conf->algorithm) {
1309                 case ALGORITHM_LEFT_ASYMMETRIC:
1310                         pd_idx = data_disks - stripe % raid_disks;
1311                         if (*dd_idx >= pd_idx)
1312                                 (*dd_idx)++;
1313                         break;
1314                 case ALGORITHM_RIGHT_ASYMMETRIC:
1315                         pd_idx = stripe % raid_disks;
1316                         if (*dd_idx >= pd_idx)
1317                                 (*dd_idx)++;
1318                         break;
1319                 case ALGORITHM_LEFT_SYMMETRIC:
1320                         pd_idx = data_disks - stripe % raid_disks;
1321                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1322                         break;
1323                 case ALGORITHM_RIGHT_SYMMETRIC:
1324                         pd_idx = stripe % raid_disks;
1325                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1326                         break;
1327                 case ALGORITHM_PARITY_0:
1328                         pd_idx = 0;
1329                         (*dd_idx)++;
1330                         break;
1331                 case ALGORITHM_PARITY_N:
1332                         pd_idx = data_disks;
1333                         break;
1334                 default:
1335                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1336                                 conf->algorithm);
1337                         BUG();
1338                 }
1339                 break;
1340         case 6:
1341
1342                 switch (conf->algorithm) {
1343                 case ALGORITHM_LEFT_ASYMMETRIC:
1344                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1345                         qd_idx = pd_idx + 1;
1346                         if (pd_idx == raid_disks-1) {
1347                                 (*dd_idx)++;    /* Q D D D P */
1348                                 qd_idx = 0;
1349                         } else if (*dd_idx >= pd_idx)
1350                                 (*dd_idx) += 2; /* D D P Q D */
1351                         break;
1352                 case ALGORITHM_RIGHT_ASYMMETRIC:
1353                         pd_idx = stripe % raid_disks;
1354                         qd_idx = pd_idx + 1;
1355                         if (pd_idx == raid_disks-1) {
1356                                 (*dd_idx)++;    /* Q D D D P */
1357                                 qd_idx = 0;
1358                         } else if (*dd_idx >= pd_idx)
1359                                 (*dd_idx) += 2; /* D D P Q D */
1360                         break;
1361                 case ALGORITHM_LEFT_SYMMETRIC:
1362                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1363                         qd_idx = (pd_idx + 1) % raid_disks;
1364                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1365                         break;
1366                 case ALGORITHM_RIGHT_SYMMETRIC:
1367                         pd_idx = stripe % raid_disks;
1368                         qd_idx = (pd_idx + 1) % raid_disks;
1369                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1370                         break;
1371
1372                 case ALGORITHM_PARITY_0:
1373                         pd_idx = 0;
1374                         qd_idx = 1;
1375                         (*dd_idx) += 2;
1376                         break;
1377                 case ALGORITHM_PARITY_N:
1378                         pd_idx = data_disks;
1379                         qd_idx = data_disks + 1;
1380                         break;
1381
1382                 case ALGORITHM_ROTATING_ZERO_RESTART:
1383                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
1384                          * of blocks for computing Q is different.
1385                          */
1386                         pd_idx = stripe % raid_disks;
1387                         qd_idx = pd_idx + 1;
1388                         if (pd_idx == raid_disks-1) {
1389                                 (*dd_idx)++;    /* Q D D D P */
1390                                 qd_idx = 0;
1391                         } else if (*dd_idx >= pd_idx)
1392                                 (*dd_idx) += 2; /* D D P Q D */
1393                         ddf_layout = 1;
1394                         break;
1395
1396                 case ALGORITHM_ROTATING_N_RESTART:
1397                         /* Same a left_asymmetric, by first stripe is
1398                          * D D D P Q  rather than
1399                          * Q D D D P
1400                          */
1401                         pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1402                         qd_idx = pd_idx + 1;
1403                         if (pd_idx == raid_disks-1) {
1404                                 (*dd_idx)++;    /* Q D D D P */
1405                                 qd_idx = 0;
1406                         } else if (*dd_idx >= pd_idx)
1407                                 (*dd_idx) += 2; /* D D P Q D */
1408                         ddf_layout = 1;
1409                         break;
1410
1411                 case ALGORITHM_ROTATING_N_CONTINUE:
1412                         /* Same as left_symmetric but Q is before P */
1413                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1414                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1415                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1416                         ddf_layout = 1;
1417                         break;
1418
1419                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1420                         /* RAID5 left_asymmetric, with Q on last device */
1421                         pd_idx = data_disks - stripe % (raid_disks-1);
1422                         if (*dd_idx >= pd_idx)
1423                                 (*dd_idx)++;
1424                         qd_idx = raid_disks - 1;
1425                         break;
1426
1427                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1428                         pd_idx = stripe % (raid_disks-1);
1429                         if (*dd_idx >= pd_idx)
1430                                 (*dd_idx)++;
1431                         qd_idx = raid_disks - 1;
1432                         break;
1433
1434                 case ALGORITHM_LEFT_SYMMETRIC_6:
1435                         pd_idx = data_disks - stripe % (raid_disks-1);
1436                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1437                         qd_idx = raid_disks - 1;
1438                         break;
1439
1440                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1441                         pd_idx = stripe % (raid_disks-1);
1442                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1443                         qd_idx = raid_disks - 1;
1444                         break;
1445
1446                 case ALGORITHM_PARITY_0_6:
1447                         pd_idx = 0;
1448                         (*dd_idx)++;
1449                         qd_idx = raid_disks - 1;
1450                         break;
1451
1452
1453                 default:
1454                         printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1455                                conf->algorithm);
1456                         BUG();
1457                 }
1458                 break;
1459         }
1460
1461         if (sh) {
1462                 sh->pd_idx = pd_idx;
1463                 sh->qd_idx = qd_idx;
1464                 sh->ddf_layout = ddf_layout;
1465         }
1466         /*
1467          * Finally, compute the new sector number
1468          */
1469         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1470         return new_sector;
1471 }
1472
1473
1474 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1475 {
1476         raid5_conf_t *conf = sh->raid_conf;
1477         int raid_disks = sh->disks;
1478         int data_disks = raid_disks - conf->max_degraded;
1479         sector_t new_sector = sh->sector, check;
1480         int sectors_per_chunk = conf->chunk_size >> 9;
1481         sector_t stripe;
1482         int chunk_offset;
1483         int chunk_number, dummy1, dd_idx = i;
1484         sector_t r_sector;
1485         struct stripe_head sh2;
1486
1487
1488         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1489         stripe = new_sector;
1490         BUG_ON(new_sector != stripe);
1491
1492         if (i == sh->pd_idx)
1493                 return 0;
1494         switch(conf->level) {
1495         case 4: break;
1496         case 5:
1497                 switch (conf->algorithm) {
1498                 case ALGORITHM_LEFT_ASYMMETRIC:
1499                 case ALGORITHM_RIGHT_ASYMMETRIC:
1500                         if (i > sh->pd_idx)
1501                                 i--;
1502                         break;
1503                 case ALGORITHM_LEFT_SYMMETRIC:
1504                 case ALGORITHM_RIGHT_SYMMETRIC:
1505                         if (i < sh->pd_idx)
1506                                 i += raid_disks;
1507                         i -= (sh->pd_idx + 1);
1508                         break;
1509                 case ALGORITHM_PARITY_0:
1510                         i -= 1;
1511                         break;
1512                 case ALGORITHM_PARITY_N:
1513                         break;
1514                 default:
1515                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1516                                conf->algorithm);
1517                         BUG();
1518                 }
1519                 break;
1520         case 6:
1521                 if (i == sh->qd_idx)
1522                         return 0; /* It is the Q disk */
1523                 switch (conf->algorithm) {
1524                 case ALGORITHM_LEFT_ASYMMETRIC:
1525                 case ALGORITHM_RIGHT_ASYMMETRIC:
1526                 case ALGORITHM_ROTATING_ZERO_RESTART:
1527                 case ALGORITHM_ROTATING_N_RESTART:
1528                         if (sh->pd_idx == raid_disks-1)
1529                                 i--;    /* Q D D D P */
1530                         else if (i > sh->pd_idx)
1531                                 i -= 2; /* D D P Q D */
1532                         break;
1533                 case ALGORITHM_LEFT_SYMMETRIC:
1534                 case ALGORITHM_RIGHT_SYMMETRIC:
1535                         if (sh->pd_idx == raid_disks-1)
1536                                 i--; /* Q D D D P */
1537                         else {
1538                                 /* D D P Q D */
1539                                 if (i < sh->pd_idx)
1540                                         i += raid_disks;
1541                                 i -= (sh->pd_idx + 2);
1542                         }
1543                         break;
1544                 case ALGORITHM_PARITY_0:
1545                         i -= 2;
1546                         break;
1547                 case ALGORITHM_PARITY_N:
1548                         break;
1549                 case ALGORITHM_ROTATING_N_CONTINUE:
1550                         if (sh->pd_idx == 0)
1551                                 i--;    /* P D D D Q */
1552                         else if (i > sh->pd_idx)
1553                                 i -= 2; /* D D Q P D */
1554                         break;
1555                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1556                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1557                         if (i > sh->pd_idx)
1558                                 i--;
1559                         break;
1560                 case ALGORITHM_LEFT_SYMMETRIC_6:
1561                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1562                         if (i < sh->pd_idx)
1563                                 i += data_disks + 1;
1564                         i -= (sh->pd_idx + 1);
1565                         break;
1566                 case ALGORITHM_PARITY_0_6:
1567                         i -= 1;
1568                         break;
1569                 default:
1570                         printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1571                                conf->algorithm);
1572                         BUG();
1573                 }
1574                 break;
1575         }
1576
1577         chunk_number = stripe * data_disks + i;
1578         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1579
1580         check = raid5_compute_sector(conf, r_sector,
1581                                      (raid_disks != conf->raid_disks),
1582                                      &dummy1, &sh2);
1583         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1584                 || sh2.qd_idx != sh->qd_idx) {
1585                 printk(KERN_ERR "compute_blocknr: map not correct\n");
1586                 return 0;
1587         }
1588         return r_sector;
1589 }
1590
1591
1592
1593 /*
1594  * Copy data between a page in the stripe cache, and one or more bion
1595  * The page could align with the middle of the bio, or there could be
1596  * several bion, each with several bio_vecs, which cover part of the page
1597  * Multiple bion are linked together on bi_next.  There may be extras
1598  * at the end of this list.  We ignore them.
1599  */
1600 static void copy_data(int frombio, struct bio *bio,
1601                      struct page *page,
1602                      sector_t sector)
1603 {
1604         char *pa = page_address(page);
1605         struct bio_vec *bvl;
1606         int i;
1607         int page_offset;
1608
1609         if (bio->bi_sector >= sector)
1610                 page_offset = (signed)(bio->bi_sector - sector) * 512;
1611         else
1612                 page_offset = (signed)(sector - bio->bi_sector) * -512;
1613         bio_for_each_segment(bvl, bio, i) {
1614                 int len = bio_iovec_idx(bio,i)->bv_len;
1615                 int clen;
1616                 int b_offset = 0;
1617
1618                 if (page_offset < 0) {
1619                         b_offset = -page_offset;
1620                         page_offset += b_offset;
1621                         len -= b_offset;
1622                 }
1623
1624                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1625                         clen = STRIPE_SIZE - page_offset;
1626                 else clen = len;
1627
1628                 if (clen > 0) {
1629                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1630                         if (frombio)
1631                                 memcpy(pa+page_offset, ba+b_offset, clen);
1632                         else
1633                                 memcpy(ba+b_offset, pa+page_offset, clen);
1634                         __bio_kunmap_atomic(ba, KM_USER0);
1635                 }
1636                 if (clen < len) /* hit end of page */
1637                         break;
1638                 page_offset +=  len;
1639         }
1640 }
1641
1642 #define check_xor()     do {                                              \
1643                                 if (count == MAX_XOR_BLOCKS) {            \
1644                                 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1645                                 count = 0;                                \
1646                            }                                              \
1647                         } while(0)
1648
1649 static void compute_parity6(struct stripe_head *sh, int method)
1650 {
1651         raid5_conf_t *conf = sh->raid_conf;
1652         int i, pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1653         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1654         struct bio *chosen;
1655         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1656         void *ptrs[syndrome_disks+2];
1657
1658         pd_idx = sh->pd_idx;
1659         qd_idx = sh->qd_idx;
1660         d0_idx = raid6_d0(sh);
1661
1662         pr_debug("compute_parity, stripe %llu, method %d\n",
1663                 (unsigned long long)sh->sector, method);
1664
1665         switch(method) {
1666         case READ_MODIFY_WRITE:
1667                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
1668         case RECONSTRUCT_WRITE:
1669                 for (i= disks; i-- ;)
1670                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1671                                 chosen = sh->dev[i].towrite;
1672                                 sh->dev[i].towrite = NULL;
1673
1674                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1675                                         wake_up(&conf->wait_for_overlap);
1676
1677                                 BUG_ON(sh->dev[i].written);
1678                                 sh->dev[i].written = chosen;
1679                         }
1680                 break;
1681         case CHECK_PARITY:
1682                 BUG();          /* Not implemented yet */
1683         }
1684
1685         for (i = disks; i--;)
1686                 if (sh->dev[i].written) {
1687                         sector_t sector = sh->dev[i].sector;
1688                         struct bio *wbi = sh->dev[i].written;
1689                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1690                                 copy_data(1, wbi, sh->dev[i].page, sector);
1691                                 wbi = r5_next_bio(wbi, sector);
1692                         }
1693
1694                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1695                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1696                 }
1697
1698         /* Note that unlike RAID-5, the ordering of the disks matters greatly.*/
1699
1700         for (i = 0; i < disks; i++)
1701                 ptrs[i] = (void *)raid6_empty_zero_page;
1702
1703         count = 0;
1704         i = d0_idx;
1705         do {
1706                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1707
1708                 ptrs[slot] = page_address(sh->dev[i].page);
1709                 if (slot < syndrome_disks &&
1710                     !test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
1711                         printk(KERN_ERR "block %d/%d not uptodate "
1712                                "on parity calc\n", i, count);
1713                         BUG();
1714                 }
1715
1716                 i = raid6_next_disk(i, disks);
1717         } while (i != d0_idx);
1718         BUG_ON(count != syndrome_disks);
1719
1720         raid6_call.gen_syndrome(syndrome_disks+2, STRIPE_SIZE, ptrs);
1721
1722         switch(method) {
1723         case RECONSTRUCT_WRITE:
1724                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1725                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1726                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1727                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
1728                 break;
1729         case UPDATE_PARITY:
1730                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1731                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1732                 break;
1733         }
1734 }
1735
1736
1737 /* Compute one missing block */
1738 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1739 {
1740         int i, count, disks = sh->disks;
1741         void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1742         int qd_idx = sh->qd_idx;
1743
1744         pr_debug("compute_block_1, stripe %llu, idx %d\n",
1745                 (unsigned long long)sh->sector, dd_idx);
1746
1747         if ( dd_idx == qd_idx ) {
1748                 /* We're actually computing the Q drive */
1749                 compute_parity6(sh, UPDATE_PARITY);
1750         } else {
1751                 dest = page_address(sh->dev[dd_idx].page);
1752                 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1753                 count = 0;
1754                 for (i = disks ; i--; ) {
1755                         if (i == dd_idx || i == qd_idx)
1756                                 continue;
1757                         p = page_address(sh->dev[i].page);
1758                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1759                                 ptr[count++] = p;
1760                         else
1761                                 printk("compute_block() %d, stripe %llu, %d"
1762                                        " not present\n", dd_idx,
1763                                        (unsigned long long)sh->sector, i);
1764
1765                         check_xor();
1766                 }
1767                 if (count)
1768                         xor_blocks(count, STRIPE_SIZE, dest, ptr);
1769                 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1770                 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1771         }
1772 }
1773
1774 /* Compute two missing blocks */
1775 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1776 {
1777         int i, count, disks = sh->disks;
1778         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1779         int d0_idx = raid6_d0(sh);
1780         int faila = -1, failb = -1;
1781         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1782         void *ptrs[syndrome_disks+2];
1783
1784         for (i = 0; i < disks ; i++)
1785                 ptrs[i] = (void *)raid6_empty_zero_page;
1786         count = 0;
1787         i = d0_idx;
1788         do {
1789                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1790
1791                 ptrs[slot] = page_address(sh->dev[i].page);
1792
1793                 if (i == dd_idx1)
1794                         faila = slot;
1795                 if (i == dd_idx2)
1796                         failb = slot;
1797                 i = raid6_next_disk(i, disks);
1798         } while (i != d0_idx);
1799         BUG_ON(count != syndrome_disks);
1800
1801         BUG_ON(faila == failb);
1802         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1803
1804         pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1805                  (unsigned long long)sh->sector, dd_idx1, dd_idx2,
1806                  faila, failb);
1807
1808         if (failb == syndrome_disks+1) {
1809                 /* Q disk is one of the missing disks */
1810                 if (faila == syndrome_disks) {
1811                         /* Missing P+Q, just recompute */
1812                         compute_parity6(sh, UPDATE_PARITY);
1813                         return;
1814                 } else {
1815                         /* We're missing D+Q; recompute D from P */
1816                         compute_block_1(sh, ((dd_idx1 == sh->qd_idx) ?
1817                                              dd_idx2 : dd_idx1),
1818                                         0);
1819                         compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1820                         return;
1821                 }
1822         }
1823
1824         /* We're missing D+P or D+D; */
1825         if (failb == syndrome_disks) {
1826                 /* We're missing D+P. */
1827                 raid6_datap_recov(syndrome_disks+2, STRIPE_SIZE, faila, ptrs);
1828         } else {
1829                 /* We're missing D+D. */
1830                 raid6_2data_recov(syndrome_disks+2, STRIPE_SIZE, faila, failb,
1831                                   ptrs);
1832         }
1833
1834         /* Both the above update both missing blocks */
1835         set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1836         set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1837 }
1838
1839 static void
1840 schedule_reconstruction5(struct stripe_head *sh, struct stripe_head_state *s,
1841                          int rcw, int expand)
1842 {
1843         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1844
1845         if (rcw) {
1846                 /* if we are not expanding this is a proper write request, and
1847                  * there will be bios with new data to be drained into the
1848                  * stripe cache
1849                  */
1850                 if (!expand) {
1851                         sh->reconstruct_state = reconstruct_state_drain_run;
1852                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1853                 } else
1854                         sh->reconstruct_state = reconstruct_state_run;
1855
1856                 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1857
1858                 for (i = disks; i--; ) {
1859                         struct r5dev *dev = &sh->dev[i];
1860
1861                         if (dev->towrite) {
1862                                 set_bit(R5_LOCKED, &dev->flags);
1863                                 set_bit(R5_Wantdrain, &dev->flags);
1864                                 if (!expand)
1865                                         clear_bit(R5_UPTODATE, &dev->flags);
1866                                 s->locked++;
1867                         }
1868                 }
1869                 if (s->locked + 1 == disks)
1870                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1871                                 atomic_inc(&sh->raid_conf->pending_full_writes);
1872         } else {
1873                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1874                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1875
1876                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
1877                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
1878                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1879                 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1880
1881                 for (i = disks; i--; ) {
1882                         struct r5dev *dev = &sh->dev[i];
1883                         if (i == pd_idx)
1884                                 continue;
1885
1886                         if (dev->towrite &&
1887                             (test_bit(R5_UPTODATE, &dev->flags) ||
1888                              test_bit(R5_Wantcompute, &dev->flags))) {
1889                                 set_bit(R5_Wantdrain, &dev->flags);
1890                                 set_bit(R5_LOCKED, &dev->flags);
1891                                 clear_bit(R5_UPTODATE, &dev->flags);
1892                                 s->locked++;
1893                         }
1894                 }
1895         }
1896
1897         /* keep the parity disk locked while asynchronous operations
1898          * are in flight
1899          */
1900         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1901         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1902         s->locked++;
1903
1904         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
1905                 __func__, (unsigned long long)sh->sector,
1906                 s->locked, s->ops_request);
1907 }
1908
1909 /*
1910  * Each stripe/dev can have one or more bion attached.
1911  * toread/towrite point to the first in a chain.
1912  * The bi_next chain must be in order.
1913  */
1914 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1915 {
1916         struct bio **bip;
1917         raid5_conf_t *conf = sh->raid_conf;
1918         int firstwrite=0;
1919
1920         pr_debug("adding bh b#%llu to stripe s#%llu\n",
1921                 (unsigned long long)bi->bi_sector,
1922                 (unsigned long long)sh->sector);
1923
1924
1925         spin_lock(&sh->lock);
1926         spin_lock_irq(&conf->device_lock);
1927         if (forwrite) {
1928                 bip = &sh->dev[dd_idx].towrite;
1929                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1930                         firstwrite = 1;
1931         } else
1932                 bip = &sh->dev[dd_idx].toread;
1933         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1934                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1935                         goto overlap;
1936                 bip = & (*bip)->bi_next;
1937         }
1938         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1939                 goto overlap;
1940
1941         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1942         if (*bip)
1943                 bi->bi_next = *bip;
1944         *bip = bi;
1945         bi->bi_phys_segments++;
1946         spin_unlock_irq(&conf->device_lock);
1947         spin_unlock(&sh->lock);
1948
1949         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1950                 (unsigned long long)bi->bi_sector,
1951                 (unsigned long long)sh->sector, dd_idx);
1952
1953         if (conf->mddev->bitmap && firstwrite) {
1954                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1955                                   STRIPE_SECTORS, 0);
1956                 sh->bm_seq = conf->seq_flush+1;
1957                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1958         }
1959
1960         if (forwrite) {
1961                 /* check if page is covered */
1962                 sector_t sector = sh->dev[dd_idx].sector;
1963                 for (bi=sh->dev[dd_idx].towrite;
1964                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1965                              bi && bi->bi_sector <= sector;
1966                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1967                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1968                                 sector = bi->bi_sector + (bi->bi_size>>9);
1969                 }
1970                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1971                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1972         }
1973         return 1;
1974
1975  overlap:
1976         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1977         spin_unlock_irq(&conf->device_lock);
1978         spin_unlock(&sh->lock);
1979         return 0;
1980 }
1981
1982 static void end_reshape(raid5_conf_t *conf);
1983
1984 static int page_is_zero(struct page *p)
1985 {
1986         char *a = page_address(p);
1987         return ((*(u32*)a) == 0 &&
1988                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1989 }
1990
1991 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
1992                             struct stripe_head *sh)
1993 {
1994         int sectors_per_chunk = conf->chunk_size >> 9;
1995         int dd_idx;
1996         int chunk_offset = sector_div(stripe, sectors_per_chunk);
1997         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
1998
1999         raid5_compute_sector(conf,
2000                              stripe * (disks - conf->max_degraded)
2001                              *sectors_per_chunk + chunk_offset,
2002                              previous,
2003                              &dd_idx, sh);
2004 }
2005
2006 static void
2007 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2008                                 struct stripe_head_state *s, int disks,
2009                                 struct bio **return_bi)
2010 {
2011         int i;
2012         for (i = disks; i--; ) {
2013                 struct bio *bi;
2014                 int bitmap_end = 0;
2015
2016                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2017                         mdk_rdev_t *rdev;
2018                         rcu_read_lock();
2019                         rdev = rcu_dereference(conf->disks[i].rdev);
2020                         if (rdev && test_bit(In_sync, &rdev->flags))
2021                                 /* multiple read failures in one stripe */
2022                                 md_error(conf->mddev, rdev);
2023                         rcu_read_unlock();
2024                 }
2025                 spin_lock_irq(&conf->device_lock);
2026                 /* fail all writes first */
2027                 bi = sh->dev[i].towrite;
2028                 sh->dev[i].towrite = NULL;
2029                 if (bi) {
2030                         s->to_write--;
2031                         bitmap_end = 1;
2032                 }
2033
2034                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2035                         wake_up(&conf->wait_for_overlap);
2036
2037                 while (bi && bi->bi_sector <
2038                         sh->dev[i].sector + STRIPE_SECTORS) {
2039                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2040                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2041                         if (!raid5_dec_bi_phys_segments(bi)) {
2042                                 md_write_end(conf->mddev);
2043                                 bi->bi_next = *return_bi;
2044                                 *return_bi = bi;
2045                         }
2046                         bi = nextbi;
2047                 }
2048                 /* and fail all 'written' */
2049                 bi = sh->dev[i].written;
2050                 sh->dev[i].written = NULL;
2051                 if (bi) bitmap_end = 1;
2052                 while (bi && bi->bi_sector <
2053                        sh->dev[i].sector + STRIPE_SECTORS) {
2054                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2055                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2056                         if (!raid5_dec_bi_phys_segments(bi)) {
2057                                 md_write_end(conf->mddev);
2058                                 bi->bi_next = *return_bi;
2059                                 *return_bi = bi;
2060                         }
2061                         bi = bi2;
2062                 }
2063
2064                 /* fail any reads if this device is non-operational and
2065                  * the data has not reached the cache yet.
2066                  */
2067                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2068                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2069                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
2070                         bi = sh->dev[i].toread;
2071                         sh->dev[i].toread = NULL;
2072                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2073                                 wake_up(&conf->wait_for_overlap);
2074                         if (bi) s->to_read--;
2075                         while (bi && bi->bi_sector <
2076                                sh->dev[i].sector + STRIPE_SECTORS) {
2077                                 struct bio *nextbi =
2078                                         r5_next_bio(bi, sh->dev[i].sector);
2079                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2080                                 if (!raid5_dec_bi_phys_segments(bi)) {
2081                                         bi->bi_next = *return_bi;
2082                                         *return_bi = bi;
2083                                 }
2084                                 bi = nextbi;
2085                         }
2086                 }
2087                 spin_unlock_irq(&conf->device_lock);
2088                 if (bitmap_end)
2089                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2090                                         STRIPE_SECTORS, 0, 0);
2091         }
2092
2093         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2094                 if (atomic_dec_and_test(&conf->pending_full_writes))
2095                         md_wakeup_thread(conf->mddev->thread);
2096 }
2097
2098 /* fetch_block5 - checks the given member device to see if its data needs
2099  * to be read or computed to satisfy a request.
2100  *
2101  * Returns 1 when no more member devices need to be checked, otherwise returns
2102  * 0 to tell the loop in handle_stripe_fill5 to continue
2103  */
2104 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2105                         int disk_idx, int disks)
2106 {
2107         struct r5dev *dev = &sh->dev[disk_idx];
2108         struct r5dev *failed_dev = &sh->dev[s->failed_num];
2109
2110         /* is the data in this block needed, and can we get it? */
2111         if (!test_bit(R5_LOCKED, &dev->flags) &&
2112             !test_bit(R5_UPTODATE, &dev->flags) &&
2113             (dev->toread ||
2114              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2115              s->syncing || s->expanding ||
2116              (s->failed &&
2117               (failed_dev->toread ||
2118                (failed_dev->towrite &&
2119                 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2120                 /* We would like to get this block, possibly by computing it,
2121                  * otherwise read it if the backing disk is insync
2122                  */
2123                 if ((s->uptodate == disks - 1) &&
2124                     (s->failed && disk_idx == s->failed_num)) {
2125                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2126                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2127                         set_bit(R5_Wantcompute, &dev->flags);
2128                         sh->ops.target = disk_idx;
2129                         s->req_compute = 1;
2130                         /* Careful: from this point on 'uptodate' is in the eye
2131                          * of raid5_run_ops which services 'compute' operations
2132                          * before writes. R5_Wantcompute flags a block that will
2133                          * be R5_UPTODATE by the time it is needed for a
2134                          * subsequent operation.
2135                          */
2136                         s->uptodate++;
2137                         return 1; /* uptodate + compute == disks */
2138                 } else if (test_bit(R5_Insync, &dev->flags)) {
2139                         set_bit(R5_LOCKED, &dev->flags);
2140                         set_bit(R5_Wantread, &dev->flags);
2141                         s->locked++;
2142                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2143                                 s->syncing);
2144                 }
2145         }
2146
2147         return 0;
2148 }
2149
2150 /**
2151  * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2152  */
2153 static void handle_stripe_fill5(struct stripe_head *sh,
2154                         struct stripe_head_state *s, int disks)
2155 {
2156         int i;
2157
2158         /* look for blocks to read/compute, skip this if a compute
2159          * is already in flight, or if the stripe contents are in the
2160          * midst of changing due to a write
2161          */
2162         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2163             !sh->reconstruct_state)
2164                 for (i = disks; i--; )
2165                         if (fetch_block5(sh, s, i, disks))
2166                                 break;
2167         set_bit(STRIPE_HANDLE, &sh->state);
2168 }
2169
2170 static void handle_stripe_fill6(struct stripe_head *sh,
2171                         struct stripe_head_state *s, struct r6_state *r6s,
2172                         int disks)
2173 {
2174         int i;
2175         for (i = disks; i--; ) {
2176                 struct r5dev *dev = &sh->dev[i];
2177                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2178                     !test_bit(R5_UPTODATE, &dev->flags) &&
2179                     (dev->toread || (dev->towrite &&
2180                      !test_bit(R5_OVERWRITE, &dev->flags)) ||
2181                      s->syncing || s->expanding ||
2182                      (s->failed >= 1 &&
2183                       (sh->dev[r6s->failed_num[0]].toread ||
2184                        s->to_write)) ||
2185                      (s->failed >= 2 &&
2186                       (sh->dev[r6s->failed_num[1]].toread ||
2187                        s->to_write)))) {
2188                         /* we would like to get this block, possibly
2189                          * by computing it, but we might not be able to
2190                          */
2191                         if ((s->uptodate == disks - 1) &&
2192                             (s->failed && (i == r6s->failed_num[0] ||
2193                                            i == r6s->failed_num[1]))) {
2194                                 pr_debug("Computing stripe %llu block %d\n",
2195                                        (unsigned long long)sh->sector, i);
2196                                 compute_block_1(sh, i, 0);
2197                                 s->uptodate++;
2198                         } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2199                                 /* Computing 2-failure is *very* expensive; only
2200                                  * do it if failed >= 2
2201                                  */
2202                                 int other;
2203                                 for (other = disks; other--; ) {
2204                                         if (other == i)
2205                                                 continue;
2206                                         if (!test_bit(R5_UPTODATE,
2207                                               &sh->dev[other].flags))
2208                                                 break;
2209                                 }
2210                                 BUG_ON(other < 0);
2211                                 pr_debug("Computing stripe %llu blocks %d,%d\n",
2212                                        (unsigned long long)sh->sector,
2213                                        i, other);
2214                                 compute_block_2(sh, i, other);
2215                                 s->uptodate += 2;
2216                         } else if (test_bit(R5_Insync, &dev->flags)) {
2217                                 set_bit(R5_LOCKED, &dev->flags);
2218                                 set_bit(R5_Wantread, &dev->flags);
2219                                 s->locked++;
2220                                 pr_debug("Reading block %d (sync=%d)\n",
2221                                         i, s->syncing);
2222                         }
2223                 }
2224         }
2225         set_bit(STRIPE_HANDLE, &sh->state);
2226 }
2227
2228
2229 /* handle_stripe_clean_event
2230  * any written block on an uptodate or failed drive can be returned.
2231  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2232  * never LOCKED, so we don't need to test 'failed' directly.
2233  */
2234 static void handle_stripe_clean_event(raid5_conf_t *conf,
2235         struct stripe_head *sh, int disks, struct bio **return_bi)
2236 {
2237         int i;
2238         struct r5dev *dev;
2239
2240         for (i = disks; i--; )
2241                 if (sh->dev[i].written) {
2242                         dev = &sh->dev[i];
2243                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2244                                 test_bit(R5_UPTODATE, &dev->flags)) {
2245                                 /* We can return any write requests */
2246                                 struct bio *wbi, *wbi2;
2247                                 int bitmap_end = 0;
2248                                 pr_debug("Return write for disc %d\n", i);
2249                                 spin_lock_irq(&conf->device_lock);
2250                                 wbi = dev->written;
2251                                 dev->written = NULL;
2252                                 while (wbi && wbi->bi_sector <
2253                                         dev->sector + STRIPE_SECTORS) {
2254                                         wbi2 = r5_next_bio(wbi, dev->sector);
2255                                         if (!raid5_dec_bi_phys_segments(wbi)) {
2256                                                 md_write_end(conf->mddev);
2257                                                 wbi->bi_next = *return_bi;
2258                                                 *return_bi = wbi;
2259                                         }
2260                                         wbi = wbi2;
2261                                 }
2262                                 if (dev->towrite == NULL)
2263                                         bitmap_end = 1;
2264                                 spin_unlock_irq(&conf->device_lock);
2265                                 if (bitmap_end)
2266                                         bitmap_endwrite(conf->mddev->bitmap,
2267                                                         sh->sector,
2268                                                         STRIPE_SECTORS,
2269                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2270                                                         0);
2271                         }
2272                 }
2273
2274         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2275                 if (atomic_dec_and_test(&conf->pending_full_writes))
2276                         md_wakeup_thread(conf->mddev->thread);
2277 }
2278
2279 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2280                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2281 {
2282         int rmw = 0, rcw = 0, i;
2283         for (i = disks; i--; ) {
2284                 /* would I have to read this buffer for read_modify_write */
2285                 struct r5dev *dev = &sh->dev[i];
2286                 if ((dev->towrite || i == sh->pd_idx) &&
2287                     !test_bit(R5_LOCKED, &dev->flags) &&
2288                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2289                       test_bit(R5_Wantcompute, &dev->flags))) {
2290                         if (test_bit(R5_Insync, &dev->flags))
2291                                 rmw++;
2292                         else
2293                                 rmw += 2*disks;  /* cannot read it */
2294                 }
2295                 /* Would I have to read this buffer for reconstruct_write */
2296                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2297                     !test_bit(R5_LOCKED, &dev->flags) &&
2298                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2299                     test_bit(R5_Wantcompute, &dev->flags))) {
2300                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2301                         else
2302                                 rcw += 2*disks;
2303                 }
2304         }
2305         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2306                 (unsigned long long)sh->sector, rmw, rcw);
2307         set_bit(STRIPE_HANDLE, &sh->state);
2308         if (rmw < rcw && rmw > 0)
2309                 /* prefer read-modify-write, but need to get some data */
2310                 for (i = disks; i--; ) {
2311                         struct r5dev *dev = &sh->dev[i];
2312                         if ((dev->towrite || i == sh->pd_idx) &&
2313                             !test_bit(R5_LOCKED, &dev->flags) &&
2314                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2315                             test_bit(R5_Wantcompute, &dev->flags)) &&
2316                             test_bit(R5_Insync, &dev->flags)) {
2317                                 if (
2318                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2319                                         pr_debug("Read_old block "
2320                                                 "%d for r-m-w\n", i);
2321                                         set_bit(R5_LOCKED, &dev->flags);
2322                                         set_bit(R5_Wantread, &dev->flags);
2323                                         s->locked++;
2324                                 } else {
2325                                         set_bit(STRIPE_DELAYED, &sh->state);
2326                                         set_bit(STRIPE_HANDLE, &sh->state);
2327                                 }
2328                         }
2329                 }
2330         if (rcw <= rmw && rcw > 0)
2331                 /* want reconstruct write, but need to get some data */
2332                 for (i = disks; i--; ) {
2333                         struct r5dev *dev = &sh->dev[i];
2334                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2335                             i != sh->pd_idx &&
2336                             !test_bit(R5_LOCKED, &dev->flags) &&
2337                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2338                             test_bit(R5_Wantcompute, &dev->flags)) &&
2339                             test_bit(R5_Insync, &dev->flags)) {
2340                                 if (
2341                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2342                                         pr_debug("Read_old block "
2343                                                 "%d for Reconstruct\n", i);
2344                                         set_bit(R5_LOCKED, &dev->flags);
2345                                         set_bit(R5_Wantread, &dev->flags);
2346                                         s->locked++;
2347                                 } else {
2348                                         set_bit(STRIPE_DELAYED, &sh->state);
2349                                         set_bit(STRIPE_HANDLE, &sh->state);
2350                                 }
2351                         }
2352                 }
2353         /* now if nothing is locked, and if we have enough data,
2354          * we can start a write request
2355          */
2356         /* since handle_stripe can be called at any time we need to handle the
2357          * case where a compute block operation has been submitted and then a
2358          * subsequent call wants to start a write request.  raid5_run_ops only
2359          * handles the case where compute block and postxor are requested
2360          * simultaneously.  If this is not the case then new writes need to be
2361          * held off until the compute completes.
2362          */
2363         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2364             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2365             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2366                 schedule_reconstruction5(sh, s, rcw == 0, 0);
2367 }
2368
2369 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2370                 struct stripe_head *sh, struct stripe_head_state *s,
2371                 struct r6_state *r6s, int disks)
2372 {
2373         int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2374         int qd_idx = r6s->qd_idx;
2375         for (i = disks; i--; ) {
2376                 struct r5dev *dev = &sh->dev[i];
2377                 /* Would I have to read this buffer for reconstruct_write */
2378                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2379                     && i != pd_idx && i != qd_idx
2380                     && (!test_bit(R5_LOCKED, &dev->flags)
2381                             ) &&
2382                     !test_bit(R5_UPTODATE, &dev->flags)) {
2383                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2384                         else {
2385                                 pr_debug("raid6: must_compute: "
2386                                         "disk %d flags=%#lx\n", i, dev->flags);
2387                                 must_compute++;
2388                         }
2389                 }
2390         }
2391         pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2392                (unsigned long long)sh->sector, rcw, must_compute);
2393         set_bit(STRIPE_HANDLE, &sh->state);
2394
2395         if (rcw > 0)
2396                 /* want reconstruct write, but need to get some data */
2397                 for (i = disks; i--; ) {
2398                         struct r5dev *dev = &sh->dev[i];
2399                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2400                             && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2401                             && !test_bit(R5_LOCKED, &dev->flags) &&
2402                             !test_bit(R5_UPTODATE, &dev->flags) &&
2403                             test_bit(R5_Insync, &dev->flags)) {
2404                                 if (
2405                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2406                                         pr_debug("Read_old stripe %llu "
2407                                                 "block %d for Reconstruct\n",
2408                                              (unsigned long long)sh->sector, i);
2409                                         set_bit(R5_LOCKED, &dev->flags);
2410                                         set_bit(R5_Wantread, &dev->flags);
2411                                         s->locked++;
2412                                 } else {
2413                                         pr_debug("Request delayed stripe %llu "
2414                                                 "block %d for Reconstruct\n",
2415                                              (unsigned long long)sh->sector, i);
2416                                         set_bit(STRIPE_DELAYED, &sh->state);
2417                                         set_bit(STRIPE_HANDLE, &sh->state);
2418                                 }
2419                         }
2420                 }
2421         /* now if nothing is locked, and if we have enough data, we can start a
2422          * write request
2423          */
2424         if (s->locked == 0 && rcw == 0 &&
2425             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2426                 if (must_compute > 0) {
2427                         /* We have failed blocks and need to compute them */
2428                         switch (s->failed) {
2429                         case 0:
2430                                 BUG();
2431                         case 1:
2432                                 compute_block_1(sh, r6s->failed_num[0], 0);
2433                                 break;
2434                         case 2:
2435                                 compute_block_2(sh, r6s->failed_num[0],
2436                                                 r6s->failed_num[1]);
2437                                 break;
2438                         default: /* This request should have been failed? */
2439                                 BUG();
2440                         }
2441                 }
2442
2443                 pr_debug("Computing parity for stripe %llu\n",
2444                         (unsigned long long)sh->sector);
2445                 compute_parity6(sh, RECONSTRUCT_WRITE);
2446                 /* now every locked buffer is ready to be written */
2447                 for (i = disks; i--; )
2448                         if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2449                                 pr_debug("Writing stripe %llu block %d\n",
2450                                        (unsigned long long)sh->sector, i);
2451                                 s->locked++;
2452                                 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2453                         }
2454                 if (s->locked == disks)
2455                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2456                                 atomic_inc(&conf->pending_full_writes);
2457                 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2458                 set_bit(STRIPE_INSYNC, &sh->state);
2459
2460                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2461                         atomic_dec(&conf->preread_active_stripes);
2462                         if (atomic_read(&conf->preread_active_stripes) <
2463                             IO_THRESHOLD)
2464                                 md_wakeup_thread(conf->mddev->thread);
2465                 }
2466         }
2467 }
2468
2469 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2470                                 struct stripe_head_state *s, int disks)
2471 {
2472         struct r5dev *dev = NULL;
2473
2474         set_bit(STRIPE_HANDLE, &sh->state);
2475
2476         switch (sh->check_state) {
2477         case check_state_idle:
2478                 /* start a new check operation if there are no failures */
2479                 if (s->failed == 0) {
2480                         BUG_ON(s->uptodate != disks);
2481                         sh->check_state = check_state_run;
2482                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2483                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2484                         s->uptodate--;
2485                         break;
2486                 }
2487                 dev = &sh->dev[s->failed_num];
2488                 /* fall through */
2489         case check_state_compute_result:
2490                 sh->check_state = check_state_idle;
2491                 if (!dev)
2492                         dev = &sh->dev[sh->pd_idx];
2493
2494                 /* check that a write has not made the stripe insync */
2495                 if (test_bit(STRIPE_INSYNC, &sh->state))
2496                         break;
2497
2498                 /* either failed parity check, or recovery is happening */
2499                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2500                 BUG_ON(s->uptodate != disks);
2501
2502                 set_bit(R5_LOCKED, &dev->flags);
2503                 s->locked++;
2504                 set_bit(R5_Wantwrite, &dev->flags);
2505
2506                 clear_bit(STRIPE_DEGRADED, &sh->state);
2507                 set_bit(STRIPE_INSYNC, &sh->state);
2508                 break;
2509         case check_state_run:
2510                 break; /* we will be called again upon completion */
2511         case check_state_check_result:
2512                 sh->check_state = check_state_idle;
2513
2514                 /* if a failure occurred during the check operation, leave
2515                  * STRIPE_INSYNC not set and let the stripe be handled again
2516                  */
2517                 if (s->failed)
2518                         break;
2519
2520                 /* handle a successful check operation, if parity is correct
2521                  * we are done.  Otherwise update the mismatch count and repair
2522                  * parity if !MD_RECOVERY_CHECK
2523                  */
2524                 if (sh->ops.zero_sum_result == 0)
2525                         /* parity is correct (on disc,
2526                          * not in buffer any more)
2527                          */
2528                         set_bit(STRIPE_INSYNC, &sh->state);
2529                 else {
2530                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2531                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2532                                 /* don't try to repair!! */
2533                                 set_bit(STRIPE_INSYNC, &sh->state);
2534                         else {
2535                                 sh->check_state = check_state_compute_run;
2536                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2537                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2538                                 set_bit(R5_Wantcompute,
2539                                         &sh->dev[sh->pd_idx].flags);
2540                                 sh->ops.target = sh->pd_idx;
2541                                 s->uptodate++;
2542                         }
2543                 }
2544                 break;
2545         case check_state_compute_run:
2546                 break;
2547         default:
2548                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2549                        __func__, sh->check_state,
2550                        (unsigned long long) sh->sector);
2551                 BUG();
2552         }
2553 }
2554
2555
2556 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2557                                 struct stripe_head_state *s,
2558                                 struct r6_state *r6s, struct page *tmp_page,
2559                                 int disks)
2560 {
2561         int update_p = 0, update_q = 0;
2562         struct r5dev *dev;
2563         int pd_idx = sh->pd_idx;
2564         int qd_idx = r6s->qd_idx;
2565
2566         set_bit(STRIPE_HANDLE, &sh->state);
2567
2568         BUG_ON(s->failed > 2);
2569         BUG_ON(s->uptodate < disks);
2570         /* Want to check and possibly repair P and Q.
2571          * However there could be one 'failed' device, in which
2572          * case we can only check one of them, possibly using the
2573          * other to generate missing data
2574          */
2575
2576         /* If !tmp_page, we cannot do the calculations,
2577          * but as we have set STRIPE_HANDLE, we will soon be called
2578          * by stripe_handle with a tmp_page - just wait until then.
2579          */
2580         if (tmp_page) {
2581                 if (s->failed == r6s->q_failed) {
2582                         /* The only possible failed device holds 'Q', so it
2583                          * makes sense to check P (If anything else were failed,
2584                          * we would have used P to recreate it).
2585                          */
2586                         compute_block_1(sh, pd_idx, 1);
2587                         if (!page_is_zero(sh->dev[pd_idx].page)) {
2588                                 compute_block_1(sh, pd_idx, 0);
2589                                 update_p = 1;
2590                         }
2591                 }
2592                 if (!r6s->q_failed && s->failed < 2) {
2593                         /* q is not failed, and we didn't use it to generate
2594                          * anything, so it makes sense to check it
2595                          */
2596                         memcpy(page_address(tmp_page),
2597                                page_address(sh->dev[qd_idx].page),
2598                                STRIPE_SIZE);
2599                         compute_parity6(sh, UPDATE_PARITY);
2600                         if (memcmp(page_address(tmp_page),
2601                                    page_address(sh->dev[qd_idx].page),
2602                                    STRIPE_SIZE) != 0) {
2603                                 clear_bit(STRIPE_INSYNC, &sh->state);
2604                                 update_q = 1;
2605                         }
2606                 }
2607                 if (update_p || update_q) {
2608                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2609                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2610                                 /* don't try to repair!! */
2611                                 update_p = update_q = 0;
2612                 }
2613
2614                 /* now write out any block on a failed drive,
2615                  * or P or Q if they need it
2616                  */
2617
2618                 if (s->failed == 2) {
2619                         dev = &sh->dev[r6s->failed_num[1]];
2620                         s->locked++;
2621                         set_bit(R5_LOCKED, &dev->flags);
2622                         set_bit(R5_Wantwrite, &dev->flags);
2623                 }
2624                 if (s->failed >= 1) {
2625                         dev = &sh->dev[r6s->failed_num[0]];
2626                         s->locked++;
2627                         set_bit(R5_LOCKED, &dev->flags);
2628                         set_bit(R5_Wantwrite, &dev->flags);
2629                 }
2630
2631                 if (update_p) {
2632                         dev = &sh->dev[pd_idx];
2633                         s->locked++;
2634                         set_bit(R5_LOCKED, &dev->flags);
2635                         set_bit(R5_Wantwrite, &dev->flags);
2636                 }
2637                 if (update_q) {
2638                         dev = &sh->dev[qd_idx];
2639                         s->locked++;
2640                         set_bit(R5_LOCKED, &dev->flags);
2641                         set_bit(R5_Wantwrite, &dev->flags);
2642                 }
2643                 clear_bit(STRIPE_DEGRADED, &sh->state);
2644
2645                 set_bit(STRIPE_INSYNC, &sh->state);
2646         }
2647 }
2648
2649 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2650                                 struct r6_state *r6s)
2651 {
2652         int i;
2653
2654         /* We have read all the blocks in this stripe and now we need to
2655          * copy some of them into a target stripe for expand.
2656          */
2657         struct dma_async_tx_descriptor *tx = NULL;
2658         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2659         for (i = 0; i < sh->disks; i++)
2660                 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2661                         int dd_idx, j;
2662                         struct stripe_head *sh2;
2663
2664                         sector_t bn = compute_blocknr(sh, i);
2665                         sector_t s = raid5_compute_sector(conf, bn, 0,
2666                                                           &dd_idx, NULL);
2667                         sh2 = get_active_stripe(conf, s, 0, 1);
2668                         if (sh2 == NULL)
2669                                 /* so far only the early blocks of this stripe
2670                                  * have been requested.  When later blocks
2671                                  * get requested, we will try again
2672                                  */
2673                                 continue;
2674                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2675                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2676                                 /* must have already done this block */
2677                                 release_stripe(sh2);
2678                                 continue;
2679                         }
2680
2681                         /* place all the copies on one channel */
2682                         tx = async_memcpy(sh2->dev[dd_idx].page,
2683                                 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2684                                 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2685
2686                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2687                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2688                         for (j = 0; j < conf->raid_disks; j++)
2689                                 if (j != sh2->pd_idx &&
2690                                     (!r6s || j != sh2->qd_idx) &&
2691                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2692                                         break;
2693                         if (j == conf->raid_disks) {
2694                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2695                                 set_bit(STRIPE_HANDLE, &sh2->state);
2696                         }
2697                         release_stripe(sh2);
2698
2699                 }
2700         /* done submitting copies, wait for them to complete */
2701         if (tx) {
2702                 async_tx_ack(tx);
2703                 dma_wait_for_async_tx(tx);
2704         }
2705 }
2706
2707
2708 /*
2709  * handle_stripe - do things to a stripe.
2710  *
2711  * We lock the stripe and then examine the state of various bits
2712  * to see what needs to be done.
2713  * Possible results:
2714  *    return some read request which now have data
2715  *    return some write requests which are safely on disc
2716  *    schedule a read on some buffers
2717  *    schedule a write of some buffers
2718  *    return confirmation of parity correctness
2719  *
2720  * buffers are taken off read_list or write_list, and bh_cache buffers
2721  * get BH_Lock set before the stripe lock is released.
2722  *
2723  */
2724
2725 static bool handle_stripe5(struct stripe_head *sh)
2726 {
2727         raid5_conf_t *conf = sh->raid_conf;
2728         int disks = sh->disks, i;
2729         struct bio *return_bi = NULL;
2730         struct stripe_head_state s;
2731         struct r5dev *dev;
2732         mdk_rdev_t *blocked_rdev = NULL;
2733         int prexor;
2734
2735         memset(&s, 0, sizeof(s));
2736         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2737                  "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2738                  atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2739                  sh->reconstruct_state);
2740
2741         spin_lock(&sh->lock);
2742         clear_bit(STRIPE_HANDLE, &sh->state);
2743         clear_bit(STRIPE_DELAYED, &sh->state);
2744
2745         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2746         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2747         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2748
2749         /* Now to look around and see what can be done */
2750         rcu_read_lock();
2751         for (i=disks; i--; ) {
2752                 mdk_rdev_t *rdev;
2753                 struct r5dev *dev = &sh->dev[i];
2754                 clear_bit(R5_Insync, &dev->flags);
2755
2756                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2757                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2758                         dev->towrite, dev->written);
2759
2760                 /* maybe we can request a biofill operation
2761                  *
2762                  * new wantfill requests are only permitted while
2763                  * ops_complete_biofill is guaranteed to be inactive
2764                  */
2765                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2766                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2767                         set_bit(R5_Wantfill, &dev->flags);
2768
2769                 /* now count some things */
2770                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2771                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2772                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2773
2774                 if (test_bit(R5_Wantfill, &dev->flags))
2775                         s.to_fill++;
2776                 else if (dev->toread)
2777                         s.to_read++;
2778                 if (dev->towrite) {
2779                         s.to_write++;
2780                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2781                                 s.non_overwrite++;
2782                 }
2783                 if (dev->written)
2784                         s.written++;
2785                 rdev = rcu_dereference(conf->disks[i].rdev);
2786                 if (blocked_rdev == NULL &&
2787                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2788                         blocked_rdev = rdev;
2789                         atomic_inc(&rdev->nr_pending);
2790                 }
2791                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2792                         /* The ReadError flag will just be confusing now */
2793                         clear_bit(R5_ReadError, &dev->flags);
2794                         clear_bit(R5_ReWrite, &dev->flags);
2795                 }
2796                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2797                     || test_bit(R5_ReadError, &dev->flags)) {
2798                         s.failed++;
2799                         s.failed_num = i;
2800                 } else
2801                         set_bit(R5_Insync, &dev->flags);
2802         }
2803         rcu_read_unlock();
2804
2805         if (unlikely(blocked_rdev)) {
2806                 if (s.syncing || s.expanding || s.expanded ||
2807                     s.to_write || s.written) {
2808                         set_bit(STRIPE_HANDLE, &sh->state);
2809                         goto unlock;
2810                 }
2811                 /* There is nothing for the blocked_rdev to block */
2812                 rdev_dec_pending(blocked_rdev, conf->mddev);
2813                 blocked_rdev = NULL;
2814         }
2815
2816         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
2817                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
2818                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
2819         }
2820
2821         pr_debug("locked=%d uptodate=%d to_read=%d"
2822                 " to_write=%d failed=%d failed_num=%d\n",
2823                 s.locked, s.uptodate, s.to_read, s.to_write,
2824                 s.failed, s.failed_num);
2825         /* check if the array has lost two devices and, if so, some requests might
2826          * need to be failed
2827          */
2828         if (s.failed > 1 && s.to_read+s.to_write+s.written)
2829                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
2830         if (s.failed > 1 && s.syncing) {
2831                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2832                 clear_bit(STRIPE_SYNCING, &sh->state);
2833                 s.syncing = 0;
2834         }
2835
2836         /* might be able to return some write requests if the parity block
2837          * is safe, or on a failed drive
2838          */
2839         dev = &sh->dev[sh->pd_idx];
2840         if ( s.written &&
2841              ((test_bit(R5_Insync, &dev->flags) &&
2842                !test_bit(R5_LOCKED, &dev->flags) &&
2843                test_bit(R5_UPTODATE, &dev->flags)) ||
2844                (s.failed == 1 && s.failed_num == sh->pd_idx)))
2845                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
2846
2847         /* Now we might consider reading some blocks, either to check/generate
2848          * parity, or to satisfy requests
2849          * or to load a block that is being partially written.
2850          */
2851         if (s.to_read || s.non_overwrite ||
2852             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
2853                 handle_stripe_fill5(sh, &s, disks);
2854
2855         /* Now we check to see if any write operations have recently
2856          * completed
2857          */
2858         prexor = 0;
2859         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
2860                 prexor = 1;
2861         if (sh->reconstruct_state == reconstruct_state_drain_result ||
2862             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
2863                 sh->reconstruct_state = reconstruct_state_idle;
2864
2865                 /* All the 'written' buffers and the parity block are ready to
2866                  * be written back to disk
2867                  */
2868                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2869                 for (i = disks; i--; ) {
2870                         dev = &sh->dev[i];
2871                         if (test_bit(R5_LOCKED, &dev->flags) &&
2872                                 (i == sh->pd_idx || dev->written)) {
2873                                 pr_debug("Writing block %d\n", i);
2874                                 set_bit(R5_Wantwrite, &dev->flags);
2875                                 if (prexor)
2876                                         continue;
2877                                 if (!test_bit(R5_Insync, &dev->flags) ||
2878                                     (i == sh->pd_idx && s.failed == 0))
2879                                         set_bit(STRIPE_INSYNC, &sh->state);
2880                         }
2881                 }
2882                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2883                         atomic_dec(&conf->preread_active_stripes);
2884                         if (atomic_read(&conf->preread_active_stripes) <
2885                                 IO_THRESHOLD)
2886                                 md_wakeup_thread(conf->mddev->thread);
2887                 }
2888         }
2889
2890         /* Now to consider new write requests and what else, if anything
2891          * should be read.  We do not handle new writes when:
2892          * 1/ A 'write' operation (copy+xor) is already in flight.
2893          * 2/ A 'check' operation is in flight, as it may clobber the parity
2894          *    block.
2895          */
2896         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
2897                 handle_stripe_dirtying5(conf, sh, &s, disks);
2898
2899         /* maybe we need to check and possibly fix the parity for this stripe
2900          * Any reads will already have been scheduled, so we just see if enough
2901          * data is available.  The parity check is held off while parity
2902          * dependent operations are in flight.
2903          */
2904         if (sh->check_state ||
2905             (s.syncing && s.locked == 0 &&
2906              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
2907              !test_bit(STRIPE_INSYNC, &sh->state)))
2908                 handle_parity_checks5(conf, sh, &s, disks);
2909
2910         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2911                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2912                 clear_bit(STRIPE_SYNCING, &sh->state);
2913         }
2914
2915         /* If the failed drive is just a ReadError, then we might need to progress
2916          * the repair/check process
2917          */
2918         if (s.failed == 1 && !conf->mddev->ro &&
2919             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2920             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2921             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2922                 ) {
2923                 dev = &sh->dev[s.failed_num];
2924                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2925                         set_bit(R5_Wantwrite, &dev->flags);
2926                         set_bit(R5_ReWrite, &dev->flags);
2927                         set_bit(R5_LOCKED, &dev->flags);
2928                         s.locked++;
2929                 } else {
2930                         /* let's read it back */
2931                         set_bit(R5_Wantread, &dev->flags);
2932                         set_bit(R5_LOCKED, &dev->flags);
2933                         s.locked++;
2934                 }
2935         }
2936
2937         /* Finish reconstruct operations initiated by the expansion process */
2938         if (sh->reconstruct_state == reconstruct_state_result) {
2939                 sh->reconstruct_state = reconstruct_state_idle;
2940                 clear_bit(STRIPE_EXPANDING, &sh->state);
2941                 for (i = conf->raid_disks; i--; ) {
2942                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2943                         set_bit(R5_LOCKED, &sh->dev[i].flags);
2944                         s.locked++;
2945                 }
2946         }
2947
2948         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2949             !sh->reconstruct_state) {
2950                 /* Need to write out all blocks after computing parity */
2951                 sh->disks = conf->raid_disks;
2952                 stripe_set_idx(sh->sector, conf, 0, sh);
2953                 schedule_reconstruction5(sh, &s, 1, 1);
2954         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
2955                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2956                 atomic_dec(&conf->reshape_stripes);
2957                 wake_up(&conf->wait_for_overlap);
2958                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2959         }
2960
2961         if (s.expanding && s.locked == 0 &&
2962             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
2963                 handle_stripe_expansion(conf, sh, NULL);
2964
2965  unlock:
2966         spin_unlock(&sh->lock);
2967
2968         /* wait for this device to become unblocked */
2969         if (unlikely(blocked_rdev))
2970                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
2971
2972         if (s.ops_request)
2973                 raid5_run_ops(sh, s.ops_request);
2974
2975         ops_run_io(sh, &s);
2976
2977         return_io(return_bi);
2978
2979         return blocked_rdev == NULL;
2980 }
2981
2982 static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2983 {
2984         raid5_conf_t *conf = sh->raid_conf;
2985         int disks = sh->disks;
2986         struct bio *return_bi = NULL;
2987         int i, pd_idx = sh->pd_idx;
2988         struct stripe_head_state s;
2989         struct r6_state r6s;
2990         struct r5dev *dev, *pdev, *qdev;
2991         mdk_rdev_t *blocked_rdev = NULL;
2992
2993         r6s.qd_idx = sh->qd_idx;
2994         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2995                 "pd_idx=%d, qd_idx=%d\n",
2996                (unsigned long long)sh->sector, sh->state,
2997                atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2998         memset(&s, 0, sizeof(s));
2999
3000         spin_lock(&sh->lock);
3001         clear_bit(STRIPE_HANDLE, &sh->state);
3002         clear_bit(STRIPE_DELAYED, &sh->state);
3003
3004         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3005         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3006         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3007         /* Now to look around and see what can be done */
3008
3009         rcu_read_lock();
3010         for (i=disks; i--; ) {
3011                 mdk_rdev_t *rdev;
3012                 dev = &sh->dev[i];
3013                 clear_bit(R5_Insync, &dev->flags);
3014
3015                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3016                         i, dev->flags, dev->toread, dev->towrite, dev->written);
3017                 /* maybe we can reply to a read */
3018                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
3019                         struct bio *rbi, *rbi2;
3020                         pr_debug("Return read for disc %d\n", i);
3021                         spin_lock_irq(&conf->device_lock);
3022                         rbi = dev->toread;
3023                         dev->toread = NULL;
3024                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
3025                                 wake_up(&conf->wait_for_overlap);
3026                         spin_unlock_irq(&conf->device_lock);
3027                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
3028                                 copy_data(0, rbi, dev->page, dev->sector);
3029                                 rbi2 = r5_next_bio(rbi, dev->sector);
3030                                 spin_lock_irq(&conf->device_lock);
3031                                 if (!raid5_dec_bi_phys_segments(rbi)) {
3032                                         rbi->bi_next = return_bi;
3033                                         return_bi = rbi;
3034                                 }
3035                                 spin_unlock_irq(&conf->device_lock);
3036                                 rbi = rbi2;
3037                         }
3038                 }
3039
3040                 /* now count some things */
3041                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3042                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3043
3044
3045                 if (dev->toread)
3046                         s.to_read++;
3047                 if (dev->towrite) {
3048                         s.to_write++;
3049                         if (!test_bit(R5_OVERWRITE, &dev->flags))
3050                                 s.non_overwrite++;
3051                 }
3052                 if (dev->written)
3053                         s.written++;
3054                 rdev = rcu_dereference(conf->disks[i].rdev);
3055                 if (blocked_rdev == NULL &&
3056                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3057                         blocked_rdev = rdev;
3058                         atomic_inc(&rdev->nr_pending);
3059                 }
3060                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3061                         /* The ReadError flag will just be confusing now */
3062                         clear_bit(R5_ReadError, &dev->flags);
3063                         clear_bit(R5_ReWrite, &dev->flags);
3064                 }
3065                 if (!rdev || !test_bit(In_sync, &rdev->flags)
3066                     || test_bit(R5_ReadError, &dev->flags)) {
3067                         if (s.failed < 2)
3068                                 r6s.failed_num[s.failed] = i;
3069                         s.failed++;
3070                 } else
3071                         set_bit(R5_Insync, &dev->flags);
3072         }
3073         rcu_read_unlock();
3074
3075         if (unlikely(blocked_rdev)) {
3076                 if (s.syncing || s.expanding || s.expanded ||
3077                     s.to_write || s.written) {
3078                         set_bit(STRIPE_HANDLE, &sh->state);
3079                         goto unlock;
3080                 }
3081                 /* There is nothing for the blocked_rdev to block */
3082                 rdev_dec_pending(blocked_rdev, conf->mddev);
3083                 blocked_rdev = NULL;
3084         }
3085
3086         pr_debug("locked=%d uptodate=%d to_read=%d"
3087                " to_write=%d failed=%d failed_num=%d,%d\n",
3088                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3089                r6s.failed_num[0], r6s.failed_num[1]);
3090         /* check if the array has lost >2 devices and, if so, some requests
3091          * might need to be failed
3092          */
3093         if (s.failed > 2 && s.to_read+s.to_write+s.written)
3094                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3095         if (s.failed > 2 && s.syncing) {
3096                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3097                 clear_bit(STRIPE_SYNCING, &sh->state);
3098                 s.syncing = 0;
3099         }
3100
3101         /*
3102          * might be able to return some write requests if the parity blocks
3103          * are safe, or on a failed drive
3104          */
3105         pdev = &sh->dev[pd_idx];
3106         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3107                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3108         qdev = &sh->dev[r6s.qd_idx];
3109         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
3110                 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
3111
3112         if ( s.written &&
3113              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3114                              && !test_bit(R5_LOCKED, &pdev->flags)
3115                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3116              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3117                              && !test_bit(R5_LOCKED, &qdev->flags)
3118                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3119                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3120
3121         /* Now we might consider reading some blocks, either to check/generate
3122          * parity, or to satisfy requests
3123          * or to load a block that is being partially written.
3124          */
3125         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3126             (s.syncing && (s.uptodate < disks)) || s.expanding)
3127                 handle_stripe_fill6(sh, &s, &r6s, disks);
3128
3129         /* now to consider writing and what else, if anything should be read */
3130         if (s.to_write)
3131                 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3132
3133         /* maybe we need to check and possibly fix the parity for this stripe
3134          * Any reads will already have been scheduled, so we just see if enough
3135          * data is available
3136          */
3137         if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3138                 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3139
3140         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3141                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3142                 clear_bit(STRIPE_SYNCING, &sh->state);
3143         }
3144
3145         /* If the failed drives are just a ReadError, then we might need
3146          * to progress the repair/check process
3147          */
3148         if (s.failed <= 2 && !conf->mddev->ro)
3149                 for (i = 0; i < s.failed; i++) {
3150                         dev = &sh->dev[r6s.failed_num[i]];
3151                         if (test_bit(R5_ReadError, &dev->flags)
3152                             && !test_bit(R5_LOCKED, &dev->flags)
3153                             && test_bit(R5_UPTODATE, &dev->flags)
3154                                 ) {
3155                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3156                                         set_bit(R5_Wantwrite, &dev->flags);
3157                                         set_bit(R5_ReWrite, &dev->flags);
3158                                         set_bit(R5_LOCKED, &dev->flags);
3159                                 } else {
3160                                         /* let's read it back */
3161                                         set_bit(R5_Wantread, &dev->flags);
3162                                         set_bit(R5_LOCKED, &dev->flags);
3163                                 }
3164                         }
3165                 }
3166
3167         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3168                 /* Need to write out all blocks after computing P&Q */
3169                 sh->disks = conf->raid_disks;
3170                 stripe_set_idx(sh->sector, conf, 0, sh);
3171                 compute_parity6(sh, RECONSTRUCT_WRITE);
3172                 for (i = conf->raid_disks ; i-- ;  ) {
3173                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3174                         s.locked++;
3175                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3176                 }
3177                 clear_bit(STRIPE_EXPANDING, &sh->state);
3178         } else if (s.expanded) {
3179                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3180                 atomic_dec(&conf->reshape_stripes);
3181                 wake_up(&conf->wait_for_overlap);
3182                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3183         }
3184
3185         if (s.expanding && s.locked == 0 &&
3186             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3187                 handle_stripe_expansion(conf, sh, &r6s);
3188
3189  unlock:
3190         spin_unlock(&sh->lock);
3191
3192         /* wait for this device to become unblocked */
3193         if (unlikely(blocked_rdev))
3194                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3195
3196         ops_run_io(sh, &s);
3197
3198         return_io(return_bi);
3199
3200         return blocked_rdev == NULL;
3201 }
3202
3203 /* returns true if the stripe was handled */
3204 static bool handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3205 {
3206         if (sh->raid_conf->level == 6)
3207                 return handle_stripe6(sh, tmp_page);
3208         else
3209                 return handle_stripe5(sh);
3210 }
3211
3212
3213
3214 static void raid5_activate_delayed(raid5_conf_t *conf)
3215 {
3216         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3217                 while (!list_empty(&conf->delayed_list)) {
3218                         struct list_head *l = conf->delayed_list.next;
3219                         struct stripe_head *sh;
3220                         sh = list_entry(l, struct stripe_head, lru);
3221                         list_del_init(l);
3222                         clear_bit(STRIPE_DELAYED, &sh->state);
3223                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3224                                 atomic_inc(&conf->preread_active_stripes);
3225                         list_add_tail(&sh->lru, &conf->hold_list);
3226                 }
3227         } else
3228                 blk_plug_device(conf->mddev->queue);
3229 }
3230
3231 static void activate_bit_delay(raid5_conf_t *conf)
3232 {
3233         /* device_lock is held */
3234         struct list_head head;
3235         list_add(&head, &conf->bitmap_list);
3236         list_del_init(&conf->bitmap_list);
3237         while (!list_empty(&head)) {
3238                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3239                 list_del_init(&sh->lru);
3240                 atomic_inc(&sh->count);
3241                 __release_stripe(conf, sh);
3242         }
3243 }
3244
3245 static void unplug_slaves(mddev_t *mddev)
3246 {
3247         raid5_conf_t *conf = mddev_to_conf(mddev);
3248         int i;
3249
3250         rcu_read_lock();
3251         for (i=0; i<mddev->raid_disks; i++) {
3252                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3253                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3254                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3255
3256                         atomic_inc(&rdev->nr_pending);
3257                         rcu_read_unlock();
3258
3259                         blk_unplug(r_queue);
3260
3261                         rdev_dec_pending(rdev, mddev);
3262                         rcu_read_lock();
3263                 }
3264         }
3265         rcu_read_unlock();
3266 }
3267
3268 static void raid5_unplug_device(struct request_queue *q)
3269 {
3270         mddev_t *mddev = q->queuedata;
3271         raid5_conf_t *conf = mddev_to_conf(mddev);
3272         unsigned long flags;
3273
3274         spin_lock_irqsave(&conf->device_lock, flags);
3275
3276         if (blk_remove_plug(q)) {
3277                 conf->seq_flush++;
3278                 raid5_activate_delayed(conf);
3279         }
3280         md_wakeup_thread(mddev->thread);
3281
3282         spin_unlock_irqrestore(&conf->device_lock, flags);
3283
3284         unplug_slaves(mddev);
3285 }
3286
3287 static int raid5_congested(void *data, int bits)
3288 {
3289         mddev_t *mddev = data;
3290         raid5_conf_t *conf = mddev_to_conf(mddev);
3291
3292         /* No difference between reads and writes.  Just check
3293          * how busy the stripe_cache is
3294          */
3295         if (conf->inactive_blocked)
3296                 return 1;
3297         if (conf->quiesce)
3298                 return 1;
3299         if (list_empty_careful(&conf->inactive_list))
3300                 return 1;
3301
3302         return 0;
3303 }
3304
3305 /* We want read requests to align with chunks where possible,
3306  * but write requests don't need to.
3307  */
3308 static int raid5_mergeable_bvec(struct request_queue *q,
3309                                 struct bvec_merge_data *bvm,
3310                                 struct bio_vec *biovec)
3311 {
3312         mddev_t *mddev = q->queuedata;
3313         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3314         int max;
3315         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3316         unsigned int bio_sectors = bvm->bi_size >> 9;
3317
3318         if ((bvm->bi_rw & 1) == WRITE)
3319                 return biovec->bv_len; /* always allow writes to be mergeable */
3320
3321         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3322         if (max < 0) max = 0;
3323         if (max <= biovec->bv_len && bio_sectors == 0)
3324                 return biovec->bv_len;
3325         else
3326                 return max;
3327 }
3328
3329
3330 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3331 {
3332         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3333         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3334         unsigned int bio_sectors = bio->bi_size >> 9;
3335
3336         return  chunk_sectors >=
3337                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3338 }
3339
3340 /*
3341  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3342  *  later sampled by raid5d.
3343  */
3344 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3345 {
3346         unsigned long flags;
3347
3348         spin_lock_irqsave(&conf->device_lock, flags);
3349
3350         bi->bi_next = conf->retry_read_aligned_list;
3351         conf->retry_read_aligned_list = bi;
3352
3353         spin_unlock_irqrestore(&conf->device_lock, flags);
3354         md_wakeup_thread(conf->mddev->thread);
3355 }
3356
3357
3358 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3359 {
3360         struct bio *bi;
3361
3362         bi = conf->retry_read_aligned;
3363         if (bi) {
3364                 conf->retry_read_aligned = NULL;
3365                 return bi;
3366         }
3367         bi = conf->retry_read_aligned_list;
3368         if(bi) {
3369                 conf->retry_read_aligned_list = bi->bi_next;
3370                 bi->bi_next = NULL;
3371                 /*
3372                  * this sets the active strip count to 1 and the processed
3373                  * strip count to zero (upper 8 bits)
3374                  */
3375                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3376         }
3377
3378         return bi;
3379 }
3380
3381
3382 /*
3383  *  The "raid5_align_endio" should check if the read succeeded and if it
3384  *  did, call bio_endio on the original bio (having bio_put the new bio
3385  *  first).
3386  *  If the read failed..
3387  */
3388 static void raid5_align_endio(struct bio *bi, int error)
3389 {
3390         struct bio* raid_bi  = bi->bi_private;
3391         mddev_t *mddev;
3392         raid5_conf_t *conf;
3393         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3394         mdk_rdev_t *rdev;
3395
3396         bio_put(bi);
3397
3398         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3399         conf = mddev_to_conf(mddev);
3400         rdev = (void*)raid_bi->bi_next;
3401         raid_bi->bi_next = NULL;
3402
3403         rdev_dec_pending(rdev, conf->mddev);
3404
3405         if (!error && uptodate) {
3406                 bio_endio(raid_bi, 0);
3407                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3408                         wake_up(&conf->wait_for_stripe);
3409                 return;
3410         }
3411
3412
3413         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3414
3415         add_bio_to_retry(raid_bi, conf);
3416 }
3417
3418 static int bio_fits_rdev(struct bio *bi)
3419 {
3420         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3421
3422         if ((bi->bi_size>>9) > q->max_sectors)
3423                 return 0;
3424         blk_recount_segments(q, bi);
3425         if (bi->bi_phys_segments > q->max_phys_segments)
3426                 return 0;
3427
3428         if (q->merge_bvec_fn)
3429                 /* it's too hard to apply the merge_bvec_fn at this stage,
3430                  * just just give up
3431                  */
3432                 return 0;
3433
3434         return 1;
3435 }
3436
3437
3438 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3439 {
3440         mddev_t *mddev = q->queuedata;
3441         raid5_conf_t *conf = mddev_to_conf(mddev);
3442         unsigned int dd_idx;
3443         struct bio* align_bi;
3444         mdk_rdev_t *rdev;
3445
3446         if (!in_chunk_boundary(mddev, raid_bio)) {
3447                 pr_debug("chunk_aligned_read : non aligned\n");
3448                 return 0;
3449         }
3450         /*
3451          * use bio_clone to make a copy of the bio
3452          */
3453         align_bi = bio_clone(raid_bio, GFP_NOIO);
3454         if (!align_bi)
3455                 return 0;
3456         /*
3457          *   set bi_end_io to a new function, and set bi_private to the
3458          *     original bio.
3459          */
3460         align_bi->bi_end_io  = raid5_align_endio;
3461         align_bi->bi_private = raid_bio;
3462         /*
3463          *      compute position
3464          */
3465         align_bi->bi_sector =  raid5_compute_sector(conf, raid_bio->bi_sector,
3466                                                     0,
3467                                                     &dd_idx, NULL);
3468
3469         rcu_read_lock();
3470         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3471         if (rdev && test_bit(In_sync, &rdev->flags)) {
3472                 atomic_inc(&rdev->nr_pending);
3473                 rcu_read_unlock();
3474                 raid_bio->bi_next = (void*)rdev;
3475                 align_bi->bi_bdev =  rdev->bdev;
3476                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3477                 align_bi->bi_sector += rdev->data_offset;
3478
3479                 if (!bio_fits_rdev(align_bi)) {
3480                         /* too big in some way */
3481                         bio_put(align_bi);
3482                         rdev_dec_pending(rdev, mddev);
3483                         return 0;
3484                 }
3485
3486                 spin_lock_irq(&conf->device_lock);
3487                 wait_event_lock_irq(conf->wait_for_stripe,
3488                                     conf->quiesce == 0,
3489                                     conf->device_lock, /* nothing */);
3490                 atomic_inc(&conf->active_aligned_reads);
3491                 spin_unlock_irq(&conf->device_lock);
3492
3493                 generic_make_request(align_bi);
3494                 return 1;
3495         } else {
3496                 rcu_read_unlock();
3497                 bio_put(align_bi);
3498                 return 0;
3499         }
3500 }
3501
3502 /* __get_priority_stripe - get the next stripe to process
3503  *
3504  * Full stripe writes are allowed to pass preread active stripes up until
3505  * the bypass_threshold is exceeded.  In general the bypass_count
3506  * increments when the handle_list is handled before the hold_list; however, it
3507  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3508  * stripe with in flight i/o.  The bypass_count will be reset when the
3509  * head of the hold_list has changed, i.e. the head was promoted to the
3510  * handle_list.
3511  */
3512 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3513 {
3514         struct stripe_head *sh;
3515
3516         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3517                   __func__,
3518                   list_empty(&conf->handle_list) ? "empty" : "busy",
3519                   list_empty(&conf->hold_list) ? "empty" : "busy",
3520                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
3521
3522         if (!list_empty(&conf->handle_list)) {
3523                 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3524
3525                 if (list_empty(&conf->hold_list))
3526                         conf->bypass_count = 0;
3527                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3528                         if (conf->hold_list.next == conf->last_hold)
3529                                 conf->bypass_count++;
3530                         else {
3531                                 conf->last_hold = conf->hold_list.next;
3532                                 conf->bypass_count -= conf->bypass_threshold;
3533                                 if (conf->bypass_count < 0)
3534                                         conf->bypass_count = 0;
3535                         }
3536                 }
3537         } else if (!list_empty(&conf->hold_list) &&
3538                    ((conf->bypass_threshold &&
3539                      conf->bypass_count > conf->bypass_threshold) ||
3540                     atomic_read(&conf->pending_full_writes) == 0)) {
3541                 sh = list_entry(conf->hold_list.next,
3542                                 typeof(*sh), lru);
3543                 conf->bypass_count -= conf->bypass_threshold;
3544                 if (conf->bypass_count < 0)
3545                         conf->bypass_count = 0;
3546         } else
3547                 return NULL;
3548
3549         list_del_init(&sh->lru);
3550         atomic_inc(&sh->count);
3551         BUG_ON(atomic_read(&sh->count) != 1);
3552         return sh;
3553 }
3554
3555 static int make_request(struct request_queue *q, struct bio * bi)
3556 {
3557         mddev_t *mddev = q->queuedata;
3558         raid5_conf_t *conf = mddev_to_conf(mddev);
3559         int dd_idx;
3560         sector_t new_sector;
3561         sector_t logical_sector, last_sector;
3562         struct stripe_head *sh;
3563         const int rw = bio_data_dir(bi);
3564         int cpu, remaining;
3565
3566         if (unlikely(bio_barrier(bi))) {
3567                 bio_endio(bi, -EOPNOTSUPP);
3568                 return 0;
3569         }
3570
3571         md_write_start(mddev, bi);
3572
3573         cpu = part_stat_lock();
3574         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
3575         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
3576                       bio_sectors(bi));
3577         part_stat_unlock();
3578
3579         if (rw == READ &&
3580              mddev->reshape_position == MaxSector &&
3581              chunk_aligned_read(q,bi))
3582                 return 0;
3583
3584         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3585         last_sector = bi->bi_sector + (bi->bi_size>>9);
3586         bi->bi_next = NULL;
3587         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3588
3589         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3590                 DEFINE_WAIT(w);
3591                 int disks, data_disks;
3592                 int previous;
3593
3594         retry:
3595                 previous = 0;
3596                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3597                 if (likely(conf->expand_progress == MaxSector))
3598                         disks = conf->raid_disks;
3599                 else {
3600                         /* spinlock is needed as expand_progress may be
3601                          * 64bit on a 32bit platform, and so it might be
3602                          * possible to see a half-updated value
3603                          * Ofcourse expand_progress could change after
3604                          * the lock is dropped, so once we get a reference
3605                          * to the stripe that we think it is, we will have
3606                          * to check again.
3607                          */
3608                         spin_lock_irq(&conf->device_lock);
3609                         disks = conf->raid_disks;
3610                         if (logical_sector >= conf->expand_progress) {
3611                                 disks = conf->previous_raid_disks;
3612                                 previous = 1;
3613                         } else {
3614                                 if (logical_sector >= conf->expand_lo) {
3615                                         spin_unlock_irq(&conf->device_lock);
3616                                         schedule();
3617                                         goto retry;
3618                                 }
3619                         }
3620                         spin_unlock_irq(&conf->device_lock);
3621                 }
3622                 data_disks = disks - conf->max_degraded;
3623
3624                 new_sector = raid5_compute_sector(conf, logical_sector,
3625                                                   previous,
3626                                                   &dd_idx, NULL);
3627                 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3628                         (unsigned long long)new_sector, 
3629                         (unsigned long long)logical_sector);
3630
3631                 sh = get_active_stripe(conf, new_sector, previous,
3632                                        (bi->bi_rw&RWA_MASK));
3633                 if (sh) {
3634                         if (unlikely(conf->expand_progress != MaxSector)) {
3635                                 /* expansion might have moved on while waiting for a
3636                                  * stripe, so we must do the range check again.
3637                                  * Expansion could still move past after this
3638                                  * test, but as we are holding a reference to
3639                                  * 'sh', we know that if that happens,
3640                                  *  STRIPE_EXPANDING will get set and the expansion
3641                                  * won't proceed until we finish with the stripe.
3642                                  */
3643                                 int must_retry = 0;
3644                                 spin_lock_irq(&conf->device_lock);
3645                                 if (logical_sector <  conf->expand_progress &&
3646                                     disks == conf->previous_raid_disks)
3647                                         /* mismatch, need to try again */
3648                                         must_retry = 1;
3649                                 spin_unlock_irq(&conf->device_lock);
3650                                 if (must_retry) {
3651                                         release_stripe(sh);
3652                                         goto retry;
3653                                 }
3654                         }
3655                         /* FIXME what if we get a false positive because these
3656                          * are being updated.
3657                          */
3658                         if (logical_sector >= mddev->suspend_lo &&
3659                             logical_sector < mddev->suspend_hi) {
3660                                 release_stripe(sh);
3661                                 schedule();
3662                                 goto retry;
3663                         }
3664
3665                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3666                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3667                                 /* Stripe is busy expanding or
3668                                  * add failed due to overlap.  Flush everything
3669                                  * and wait a while
3670                                  */
3671                                 raid5_unplug_device(mddev->queue);
3672                                 release_stripe(sh);
3673                                 schedule();
3674                                 goto retry;
3675                         }
3676                         finish_wait(&conf->wait_for_overlap, &w);
3677                         set_bit(STRIPE_HANDLE, &sh->state);
3678                         clear_bit(STRIPE_DELAYED, &sh->state);
3679                         release_stripe(sh);
3680                 } else {
3681                         /* cannot get stripe for read-ahead, just give-up */
3682                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3683                         finish_wait(&conf->wait_for_overlap, &w);
3684                         break;
3685                 }
3686                         
3687         }
3688         spin_lock_irq(&conf->device_lock);
3689         remaining = raid5_dec_bi_phys_segments(bi);
3690         spin_unlock_irq(&conf->device_lock);
3691         if (remaining == 0) {
3692
3693                 if ( rw == WRITE )
3694                         md_write_end(mddev);
3695
3696                 bio_endio(bi, 0);
3697         }
3698         return 0;
3699 }
3700
3701 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
3702
3703 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3704 {
3705         /* reshaping is quite different to recovery/resync so it is
3706          * handled quite separately ... here.
3707          *
3708          * On each call to sync_request, we gather one chunk worth of
3709          * destination stripes and flag them as expanding.
3710          * Then we find all the source stripes and request reads.
3711          * As the reads complete, handle_stripe will copy the data
3712          * into the destination stripe and release that stripe.
3713          */
3714         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3715         struct stripe_head *sh;
3716         sector_t first_sector, last_sector;
3717         int raid_disks = conf->previous_raid_disks;
3718         int data_disks = raid_disks - conf->max_degraded;
3719         int new_data_disks = conf->raid_disks - conf->max_degraded;
3720         int i;
3721         int dd_idx;
3722         sector_t writepos, safepos, gap;
3723
3724         if (sector_nr == 0 &&
3725             conf->expand_progress != 0) {
3726                 /* restarting in the middle, skip the initial sectors */
3727                 sector_nr = conf->expand_progress;
3728                 sector_div(sector_nr, new_data_disks);
3729                 *skipped = 1;
3730                 return sector_nr;
3731         }
3732
3733         /* we update the metadata when there is more than 3Meg
3734          * in the block range (that is rather arbitrary, should
3735          * probably be time based) or when the data about to be
3736          * copied would over-write the source of the data at
3737          * the front of the range.
3738          * i.e. one new_stripe forward from expand_progress new_maps
3739          * to after where expand_lo old_maps to
3740          */
3741         writepos = conf->expand_progress +
3742                 conf->chunk_size/512*(new_data_disks);
3743         sector_div(writepos, new_data_disks);
3744         safepos = conf->expand_lo;
3745         sector_div(safepos, data_disks);
3746         gap = conf->expand_progress - conf->expand_lo;
3747
3748         if (writepos >= safepos ||
3749             gap > (new_data_disks)*3000*2 /*3Meg*/) {
3750                 /* Cannot proceed until we've updated the superblock... */
3751                 wait_event(conf->wait_for_overlap,
3752                            atomic_read(&conf->reshape_stripes)==0);
3753                 mddev->reshape_position = conf->expand_progress;
3754                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3755                 md_wakeup_thread(mddev->thread);
3756                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3757                            kthread_should_stop());
3758                 spin_lock_irq(&conf->device_lock);
3759                 conf->expand_lo = mddev->reshape_position;
3760                 spin_unlock_irq(&conf->device_lock);
3761                 wake_up(&conf->wait_for_overlap);
3762         }
3763
3764         for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3765                 int j;
3766                 int skipped = 0;
3767                 sh = get_active_stripe(conf, sector_nr+i, 0, 0);
3768                 set_bit(STRIPE_EXPANDING, &sh->state);
3769                 atomic_inc(&conf->reshape_stripes);
3770                 /* If any of this stripe is beyond the end of the old
3771                  * array, then we need to zero those blocks
3772                  */
3773                 for (j=sh->disks; j--;) {
3774                         sector_t s;
3775                         if (j == sh->pd_idx)
3776                                 continue;
3777                         if (conf->level == 6 &&
3778                             j == sh->qd_idx)
3779                                 continue;
3780                         s = compute_blocknr(sh, j);
3781                         if (s < raid5_size(mddev, 0, 0)) {
3782                                 skipped = 1;
3783                                 continue;
3784                         }
3785                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3786                         set_bit(R5_Expanded, &sh->dev[j].flags);
3787                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
3788                 }
3789                 if (!skipped) {
3790                         set_bit(STRIPE_EXPAND_READY, &sh->state);
3791                         set_bit(STRIPE_HANDLE, &sh->state);
3792                 }
3793                 release_stripe(sh);
3794         }
3795         spin_lock_irq(&conf->device_lock);
3796         conf->expand_progress = (sector_nr + i) * new_data_disks;
3797         spin_unlock_irq(&conf->device_lock);
3798         /* Ok, those stripe are ready. We can start scheduling
3799          * reads on the source stripes.
3800          * The source stripes are determined by mapping the first and last
3801          * block on the destination stripes.
3802          */
3803         first_sector =
3804                 raid5_compute_sector(conf, sector_nr*(new_data_disks),
3805                                      1, &dd_idx, NULL);
3806         last_sector =
3807                 raid5_compute_sector(conf, ((sector_nr+conf->chunk_size/512)
3808                                             *(new_data_disks) - 1),
3809                                      1, &dd_idx, NULL);
3810         if (last_sector >= mddev->dev_sectors)
3811                 last_sector = mddev->dev_sectors - 1;
3812         while (first_sector <= last_sector) {
3813                 sh = get_active_stripe(conf, first_sector, 1, 0);
3814                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3815                 set_bit(STRIPE_HANDLE, &sh->state);
3816                 release_stripe(sh);
3817                 first_sector += STRIPE_SECTORS;
3818         }
3819         /* If this takes us to the resync_max point where we have to pause,
3820          * then we need to write out the superblock.
3821          */
3822         sector_nr += conf->chunk_size>>9;
3823         if (sector_nr >= mddev->resync_max) {
3824                 /* Cannot proceed until we've updated the superblock... */
3825                 wait_event(conf->wait_for_overlap,
3826                            atomic_read(&conf->reshape_stripes) == 0);
3827                 mddev->reshape_position = conf->expand_progress;
3828                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3829                 md_wakeup_thread(mddev->thread);
3830                 wait_event(mddev->sb_wait,
3831                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3832                            || kthread_should_stop());
3833                 spin_lock_irq(&conf->device_lock);
3834                 conf->expand_lo = mddev->reshape_position;
3835                 spin_unlock_irq(&conf->device_lock);
3836                 wake_up(&conf->wait_for_overlap);
3837         }
3838         return conf->chunk_size>>9;
3839 }
3840
3841 /* FIXME go_faster isn't used */
3842 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3843 {
3844         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3845         struct stripe_head *sh;
3846         sector_t max_sector = mddev->dev_sectors;
3847         int sync_blocks;
3848         int still_degraded = 0;
3849         int i;
3850
3851         if (sector_nr >= max_sector) {
3852                 /* just being told to finish up .. nothing much to do */
3853                 unplug_slaves(mddev);
3854                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3855                         end_reshape(conf);
3856                         return 0;
3857                 }
3858
3859                 if (mddev->curr_resync < max_sector) /* aborted */
3860                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3861                                         &sync_blocks, 1);
3862                 else /* completed sync */
3863                         conf->fullsync = 0;
3864                 bitmap_close_sync(mddev->bitmap);
3865
3866                 return 0;
3867         }
3868
3869         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3870                 return reshape_request(mddev, sector_nr, skipped);
3871
3872         /* No need to check resync_max as we never do more than one
3873          * stripe, and as resync_max will always be on a chunk boundary,
3874          * if the check in md_do_sync didn't fire, there is no chance
3875          * of overstepping resync_max here
3876          */
3877
3878         /* if there is too many failed drives and we are trying
3879          * to resync, then assert that we are finished, because there is
3880          * nothing we can do.
3881          */
3882         if (mddev->degraded >= conf->max_degraded &&
3883             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3884                 sector_t rv = mddev->dev_sectors - sector_nr;
3885                 *skipped = 1;
3886                 return rv;
3887         }
3888         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3889             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3890             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3891                 /* we can skip this block, and probably more */
3892                 sync_blocks /= STRIPE_SECTORS;
3893                 *skipped = 1;
3894                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3895         }
3896
3897
3898         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3899
3900         sh = get_active_stripe(conf, sector_nr, 0, 1);
3901         if (sh == NULL) {
3902                 sh = get_active_stripe(conf, sector_nr, 0, 0);
3903                 /* make sure we don't swamp the stripe cache if someone else
3904                  * is trying to get access
3905                  */
3906                 schedule_timeout_uninterruptible(1);
3907         }
3908         /* Need to check if array will still be degraded after recovery/resync
3909          * We don't need to check the 'failed' flag as when that gets set,
3910          * recovery aborts.
3911          */
3912         for (i=0; i<mddev->raid_disks; i++)
3913                 if (conf->disks[i].rdev == NULL)
3914                         still_degraded = 1;
3915
3916         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3917
3918         spin_lock(&sh->lock);
3919         set_bit(STRIPE_SYNCING, &sh->state);
3920         clear_bit(STRIPE_INSYNC, &sh->state);
3921         spin_unlock(&sh->lock);
3922
3923         /* wait for any blocked device to be handled */
3924         while(unlikely(!handle_stripe(sh, NULL)))
3925                 ;
3926         release_stripe(sh);
3927
3928         return STRIPE_SECTORS;
3929 }
3930
3931 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3932 {
3933         /* We may not be able to submit a whole bio at once as there
3934          * may not be enough stripe_heads available.
3935          * We cannot pre-allocate enough stripe_heads as we may need
3936          * more than exist in the cache (if we allow ever large chunks).
3937          * So we do one stripe head at a time and record in
3938          * ->bi_hw_segments how many have been done.
3939          *
3940          * We *know* that this entire raid_bio is in one chunk, so
3941          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3942          */
3943         struct stripe_head *sh;
3944         int dd_idx;
3945         sector_t sector, logical_sector, last_sector;
3946         int scnt = 0;
3947         int remaining;
3948         int handled = 0;
3949
3950         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3951         sector = raid5_compute_sector(conf, logical_sector,
3952                                       0, &dd_idx, NULL);
3953         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3954
3955         for (; logical_sector < last_sector;
3956              logical_sector += STRIPE_SECTORS,
3957                      sector += STRIPE_SECTORS,
3958                      scnt++) {
3959
3960                 if (scnt < raid5_bi_hw_segments(raid_bio))
3961                         /* already done this stripe */
3962                         continue;
3963
3964                 sh = get_active_stripe(conf, sector, 0, 1);
3965
3966                 if (!sh) {
3967                         /* failed to get a stripe - must wait */
3968                         raid5_set_bi_hw_segments(raid_bio, scnt);
3969                         conf->retry_read_aligned = raid_bio;
3970                         return handled;
3971                 }
3972
3973                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3974                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3975                         release_stripe(sh);
3976                         raid5_set_bi_hw_segments(raid_bio, scnt);
3977                         conf->retry_read_aligned = raid_bio;
3978                         return handled;
3979                 }
3980
3981                 handle_stripe(sh, NULL);
3982                 release_stripe(sh);
3983                 handled++;
3984         }
3985         spin_lock_irq(&conf->device_lock);
3986         remaining = raid5_dec_bi_phys_segments(raid_bio);
3987         spin_unlock_irq(&conf->device_lock);
3988         if (remaining == 0)
3989                 bio_endio(raid_bio, 0);
3990         if (atomic_dec_and_test(&conf->active_aligned_reads))
3991                 wake_up(&conf->wait_for_stripe);
3992         return handled;
3993 }
3994
3995
3996
3997 /*
3998  * This is our raid5 kernel thread.
3999  *
4000  * We scan the hash table for stripes which can be handled now.
4001  * During the scan, completed stripes are saved for us by the interrupt
4002  * handler, so that they will not have to wait for our next wakeup.
4003  */
4004 static void raid5d(mddev_t *mddev)
4005 {
4006         struct stripe_head *sh;
4007         raid5_conf_t *conf = mddev_to_conf(mddev);
4008         int handled;
4009
4010         pr_debug("+++ raid5d active\n");
4011
4012         md_check_recovery(mddev);
4013
4014         handled = 0;
4015         spin_lock_irq(&conf->device_lock);
4016         while (1) {
4017                 struct bio *bio;
4018
4019                 if (conf->seq_flush != conf->seq_write) {
4020                         int seq = conf->seq_flush;
4021                         spin_unlock_irq(&conf->device_lock);
4022                         bitmap_unplug(mddev->bitmap);
4023                         spin_lock_irq(&conf->device_lock);
4024                         conf->seq_write = seq;
4025                         activate_bit_delay(conf);
4026                 }
4027
4028                 while ((bio = remove_bio_from_retry(conf))) {
4029                         int ok;
4030                         spin_unlock_irq(&conf->device_lock);
4031                         ok = retry_aligned_read(conf, bio);
4032                         spin_lock_irq(&conf->device_lock);
4033                         if (!ok)
4034                                 break;
4035                         handled++;
4036                 }
4037
4038                 sh = __get_priority_stripe(conf);
4039
4040                 if (!sh)
4041                         break;
4042                 spin_unlock_irq(&conf->device_lock);
4043                 
4044                 handled++;
4045                 handle_stripe(sh, conf->spare_page);
4046                 release_stripe(sh);
4047
4048                 spin_lock_irq(&conf->device_lock);
4049         }
4050         pr_debug("%d stripes handled\n", handled);
4051
4052         spin_unlock_irq(&conf->device_lock);
4053
4054         async_tx_issue_pending_all();
4055         unplug_slaves(mddev);
4056
4057         pr_debug("--- raid5d inactive\n");
4058 }
4059
4060 static ssize_t
4061 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4062 {
4063         raid5_conf_t *conf = mddev_to_conf(mddev);
4064         if (conf)
4065                 return sprintf(page, "%d\n", conf->max_nr_stripes);
4066         else
4067                 return 0;
4068 }
4069
4070 static ssize_t
4071 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4072 {
4073         raid5_conf_t *conf = mddev_to_conf(mddev);
4074         unsigned long new;
4075         int err;
4076
4077         if (len >= PAGE_SIZE)
4078                 return -EINVAL;
4079         if (!conf)
4080                 return -ENODEV;
4081
4082         if (strict_strtoul(page, 10, &new))
4083                 return -EINVAL;
4084         if (new <= 16 || new > 32768)
4085                 return -EINVAL;
4086         while (new < conf->max_nr_stripes) {
4087                 if (drop_one_stripe(conf))
4088                         conf->max_nr_stripes--;
4089                 else
4090                         break;
4091         }
4092         err = md_allow_write(mddev);
4093         if (err)
4094                 return err;
4095         while (new > conf->max_nr_stripes) {
4096                 if (grow_one_stripe(conf))
4097                         conf->max_nr_stripes++;
4098                 else break;
4099         }
4100         return len;
4101 }
4102
4103 static struct md_sysfs_entry
4104 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4105                                 raid5_show_stripe_cache_size,
4106                                 raid5_store_stripe_cache_size);
4107
4108 static ssize_t
4109 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4110 {
4111         raid5_conf_t *conf = mddev_to_conf(mddev);
4112         if (conf)
4113                 return sprintf(page, "%d\n", conf->bypass_threshold);
4114         else
4115                 return 0;
4116 }
4117
4118 static ssize_t
4119 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4120 {
4121         raid5_conf_t *conf = mddev_to_conf(mddev);
4122         unsigned long new;
4123         if (len >= PAGE_SIZE)
4124                 return -EINVAL;
4125         if (!conf)
4126                 return -ENODEV;
4127
4128         if (strict_strtoul(page, 10, &new))
4129                 return -EINVAL;
4130         if (new > conf->max_nr_stripes)
4131                 return -EINVAL;
4132         conf->bypass_threshold = new;
4133         return len;
4134 }
4135
4136 static struct md_sysfs_entry
4137 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4138                                         S_IRUGO | S_IWUSR,
4139                                         raid5_show_preread_threshold,
4140                                         raid5_store_preread_threshold);
4141
4142 static ssize_t
4143 stripe_cache_active_show(mddev_t *mddev, char *page)
4144 {
4145         raid5_conf_t *conf = mddev_to_conf(mddev);
4146         if (conf)
4147                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4148         else
4149                 return 0;
4150 }
4151
4152 static struct md_sysfs_entry
4153 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4154
4155 static struct attribute *raid5_attrs[] =  {
4156         &raid5_stripecache_size.attr,
4157         &raid5_stripecache_active.attr,
4158         &raid5_preread_bypass_threshold.attr,
4159         NULL,
4160 };
4161 static struct attribute_group raid5_attrs_group = {
4162         .name = NULL,
4163         .attrs = raid5_attrs,
4164 };
4165
4166 static sector_t
4167 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4168 {
4169         raid5_conf_t *conf = mddev_to_conf(mddev);
4170
4171         if (!sectors)
4172                 sectors = mddev->dev_sectors;
4173         if (!raid_disks)
4174                 raid_disks = conf->previous_raid_disks;
4175
4176         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4177         return sectors * (raid_disks - conf->max_degraded);
4178 }
4179
4180 static raid5_conf_t *setup_conf(mddev_t *mddev)
4181 {
4182         raid5_conf_t *conf;
4183         int raid_disk, memory;
4184         mdk_rdev_t *rdev;
4185         struct disk_info *disk;
4186
4187         if (mddev->new_level != 5
4188             && mddev->new_level != 4
4189             && mddev->new_level != 6) {
4190                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4191                        mdname(mddev), mddev->new_level);
4192                 return ERR_PTR(-EIO);
4193         }
4194         if ((mddev->new_level == 5
4195              && !algorithm_valid_raid5(mddev->new_layout)) ||
4196             (mddev->new_level == 6
4197              && !algorithm_valid_raid6(mddev->new_layout))) {
4198                 printk(KERN_ERR "raid5: %s: layout %d not supported\n",
4199                        mdname(mddev), mddev->new_layout);
4200                 return ERR_PTR(-EIO);
4201         }
4202         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4203                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4204                        mdname(mddev), mddev->raid_disks);
4205                 return ERR_PTR(-EINVAL);
4206         }
4207
4208         if (!mddev->new_chunk || mddev->new_chunk % PAGE_SIZE) {
4209                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4210                         mddev->new_chunk, mdname(mddev));
4211                 return ERR_PTR(-EINVAL);
4212         }
4213
4214         conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4215         if (conf == NULL)
4216                 goto abort;
4217
4218         conf->raid_disks = mddev->raid_disks;
4219         if (mddev->reshape_position == MaxSector)
4220                 conf->previous_raid_disks = mddev->raid_disks;
4221         else
4222                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4223
4224         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4225                               GFP_KERNEL);
4226         if (!conf->disks)
4227                 goto abort;
4228
4229         conf->mddev = mddev;
4230
4231         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4232                 goto abort;
4233
4234         if (mddev->new_level == 6) {
4235                 conf->spare_page = alloc_page(GFP_KERNEL);
4236                 if (!conf->spare_page)
4237                         goto abort;
4238         }
4239         spin_lock_init(&conf->device_lock);
4240         init_waitqueue_head(&conf->wait_for_stripe);
4241         init_waitqueue_head(&conf->wait_for_overlap);
4242         INIT_LIST_HEAD(&conf->handle_list);
4243         INIT_LIST_HEAD(&conf->hold_list);
4244         INIT_LIST_HEAD(&conf->delayed_list);
4245         INIT_LIST_HEAD(&conf->bitmap_list);
4246         INIT_LIST_HEAD(&conf->inactive_list);
4247         atomic_set(&conf->active_stripes, 0);
4248         atomic_set(&conf->preread_active_stripes, 0);
4249         atomic_set(&conf->active_aligned_reads, 0);
4250         conf->bypass_threshold = BYPASS_THRESHOLD;
4251
4252         pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4253
4254         list_for_each_entry(rdev, &mddev->disks, same_set) {
4255                 raid_disk = rdev->raid_disk;
4256                 if (raid_disk >= conf->raid_disks
4257                     || raid_disk < 0)
4258                         continue;
4259                 disk = conf->disks + raid_disk;
4260
4261                 disk->rdev = rdev;
4262
4263                 if (test_bit(In_sync, &rdev->flags)) {
4264                         char b[BDEVNAME_SIZE];
4265                         printk(KERN_INFO "raid5: device %s operational as raid"
4266                                 " disk %d\n", bdevname(rdev->bdev,b),
4267                                 raid_disk);
4268                 } else
4269                         /* Cannot rely on bitmap to complete recovery */
4270                         conf->fullsync = 1;
4271         }
4272
4273         conf->chunk_size = mddev->new_chunk;
4274         conf->level = mddev->new_level;
4275         if (conf->level == 6)
4276                 conf->max_degraded = 2;
4277         else
4278                 conf->max_degraded = 1;
4279         conf->algorithm = mddev->new_layout;
4280         conf->max_nr_stripes = NR_STRIPES;
4281         conf->expand_progress = mddev->reshape_position;
4282
4283         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4284                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4285         if (grow_stripes(conf, conf->max_nr_stripes)) {
4286                 printk(KERN_ERR
4287                         "raid5: couldn't allocate %dkB for buffers\n", memory);
4288                 goto abort;
4289         } else
4290                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4291                         memory, mdname(mddev));
4292
4293         conf->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4294         if (!conf->thread) {
4295                 printk(KERN_ERR
4296                        "raid5: couldn't allocate thread for %s\n",
4297                        mdname(mddev));
4298                 goto abort;
4299         }
4300
4301         return conf;
4302
4303  abort:
4304         if (conf) {
4305                 shrink_stripes(conf);
4306                 safe_put_page(conf->spare_page);
4307                 kfree(conf->disks);
4308                 kfree(conf->stripe_hashtbl);
4309                 kfree(conf);
4310                 return ERR_PTR(-EIO);
4311         } else
4312                 return ERR_PTR(-ENOMEM);
4313 }
4314
4315 static int run(mddev_t *mddev)
4316 {
4317         raid5_conf_t *conf;
4318         int working_disks = 0;
4319         mdk_rdev_t *rdev;
4320
4321         if (mddev->reshape_position != MaxSector) {
4322                 /* Check that we can continue the reshape.
4323                  * Currently only disks can change, it must
4324                  * increase, and we must be past the point where
4325                  * a stripe over-writes itself
4326                  */
4327                 sector_t here_new, here_old;
4328                 int old_disks;
4329                 int max_degraded = (mddev->level == 6 ? 2 : 1);
4330
4331                 if (mddev->new_level != mddev->level ||
4332                     mddev->new_layout != mddev->layout ||
4333                     mddev->new_chunk != mddev->chunk_size) {
4334                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4335                                "required - aborting.\n",
4336                                mdname(mddev));
4337                         return -EINVAL;
4338                 }
4339                 if (mddev->delta_disks <= 0) {
4340                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4341                                "(reduce disks) required - aborting.\n",
4342                                mdname(mddev));
4343                         return -EINVAL;
4344                 }
4345                 old_disks = mddev->raid_disks - mddev->delta_disks;
4346                 /* reshape_position must be on a new-stripe boundary, and one
4347                  * further up in new geometry must map after here in old
4348                  * geometry.
4349                  */
4350                 here_new = mddev->reshape_position;
4351                 if (sector_div(here_new, (mddev->chunk_size>>9)*
4352                                (mddev->raid_disks - max_degraded))) {
4353                         printk(KERN_ERR "raid5: reshape_position not "
4354                                "on a stripe boundary\n");
4355                         return -EINVAL;
4356                 }
4357                 /* here_new is the stripe we will write to */
4358                 here_old = mddev->reshape_position;
4359                 sector_div(here_old, (mddev->chunk_size>>9)*
4360                            (old_disks-max_degraded));
4361                 /* here_old is the first stripe that we might need to read
4362                  * from */
4363                 if (here_new >= here_old) {
4364                         /* Reading from the same stripe as writing to - bad */
4365                         printk(KERN_ERR "raid5: reshape_position too early for "
4366                                "auto-recovery - aborting.\n");
4367                         return -EINVAL;
4368                 }
4369                 printk(KERN_INFO "raid5: reshape will continue\n");
4370                 /* OK, we should be able to continue; */
4371         } else {
4372                 BUG_ON(mddev->level != mddev->new_level);
4373                 BUG_ON(mddev->layout != mddev->new_layout);
4374                 BUG_ON(mddev->chunk_size != mddev->new_chunk);
4375                 BUG_ON(mddev->delta_disks != 0);
4376         }
4377
4378         if (mddev->private == NULL)
4379                 conf = setup_conf(mddev);
4380         else
4381                 conf = mddev->private;
4382
4383         if (IS_ERR(conf))
4384                 return PTR_ERR(conf);
4385
4386         mddev->thread = conf->thread;
4387         conf->thread = NULL;
4388         mddev->private = conf;
4389
4390         /*
4391          * 0 for a fully functional array, 1 or 2 for a degraded array.
4392          */
4393         list_for_each_entry(rdev, &mddev->disks, same_set)
4394                 if (rdev->raid_disk >= 0 &&
4395                     test_bit(In_sync, &rdev->flags))
4396                         working_disks++;
4397
4398         mddev->degraded = conf->raid_disks - working_disks;
4399
4400         if (mddev->degraded > conf->max_degraded) {
4401                 printk(KERN_ERR "raid5: not enough operational devices for %s"
4402                         " (%d/%d failed)\n",
4403                         mdname(mddev), mddev->degraded, conf->raid_disks);
4404                 goto abort;
4405         }
4406
4407         /* device size must be a multiple of chunk size */
4408         mddev->dev_sectors &= ~(mddev->chunk_size / 512 - 1);
4409         mddev->resync_max_sectors = mddev->dev_sectors;
4410
4411         if (mddev->degraded > 0 &&
4412             mddev->recovery_cp != MaxSector) {
4413                 if (mddev->ok_start_degraded)
4414                         printk(KERN_WARNING
4415                                "raid5: starting dirty degraded array: %s"
4416                                "- data corruption possible.\n",
4417                                mdname(mddev));
4418                 else {
4419                         printk(KERN_ERR
4420                                "raid5: cannot start dirty degraded array for %s\n",
4421                                mdname(mddev));
4422                         goto abort;
4423                 }
4424         }
4425
4426         if (mddev->degraded == 0)
4427                 printk("raid5: raid level %d set %s active with %d out of %d"
4428                         " devices, algorithm %d\n", conf->level, mdname(mddev), 
4429                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4430                         conf->algorithm);
4431         else
4432                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4433                         " out of %d devices, algorithm %d\n", conf->level,
4434                         mdname(mddev), mddev->raid_disks - mddev->degraded,
4435                         mddev->raid_disks, conf->algorithm);
4436
4437         print_raid5_conf(conf);
4438
4439         if (conf->expand_progress != MaxSector) {
4440                 printk("...ok start reshape thread\n");
4441                 conf->expand_lo = conf->expand_progress;
4442                 atomic_set(&conf->reshape_stripes, 0);
4443                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4444                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4445                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4446                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4447                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4448                                                         "%s_reshape");
4449         }
4450
4451         /* read-ahead size must cover two whole stripes, which is
4452          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4453          */
4454         {
4455                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4456                 int stripe = data_disks *
4457                         (mddev->chunk_size / PAGE_SIZE);
4458                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4459                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4460         }
4461
4462         /* Ok, everything is just fine now */
4463         if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4464                 printk(KERN_WARNING
4465                        "raid5: failed to create sysfs attributes for %s\n",
4466                        mdname(mddev));
4467
4468         mddev->queue->queue_lock = &conf->device_lock;
4469
4470         mddev->queue->unplug_fn = raid5_unplug_device;
4471         mddev->queue->backing_dev_info.congested_data = mddev;
4472         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4473
4474         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4475
4476         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4477
4478         return 0;
4479 abort:
4480         md_unregister_thread(mddev->thread);
4481         mddev->thread = NULL;
4482         if (conf) {
4483                 shrink_stripes(conf);
4484                 print_raid5_conf(conf);
4485                 safe_put_page(conf->spare_page);
4486                 kfree(conf->disks);
4487                 kfree(conf->stripe_hashtbl);
4488                 kfree(conf);
4489         }
4490         mddev->private = NULL;
4491         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4492         return -EIO;
4493 }
4494
4495
4496
4497 static int stop(mddev_t *mddev)
4498 {
4499         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4500
4501         md_unregister_thread(mddev->thread);
4502         mddev->thread = NULL;
4503         shrink_stripes(conf);
4504         kfree(conf->stripe_hashtbl);
4505         mddev->queue->backing_dev_info.congested_fn = NULL;
4506         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4507         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4508         kfree(conf->disks);
4509         kfree(conf);
4510         mddev->private = NULL;
4511         return 0;
4512 }
4513
4514 #ifdef DEBUG
4515 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
4516 {
4517         int i;
4518
4519         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4520                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4521         seq_printf(seq, "sh %llu,  count %d.\n",
4522                    (unsigned long long)sh->sector, atomic_read(&sh->count));
4523         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4524         for (i = 0; i < sh->disks; i++) {
4525                 seq_printf(seq, "(cache%d: %p %ld) ",
4526                            i, sh->dev[i].page, sh->dev[i].flags);
4527         }
4528         seq_printf(seq, "\n");
4529 }
4530
4531 static void printall(struct seq_file *seq, raid5_conf_t *conf)
4532 {
4533         struct stripe_head *sh;
4534         struct hlist_node *hn;
4535         int i;
4536
4537         spin_lock_irq(&conf->device_lock);
4538         for (i = 0; i < NR_HASH; i++) {
4539                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4540                         if (sh->raid_conf != conf)
4541                                 continue;
4542                         print_sh(seq, sh);
4543                 }
4544         }
4545         spin_unlock_irq(&conf->device_lock);
4546 }
4547 #endif
4548
4549 static void status(struct seq_file *seq, mddev_t *mddev)
4550 {
4551         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4552         int i;
4553
4554         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4555         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4556         for (i = 0; i < conf->raid_disks; i++)
4557                 seq_printf (seq, "%s",
4558                                conf->disks[i].rdev &&
4559                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4560         seq_printf (seq, "]");
4561 #ifdef DEBUG
4562         seq_printf (seq, "\n");
4563         printall(seq, conf);
4564 #endif
4565 }
4566
4567 static void print_raid5_conf (raid5_conf_t *conf)
4568 {
4569         int i;
4570         struct disk_info *tmp;
4571
4572         printk("RAID5 conf printout:\n");
4573         if (!conf) {
4574                 printk("(conf==NULL)\n");
4575                 return;
4576         }
4577         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4578                  conf->raid_disks - conf->mddev->degraded);
4579
4580         for (i = 0; i < conf->raid_disks; i++) {
4581                 char b[BDEVNAME_SIZE];
4582                 tmp = conf->disks + i;
4583                 if (tmp->rdev)
4584                 printk(" disk %d, o:%d, dev:%s\n",
4585                         i, !test_bit(Faulty, &tmp->rdev->flags),
4586                         bdevname(tmp->rdev->bdev,b));
4587         }
4588 }
4589
4590 static int raid5_spare_active(mddev_t *mddev)
4591 {
4592         int i;
4593         raid5_conf_t *conf = mddev->private;
4594         struct disk_info *tmp;
4595
4596         for (i = 0; i < conf->raid_disks; i++) {
4597                 tmp = conf->disks + i;
4598                 if (tmp->rdev
4599                     && !test_bit(Faulty, &tmp->rdev->flags)
4600                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4601                         unsigned long flags;
4602                         spin_lock_irqsave(&conf->device_lock, flags);
4603                         mddev->degraded--;
4604                         spin_unlock_irqrestore(&conf->device_lock, flags);
4605                 }
4606         }
4607         print_raid5_conf(conf);
4608         return 0;
4609 }
4610
4611 static int raid5_remove_disk(mddev_t *mddev, int number)
4612 {
4613         raid5_conf_t *conf = mddev->private;
4614         int err = 0;
4615         mdk_rdev_t *rdev;
4616         struct disk_info *p = conf->disks + number;
4617
4618         print_raid5_conf(conf);
4619         rdev = p->rdev;
4620         if (rdev) {
4621                 if (test_bit(In_sync, &rdev->flags) ||
4622                     atomic_read(&rdev->nr_pending)) {
4623                         err = -EBUSY;
4624                         goto abort;
4625                 }
4626                 /* Only remove non-faulty devices if recovery
4627                  * isn't possible.
4628                  */
4629                 if (!test_bit(Faulty, &rdev->flags) &&
4630                     mddev->degraded <= conf->max_degraded) {
4631                         err = -EBUSY;
4632                         goto abort;
4633                 }
4634                 p->rdev = NULL;
4635                 synchronize_rcu();
4636                 if (atomic_read(&rdev->nr_pending)) {
4637                         /* lost the race, try later */
4638                         err = -EBUSY;
4639                         p->rdev = rdev;
4640                 }
4641         }
4642 abort:
4643
4644         print_raid5_conf(conf);
4645         return err;
4646 }
4647
4648 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4649 {
4650         raid5_conf_t *conf = mddev->private;
4651         int err = -EEXIST;
4652         int disk;
4653         struct disk_info *p;
4654         int first = 0;
4655         int last = conf->raid_disks - 1;
4656
4657         if (mddev->degraded > conf->max_degraded)
4658                 /* no point adding a device */
4659                 return -EINVAL;
4660
4661         if (rdev->raid_disk >= 0)
4662                 first = last = rdev->raid_disk;
4663
4664         /*
4665          * find the disk ... but prefer rdev->saved_raid_disk
4666          * if possible.
4667          */
4668         if (rdev->saved_raid_disk >= 0 &&
4669             rdev->saved_raid_disk >= first &&
4670             conf->disks[rdev->saved_raid_disk].rdev == NULL)
4671                 disk = rdev->saved_raid_disk;
4672         else
4673                 disk = first;
4674         for ( ; disk <= last ; disk++)
4675                 if ((p=conf->disks + disk)->rdev == NULL) {
4676                         clear_bit(In_sync, &rdev->flags);
4677                         rdev->raid_disk = disk;
4678                         err = 0;
4679                         if (rdev->saved_raid_disk != disk)
4680                                 conf->fullsync = 1;
4681                         rcu_assign_pointer(p->rdev, rdev);
4682                         break;
4683                 }
4684         print_raid5_conf(conf);
4685         return err;
4686 }
4687
4688 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4689 {
4690         /* no resync is happening, and there is enough space
4691          * on all devices, so we can resize.
4692          * We need to make sure resync covers any new space.
4693          * If the array is shrinking we should possibly wait until
4694          * any io in the removed space completes, but it hardly seems
4695          * worth it.
4696          */
4697         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4698         md_set_array_sectors(mddev, raid5_size(mddev, sectors,
4699                                                mddev->raid_disks));
4700         if (mddev->array_sectors >
4701             raid5_size(mddev, sectors, mddev->raid_disks))
4702                 return -EINVAL;
4703         set_capacity(mddev->gendisk, mddev->array_sectors);
4704         mddev->changed = 1;
4705         if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
4706                 mddev->recovery_cp = mddev->dev_sectors;
4707                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4708         }
4709         mddev->dev_sectors = sectors;
4710         mddev->resync_max_sectors = sectors;
4711         return 0;
4712 }
4713
4714 #ifdef CONFIG_MD_RAID5_RESHAPE
4715 static int raid5_check_reshape(mddev_t *mddev)
4716 {
4717         raid5_conf_t *conf = mddev_to_conf(mddev);
4718         int err;
4719
4720         if (mddev->delta_disks < 0 ||
4721             mddev->new_level != mddev->level)
4722                 return -EINVAL; /* Cannot shrink array or change level yet */
4723         if (mddev->delta_disks == 0)
4724                 return 0; /* nothing to do */
4725         if (mddev->bitmap)
4726                 /* Cannot grow a bitmap yet */
4727                 return -EBUSY;
4728
4729         /* Can only proceed if there are plenty of stripe_heads.
4730          * We need a minimum of one full stripe,, and for sensible progress
4731          * it is best to have about 4 times that.
4732          * If we require 4 times, then the default 256 4K stripe_heads will
4733          * allow for chunk sizes up to 256K, which is probably OK.
4734          * If the chunk size is greater, user-space should request more
4735          * stripe_heads first.
4736          */
4737         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4738             (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4739                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
4740                        (mddev->chunk_size / STRIPE_SIZE)*4);
4741                 return -ENOSPC;
4742         }
4743
4744         err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4745         if (err)
4746                 return err;
4747
4748         if (mddev->degraded > conf->max_degraded)
4749                 return -EINVAL;
4750         /* looks like we might be able to manage this */
4751         return 0;
4752 }
4753
4754 static int raid5_start_reshape(mddev_t *mddev)
4755 {
4756         raid5_conf_t *conf = mddev_to_conf(mddev);
4757         mdk_rdev_t *rdev;
4758         int spares = 0;
4759         int added_devices = 0;
4760         unsigned long flags;
4761
4762         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4763                 return -EBUSY;
4764
4765         list_for_each_entry(rdev, &mddev->disks, same_set)
4766                 if (rdev->raid_disk < 0 &&
4767                     !test_bit(Faulty, &rdev->flags))
4768                         spares++;
4769
4770         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4771                 /* Not enough devices even to make a degraded array
4772                  * of that size
4773                  */
4774                 return -EINVAL;
4775
4776         atomic_set(&conf->reshape_stripes, 0);
4777         spin_lock_irq(&conf->device_lock);
4778         conf->previous_raid_disks = conf->raid_disks;
4779         conf->raid_disks += mddev->delta_disks;
4780         conf->expand_progress = 0;
4781         conf->expand_lo = 0;
4782         spin_unlock_irq(&conf->device_lock);
4783
4784         /* Add some new drives, as many as will fit.
4785          * We know there are enough to make the newly sized array work.
4786          */
4787         list_for_each_entry(rdev, &mddev->disks, same_set)
4788                 if (rdev->raid_disk < 0 &&
4789                     !test_bit(Faulty, &rdev->flags)) {
4790                         if (raid5_add_disk(mddev, rdev) == 0) {
4791                                 char nm[20];
4792                                 set_bit(In_sync, &rdev->flags);
4793                                 added_devices++;
4794                                 rdev->recovery_offset = 0;
4795                                 sprintf(nm, "rd%d", rdev->raid_disk);
4796                                 if (sysfs_create_link(&mddev->kobj,
4797                                                       &rdev->kobj, nm))
4798                                         printk(KERN_WARNING
4799                                                "raid5: failed to create "
4800                                                " link %s for %s\n",
4801                                                nm, mdname(mddev));
4802                         } else
4803                                 break;
4804                 }
4805
4806         spin_lock_irqsave(&conf->device_lock, flags);
4807         mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4808         spin_unlock_irqrestore(&conf->device_lock, flags);
4809         mddev->raid_disks = conf->raid_disks;
4810         mddev->reshape_position = 0;
4811         set_bit(MD_CHANGE_DEVS, &mddev->flags);
4812
4813         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4814         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4815         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4816         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4817         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4818                                                 "%s_reshape");
4819         if (!mddev->sync_thread) {
4820                 mddev->recovery = 0;
4821                 spin_lock_irq(&conf->device_lock);
4822                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4823                 conf->expand_progress = MaxSector;
4824                 spin_unlock_irq(&conf->device_lock);
4825                 return -EAGAIN;
4826         }
4827         md_wakeup_thread(mddev->sync_thread);
4828         md_new_event(mddev);
4829         return 0;
4830 }
4831 #endif
4832
4833 static void end_reshape(raid5_conf_t *conf)
4834 {
4835         struct block_device *bdev;
4836
4837         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4838                 mddev_t *mddev = conf->mddev;
4839
4840                 md_set_array_sectors_lock(mddev, raid5_size(mddev, 0,
4841                                                        conf->raid_disks));
4842                 set_capacity(mddev->gendisk, mddev->array_sectors);
4843                 mddev->changed = 1;
4844                 conf->previous_raid_disks = conf->raid_disks;
4845
4846                 bdev = bdget_disk(conf->mddev->gendisk, 0);
4847                 if (bdev) {
4848                         mutex_lock(&bdev->bd_inode->i_mutex);
4849                         i_size_write(bdev->bd_inode,
4850                                      (loff_t)conf->mddev->array_sectors << 9);
4851                         mutex_unlock(&bdev->bd_inode->i_mutex);
4852                         bdput(bdev);
4853                 }
4854                 spin_lock_irq(&conf->device_lock);
4855                 conf->expand_progress = MaxSector;
4856                 spin_unlock_irq(&conf->device_lock);
4857                 conf->mddev->reshape_position = MaxSector;
4858
4859                 /* read-ahead size must cover two whole stripes, which is
4860                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4861                  */
4862                 {
4863                         int data_disks = conf->previous_raid_disks - conf->max_degraded;
4864                         int stripe = data_disks *
4865                                 (conf->mddev->chunk_size / PAGE_SIZE);
4866                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4867                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4868                 }
4869         }
4870 }
4871
4872 static void raid5_quiesce(mddev_t *mddev, int state)
4873 {
4874         raid5_conf_t *conf = mddev_to_conf(mddev);
4875
4876         switch(state) {
4877         case 2: /* resume for a suspend */
4878                 wake_up(&conf->wait_for_overlap);
4879                 break;
4880
4881         case 1: /* stop all writes */
4882                 spin_lock_irq(&conf->device_lock);
4883                 conf->quiesce = 1;
4884                 wait_event_lock_irq(conf->wait_for_stripe,
4885                                     atomic_read(&conf->active_stripes) == 0 &&
4886                                     atomic_read(&conf->active_aligned_reads) == 0,
4887                                     conf->device_lock, /* nothing */);
4888                 spin_unlock_irq(&conf->device_lock);
4889                 break;
4890
4891         case 0: /* re-enable writes */
4892                 spin_lock_irq(&conf->device_lock);
4893                 conf->quiesce = 0;
4894                 wake_up(&conf->wait_for_stripe);
4895                 wake_up(&conf->wait_for_overlap);
4896                 spin_unlock_irq(&conf->device_lock);
4897                 break;
4898         }
4899 }
4900
4901
4902 static void *raid5_takeover_raid1(mddev_t *mddev)
4903 {
4904         int chunksect;
4905
4906         if (mddev->raid_disks != 2 ||
4907             mddev->degraded > 1)
4908                 return ERR_PTR(-EINVAL);
4909
4910         /* Should check if there are write-behind devices? */
4911
4912         chunksect = 64*2; /* 64K by default */
4913
4914         /* The array must be an exact multiple of chunksize */
4915         while (chunksect && (mddev->array_sectors & (chunksect-1)))
4916                 chunksect >>= 1;
4917
4918         if ((chunksect<<9) < STRIPE_SIZE)
4919                 /* array size does not allow a suitable chunk size */
4920                 return ERR_PTR(-EINVAL);
4921
4922         mddev->new_level = 5;
4923         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
4924         mddev->new_chunk = chunksect << 9;
4925
4926         return setup_conf(mddev);
4927 }
4928
4929 static void *raid5_takeover_raid6(mddev_t *mddev)
4930 {
4931         int new_layout;
4932
4933         switch (mddev->layout) {
4934         case ALGORITHM_LEFT_ASYMMETRIC_6:
4935                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
4936                 break;
4937         case ALGORITHM_RIGHT_ASYMMETRIC_6:
4938                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
4939                 break;
4940         case ALGORITHM_LEFT_SYMMETRIC_6:
4941                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
4942                 break;
4943         case ALGORITHM_RIGHT_SYMMETRIC_6:
4944                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
4945                 break;
4946         case ALGORITHM_PARITY_0_6:
4947                 new_layout = ALGORITHM_PARITY_0;
4948                 break;
4949         case ALGORITHM_PARITY_N:
4950                 new_layout = ALGORITHM_PARITY_N;
4951                 break;
4952         default:
4953                 return ERR_PTR(-EINVAL);
4954         }
4955         mddev->new_level = 5;
4956         mddev->new_layout = new_layout;
4957         mddev->delta_disks = -1;
4958         mddev->raid_disks -= 1;
4959         return setup_conf(mddev);
4960 }
4961
4962
4963 static int raid5_reconfig(mddev_t *mddev, int new_layout, int new_chunk)
4964 {
4965         /* Currently the layout and chunk size can only be changed
4966          * for a 2-drive raid array, as in that case no data shuffling
4967          * is required.
4968          * Later we might validate these and set new_* so a reshape
4969          * can complete the change.
4970          */
4971         raid5_conf_t *conf = mddev_to_conf(mddev);
4972
4973         if (new_layout >= 0 && !algorithm_valid_raid5(new_layout))
4974                 return -EINVAL;
4975         if (new_chunk > 0) {
4976                 if (new_chunk & (new_chunk-1))
4977                         /* not a power of 2 */
4978                         return -EINVAL;
4979                 if (new_chunk < PAGE_SIZE)
4980                         return -EINVAL;
4981                 if (mddev->array_sectors & ((new_chunk>>9)-1))
4982                         /* not factor of array size */
4983                         return -EINVAL;
4984         }
4985
4986         /* They look valid */
4987
4988         if (mddev->raid_disks != 2)
4989                 return -EINVAL;
4990
4991         if (new_layout >= 0) {
4992                 conf->algorithm = new_layout;
4993                 mddev->layout = mddev->new_layout = new_layout;
4994         }
4995         if (new_chunk > 0) {
4996                 conf->chunk_size = new_chunk;
4997                 mddev->chunk_size = mddev->new_chunk = new_chunk;
4998         }
4999         set_bit(MD_CHANGE_DEVS, &mddev->flags);
5000         md_wakeup_thread(mddev->thread);
5001         return 0;
5002 }
5003
5004 static void *raid5_takeover(mddev_t *mddev)
5005 {
5006         /* raid5 can take over:
5007          *  raid0 - if all devices are the same - make it a raid4 layout
5008          *  raid1 - if there are two drives.  We need to know the chunk size
5009          *  raid4 - trivial - just use a raid4 layout.
5010          *  raid6 - Providing it is a *_6 layout
5011          *
5012          * For now, just do raid1
5013          */
5014
5015         if (mddev->level == 1)
5016                 return raid5_takeover_raid1(mddev);
5017         if (mddev->level == 4) {
5018                 mddev->new_layout = ALGORITHM_PARITY_N;
5019                 mddev->new_level = 5;
5020                 return setup_conf(mddev);
5021         }
5022         if (mddev->level == 6)
5023                 return raid5_takeover_raid6(mddev);
5024
5025         return ERR_PTR(-EINVAL);
5026 }
5027
5028
5029 static struct mdk_personality raid5_personality;
5030
5031 static void *raid6_takeover(mddev_t *mddev)
5032 {
5033         /* Currently can only take over a raid5.  We map the
5034          * personality to an equivalent raid6 personality
5035          * with the Q block at the end.
5036          */
5037         int new_layout;
5038
5039         if (mddev->pers != &raid5_personality)
5040                 return ERR_PTR(-EINVAL);
5041         if (mddev->degraded > 1)
5042                 return ERR_PTR(-EINVAL);
5043         if (mddev->raid_disks > 253)
5044                 return ERR_PTR(-EINVAL);
5045         if (mddev->raid_disks < 3)
5046                 return ERR_PTR(-EINVAL);
5047
5048         switch (mddev->layout) {
5049         case ALGORITHM_LEFT_ASYMMETRIC:
5050                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5051                 break;
5052         case ALGORITHM_RIGHT_ASYMMETRIC:
5053                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5054                 break;
5055         case ALGORITHM_LEFT_SYMMETRIC:
5056                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5057                 break;
5058         case ALGORITHM_RIGHT_SYMMETRIC:
5059                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5060                 break;
5061         case ALGORITHM_PARITY_0:
5062                 new_layout = ALGORITHM_PARITY_0_6;
5063                 break;
5064         case ALGORITHM_PARITY_N:
5065                 new_layout = ALGORITHM_PARITY_N;
5066                 break;
5067         default:
5068                 return ERR_PTR(-EINVAL);
5069         }
5070         mddev->new_level = 6;
5071         mddev->new_layout = new_layout;
5072         mddev->delta_disks = 1;
5073         mddev->raid_disks += 1;
5074         return setup_conf(mddev);
5075 }
5076
5077
5078 static struct mdk_personality raid6_personality =
5079 {
5080         .name           = "raid6",
5081         .level          = 6,
5082         .owner          = THIS_MODULE,
5083         .make_request   = make_request,
5084         .run            = run,
5085         .stop           = stop,
5086         .status         = status,
5087         .error_handler  = error,
5088         .hot_add_disk   = raid5_add_disk,
5089         .hot_remove_disk= raid5_remove_disk,
5090         .spare_active   = raid5_spare_active,
5091         .sync_request   = sync_request,
5092         .resize         = raid5_resize,
5093         .size           = raid5_size,
5094 #ifdef CONFIG_MD_RAID5_RESHAPE
5095         .check_reshape  = raid5_check_reshape,
5096         .start_reshape  = raid5_start_reshape,
5097 #endif
5098         .quiesce        = raid5_quiesce,
5099         .takeover       = raid6_takeover,
5100 };
5101 static struct mdk_personality raid5_personality =
5102 {
5103         .name           = "raid5",
5104         .level          = 5,
5105         .owner          = THIS_MODULE,
5106         .make_request   = make_request,
5107         .run            = run,
5108         .stop           = stop,
5109         .status         = status,
5110         .error_handler  = error,
5111         .hot_add_disk   = raid5_add_disk,
5112         .hot_remove_disk= raid5_remove_disk,
5113         .spare_active   = raid5_spare_active,
5114         .sync_request   = sync_request,
5115         .resize         = raid5_resize,
5116         .size           = raid5_size,
5117 #ifdef CONFIG_MD_RAID5_RESHAPE
5118         .check_reshape  = raid5_check_reshape,
5119         .start_reshape  = raid5_start_reshape,
5120 #endif
5121         .quiesce        = raid5_quiesce,
5122         .takeover       = raid5_takeover,
5123         .reconfig       = raid5_reconfig,
5124 };
5125
5126 static struct mdk_personality raid4_personality =
5127 {
5128         .name           = "raid4",
5129         .level          = 4,
5130         .owner          = THIS_MODULE,
5131         .make_request   = make_request,
5132         .run            = run,
5133         .stop           = stop,
5134         .status         = status,
5135         .error_handler  = error,
5136         .hot_add_disk   = raid5_add_disk,
5137         .hot_remove_disk= raid5_remove_disk,
5138         .spare_active   = raid5_spare_active,
5139         .sync_request   = sync_request,
5140         .resize         = raid5_resize,
5141         .size           = raid5_size,
5142 #ifdef CONFIG_MD_RAID5_RESHAPE
5143         .check_reshape  = raid5_check_reshape,
5144         .start_reshape  = raid5_start_reshape,
5145 #endif
5146         .quiesce        = raid5_quiesce,
5147 };
5148
5149 static int __init raid5_init(void)
5150 {
5151         register_md_personality(&raid6_personality);
5152         register_md_personality(&raid5_personality);
5153         register_md_personality(&raid4_personality);
5154         return 0;
5155 }
5156
5157 static void raid5_exit(void)
5158 {
5159         unregister_md_personality(&raid6_personality);
5160         unregister_md_personality(&raid5_personality);
5161         unregister_md_personality(&raid4_personality);
5162 }
5163
5164 module_init(raid5_init);
5165 module_exit(raid5_exit);
5166 MODULE_LICENSE("GPL");
5167 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5168 MODULE_ALIAS("md-raid5");
5169 MODULE_ALIAS("md-raid4");
5170 MODULE_ALIAS("md-level-5");
5171 MODULE_ALIAS("md-level-4");
5172 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5173 MODULE_ALIAS("md-raid6");
5174 MODULE_ALIAS("md-level-6");
5175
5176 /* This used to be two separate modules, they were: */
5177 MODULE_ALIAS("raid5");
5178 MODULE_ALIAS("raid6");