md: raid-1/10: fix RW bits manipulation
[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, int previous);
303 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
304                             struct stripe_head *sh);
305
306 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
307 {
308         raid5_conf_t *conf = sh->raid_conf;
309         int i;
310
311         BUG_ON(atomic_read(&sh->count) != 0);
312         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
313         BUG_ON(stripe_operations_active(sh));
314
315         CHECK_DEVLOCK();
316         pr_debug("init_stripe called, stripe %llu\n",
317                 (unsigned long long)sh->sector);
318
319         remove_hash(sh);
320
321         sh->generation = conf->generation - previous;
322         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
323         sh->sector = sector;
324         stripe_set_idx(sector, conf, previous, sh);
325         sh->state = 0;
326
327
328         for (i = sh->disks; i--; ) {
329                 struct r5dev *dev = &sh->dev[i];
330
331                 if (dev->toread || dev->read || dev->towrite || dev->written ||
332                     test_bit(R5_LOCKED, &dev->flags)) {
333                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
334                                (unsigned long long)sh->sector, i, dev->toread,
335                                dev->read, dev->towrite, dev->written,
336                                test_bit(R5_LOCKED, &dev->flags));
337                         BUG();
338                 }
339                 dev->flags = 0;
340                 raid5_build_block(sh, i, previous);
341         }
342         insert_hash(conf, sh);
343 }
344
345 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
346                                          short generation)
347 {
348         struct stripe_head *sh;
349         struct hlist_node *hn;
350
351         CHECK_DEVLOCK();
352         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
353         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
354                 if (sh->sector == sector && sh->generation == generation)
355                         return sh;
356         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
357         return NULL;
358 }
359
360 static void unplug_slaves(mddev_t *mddev);
361 static void raid5_unplug_device(struct request_queue *q);
362
363 static struct stripe_head *
364 get_active_stripe(raid5_conf_t *conf, sector_t sector,
365                   int previous, int noblock, int noquiesce)
366 {
367         struct stripe_head *sh;
368
369         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
370
371         spin_lock_irq(&conf->device_lock);
372
373         do {
374                 wait_event_lock_irq(conf->wait_for_stripe,
375                                     conf->quiesce == 0 || noquiesce,
376                                     conf->device_lock, /* nothing */);
377                 sh = __find_stripe(conf, sector, conf->generation - previous);
378                 if (!sh) {
379                         if (!conf->inactive_blocked)
380                                 sh = get_free_stripe(conf);
381                         if (noblock && sh == NULL)
382                                 break;
383                         if (!sh) {
384                                 conf->inactive_blocked = 1;
385                                 wait_event_lock_irq(conf->wait_for_stripe,
386                                                     !list_empty(&conf->inactive_list) &&
387                                                     (atomic_read(&conf->active_stripes)
388                                                      < (conf->max_nr_stripes *3/4)
389                                                      || !conf->inactive_blocked),
390                                                     conf->device_lock,
391                                                     raid5_unplug_device(conf->mddev->queue)
392                                         );
393                                 conf->inactive_blocked = 0;
394                         } else
395                                 init_stripe(sh, sector, previous);
396                 } else {
397                         if (atomic_read(&sh->count)) {
398                                 BUG_ON(!list_empty(&sh->lru)
399                                     && !test_bit(STRIPE_EXPANDING, &sh->state));
400                         } else {
401                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
402                                         atomic_inc(&conf->active_stripes);
403                                 if (list_empty(&sh->lru) &&
404                                     !test_bit(STRIPE_EXPANDING, &sh->state))
405                                         BUG();
406                                 list_del_init(&sh->lru);
407                         }
408                 }
409         } while (sh == NULL);
410
411         if (sh)
412                 atomic_inc(&sh->count);
413
414         spin_unlock_irq(&conf->device_lock);
415         return sh;
416 }
417
418 static void
419 raid5_end_read_request(struct bio *bi, int error);
420 static void
421 raid5_end_write_request(struct bio *bi, int error);
422
423 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
424 {
425         raid5_conf_t *conf = sh->raid_conf;
426         int i, disks = sh->disks;
427
428         might_sleep();
429
430         for (i = disks; i--; ) {
431                 int rw;
432                 struct bio *bi;
433                 mdk_rdev_t *rdev;
434                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
435                         rw = WRITE;
436                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
437                         rw = READ;
438                 else
439                         continue;
440
441                 bi = &sh->dev[i].req;
442
443                 bi->bi_rw = rw;
444                 if (rw == WRITE)
445                         bi->bi_end_io = raid5_end_write_request;
446                 else
447                         bi->bi_end_io = raid5_end_read_request;
448
449                 rcu_read_lock();
450                 rdev = rcu_dereference(conf->disks[i].rdev);
451                 if (rdev && test_bit(Faulty, &rdev->flags))
452                         rdev = NULL;
453                 if (rdev)
454                         atomic_inc(&rdev->nr_pending);
455                 rcu_read_unlock();
456
457                 if (rdev) {
458                         if (s->syncing || s->expanding || s->expanded)
459                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
460
461                         set_bit(STRIPE_IO_STARTED, &sh->state);
462
463                         bi->bi_bdev = rdev->bdev;
464                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
465                                 __func__, (unsigned long long)sh->sector,
466                                 bi->bi_rw, i);
467                         atomic_inc(&sh->count);
468                         bi->bi_sector = sh->sector + rdev->data_offset;
469                         bi->bi_flags = 1 << BIO_UPTODATE;
470                         bi->bi_vcnt = 1;
471                         bi->bi_max_vecs = 1;
472                         bi->bi_idx = 0;
473                         bi->bi_io_vec = &sh->dev[i].vec;
474                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
475                         bi->bi_io_vec[0].bv_offset = 0;
476                         bi->bi_size = STRIPE_SIZE;
477                         bi->bi_next = NULL;
478                         if (rw == WRITE &&
479                             test_bit(R5_ReWrite, &sh->dev[i].flags))
480                                 atomic_add(STRIPE_SECTORS,
481                                         &rdev->corrected_errors);
482                         generic_make_request(bi);
483                 } else {
484                         if (rw == WRITE)
485                                 set_bit(STRIPE_DEGRADED, &sh->state);
486                         pr_debug("skip op %ld on disc %d for sector %llu\n",
487                                 bi->bi_rw, i, (unsigned long long)sh->sector);
488                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
489                         set_bit(STRIPE_HANDLE, &sh->state);
490                 }
491         }
492 }
493
494 static struct dma_async_tx_descriptor *
495 async_copy_data(int frombio, struct bio *bio, struct page *page,
496         sector_t sector, struct dma_async_tx_descriptor *tx)
497 {
498         struct bio_vec *bvl;
499         struct page *bio_page;
500         int i;
501         int page_offset;
502
503         if (bio->bi_sector >= sector)
504                 page_offset = (signed)(bio->bi_sector - sector) * 512;
505         else
506                 page_offset = (signed)(sector - bio->bi_sector) * -512;
507         bio_for_each_segment(bvl, bio, i) {
508                 int len = bio_iovec_idx(bio, i)->bv_len;
509                 int clen;
510                 int b_offset = 0;
511
512                 if (page_offset < 0) {
513                         b_offset = -page_offset;
514                         page_offset += b_offset;
515                         len -= b_offset;
516                 }
517
518                 if (len > 0 && page_offset + len > STRIPE_SIZE)
519                         clen = STRIPE_SIZE - page_offset;
520                 else
521                         clen = len;
522
523                 if (clen > 0) {
524                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
525                         bio_page = bio_iovec_idx(bio, i)->bv_page;
526                         if (frombio)
527                                 tx = async_memcpy(page, bio_page, page_offset,
528                                         b_offset, clen,
529                                         ASYNC_TX_DEP_ACK,
530                                         tx, NULL, NULL);
531                         else
532                                 tx = async_memcpy(bio_page, page, b_offset,
533                                         page_offset, clen,
534                                         ASYNC_TX_DEP_ACK,
535                                         tx, NULL, NULL);
536                 }
537                 if (clen < len) /* hit end of page */
538                         break;
539                 page_offset +=  len;
540         }
541
542         return tx;
543 }
544
545 static void ops_complete_biofill(void *stripe_head_ref)
546 {
547         struct stripe_head *sh = stripe_head_ref;
548         struct bio *return_bi = NULL;
549         raid5_conf_t *conf = sh->raid_conf;
550         int i;
551
552         pr_debug("%s: stripe %llu\n", __func__,
553                 (unsigned long long)sh->sector);
554
555         /* clear completed biofills */
556         spin_lock_irq(&conf->device_lock);
557         for (i = sh->disks; i--; ) {
558                 struct r5dev *dev = &sh->dev[i];
559
560                 /* acknowledge completion of a biofill operation */
561                 /* and check if we need to reply to a read request,
562                  * new R5_Wantfill requests are held off until
563                  * !STRIPE_BIOFILL_RUN
564                  */
565                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
566                         struct bio *rbi, *rbi2;
567
568                         BUG_ON(!dev->read);
569                         rbi = dev->read;
570                         dev->read = NULL;
571                         while (rbi && rbi->bi_sector <
572                                 dev->sector + STRIPE_SECTORS) {
573                                 rbi2 = r5_next_bio(rbi, dev->sector);
574                                 if (!raid5_dec_bi_phys_segments(rbi)) {
575                                         rbi->bi_next = return_bi;
576                                         return_bi = rbi;
577                                 }
578                                 rbi = rbi2;
579                         }
580                 }
581         }
582         spin_unlock_irq(&conf->device_lock);
583         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
584
585         return_io(return_bi);
586
587         set_bit(STRIPE_HANDLE, &sh->state);
588         release_stripe(sh);
589 }
590
591 static void ops_run_biofill(struct stripe_head *sh)
592 {
593         struct dma_async_tx_descriptor *tx = NULL;
594         raid5_conf_t *conf = sh->raid_conf;
595         int i;
596
597         pr_debug("%s: stripe %llu\n", __func__,
598                 (unsigned long long)sh->sector);
599
600         for (i = sh->disks; i--; ) {
601                 struct r5dev *dev = &sh->dev[i];
602                 if (test_bit(R5_Wantfill, &dev->flags)) {
603                         struct bio *rbi;
604                         spin_lock_irq(&conf->device_lock);
605                         dev->read = rbi = dev->toread;
606                         dev->toread = NULL;
607                         spin_unlock_irq(&conf->device_lock);
608                         while (rbi && rbi->bi_sector <
609                                 dev->sector + STRIPE_SECTORS) {
610                                 tx = async_copy_data(0, rbi, dev->page,
611                                         dev->sector, tx);
612                                 rbi = r5_next_bio(rbi, dev->sector);
613                         }
614                 }
615         }
616
617         atomic_inc(&sh->count);
618         async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
619                 ops_complete_biofill, sh);
620 }
621
622 static void ops_complete_compute5(void *stripe_head_ref)
623 {
624         struct stripe_head *sh = stripe_head_ref;
625         int target = sh->ops.target;
626         struct r5dev *tgt = &sh->dev[target];
627
628         pr_debug("%s: stripe %llu\n", __func__,
629                 (unsigned long long)sh->sector);
630
631         set_bit(R5_UPTODATE, &tgt->flags);
632         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
633         clear_bit(R5_Wantcompute, &tgt->flags);
634         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
635         if (sh->check_state == check_state_compute_run)
636                 sh->check_state = check_state_compute_result;
637         set_bit(STRIPE_HANDLE, &sh->state);
638         release_stripe(sh);
639 }
640
641 static struct dma_async_tx_descriptor *ops_run_compute5(struct stripe_head *sh)
642 {
643         /* kernel stack size limits the total number of disks */
644         int disks = sh->disks;
645         struct page *xor_srcs[disks];
646         int target = sh->ops.target;
647         struct r5dev *tgt = &sh->dev[target];
648         struct page *xor_dest = tgt->page;
649         int count = 0;
650         struct dma_async_tx_descriptor *tx;
651         int i;
652
653         pr_debug("%s: stripe %llu block: %d\n",
654                 __func__, (unsigned long long)sh->sector, target);
655         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
656
657         for (i = disks; i--; )
658                 if (i != target)
659                         xor_srcs[count++] = sh->dev[i].page;
660
661         atomic_inc(&sh->count);
662
663         if (unlikely(count == 1))
664                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
665                         0, NULL, ops_complete_compute5, sh);
666         else
667                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
668                         ASYNC_TX_XOR_ZERO_DST, NULL,
669                         ops_complete_compute5, sh);
670
671         return tx;
672 }
673
674 static void ops_complete_prexor(void *stripe_head_ref)
675 {
676         struct stripe_head *sh = stripe_head_ref;
677
678         pr_debug("%s: stripe %llu\n", __func__,
679                 (unsigned long long)sh->sector);
680 }
681
682 static struct dma_async_tx_descriptor *
683 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
684 {
685         /* kernel stack size limits the total number of disks */
686         int disks = sh->disks;
687         struct page *xor_srcs[disks];
688         int count = 0, pd_idx = sh->pd_idx, i;
689
690         /* existing parity data subtracted */
691         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
692
693         pr_debug("%s: stripe %llu\n", __func__,
694                 (unsigned long long)sh->sector);
695
696         for (i = disks; i--; ) {
697                 struct r5dev *dev = &sh->dev[i];
698                 /* Only process blocks that are known to be uptodate */
699                 if (test_bit(R5_Wantdrain, &dev->flags))
700                         xor_srcs[count++] = dev->page;
701         }
702
703         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
704                 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
705                 ops_complete_prexor, sh);
706
707         return tx;
708 }
709
710 static struct dma_async_tx_descriptor *
711 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
712 {
713         int disks = sh->disks;
714         int i;
715
716         pr_debug("%s: stripe %llu\n", __func__,
717                 (unsigned long long)sh->sector);
718
719         for (i = disks; i--; ) {
720                 struct r5dev *dev = &sh->dev[i];
721                 struct bio *chosen;
722
723                 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
724                         struct bio *wbi;
725
726                         spin_lock(&sh->lock);
727                         chosen = dev->towrite;
728                         dev->towrite = NULL;
729                         BUG_ON(dev->written);
730                         wbi = dev->written = chosen;
731                         spin_unlock(&sh->lock);
732
733                         while (wbi && wbi->bi_sector <
734                                 dev->sector + STRIPE_SECTORS) {
735                                 tx = async_copy_data(1, wbi, dev->page,
736                                         dev->sector, tx);
737                                 wbi = r5_next_bio(wbi, dev->sector);
738                         }
739                 }
740         }
741
742         return tx;
743 }
744
745 static void ops_complete_postxor(void *stripe_head_ref)
746 {
747         struct stripe_head *sh = stripe_head_ref;
748         int disks = sh->disks, i, pd_idx = sh->pd_idx;
749
750         pr_debug("%s: stripe %llu\n", __func__,
751                 (unsigned long long)sh->sector);
752
753         for (i = disks; i--; ) {
754                 struct r5dev *dev = &sh->dev[i];
755                 if (dev->written || i == pd_idx)
756                         set_bit(R5_UPTODATE, &dev->flags);
757         }
758
759         if (sh->reconstruct_state == reconstruct_state_drain_run)
760                 sh->reconstruct_state = reconstruct_state_drain_result;
761         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
762                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
763         else {
764                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
765                 sh->reconstruct_state = reconstruct_state_result;
766         }
767
768         set_bit(STRIPE_HANDLE, &sh->state);
769         release_stripe(sh);
770 }
771
772 static void
773 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
774 {
775         /* kernel stack size limits the total number of disks */
776         int disks = sh->disks;
777         struct page *xor_srcs[disks];
778
779         int count = 0, pd_idx = sh->pd_idx, i;
780         struct page *xor_dest;
781         int prexor = 0;
782         unsigned long flags;
783
784         pr_debug("%s: stripe %llu\n", __func__,
785                 (unsigned long long)sh->sector);
786
787         /* check if prexor is active which means only process blocks
788          * that are part of a read-modify-write (written)
789          */
790         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
791                 prexor = 1;
792                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
793                 for (i = disks; i--; ) {
794                         struct r5dev *dev = &sh->dev[i];
795                         if (dev->written)
796                                 xor_srcs[count++] = dev->page;
797                 }
798         } else {
799                 xor_dest = sh->dev[pd_idx].page;
800                 for (i = disks; i--; ) {
801                         struct r5dev *dev = &sh->dev[i];
802                         if (i != pd_idx)
803                                 xor_srcs[count++] = dev->page;
804                 }
805         }
806
807         /* 1/ if we prexor'd then the dest is reused as a source
808          * 2/ if we did not prexor then we are redoing the parity
809          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
810          * for the synchronous xor case
811          */
812         flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
813                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
814
815         atomic_inc(&sh->count);
816
817         if (unlikely(count == 1)) {
818                 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
819                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
820                         flags, tx, ops_complete_postxor, sh);
821         } else
822                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
823                         flags, tx, ops_complete_postxor, sh);
824 }
825
826 static void ops_complete_check(void *stripe_head_ref)
827 {
828         struct stripe_head *sh = stripe_head_ref;
829
830         pr_debug("%s: stripe %llu\n", __func__,
831                 (unsigned long long)sh->sector);
832
833         sh->check_state = check_state_check_result;
834         set_bit(STRIPE_HANDLE, &sh->state);
835         release_stripe(sh);
836 }
837
838 static void ops_run_check(struct stripe_head *sh)
839 {
840         /* kernel stack size limits the total number of disks */
841         int disks = sh->disks;
842         struct page *xor_srcs[disks];
843         struct dma_async_tx_descriptor *tx;
844
845         int count = 0, pd_idx = sh->pd_idx, i;
846         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
847
848         pr_debug("%s: stripe %llu\n", __func__,
849                 (unsigned long long)sh->sector);
850
851         for (i = disks; i--; ) {
852                 struct r5dev *dev = &sh->dev[i];
853                 if (i != pd_idx)
854                         xor_srcs[count++] = dev->page;
855         }
856
857         tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
858                 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
859
860         atomic_inc(&sh->count);
861         tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
862                 ops_complete_check, sh);
863 }
864
865 static void raid5_run_ops(struct stripe_head *sh, unsigned long ops_request)
866 {
867         int overlap_clear = 0, i, disks = sh->disks;
868         struct dma_async_tx_descriptor *tx = NULL;
869
870         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
871                 ops_run_biofill(sh);
872                 overlap_clear++;
873         }
874
875         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
876                 tx = ops_run_compute5(sh);
877                 /* terminate the chain if postxor is not set to be run */
878                 if (tx && !test_bit(STRIPE_OP_POSTXOR, &ops_request))
879                         async_tx_ack(tx);
880         }
881
882         if (test_bit(STRIPE_OP_PREXOR, &ops_request))
883                 tx = ops_run_prexor(sh, tx);
884
885         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
886                 tx = ops_run_biodrain(sh, tx);
887                 overlap_clear++;
888         }
889
890         if (test_bit(STRIPE_OP_POSTXOR, &ops_request))
891                 ops_run_postxor(sh, tx);
892
893         if (test_bit(STRIPE_OP_CHECK, &ops_request))
894                 ops_run_check(sh);
895
896         if (overlap_clear)
897                 for (i = disks; i--; ) {
898                         struct r5dev *dev = &sh->dev[i];
899                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
900                                 wake_up(&sh->raid_conf->wait_for_overlap);
901                 }
902 }
903
904 static int grow_one_stripe(raid5_conf_t *conf)
905 {
906         struct stripe_head *sh;
907         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
908         if (!sh)
909                 return 0;
910         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
911         sh->raid_conf = conf;
912         spin_lock_init(&sh->lock);
913
914         if (grow_buffers(sh, conf->raid_disks)) {
915                 shrink_buffers(sh, conf->raid_disks);
916                 kmem_cache_free(conf->slab_cache, sh);
917                 return 0;
918         }
919         sh->disks = conf->raid_disks;
920         /* we just created an active stripe so... */
921         atomic_set(&sh->count, 1);
922         atomic_inc(&conf->active_stripes);
923         INIT_LIST_HEAD(&sh->lru);
924         release_stripe(sh);
925         return 1;
926 }
927
928 static int grow_stripes(raid5_conf_t *conf, int num)
929 {
930         struct kmem_cache *sc;
931         int devs = conf->raid_disks;
932
933         sprintf(conf->cache_name[0],
934                 "raid%d-%s", conf->level, mdname(conf->mddev));
935         sprintf(conf->cache_name[1],
936                 "raid%d-%s-alt", conf->level, mdname(conf->mddev));
937         conf->active_name = 0;
938         sc = kmem_cache_create(conf->cache_name[conf->active_name],
939                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
940                                0, 0, NULL);
941         if (!sc)
942                 return 1;
943         conf->slab_cache = sc;
944         conf->pool_size = devs;
945         while (num--)
946                 if (!grow_one_stripe(conf))
947                         return 1;
948         return 0;
949 }
950
951 static int resize_stripes(raid5_conf_t *conf, int newsize)
952 {
953         /* Make all the stripes able to hold 'newsize' devices.
954          * New slots in each stripe get 'page' set to a new page.
955          *
956          * This happens in stages:
957          * 1/ create a new kmem_cache and allocate the required number of
958          *    stripe_heads.
959          * 2/ gather all the old stripe_heads and tranfer the pages across
960          *    to the new stripe_heads.  This will have the side effect of
961          *    freezing the array as once all stripe_heads have been collected,
962          *    no IO will be possible.  Old stripe heads are freed once their
963          *    pages have been transferred over, and the old kmem_cache is
964          *    freed when all stripes are done.
965          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
966          *    we simple return a failre status - no need to clean anything up.
967          * 4/ allocate new pages for the new slots in the new stripe_heads.
968          *    If this fails, we don't bother trying the shrink the
969          *    stripe_heads down again, we just leave them as they are.
970          *    As each stripe_head is processed the new one is released into
971          *    active service.
972          *
973          * Once step2 is started, we cannot afford to wait for a write,
974          * so we use GFP_NOIO allocations.
975          */
976         struct stripe_head *osh, *nsh;
977         LIST_HEAD(newstripes);
978         struct disk_info *ndisks;
979         int err;
980         struct kmem_cache *sc;
981         int i;
982
983         if (newsize <= conf->pool_size)
984                 return 0; /* never bother to shrink */
985
986         err = md_allow_write(conf->mddev);
987         if (err)
988                 return err;
989
990         /* Step 1 */
991         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
992                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
993                                0, 0, NULL);
994         if (!sc)
995                 return -ENOMEM;
996
997         for (i = conf->max_nr_stripes; i; i--) {
998                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
999                 if (!nsh)
1000                         break;
1001
1002                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1003
1004                 nsh->raid_conf = conf;
1005                 spin_lock_init(&nsh->lock);
1006
1007                 list_add(&nsh->lru, &newstripes);
1008         }
1009         if (i) {
1010                 /* didn't get enough, give up */
1011                 while (!list_empty(&newstripes)) {
1012                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1013                         list_del(&nsh->lru);
1014                         kmem_cache_free(sc, nsh);
1015                 }
1016                 kmem_cache_destroy(sc);
1017                 return -ENOMEM;
1018         }
1019         /* Step 2 - Must use GFP_NOIO now.
1020          * OK, we have enough stripes, start collecting inactive
1021          * stripes and copying them over
1022          */
1023         list_for_each_entry(nsh, &newstripes, lru) {
1024                 spin_lock_irq(&conf->device_lock);
1025                 wait_event_lock_irq(conf->wait_for_stripe,
1026                                     !list_empty(&conf->inactive_list),
1027                                     conf->device_lock,
1028                                     unplug_slaves(conf->mddev)
1029                         );
1030                 osh = get_free_stripe(conf);
1031                 spin_unlock_irq(&conf->device_lock);
1032                 atomic_set(&nsh->count, 1);
1033                 for(i=0; i<conf->pool_size; i++)
1034                         nsh->dev[i].page = osh->dev[i].page;
1035                 for( ; i<newsize; i++)
1036                         nsh->dev[i].page = NULL;
1037                 kmem_cache_free(conf->slab_cache, osh);
1038         }
1039         kmem_cache_destroy(conf->slab_cache);
1040
1041         /* Step 3.
1042          * At this point, we are holding all the stripes so the array
1043          * is completely stalled, so now is a good time to resize
1044          * conf->disks.
1045          */
1046         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1047         if (ndisks) {
1048                 for (i=0; i<conf->raid_disks; i++)
1049                         ndisks[i] = conf->disks[i];
1050                 kfree(conf->disks);
1051                 conf->disks = ndisks;
1052         } else
1053                 err = -ENOMEM;
1054
1055         /* Step 4, return new stripes to service */
1056         while(!list_empty(&newstripes)) {
1057                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1058                 list_del_init(&nsh->lru);
1059                 for (i=conf->raid_disks; i < newsize; i++)
1060                         if (nsh->dev[i].page == NULL) {
1061                                 struct page *p = alloc_page(GFP_NOIO);
1062                                 nsh->dev[i].page = p;
1063                                 if (!p)
1064                                         err = -ENOMEM;
1065                         }
1066                 release_stripe(nsh);
1067         }
1068         /* critical section pass, GFP_NOIO no longer needed */
1069
1070         conf->slab_cache = sc;
1071         conf->active_name = 1-conf->active_name;
1072         conf->pool_size = newsize;
1073         return err;
1074 }
1075
1076 static int drop_one_stripe(raid5_conf_t *conf)
1077 {
1078         struct stripe_head *sh;
1079
1080         spin_lock_irq(&conf->device_lock);
1081         sh = get_free_stripe(conf);
1082         spin_unlock_irq(&conf->device_lock);
1083         if (!sh)
1084                 return 0;
1085         BUG_ON(atomic_read(&sh->count));
1086         shrink_buffers(sh, conf->pool_size);
1087         kmem_cache_free(conf->slab_cache, sh);
1088         atomic_dec(&conf->active_stripes);
1089         return 1;
1090 }
1091
1092 static void shrink_stripes(raid5_conf_t *conf)
1093 {
1094         while (drop_one_stripe(conf))
1095                 ;
1096
1097         if (conf->slab_cache)
1098                 kmem_cache_destroy(conf->slab_cache);
1099         conf->slab_cache = NULL;
1100 }
1101
1102 static void raid5_end_read_request(struct bio * bi, int error)
1103 {
1104         struct stripe_head *sh = bi->bi_private;
1105         raid5_conf_t *conf = sh->raid_conf;
1106         int disks = sh->disks, i;
1107         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1108         char b[BDEVNAME_SIZE];
1109         mdk_rdev_t *rdev;
1110
1111
1112         for (i=0 ; i<disks; i++)
1113                 if (bi == &sh->dev[i].req)
1114                         break;
1115
1116         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1117                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1118                 uptodate);
1119         if (i == disks) {
1120                 BUG();
1121                 return;
1122         }
1123
1124         if (uptodate) {
1125                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1126                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1127                         rdev = conf->disks[i].rdev;
1128                         printk_rl(KERN_INFO "raid5:%s: read error corrected"
1129                                   " (%lu sectors at %llu on %s)\n",
1130                                   mdname(conf->mddev), STRIPE_SECTORS,
1131                                   (unsigned long long)(sh->sector
1132                                                        + rdev->data_offset),
1133                                   bdevname(rdev->bdev, b));
1134                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1135                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1136                 }
1137                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1138                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1139         } else {
1140                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1141                 int retry = 0;
1142                 rdev = conf->disks[i].rdev;
1143
1144                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1145                 atomic_inc(&rdev->read_errors);
1146                 if (conf->mddev->degraded)
1147                         printk_rl(KERN_WARNING
1148                                   "raid5:%s: read error not correctable "
1149                                   "(sector %llu on %s).\n",
1150                                   mdname(conf->mddev),
1151                                   (unsigned long long)(sh->sector
1152                                                        + rdev->data_offset),
1153                                   bdn);
1154                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1155                         /* Oh, no!!! */
1156                         printk_rl(KERN_WARNING
1157                                   "raid5:%s: read error NOT corrected!! "
1158                                   "(sector %llu on %s).\n",
1159                                   mdname(conf->mddev),
1160                                   (unsigned long long)(sh->sector
1161                                                        + rdev->data_offset),
1162                                   bdn);
1163                 else if (atomic_read(&rdev->read_errors)
1164                          > conf->max_nr_stripes)
1165                         printk(KERN_WARNING
1166                                "raid5:%s: Too many read errors, failing device %s.\n",
1167                                mdname(conf->mddev), bdn);
1168                 else
1169                         retry = 1;
1170                 if (retry)
1171                         set_bit(R5_ReadError, &sh->dev[i].flags);
1172                 else {
1173                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1174                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1175                         md_error(conf->mddev, rdev);
1176                 }
1177         }
1178         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1179         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1180         set_bit(STRIPE_HANDLE, &sh->state);
1181         release_stripe(sh);
1182 }
1183
1184 static void raid5_end_write_request(struct bio *bi, int error)
1185 {
1186         struct stripe_head *sh = bi->bi_private;
1187         raid5_conf_t *conf = sh->raid_conf;
1188         int disks = sh->disks, i;
1189         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1190
1191         for (i=0 ; i<disks; i++)
1192                 if (bi == &sh->dev[i].req)
1193                         break;
1194
1195         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1196                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1197                 uptodate);
1198         if (i == disks) {
1199                 BUG();
1200                 return;
1201         }
1202
1203         if (!uptodate)
1204                 md_error(conf->mddev, conf->disks[i].rdev);
1205
1206         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1207         
1208         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1209         set_bit(STRIPE_HANDLE, &sh->state);
1210         release_stripe(sh);
1211 }
1212
1213
1214 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1215         
1216 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1217 {
1218         struct r5dev *dev = &sh->dev[i];
1219
1220         bio_init(&dev->req);
1221         dev->req.bi_io_vec = &dev->vec;
1222         dev->req.bi_vcnt++;
1223         dev->req.bi_max_vecs++;
1224         dev->vec.bv_page = dev->page;
1225         dev->vec.bv_len = STRIPE_SIZE;
1226         dev->vec.bv_offset = 0;
1227
1228         dev->req.bi_sector = sh->sector;
1229         dev->req.bi_private = sh;
1230
1231         dev->flags = 0;
1232         dev->sector = compute_blocknr(sh, i, previous);
1233 }
1234
1235 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1236 {
1237         char b[BDEVNAME_SIZE];
1238         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1239         pr_debug("raid5: error called\n");
1240
1241         if (!test_bit(Faulty, &rdev->flags)) {
1242                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1243                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1244                         unsigned long flags;
1245                         spin_lock_irqsave(&conf->device_lock, flags);
1246                         mddev->degraded++;
1247                         spin_unlock_irqrestore(&conf->device_lock, flags);
1248                         /*
1249                          * if recovery was running, make sure it aborts.
1250                          */
1251                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1252                 }
1253                 set_bit(Faulty, &rdev->flags);
1254                 printk(KERN_ALERT
1255                        "raid5: Disk failure on %s, disabling device.\n"
1256                        "raid5: Operation continuing on %d devices.\n",
1257                        bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1258         }
1259 }
1260
1261 /*
1262  * Input: a 'big' sector number,
1263  * Output: index of the data and parity disk, and the sector # in them.
1264  */
1265 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1266                                      int previous, int *dd_idx,
1267                                      struct stripe_head *sh)
1268 {
1269         long stripe;
1270         unsigned long chunk_number;
1271         unsigned int chunk_offset;
1272         int pd_idx, qd_idx;
1273         int ddf_layout = 0;
1274         sector_t new_sector;
1275         int algorithm = previous ? conf->prev_algo
1276                                  : conf->algorithm;
1277         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1278                                          : conf->chunk_sectors;
1279         int raid_disks = previous ? conf->previous_raid_disks
1280                                   : conf->raid_disks;
1281         int data_disks = raid_disks - conf->max_degraded;
1282
1283         /* First compute the information on this sector */
1284
1285         /*
1286          * Compute the chunk number and the sector offset inside the chunk
1287          */
1288         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1289         chunk_number = r_sector;
1290         BUG_ON(r_sector != chunk_number);
1291
1292         /*
1293          * Compute the stripe number
1294          */
1295         stripe = chunk_number / data_disks;
1296
1297         /*
1298          * Compute the data disk and parity disk indexes inside the stripe
1299          */
1300         *dd_idx = chunk_number % data_disks;
1301
1302         /*
1303          * Select the parity disk based on the user selected algorithm.
1304          */
1305         pd_idx = qd_idx = ~0;
1306         switch(conf->level) {
1307         case 4:
1308                 pd_idx = data_disks;
1309                 break;
1310         case 5:
1311                 switch (algorithm) {
1312                 case ALGORITHM_LEFT_ASYMMETRIC:
1313                         pd_idx = data_disks - stripe % raid_disks;
1314                         if (*dd_idx >= pd_idx)
1315                                 (*dd_idx)++;
1316                         break;
1317                 case ALGORITHM_RIGHT_ASYMMETRIC:
1318                         pd_idx = stripe % raid_disks;
1319                         if (*dd_idx >= pd_idx)
1320                                 (*dd_idx)++;
1321                         break;
1322                 case ALGORITHM_LEFT_SYMMETRIC:
1323                         pd_idx = data_disks - stripe % raid_disks;
1324                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1325                         break;
1326                 case ALGORITHM_RIGHT_SYMMETRIC:
1327                         pd_idx = stripe % raid_disks;
1328                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1329                         break;
1330                 case ALGORITHM_PARITY_0:
1331                         pd_idx = 0;
1332                         (*dd_idx)++;
1333                         break;
1334                 case ALGORITHM_PARITY_N:
1335                         pd_idx = data_disks;
1336                         break;
1337                 default:
1338                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1339                                 algorithm);
1340                         BUG();
1341                 }
1342                 break;
1343         case 6:
1344
1345                 switch (algorithm) {
1346                 case ALGORITHM_LEFT_ASYMMETRIC:
1347                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1348                         qd_idx = pd_idx + 1;
1349                         if (pd_idx == raid_disks-1) {
1350                                 (*dd_idx)++;    /* Q D D D P */
1351                                 qd_idx = 0;
1352                         } else if (*dd_idx >= pd_idx)
1353                                 (*dd_idx) += 2; /* D D P Q D */
1354                         break;
1355                 case ALGORITHM_RIGHT_ASYMMETRIC:
1356                         pd_idx = stripe % raid_disks;
1357                         qd_idx = pd_idx + 1;
1358                         if (pd_idx == raid_disks-1) {
1359                                 (*dd_idx)++;    /* Q D D D P */
1360                                 qd_idx = 0;
1361                         } else if (*dd_idx >= pd_idx)
1362                                 (*dd_idx) += 2; /* D D P Q D */
1363                         break;
1364                 case ALGORITHM_LEFT_SYMMETRIC:
1365                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1366                         qd_idx = (pd_idx + 1) % raid_disks;
1367                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1368                         break;
1369                 case ALGORITHM_RIGHT_SYMMETRIC:
1370                         pd_idx = stripe % raid_disks;
1371                         qd_idx = (pd_idx + 1) % raid_disks;
1372                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1373                         break;
1374
1375                 case ALGORITHM_PARITY_0:
1376                         pd_idx = 0;
1377                         qd_idx = 1;
1378                         (*dd_idx) += 2;
1379                         break;
1380                 case ALGORITHM_PARITY_N:
1381                         pd_idx = data_disks;
1382                         qd_idx = data_disks + 1;
1383                         break;
1384
1385                 case ALGORITHM_ROTATING_ZERO_RESTART:
1386                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
1387                          * of blocks for computing Q is different.
1388                          */
1389                         pd_idx = stripe % raid_disks;
1390                         qd_idx = pd_idx + 1;
1391                         if (pd_idx == raid_disks-1) {
1392                                 (*dd_idx)++;    /* Q D D D P */
1393                                 qd_idx = 0;
1394                         } else if (*dd_idx >= pd_idx)
1395                                 (*dd_idx) += 2; /* D D P Q D */
1396                         ddf_layout = 1;
1397                         break;
1398
1399                 case ALGORITHM_ROTATING_N_RESTART:
1400                         /* Same a left_asymmetric, by first stripe is
1401                          * D D D P Q  rather than
1402                          * Q D D D P
1403                          */
1404                         pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1405                         qd_idx = pd_idx + 1;
1406                         if (pd_idx == raid_disks-1) {
1407                                 (*dd_idx)++;    /* Q D D D P */
1408                                 qd_idx = 0;
1409                         } else if (*dd_idx >= pd_idx)
1410                                 (*dd_idx) += 2; /* D D P Q D */
1411                         ddf_layout = 1;
1412                         break;
1413
1414                 case ALGORITHM_ROTATING_N_CONTINUE:
1415                         /* Same as left_symmetric but Q is before P */
1416                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1417                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1418                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1419                         ddf_layout = 1;
1420                         break;
1421
1422                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1423                         /* RAID5 left_asymmetric, with Q on last device */
1424                         pd_idx = data_disks - stripe % (raid_disks-1);
1425                         if (*dd_idx >= pd_idx)
1426                                 (*dd_idx)++;
1427                         qd_idx = raid_disks - 1;
1428                         break;
1429
1430                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1431                         pd_idx = stripe % (raid_disks-1);
1432                         if (*dd_idx >= pd_idx)
1433                                 (*dd_idx)++;
1434                         qd_idx = raid_disks - 1;
1435                         break;
1436
1437                 case ALGORITHM_LEFT_SYMMETRIC_6:
1438                         pd_idx = data_disks - stripe % (raid_disks-1);
1439                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1440                         qd_idx = raid_disks - 1;
1441                         break;
1442
1443                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1444                         pd_idx = stripe % (raid_disks-1);
1445                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1446                         qd_idx = raid_disks - 1;
1447                         break;
1448
1449                 case ALGORITHM_PARITY_0_6:
1450                         pd_idx = 0;
1451                         (*dd_idx)++;
1452                         qd_idx = raid_disks - 1;
1453                         break;
1454
1455
1456                 default:
1457                         printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1458                                algorithm);
1459                         BUG();
1460                 }
1461                 break;
1462         }
1463
1464         if (sh) {
1465                 sh->pd_idx = pd_idx;
1466                 sh->qd_idx = qd_idx;
1467                 sh->ddf_layout = ddf_layout;
1468         }
1469         /*
1470          * Finally, compute the new sector number
1471          */
1472         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1473         return new_sector;
1474 }
1475
1476
1477 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1478 {
1479         raid5_conf_t *conf = sh->raid_conf;
1480         int raid_disks = sh->disks;
1481         int data_disks = raid_disks - conf->max_degraded;
1482         sector_t new_sector = sh->sector, check;
1483         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1484                                          : conf->chunk_sectors;
1485         int algorithm = previous ? conf->prev_algo
1486                                  : conf->algorithm;
1487         sector_t stripe;
1488         int chunk_offset;
1489         int chunk_number, dummy1, dd_idx = i;
1490         sector_t r_sector;
1491         struct stripe_head sh2;
1492
1493
1494         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1495         stripe = new_sector;
1496         BUG_ON(new_sector != stripe);
1497
1498         if (i == sh->pd_idx)
1499                 return 0;
1500         switch(conf->level) {
1501         case 4: break;
1502         case 5:
1503                 switch (algorithm) {
1504                 case ALGORITHM_LEFT_ASYMMETRIC:
1505                 case ALGORITHM_RIGHT_ASYMMETRIC:
1506                         if (i > sh->pd_idx)
1507                                 i--;
1508                         break;
1509                 case ALGORITHM_LEFT_SYMMETRIC:
1510                 case ALGORITHM_RIGHT_SYMMETRIC:
1511                         if (i < sh->pd_idx)
1512                                 i += raid_disks;
1513                         i -= (sh->pd_idx + 1);
1514                         break;
1515                 case ALGORITHM_PARITY_0:
1516                         i -= 1;
1517                         break;
1518                 case ALGORITHM_PARITY_N:
1519                         break;
1520                 default:
1521                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1522                                algorithm);
1523                         BUG();
1524                 }
1525                 break;
1526         case 6:
1527                 if (i == sh->qd_idx)
1528                         return 0; /* It is the Q disk */
1529                 switch (algorithm) {
1530                 case ALGORITHM_LEFT_ASYMMETRIC:
1531                 case ALGORITHM_RIGHT_ASYMMETRIC:
1532                 case ALGORITHM_ROTATING_ZERO_RESTART:
1533                 case ALGORITHM_ROTATING_N_RESTART:
1534                         if (sh->pd_idx == raid_disks-1)
1535                                 i--;    /* Q D D D P */
1536                         else if (i > sh->pd_idx)
1537                                 i -= 2; /* D D P Q D */
1538                         break;
1539                 case ALGORITHM_LEFT_SYMMETRIC:
1540                 case ALGORITHM_RIGHT_SYMMETRIC:
1541                         if (sh->pd_idx == raid_disks-1)
1542                                 i--; /* Q D D D P */
1543                         else {
1544                                 /* D D P Q D */
1545                                 if (i < sh->pd_idx)
1546                                         i += raid_disks;
1547                                 i -= (sh->pd_idx + 2);
1548                         }
1549                         break;
1550                 case ALGORITHM_PARITY_0:
1551                         i -= 2;
1552                         break;
1553                 case ALGORITHM_PARITY_N:
1554                         break;
1555                 case ALGORITHM_ROTATING_N_CONTINUE:
1556                         if (sh->pd_idx == 0)
1557                                 i--;    /* P D D D Q */
1558                         else if (i > sh->pd_idx)
1559                                 i -= 2; /* D D Q P D */
1560                         break;
1561                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1562                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1563                         if (i > sh->pd_idx)
1564                                 i--;
1565                         break;
1566                 case ALGORITHM_LEFT_SYMMETRIC_6:
1567                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1568                         if (i < sh->pd_idx)
1569                                 i += data_disks + 1;
1570                         i -= (sh->pd_idx + 1);
1571                         break;
1572                 case ALGORITHM_PARITY_0_6:
1573                         i -= 1;
1574                         break;
1575                 default:
1576                         printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1577                                algorithm);
1578                         BUG();
1579                 }
1580                 break;
1581         }
1582
1583         chunk_number = stripe * data_disks + i;
1584         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1585
1586         check = raid5_compute_sector(conf, r_sector,
1587                                      previous, &dummy1, &sh2);
1588         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1589                 || sh2.qd_idx != sh->qd_idx) {
1590                 printk(KERN_ERR "compute_blocknr: map not correct\n");
1591                 return 0;
1592         }
1593         return r_sector;
1594 }
1595
1596
1597
1598 /*
1599  * Copy data between a page in the stripe cache, and one or more bion
1600  * The page could align with the middle of the bio, or there could be
1601  * several bion, each with several bio_vecs, which cover part of the page
1602  * Multiple bion are linked together on bi_next.  There may be extras
1603  * at the end of this list.  We ignore them.
1604  */
1605 static void copy_data(int frombio, struct bio *bio,
1606                      struct page *page,
1607                      sector_t sector)
1608 {
1609         char *pa = page_address(page);
1610         struct bio_vec *bvl;
1611         int i;
1612         int page_offset;
1613
1614         if (bio->bi_sector >= sector)
1615                 page_offset = (signed)(bio->bi_sector - sector) * 512;
1616         else
1617                 page_offset = (signed)(sector - bio->bi_sector) * -512;
1618         bio_for_each_segment(bvl, bio, i) {
1619                 int len = bio_iovec_idx(bio,i)->bv_len;
1620                 int clen;
1621                 int b_offset = 0;
1622
1623                 if (page_offset < 0) {
1624                         b_offset = -page_offset;
1625                         page_offset += b_offset;
1626                         len -= b_offset;
1627                 }
1628
1629                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1630                         clen = STRIPE_SIZE - page_offset;
1631                 else clen = len;
1632
1633                 if (clen > 0) {
1634                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1635                         if (frombio)
1636                                 memcpy(pa+page_offset, ba+b_offset, clen);
1637                         else
1638                                 memcpy(ba+b_offset, pa+page_offset, clen);
1639                         __bio_kunmap_atomic(ba, KM_USER0);
1640                 }
1641                 if (clen < len) /* hit end of page */
1642                         break;
1643                 page_offset +=  len;
1644         }
1645 }
1646
1647 #define check_xor()     do {                                              \
1648                                 if (count == MAX_XOR_BLOCKS) {            \
1649                                 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1650                                 count = 0;                                \
1651                            }                                              \
1652                         } while(0)
1653
1654 static void compute_parity6(struct stripe_head *sh, int method)
1655 {
1656         raid5_conf_t *conf = sh->raid_conf;
1657         int i, pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1658         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1659         struct bio *chosen;
1660         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1661         void *ptrs[syndrome_disks+2];
1662
1663         pd_idx = sh->pd_idx;
1664         qd_idx = sh->qd_idx;
1665         d0_idx = raid6_d0(sh);
1666
1667         pr_debug("compute_parity, stripe %llu, method %d\n",
1668                 (unsigned long long)sh->sector, method);
1669
1670         switch(method) {
1671         case READ_MODIFY_WRITE:
1672                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
1673         case RECONSTRUCT_WRITE:
1674                 for (i= disks; i-- ;)
1675                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1676                                 chosen = sh->dev[i].towrite;
1677                                 sh->dev[i].towrite = NULL;
1678
1679                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1680                                         wake_up(&conf->wait_for_overlap);
1681
1682                                 BUG_ON(sh->dev[i].written);
1683                                 sh->dev[i].written = chosen;
1684                         }
1685                 break;
1686         case CHECK_PARITY:
1687                 BUG();          /* Not implemented yet */
1688         }
1689
1690         for (i = disks; i--;)
1691                 if (sh->dev[i].written) {
1692                         sector_t sector = sh->dev[i].sector;
1693                         struct bio *wbi = sh->dev[i].written;
1694                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1695                                 copy_data(1, wbi, sh->dev[i].page, sector);
1696                                 wbi = r5_next_bio(wbi, sector);
1697                         }
1698
1699                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1700                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1701                 }
1702
1703         /* Note that unlike RAID-5, the ordering of the disks matters greatly.*/
1704
1705         for (i = 0; i < disks; i++)
1706                 ptrs[i] = (void *)raid6_empty_zero_page;
1707
1708         count = 0;
1709         i = d0_idx;
1710         do {
1711                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1712
1713                 ptrs[slot] = page_address(sh->dev[i].page);
1714                 if (slot < syndrome_disks &&
1715                     !test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
1716                         printk(KERN_ERR "block %d/%d not uptodate "
1717                                "on parity calc\n", i, count);
1718                         BUG();
1719                 }
1720
1721                 i = raid6_next_disk(i, disks);
1722         } while (i != d0_idx);
1723         BUG_ON(count != syndrome_disks);
1724
1725         raid6_call.gen_syndrome(syndrome_disks+2, STRIPE_SIZE, ptrs);
1726
1727         switch(method) {
1728         case RECONSTRUCT_WRITE:
1729                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1730                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1731                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1732                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
1733                 break;
1734         case UPDATE_PARITY:
1735                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1736                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1737                 break;
1738         }
1739 }
1740
1741
1742 /* Compute one missing block */
1743 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1744 {
1745         int i, count, disks = sh->disks;
1746         void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1747         int qd_idx = sh->qd_idx;
1748
1749         pr_debug("compute_block_1, stripe %llu, idx %d\n",
1750                 (unsigned long long)sh->sector, dd_idx);
1751
1752         if ( dd_idx == qd_idx ) {
1753                 /* We're actually computing the Q drive */
1754                 compute_parity6(sh, UPDATE_PARITY);
1755         } else {
1756                 dest = page_address(sh->dev[dd_idx].page);
1757                 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1758                 count = 0;
1759                 for (i = disks ; i--; ) {
1760                         if (i == dd_idx || i == qd_idx)
1761                                 continue;
1762                         p = page_address(sh->dev[i].page);
1763                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1764                                 ptr[count++] = p;
1765                         else
1766                                 printk("compute_block() %d, stripe %llu, %d"
1767                                        " not present\n", dd_idx,
1768                                        (unsigned long long)sh->sector, i);
1769
1770                         check_xor();
1771                 }
1772                 if (count)
1773                         xor_blocks(count, STRIPE_SIZE, dest, ptr);
1774                 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1775                 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1776         }
1777 }
1778
1779 /* Compute two missing blocks */
1780 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1781 {
1782         int i, count, disks = sh->disks;
1783         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1784         int d0_idx = raid6_d0(sh);
1785         int faila = -1, failb = -1;
1786         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1787         void *ptrs[syndrome_disks+2];
1788
1789         for (i = 0; i < disks ; i++)
1790                 ptrs[i] = (void *)raid6_empty_zero_page;
1791         count = 0;
1792         i = d0_idx;
1793         do {
1794                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1795
1796                 ptrs[slot] = page_address(sh->dev[i].page);
1797
1798                 if (i == dd_idx1)
1799                         faila = slot;
1800                 if (i == dd_idx2)
1801                         failb = slot;
1802                 i = raid6_next_disk(i, disks);
1803         } while (i != d0_idx);
1804         BUG_ON(count != syndrome_disks);
1805
1806         BUG_ON(faila == failb);
1807         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1808
1809         pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1810                  (unsigned long long)sh->sector, dd_idx1, dd_idx2,
1811                  faila, failb);
1812
1813         if (failb == syndrome_disks+1) {
1814                 /* Q disk is one of the missing disks */
1815                 if (faila == syndrome_disks) {
1816                         /* Missing P+Q, just recompute */
1817                         compute_parity6(sh, UPDATE_PARITY);
1818                         return;
1819                 } else {
1820                         /* We're missing D+Q; recompute D from P */
1821                         compute_block_1(sh, ((dd_idx1 == sh->qd_idx) ?
1822                                              dd_idx2 : dd_idx1),
1823                                         0);
1824                         compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1825                         return;
1826                 }
1827         }
1828
1829         /* We're missing D+P or D+D; */
1830         if (failb == syndrome_disks) {
1831                 /* We're missing D+P. */
1832                 raid6_datap_recov(syndrome_disks+2, STRIPE_SIZE, faila, ptrs);
1833         } else {
1834                 /* We're missing D+D. */
1835                 raid6_2data_recov(syndrome_disks+2, STRIPE_SIZE, faila, failb,
1836                                   ptrs);
1837         }
1838
1839         /* Both the above update both missing blocks */
1840         set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1841         set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1842 }
1843
1844 static void
1845 schedule_reconstruction5(struct stripe_head *sh, struct stripe_head_state *s,
1846                          int rcw, int expand)
1847 {
1848         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1849
1850         if (rcw) {
1851                 /* if we are not expanding this is a proper write request, and
1852                  * there will be bios with new data to be drained into the
1853                  * stripe cache
1854                  */
1855                 if (!expand) {
1856                         sh->reconstruct_state = reconstruct_state_drain_run;
1857                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1858                 } else
1859                         sh->reconstruct_state = reconstruct_state_run;
1860
1861                 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1862
1863                 for (i = disks; i--; ) {
1864                         struct r5dev *dev = &sh->dev[i];
1865
1866                         if (dev->towrite) {
1867                                 set_bit(R5_LOCKED, &dev->flags);
1868                                 set_bit(R5_Wantdrain, &dev->flags);
1869                                 if (!expand)
1870                                         clear_bit(R5_UPTODATE, &dev->flags);
1871                                 s->locked++;
1872                         }
1873                 }
1874                 if (s->locked + 1 == disks)
1875                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1876                                 atomic_inc(&sh->raid_conf->pending_full_writes);
1877         } else {
1878                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1879                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1880
1881                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
1882                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
1883                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1884                 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1885
1886                 for (i = disks; i--; ) {
1887                         struct r5dev *dev = &sh->dev[i];
1888                         if (i == pd_idx)
1889                                 continue;
1890
1891                         if (dev->towrite &&
1892                             (test_bit(R5_UPTODATE, &dev->flags) ||
1893                              test_bit(R5_Wantcompute, &dev->flags))) {
1894                                 set_bit(R5_Wantdrain, &dev->flags);
1895                                 set_bit(R5_LOCKED, &dev->flags);
1896                                 clear_bit(R5_UPTODATE, &dev->flags);
1897                                 s->locked++;
1898                         }
1899                 }
1900         }
1901
1902         /* keep the parity disk locked while asynchronous operations
1903          * are in flight
1904          */
1905         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1906         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1907         s->locked++;
1908
1909         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
1910                 __func__, (unsigned long long)sh->sector,
1911                 s->locked, s->ops_request);
1912 }
1913
1914 /*
1915  * Each stripe/dev can have one or more bion attached.
1916  * toread/towrite point to the first in a chain.
1917  * The bi_next chain must be in order.
1918  */
1919 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1920 {
1921         struct bio **bip;
1922         raid5_conf_t *conf = sh->raid_conf;
1923         int firstwrite=0;
1924
1925         pr_debug("adding bh b#%llu to stripe s#%llu\n",
1926                 (unsigned long long)bi->bi_sector,
1927                 (unsigned long long)sh->sector);
1928
1929
1930         spin_lock(&sh->lock);
1931         spin_lock_irq(&conf->device_lock);
1932         if (forwrite) {
1933                 bip = &sh->dev[dd_idx].towrite;
1934                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1935                         firstwrite = 1;
1936         } else
1937                 bip = &sh->dev[dd_idx].toread;
1938         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1939                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1940                         goto overlap;
1941                 bip = & (*bip)->bi_next;
1942         }
1943         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1944                 goto overlap;
1945
1946         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1947         if (*bip)
1948                 bi->bi_next = *bip;
1949         *bip = bi;
1950         bi->bi_phys_segments++;
1951         spin_unlock_irq(&conf->device_lock);
1952         spin_unlock(&sh->lock);
1953
1954         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1955                 (unsigned long long)bi->bi_sector,
1956                 (unsigned long long)sh->sector, dd_idx);
1957
1958         if (conf->mddev->bitmap && firstwrite) {
1959                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1960                                   STRIPE_SECTORS, 0);
1961                 sh->bm_seq = conf->seq_flush+1;
1962                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1963         }
1964
1965         if (forwrite) {
1966                 /* check if page is covered */
1967                 sector_t sector = sh->dev[dd_idx].sector;
1968                 for (bi=sh->dev[dd_idx].towrite;
1969                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1970                              bi && bi->bi_sector <= sector;
1971                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1972                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1973                                 sector = bi->bi_sector + (bi->bi_size>>9);
1974                 }
1975                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1976                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1977         }
1978         return 1;
1979
1980  overlap:
1981         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1982         spin_unlock_irq(&conf->device_lock);
1983         spin_unlock(&sh->lock);
1984         return 0;
1985 }
1986
1987 static void end_reshape(raid5_conf_t *conf);
1988
1989 static int page_is_zero(struct page *p)
1990 {
1991         char *a = page_address(p);
1992         return ((*(u32*)a) == 0 &&
1993                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1994 }
1995
1996 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
1997                             struct stripe_head *sh)
1998 {
1999         int sectors_per_chunk =
2000                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2001         int dd_idx;
2002         int chunk_offset = sector_div(stripe, sectors_per_chunk);
2003         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2004
2005         raid5_compute_sector(conf,
2006                              stripe * (disks - conf->max_degraded)
2007                              *sectors_per_chunk + chunk_offset,
2008                              previous,
2009                              &dd_idx, sh);
2010 }
2011
2012 static void
2013 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2014                                 struct stripe_head_state *s, int disks,
2015                                 struct bio **return_bi)
2016 {
2017         int i;
2018         for (i = disks; i--; ) {
2019                 struct bio *bi;
2020                 int bitmap_end = 0;
2021
2022                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2023                         mdk_rdev_t *rdev;
2024                         rcu_read_lock();
2025                         rdev = rcu_dereference(conf->disks[i].rdev);
2026                         if (rdev && test_bit(In_sync, &rdev->flags))
2027                                 /* multiple read failures in one stripe */
2028                                 md_error(conf->mddev, rdev);
2029                         rcu_read_unlock();
2030                 }
2031                 spin_lock_irq(&conf->device_lock);
2032                 /* fail all writes first */
2033                 bi = sh->dev[i].towrite;
2034                 sh->dev[i].towrite = NULL;
2035                 if (bi) {
2036                         s->to_write--;
2037                         bitmap_end = 1;
2038                 }
2039
2040                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2041                         wake_up(&conf->wait_for_overlap);
2042
2043                 while (bi && bi->bi_sector <
2044                         sh->dev[i].sector + STRIPE_SECTORS) {
2045                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2046                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2047                         if (!raid5_dec_bi_phys_segments(bi)) {
2048                                 md_write_end(conf->mddev);
2049                                 bi->bi_next = *return_bi;
2050                                 *return_bi = bi;
2051                         }
2052                         bi = nextbi;
2053                 }
2054                 /* and fail all 'written' */
2055                 bi = sh->dev[i].written;
2056                 sh->dev[i].written = NULL;
2057                 if (bi) bitmap_end = 1;
2058                 while (bi && bi->bi_sector <
2059                        sh->dev[i].sector + STRIPE_SECTORS) {
2060                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2061                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2062                         if (!raid5_dec_bi_phys_segments(bi)) {
2063                                 md_write_end(conf->mddev);
2064                                 bi->bi_next = *return_bi;
2065                                 *return_bi = bi;
2066                         }
2067                         bi = bi2;
2068                 }
2069
2070                 /* fail any reads if this device is non-operational and
2071                  * the data has not reached the cache yet.
2072                  */
2073                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2074                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2075                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
2076                         bi = sh->dev[i].toread;
2077                         sh->dev[i].toread = NULL;
2078                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2079                                 wake_up(&conf->wait_for_overlap);
2080                         if (bi) s->to_read--;
2081                         while (bi && bi->bi_sector <
2082                                sh->dev[i].sector + STRIPE_SECTORS) {
2083                                 struct bio *nextbi =
2084                                         r5_next_bio(bi, sh->dev[i].sector);
2085                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2086                                 if (!raid5_dec_bi_phys_segments(bi)) {
2087                                         bi->bi_next = *return_bi;
2088                                         *return_bi = bi;
2089                                 }
2090                                 bi = nextbi;
2091                         }
2092                 }
2093                 spin_unlock_irq(&conf->device_lock);
2094                 if (bitmap_end)
2095                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2096                                         STRIPE_SECTORS, 0, 0);
2097         }
2098
2099         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2100                 if (atomic_dec_and_test(&conf->pending_full_writes))
2101                         md_wakeup_thread(conf->mddev->thread);
2102 }
2103
2104 /* fetch_block5 - checks the given member device to see if its data needs
2105  * to be read or computed to satisfy a request.
2106  *
2107  * Returns 1 when no more member devices need to be checked, otherwise returns
2108  * 0 to tell the loop in handle_stripe_fill5 to continue
2109  */
2110 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2111                         int disk_idx, int disks)
2112 {
2113         struct r5dev *dev = &sh->dev[disk_idx];
2114         struct r5dev *failed_dev = &sh->dev[s->failed_num];
2115
2116         /* is the data in this block needed, and can we get it? */
2117         if (!test_bit(R5_LOCKED, &dev->flags) &&
2118             !test_bit(R5_UPTODATE, &dev->flags) &&
2119             (dev->toread ||
2120              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2121              s->syncing || s->expanding ||
2122              (s->failed &&
2123               (failed_dev->toread ||
2124                (failed_dev->towrite &&
2125                 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2126                 /* We would like to get this block, possibly by computing it,
2127                  * otherwise read it if the backing disk is insync
2128                  */
2129                 if ((s->uptodate == disks - 1) &&
2130                     (s->failed && disk_idx == s->failed_num)) {
2131                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2132                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2133                         set_bit(R5_Wantcompute, &dev->flags);
2134                         sh->ops.target = disk_idx;
2135                         s->req_compute = 1;
2136                         /* Careful: from this point on 'uptodate' is in the eye
2137                          * of raid5_run_ops which services 'compute' operations
2138                          * before writes. R5_Wantcompute flags a block that will
2139                          * be R5_UPTODATE by the time it is needed for a
2140                          * subsequent operation.
2141                          */
2142                         s->uptodate++;
2143                         return 1; /* uptodate + compute == disks */
2144                 } else if (test_bit(R5_Insync, &dev->flags)) {
2145                         set_bit(R5_LOCKED, &dev->flags);
2146                         set_bit(R5_Wantread, &dev->flags);
2147                         s->locked++;
2148                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2149                                 s->syncing);
2150                 }
2151         }
2152
2153         return 0;
2154 }
2155
2156 /**
2157  * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2158  */
2159 static void handle_stripe_fill5(struct stripe_head *sh,
2160                         struct stripe_head_state *s, int disks)
2161 {
2162         int i;
2163
2164         /* look for blocks to read/compute, skip this if a compute
2165          * is already in flight, or if the stripe contents are in the
2166          * midst of changing due to a write
2167          */
2168         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2169             !sh->reconstruct_state)
2170                 for (i = disks; i--; )
2171                         if (fetch_block5(sh, s, i, disks))
2172                                 break;
2173         set_bit(STRIPE_HANDLE, &sh->state);
2174 }
2175
2176 static void handle_stripe_fill6(struct stripe_head *sh,
2177                         struct stripe_head_state *s, struct r6_state *r6s,
2178                         int disks)
2179 {
2180         int i;
2181         for (i = disks; i--; ) {
2182                 struct r5dev *dev = &sh->dev[i];
2183                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2184                     !test_bit(R5_UPTODATE, &dev->flags) &&
2185                     (dev->toread || (dev->towrite &&
2186                      !test_bit(R5_OVERWRITE, &dev->flags)) ||
2187                      s->syncing || s->expanding ||
2188                      (s->failed >= 1 &&
2189                       (sh->dev[r6s->failed_num[0]].toread ||
2190                        s->to_write)) ||
2191                      (s->failed >= 2 &&
2192                       (sh->dev[r6s->failed_num[1]].toread ||
2193                        s->to_write)))) {
2194                         /* we would like to get this block, possibly
2195                          * by computing it, but we might not be able to
2196                          */
2197                         if ((s->uptodate == disks - 1) &&
2198                             (s->failed && (i == r6s->failed_num[0] ||
2199                                            i == r6s->failed_num[1]))) {
2200                                 pr_debug("Computing stripe %llu block %d\n",
2201                                        (unsigned long long)sh->sector, i);
2202                                 compute_block_1(sh, i, 0);
2203                                 s->uptodate++;
2204                         } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2205                                 /* Computing 2-failure is *very* expensive; only
2206                                  * do it if failed >= 2
2207                                  */
2208                                 int other;
2209                                 for (other = disks; other--; ) {
2210                                         if (other == i)
2211                                                 continue;
2212                                         if (!test_bit(R5_UPTODATE,
2213                                               &sh->dev[other].flags))
2214                                                 break;
2215                                 }
2216                                 BUG_ON(other < 0);
2217                                 pr_debug("Computing stripe %llu blocks %d,%d\n",
2218                                        (unsigned long long)sh->sector,
2219                                        i, other);
2220                                 compute_block_2(sh, i, other);
2221                                 s->uptodate += 2;
2222                         } else if (test_bit(R5_Insync, &dev->flags)) {
2223                                 set_bit(R5_LOCKED, &dev->flags);
2224                                 set_bit(R5_Wantread, &dev->flags);
2225                                 s->locked++;
2226                                 pr_debug("Reading block %d (sync=%d)\n",
2227                                         i, s->syncing);
2228                         }
2229                 }
2230         }
2231         set_bit(STRIPE_HANDLE, &sh->state);
2232 }
2233
2234
2235 /* handle_stripe_clean_event
2236  * any written block on an uptodate or failed drive can be returned.
2237  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2238  * never LOCKED, so we don't need to test 'failed' directly.
2239  */
2240 static void handle_stripe_clean_event(raid5_conf_t *conf,
2241         struct stripe_head *sh, int disks, struct bio **return_bi)
2242 {
2243         int i;
2244         struct r5dev *dev;
2245
2246         for (i = disks; i--; )
2247                 if (sh->dev[i].written) {
2248                         dev = &sh->dev[i];
2249                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2250                                 test_bit(R5_UPTODATE, &dev->flags)) {
2251                                 /* We can return any write requests */
2252                                 struct bio *wbi, *wbi2;
2253                                 int bitmap_end = 0;
2254                                 pr_debug("Return write for disc %d\n", i);
2255                                 spin_lock_irq(&conf->device_lock);
2256                                 wbi = dev->written;
2257                                 dev->written = NULL;
2258                                 while (wbi && wbi->bi_sector <
2259                                         dev->sector + STRIPE_SECTORS) {
2260                                         wbi2 = r5_next_bio(wbi, dev->sector);
2261                                         if (!raid5_dec_bi_phys_segments(wbi)) {
2262                                                 md_write_end(conf->mddev);
2263                                                 wbi->bi_next = *return_bi;
2264                                                 *return_bi = wbi;
2265                                         }
2266                                         wbi = wbi2;
2267                                 }
2268                                 if (dev->towrite == NULL)
2269                                         bitmap_end = 1;
2270                                 spin_unlock_irq(&conf->device_lock);
2271                                 if (bitmap_end)
2272                                         bitmap_endwrite(conf->mddev->bitmap,
2273                                                         sh->sector,
2274                                                         STRIPE_SECTORS,
2275                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2276                                                         0);
2277                         }
2278                 }
2279
2280         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2281                 if (atomic_dec_and_test(&conf->pending_full_writes))
2282                         md_wakeup_thread(conf->mddev->thread);
2283 }
2284
2285 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2286                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2287 {
2288         int rmw = 0, rcw = 0, i;
2289         for (i = disks; i--; ) {
2290                 /* would I have to read this buffer for read_modify_write */
2291                 struct r5dev *dev = &sh->dev[i];
2292                 if ((dev->towrite || i == sh->pd_idx) &&
2293                     !test_bit(R5_LOCKED, &dev->flags) &&
2294                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2295                       test_bit(R5_Wantcompute, &dev->flags))) {
2296                         if (test_bit(R5_Insync, &dev->flags))
2297                                 rmw++;
2298                         else
2299                                 rmw += 2*disks;  /* cannot read it */
2300                 }
2301                 /* Would I have to read this buffer for reconstruct_write */
2302                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2303                     !test_bit(R5_LOCKED, &dev->flags) &&
2304                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2305                     test_bit(R5_Wantcompute, &dev->flags))) {
2306                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2307                         else
2308                                 rcw += 2*disks;
2309                 }
2310         }
2311         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2312                 (unsigned long long)sh->sector, rmw, rcw);
2313         set_bit(STRIPE_HANDLE, &sh->state);
2314         if (rmw < rcw && rmw > 0)
2315                 /* prefer read-modify-write, but need to get some data */
2316                 for (i = disks; i--; ) {
2317                         struct r5dev *dev = &sh->dev[i];
2318                         if ((dev->towrite || i == sh->pd_idx) &&
2319                             !test_bit(R5_LOCKED, &dev->flags) &&
2320                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2321                             test_bit(R5_Wantcompute, &dev->flags)) &&
2322                             test_bit(R5_Insync, &dev->flags)) {
2323                                 if (
2324                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2325                                         pr_debug("Read_old block "
2326                                                 "%d for r-m-w\n", i);
2327                                         set_bit(R5_LOCKED, &dev->flags);
2328                                         set_bit(R5_Wantread, &dev->flags);
2329                                         s->locked++;
2330                                 } else {
2331                                         set_bit(STRIPE_DELAYED, &sh->state);
2332                                         set_bit(STRIPE_HANDLE, &sh->state);
2333                                 }
2334                         }
2335                 }
2336         if (rcw <= rmw && rcw > 0)
2337                 /* want reconstruct write, but need to get some data */
2338                 for (i = disks; i--; ) {
2339                         struct r5dev *dev = &sh->dev[i];
2340                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2341                             i != sh->pd_idx &&
2342                             !test_bit(R5_LOCKED, &dev->flags) &&
2343                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2344                             test_bit(R5_Wantcompute, &dev->flags)) &&
2345                             test_bit(R5_Insync, &dev->flags)) {
2346                                 if (
2347                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2348                                         pr_debug("Read_old block "
2349                                                 "%d for Reconstruct\n", i);
2350                                         set_bit(R5_LOCKED, &dev->flags);
2351                                         set_bit(R5_Wantread, &dev->flags);
2352                                         s->locked++;
2353                                 } else {
2354                                         set_bit(STRIPE_DELAYED, &sh->state);
2355                                         set_bit(STRIPE_HANDLE, &sh->state);
2356                                 }
2357                         }
2358                 }
2359         /* now if nothing is locked, and if we have enough data,
2360          * we can start a write request
2361          */
2362         /* since handle_stripe can be called at any time we need to handle the
2363          * case where a compute block operation has been submitted and then a
2364          * subsequent call wants to start a write request.  raid5_run_ops only
2365          * handles the case where compute block and postxor are requested
2366          * simultaneously.  If this is not the case then new writes need to be
2367          * held off until the compute completes.
2368          */
2369         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2370             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2371             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2372                 schedule_reconstruction5(sh, s, rcw == 0, 0);
2373 }
2374
2375 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2376                 struct stripe_head *sh, struct stripe_head_state *s,
2377                 struct r6_state *r6s, int disks)
2378 {
2379         int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2380         int qd_idx = sh->qd_idx;
2381         for (i = disks; i--; ) {
2382                 struct r5dev *dev = &sh->dev[i];
2383                 /* Would I have to read this buffer for reconstruct_write */
2384                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2385                     && i != pd_idx && i != qd_idx
2386                     && (!test_bit(R5_LOCKED, &dev->flags)
2387                             ) &&
2388                     !test_bit(R5_UPTODATE, &dev->flags)) {
2389                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2390                         else {
2391                                 pr_debug("raid6: must_compute: "
2392                                         "disk %d flags=%#lx\n", i, dev->flags);
2393                                 must_compute++;
2394                         }
2395                 }
2396         }
2397         pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2398                (unsigned long long)sh->sector, rcw, must_compute);
2399         set_bit(STRIPE_HANDLE, &sh->state);
2400
2401         if (rcw > 0)
2402                 /* want reconstruct write, but need to get some data */
2403                 for (i = disks; i--; ) {
2404                         struct r5dev *dev = &sh->dev[i];
2405                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2406                             && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2407                             && !test_bit(R5_LOCKED, &dev->flags) &&
2408                             !test_bit(R5_UPTODATE, &dev->flags) &&
2409                             test_bit(R5_Insync, &dev->flags)) {
2410                                 if (
2411                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2412                                         pr_debug("Read_old stripe %llu "
2413                                                 "block %d for Reconstruct\n",
2414                                              (unsigned long long)sh->sector, i);
2415                                         set_bit(R5_LOCKED, &dev->flags);
2416                                         set_bit(R5_Wantread, &dev->flags);
2417                                         s->locked++;
2418                                 } else {
2419                                         pr_debug("Request delayed stripe %llu "
2420                                                 "block %d for Reconstruct\n",
2421                                              (unsigned long long)sh->sector, i);
2422                                         set_bit(STRIPE_DELAYED, &sh->state);
2423                                         set_bit(STRIPE_HANDLE, &sh->state);
2424                                 }
2425                         }
2426                 }
2427         /* now if nothing is locked, and if we have enough data, we can start a
2428          * write request
2429          */
2430         if (s->locked == 0 && rcw == 0 &&
2431             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2432                 if (must_compute > 0) {
2433                         /* We have failed blocks and need to compute them */
2434                         switch (s->failed) {
2435                         case 0:
2436                                 BUG();
2437                         case 1:
2438                                 compute_block_1(sh, r6s->failed_num[0], 0);
2439                                 break;
2440                         case 2:
2441                                 compute_block_2(sh, r6s->failed_num[0],
2442                                                 r6s->failed_num[1]);
2443                                 break;
2444                         default: /* This request should have been failed? */
2445                                 BUG();
2446                         }
2447                 }
2448
2449                 pr_debug("Computing parity for stripe %llu\n",
2450                         (unsigned long long)sh->sector);
2451                 compute_parity6(sh, RECONSTRUCT_WRITE);
2452                 /* now every locked buffer is ready to be written */
2453                 for (i = disks; i--; )
2454                         if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2455                                 pr_debug("Writing stripe %llu block %d\n",
2456                                        (unsigned long long)sh->sector, i);
2457                                 s->locked++;
2458                                 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2459                         }
2460                 if (s->locked == disks)
2461                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2462                                 atomic_inc(&conf->pending_full_writes);
2463                 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2464                 set_bit(STRIPE_INSYNC, &sh->state);
2465
2466                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2467                         atomic_dec(&conf->preread_active_stripes);
2468                         if (atomic_read(&conf->preread_active_stripes) <
2469                             IO_THRESHOLD)
2470                                 md_wakeup_thread(conf->mddev->thread);
2471                 }
2472         }
2473 }
2474
2475 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2476                                 struct stripe_head_state *s, int disks)
2477 {
2478         struct r5dev *dev = NULL;
2479
2480         set_bit(STRIPE_HANDLE, &sh->state);
2481
2482         switch (sh->check_state) {
2483         case check_state_idle:
2484                 /* start a new check operation if there are no failures */
2485                 if (s->failed == 0) {
2486                         BUG_ON(s->uptodate != disks);
2487                         sh->check_state = check_state_run;
2488                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2489                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2490                         s->uptodate--;
2491                         break;
2492                 }
2493                 dev = &sh->dev[s->failed_num];
2494                 /* fall through */
2495         case check_state_compute_result:
2496                 sh->check_state = check_state_idle;
2497                 if (!dev)
2498                         dev = &sh->dev[sh->pd_idx];
2499
2500                 /* check that a write has not made the stripe insync */
2501                 if (test_bit(STRIPE_INSYNC, &sh->state))
2502                         break;
2503
2504                 /* either failed parity check, or recovery is happening */
2505                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2506                 BUG_ON(s->uptodate != disks);
2507
2508                 set_bit(R5_LOCKED, &dev->flags);
2509                 s->locked++;
2510                 set_bit(R5_Wantwrite, &dev->flags);
2511
2512                 clear_bit(STRIPE_DEGRADED, &sh->state);
2513                 set_bit(STRIPE_INSYNC, &sh->state);
2514                 break;
2515         case check_state_run:
2516                 break; /* we will be called again upon completion */
2517         case check_state_check_result:
2518                 sh->check_state = check_state_idle;
2519
2520                 /* if a failure occurred during the check operation, leave
2521                  * STRIPE_INSYNC not set and let the stripe be handled again
2522                  */
2523                 if (s->failed)
2524                         break;
2525
2526                 /* handle a successful check operation, if parity is correct
2527                  * we are done.  Otherwise update the mismatch count and repair
2528                  * parity if !MD_RECOVERY_CHECK
2529                  */
2530                 if (sh->ops.zero_sum_result == 0)
2531                         /* parity is correct (on disc,
2532                          * not in buffer any more)
2533                          */
2534                         set_bit(STRIPE_INSYNC, &sh->state);
2535                 else {
2536                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2537                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2538                                 /* don't try to repair!! */
2539                                 set_bit(STRIPE_INSYNC, &sh->state);
2540                         else {
2541                                 sh->check_state = check_state_compute_run;
2542                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2543                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2544                                 set_bit(R5_Wantcompute,
2545                                         &sh->dev[sh->pd_idx].flags);
2546                                 sh->ops.target = sh->pd_idx;
2547                                 s->uptodate++;
2548                         }
2549                 }
2550                 break;
2551         case check_state_compute_run:
2552                 break;
2553         default:
2554                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2555                        __func__, sh->check_state,
2556                        (unsigned long long) sh->sector);
2557                 BUG();
2558         }
2559 }
2560
2561
2562 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2563                                 struct stripe_head_state *s,
2564                                 struct r6_state *r6s, struct page *tmp_page,
2565                                 int disks)
2566 {
2567         int update_p = 0, update_q = 0;
2568         struct r5dev *dev;
2569         int pd_idx = sh->pd_idx;
2570         int qd_idx = sh->qd_idx;
2571
2572         set_bit(STRIPE_HANDLE, &sh->state);
2573
2574         BUG_ON(s->failed > 2);
2575         BUG_ON(s->uptodate < disks);
2576         /* Want to check and possibly repair P and Q.
2577          * However there could be one 'failed' device, in which
2578          * case we can only check one of them, possibly using the
2579          * other to generate missing data
2580          */
2581
2582         /* If !tmp_page, we cannot do the calculations,
2583          * but as we have set STRIPE_HANDLE, we will soon be called
2584          * by stripe_handle with a tmp_page - just wait until then.
2585          */
2586         if (tmp_page) {
2587                 if (s->failed == r6s->q_failed) {
2588                         /* The only possible failed device holds 'Q', so it
2589                          * makes sense to check P (If anything else were failed,
2590                          * we would have used P to recreate it).
2591                          */
2592                         compute_block_1(sh, pd_idx, 1);
2593                         if (!page_is_zero(sh->dev[pd_idx].page)) {
2594                                 compute_block_1(sh, pd_idx, 0);
2595                                 update_p = 1;
2596                         }
2597                 }
2598                 if (!r6s->q_failed && s->failed < 2) {
2599                         /* q is not failed, and we didn't use it to generate
2600                          * anything, so it makes sense to check it
2601                          */
2602                         memcpy(page_address(tmp_page),
2603                                page_address(sh->dev[qd_idx].page),
2604                                STRIPE_SIZE);
2605                         compute_parity6(sh, UPDATE_PARITY);
2606                         if (memcmp(page_address(tmp_page),
2607                                    page_address(sh->dev[qd_idx].page),
2608                                    STRIPE_SIZE) != 0) {
2609                                 clear_bit(STRIPE_INSYNC, &sh->state);
2610                                 update_q = 1;
2611                         }
2612                 }
2613                 if (update_p || update_q) {
2614                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2615                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2616                                 /* don't try to repair!! */
2617                                 update_p = update_q = 0;
2618                 }
2619
2620                 /* now write out any block on a failed drive,
2621                  * or P or Q if they need it
2622                  */
2623
2624                 if (s->failed == 2) {
2625                         dev = &sh->dev[r6s->failed_num[1]];
2626                         s->locked++;
2627                         set_bit(R5_LOCKED, &dev->flags);
2628                         set_bit(R5_Wantwrite, &dev->flags);
2629                 }
2630                 if (s->failed >= 1) {
2631                         dev = &sh->dev[r6s->failed_num[0]];
2632                         s->locked++;
2633                         set_bit(R5_LOCKED, &dev->flags);
2634                         set_bit(R5_Wantwrite, &dev->flags);
2635                 }
2636
2637                 if (update_p) {
2638                         dev = &sh->dev[pd_idx];
2639                         s->locked++;
2640                         set_bit(R5_LOCKED, &dev->flags);
2641                         set_bit(R5_Wantwrite, &dev->flags);
2642                 }
2643                 if (update_q) {
2644                         dev = &sh->dev[qd_idx];
2645                         s->locked++;
2646                         set_bit(R5_LOCKED, &dev->flags);
2647                         set_bit(R5_Wantwrite, &dev->flags);
2648                 }
2649                 clear_bit(STRIPE_DEGRADED, &sh->state);
2650
2651                 set_bit(STRIPE_INSYNC, &sh->state);
2652         }
2653 }
2654
2655 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2656                                 struct r6_state *r6s)
2657 {
2658         int i;
2659
2660         /* We have read all the blocks in this stripe and now we need to
2661          * copy some of them into a target stripe for expand.
2662          */
2663         struct dma_async_tx_descriptor *tx = NULL;
2664         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2665         for (i = 0; i < sh->disks; i++)
2666                 if (i != sh->pd_idx && i != sh->qd_idx) {
2667                         int dd_idx, j;
2668                         struct stripe_head *sh2;
2669
2670                         sector_t bn = compute_blocknr(sh, i, 1);
2671                         sector_t s = raid5_compute_sector(conf, bn, 0,
2672                                                           &dd_idx, NULL);
2673                         sh2 = get_active_stripe(conf, s, 0, 1, 1);
2674                         if (sh2 == NULL)
2675                                 /* so far only the early blocks of this stripe
2676                                  * have been requested.  When later blocks
2677                                  * get requested, we will try again
2678                                  */
2679                                 continue;
2680                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2681                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2682                                 /* must have already done this block */
2683                                 release_stripe(sh2);
2684                                 continue;
2685                         }
2686
2687                         /* place all the copies on one channel */
2688                         tx = async_memcpy(sh2->dev[dd_idx].page,
2689                                 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2690                                 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2691
2692                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2693                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2694                         for (j = 0; j < conf->raid_disks; j++)
2695                                 if (j != sh2->pd_idx &&
2696                                     (!r6s || j != sh2->qd_idx) &&
2697                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2698                                         break;
2699                         if (j == conf->raid_disks) {
2700                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2701                                 set_bit(STRIPE_HANDLE, &sh2->state);
2702                         }
2703                         release_stripe(sh2);
2704
2705                 }
2706         /* done submitting copies, wait for them to complete */
2707         if (tx) {
2708                 async_tx_ack(tx);
2709                 dma_wait_for_async_tx(tx);
2710         }
2711 }
2712
2713
2714 /*
2715  * handle_stripe - do things to a stripe.
2716  *
2717  * We lock the stripe and then examine the state of various bits
2718  * to see what needs to be done.
2719  * Possible results:
2720  *    return some read request which now have data
2721  *    return some write requests which are safely on disc
2722  *    schedule a read on some buffers
2723  *    schedule a write of some buffers
2724  *    return confirmation of parity correctness
2725  *
2726  * buffers are taken off read_list or write_list, and bh_cache buffers
2727  * get BH_Lock set before the stripe lock is released.
2728  *
2729  */
2730
2731 static bool handle_stripe5(struct stripe_head *sh)
2732 {
2733         raid5_conf_t *conf = sh->raid_conf;
2734         int disks = sh->disks, i;
2735         struct bio *return_bi = NULL;
2736         struct stripe_head_state s;
2737         struct r5dev *dev;
2738         mdk_rdev_t *blocked_rdev = NULL;
2739         int prexor;
2740
2741         memset(&s, 0, sizeof(s));
2742         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2743                  "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2744                  atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2745                  sh->reconstruct_state);
2746
2747         spin_lock(&sh->lock);
2748         clear_bit(STRIPE_HANDLE, &sh->state);
2749         clear_bit(STRIPE_DELAYED, &sh->state);
2750
2751         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2752         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2753         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2754
2755         /* Now to look around and see what can be done */
2756         rcu_read_lock();
2757         for (i=disks; i--; ) {
2758                 mdk_rdev_t *rdev;
2759
2760                 dev = &sh->dev[i];
2761                 clear_bit(R5_Insync, &dev->flags);
2762
2763                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2764                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2765                         dev->towrite, dev->written);
2766
2767                 /* maybe we can request a biofill operation
2768                  *
2769                  * new wantfill requests are only permitted while
2770                  * ops_complete_biofill is guaranteed to be inactive
2771                  */
2772                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2773                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2774                         set_bit(R5_Wantfill, &dev->flags);
2775
2776                 /* now count some things */
2777                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2778                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2779                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2780
2781                 if (test_bit(R5_Wantfill, &dev->flags))
2782                         s.to_fill++;
2783                 else if (dev->toread)
2784                         s.to_read++;
2785                 if (dev->towrite) {
2786                         s.to_write++;
2787                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2788                                 s.non_overwrite++;
2789                 }
2790                 if (dev->written)
2791                         s.written++;
2792                 rdev = rcu_dereference(conf->disks[i].rdev);
2793                 if (blocked_rdev == NULL &&
2794                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2795                         blocked_rdev = rdev;
2796                         atomic_inc(&rdev->nr_pending);
2797                 }
2798                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2799                         /* The ReadError flag will just be confusing now */
2800                         clear_bit(R5_ReadError, &dev->flags);
2801                         clear_bit(R5_ReWrite, &dev->flags);
2802                 }
2803                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2804                     || test_bit(R5_ReadError, &dev->flags)) {
2805                         s.failed++;
2806                         s.failed_num = i;
2807                 } else
2808                         set_bit(R5_Insync, &dev->flags);
2809         }
2810         rcu_read_unlock();
2811
2812         if (unlikely(blocked_rdev)) {
2813                 if (s.syncing || s.expanding || s.expanded ||
2814                     s.to_write || s.written) {
2815                         set_bit(STRIPE_HANDLE, &sh->state);
2816                         goto unlock;
2817                 }
2818                 /* There is nothing for the blocked_rdev to block */
2819                 rdev_dec_pending(blocked_rdev, conf->mddev);
2820                 blocked_rdev = NULL;
2821         }
2822
2823         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
2824                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
2825                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
2826         }
2827
2828         pr_debug("locked=%d uptodate=%d to_read=%d"
2829                 " to_write=%d failed=%d failed_num=%d\n",
2830                 s.locked, s.uptodate, s.to_read, s.to_write,
2831                 s.failed, s.failed_num);
2832         /* check if the array has lost two devices and, if so, some requests might
2833          * need to be failed
2834          */
2835         if (s.failed > 1 && s.to_read+s.to_write+s.written)
2836                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
2837         if (s.failed > 1 && s.syncing) {
2838                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2839                 clear_bit(STRIPE_SYNCING, &sh->state);
2840                 s.syncing = 0;
2841         }
2842
2843         /* might be able to return some write requests if the parity block
2844          * is safe, or on a failed drive
2845          */
2846         dev = &sh->dev[sh->pd_idx];
2847         if ( s.written &&
2848              ((test_bit(R5_Insync, &dev->flags) &&
2849                !test_bit(R5_LOCKED, &dev->flags) &&
2850                test_bit(R5_UPTODATE, &dev->flags)) ||
2851                (s.failed == 1 && s.failed_num == sh->pd_idx)))
2852                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
2853
2854         /* Now we might consider reading some blocks, either to check/generate
2855          * parity, or to satisfy requests
2856          * or to load a block that is being partially written.
2857          */
2858         if (s.to_read || s.non_overwrite ||
2859             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
2860                 handle_stripe_fill5(sh, &s, disks);
2861
2862         /* Now we check to see if any write operations have recently
2863          * completed
2864          */
2865         prexor = 0;
2866         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
2867                 prexor = 1;
2868         if (sh->reconstruct_state == reconstruct_state_drain_result ||
2869             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
2870                 sh->reconstruct_state = reconstruct_state_idle;
2871
2872                 /* All the 'written' buffers and the parity block are ready to
2873                  * be written back to disk
2874                  */
2875                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2876                 for (i = disks; i--; ) {
2877                         dev = &sh->dev[i];
2878                         if (test_bit(R5_LOCKED, &dev->flags) &&
2879                                 (i == sh->pd_idx || dev->written)) {
2880                                 pr_debug("Writing block %d\n", i);
2881                                 set_bit(R5_Wantwrite, &dev->flags);
2882                                 if (prexor)
2883                                         continue;
2884                                 if (!test_bit(R5_Insync, &dev->flags) ||
2885                                     (i == sh->pd_idx && s.failed == 0))
2886                                         set_bit(STRIPE_INSYNC, &sh->state);
2887                         }
2888                 }
2889                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2890                         atomic_dec(&conf->preread_active_stripes);
2891                         if (atomic_read(&conf->preread_active_stripes) <
2892                                 IO_THRESHOLD)
2893                                 md_wakeup_thread(conf->mddev->thread);
2894                 }
2895         }
2896
2897         /* Now to consider new write requests and what else, if anything
2898          * should be read.  We do not handle new writes when:
2899          * 1/ A 'write' operation (copy+xor) is already in flight.
2900          * 2/ A 'check' operation is in flight, as it may clobber the parity
2901          *    block.
2902          */
2903         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
2904                 handle_stripe_dirtying5(conf, sh, &s, disks);
2905
2906         /* maybe we need to check and possibly fix the parity for this stripe
2907          * Any reads will already have been scheduled, so we just see if enough
2908          * data is available.  The parity check is held off while parity
2909          * dependent operations are in flight.
2910          */
2911         if (sh->check_state ||
2912             (s.syncing && s.locked == 0 &&
2913              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
2914              !test_bit(STRIPE_INSYNC, &sh->state)))
2915                 handle_parity_checks5(conf, sh, &s, disks);
2916
2917         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2918                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2919                 clear_bit(STRIPE_SYNCING, &sh->state);
2920         }
2921
2922         /* If the failed drive is just a ReadError, then we might need to progress
2923          * the repair/check process
2924          */
2925         if (s.failed == 1 && !conf->mddev->ro &&
2926             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2927             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2928             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2929                 ) {
2930                 dev = &sh->dev[s.failed_num];
2931                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2932                         set_bit(R5_Wantwrite, &dev->flags);
2933                         set_bit(R5_ReWrite, &dev->flags);
2934                         set_bit(R5_LOCKED, &dev->flags);
2935                         s.locked++;
2936                 } else {
2937                         /* let's read it back */
2938                         set_bit(R5_Wantread, &dev->flags);
2939                         set_bit(R5_LOCKED, &dev->flags);
2940                         s.locked++;
2941                 }
2942         }
2943
2944         /* Finish reconstruct operations initiated by the expansion process */
2945         if (sh->reconstruct_state == reconstruct_state_result) {
2946                 struct stripe_head *sh2
2947                         = get_active_stripe(conf, sh->sector, 1, 1, 1);
2948                 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
2949                         /* sh cannot be written until sh2 has been read.
2950                          * so arrange for sh to be delayed a little
2951                          */
2952                         set_bit(STRIPE_DELAYED, &sh->state);
2953                         set_bit(STRIPE_HANDLE, &sh->state);
2954                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
2955                                               &sh2->state))
2956                                 atomic_inc(&conf->preread_active_stripes);
2957                         release_stripe(sh2);
2958                         goto unlock;
2959                 }
2960                 if (sh2)
2961                         release_stripe(sh2);
2962
2963                 sh->reconstruct_state = reconstruct_state_idle;
2964                 clear_bit(STRIPE_EXPANDING, &sh->state);
2965                 for (i = conf->raid_disks; i--; ) {
2966                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2967                         set_bit(R5_LOCKED, &sh->dev[i].flags);
2968                         s.locked++;
2969                 }
2970         }
2971
2972         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2973             !sh->reconstruct_state) {
2974                 /* Need to write out all blocks after computing parity */
2975                 sh->disks = conf->raid_disks;
2976                 stripe_set_idx(sh->sector, conf, 0, sh);
2977                 schedule_reconstruction5(sh, &s, 1, 1);
2978         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
2979                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2980                 atomic_dec(&conf->reshape_stripes);
2981                 wake_up(&conf->wait_for_overlap);
2982                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2983         }
2984
2985         if (s.expanding && s.locked == 0 &&
2986             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
2987                 handle_stripe_expansion(conf, sh, NULL);
2988
2989  unlock:
2990         spin_unlock(&sh->lock);
2991
2992         /* wait for this device to become unblocked */
2993         if (unlikely(blocked_rdev))
2994                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
2995
2996         if (s.ops_request)
2997                 raid5_run_ops(sh, s.ops_request);
2998
2999         ops_run_io(sh, &s);
3000
3001         return_io(return_bi);
3002
3003         return blocked_rdev == NULL;
3004 }
3005
3006 static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
3007 {
3008         raid5_conf_t *conf = sh->raid_conf;
3009         int disks = sh->disks;
3010         struct bio *return_bi = NULL;
3011         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3012         struct stripe_head_state s;
3013         struct r6_state r6s;
3014         struct r5dev *dev, *pdev, *qdev;
3015         mdk_rdev_t *blocked_rdev = NULL;
3016
3017         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3018                 "pd_idx=%d, qd_idx=%d\n",
3019                (unsigned long long)sh->sector, sh->state,
3020                atomic_read(&sh->count), pd_idx, qd_idx);
3021         memset(&s, 0, sizeof(s));
3022
3023         spin_lock(&sh->lock);
3024         clear_bit(STRIPE_HANDLE, &sh->state);
3025         clear_bit(STRIPE_DELAYED, &sh->state);
3026
3027         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3028         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3029         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3030         /* Now to look around and see what can be done */
3031
3032         rcu_read_lock();
3033         for (i=disks; i--; ) {
3034                 mdk_rdev_t *rdev;
3035                 dev = &sh->dev[i];
3036                 clear_bit(R5_Insync, &dev->flags);
3037
3038                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3039                         i, dev->flags, dev->toread, dev->towrite, dev->written);
3040                 /* maybe we can reply to a read */
3041                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
3042                         struct bio *rbi, *rbi2;
3043                         pr_debug("Return read for disc %d\n", i);
3044                         spin_lock_irq(&conf->device_lock);
3045                         rbi = dev->toread;
3046                         dev->toread = NULL;
3047                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
3048                                 wake_up(&conf->wait_for_overlap);
3049                         spin_unlock_irq(&conf->device_lock);
3050                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
3051                                 copy_data(0, rbi, dev->page, dev->sector);
3052                                 rbi2 = r5_next_bio(rbi, dev->sector);
3053                                 spin_lock_irq(&conf->device_lock);
3054                                 if (!raid5_dec_bi_phys_segments(rbi)) {
3055                                         rbi->bi_next = return_bi;
3056                                         return_bi = rbi;
3057                                 }
3058                                 spin_unlock_irq(&conf->device_lock);
3059                                 rbi = rbi2;
3060                         }
3061                 }
3062
3063                 /* now count some things */
3064                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3065                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3066
3067
3068                 if (dev->toread)
3069                         s.to_read++;
3070                 if (dev->towrite) {
3071                         s.to_write++;
3072                         if (!test_bit(R5_OVERWRITE, &dev->flags))
3073                                 s.non_overwrite++;
3074                 }
3075                 if (dev->written)
3076                         s.written++;
3077                 rdev = rcu_dereference(conf->disks[i].rdev);
3078                 if (blocked_rdev == NULL &&
3079                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3080                         blocked_rdev = rdev;
3081                         atomic_inc(&rdev->nr_pending);
3082                 }
3083                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3084                         /* The ReadError flag will just be confusing now */
3085                         clear_bit(R5_ReadError, &dev->flags);
3086                         clear_bit(R5_ReWrite, &dev->flags);
3087                 }
3088                 if (!rdev || !test_bit(In_sync, &rdev->flags)
3089                     || test_bit(R5_ReadError, &dev->flags)) {
3090                         if (s.failed < 2)
3091                                 r6s.failed_num[s.failed] = i;
3092                         s.failed++;
3093                 } else
3094                         set_bit(R5_Insync, &dev->flags);
3095         }
3096         rcu_read_unlock();
3097
3098         if (unlikely(blocked_rdev)) {
3099                 if (s.syncing || s.expanding || s.expanded ||
3100                     s.to_write || s.written) {
3101                         set_bit(STRIPE_HANDLE, &sh->state);
3102                         goto unlock;
3103                 }
3104                 /* There is nothing for the blocked_rdev to block */
3105                 rdev_dec_pending(blocked_rdev, conf->mddev);
3106                 blocked_rdev = NULL;
3107         }
3108
3109         pr_debug("locked=%d uptodate=%d to_read=%d"
3110                " to_write=%d failed=%d failed_num=%d,%d\n",
3111                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3112                r6s.failed_num[0], r6s.failed_num[1]);
3113         /* check if the array has lost >2 devices and, if so, some requests
3114          * might need to be failed
3115          */
3116         if (s.failed > 2 && s.to_read+s.to_write+s.written)
3117                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3118         if (s.failed > 2 && s.syncing) {
3119                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3120                 clear_bit(STRIPE_SYNCING, &sh->state);
3121                 s.syncing = 0;
3122         }
3123
3124         /*
3125          * might be able to return some write requests if the parity blocks
3126          * are safe, or on a failed drive
3127          */
3128         pdev = &sh->dev[pd_idx];
3129         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3130                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3131         qdev = &sh->dev[qd_idx];
3132         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3133                 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3134
3135         if ( s.written &&
3136              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3137                              && !test_bit(R5_LOCKED, &pdev->flags)
3138                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3139              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3140                              && !test_bit(R5_LOCKED, &qdev->flags)
3141                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3142                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3143
3144         /* Now we might consider reading some blocks, either to check/generate
3145          * parity, or to satisfy requests
3146          * or to load a block that is being partially written.
3147          */
3148         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3149             (s.syncing && (s.uptodate < disks)) || s.expanding)
3150                 handle_stripe_fill6(sh, &s, &r6s, disks);
3151
3152         /* now to consider writing and what else, if anything should be read */
3153         if (s.to_write)
3154                 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3155
3156         /* maybe we need to check and possibly fix the parity for this stripe
3157          * Any reads will already have been scheduled, so we just see if enough
3158          * data is available
3159          */
3160         if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3161                 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3162
3163         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3164                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3165                 clear_bit(STRIPE_SYNCING, &sh->state);
3166         }
3167
3168         /* If the failed drives are just a ReadError, then we might need
3169          * to progress the repair/check process
3170          */
3171         if (s.failed <= 2 && !conf->mddev->ro)
3172                 for (i = 0; i < s.failed; i++) {
3173                         dev = &sh->dev[r6s.failed_num[i]];
3174                         if (test_bit(R5_ReadError, &dev->flags)
3175                             && !test_bit(R5_LOCKED, &dev->flags)
3176                             && test_bit(R5_UPTODATE, &dev->flags)
3177                                 ) {
3178                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3179                                         set_bit(R5_Wantwrite, &dev->flags);
3180                                         set_bit(R5_ReWrite, &dev->flags);
3181                                         set_bit(R5_LOCKED, &dev->flags);
3182                                 } else {
3183                                         /* let's read it back */
3184                                         set_bit(R5_Wantread, &dev->flags);
3185                                         set_bit(R5_LOCKED, &dev->flags);
3186                                 }
3187                         }
3188                 }
3189
3190         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3191                 struct stripe_head *sh2
3192                         = get_active_stripe(conf, sh->sector, 1, 1, 1);
3193                 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3194                         /* sh cannot be written until sh2 has been read.
3195                          * so arrange for sh to be delayed a little
3196                          */
3197                         set_bit(STRIPE_DELAYED, &sh->state);
3198                         set_bit(STRIPE_HANDLE, &sh->state);
3199                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3200                                               &sh2->state))
3201                                 atomic_inc(&conf->preread_active_stripes);
3202                         release_stripe(sh2);
3203                         goto unlock;
3204                 }
3205                 if (sh2)
3206                         release_stripe(sh2);
3207
3208                 /* Need to write out all blocks after computing P&Q */
3209                 sh->disks = conf->raid_disks;
3210                 stripe_set_idx(sh->sector, conf, 0, sh);
3211                 compute_parity6(sh, RECONSTRUCT_WRITE);
3212                 for (i = conf->raid_disks ; i-- ;  ) {
3213                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3214                         s.locked++;
3215                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3216                 }
3217                 clear_bit(STRIPE_EXPANDING, &sh->state);
3218         } else if (s.expanded) {
3219                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3220                 atomic_dec(&conf->reshape_stripes);
3221                 wake_up(&conf->wait_for_overlap);
3222                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3223         }
3224
3225         if (s.expanding && s.locked == 0 &&
3226             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3227                 handle_stripe_expansion(conf, sh, &r6s);
3228
3229  unlock:
3230         spin_unlock(&sh->lock);
3231
3232         /* wait for this device to become unblocked */
3233         if (unlikely(blocked_rdev))
3234                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3235
3236         ops_run_io(sh, &s);
3237
3238         return_io(return_bi);
3239
3240         return blocked_rdev == NULL;
3241 }
3242
3243 /* returns true if the stripe was handled */
3244 static bool handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3245 {
3246         if (sh->raid_conf->level == 6)
3247                 return handle_stripe6(sh, tmp_page);
3248         else
3249                 return handle_stripe5(sh);
3250 }
3251
3252
3253
3254 static void raid5_activate_delayed(raid5_conf_t *conf)
3255 {
3256         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3257                 while (!list_empty(&conf->delayed_list)) {
3258                         struct list_head *l = conf->delayed_list.next;
3259                         struct stripe_head *sh;
3260                         sh = list_entry(l, struct stripe_head, lru);
3261                         list_del_init(l);
3262                         clear_bit(STRIPE_DELAYED, &sh->state);
3263                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3264                                 atomic_inc(&conf->preread_active_stripes);
3265                         list_add_tail(&sh->lru, &conf->hold_list);
3266                 }
3267         } else
3268                 blk_plug_device(conf->mddev->queue);
3269 }
3270
3271 static void activate_bit_delay(raid5_conf_t *conf)
3272 {
3273         /* device_lock is held */
3274         struct list_head head;
3275         list_add(&head, &conf->bitmap_list);
3276         list_del_init(&conf->bitmap_list);
3277         while (!list_empty(&head)) {
3278                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3279                 list_del_init(&sh->lru);
3280                 atomic_inc(&sh->count);
3281                 __release_stripe(conf, sh);
3282         }
3283 }
3284
3285 static void unplug_slaves(mddev_t *mddev)
3286 {
3287         raid5_conf_t *conf = mddev->private;
3288         int i;
3289
3290         rcu_read_lock();
3291         for (i = 0; i < conf->raid_disks; i++) {
3292                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3293                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3294                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3295
3296                         atomic_inc(&rdev->nr_pending);
3297                         rcu_read_unlock();
3298
3299                         blk_unplug(r_queue);
3300
3301                         rdev_dec_pending(rdev, mddev);
3302                         rcu_read_lock();
3303                 }
3304         }
3305         rcu_read_unlock();
3306 }
3307
3308 static void raid5_unplug_device(struct request_queue *q)
3309 {
3310         mddev_t *mddev = q->queuedata;
3311         raid5_conf_t *conf = mddev->private;
3312         unsigned long flags;
3313
3314         spin_lock_irqsave(&conf->device_lock, flags);
3315
3316         if (blk_remove_plug(q)) {
3317                 conf->seq_flush++;
3318                 raid5_activate_delayed(conf);
3319         }
3320         md_wakeup_thread(mddev->thread);
3321
3322         spin_unlock_irqrestore(&conf->device_lock, flags);
3323
3324         unplug_slaves(mddev);
3325 }
3326
3327 static int raid5_congested(void *data, int bits)
3328 {
3329         mddev_t *mddev = data;
3330         raid5_conf_t *conf = mddev->private;
3331
3332         /* No difference between reads and writes.  Just check
3333          * how busy the stripe_cache is
3334          */
3335
3336         if (mddev_congested(mddev, bits))
3337                 return 1;
3338         if (conf->inactive_blocked)
3339                 return 1;
3340         if (conf->quiesce)
3341                 return 1;
3342         if (list_empty_careful(&conf->inactive_list))
3343                 return 1;
3344
3345         return 0;
3346 }
3347
3348 /* We want read requests to align with chunks where possible,
3349  * but write requests don't need to.
3350  */
3351 static int raid5_mergeable_bvec(struct request_queue *q,
3352                                 struct bvec_merge_data *bvm,
3353                                 struct bio_vec *biovec)
3354 {
3355         mddev_t *mddev = q->queuedata;
3356         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3357         int max;
3358         unsigned int chunk_sectors = mddev->chunk_sectors;
3359         unsigned int bio_sectors = bvm->bi_size >> 9;
3360
3361         if ((bvm->bi_rw & 1) == WRITE)
3362                 return biovec->bv_len; /* always allow writes to be mergeable */
3363
3364         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3365                 chunk_sectors = mddev->new_chunk_sectors;
3366         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3367         if (max < 0) max = 0;
3368         if (max <= biovec->bv_len && bio_sectors == 0)
3369                 return biovec->bv_len;
3370         else
3371                 return max;
3372 }
3373
3374
3375 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3376 {
3377         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3378         unsigned int chunk_sectors = mddev->chunk_sectors;
3379         unsigned int bio_sectors = bio->bi_size >> 9;
3380
3381         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3382                 chunk_sectors = mddev->new_chunk_sectors;
3383         return  chunk_sectors >=
3384                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3385 }
3386
3387 /*
3388  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3389  *  later sampled by raid5d.
3390  */
3391 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3392 {
3393         unsigned long flags;
3394
3395         spin_lock_irqsave(&conf->device_lock, flags);
3396
3397         bi->bi_next = conf->retry_read_aligned_list;
3398         conf->retry_read_aligned_list = bi;
3399
3400         spin_unlock_irqrestore(&conf->device_lock, flags);
3401         md_wakeup_thread(conf->mddev->thread);
3402 }
3403
3404
3405 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3406 {
3407         struct bio *bi;
3408
3409         bi = conf->retry_read_aligned;
3410         if (bi) {
3411                 conf->retry_read_aligned = NULL;
3412                 return bi;
3413         }
3414         bi = conf->retry_read_aligned_list;
3415         if(bi) {
3416                 conf->retry_read_aligned_list = bi->bi_next;
3417                 bi->bi_next = NULL;
3418                 /*
3419                  * this sets the active strip count to 1 and the processed
3420                  * strip count to zero (upper 8 bits)
3421                  */
3422                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3423         }
3424
3425         return bi;
3426 }
3427
3428
3429 /*
3430  *  The "raid5_align_endio" should check if the read succeeded and if it
3431  *  did, call bio_endio on the original bio (having bio_put the new bio
3432  *  first).
3433  *  If the read failed..
3434  */
3435 static void raid5_align_endio(struct bio *bi, int error)
3436 {
3437         struct bio* raid_bi  = bi->bi_private;
3438         mddev_t *mddev;
3439         raid5_conf_t *conf;
3440         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3441         mdk_rdev_t *rdev;
3442
3443         bio_put(bi);
3444
3445         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3446         conf = mddev->private;
3447         rdev = (void*)raid_bi->bi_next;
3448         raid_bi->bi_next = NULL;
3449
3450         rdev_dec_pending(rdev, conf->mddev);
3451
3452         if (!error && uptodate) {
3453                 bio_endio(raid_bi, 0);
3454                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3455                         wake_up(&conf->wait_for_stripe);
3456                 return;
3457         }
3458
3459
3460         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3461
3462         add_bio_to_retry(raid_bi, conf);
3463 }
3464
3465 static int bio_fits_rdev(struct bio *bi)
3466 {
3467         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3468
3469         if ((bi->bi_size>>9) > queue_max_sectors(q))
3470                 return 0;
3471         blk_recount_segments(q, bi);
3472         if (bi->bi_phys_segments > queue_max_phys_segments(q))
3473                 return 0;
3474
3475         if (q->merge_bvec_fn)
3476                 /* it's too hard to apply the merge_bvec_fn at this stage,
3477                  * just just give up
3478                  */
3479                 return 0;
3480
3481         return 1;
3482 }
3483
3484
3485 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3486 {
3487         mddev_t *mddev = q->queuedata;
3488         raid5_conf_t *conf = mddev->private;
3489         unsigned int dd_idx;
3490         struct bio* align_bi;
3491         mdk_rdev_t *rdev;
3492
3493         if (!in_chunk_boundary(mddev, raid_bio)) {
3494                 pr_debug("chunk_aligned_read : non aligned\n");
3495                 return 0;
3496         }
3497         /*
3498          * use bio_clone to make a copy of the bio
3499          */
3500         align_bi = bio_clone(raid_bio, GFP_NOIO);
3501         if (!align_bi)
3502                 return 0;
3503         /*
3504          *   set bi_end_io to a new function, and set bi_private to the
3505          *     original bio.
3506          */
3507         align_bi->bi_end_io  = raid5_align_endio;
3508         align_bi->bi_private = raid_bio;
3509         /*
3510          *      compute position
3511          */
3512         align_bi->bi_sector =  raid5_compute_sector(conf, raid_bio->bi_sector,
3513                                                     0,
3514                                                     &dd_idx, NULL);
3515
3516         rcu_read_lock();
3517         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3518         if (rdev && test_bit(In_sync, &rdev->flags)) {
3519                 atomic_inc(&rdev->nr_pending);
3520                 rcu_read_unlock();
3521                 raid_bio->bi_next = (void*)rdev;
3522                 align_bi->bi_bdev =  rdev->bdev;
3523                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3524                 align_bi->bi_sector += rdev->data_offset;
3525
3526                 if (!bio_fits_rdev(align_bi)) {
3527                         /* too big in some way */
3528                         bio_put(align_bi);
3529                         rdev_dec_pending(rdev, mddev);
3530                         return 0;
3531                 }
3532
3533                 spin_lock_irq(&conf->device_lock);
3534                 wait_event_lock_irq(conf->wait_for_stripe,
3535                                     conf->quiesce == 0,
3536                                     conf->device_lock, /* nothing */);
3537                 atomic_inc(&conf->active_aligned_reads);
3538                 spin_unlock_irq(&conf->device_lock);
3539
3540                 generic_make_request(align_bi);
3541                 return 1;
3542         } else {
3543                 rcu_read_unlock();
3544                 bio_put(align_bi);
3545                 return 0;
3546         }
3547 }
3548
3549 /* __get_priority_stripe - get the next stripe to process
3550  *
3551  * Full stripe writes are allowed to pass preread active stripes up until
3552  * the bypass_threshold is exceeded.  In general the bypass_count
3553  * increments when the handle_list is handled before the hold_list; however, it
3554  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3555  * stripe with in flight i/o.  The bypass_count will be reset when the
3556  * head of the hold_list has changed, i.e. the head was promoted to the
3557  * handle_list.
3558  */
3559 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3560 {
3561         struct stripe_head *sh;
3562
3563         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3564                   __func__,
3565                   list_empty(&conf->handle_list) ? "empty" : "busy",
3566                   list_empty(&conf->hold_list) ? "empty" : "busy",
3567                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
3568
3569         if (!list_empty(&conf->handle_list)) {
3570                 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3571
3572                 if (list_empty(&conf->hold_list))
3573                         conf->bypass_count = 0;
3574                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3575                         if (conf->hold_list.next == conf->last_hold)
3576                                 conf->bypass_count++;
3577                         else {
3578                                 conf->last_hold = conf->hold_list.next;
3579                                 conf->bypass_count -= conf->bypass_threshold;
3580                                 if (conf->bypass_count < 0)
3581                                         conf->bypass_count = 0;
3582                         }
3583                 }
3584         } else if (!list_empty(&conf->hold_list) &&
3585                    ((conf->bypass_threshold &&
3586                      conf->bypass_count > conf->bypass_threshold) ||
3587                     atomic_read(&conf->pending_full_writes) == 0)) {
3588                 sh = list_entry(conf->hold_list.next,
3589                                 typeof(*sh), lru);
3590                 conf->bypass_count -= conf->bypass_threshold;
3591                 if (conf->bypass_count < 0)
3592                         conf->bypass_count = 0;
3593         } else
3594                 return NULL;
3595
3596         list_del_init(&sh->lru);
3597         atomic_inc(&sh->count);
3598         BUG_ON(atomic_read(&sh->count) != 1);
3599         return sh;
3600 }
3601
3602 static int make_request(struct request_queue *q, struct bio * bi)
3603 {
3604         mddev_t *mddev = q->queuedata;
3605         raid5_conf_t *conf = mddev->private;
3606         int dd_idx;
3607         sector_t new_sector;
3608         sector_t logical_sector, last_sector;
3609         struct stripe_head *sh;
3610         const int rw = bio_data_dir(bi);
3611         int cpu, remaining;
3612
3613         if (unlikely(bio_rw_flagged(bi, BIO_RW_BARRIER))) {
3614                 bio_endio(bi, -EOPNOTSUPP);
3615                 return 0;
3616         }
3617
3618         md_write_start(mddev, bi);
3619
3620         cpu = part_stat_lock();
3621         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
3622         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
3623                       bio_sectors(bi));
3624         part_stat_unlock();
3625
3626         if (rw == READ &&
3627              mddev->reshape_position == MaxSector &&
3628              chunk_aligned_read(q,bi))
3629                 return 0;
3630
3631         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3632         last_sector = bi->bi_sector + (bi->bi_size>>9);
3633         bi->bi_next = NULL;
3634         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3635
3636         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3637                 DEFINE_WAIT(w);
3638                 int disks, data_disks;
3639                 int previous;
3640
3641         retry:
3642                 previous = 0;
3643                 disks = conf->raid_disks;
3644                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3645                 if (unlikely(conf->reshape_progress != MaxSector)) {
3646                         /* spinlock is needed as reshape_progress may be
3647                          * 64bit on a 32bit platform, and so it might be
3648                          * possible to see a half-updated value
3649                          * Ofcourse reshape_progress could change after
3650                          * the lock is dropped, so once we get a reference
3651                          * to the stripe that we think it is, we will have
3652                          * to check again.
3653                          */
3654                         spin_lock_irq(&conf->device_lock);
3655                         if (mddev->delta_disks < 0
3656                             ? logical_sector < conf->reshape_progress
3657                             : logical_sector >= conf->reshape_progress) {
3658                                 disks = conf->previous_raid_disks;
3659                                 previous = 1;
3660                         } else {
3661                                 if (mddev->delta_disks < 0
3662                                     ? logical_sector < conf->reshape_safe
3663                                     : logical_sector >= conf->reshape_safe) {
3664                                         spin_unlock_irq(&conf->device_lock);
3665                                         schedule();
3666                                         goto retry;
3667                                 }
3668                         }
3669                         spin_unlock_irq(&conf->device_lock);
3670                 }
3671                 data_disks = disks - conf->max_degraded;
3672
3673                 new_sector = raid5_compute_sector(conf, logical_sector,
3674                                                   previous,
3675                                                   &dd_idx, NULL);
3676                 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3677                         (unsigned long long)new_sector, 
3678                         (unsigned long long)logical_sector);
3679
3680                 sh = get_active_stripe(conf, new_sector, previous,
3681                                        (bi->bi_rw&RWA_MASK), 0);
3682                 if (sh) {
3683                         if (unlikely(previous)) {
3684                                 /* expansion might have moved on while waiting for a
3685                                  * stripe, so we must do the range check again.
3686                                  * Expansion could still move past after this
3687                                  * test, but as we are holding a reference to
3688                                  * 'sh', we know that if that happens,
3689                                  *  STRIPE_EXPANDING will get set and the expansion
3690                                  * won't proceed until we finish with the stripe.
3691                                  */
3692                                 int must_retry = 0;
3693                                 spin_lock_irq(&conf->device_lock);
3694                                 if (mddev->delta_disks < 0
3695                                     ? logical_sector >= conf->reshape_progress
3696                                     : logical_sector < conf->reshape_progress)
3697                                         /* mismatch, need to try again */
3698                                         must_retry = 1;
3699                                 spin_unlock_irq(&conf->device_lock);
3700                                 if (must_retry) {
3701                                         release_stripe(sh);
3702                                         schedule();
3703                                         goto retry;
3704                                 }
3705                         }
3706
3707                         if (bio_data_dir(bi) == WRITE &&
3708                             logical_sector >= mddev->suspend_lo &&
3709                             logical_sector < mddev->suspend_hi) {
3710                                 release_stripe(sh);
3711                                 /* As the suspend_* range is controlled by
3712                                  * userspace, we want an interruptible
3713                                  * wait.
3714                                  */
3715                                 flush_signals(current);
3716                                 prepare_to_wait(&conf->wait_for_overlap,
3717                                                 &w, TASK_INTERRUPTIBLE);
3718                                 if (logical_sector >= mddev->suspend_lo &&
3719                                     logical_sector < mddev->suspend_hi)
3720                                         schedule();
3721                                 goto retry;
3722                         }
3723
3724                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3725                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3726                                 /* Stripe is busy expanding or
3727                                  * add failed due to overlap.  Flush everything
3728                                  * and wait a while
3729                                  */
3730                                 raid5_unplug_device(mddev->queue);
3731                                 release_stripe(sh);
3732                                 schedule();
3733                                 goto retry;
3734                         }
3735                         finish_wait(&conf->wait_for_overlap, &w);
3736                         set_bit(STRIPE_HANDLE, &sh->state);
3737                         clear_bit(STRIPE_DELAYED, &sh->state);
3738                         release_stripe(sh);
3739                 } else {
3740                         /* cannot get stripe for read-ahead, just give-up */
3741                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3742                         finish_wait(&conf->wait_for_overlap, &w);
3743                         break;
3744                 }
3745                         
3746         }
3747         spin_lock_irq(&conf->device_lock);
3748         remaining = raid5_dec_bi_phys_segments(bi);
3749         spin_unlock_irq(&conf->device_lock);
3750         if (remaining == 0) {
3751
3752                 if ( rw == WRITE )
3753                         md_write_end(mddev);
3754
3755                 bio_endio(bi, 0);
3756         }
3757         return 0;
3758 }
3759
3760 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
3761
3762 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3763 {
3764         /* reshaping is quite different to recovery/resync so it is
3765          * handled quite separately ... here.
3766          *
3767          * On each call to sync_request, we gather one chunk worth of
3768          * destination stripes and flag them as expanding.
3769          * Then we find all the source stripes and request reads.
3770          * As the reads complete, handle_stripe will copy the data
3771          * into the destination stripe and release that stripe.
3772          */
3773         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3774         struct stripe_head *sh;
3775         sector_t first_sector, last_sector;
3776         int raid_disks = conf->previous_raid_disks;
3777         int data_disks = raid_disks - conf->max_degraded;
3778         int new_data_disks = conf->raid_disks - conf->max_degraded;
3779         int i;
3780         int dd_idx;
3781         sector_t writepos, readpos, safepos;
3782         sector_t stripe_addr;
3783         int reshape_sectors;
3784         struct list_head stripes;
3785
3786         if (sector_nr == 0) {
3787                 /* If restarting in the middle, skip the initial sectors */
3788                 if (mddev->delta_disks < 0 &&
3789                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3790                         sector_nr = raid5_size(mddev, 0, 0)
3791                                 - conf->reshape_progress;
3792                 } else if (mddev->delta_disks >= 0 &&
3793                            conf->reshape_progress > 0)
3794                         sector_nr = conf->reshape_progress;
3795                 sector_div(sector_nr, new_data_disks);
3796                 if (sector_nr) {
3797                         *skipped = 1;
3798                         return sector_nr;
3799                 }
3800         }
3801
3802         /* We need to process a full chunk at a time.
3803          * If old and new chunk sizes differ, we need to process the
3804          * largest of these
3805          */
3806         if (mddev->new_chunk_sectors > mddev->chunk_sectors)
3807                 reshape_sectors = mddev->new_chunk_sectors;
3808         else
3809                 reshape_sectors = mddev->chunk_sectors;
3810
3811         /* we update the metadata when there is more than 3Meg
3812          * in the block range (that is rather arbitrary, should
3813          * probably be time based) or when the data about to be
3814          * copied would over-write the source of the data at
3815          * the front of the range.
3816          * i.e. one new_stripe along from reshape_progress new_maps
3817          * to after where reshape_safe old_maps to
3818          */
3819         writepos = conf->reshape_progress;
3820         sector_div(writepos, new_data_disks);
3821         readpos = conf->reshape_progress;
3822         sector_div(readpos, data_disks);
3823         safepos = conf->reshape_safe;
3824         sector_div(safepos, data_disks);
3825         if (mddev->delta_disks < 0) {
3826                 writepos -= min_t(sector_t, reshape_sectors, writepos);
3827                 readpos += reshape_sectors;
3828                 safepos += reshape_sectors;
3829         } else {
3830                 writepos += reshape_sectors;
3831                 readpos -= min_t(sector_t, reshape_sectors, readpos);
3832                 safepos -= min_t(sector_t, reshape_sectors, safepos);
3833         }
3834
3835         /* 'writepos' is the most advanced device address we might write.
3836          * 'readpos' is the least advanced device address we might read.
3837          * 'safepos' is the least address recorded in the metadata as having
3838          *     been reshaped.
3839          * If 'readpos' is behind 'writepos', then there is no way that we can
3840          * ensure safety in the face of a crash - that must be done by userspace
3841          * making a backup of the data.  So in that case there is no particular
3842          * rush to update metadata.
3843          * Otherwise if 'safepos' is behind 'writepos', then we really need to
3844          * update the metadata to advance 'safepos' to match 'readpos' so that
3845          * we can be safe in the event of a crash.
3846          * So we insist on updating metadata if safepos is behind writepos and
3847          * readpos is beyond writepos.
3848          * In any case, update the metadata every 10 seconds.
3849          * Maybe that number should be configurable, but I'm not sure it is
3850          * worth it.... maybe it could be a multiple of safemode_delay???
3851          */
3852         if ((mddev->delta_disks < 0
3853              ? (safepos > writepos && readpos < writepos)
3854              : (safepos < writepos && readpos > writepos)) ||
3855             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
3856                 /* Cannot proceed until we've updated the superblock... */
3857                 wait_event(conf->wait_for_overlap,
3858                            atomic_read(&conf->reshape_stripes)==0);
3859                 mddev->reshape_position = conf->reshape_progress;
3860                 mddev->curr_resync_completed = mddev->curr_resync;
3861                 conf->reshape_checkpoint = jiffies;
3862                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3863                 md_wakeup_thread(mddev->thread);
3864                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3865                            kthread_should_stop());
3866                 spin_lock_irq(&conf->device_lock);
3867                 conf->reshape_safe = mddev->reshape_position;
3868                 spin_unlock_irq(&conf->device_lock);
3869                 wake_up(&conf->wait_for_overlap);
3870                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3871         }
3872
3873         if (mddev->delta_disks < 0) {
3874                 BUG_ON(conf->reshape_progress == 0);
3875                 stripe_addr = writepos;
3876                 BUG_ON((mddev->dev_sectors &
3877                         ~((sector_t)reshape_sectors - 1))
3878                        - reshape_sectors - stripe_addr
3879                        != sector_nr);
3880         } else {
3881                 BUG_ON(writepos != sector_nr + reshape_sectors);
3882                 stripe_addr = sector_nr;
3883         }
3884         INIT_LIST_HEAD(&stripes);
3885         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
3886                 int j;
3887                 int skipped_disk = 0;
3888                 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
3889                 set_bit(STRIPE_EXPANDING, &sh->state);
3890                 atomic_inc(&conf->reshape_stripes);
3891                 /* If any of this stripe is beyond the end of the old
3892                  * array, then we need to zero those blocks
3893                  */
3894                 for (j=sh->disks; j--;) {
3895                         sector_t s;
3896                         if (j == sh->pd_idx)
3897                                 continue;
3898                         if (conf->level == 6 &&
3899                             j == sh->qd_idx)
3900                                 continue;
3901                         s = compute_blocknr(sh, j, 0);
3902                         if (s < raid5_size(mddev, 0, 0)) {
3903                                 skipped_disk = 1;
3904                                 continue;
3905                         }
3906                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3907                         set_bit(R5_Expanded, &sh->dev[j].flags);
3908                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
3909                 }
3910                 if (!skipped_disk) {
3911                         set_bit(STRIPE_EXPAND_READY, &sh->state);
3912                         set_bit(STRIPE_HANDLE, &sh->state);
3913                 }
3914                 list_add(&sh->lru, &stripes);
3915         }
3916         spin_lock_irq(&conf->device_lock);
3917         if (mddev->delta_disks < 0)
3918                 conf->reshape_progress -= reshape_sectors * new_data_disks;
3919         else
3920                 conf->reshape_progress += reshape_sectors * new_data_disks;
3921         spin_unlock_irq(&conf->device_lock);
3922         /* Ok, those stripe are ready. We can start scheduling
3923          * reads on the source stripes.
3924          * The source stripes are determined by mapping the first and last
3925          * block on the destination stripes.
3926          */
3927         first_sector =
3928                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
3929                                      1, &dd_idx, NULL);
3930         last_sector =
3931                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
3932                                             * new_data_disks - 1),
3933                                      1, &dd_idx, NULL);
3934         if (last_sector >= mddev->dev_sectors)
3935                 last_sector = mddev->dev_sectors - 1;
3936         while (first_sector <= last_sector) {
3937                 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
3938                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3939                 set_bit(STRIPE_HANDLE, &sh->state);
3940                 release_stripe(sh);
3941                 first_sector += STRIPE_SECTORS;
3942         }
3943         /* Now that the sources are clearly marked, we can release
3944          * the destination stripes
3945          */
3946         while (!list_empty(&stripes)) {
3947                 sh = list_entry(stripes.next, struct stripe_head, lru);
3948                 list_del_init(&sh->lru);
3949                 release_stripe(sh);
3950         }
3951         /* If this takes us to the resync_max point where we have to pause,
3952          * then we need to write out the superblock.
3953          */
3954         sector_nr += reshape_sectors;
3955         if ((sector_nr - mddev->curr_resync_completed) * 2
3956             >= mddev->resync_max - mddev->curr_resync_completed) {
3957                 /* Cannot proceed until we've updated the superblock... */
3958                 wait_event(conf->wait_for_overlap,
3959                            atomic_read(&conf->reshape_stripes) == 0);
3960                 mddev->reshape_position = conf->reshape_progress;
3961                 mddev->curr_resync_completed = mddev->curr_resync + reshape_sectors;
3962                 conf->reshape_checkpoint = jiffies;
3963                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3964                 md_wakeup_thread(mddev->thread);
3965                 wait_event(mddev->sb_wait,
3966                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3967                            || kthread_should_stop());
3968                 spin_lock_irq(&conf->device_lock);
3969                 conf->reshape_safe = mddev->reshape_position;
3970                 spin_unlock_irq(&conf->device_lock);
3971                 wake_up(&conf->wait_for_overlap);
3972                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3973         }
3974         return reshape_sectors;
3975 }
3976
3977 /* FIXME go_faster isn't used */
3978 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3979 {
3980         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3981         struct stripe_head *sh;
3982         sector_t max_sector = mddev->dev_sectors;
3983         int sync_blocks;
3984         int still_degraded = 0;
3985         int i;
3986
3987         if (sector_nr >= max_sector) {
3988                 /* just being told to finish up .. nothing much to do */
3989                 unplug_slaves(mddev);
3990
3991                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3992                         end_reshape(conf);
3993                         return 0;
3994                 }
3995
3996                 if (mddev->curr_resync < max_sector) /* aborted */
3997                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3998                                         &sync_blocks, 1);
3999                 else /* completed sync */
4000                         conf->fullsync = 0;
4001                 bitmap_close_sync(mddev->bitmap);
4002
4003                 return 0;
4004         }
4005
4006         /* Allow raid5_quiesce to complete */
4007         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4008
4009         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4010                 return reshape_request(mddev, sector_nr, skipped);
4011
4012         /* No need to check resync_max as we never do more than one
4013          * stripe, and as resync_max will always be on a chunk boundary,
4014          * if the check in md_do_sync didn't fire, there is no chance
4015          * of overstepping resync_max here
4016          */
4017
4018         /* if there is too many failed drives and we are trying
4019          * to resync, then assert that we are finished, because there is
4020          * nothing we can do.
4021          */
4022         if (mddev->degraded >= conf->max_degraded &&
4023             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4024                 sector_t rv = mddev->dev_sectors - sector_nr;
4025                 *skipped = 1;
4026                 return rv;
4027         }
4028         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4029             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4030             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4031                 /* we can skip this block, and probably more */
4032                 sync_blocks /= STRIPE_SECTORS;
4033                 *skipped = 1;
4034                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4035         }
4036
4037
4038         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4039
4040         sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4041         if (sh == NULL) {
4042                 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4043                 /* make sure we don't swamp the stripe cache if someone else
4044                  * is trying to get access
4045                  */
4046                 schedule_timeout_uninterruptible(1);
4047         }
4048         /* Need to check if array will still be degraded after recovery/resync
4049          * We don't need to check the 'failed' flag as when that gets set,
4050          * recovery aborts.
4051          */
4052         for (i = 0; i < conf->raid_disks; i++)
4053                 if (conf->disks[i].rdev == NULL)
4054                         still_degraded = 1;
4055
4056         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4057
4058         spin_lock(&sh->lock);
4059         set_bit(STRIPE_SYNCING, &sh->state);
4060         clear_bit(STRIPE_INSYNC, &sh->state);
4061         spin_unlock(&sh->lock);
4062
4063         /* wait for any blocked device to be handled */
4064         while(unlikely(!handle_stripe(sh, NULL)))
4065                 ;
4066         release_stripe(sh);
4067
4068         return STRIPE_SECTORS;
4069 }
4070
4071 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4072 {
4073         /* We may not be able to submit a whole bio at once as there
4074          * may not be enough stripe_heads available.
4075          * We cannot pre-allocate enough stripe_heads as we may need
4076          * more than exist in the cache (if we allow ever large chunks).
4077          * So we do one stripe head at a time and record in
4078          * ->bi_hw_segments how many have been done.
4079          *
4080          * We *know* that this entire raid_bio is in one chunk, so
4081          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4082          */
4083         struct stripe_head *sh;
4084         int dd_idx;
4085         sector_t sector, logical_sector, last_sector;
4086         int scnt = 0;
4087         int remaining;
4088         int handled = 0;
4089
4090         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4091         sector = raid5_compute_sector(conf, logical_sector,
4092                                       0, &dd_idx, NULL);
4093         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4094
4095         for (; logical_sector < last_sector;
4096              logical_sector += STRIPE_SECTORS,
4097                      sector += STRIPE_SECTORS,
4098                      scnt++) {
4099
4100                 if (scnt < raid5_bi_hw_segments(raid_bio))
4101                         /* already done this stripe */
4102                         continue;
4103
4104                 sh = get_active_stripe(conf, sector, 0, 1, 0);
4105
4106                 if (!sh) {
4107                         /* failed to get a stripe - must wait */
4108                         raid5_set_bi_hw_segments(raid_bio, scnt);
4109                         conf->retry_read_aligned = raid_bio;
4110                         return handled;
4111                 }
4112
4113                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4114                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4115                         release_stripe(sh);
4116                         raid5_set_bi_hw_segments(raid_bio, scnt);
4117                         conf->retry_read_aligned = raid_bio;
4118                         return handled;
4119                 }
4120
4121                 handle_stripe(sh, NULL);
4122                 release_stripe(sh);
4123                 handled++;
4124         }
4125         spin_lock_irq(&conf->device_lock);
4126         remaining = raid5_dec_bi_phys_segments(raid_bio);
4127         spin_unlock_irq(&conf->device_lock);
4128         if (remaining == 0)
4129                 bio_endio(raid_bio, 0);
4130         if (atomic_dec_and_test(&conf->active_aligned_reads))
4131                 wake_up(&conf->wait_for_stripe);
4132         return handled;
4133 }
4134
4135
4136
4137 /*
4138  * This is our raid5 kernel thread.
4139  *
4140  * We scan the hash table for stripes which can be handled now.
4141  * During the scan, completed stripes are saved for us by the interrupt
4142  * handler, so that they will not have to wait for our next wakeup.
4143  */
4144 static void raid5d(mddev_t *mddev)
4145 {
4146         struct stripe_head *sh;
4147         raid5_conf_t *conf = mddev->private;
4148         int handled;
4149
4150         pr_debug("+++ raid5d active\n");
4151
4152         md_check_recovery(mddev);
4153
4154         handled = 0;
4155         spin_lock_irq(&conf->device_lock);
4156         while (1) {
4157                 struct bio *bio;
4158
4159                 if (conf->seq_flush != conf->seq_write) {
4160                         int seq = conf->seq_flush;
4161                         spin_unlock_irq(&conf->device_lock);
4162                         bitmap_unplug(mddev->bitmap);
4163                         spin_lock_irq(&conf->device_lock);
4164                         conf->seq_write = seq;
4165                         activate_bit_delay(conf);
4166                 }
4167
4168                 while ((bio = remove_bio_from_retry(conf))) {
4169                         int ok;
4170                         spin_unlock_irq(&conf->device_lock);
4171                         ok = retry_aligned_read(conf, bio);
4172                         spin_lock_irq(&conf->device_lock);
4173                         if (!ok)
4174                                 break;
4175                         handled++;
4176                 }
4177
4178                 sh = __get_priority_stripe(conf);
4179
4180                 if (!sh)
4181                         break;
4182                 spin_unlock_irq(&conf->device_lock);
4183                 
4184                 handled++;
4185                 handle_stripe(sh, conf->spare_page);
4186                 release_stripe(sh);
4187
4188                 spin_lock_irq(&conf->device_lock);
4189         }
4190         pr_debug("%d stripes handled\n", handled);
4191
4192         spin_unlock_irq(&conf->device_lock);
4193
4194         async_tx_issue_pending_all();
4195         unplug_slaves(mddev);
4196
4197         pr_debug("--- raid5d inactive\n");
4198 }
4199
4200 static ssize_t
4201 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4202 {
4203         raid5_conf_t *conf = mddev->private;
4204         if (conf)
4205                 return sprintf(page, "%d\n", conf->max_nr_stripes);
4206         else
4207                 return 0;
4208 }
4209
4210 static ssize_t
4211 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4212 {
4213         raid5_conf_t *conf = mddev->private;
4214         unsigned long new;
4215         int err;
4216
4217         if (len >= PAGE_SIZE)
4218                 return -EINVAL;
4219         if (!conf)
4220                 return -ENODEV;
4221
4222         if (strict_strtoul(page, 10, &new))
4223                 return -EINVAL;
4224         if (new <= 16 || new > 32768)
4225                 return -EINVAL;
4226         while (new < conf->max_nr_stripes) {
4227                 if (drop_one_stripe(conf))
4228                         conf->max_nr_stripes--;
4229                 else
4230                         break;
4231         }
4232         err = md_allow_write(mddev);
4233         if (err)
4234                 return err;
4235         while (new > conf->max_nr_stripes) {
4236                 if (grow_one_stripe(conf))
4237                         conf->max_nr_stripes++;
4238                 else break;
4239         }
4240         return len;
4241 }
4242
4243 static struct md_sysfs_entry
4244 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4245                                 raid5_show_stripe_cache_size,
4246                                 raid5_store_stripe_cache_size);
4247
4248 static ssize_t
4249 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4250 {
4251         raid5_conf_t *conf = mddev->private;
4252         if (conf)
4253                 return sprintf(page, "%d\n", conf->bypass_threshold);
4254         else
4255                 return 0;
4256 }
4257
4258 static ssize_t
4259 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4260 {
4261         raid5_conf_t *conf = mddev->private;
4262         unsigned long new;
4263         if (len >= PAGE_SIZE)
4264                 return -EINVAL;
4265         if (!conf)
4266                 return -ENODEV;
4267
4268         if (strict_strtoul(page, 10, &new))
4269                 return -EINVAL;
4270         if (new > conf->max_nr_stripes)
4271                 return -EINVAL;
4272         conf->bypass_threshold = new;
4273         return len;
4274 }
4275
4276 static struct md_sysfs_entry
4277 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4278                                         S_IRUGO | S_IWUSR,
4279                                         raid5_show_preread_threshold,
4280                                         raid5_store_preread_threshold);
4281
4282 static ssize_t
4283 stripe_cache_active_show(mddev_t *mddev, char *page)
4284 {
4285         raid5_conf_t *conf = mddev->private;
4286         if (conf)
4287                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4288         else
4289                 return 0;
4290 }
4291
4292 static struct md_sysfs_entry
4293 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4294
4295 static struct attribute *raid5_attrs[] =  {
4296         &raid5_stripecache_size.attr,
4297         &raid5_stripecache_active.attr,
4298         &raid5_preread_bypass_threshold.attr,
4299         NULL,
4300 };
4301 static struct attribute_group raid5_attrs_group = {
4302         .name = NULL,
4303         .attrs = raid5_attrs,
4304 };
4305
4306 static sector_t
4307 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4308 {
4309         raid5_conf_t *conf = mddev->private;
4310
4311         if (!sectors)
4312                 sectors = mddev->dev_sectors;
4313         if (!raid_disks) {
4314                 /* size is defined by the smallest of previous and new size */
4315                 if (conf->raid_disks < conf->previous_raid_disks)
4316                         raid_disks = conf->raid_disks;
4317                 else
4318                         raid_disks = conf->previous_raid_disks;
4319         }
4320
4321         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4322         sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4323         return sectors * (raid_disks - conf->max_degraded);
4324 }
4325
4326 static void free_conf(raid5_conf_t *conf)
4327 {
4328         shrink_stripes(conf);
4329         safe_put_page(conf->spare_page);
4330         kfree(conf->disks);
4331         kfree(conf->stripe_hashtbl);
4332         kfree(conf);
4333 }
4334
4335 static raid5_conf_t *setup_conf(mddev_t *mddev)
4336 {
4337         raid5_conf_t *conf;
4338         int raid_disk, memory;
4339         mdk_rdev_t *rdev;
4340         struct disk_info *disk;
4341
4342         if (mddev->new_level != 5
4343             && mddev->new_level != 4
4344             && mddev->new_level != 6) {
4345                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4346                        mdname(mddev), mddev->new_level);
4347                 return ERR_PTR(-EIO);
4348         }
4349         if ((mddev->new_level == 5
4350              && !algorithm_valid_raid5(mddev->new_layout)) ||
4351             (mddev->new_level == 6
4352              && !algorithm_valid_raid6(mddev->new_layout))) {
4353                 printk(KERN_ERR "raid5: %s: layout %d not supported\n",
4354                        mdname(mddev), mddev->new_layout);
4355                 return ERR_PTR(-EIO);
4356         }
4357         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4358                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4359                        mdname(mddev), mddev->raid_disks);
4360                 return ERR_PTR(-EINVAL);
4361         }
4362
4363         if (!mddev->new_chunk_sectors ||
4364             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4365             !is_power_of_2(mddev->new_chunk_sectors)) {
4366                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4367                        mddev->new_chunk_sectors << 9, mdname(mddev));
4368                 return ERR_PTR(-EINVAL);
4369         }
4370
4371         conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4372         if (conf == NULL)
4373                 goto abort;
4374
4375         conf->raid_disks = mddev->raid_disks;
4376         if (mddev->reshape_position == MaxSector)
4377                 conf->previous_raid_disks = mddev->raid_disks;
4378         else
4379                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4380
4381         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4382                               GFP_KERNEL);
4383         if (!conf->disks)
4384                 goto abort;
4385
4386         conf->mddev = mddev;
4387
4388         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4389                 goto abort;
4390
4391         if (mddev->new_level == 6) {
4392                 conf->spare_page = alloc_page(GFP_KERNEL);
4393                 if (!conf->spare_page)
4394                         goto abort;
4395         }
4396         spin_lock_init(&conf->device_lock);
4397         init_waitqueue_head(&conf->wait_for_stripe);
4398         init_waitqueue_head(&conf->wait_for_overlap);
4399         INIT_LIST_HEAD(&conf->handle_list);
4400         INIT_LIST_HEAD(&conf->hold_list);
4401         INIT_LIST_HEAD(&conf->delayed_list);
4402         INIT_LIST_HEAD(&conf->bitmap_list);
4403         INIT_LIST_HEAD(&conf->inactive_list);
4404         atomic_set(&conf->active_stripes, 0);
4405         atomic_set(&conf->preread_active_stripes, 0);
4406         atomic_set(&conf->active_aligned_reads, 0);
4407         conf->bypass_threshold = BYPASS_THRESHOLD;
4408
4409         pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4410
4411         list_for_each_entry(rdev, &mddev->disks, same_set) {
4412                 raid_disk = rdev->raid_disk;
4413                 if (raid_disk >= conf->raid_disks
4414                     || raid_disk < 0)
4415                         continue;
4416                 disk = conf->disks + raid_disk;
4417
4418                 disk->rdev = rdev;
4419
4420                 if (test_bit(In_sync, &rdev->flags)) {
4421                         char b[BDEVNAME_SIZE];
4422                         printk(KERN_INFO "raid5: device %s operational as raid"
4423                                 " disk %d\n", bdevname(rdev->bdev,b),
4424                                 raid_disk);
4425                 } else
4426                         /* Cannot rely on bitmap to complete recovery */
4427                         conf->fullsync = 1;
4428         }
4429
4430         conf->chunk_sectors = mddev->new_chunk_sectors;
4431         conf->level = mddev->new_level;
4432         if (conf->level == 6)
4433                 conf->max_degraded = 2;
4434         else
4435                 conf->max_degraded = 1;
4436         conf->algorithm = mddev->new_layout;
4437         conf->max_nr_stripes = NR_STRIPES;
4438         conf->reshape_progress = mddev->reshape_position;
4439         if (conf->reshape_progress != MaxSector) {
4440                 conf->prev_chunk_sectors = mddev->chunk_sectors;
4441                 conf->prev_algo = mddev->layout;
4442         }
4443
4444         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4445                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4446         if (grow_stripes(conf, conf->max_nr_stripes)) {
4447                 printk(KERN_ERR
4448                         "raid5: couldn't allocate %dkB for buffers\n", memory);
4449                 goto abort;
4450         } else
4451                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4452                         memory, mdname(mddev));
4453
4454         conf->thread = md_register_thread(raid5d, mddev, NULL);
4455         if (!conf->thread) {
4456                 printk(KERN_ERR
4457                        "raid5: couldn't allocate thread for %s\n",
4458                        mdname(mddev));
4459                 goto abort;
4460         }
4461
4462         return conf;
4463
4464  abort:
4465         if (conf) {
4466                 free_conf(conf);
4467                 return ERR_PTR(-EIO);
4468         } else
4469                 return ERR_PTR(-ENOMEM);
4470 }
4471
4472 static int run(mddev_t *mddev)
4473 {
4474         raid5_conf_t *conf;
4475         int working_disks = 0, chunk_size;
4476         mdk_rdev_t *rdev;
4477
4478         if (mddev->recovery_cp != MaxSector)
4479                 printk(KERN_NOTICE "raid5: %s is not clean"
4480                        " -- starting background reconstruction\n",
4481                        mdname(mddev));
4482         if (mddev->reshape_position != MaxSector) {
4483                 /* Check that we can continue the reshape.
4484                  * Currently only disks can change, it must
4485                  * increase, and we must be past the point where
4486                  * a stripe over-writes itself
4487                  */
4488                 sector_t here_new, here_old;
4489                 int old_disks;
4490                 int max_degraded = (mddev->level == 6 ? 2 : 1);
4491
4492                 if (mddev->new_level != mddev->level) {
4493                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4494                                "required - aborting.\n",
4495                                mdname(mddev));
4496                         return -EINVAL;
4497                 }
4498                 old_disks = mddev->raid_disks - mddev->delta_disks;
4499                 /* reshape_position must be on a new-stripe boundary, and one
4500                  * further up in new geometry must map after here in old
4501                  * geometry.
4502                  */
4503                 here_new = mddev->reshape_position;
4504                 if (sector_div(here_new, mddev->new_chunk_sectors *
4505                                (mddev->raid_disks - max_degraded))) {
4506                         printk(KERN_ERR "raid5: reshape_position not "
4507                                "on a stripe boundary\n");
4508                         return -EINVAL;
4509                 }
4510                 /* here_new is the stripe we will write to */
4511                 here_old = mddev->reshape_position;
4512                 sector_div(here_old, mddev->chunk_sectors *
4513                            (old_disks-max_degraded));
4514                 /* here_old is the first stripe that we might need to read
4515                  * from */
4516                 if (mddev->delta_disks == 0) {
4517                         /* We cannot be sure it is safe to start an in-place
4518                          * reshape.  It is only safe if user-space if monitoring
4519                          * and taking constant backups.
4520                          * mdadm always starts a situation like this in
4521                          * readonly mode so it can take control before
4522                          * allowing any writes.  So just check for that.
4523                          */
4524                         if ((here_new * mddev->new_chunk_sectors != 
4525                              here_old * mddev->chunk_sectors) ||
4526                             mddev->ro == 0) {
4527                                 printk(KERN_ERR "raid5: in-place reshape must be started"
4528                                        " in read-only mode - aborting\n");
4529                                 return -EINVAL;
4530                         }
4531                 } else if (mddev->delta_disks < 0
4532                     ? (here_new * mddev->new_chunk_sectors <=
4533                        here_old * mddev->chunk_sectors)
4534                     : (here_new * mddev->new_chunk_sectors >=
4535                        here_old * mddev->chunk_sectors)) {
4536                         /* Reading from the same stripe as writing to - bad */
4537                         printk(KERN_ERR "raid5: reshape_position too early for "
4538                                "auto-recovery - aborting.\n");
4539                         return -EINVAL;
4540                 }
4541                 printk(KERN_INFO "raid5: reshape will continue\n");
4542                 /* OK, we should be able to continue; */
4543         } else {
4544                 BUG_ON(mddev->level != mddev->new_level);
4545                 BUG_ON(mddev->layout != mddev->new_layout);
4546                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4547                 BUG_ON(mddev->delta_disks != 0);
4548         }
4549
4550         if (mddev->private == NULL)
4551                 conf = setup_conf(mddev);
4552         else
4553                 conf = mddev->private;
4554
4555         if (IS_ERR(conf))
4556                 return PTR_ERR(conf);
4557
4558         mddev->thread = conf->thread;
4559         conf->thread = NULL;
4560         mddev->private = conf;
4561
4562         /*
4563          * 0 for a fully functional array, 1 or 2 for a degraded array.
4564          */
4565         list_for_each_entry(rdev, &mddev->disks, same_set)
4566                 if (rdev->raid_disk >= 0 &&
4567                     test_bit(In_sync, &rdev->flags))
4568                         working_disks++;
4569
4570         mddev->degraded = conf->raid_disks - working_disks;
4571
4572         if (mddev->degraded > conf->max_degraded) {
4573                 printk(KERN_ERR "raid5: not enough operational devices for %s"
4574                         " (%d/%d failed)\n",
4575                         mdname(mddev), mddev->degraded, conf->raid_disks);
4576                 goto abort;
4577         }
4578
4579         /* device size must be a multiple of chunk size */
4580         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
4581         mddev->resync_max_sectors = mddev->dev_sectors;
4582
4583         if (mddev->degraded > 0 &&
4584             mddev->recovery_cp != MaxSector) {
4585                 if (mddev->ok_start_degraded)
4586                         printk(KERN_WARNING
4587                                "raid5: starting dirty degraded array: %s"
4588                                "- data corruption possible.\n",
4589                                mdname(mddev));
4590                 else {
4591                         printk(KERN_ERR
4592                                "raid5: cannot start dirty degraded array for %s\n",
4593                                mdname(mddev));
4594                         goto abort;
4595                 }
4596         }
4597
4598         if (mddev->degraded == 0)
4599                 printk("raid5: raid level %d set %s active with %d out of %d"
4600                        " devices, algorithm %d\n", conf->level, mdname(mddev),
4601                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4602                        mddev->new_layout);
4603         else
4604                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4605                         " out of %d devices, algorithm %d\n", conf->level,
4606                         mdname(mddev), mddev->raid_disks - mddev->degraded,
4607                         mddev->raid_disks, mddev->new_layout);
4608
4609         print_raid5_conf(conf);
4610
4611         if (conf->reshape_progress != MaxSector) {
4612                 printk("...ok start reshape thread\n");
4613                 conf->reshape_safe = conf->reshape_progress;
4614                 atomic_set(&conf->reshape_stripes, 0);
4615                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4616                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4617                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4618                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4619                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4620                                                         "reshape");
4621         }
4622
4623         /* read-ahead size must cover two whole stripes, which is
4624          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4625          */
4626         {
4627                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4628                 int stripe = data_disks *
4629                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
4630                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4631                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4632         }
4633
4634         /* Ok, everything is just fine now */
4635         if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4636                 printk(KERN_WARNING
4637                        "raid5: failed to create sysfs attributes for %s\n",
4638                        mdname(mddev));
4639
4640         mddev->queue->queue_lock = &conf->device_lock;
4641
4642         mddev->queue->unplug_fn = raid5_unplug_device;
4643         mddev->queue->backing_dev_info.congested_data = mddev;
4644         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4645
4646         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4647
4648         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4649         chunk_size = mddev->chunk_sectors << 9;
4650         blk_queue_io_min(mddev->queue, chunk_size);
4651         blk_queue_io_opt(mddev->queue, chunk_size *
4652                          (conf->raid_disks - conf->max_degraded));
4653
4654         list_for_each_entry(rdev, &mddev->disks, same_set)
4655                 disk_stack_limits(mddev->gendisk, rdev->bdev,
4656                                   rdev->data_offset << 9);
4657
4658         return 0;
4659 abort:
4660         md_unregister_thread(mddev->thread);
4661         mddev->thread = NULL;
4662         if (conf) {
4663                 print_raid5_conf(conf);
4664                 free_conf(conf);
4665         }
4666         mddev->private = NULL;
4667         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4668         return -EIO;
4669 }
4670
4671
4672
4673 static int stop(mddev_t *mddev)
4674 {
4675         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4676
4677         md_unregister_thread(mddev->thread);
4678         mddev->thread = NULL;
4679         mddev->queue->backing_dev_info.congested_fn = NULL;
4680         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4681         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4682         free_conf(conf);
4683         mddev->private = NULL;
4684         return 0;
4685 }
4686
4687 #ifdef DEBUG
4688 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
4689 {
4690         int i;
4691
4692         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4693                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4694         seq_printf(seq, "sh %llu,  count %d.\n",
4695                    (unsigned long long)sh->sector, atomic_read(&sh->count));
4696         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4697         for (i = 0; i < sh->disks; i++) {
4698                 seq_printf(seq, "(cache%d: %p %ld) ",
4699                            i, sh->dev[i].page, sh->dev[i].flags);
4700         }
4701         seq_printf(seq, "\n");
4702 }
4703
4704 static void printall(struct seq_file *seq, raid5_conf_t *conf)
4705 {
4706         struct stripe_head *sh;
4707         struct hlist_node *hn;
4708         int i;
4709
4710         spin_lock_irq(&conf->device_lock);
4711         for (i = 0; i < NR_HASH; i++) {
4712                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4713                         if (sh->raid_conf != conf)
4714                                 continue;
4715                         print_sh(seq, sh);
4716                 }
4717         }
4718         spin_unlock_irq(&conf->device_lock);
4719 }
4720 #endif
4721
4722 static void status(struct seq_file *seq, mddev_t *mddev)
4723 {
4724         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4725         int i;
4726
4727         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
4728                 mddev->chunk_sectors / 2, mddev->layout);
4729         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4730         for (i = 0; i < conf->raid_disks; i++)
4731                 seq_printf (seq, "%s",
4732                                conf->disks[i].rdev &&
4733                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4734         seq_printf (seq, "]");
4735 #ifdef DEBUG
4736         seq_printf (seq, "\n");
4737         printall(seq, conf);
4738 #endif
4739 }
4740
4741 static void print_raid5_conf (raid5_conf_t *conf)
4742 {
4743         int i;
4744         struct disk_info *tmp;
4745
4746         printk("RAID5 conf printout:\n");
4747         if (!conf) {
4748                 printk("(conf==NULL)\n");
4749                 return;
4750         }
4751         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4752                  conf->raid_disks - conf->mddev->degraded);
4753
4754         for (i = 0; i < conf->raid_disks; i++) {
4755                 char b[BDEVNAME_SIZE];
4756                 tmp = conf->disks + i;
4757                 if (tmp->rdev)
4758                 printk(" disk %d, o:%d, dev:%s\n",
4759                         i, !test_bit(Faulty, &tmp->rdev->flags),
4760                         bdevname(tmp->rdev->bdev,b));
4761         }
4762 }
4763
4764 static int raid5_spare_active(mddev_t *mddev)
4765 {
4766         int i;
4767         raid5_conf_t *conf = mddev->private;
4768         struct disk_info *tmp;
4769
4770         for (i = 0; i < conf->raid_disks; i++) {
4771                 tmp = conf->disks + i;
4772                 if (tmp->rdev
4773                     && !test_bit(Faulty, &tmp->rdev->flags)
4774                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4775                         unsigned long flags;
4776                         spin_lock_irqsave(&conf->device_lock, flags);
4777                         mddev->degraded--;
4778                         spin_unlock_irqrestore(&conf->device_lock, flags);
4779                 }
4780         }
4781         print_raid5_conf(conf);
4782         return 0;
4783 }
4784
4785 static int raid5_remove_disk(mddev_t *mddev, int number)
4786 {
4787         raid5_conf_t *conf = mddev->private;
4788         int err = 0;
4789         mdk_rdev_t *rdev;
4790         struct disk_info *p = conf->disks + number;
4791
4792         print_raid5_conf(conf);
4793         rdev = p->rdev;
4794         if (rdev) {
4795                 if (number >= conf->raid_disks &&
4796                     conf->reshape_progress == MaxSector)
4797                         clear_bit(In_sync, &rdev->flags);
4798
4799                 if (test_bit(In_sync, &rdev->flags) ||
4800                     atomic_read(&rdev->nr_pending)) {
4801                         err = -EBUSY;
4802                         goto abort;
4803                 }
4804                 /* Only remove non-faulty devices if recovery
4805                  * isn't possible.
4806                  */
4807                 if (!test_bit(Faulty, &rdev->flags) &&
4808                     mddev->degraded <= conf->max_degraded &&
4809                     number < conf->raid_disks) {
4810                         err = -EBUSY;
4811                         goto abort;
4812                 }
4813                 p->rdev = NULL;
4814                 synchronize_rcu();
4815                 if (atomic_read(&rdev->nr_pending)) {
4816                         /* lost the race, try later */
4817                         err = -EBUSY;
4818                         p->rdev = rdev;
4819                 }
4820         }
4821 abort:
4822
4823         print_raid5_conf(conf);
4824         return err;
4825 }
4826
4827 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4828 {
4829         raid5_conf_t *conf = mddev->private;
4830         int err = -EEXIST;
4831         int disk;
4832         struct disk_info *p;
4833         int first = 0;
4834         int last = conf->raid_disks - 1;
4835
4836         if (mddev->degraded > conf->max_degraded)
4837                 /* no point adding a device */
4838                 return -EINVAL;
4839
4840         if (rdev->raid_disk >= 0)
4841                 first = last = rdev->raid_disk;
4842
4843         /*
4844          * find the disk ... but prefer rdev->saved_raid_disk
4845          * if possible.
4846          */
4847         if (rdev->saved_raid_disk >= 0 &&
4848             rdev->saved_raid_disk >= first &&
4849             conf->disks[rdev->saved_raid_disk].rdev == NULL)
4850                 disk = rdev->saved_raid_disk;
4851         else
4852                 disk = first;
4853         for ( ; disk <= last ; disk++)
4854                 if ((p=conf->disks + disk)->rdev == NULL) {
4855                         clear_bit(In_sync, &rdev->flags);
4856                         rdev->raid_disk = disk;
4857                         err = 0;
4858                         if (rdev->saved_raid_disk != disk)
4859                                 conf->fullsync = 1;
4860                         rcu_assign_pointer(p->rdev, rdev);
4861                         break;
4862                 }
4863         print_raid5_conf(conf);
4864         return err;
4865 }
4866
4867 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4868 {
4869         /* no resync is happening, and there is enough space
4870          * on all devices, so we can resize.
4871          * We need to make sure resync covers any new space.
4872          * If the array is shrinking we should possibly wait until
4873          * any io in the removed space completes, but it hardly seems
4874          * worth it.
4875          */
4876         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4877         md_set_array_sectors(mddev, raid5_size(mddev, sectors,
4878                                                mddev->raid_disks));
4879         if (mddev->array_sectors >
4880             raid5_size(mddev, sectors, mddev->raid_disks))
4881                 return -EINVAL;
4882         set_capacity(mddev->gendisk, mddev->array_sectors);
4883         mddev->changed = 1;
4884         revalidate_disk(mddev->gendisk);
4885         if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
4886                 mddev->recovery_cp = mddev->dev_sectors;
4887                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4888         }
4889         mddev->dev_sectors = sectors;
4890         mddev->resync_max_sectors = sectors;
4891         return 0;
4892 }
4893
4894 static int check_stripe_cache(mddev_t *mddev)
4895 {
4896         /* Can only proceed if there are plenty of stripe_heads.
4897          * We need a minimum of one full stripe,, and for sensible progress
4898          * it is best to have about 4 times that.
4899          * If we require 4 times, then the default 256 4K stripe_heads will
4900          * allow for chunk sizes up to 256K, which is probably OK.
4901          * If the chunk size is greater, user-space should request more
4902          * stripe_heads first.
4903          */
4904         raid5_conf_t *conf = mddev->private;
4905         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
4906             > conf->max_nr_stripes ||
4907             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
4908             > conf->max_nr_stripes) {
4909                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
4910                        ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
4911                         / STRIPE_SIZE)*4);
4912                 return 0;
4913         }
4914         return 1;
4915 }
4916
4917 static int check_reshape(mddev_t *mddev)
4918 {
4919         raid5_conf_t *conf = mddev->private;
4920
4921         if (mddev->delta_disks == 0 &&
4922             mddev->new_layout == mddev->layout &&
4923             mddev->new_chunk_sectors == mddev->chunk_sectors)
4924                 return 0; /* nothing to do */
4925         if (mddev->bitmap)
4926                 /* Cannot grow a bitmap yet */
4927                 return -EBUSY;
4928         if (mddev->degraded > conf->max_degraded)
4929                 return -EINVAL;
4930         if (mddev->delta_disks < 0) {
4931                 /* We might be able to shrink, but the devices must
4932                  * be made bigger first.
4933                  * For raid6, 4 is the minimum size.
4934                  * Otherwise 2 is the minimum
4935                  */
4936                 int min = 2;
4937                 if (mddev->level == 6)
4938                         min = 4;
4939                 if (mddev->raid_disks + mddev->delta_disks < min)
4940                         return -EINVAL;
4941         }
4942
4943         if (!check_stripe_cache(mddev))
4944                 return -ENOSPC;
4945
4946         return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4947 }
4948
4949 static int raid5_start_reshape(mddev_t *mddev)
4950 {
4951         raid5_conf_t *conf = mddev->private;
4952         mdk_rdev_t *rdev;
4953         int spares = 0;
4954         int added_devices = 0;
4955         unsigned long flags;
4956
4957         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4958                 return -EBUSY;
4959
4960         if (!check_stripe_cache(mddev))
4961                 return -ENOSPC;
4962
4963         list_for_each_entry(rdev, &mddev->disks, same_set)
4964                 if (rdev->raid_disk < 0 &&
4965                     !test_bit(Faulty, &rdev->flags))
4966                         spares++;
4967
4968         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4969                 /* Not enough devices even to make a degraded array
4970                  * of that size
4971                  */
4972                 return -EINVAL;
4973
4974         /* Refuse to reduce size of the array.  Any reductions in
4975          * array size must be through explicit setting of array_size
4976          * attribute.
4977          */
4978         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
4979             < mddev->array_sectors) {
4980                 printk(KERN_ERR "md: %s: array size must be reduced "
4981                        "before number of disks\n", mdname(mddev));
4982                 return -EINVAL;
4983         }
4984
4985         atomic_set(&conf->reshape_stripes, 0);
4986         spin_lock_irq(&conf->device_lock);
4987         conf->previous_raid_disks = conf->raid_disks;
4988         conf->raid_disks += mddev->delta_disks;
4989         conf->prev_chunk_sectors = conf->chunk_sectors;
4990         conf->chunk_sectors = mddev->new_chunk_sectors;
4991         conf->prev_algo = conf->algorithm;
4992         conf->algorithm = mddev->new_layout;
4993         if (mddev->delta_disks < 0)
4994                 conf->reshape_progress = raid5_size(mddev, 0, 0);
4995         else
4996                 conf->reshape_progress = 0;
4997         conf->reshape_safe = conf->reshape_progress;
4998         conf->generation++;
4999         spin_unlock_irq(&conf->device_lock);
5000
5001         /* Add some new drives, as many as will fit.
5002          * We know there are enough to make the newly sized array work.
5003          */
5004         list_for_each_entry(rdev, &mddev->disks, same_set)
5005                 if (rdev->raid_disk < 0 &&
5006                     !test_bit(Faulty, &rdev->flags)) {
5007                         if (raid5_add_disk(mddev, rdev) == 0) {
5008                                 char nm[20];
5009                                 set_bit(In_sync, &rdev->flags);
5010                                 added_devices++;
5011                                 rdev->recovery_offset = 0;
5012                                 sprintf(nm, "rd%d", rdev->raid_disk);
5013                                 if (sysfs_create_link(&mddev->kobj,
5014                                                       &rdev->kobj, nm))
5015                                         printk(KERN_WARNING
5016                                                "raid5: failed to create "
5017                                                " link %s for %s\n",
5018                                                nm, mdname(mddev));
5019                         } else
5020                                 break;
5021                 }
5022
5023         if (mddev->delta_disks > 0) {
5024                 spin_lock_irqsave(&conf->device_lock, flags);
5025                 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks)
5026                         - added_devices;
5027                 spin_unlock_irqrestore(&conf->device_lock, flags);
5028         }
5029         mddev->raid_disks = conf->raid_disks;
5030         mddev->reshape_position = conf->reshape_progress;
5031         set_bit(MD_CHANGE_DEVS, &mddev->flags);
5032
5033         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5034         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5035         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5036         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5037         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5038                                                 "reshape");
5039         if (!mddev->sync_thread) {
5040                 mddev->recovery = 0;
5041                 spin_lock_irq(&conf->device_lock);
5042                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5043                 conf->reshape_progress = MaxSector;
5044                 spin_unlock_irq(&conf->device_lock);
5045                 return -EAGAIN;
5046         }
5047         conf->reshape_checkpoint = jiffies;
5048         md_wakeup_thread(mddev->sync_thread);
5049         md_new_event(mddev);
5050         return 0;
5051 }
5052
5053 /* This is called from the reshape thread and should make any
5054  * changes needed in 'conf'
5055  */
5056 static void end_reshape(raid5_conf_t *conf)
5057 {
5058
5059         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5060
5061                 spin_lock_irq(&conf->device_lock);
5062                 conf->previous_raid_disks = conf->raid_disks;
5063                 conf->reshape_progress = MaxSector;
5064                 spin_unlock_irq(&conf->device_lock);
5065                 wake_up(&conf->wait_for_overlap);
5066
5067                 /* read-ahead size must cover two whole stripes, which is
5068                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5069                  */
5070                 {
5071                         int data_disks = conf->raid_disks - conf->max_degraded;
5072                         int stripe = data_disks * ((conf->chunk_sectors << 9)
5073                                                    / PAGE_SIZE);
5074                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5075                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5076                 }
5077         }
5078 }
5079
5080 /* This is called from the raid5d thread with mddev_lock held.
5081  * It makes config changes to the device.
5082  */
5083 static void raid5_finish_reshape(mddev_t *mddev)
5084 {
5085         raid5_conf_t *conf = mddev->private;
5086
5087         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5088
5089                 if (mddev->delta_disks > 0) {
5090                         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5091                         set_capacity(mddev->gendisk, mddev->array_sectors);
5092                         mddev->changed = 1;
5093                         revalidate_disk(mddev->gendisk);
5094                 } else {
5095                         int d;
5096                         mddev->degraded = conf->raid_disks;
5097                         for (d = 0; d < conf->raid_disks ; d++)
5098                                 if (conf->disks[d].rdev &&
5099                                     test_bit(In_sync,
5100                                              &conf->disks[d].rdev->flags))
5101                                         mddev->degraded--;
5102                         for (d = conf->raid_disks ;
5103                              d < conf->raid_disks - mddev->delta_disks;
5104                              d++) {
5105                                 mdk_rdev_t *rdev = conf->disks[d].rdev;
5106                                 if (rdev && raid5_remove_disk(mddev, d) == 0) {
5107                                         char nm[20];
5108                                         sprintf(nm, "rd%d", rdev->raid_disk);
5109                                         sysfs_remove_link(&mddev->kobj, nm);
5110                                         rdev->raid_disk = -1;
5111                                 }
5112                         }
5113                 }
5114                 mddev->layout = conf->algorithm;
5115                 mddev->chunk_sectors = conf->chunk_sectors;
5116                 mddev->reshape_position = MaxSector;
5117                 mddev->delta_disks = 0;
5118         }
5119 }
5120
5121 static void raid5_quiesce(mddev_t *mddev, int state)
5122 {
5123         raid5_conf_t *conf = mddev->private;
5124
5125         switch(state) {
5126         case 2: /* resume for a suspend */
5127                 wake_up(&conf->wait_for_overlap);
5128                 break;
5129
5130         case 1: /* stop all writes */
5131                 spin_lock_irq(&conf->device_lock);
5132                 /* '2' tells resync/reshape to pause so that all
5133                  * active stripes can drain
5134                  */
5135                 conf->quiesce = 2;
5136                 wait_event_lock_irq(conf->wait_for_stripe,
5137                                     atomic_read(&conf->active_stripes) == 0 &&
5138                                     atomic_read(&conf->active_aligned_reads) == 0,
5139                                     conf->device_lock, /* nothing */);
5140                 conf->quiesce = 1;
5141                 spin_unlock_irq(&conf->device_lock);
5142                 /* allow reshape to continue */
5143                 wake_up(&conf->wait_for_overlap);
5144                 break;
5145
5146         case 0: /* re-enable writes */
5147                 spin_lock_irq(&conf->device_lock);
5148                 conf->quiesce = 0;
5149                 wake_up(&conf->wait_for_stripe);
5150                 wake_up(&conf->wait_for_overlap);
5151                 spin_unlock_irq(&conf->device_lock);
5152                 break;
5153         }
5154 }
5155
5156
5157 static void *raid5_takeover_raid1(mddev_t *mddev)
5158 {
5159         int chunksect;
5160
5161         if (mddev->raid_disks != 2 ||
5162             mddev->degraded > 1)
5163                 return ERR_PTR(-EINVAL);
5164
5165         /* Should check if there are write-behind devices? */
5166
5167         chunksect = 64*2; /* 64K by default */
5168
5169         /* The array must be an exact multiple of chunksize */
5170         while (chunksect && (mddev->array_sectors & (chunksect-1)))
5171                 chunksect >>= 1;
5172
5173         if ((chunksect<<9) < STRIPE_SIZE)
5174                 /* array size does not allow a suitable chunk size */
5175                 return ERR_PTR(-EINVAL);
5176
5177         mddev->new_level = 5;
5178         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5179         mddev->new_chunk_sectors = chunksect;
5180
5181         return setup_conf(mddev);
5182 }
5183
5184 static void *raid5_takeover_raid6(mddev_t *mddev)
5185 {
5186         int new_layout;
5187
5188         switch (mddev->layout) {
5189         case ALGORITHM_LEFT_ASYMMETRIC_6:
5190                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5191                 break;
5192         case ALGORITHM_RIGHT_ASYMMETRIC_6:
5193                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5194                 break;
5195         case ALGORITHM_LEFT_SYMMETRIC_6:
5196                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5197                 break;
5198         case ALGORITHM_RIGHT_SYMMETRIC_6:
5199                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5200                 break;
5201         case ALGORITHM_PARITY_0_6:
5202                 new_layout = ALGORITHM_PARITY_0;
5203                 break;
5204         case ALGORITHM_PARITY_N:
5205                 new_layout = ALGORITHM_PARITY_N;
5206                 break;
5207         default:
5208                 return ERR_PTR(-EINVAL);
5209         }
5210         mddev->new_level = 5;
5211         mddev->new_layout = new_layout;
5212         mddev->delta_disks = -1;
5213         mddev->raid_disks -= 1;
5214         return setup_conf(mddev);
5215 }
5216
5217
5218 static int raid5_check_reshape(mddev_t *mddev)
5219 {
5220         /* For a 2-drive array, the layout and chunk size can be changed
5221          * immediately as not restriping is needed.
5222          * For larger arrays we record the new value - after validation
5223          * to be used by a reshape pass.
5224          */
5225         raid5_conf_t *conf = mddev->private;
5226         int new_chunk = mddev->new_chunk_sectors;
5227
5228         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5229                 return -EINVAL;
5230         if (new_chunk > 0) {
5231                 if (!is_power_of_2(new_chunk))
5232                         return -EINVAL;
5233                 if (new_chunk < (PAGE_SIZE>>9))
5234                         return -EINVAL;
5235                 if (mddev->array_sectors & (new_chunk-1))
5236                         /* not factor of array size */
5237                         return -EINVAL;
5238         }
5239
5240         /* They look valid */
5241
5242         if (mddev->raid_disks == 2) {
5243                 /* can make the change immediately */
5244                 if (mddev->new_layout >= 0) {
5245                         conf->algorithm = mddev->new_layout;
5246                         mddev->layout = mddev->new_layout;
5247                 }
5248                 if (new_chunk > 0) {
5249                         conf->chunk_sectors = new_chunk ;
5250                         mddev->chunk_sectors = new_chunk;
5251                 }
5252                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5253                 md_wakeup_thread(mddev->thread);
5254         }
5255         return check_reshape(mddev);
5256 }
5257
5258 static int raid6_check_reshape(mddev_t *mddev)
5259 {
5260         int new_chunk = mddev->new_chunk_sectors;
5261
5262         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5263                 return -EINVAL;
5264         if (new_chunk > 0) {
5265                 if (!is_power_of_2(new_chunk))
5266                         return -EINVAL;
5267                 if (new_chunk < (PAGE_SIZE >> 9))
5268                         return -EINVAL;
5269                 if (mddev->array_sectors & (new_chunk-1))
5270                         /* not factor of array size */
5271                         return -EINVAL;
5272         }
5273
5274         /* They look valid */
5275         return check_reshape(mddev);
5276 }
5277
5278 static void *raid5_takeover(mddev_t *mddev)
5279 {
5280         /* raid5 can take over:
5281          *  raid0 - if all devices are the same - make it a raid4 layout
5282          *  raid1 - if there are two drives.  We need to know the chunk size
5283          *  raid4 - trivial - just use a raid4 layout.
5284          *  raid6 - Providing it is a *_6 layout
5285          */
5286
5287         if (mddev->level == 1)
5288                 return raid5_takeover_raid1(mddev);
5289         if (mddev->level == 4) {
5290                 mddev->new_layout = ALGORITHM_PARITY_N;
5291                 mddev->new_level = 5;
5292                 return setup_conf(mddev);
5293         }
5294         if (mddev->level == 6)
5295                 return raid5_takeover_raid6(mddev);
5296
5297         return ERR_PTR(-EINVAL);
5298 }
5299
5300
5301 static struct mdk_personality raid5_personality;
5302
5303 static void *raid6_takeover(mddev_t *mddev)
5304 {
5305         /* Currently can only take over a raid5.  We map the
5306          * personality to an equivalent raid6 personality
5307          * with the Q block at the end.
5308          */
5309         int new_layout;
5310
5311         if (mddev->pers != &raid5_personality)
5312                 return ERR_PTR(-EINVAL);
5313         if (mddev->degraded > 1)
5314                 return ERR_PTR(-EINVAL);
5315         if (mddev->raid_disks > 253)
5316                 return ERR_PTR(-EINVAL);
5317         if (mddev->raid_disks < 3)
5318                 return ERR_PTR(-EINVAL);
5319
5320         switch (mddev->layout) {
5321         case ALGORITHM_LEFT_ASYMMETRIC:
5322                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5323                 break;
5324         case ALGORITHM_RIGHT_ASYMMETRIC:
5325                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5326                 break;