e1ee181b79bb784688e065cae55138461cea64d2
[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/async_tx.h>
49 #include <linux/seq_file.h>
50 #include "md.h"
51 #include "raid5.h"
52 #include "raid6.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 #if !RAID6_USE_EMPTY_ZERO_PAGE
98 /* In .bss so it's zeroed */
99 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
100 #endif
101
102 /*
103  * We maintain a biased count of active stripes in the bottom 16 bits of
104  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
105  */
106 static inline int raid5_bi_phys_segments(struct bio *bio)
107 {
108         return bio->bi_phys_segments & 0xffff;
109 }
110
111 static inline int raid5_bi_hw_segments(struct bio *bio)
112 {
113         return (bio->bi_phys_segments >> 16) & 0xffff;
114 }
115
116 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
117 {
118         --bio->bi_phys_segments;
119         return raid5_bi_phys_segments(bio);
120 }
121
122 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
123 {
124         unsigned short val = raid5_bi_hw_segments(bio);
125
126         --val;
127         bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
128         return val;
129 }
130
131 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
132 {
133         bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16);
134 }
135
136 /* Find first data disk in a raid6 stripe */
137 static inline int raid6_d0(struct stripe_head *sh)
138 {
139         if (sh->ddf_layout)
140                 /* ddf always start from first device */
141                 return 0;
142         /* md starts just after Q block */
143         if (sh->qd_idx == sh->disks - 1)
144                 return 0;
145         else
146                 return sh->qd_idx + 1;
147 }
148 static inline int raid6_next_disk(int disk, int raid_disks)
149 {
150         disk++;
151         return (disk < raid_disks) ? disk : 0;
152 }
153
154 /* When walking through the disks in a raid5, starting at raid6_d0,
155  * We need to map each disk to a 'slot', where the data disks are slot
156  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
157  * is raid_disks-1.  This help does that mapping.
158  */
159 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
160                              int *count, int syndrome_disks)
161 {
162         int slot;
163
164         if (idx == sh->pd_idx)
165                 return syndrome_disks;
166         if (idx == sh->qd_idx)
167                 return syndrome_disks + 1;
168         slot = (*count)++;
169         return slot;
170 }
171
172 static void return_io(struct bio *return_bi)
173 {
174         struct bio *bi = return_bi;
175         while (bi) {
176
177                 return_bi = bi->bi_next;
178                 bi->bi_next = NULL;
179                 bi->bi_size = 0;
180                 bio_endio(bi, 0);
181                 bi = return_bi;
182         }
183 }
184
185 static void print_raid5_conf (raid5_conf_t *conf);
186
187 static int stripe_operations_active(struct stripe_head *sh)
188 {
189         return sh->check_state || sh->reconstruct_state ||
190                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
191                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
192 }
193
194 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
195 {
196         if (atomic_dec_and_test(&sh->count)) {
197                 BUG_ON(!list_empty(&sh->lru));
198                 BUG_ON(atomic_read(&conf->active_stripes)==0);
199                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
200                         if (test_bit(STRIPE_DELAYED, &sh->state)) {
201                                 list_add_tail(&sh->lru, &conf->delayed_list);
202                                 blk_plug_device(conf->mddev->queue);
203                         } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
204                                    sh->bm_seq - conf->seq_write > 0) {
205                                 list_add_tail(&sh->lru, &conf->bitmap_list);
206                                 blk_plug_device(conf->mddev->queue);
207                         } else {
208                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
209                                 list_add_tail(&sh->lru, &conf->handle_list);
210                         }
211                         md_wakeup_thread(conf->mddev->thread);
212                 } else {
213                         BUG_ON(stripe_operations_active(sh));
214                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
215                                 atomic_dec(&conf->preread_active_stripes);
216                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
217                                         md_wakeup_thread(conf->mddev->thread);
218                         }
219                         atomic_dec(&conf->active_stripes);
220                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
221                                 list_add_tail(&sh->lru, &conf->inactive_list);
222                                 wake_up(&conf->wait_for_stripe);
223                                 if (conf->retry_read_aligned)
224                                         md_wakeup_thread(conf->mddev->thread);
225                         }
226                 }
227         }
228 }
229
230 static void release_stripe(struct stripe_head *sh)
231 {
232         raid5_conf_t *conf = sh->raid_conf;
233         unsigned long flags;
234
235         spin_lock_irqsave(&conf->device_lock, flags);
236         __release_stripe(conf, sh);
237         spin_unlock_irqrestore(&conf->device_lock, flags);
238 }
239
240 static inline void remove_hash(struct stripe_head *sh)
241 {
242         pr_debug("remove_hash(), stripe %llu\n",
243                 (unsigned long long)sh->sector);
244
245         hlist_del_init(&sh->hash);
246 }
247
248 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
249 {
250         struct hlist_head *hp = stripe_hash(conf, sh->sector);
251
252         pr_debug("insert_hash(), stripe %llu\n",
253                 (unsigned long long)sh->sector);
254
255         CHECK_DEVLOCK();
256         hlist_add_head(&sh->hash, hp);
257 }
258
259
260 /* find an idle stripe, make sure it is unhashed, and return it. */
261 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
262 {
263         struct stripe_head *sh = NULL;
264         struct list_head *first;
265
266         CHECK_DEVLOCK();
267         if (list_empty(&conf->inactive_list))
268                 goto out;
269         first = conf->inactive_list.next;
270         sh = list_entry(first, struct stripe_head, lru);
271         list_del_init(first);
272         remove_hash(sh);
273         atomic_inc(&conf->active_stripes);
274 out:
275         return sh;
276 }
277
278 static void shrink_buffers(struct stripe_head *sh, int num)
279 {
280         struct page *p;
281         int i;
282
283         for (i=0; i<num ; i++) {
284                 p = sh->dev[i].page;
285                 if (!p)
286                         continue;
287                 sh->dev[i].page = NULL;
288                 put_page(p);
289         }
290 }
291
292 static int grow_buffers(struct stripe_head *sh, int num)
293 {
294         int i;
295
296         for (i=0; i<num; i++) {
297                 struct page *page;
298
299                 if (!(page = alloc_page(GFP_KERNEL))) {
300                         return 1;
301                 }
302                 sh->dev[i].page = page;
303         }
304         return 0;
305 }
306
307 static void raid5_build_block(struct stripe_head *sh, int i);
308 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
309                             struct stripe_head *sh);
310
311 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
312 {
313         raid5_conf_t *conf = sh->raid_conf;
314         int i;
315
316         BUG_ON(atomic_read(&sh->count) != 0);
317         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
318         BUG_ON(stripe_operations_active(sh));
319
320         CHECK_DEVLOCK();
321         pr_debug("init_stripe called, stripe %llu\n",
322                 (unsigned long long)sh->sector);
323
324         remove_hash(sh);
325
326         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
327         sh->sector = sector;
328         stripe_set_idx(sector, conf, previous, sh);
329         sh->state = 0;
330
331
332         for (i = sh->disks; i--; ) {
333                 struct r5dev *dev = &sh->dev[i];
334
335                 if (dev->toread || dev->read || dev->towrite || dev->written ||
336                     test_bit(R5_LOCKED, &dev->flags)) {
337                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
338                                (unsigned long long)sh->sector, i, dev->toread,
339                                dev->read, dev->towrite, dev->written,
340                                test_bit(R5_LOCKED, &dev->flags));
341                         BUG();
342                 }
343                 dev->flags = 0;
344                 raid5_build_block(sh, i);
345         }
346         insert_hash(conf, sh);
347 }
348
349 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
350 {
351         struct stripe_head *sh;
352         struct hlist_node *hn;
353
354         CHECK_DEVLOCK();
355         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
356         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
357                 if (sh->sector == sector && sh->disks == disks)
358                         return sh;
359         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
360         return NULL;
361 }
362
363 static void unplug_slaves(mddev_t *mddev);
364 static void raid5_unplug_device(struct request_queue *q);
365
366 static struct stripe_head *
367 get_active_stripe(raid5_conf_t *conf, sector_t sector,
368                   int previous, int noblock)
369 {
370         struct stripe_head *sh;
371         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
372
373         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
374
375         spin_lock_irq(&conf->device_lock);
376
377         do {
378                 wait_event_lock_irq(conf->wait_for_stripe,
379                                     conf->quiesce == 0,
380                                     conf->device_lock, /* nothing */);
381                 sh = __find_stripe(conf, sector, disks);
382                 if (!sh) {
383                         if (!conf->inactive_blocked)
384                                 sh = get_free_stripe(conf);
385                         if (noblock && sh == NULL)
386                                 break;
387                         if (!sh) {
388                                 conf->inactive_blocked = 1;
389                                 wait_event_lock_irq(conf->wait_for_stripe,
390                                                     !list_empty(&conf->inactive_list) &&
391                                                     (atomic_read(&conf->active_stripes)
392                                                      < (conf->max_nr_stripes *3/4)
393                                                      || !conf->inactive_blocked),
394                                                     conf->device_lock,
395                                                     raid5_unplug_device(conf->mddev->queue)
396                                         );
397                                 conf->inactive_blocked = 0;
398                         } else
399                                 init_stripe(sh, sector, previous);
400                 } else {
401                         if (atomic_read(&sh->count)) {
402                           BUG_ON(!list_empty(&sh->lru));
403                         } else {
404                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
405                                         atomic_inc(&conf->active_stripes);
406                                 if (list_empty(&sh->lru) &&
407                                     !test_bit(STRIPE_EXPANDING, &sh->state))
408                                         BUG();
409                                 list_del_init(&sh->lru);
410                         }
411                 }
412         } while (sh == NULL);
413
414         if (sh)
415                 atomic_inc(&sh->count);
416
417         spin_unlock_irq(&conf->device_lock);
418         return sh;
419 }
420
421 static void
422 raid5_end_read_request(struct bio *bi, int error);
423 static void
424 raid5_end_write_request(struct bio *bi, int error);
425
426 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
427 {
428         raid5_conf_t *conf = sh->raid_conf;
429         int i, disks = sh->disks;
430
431         might_sleep();
432
433         for (i = disks; i--; ) {
434                 int rw;
435                 struct bio *bi;
436                 mdk_rdev_t *rdev;
437                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
438                         rw = WRITE;
439                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
440                         rw = READ;
441                 else
442                         continue;
443
444                 bi = &sh->dev[i].req;
445
446                 bi->bi_rw = rw;
447                 if (rw == WRITE)
448                         bi->bi_end_io = raid5_end_write_request;
449                 else
450                         bi->bi_end_io = raid5_end_read_request;
451
452                 rcu_read_lock();
453                 rdev = rcu_dereference(conf->disks[i].rdev);
454                 if (rdev && test_bit(Faulty, &rdev->flags))
455                         rdev = NULL;
456                 if (rdev)
457                         atomic_inc(&rdev->nr_pending);
458                 rcu_read_unlock();
459
460                 if (rdev) {
461                         if (s->syncing || s->expanding || s->expanded)
462                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
463
464                         set_bit(STRIPE_IO_STARTED, &sh->state);
465
466                         bi->bi_bdev = rdev->bdev;
467                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
468                                 __func__, (unsigned long long)sh->sector,
469                                 bi->bi_rw, i);
470                         atomic_inc(&sh->count);
471                         bi->bi_sector = sh->sector + rdev->data_offset;
472                         bi->bi_flags = 1 << BIO_UPTODATE;
473                         bi->bi_vcnt = 1;
474                         bi->bi_max_vecs = 1;
475                         bi->bi_idx = 0;
476                         bi->bi_io_vec = &sh->dev[i].vec;
477                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
478                         bi->bi_io_vec[0].bv_offset = 0;
479                         bi->bi_size = STRIPE_SIZE;
480                         bi->bi_next = NULL;
481                         if (rw == WRITE &&
482                             test_bit(R5_ReWrite, &sh->dev[i].flags))
483                                 atomic_add(STRIPE_SECTORS,
484                                         &rdev->corrected_errors);
485                         generic_make_request(bi);
486                 } else {
487                         if (rw == WRITE)
488                                 set_bit(STRIPE_DEGRADED, &sh->state);
489                         pr_debug("skip op %ld on disc %d for sector %llu\n",
490                                 bi->bi_rw, i, (unsigned long long)sh->sector);
491                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
492                         set_bit(STRIPE_HANDLE, &sh->state);
493                 }
494         }
495 }
496
497 static struct dma_async_tx_descriptor *
498 async_copy_data(int frombio, struct bio *bio, struct page *page,
499         sector_t sector, struct dma_async_tx_descriptor *tx)
500 {
501         struct bio_vec *bvl;
502         struct page *bio_page;
503         int i;
504         int page_offset;
505
506         if (bio->bi_sector >= sector)
507                 page_offset = (signed)(bio->bi_sector - sector) * 512;
508         else
509                 page_offset = (signed)(sector - bio->bi_sector) * -512;
510         bio_for_each_segment(bvl, bio, i) {
511                 int len = bio_iovec_idx(bio, i)->bv_len;
512                 int clen;
513                 int b_offset = 0;
514
515                 if (page_offset < 0) {
516                         b_offset = -page_offset;
517                         page_offset += b_offset;
518                         len -= b_offset;
519                 }
520
521                 if (len > 0 && page_offset + len > STRIPE_SIZE)
522                         clen = STRIPE_SIZE - page_offset;
523                 else
524                         clen = len;
525
526                 if (clen > 0) {
527                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
528                         bio_page = bio_iovec_idx(bio, i)->bv_page;
529                         if (frombio)
530                                 tx = async_memcpy(page, bio_page, page_offset,
531                                         b_offset, clen,
532                                         ASYNC_TX_DEP_ACK,
533                                         tx, NULL, NULL);
534                         else
535                                 tx = async_memcpy(bio_page, page, b_offset,
536                                         page_offset, clen,
537                                         ASYNC_TX_DEP_ACK,
538                                         tx, NULL, NULL);
539                 }
540                 if (clen < len) /* hit end of page */
541                         break;
542                 page_offset +=  len;
543         }
544
545         return tx;
546 }
547
548 static void ops_complete_biofill(void *stripe_head_ref)
549 {
550         struct stripe_head *sh = stripe_head_ref;
551         struct bio *return_bi = NULL;
552         raid5_conf_t *conf = sh->raid_conf;
553         int i;
554
555         pr_debug("%s: stripe %llu\n", __func__,
556                 (unsigned long long)sh->sector);
557
558         /* clear completed biofills */
559         spin_lock_irq(&conf->device_lock);
560         for (i = sh->disks; i--; ) {
561                 struct r5dev *dev = &sh->dev[i];
562
563                 /* acknowledge completion of a biofill operation */
564                 /* and check if we need to reply to a read request,
565                  * new R5_Wantfill requests are held off until
566                  * !STRIPE_BIOFILL_RUN
567                  */
568                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
569                         struct bio *rbi, *rbi2;
570
571                         BUG_ON(!dev->read);
572                         rbi = dev->read;
573                         dev->read = NULL;
574                         while (rbi && rbi->bi_sector <
575                                 dev->sector + STRIPE_SECTORS) {
576                                 rbi2 = r5_next_bio(rbi, dev->sector);
577                                 if (!raid5_dec_bi_phys_segments(rbi)) {
578                                         rbi->bi_next = return_bi;
579                                         return_bi = rbi;
580                                 }
581                                 rbi = rbi2;
582                         }
583                 }
584         }
585         spin_unlock_irq(&conf->device_lock);
586         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
587
588         return_io(return_bi);
589
590         set_bit(STRIPE_HANDLE, &sh->state);
591         release_stripe(sh);
592 }
593
594 static void ops_run_biofill(struct stripe_head *sh)
595 {
596         struct dma_async_tx_descriptor *tx = NULL;
597         raid5_conf_t *conf = sh->raid_conf;
598         int i;
599
600         pr_debug("%s: stripe %llu\n", __func__,
601                 (unsigned long long)sh->sector);
602
603         for (i = sh->disks; i--; ) {
604                 struct r5dev *dev = &sh->dev[i];
605                 if (test_bit(R5_Wantfill, &dev->flags)) {
606                         struct bio *rbi;
607                         spin_lock_irq(&conf->device_lock);
608                         dev->read = rbi = dev->toread;
609                         dev->toread = NULL;
610                         spin_unlock_irq(&conf->device_lock);
611                         while (rbi && rbi->bi_sector <
612                                 dev->sector + STRIPE_SECTORS) {
613                                 tx = async_copy_data(0, rbi, dev->page,
614                                         dev->sector, tx);
615                                 rbi = r5_next_bio(rbi, dev->sector);
616                         }
617                 }
618         }
619
620         atomic_inc(&sh->count);
621         async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
622                 ops_complete_biofill, sh);
623 }
624
625 static void ops_complete_compute5(void *stripe_head_ref)
626 {
627         struct stripe_head *sh = stripe_head_ref;
628         int target = sh->ops.target;
629         struct r5dev *tgt = &sh->dev[target];
630
631         pr_debug("%s: stripe %llu\n", __func__,
632                 (unsigned long long)sh->sector);
633
634         set_bit(R5_UPTODATE, &tgt->flags);
635         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
636         clear_bit(R5_Wantcompute, &tgt->flags);
637         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
638         if (sh->check_state == check_state_compute_run)
639                 sh->check_state = check_state_compute_result;
640         set_bit(STRIPE_HANDLE, &sh->state);
641         release_stripe(sh);
642 }
643
644 static struct dma_async_tx_descriptor *ops_run_compute5(struct stripe_head *sh)
645 {
646         /* kernel stack size limits the total number of disks */
647         int disks = sh->disks;
648         struct page *xor_srcs[disks];
649         int target = sh->ops.target;
650         struct r5dev *tgt = &sh->dev[target];
651         struct page *xor_dest = tgt->page;
652         int count = 0;
653         struct dma_async_tx_descriptor *tx;
654         int i;
655
656         pr_debug("%s: stripe %llu block: %d\n",
657                 __func__, (unsigned long long)sh->sector, target);
658         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
659
660         for (i = disks; i--; )
661                 if (i != target)
662                         xor_srcs[count++] = sh->dev[i].page;
663
664         atomic_inc(&sh->count);
665
666         if (unlikely(count == 1))
667                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
668                         0, NULL, ops_complete_compute5, sh);
669         else
670                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
671                         ASYNC_TX_XOR_ZERO_DST, NULL,
672                         ops_complete_compute5, sh);
673
674         return tx;
675 }
676
677 static void ops_complete_prexor(void *stripe_head_ref)
678 {
679         struct stripe_head *sh = stripe_head_ref;
680
681         pr_debug("%s: stripe %llu\n", __func__,
682                 (unsigned long long)sh->sector);
683 }
684
685 static struct dma_async_tx_descriptor *
686 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
687 {
688         /* kernel stack size limits the total number of disks */
689         int disks = sh->disks;
690         struct page *xor_srcs[disks];
691         int count = 0, pd_idx = sh->pd_idx, i;
692
693         /* existing parity data subtracted */
694         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
695
696         pr_debug("%s: stripe %llu\n", __func__,
697                 (unsigned long long)sh->sector);
698
699         for (i = disks; i--; ) {
700                 struct r5dev *dev = &sh->dev[i];
701                 /* Only process blocks that are known to be uptodate */
702                 if (test_bit(R5_Wantdrain, &dev->flags))
703                         xor_srcs[count++] = dev->page;
704         }
705
706         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
707                 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
708                 ops_complete_prexor, sh);
709
710         return tx;
711 }
712
713 static struct dma_async_tx_descriptor *
714 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
715 {
716         int disks = sh->disks;
717         int i;
718
719         pr_debug("%s: stripe %llu\n", __func__,
720                 (unsigned long long)sh->sector);
721
722         for (i = disks; i--; ) {
723                 struct r5dev *dev = &sh->dev[i];
724                 struct bio *chosen;
725
726                 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
727                         struct bio *wbi;
728
729                         spin_lock(&sh->lock);
730                         chosen = dev->towrite;
731                         dev->towrite = NULL;
732                         BUG_ON(dev->written);
733                         wbi = dev->written = chosen;
734                         spin_unlock(&sh->lock);
735
736                         while (wbi && wbi->bi_sector <
737                                 dev->sector + STRIPE_SECTORS) {
738                                 tx = async_copy_data(1, wbi, dev->page,
739                                         dev->sector, tx);
740                                 wbi = r5_next_bio(wbi, dev->sector);
741                         }
742                 }
743         }
744
745         return tx;
746 }
747
748 static void ops_complete_postxor(void *stripe_head_ref)
749 {
750         struct stripe_head *sh = stripe_head_ref;
751         int disks = sh->disks, i, pd_idx = sh->pd_idx;
752
753         pr_debug("%s: stripe %llu\n", __func__,
754                 (unsigned long long)sh->sector);
755
756         for (i = disks; i--; ) {
757                 struct r5dev *dev = &sh->dev[i];
758                 if (dev->written || i == pd_idx)
759                         set_bit(R5_UPTODATE, &dev->flags);
760         }
761
762         if (sh->reconstruct_state == reconstruct_state_drain_run)
763                 sh->reconstruct_state = reconstruct_state_drain_result;
764         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
765                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
766         else {
767                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
768                 sh->reconstruct_state = reconstruct_state_result;
769         }
770
771         set_bit(STRIPE_HANDLE, &sh->state);
772         release_stripe(sh);
773 }
774
775 static void
776 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
777 {
778         /* kernel stack size limits the total number of disks */
779         int disks = sh->disks;
780         struct page *xor_srcs[disks];
781
782         int count = 0, pd_idx = sh->pd_idx, i;
783         struct page *xor_dest;
784         int prexor = 0;
785         unsigned long flags;
786
787         pr_debug("%s: stripe %llu\n", __func__,
788                 (unsigned long long)sh->sector);
789
790         /* check if prexor is active which means only process blocks
791          * that are part of a read-modify-write (written)
792          */
793         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
794                 prexor = 1;
795                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
796                 for (i = disks; i--; ) {
797                         struct r5dev *dev = &sh->dev[i];
798                         if (dev->written)
799                                 xor_srcs[count++] = dev->page;
800                 }
801         } else {
802                 xor_dest = sh->dev[pd_idx].page;
803                 for (i = disks; i--; ) {
804                         struct r5dev *dev = &sh->dev[i];
805                         if (i != pd_idx)
806                                 xor_srcs[count++] = dev->page;
807                 }
808         }
809
810         /* 1/ if we prexor'd then the dest is reused as a source
811          * 2/ if we did not prexor then we are redoing the parity
812          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
813          * for the synchronous xor case
814          */
815         flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
816                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
817
818         atomic_inc(&sh->count);
819
820         if (unlikely(count == 1)) {
821                 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
822                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
823                         flags, tx, ops_complete_postxor, sh);
824         } else
825                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
826                         flags, tx, ops_complete_postxor, sh);
827 }
828
829 static void ops_complete_check(void *stripe_head_ref)
830 {
831         struct stripe_head *sh = stripe_head_ref;
832
833         pr_debug("%s: stripe %llu\n", __func__,
834                 (unsigned long long)sh->sector);
835
836         sh->check_state = check_state_check_result;
837         set_bit(STRIPE_HANDLE, &sh->state);
838         release_stripe(sh);
839 }
840
841 static void ops_run_check(struct stripe_head *sh)
842 {
843         /* kernel stack size limits the total number of disks */
844         int disks = sh->disks;
845         struct page *xor_srcs[disks];
846         struct dma_async_tx_descriptor *tx;
847
848         int count = 0, pd_idx = sh->pd_idx, i;
849         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
850
851         pr_debug("%s: stripe %llu\n", __func__,
852                 (unsigned long long)sh->sector);
853
854         for (i = disks; i--; ) {
855                 struct r5dev *dev = &sh->dev[i];
856                 if (i != pd_idx)
857                         xor_srcs[count++] = dev->page;
858         }
859
860         tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
861                 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
862
863         atomic_inc(&sh->count);
864         tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
865                 ops_complete_check, sh);
866 }
867
868 static void raid5_run_ops(struct stripe_head *sh, unsigned long ops_request)
869 {
870         int overlap_clear = 0, i, disks = sh->disks;
871         struct dma_async_tx_descriptor *tx = NULL;
872
873         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
874                 ops_run_biofill(sh);
875                 overlap_clear++;
876         }
877
878         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
879                 tx = ops_run_compute5(sh);
880                 /* terminate the chain if postxor is not set to be run */
881                 if (tx && !test_bit(STRIPE_OP_POSTXOR, &ops_request))
882                         async_tx_ack(tx);
883         }
884
885         if (test_bit(STRIPE_OP_PREXOR, &ops_request))
886                 tx = ops_run_prexor(sh, tx);
887
888         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
889                 tx = ops_run_biodrain(sh, tx);
890                 overlap_clear++;
891         }
892
893         if (test_bit(STRIPE_OP_POSTXOR, &ops_request))
894                 ops_run_postxor(sh, tx);
895
896         if (test_bit(STRIPE_OP_CHECK, &ops_request))
897                 ops_run_check(sh);
898
899         if (overlap_clear)
900                 for (i = disks; i--; ) {
901                         struct r5dev *dev = &sh->dev[i];
902                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
903                                 wake_up(&sh->raid_conf->wait_for_overlap);
904                 }
905 }
906
907 static int grow_one_stripe(raid5_conf_t *conf)
908 {
909         struct stripe_head *sh;
910         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
911         if (!sh)
912                 return 0;
913         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
914         sh->raid_conf = conf;
915         spin_lock_init(&sh->lock);
916
917         if (grow_buffers(sh, conf->raid_disks)) {
918                 shrink_buffers(sh, conf->raid_disks);
919                 kmem_cache_free(conf->slab_cache, sh);
920                 return 0;
921         }
922         sh->disks = conf->raid_disks;
923         /* we just created an active stripe so... */
924         atomic_set(&sh->count, 1);
925         atomic_inc(&conf->active_stripes);
926         INIT_LIST_HEAD(&sh->lru);
927         release_stripe(sh);
928         return 1;
929 }
930
931 static int grow_stripes(raid5_conf_t *conf, int num)
932 {
933         struct kmem_cache *sc;
934         int devs = conf->raid_disks;
935
936         sprintf(conf->cache_name[0],
937                 "raid%d-%s", conf->level, mdname(conf->mddev));
938         sprintf(conf->cache_name[1],
939                 "raid%d-%s-alt", conf->level, mdname(conf->mddev));
940         conf->active_name = 0;
941         sc = kmem_cache_create(conf->cache_name[conf->active_name],
942                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
943                                0, 0, NULL);
944         if (!sc)
945                 return 1;
946         conf->slab_cache = sc;
947         conf->pool_size = devs;
948         while (num--)
949                 if (!grow_one_stripe(conf))
950                         return 1;
951         return 0;
952 }
953
954 #ifdef CONFIG_MD_RAID5_RESHAPE
955 static int resize_stripes(raid5_conf_t *conf, int newsize)
956 {
957         /* Make all the stripes able to hold 'newsize' devices.
958          * New slots in each stripe get 'page' set to a new page.
959          *
960          * This happens in stages:
961          * 1/ create a new kmem_cache and allocate the required number of
962          *    stripe_heads.
963          * 2/ gather all the old stripe_heads and tranfer the pages across
964          *    to the new stripe_heads.  This will have the side effect of
965          *    freezing the array as once all stripe_heads have been collected,
966          *    no IO will be possible.  Old stripe heads are freed once their
967          *    pages have been transferred over, and the old kmem_cache is
968          *    freed when all stripes are done.
969          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
970          *    we simple return a failre status - no need to clean anything up.
971          * 4/ allocate new pages for the new slots in the new stripe_heads.
972          *    If this fails, we don't bother trying the shrink the
973          *    stripe_heads down again, we just leave them as they are.
974          *    As each stripe_head is processed the new one is released into
975          *    active service.
976          *
977          * Once step2 is started, we cannot afford to wait for a write,
978          * so we use GFP_NOIO allocations.
979          */
980         struct stripe_head *osh, *nsh;
981         LIST_HEAD(newstripes);
982         struct disk_info *ndisks;
983         int err;
984         struct kmem_cache *sc;
985         int i;
986
987         if (newsize <= conf->pool_size)
988                 return 0; /* never bother to shrink */
989
990         err = md_allow_write(conf->mddev);
991         if (err)
992                 return err;
993
994         /* Step 1 */
995         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
996                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
997                                0, 0, NULL);
998         if (!sc)
999                 return -ENOMEM;
1000
1001         for (i = conf->max_nr_stripes; i; i--) {
1002                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1003                 if (!nsh)
1004                         break;
1005
1006                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1007
1008                 nsh->raid_conf = conf;
1009                 spin_lock_init(&nsh->lock);
1010
1011                 list_add(&nsh->lru, &newstripes);
1012         }
1013         if (i) {
1014                 /* didn't get enough, give up */
1015                 while (!list_empty(&newstripes)) {
1016                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1017                         list_del(&nsh->lru);
1018                         kmem_cache_free(sc, nsh);
1019                 }
1020                 kmem_cache_destroy(sc);
1021                 return -ENOMEM;
1022         }
1023         /* Step 2 - Must use GFP_NOIO now.
1024          * OK, we have enough stripes, start collecting inactive
1025          * stripes and copying them over
1026          */
1027         list_for_each_entry(nsh, &newstripes, lru) {
1028                 spin_lock_irq(&conf->device_lock);
1029                 wait_event_lock_irq(conf->wait_for_stripe,
1030                                     !list_empty(&conf->inactive_list),
1031                                     conf->device_lock,
1032                                     unplug_slaves(conf->mddev)
1033                         );
1034                 osh = get_free_stripe(conf);
1035                 spin_unlock_irq(&conf->device_lock);
1036                 atomic_set(&nsh->count, 1);
1037                 for(i=0; i<conf->pool_size; i++)
1038                         nsh->dev[i].page = osh->dev[i].page;
1039                 for( ; i<newsize; i++)
1040                         nsh->dev[i].page = NULL;
1041                 kmem_cache_free(conf->slab_cache, osh);
1042         }
1043         kmem_cache_destroy(conf->slab_cache);
1044
1045         /* Step 3.
1046          * At this point, we are holding all the stripes so the array
1047          * is completely stalled, so now is a good time to resize
1048          * conf->disks.
1049          */
1050         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1051         if (ndisks) {
1052                 for (i=0; i<conf->raid_disks; i++)
1053                         ndisks[i] = conf->disks[i];
1054                 kfree(conf->disks);
1055                 conf->disks = ndisks;
1056         } else
1057                 err = -ENOMEM;
1058
1059         /* Step 4, return new stripes to service */
1060         while(!list_empty(&newstripes)) {
1061                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1062                 list_del_init(&nsh->lru);
1063                 for (i=conf->raid_disks; i < newsize; i++)
1064                         if (nsh->dev[i].page == NULL) {
1065                                 struct page *p = alloc_page(GFP_NOIO);
1066                                 nsh->dev[i].page = p;
1067                                 if (!p)
1068                                         err = -ENOMEM;
1069                         }
1070                 release_stripe(nsh);
1071         }
1072         /* critical section pass, GFP_NOIO no longer needed */
1073
1074         conf->slab_cache = sc;
1075         conf->active_name = 1-conf->active_name;
1076         conf->pool_size = newsize;
1077         return err;
1078 }
1079 #endif
1080
1081 static int drop_one_stripe(raid5_conf_t *conf)
1082 {
1083         struct stripe_head *sh;
1084
1085         spin_lock_irq(&conf->device_lock);
1086         sh = get_free_stripe(conf);
1087         spin_unlock_irq(&conf->device_lock);
1088         if (!sh)
1089                 return 0;
1090         BUG_ON(atomic_read(&sh->count));
1091         shrink_buffers(sh, conf->pool_size);
1092         kmem_cache_free(conf->slab_cache, sh);
1093         atomic_dec(&conf->active_stripes);
1094         return 1;
1095 }
1096
1097 static void shrink_stripes(raid5_conf_t *conf)
1098 {
1099         while (drop_one_stripe(conf))
1100                 ;
1101
1102         if (conf->slab_cache)
1103                 kmem_cache_destroy(conf->slab_cache);
1104         conf->slab_cache = NULL;
1105 }
1106
1107 static void raid5_end_read_request(struct bio * bi, int error)
1108 {
1109         struct stripe_head *sh = bi->bi_private;
1110         raid5_conf_t *conf = sh->raid_conf;
1111         int disks = sh->disks, i;
1112         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1113         char b[BDEVNAME_SIZE];
1114         mdk_rdev_t *rdev;
1115
1116
1117         for (i=0 ; i<disks; i++)
1118                 if (bi == &sh->dev[i].req)
1119                         break;
1120
1121         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1122                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1123                 uptodate);
1124         if (i == disks) {
1125                 BUG();
1126                 return;
1127         }
1128
1129         if (uptodate) {
1130                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1131                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1132                         rdev = conf->disks[i].rdev;
1133                         printk_rl(KERN_INFO "raid5:%s: read error corrected"
1134                                   " (%lu sectors at %llu on %s)\n",
1135                                   mdname(conf->mddev), STRIPE_SECTORS,
1136                                   (unsigned long long)(sh->sector
1137                                                        + rdev->data_offset),
1138                                   bdevname(rdev->bdev, b));
1139                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1140                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1141                 }
1142                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1143                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1144         } else {
1145                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1146                 int retry = 0;
1147                 rdev = conf->disks[i].rdev;
1148
1149                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1150                 atomic_inc(&rdev->read_errors);
1151                 if (conf->mddev->degraded)
1152                         printk_rl(KERN_WARNING
1153                                   "raid5:%s: read error not correctable "
1154                                   "(sector %llu on %s).\n",
1155                                   mdname(conf->mddev),
1156                                   (unsigned long long)(sh->sector
1157                                                        + rdev->data_offset),
1158                                   bdn);
1159                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1160                         /* Oh, no!!! */
1161                         printk_rl(KERN_WARNING
1162                                   "raid5:%s: read error NOT corrected!! "
1163                                   "(sector %llu on %s).\n",
1164                                   mdname(conf->mddev),
1165                                   (unsigned long long)(sh->sector
1166                                                        + rdev->data_offset),
1167                                   bdn);
1168                 else if (atomic_read(&rdev->read_errors)
1169                          > conf->max_nr_stripes)
1170                         printk(KERN_WARNING
1171                                "raid5:%s: Too many read errors, failing device %s.\n",
1172                                mdname(conf->mddev), bdn);
1173                 else
1174                         retry = 1;
1175                 if (retry)
1176                         set_bit(R5_ReadError, &sh->dev[i].flags);
1177                 else {
1178                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1179                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1180                         md_error(conf->mddev, rdev);
1181                 }
1182         }
1183         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1184         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1185         set_bit(STRIPE_HANDLE, &sh->state);
1186         release_stripe(sh);
1187 }
1188
1189 static void raid5_end_write_request(struct bio *bi, int error)
1190 {
1191         struct stripe_head *sh = bi->bi_private;
1192         raid5_conf_t *conf = sh->raid_conf;
1193         int disks = sh->disks, i;
1194         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1195
1196         for (i=0 ; i<disks; i++)
1197                 if (bi == &sh->dev[i].req)
1198                         break;
1199
1200         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1201                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1202                 uptodate);
1203         if (i == disks) {
1204                 BUG();
1205                 return;
1206         }
1207
1208         if (!uptodate)
1209                 md_error(conf->mddev, conf->disks[i].rdev);
1210
1211         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1212         
1213         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1214         set_bit(STRIPE_HANDLE, &sh->state);
1215         release_stripe(sh);
1216 }
1217
1218
1219 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1220         
1221 static void raid5_build_block(struct stripe_head *sh, int i)
1222 {
1223         struct r5dev *dev = &sh->dev[i];
1224
1225         bio_init(&dev->req);
1226         dev->req.bi_io_vec = &dev->vec;
1227         dev->req.bi_vcnt++;
1228         dev->req.bi_max_vecs++;
1229         dev->vec.bv_page = dev->page;
1230         dev->vec.bv_len = STRIPE_SIZE;
1231         dev->vec.bv_offset = 0;
1232
1233         dev->req.bi_sector = sh->sector;
1234         dev->req.bi_private = sh;
1235
1236         dev->flags = 0;
1237         dev->sector = compute_blocknr(sh, i);
1238 }
1239
1240 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1241 {
1242         char b[BDEVNAME_SIZE];
1243         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1244         pr_debug("raid5: error called\n");
1245
1246         if (!test_bit(Faulty, &rdev->flags)) {
1247                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1248                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1249                         unsigned long flags;
1250                         spin_lock_irqsave(&conf->device_lock, flags);
1251                         mddev->degraded++;
1252                         spin_unlock_irqrestore(&conf->device_lock, flags);
1253                         /*
1254                          * if recovery was running, make sure it aborts.
1255                          */
1256                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1257                 }
1258                 set_bit(Faulty, &rdev->flags);
1259                 printk(KERN_ALERT
1260                        "raid5: Disk failure on %s, disabling device.\n"
1261                        "raid5: Operation continuing on %d devices.\n",
1262                        bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1263         }
1264 }
1265
1266 /*
1267  * Input: a 'big' sector number,
1268  * Output: index of the data and parity disk, and the sector # in them.
1269  */
1270 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1271                                      int previous, int *dd_idx,
1272                                      struct stripe_head *sh)
1273 {
1274         long stripe;
1275         unsigned long chunk_number;
1276         unsigned int chunk_offset;
1277         int pd_idx, qd_idx;
1278         int ddf_layout = 0;
1279         sector_t new_sector;
1280         int sectors_per_chunk = conf->chunk_size >> 9;
1281         int raid_disks = previous ? conf->previous_raid_disks
1282                                   : conf->raid_disks;
1283         int data_disks = raid_disks - conf->max_degraded;
1284
1285         /* First compute the information on this sector */
1286
1287         /*
1288          * Compute the chunk number and the sector offset inside the chunk
1289          */
1290         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1291         chunk_number = r_sector;
1292         BUG_ON(r_sector != chunk_number);
1293
1294         /*
1295          * Compute the stripe number
1296          */
1297         stripe = chunk_number / data_disks;
1298
1299         /*
1300          * Compute the data disk and parity disk indexes inside the stripe
1301          */
1302         *dd_idx = chunk_number % data_disks;
1303
1304         /*
1305          * Select the parity disk based on the user selected algorithm.
1306          */
1307         pd_idx = qd_idx = ~0;
1308         switch(conf->level) {
1309         case 4:
1310                 pd_idx = data_disks;
1311                 break;
1312         case 5:
1313                 switch (conf->algorithm) {
1314                 case ALGORITHM_LEFT_ASYMMETRIC:
1315                         pd_idx = data_disks - stripe % raid_disks;
1316                         if (*dd_idx >= pd_idx)
1317                                 (*dd_idx)++;
1318                         break;
1319                 case ALGORITHM_RIGHT_ASYMMETRIC:
1320                         pd_idx = stripe % raid_disks;
1321                         if (*dd_idx >= pd_idx)
1322                                 (*dd_idx)++;
1323                         break;
1324                 case ALGORITHM_LEFT_SYMMETRIC:
1325                         pd_idx = data_disks - stripe % raid_disks;
1326                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1327                         break;
1328                 case ALGORITHM_RIGHT_SYMMETRIC:
1329                         pd_idx = stripe % raid_disks;
1330                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1331                         break;
1332                 case ALGORITHM_PARITY_0:
1333                         pd_idx = 0;
1334                         (*dd_idx)++;
1335                         break;
1336                 case ALGORITHM_PARITY_N:
1337                         pd_idx = data_disks;
1338                         break;
1339                 default:
1340                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1341                                 conf->algorithm);
1342                         BUG();
1343                 }
1344                 break;
1345         case 6:
1346
1347                 switch (conf->algorithm) {
1348                 case ALGORITHM_LEFT_ASYMMETRIC:
1349                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1350                         qd_idx = pd_idx + 1;
1351                         if (pd_idx == raid_disks-1) {
1352                                 (*dd_idx)++;    /* Q D D D P */
1353                                 qd_idx = 0;
1354                         } else if (*dd_idx >= pd_idx)
1355                                 (*dd_idx) += 2; /* D D P Q D */
1356                         break;
1357                 case ALGORITHM_RIGHT_ASYMMETRIC:
1358                         pd_idx = stripe % raid_disks;
1359                         qd_idx = pd_idx + 1;
1360                         if (pd_idx == raid_disks-1) {
1361                                 (*dd_idx)++;    /* Q D D D P */
1362                                 qd_idx = 0;
1363                         } else if (*dd_idx >= pd_idx)
1364                                 (*dd_idx) += 2; /* D D P Q D */
1365                         break;
1366                 case ALGORITHM_LEFT_SYMMETRIC:
1367                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1368                         qd_idx = (pd_idx + 1) % raid_disks;
1369                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1370                         break;
1371                 case ALGORITHM_RIGHT_SYMMETRIC:
1372                         pd_idx = stripe % raid_disks;
1373                         qd_idx = (pd_idx + 1) % raid_disks;
1374                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1375                         break;
1376
1377                 case ALGORITHM_PARITY_0:
1378                         pd_idx = 0;
1379                         qd_idx = 1;
1380                         (*dd_idx) += 2;
1381                         break;
1382                 case ALGORITHM_PARITY_N:
1383                         pd_idx = data_disks;
1384                         qd_idx = data_disks + 1;
1385                         break;
1386
1387                 case ALGORITHM_ROTATING_ZERO_RESTART:
1388                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
1389                          * of blocks for computing Q is different.
1390                          */
1391                         pd_idx = stripe % raid_disks;
1392                         qd_idx = pd_idx + 1;
1393                         if (pd_idx == raid_disks-1) {
1394                                 (*dd_idx)++;    /* Q D D D P */
1395                                 qd_idx = 0;
1396                         } else if (*dd_idx >= pd_idx)
1397                                 (*dd_idx) += 2; /* D D P Q D */
1398                         ddf_layout = 1;
1399                         break;
1400
1401                 case ALGORITHM_ROTATING_N_RESTART:
1402                         /* Same a left_asymmetric, by first stripe is
1403                          * D D D P Q  rather than
1404                          * Q D D D P
1405                          */
1406                         pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1407                         qd_idx = pd_idx + 1;
1408                         if (pd_idx == raid_disks-1) {
1409                                 (*dd_idx)++;    /* Q D D D P */
1410                                 qd_idx = 0;
1411                         } else if (*dd_idx >= pd_idx)
1412                                 (*dd_idx) += 2; /* D D P Q D */
1413                         ddf_layout = 1;
1414                         break;
1415
1416                 case ALGORITHM_ROTATING_N_CONTINUE:
1417                         /* Same as left_symmetric but Q is before P */
1418                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1419                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1420                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1421                         ddf_layout = 1;
1422                         break;
1423
1424                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1425                         /* RAID5 left_asymmetric, with Q on last device */
1426                         pd_idx = data_disks - stripe % (raid_disks-1);
1427                         if (*dd_idx >= pd_idx)
1428                                 (*dd_idx)++;
1429                         qd_idx = raid_disks - 1;
1430                         break;
1431
1432                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1433                         pd_idx = stripe % (raid_disks-1);
1434                         if (*dd_idx >= pd_idx)
1435                                 (*dd_idx)++;
1436                         qd_idx = raid_disks - 1;
1437                         break;
1438
1439                 case ALGORITHM_LEFT_SYMMETRIC_6:
1440                         pd_idx = data_disks - stripe % (raid_disks-1);
1441                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1442                         qd_idx = raid_disks - 1;
1443                         break;
1444
1445                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1446                         pd_idx = stripe % (raid_disks-1);
1447                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1448                         qd_idx = raid_disks - 1;
1449                         break;
1450
1451                 case ALGORITHM_PARITY_0_6:
1452                         pd_idx = 0;
1453                         (*dd_idx)++;
1454                         qd_idx = raid_disks - 1;
1455                         break;
1456
1457
1458                 default:
1459                         printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1460                                conf->algorithm);
1461                         BUG();
1462                 }
1463                 break;
1464         }
1465
1466         if (sh) {
1467                 sh->pd_idx = pd_idx;
1468                 sh->qd_idx = qd_idx;
1469                 sh->ddf_layout = ddf_layout;
1470         }
1471         /*
1472          * Finally, compute the new sector number
1473          */
1474         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1475         return new_sector;
1476 }
1477
1478
1479 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1480 {
1481         raid5_conf_t *conf = sh->raid_conf;
1482         int raid_disks = sh->disks;
1483         int data_disks = raid_disks - conf->max_degraded;
1484         sector_t new_sector = sh->sector, check;
1485         int sectors_per_chunk = conf->chunk_size >> 9;
1486         sector_t stripe;
1487         int chunk_offset;
1488         int chunk_number, dummy1, dd_idx = i;
1489         sector_t r_sector;
1490         struct stripe_head sh2;
1491
1492
1493         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1494         stripe = new_sector;
1495         BUG_ON(new_sector != stripe);
1496
1497         if (i == sh->pd_idx)
1498                 return 0;
1499         switch(conf->level) {
1500         case 4: break;
1501         case 5:
1502                 switch (conf->algorithm) {
1503                 case ALGORITHM_LEFT_ASYMMETRIC:
1504                 case ALGORITHM_RIGHT_ASYMMETRIC:
1505                         if (i > sh->pd_idx)
1506                                 i--;
1507                         break;
1508                 case ALGORITHM_LEFT_SYMMETRIC:
1509                 case ALGORITHM_RIGHT_SYMMETRIC:
1510                         if (i < sh->pd_idx)
1511                                 i += raid_disks;
1512                         i -= (sh->pd_idx + 1);
1513                         break;
1514                 case ALGORITHM_PARITY_0:
1515                         i -= 1;
1516                         break;
1517                 case ALGORITHM_PARITY_N:
1518                         break;
1519                 default:
1520                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1521                                conf->algorithm);
1522                         BUG();
1523                 }
1524                 break;
1525         case 6:
1526                 if (i == sh->qd_idx)
1527                         return 0; /* It is the Q disk */
1528                 switch (conf->algorithm) {
1529                 case ALGORITHM_LEFT_ASYMMETRIC:
1530                 case ALGORITHM_RIGHT_ASYMMETRIC:
1531                 case ALGORITHM_ROTATING_ZERO_RESTART:
1532                 case ALGORITHM_ROTATING_N_RESTART:
1533                         if (sh->pd_idx == raid_disks-1)
1534                                 i--;    /* Q D D D P */
1535                         else if (i > sh->pd_idx)
1536                                 i -= 2; /* D D P Q D */
1537                         break;
1538                 case ALGORITHM_LEFT_SYMMETRIC:
1539                 case ALGORITHM_RIGHT_SYMMETRIC:
1540                         if (sh->pd_idx == raid_disks-1)
1541                                 i--; /* Q D D D P */
1542                         else {
1543                                 /* D D P Q D */
1544                                 if (i < sh->pd_idx)
1545                                         i += raid_disks;
1546                                 i -= (sh->pd_idx + 2);
1547                         }
1548                         break;
1549                 case ALGORITHM_PARITY_0:
1550                         i -= 2;
1551                         break;
1552                 case ALGORITHM_PARITY_N:
1553                         break;
1554                 case ALGORITHM_ROTATING_N_CONTINUE:
1555                         if (sh->pd_idx == 0)
1556                                 i--;    /* P D D D Q */
1557                         else if (i > sh->pd_idx)
1558                                 i -= 2; /* D D Q P D */
1559                         break;
1560                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1561                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1562                         if (i > sh->pd_idx)
1563                                 i--;
1564                         break;
1565                 case ALGORITHM_LEFT_SYMMETRIC_6:
1566                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1567                         if (i < sh->pd_idx)
1568                                 i += data_disks + 1;
1569                         i -= (sh->pd_idx + 1);
1570                         break;
1571                 case ALGORITHM_PARITY_0_6:
1572                         i -= 1;
1573                         break;
1574                 default:
1575                         printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1576                                conf->algorithm);
1577                         BUG();
1578                 }
1579                 break;
1580         }
1581
1582         chunk_number = stripe * data_disks + i;
1583         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1584
1585         check = raid5_compute_sector(conf, r_sector,
1586                                      (raid_disks != conf->raid_disks),
1587                                      &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 = conf->chunk_size >> 9;
2000         int dd_idx;
2001         int chunk_offset = sector_div(stripe, sectors_per_chunk);
2002         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2003
2004         raid5_compute_sector(conf,
2005                              stripe * (disks - conf->max_degraded)
2006                              *sectors_per_chunk + chunk_offset,
2007                              previous,
2008                              &dd_idx, sh);
2009 }
2010
2011 static void
2012 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2013                                 struct stripe_head_state *s, int disks,
2014                                 struct bio **return_bi)
2015 {
2016         int i;
2017         for (i = disks; i--; ) {
2018                 struct bio *bi;
2019                 int bitmap_end = 0;
2020
2021                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2022                         mdk_rdev_t *rdev;
2023                         rcu_read_lock();
2024                         rdev = rcu_dereference(conf->disks[i].rdev);
2025                         if (rdev && test_bit(In_sync, &rdev->flags))
2026                                 /* multiple read failures in one stripe */
2027                                 md_error(conf->mddev, rdev);
2028                         rcu_read_unlock();
2029                 }
2030                 spin_lock_irq(&conf->device_lock);
2031                 /* fail all writes first */
2032                 bi = sh->dev[i].towrite;
2033                 sh->dev[i].towrite = NULL;
2034                 if (bi) {
2035                         s->to_write--;
2036                         bitmap_end = 1;
2037                 }
2038
2039                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2040                         wake_up(&conf->wait_for_overlap);
2041
2042                 while (bi && bi->bi_sector <
2043                         sh->dev[i].sector + STRIPE_SECTORS) {
2044                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2045                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2046                         if (!raid5_dec_bi_phys_segments(bi)) {
2047                                 md_write_end(conf->mddev);
2048                                 bi->bi_next = *return_bi;
2049                                 *return_bi = bi;
2050                         }
2051                         bi = nextbi;
2052                 }
2053                 /* and fail all 'written' */
2054                 bi = sh->dev[i].written;
2055                 sh->dev[i].written = NULL;
2056                 if (bi) bitmap_end = 1;
2057                 while (bi && bi->bi_sector <
2058                        sh->dev[i].sector + STRIPE_SECTORS) {
2059                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2060                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2061                         if (!raid5_dec_bi_phys_segments(bi)) {
2062                                 md_write_end(conf->mddev);
2063                                 bi->bi_next = *return_bi;
2064                                 *return_bi = bi;
2065                         }
2066                         bi = bi2;
2067                 }
2068
2069                 /* fail any reads if this device is non-operational and
2070                  * the data has not reached the cache yet.
2071                  */
2072                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2073                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2074                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
2075                         bi = sh->dev[i].toread;
2076                         sh->dev[i].toread = NULL;
2077                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2078                                 wake_up(&conf->wait_for_overlap);
2079                         if (bi) s->to_read--;
2080                         while (bi && bi->bi_sector <
2081                                sh->dev[i].sector + STRIPE_SECTORS) {
2082                                 struct bio *nextbi =
2083                                         r5_next_bio(bi, sh->dev[i].sector);
2084                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2085                                 if (!raid5_dec_bi_phys_segments(bi)) {
2086                                         bi->bi_next = *return_bi;
2087                                         *return_bi = bi;
2088                                 }
2089                                 bi = nextbi;
2090                         }
2091                 }
2092                 spin_unlock_irq(&conf->device_lock);
2093                 if (bitmap_end)
2094                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2095                                         STRIPE_SECTORS, 0, 0);
2096         }
2097
2098         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2099                 if (atomic_dec_and_test(&conf->pending_full_writes))
2100                         md_wakeup_thread(conf->mddev->thread);
2101 }
2102
2103 /* fetch_block5 - checks the given member device to see if its data needs
2104  * to be read or computed to satisfy a request.
2105  *
2106  * Returns 1 when no more member devices need to be checked, otherwise returns
2107  * 0 to tell the loop in handle_stripe_fill5 to continue
2108  */
2109 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2110                         int disk_idx, int disks)
2111 {
2112         struct r5dev *dev = &sh->dev[disk_idx];
2113         struct r5dev *failed_dev = &sh->dev[s->failed_num];
2114
2115         /* is the data in this block needed, and can we get it? */
2116         if (!test_bit(R5_LOCKED, &dev->flags) &&
2117             !test_bit(R5_UPTODATE, &dev->flags) &&
2118             (dev->toread ||
2119              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2120              s->syncing || s->expanding ||
2121              (s->failed &&
2122               (failed_dev->toread ||
2123                (failed_dev->towrite &&
2124                 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2125                 /* We would like to get this block, possibly by computing it,
2126                  * otherwise read it if the backing disk is insync
2127                  */
2128                 if ((s->uptodate == disks - 1) &&
2129                     (s->failed && disk_idx == s->failed_num)) {
2130                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2131                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2132                         set_bit(R5_Wantcompute, &dev->flags);
2133                         sh->ops.target = disk_idx;
2134                         s->req_compute = 1;
2135                         /* Careful: from this point on 'uptodate' is in the eye
2136                          * of raid5_run_ops which services 'compute' operations
2137                          * before writes. R5_Wantcompute flags a block that will
2138                          * be R5_UPTODATE by the time it is needed for a
2139                          * subsequent operation.
2140                          */
2141                         s->uptodate++;
2142                         return 1; /* uptodate + compute == disks */
2143                 } else if (test_bit(R5_Insync, &dev->flags)) {
2144                         set_bit(R5_LOCKED, &dev->flags);
2145                         set_bit(R5_Wantread, &dev->flags);
2146                         s->locked++;
2147                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2148                                 s->syncing);
2149                 }
2150         }
2151
2152         return 0;
2153 }
2154
2155 /**
2156  * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2157  */
2158 static void handle_stripe_fill5(struct stripe_head *sh,
2159                         struct stripe_head_state *s, int disks)
2160 {
2161         int i;
2162
2163         /* look for blocks to read/compute, skip this if a compute
2164          * is already in flight, or if the stripe contents are in the
2165          * midst of changing due to a write
2166          */
2167         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2168             !sh->reconstruct_state)
2169                 for (i = disks; i--; )
2170                         if (fetch_block5(sh, s, i, disks))
2171                                 break;
2172         set_bit(STRIPE_HANDLE, &sh->state);
2173 }
2174
2175 static void handle_stripe_fill6(struct stripe_head *sh,
2176                         struct stripe_head_state *s, struct r6_state *r6s,
2177                         int disks)
2178 {
2179         int i;
2180         for (i = disks; i--; ) {
2181                 struct r5dev *dev = &sh->dev[i];
2182                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2183                     !test_bit(R5_UPTODATE, &dev->flags) &&
2184                     (dev->toread || (dev->towrite &&
2185                      !test_bit(R5_OVERWRITE, &dev->flags)) ||
2186                      s->syncing || s->expanding ||
2187                      (s->failed >= 1 &&
2188                       (sh->dev[r6s->failed_num[0]].toread ||
2189                        s->to_write)) ||
2190                      (s->failed >= 2 &&
2191                       (sh->dev[r6s->failed_num[1]].toread ||
2192                        s->to_write)))) {
2193                         /* we would like to get this block, possibly
2194                          * by computing it, but we might not be able to
2195                          */
2196                         if ((s->uptodate == disks - 1) &&
2197                             (s->failed && (i == r6s->failed_num[0] ||
2198                                            i == r6s->failed_num[1]))) {
2199                                 pr_debug("Computing stripe %llu block %d\n",
2200                                        (unsigned long long)sh->sector, i);
2201                                 compute_block_1(sh, i, 0);
2202                                 s->uptodate++;
2203                         } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2204                                 /* Computing 2-failure is *very* expensive; only
2205                                  * do it if failed >= 2
2206                                  */
2207                                 int other;
2208                                 for (other = disks; other--; ) {
2209                                         if (other == i)
2210                                                 continue;
2211                                         if (!test_bit(R5_UPTODATE,
2212                                               &sh->dev[other].flags))
2213                                                 break;
2214                                 }
2215                                 BUG_ON(other < 0);
2216                                 pr_debug("Computing stripe %llu blocks %d,%d\n",
2217                                        (unsigned long long)sh->sector,
2218                                        i, other);
2219                                 compute_block_2(sh, i, other);
2220                                 s->uptodate += 2;
2221                         } else if (test_bit(R5_Insync, &dev->flags)) {
2222                                 set_bit(R5_LOCKED, &dev->flags);
2223                                 set_bit(R5_Wantread, &dev->flags);
2224                                 s->locked++;
2225                                 pr_debug("Reading block %d (sync=%d)\n",
2226                                         i, s->syncing);
2227                         }
2228                 }
2229         }
2230         set_bit(STRIPE_HANDLE, &sh->state);
2231 }
2232
2233
2234 /* handle_stripe_clean_event
2235  * any written block on an uptodate or failed drive can be returned.
2236  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2237  * never LOCKED, so we don't need to test 'failed' directly.
2238  */
2239 static void handle_stripe_clean_event(raid5_conf_t *conf,
2240         struct stripe_head *sh, int disks, struct bio **return_bi)
2241 {
2242         int i;
2243         struct r5dev *dev;
2244
2245         for (i = disks; i--; )
2246                 if (sh->dev[i].written) {
2247                         dev = &sh->dev[i];
2248                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2249                                 test_bit(R5_UPTODATE, &dev->flags)) {
2250                                 /* We can return any write requests */
2251                                 struct bio *wbi, *wbi2;
2252                                 int bitmap_end = 0;
2253                                 pr_debug("Return write for disc %d\n", i);
2254                                 spin_lock_irq(&conf->device_lock);
2255                                 wbi = dev->written;
2256                                 dev->written = NULL;
2257                                 while (wbi && wbi->bi_sector <
2258                                         dev->sector + STRIPE_SECTORS) {
2259                                         wbi2 = r5_next_bio(wbi, dev->sector);
2260                                         if (!raid5_dec_bi_phys_segments(wbi)) {
2261                                                 md_write_end(conf->mddev);
2262                                                 wbi->bi_next = *return_bi;
2263                                                 *return_bi = wbi;
2264                                         }
2265                                         wbi = wbi2;
2266                                 }
2267                                 if (dev->towrite == NULL)
2268                                         bitmap_end = 1;
2269                                 spin_unlock_irq(&conf->device_lock);
2270                                 if (bitmap_end)
2271                                         bitmap_endwrite(conf->mddev->bitmap,
2272                                                         sh->sector,
2273                                                         STRIPE_SECTORS,
2274                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2275                                                         0);
2276                         }
2277                 }
2278
2279         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2280                 if (atomic_dec_and_test(&conf->pending_full_writes))
2281                         md_wakeup_thread(conf->mddev->thread);
2282 }
2283
2284 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2285                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2286 {
2287         int rmw = 0, rcw = 0, i;
2288         for (i = disks; i--; ) {
2289                 /* would I have to read this buffer for read_modify_write */
2290                 struct r5dev *dev = &sh->dev[i];
2291                 if ((dev->towrite || i == sh->pd_idx) &&
2292                     !test_bit(R5_LOCKED, &dev->flags) &&
2293                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2294                       test_bit(R5_Wantcompute, &dev->flags))) {
2295                         if (test_bit(R5_Insync, &dev->flags))
2296                                 rmw++;
2297                         else
2298                                 rmw += 2*disks;  /* cannot read it */
2299                 }
2300                 /* Would I have to read this buffer for reconstruct_write */
2301                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2302                     !test_bit(R5_LOCKED, &dev->flags) &&
2303                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2304                     test_bit(R5_Wantcompute, &dev->flags))) {
2305                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2306                         else
2307                                 rcw += 2*disks;
2308                 }
2309         }
2310         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2311                 (unsigned long long)sh->sector, rmw, rcw);
2312         set_bit(STRIPE_HANDLE, &sh->state);
2313         if (rmw < rcw && rmw > 0)
2314                 /* prefer read-modify-write, but need to get some data */
2315                 for (i = disks; i--; ) {
2316                         struct r5dev *dev = &sh->dev[i];
2317                         if ((dev->towrite || i == sh->pd_idx) &&
2318                             !test_bit(R5_LOCKED, &dev->flags) &&
2319                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2320                             test_bit(R5_Wantcompute, &dev->flags)) &&
2321                             test_bit(R5_Insync, &dev->flags)) {
2322                                 if (
2323                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2324                                         pr_debug("Read_old block "
2325                                                 "%d for r-m-w\n", i);
2326                                         set_bit(R5_LOCKED, &dev->flags);
2327                                         set_bit(R5_Wantread, &dev->flags);
2328                                         s->locked++;
2329                                 } else {
2330                                         set_bit(STRIPE_DELAYED, &sh->state);
2331                                         set_bit(STRIPE_HANDLE, &sh->state);
2332                                 }
2333                         }
2334                 }
2335         if (rcw <= rmw && rcw > 0)
2336                 /* want reconstruct write, but need to get some data */
2337                 for (i = disks; i--; ) {
2338                         struct r5dev *dev = &sh->dev[i];
2339                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2340                             i != sh->pd_idx &&
2341                             !test_bit(R5_LOCKED, &dev->flags) &&
2342                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2343                             test_bit(R5_Wantcompute, &dev->flags)) &&
2344                             test_bit(R5_Insync, &dev->flags)) {
2345                                 if (
2346                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2347                                         pr_debug("Read_old block "
2348                                                 "%d for Reconstruct\n", i);
2349                                         set_bit(R5_LOCKED, &dev->flags);
2350                                         set_bit(R5_Wantread, &dev->flags);
2351                                         s->locked++;
2352                                 } else {
2353                                         set_bit(STRIPE_DELAYED, &sh->state);
2354                                         set_bit(STRIPE_HANDLE, &sh->state);
2355                                 }
2356                         }
2357                 }
2358         /* now if nothing is locked, and if we have enough data,
2359          * we can start a write request
2360          */
2361         /* since handle_stripe can be called at any time we need to handle the
2362          * case where a compute block operation has been submitted and then a
2363          * subsequent call wants to start a write request.  raid5_run_ops only
2364          * handles the case where compute block and postxor are requested
2365          * simultaneously.  If this is not the case then new writes need to be
2366          * held off until the compute completes.
2367          */
2368         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2369             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2370             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2371                 schedule_reconstruction5(sh, s, rcw == 0, 0);
2372 }
2373
2374 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2375                 struct stripe_head *sh, struct stripe_head_state *s,
2376                 struct r6_state *r6s, int disks)
2377 {
2378         int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2379         int qd_idx = r6s->qd_idx;
2380         for (i = disks; i--; ) {
2381                 struct r5dev *dev = &sh->dev[i];
2382                 /* Would I have to read this buffer for reconstruct_write */
2383                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2384                     && i != pd_idx && i != qd_idx
2385                     && (!test_bit(R5_LOCKED, &dev->flags)
2386                             ) &&
2387                     !test_bit(R5_UPTODATE, &dev->flags)) {
2388                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2389                         else {
2390                                 pr_debug("raid6: must_compute: "
2391                                         "disk %d flags=%#lx\n", i, dev->flags);
2392                                 must_compute++;
2393                         }
2394                 }
2395         }
2396         pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2397                (unsigned long long)sh->sector, rcw, must_compute);
2398         set_bit(STRIPE_HANDLE, &sh->state);
2399
2400         if (rcw > 0)
2401                 /* want reconstruct write, but need to get some data */
2402                 for (i = disks; i--; ) {
2403                         struct r5dev *dev = &sh->dev[i];
2404                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2405                             && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2406                             && !test_bit(R5_LOCKED, &dev->flags) &&
2407                             !test_bit(R5_UPTODATE, &dev->flags) &&
2408                             test_bit(R5_Insync, &dev->flags)) {
2409                                 if (
2410                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2411                                         pr_debug("Read_old stripe %llu "
2412                                                 "block %d for Reconstruct\n",
2413                                              (unsigned long long)sh->sector, i);
2414                                         set_bit(R5_LOCKED, &dev->flags);
2415                                         set_bit(R5_Wantread, &dev->flags);
2416                                         s->locked++;
2417                                 } else {
2418                                         pr_debug("Request delayed stripe %llu "
2419                                                 "block %d for Reconstruct\n",
2420                                              (unsigned long long)sh->sector, i);
2421                                         set_bit(STRIPE_DELAYED, &sh->state);
2422                                         set_bit(STRIPE_HANDLE, &sh->state);
2423                                 }
2424                         }
2425                 }
2426         /* now if nothing is locked, and if we have enough data, we can start a
2427          * write request
2428          */
2429         if (s->locked == 0 && rcw == 0 &&
2430             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2431                 if (must_compute > 0) {
2432                         /* We have failed blocks and need to compute them */
2433                         switch (s->failed) {
2434                         case 0:
2435                                 BUG();
2436                         case 1:
2437                                 compute_block_1(sh, r6s->failed_num[0], 0);
2438                                 break;
2439                         case 2:
2440                                 compute_block_2(sh, r6s->failed_num[0],
2441                                                 r6s->failed_num[1]);
2442                                 break;
2443                         default: /* This request should have been failed? */
2444                                 BUG();
2445                         }
2446                 }
2447
2448                 pr_debug("Computing parity for stripe %llu\n",
2449                         (unsigned long long)sh->sector);
2450                 compute_parity6(sh, RECONSTRUCT_WRITE);
2451                 /* now every locked buffer is ready to be written */
2452                 for (i = disks; i--; )
2453                         if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2454                                 pr_debug("Writing stripe %llu block %d\n",
2455                                        (unsigned long long)sh->sector, i);
2456                                 s->locked++;
2457                                 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2458                         }
2459                 if (s->locked == disks)
2460                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2461                                 atomic_inc(&conf->pending_full_writes);
2462                 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2463                 set_bit(STRIPE_INSYNC, &sh->state);
2464
2465                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2466                         atomic_dec(&conf->preread_active_stripes);
2467                         if (atomic_read(&conf->preread_active_stripes) <
2468                             IO_THRESHOLD)
2469                                 md_wakeup_thread(conf->mddev->thread);
2470                 }
2471         }
2472 }
2473
2474 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2475                                 struct stripe_head_state *s, int disks)
2476 {
2477         struct r5dev *dev = NULL;
2478
2479         set_bit(STRIPE_HANDLE, &sh->state);
2480
2481         switch (sh->check_state) {
2482         case check_state_idle:
2483                 /* start a new check operation if there are no failures */
2484                 if (s->failed == 0) {
2485                         BUG_ON(s->uptodate != disks);
2486                         sh->check_state = check_state_run;
2487                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2488                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2489                         s->uptodate--;
2490                         break;
2491                 }
2492                 dev = &sh->dev[s->failed_num];
2493                 /* fall through */
2494         case check_state_compute_result:
2495                 sh->check_state = check_state_idle;
2496                 if (!dev)
2497                         dev = &sh->dev[sh->pd_idx];
2498
2499                 /* check that a write has not made the stripe insync */
2500                 if (test_bit(STRIPE_INSYNC, &sh->state))
2501                         break;
2502
2503                 /* either failed parity check, or recovery is happening */
2504                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2505                 BUG_ON(s->uptodate != disks);
2506
2507                 set_bit(R5_LOCKED, &dev->flags);
2508                 s->locked++;
2509                 set_bit(R5_Wantwrite, &dev->flags);
2510
2511                 clear_bit(STRIPE_DEGRADED, &sh->state);
2512                 set_bit(STRIPE_INSYNC, &sh->state);
2513                 break;
2514         case check_state_run:
2515                 break; /* we will be called again upon completion */
2516         case check_state_check_result:
2517                 sh->check_state = check_state_idle;
2518
2519                 /* if a failure occurred during the check operation, leave
2520                  * STRIPE_INSYNC not set and let the stripe be handled again
2521                  */
2522                 if (s->failed)
2523                         break;
2524
2525                 /* handle a successful check operation, if parity is correct
2526                  * we are done.  Otherwise update the mismatch count and repair
2527                  * parity if !MD_RECOVERY_CHECK
2528                  */
2529                 if (sh->ops.zero_sum_result == 0)
2530                         /* parity is correct (on disc,
2531                          * not in buffer any more)
2532                          */
2533                         set_bit(STRIPE_INSYNC, &sh->state);
2534                 else {
2535                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2536                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2537                                 /* don't try to repair!! */
2538                                 set_bit(STRIPE_INSYNC, &sh->state);
2539                         else {
2540                                 sh->check_state = check_state_compute_run;
2541                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2542                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2543                                 set_bit(R5_Wantcompute,
2544                                         &sh->dev[sh->pd_idx].flags);
2545                                 sh->ops.target = sh->pd_idx;
2546                                 s->uptodate++;
2547                         }
2548                 }
2549                 break;
2550         case check_state_compute_run:
2551                 break;
2552         default:
2553                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2554                        __func__, sh->check_state,
2555                        (unsigned long long) sh->sector);
2556                 BUG();
2557         }
2558 }
2559
2560
2561 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2562                                 struct stripe_head_state *s,
2563                                 struct r6_state *r6s, struct page *tmp_page,
2564                                 int disks)
2565 {
2566         int update_p = 0, update_q = 0;
2567         struct r5dev *dev;
2568         int pd_idx = sh->pd_idx;
2569         int qd_idx = r6s->qd_idx;
2570
2571         set_bit(STRIPE_HANDLE, &sh->state);
2572
2573         BUG_ON(s->failed > 2);
2574         BUG_ON(s->uptodate < disks);
2575         /* Want to check and possibly repair P and Q.
2576          * However there could be one 'failed' device, in which
2577          * case we can only check one of them, possibly using the
2578          * other to generate missing data
2579          */
2580
2581         /* If !tmp_page, we cannot do the calculations,
2582          * but as we have set STRIPE_HANDLE, we will soon be called
2583          * by stripe_handle with a tmp_page - just wait until then.
2584          */
2585         if (tmp_page) {
2586                 if (s->failed == r6s->q_failed) {
2587                         /* The only possible failed device holds 'Q', so it
2588                          * makes sense to check P (If anything else were failed,
2589                          * we would have used P to recreate it).
2590                          */
2591                         compute_block_1(sh, pd_idx, 1);
2592                         if (!page_is_zero(sh->dev[pd_idx].page)) {
2593                                 compute_block_1(sh, pd_idx, 0);
2594                                 update_p = 1;
2595                         }
2596                 }
2597                 if (!r6s->q_failed && s->failed < 2) {
2598                         /* q is not failed, and we didn't use it to generate
2599                          * anything, so it makes sense to check it
2600                          */
2601                         memcpy(page_address(tmp_page),
2602                                page_address(sh->dev[qd_idx].page),
2603                                STRIPE_SIZE);
2604                         compute_parity6(sh, UPDATE_PARITY);
2605                         if (memcmp(page_address(tmp_page),
2606                                    page_address(sh->dev[qd_idx].page),
2607                                    STRIPE_SIZE) != 0) {
2608                                 clear_bit(STRIPE_INSYNC, &sh->state);
2609                                 update_q = 1;
2610                         }
2611                 }
2612                 if (update_p || update_q) {
2613                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2614                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2615                                 /* don't try to repair!! */
2616                                 update_p = update_q = 0;
2617                 }
2618
2619                 /* now write out any block on a failed drive,
2620                  * or P or Q if they need it
2621                  */
2622
2623                 if (s->failed == 2) {
2624                         dev = &sh->dev[r6s->failed_num[1]];
2625                         s->locked++;
2626                         set_bit(R5_LOCKED, &dev->flags);
2627                         set_bit(R5_Wantwrite, &dev->flags);
2628                 }
2629                 if (s->failed >= 1) {
2630                         dev = &sh->dev[r6s->failed_num[0]];
2631                         s->locked++;
2632                         set_bit(R5_LOCKED, &dev->flags);
2633                         set_bit(R5_Wantwrite, &dev->flags);
2634                 }
2635
2636                 if (update_p) {
2637                         dev = &sh->dev[pd_idx];
2638                         s->locked++;
2639                         set_bit(R5_LOCKED, &dev->flags);
2640                         set_bit(R5_Wantwrite, &dev->flags);
2641                 }
2642                 if (update_q) {
2643                         dev = &sh->dev[qd_idx];
2644                         s->locked++;
2645                         set_bit(R5_LOCKED, &dev->flags);
2646                         set_bit(R5_Wantwrite, &dev->flags);
2647                 }
2648                 clear_bit(STRIPE_DEGRADED, &sh->state);
2649
2650                 set_bit(STRIPE_INSYNC, &sh->state);
2651         }
2652 }
2653
2654 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2655                                 struct r6_state *r6s)
2656 {
2657         int i;
2658
2659         /* We have read all the blocks in this stripe and now we need to
2660          * copy some of them into a target stripe for expand.
2661          */
2662         struct dma_async_tx_descriptor *tx = NULL;
2663         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2664         for (i = 0; i < sh->disks; i++)
2665                 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2666                         int dd_idx, j;
2667                         struct stripe_head *sh2;
2668
2669                         sector_t bn = compute_blocknr(sh, i);
2670                         sector_t s = raid5_compute_sector(conf, bn, 0,
2671                                                           &dd_idx, NULL);
2672                         sh2 = get_active_stripe(conf, s, 0, 1);
2673                         if (sh2 == NULL)
2674                                 /* so far only the early blocks of this stripe
2675                                  * have been requested.  When later blocks
2676                                  * get requested, we will try again
2677                                  */
2678                                 continue;
2679                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2680                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2681                                 /* must have already done this block */
2682                                 release_stripe(sh2);
2683                                 continue;
2684                         }
2685
2686                         /* place all the copies on one channel */
2687                         tx = async_memcpy(sh2->dev[dd_idx].page,
2688                                 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2689                                 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2690
2691                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2692                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2693                         for (j = 0; j < conf->raid_disks; j++)
2694                                 if (j != sh2->pd_idx &&
2695                                     (!r6s || j != sh2->qd_idx) &&
2696                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2697                                         break;
2698                         if (j == conf->raid_disks) {
2699                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2700                                 set_bit(STRIPE_HANDLE, &sh2->state);
2701                         }
2702                         release_stripe(sh2);
2703
2704                 }
2705         /* done submitting copies, wait for them to complete */
2706         if (tx) {
2707                 async_tx_ack(tx);
2708                 dma_wait_for_async_tx(tx);
2709         }
2710 }
2711
2712
2713 /*
2714  * handle_stripe - do things to a stripe.
2715  *
2716  * We lock the stripe and then examine the state of various bits
2717  * to see what needs to be done.
2718  * Possible results:
2719  *    return some read request which now have data
2720  *    return some write requests which are safely on disc
2721  *    schedule a read on some buffers
2722  *    schedule a write of some buffers
2723  *    return confirmation of parity correctness
2724  *
2725  * buffers are taken off read_list or write_list, and bh_cache buffers
2726  * get BH_Lock set before the stripe lock is released.
2727  *
2728  */
2729
2730 static bool handle_stripe5(struct stripe_head *sh)
2731 {
2732         raid5_conf_t *conf = sh->raid_conf;
2733         int disks = sh->disks, i;
2734         struct bio *return_bi = NULL;
2735         struct stripe_head_state s;
2736         struct r5dev *dev;
2737         mdk_rdev_t *blocked_rdev = NULL;
2738         int prexor;
2739
2740         memset(&s, 0, sizeof(s));
2741         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2742                  "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2743                  atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2744                  sh->reconstruct_state);
2745
2746         spin_lock(&sh->lock);
2747         clear_bit(STRIPE_HANDLE, &sh->state);
2748         clear_bit(STRIPE_DELAYED, &sh->state);
2749
2750         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2751         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2752         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2753
2754         /* Now to look around and see what can be done */
2755         rcu_read_lock();
2756         for (i=disks; i--; ) {
2757                 mdk_rdev_t *rdev;
2758                 struct r5dev *dev = &sh->dev[i];
2759                 clear_bit(R5_Insync, &dev->flags);
2760
2761                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2762                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2763                         dev->towrite, dev->written);
2764
2765                 /* maybe we can request a biofill operation
2766                  *
2767                  * new wantfill requests are only permitted while
2768                  * ops_complete_biofill is guaranteed to be inactive
2769                  */
2770                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2771                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2772                         set_bit(R5_Wantfill, &dev->flags);
2773
2774                 /* now count some things */
2775                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2776                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2777                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2778
2779                 if (test_bit(R5_Wantfill, &dev->flags))
2780                         s.to_fill++;
2781                 else if (dev->toread)
2782                         s.to_read++;
2783                 if (dev->towrite) {
2784                         s.to_write++;
2785                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2786                                 s.non_overwrite++;
2787                 }
2788                 if (dev->written)
2789                         s.written++;
2790                 rdev = rcu_dereference(conf->disks[i].rdev);
2791                 if (blocked_rdev == NULL &&
2792                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2793                         blocked_rdev = rdev;
2794                         atomic_inc(&rdev->nr_pending);
2795                 }
2796                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2797                         /* The ReadError flag will just be confusing now */
2798                         clear_bit(R5_ReadError, &dev->flags);
2799                         clear_bit(R5_ReWrite, &dev->flags);
2800                 }
2801                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2802                     || test_bit(R5_ReadError, &dev->flags)) {
2803                         s.failed++;
2804                         s.failed_num = i;
2805                 } else
2806                         set_bit(R5_Insync, &dev->flags);
2807         }
2808         rcu_read_unlock();
2809
2810         if (unlikely(blocked_rdev)) {
2811                 if (s.syncing || s.expanding || s.expanded ||
2812                     s.to_write || s.written) {
2813                         set_bit(STRIPE_HANDLE, &sh->state);
2814                         goto unlock;
2815                 }
2816                 /* There is nothing for the blocked_rdev to block */
2817                 rdev_dec_pending(blocked_rdev, conf->mddev);
2818                 blocked_rdev = NULL;
2819         }
2820
2821         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
2822                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
2823                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
2824         }
2825
2826         pr_debug("locked=%d uptodate=%d to_read=%d"
2827                 " to_write=%d failed=%d failed_num=%d\n",
2828                 s.locked, s.uptodate, s.to_read, s.to_write,
2829                 s.failed, s.failed_num);
2830         /* check if the array has lost two devices and, if so, some requests might
2831          * need to be failed
2832          */
2833         if (s.failed > 1 && s.to_read+s.to_write+s.written)
2834                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
2835         if (s.failed > 1 && s.syncing) {
2836                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2837                 clear_bit(STRIPE_SYNCING, &sh->state);
2838                 s.syncing = 0;
2839         }
2840
2841         /* might be able to return some write requests if the parity block
2842          * is safe, or on a failed drive
2843          */
2844         dev = &sh->dev[sh->pd_idx];
2845         if ( s.written &&
2846              ((test_bit(R5_Insync, &dev->flags) &&
2847                !test_bit(R5_LOCKED, &dev->flags) &&
2848                test_bit(R5_UPTODATE, &dev->flags)) ||
2849                (s.failed == 1 && s.failed_num == sh->pd_idx)))
2850                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
2851
2852         /* Now we might consider reading some blocks, either to check/generate
2853          * parity, or to satisfy requests
2854          * or to load a block that is being partially written.
2855          */
2856         if (s.to_read || s.non_overwrite ||
2857             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
2858                 handle_stripe_fill5(sh, &s, disks);
2859
2860         /* Now we check to see if any write operations have recently
2861          * completed
2862          */
2863         prexor = 0;
2864         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
2865                 prexor = 1;
2866         if (sh->reconstruct_state == reconstruct_state_drain_result ||
2867             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
2868                 sh->reconstruct_state = reconstruct_state_idle;
2869
2870                 /* All the 'written' buffers and the parity block are ready to
2871                  * be written back to disk
2872                  */
2873                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2874                 for (i = disks; i--; ) {
2875                         dev = &sh->dev[i];
2876                         if (test_bit(R5_LOCKED, &dev->flags) &&
2877                                 (i == sh->pd_idx || dev->written)) {
2878                                 pr_debug("Writing block %d\n", i);
2879                                 set_bit(R5_Wantwrite, &dev->flags);
2880                                 if (prexor)
2881                                         continue;
2882                                 if (!test_bit(R5_Insync, &dev->flags) ||
2883                                     (i == sh->pd_idx && s.failed == 0))
2884                                         set_bit(STRIPE_INSYNC, &sh->state);
2885                         }
2886                 }
2887                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2888                         atomic_dec(&conf->preread_active_stripes);
2889                         if (atomic_read(&conf->preread_active_stripes) <
2890                                 IO_THRESHOLD)
2891                                 md_wakeup_thread(conf->mddev->thread);
2892                 }
2893         }
2894
2895         /* Now to consider new write requests and what else, if anything
2896          * should be read.  We do not handle new writes when:
2897          * 1/ A 'write' operation (copy+xor) is already in flight.
2898          * 2/ A 'check' operation is in flight, as it may clobber the parity
2899          *    block.
2900          */
2901         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
2902                 handle_stripe_dirtying5(conf, sh, &s, disks);
2903
2904         /* maybe we need to check and possibly fix the parity for this stripe
2905          * Any reads will already have been scheduled, so we just see if enough
2906          * data is available.  The parity check is held off while parity
2907          * dependent operations are in flight.
2908          */
2909         if (sh->check_state ||
2910             (s.syncing && s.locked == 0 &&
2911              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
2912              !test_bit(STRIPE_INSYNC, &sh->state)))
2913                 handle_parity_checks5(conf, sh, &s, disks);
2914
2915         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2916                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2917                 clear_bit(STRIPE_SYNCING, &sh->state);
2918         }
2919
2920         /* If the failed drive is just a ReadError, then we might need to progress
2921          * the repair/check process
2922          */
2923         if (s.failed == 1 && !conf->mddev->ro &&
2924             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2925             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2926             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2927                 ) {
2928                 dev = &sh->dev[s.failed_num];
2929                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2930                         set_bit(R5_Wantwrite, &dev->flags);
2931                         set_bit(R5_ReWrite, &dev->flags);
2932                         set_bit(R5_LOCKED, &dev->flags);
2933                         s.locked++;
2934                 } else {
2935                         /* let's read it back */
2936                         set_bit(R5_Wantread, &dev->flags);
2937                         set_bit(R5_LOCKED, &dev->flags);
2938                         s.locked++;
2939                 }
2940         }
2941
2942         /* Finish reconstruct operations initiated by the expansion process */
2943         if (sh->reconstruct_state == reconstruct_state_result) {
2944                 sh->reconstruct_state = reconstruct_state_idle;
2945                 clear_bit(STRIPE_EXPANDING, &sh->state);
2946                 for (i = conf->raid_disks; i--; ) {
2947                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2948                         set_bit(R5_LOCKED, &sh->dev[i].flags);
2949                         s.locked++;
2950                 }
2951         }
2952
2953         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2954             !sh->reconstruct_state) {
2955                 /* Need to write out all blocks after computing parity */
2956                 sh->disks = conf->raid_disks;
2957                 stripe_set_idx(sh->sector, conf, 0, sh);
2958                 schedule_reconstruction5(sh, &s, 1, 1);
2959         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
2960                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2961                 atomic_dec(&conf->reshape_stripes);
2962                 wake_up(&conf->wait_for_overlap);
2963                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2964         }
2965
2966         if (s.expanding && s.locked == 0 &&
2967             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
2968                 handle_stripe_expansion(conf, sh, NULL);
2969
2970  unlock:
2971         spin_unlock(&sh->lock);
2972
2973         /* wait for this device to become unblocked */
2974         if (unlikely(blocked_rdev))
2975                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
2976
2977         if (s.ops_request)
2978                 raid5_run_ops(sh, s.ops_request);
2979
2980         ops_run_io(sh, &s);
2981
2982         return_io(return_bi);
2983
2984         return blocked_rdev == NULL;
2985 }
2986
2987 static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2988 {
2989         raid5_conf_t *conf = sh->raid_conf;
2990         int disks = sh->disks;
2991         struct bio *return_bi = NULL;
2992         int i, pd_idx = sh->pd_idx;
2993         struct stripe_head_state s;
2994         struct r6_state r6s;
2995         struct r5dev *dev, *pdev, *qdev;
2996         mdk_rdev_t *blocked_rdev = NULL;
2997
2998         r6s.qd_idx = sh->qd_idx;
2999         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3000                 "pd_idx=%d, qd_idx=%d\n",
3001                (unsigned long long)sh->sector, sh->state,
3002                atomic_read(&sh->count), pd_idx, r6s.qd_idx);
3003         memset(&s, 0, sizeof(s));
3004
3005         spin_lock(&sh->lock);
3006         clear_bit(STRIPE_HANDLE, &sh->state);
3007         clear_bit(STRIPE_DELAYED, &sh->state);
3008
3009         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3010         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3011         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3012         /* Now to look around and see what can be done */
3013
3014         rcu_read_lock();
3015         for (i=disks; i--; ) {
3016                 mdk_rdev_t *rdev;
3017                 dev = &sh->dev[i];
3018                 clear_bit(R5_Insync, &dev->flags);
3019
3020                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3021                         i, dev->flags, dev->toread, dev->towrite, dev->written);
3022                 /* maybe we can reply to a read */
3023                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
3024                         struct bio *rbi, *rbi2;
3025                         pr_debug("Return read for disc %d\n", i);
3026                         spin_lock_irq(&conf->device_lock);
3027                         rbi = dev->toread;
3028                         dev->toread = NULL;
3029                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
3030                                 wake_up(&conf->wait_for_overlap);
3031                         spin_unlock_irq(&conf->device_lock);
3032                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
3033                                 copy_data(0, rbi, dev->page, dev->sector);
3034                                 rbi2 = r5_next_bio(rbi, dev->sector);
3035                                 spin_lock_irq(&conf->device_lock);
3036                                 if (!raid5_dec_bi_phys_segments(rbi)) {
3037                                         rbi->bi_next = return_bi;
3038                                         return_bi = rbi;
3039                                 }
3040                                 spin_unlock_irq(&conf->device_lock);
3041                                 rbi = rbi2;
3042                         }
3043                 }
3044
3045                 /* now count some things */
3046                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3047                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3048
3049
3050                 if (dev->toread)
3051                         s.to_read++;
3052                 if (dev->towrite) {
3053                         s.to_write++;
3054                         if (!test_bit(R5_OVERWRITE, &dev->flags))
3055                                 s.non_overwrite++;
3056                 }
3057                 if (dev->written)
3058                         s.written++;
3059                 rdev = rcu_dereference(conf->disks[i].rdev);
3060                 if (blocked_rdev == NULL &&
3061                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3062                         blocked_rdev = rdev;
3063                         atomic_inc(&rdev->nr_pending);
3064                 }
3065                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3066                         /* The ReadError flag will just be confusing now */
3067                         clear_bit(R5_ReadError, &dev->flags);
3068                         clear_bit(R5_ReWrite, &dev->flags);
3069                 }
3070                 if (!rdev || !test_bit(In_sync, &rdev->flags)
3071                     || test_bit(R5_ReadError, &dev->flags)) {
3072                         if (s.failed < 2)
3073                                 r6s.failed_num[s.failed] = i;
3074                         s.failed++;
3075                 } else
3076                         set_bit(R5_Insync, &dev->flags);
3077         }
3078         rcu_read_unlock();
3079
3080         if (unlikely(blocked_rdev)) {
3081                 if (s.syncing || s.expanding || s.expanded ||
3082                     s.to_write || s.written) {
3083                         set_bit(STRIPE_HANDLE, &sh->state);
3084                         goto unlock;
3085                 }
3086                 /* There is nothing for the blocked_rdev to block */
3087                 rdev_dec_pending(blocked_rdev, conf->mddev);
3088                 blocked_rdev = NULL;
3089         }
3090
3091         pr_debug("locked=%d uptodate=%d to_read=%d"
3092                " to_write=%d failed=%d failed_num=%d,%d\n",
3093                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3094                r6s.failed_num[0], r6s.failed_num[1]);
3095         /* check if the array has lost >2 devices and, if so, some requests
3096          * might need to be failed
3097          */
3098         if (s.failed > 2 && s.to_read+s.to_write+s.written)
3099                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3100         if (s.failed > 2 && s.syncing) {
3101                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3102                 clear_bit(STRIPE_SYNCING, &sh->state);
3103                 s.syncing = 0;
3104         }
3105
3106         /*
3107          * might be able to return some write requests if the parity blocks
3108          * are safe, or on a failed drive
3109          */
3110         pdev = &sh->dev[pd_idx];
3111         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3112                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3113         qdev = &sh->dev[r6s.qd_idx];
3114         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
3115                 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
3116
3117         if ( s.written &&
3118              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3119                              && !test_bit(R5_LOCKED, &pdev->flags)
3120                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3121              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3122                              && !test_bit(R5_LOCKED, &qdev->flags)
3123                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3124                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3125
3126         /* Now we might consider reading some blocks, either to check/generate
3127          * parity, or to satisfy requests
3128          * or to load a block that is being partially written.
3129          */
3130         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3131             (s.syncing && (s.uptodate < disks)) || s.expanding)
3132                 handle_stripe_fill6(sh, &s, &r6s, disks);
3133
3134         /* now to consider writing and what else, if anything should be read */
3135         if (s.to_write)
3136                 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3137
3138         /* maybe we need to check and possibly fix the parity for this stripe
3139          * Any reads will already have been scheduled, so we just see if enough
3140          * data is available
3141          */
3142         if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3143                 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3144
3145         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3146                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3147                 clear_bit(STRIPE_SYNCING, &sh->state);
3148         }
3149
3150         /* If the failed drives are just a ReadError, then we might need
3151          * to progress the repair/check process
3152          */
3153         if (s.failed <= 2 && !conf->mddev->ro)
3154                 for (i = 0; i < s.failed; i++) {
3155                         dev = &sh->dev[r6s.failed_num[i]];
3156                         if (test_bit(R5_ReadError, &dev->flags)
3157                             && !test_bit(R5_LOCKED, &dev->flags)
3158                             && test_bit(R5_UPTODATE, &dev->flags)
3159                                 ) {
3160                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3161                                         set_bit(R5_Wantwrite, &dev->flags);
3162                                         set_bit(R5_ReWrite, &dev->flags);
3163                                         set_bit(R5_LOCKED, &dev->flags);
3164                                 } else {
3165                                         /* let's read it back */
3166                                         set_bit(R5_Wantread, &dev->flags);
3167                                         set_bit(R5_LOCKED, &dev->flags);
3168                                 }
3169                         }
3170                 }
3171
3172         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3173                 /* Need to write out all blocks after computing P&Q */
3174                 sh->disks = conf->raid_disks;
3175                 stripe_set_idx(sh->sector, conf, 0, sh);
3176                 compute_parity6(sh, RECONSTRUCT_WRITE);
3177                 for (i = conf->raid_disks ; i-- ;  ) {
3178                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3179                         s.locked++;
3180                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3181                 }
3182                 clear_bit(STRIPE_EXPANDING, &sh->state);
3183         } else if (s.expanded) {
3184                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3185                 atomic_dec(&conf->reshape_stripes);
3186                 wake_up(&conf->wait_for_overlap);
3187                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3188         }
3189
3190         if (s.expanding && s.locked == 0 &&
3191             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3192                 handle_stripe_expansion(conf, sh, &r6s);
3193
3194  unlock:
3195         spin_unlock(&sh->lock);
3196
3197         /* wait for this device to become unblocked */
3198         if (unlikely(blocked_rdev))
3199                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3200
3201         ops_run_io(sh, &s);
3202
3203         return_io(return_bi);
3204
3205         return blocked_rdev == NULL;
3206 }
3207
3208 /* returns true if the stripe was handled */
3209 static bool handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3210 {
3211         if (sh->raid_conf->level == 6)
3212                 return handle_stripe6(sh, tmp_page);
3213         else
3214                 return handle_stripe5(sh);
3215 }
3216
3217
3218
3219 static void raid5_activate_delayed(raid5_conf_t *conf)
3220 {
3221         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3222                 while (!list_empty(&conf->delayed_list)) {
3223                         struct list_head *l = conf->delayed_list.next;
3224                         struct stripe_head *sh;
3225                         sh = list_entry(l, struct stripe_head, lru);
3226                         list_del_init(l);
3227                         clear_bit(STRIPE_DELAYED, &sh->state);
3228                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3229                                 atomic_inc(&conf->preread_active_stripes);
3230                         list_add_tail(&sh->lru, &conf->hold_list);
3231                 }
3232         } else
3233                 blk_plug_device(conf->mddev->queue);
3234 }
3235
3236 static void activate_bit_delay(raid5_conf_t *conf)
3237 {
3238         /* device_lock is held */
3239         struct list_head head;
3240         list_add(&head, &conf->bitmap_list);
3241         list_del_init(&conf->bitmap_list);
3242         while (!list_empty(&head)) {
3243                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3244                 list_del_init(&sh->lru);
3245                 atomic_inc(&sh->count);
3246                 __release_stripe(conf, sh);
3247         }
3248 }
3249
3250 static void unplug_slaves(mddev_t *mddev)
3251 {
3252         raid5_conf_t *conf = mddev_to_conf(mddev);
3253         int i;
3254
3255         rcu_read_lock();
3256         for (i=0; i<mddev->raid_disks; i++) {
3257                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3258                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3259                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3260
3261                         atomic_inc(&rdev->nr_pending);
3262                         rcu_read_unlock();
3263
3264                         blk_unplug(r_queue);
3265
3266                         rdev_dec_pending(rdev, mddev);
3267                         rcu_read_lock();
3268                 }
3269         }
3270         rcu_read_unlock();
3271 }
3272
3273 static void raid5_unplug_device(struct request_queue *q)
3274 {
3275         mddev_t *mddev = q->queuedata;
3276         raid5_conf_t *conf = mddev_to_conf(mddev);
3277         unsigned long flags;
3278
3279         spin_lock_irqsave(&conf->device_lock, flags);
3280
3281         if (blk_remove_plug(q)) {
3282                 conf->seq_flush++;
3283                 raid5_activate_delayed(conf);
3284         }
3285         md_wakeup_thread(mddev->thread);
3286
3287         spin_unlock_irqrestore(&conf->device_lock, flags);
3288
3289         unplug_slaves(mddev);
3290 }
3291
3292 static int raid5_congested(void *data, int bits)
3293 {
3294         mddev_t *mddev = data;
3295         raid5_conf_t *conf = mddev_to_conf(mddev);
3296
3297         /* No difference between reads and writes.  Just check
3298          * how busy the stripe_cache is
3299          */
3300         if (conf->inactive_blocked)
3301                 return 1;
3302         if (conf->quiesce)
3303                 return 1;
3304         if (list_empty_careful(&conf->inactive_list))
3305                 return 1;
3306
3307         return 0;
3308 }
3309
3310 /* We want read requests to align with chunks where possible,
3311  * but write requests don't need to.
3312  */
3313 static int raid5_mergeable_bvec(struct request_queue *q,
3314                                 struct bvec_merge_data *bvm,
3315                                 struct bio_vec *biovec)
3316 {
3317         mddev_t *mddev = q->queuedata;
3318         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3319         int max;
3320         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3321         unsigned int bio_sectors = bvm->bi_size >> 9;
3322
3323         if ((bvm->bi_rw & 1) == WRITE)
3324                 return biovec->bv_len; /* always allow writes to be mergeable */
3325
3326         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3327         if (max < 0) max = 0;
3328         if (max <= biovec->bv_len && bio_sectors == 0)
3329                 return biovec->bv_len;
3330         else
3331                 return max;
3332 }
3333
3334
3335 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3336 {
3337         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3338         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3339         unsigned int bio_sectors = bio->bi_size >> 9;
3340
3341         return  chunk_sectors >=
3342                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3343 }
3344
3345 /*
3346  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3347  *  later sampled by raid5d.
3348  */
3349 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3350 {
3351         unsigned long flags;
3352
3353         spin_lock_irqsave(&conf->device_lock, flags);
3354
3355         bi->bi_next = conf->retry_read_aligned_list;
3356         conf->retry_read_aligned_list = bi;
3357
3358         spin_unlock_irqrestore(&conf->device_lock, flags);
3359         md_wakeup_thread(conf->mddev->thread);
3360 }
3361
3362
3363 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3364 {
3365         struct bio *bi;
3366
3367         bi = conf->retry_read_aligned;
3368         if (bi) {
3369                 conf->retry_read_aligned = NULL;
3370                 return bi;
3371         }
3372         bi = conf->retry_read_aligned_list;
3373         if(bi) {
3374                 conf->retry_read_aligned_list = bi->bi_next;
3375                 bi->bi_next = NULL;
3376                 /*
3377                  * this sets the active strip count to 1 and the processed
3378                  * strip count to zero (upper 8 bits)
3379                  */
3380                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3381         }
3382
3383         return bi;
3384 }
3385
3386
3387 /*
3388  *  The "raid5_align_endio" should check if the read succeeded and if it
3389  *  did, call bio_endio on the original bio (having bio_put the new bio
3390  *  first).
3391  *  If the read failed..
3392  */
3393 static void raid5_align_endio(struct bio *bi, int error)
3394 {
3395         struct bio* raid_bi  = bi->bi_private;
3396         mddev_t *mddev;
3397         raid5_conf_t *conf;
3398         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3399         mdk_rdev_t *rdev;
3400
3401         bio_put(bi);
3402
3403         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3404         conf = mddev_to_conf(mddev);
3405         rdev = (void*)raid_bi->bi_next;
3406         raid_bi->bi_next = NULL;
3407
3408         rdev_dec_pending(rdev, conf->mddev);
3409
3410         if (!error && uptodate) {
3411                 bio_endio(raid_bi, 0);
3412                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3413                         wake_up(&conf->wait_for_stripe);
3414                 return;
3415         }
3416
3417
3418         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3419
3420         add_bio_to_retry(raid_bi, conf);
3421 }
3422
3423 static int bio_fits_rdev(struct bio *bi)
3424 {
3425         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3426
3427         if ((bi->bi_size>>9) > q->max_sectors)
3428                 return 0;
3429         blk_recount_segments(q, bi);
3430         if (bi->bi_phys_segments > q->max_phys_segments)
3431                 return 0;
3432
3433         if (q->merge_bvec_fn)
3434                 /* it's too hard to apply the merge_bvec_fn at this stage,
3435                  * just just give up
3436                  */
3437                 return 0;
3438
3439         return 1;
3440 }
3441
3442
3443 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3444 {
3445         mddev_t *mddev = q->queuedata;
3446         raid5_conf_t *conf = mddev_to_conf(mddev);
3447         unsigned int dd_idx;
3448         struct bio* align_bi;
3449         mdk_rdev_t *rdev;
3450
3451         if (!in_chunk_boundary(mddev, raid_bio)) {
3452                 pr_debug("chunk_aligned_read : non aligned\n");
3453                 return 0;
3454         }
3455         /*
3456          * use bio_clone to make a copy of the bio
3457          */
3458         align_bi = bio_clone(raid_bio, GFP_NOIO);
3459         if (!align_bi)
3460                 return 0;
3461         /*
3462          *   set bi_end_io to a new function, and set bi_private to the
3463          *     original bio.
3464          */
3465         align_bi->bi_end_io  = raid5_align_endio;
3466         align_bi->bi_private = raid_bio;
3467         /*
3468          *      compute position
3469          */
3470         align_bi->bi_sector =  raid5_compute_sector(conf, raid_bio->bi_sector,
3471                                                     0,
3472                                                     &dd_idx, NULL);
3473
3474         rcu_read_lock();
3475         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3476         if (rdev && test_bit(In_sync, &rdev->flags)) {
3477                 atomic_inc(&rdev->nr_pending);
3478                 rcu_read_unlock();
3479                 raid_bio->bi_next = (void*)rdev;
3480                 align_bi->bi_bdev =  rdev->bdev;
3481                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3482                 align_bi->bi_sector += rdev->data_offset;
3483
3484                 if (!bio_fits_rdev(align_bi)) {
3485                         /* too big in some way */
3486                         bio_put(align_bi);
3487                         rdev_dec_pending(rdev, mddev);
3488                         return 0;
3489                 }
3490
3491                 spin_lock_irq(&conf->device_lock);
3492                 wait_event_lock_irq(conf->wait_for_stripe,
3493                                     conf->quiesce == 0,
3494                                     conf->device_lock, /* nothing */);
3495                 atomic_inc(&conf->active_aligned_reads);
3496                 spin_unlock_irq(&conf->device_lock);
3497
3498                 generic_make_request(align_bi);
3499                 return 1;
3500         } else {
3501                 rcu_read_unlock();
3502                 bio_put(align_bi);
3503                 return 0;
3504         }
3505 }
3506
3507 /* __get_priority_stripe - get the next stripe to process
3508  *
3509  * Full stripe writes are allowed to pass preread active stripes up until
3510  * the bypass_threshold is exceeded.  In general the bypass_count
3511  * increments when the handle_list is handled before the hold_list; however, it
3512  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3513  * stripe with in flight i/o.  The bypass_count will be reset when the
3514  * head of the hold_list has changed, i.e. the head was promoted to the
3515  * handle_list.
3516  */
3517 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3518 {
3519         struct stripe_head *sh;
3520
3521         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3522                   __func__,
3523                   list_empty(&conf->handle_list) ? "empty" : "busy",
3524                   list_empty(&conf->hold_list) ? "empty" : "busy",
3525                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
3526
3527         if (!list_empty(&conf->handle_list)) {
3528                 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3529
3530                 if (list_empty(&conf->hold_list))
3531                         conf->bypass_count = 0;
3532                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3533                         if (conf->hold_list.next == conf->last_hold)
3534                                 conf->bypass_count++;
3535                         else {
3536                                 conf->last_hold = conf->hold_list.next;
3537                                 conf->bypass_count -= conf->bypass_threshold;
3538                                 if (conf->bypass_count < 0)
3539                                         conf->bypass_count = 0;
3540                         }
3541                 }
3542         } else if (!list_empty(&conf->hold_list) &&
3543                    ((conf->bypass_threshold &&
3544                      conf->bypass_count > conf->bypass_threshold) ||
3545                     atomic_read(&conf->pending_full_writes) == 0)) {
3546                 sh = list_entry(conf->hold_list.next,
3547                                 typeof(*sh), lru);
3548                 conf->bypass_count -= conf->bypass_threshold;
3549                 if (conf->bypass_count < 0)
3550                         conf->bypass_count = 0;
3551         } else
3552                 return NULL;
3553
3554         list_del_init(&sh->lru);
3555         atomic_inc(&sh->count);
3556         BUG_ON(atomic_read(&sh->count) != 1);
3557         return sh;
3558 }
3559
3560 static int make_request(struct request_queue *q, struct bio * bi)
3561 {
3562         mddev_t *mddev = q->queuedata;
3563         raid5_conf_t *conf = mddev_to_conf(mddev);
3564         int dd_idx;
3565         sector_t new_sector;
3566         sector_t logical_sector, last_sector;
3567         struct stripe_head *sh;
3568         const int rw = bio_data_dir(bi);
3569         int cpu, remaining;
3570
3571         if (unlikely(bio_barrier(bi))) {
3572                 bio_endio(bi, -EOPNOTSUPP);
3573                 return 0;
3574         }
3575
3576         md_write_start(mddev, bi);
3577
3578         cpu = part_stat_lock();
3579         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
3580         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
3581                       bio_sectors(bi));
3582         part_stat_unlock();
3583
3584         if (rw == READ &&
3585              mddev->reshape_position == MaxSector &&
3586              chunk_aligned_read(q,bi))
3587                 return 0;
3588
3589         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3590         last_sector = bi->bi_sector + (bi->bi_size>>9);
3591         bi->bi_next = NULL;
3592         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3593
3594         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3595                 DEFINE_WAIT(w);
3596                 int disks, data_disks;
3597                 int previous;
3598
3599         retry:
3600                 previous = 0;
3601                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3602                 if (likely(conf->expand_progress == MaxSector))
3603                         disks = conf->raid_disks;
3604                 else {
3605                         /* spinlock is needed as expand_progress may be
3606                          * 64bit on a 32bit platform, and so it might be
3607                          * possible to see a half-updated value
3608                          * Ofcourse expand_progress could change after
3609                          * the lock is dropped, so once we get a reference
3610                          * to the stripe that we think it is, we will have
3611                          * to check again.
3612                          */
3613                         spin_lock_irq(&conf->device_lock);
3614                         disks = conf->raid_disks;
3615                         if (logical_sector >= conf->expand_progress) {
3616                                 disks = conf->previous_raid_disks;
3617                                 previous = 1;
3618                         } else {
3619                                 if (logical_sector >= conf->expand_lo) {
3620                                         spin_unlock_irq(&conf->device_lock);
3621                                         schedule();
3622                                         goto retry;
3623                                 }
3624                         }
3625                         spin_unlock_irq(&conf->device_lock);
3626                 }
3627                 data_disks = disks - conf->max_degraded;
3628
3629                 new_sector = raid5_compute_sector(conf, logical_sector,
3630                                                   previous,
3631                                                   &dd_idx, NULL);
3632                 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3633                         (unsigned long long)new_sector, 
3634                         (unsigned long long)logical_sector);
3635
3636                 sh = get_active_stripe(conf, new_sector, previous,
3637                                        (bi->bi_rw&RWA_MASK));
3638                 if (sh) {
3639                         if (unlikely(conf->expand_progress != MaxSector)) {
3640                                 /* expansion might have moved on while waiting for a
3641                                  * stripe, so we must do the range check again.
3642                                  * Expansion could still move past after this
3643                                  * test, but as we are holding a reference to
3644                                  * 'sh', we know that if that happens,
3645                                  *  STRIPE_EXPANDING will get set and the expansion
3646                                  * won't proceed until we finish with the stripe.
3647                                  */
3648                                 int must_retry = 0;
3649                                 spin_lock_irq(&conf->device_lock);
3650                                 if (logical_sector <  conf->expand_progress &&
3651                                     disks == conf->previous_raid_disks)
3652                                         /* mismatch, need to try again */
3653                                         must_retry = 1;
3654                                 spin_unlock_irq(&conf->device_lock);
3655                                 if (must_retry) {
3656                                         release_stripe(sh);
3657                                         goto retry;
3658                                 }
3659                         }
3660                         /* FIXME what if we get a false positive because these
3661                          * are being updated.
3662                          */
3663                         if (logical_sector >= mddev->suspend_lo &&
3664                             logical_sector < mddev->suspend_hi) {
3665                                 release_stripe(sh);
3666                                 schedule();
3667                                 goto retry;
3668                         }
3669
3670                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3671                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3672                                 /* Stripe is busy expanding or
3673                                  * add failed due to overlap.  Flush everything
3674                                  * and wait a while
3675                                  */
3676                                 raid5_unplug_device(mddev->queue);
3677                                 release_stripe(sh);
3678                                 schedule();
3679                                 goto retry;
3680                         }
3681                         finish_wait(&conf->wait_for_overlap, &w);
3682                         set_bit(STRIPE_HANDLE, &sh->state);
3683                         clear_bit(STRIPE_DELAYED, &sh->state);
3684                         release_stripe(sh);
3685                 } else {
3686                         /* cannot get stripe for read-ahead, just give-up */
3687                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3688                         finish_wait(&conf->wait_for_overlap, &w);
3689                         break;
3690                 }
3691                         
3692         }
3693         spin_lock_irq(&conf->device_lock);
3694         remaining = raid5_dec_bi_phys_segments(bi);
3695         spin_unlock_irq(&conf->device_lock);
3696         if (remaining == 0) {
3697
3698                 if ( rw == WRITE )
3699                         md_write_end(mddev);
3700
3701                 bio_endio(bi, 0);
3702         }
3703         return 0;
3704 }
3705
3706 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
3707
3708 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3709 {
3710         /* reshaping is quite different to recovery/resync so it is
3711          * handled quite separately ... here.
3712          *
3713          * On each call to sync_request, we gather one chunk worth of
3714          * destination stripes and flag them as expanding.
3715          * Then we find all the source stripes and request reads.
3716          * As the reads complete, handle_stripe will copy the data
3717          * into the destination stripe and release that stripe.
3718          */
3719         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3720         struct stripe_head *sh;
3721         sector_t first_sector, last_sector;
3722         int raid_disks = conf->previous_raid_disks;
3723         int data_disks = raid_disks - conf->max_degraded;
3724         int new_data_disks = conf->raid_disks - conf->max_degraded;
3725         int i;
3726         int dd_idx;
3727         sector_t writepos, safepos, gap;
3728
3729         if (sector_nr == 0 &&
3730             conf->expand_progress != 0) {
3731                 /* restarting in the middle, skip the initial sectors */
3732                 sector_nr = conf->expand_progress;
3733                 sector_div(sector_nr, new_data_disks);
3734                 *skipped = 1;
3735                 return sector_nr;
3736         }
3737
3738         /* we update the metadata when there is more than 3Meg
3739          * in the block range (that is rather arbitrary, should
3740          * probably be time based) or when the data about to be
3741          * copied would over-write the source of the data at
3742          * the front of the range.
3743          * i.e. one new_stripe forward from expand_progress new_maps
3744          * to after where expand_lo old_maps to
3745          */
3746         writepos = conf->expand_progress +
3747                 conf->chunk_size/512*(new_data_disks);
3748         sector_div(writepos, new_data_disks);
3749         safepos = conf->expand_lo;
3750         sector_div(safepos, data_disks);
3751         gap = conf->expand_progress - conf->expand_lo;
3752
3753         if (writepos >= safepos ||
3754             gap > (new_data_disks)*3000*2 /*3Meg*/) {
3755                 /* Cannot proceed until we've updated the superblock... */
3756                 wait_event(conf->wait_for_overlap,
3757                            atomic_read(&conf->reshape_stripes)==0);
3758                 mddev->reshape_position = conf->expand_progress;
3759                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3760                 md_wakeup_thread(mddev->thread);
3761                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3762                            kthread_should_stop());
3763                 spin_lock_irq(&conf->device_lock);
3764                 conf->expand_lo = mddev->reshape_position;
3765                 spin_unlock_irq(&conf->device_lock);
3766                 wake_up(&conf->wait_for_overlap);
3767         }
3768
3769         for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3770                 int j;
3771                 int skipped = 0;
3772                 sh = get_active_stripe(conf, sector_nr+i, 0, 0);
3773                 set_bit(STRIPE_EXPANDING, &sh->state);
3774                 atomic_inc(&conf->reshape_stripes);
3775                 /* If any of this stripe is beyond the end of the old
3776                  * array, then we need to zero those blocks
3777                  */
3778                 for (j=sh->disks; j--;) {
3779                         sector_t s;
3780                         if (j == sh->pd_idx)
3781                                 continue;
3782                         if (conf->level == 6 &&
3783                             j == sh->qd_idx)
3784                                 continue;
3785                         s = compute_blocknr(sh, j);
3786                         if (s < raid5_size(mddev, 0, 0)) {
3787                                 skipped = 1;
3788                                 continue;
3789                         }
3790                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3791                         set_bit(R5_Expanded, &sh->dev[j].flags);
3792                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
3793                 }
3794                 if (!skipped) {
3795                         set_bit(STRIPE_EXPAND_READY, &sh->state);
3796                         set_bit(STRIPE_HANDLE, &sh->state);
3797                 }
3798                 release_stripe(sh);
3799         }
3800         spin_lock_irq(&conf->device_lock);
3801         conf->expand_progress = (sector_nr + i) * new_data_disks;
3802         spin_unlock_irq(&conf->device_lock);
3803         /* Ok, those stripe are ready. We can start scheduling
3804          * reads on the source stripes.
3805          * The source stripes are determined by mapping the first and last
3806          * block on the destination stripes.
3807          */
3808         first_sector =
3809                 raid5_compute_sector(conf, sector_nr*(new_data_disks),
3810                                      1, &dd_idx, NULL);
3811         last_sector =
3812                 raid5_compute_sector(conf, ((sector_nr+conf->chunk_size/512)
3813                                             *(new_data_disks) - 1),
3814                                      1, &dd_idx, NULL);
3815         if (last_sector >= mddev->dev_sectors)
3816                 last_sector = mddev->dev_sectors - 1;
3817         while (first_sector <= last_sector) {
3818                 sh = get_active_stripe(conf, first_sector, 1, 0);
3819                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3820                 set_bit(STRIPE_HANDLE, &sh->state);
3821                 release_stripe(sh);
3822                 first_sector += STRIPE_SECTORS;
3823         }
3824         /* If this takes us to the resync_max point where we have to pause,
3825          * then we need to write out the superblock.
3826          */
3827         sector_nr += conf->chunk_size>>9;
3828         if (sector_nr >= mddev->resync_max) {
3829                 /* Cannot proceed until we've updated the superblock... */
3830                 wait_event(conf->wait_for_overlap,
3831                            atomic_read(&conf->reshape_stripes) == 0);
3832                 mddev->reshape_position = conf->expand_progress;
3833                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3834                 md_wakeup_thread(mddev->thread);
3835                 wait_event(mddev->sb_wait,
3836                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3837                            || kthread_should_stop());
3838                 spin_lock_irq(&conf->device_lock);
3839                 conf->expand_lo = mddev->reshape_position;
3840                 spin_unlock_irq(&conf->device_lock);
3841                 wake_up(&conf->wait_for_overlap);
3842         }
3843         return conf->chunk_size>>9;
3844 }
3845
3846 /* FIXME go_faster isn't used */
3847 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3848 {
3849         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3850         struct stripe_head *sh;
3851         sector_t max_sector = mddev->dev_sectors;
3852         int sync_blocks;
3853         int still_degraded = 0;
3854         int i;
3855
3856         if (sector_nr >= max_sector) {
3857                 /* just being told to finish up .. nothing much to do */
3858                 unplug_slaves(mddev);
3859                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3860                         end_reshape(conf);
3861                         return 0;
3862                 }
3863
3864                 if (mddev->curr_resync < max_sector) /* aborted */
3865                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3866                                         &sync_blocks, 1);
3867                 else /* completed sync */
3868                         conf->fullsync = 0;
3869                 bitmap_close_sync(mddev->bitmap);
3870
3871                 return 0;
3872         }
3873
3874         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3875                 return reshape_request(mddev, sector_nr, skipped);
3876
3877         /* No need to check resync_max as we never do more than one
3878          * stripe, and as resync_max will always be on a chunk boundary,
3879          * if the check in md_do_sync didn't fire, there is no chance
3880          * of overstepping resync_max here
3881          */
3882
3883         /* if there is too many failed drives and we are trying
3884          * to resync, then assert that we are finished, because there is
3885          * nothing we can do.
3886          */
3887         if (mddev->degraded >= conf->max_degraded &&
3888             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3889                 sector_t rv = mddev->dev_sectors - sector_nr;
3890                 *skipped = 1;
3891                 return rv;
3892         }
3893         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3894             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3895             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3896                 /* we can skip this block, and probably more */
3897                 sync_blocks /= STRIPE_SECTORS;
3898                 *skipped = 1;
3899                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3900         }
3901
3902
3903         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3904
3905         sh = get_active_stripe(conf, sector_nr, 0, 1);
3906         if (sh == NULL) {
3907                 sh = get_active_stripe(conf, sector_nr, 0, 0);
3908                 /* make sure we don't swamp the stripe cache if someone else
3909                  * is trying to get access
3910                  */
3911                 schedule_timeout_uninterruptible(1);
3912         }
3913         /* Need to check if array will still be degraded after recovery/resync
3914          * We don't need to check the 'failed' flag as when that gets set,
3915          * recovery aborts.
3916          */
3917         for (i=0; i<mddev->raid_disks; i++)
3918                 if (conf->disks[i].rdev == NULL)
3919                         still_degraded = 1;
3920
3921         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3922
3923         spin_lock(&sh->lock);
3924         set_bit(STRIPE_SYNCING, &sh->state);
3925         clear_bit(STRIPE_INSYNC, &sh->state);
3926         spin_unlock(&sh->lock);
3927
3928         /* wait for any blocked device to be handled */
3929         while(unlikely(!handle_stripe(sh, NULL)))
3930                 ;
3931         release_stripe(sh);
3932
3933         return STRIPE_SECTORS;
3934 }
3935
3936 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3937 {
3938         /* We may not be able to submit a whole bio at once as there
3939          * may not be enough stripe_heads available.
3940          * We cannot pre-allocate enough stripe_heads as we may need
3941          * more than exist in the cache (if we allow ever large chunks).
3942          * So we do one stripe head at a time and record in
3943          * ->bi_hw_segments how many have been done.
3944          *
3945          * We *know* that this entire raid_bio is in one chunk, so
3946          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3947          */
3948         struct stripe_head *sh;
3949         int dd_idx;
3950         sector_t sector, logical_sector, last_sector;
3951         int scnt = 0;
3952         int remaining;
3953         int handled = 0;
3954
3955         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3956         sector = raid5_compute_sector(conf, logical_sector,
3957                                       0, &dd_idx, NULL);
3958         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3959
3960         for (; logical_sector < last_sector;
3961              logical_sector += STRIPE_SECTORS,
3962                      sector += STRIPE_SECTORS,
3963                      scnt++) {
3964
3965                 if (scnt < raid5_bi_hw_segments(raid_bio))
3966                         /* already done this stripe */
3967                         continue;
3968
3969                 sh = get_active_stripe(conf, sector, 0, 1);
3970
3971                 if (!sh) {
3972                         /* failed to get a stripe - must wait */
3973                         raid5_set_bi_hw_segments(raid_bio, scnt);
3974                         conf->retry_read_aligned = raid_bio;
3975                         return handled;
3976                 }
3977
3978                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3979                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3980                         release_stripe(sh);
3981                         raid5_set_bi_hw_segments(raid_bio, scnt);
3982                         conf->retry_read_aligned = raid_bio;
3983                         return handled;
3984                 }
3985
3986                 handle_stripe(sh, NULL);
3987                 release_stripe(sh);
3988                 handled++;
3989         }
3990         spin_lock_irq(&conf->device_lock);
3991         remaining = raid5_dec_bi_phys_segments(raid_bio);
3992         spin_unlock_irq(&conf->device_lock);
3993         if (remaining == 0)
3994                 bio_endio(raid_bio, 0);
3995         if (atomic_dec_and_test(&conf->active_aligned_reads))
3996                 wake_up(&conf->wait_for_stripe);
3997         return handled;
3998 }
3999
4000
4001
4002 /*
4003  * This is our raid5 kernel thread.
4004  *
4005  * We scan the hash table for stripes which can be handled now.
4006  * During the scan, completed stripes are saved for us by the interrupt
4007  * handler, so that they will not have to wait for our next wakeup.
4008  */
4009 static void raid5d(mddev_t *mddev)
4010 {
4011         struct stripe_head *sh;
4012         raid5_conf_t *conf = mddev_to_conf(mddev);
4013         int handled;
4014
4015         pr_debug("+++ raid5d active\n");
4016
4017         md_check_recovery(mddev);
4018
4019         handled = 0;
4020         spin_lock_irq(&conf->device_lock);
4021         while (1) {
4022                 struct bio *bio;
4023
4024                 if (conf->seq_flush != conf->seq_write) {
4025                         int seq = conf->seq_flush;
4026                         spin_unlock_irq(&conf->device_lock);
4027                         bitmap_unplug(mddev->bitmap);
4028                         spin_lock_irq(&conf->device_lock);
4029                         conf->seq_write = seq;
4030                         activate_bit_delay(conf);
4031                 }
4032
4033                 while ((bio = remove_bio_from_retry(conf))) {
4034                         int ok;
4035                         spin_unlock_irq(&conf->device_lock);
4036                         ok = retry_aligned_read(conf, bio);
4037                         spin_lock_irq(&conf->device_lock);
4038                         if (!ok)
4039                                 break;
4040                         handled++;
4041                 }
4042
4043                 sh = __get_priority_stripe(conf);
4044
4045                 if (!sh)
4046                         break;
4047                 spin_unlock_irq(&conf->device_lock);
4048                 
4049                 handled++;
4050                 handle_stripe(sh, conf->spare_page);
4051                 release_stripe(sh);
4052
4053                 spin_lock_irq(&conf->device_lock);
4054         }
4055         pr_debug("%d stripes handled\n", handled);
4056
4057         spin_unlock_irq(&conf->device_lock);
4058
4059         async_tx_issue_pending_all();
4060         unplug_slaves(mddev);
4061
4062         pr_debug("--- raid5d inactive\n");
4063 }
4064
4065 static ssize_t
4066 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4067 {
4068         raid5_conf_t *conf = mddev_to_conf(mddev);
4069         if (conf)
4070                 return sprintf(page, "%d\n", conf->max_nr_stripes);
4071         else
4072                 return 0;
4073 }
4074
4075 static ssize_t
4076 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4077 {
4078         raid5_conf_t *conf = mddev_to_conf(mddev);
4079         unsigned long new;
4080         int err;
4081
4082         if (len >= PAGE_SIZE)
4083                 return -EINVAL;
4084         if (!conf)
4085                 return -ENODEV;
4086
4087         if (strict_strtoul(page, 10, &new))
4088                 return -EINVAL;
4089         if (new <= 16 || new > 32768)
4090                 return -EINVAL;
4091         while (new < conf->max_nr_stripes) {
4092                 if (drop_one_stripe(conf))
4093                         conf->max_nr_stripes--;
4094                 else
4095                         break;
4096         }
4097         err = md_allow_write(mddev);
4098         if (err)
4099                 return err;
4100         while (new > conf->max_nr_stripes) {
4101                 if (grow_one_stripe(conf))
4102                         conf->max_nr_stripes++;
4103                 else break;
4104         }
4105         return len;
4106 }
4107
4108 static struct md_sysfs_entry
4109 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4110                                 raid5_show_stripe_cache_size,
4111                                 raid5_store_stripe_cache_size);
4112
4113 static ssize_t
4114 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4115 {
4116         raid5_conf_t *conf = mddev_to_conf(mddev);
4117         if (conf)
4118                 return sprintf(page, "%d\n", conf->bypass_threshold);
4119         else
4120                 return 0;
4121 }
4122
4123 static ssize_t
4124 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4125 {
4126         raid5_conf_t *conf = mddev_to_conf(mddev);
4127         unsigned long new;
4128         if (len >= PAGE_SIZE)
4129                 return -EINVAL;
4130         if (!conf)
4131                 return -ENODEV;
4132
4133         if (strict_strtoul(page, 10, &new))
4134                 return -EINVAL;
4135         if (new > conf->max_nr_stripes)
4136                 return -EINVAL;
4137         conf->bypass_threshold = new;
4138         return len;
4139 }
4140
4141 static struct md_sysfs_entry
4142 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4143                                         S_IRUGO | S_IWUSR,
4144                                         raid5_show_preread_threshold,
4145                                         raid5_store_preread_threshold);
4146
4147 static ssize_t
4148 stripe_cache_active_show(mddev_t *mddev, char *page)
4149 {
4150         raid5_conf_t *conf = mddev_to_conf(mddev);
4151         if (conf)
4152                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4153         else
4154                 return 0;
4155 }
4156
4157 static struct md_sysfs_entry
4158 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4159
4160 static struct attribute *raid5_attrs[] =  {
4161         &raid5_stripecache_size.attr,
4162         &raid5_stripecache_active.attr,
4163         &raid5_preread_bypass_threshold.attr,
4164         NULL,
4165 };
4166 static struct attribute_group raid5_attrs_group = {
4167         .name = NULL,
4168         .attrs = raid5_attrs,
4169 };
4170
4171 static sector_t
4172 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4173 {
4174         raid5_conf_t *conf = mddev_to_conf(mddev);
4175
4176         if (!sectors)
4177                 sectors = mddev->dev_sectors;
4178         if (!raid_disks)
4179                 raid_disks = conf->previous_raid_disks;
4180
4181         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4182         return sectors * (raid_disks - conf->max_degraded);
4183 }
4184
4185 static raid5_conf_t *setup_conf(mddev_t *mddev)
4186 {
4187         raid5_conf_t *conf;
4188         int raid_disk, memory;
4189         mdk_rdev_t *rdev;
4190         struct disk_info *disk;
4191
4192         if (mddev->new_level != 5
4193             && mddev->new_level != 4
4194             && mddev->new_level != 6) {
4195                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4196                        mdname(mddev), mddev->new_level);
4197                 return ERR_PTR(-EIO);
4198         }
4199         if ((mddev->new_level == 5
4200              && !algorithm_valid_raid5(mddev->new_layout)) ||
4201             (mddev->new_level == 6
4202              && !algorithm_valid_raid6(mddev->new_layout))) {
4203                 printk(KERN_ERR "raid5: %s: layout %d not supported\n",
4204                        mdname(mddev), mddev->new_layout);
4205                 return ERR_PTR(-EIO);
4206         }
4207         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4208                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4209                        mdname(mddev), mddev->raid_disks);
4210                 return ERR_PTR(-EINVAL);
4211         }
4212
4213         if (!mddev->new_chunk || mddev->new_chunk % PAGE_SIZE) {
4214                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4215                         mddev->new_chunk, mdname(mddev));
4216                 return ERR_PTR(-EINVAL);
4217         }
4218
4219         conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4220         if (conf == NULL)
4221                 goto abort;
4222
4223         conf->raid_disks = mddev->raid_disks;
4224         if (mddev->reshape_position == MaxSector)
4225                 conf->previous_raid_disks = mddev->raid_disks;
4226         else
4227                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4228
4229         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4230                               GFP_KERNEL);
4231         if (!conf->disks)
4232                 goto abort;
4233
4234         conf->mddev = mddev;
4235
4236         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4237                 goto abort;
4238
4239         if (mddev->new_level == 6) {
4240                 conf->spare_page = alloc_page(GFP_KERNEL);
4241                 if (!conf->spare_page)
4242                         goto abort;
4243         }
4244         spin_lock_init(&conf->device_lock);
4245         init_waitqueue_head(&conf->wait_for_stripe);
4246         init_waitqueue_head(&conf->wait_for_overlap);
4247         INIT_LIST_HEAD(&conf->handle_list);
4248         INIT_LIST_HEAD(&conf->hold_list);
4249         INIT_LIST_HEAD(&conf->delayed_list);
4250         INIT_LIST_HEAD(&conf->bitmap_list);
4251         INIT_LIST_HEAD(&conf->inactive_list);
4252         atomic_set(&conf->active_stripes, 0);
4253         atomic_set(&conf->preread_active_stripes, 0);
4254         atomic_set(&conf->active_aligned_reads, 0);
4255         conf->bypass_threshold = BYPASS_THRESHOLD;
4256
4257         pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4258
4259         list_for_each_entry(rdev, &mddev->disks, same_set) {
4260                 raid_disk = rdev->raid_disk;
4261                 if (raid_disk >= conf->raid_disks
4262                     || raid_disk < 0)
4263                         continue;
4264                 disk = conf->disks + raid_disk;
4265
4266                 disk->rdev = rdev;
4267
4268                 if (test_bit(In_sync, &rdev->flags)) {
4269                         char b[BDEVNAME_SIZE];
4270                         printk(KERN_INFO "raid5: device %s operational as raid"
4271                                 " disk %d\n", bdevname(rdev->bdev,b),
4272                                 raid_disk);
4273                 } else
4274                         /* Cannot rely on bitmap to complete recovery */
4275                         conf->fullsync = 1;
4276         }
4277
4278         conf->chunk_size = mddev->new_chunk;
4279         conf->level = mddev->new_level;
4280         if (conf->level == 6)
4281                 conf->max_degraded = 2;
4282         else
4283                 conf->max_degraded = 1;
4284         conf->algorithm = mddev->new_layout;
4285         conf->max_nr_stripes = NR_STRIPES;
4286         conf->expand_progress = mddev->reshape_position;
4287
4288         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4289                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4290         if (grow_stripes(conf, conf->max_nr_stripes)) {
4291                 printk(KERN_ERR
4292                         "raid5: couldn't allocate %dkB for buffers\n", memory);
4293                 goto abort;
4294         } else
4295                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4296                         memory, mdname(mddev));
4297
4298         conf->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4299         if (!conf->thread) {
4300                 printk(KERN_ERR
4301                        "raid5: couldn't allocate thread for %s\n",
4302                        mdname(mddev));
4303                 goto abort;
4304         }
4305
4306         return conf;
4307
4308  abort:
4309         if (conf) {
4310                 shrink_stripes(conf);
4311                 safe_put_page(conf->spare_page);
4312                 kfree(conf->disks);
4313                 kfree(conf->stripe_hashtbl);
4314                 kfree(conf);
4315                 return ERR_PTR(-EIO);
4316         } else
4317                 return ERR_PTR(-ENOMEM);
4318 }
4319
4320 static int run(mddev_t *mddev)
4321 {
4322         raid5_conf_t *conf;
4323         int working_disks = 0;
4324         mdk_rdev_t *rdev;
4325
4326         if (mddev->reshape_position != MaxSector) {
4327                 /* Check that we can continue the reshape.
4328                  * Currently only disks can change, it must
4329                  * increase, and we must be past the point where
4330                  * a stripe over-writes itself
4331                  */
4332                 sector_t here_new, here_old;
4333                 int old_disks;
4334                 int max_degraded = (mddev->level == 6 ? 2 : 1);
4335
4336                 if (mddev->new_level != mddev->level ||
4337                     mddev->new_layout != mddev->layout ||
4338                     mddev->new_chunk != mddev->chunk_size) {
4339                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4340                                "required - aborting.\n",
4341                                mdname(mddev));
4342                         return -EINVAL;
4343                 }
4344                 if (mddev->delta_disks <= 0) {
4345                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4346                                "(reduce disks) required - aborting.\n",
4347                                mdname(mddev));
4348                         return -EINVAL;
4349                 }
4350                 old_disks = mddev->raid_disks - mddev->delta_disks;
4351                 /* reshape_position must be on a new-stripe boundary, and one
4352                  * further up in new geometry must map after here in old
4353                  * geometry.
4354                  */
4355                 here_new = mddev->reshape_position;
4356                 if (sector_div(here_new, (mddev->chunk_size>>9)*
4357                                (mddev->raid_disks - max_degraded))) {
4358                         printk(KERN_ERR "raid5: reshape_position not "
4359                                "on a stripe boundary\n");
4360                         return -EINVAL;
4361                 }
4362                 /* here_new is the stripe we will write to */
4363                 here_old = mddev->reshape_position;
4364                 sector_div(here_old, (mddev->chunk_size>>9)*
4365                            (old_disks-max_degraded));
4366                 /* here_old is the first stripe that we might need to read
4367                  * from */
4368                 if (here_new >= here_old) {
4369                         /* Reading from the same stripe as writing to - bad */
4370                         printk(KERN_ERR "raid5: reshape_position too early for "
4371                                "auto-recovery - aborting.\n");
4372                         return -EINVAL;
4373                 }
4374                 printk(KERN_INFO "raid5: reshape will continue\n");
4375                 /* OK, we should be able to continue; */
4376         } else {
4377                 BUG_ON(mddev->level != mddev->new_level);
4378                 BUG_ON(mddev->layout != mddev->new_layout);
4379                 BUG_ON(mddev->chunk_size != mddev->new_chunk);
4380                 BUG_ON(mddev->delta_disks != 0);
4381         }
4382
4383         if (mddev->private == NULL)
4384                 conf = setup_conf(mddev);
4385         else
4386                 conf = mddev->private;
4387
4388         if (IS_ERR(conf))
4389                 return PTR_ERR(conf);
4390
4391         mddev->thread = conf->thread;
4392         conf->thread = NULL;
4393         mddev->private = conf;
4394
4395         /*
4396          * 0 for a fully functional array, 1 or 2 for a degraded array.
4397          */
4398         list_for_each_entry(rdev, &mddev->disks, same_set)
4399                 if (rdev->raid_disk >= 0 &&
4400                     test_bit(In_sync, &rdev->flags))
4401                         working_disks++;
4402
4403         mddev->degraded = conf->raid_disks - working_disks;
4404
4405         if (mddev->degraded > conf->max_degraded) {
4406                 printk(KERN_ERR "raid5: not enough operational devices for %s"
4407                         " (%d/%d failed)\n",
4408                         mdname(mddev), mddev->degraded, conf->raid_disks);
4409                 goto abort;
4410         }
4411
4412         /* device size must be a multiple of chunk size */
4413         mddev->dev_sectors &= ~(mddev->chunk_size / 512 - 1);
4414         mddev->resync_max_sectors = mddev->dev_sectors;
4415
4416         if (mddev->degraded > 0 &&
4417             mddev->recovery_cp != MaxSector) {
4418                 if (mddev->ok_start_degraded)
4419                         printk(KERN_WARNING
4420                                "raid5: starting dirty degraded array: %s"
4421                                "- data corruption possible.\n",
4422                                mdname(mddev));
4423                 else {
4424                         printk(KERN_ERR
4425                                "raid5: cannot start dirty degraded array for %s\n",
4426                                mdname(mddev));
4427                         goto abort;
4428                 }
4429         }
4430
4431         if (mddev->degraded == 0)
4432                 printk("raid5: raid level %d set %s active with %d out of %d"
4433                         " devices, algorithm %d\n", conf->level, mdname(mddev), 
4434                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4435                         conf->algorithm);
4436         else
4437                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4438                         " out of %d devices, algorithm %d\n", conf->level,
4439                         mdname(mddev), mddev->raid_disks - mddev->degraded,
4440                         mddev->raid_disks, conf->algorithm);
4441
4442         print_raid5_conf(conf);
4443
4444         if (conf->expand_progress != MaxSector) {
4445                 printk("...ok start reshape thread\n");
4446                 conf->expand_lo = conf->expand_progress;
4447                 atomic_set(&conf->reshape_stripes, 0);
4448                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4449                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4450                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4451                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4452                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4453                                                         "%s_reshape");
4454         }
4455
4456         /* read-ahead size must cover two whole stripes, which is
4457          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4458          */
4459         {
4460                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4461                 int stripe = data_disks *
4462                         (mddev->chunk_size / PAGE_SIZE);
4463                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4464                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4465         }
4466
4467         /* Ok, everything is just fine now */
4468         if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4469                 printk(KERN_WARNING
4470                        "raid5: failed to create sysfs attributes for %s\n",
4471                        mdname(mddev));
4472
4473         mddev->queue->queue_lock = &conf->device_lock;
4474
4475         mddev->queue->unplug_fn = raid5_unplug_device;
4476         mddev->queue->backing_dev_info.congested_data = mddev;
4477         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4478
4479         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4480
4481         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4482
4483         return 0;
4484 abort:
4485         md_unregister_thread(mddev->thread);
4486         mddev->thread = NULL;
4487         if (conf) {
4488                 shrink_stripes(conf);
4489                 print_raid5_conf(conf);
4490                 safe_put_page(conf->spare_page);
4491                 kfree(conf->disks);
4492                 kfree(conf->stripe_hashtbl);
4493                 kfree(conf);
4494         }
4495         mddev->private = NULL;
4496         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4497         return -EIO;
4498 }
4499
4500
4501
4502 static int stop(mddev_t *mddev)
4503 {
4504         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4505
4506         md_unregister_thread(mddev->thread);
4507         mddev->thread = NULL;
4508         shrink_stripes(conf);
4509         kfree(conf->stripe_hashtbl);
4510         mddev->queue->backing_dev_info.congested_fn = NULL;
4511         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4512         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4513         kfree(conf->disks);
4514         kfree(conf);
4515         mddev->private = NULL;
4516         return 0;
4517 }
4518
4519 #ifdef DEBUG
4520 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
4521 {
4522         int i;
4523
4524         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4525                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4526         seq_printf(seq, "sh %llu,  count %d.\n",
4527                    (unsigned long long)sh->sector, atomic_read(&sh->count));
4528         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4529         for (i = 0; i < sh->disks; i++) {
4530                 seq_printf(seq, "(cache%d: %p %ld) ",
4531                            i, sh->dev[i].page, sh->dev[i].flags);
4532         }
4533         seq_printf(seq, "\n");
4534 }
4535
4536 static void printall(struct seq_file *seq, raid5_conf_t *conf)
4537 {
4538         struct stripe_head *sh;
4539         struct hlist_node *hn;
4540         int i;
4541
4542         spin_lock_irq(&conf->device_lock);
4543         for (i = 0; i < NR_HASH; i++) {
4544                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4545                         if (sh->raid_conf != conf)
4546                                 continue;
4547                         print_sh(seq, sh);
4548                 }
4549         }
4550         spin_unlock_irq(&conf->device_lock);
4551 }
4552 #endif
4553
4554 static void status(struct seq_file *seq, mddev_t *mddev)
4555 {
4556         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4557         int i;
4558
4559         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4560         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4561         for (i = 0; i < conf->raid_disks; i++)
4562                 seq_printf (seq, "%s",
4563                                conf->disks[i].rdev &&
4564                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4565         seq_printf (seq, "]");
4566 #ifdef DEBUG
4567         seq_printf (seq, "\n");
4568         printall(seq, conf);
4569 #endif
4570 }
4571
4572 static void print_raid5_conf (raid5_conf_t *conf)
4573 {
4574         int i;
4575         struct disk_info *tmp;
4576
4577         printk("RAID5 conf printout:\n");
4578         if (!conf) {
4579                 printk("(conf==NULL)\n");
4580                 return;
4581         }
4582         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4583                  conf->raid_disks - conf->mddev->degraded);
4584
4585         for (i = 0; i < conf->raid_disks; i++) {
4586                 char b[BDEVNAME_SIZE];
4587                 tmp = conf->disks + i;
4588                 if (tmp->rdev)
4589                 printk(" disk %d, o:%d, dev:%s\n",
4590                         i, !test_bit(Faulty, &tmp->rdev->flags),
4591                         bdevname(tmp->rdev->bdev,b));
4592         }
4593 }
4594
4595 static int raid5_spare_active(mddev_t *mddev)
4596 {
4597         int i;
4598         raid5_conf_t *conf = mddev->private;
4599         struct disk_info *tmp;
4600
4601         for (i = 0; i < conf->raid_disks; i++) {
4602                 tmp = conf->disks + i;
4603                 if (tmp->rdev
4604                     && !test_bit(Faulty, &tmp->rdev->flags)
4605                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4606                         unsigned long flags;
4607                         spin_lock_irqsave(&conf->device_lock, flags);
4608                         mddev->degraded--;
4609                         spin_unlock_irqrestore(&conf->device_lock, flags);
4610                 }
4611         }
4612         print_raid5_conf(conf);
4613         return 0;
4614 }
4615
4616 static int raid5_remove_disk(mddev_t *mddev, int number)
4617 {
4618         raid5_conf_t *conf = mddev->private;
4619         int err = 0;
4620         mdk_rdev_t *rdev;
4621         struct disk_info *p = conf->disks + number;
4622
4623         print_raid5_conf(conf);
4624         rdev = p->rdev;
4625         if (rdev) {
4626                 if (test_bit(In_sync, &rdev->flags) ||
4627                     atomic_read(&rdev->nr_pending)) {
4628                         err = -EBUSY;
4629                         goto abort;
4630                 }
4631                 /* Only remove non-faulty devices if recovery
4632                  * isn't possible.
4633                  */
4634                 if (!test_bit(Faulty, &rdev->flags) &&
4635                     mddev->degraded <= conf->max_degraded) {
4636                         err = -EBUSY;
4637                         goto abort;
4638                 }
4639                 p->rdev = NULL;
4640                 synchronize_rcu();
4641                 if (atomic_read(&rdev->nr_pending)) {
4642                         /* lost the race, try later */
4643                         err = -EBUSY;
4644                         p->rdev = rdev;
4645                 }
4646         }
4647 abort:
4648
4649         print_raid5_conf(conf);
4650         return err;
4651 }
4652
4653 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4654 {
4655         raid5_conf_t *conf = mddev->private;
4656         int err = -EEXIST;
4657         int disk;
4658         struct disk_info *p;
4659         int first = 0;
4660         int last = conf->raid_disks - 1;
4661
4662         if (mddev->degraded > conf->max_degraded)
4663                 /* no point adding a device */
4664                 return -EINVAL;
4665
4666         if (rdev->raid_disk >= 0)
4667                 first = last = rdev->raid_disk;
4668
4669         /*
4670          * find the disk ... but prefer rdev->saved_raid_disk
4671          * if possible.
4672          */
4673         if (rdev->saved_raid_disk >= 0 &&
4674             rdev->saved_raid_disk >= first &&
4675             conf->disks[rdev->saved_raid_disk].rdev == NULL)
4676                 disk = rdev->saved_raid_disk;
4677         else
4678                 disk = first;
4679         for ( ; disk <= last ; disk++)
4680                 if ((p=conf->disks + disk)->rdev == NULL) {
4681                         clear_bit(In_sync, &rdev->flags);
4682                         rdev->raid_disk = disk;
4683                         err = 0;
4684                         if (rdev->saved_raid_disk != disk)
4685                                 conf->fullsync = 1;
4686                         rcu_assign_pointer(p->rdev, rdev);
4687                         break;
4688                 }
4689         print_raid5_conf(conf);
4690         return err;
4691 }
4692
4693 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4694 {
4695         /* no resync is happening, and there is enough space
4696          * on all devices, so we can resize.
4697          * We need to make sure resync covers any new space.
4698          * If the array is shrinking we should possibly wait until
4699          * any io in the removed space completes, but it hardly seems
4700          * worth it.
4701          */
4702         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4703         md_set_array_sectors(mddev, raid5_size(mddev, sectors,
4704                                                mddev->raid_disks));
4705         if (mddev->array_sectors >
4706             raid5_size(mddev, sectors, mddev->raid_disks))
4707                 return -EINVAL;
4708         set_capacity(mddev->gendisk, mddev->array_sectors);
4709         mddev->changed = 1;
4710         if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
4711                 mddev->recovery_cp = mddev->dev_sectors;
4712                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4713         }
4714         mddev->dev_sectors = sectors;
4715         mddev->resync_max_sectors = sectors;
4716         return 0;
4717 }
4718
4719 #ifdef CONFIG_MD_RAID5_RESHAPE
4720 static int raid5_check_reshape(mddev_t *mddev)
4721 {
4722         raid5_conf_t *conf = mddev_to_conf(mddev);
4723         int err;
4724
4725         if (mddev->delta_disks < 0 ||
4726             mddev->new_level != mddev->level)
4727                 return -EINVAL; /* Cannot shrink array or change level yet */
4728         if (mddev->delta_disks == 0)
4729                 return 0; /* nothing to do */
4730         if (mddev->bitmap)
4731                 /* Cannot grow a bitmap yet */
4732                 return -EBUSY;
4733
4734         /* Can only proceed if there are plenty of stripe_heads.
4735          * We need a minimum of one full stripe,, and for sensible progress
4736          * it is best to have about 4 times that.
4737          * If we require 4 times, then the default 256 4K stripe_heads will
4738          * allow for chunk sizes up to 256K, which is probably OK.
4739          * If the chunk size is greater, user-space should request more
4740          * stripe_heads first.
4741          */
4742         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4743             (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4744                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
4745                        (mddev->chunk_size / STRIPE_SIZE)*4);
4746                 return -ENOSPC;
4747         }
4748
4749         err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4750         if (err)
4751                 return err;
4752
4753         if (mddev->degraded > conf->max_degraded)
4754                 return -EINVAL;
4755         /* looks like we might be able to manage this */
4756         return 0;
4757 }
4758
4759 static int raid5_start_reshape(mddev_t *mddev)
4760 {
4761         raid5_conf_t *conf = mddev_to_conf(mddev);
4762         mdk_rdev_t *rdev;
4763         int spares = 0;
4764         int added_devices = 0;
4765         unsigned long flags;
4766
4767         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4768                 return -EBUSY;
4769
4770         list_for_each_entry(rdev, &mddev->disks, same_set)
4771                 if (rdev->raid_disk < 0 &&
4772                     !test_bit(Faulty, &rdev->flags))
4773                         spares++;
4774
4775         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4776                 /* Not enough devices even to make a degraded array
4777                  * of that size
4778                  */
4779                 return -EINVAL;
4780
4781         atomic_set(&conf->reshape_stripes, 0);
4782         spin_lock_irq(&conf->device_lock);
4783         conf->previous_raid_disks = conf->raid_disks;
4784         conf->raid_disks += mddev->delta_disks;
4785         conf->expand_progress = 0;
4786         conf->expand_lo = 0;
4787         spin_unlock_irq(&conf->device_lock);
4788
4789         /* Add some new drives, as many as will fit.
4790          * We know there are enough to make the newly sized array work.
4791          */
4792         list_for_each_entry(rdev, &mddev->disks, same_set)
4793                 if (rdev->raid_disk < 0 &&
4794                     !test_bit(Faulty, &rdev->flags)) {
4795                         if (raid5_add_disk(mddev, rdev) == 0) {
4796                                 char nm[20];
4797                                 set_bit(In_sync, &rdev->flags);
4798                                 added_devices++;
4799                                 rdev->recovery_offset = 0;
4800                                 sprintf(nm, "rd%d", rdev->raid_disk);
4801                                 if (sysfs_create_link(&mddev->kobj,
4802                                                       &rdev->kobj, nm))
4803                                         printk(KERN_WARNING
4804                                                "raid5: failed to create "
4805                                                " link %s for %s\n",
4806                                                nm, mdname(mddev));
4807                         } else
4808                                 break;
4809                 }
4810
4811         spin_lock_irqsave(&conf->device_lock, flags);
4812         mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4813         spin_unlock_irqrestore(&conf->device_lock, flags);
4814         mddev->raid_disks = conf->raid_disks;
4815         mddev->reshape_position = 0;
4816         set_bit(MD_CHANGE_DEVS, &mddev->flags);
4817
4818         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4819         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4820         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4821         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4822         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4823                                                 "%s_reshape");
4824         if (!mddev->sync_thread) {
4825                 mddev->recovery = 0;
4826                 spin_lock_irq(&conf->device_lock);
4827                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4828                 conf->expand_progress = MaxSector;
4829                 spin_unlock_irq(&conf->device_lock);
4830                 return -EAGAIN;
4831         }
4832         md_wakeup_thread(mddev->sync_thread);
4833         md_new_event(mddev);
4834         return 0;
4835 }
4836 #endif
4837
4838 static void end_reshape(raid5_conf_t *conf)
4839 {
4840         struct block_device *bdev;
4841
4842         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4843                 mddev_t *mddev = conf->mddev;
4844
4845                 md_set_array_sectors_lock(mddev, raid5_size(mddev, 0,
4846                                                        conf->raid_disks));
4847                 set_capacity(mddev->gendisk, mddev->array_sectors);
4848                 mddev->changed = 1;
4849                 conf->previous_raid_disks = conf->raid_disks;
4850
4851                 bdev = bdget_disk(conf->mddev->gendisk, 0);
4852                 if (bdev) {
4853                         mutex_lock(&bdev->bd_inode->i_mutex);
4854                         i_size_write(bdev->bd_inode,
4855                                      (loff_t)conf->mddev->array_sectors << 9);
4856                         mutex_unlock(&bdev->bd_inode->i_mutex);
4857                         bdput(bdev);
4858                 }
4859                 spin_lock_irq(&conf->device_lock);
4860                 conf->expand_progress = MaxSector;
4861                 spin_unlock_irq(&conf->device_lock);
4862                 conf->mddev->reshape_position = MaxSector;
4863
4864                 /* read-ahead size must cover two whole stripes, which is
4865                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4866                  */
4867                 {
4868                         int data_disks = conf->previous_raid_disks - conf->max_degraded;
4869                         int stripe = data_disks *
4870                                 (conf->mddev->chunk_size / PAGE_SIZE);
4871                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4872                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4873                 }
4874         }
4875 }
4876
4877 static void raid5_quiesce(mddev_t *mddev, int state)
4878 {
4879         raid5_conf_t *conf = mddev_to_conf(mddev);
4880
4881         switch(state) {
4882         case 2: /* resume for a suspend */
4883                 wake_up(&conf->wait_for_overlap);
4884                 break;
4885
4886         case 1: /* stop all writes */
4887                 spin_lock_irq(&conf->device_lock);
4888                 conf->quiesce = 1;
4889                 wait_event_lock_irq(conf->wait_for_stripe,
4890                                     atomic_read(&conf->active_stripes) == 0 &&
4891                                     atomic_read(&conf->active_aligned_reads) == 0,
4892                                     conf->device_lock, /* nothing */);
4893                 spin_unlock_irq(&conf->device_lock);
4894                 break;
4895
4896         case 0: /* re-enable writes */
4897                 spin_lock_irq(&conf->device_lock);
4898                 conf->quiesce = 0;
4899                 wake_up(&conf->wait_for_stripe);
4900                 wake_up(&conf->wait_for_overlap);
4901                 spin_unlock_irq(&conf->device_lock);
4902                 break;
4903         }
4904 }
4905
4906
4907 static void *raid5_takeover_raid1(mddev_t *mddev)
4908 {
4909         int chunksect;
4910
4911         if (mddev->raid_disks != 2 ||
4912             mddev->degraded > 1)
4913                 return ERR_PTR(-EINVAL);
4914
4915         /* Should check if there are write-behind devices? */
4916
4917         chunksect = 64*2; /* 64K by default */
4918
4919         /* The array must be an exact multiple of chunksize */
4920         while (chunksect && (mddev->array_sectors & (chunksect-1)))
4921                 chunksect >>= 1;
4922
4923         if ((chunksect<<9) < STRIPE_SIZE)
4924                 /* array size does not allow a suitable chunk size */
4925                 return ERR_PTR(-EINVAL);
4926
4927         mddev->new_level = 5;
4928         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
4929         mddev->new_chunk = chunksect << 9;
4930
4931         return setup_conf(mddev);
4932 }
4933
4934 static void *raid5_takeover_raid6(mddev_t *mddev)
4935 {
4936         int new_layout;
4937
4938         switch (mddev->layout) {
4939         case ALGORITHM_LEFT_ASYMMETRIC_6:
4940                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
4941                 break;
4942         case ALGORITHM_RIGHT_ASYMMETRIC_6:
4943                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
4944                 break;
4945         case ALGORITHM_LEFT_SYMMETRIC_6:
4946                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
4947                 break;
4948         case ALGORITHM_RIGHT_SYMMETRIC_6:
4949                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
4950                 break;
4951         case ALGORITHM_PARITY_0_6:
4952                 new_layout = ALGORITHM_PARITY_0;
4953                 break;
4954         case ALGORITHM_PARITY_N:
4955                 new_layout = ALGORITHM_PARITY_N;
4956                 break;
4957         default:
4958                 return ERR_PTR(-EINVAL);
4959         }
4960         mddev->new_level = 5;
4961         mddev->new_layout = new_layout;
4962         mddev->delta_disks = -1;
4963         mddev->raid_disks -= 1;
4964         return setup_conf(mddev);
4965 }
4966
4967
4968 static int raid5_reconfig(mddev_t *mddev, int new_layout, int new_chunk)
4969 {
4970         /* Currently the layout and chunk size can only be changed
4971          * for a 2-drive raid array, as in that case no data shuffling
4972          * is required.
4973          * Later we might validate these and set new_* so a reshape
4974          * can complete the change.
4975          */
4976         raid5_conf_t *conf = mddev_to_conf(mddev);
4977
4978         if (new_layout >= 0 && !algorithm_valid_raid5(new_layout))
4979                 return -EINVAL;
4980         if (new_chunk > 0) {
4981                 if (new_chunk & (new_chunk-1))
4982                         /* not a power of 2 */
4983                         return -EINVAL;
4984                 if (new_chunk < PAGE_SIZE)
4985                         return -EINVAL;
4986                 if (mddev->array_sectors & ((new_chunk>>9)-1))
4987                         /* not factor of array size */
4988                         return -EINVAL;
4989         }
4990
4991         /* They look valid */
4992
4993         if (mddev->raid_disks != 2)
4994                 return -EINVAL;
4995
4996         if (new_layout >= 0) {
4997                 conf->algorithm = new_layout;
4998                 mddev->layout = mddev->new_layout = new_layout;
4999         }
5000         if (new_chunk > 0) {
5001                 conf->chunk_size = new_chunk;
5002                 mddev->chunk_size = mddev->new_chunk = new_chunk;
5003         }
5004         set_bit(MD_CHANGE_DEVS, &mddev->flags);
5005         md_wakeup_thread(mddev->thread);
5006         return 0;
5007 }
5008
5009 static void *raid5_takeover(mddev_t *mddev)
5010 {
5011         /* raid5 can take over:
5012          *  raid0 - if all devices are the same - make it a raid4 layout
5013          *  raid1 - if there are two drives.  We need to know the chunk size
5014          *  raid4 - trivial - just use a raid4 layout.
5015          *  raid6 - Providing it is a *_6 layout
5016          *
5017          * For now, just do raid1
5018          */
5019
5020         if (mddev->level == 1)
5021                 return raid5_takeover_raid1(mddev);
5022         if (mddev->level == 4) {
5023                 mddev->new_layout = ALGORITHM_PARITY_N;
5024                 mddev->new_level = 5;
5025                 return setup_conf(mddev);
5026         }
5027         if (mddev->level == 6)
5028                 return raid5_takeover_raid6(mddev);
5029
5030         return ERR_PTR(-EINVAL);
5031 }
5032
5033
5034 static struct mdk_personality raid5_personality;
5035
5036 static void *raid6_takeover(mddev_t *mddev)
5037 {
5038         /* Currently can only take over a raid5.  We map the
5039          * personality to an equivalent raid6 personality
5040          * with the Q block at the end.
5041          */
5042         int new_layout;
5043
5044         if (mddev->pers != &raid5_personality)
5045                 return ERR_PTR(-EINVAL);
5046         if (mddev->degraded > 1)
5047                 return ERR_PTR(-EINVAL);
5048         if (mddev->raid_disks > 253)
5049                 return ERR_PTR(-EINVAL);
5050         if (mddev->raid_disks < 3)
5051                 return ERR_PTR(-EINVAL);
5052
5053         switch (mddev->layout) {
5054         case ALGORITHM_LEFT_ASYMMETRIC:
5055                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5056                 break;
5057         case ALGORITHM_RIGHT_ASYMMETRIC:
5058                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5059                 break;
5060         case ALGORITHM_LEFT_SYMMETRIC:
5061                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5062                 break;
5063         case ALGORITHM_RIGHT_SYMMETRIC:
5064                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5065                 break;
5066         case ALGORITHM_PARITY_0:
5067                 new_layout = ALGORITHM_PARITY_0_6;
5068                 break;
5069         case ALGORITHM_PARITY_N:
5070                 new_layout = ALGORITHM_PARITY_N;
5071                 break;
5072         default:
5073                 return ERR_PTR(-EINVAL);
5074         }
5075         mddev->new_level = 6;
5076         mddev->new_layout = new_layout;
5077         mddev->delta_disks = 1;
5078         mddev->raid_disks += 1;
5079         return setup_conf(mddev);
5080 }
5081
5082
5083 static struct mdk_personality raid6_personality =
5084 {
5085         .name           = "raid6",
5086         .level          = 6,
5087         .owner          = THIS_MODULE,
5088         .make_request   = make_request,
5089         .run            = run,
5090         .stop           = stop,
5091         .status         = status,
5092         .error_handler  = error,
5093         .hot_add_disk   = raid5_add_disk,
5094         .hot_remove_disk= raid5_remove_disk,
5095         .spare_active   = raid5_spare_active,
5096         .sync_request   = sync_request,
5097         .resize         = raid5_resize,
5098         .size           = raid5_size,
5099 #ifdef CONFIG_MD_RAID5_RESHAPE
5100         .check_reshape  = raid5_check_reshape,
5101         .start_reshape  = raid5_start_reshape,
5102 #endif
5103         .quiesce        = raid5_quiesce,
5104         .takeover       = raid6_takeover,
5105 };
5106 static struct mdk_personality raid5_personality =
5107 {
5108         .name           = "raid5",
5109         .level          = 5,
5110         .owner          = THIS_MODULE,
5111         .make_request   = make_request,
5112         .run            = run,
5113         .stop           = stop,
5114         .status         = status,
5115         .error_handler  = error,
5116         .hot_add_disk   = raid5_add_disk,
5117         .hot_remove_disk= raid5_remove_disk,
5118         .spare_active   = raid5_spare_active,
5119         .sync_request   = sync_request,
5120         .resize         = raid5_resize,
5121         .size           = raid5_size,
5122 #ifdef CONFIG_MD_RAID5_RESHAPE
5123         .check_reshape  = raid5_check_reshape,
5124         .start_reshape  = raid5_start_reshape,
5125 #endif
5126         .quiesce        = raid5_quiesce,
5127         .takeover       = raid5_takeover,
5128         .reconfig       = raid5_reconfig,
5129 };
5130
5131 static struct mdk_personality raid4_personality =
5132 {
5133         .name           = "raid4",
5134         .level          = 4,
5135         .owner          = THIS_MODULE,
5136         .make_request   = make_request,
5137         .run            = run,
5138         .stop           = stop,
5139         .status         = status,
5140         .error_handler  = error,
5141         .hot_add_disk   = raid5_add_disk,
5142         .hot_remove_disk= raid5_remove_disk,
5143         .spare_active   = raid5_spare_active,
5144         .sync_request   = sync_request,
5145         .resize         = raid5_resize,
5146         .size           = raid5_size,
5147 #ifdef CONFIG_MD_RAID5_RESHAPE
5148         .check_reshape  = raid5_check_reshape,
5149         .start_reshape  = raid5_start_reshape,
5150 #endif
5151         .quiesce        = raid5_quiesce,
5152 };
5153
5154 static int __init raid5_init(void)
5155 {
5156         int e;
5157
5158         e = raid6_select_algo();
5159         if ( e )
5160                 return e;
5161         register_md_personality(&raid6_personality);
5162         register_md_personality(&raid5_personality);
5163         register_md_personality(&raid4_personality);
5164         return 0;
5165 }
5166
5167 static void raid5_exit(void)
5168 {
5169         unregister_md_personality(&raid6_personality);
5170         unregister_md_personality(&raid5_personality);
5171         unregister_md_personality(&raid4_personality);
5172 }
5173
5174 module_init(raid5_init);
5175 module_exit(raid5_exit);
5176 MODULE_LICENSE("GPL");
5177 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5178 MODULE_ALIAS("md-raid5");
5179 MODULE_ALIAS("md-raid4");
5180 MODULE_ALIAS("md-level-5");
5181 MODULE_ALIAS("md-level-4");
5182 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5183 MODULE_ALIAS("md-raid6");
5184 MODULE_ALIAS("md-level-6");
5185
5186 /* This used to be two separate modules, they were: */
5187 MODULE_ALIAS("raid5");
5188 MODULE_ALIAS("raid6");