7c22e19aca82f5cc312c8b14fe49a9fc7ec55cea
[pandora-kernel.git] / drivers / md / raid5.c
1 /*
2  * raid5.c : Multiple Devices driver for Linux
3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4  *         Copyright (C) 1999, 2000 Ingo Molnar
5  *         Copyright (C) 2002, 2003 H. Peter Anvin
6  *
7  * RAID-4/5/6 management functions.
8  * Thanks to Penguin Computing for making the RAID-6 development possible
9  * by donating a test server!
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 /*
22  * BITMAP UNPLUGGING:
23  *
24  * The sequencing for updating the bitmap reliably is a little
25  * subtle (and I got it wrong the first time) so it deserves some
26  * explanation.
27  *
28  * We group bitmap updates into batches.  Each batch has a number.
29  * We may write out several batches at once, but that isn't very important.
30  * conf->bm_write is the number of the last batch successfully written.
31  * conf->bm_flush is the number of the last batch that was closed to
32  *    new additions.
33  * When we discover that we will need to write to any block in a stripe
34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35  * the number of the batch it will be in. This is bm_flush+1.
36  * When we are ready to do a write, if that batch hasn't been written yet,
37  *   we plug the array and queue the stripe for later.
38  * When an unplug happens, we increment bm_flush, thus closing the current
39  *   batch.
40  * When we notice that bm_flush > bm_write, we write out all pending updates
41  * to the bitmap, and advance bm_write to where bm_flush was.
42  * This may occasionally write a bit out twice, but is sure never to
43  * miss any bits.
44  */
45
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/seq_file.h>
51 #include <linux/cpu.h>
52 #include "md.h"
53 #include "raid5.h"
54 #include "bitmap.h"
55
56 /*
57  * Stripe cache
58  */
59
60 #define NR_STRIPES              256
61 #define STRIPE_SIZE             PAGE_SIZE
62 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
63 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
64 #define IO_THRESHOLD            1
65 #define BYPASS_THRESHOLD        1
66 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
67 #define HASH_MASK               (NR_HASH - 1)
68
69 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
70
71 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
72  * order without overlap.  There may be several bio's per stripe+device, and
73  * a bio could span several devices.
74  * When walking this list for a particular stripe+device, we must never proceed
75  * beyond a bio that extends past this device, as the next bio might no longer
76  * be valid.
77  * This macro is used to determine the 'next' bio in the list, given the sector
78  * of the current stripe+device
79  */
80 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
81 /*
82  * The following can be used to debug the driver
83  */
84 #define RAID5_PARANOIA  1
85 #if RAID5_PARANOIA && defined(CONFIG_SMP)
86 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
87 #else
88 # define CHECK_DEVLOCK()
89 #endif
90
91 #ifdef DEBUG
92 #define inline
93 #define __inline__
94 #endif
95
96 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
97
98 /*
99  * We maintain a biased count of active stripes in the bottom 16 bits of
100  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
101  */
102 static inline int raid5_bi_phys_segments(struct bio *bio)
103 {
104         return bio->bi_phys_segments & 0xffff;
105 }
106
107 static inline int raid5_bi_hw_segments(struct bio *bio)
108 {
109         return (bio->bi_phys_segments >> 16) & 0xffff;
110 }
111
112 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
113 {
114         --bio->bi_phys_segments;
115         return raid5_bi_phys_segments(bio);
116 }
117
118 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
119 {
120         unsigned short val = raid5_bi_hw_segments(bio);
121
122         --val;
123         bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
124         return val;
125 }
126
127 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
128 {
129         bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16);
130 }
131
132 /* Find first data disk in a raid6 stripe */
133 static inline int raid6_d0(struct stripe_head *sh)
134 {
135         if (sh->ddf_layout)
136                 /* ddf always start from first device */
137                 return 0;
138         /* md starts just after Q block */
139         if (sh->qd_idx == sh->disks - 1)
140                 return 0;
141         else
142                 return sh->qd_idx + 1;
143 }
144 static inline int raid6_next_disk(int disk, int raid_disks)
145 {
146         disk++;
147         return (disk < raid_disks) ? disk : 0;
148 }
149
150 /* When walking through the disks in a raid5, starting at raid6_d0,
151  * We need to map each disk to a 'slot', where the data disks are slot
152  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
153  * is raid_disks-1.  This help does that mapping.
154  */
155 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
156                              int *count, int syndrome_disks)
157 {
158         int slot;
159
160         if (idx == sh->pd_idx)
161                 return syndrome_disks;
162         if (idx == sh->qd_idx)
163                 return syndrome_disks + 1;
164         slot = (*count)++;
165         return slot;
166 }
167
168 static void return_io(struct bio *return_bi)
169 {
170         struct bio *bi = return_bi;
171         while (bi) {
172
173                 return_bi = bi->bi_next;
174                 bi->bi_next = NULL;
175                 bi->bi_size = 0;
176                 bio_endio(bi, 0);
177                 bi = return_bi;
178         }
179 }
180
181 static void print_raid5_conf (raid5_conf_t *conf);
182
183 static int stripe_operations_active(struct stripe_head *sh)
184 {
185         return sh->check_state || sh->reconstruct_state ||
186                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
187                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
188 }
189
190 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
191 {
192         if (atomic_dec_and_test(&sh->count)) {
193                 BUG_ON(!list_empty(&sh->lru));
194                 BUG_ON(atomic_read(&conf->active_stripes)==0);
195                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
196                         if (test_bit(STRIPE_DELAYED, &sh->state)) {
197                                 list_add_tail(&sh->lru, &conf->delayed_list);
198                                 blk_plug_device(conf->mddev->queue);
199                         } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
200                                    sh->bm_seq - conf->seq_write > 0) {
201                                 list_add_tail(&sh->lru, &conf->bitmap_list);
202                                 blk_plug_device(conf->mddev->queue);
203                         } else {
204                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
205                                 list_add_tail(&sh->lru, &conf->handle_list);
206                         }
207                         md_wakeup_thread(conf->mddev->thread);
208                 } else {
209                         BUG_ON(stripe_operations_active(sh));
210                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
211                                 atomic_dec(&conf->preread_active_stripes);
212                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
213                                         md_wakeup_thread(conf->mddev->thread);
214                         }
215                         atomic_dec(&conf->active_stripes);
216                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
217                                 list_add_tail(&sh->lru, &conf->inactive_list);
218                                 wake_up(&conf->wait_for_stripe);
219                                 if (conf->retry_read_aligned)
220                                         md_wakeup_thread(conf->mddev->thread);
221                         }
222                 }
223         }
224 }
225
226 static void release_stripe(struct stripe_head *sh)
227 {
228         raid5_conf_t *conf = sh->raid_conf;
229         unsigned long flags;
230
231         spin_lock_irqsave(&conf->device_lock, flags);
232         __release_stripe(conf, sh);
233         spin_unlock_irqrestore(&conf->device_lock, flags);
234 }
235
236 static inline void remove_hash(struct stripe_head *sh)
237 {
238         pr_debug("remove_hash(), stripe %llu\n",
239                 (unsigned long long)sh->sector);
240
241         hlist_del_init(&sh->hash);
242 }
243
244 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
245 {
246         struct hlist_head *hp = stripe_hash(conf, sh->sector);
247
248         pr_debug("insert_hash(), stripe %llu\n",
249                 (unsigned long long)sh->sector);
250
251         CHECK_DEVLOCK();
252         hlist_add_head(&sh->hash, hp);
253 }
254
255
256 /* find an idle stripe, make sure it is unhashed, and return it. */
257 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
258 {
259         struct stripe_head *sh = NULL;
260         struct list_head *first;
261
262         CHECK_DEVLOCK();
263         if (list_empty(&conf->inactive_list))
264                 goto out;
265         first = conf->inactive_list.next;
266         sh = list_entry(first, struct stripe_head, lru);
267         list_del_init(first);
268         remove_hash(sh);
269         atomic_inc(&conf->active_stripes);
270 out:
271         return sh;
272 }
273
274 static void shrink_buffers(struct stripe_head *sh, int num)
275 {
276         struct page *p;
277         int i;
278
279         for (i=0; i<num ; i++) {
280                 p = sh->dev[i].page;
281                 if (!p)
282                         continue;
283                 sh->dev[i].page = NULL;
284                 put_page(p);
285         }
286 }
287
288 static int grow_buffers(struct stripe_head *sh, int num)
289 {
290         int i;
291
292         for (i=0; i<num; i++) {
293                 struct page *page;
294
295                 if (!(page = alloc_page(GFP_KERNEL))) {
296                         return 1;
297                 }
298                 sh->dev[i].page = page;
299         }
300         return 0;
301 }
302
303 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
304 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
305                             struct stripe_head *sh);
306
307 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
308 {
309         raid5_conf_t *conf = sh->raid_conf;
310         int i;
311
312         BUG_ON(atomic_read(&sh->count) != 0);
313         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
314         BUG_ON(stripe_operations_active(sh));
315
316         CHECK_DEVLOCK();
317         pr_debug("init_stripe called, stripe %llu\n",
318                 (unsigned long long)sh->sector);
319
320         remove_hash(sh);
321
322         sh->generation = conf->generation - previous;
323         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
324         sh->sector = sector;
325         stripe_set_idx(sector, conf, previous, sh);
326         sh->state = 0;
327
328
329         for (i = sh->disks; i--; ) {
330                 struct r5dev *dev = &sh->dev[i];
331
332                 if (dev->toread || dev->read || dev->towrite || dev->written ||
333                     test_bit(R5_LOCKED, &dev->flags)) {
334                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
335                                (unsigned long long)sh->sector, i, dev->toread,
336                                dev->read, dev->towrite, dev->written,
337                                test_bit(R5_LOCKED, &dev->flags));
338                         BUG();
339                 }
340                 dev->flags = 0;
341                 raid5_build_block(sh, i, previous);
342         }
343         insert_hash(conf, sh);
344 }
345
346 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
347                                          short generation)
348 {
349         struct stripe_head *sh;
350         struct hlist_node *hn;
351
352         CHECK_DEVLOCK();
353         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
354         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
355                 if (sh->sector == sector && sh->generation == generation)
356                         return sh;
357         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
358         return NULL;
359 }
360
361 static void unplug_slaves(mddev_t *mddev);
362 static void raid5_unplug_device(struct request_queue *q);
363
364 static struct stripe_head *
365 get_active_stripe(raid5_conf_t *conf, sector_t sector,
366                   int previous, int noblock)
367 {
368         struct stripe_head *sh;
369
370         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
371
372         spin_lock_irq(&conf->device_lock);
373
374         do {
375                 wait_event_lock_irq(conf->wait_for_stripe,
376                                     conf->quiesce == 0,
377                                     conf->device_lock, /* nothing */);
378                 sh = __find_stripe(conf, sector, conf->generation - previous);
379                 if (!sh) {
380                         if (!conf->inactive_blocked)
381                                 sh = get_free_stripe(conf);
382                         if (noblock && sh == NULL)
383                                 break;
384                         if (!sh) {
385                                 conf->inactive_blocked = 1;
386                                 wait_event_lock_irq(conf->wait_for_stripe,
387                                                     !list_empty(&conf->inactive_list) &&
388                                                     (atomic_read(&conf->active_stripes)
389                                                      < (conf->max_nr_stripes *3/4)
390                                                      || !conf->inactive_blocked),
391                                                     conf->device_lock,
392                                                     raid5_unplug_device(conf->mddev->queue)
393                                         );
394                                 conf->inactive_blocked = 0;
395                         } else
396                                 init_stripe(sh, sector, previous);
397                 } else {
398                         if (atomic_read(&sh->count)) {
399                                 BUG_ON(!list_empty(&sh->lru)
400                                     && !test_bit(STRIPE_EXPANDING, &sh->state));
401                         } else {
402                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
403                                         atomic_inc(&conf->active_stripes);
404                                 if (list_empty(&sh->lru) &&
405                                     !test_bit(STRIPE_EXPANDING, &sh->state))
406                                         BUG();
407                                 list_del_init(&sh->lru);
408                         }
409                 }
410         } while (sh == NULL);
411
412         if (sh)
413                 atomic_inc(&sh->count);
414
415         spin_unlock_irq(&conf->device_lock);
416         return sh;
417 }
418
419 static void
420 raid5_end_read_request(struct bio *bi, int error);
421 static void
422 raid5_end_write_request(struct bio *bi, int error);
423
424 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
425 {
426         raid5_conf_t *conf = sh->raid_conf;
427         int i, disks = sh->disks;
428
429         might_sleep();
430
431         for (i = disks; i--; ) {
432                 int rw;
433                 struct bio *bi;
434                 mdk_rdev_t *rdev;
435                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
436                         rw = WRITE;
437                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
438                         rw = READ;
439                 else
440                         continue;
441
442                 bi = &sh->dev[i].req;
443
444                 bi->bi_rw = rw;
445                 if (rw == WRITE)
446                         bi->bi_end_io = raid5_end_write_request;
447                 else
448                         bi->bi_end_io = raid5_end_read_request;
449
450                 rcu_read_lock();
451                 rdev = rcu_dereference(conf->disks[i].rdev);
452                 if (rdev && test_bit(Faulty, &rdev->flags))
453                         rdev = NULL;
454                 if (rdev)
455                         atomic_inc(&rdev->nr_pending);
456                 rcu_read_unlock();
457
458                 if (rdev) {
459                         if (s->syncing || s->expanding || s->expanded)
460                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
461
462                         set_bit(STRIPE_IO_STARTED, &sh->state);
463
464                         bi->bi_bdev = rdev->bdev;
465                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
466                                 __func__, (unsigned long long)sh->sector,
467                                 bi->bi_rw, i);
468                         atomic_inc(&sh->count);
469                         bi->bi_sector = sh->sector + rdev->data_offset;
470                         bi->bi_flags = 1 << BIO_UPTODATE;
471                         bi->bi_vcnt = 1;
472                         bi->bi_max_vecs = 1;
473                         bi->bi_idx = 0;
474                         bi->bi_io_vec = &sh->dev[i].vec;
475                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
476                         bi->bi_io_vec[0].bv_offset = 0;
477                         bi->bi_size = STRIPE_SIZE;
478                         bi->bi_next = NULL;
479                         if (rw == WRITE &&
480                             test_bit(R5_ReWrite, &sh->dev[i].flags))
481                                 atomic_add(STRIPE_SECTORS,
482                                         &rdev->corrected_errors);
483                         generic_make_request(bi);
484                 } else {
485                         if (rw == WRITE)
486                                 set_bit(STRIPE_DEGRADED, &sh->state);
487                         pr_debug("skip op %ld on disc %d for sector %llu\n",
488                                 bi->bi_rw, i, (unsigned long long)sh->sector);
489                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
490                         set_bit(STRIPE_HANDLE, &sh->state);
491                 }
492         }
493 }
494
495 static struct dma_async_tx_descriptor *
496 async_copy_data(int frombio, struct bio *bio, struct page *page,
497         sector_t sector, struct dma_async_tx_descriptor *tx)
498 {
499         struct bio_vec *bvl;
500         struct page *bio_page;
501         int i;
502         int page_offset;
503         struct async_submit_ctl submit;
504
505         if (bio->bi_sector >= sector)
506                 page_offset = (signed)(bio->bi_sector - sector) * 512;
507         else
508                 page_offset = (signed)(sector - bio->bi_sector) * -512;
509
510         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
511         bio_for_each_segment(bvl, bio, i) {
512                 int len = bio_iovec_idx(bio, i)->bv_len;
513                 int clen;
514                 int b_offset = 0;
515
516                 if (page_offset < 0) {
517                         b_offset = -page_offset;
518                         page_offset += b_offset;
519                         len -= b_offset;
520                 }
521
522                 if (len > 0 && page_offset + len > STRIPE_SIZE)
523                         clen = STRIPE_SIZE - page_offset;
524                 else
525                         clen = len;
526
527                 if (clen > 0) {
528                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
529                         bio_page = bio_iovec_idx(bio, i)->bv_page;
530                         if (frombio)
531                                 tx = async_memcpy(page, bio_page, page_offset,
532                                                   b_offset, clen, &submit);
533                         else
534                                 tx = async_memcpy(bio_page, page, b_offset,
535                                                   page_offset, clen, &submit);
536                 }
537                 /* chain the operations */
538                 submit.depend_tx = tx;
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         struct async_submit_ctl submit;
599         int i;
600
601         pr_debug("%s: stripe %llu\n", __func__,
602                 (unsigned long long)sh->sector);
603
604         for (i = sh->disks; i--; ) {
605                 struct r5dev *dev = &sh->dev[i];
606                 if (test_bit(R5_Wantfill, &dev->flags)) {
607                         struct bio *rbi;
608                         spin_lock_irq(&conf->device_lock);
609                         dev->read = rbi = dev->toread;
610                         dev->toread = NULL;
611                         spin_unlock_irq(&conf->device_lock);
612                         while (rbi && rbi->bi_sector <
613                                 dev->sector + STRIPE_SECTORS) {
614                                 tx = async_copy_data(0, rbi, dev->page,
615                                         dev->sector, tx);
616                                 rbi = r5_next_bio(rbi, dev->sector);
617                         }
618                 }
619         }
620
621         atomic_inc(&sh->count);
622         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
623         async_trigger_callback(&submit);
624 }
625
626 static void mark_target_uptodate(struct stripe_head *sh, int target)
627 {
628         struct r5dev *tgt;
629
630         if (target < 0)
631                 return;
632
633         tgt = &sh->dev[target];
634         set_bit(R5_UPTODATE, &tgt->flags);
635         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
636         clear_bit(R5_Wantcompute, &tgt->flags);
637 }
638
639 static void ops_complete_compute(void *stripe_head_ref)
640 {
641         struct stripe_head *sh = stripe_head_ref;
642
643         pr_debug("%s: stripe %llu\n", __func__,
644                 (unsigned long long)sh->sector);
645
646         /* mark the computed target(s) as uptodate */
647         mark_target_uptodate(sh, sh->ops.target);
648         mark_target_uptodate(sh, sh->ops.target2);
649
650         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
651         if (sh->check_state == check_state_compute_run)
652                 sh->check_state = check_state_compute_result;
653         set_bit(STRIPE_HANDLE, &sh->state);
654         release_stripe(sh);
655 }
656
657 /* return a pointer to the address conversion region of the scribble buffer */
658 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
659                                  struct raid5_percpu *percpu)
660 {
661         return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
662 }
663
664 static struct dma_async_tx_descriptor *
665 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
666 {
667         int disks = sh->disks;
668         struct page **xor_srcs = percpu->scribble;
669         int target = sh->ops.target;
670         struct r5dev *tgt = &sh->dev[target];
671         struct page *xor_dest = tgt->page;
672         int count = 0;
673         struct dma_async_tx_descriptor *tx;
674         struct async_submit_ctl submit;
675         int i;
676
677         pr_debug("%s: stripe %llu block: %d\n",
678                 __func__, (unsigned long long)sh->sector, target);
679         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
680
681         for (i = disks; i--; )
682                 if (i != target)
683                         xor_srcs[count++] = sh->dev[i].page;
684
685         atomic_inc(&sh->count);
686
687         init_async_submit(&submit, ASYNC_TX_XOR_ZERO_DST, NULL,
688                           ops_complete_compute, sh, to_addr_conv(sh, percpu));
689         if (unlikely(count == 1))
690                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
691         else
692                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
693
694         return tx;
695 }
696
697 /* set_syndrome_sources - populate source buffers for gen_syndrome
698  * @srcs - (struct page *) array of size sh->disks
699  * @sh - stripe_head to parse
700  *
701  * Populates srcs in proper layout order for the stripe and returns the
702  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
703  * destination buffer is recorded in srcs[count] and the Q destination
704  * is recorded in srcs[count+1]].
705  */
706 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
707 {
708         int disks = sh->disks;
709         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
710         int d0_idx = raid6_d0(sh);
711         int count;
712         int i;
713
714         for (i = 0; i < disks; i++)
715                 srcs[i] = (void *)raid6_empty_zero_page;
716
717         count = 0;
718         i = d0_idx;
719         do {
720                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
721
722                 srcs[slot] = sh->dev[i].page;
723                 i = raid6_next_disk(i, disks);
724         } while (i != d0_idx);
725         BUG_ON(count != syndrome_disks);
726
727         return count;
728 }
729
730 static struct dma_async_tx_descriptor *
731 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
732 {
733         int disks = sh->disks;
734         struct page **blocks = percpu->scribble;
735         int target;
736         int qd_idx = sh->qd_idx;
737         struct dma_async_tx_descriptor *tx;
738         struct async_submit_ctl submit;
739         struct r5dev *tgt;
740         struct page *dest;
741         int i;
742         int count;
743
744         if (sh->ops.target < 0)
745                 target = sh->ops.target2;
746         else if (sh->ops.target2 < 0)
747                 target = sh->ops.target;
748         else
749                 /* we should only have one valid target */
750                 BUG();
751         BUG_ON(target < 0);
752         pr_debug("%s: stripe %llu block: %d\n",
753                 __func__, (unsigned long long)sh->sector, target);
754
755         tgt = &sh->dev[target];
756         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
757         dest = tgt->page;
758
759         atomic_inc(&sh->count);
760
761         if (target == qd_idx) {
762                 count = set_syndrome_sources(blocks, sh);
763                 blocks[count] = NULL; /* regenerating p is not necessary */
764                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
765                 init_async_submit(&submit, 0, NULL, ops_complete_compute, sh,
766                                   to_addr_conv(sh, percpu));
767                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
768         } else {
769                 /* Compute any data- or p-drive using XOR */
770                 count = 0;
771                 for (i = disks; i-- ; ) {
772                         if (i == target || i == qd_idx)
773                                 continue;
774                         blocks[count++] = sh->dev[i].page;
775                 }
776
777                 init_async_submit(&submit, ASYNC_TX_XOR_ZERO_DST, NULL,
778                                   ops_complete_compute, sh,
779                                   to_addr_conv(sh, percpu));
780                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
781         }
782
783         return tx;
784 }
785
786 static struct dma_async_tx_descriptor *
787 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
788 {
789         int i, count, disks = sh->disks;
790         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
791         int d0_idx = raid6_d0(sh);
792         int faila = -1, failb = -1;
793         int target = sh->ops.target;
794         int target2 = sh->ops.target2;
795         struct r5dev *tgt = &sh->dev[target];
796         struct r5dev *tgt2 = &sh->dev[target2];
797         struct dma_async_tx_descriptor *tx;
798         struct page **blocks = percpu->scribble;
799         struct async_submit_ctl submit;
800
801         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
802                  __func__, (unsigned long long)sh->sector, target, target2);
803         BUG_ON(target < 0 || target2 < 0);
804         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
805         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
806
807         /* we need to open-code set_syndrome_sources to handle to the
808          * slot number conversion for 'faila' and 'failb'
809          */
810         for (i = 0; i < disks ; i++)
811                 blocks[i] = (void *)raid6_empty_zero_page;
812         count = 0;
813         i = d0_idx;
814         do {
815                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
816
817                 blocks[slot] = sh->dev[i].page;
818
819                 if (i == target)
820                         faila = slot;
821                 if (i == target2)
822                         failb = slot;
823                 i = raid6_next_disk(i, disks);
824         } while (i != d0_idx);
825         BUG_ON(count != syndrome_disks);
826
827         BUG_ON(faila == failb);
828         if (failb < faila)
829                 swap(faila, failb);
830         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
831                  __func__, (unsigned long long)sh->sector, faila, failb);
832
833         atomic_inc(&sh->count);
834
835         if (failb == syndrome_disks+1) {
836                 /* Q disk is one of the missing disks */
837                 if (faila == syndrome_disks) {
838                         /* Missing P+Q, just recompute */
839                         init_async_submit(&submit, 0, NULL, ops_complete_compute,
840                                           sh, to_addr_conv(sh, percpu));
841                         return async_gen_syndrome(blocks, 0, count+2,
842                                                   STRIPE_SIZE, &submit);
843                 } else {
844                         struct page *dest;
845                         int data_target;
846                         int qd_idx = sh->qd_idx;
847
848                         /* Missing D+Q: recompute D from P, then recompute Q */
849                         if (target == qd_idx)
850                                 data_target = target2;
851                         else
852                                 data_target = target;
853
854                         count = 0;
855                         for (i = disks; i-- ; ) {
856                                 if (i == data_target || i == qd_idx)
857                                         continue;
858                                 blocks[count++] = sh->dev[i].page;
859                         }
860                         dest = sh->dev[data_target].page;
861                         init_async_submit(&submit, ASYNC_TX_XOR_ZERO_DST, NULL,
862                                           NULL, NULL, to_addr_conv(sh, percpu));
863                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
864                                        &submit);
865
866                         count = set_syndrome_sources(blocks, sh);
867                         init_async_submit(&submit, 0, tx, ops_complete_compute,
868                                           sh, to_addr_conv(sh, percpu));
869                         return async_gen_syndrome(blocks, 0, count+2,
870                                                   STRIPE_SIZE, &submit);
871                 }
872         }
873
874         init_async_submit(&submit, 0, NULL, ops_complete_compute, sh,
875                           to_addr_conv(sh, percpu));
876         if (failb == syndrome_disks) {
877                 /* We're missing D+P. */
878                 return async_raid6_datap_recov(syndrome_disks+2, STRIPE_SIZE,
879                                                faila, blocks, &submit);
880         } else {
881                 /* We're missing D+D. */
882                 return async_raid6_2data_recov(syndrome_disks+2, STRIPE_SIZE,
883                                                faila, failb, blocks, &submit);
884         }
885 }
886
887
888 static void ops_complete_prexor(void *stripe_head_ref)
889 {
890         struct stripe_head *sh = stripe_head_ref;
891
892         pr_debug("%s: stripe %llu\n", __func__,
893                 (unsigned long long)sh->sector);
894 }
895
896 static struct dma_async_tx_descriptor *
897 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
898                struct dma_async_tx_descriptor *tx)
899 {
900         int disks = sh->disks;
901         struct page **xor_srcs = percpu->scribble;
902         int count = 0, pd_idx = sh->pd_idx, i;
903         struct async_submit_ctl submit;
904
905         /* existing parity data subtracted */
906         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
907
908         pr_debug("%s: stripe %llu\n", __func__,
909                 (unsigned long long)sh->sector);
910
911         for (i = disks; i--; ) {
912                 struct r5dev *dev = &sh->dev[i];
913                 /* Only process blocks that are known to be uptodate */
914                 if (test_bit(R5_Wantdrain, &dev->flags))
915                         xor_srcs[count++] = dev->page;
916         }
917
918         init_async_submit(&submit, ASYNC_TX_XOR_DROP_DST, tx,
919                           ops_complete_prexor, sh, to_addr_conv(sh, percpu));
920         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
921
922         return tx;
923 }
924
925 static struct dma_async_tx_descriptor *
926 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
927 {
928         int disks = sh->disks;
929         int i;
930
931         pr_debug("%s: stripe %llu\n", __func__,
932                 (unsigned long long)sh->sector);
933
934         for (i = disks; i--; ) {
935                 struct r5dev *dev = &sh->dev[i];
936                 struct bio *chosen;
937
938                 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
939                         struct bio *wbi;
940
941                         spin_lock(&sh->lock);
942                         chosen = dev->towrite;
943                         dev->towrite = NULL;
944                         BUG_ON(dev->written);
945                         wbi = dev->written = chosen;
946                         spin_unlock(&sh->lock);
947
948                         while (wbi && wbi->bi_sector <
949                                 dev->sector + STRIPE_SECTORS) {
950                                 tx = async_copy_data(1, wbi, dev->page,
951                                         dev->sector, tx);
952                                 wbi = r5_next_bio(wbi, dev->sector);
953                         }
954                 }
955         }
956
957         return tx;
958 }
959
960 static void ops_complete_reconstruct(void *stripe_head_ref)
961 {
962         struct stripe_head *sh = stripe_head_ref;
963         int disks = sh->disks;
964         int pd_idx = sh->pd_idx;
965         int qd_idx = sh->qd_idx;
966         int i;
967
968         pr_debug("%s: stripe %llu\n", __func__,
969                 (unsigned long long)sh->sector);
970
971         for (i = disks; i--; ) {
972                 struct r5dev *dev = &sh->dev[i];
973
974                 if (dev->written || i == pd_idx || i == qd_idx)
975                         set_bit(R5_UPTODATE, &dev->flags);
976         }
977
978         if (sh->reconstruct_state == reconstruct_state_drain_run)
979                 sh->reconstruct_state = reconstruct_state_drain_result;
980         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
981                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
982         else {
983                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
984                 sh->reconstruct_state = reconstruct_state_result;
985         }
986
987         set_bit(STRIPE_HANDLE, &sh->state);
988         release_stripe(sh);
989 }
990
991 static void
992 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
993                      struct dma_async_tx_descriptor *tx)
994 {
995         int disks = sh->disks;
996         struct page **xor_srcs = percpu->scribble;
997         struct async_submit_ctl submit;
998         int count = 0, pd_idx = sh->pd_idx, i;
999         struct page *xor_dest;
1000         int prexor = 0;
1001         unsigned long flags;
1002
1003         pr_debug("%s: stripe %llu\n", __func__,
1004                 (unsigned long long)sh->sector);
1005
1006         /* check if prexor is active which means only process blocks
1007          * that are part of a read-modify-write (written)
1008          */
1009         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1010                 prexor = 1;
1011                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1012                 for (i = disks; i--; ) {
1013                         struct r5dev *dev = &sh->dev[i];
1014                         if (dev->written)
1015                                 xor_srcs[count++] = dev->page;
1016                 }
1017         } else {
1018                 xor_dest = sh->dev[pd_idx].page;
1019                 for (i = disks; i--; ) {
1020                         struct r5dev *dev = &sh->dev[i];
1021                         if (i != pd_idx)
1022                                 xor_srcs[count++] = dev->page;
1023                 }
1024         }
1025
1026         /* 1/ if we prexor'd then the dest is reused as a source
1027          * 2/ if we did not prexor then we are redoing the parity
1028          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1029          * for the synchronous xor case
1030          */
1031         flags = ASYNC_TX_ACK |
1032                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1033
1034         atomic_inc(&sh->count);
1035
1036         init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1037                           to_addr_conv(sh, percpu));
1038         if (unlikely(count == 1))
1039                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1040         else
1041                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1042 }
1043
1044 static void
1045 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1046                      struct dma_async_tx_descriptor *tx)
1047 {
1048         struct async_submit_ctl submit;
1049         struct page **blocks = percpu->scribble;
1050         int count;
1051
1052         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1053
1054         count = set_syndrome_sources(blocks, sh);
1055
1056         atomic_inc(&sh->count);
1057
1058         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1059                           sh, to_addr_conv(sh, percpu));
1060         async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1061 }
1062
1063 static void ops_complete_check(void *stripe_head_ref)
1064 {
1065         struct stripe_head *sh = stripe_head_ref;
1066
1067         pr_debug("%s: stripe %llu\n", __func__,
1068                 (unsigned long long)sh->sector);
1069
1070         sh->check_state = check_state_check_result;
1071         set_bit(STRIPE_HANDLE, &sh->state);
1072         release_stripe(sh);
1073 }
1074
1075 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1076 {
1077         int disks = sh->disks;
1078         int pd_idx = sh->pd_idx;
1079         int qd_idx = sh->qd_idx;
1080         struct page *xor_dest;
1081         struct page **xor_srcs = percpu->scribble;
1082         struct dma_async_tx_descriptor *tx;
1083         struct async_submit_ctl submit;
1084         int count;
1085         int i;
1086
1087         pr_debug("%s: stripe %llu\n", __func__,
1088                 (unsigned long long)sh->sector);
1089
1090         count = 0;
1091         xor_dest = sh->dev[pd_idx].page;
1092         xor_srcs[count++] = xor_dest;
1093         for (i = disks; i--; ) {
1094                 if (i == pd_idx || i == qd_idx)
1095                         continue;
1096                 xor_srcs[count++] = sh->dev[i].page;
1097         }
1098
1099         init_async_submit(&submit, 0, NULL, NULL, NULL,
1100                           to_addr_conv(sh, percpu));
1101         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1102                            &sh->ops.zero_sum_result, &submit);
1103
1104         atomic_inc(&sh->count);
1105         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1106         tx = async_trigger_callback(&submit);
1107 }
1108
1109 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1110 {
1111         struct page **srcs = percpu->scribble;
1112         struct async_submit_ctl submit;
1113         int count;
1114
1115         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1116                 (unsigned long long)sh->sector, checkp);
1117
1118         count = set_syndrome_sources(srcs, sh);
1119         if (!checkp)
1120                 srcs[count] = NULL;
1121
1122         atomic_inc(&sh->count);
1123         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1124                           sh, to_addr_conv(sh, percpu));
1125         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1126                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1127 }
1128
1129 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1130 {
1131         int overlap_clear = 0, i, disks = sh->disks;
1132         struct dma_async_tx_descriptor *tx = NULL;
1133         raid5_conf_t *conf = sh->raid_conf;
1134         int level = conf->level;
1135         struct raid5_percpu *percpu;
1136         unsigned long cpu;
1137
1138         cpu = get_cpu();
1139         percpu = per_cpu_ptr(conf->percpu, cpu);
1140         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1141                 ops_run_biofill(sh);
1142                 overlap_clear++;
1143         }
1144
1145         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1146                 if (level < 6)
1147                         tx = ops_run_compute5(sh, percpu);
1148                 else {
1149                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
1150                                 tx = ops_run_compute6_1(sh, percpu);
1151                         else
1152                                 tx = ops_run_compute6_2(sh, percpu);
1153                 }
1154                 /* terminate the chain if reconstruct is not set to be run */
1155                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1156                         async_tx_ack(tx);
1157         }
1158
1159         if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1160                 tx = ops_run_prexor(sh, percpu, tx);
1161
1162         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1163                 tx = ops_run_biodrain(sh, tx);
1164                 overlap_clear++;
1165         }
1166
1167         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1168                 if (level < 6)
1169                         ops_run_reconstruct5(sh, percpu, tx);
1170                 else
1171                         ops_run_reconstruct6(sh, percpu, tx);
1172         }
1173
1174         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1175                 if (sh->check_state == check_state_run)
1176                         ops_run_check_p(sh, percpu);
1177                 else if (sh->check_state == check_state_run_q)
1178                         ops_run_check_pq(sh, percpu, 0);
1179                 else if (sh->check_state == check_state_run_pq)
1180                         ops_run_check_pq(sh, percpu, 1);
1181                 else
1182                         BUG();
1183         }
1184
1185         if (overlap_clear)
1186                 for (i = disks; i--; ) {
1187                         struct r5dev *dev = &sh->dev[i];
1188                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1189                                 wake_up(&sh->raid_conf->wait_for_overlap);
1190                 }
1191         put_cpu();
1192 }
1193
1194 static int grow_one_stripe(raid5_conf_t *conf)
1195 {
1196         struct stripe_head *sh;
1197         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
1198         if (!sh)
1199                 return 0;
1200         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
1201         sh->raid_conf = conf;
1202         spin_lock_init(&sh->lock);
1203
1204         if (grow_buffers(sh, conf->raid_disks)) {
1205                 shrink_buffers(sh, conf->raid_disks);
1206                 kmem_cache_free(conf->slab_cache, sh);
1207                 return 0;
1208         }
1209         sh->disks = conf->raid_disks;
1210         /* we just created an active stripe so... */
1211         atomic_set(&sh->count, 1);
1212         atomic_inc(&conf->active_stripes);
1213         INIT_LIST_HEAD(&sh->lru);
1214         release_stripe(sh);
1215         return 1;
1216 }
1217
1218 static int grow_stripes(raid5_conf_t *conf, int num)
1219 {
1220         struct kmem_cache *sc;
1221         int devs = conf->raid_disks;
1222
1223         sprintf(conf->cache_name[0],
1224                 "raid%d-%s", conf->level, mdname(conf->mddev));
1225         sprintf(conf->cache_name[1],
1226                 "raid%d-%s-alt", conf->level, mdname(conf->mddev));
1227         conf->active_name = 0;
1228         sc = kmem_cache_create(conf->cache_name[conf->active_name],
1229                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1230                                0, 0, NULL);
1231         if (!sc)
1232                 return 1;
1233         conf->slab_cache = sc;
1234         conf->pool_size = devs;
1235         while (num--)
1236                 if (!grow_one_stripe(conf))
1237                         return 1;
1238         return 0;
1239 }
1240
1241 /**
1242  * scribble_len - return the required size of the scribble region
1243  * @num - total number of disks in the array
1244  *
1245  * The size must be enough to contain:
1246  * 1/ a struct page pointer for each device in the array +2
1247  * 2/ room to convert each entry in (1) to its corresponding dma
1248  *    (dma_map_page()) or page (page_address()) address.
1249  *
1250  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1251  * calculate over all devices (not just the data blocks), using zeros in place
1252  * of the P and Q blocks.
1253  */
1254 static size_t scribble_len(int num)
1255 {
1256         size_t len;
1257
1258         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1259
1260         return len;
1261 }
1262
1263 static int resize_stripes(raid5_conf_t *conf, int newsize)
1264 {
1265         /* Make all the stripes able to hold 'newsize' devices.
1266          * New slots in each stripe get 'page' set to a new page.
1267          *
1268          * This happens in stages:
1269          * 1/ create a new kmem_cache and allocate the required number of
1270          *    stripe_heads.
1271          * 2/ gather all the old stripe_heads and tranfer the pages across
1272          *    to the new stripe_heads.  This will have the side effect of
1273          *    freezing the array as once all stripe_heads have been collected,
1274          *    no IO will be possible.  Old stripe heads are freed once their
1275          *    pages have been transferred over, and the old kmem_cache is
1276          *    freed when all stripes are done.
1277          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
1278          *    we simple return a failre status - no need to clean anything up.
1279          * 4/ allocate new pages for the new slots in the new stripe_heads.
1280          *    If this fails, we don't bother trying the shrink the
1281          *    stripe_heads down again, we just leave them as they are.
1282          *    As each stripe_head is processed the new one is released into
1283          *    active service.
1284          *
1285          * Once step2 is started, we cannot afford to wait for a write,
1286          * so we use GFP_NOIO allocations.
1287          */
1288         struct stripe_head *osh, *nsh;
1289         LIST_HEAD(newstripes);
1290         struct disk_info *ndisks;
1291         unsigned long cpu;
1292         int err;
1293         struct kmem_cache *sc;
1294         int i;
1295
1296         if (newsize <= conf->pool_size)
1297                 return 0; /* never bother to shrink */
1298
1299         err = md_allow_write(conf->mddev);
1300         if (err)
1301                 return err;
1302
1303         /* Step 1 */
1304         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1305                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1306                                0, 0, NULL);
1307         if (!sc)
1308                 return -ENOMEM;
1309
1310         for (i = conf->max_nr_stripes; i; i--) {
1311                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1312                 if (!nsh)
1313                         break;
1314
1315                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1316
1317                 nsh->raid_conf = conf;
1318                 spin_lock_init(&nsh->lock);
1319
1320                 list_add(&nsh->lru, &newstripes);
1321         }
1322         if (i) {
1323                 /* didn't get enough, give up */
1324                 while (!list_empty(&newstripes)) {
1325                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1326                         list_del(&nsh->lru);
1327                         kmem_cache_free(sc, nsh);
1328                 }
1329                 kmem_cache_destroy(sc);
1330                 return -ENOMEM;
1331         }
1332         /* Step 2 - Must use GFP_NOIO now.
1333          * OK, we have enough stripes, start collecting inactive
1334          * stripes and copying them over
1335          */
1336         list_for_each_entry(nsh, &newstripes, lru) {
1337                 spin_lock_irq(&conf->device_lock);
1338                 wait_event_lock_irq(conf->wait_for_stripe,
1339                                     !list_empty(&conf->inactive_list),
1340                                     conf->device_lock,
1341                                     unplug_slaves(conf->mddev)
1342                         );
1343                 osh = get_free_stripe(conf);
1344                 spin_unlock_irq(&conf->device_lock);
1345                 atomic_set(&nsh->count, 1);
1346                 for(i=0; i<conf->pool_size; i++)
1347                         nsh->dev[i].page = osh->dev[i].page;
1348                 for( ; i<newsize; i++)
1349                         nsh->dev[i].page = NULL;
1350                 kmem_cache_free(conf->slab_cache, osh);
1351         }
1352         kmem_cache_destroy(conf->slab_cache);
1353
1354         /* Step 3.
1355          * At this point, we are holding all the stripes so the array
1356          * is completely stalled, so now is a good time to resize
1357          * conf->disks and the scribble region
1358          */
1359         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1360         if (ndisks) {
1361                 for (i=0; i<conf->raid_disks; i++)
1362                         ndisks[i] = conf->disks[i];
1363                 kfree(conf->disks);
1364                 conf->disks = ndisks;
1365         } else
1366                 err = -ENOMEM;
1367
1368         get_online_cpus();
1369         conf->scribble_len = scribble_len(newsize);
1370         for_each_present_cpu(cpu) {
1371                 struct raid5_percpu *percpu;
1372                 void *scribble;
1373
1374                 percpu = per_cpu_ptr(conf->percpu, cpu);
1375                 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1376
1377                 if (scribble) {
1378                         kfree(percpu->scribble);
1379                         percpu->scribble = scribble;
1380                 } else {
1381                         err = -ENOMEM;
1382                         break;
1383                 }
1384         }
1385         put_online_cpus();
1386
1387         /* Step 4, return new stripes to service */
1388         while(!list_empty(&newstripes)) {
1389                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1390                 list_del_init(&nsh->lru);
1391
1392                 for (i=conf->raid_disks; i < newsize; i++)
1393                         if (nsh->dev[i].page == NULL) {
1394                                 struct page *p = alloc_page(GFP_NOIO);
1395                                 nsh->dev[i].page = p;
1396                                 if (!p)
1397                                         err = -ENOMEM;
1398                         }
1399                 release_stripe(nsh);
1400         }
1401         /* critical section pass, GFP_NOIO no longer needed */
1402
1403         conf->slab_cache = sc;
1404         conf->active_name = 1-conf->active_name;
1405         conf->pool_size = newsize;
1406         return err;
1407 }
1408
1409 static int drop_one_stripe(raid5_conf_t *conf)
1410 {
1411         struct stripe_head *sh;
1412
1413         spin_lock_irq(&conf->device_lock);
1414         sh = get_free_stripe(conf);
1415         spin_unlock_irq(&conf->device_lock);
1416         if (!sh)
1417                 return 0;
1418         BUG_ON(atomic_read(&sh->count));
1419         shrink_buffers(sh, conf->pool_size);
1420         kmem_cache_free(conf->slab_cache, sh);
1421         atomic_dec(&conf->active_stripes);
1422         return 1;
1423 }
1424
1425 static void shrink_stripes(raid5_conf_t *conf)
1426 {
1427         while (drop_one_stripe(conf))
1428                 ;
1429
1430         if (conf->slab_cache)
1431                 kmem_cache_destroy(conf->slab_cache);
1432         conf->slab_cache = NULL;
1433 }
1434
1435 static void raid5_end_read_request(struct bio * bi, int error)
1436 {
1437         struct stripe_head *sh = bi->bi_private;
1438         raid5_conf_t *conf = sh->raid_conf;
1439         int disks = sh->disks, i;
1440         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1441         char b[BDEVNAME_SIZE];
1442         mdk_rdev_t *rdev;
1443
1444
1445         for (i=0 ; i<disks; i++)
1446                 if (bi == &sh->dev[i].req)
1447                         break;
1448
1449         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1450                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1451                 uptodate);
1452         if (i == disks) {
1453                 BUG();
1454                 return;
1455         }
1456
1457         if (uptodate) {
1458                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1459                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1460                         rdev = conf->disks[i].rdev;
1461                         printk_rl(KERN_INFO "raid5:%s: read error corrected"
1462                                   " (%lu sectors at %llu on %s)\n",
1463                                   mdname(conf->mddev), STRIPE_SECTORS,
1464                                   (unsigned long long)(sh->sector
1465                                                        + rdev->data_offset),
1466                                   bdevname(rdev->bdev, b));
1467                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1468                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1469                 }
1470                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1471                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1472         } else {
1473                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1474                 int retry = 0;
1475                 rdev = conf->disks[i].rdev;
1476
1477                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1478                 atomic_inc(&rdev->read_errors);
1479                 if (conf->mddev->degraded)
1480                         printk_rl(KERN_WARNING
1481                                   "raid5:%s: read error not correctable "
1482                                   "(sector %llu on %s).\n",
1483                                   mdname(conf->mddev),
1484                                   (unsigned long long)(sh->sector
1485                                                        + rdev->data_offset),
1486                                   bdn);
1487                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1488                         /* Oh, no!!! */
1489                         printk_rl(KERN_WARNING
1490                                   "raid5:%s: read error NOT corrected!! "
1491                                   "(sector %llu on %s).\n",
1492                                   mdname(conf->mddev),
1493                                   (unsigned long long)(sh->sector
1494                                                        + rdev->data_offset),
1495                                   bdn);
1496                 else if (atomic_read(&rdev->read_errors)
1497                          > conf->max_nr_stripes)
1498                         printk(KERN_WARNING
1499                                "raid5:%s: Too many read errors, failing device %s.\n",
1500                                mdname(conf->mddev), bdn);
1501                 else
1502                         retry = 1;
1503                 if (retry)
1504                         set_bit(R5_ReadError, &sh->dev[i].flags);
1505                 else {
1506                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1507                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1508                         md_error(conf->mddev, rdev);
1509                 }
1510         }
1511         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1512         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1513         set_bit(STRIPE_HANDLE, &sh->state);
1514         release_stripe(sh);
1515 }
1516
1517 static void raid5_end_write_request(struct bio *bi, int error)
1518 {
1519         struct stripe_head *sh = bi->bi_private;
1520         raid5_conf_t *conf = sh->raid_conf;
1521         int disks = sh->disks, i;
1522         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1523
1524         for (i=0 ; i<disks; i++)
1525                 if (bi == &sh->dev[i].req)
1526                         break;
1527
1528         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1529                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1530                 uptodate);
1531         if (i == disks) {
1532                 BUG();
1533                 return;
1534         }
1535
1536         if (!uptodate)
1537                 md_error(conf->mddev, conf->disks[i].rdev);
1538
1539         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1540         
1541         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1542         set_bit(STRIPE_HANDLE, &sh->state);
1543         release_stripe(sh);
1544 }
1545
1546
1547 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1548         
1549 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1550 {
1551         struct r5dev *dev = &sh->dev[i];
1552
1553         bio_init(&dev->req);
1554         dev->req.bi_io_vec = &dev->vec;
1555         dev->req.bi_vcnt++;
1556         dev->req.bi_max_vecs++;
1557         dev->vec.bv_page = dev->page;
1558         dev->vec.bv_len = STRIPE_SIZE;
1559         dev->vec.bv_offset = 0;
1560
1561         dev->req.bi_sector = sh->sector;
1562         dev->req.bi_private = sh;
1563
1564         dev->flags = 0;
1565         dev->sector = compute_blocknr(sh, i, previous);
1566 }
1567
1568 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1569 {
1570         char b[BDEVNAME_SIZE];
1571         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1572         pr_debug("raid5: error called\n");
1573
1574         if (!test_bit(Faulty, &rdev->flags)) {
1575                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1576                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1577                         unsigned long flags;
1578                         spin_lock_irqsave(&conf->device_lock, flags);
1579                         mddev->degraded++;
1580                         spin_unlock_irqrestore(&conf->device_lock, flags);
1581                         /*
1582                          * if recovery was running, make sure it aborts.
1583                          */
1584                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1585                 }
1586                 set_bit(Faulty, &rdev->flags);
1587                 printk(KERN_ALERT
1588                        "raid5: Disk failure on %s, disabling device.\n"
1589                        "raid5: Operation continuing on %d devices.\n",
1590                        bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1591         }
1592 }
1593
1594 /*
1595  * Input: a 'big' sector number,
1596  * Output: index of the data and parity disk, and the sector # in them.
1597  */
1598 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1599                                      int previous, int *dd_idx,
1600                                      struct stripe_head *sh)
1601 {
1602         long stripe;
1603         unsigned long chunk_number;
1604         unsigned int chunk_offset;
1605         int pd_idx, qd_idx;
1606         int ddf_layout = 0;
1607         sector_t new_sector;
1608         int algorithm = previous ? conf->prev_algo
1609                                  : conf->algorithm;
1610         int sectors_per_chunk = previous ? (conf->prev_chunk >> 9)
1611                                          : (conf->chunk_size >> 9);
1612         int raid_disks = previous ? conf->previous_raid_disks
1613                                   : conf->raid_disks;
1614         int data_disks = raid_disks - conf->max_degraded;
1615
1616         /* First compute the information on this sector */
1617
1618         /*
1619          * Compute the chunk number and the sector offset inside the chunk
1620          */
1621         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1622         chunk_number = r_sector;
1623         BUG_ON(r_sector != chunk_number);
1624
1625         /*
1626          * Compute the stripe number
1627          */
1628         stripe = chunk_number / data_disks;
1629
1630         /*
1631          * Compute the data disk and parity disk indexes inside the stripe
1632          */
1633         *dd_idx = chunk_number % data_disks;
1634
1635         /*
1636          * Select the parity disk based on the user selected algorithm.
1637          */
1638         pd_idx = qd_idx = ~0;
1639         switch(conf->level) {
1640         case 4:
1641                 pd_idx = data_disks;
1642                 break;
1643         case 5:
1644                 switch (algorithm) {
1645                 case ALGORITHM_LEFT_ASYMMETRIC:
1646                         pd_idx = data_disks - stripe % raid_disks;
1647                         if (*dd_idx >= pd_idx)
1648                                 (*dd_idx)++;
1649                         break;
1650                 case ALGORITHM_RIGHT_ASYMMETRIC:
1651                         pd_idx = stripe % raid_disks;
1652                         if (*dd_idx >= pd_idx)
1653                                 (*dd_idx)++;
1654                         break;
1655                 case ALGORITHM_LEFT_SYMMETRIC:
1656                         pd_idx = data_disks - stripe % raid_disks;
1657                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1658                         break;
1659                 case ALGORITHM_RIGHT_SYMMETRIC:
1660                         pd_idx = stripe % raid_disks;
1661                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1662                         break;
1663                 case ALGORITHM_PARITY_0:
1664                         pd_idx = 0;
1665                         (*dd_idx)++;
1666                         break;
1667                 case ALGORITHM_PARITY_N:
1668                         pd_idx = data_disks;
1669                         break;
1670                 default:
1671                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1672                                 algorithm);
1673                         BUG();
1674                 }
1675                 break;
1676         case 6:
1677
1678                 switch (algorithm) {
1679                 case ALGORITHM_LEFT_ASYMMETRIC:
1680                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1681                         qd_idx = pd_idx + 1;
1682                         if (pd_idx == raid_disks-1) {
1683                                 (*dd_idx)++;    /* Q D D D P */
1684                                 qd_idx = 0;
1685                         } else if (*dd_idx >= pd_idx)
1686                                 (*dd_idx) += 2; /* D D P Q D */
1687                         break;
1688                 case ALGORITHM_RIGHT_ASYMMETRIC:
1689                         pd_idx = stripe % raid_disks;
1690                         qd_idx = pd_idx + 1;
1691                         if (pd_idx == raid_disks-1) {
1692                                 (*dd_idx)++;    /* Q D D D P */
1693                                 qd_idx = 0;
1694                         } else if (*dd_idx >= pd_idx)
1695                                 (*dd_idx) += 2; /* D D P Q D */
1696                         break;
1697                 case ALGORITHM_LEFT_SYMMETRIC:
1698                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1699                         qd_idx = (pd_idx + 1) % raid_disks;
1700                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1701                         break;
1702                 case ALGORITHM_RIGHT_SYMMETRIC:
1703                         pd_idx = stripe % raid_disks;
1704                         qd_idx = (pd_idx + 1) % raid_disks;
1705                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1706                         break;
1707
1708                 case ALGORITHM_PARITY_0:
1709                         pd_idx = 0;
1710                         qd_idx = 1;
1711                         (*dd_idx) += 2;
1712                         break;
1713                 case ALGORITHM_PARITY_N:
1714                         pd_idx = data_disks;
1715                         qd_idx = data_disks + 1;
1716                         break;
1717
1718                 case ALGORITHM_ROTATING_ZERO_RESTART:
1719                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
1720                          * of blocks for computing Q is different.
1721                          */
1722                         pd_idx = stripe % raid_disks;
1723                         qd_idx = pd_idx + 1;
1724                         if (pd_idx == raid_disks-1) {
1725                                 (*dd_idx)++;    /* Q D D D P */
1726                                 qd_idx = 0;
1727                         } else if (*dd_idx >= pd_idx)
1728                                 (*dd_idx) += 2; /* D D P Q D */
1729                         ddf_layout = 1;
1730                         break;
1731
1732                 case ALGORITHM_ROTATING_N_RESTART:
1733                         /* Same a left_asymmetric, by first stripe is
1734                          * D D D P Q  rather than
1735                          * Q D D D P
1736                          */
1737                         pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1738                         qd_idx = pd_idx + 1;
1739                         if (pd_idx == raid_disks-1) {
1740                                 (*dd_idx)++;    /* Q D D D P */
1741                                 qd_idx = 0;
1742                         } else if (*dd_idx >= pd_idx)
1743                                 (*dd_idx) += 2; /* D D P Q D */
1744                         ddf_layout = 1;
1745                         break;
1746
1747                 case ALGORITHM_ROTATING_N_CONTINUE:
1748                         /* Same as left_symmetric but Q is before P */
1749                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1750                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1751                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1752                         ddf_layout = 1;
1753                         break;
1754
1755                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1756                         /* RAID5 left_asymmetric, with Q on last device */
1757                         pd_idx = data_disks - stripe % (raid_disks-1);
1758                         if (*dd_idx >= pd_idx)
1759                                 (*dd_idx)++;
1760                         qd_idx = raid_disks - 1;
1761                         break;
1762
1763                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1764                         pd_idx = stripe % (raid_disks-1);
1765                         if (*dd_idx >= pd_idx)
1766                                 (*dd_idx)++;
1767                         qd_idx = raid_disks - 1;
1768                         break;
1769
1770                 case ALGORITHM_LEFT_SYMMETRIC_6:
1771                         pd_idx = data_disks - stripe % (raid_disks-1);
1772                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1773                         qd_idx = raid_disks - 1;
1774                         break;
1775
1776                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1777                         pd_idx = stripe % (raid_disks-1);
1778                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1779                         qd_idx = raid_disks - 1;
1780                         break;
1781
1782                 case ALGORITHM_PARITY_0_6:
1783                         pd_idx = 0;
1784                         (*dd_idx)++;
1785                         qd_idx = raid_disks - 1;
1786                         break;
1787
1788
1789                 default:
1790                         printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1791                                algorithm);
1792                         BUG();
1793                 }
1794                 break;
1795         }
1796
1797         if (sh) {
1798                 sh->pd_idx = pd_idx;
1799                 sh->qd_idx = qd_idx;
1800                 sh->ddf_layout = ddf_layout;
1801         }
1802         /*
1803          * Finally, compute the new sector number
1804          */
1805         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1806         return new_sector;
1807 }
1808
1809
1810 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1811 {
1812         raid5_conf_t *conf = sh->raid_conf;
1813         int raid_disks = sh->disks;
1814         int data_disks = raid_disks - conf->max_degraded;
1815         sector_t new_sector = sh->sector, check;
1816         int sectors_per_chunk = previous ? (conf->prev_chunk >> 9)
1817                                          : (conf->chunk_size >> 9);
1818         int algorithm = previous ? conf->prev_algo
1819                                  : conf->algorithm;
1820         sector_t stripe;
1821         int chunk_offset;
1822         int chunk_number, dummy1, dd_idx = i;
1823         sector_t r_sector;
1824         struct stripe_head sh2;
1825
1826
1827         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1828         stripe = new_sector;
1829         BUG_ON(new_sector != stripe);
1830
1831         if (i == sh->pd_idx)
1832                 return 0;
1833         switch(conf->level) {
1834         case 4: break;
1835         case 5:
1836                 switch (algorithm) {
1837                 case ALGORITHM_LEFT_ASYMMETRIC:
1838                 case ALGORITHM_RIGHT_ASYMMETRIC:
1839                         if (i > sh->pd_idx)
1840                                 i--;
1841                         break;
1842                 case ALGORITHM_LEFT_SYMMETRIC:
1843                 case ALGORITHM_RIGHT_SYMMETRIC:
1844                         if (i < sh->pd_idx)
1845                                 i += raid_disks;
1846                         i -= (sh->pd_idx + 1);
1847                         break;
1848                 case ALGORITHM_PARITY_0:
1849                         i -= 1;
1850                         break;
1851                 case ALGORITHM_PARITY_N:
1852                         break;
1853                 default:
1854                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1855                                algorithm);
1856                         BUG();
1857                 }
1858                 break;
1859         case 6:
1860                 if (i == sh->qd_idx)
1861                         return 0; /* It is the Q disk */
1862                 switch (algorithm) {
1863                 case ALGORITHM_LEFT_ASYMMETRIC:
1864                 case ALGORITHM_RIGHT_ASYMMETRIC:
1865                 case ALGORITHM_ROTATING_ZERO_RESTART:
1866                 case ALGORITHM_ROTATING_N_RESTART:
1867                         if (sh->pd_idx == raid_disks-1)
1868                                 i--;    /* Q D D D P */
1869                         else if (i > sh->pd_idx)
1870                                 i -= 2; /* D D P Q D */
1871                         break;
1872                 case ALGORITHM_LEFT_SYMMETRIC:
1873                 case ALGORITHM_RIGHT_SYMMETRIC:
1874                         if (sh->pd_idx == raid_disks-1)
1875                                 i--; /* Q D D D P */
1876                         else {
1877                                 /* D D P Q D */
1878                                 if (i < sh->pd_idx)
1879                                         i += raid_disks;
1880                                 i -= (sh->pd_idx + 2);
1881                         }
1882                         break;
1883                 case ALGORITHM_PARITY_0:
1884                         i -= 2;
1885                         break;
1886                 case ALGORITHM_PARITY_N:
1887                         break;
1888                 case ALGORITHM_ROTATING_N_CONTINUE:
1889                         if (sh->pd_idx == 0)
1890                                 i--;    /* P D D D Q */
1891                         else if (i > sh->pd_idx)
1892                                 i -= 2; /* D D Q P D */
1893                         break;
1894                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1895                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1896                         if (i > sh->pd_idx)
1897                                 i--;
1898                         break;
1899                 case ALGORITHM_LEFT_SYMMETRIC_6:
1900                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1901                         if (i < sh->pd_idx)
1902                                 i += data_disks + 1;
1903                         i -= (sh->pd_idx + 1);
1904                         break;
1905                 case ALGORITHM_PARITY_0_6:
1906                         i -= 1;
1907                         break;
1908                 default:
1909                         printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1910                                algorithm);
1911                         BUG();
1912                 }
1913                 break;
1914         }
1915
1916         chunk_number = stripe * data_disks + i;
1917         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1918
1919         check = raid5_compute_sector(conf, r_sector,
1920                                      previous, &dummy1, &sh2);
1921         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1922                 || sh2.qd_idx != sh->qd_idx) {
1923                 printk(KERN_ERR "compute_blocknr: map not correct\n");
1924                 return 0;
1925         }
1926         return r_sector;
1927 }
1928
1929
1930 static void
1931 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
1932                          int rcw, int expand)
1933 {
1934         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1935         raid5_conf_t *conf = sh->raid_conf;
1936         int level = conf->level;
1937
1938         if (rcw) {
1939                 /* if we are not expanding this is a proper write request, and
1940                  * there will be bios with new data to be drained into the
1941                  * stripe cache
1942                  */
1943                 if (!expand) {
1944                         sh->reconstruct_state = reconstruct_state_drain_run;
1945                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1946                 } else
1947                         sh->reconstruct_state = reconstruct_state_run;
1948
1949                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
1950
1951                 for (i = disks; i--; ) {
1952                         struct r5dev *dev = &sh->dev[i];
1953
1954                         if (dev->towrite) {
1955                                 set_bit(R5_LOCKED, &dev->flags);
1956                                 set_bit(R5_Wantdrain, &dev->flags);
1957                                 if (!expand)
1958                                         clear_bit(R5_UPTODATE, &dev->flags);
1959                                 s->locked++;
1960                         }
1961                 }
1962                 if (s->locked + conf->max_degraded == disks)
1963                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1964                                 atomic_inc(&conf->pending_full_writes);
1965         } else {
1966                 BUG_ON(level == 6);
1967                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1968                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1969
1970                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
1971                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
1972                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1973                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
1974
1975                 for (i = disks; i--; ) {
1976                         struct r5dev *dev = &sh->dev[i];
1977                         if (i == pd_idx)
1978                                 continue;
1979
1980                         if (dev->towrite &&
1981                             (test_bit(R5_UPTODATE, &dev->flags) ||
1982                              test_bit(R5_Wantcompute, &dev->flags))) {
1983                                 set_bit(R5_Wantdrain, &dev->flags);
1984                                 set_bit(R5_LOCKED, &dev->flags);
1985                                 clear_bit(R5_UPTODATE, &dev->flags);
1986                                 s->locked++;
1987                         }
1988                 }
1989         }
1990
1991         /* keep the parity disk(s) locked while asynchronous operations
1992          * are in flight
1993          */
1994         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1995         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1996         s->locked++;
1997
1998         if (level == 6) {
1999                 int qd_idx = sh->qd_idx;
2000                 struct r5dev *dev = &sh->dev[qd_idx];
2001
2002                 set_bit(R5_LOCKED, &dev->flags);
2003                 clear_bit(R5_UPTODATE, &dev->flags);
2004                 s->locked++;
2005         }
2006
2007         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2008                 __func__, (unsigned long long)sh->sector,
2009                 s->locked, s->ops_request);
2010 }
2011
2012 /*
2013  * Each stripe/dev can have one or more bion attached.
2014  * toread/towrite point to the first in a chain.
2015  * The bi_next chain must be in order.
2016  */
2017 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2018 {
2019         struct bio **bip;
2020         raid5_conf_t *conf = sh->raid_conf;
2021         int firstwrite=0;
2022
2023         pr_debug("adding bh b#%llu to stripe s#%llu\n",
2024                 (unsigned long long)bi->bi_sector,
2025                 (unsigned long long)sh->sector);
2026
2027
2028         spin_lock(&sh->lock);
2029         spin_lock_irq(&conf->device_lock);
2030         if (forwrite) {
2031                 bip = &sh->dev[dd_idx].towrite;
2032                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2033                         firstwrite = 1;
2034         } else
2035                 bip = &sh->dev[dd_idx].toread;
2036         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2037                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2038                         goto overlap;
2039                 bip = & (*bip)->bi_next;
2040         }
2041         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2042                 goto overlap;
2043
2044         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2045         if (*bip)
2046                 bi->bi_next = *bip;
2047         *bip = bi;
2048         bi->bi_phys_segments++;
2049         spin_unlock_irq(&conf->device_lock);
2050         spin_unlock(&sh->lock);
2051
2052         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2053                 (unsigned long long)bi->bi_sector,
2054                 (unsigned long long)sh->sector, dd_idx);
2055
2056         if (conf->mddev->bitmap && firstwrite) {
2057                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2058                                   STRIPE_SECTORS, 0);
2059                 sh->bm_seq = conf->seq_flush+1;
2060                 set_bit(STRIPE_BIT_DELAY, &sh->state);
2061         }
2062
2063         if (forwrite) {
2064                 /* check if page is covered */
2065                 sector_t sector = sh->dev[dd_idx].sector;
2066                 for (bi=sh->dev[dd_idx].towrite;
2067                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2068                              bi && bi->bi_sector <= sector;
2069                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2070                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2071                                 sector = bi->bi_sector + (bi->bi_size>>9);
2072                 }
2073                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2074                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2075         }
2076         return 1;
2077
2078  overlap:
2079         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2080         spin_unlock_irq(&conf->device_lock);
2081         spin_unlock(&sh->lock);
2082         return 0;
2083 }
2084
2085 static void end_reshape(raid5_conf_t *conf);
2086
2087 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2088                             struct stripe_head *sh)
2089 {
2090         int sectors_per_chunk =
2091                 previous ? (conf->prev_chunk >> 9)
2092                          : (conf->chunk_size >> 9);
2093         int dd_idx;
2094         int chunk_offset = sector_div(stripe, sectors_per_chunk);
2095         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2096
2097         raid5_compute_sector(conf,
2098                              stripe * (disks - conf->max_degraded)
2099                              *sectors_per_chunk + chunk_offset,
2100                              previous,
2101                              &dd_idx, sh);
2102 }
2103
2104 static void
2105 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2106                                 struct stripe_head_state *s, int disks,
2107                                 struct bio **return_bi)
2108 {
2109         int i;
2110         for (i = disks; i--; ) {
2111                 struct bio *bi;
2112                 int bitmap_end = 0;
2113
2114                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2115                         mdk_rdev_t *rdev;
2116                         rcu_read_lock();
2117                         rdev = rcu_dereference(conf->disks[i].rdev);
2118                         if (rdev && test_bit(In_sync, &rdev->flags))
2119                                 /* multiple read failures in one stripe */
2120                                 md_error(conf->mddev, rdev);
2121                         rcu_read_unlock();
2122                 }
2123                 spin_lock_irq(&conf->device_lock);
2124                 /* fail all writes first */
2125                 bi = sh->dev[i].towrite;
2126                 sh->dev[i].towrite = NULL;
2127                 if (bi) {
2128                         s->to_write--;
2129                         bitmap_end = 1;
2130                 }
2131
2132                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2133                         wake_up(&conf->wait_for_overlap);
2134
2135                 while (bi && bi->bi_sector <
2136                         sh->dev[i].sector + STRIPE_SECTORS) {
2137                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2138                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2139                         if (!raid5_dec_bi_phys_segments(bi)) {
2140                                 md_write_end(conf->mddev);
2141                                 bi->bi_next = *return_bi;
2142                                 *return_bi = bi;
2143                         }
2144                         bi = nextbi;
2145                 }
2146                 /* and fail all 'written' */
2147                 bi = sh->dev[i].written;
2148                 sh->dev[i].written = NULL;
2149                 if (bi) bitmap_end = 1;
2150                 while (bi && bi->bi_sector <
2151                        sh->dev[i].sector + STRIPE_SECTORS) {
2152                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2153                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2154                         if (!raid5_dec_bi_phys_segments(bi)) {
2155                                 md_write_end(conf->mddev);
2156                                 bi->bi_next = *return_bi;
2157                                 *return_bi = bi;
2158                         }
2159                         bi = bi2;
2160                 }
2161
2162                 /* fail any reads if this device is non-operational and
2163                  * the data has not reached the cache yet.
2164                  */
2165                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2166                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2167                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
2168                         bi = sh->dev[i].toread;
2169                         sh->dev[i].toread = NULL;
2170                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2171                                 wake_up(&conf->wait_for_overlap);
2172                         if (bi) s->to_read--;
2173                         while (bi && bi->bi_sector <
2174                                sh->dev[i].sector + STRIPE_SECTORS) {
2175                                 struct bio *nextbi =
2176                                         r5_next_bio(bi, sh->dev[i].sector);
2177                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2178                                 if (!raid5_dec_bi_phys_segments(bi)) {
2179                                         bi->bi_next = *return_bi;
2180                                         *return_bi = bi;
2181                                 }
2182                                 bi = nextbi;
2183                         }
2184                 }
2185                 spin_unlock_irq(&conf->device_lock);
2186                 if (bitmap_end)
2187                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2188                                         STRIPE_SECTORS, 0, 0);
2189         }
2190
2191         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2192                 if (atomic_dec_and_test(&conf->pending_full_writes))
2193                         md_wakeup_thread(conf->mddev->thread);
2194 }
2195
2196 /* fetch_block5 - checks the given member device to see if its data needs
2197  * to be read or computed to satisfy a request.
2198  *
2199  * Returns 1 when no more member devices need to be checked, otherwise returns
2200  * 0 to tell the loop in handle_stripe_fill5 to continue
2201  */
2202 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2203                         int disk_idx, int disks)
2204 {
2205         struct r5dev *dev = &sh->dev[disk_idx];
2206         struct r5dev *failed_dev = &sh->dev[s->failed_num];
2207
2208         /* is the data in this block needed, and can we get it? */
2209         if (!test_bit(R5_LOCKED, &dev->flags) &&
2210             !test_bit(R5_UPTODATE, &dev->flags) &&
2211             (dev->toread ||
2212              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2213              s->syncing || s->expanding ||
2214              (s->failed &&
2215               (failed_dev->toread ||
2216                (failed_dev->towrite &&
2217                 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2218                 /* We would like to get this block, possibly by computing it,
2219                  * otherwise read it if the backing disk is insync
2220                  */
2221                 if ((s->uptodate == disks - 1) &&
2222                     (s->failed && disk_idx == s->failed_num)) {
2223                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2224                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2225                         set_bit(R5_Wantcompute, &dev->flags);
2226                         sh->ops.target = disk_idx;
2227                         sh->ops.target2 = -1;
2228                         s->req_compute = 1;
2229                         /* Careful: from this point on 'uptodate' is in the eye
2230                          * of raid_run_ops which services 'compute' operations
2231                          * before writes. R5_Wantcompute flags a block that will
2232                          * be R5_UPTODATE by the time it is needed for a
2233                          * subsequent operation.
2234                          */
2235                         s->uptodate++;
2236                         return 1; /* uptodate + compute == disks */
2237                 } else if (test_bit(R5_Insync, &dev->flags)) {
2238                         set_bit(R5_LOCKED, &dev->flags);
2239                         set_bit(R5_Wantread, &dev->flags);
2240                         s->locked++;
2241                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2242                                 s->syncing);
2243                 }
2244         }
2245
2246         return 0;
2247 }
2248
2249 /**
2250  * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2251  */
2252 static void handle_stripe_fill5(struct stripe_head *sh,
2253                         struct stripe_head_state *s, int disks)
2254 {
2255         int i;
2256
2257         /* look for blocks to read/compute, skip this if a compute
2258          * is already in flight, or if the stripe contents are in the
2259          * midst of changing due to a write
2260          */
2261         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2262             !sh->reconstruct_state)
2263                 for (i = disks; i--; )
2264                         if (fetch_block5(sh, s, i, disks))
2265                                 break;
2266         set_bit(STRIPE_HANDLE, &sh->state);
2267 }
2268
2269 /* fetch_block6 - checks the given member device to see if its data needs
2270  * to be read or computed to satisfy a request.
2271  *
2272  * Returns 1 when no more member devices need to be checked, otherwise returns
2273  * 0 to tell the loop in handle_stripe_fill6 to continue
2274  */
2275 static int fetch_block6(struct stripe_head *sh, struct stripe_head_state *s,
2276                          struct r6_state *r6s, int disk_idx, int disks)
2277 {
2278         struct r5dev *dev = &sh->dev[disk_idx];
2279         struct r5dev *fdev[2] = { &sh->dev[r6s->failed_num[0]],
2280                                   &sh->dev[r6s->failed_num[1]] };
2281
2282         if (!test_bit(R5_LOCKED, &dev->flags) &&
2283             !test_bit(R5_UPTODATE, &dev->flags) &&
2284             (dev->toread ||
2285              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2286              s->syncing || s->expanding ||
2287              (s->failed >= 1 &&
2288               (fdev[0]->toread || s->to_write)) ||
2289              (s->failed >= 2 &&
2290               (fdev[1]->toread || s->to_write)))) {
2291                 /* we would like to get this block, possibly by computing it,
2292                  * otherwise read it if the backing disk is insync
2293                  */
2294                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2295                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2296                 if ((s->uptodate == disks - 1) &&
2297                     (s->failed && (disk_idx == r6s->failed_num[0] ||
2298                                    disk_idx == r6s->failed_num[1]))) {
2299                         /* have disk failed, and we're requested to fetch it;
2300                          * do compute it
2301                          */
2302                         pr_debug("Computing stripe %llu block %d\n",
2303                                (unsigned long long)sh->sector, disk_idx);
2304                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2305                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2306                         set_bit(R5_Wantcompute, &dev->flags);
2307                         sh->ops.target = disk_idx;
2308                         sh->ops.target2 = -1; /* no 2nd target */
2309                         s->req_compute = 1;
2310                         s->uptodate++;
2311                         return 1;
2312                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2313                         /* Computing 2-failure is *very* expensive; only
2314                          * do it if failed >= 2
2315                          */
2316                         int other;
2317                         for (other = disks; other--; ) {
2318                                 if (other == disk_idx)
2319                                         continue;
2320                                 if (!test_bit(R5_UPTODATE,
2321                                       &sh->dev[other].flags))
2322                                         break;
2323                         }
2324                         BUG_ON(other < 0);
2325                         pr_debug("Computing stripe %llu blocks %d,%d\n",
2326                                (unsigned long long)sh->sector,
2327                                disk_idx, other);
2328                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2329                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2330                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2331                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
2332                         sh->ops.target = disk_idx;
2333                         sh->ops.target2 = other;
2334                         s->uptodate += 2;
2335                         s->req_compute = 1;
2336                         return 1;
2337                 } else if (test_bit(R5_Insync, &dev->flags)) {
2338                         set_bit(R5_LOCKED, &dev->flags);
2339                         set_bit(R5_Wantread, &dev->flags);
2340                         s->locked++;
2341                         pr_debug("Reading block %d (sync=%d)\n",
2342                                 disk_idx, s->syncing);
2343                 }
2344         }
2345
2346         return 0;
2347 }
2348
2349 /**
2350  * handle_stripe_fill6 - read or compute data to satisfy pending requests.
2351  */
2352 static void handle_stripe_fill6(struct stripe_head *sh,
2353                         struct stripe_head_state *s, struct r6_state *r6s,
2354                         int disks)
2355 {
2356         int i;
2357
2358         /* look for blocks to read/compute, skip this if a compute
2359          * is already in flight, or if the stripe contents are in the
2360          * midst of changing due to a write
2361          */
2362         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2363             !sh->reconstruct_state)
2364                 for (i = disks; i--; )
2365                         if (fetch_block6(sh, s, r6s, i, disks))
2366                                 break;
2367         set_bit(STRIPE_HANDLE, &sh->state);
2368 }
2369
2370
2371 /* handle_stripe_clean_event
2372  * any written block on an uptodate or failed drive can be returned.
2373  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2374  * never LOCKED, so we don't need to test 'failed' directly.
2375  */
2376 static void handle_stripe_clean_event(raid5_conf_t *conf,
2377         struct stripe_head *sh, int disks, struct bio **return_bi)
2378 {
2379         int i;
2380         struct r5dev *dev;
2381
2382         for (i = disks; i--; )
2383                 if (sh->dev[i].written) {
2384                         dev = &sh->dev[i];
2385                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2386                                 test_bit(R5_UPTODATE, &dev->flags)) {
2387                                 /* We can return any write requests */
2388                                 struct bio *wbi, *wbi2;
2389                                 int bitmap_end = 0;
2390                                 pr_debug("Return write for disc %d\n", i);
2391                                 spin_lock_irq(&conf->device_lock);
2392                                 wbi = dev->written;
2393                                 dev->written = NULL;
2394                                 while (wbi && wbi->bi_sector <
2395                                         dev->sector + STRIPE_SECTORS) {
2396                                         wbi2 = r5_next_bio(wbi, dev->sector);
2397                                         if (!raid5_dec_bi_phys_segments(wbi)) {
2398                                                 md_write_end(conf->mddev);
2399                                                 wbi->bi_next = *return_bi;
2400                                                 *return_bi = wbi;
2401                                         }
2402                                         wbi = wbi2;
2403                                 }
2404                                 if (dev->towrite == NULL)
2405                                         bitmap_end = 1;
2406                                 spin_unlock_irq(&conf->device_lock);
2407                                 if (bitmap_end)
2408                                         bitmap_endwrite(conf->mddev->bitmap,
2409                                                         sh->sector,
2410                                                         STRIPE_SECTORS,
2411                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2412                                                         0);
2413                         }
2414                 }
2415
2416         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2417                 if (atomic_dec_and_test(&conf->pending_full_writes))
2418                         md_wakeup_thread(conf->mddev->thread);
2419 }
2420
2421 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2422                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2423 {
2424         int rmw = 0, rcw = 0, i;
2425         for (i = disks; i--; ) {
2426                 /* would I have to read this buffer for read_modify_write */
2427                 struct r5dev *dev = &sh->dev[i];
2428                 if ((dev->towrite || i == sh->pd_idx) &&
2429                     !test_bit(R5_LOCKED, &dev->flags) &&
2430                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2431                       test_bit(R5_Wantcompute, &dev->flags))) {
2432                         if (test_bit(R5_Insync, &dev->flags))
2433                                 rmw++;
2434                         else
2435                                 rmw += 2*disks;  /* cannot read it */
2436                 }
2437                 /* Would I have to read this buffer for reconstruct_write */
2438                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2439                     !test_bit(R5_LOCKED, &dev->flags) &&
2440                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2441                     test_bit(R5_Wantcompute, &dev->flags))) {
2442                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2443                         else
2444                                 rcw += 2*disks;
2445                 }
2446         }
2447         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2448                 (unsigned long long)sh->sector, rmw, rcw);
2449         set_bit(STRIPE_HANDLE, &sh->state);
2450         if (rmw < rcw && rmw > 0)
2451                 /* prefer read-modify-write, but need to get some data */
2452                 for (i = disks; i--; ) {
2453                         struct r5dev *dev = &sh->dev[i];
2454                         if ((dev->towrite || i == sh->pd_idx) &&
2455                             !test_bit(R5_LOCKED, &dev->flags) &&
2456                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2457                             test_bit(R5_Wantcompute, &dev->flags)) &&
2458                             test_bit(R5_Insync, &dev->flags)) {
2459                                 if (
2460                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2461                                         pr_debug("Read_old block "
2462                                                 "%d for r-m-w\n", i);
2463                                         set_bit(R5_LOCKED, &dev->flags);
2464                                         set_bit(R5_Wantread, &dev->flags);
2465                                         s->locked++;
2466                                 } else {
2467                                         set_bit(STRIPE_DELAYED, &sh->state);
2468                                         set_bit(STRIPE_HANDLE, &sh->state);
2469                                 }
2470                         }
2471                 }
2472         if (rcw <= rmw && rcw > 0)
2473                 /* want reconstruct write, but need to get some data */
2474                 for (i = disks; i--; ) {
2475                         struct r5dev *dev = &sh->dev[i];
2476                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2477                             i != sh->pd_idx &&
2478                             !test_bit(R5_LOCKED, &dev->flags) &&
2479                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2480                             test_bit(R5_Wantcompute, &dev->flags)) &&
2481                             test_bit(R5_Insync, &dev->flags)) {
2482                                 if (
2483                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2484                                         pr_debug("Read_old block "
2485                                                 "%d for Reconstruct\n", i);
2486                                         set_bit(R5_LOCKED, &dev->flags);
2487                                         set_bit(R5_Wantread, &dev->flags);
2488                                         s->locked++;
2489                                 } else {
2490                                         set_bit(STRIPE_DELAYED, &sh->state);
2491                                         set_bit(STRIPE_HANDLE, &sh->state);
2492                                 }
2493                         }
2494                 }
2495         /* now if nothing is locked, and if we have enough data,
2496          * we can start a write request
2497          */
2498         /* since handle_stripe can be called at any time we need to handle the
2499          * case where a compute block operation has been submitted and then a
2500          * subsequent call wants to start a write request.  raid_run_ops only
2501          * handles the case where compute block and reconstruct are requested
2502          * simultaneously.  If this is not the case then new writes need to be
2503          * held off until the compute completes.
2504          */
2505         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2506             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2507             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2508                 schedule_reconstruction(sh, s, rcw == 0, 0);
2509 }
2510
2511 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2512                 struct stripe_head *sh, struct stripe_head_state *s,
2513                 struct r6_state *r6s, int disks)
2514 {
2515         int rcw = 0, pd_idx = sh->pd_idx, i;
2516         int qd_idx = sh->qd_idx;
2517
2518         set_bit(STRIPE_HANDLE, &sh->state);
2519         for (i = disks; i--; ) {
2520                 struct r5dev *dev = &sh->dev[i];
2521                 /* check if we haven't enough data */
2522                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2523                     i != pd_idx && i != qd_idx &&
2524                     !test_bit(R5_LOCKED, &dev->flags) &&
2525                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2526                       test_bit(R5_Wantcompute, &dev->flags))) {
2527                         rcw++;
2528                         if (!test_bit(R5_Insync, &dev->flags))
2529                                 continue; /* it's a failed drive */
2530
2531                         if (
2532                           test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2533                                 pr_debug("Read_old stripe %llu "
2534                                         "block %d for Reconstruct\n",
2535                                      (unsigned long long)sh->sector, i);
2536                                 set_bit(R5_LOCKED, &dev->flags);
2537                                 set_bit(R5_Wantread, &dev->flags);
2538                                 s->locked++;
2539                         } else {
2540                                 pr_debug("Request delayed stripe %llu "
2541                                         "block %d for Reconstruct\n",
2542                                      (unsigned long long)sh->sector, i);
2543                                 set_bit(STRIPE_DELAYED, &sh->state);
2544                                 set_bit(STRIPE_HANDLE, &sh->state);
2545                         }
2546                 }
2547         }
2548         /* now if nothing is locked, and if we have enough data, we can start a
2549          * write request
2550          */
2551         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2552             s->locked == 0 && rcw == 0 &&
2553             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2554                 schedule_reconstruction(sh, s, 1, 0);
2555         }
2556 }
2557
2558 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2559                                 struct stripe_head_state *s, int disks)
2560 {
2561         struct r5dev *dev = NULL;
2562
2563         set_bit(STRIPE_HANDLE, &sh->state);
2564
2565         switch (sh->check_state) {
2566         case check_state_idle:
2567                 /* start a new check operation if there are no failures */
2568                 if (s->failed == 0) {
2569                         BUG_ON(s->uptodate != disks);
2570                         sh->check_state = check_state_run;
2571                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2572                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2573                         s->uptodate--;
2574                         break;
2575                 }
2576                 dev = &sh->dev[s->failed_num];
2577                 /* fall through */
2578         case check_state_compute_result:
2579                 sh->check_state = check_state_idle;
2580                 if (!dev)
2581                         dev = &sh->dev[sh->pd_idx];
2582
2583                 /* check that a write has not made the stripe insync */
2584                 if (test_bit(STRIPE_INSYNC, &sh->state))
2585                         break;
2586
2587                 /* either failed parity check, or recovery is happening */
2588                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2589                 BUG_ON(s->uptodate != disks);
2590
2591                 set_bit(R5_LOCKED, &dev->flags);
2592                 s->locked++;
2593                 set_bit(R5_Wantwrite, &dev->flags);
2594
2595                 clear_bit(STRIPE_DEGRADED, &sh->state);
2596                 set_bit(STRIPE_INSYNC, &sh->state);
2597                 break;
2598         case check_state_run:
2599                 break; /* we will be called again upon completion */
2600         case check_state_check_result:
2601                 sh->check_state = check_state_idle;
2602
2603                 /* if a failure occurred during the check operation, leave
2604                  * STRIPE_INSYNC not set and let the stripe be handled again
2605                  */
2606                 if (s->failed)
2607                         break;
2608
2609                 /* handle a successful check operation, if parity is correct
2610                  * we are done.  Otherwise update the mismatch count and repair
2611                  * parity if !MD_RECOVERY_CHECK
2612                  */
2613                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2614                         /* parity is correct (on disc,
2615                          * not in buffer any more)
2616                          */
2617                         set_bit(STRIPE_INSYNC, &sh->state);
2618                 else {
2619                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2620                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2621                                 /* don't try to repair!! */
2622                                 set_bit(STRIPE_INSYNC, &sh->state);
2623                         else {
2624                                 sh->check_state = check_state_compute_run;
2625                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2626                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2627                                 set_bit(R5_Wantcompute,
2628                                         &sh->dev[sh->pd_idx].flags);
2629                                 sh->ops.target = sh->pd_idx;
2630                                 sh->ops.target2 = -1;
2631                                 s->uptodate++;
2632                         }
2633                 }
2634                 break;
2635         case check_state_compute_run:
2636                 break;
2637         default:
2638                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2639                        __func__, sh->check_state,
2640                        (unsigned long long) sh->sector);
2641                 BUG();
2642         }
2643 }
2644
2645
2646 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2647                                   struct stripe_head_state *s,
2648                                   struct r6_state *r6s, int disks)
2649 {
2650         int pd_idx = sh->pd_idx;
2651         int qd_idx = sh->qd_idx;
2652         struct r5dev *dev;
2653
2654         set_bit(STRIPE_HANDLE, &sh->state);
2655
2656         BUG_ON(s->failed > 2);
2657
2658         /* Want to check and possibly repair P and Q.
2659          * However there could be one 'failed' device, in which
2660          * case we can only check one of them, possibly using the
2661          * other to generate missing data
2662          */
2663
2664         switch (sh->check_state) {
2665         case check_state_idle:
2666                 /* start a new check operation if there are < 2 failures */
2667                 if (s->failed == r6s->q_failed) {
2668                         /* The only possible failed device holds Q, so it
2669                          * makes sense to check P (If anything else were failed,
2670                          * we would have used P to recreate it).
2671                          */
2672                         sh->check_state = check_state_run;
2673                 }
2674                 if (!r6s->q_failed && s->failed < 2) {
2675                         /* Q is not failed, and we didn't use it to generate
2676                          * anything, so it makes sense to check it
2677                          */
2678                         if (sh->check_state == check_state_run)
2679                                 sh->check_state = check_state_run_pq;
2680                         else
2681                                 sh->check_state = check_state_run_q;
2682                 }
2683
2684                 /* discard potentially stale zero_sum_result */
2685                 sh->ops.zero_sum_result = 0;
2686
2687                 if (sh->check_state == check_state_run) {
2688                         /* async_xor_zero_sum destroys the contents of P */
2689                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2690                         s->uptodate--;
2691                 }
2692                 if (sh->check_state >= check_state_run &&
2693                     sh->check_state <= check_state_run_pq) {
2694                         /* async_syndrome_zero_sum preserves P and Q, so
2695                          * no need to mark them !uptodate here
2696                          */
2697                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2698                         break;
2699                 }
2700
2701                 /* we have 2-disk failure */
2702                 BUG_ON(s->failed != 2);
2703                 /* fall through */
2704         case check_state_compute_result:
2705                 sh->check_state = check_state_idle;
2706
2707                 /* check that a write has not made the stripe insync */
2708                 if (test_bit(STRIPE_INSYNC, &sh->state))
2709                         break;
2710
2711                 /* now write out any block on a failed drive,
2712                  * or P or Q if they were recomputed
2713                  */
2714                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2715                 if (s->failed == 2) {
2716                         dev = &sh->dev[r6s->failed_num[1]];
2717                         s->locked++;
2718                         set_bit(R5_LOCKED, &dev->flags);
2719                         set_bit(R5_Wantwrite, &dev->flags);
2720                 }
2721                 if (s->failed >= 1) {
2722                         dev = &sh->dev[r6s->failed_num[0]];
2723                         s->locked++;
2724                         set_bit(R5_LOCKED, &dev->flags);
2725                         set_bit(R5_Wantwrite, &dev->flags);
2726                 }
2727                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2728                         dev = &sh->dev[pd_idx];
2729                         s->locked++;
2730                         set_bit(R5_LOCKED, &dev->flags);
2731                         set_bit(R5_Wantwrite, &dev->flags);
2732                 }
2733                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2734                         dev = &sh->dev[qd_idx];
2735                         s->locked++;
2736                         set_bit(R5_LOCKED, &dev->flags);
2737                         set_bit(R5_Wantwrite, &dev->flags);
2738                 }
2739                 clear_bit(STRIPE_DEGRADED, &sh->state);
2740
2741                 set_bit(STRIPE_INSYNC, &sh->state);
2742                 break;
2743         case check_state_run:
2744         case check_state_run_q:
2745         case check_state_run_pq:
2746                 break; /* we will be called again upon completion */
2747         case check_state_check_result:
2748                 sh->check_state = check_state_idle;
2749
2750                 /* handle a successful check operation, if parity is correct
2751                  * we are done.  Otherwise update the mismatch count and repair
2752                  * parity if !MD_RECOVERY_CHECK
2753                  */
2754                 if (sh->ops.zero_sum_result == 0) {
2755                         /* both parities are correct */
2756                         if (!s->failed)
2757                                 set_bit(STRIPE_INSYNC, &sh->state);
2758                         else {
2759                                 /* in contrast to the raid5 case we can validate
2760                                  * parity, but still have a failure to write
2761                                  * back
2762                                  */
2763                                 sh->check_state = check_state_compute_result;
2764                                 /* Returning at this point means that we may go
2765                                  * off and bring p and/or q uptodate again so
2766                                  * we make sure to check zero_sum_result again
2767                                  * to verify if p or q need writeback
2768                                  */
2769                         }
2770                 } else {
2771                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2772                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2773                                 /* don't try to repair!! */
2774                                 set_bit(STRIPE_INSYNC, &sh->state);
2775                         else {
2776                                 int *target = &sh->ops.target;
2777
2778                                 sh->ops.target = -1;
2779                                 sh->ops.target2 = -1;
2780                                 sh->check_state = check_state_compute_run;
2781                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2782                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2783                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2784                                         set_bit(R5_Wantcompute,
2785                                                 &sh->dev[pd_idx].flags);
2786                                         *target = pd_idx;
2787                                         target = &sh->ops.target2;
2788                                         s->uptodate++;
2789                                 }
2790                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2791                                         set_bit(R5_Wantcompute,
2792                                                 &sh->dev[qd_idx].flags);
2793                                         *target = qd_idx;
2794                                         s->uptodate++;
2795                                 }
2796                         }
2797                 }
2798                 break;
2799         case check_state_compute_run:
2800                 break;
2801         default:
2802                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2803                        __func__, sh->check_state,
2804                        (unsigned long long) sh->sector);
2805                 BUG();
2806         }
2807 }
2808
2809 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2810                                 struct r6_state *r6s)
2811 {
2812         int i;
2813
2814         /* We have read all the blocks in this stripe and now we need to
2815          * copy some of them into a target stripe for expand.
2816          */
2817         struct dma_async_tx_descriptor *tx = NULL;
2818         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2819         for (i = 0; i < sh->disks; i++)
2820                 if (i != sh->pd_idx && i != sh->qd_idx) {
2821                         int dd_idx, j;
2822                         struct stripe_head *sh2;
2823                         struct async_submit_ctl submit;
2824
2825                         sector_t bn = compute_blocknr(sh, i, 1);
2826                         sector_t s = raid5_compute_sector(conf, bn, 0,
2827                                                           &dd_idx, NULL);
2828                         sh2 = get_active_stripe(conf, s, 0, 1);
2829                         if (sh2 == NULL)
2830                                 /* so far only the early blocks of this stripe
2831                                  * have been requested.  When later blocks
2832                                  * get requested, we will try again
2833                                  */
2834                                 continue;
2835                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2836                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2837                                 /* must have already done this block */
2838                                 release_stripe(sh2);
2839                                 continue;
2840                         }
2841
2842                         /* place all the copies on one channel */
2843                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2844                         tx = async_memcpy(sh2->dev[dd_idx].page,
2845                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
2846                                           &submit);
2847
2848                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2849                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2850                         for (j = 0; j < conf->raid_disks; j++)
2851                                 if (j != sh2->pd_idx &&
2852                                     (!r6s || j != sh2->qd_idx) &&
2853                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2854                                         break;
2855                         if (j == conf->raid_disks) {
2856                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2857                                 set_bit(STRIPE_HANDLE, &sh2->state);
2858                         }
2859                         release_stripe(sh2);
2860
2861                 }
2862         /* done submitting copies, wait for them to complete */
2863         if (tx) {
2864                 async_tx_ack(tx);
2865                 dma_wait_for_async_tx(tx);
2866         }
2867 }
2868
2869
2870 /*
2871  * handle_stripe - do things to a stripe.
2872  *
2873  * We lock the stripe and then examine the state of various bits
2874  * to see what needs to be done.
2875  * Possible results:
2876  *    return some read request which now have data
2877  *    return some write requests which are safely on disc
2878  *    schedule a read on some buffers
2879  *    schedule a write of some buffers
2880  *    return confirmation of parity correctness
2881  *
2882  * buffers are taken off read_list or write_list, and bh_cache buffers
2883  * get BH_Lock set before the stripe lock is released.
2884  *
2885  */
2886
2887 static bool handle_stripe5(struct stripe_head *sh)
2888 {
2889         raid5_conf_t *conf = sh->raid_conf;
2890         int disks = sh->disks, i;
2891         struct bio *return_bi = NULL;
2892         struct stripe_head_state s;
2893         struct r5dev *dev;
2894         mdk_rdev_t *blocked_rdev = NULL;
2895         int prexor;
2896
2897         memset(&s, 0, sizeof(s));
2898         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2899                  "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2900                  atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2901                  sh->reconstruct_state);
2902
2903         spin_lock(&sh->lock);
2904         clear_bit(STRIPE_HANDLE, &sh->state);
2905         clear_bit(STRIPE_DELAYED, &sh->state);
2906
2907         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2908         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2909         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2910
2911         /* Now to look around and see what can be done */
2912         rcu_read_lock();
2913         for (i=disks; i--; ) {
2914                 mdk_rdev_t *rdev;
2915                 struct r5dev *dev = &sh->dev[i];
2916                 clear_bit(R5_Insync, &dev->flags);
2917
2918                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2919                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2920                         dev->towrite, dev->written);
2921
2922                 /* maybe we can request a biofill operation
2923                  *
2924                  * new wantfill requests are only permitted while
2925                  * ops_complete_biofill is guaranteed to be inactive
2926                  */
2927                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2928                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2929                         set_bit(R5_Wantfill, &dev->flags);
2930
2931                 /* now count some things */
2932                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2933                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2934                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2935
2936                 if (test_bit(R5_Wantfill, &dev->flags))
2937                         s.to_fill++;
2938                 else if (dev->toread)
2939                         s.to_read++;
2940                 if (dev->towrite) {
2941                         s.to_write++;
2942                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2943                                 s.non_overwrite++;
2944                 }
2945                 if (dev->written)
2946                         s.written++;
2947                 rdev = rcu_dereference(conf->disks[i].rdev);
2948                 if (blocked_rdev == NULL &&
2949                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2950                         blocked_rdev = rdev;
2951                         atomic_inc(&rdev->nr_pending);
2952                 }
2953                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2954                         /* The ReadError flag will just be confusing now */
2955                         clear_bit(R5_ReadError, &dev->flags);
2956                         clear_bit(R5_ReWrite, &dev->flags);
2957                 }
2958                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2959                     || test_bit(R5_ReadError, &dev->flags)) {
2960                         s.failed++;
2961                         s.failed_num = i;
2962                 } else
2963                         set_bit(R5_Insync, &dev->flags);
2964         }
2965         rcu_read_unlock();
2966
2967         if (unlikely(blocked_rdev)) {
2968                 if (s.syncing || s.expanding || s.expanded ||
2969                     s.to_write || s.written) {
2970                         set_bit(STRIPE_HANDLE, &sh->state);
2971                         goto unlock;
2972                 }
2973                 /* There is nothing for the blocked_rdev to block */
2974                 rdev_dec_pending(blocked_rdev, conf->mddev);
2975                 blocked_rdev = NULL;
2976         }
2977
2978         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
2979                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
2980                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
2981         }
2982
2983         pr_debug("locked=%d uptodate=%d to_read=%d"
2984                 " to_write=%d failed=%d failed_num=%d\n",
2985                 s.locked, s.uptodate, s.to_read, s.to_write,
2986                 s.failed, s.failed_num);
2987         /* check if the array has lost two devices and, if so, some requests might
2988          * need to be failed
2989          */
2990         if (s.failed > 1 && s.to_read+s.to_write+s.written)
2991                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
2992         if (s.failed > 1 && s.syncing) {
2993                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2994                 clear_bit(STRIPE_SYNCING, &sh->state);
2995                 s.syncing = 0;
2996         }
2997
2998         /* might be able to return some write requests if the parity block
2999          * is safe, or on a failed drive
3000          */
3001         dev = &sh->dev[sh->pd_idx];
3002         if ( s.written &&
3003              ((test_bit(R5_Insync, &dev->flags) &&
3004                !test_bit(R5_LOCKED, &dev->flags) &&
3005                test_bit(R5_UPTODATE, &dev->flags)) ||
3006                (s.failed == 1 && s.failed_num == sh->pd_idx)))
3007                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3008
3009         /* Now we might consider reading some blocks, either to check/generate
3010          * parity, or to satisfy requests
3011          * or to load a block that is being partially written.
3012          */
3013         if (s.to_read || s.non_overwrite ||
3014             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3015                 handle_stripe_fill5(sh, &s, disks);
3016
3017         /* Now we check to see if any write operations have recently
3018          * completed
3019          */
3020         prexor = 0;
3021         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3022                 prexor = 1;
3023         if (sh->reconstruct_state == reconstruct_state_drain_result ||
3024             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3025                 sh->reconstruct_state = reconstruct_state_idle;
3026
3027                 /* All the 'written' buffers and the parity block are ready to
3028                  * be written back to disk
3029                  */
3030                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3031                 for (i = disks; i--; ) {
3032                         dev = &sh->dev[i];
3033                         if (test_bit(R5_LOCKED, &dev->flags) &&
3034                                 (i == sh->pd_idx || dev->written)) {
3035                                 pr_debug("Writing block %d\n", i);
3036                                 set_bit(R5_Wantwrite, &dev->flags);
3037                                 if (prexor)
3038                                         continue;
3039                                 if (!test_bit(R5_Insync, &dev->flags) ||
3040                                     (i == sh->pd_idx && s.failed == 0))
3041                                         set_bit(STRIPE_INSYNC, &sh->state);
3042                         }
3043                 }
3044                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3045                         atomic_dec(&conf->preread_active_stripes);
3046                         if (atomic_read(&conf->preread_active_stripes) <
3047                                 IO_THRESHOLD)
3048                                 md_wakeup_thread(conf->mddev->thread);
3049                 }
3050         }
3051
3052         /* Now to consider new write requests and what else, if anything
3053          * should be read.  We do not handle new writes when:
3054          * 1/ A 'write' operation (copy+xor) is already in flight.
3055          * 2/ A 'check' operation is in flight, as it may clobber the parity
3056          *    block.
3057          */
3058         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3059                 handle_stripe_dirtying5(conf, sh, &s, disks);
3060
3061         /* maybe we need to check and possibly fix the parity for this stripe
3062          * Any reads will already have been scheduled, so we just see if enough
3063          * data is available.  The parity check is held off while parity
3064          * dependent operations are in flight.
3065          */
3066         if (sh->check_state ||
3067             (s.syncing && s.locked == 0 &&
3068              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3069              !test_bit(STRIPE_INSYNC, &sh->state)))
3070                 handle_parity_checks5(conf, sh, &s, disks);
3071
3072         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3073                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3074                 clear_bit(STRIPE_SYNCING, &sh->state);
3075         }
3076
3077         /* If the failed drive is just a ReadError, then we might need to progress
3078          * the repair/check process
3079          */
3080         if (s.failed == 1 && !conf->mddev->ro &&
3081             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
3082             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
3083             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
3084                 ) {
3085                 dev = &sh->dev[s.failed_num];
3086                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3087                         set_bit(R5_Wantwrite, &dev->flags);
3088                         set_bit(R5_ReWrite, &dev->flags);
3089                         set_bit(R5_LOCKED, &dev->flags);
3090                         s.locked++;
3091                 } else {
3092                         /* let's read it back */
3093                         set_bit(R5_Wantread, &dev->flags);
3094                         set_bit(R5_LOCKED, &dev->flags);
3095                         s.locked++;
3096                 }
3097         }
3098
3099         /* Finish reconstruct operations initiated by the expansion process */
3100         if (sh->reconstruct_state == reconstruct_state_result) {
3101                 struct stripe_head *sh2
3102                         = get_active_stripe(conf, sh->sector, 1, 1);
3103                 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3104                         /* sh cannot be written until sh2 has been read.
3105                          * so arrange for sh to be delayed a little
3106                          */
3107                         set_bit(STRIPE_DELAYED, &sh->state);
3108                         set_bit(STRIPE_HANDLE, &sh->state);
3109                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3110                                               &sh2->state))
3111                                 atomic_inc(&conf->preread_active_stripes);
3112                         release_stripe(sh2);
3113                         goto unlock;
3114                 }
3115                 if (sh2)
3116                         release_stripe(sh2);
3117
3118                 sh->reconstruct_state = reconstruct_state_idle;
3119                 clear_bit(STRIPE_EXPANDING, &sh->state);
3120                 for (i = conf->raid_disks; i--; ) {
3121                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3122                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3123                         s.locked++;
3124                 }
3125         }
3126
3127         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3128             !sh->reconstruct_state) {
3129                 /* Need to write out all blocks after computing parity */
3130                 sh->disks = conf->raid_disks;
3131                 stripe_set_idx(sh->sector, conf, 0, sh);
3132                 schedule_reconstruction(sh, &s, 1, 1);
3133         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3134                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3135                 atomic_dec(&conf->reshape_stripes);
3136                 wake_up(&conf->wait_for_overlap);
3137                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3138         }
3139
3140         if (s.expanding && s.locked == 0 &&
3141             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3142                 handle_stripe_expansion(conf, sh, NULL);
3143
3144  unlock:
3145         spin_unlock(&sh->lock);
3146
3147         /* wait for this device to become unblocked */
3148         if (unlikely(blocked_rdev))
3149                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3150
3151         if (s.ops_request)
3152                 raid_run_ops(sh, s.ops_request);
3153
3154         ops_run_io(sh, &s);
3155
3156         return_io(return_bi);
3157
3158         return blocked_rdev == NULL;
3159 }
3160
3161 static bool handle_stripe6(struct stripe_head *sh)
3162 {
3163         raid5_conf_t *conf = sh->raid_conf;
3164         int disks = sh->disks;
3165         struct bio *return_bi = NULL;
3166         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3167         struct stripe_head_state s;
3168         struct r6_state r6s;
3169         struct r5dev *dev, *pdev, *qdev;
3170         mdk_rdev_t *blocked_rdev = NULL;
3171
3172         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3173                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3174                (unsigned long long)sh->sector, sh->state,
3175                atomic_read(&sh->count), pd_idx, qd_idx,
3176                sh->check_state, sh->reconstruct_state);
3177         memset(&s, 0, sizeof(s));
3178
3179         spin_lock(&sh->lock);
3180         clear_bit(STRIPE_HANDLE, &sh->state);
3181         clear_bit(STRIPE_DELAYED, &sh->state);
3182
3183         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3184         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3185         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3186         /* Now to look around and see what can be done */
3187
3188         rcu_read_lock();
3189         for (i=disks; i--; ) {
3190                 mdk_rdev_t *rdev;
3191                 dev = &sh->dev[i];
3192                 clear_bit(R5_Insync, &dev->flags);
3193
3194                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3195                         i, dev->flags, dev->toread, dev->towrite, dev->written);
3196                 /* maybe we can reply to a read
3197                  *
3198                  * new wantfill requests are only permitted while
3199                  * ops_complete_biofill is guaranteed to be inactive
3200                  */
3201                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3202                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3203                         set_bit(R5_Wantfill, &dev->flags);
3204
3205                 /* now count some things */
3206                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3207                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3208                 if (test_bit(R5_Wantcompute, &dev->flags))
3209                         BUG_ON(++s.compute > 2);
3210
3211                 if (test_bit(R5_Wantfill, &dev->flags)) {
3212                         s.to_fill++;
3213                 } else if (dev->toread)
3214                         s.to_read++;
3215                 if (dev->towrite) {
3216                         s.to_write++;
3217                         if (!test_bit(R5_OVERWRITE, &dev->flags))
3218                                 s.non_overwrite++;
3219                 }
3220                 if (dev->written)
3221                         s.written++;
3222                 rdev = rcu_dereference(conf->disks[i].rdev);
3223                 if (blocked_rdev == NULL &&
3224                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3225                         blocked_rdev = rdev;
3226                         atomic_inc(&rdev->nr_pending);
3227                 }
3228                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3229                         /* The ReadError flag will just be confusing now */
3230                         clear_bit(R5_ReadError, &dev->flags);
3231                         clear_bit(R5_ReWrite, &dev->flags);
3232                 }
3233                 if (!rdev || !test_bit(In_sync, &rdev->flags)
3234                     || test_bit(R5_ReadError, &dev->flags)) {
3235                         if (s.failed < 2)
3236                                 r6s.failed_num[s.failed] = i;
3237                         s.failed++;
3238                 } else
3239                         set_bit(R5_Insync, &dev->flags);
3240         }
3241         rcu_read_unlock();
3242
3243         if (unlikely(blocked_rdev)) {
3244                 if (s.syncing || s.expanding || s.expanded ||
3245                     s.to_write || s.written) {
3246                         set_bit(STRIPE_HANDLE, &sh->state);
3247                         goto unlock;
3248                 }
3249                 /* There is nothing for the blocked_rdev to block */
3250                 rdev_dec_pending(blocked_rdev, conf->mddev);
3251                 blocked_rdev = NULL;
3252         }
3253
3254         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3255                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3256                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3257         }
3258
3259         pr_debug("locked=%d uptodate=%d to_read=%d"
3260                " to_write=%d failed=%d failed_num=%d,%d\n",
3261                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3262                r6s.failed_num[0], r6s.failed_num[1]);
3263         /* check if the array has lost >2 devices and, if so, some requests
3264          * might need to be failed
3265          */
3266         if (s.failed > 2 && s.to_read+s.to_write+s.written)
3267                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3268         if (s.failed > 2 && s.syncing) {
3269                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3270                 clear_bit(STRIPE_SYNCING, &sh->state);
3271                 s.syncing = 0;
3272         }
3273
3274         /*
3275          * might be able to return some write requests if the parity blocks
3276          * are safe, or on a failed drive
3277          */
3278         pdev = &sh->dev[pd_idx];
3279         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3280                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3281         qdev = &sh->dev[qd_idx];
3282         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3283                 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3284
3285         if ( s.written &&
3286              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3287                              && !test_bit(R5_LOCKED, &pdev->flags)
3288                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3289              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3290                              && !test_bit(R5_LOCKED, &qdev->flags)
3291                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3292                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3293
3294         /* Now we might consider reading some blocks, either to check/generate
3295          * parity, or to satisfy requests
3296          * or to load a block that is being partially written.
3297          */
3298         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3299             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3300                 handle_stripe_fill6(sh, &s, &r6s, disks);
3301
3302         /* Now we check to see if any write operations have recently
3303          * completed
3304          */
3305         if (sh->reconstruct_state == reconstruct_state_drain_result) {
3306                 int qd_idx = sh->qd_idx;
3307
3308                 sh->reconstruct_state = reconstruct_state_idle;
3309                 /* All the 'written' buffers and the parity blocks are ready to
3310                  * be written back to disk
3311                  */
3312                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3313                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags));
3314                 for (i = disks; i--; ) {
3315                         dev = &sh->dev[i];
3316                         if (test_bit(R5_LOCKED, &dev->flags) &&
3317                             (i == sh->pd_idx || i == qd_idx ||
3318                              dev->written)) {
3319                                 pr_debug("Writing block %d\n", i);
3320                                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3321                                 set_bit(R5_Wantwrite, &dev->flags);
3322                                 if (!test_bit(R5_Insync, &dev->flags) ||
3323                                     ((i == sh->pd_idx || i == qd_idx) &&
3324                                       s.failed == 0))
3325                                         set_bit(STRIPE_INSYNC, &sh->state);
3326                         }
3327                 }
3328                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3329                         atomic_dec(&conf->preread_active_stripes);
3330                         if (atomic_read(&conf->preread_active_stripes) <
3331                                 IO_THRESHOLD)
3332                                 md_wakeup_thread(conf->mddev->thread);
3333                 }
3334         }
3335
3336         /* Now to consider new write requests and what else, if anything
3337          * should be read.  We do not handle new writes when:
3338          * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
3339          * 2/ A 'check' operation is in flight, as it may clobber the parity
3340          *    block.
3341          */
3342         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3343                 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3344
3345         /* maybe we need to check and possibly fix the parity for this stripe
3346          * Any reads will already have been scheduled, so we just see if enough
3347          * data is available.  The parity check is held off while parity
3348          * dependent operations are in flight.
3349          */
3350         if (sh->check_state ||
3351             (s.syncing && s.locked == 0 &&
3352              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3353              !test_bit(STRIPE_INSYNC, &sh->state)))
3354                 handle_parity_checks6(conf, sh, &s, &r6s, disks);
3355
3356         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3357                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3358                 clear_bit(STRIPE_SYNCING, &sh->state);
3359         }
3360
3361         /* If the failed drives are just a ReadError, then we might need
3362          * to progress the repair/check process
3363          */
3364         if (s.failed <= 2 && !conf->mddev->ro)
3365                 for (i = 0; i < s.failed; i++) {
3366                         dev = &sh->dev[r6s.failed_num[i]];
3367                         if (test_bit(R5_ReadError, &dev->flags)
3368      &nb