block: convert blkdev_issue_flush() to use empty barriers
[linux-2.6.git] / drivers / md / raid5.c
1 /*
2  * raid5.c : Multiple Devices driver for Linux
3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4  *         Copyright (C) 1999, 2000 Ingo Molnar
5  *         Copyright (C) 2002, 2003 H. Peter Anvin
6  *
7  * RAID-4/5/6 management functions.
8  * Thanks to Penguin Computing for making the RAID-6 development possible
9  * by donating a test server!
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 /*
22  * BITMAP UNPLUGGING:
23  *
24  * The sequencing for updating the bitmap reliably is a little
25  * subtle (and I got it wrong the first time) so it deserves some
26  * explanation.
27  *
28  * We group bitmap updates into batches.  Each batch has a number.
29  * We may write out several batches at once, but that isn't very important.
30  * conf->bm_write is the number of the last batch successfully written.
31  * conf->bm_flush is the number of the last batch that was closed to
32  *    new additions.
33  * When we discover that we will need to write to any block in a stripe
34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35  * the number of the batch it will be in. This is bm_flush+1.
36  * When we are ready to do a write, if that batch hasn't been written yet,
37  *   we plug the array and queue the stripe for later.
38  * When an unplug happens, we increment bm_flush, thus closing the current
39  *   batch.
40  * When we notice that bm_flush > bm_write, we write out all pending updates
41  * to the bitmap, and advance bm_write to where bm_flush was.
42  * This may occasionally write a bit out twice, but is sure never to
43  * miss any bits.
44  */
45
46 #include <linux/module.h>
47 #include <linux/slab.h>
48 #include <linux/highmem.h>
49 #include <linux/bitops.h>
50 #include <linux/kthread.h>
51 #include <asm/atomic.h>
52 #include "raid6.h"
53
54 #include <linux/raid/bitmap.h>
55 #include <linux/async_tx.h>
56
57 /*
58  * Stripe cache
59  */
60
61 #define NR_STRIPES              256
62 #define STRIPE_SIZE             PAGE_SIZE
63 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
64 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
65 #define IO_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 #if !RAID6_USE_EMPTY_ZERO_PAGE
97 /* In .bss so it's zeroed */
98 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
99 #endif
100
101 static inline int raid6_next_disk(int disk, int raid_disks)
102 {
103         disk++;
104         return (disk < raid_disks) ? disk : 0;
105 }
106
107 static void return_io(struct bio *return_bi)
108 {
109         struct bio *bi = return_bi;
110         while (bi) {
111
112                 return_bi = bi->bi_next;
113                 bi->bi_next = NULL;
114                 bi->bi_size = 0;
115                 bi->bi_end_io(bi,
116                               test_bit(BIO_UPTODATE, &bi->bi_flags)
117                                 ? 0 : -EIO);
118                 bi = return_bi;
119         }
120 }
121
122 static void print_raid5_conf (raid5_conf_t *conf);
123
124 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
125 {
126         if (atomic_dec_and_test(&sh->count)) {
127                 BUG_ON(!list_empty(&sh->lru));
128                 BUG_ON(atomic_read(&conf->active_stripes)==0);
129                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
130                         if (test_bit(STRIPE_DELAYED, &sh->state)) {
131                                 list_add_tail(&sh->lru, &conf->delayed_list);
132                                 blk_plug_device(conf->mddev->queue);
133                         } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
134                                    sh->bm_seq - conf->seq_write > 0) {
135                                 list_add_tail(&sh->lru, &conf->bitmap_list);
136                                 blk_plug_device(conf->mddev->queue);
137                         } else {
138                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
139                                 list_add_tail(&sh->lru, &conf->handle_list);
140                         }
141                         md_wakeup_thread(conf->mddev->thread);
142                 } else {
143                         BUG_ON(sh->ops.pending);
144                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
145                                 atomic_dec(&conf->preread_active_stripes);
146                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
147                                         md_wakeup_thread(conf->mddev->thread);
148                         }
149                         atomic_dec(&conf->active_stripes);
150                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
151                                 list_add_tail(&sh->lru, &conf->inactive_list);
152                                 wake_up(&conf->wait_for_stripe);
153                                 if (conf->retry_read_aligned)
154                                         md_wakeup_thread(conf->mddev->thread);
155                         }
156                 }
157         }
158 }
159 static void release_stripe(struct stripe_head *sh)
160 {
161         raid5_conf_t *conf = sh->raid_conf;
162         unsigned long flags;
163
164         spin_lock_irqsave(&conf->device_lock, flags);
165         __release_stripe(conf, sh);
166         spin_unlock_irqrestore(&conf->device_lock, flags);
167 }
168
169 static inline void remove_hash(struct stripe_head *sh)
170 {
171         pr_debug("remove_hash(), stripe %llu\n",
172                 (unsigned long long)sh->sector);
173
174         hlist_del_init(&sh->hash);
175 }
176
177 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
178 {
179         struct hlist_head *hp = stripe_hash(conf, sh->sector);
180
181         pr_debug("insert_hash(), stripe %llu\n",
182                 (unsigned long long)sh->sector);
183
184         CHECK_DEVLOCK();
185         hlist_add_head(&sh->hash, hp);
186 }
187
188
189 /* find an idle stripe, make sure it is unhashed, and return it. */
190 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
191 {
192         struct stripe_head *sh = NULL;
193         struct list_head *first;
194
195         CHECK_DEVLOCK();
196         if (list_empty(&conf->inactive_list))
197                 goto out;
198         first = conf->inactive_list.next;
199         sh = list_entry(first, struct stripe_head, lru);
200         list_del_init(first);
201         remove_hash(sh);
202         atomic_inc(&conf->active_stripes);
203 out:
204         return sh;
205 }
206
207 static void shrink_buffers(struct stripe_head *sh, int num)
208 {
209         struct page *p;
210         int i;
211
212         for (i=0; i<num ; i++) {
213                 p = sh->dev[i].page;
214                 if (!p)
215                         continue;
216                 sh->dev[i].page = NULL;
217                 put_page(p);
218         }
219 }
220
221 static int grow_buffers(struct stripe_head *sh, int num)
222 {
223         int i;
224
225         for (i=0; i<num; i++) {
226                 struct page *page;
227
228                 if (!(page = alloc_page(GFP_KERNEL))) {
229                         return 1;
230                 }
231                 sh->dev[i].page = page;
232         }
233         return 0;
234 }
235
236 static void raid5_build_block (struct stripe_head *sh, int i);
237
238 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks)
239 {
240         raid5_conf_t *conf = sh->raid_conf;
241         int i;
242
243         BUG_ON(atomic_read(&sh->count) != 0);
244         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
245         BUG_ON(sh->ops.pending || sh->ops.ack || sh->ops.complete);
246
247         CHECK_DEVLOCK();
248         pr_debug("init_stripe called, stripe %llu\n",
249                 (unsigned long long)sh->sector);
250
251         remove_hash(sh);
252
253         sh->sector = sector;
254         sh->pd_idx = pd_idx;
255         sh->state = 0;
256
257         sh->disks = disks;
258
259         for (i = sh->disks; i--; ) {
260                 struct r5dev *dev = &sh->dev[i];
261
262                 if (dev->toread || dev->read || dev->towrite || dev->written ||
263                     test_bit(R5_LOCKED, &dev->flags)) {
264                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
265                                (unsigned long long)sh->sector, i, dev->toread,
266                                dev->read, dev->towrite, dev->written,
267                                test_bit(R5_LOCKED, &dev->flags));
268                         BUG();
269                 }
270                 dev->flags = 0;
271                 raid5_build_block(sh, i);
272         }
273         insert_hash(conf, sh);
274 }
275
276 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
277 {
278         struct stripe_head *sh;
279         struct hlist_node *hn;
280
281         CHECK_DEVLOCK();
282         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
283         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
284                 if (sh->sector == sector && sh->disks == disks)
285                         return sh;
286         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
287         return NULL;
288 }
289
290 static void unplug_slaves(mddev_t *mddev);
291 static void raid5_unplug_device(struct request_queue *q);
292
293 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks,
294                                              int pd_idx, int noblock)
295 {
296         struct stripe_head *sh;
297
298         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
299
300         spin_lock_irq(&conf->device_lock);
301
302         do {
303                 wait_event_lock_irq(conf->wait_for_stripe,
304                                     conf->quiesce == 0,
305                                     conf->device_lock, /* nothing */);
306                 sh = __find_stripe(conf, sector, disks);
307                 if (!sh) {
308                         if (!conf->inactive_blocked)
309                                 sh = get_free_stripe(conf);
310                         if (noblock && sh == NULL)
311                                 break;
312                         if (!sh) {
313                                 conf->inactive_blocked = 1;
314                                 wait_event_lock_irq(conf->wait_for_stripe,
315                                                     !list_empty(&conf->inactive_list) &&
316                                                     (atomic_read(&conf->active_stripes)
317                                                      < (conf->max_nr_stripes *3/4)
318                                                      || !conf->inactive_blocked),
319                                                     conf->device_lock,
320                                                     raid5_unplug_device(conf->mddev->queue)
321                                         );
322                                 conf->inactive_blocked = 0;
323                         } else
324                                 init_stripe(sh, sector, pd_idx, disks);
325                 } else {
326                         if (atomic_read(&sh->count)) {
327                           BUG_ON(!list_empty(&sh->lru));
328                         } else {
329                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
330                                         atomic_inc(&conf->active_stripes);
331                                 if (list_empty(&sh->lru) &&
332                                     !test_bit(STRIPE_EXPANDING, &sh->state))
333                                         BUG();
334                                 list_del_init(&sh->lru);
335                         }
336                 }
337         } while (sh == NULL);
338
339         if (sh)
340                 atomic_inc(&sh->count);
341
342         spin_unlock_irq(&conf->device_lock);
343         return sh;
344 }
345
346 /* test_and_ack_op() ensures that we only dequeue an operation once */
347 #define test_and_ack_op(op, pend) \
348 do {                                                    \
349         if (test_bit(op, &sh->ops.pending) &&           \
350                 !test_bit(op, &sh->ops.complete)) {     \
351                 if (test_and_set_bit(op, &sh->ops.ack)) \
352                         clear_bit(op, &pend);           \
353                 else                                    \
354                         ack++;                          \
355         } else                                          \
356                 clear_bit(op, &pend);                   \
357 } while (0)
358
359 /* find new work to run, do not resubmit work that is already
360  * in flight
361  */
362 static unsigned long get_stripe_work(struct stripe_head *sh)
363 {
364         unsigned long pending;
365         int ack = 0;
366
367         pending = sh->ops.pending;
368
369         test_and_ack_op(STRIPE_OP_BIOFILL, pending);
370         test_and_ack_op(STRIPE_OP_COMPUTE_BLK, pending);
371         test_and_ack_op(STRIPE_OP_PREXOR, pending);
372         test_and_ack_op(STRIPE_OP_BIODRAIN, pending);
373         test_and_ack_op(STRIPE_OP_POSTXOR, pending);
374         test_and_ack_op(STRIPE_OP_CHECK, pending);
375         if (test_and_clear_bit(STRIPE_OP_IO, &sh->ops.pending))
376                 ack++;
377
378         sh->ops.count -= ack;
379         BUG_ON(sh->ops.count < 0);
380
381         return pending;
382 }
383
384 static void
385 raid5_end_read_request(struct bio *bi, int error);
386 static void
387 raid5_end_write_request(struct bio *bi, int error);
388
389 static void ops_run_io(struct stripe_head *sh)
390 {
391         raid5_conf_t *conf = sh->raid_conf;
392         int i, disks = sh->disks;
393
394         might_sleep();
395
396         for (i = disks; i--; ) {
397                 int rw;
398                 struct bio *bi;
399                 mdk_rdev_t *rdev;
400                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
401                         rw = WRITE;
402                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
403                         rw = READ;
404                 else
405                         continue;
406
407                 bi = &sh->dev[i].req;
408
409                 bi->bi_rw = rw;
410                 if (rw == WRITE)
411                         bi->bi_end_io = raid5_end_write_request;
412                 else
413                         bi->bi_end_io = raid5_end_read_request;
414
415                 rcu_read_lock();
416                 rdev = rcu_dereference(conf->disks[i].rdev);
417                 if (rdev && test_bit(Faulty, &rdev->flags))
418                         rdev = NULL;
419                 if (rdev)
420                         atomic_inc(&rdev->nr_pending);
421                 rcu_read_unlock();
422
423                 if (rdev) {
424                         if (test_bit(STRIPE_SYNCING, &sh->state) ||
425                                 test_bit(STRIPE_EXPAND_SOURCE, &sh->state) ||
426                                 test_bit(STRIPE_EXPAND_READY, &sh->state))
427                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
428
429                         bi->bi_bdev = rdev->bdev;
430                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
431                                 __FUNCTION__, (unsigned long long)sh->sector,
432                                 bi->bi_rw, i);
433                         atomic_inc(&sh->count);
434                         bi->bi_sector = sh->sector + rdev->data_offset;
435                         bi->bi_flags = 1 << BIO_UPTODATE;
436                         bi->bi_vcnt = 1;
437                         bi->bi_max_vecs = 1;
438                         bi->bi_idx = 0;
439                         bi->bi_io_vec = &sh->dev[i].vec;
440                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
441                         bi->bi_io_vec[0].bv_offset = 0;
442                         bi->bi_size = STRIPE_SIZE;
443                         bi->bi_next = NULL;
444                         if (rw == WRITE &&
445                             test_bit(R5_ReWrite, &sh->dev[i].flags))
446                                 atomic_add(STRIPE_SECTORS,
447                                         &rdev->corrected_errors);
448                         generic_make_request(bi);
449                 } else {
450                         if (rw == WRITE)
451                                 set_bit(STRIPE_DEGRADED, &sh->state);
452                         pr_debug("skip op %ld on disc %d for sector %llu\n",
453                                 bi->bi_rw, i, (unsigned long long)sh->sector);
454                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
455                         set_bit(STRIPE_HANDLE, &sh->state);
456                 }
457         }
458 }
459
460 static struct dma_async_tx_descriptor *
461 async_copy_data(int frombio, struct bio *bio, struct page *page,
462         sector_t sector, struct dma_async_tx_descriptor *tx)
463 {
464         struct bio_vec *bvl;
465         struct page *bio_page;
466         int i;
467         int page_offset;
468
469         if (bio->bi_sector >= sector)
470                 page_offset = (signed)(bio->bi_sector - sector) * 512;
471         else
472                 page_offset = (signed)(sector - bio->bi_sector) * -512;
473         bio_for_each_segment(bvl, bio, i) {
474                 int len = bio_iovec_idx(bio, i)->bv_len;
475                 int clen;
476                 int b_offset = 0;
477
478                 if (page_offset < 0) {
479                         b_offset = -page_offset;
480                         page_offset += b_offset;
481                         len -= b_offset;
482                 }
483
484                 if (len > 0 && page_offset + len > STRIPE_SIZE)
485                         clen = STRIPE_SIZE - page_offset;
486                 else
487                         clen = len;
488
489                 if (clen > 0) {
490                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
491                         bio_page = bio_iovec_idx(bio, i)->bv_page;
492                         if (frombio)
493                                 tx = async_memcpy(page, bio_page, page_offset,
494                                         b_offset, clen,
495                                         ASYNC_TX_DEP_ACK,
496                                         tx, NULL, NULL);
497                         else
498                                 tx = async_memcpy(bio_page, page, b_offset,
499                                         page_offset, clen,
500                                         ASYNC_TX_DEP_ACK,
501                                         tx, NULL, NULL);
502                 }
503                 if (clen < len) /* hit end of page */
504                         break;
505                 page_offset +=  len;
506         }
507
508         return tx;
509 }
510
511 static void ops_complete_biofill(void *stripe_head_ref)
512 {
513         struct stripe_head *sh = stripe_head_ref;
514         struct bio *return_bi = NULL;
515         raid5_conf_t *conf = sh->raid_conf;
516         int i;
517
518         pr_debug("%s: stripe %llu\n", __FUNCTION__,
519                 (unsigned long long)sh->sector);
520
521         /* clear completed biofills */
522         for (i = sh->disks; i--; ) {
523                 struct r5dev *dev = &sh->dev[i];
524
525                 /* acknowledge completion of a biofill operation */
526                 /* and check if we need to reply to a read request,
527                  * new R5_Wantfill requests are held off until
528                  * !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending)
529                  */
530                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
531                         struct bio *rbi, *rbi2;
532
533                         /* The access to dev->read is outside of the
534                          * spin_lock_irq(&conf->device_lock), but is protected
535                          * by the STRIPE_OP_BIOFILL pending bit
536                          */
537                         BUG_ON(!dev->read);
538                         rbi = dev->read;
539                         dev->read = NULL;
540                         while (rbi && rbi->bi_sector <
541                                 dev->sector + STRIPE_SECTORS) {
542                                 rbi2 = r5_next_bio(rbi, dev->sector);
543                                 spin_lock_irq(&conf->device_lock);
544                                 if (--rbi->bi_phys_segments == 0) {
545                                         rbi->bi_next = return_bi;
546                                         return_bi = rbi;
547                                 }
548                                 spin_unlock_irq(&conf->device_lock);
549                                 rbi = rbi2;
550                         }
551                 }
552         }
553         clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack);
554         clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending);
555
556         return_io(return_bi);
557
558         set_bit(STRIPE_HANDLE, &sh->state);
559         release_stripe(sh);
560 }
561
562 static void ops_run_biofill(struct stripe_head *sh)
563 {
564         struct dma_async_tx_descriptor *tx = NULL;
565         raid5_conf_t *conf = sh->raid_conf;
566         int i;
567
568         pr_debug("%s: stripe %llu\n", __FUNCTION__,
569                 (unsigned long long)sh->sector);
570
571         for (i = sh->disks; i--; ) {
572                 struct r5dev *dev = &sh->dev[i];
573                 if (test_bit(R5_Wantfill, &dev->flags)) {
574                         struct bio *rbi;
575                         spin_lock_irq(&conf->device_lock);
576                         dev->read = rbi = dev->toread;
577                         dev->toread = NULL;
578                         spin_unlock_irq(&conf->device_lock);
579                         while (rbi && rbi->bi_sector <
580                                 dev->sector + STRIPE_SECTORS) {
581                                 tx = async_copy_data(0, rbi, dev->page,
582                                         dev->sector, tx);
583                                 rbi = r5_next_bio(rbi, dev->sector);
584                         }
585                 }
586         }
587
588         atomic_inc(&sh->count);
589         async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
590                 ops_complete_biofill, sh);
591 }
592
593 static void ops_complete_compute5(void *stripe_head_ref)
594 {
595         struct stripe_head *sh = stripe_head_ref;
596         int target = sh->ops.target;
597         struct r5dev *tgt = &sh->dev[target];
598
599         pr_debug("%s: stripe %llu\n", __FUNCTION__,
600                 (unsigned long long)sh->sector);
601
602         set_bit(R5_UPTODATE, &tgt->flags);
603         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
604         clear_bit(R5_Wantcompute, &tgt->flags);
605         set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
606         set_bit(STRIPE_HANDLE, &sh->state);
607         release_stripe(sh);
608 }
609
610 static struct dma_async_tx_descriptor *
611 ops_run_compute5(struct stripe_head *sh, unsigned long pending)
612 {
613         /* kernel stack size limits the total number of disks */
614         int disks = sh->disks;
615         struct page *xor_srcs[disks];
616         int target = sh->ops.target;
617         struct r5dev *tgt = &sh->dev[target];
618         struct page *xor_dest = tgt->page;
619         int count = 0;
620         struct dma_async_tx_descriptor *tx;
621         int i;
622
623         pr_debug("%s: stripe %llu block: %d\n",
624                 __FUNCTION__, (unsigned long long)sh->sector, target);
625         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
626
627         for (i = disks; i--; )
628                 if (i != target)
629                         xor_srcs[count++] = sh->dev[i].page;
630
631         atomic_inc(&sh->count);
632
633         if (unlikely(count == 1))
634                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
635                         0, NULL, ops_complete_compute5, sh);
636         else
637                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
638                         ASYNC_TX_XOR_ZERO_DST, NULL,
639                         ops_complete_compute5, sh);
640
641         /* ack now if postxor is not set to be run */
642         if (tx && !test_bit(STRIPE_OP_POSTXOR, &pending))
643                 async_tx_ack(tx);
644
645         return tx;
646 }
647
648 static void ops_complete_prexor(void *stripe_head_ref)
649 {
650         struct stripe_head *sh = stripe_head_ref;
651
652         pr_debug("%s: stripe %llu\n", __FUNCTION__,
653                 (unsigned long long)sh->sector);
654
655         set_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
656 }
657
658 static struct dma_async_tx_descriptor *
659 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
660 {
661         /* kernel stack size limits the total number of disks */
662         int disks = sh->disks;
663         struct page *xor_srcs[disks];
664         int count = 0, pd_idx = sh->pd_idx, i;
665
666         /* existing parity data subtracted */
667         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
668
669         pr_debug("%s: stripe %llu\n", __FUNCTION__,
670                 (unsigned long long)sh->sector);
671
672         for (i = disks; i--; ) {
673                 struct r5dev *dev = &sh->dev[i];
674                 /* Only process blocks that are known to be uptodate */
675                 if (dev->towrite && test_bit(R5_Wantprexor, &dev->flags))
676                         xor_srcs[count++] = dev->page;
677         }
678
679         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
680                 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
681                 ops_complete_prexor, sh);
682
683         return tx;
684 }
685
686 static struct dma_async_tx_descriptor *
687 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
688 {
689         int disks = sh->disks;
690         int pd_idx = sh->pd_idx, i;
691
692         /* check if prexor is active which means only process blocks
693          * that are part of a read-modify-write (Wantprexor)
694          */
695         int prexor = test_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
696
697         pr_debug("%s: stripe %llu\n", __FUNCTION__,
698                 (unsigned long long)sh->sector);
699
700         for (i = disks; i--; ) {
701                 struct r5dev *dev = &sh->dev[i];
702                 struct bio *chosen;
703                 int towrite;
704
705                 towrite = 0;
706                 if (prexor) { /* rmw */
707                         if (dev->towrite &&
708                             test_bit(R5_Wantprexor, &dev->flags))
709                                 towrite = 1;
710                 } else { /* rcw */
711                         if (i != pd_idx && dev->towrite &&
712                                 test_bit(R5_LOCKED, &dev->flags))
713                                 towrite = 1;
714                 }
715
716                 if (towrite) {
717                         struct bio *wbi;
718
719                         spin_lock(&sh->lock);
720                         chosen = dev->towrite;
721                         dev->towrite = NULL;
722                         BUG_ON(dev->written);
723                         wbi = dev->written = chosen;
724                         spin_unlock(&sh->lock);
725
726                         while (wbi && wbi->bi_sector <
727                                 dev->sector + STRIPE_SECTORS) {
728                                 tx = async_copy_data(1, wbi, dev->page,
729                                         dev->sector, tx);
730                                 wbi = r5_next_bio(wbi, dev->sector);
731                         }
732                 }
733         }
734
735         return tx;
736 }
737
738 static void ops_complete_postxor(void *stripe_head_ref)
739 {
740         struct stripe_head *sh = stripe_head_ref;
741
742         pr_debug("%s: stripe %llu\n", __FUNCTION__,
743                 (unsigned long long)sh->sector);
744
745         set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
746         set_bit(STRIPE_HANDLE, &sh->state);
747         release_stripe(sh);
748 }
749
750 static void ops_complete_write(void *stripe_head_ref)
751 {
752         struct stripe_head *sh = stripe_head_ref;
753         int disks = sh->disks, i, pd_idx = sh->pd_idx;
754
755         pr_debug("%s: stripe %llu\n", __FUNCTION__,
756                 (unsigned long long)sh->sector);
757
758         for (i = disks; i--; ) {
759                 struct r5dev *dev = &sh->dev[i];
760                 if (dev->written || i == pd_idx)
761                         set_bit(R5_UPTODATE, &dev->flags);
762         }
763
764         set_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
765         set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
766
767         set_bit(STRIPE_HANDLE, &sh->state);
768         release_stripe(sh);
769 }
770
771 static void
772 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
773 {
774         /* kernel stack size limits the total number of disks */
775         int disks = sh->disks;
776         struct page *xor_srcs[disks];
777
778         int count = 0, pd_idx = sh->pd_idx, i;
779         struct page *xor_dest;
780         int prexor = test_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
781         unsigned long flags;
782         dma_async_tx_callback callback;
783
784         pr_debug("%s: stripe %llu\n", __FUNCTION__,
785                 (unsigned long long)sh->sector);
786
787         /* check if prexor is active which means only process blocks
788          * that are part of a read-modify-write (written)
789          */
790         if (prexor) {
791                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
792                 for (i = disks; i--; ) {
793                         struct r5dev *dev = &sh->dev[i];
794                         if (dev->written)
795                                 xor_srcs[count++] = dev->page;
796                 }
797         } else {
798                 xor_dest = sh->dev[pd_idx].page;
799                 for (i = disks; i--; ) {
800                         struct r5dev *dev = &sh->dev[i];
801                         if (i != pd_idx)
802                                 xor_srcs[count++] = dev->page;
803                 }
804         }
805
806         /* check whether this postxor is part of a write */
807         callback = test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending) ?
808                 ops_complete_write : ops_complete_postxor;
809
810         /* 1/ if we prexor'd then the dest is reused as a source
811          * 2/ if we did not prexor then we are redoing the parity
812          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
813          * for the synchronous xor case
814          */
815         flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
816                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
817
818         atomic_inc(&sh->count);
819
820         if (unlikely(count == 1)) {
821                 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
822                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
823                         flags, tx, callback, sh);
824         } else
825                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
826                         flags, tx, callback, sh);
827 }
828
829 static void ops_complete_check(void *stripe_head_ref)
830 {
831         struct stripe_head *sh = stripe_head_ref;
832         int pd_idx = sh->pd_idx;
833
834         pr_debug("%s: stripe %llu\n", __FUNCTION__,
835                 (unsigned long long)sh->sector);
836
837         if (test_and_clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending) &&
838                 sh->ops.zero_sum_result == 0)
839                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
840
841         set_bit(STRIPE_OP_CHECK, &sh->ops.complete);
842         set_bit(STRIPE_HANDLE, &sh->state);
843         release_stripe(sh);
844 }
845
846 static void ops_run_check(struct stripe_head *sh)
847 {
848         /* kernel stack size limits the total number of disks */
849         int disks = sh->disks;
850         struct page *xor_srcs[disks];
851         struct dma_async_tx_descriptor *tx;
852
853         int count = 0, pd_idx = sh->pd_idx, i;
854         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
855
856         pr_debug("%s: stripe %llu\n", __FUNCTION__,
857                 (unsigned long long)sh->sector);
858
859         for (i = disks; i--; ) {
860                 struct r5dev *dev = &sh->dev[i];
861                 if (i != pd_idx)
862                         xor_srcs[count++] = dev->page;
863         }
864
865         tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
866                 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
867
868         if (tx)
869                 set_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
870         else
871                 clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
872
873         atomic_inc(&sh->count);
874         tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
875                 ops_complete_check, sh);
876 }
877
878 static void raid5_run_ops(struct stripe_head *sh, unsigned long pending)
879 {
880         int overlap_clear = 0, i, disks = sh->disks;
881         struct dma_async_tx_descriptor *tx = NULL;
882
883         if (test_bit(STRIPE_OP_BIOFILL, &pending)) {
884                 ops_run_biofill(sh);
885                 overlap_clear++;
886         }
887
888         if (test_bit(STRIPE_OP_COMPUTE_BLK, &pending))
889                 tx = ops_run_compute5(sh, pending);
890
891         if (test_bit(STRIPE_OP_PREXOR, &pending))
892                 tx = ops_run_prexor(sh, tx);
893
894         if (test_bit(STRIPE_OP_BIODRAIN, &pending)) {
895                 tx = ops_run_biodrain(sh, tx);
896                 overlap_clear++;
897         }
898
899         if (test_bit(STRIPE_OP_POSTXOR, &pending))
900                 ops_run_postxor(sh, tx);
901
902         if (test_bit(STRIPE_OP_CHECK, &pending))
903                 ops_run_check(sh);
904
905         if (test_bit(STRIPE_OP_IO, &pending))
906                 ops_run_io(sh);
907
908         if (overlap_clear)
909                 for (i = disks; i--; ) {
910                         struct r5dev *dev = &sh->dev[i];
911                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
912                                 wake_up(&sh->raid_conf->wait_for_overlap);
913                 }
914 }
915
916 static int grow_one_stripe(raid5_conf_t *conf)
917 {
918         struct stripe_head *sh;
919         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
920         if (!sh)
921                 return 0;
922         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
923         sh->raid_conf = conf;
924         spin_lock_init(&sh->lock);
925
926         if (grow_buffers(sh, conf->raid_disks)) {
927                 shrink_buffers(sh, conf->raid_disks);
928                 kmem_cache_free(conf->slab_cache, sh);
929                 return 0;
930         }
931         sh->disks = conf->raid_disks;
932         /* we just created an active stripe so... */
933         atomic_set(&sh->count, 1);
934         atomic_inc(&conf->active_stripes);
935         INIT_LIST_HEAD(&sh->lru);
936         release_stripe(sh);
937         return 1;
938 }
939
940 static int grow_stripes(raid5_conf_t *conf, int num)
941 {
942         struct kmem_cache *sc;
943         int devs = conf->raid_disks;
944
945         sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
946         sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
947         conf->active_name = 0;
948         sc = kmem_cache_create(conf->cache_name[conf->active_name],
949                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
950                                0, 0, NULL);
951         if (!sc)
952                 return 1;
953         conf->slab_cache = sc;
954         conf->pool_size = devs;
955         while (num--)
956                 if (!grow_one_stripe(conf))
957                         return 1;
958         return 0;
959 }
960
961 #ifdef CONFIG_MD_RAID5_RESHAPE
962 static int resize_stripes(raid5_conf_t *conf, int newsize)
963 {
964         /* Make all the stripes able to hold 'newsize' devices.
965          * New slots in each stripe get 'page' set to a new page.
966          *
967          * This happens in stages:
968          * 1/ create a new kmem_cache and allocate the required number of
969          *    stripe_heads.
970          * 2/ gather all the old stripe_heads and tranfer the pages across
971          *    to the new stripe_heads.  This will have the side effect of
972          *    freezing the array as once all stripe_heads have been collected,
973          *    no IO will be possible.  Old stripe heads are freed once their
974          *    pages have been transferred over, and the old kmem_cache is
975          *    freed when all stripes are done.
976          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
977          *    we simple return a failre status - no need to clean anything up.
978          * 4/ allocate new pages for the new slots in the new stripe_heads.
979          *    If this fails, we don't bother trying the shrink the
980          *    stripe_heads down again, we just leave them as they are.
981          *    As each stripe_head is processed the new one is released into
982          *    active service.
983          *
984          * Once step2 is started, we cannot afford to wait for a write,
985          * so we use GFP_NOIO allocations.
986          */
987         struct stripe_head *osh, *nsh;
988         LIST_HEAD(newstripes);
989         struct disk_info *ndisks;
990         int err = 0;
991         struct kmem_cache *sc;
992         int i;
993
994         if (newsize <= conf->pool_size)
995                 return 0; /* never bother to shrink */
996
997         md_allow_write(conf->mddev);
998
999         /* Step 1 */
1000         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1001                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1002                                0, 0, NULL);
1003         if (!sc)
1004                 return -ENOMEM;
1005
1006         for (i = conf->max_nr_stripes; i; i--) {
1007                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1008                 if (!nsh)
1009                         break;
1010
1011                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1012
1013                 nsh->raid_conf = conf;
1014                 spin_lock_init(&nsh->lock);
1015
1016                 list_add(&nsh->lru, &newstripes);
1017         }
1018         if (i) {
1019                 /* didn't get enough, give up */
1020                 while (!list_empty(&newstripes)) {
1021                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1022                         list_del(&nsh->lru);
1023                         kmem_cache_free(sc, nsh);
1024                 }
1025                 kmem_cache_destroy(sc);
1026                 return -ENOMEM;
1027         }
1028         /* Step 2 - Must use GFP_NOIO now.
1029          * OK, we have enough stripes, start collecting inactive
1030          * stripes and copying them over
1031          */
1032         list_for_each_entry(nsh, &newstripes, lru) {
1033                 spin_lock_irq(&conf->device_lock);
1034                 wait_event_lock_irq(conf->wait_for_stripe,
1035                                     !list_empty(&conf->inactive_list),
1036                                     conf->device_lock,
1037                                     unplug_slaves(conf->mddev)
1038                         );
1039                 osh = get_free_stripe(conf);
1040                 spin_unlock_irq(&conf->device_lock);
1041                 atomic_set(&nsh->count, 1);
1042                 for(i=0; i<conf->pool_size; i++)
1043                         nsh->dev[i].page = osh->dev[i].page;
1044                 for( ; i<newsize; i++)
1045                         nsh->dev[i].page = NULL;
1046                 kmem_cache_free(conf->slab_cache, osh);
1047         }
1048         kmem_cache_destroy(conf->slab_cache);
1049
1050         /* Step 3.
1051          * At this point, we are holding all the stripes so the array
1052          * is completely stalled, so now is a good time to resize
1053          * conf->disks.
1054          */
1055         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1056         if (ndisks) {
1057                 for (i=0; i<conf->raid_disks; i++)
1058                         ndisks[i] = conf->disks[i];
1059                 kfree(conf->disks);
1060                 conf->disks = ndisks;
1061         } else
1062                 err = -ENOMEM;
1063
1064         /* Step 4, return new stripes to service */
1065         while(!list_empty(&newstripes)) {
1066                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1067                 list_del_init(&nsh->lru);
1068                 for (i=conf->raid_disks; i < newsize; i++)
1069                         if (nsh->dev[i].page == NULL) {
1070                                 struct page *p = alloc_page(GFP_NOIO);
1071                                 nsh->dev[i].page = p;
1072                                 if (!p)
1073                                         err = -ENOMEM;
1074                         }
1075                 release_stripe(nsh);
1076         }
1077         /* critical section pass, GFP_NOIO no longer needed */
1078
1079         conf->slab_cache = sc;
1080         conf->active_name = 1-conf->active_name;
1081         conf->pool_size = newsize;
1082         return err;
1083 }
1084 #endif
1085
1086 static int drop_one_stripe(raid5_conf_t *conf)
1087 {
1088         struct stripe_head *sh;
1089
1090         spin_lock_irq(&conf->device_lock);
1091         sh = get_free_stripe(conf);
1092         spin_unlock_irq(&conf->device_lock);
1093         if (!sh)
1094                 return 0;
1095         BUG_ON(atomic_read(&sh->count));
1096         shrink_buffers(sh, conf->pool_size);
1097         kmem_cache_free(conf->slab_cache, sh);
1098         atomic_dec(&conf->active_stripes);
1099         return 1;
1100 }
1101
1102 static void shrink_stripes(raid5_conf_t *conf)
1103 {
1104         while (drop_one_stripe(conf))
1105                 ;
1106
1107         if (conf->slab_cache)
1108                 kmem_cache_destroy(conf->slab_cache);
1109         conf->slab_cache = NULL;
1110 }
1111
1112 static void raid5_end_read_request(struct bio * bi, int error)
1113 {
1114         struct stripe_head *sh = bi->bi_private;
1115         raid5_conf_t *conf = sh->raid_conf;
1116         int disks = sh->disks, i;
1117         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1118         char b[BDEVNAME_SIZE];
1119         mdk_rdev_t *rdev;
1120
1121
1122         for (i=0 ; i<disks; i++)
1123                 if (bi == &sh->dev[i].req)
1124                         break;
1125
1126         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1127                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1128                 uptodate);
1129         if (i == disks) {
1130                 BUG();
1131                 return;
1132         }
1133
1134         if (uptodate) {
1135                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1136                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1137                         rdev = conf->disks[i].rdev;
1138                         printk(KERN_INFO "raid5:%s: read error corrected (%lu sectors at %llu on %s)\n",
1139                                mdname(conf->mddev), STRIPE_SECTORS,
1140                                (unsigned long long)sh->sector + rdev->data_offset,
1141                                bdevname(rdev->bdev, b));
1142                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1143                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1144                 }
1145                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1146                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1147         } else {
1148                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1149                 int retry = 0;
1150                 rdev = conf->disks[i].rdev;
1151
1152                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1153                 atomic_inc(&rdev->read_errors);
1154                 if (conf->mddev->degraded)
1155                         printk(KERN_WARNING "raid5:%s: read error not correctable (sector %llu on %s).\n",
1156                                mdname(conf->mddev),
1157                                (unsigned long long)sh->sector + rdev->data_offset,
1158                                bdn);
1159                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1160                         /* Oh, no!!! */
1161                         printk(KERN_WARNING "raid5:%s: read error NOT corrected!! (sector %llu on %s).\n",
1162                                mdname(conf->mddev),
1163                                (unsigned long long)sh->sector + rdev->data_offset,
1164                                bdn);
1165                 else if (atomic_read(&rdev->read_errors)
1166                          > conf->max_nr_stripes)
1167                         printk(KERN_WARNING
1168                                "raid5:%s: Too many read errors, failing device %s.\n",
1169                                mdname(conf->mddev), bdn);
1170                 else
1171                         retry = 1;
1172                 if (retry)
1173                         set_bit(R5_ReadError, &sh->dev[i].flags);
1174                 else {
1175                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1176                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1177                         md_error(conf->mddev, rdev);
1178                 }
1179         }
1180         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1181         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1182         set_bit(STRIPE_HANDLE, &sh->state);
1183         release_stripe(sh);
1184 }
1185
1186 static void raid5_end_write_request (struct bio *bi, int error)
1187 {
1188         struct stripe_head *sh = bi->bi_private;
1189         raid5_conf_t *conf = sh->raid_conf;
1190         int disks = sh->disks, i;
1191         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1192
1193         for (i=0 ; i<disks; i++)
1194                 if (bi == &sh->dev[i].req)
1195                         break;
1196
1197         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1198                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1199                 uptodate);
1200         if (i == disks) {
1201                 BUG();
1202                 return;
1203         }
1204
1205         if (!uptodate)
1206                 md_error(conf->mddev, conf->disks[i].rdev);
1207
1208         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1209         
1210         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1211         set_bit(STRIPE_HANDLE, &sh->state);
1212         release_stripe(sh);
1213 }
1214
1215
1216 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1217         
1218 static void raid5_build_block (struct stripe_head *sh, int i)
1219 {
1220         struct r5dev *dev = &sh->dev[i];
1221
1222         bio_init(&dev->req);
1223         dev->req.bi_io_vec = &dev->vec;
1224         dev->req.bi_vcnt++;
1225         dev->req.bi_max_vecs++;
1226         dev->vec.bv_page = dev->page;
1227         dev->vec.bv_len = STRIPE_SIZE;
1228         dev->vec.bv_offset = 0;
1229
1230         dev->req.bi_sector = sh->sector;
1231         dev->req.bi_private = sh;
1232
1233         dev->flags = 0;
1234         dev->sector = compute_blocknr(sh, i);
1235 }
1236
1237 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1238 {
1239         char b[BDEVNAME_SIZE];
1240         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1241         pr_debug("raid5: error called\n");
1242
1243         if (!test_bit(Faulty, &rdev->flags)) {
1244                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1245                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1246                         unsigned long flags;
1247                         spin_lock_irqsave(&conf->device_lock, flags);
1248                         mddev->degraded++;
1249                         spin_unlock_irqrestore(&conf->device_lock, flags);
1250                         /*
1251                          * if recovery was running, make sure it aborts.
1252                          */
1253                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
1254                 }
1255                 set_bit(Faulty, &rdev->flags);
1256                 printk (KERN_ALERT
1257                         "raid5: Disk failure on %s, disabling device."
1258                         " Operation continuing on %d devices\n",
1259                         bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1260         }
1261 }
1262
1263 /*
1264  * Input: a 'big' sector number,
1265  * Output: index of the data and parity disk, and the sector # in them.
1266  */
1267 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1268                         unsigned int data_disks, unsigned int * dd_idx,
1269                         unsigned int * pd_idx, raid5_conf_t *conf)
1270 {
1271         long stripe;
1272         unsigned long chunk_number;
1273         unsigned int chunk_offset;
1274         sector_t new_sector;
1275         int sectors_per_chunk = conf->chunk_size >> 9;
1276
1277         /* First compute the information on this sector */
1278
1279         /*
1280          * Compute the chunk number and the sector offset inside the chunk
1281          */
1282         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1283         chunk_number = r_sector;
1284         BUG_ON(r_sector != chunk_number);
1285
1286         /*
1287          * Compute the stripe number
1288          */
1289         stripe = chunk_number / data_disks;
1290
1291         /*
1292          * Compute the data disk and parity disk indexes inside the stripe
1293          */
1294         *dd_idx = chunk_number % data_disks;
1295
1296         /*
1297          * Select the parity disk based on the user selected algorithm.
1298          */
1299         switch(conf->level) {
1300         case 4:
1301                 *pd_idx = data_disks;
1302                 break;
1303         case 5:
1304                 switch (conf->algorithm) {
1305                 case ALGORITHM_LEFT_ASYMMETRIC:
1306                         *pd_idx = data_disks - stripe % raid_disks;
1307                         if (*dd_idx >= *pd_idx)
1308                                 (*dd_idx)++;
1309                         break;
1310                 case ALGORITHM_RIGHT_ASYMMETRIC:
1311                         *pd_idx = stripe % raid_disks;
1312                         if (*dd_idx >= *pd_idx)
1313                                 (*dd_idx)++;
1314                         break;
1315                 case ALGORITHM_LEFT_SYMMETRIC:
1316                         *pd_idx = data_disks - stripe % raid_disks;
1317                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1318                         break;
1319                 case ALGORITHM_RIGHT_SYMMETRIC:
1320                         *pd_idx = stripe % raid_disks;
1321                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1322                         break;
1323                 default:
1324                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1325                                 conf->algorithm);
1326                 }
1327                 break;
1328         case 6:
1329
1330                 /**** FIX THIS ****/
1331                 switch (conf->algorithm) {
1332                 case ALGORITHM_LEFT_ASYMMETRIC:
1333                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1334                         if (*pd_idx == raid_disks-1)
1335                                 (*dd_idx)++;    /* Q D D D P */
1336                         else if (*dd_idx >= *pd_idx)
1337                                 (*dd_idx) += 2; /* D D P Q D */
1338                         break;
1339                 case ALGORITHM_RIGHT_ASYMMETRIC:
1340                         *pd_idx = stripe % raid_disks;
1341                         if (*pd_idx == raid_disks-1)
1342                                 (*dd_idx)++;    /* Q D D D P */
1343                         else if (*dd_idx >= *pd_idx)
1344                                 (*dd_idx) += 2; /* D D P Q D */
1345                         break;
1346                 case ALGORITHM_LEFT_SYMMETRIC:
1347                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1348                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1349                         break;
1350                 case ALGORITHM_RIGHT_SYMMETRIC:
1351                         *pd_idx = stripe % raid_disks;
1352                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1353                         break;
1354                 default:
1355                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1356                                 conf->algorithm);
1357                 }
1358                 break;
1359         }
1360
1361         /*
1362          * Finally, compute the new sector number
1363          */
1364         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1365         return new_sector;
1366 }
1367
1368
1369 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1370 {
1371         raid5_conf_t *conf = sh->raid_conf;
1372         int raid_disks = sh->disks;
1373         int data_disks = raid_disks - conf->max_degraded;
1374         sector_t new_sector = sh->sector, check;
1375         int sectors_per_chunk = conf->chunk_size >> 9;
1376         sector_t stripe;
1377         int chunk_offset;
1378         int chunk_number, dummy1, dummy2, dd_idx = i;
1379         sector_t r_sector;
1380
1381
1382         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1383         stripe = new_sector;
1384         BUG_ON(new_sector != stripe);
1385
1386         if (i == sh->pd_idx)
1387                 return 0;
1388         switch(conf->level) {
1389         case 4: break;
1390         case 5:
1391                 switch (conf->algorithm) {
1392                 case ALGORITHM_LEFT_ASYMMETRIC:
1393                 case ALGORITHM_RIGHT_ASYMMETRIC:
1394                         if (i > sh->pd_idx)
1395                                 i--;
1396                         break;
1397                 case ALGORITHM_LEFT_SYMMETRIC:
1398                 case ALGORITHM_RIGHT_SYMMETRIC:
1399                         if (i < sh->pd_idx)
1400                                 i += raid_disks;
1401                         i -= (sh->pd_idx + 1);
1402                         break;
1403                 default:
1404                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1405                                conf->algorithm);
1406                 }
1407                 break;
1408         case 6:
1409                 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1410                         return 0; /* It is the Q disk */
1411                 switch (conf->algorithm) {
1412                 case ALGORITHM_LEFT_ASYMMETRIC:
1413                 case ALGORITHM_RIGHT_ASYMMETRIC:
1414                         if (sh->pd_idx == raid_disks-1)
1415                                 i--;    /* Q D D D P */
1416                         else if (i > sh->pd_idx)
1417                                 i -= 2; /* D D P Q D */
1418                         break;
1419                 case ALGORITHM_LEFT_SYMMETRIC:
1420                 case ALGORITHM_RIGHT_SYMMETRIC:
1421                         if (sh->pd_idx == raid_disks-1)
1422                                 i--; /* Q D D D P */
1423                         else {
1424                                 /* D D P Q D */
1425                                 if (i < sh->pd_idx)
1426                                         i += raid_disks;
1427                                 i -= (sh->pd_idx + 2);
1428                         }
1429                         break;
1430                 default:
1431                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1432                                 conf->algorithm);
1433                 }
1434                 break;
1435         }
1436
1437         chunk_number = stripe * data_disks + i;
1438         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1439
1440         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1441         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1442                 printk(KERN_ERR "compute_blocknr: map not correct\n");
1443                 return 0;
1444         }
1445         return r_sector;
1446 }
1447
1448
1449
1450 /*
1451  * Copy data between a page in the stripe cache, and one or more bion
1452  * The page could align with the middle of the bio, or there could be
1453  * several bion, each with several bio_vecs, which cover part of the page
1454  * Multiple bion are linked together on bi_next.  There may be extras
1455  * at the end of this list.  We ignore them.
1456  */
1457 static void copy_data(int frombio, struct bio *bio,
1458                      struct page *page,
1459                      sector_t sector)
1460 {
1461         char *pa = page_address(page);
1462         struct bio_vec *bvl;
1463         int i;
1464         int page_offset;
1465
1466         if (bio->bi_sector >= sector)
1467                 page_offset = (signed)(bio->bi_sector - sector) * 512;
1468         else
1469                 page_offset = (signed)(sector - bio->bi_sector) * -512;
1470         bio_for_each_segment(bvl, bio, i) {
1471                 int len = bio_iovec_idx(bio,i)->bv_len;
1472                 int clen;
1473                 int b_offset = 0;
1474
1475                 if (page_offset < 0) {
1476                         b_offset = -page_offset;
1477                         page_offset += b_offset;
1478                         len -= b_offset;
1479                 }
1480
1481                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1482                         clen = STRIPE_SIZE - page_offset;
1483                 else clen = len;
1484
1485                 if (clen > 0) {
1486                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1487                         if (frombio)
1488                                 memcpy(pa+page_offset, ba+b_offset, clen);
1489                         else
1490                                 memcpy(ba+b_offset, pa+page_offset, clen);
1491                         __bio_kunmap_atomic(ba, KM_USER0);
1492                 }
1493                 if (clen < len) /* hit end of page */
1494                         break;
1495                 page_offset +=  len;
1496         }
1497 }
1498
1499 #define check_xor()     do {                                              \
1500                                 if (count == MAX_XOR_BLOCKS) {            \
1501                                 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1502                                 count = 0;                                \
1503                            }                                              \
1504                         } while(0)
1505
1506 static void compute_parity6(struct stripe_head *sh, int method)
1507 {
1508         raid6_conf_t *conf = sh->raid_conf;
1509         int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1510         struct bio *chosen;
1511         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1512         void *ptrs[disks];
1513
1514         qd_idx = raid6_next_disk(pd_idx, disks);
1515         d0_idx = raid6_next_disk(qd_idx, disks);
1516
1517         pr_debug("compute_parity, stripe %llu, method %d\n",
1518                 (unsigned long long)sh->sector, method);
1519
1520         switch(method) {
1521         case READ_MODIFY_WRITE:
1522                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
1523         case RECONSTRUCT_WRITE:
1524                 for (i= disks; i-- ;)
1525                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1526                                 chosen = sh->dev[i].towrite;
1527                                 sh->dev[i].towrite = NULL;
1528
1529                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1530                                         wake_up(&conf->wait_for_overlap);
1531
1532                                 BUG_ON(sh->dev[i].written);
1533                                 sh->dev[i].written = chosen;
1534                         }
1535                 break;
1536         case CHECK_PARITY:
1537                 BUG();          /* Not implemented yet */
1538         }
1539
1540         for (i = disks; i--;)
1541                 if (sh->dev[i].written) {
1542                         sector_t sector = sh->dev[i].sector;
1543                         struct bio *wbi = sh->dev[i].written;
1544                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1545                                 copy_data(1, wbi, sh->dev[i].page, sector);
1546                                 wbi = r5_next_bio(wbi, sector);
1547                         }
1548
1549                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1550                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1551                 }
1552
1553 //      switch(method) {
1554 //      case RECONSTRUCT_WRITE:
1555 //      case CHECK_PARITY:
1556 //      case UPDATE_PARITY:
1557                 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1558                 /* FIX: Is this ordering of drives even remotely optimal? */
1559                 count = 0;
1560                 i = d0_idx;
1561                 do {
1562                         ptrs[count++] = page_address(sh->dev[i].page);
1563                         if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1564                                 printk("block %d/%d not uptodate on parity calc\n", i,count);
1565                         i = raid6_next_disk(i, disks);
1566                 } while ( i != d0_idx );
1567 //              break;
1568 //      }
1569
1570         raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1571
1572         switch(method) {
1573         case RECONSTRUCT_WRITE:
1574                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1575                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1576                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1577                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
1578                 break;
1579         case UPDATE_PARITY:
1580                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1581                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1582                 break;
1583         }
1584 }
1585
1586
1587 /* Compute one missing block */
1588 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1589 {
1590         int i, count, disks = sh->disks;
1591         void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1592         int pd_idx = sh->pd_idx;
1593         int qd_idx = raid6_next_disk(pd_idx, disks);
1594
1595         pr_debug("compute_block_1, stripe %llu, idx %d\n",
1596                 (unsigned long long)sh->sector, dd_idx);
1597
1598         if ( dd_idx == qd_idx ) {
1599                 /* We're actually computing the Q drive */
1600                 compute_parity6(sh, UPDATE_PARITY);
1601         } else {
1602                 dest = page_address(sh->dev[dd_idx].page);
1603                 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1604                 count = 0;
1605                 for (i = disks ; i--; ) {
1606                         if (i == dd_idx || i == qd_idx)
1607                                 continue;
1608                         p = page_address(sh->dev[i].page);
1609                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1610                                 ptr[count++] = p;
1611                         else
1612                                 printk("compute_block() %d, stripe %llu, %d"
1613                                        " not present\n", dd_idx,
1614                                        (unsigned long long)sh->sector, i);
1615
1616                         check_xor();
1617                 }
1618                 if (count)
1619                         xor_blocks(count, STRIPE_SIZE, dest, ptr);
1620                 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1621                 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1622         }
1623 }
1624
1625 /* Compute two missing blocks */
1626 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1627 {
1628         int i, count, disks = sh->disks;
1629         int pd_idx = sh->pd_idx;
1630         int qd_idx = raid6_next_disk(pd_idx, disks);
1631         int d0_idx = raid6_next_disk(qd_idx, disks);
1632         int faila, failb;
1633
1634         /* faila and failb are disk numbers relative to d0_idx */
1635         /* pd_idx become disks-2 and qd_idx become disks-1 */
1636         faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1637         failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1638
1639         BUG_ON(faila == failb);
1640         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1641
1642         pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1643                (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1644
1645         if ( failb == disks-1 ) {
1646                 /* Q disk is one of the missing disks */
1647                 if ( faila == disks-2 ) {
1648                         /* Missing P+Q, just recompute */
1649                         compute_parity6(sh, UPDATE_PARITY);
1650                         return;
1651                 } else {
1652                         /* We're missing D+Q; recompute D from P */
1653                         compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1654                         compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1655                         return;
1656                 }
1657         }
1658
1659         /* We're missing D+P or D+D; build pointer table */
1660         {
1661                 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1662                 void *ptrs[disks];
1663
1664                 count = 0;
1665                 i = d0_idx;
1666                 do {
1667                         ptrs[count++] = page_address(sh->dev[i].page);
1668                         i = raid6_next_disk(i, disks);
1669                         if (i != dd_idx1 && i != dd_idx2 &&
1670                             !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1671                                 printk("compute_2 with missing block %d/%d\n", count, i);
1672                 } while ( i != d0_idx );
1673
1674                 if ( failb == disks-2 ) {
1675                         /* We're missing D+P. */
1676                         raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1677                 } else {
1678                         /* We're missing D+D. */
1679                         raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1680                 }
1681
1682                 /* Both the above update both missing blocks */
1683                 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1684                 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1685         }
1686 }
1687
1688 static int
1689 handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
1690 {
1691         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1692         int locked = 0;
1693
1694         if (rcw) {
1695                 /* if we are not expanding this is a proper write request, and
1696                  * there will be bios with new data to be drained into the
1697                  * stripe cache
1698                  */
1699                 if (!expand) {
1700                         set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1701                         sh->ops.count++;
1702                 }
1703
1704                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1705                 sh->ops.count++;
1706
1707                 for (i = disks; i--; ) {
1708                         struct r5dev *dev = &sh->dev[i];
1709
1710                         if (dev->towrite) {
1711                                 set_bit(R5_LOCKED, &dev->flags);
1712                                 if (!expand)
1713                                         clear_bit(R5_UPTODATE, &dev->flags);
1714                                 locked++;
1715                         }
1716                 }
1717         } else {
1718                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1719                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1720
1721                 set_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
1722                 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1723                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1724
1725                 sh->ops.count += 3;
1726
1727                 for (i = disks; i--; ) {
1728                         struct r5dev *dev = &sh->dev[i];
1729                         if (i == pd_idx)
1730                                 continue;
1731
1732                         /* For a read-modify write there may be blocks that are
1733                          * locked for reading while others are ready to be
1734                          * written so we distinguish these blocks by the
1735                          * R5_Wantprexor bit
1736                          */
1737                         if (dev->towrite &&
1738                             (test_bit(R5_UPTODATE, &dev->flags) ||
1739                             test_bit(R5_Wantcompute, &dev->flags))) {
1740                                 set_bit(R5_Wantprexor, &dev->flags);
1741                                 set_bit(R5_LOCKED, &dev->flags);
1742                                 clear_bit(R5_UPTODATE, &dev->flags);
1743                                 locked++;
1744                         }
1745                 }
1746         }
1747
1748         /* keep the parity disk locked while asynchronous operations
1749          * are in flight
1750          */
1751         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1752         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1753         locked++;
1754
1755         pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1756                 __FUNCTION__, (unsigned long long)sh->sector,
1757                 locked, sh->ops.pending);
1758
1759         return locked;
1760 }
1761
1762 /*
1763  * Each stripe/dev can have one or more bion attached.
1764  * toread/towrite point to the first in a chain.
1765  * The bi_next chain must be in order.
1766  */
1767 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1768 {
1769         struct bio **bip;
1770         raid5_conf_t *conf = sh->raid_conf;
1771         int firstwrite=0;
1772
1773         pr_debug("adding bh b#%llu to stripe s#%llu\n",
1774                 (unsigned long long)bi->bi_sector,
1775                 (unsigned long long)sh->sector);
1776
1777
1778         spin_lock(&sh->lock);
1779         spin_lock_irq(&conf->device_lock);
1780         if (forwrite) {
1781                 bip = &sh->dev[dd_idx].towrite;
1782                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1783                         firstwrite = 1;
1784         } else
1785                 bip = &sh->dev[dd_idx].toread;
1786         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1787                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1788                         goto overlap;
1789                 bip = & (*bip)->bi_next;
1790         }
1791         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1792                 goto overlap;
1793
1794         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1795         if (*bip)
1796                 bi->bi_next = *bip;
1797         *bip = bi;
1798         bi->bi_phys_segments ++;
1799         spin_unlock_irq(&conf->device_lock);
1800         spin_unlock(&sh->lock);
1801
1802         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1803                 (unsigned long long)bi->bi_sector,
1804                 (unsigned long long)sh->sector, dd_idx);
1805
1806         if (conf->mddev->bitmap && firstwrite) {
1807                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1808                                   STRIPE_SECTORS, 0);
1809                 sh->bm_seq = conf->seq_flush+1;
1810                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1811         }
1812
1813         if (forwrite) {
1814                 /* check if page is covered */
1815                 sector_t sector = sh->dev[dd_idx].sector;
1816                 for (bi=sh->dev[dd_idx].towrite;
1817                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1818                              bi && bi->bi_sector <= sector;
1819                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1820                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1821                                 sector = bi->bi_sector + (bi->bi_size>>9);
1822                 }
1823                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1824                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1825         }
1826         return 1;
1827
1828  overlap:
1829         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1830         spin_unlock_irq(&conf->device_lock);
1831         spin_unlock(&sh->lock);
1832         return 0;
1833 }
1834
1835 static void end_reshape(raid5_conf_t *conf);
1836
1837 static int page_is_zero(struct page *p)
1838 {
1839         char *a = page_address(p);
1840         return ((*(u32*)a) == 0 &&
1841                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1842 }
1843
1844 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1845 {
1846         int sectors_per_chunk = conf->chunk_size >> 9;
1847         int pd_idx, dd_idx;
1848         int chunk_offset = sector_div(stripe, sectors_per_chunk);
1849
1850         raid5_compute_sector(stripe * (disks - conf->max_degraded)
1851                              *sectors_per_chunk + chunk_offset,
1852                              disks, disks - conf->max_degraded,
1853                              &dd_idx, &pd_idx, conf);
1854         return pd_idx;
1855 }
1856
1857 static void
1858 handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1859                                 struct stripe_head_state *s, int disks,
1860                                 struct bio **return_bi)
1861 {
1862         int i;
1863         for (i = disks; i--; ) {
1864                 struct bio *bi;
1865                 int bitmap_end = 0;
1866
1867                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1868                         mdk_rdev_t *rdev;
1869                         rcu_read_lock();
1870                         rdev = rcu_dereference(conf->disks[i].rdev);
1871                         if (rdev && test_bit(In_sync, &rdev->flags))
1872                                 /* multiple read failures in one stripe */
1873                                 md_error(conf->mddev, rdev);
1874                         rcu_read_unlock();
1875                 }
1876                 spin_lock_irq(&conf->device_lock);
1877                 /* fail all writes first */
1878                 bi = sh->dev[i].towrite;
1879                 sh->dev[i].towrite = NULL;
1880                 if (bi) {
1881                         s->to_write--;
1882                         bitmap_end = 1;
1883                 }
1884
1885                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1886                         wake_up(&conf->wait_for_overlap);
1887
1888                 while (bi && bi->bi_sector <
1889                         sh->dev[i].sector + STRIPE_SECTORS) {
1890                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1891                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1892                         if (--bi->bi_phys_segments == 0) {
1893                                 md_write_end(conf->mddev);
1894                                 bi->bi_next = *return_bi;
1895                                 *return_bi = bi;
1896                         }
1897                         bi = nextbi;
1898                 }
1899                 /* and fail all 'written' */
1900                 bi = sh->dev[i].written;
1901                 sh->dev[i].written = NULL;
1902                 if (bi) bitmap_end = 1;
1903                 while (bi && bi->bi_sector <
1904                        sh->dev[i].sector + STRIPE_SECTORS) {
1905                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1906                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1907                         if (--bi->bi_phys_segments == 0) {
1908                                 md_write_end(conf->mddev);
1909                                 bi->bi_next = *return_bi;
1910                                 *return_bi = bi;
1911                         }
1912                         bi = bi2;
1913                 }
1914
1915                 /* fail any reads if this device is non-operational and
1916                  * the data has not reached the cache yet.
1917                  */
1918                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
1919                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1920                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
1921                         bi = sh->dev[i].toread;
1922                         sh->dev[i].toread = NULL;
1923                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1924                                 wake_up(&conf->wait_for_overlap);
1925                         if (bi) s->to_read--;
1926                         while (bi && bi->bi_sector <
1927                                sh->dev[i].sector + STRIPE_SECTORS) {
1928                                 struct bio *nextbi =
1929                                         r5_next_bio(bi, sh->dev[i].sector);
1930                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1931                                 if (--bi->bi_phys_segments == 0) {
1932                                         bi->bi_next = *return_bi;
1933                                         *return_bi = bi;
1934                                 }
1935                                 bi = nextbi;
1936                         }
1937                 }
1938                 spin_unlock_irq(&conf->device_lock);
1939                 if (bitmap_end)
1940                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1941                                         STRIPE_SECTORS, 0, 0);
1942         }
1943
1944 }
1945
1946 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
1947  * to process
1948  */
1949 static int __handle_issuing_new_read_requests5(struct stripe_head *sh,
1950                         struct stripe_head_state *s, int disk_idx, int disks)
1951 {
1952         struct r5dev *dev = &sh->dev[disk_idx];
1953         struct r5dev *failed_dev = &sh->dev[s->failed_num];
1954
1955         /* don't schedule compute operations or reads on the parity block while
1956          * a check is in flight
1957          */
1958         if ((disk_idx == sh->pd_idx) &&
1959              test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
1960                 return ~0;
1961
1962         /* is the data in this block needed, and can we get it? */
1963         if (!test_bit(R5_LOCKED, &dev->flags) &&
1964             !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread ||
1965             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1966              s->syncing || s->expanding || (s->failed &&
1967              (failed_dev->toread || (failed_dev->towrite &&
1968              !test_bit(R5_OVERWRITE, &failed_dev->flags)
1969              ))))) {
1970                 /* 1/ We would like to get this block, possibly by computing it,
1971                  * but we might not be able to.
1972                  *
1973                  * 2/ Since parity check operations potentially make the parity
1974                  * block !uptodate it will need to be refreshed before any
1975                  * compute operations on data disks are scheduled.
1976                  *
1977                  * 3/ We hold off parity block re-reads until check operations
1978                  * have quiesced.
1979                  */
1980                 if ((s->uptodate == disks - 1) &&
1981                     !test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
1982                         set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
1983                         set_bit(R5_Wantcompute, &dev->flags);
1984                         sh->ops.target = disk_idx;
1985                         s->req_compute = 1;
1986                         sh->ops.count++;
1987                         /* Careful: from this point on 'uptodate' is in the eye
1988                          * of raid5_run_ops which services 'compute' operations
1989                          * before writes. R5_Wantcompute flags a block that will
1990                          * be R5_UPTODATE by the time it is needed for a
1991                          * subsequent operation.
1992                          */
1993                         s->uptodate++;
1994                         return 0; /* uptodate + compute == disks */
1995                 } else if ((s->uptodate < disks - 1) &&
1996                         test_bit(R5_Insync, &dev->flags)) {
1997                         /* Note: we hold off compute operations while checks are
1998                          * in flight, but we still prefer 'compute' over 'read'
1999                          * hence we only read if (uptodate < * disks-1)
2000                          */
2001                         set_bit(R5_LOCKED, &dev->flags);
2002                         set_bit(R5_Wantread, &dev->flags);
2003                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2004                                 sh->ops.count++;
2005                         s->locked++;
2006                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2007                                 s->syncing);
2008                 }
2009         }
2010
2011         return ~0;
2012 }
2013
2014 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2015                         struct stripe_head_state *s, int disks)
2016 {
2017         int i;
2018
2019         /* Clear completed compute operations.  Parity recovery
2020          * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled
2021          * later on in this routine
2022          */
2023         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2024                 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2025                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2026                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2027                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2028         }
2029
2030         /* look for blocks to read/compute, skip this if a compute
2031          * is already in flight, or if the stripe contents are in the
2032          * midst of changing due to a write
2033          */
2034         if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2035                 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2036                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2037                 for (i = disks; i--; )
2038                         if (__handle_issuing_new_read_requests5(
2039                                 sh, s, i, disks) == 0)
2040                                 break;
2041         }
2042         set_bit(STRIPE_HANDLE, &sh->state);
2043 }
2044
2045 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2046                         struct stripe_head_state *s, struct r6_state *r6s,
2047                         int disks)
2048 {
2049         int i;
2050         for (i = disks; i--; ) {
2051                 struct r5dev *dev = &sh->dev[i];
2052                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2053                     !test_bit(R5_UPTODATE, &dev->flags) &&
2054                     (dev->toread || (dev->towrite &&
2055                      !test_bit(R5_OVERWRITE, &dev->flags)) ||
2056                      s->syncing || s->expanding ||
2057                      (s->failed >= 1 &&
2058                       (sh->dev[r6s->failed_num[0]].toread ||
2059                        s->to_write)) ||
2060                      (s->failed >= 2 &&
2061                       (sh->dev[r6s->failed_num[1]].toread ||
2062                        s->to_write)))) {
2063                         /* we would like to get this block, possibly
2064                          * by computing it, but we might not be able to
2065                          */
2066                         if (s->uptodate == disks-1) {
2067                                 pr_debug("Computing stripe %llu block %d\n",
2068                                        (unsigned long long)sh->sector, i);
2069                                 compute_block_1(sh, i, 0);
2070                                 s->uptodate++;
2071                         } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2072                                 /* Computing 2-failure is *very* expensive; only
2073                                  * do it if failed >= 2
2074                                  */
2075                                 int other;
2076                                 for (other = disks; other--; ) {
2077                                         if (other == i)
2078                                                 continue;
2079                                         if (!test_bit(R5_UPTODATE,
2080                                               &sh->dev[other].flags))
2081                                                 break;
2082                                 }
2083                                 BUG_ON(other < 0);
2084                                 pr_debug("Computing stripe %llu blocks %d,%d\n",
2085                                        (unsigned long long)sh->sector,
2086                                        i, other);
2087                                 compute_block_2(sh, i, other);
2088                                 s->uptodate += 2;
2089                         } else if (test_bit(R5_Insync, &dev->flags)) {
2090                                 set_bit(R5_LOCKED, &dev->flags);
2091                                 set_bit(R5_Wantread, &dev->flags);
2092                                 s->locked++;
2093                                 pr_debug("Reading block %d (sync=%d)\n",
2094                                         i, s->syncing);
2095                         }
2096                 }
2097         }
2098         set_bit(STRIPE_HANDLE, &sh->state);
2099 }
2100
2101
2102 /* handle_completed_write_requests
2103  * any written block on an uptodate or failed drive can be returned.
2104  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2105  * never LOCKED, so we don't need to test 'failed' directly.
2106  */
2107 static void handle_completed_write_requests(raid5_conf_t *conf,
2108         struct stripe_head *sh, int disks, struct bio **return_bi)
2109 {
2110         int i;
2111         struct r5dev *dev;
2112
2113         for (i = disks; i--; )
2114                 if (sh->dev[i].written) {
2115                         dev = &sh->dev[i];
2116                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2117                                 test_bit(R5_UPTODATE, &dev->flags)) {
2118                                 /* We can return any write requests */
2119                                 struct bio *wbi, *wbi2;
2120                                 int bitmap_end = 0;
2121                                 pr_debug("Return write for disc %d\n", i);
2122                                 spin_lock_irq(&conf->device_lock);
2123                                 wbi = dev->written;
2124                                 dev->written = NULL;
2125                                 while (wbi && wbi->bi_sector <
2126                                         dev->sector + STRIPE_SECTORS) {
2127                                         wbi2 = r5_next_bio(wbi, dev->sector);
2128                                         if (--wbi->bi_phys_segments == 0) {
2129                                                 md_write_end(conf->mddev);
2130                                                 wbi->bi_next = *return_bi;
2131                                                 *return_bi = wbi;
2132                                         }
2133                                         wbi = wbi2;
2134                                 }
2135                                 if (dev->towrite == NULL)
2136                                         bitmap_end = 1;
2137                                 spin_unlock_irq(&conf->device_lock);
2138                                 if (bitmap_end)
2139                                         bitmap_endwrite(conf->mddev->bitmap,
2140                                                         sh->sector,
2141                                                         STRIPE_SECTORS,
2142                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2143                                                         0);
2144                         }
2145                 }
2146 }
2147
2148 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2149                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2150 {
2151         int rmw = 0, rcw = 0, i;
2152         for (i = disks; i--; ) {
2153                 /* would I have to read this buffer for read_modify_write */
2154                 struct r5dev *dev = &sh->dev[i];
2155                 if ((dev->towrite || i == sh->pd_idx) &&
2156                     !test_bit(R5_LOCKED, &dev->flags) &&
2157                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2158                       test_bit(R5_Wantcompute, &dev->flags))) {
2159                         if (test_bit(R5_Insync, &dev->flags))
2160                                 rmw++;
2161                         else
2162                                 rmw += 2*disks;  /* cannot read it */
2163                 }
2164                 /* Would I have to read this buffer for reconstruct_write */
2165                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2166                     !test_bit(R5_LOCKED, &dev->flags) &&
2167                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2168                     test_bit(R5_Wantcompute, &dev->flags))) {
2169                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2170                         else
2171                                 rcw += 2*disks;
2172                 }
2173         }
2174         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2175                 (unsigned long long)sh->sector, rmw, rcw);
2176         set_bit(STRIPE_HANDLE, &sh->state);
2177         if (rmw < rcw && rmw > 0)
2178                 /* prefer read-modify-write, but need to get some data */
2179                 for (i = disks; i--; ) {
2180                         struct r5dev *dev = &sh->dev[i];
2181                         if ((dev->towrite || i == sh->pd_idx) &&
2182                             !test_bit(R5_LOCKED, &dev->flags) &&
2183                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2184                             test_bit(R5_Wantcompute, &dev->flags)) &&
2185                             test_bit(R5_Insync, &dev->flags)) {
2186                                 if (
2187                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2188                                         pr_debug("Read_old block "
2189                                                 "%d for r-m-w\n", i);
2190                                         set_bit(R5_LOCKED, &dev->flags);
2191                                         set_bit(R5_Wantread, &dev->flags);
2192                                         if (!test_and_set_bit(
2193                                                 STRIPE_OP_IO, &sh->ops.pending))
2194                                                 sh->ops.count++;
2195                                         s->locked++;
2196                                 } else {
2197                                         set_bit(STRIPE_DELAYED, &sh->state);
2198                                         set_bit(STRIPE_HANDLE, &sh->state);
2199                                 }
2200                         }
2201                 }
2202         if (rcw <= rmw && rcw > 0)
2203                 /* want reconstruct write, but need to get some data */
2204                 for (i = disks; i--; ) {
2205                         struct r5dev *dev = &sh->dev[i];
2206                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2207                             i != sh->pd_idx &&
2208                             !test_bit(R5_LOCKED, &dev->flags) &&
2209                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2210                             test_bit(R5_Wantcompute, &dev->flags)) &&
2211                             test_bit(R5_Insync, &dev->flags)) {
2212                                 if (
2213                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2214                                         pr_debug("Read_old block "
2215                                                 "%d for Reconstruct\n", i);
2216                                         set_bit(R5_LOCKED, &dev->flags);
2217                                         set_bit(R5_Wantread, &dev->flags);
2218                                         if (!test_and_set_bit(
2219                                                 STRIPE_OP_IO, &sh->ops.pending))
2220                                                 sh->ops.count++;
2221                                         s->locked++;
2222                                 } else {
2223                                         set_bit(STRIPE_DELAYED, &sh->state);
2224                                         set_bit(STRIPE_HANDLE, &sh->state);
2225                                 }
2226                         }
2227                 }
2228         /* now if nothing is locked, and if we have enough data,
2229          * we can start a write request
2230          */
2231         /* since handle_stripe can be called at any time we need to handle the
2232          * case where a compute block operation has been submitted and then a
2233          * subsequent call wants to start a write request.  raid5_run_ops only
2234          * handles the case where compute block and postxor are requested
2235          * simultaneously.  If this is not the case then new writes need to be
2236          * held off until the compute completes.
2237          */
2238         if ((s->req_compute ||
2239             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) &&
2240                 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2241                 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2242                 s->locked += handle_write_operations5(sh, rcw == 0, 0);
2243 }
2244
2245 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2246                 struct stripe_head *sh, struct stripe_head_state *s,
2247                 struct r6_state *r6s, int disks)
2248 {
2249         int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2250         int qd_idx = r6s->qd_idx;
2251         for (i = disks; i--; ) {
2252                 struct r5dev *dev = &sh->dev[i];
2253                 /* Would I have to read this buffer for reconstruct_write */
2254                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2255                     && i != pd_idx && i != qd_idx
2256                     && (!test_bit(R5_LOCKED, &dev->flags)
2257                             ) &&
2258                     !test_bit(R5_UPTODATE, &dev->flags)) {
2259                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2260                         else {
2261                                 pr_debug("raid6: must_compute: "
2262                                         "disk %d flags=%#lx\n", i, dev->flags);
2263                                 must_compute++;
2264                         }
2265                 }
2266         }
2267         pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2268                (unsigned long long)sh->sector, rcw, must_compute);
2269         set_bit(STRIPE_HANDLE, &sh->state);
2270
2271         if (rcw > 0)
2272                 /* want reconstruct write, but need to get some data */
2273                 for (i = disks; i--; ) {
2274                         struct r5dev *dev = &sh->dev[i];
2275                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2276                             && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2277                             && !test_bit(R5_LOCKED, &dev->flags) &&
2278                             !test_bit(R5_UPTODATE, &dev->flags) &&
2279                             test_bit(R5_Insync, &dev->flags)) {
2280                                 if (
2281                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2282                                         pr_debug("Read_old stripe %llu "
2283                                                 "block %d for Reconstruct\n",
2284                                              (unsigned long long)sh->sector, i);
2285                                         set_bit(R5_LOCKED, &dev->flags);
2286                                         set_bit(R5_Wantread, &dev->flags);
2287                                         s->locked++;
2288                                 } else {
2289                                         pr_debug("Request delayed stripe %llu "
2290                                                 "block %d for Reconstruct\n",
2291                                              (unsigned long long)sh->sector, i);
2292                                         set_bit(STRIPE_DELAYED, &sh->state);
2293                                         set_bit(STRIPE_HANDLE, &sh->state);
2294                                 }
2295                         }
2296                 }
2297         /* now if nothing is locked, and if we have enough data, we can start a
2298          * write request
2299          */
2300         if (s->locked == 0 && rcw == 0 &&
2301             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2302                 if (must_compute > 0) {
2303                         /* We have failed blocks and need to compute them */
2304                         switch (s->failed) {
2305                         case 0:
2306                                 BUG();
2307                         case 1:
2308                                 compute_block_1(sh, r6s->failed_num[0], 0);
2309                                 break;
2310                         case 2:
2311                                 compute_block_2(sh, r6s->failed_num[0],
2312                                                 r6s->failed_num[1]);
2313                                 break;
2314                         default: /* This request should have been failed? */
2315                                 BUG();
2316                         }
2317                 }
2318
2319                 pr_debug("Computing parity for stripe %llu\n",
2320                         (unsigned long long)sh->sector);
2321                 compute_parity6(sh, RECONSTRUCT_WRITE);
2322                 /* now every locked buffer is ready to be written */
2323                 for (i = disks; i--; )
2324                         if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2325                                 pr_debug("Writing stripe %llu block %d\n",
2326                                        (unsigned long long)sh->sector, i);
2327                                 s->locked++;
2328                                 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2329                         }
2330                 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2331                 set_bit(STRIPE_INSYNC, &sh->state);
2332
2333                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2334                         atomic_dec(&conf->preread_active_stripes);
2335                         if (atomic_read(&conf->preread_active_stripes) <
2336                             IO_THRESHOLD)
2337                                 md_wakeup_thread(conf->mddev->thread);
2338                 }
2339         }
2340 }
2341
2342 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2343                                 struct stripe_head_state *s, int disks)
2344 {
2345         set_bit(STRIPE_HANDLE, &sh->state);
2346         /* Take one of the following actions:
2347          * 1/ start a check parity operation if (uptodate == disks)
2348          * 2/ finish a check parity operation and act on the result
2349          * 3/ skip to the writeback section if we previously
2350          *    initiated a recovery operation
2351          */
2352         if (s->failed == 0 &&
2353             !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2354                 if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2355                         BUG_ON(s->uptodate != disks);
2356                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2357                         sh->ops.count++;
2358                         s->uptodate--;
2359                 } else if (
2360                        test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) {
2361                         clear_bit(STRIPE_OP_CHECK, &sh->ops.ack);
2362                         clear_bit(STRIPE_OP_CHECK, &sh->ops.pending);
2363
2364                         if (sh->ops.zero_sum_result == 0)
2365                                 /* parity is correct (on disc,
2366                                  * not in buffer any more)
2367                                  */
2368                                 set_bit(STRIPE_INSYNC, &sh->state);
2369                         else {
2370                                 conf->mddev->resync_mismatches +=
2371                                         STRIPE_SECTORS;
2372                                 if (test_bit(
2373                                      MD_RECOVERY_CHECK, &conf->mddev->recovery))
2374                                         /* don't try to repair!! */
2375                                         set_bit(STRIPE_INSYNC, &sh->state);
2376                                 else {
2377                                         set_bit(STRIPE_OP_COMPUTE_BLK,
2378                                                 &sh->ops.pending);
2379                                         set_bit(STRIPE_OP_MOD_REPAIR_PD,
2380                                                 &sh->ops.pending);
2381                                         set_bit(R5_Wantcompute,
2382                                                 &sh->dev[sh->pd_idx].flags);
2383                                         sh->ops.target = sh->pd_idx;
2384                                         sh->ops.count++;
2385                                         s->uptodate++;
2386                                 }
2387                         }
2388                 }
2389         }
2390
2391         /* check if we can clear a parity disk reconstruct */
2392         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2393                 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2394
2395                 clear_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending);
2396                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2397                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2398                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2399         }
2400
2401         /* Wait for check parity and compute block operations to complete
2402          * before write-back
2403          */
2404         if (!test_bit(STRIPE_INSYNC, &sh->state) &&
2405                 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) &&
2406                 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) {
2407                 struct r5dev *dev;
2408                 /* either failed parity check, or recovery is happening */
2409                 if (s->failed == 0)
2410                         s->failed_num = sh->pd_idx;
2411                 dev = &sh->dev[s->failed_num];
2412                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2413                 BUG_ON(s->uptodate != disks);
2414
2415                 set_bit(R5_LOCKED, &dev->flags);
2416                 set_bit(R5_Wantwrite, &dev->flags);
2417                 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2418                         sh->ops.count++;
2419
2420                 clear_bit(STRIPE_DEGRADED, &sh->state);
2421                 s->locked++;
2422                 set_bit(STRIPE_INSYNC, &sh->state);
2423         }
2424 }
2425
2426
2427 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2428                                 struct stripe_head_state *s,
2429                                 struct r6_state *r6s, struct page *tmp_page,
2430                                 int disks)
2431 {
2432         int update_p = 0, update_q = 0;
2433         struct r5dev *dev;
2434         int pd_idx = sh->pd_idx;
2435         int qd_idx = r6s->qd_idx;
2436
2437         set_bit(STRIPE_HANDLE, &sh->state);
2438
2439         BUG_ON(s->failed > 2);
2440         BUG_ON(s->uptodate < disks);
2441         /* Want to check and possibly repair P and Q.
2442          * However there could be one 'failed' device, in which
2443          * case we can only check one of them, possibly using the
2444          * other to generate missing data
2445          */
2446
2447         /* If !tmp_page, we cannot do the calculations,
2448          * but as we have set STRIPE_HANDLE, we will soon be called
2449          * by stripe_handle with a tmp_page - just wait until then.
2450          */
2451         if (tmp_page) {
2452                 if (s->failed == r6s->q_failed) {
2453                         /* The only possible failed device holds 'Q', so it
2454                          * makes sense to check P (If anything else were failed,
2455                          * we would have used P to recreate it).
2456                          */
2457                         compute_block_1(sh, pd_idx, 1);
2458                         if (!page_is_zero(sh->dev[pd_idx].page)) {
2459                                 compute_block_1(sh, pd_idx, 0);
2460                                 update_p = 1;
2461                         }
2462                 }
2463                 if (!r6s->q_failed && s->failed < 2) {
2464                         /* q is not failed, and we didn't use it to generate
2465                          * anything, so it makes sense to check it
2466                          */
2467                         memcpy(page_address(tmp_page),
2468                                page_address(sh->dev[qd_idx].page),
2469                                STRIPE_SIZE);
2470                         compute_parity6(sh, UPDATE_PARITY);
2471                         if (memcmp(page_address(tmp_page),
2472                                    page_address(sh->dev[qd_idx].page),
2473                                    STRIPE_SIZE) != 0) {
2474                                 clear_bit(STRIPE_INSYNC, &sh->state);
2475                                 update_q = 1;
2476                         }
2477                 }
2478                 if (update_p || update_q) {
2479                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2480                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2481                                 /* don't try to repair!! */
2482                                 update_p = update_q = 0;
2483                 }
2484
2485                 /* now write out any block on a failed drive,
2486                  * or P or Q if they need it
2487                  */
2488
2489                 if (s->failed == 2) {
2490                         dev = &sh->dev[r6s->failed_num[1]];
2491                         s->locked++;
2492                         set_bit(R5_LOCKED, &dev->flags);
2493                         set_bit(R5_Wantwrite, &dev->flags);
2494                 }
2495                 if (s->failed >= 1) {
2496                         dev = &sh->dev[r6s->failed_num[0]];
2497                         s->locked++;
2498                         set_bit(R5_LOCKED, &dev->flags);
2499                         set_bit(R5_Wantwrite, &dev->flags);
2500                 }
2501
2502                 if (update_p) {
2503                         dev = &sh->dev[pd_idx];
2504                         s->locked++;
2505                         set_bit(R5_LOCKED, &dev->flags);
2506                         set_bit(R5_Wantwrite, &dev->flags);
2507                 }
2508                 if (update_q) {
2509                         dev = &sh->dev[qd_idx];
2510                         s->locked++;
2511                         set_bit(R5_LOCKED, &dev->flags);
2512                         set_bit(R5_Wantwrite, &dev->flags);
2513                 }
2514                 clear_bit(STRIPE_DEGRADED, &sh->state);
2515
2516                 set_bit(STRIPE_INSYNC, &sh->state);
2517         }
2518 }
2519
2520 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2521                                 struct r6_state *r6s)
2522 {
2523         int i;
2524
2525         /* We have read all the blocks in this stripe and now we need to
2526          * copy some of them into a target stripe for expand.
2527          */
2528         struct dma_async_tx_descriptor *tx = NULL;
2529         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2530         for (i = 0; i < sh->disks; i++)
2531                 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2532                         int dd_idx, pd_idx, j;
2533                         struct stripe_head *sh2;
2534
2535                         sector_t bn = compute_blocknr(sh, i);
2536                         sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2537                                                 conf->raid_disks -
2538                                                 conf->max_degraded, &dd_idx,
2539                                                 &pd_idx, conf);
2540                         sh2 = get_active_stripe(conf, s, conf->raid_disks,
2541                                                 pd_idx, 1);
2542                         if (sh2 == NULL)
2543                                 /* so far only the early blocks of this stripe
2544                                  * have been requested.  When later blocks
2545                                  * get requested, we will try again
2546                                  */
2547                                 continue;
2548                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2549                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2550                                 /* must have already done this block */
2551                                 release_stripe(sh2);
2552                                 continue;
2553                         }
2554
2555                         /* place all the copies on one channel */
2556                         tx = async_memcpy(sh2->dev[dd_idx].page,
2557                                 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2558                                 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2559
2560                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2561                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2562                         for (j = 0; j < conf->raid_disks; j++)
2563                                 if (j != sh2->pd_idx &&
2564                                     (!r6s || j != raid6_next_disk(sh2->pd_idx,
2565                                                                  sh2->disks)) &&
2566                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2567                                         break;
2568                         if (j == conf->raid_disks) {
2569                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2570                                 set_bit(STRIPE_HANDLE, &sh2->state);
2571                         }
2572                         release_stripe(sh2);
2573
2574                 }
2575         /* done submitting copies, wait for them to complete */
2576         if (tx) {
2577                 async_tx_ack(tx);
2578                 dma_wait_for_async_tx(tx);
2579         }
2580 }
2581
2582 /*
2583  * handle_stripe - do things to a stripe.
2584  *
2585  * We lock the stripe and then examine the state of various bits
2586  * to see what needs to be done.
2587  * Possible results:
2588  *    return some read request which now have data
2589  *    return some write requests which are safely on disc
2590  *    schedule a read on some buffers
2591  *    schedule a write of some buffers
2592  *    return confirmation of parity correctness
2593  *
2594  * buffers are taken off read_list or write_list, and bh_cache buffers
2595  * get BH_Lock set before the stripe lock is released.
2596  *
2597  */
2598
2599 static void handle_stripe5(struct stripe_head *sh)
2600 {
2601         raid5_conf_t *conf = sh->raid_conf;
2602         int disks = sh->disks, i;
2603         struct bio *return_bi = NULL;
2604         struct stripe_head_state s;
2605         struct r5dev *dev;
2606         unsigned long pending = 0;
2607
2608         memset(&s, 0, sizeof(s));
2609         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2610                 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2611                 atomic_read(&sh->count), sh->pd_idx,
2612                 sh->ops.pending, sh->ops.ack, sh->ops.complete);
2613
2614         spin_lock(&sh->lock);
2615         clear_bit(STRIPE_HANDLE, &sh->state);
2616         clear_bit(STRIPE_DELAYED, &sh->state);
2617
2618         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2619         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2620         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2621         /* Now to look around and see what can be done */
2622
2623         rcu_read_lock();
2624         for (i=disks; i--; ) {
2625                 mdk_rdev_t *rdev;
2626                 struct r5dev *dev = &sh->dev[i];
2627                 clear_bit(R5_Insync, &dev->flags);
2628
2629                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2630                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2631                         dev->towrite, dev->written);
2632
2633                 /* maybe we can request a biofill operation
2634                  *
2635                  * new wantfill requests are only permitted while
2636                  * STRIPE_OP_BIOFILL is clear
2637                  */
2638                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2639                         !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2640                         set_bit(R5_Wantfill, &dev->flags);
2641
2642                 /* now count some things */
2643                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2644                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2645                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2646
2647                 if (test_bit(R5_Wantfill, &dev->flags))
2648                         s.to_fill++;
2649                 else if (dev->toread)
2650                         s.to_read++;
2651                 if (dev->towrite) {
2652                         s.to_write++;
2653                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2654                                 s.non_overwrite++;
2655                 }
2656                 if (dev->written)
2657                         s.written++;
2658                 rdev = rcu_dereference(conf->disks[i].rdev);
2659                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2660                         /* The ReadError flag will just be confusing now */
2661                         clear_bit(R5_ReadError, &dev->flags);
2662                         clear_bit(R5_ReWrite, &dev->flags);
2663                 }
2664                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2665                     || test_bit(R5_ReadError, &dev->flags)) {
2666                         s.failed++;
2667                         s.failed_num = i;
2668                 } else
2669                         set_bit(R5_Insync, &dev->flags);
2670         }
2671         rcu_read_unlock();
2672
2673         if (s.to_fill && !test_and_set_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2674                 sh->ops.count++;
2675
2676         pr_debug("locked=%d uptodate=%d to_read=%d"
2677                 " to_write=%d failed=%d failed_num=%d\n",
2678                 s.locked, s.uptodate, s.to_read, s.to_write,
2679                 s.failed, s.failed_num);
2680         /* check if the array has lost two devices and, if so, some requests might
2681          * need to be failed
2682          */
2683         if (s.failed > 1 && s.to_read+s.to_write+s.written)
2684                 handle_requests_to_failed_array(conf, sh, &s, disks,
2685                                                 &return_bi);
2686         if (s.failed > 1 && s.syncing) {
2687                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2688                 clear_bit(STRIPE_SYNCING, &sh->state);
2689                 s.syncing = 0;
2690         }
2691
2692         /* might be able to return some write requests if the parity block
2693          * is safe, or on a failed drive
2694          */
2695         dev = &sh->dev[sh->pd_idx];
2696         if ( s.written &&
2697              ((test_bit(R5_Insync, &dev->flags) &&
2698                !test_bit(R5_LOCKED, &dev->flags) &&
2699                test_bit(R5_UPTODATE, &dev->flags)) ||
2700                (s.failed == 1 && s.failed_num == sh->pd_idx)))
2701                 handle_completed_write_requests(conf, sh, disks, &return_bi);
2702
2703         /* Now we might consider reading some blocks, either to check/generate
2704          * parity, or to satisfy requests
2705          * or to load a block that is being partially written.
2706          */
2707         if (s.to_read || s.non_overwrite ||
2708             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding ||
2709             test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2710                 handle_issuing_new_read_requests5(sh, &s, disks);
2711
2712         /* Now we check to see if any write operations have recently
2713          * completed
2714          */
2715
2716         /* leave prexor set until postxor is done, allows us to distinguish
2717          * a rmw from a rcw during biodrain
2718          */
2719         if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2720                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2721
2722                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2723                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2724                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2725
2726                 for (i = disks; i--; )
2727                         clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2728         }
2729
2730         /* if only POSTXOR is set then this is an 'expand' postxor */
2731         if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2732                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2733
2734                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2735                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2736                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2737
2738                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2739                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2740                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2741
2742                 /* All the 'written' buffers and the parity block are ready to
2743                  * be written back to disk
2744                  */
2745                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2746                 for (i = disks; i--; ) {
2747                         dev = &sh->dev[i];
2748                         if (test_bit(R5_LOCKED, &dev->flags) &&
2749                                 (i == sh->pd_idx || dev->written)) {
2750                                 pr_debug("Writing block %d\n", i);
2751                                 set_bit(R5_Wantwrite, &dev->flags);
2752                                 if (!test_and_set_bit(
2753                                     STRIPE_OP_IO, &sh->ops.pending))
2754                                         sh->ops.count++;
2755                                 if (!test_bit(R5_Insync, &dev->flags) ||
2756                                     (i == sh->pd_idx && s.failed == 0))
2757                                         set_bit(STRIPE_INSYNC, &sh->state);
2758                         }
2759                 }
2760                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2761                         atomic_dec(&conf->preread_active_stripes);
2762                         if (atomic_read(&conf->preread_active_stripes) <
2763                                 IO_THRESHOLD)
2764                                 md_wakeup_thread(conf->mddev->thread);
2765                 }
2766         }
2767
2768         /* Now to consider new write requests and what else, if anything
2769          * should be read.  We do not handle new writes when:
2770          * 1/ A 'write' operation (copy+xor) is already in flight.
2771          * 2/ A 'check' operation is in flight, as it may clobber the parity
2772          *    block.
2773          */
2774         if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2775                           !test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2776                 handle_issuing_new_write_requests5(conf, sh, &s, disks);
2777
2778         /* maybe we need to check and possibly fix the parity for this stripe
2779          * Any reads will already have been scheduled, so we just see if enough
2780          * data is available.  The parity check is held off while parity
2781          * dependent operations are in flight.
2782          */
2783         if ((s.syncing && s.locked == 0 &&
2784              !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2785              !test_bit(STRIPE_INSYNC, &sh->state)) ||
2786               test_bit(STRIPE_OP_CHECK, &sh->ops.pending) ||
2787               test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending))
2788                 handle_parity_checks5(conf, sh, &s, disks);
2789
2790         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2791                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2792                 clear_bit(STRIPE_SYNCING, &sh->state);
2793         }
2794
2795         /* If the failed drive is just a ReadError, then we might need to progress
2796          * the repair/check process
2797          */
2798         if (s.failed == 1 && !conf->mddev->ro &&
2799             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2800             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2801             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2802                 ) {
2803                 dev = &sh->dev[s.failed_num];
2804                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2805                         set_bit(R5_Wantwrite, &dev->flags);
2806                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2807                                 sh->ops.count++;
2808                         set_bit(R5_ReWrite, &dev->flags);
2809                         set_bit(R5_LOCKED, &dev->flags);
2810                         s.locked++;
2811                 } else {
2812                         /* let's read it back */
2813                         set_bit(R5_Wantread, &dev->flags);
2814                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2815                                 sh->ops.count++;
2816                         set_bit(R5_LOCKED, &dev->flags);
2817                         s.locked++;
2818                 }
2819         }
2820
2821         /* Finish postxor operations initiated by the expansion
2822          * process
2823          */
2824         if (test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete) &&
2825                 !test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending)) {
2826
2827                 clear_bit(STRIPE_EXPANDING, &sh->state);
2828
2829                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2830                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2831                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2832
2833                 for (i = conf->raid_disks; i--; ) {
2834                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2835                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2836                                 sh->ops.count++;
2837                 }
2838         }
2839
2840         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2841                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2842                 /* Need to write out all blocks after computing parity */
2843                 sh->disks = conf->raid_disks;
2844                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2845                         conf->raid_disks);
2846                 s.locked += handle_write_operations5(sh, 1, 1);
2847         } else if (s.expanded &&
2848                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2849                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2850                 atomic_dec(&conf->reshape_stripes);
2851                 wake_up(&conf->wait_for_overlap);
2852                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2853         }
2854
2855         if (s.expanding && s.locked == 0)
2856                 handle_stripe_expansion(conf, sh, NULL);
2857
2858         if (sh->ops.count)
2859                 pending = get_stripe_work(sh);
2860
2861         spin_unlock(&sh->lock);
2862
2863         if (pending)
2864                 raid5_run_ops(sh, pending);
2865
2866         return_io(return_bi);
2867
2868 }
2869
2870 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2871 {
2872         raid6_conf_t *conf = sh->raid_conf;
2873         int disks = sh->disks;
2874         struct bio *return_bi = NULL;
2875         int i, pd_idx = sh->pd_idx;
2876         struct stripe_head_state s;
2877         struct r6_state r6s;
2878         struct r5dev *dev, *pdev, *qdev;
2879
2880         r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2881         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2882                 "pd_idx=%d, qd_idx=%d\n",
2883                (unsigned long long)sh->sector, sh->state,
2884                atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2885         memset(&s, 0, sizeof(s));
2886
2887         spin_lock(&sh->lock);
2888         clear_bit(STRIPE_HANDLE, &sh->state);
2889         clear_bit(STRIPE_DELAYED, &sh->state);
2890
2891         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2892         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2893         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2894         /* Now to look around and see what can be done */
2895
2896         rcu_read_lock();
2897         for (i=disks; i--; ) {
2898                 mdk_rdev_t *rdev;
2899                 dev = &sh->dev[i];
2900                 clear_bit(R5_Insync, &dev->flags);
2901
2902                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2903                         i, dev->flags, dev->toread, dev->towrite, dev->written);
2904                 /* maybe we can reply to a read */
2905                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2906                         struct bio *rbi, *rbi2;
2907                         pr_debug("Return read for disc %d\n", i);
2908                         spin_lock_irq(&conf->device_lock);
2909                         rbi = dev->toread;
2910                         dev->toread = NULL;
2911                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2912                                 wake_up(&conf->wait_for_overlap);
2913                         spin_unlock_irq(&conf->device_lock);
2914                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2915                                 copy_data(0, rbi, dev->page, dev->sector);
2916                                 rbi2 = r5_next_bio(rbi, dev->sector);
2917                                 spin_lock_irq(&conf->device_lock);
2918                                 if (--rbi->bi_phys_segments == 0) {
2919                                         rbi->bi_next = return_bi;
2920                                         return_bi = rbi;
2921                                 }
2922                                 spin_unlock_irq(&conf->device_lock);
2923                                 rbi = rbi2;
2924                         }
2925                 }
2926
2927                 /* now count some things */
2928                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2929                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2930
2931
2932                 if (dev->toread)
2933                         s.to_read++;
2934                 if (dev->towrite) {
2935                         s.to_write++;
2936                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2937                                 s.non_overwrite++;
2938                 }
2939                 if (dev->written)
2940                         s.written++;
2941                 rdev = rcu_dereference(conf->disks[i].rdev);
2942                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2943                         /* The ReadError flag will just be confusing now */
2944                         clear_bit(R5_ReadError, &dev->flags);
2945                         clear_bit(R5_ReWrite, &dev->flags);
2946                 }
2947                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2948                     || test_bit(R5_ReadError, &dev->flags)) {
2949                         if (s.failed < 2)
2950                                 r6s.failed_num[s.failed] = i;
2951                         s.failed++;
2952                 } else
2953                         set_bit(R5_Insync, &dev->flags);
2954         }
2955         rcu_read_unlock();
2956         pr_debug("locked=%d uptodate=%d to_read=%d"
2957                " to_write=%d failed=%d failed_num=%d,%d\n",
2958                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
2959                r6s.failed_num[0], r6s.failed_num[1]);
2960         /* check if the array has lost >2 devices and, if so, some requests
2961          * might need to be failed
2962          */
2963         if (s.failed > 2 && s.to_read+s.to_write+s.written)
2964                 handle_requests_to_failed_array(conf, sh, &s, disks,
2965                                                 &return_bi);
2966         if (s.failed > 2 && s.syncing) {
2967                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2968                 clear_bit(STRIPE_SYNCING, &sh->state);
2969                 s.syncing = 0;
2970         }
2971
2972         /*
2973          * might be able to return some write requests if the parity blocks
2974          * are safe, or on a failed drive
2975          */
2976         pdev = &sh->dev[pd_idx];
2977         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
2978                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
2979         qdev = &sh->dev[r6s.qd_idx];
2980         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
2981                 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
2982
2983         if ( s.written &&
2984              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
2985                              && !test_bit(R5_LOCKED, &pdev->flags)
2986                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
2987              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
2988                              && !test_bit(R5_LOCKED, &qdev->flags)
2989                              && test_bit(R5_UPTODATE, &qdev->flags)))))
2990                 handle_completed_write_requests(conf, sh, disks, &return_bi);
2991
2992         /* Now we might consider reading some blocks, either to check/generate
2993          * parity, or to satisfy requests
2994          * or to load a block that is being partially written.
2995          */
2996         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
2997             (s.syncing && (s.uptodate < disks)) || s.expanding)
2998                 handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
2999
3000         /* now to consider writing and what else, if anything should be read */
3001         if (s.to_write)
3002                 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
3003
3004         /* maybe we need to check and possibly fix the parity for this stripe
3005          * Any reads will already have been scheduled, so we just see if enough
3006          * data is available
3007          */
3008         if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3009                 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3010
3011         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3012                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3013                 clear_bit(STRIPE_SYNCING, &sh->state);
3014         }
3015
3016         /* If the failed drives are just a ReadError, then we might need
3017          * to progress the repair/check process
3018          */
3019         if (s.failed <= 2 && !conf->mddev->ro)
3020                 for (i = 0; i < s.failed; i++) {
3021                         dev = &sh->dev[r6s.failed_num[i]];
3022                         if (test_bit(R5_ReadError, &dev->flags)
3023                             && !test_bit(R5_LOCKED, &dev->flags)
3024                             && test_bit(R5_UPTODATE, &dev->flags)
3025                                 ) {
3026                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3027                                         set_bit(R5_Wantwrite, &dev->flags);
3028                                         set_bit(R5_ReWrite, &dev->flags);
3029                                         set_bit(R5_LOCKED, &dev->flags);
3030                                 } else {
3031                                         /* let's read it back */
3032                                         set_bit(R5_Wantread, &dev->flags);
3033                                         set_bit(R5_LOCKED, &dev->flags);
3034                                 }
3035                         }
3036                 }
3037
3038         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3039                 /* Need to write out all blocks after computing P&Q */
3040                 sh->disks = conf->raid_disks;
3041                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
3042                                              conf->raid_disks);
3043                 compute_parity6(sh, RECONSTRUCT_WRITE);
3044                 for (i = conf->raid_disks ; i-- ;  ) {
3045                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3046                         s.locked++;
3047                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3048                 }
3049                 clear_bit(STRIPE_EXPANDING, &sh->state);
3050         } else if (s.expanded) {
3051                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3052                 atomic_dec(&conf->reshape_stripes);
3053                 wake_up(&conf->wait_for_overlap);
3054                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3055         }
3056
3057         if (s.expanding && s.locked == 0)
3058                 handle_stripe_expansion(conf, sh, &r6s);
3059
3060         spin_unlock(&sh->lock);
3061
3062         return_io(return_bi);
3063
3064         for (i=disks; i-- ;) {
3065                 int rw;
3066                 struct bio *bi;
3067                 mdk_rdev_t *rdev;
3068                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
3069                         rw = WRITE;
3070                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
3071                         rw = READ;
3072                 else
3073                         continue;
3074
3075                 bi = &sh->dev[i].req;
3076
3077                 bi->bi_rw = rw;
3078                 if (rw == WRITE)
3079                         bi->bi_end_io = raid5_end_write_request;
3080                 else
3081                         bi->bi_end_io = raid5_end_read_request;
3082
3083                 rcu_read_lock();
3084                 rdev = rcu_dereference(conf->disks[i].rdev);
3085                 if (rdev && test_bit(Faulty, &rdev->flags))
3086                         rdev = NULL;
3087                 if (rdev)
3088                         atomic_inc(&rdev->nr_pending);
3089                 rcu_read_unlock();
3090
3091                 if (rdev) {
3092                         if (s.syncing || s.expanding || s.expanded)
3093                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
3094
3095                         bi->bi_bdev = rdev->bdev;
3096                         pr_debug("for %llu schedule op %ld on disc %d\n",
3097                                 (unsigned long long)sh->sector, bi->bi_rw, i);
3098                         atomic_inc(&sh->count);
3099                         bi->bi_sector = sh->sector + rdev->data_offset;
3100                         bi->bi_flags = 1 << BIO_UPTODATE;
3101                         bi->bi_vcnt = 1;
3102                         bi->bi_max_vecs = 1;
3103                         bi->bi_idx = 0;
3104                         bi->bi_io_vec = &sh->dev[i].vec;
3105                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
3106                         bi->bi_io_vec[0].bv_offset = 0;
3107                         bi->bi_size = STRIPE_SIZE;
3108                         bi->bi_next = NULL;
3109                         if (rw == WRITE &&
3110                             test_bit(R5_ReWrite, &sh->dev[i].flags))
3111                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
3112                         generic_make_request(bi);
3113                 } else {
3114                         if (rw == WRITE)
3115                                 set_bit(STRIPE_DEGRADED, &sh->state);
3116                         pr_debug("skip op %ld on disc %d for sector %llu\n",
3117                                 bi->bi_rw, i, (unsigned long long)sh->sector);
3118                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
3119                         set_bit(STRIPE_HANDLE, &sh->state);
3120                 }
3121         }
3122 }
3123
3124 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3125 {
3126         if (sh->raid_conf->level == 6)
3127                 handle_stripe6(sh, tmp_page);
3128         else
3129                 handle_stripe5(sh);
3130 }
3131
3132
3133
3134 static void raid5_activate_delayed(raid5_conf_t *conf)
3135 {
3136         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3137                 while (!list_empty(&conf->delayed_list)) {
3138                         struct list_head *l = conf->delayed_list.next;
3139                         struct stripe_head *sh;
3140                         sh = list_entry(l, struct stripe_head, lru);
3141                         list_del_init(l);
3142                         clear_bit(STRIPE_DELAYED, &sh->state);
3143                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3144                                 atomic_inc(&conf->preread_active_stripes);
3145                         list_add_tail(&sh->lru, &conf->handle_list);
3146                 }
3147         }
3148 }
3149
3150 static void activate_bit_delay(raid5_conf_t *conf)
3151 {
3152         /* device_lock is held */
3153         struct list_head head;
3154         list_add(&head, &conf->bitmap_list);
3155         list_del_init(&conf->bitmap_list);
3156         while (!list_empty(&head)) {
3157                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3158                 list_del_init(&sh->lru);
3159                 atomic_inc(&sh->count);
3160                 __release_stripe(conf, sh);
3161         }
3162 }
3163
3164 static void unplug_slaves(mddev_t *mddev)
3165 {
3166         raid5_conf_t *conf = mddev_to_conf(mddev);
3167         int i;
3168
3169         rcu_read_lock();
3170         for (i=0; i<mddev->raid_disks; i++) {
3171                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3172                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3173                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3174
3175                         atomic_inc(&rdev->nr_pending);
3176                         rcu_read_unlock();
3177
3178                         if (r_queue->unplug_fn)
3179                                 r_queue->unplug_fn(r_queue);
3180
3181                         rdev_dec_pending(rdev, mddev);
3182                         rcu_read_lock();
3183                 }
3184         }
3185         rcu_read_unlock();
3186 }
3187
3188 static void raid5_unplug_device(struct request_queue *q)
3189 {
3190         mddev_t *mddev = q->queuedata;
3191         raid5_conf_t *conf = mddev_to_conf(mddev);
3192         unsigned long flags;
3193
3194         spin_lock_irqsave(&conf->device_lock, flags);
3195
3196         if (blk_remove_plug(q)) {
3197                 conf->seq_flush++;
3198                 raid5_activate_delayed(conf);
3199         }
3200         md_wakeup_thread(mddev->thread);
3201
3202         spin_unlock_irqrestore(&conf->device_lock, flags);
3203
3204         unplug_slaves(mddev);
3205 }
3206
3207 static int raid5_congested(void *data, int bits)
3208 {
3209         mddev_t *mddev = data;
3210         raid5_conf_t *conf = mddev_to_conf(mddev);
3211
3212         /* No difference between reads and writes.  Just check
3213          * how busy the stripe_cache is
3214          */
3215         if (conf->inactive_blocked)
3216                 return 1;
3217         if (conf->quiesce)
3218                 return 1;
3219         if (list_empty_careful(&conf->inactive_list))
3220                 return 1;
3221
3222         return 0;
3223 }
3224
3225 /* We want read requests to align with chunks where possible,
3226  * but write requests don't need to.
3227  */
3228 static int raid5_mergeable_bvec(struct request_queue *q, struct bio *bio, struct bio_vec *biovec)
3229 {
3230         mddev_t *mddev = q->queuedata;
3231         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3232         int max;
3233         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3234         unsigned int bio_sectors = bio->bi_size >> 9;
3235
3236         if (bio_data_dir(bio) == WRITE)
3237                 return biovec->bv_len; /* always allow writes to be mergeable */
3238
3239         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3240         if (max < 0) max = 0;
3241         if (max <= biovec->bv_len && bio_sectors == 0)
3242                 return biovec->bv_len;
3243         else
3244                 return max;
3245 }
3246
3247
3248 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3249 {
3250         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3251         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3252         unsigned int bio_sectors = bio->bi_size >> 9;
3253
3254         return  chunk_sectors >=
3255                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3256 }
3257
3258 /*
3259  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3260  *  later sampled by raid5d.
3261  */
3262 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3263 {
3264         unsigned long flags;
3265
3266         spin_lock_irqsave(&conf->device_lock, flags);
3267
3268         bi->bi_next = conf->retry_read_aligned_list;
3269         conf->retry_read_aligned_list = bi;
3270
3271         spin_unlock_irqrestore(&conf->device_lock, flags);
3272         md_wakeup_thread(conf->mddev->thread);
3273 }
3274
3275
3276 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3277 {
3278         struct bio *bi;
3279
3280         bi = conf->retry_read_aligned;
3281         if (bi) {
3282                 conf->retry_read_aligned = NULL;
3283                 return bi;
3284         }
3285         bi = conf->retry_read_aligned_list;
3286         if(bi) {
3287                 conf->retry_read_aligned_list = bi->bi_next;
3288                 bi->bi_next = NULL;
3289                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3290                 bi->bi_hw_segments = 0; /* count of processed stripes */
3291         }
3292
3293         return bi;
3294 }
3295
3296
3297 /*
3298  *  The "raid5_align_endio" should check if the read succeeded and if it
3299  *  did, call bio_endio on the original bio (having bio_put the new bio
3300  *  first).
3301  *  If the read failed..
3302  */
3303 static void raid5_align_endio(struct bio *bi, int error)
3304 {
3305         struct bio* raid_bi  = bi->bi_private;
3306         mddev_t *mddev;
3307         raid5_conf_t *conf;
3308         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3309         mdk_rdev_t *rdev;
3310
3311         bio_put(bi);
3312
3313         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3314         conf = mddev_to_conf(mddev);
3315         rdev = (void*)raid_bi->bi_next;
3316         raid_bi->bi_next = NULL;
3317
3318         rdev_dec_pending(rdev, conf->mddev);
3319
3320         if (!error && uptodate) {
3321                 bio_endio(raid_bi, 0);
3322                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3323                         wake_up(&conf->wait_for_stripe);
3324                 return;
3325         }
3326
3327
3328         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3329
3330         add_bio_to_retry(raid_bi, conf);
3331 }
3332
3333 static int bio_fits_rdev(struct bio *bi)
3334 {
3335         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3336
3337         if ((bi->bi_size>>9) > q->max_sectors)
3338                 return 0;
3339         blk_recount_segments(q, bi);
3340         if (bi->bi_phys_segments > q->max_phys_segments ||
3341             bi->bi_hw_segments > q->max_hw_segments)
3342                 return 0;
3343
3344         if (q->merge_bvec_fn)
3345                 /* it's too hard to apply the merge_bvec_fn at this stage,
3346                  * just just give up
3347                  */
3348                 return 0;
3349
3350         return 1;
3351 }
3352
3353
3354 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3355 {
3356         mddev_t *mddev = q->queuedata;
3357         raid5_conf_t *conf = mddev_to_conf(mddev);
3358         const unsigned int raid_disks = conf->raid_disks;
3359         const unsigned int data_disks = raid_disks - conf->max_degraded;
3360         unsigned int dd_idx, pd_idx;
3361         struct bio* align_bi;
3362         mdk_rdev_t *rdev;
3363
3364         if (!in_chunk_boundary(mddev, raid_bio)) {
3365                 pr_debug("chunk_aligned_read : non aligned\n");
3366                 return 0;
3367         }
3368         /*
3369          * use bio_clone to make a copy of the bio
3370          */
3371         align_bi = bio_clone(raid_bio, GFP_NOIO);
3372         if (!align_bi)
3373                 return 0;
3374         /*
3375          *   set bi_end_io to a new function, and set bi_private to the
3376          *     original bio.
3377          */
3378         align_bi->bi_end_io  = raid5_align_endio;
3379         align_bi->bi_private = raid_bio;
3380         /*
3381          *      compute position
3382          */
3383         align_bi->bi_sector =  raid5_compute_sector(raid_bio->bi_sector,
3384                                         raid_disks,
3385                                         data_disks,
3386                                         &dd_idx,
3387                                         &pd_idx,
3388                                         conf);
3389
3390         rcu_read_lock();
3391         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3392         if (rdev && test_bit(In_sync, &rdev->flags)) {
3393                 atomic_inc(&rdev->nr_pending);
3394                 rcu_read_unlock();
3395                 raid_bio->bi_next = (void*)rdev;
3396                 align_bi->bi_bdev =  rdev->bdev;
3397                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3398                 align_bi->bi_sector += rdev->data_offset;
3399
3400                 if (!bio_fits_rdev(align_bi)) {
3401                         /* too big in some way */
3402                         bio_put(align_bi);
3403                         rdev_dec_pending(rdev, mddev);
3404                         return 0;
3405                 }
3406
3407                 spin_lock_irq(&conf->device_lock);
3408                 wait_event_lock_irq(conf->wait_for_stripe,
3409                                     conf->quiesce == 0,
3410                                     conf->device_lock, /* nothing */);
3411                 atomic_inc(&conf->active_aligned_reads);
3412                 spin_unlock_irq(&conf->device_lock);
3413
3414                 generic_make_request(align_bi);
3415                 return 1;
3416         } else {
3417                 rcu_read_unlock();
3418                 bio_put(align_bi);
3419                 return 0;
3420         }
3421 }
3422
3423
3424 static int make_request(struct request_queue *q, struct bio * bi)
3425 {
3426         mddev_t *mddev = q->queuedata;
3427         raid5_conf_t *conf = mddev_to_conf(mddev);
3428         unsigned int dd_idx, pd_idx;
3429         sector_t new_sector;
3430         sector_t logical_sector, last_sector;
3431         struct stripe_head *sh;
3432         const int rw = bio_data_dir(bi);
3433         int remaining;
3434
3435         if (unlikely(bio_barrier(bi))) {
3436                 bio_endio(bi, -EOPNOTSUPP);
3437                 return 0;
3438         }
3439
3440         md_write_start(mddev, bi);
3441
3442         disk_stat_inc(mddev->gendisk, ios[rw]);
3443         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3444
3445         if (rw == READ &&
3446              mddev->reshape_position == MaxSector &&
3447              chunk_aligned_read(q,bi))
3448                 return 0;
3449
3450         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3451         last_sector = bi->bi_sector + (bi->bi_size>>9);
3452         bi->bi_next = NULL;
3453         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3454
3455         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3456                 DEFINE_WAIT(w);
3457                 int disks, data_disks;
3458
3459         retry:
3460                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3461                 if (likely(conf->expand_progress == MaxSector))
3462                         disks = conf->raid_disks;
3463                 else {
3464                         /* spinlock is needed as expand_progress may be
3465                          * 64bit on a 32bit platform, and so it might be
3466                          * possible to see a half-updated value
3467                          * Ofcourse expand_progress could change after
3468                          * the lock is dropped, so once we get a reference
3469                          * to the stripe that we think it is, we will have
3470                          * to check again.
3471                          */
3472                         spin_lock_irq(&conf->device_lock);
3473                         disks = conf->raid_disks;
3474                         if (logical_sector >= conf->expand_progress)
3475                                 disks = conf->previous_raid_disks;
3476                         else {
3477                                 if (logical_sector >= conf->expand_lo) {
3478                                         spin_unlock_irq(&conf->device_lock);
3479                                         schedule();
3480                                         goto retry;
3481                                 }
3482                         }
3483                         spin_unlock_irq(&conf->device_lock);
3484                 }
3485                 data_disks = disks - conf->max_degraded;
3486
3487                 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3488                                                   &dd_idx, &pd_idx, conf);
3489                 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3490                         (unsigned long long)new_sector, 
3491                         (unsigned long long)logical_sector);
3492
3493                 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3494                 if (sh) {
3495                         if (unlikely(conf->expand_progress != MaxSector)) {
3496                                 /* expansion might have moved on while waiting for a
3497                                  * stripe, so we must do the range check again.
3498                                  * Expansion could still move past after this
3499                                  * test, but as we are holding a reference to
3500                                  * 'sh', we know that if that happens,
3501                                  *  STRIPE_EXPANDING will get set and the expansion
3502                                  * won't proceed until we finish with the stripe.
3503                                  */
3504                                 int must_retry = 0;
3505                                 spin_lock_irq(&conf->device_lock);
3506                                 if (logical_sector <  conf->expand_progress &&
3507                                     disks == conf->previous_raid_disks)
3508                                         /* mismatch, need to try again */
3509                                         must_retry = 1;
3510                                 spin_unlock_irq(&conf->device_lock);
3511                                 if (must_retry) {
3512                                         release_stripe(sh);
3513                                         goto retry;
3514                                 }
3515                         }
3516                         /* FIXME what if we get a false positive because these
3517                          * are being updated.
3518                          */
3519                         if (logical_sector >= mddev->suspend_lo &&
3520                             logical_sector < mddev->suspend_hi) {
3521                                 release_stripe(sh);
3522                                 schedule();
3523                                 goto retry;
3524                         }
3525
3526                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3527                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3528                                 /* Stripe is busy expanding or
3529                                  * add failed due to overlap.  Flush everything
3530                                  * and wait a while
3531                                  */
3532                                 raid5_unplug_device(mddev->queue);
3533                                 release_stripe(sh);
3534                                 schedule();
3535                                 goto retry;
3536                         }
3537                         finish_wait(&conf->wait_for_overlap, &w);
3538                         handle_stripe(sh, NULL);
3539                         release_stripe(sh);
3540                 } else {
3541                         /* cannot get stripe for read-ahead, just give-up */
3542                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3543                         finish_wait(&conf->wait_for_overlap, &w);
3544                         break;
3545                 }
3546                         
3547         }
3548         spin_lock_irq(&conf->device_lock);
3549         remaining = --bi->bi_phys_segments;
3550         spin_unlock_irq(&conf->device_lock);
3551         if (remaining == 0) {
3552
3553                 if ( rw == WRITE )
3554                         md_write_end(mddev);
3555
3556                 bi->bi_end_io(bi,
3557                               test_bit(BIO_UPTODATE, &bi->bi_flags)
3558                                 ? 0 : -EIO);
3559         }
3560         return 0;
3561 }
3562
3563 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3564 {
3565         /* reshaping is quite different to recovery/resync so it is
3566          * handled quite separately ... here.
3567          *
3568          * On each call to sync_request, we gather one chunk worth of
3569          * destination stripes and flag them as expanding.
3570          * Then we find all the source stripes and request reads.
3571          * As the reads complete, handle_stripe will copy the data
3572          * into the destination stripe and release that stripe.
3573          */
3574         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3575         struct stripe_head *sh;
3576         int pd_idx;
3577         sector_t first_sector, last_sector;
3578         int raid_disks = conf->previous_raid_disks;
3579         int data_disks = raid_disks - conf->max_degraded;
3580         int new_data_disks = conf->raid_disks - conf->max_degraded;
3581         int i;
3582         int dd_idx;
3583         sector_t writepos, safepos, gap;
3584
3585         if (sector_nr == 0 &&
3586             conf->expand_progress != 0) {
3587                 /* restarting in the middle, skip the initial sectors */
3588                 sector_nr = conf->expand_progress;
3589                 sector_div(sector_nr, new_data_disks);
3590                 *skipped = 1;
3591                 return sector_nr;
3592         }
3593
3594         /* we update the metadata when there is more than 3Meg
3595          * in the block range (that is rather arbitrary, should
3596          * probably be time based) or when the data about to be
3597          * copied would over-write the source of the data at
3598          * the front of the range.
3599          * i.e. one new_stripe forward from expand_progress new_maps
3600          * to after where expand_lo old_maps to
3601          */
3602         writepos = conf->expand_progress +
3603                 conf->chunk_size/512*(new_data_disks);
3604         sector_div(writepos, new_data_disks);
3605         safepos = conf->expand_lo;
3606         sector_div(safepos, data_disks);
3607         gap = conf->expand_progress - conf->expand_lo;
3608
3609         if (writepos >= safepos ||
3610             gap > (new_data_disks)*3000*2 /*3Meg*/) {
3611                 /* Cannot proceed until we've updated the superblock... */
3612                 wait_event(conf->wait_for_overlap,
3613                            atomic_read(&conf->reshape_stripes)==0);
3614                 mddev->reshape_position = conf->expand_progress;
3615                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3616                 md_wakeup_thread(mddev->thread);
3617                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3618                            kthread_should_stop());
3619                 spin_lock_irq(&conf->device_lock);
3620                 conf->expand_lo = mddev->reshape_position;
3621                 spin_unlock_irq(&conf->device_lock);
3622                 wake_up(&conf->wait_for_overlap);
3623         }
3624
3625         for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3626                 int j;
3627                 int skipped = 0;
3628                 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3629                 sh = get_active_stripe(conf, sector_nr+i,
3630                                        conf->raid_disks, pd_idx, 0);
3631                 set_bit(STRIPE_EXPANDING, &sh->state);
3632                 atomic_inc(&conf->reshape_stripes);
3633                 /* If any of this stripe is beyond the end of the old
3634                  * array, then we need to zero those blocks
3635                  */
3636                 for (j=sh->disks; j--;) {
3637                         sector_t s;
3638                         if (j == sh->pd_idx)
3639                                 continue;
3640                         if (conf->level == 6 &&
3641                             j == raid6_next_disk(sh->pd_idx, sh->disks))
3642                                 continue;
3643                         s = compute_blocknr(sh, j);
3644                         if (s < (mddev->array_size<<1)) {
3645                                 skipped = 1;
3646                                 continue;
3647                         }
3648                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3649                         set_bit(R5_Expanded, &sh->dev[j].flags);
3650                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
3651                 }
3652                 if (!skipped) {
3653                         set_bit(STRIPE_EXPAND_READY, &sh->state);
3654                         set_bit(STRIPE_HANDLE, &sh->state);
3655                 }
3656                 release_stripe(sh);
3657         }
3658         spin_lock_irq(&conf->device_lock);
3659         conf->expand_progress = (sector_nr + i) * new_data_disks;
3660         spin_unlock_irq(&conf->device_lock);
3661         /* Ok, those stripe are ready. We can start scheduling
3662          * reads on the source stripes.
3663          * The source stripes are determined by mapping the first and last
3664          * block on the destination stripes.
3665          */
3666         first_sector =
3667                 raid5_compute_sector(sector_nr*(new_data_disks),
3668                                      raid_disks, data_disks,
3669                                      &dd_idx, &pd_idx, conf);
3670         last_sector =
3671                 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3672                                      *(new_data_disks) -1,
3673                                      raid_disks, data_disks,
3674                                      &dd_idx, &pd_idx, conf);
3675         if (last_sector >= (mddev->size<<1))
3676                 last_sector = (mddev->size<<1)-1;
3677         while (first_sector <= last_sector) {
3678                 pd_idx = stripe_to_pdidx(first_sector, conf,
3679                                          conf->previous_raid_disks);
3680                 sh = get_active_stripe(conf, first_sector,
3681                                        conf->previous_raid_disks, pd_idx, 0);
3682                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3683                 set_bit(STRIPE_HANDLE, &sh->state);
3684                 release_stripe(sh);
3685                 first_sector += STRIPE_SECTORS;
3686         }
3687         return conf->chunk_size>>9;
3688 }
3689
3690 /* FIXME go_faster isn't used */
3691 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3692 {
3693         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3694         struct stripe_head *sh;
3695         int pd_idx;
3696         int raid_disks = conf->raid_disks;
3697         sector_t max_sector = mddev->size << 1;
3698         int sync_blocks;
3699         int still_degraded = 0;
3700         int i;
3701
3702         if (sector_nr >= max_sector) {
3703                 /* just being told to finish up .. nothing much to do */
3704                 unplug_slaves(mddev);
3705                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3706                         end_reshape(conf);
3707                         return 0;
3708                 }
3709
3710                 if (mddev->curr_resync < max_sector) /* aborted */
3711                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3712                                         &sync_blocks, 1);
3713                 else /* completed sync */
3714                         conf->fullsync = 0;
3715                 bitmap_close_sync(mddev->bitmap);
3716
3717                 return 0;
3718         }
3719
3720         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3721                 return reshape_request(mddev, sector_nr, skipped);
3722
3723         /* if there is too many failed drives and we are trying
3724          * to resync, then assert that we are finished, because there is
3725          * nothing we can do.
3726          */
3727         if (mddev->degraded >= conf->max_degraded &&
3728             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3729                 sector_t rv = (mddev->size << 1) - sector_nr;
3730                 *skipped = 1;
3731                 return rv;
3732         }
3733         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3734             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3735             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3736                 /* we can skip this block, and probably more */
3737                 sync_blocks /= STRIPE_SECTORS;
3738                 *skipped = 1;
3739                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3740         }
3741
3742         pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3743         sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3744         if (sh == NULL) {
3745                 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3746                 /* make sure we don't swamp the stripe cache if someone else
3747                  * is trying to get access
3748                  */
3749                 schedule_timeout_uninterruptible(1);
3750         }
3751         /* Need to check if array will still be degraded after recovery/resync
3752          * We don't need to check the 'failed' flag as when that gets set,
3753          * recovery aborts.
3754          */
3755         for (i=0; i<mddev->raid_disks; i++)
3756                 if (conf->disks[i].rdev == NULL)
3757                         still_degraded = 1;
3758
3759         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3760
3761         spin_lock(&sh->lock);
3762         set_bit(STRIPE_SYNCING, &sh->state);
3763         clear_bit(STRIPE_INSYNC, &sh->state);
3764         spin_unlock(&sh->lock);
3765
3766         handle_stripe(sh, NULL);
3767         release_stripe(sh);
3768
3769         return STRIPE_SECTORS;
3770 }
3771
3772 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3773 {
3774         /* We may not be able to submit a whole bio at once as there
3775          * may not be enough stripe_heads available.
3776          * We cannot pre-allocate enough stripe_heads as we may need
3777          * more than exist in the cache (if we allow ever large chunks).
3778          * So we do one stripe head at a time and record in
3779          * ->bi_hw_segments how many have been done.
3780          *
3781          * We *know* that this entire raid_bio is in one chunk, so
3782          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3783          */
3784         struct stripe_head *sh;
3785         int dd_idx, pd_idx;
3786         sector_t sector, logical_sector, last_sector;
3787         int scnt = 0;
3788         int remaining;
3789         int handled = 0;
3790
3791         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3792         sector = raid5_compute_sector(  logical_sector,
3793                                         conf->raid_disks,
3794                                         conf->raid_disks - conf->max_degraded,
3795                                         &dd_idx,
3796                                         &pd_idx,
3797                                         conf);
3798         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3799
3800         for (; logical_sector < last_sector;
3801              logical_sector += STRIPE_SECTORS,
3802                      sector += STRIPE_SECTORS,
3803                      scnt++) {
3804
3805                 if (scnt < raid_bio->bi_hw_segments)
3806                         /* already done this stripe */
3807                         continue;
3808
3809                 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3810
3811                 if (!sh) {
3812                         /* failed to get a stripe - must wait */
3813                         raid_bio->bi_hw_segments = scnt;
3814                         conf->retry_read_aligned = raid_bio;
3815                         return handled;
3816                 }
3817
3818                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3819                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3820                         release_stripe(sh);
3821                         raid_bio->bi_hw_segments = scnt;
3822                         conf->retry_read_aligned = raid_bio;
3823                         return handled;
3824                 }
3825
3826                 handle_stripe(sh, NULL);
3827                 release_stripe(sh);
3828                 handled++;
3829         }
3830         spin_lock_irq(&conf->device_lock);
3831         remaining = --raid_bio->bi_phys_segments;
3832         spin_unlock_irq(&conf->device_lock);
3833         if (remaining == 0) {
3834
3835                 raid_bio->bi_end_io(raid_bio,
3836                               test_bit(BIO_UPTODATE, &raid_bio->bi_flags)
3837                                 ? 0 : -EIO);
3838         }
3839         if (atomic_dec_and_test(&conf->active_aligned_reads))
3840                 wake_up(&conf->wait_for_stripe);
3841         return handled;
3842 }
3843
3844
3845
3846 /*
3847  * This is our raid5 kernel thread.
3848  *
3849  * We scan the hash table for stripes which can be handled now.
3850  * During the scan, completed stripes are saved for us by the interrupt
3851  * handler, so that they will not have to wait for our next wakeup.
3852  */
3853 static void raid5d (mddev_t *mddev)
3854 {
3855         struct stripe_head *sh;
3856         raid5_conf_t *conf = mddev_to_conf(mddev);
3857         int handled;
3858
3859         pr_debug("+++ raid5d active\n");
3860
3861         md_check_recovery(mddev);
3862
3863         handled = 0;
3864         spin_lock_irq(&conf->device_lock);
3865         while (1) {
3866                 struct list_head *first;
3867                 struct bio *bio;
3868
3869                 if (conf->seq_flush != conf->seq_write) {
3870                         int seq = conf->seq_flush;
3871                         spin_unlock_irq(&conf->device_lock);
3872                         bitmap_unplug(mddev->bitmap);
3873                         spin_lock_irq(&conf->device_lock);
3874                         conf->seq_write = seq;
3875                         activate_bit_delay(conf);
3876                 }
3877
3878                 if (list_empty(&conf->handle_list) &&
3879                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
3880                     !blk_queue_plugged(mddev->queue) &&
3881                     !list_empty(&conf->delayed_list))
3882                         raid5_activate_delayed(conf);
3883
3884                 while ((bio = remove_bio_from_retry(conf))) {
3885                         int ok;
3886                         spin_unlock_irq(&conf->device_lock);
3887                         ok = retry_aligned_read(conf, bio);
3888                         spin_lock_irq(&conf->device_lock);
3889                         if (!ok)
3890                                 break;
3891                         handled++;
3892                 }
3893
3894                 if (list_empty(&conf->handle_list)) {
3895                         async_tx_issue_pending_all();
3896                         break;
3897                 }
3898
3899                 first = conf->handle_list.next;
3900                 sh = list_entry(first, struct stripe_head, lru);
3901
3902                 list_del_init(first);
3903                 atomic_inc(&sh->count);
3904                 BUG_ON(atomic_read(&sh->count)!= 1);
3905                 spin_unlock_irq(&conf->device_lock);
3906                 
3907                 handled++;
3908                 handle_stripe(sh, conf->spare_page);
3909                 release_stripe(sh);
3910
3911                 spin_lock_irq(&conf->device_lock);
3912         }
3913         pr_debug("%d stripes handled\n", handled);
3914
3915         spin_unlock_irq(&conf->device_lock);
3916
3917         unplug_slaves(mddev);
3918
3919         pr_debug("--- raid5d inactive\n");
3920 }
3921
3922 static ssize_t
3923 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
3924 {
3925         raid5_conf_t *conf = mddev_to_conf(mddev);
3926         if (conf)
3927                 return sprintf(page, "%d\n", conf->max_nr_stripes);
3928         else
3929                 return 0;
3930 }
3931
3932 static ssize_t
3933 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
3934 {
3935         raid5_conf_t *conf = mddev_to_conf(mddev);
3936         char *end;
3937         int new;
3938         if (len >= PAGE_SIZE)
3939                 return -EINVAL;
3940         if (!conf)
3941                 return -ENODEV;
3942
3943         new = simple_strtoul(page, &end, 10);
3944         if (!*page || (*end && *end != '\n') )
3945                 return -EINVAL;
3946         if (new <= 16 || new > 32768)
3947                 return -EINVAL;
3948         while (new < conf->max_nr_stripes) {
3949                 if (drop_one_stripe(conf))
3950                         conf->max_nr_stripes--;
3951                 else
3952                         break;
3953         }
3954         md_allow_write(mddev);
3955         while (new > conf->max_nr_stripes) {
3956                 if (grow_one_stripe(conf))
3957                         conf->max_nr_stripes++;
3958                 else break;
3959         }
3960         return len;
3961 }
3962
3963 static struct md_sysfs_entry
3964 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
3965                                 raid5_show_stripe_cache_size,
3966                                 raid5_store_stripe_cache_size);
3967
3968 static ssize_t
3969 stripe_cache_active_show(mddev_t *mddev, char *page)
3970 {
3971         raid5_conf_t *conf = mddev_to_conf(mddev);
3972         if (conf)
3973                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
3974         else
3975                 return 0;
3976 }
3977
3978 static struct md_sysfs_entry
3979 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
3980
3981 static struct attribute *raid5_attrs[] =  {
3982         &raid5_stripecache_size.attr,
3983         &raid5_stripecache_active.attr,
3984         NULL,
3985 };
3986 static struct attribute_group raid5_attrs_group = {
3987         .name = NULL,
3988         .attrs = raid5_attrs,
3989 };
3990
3991 static int run(mddev_t *mddev)
3992 {
3993         raid5_conf_t *conf;
3994         int raid_disk, memory;
3995         mdk_rdev_t *rdev;
3996         struct disk_info *disk;
3997         struct list_head *tmp;
3998         int working_disks = 0;
3999
4000         if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4001                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4002                        mdname(mddev), mddev->level);
4003                 return -EIO;
4004         }
4005
4006         if (mddev->reshape_position != MaxSector) {
4007                 /* Check that we can continue the reshape.
4008                  * Currently only disks can change, it must
4009                  * increase, and we must be past the point where
4010                  * a stripe over-writes itself
4011                  */
4012                 sector_t here_new, here_old;
4013                 int old_disks;
4014                 int max_degraded = (mddev->level == 5 ? 1 : 2);
4015
4016                 if (mddev->new_level != mddev->level ||
4017                     mddev->new_layout != mddev->layout ||
4018                     mddev->new_chunk != mddev->chunk_size) {
4019                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4020                                "required - aborting.\n",
4021                                mdname(mddev));
4022                         return -EINVAL;
4023                 }
4024                 if (mddev->delta_disks <= 0) {
4025                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4026                                "(reduce disks) required - aborting.\n",
4027                                mdname(mddev));
4028                         return -EINVAL;
4029                 }
4030                 old_disks = mddev->raid_disks - mddev->delta_disks;
4031                 /* reshape_position must be on a new-stripe boundary, and one
4032                  * further up in new geometry must map after here in old
4033                  * geometry.
4034                  */
4035                 here_new = mddev->reshape_position;
4036                 if (sector_div(here_new, (mddev->chunk_size>>9)*
4037                                (mddev->raid_disks - max_degraded))) {
4038                         printk(KERN_ERR "raid5: reshape_position not "
4039                                "on a stripe boundary\n");
4040                         return -EINVAL;
4041                 }
4042                 /* here_new is the stripe we will write to */
4043                 here_old = mddev->reshape_position;
4044                 sector_div(here_old, (mddev->chunk_size>>9)*
4045                            (old_disks-max_degraded));
4046                 /* here_old is the first stripe that we might need to read
4047                  * from */
4048                 if (here_new >= here_old) {
4049                         /* Reading from the same stripe as writing to - bad */
4050                         printk(KERN_ERR "raid5: reshape_position too early for "
4051                                "auto-recovery - aborting.\n");
4052                         return -EINVAL;
4053                 }
4054                 printk(KERN_INFO "raid5: reshape will continue\n");
4055                 /* OK, we should be able to continue; */
4056         }
4057
4058
4059         mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4060         if ((conf = mddev->private) == NULL)
4061                 goto abort;
4062         if (mddev->reshape_position == MaxSector) {
4063                 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4064         } else {
4065                 conf->raid_disks = mddev->raid_disks;
4066                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4067         }
4068
4069         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4070                               GFP_KERNEL);
4071         if (!conf->disks)
4072                 goto abort;
4073
4074         conf->mddev = mddev;
4075
4076         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4077                 goto abort;
4078
4079         if (mddev->level == 6) {
4080                 conf->spare_page = alloc_page(GFP_KERNEL);
4081                 if (!conf->spare_page)
4082                         goto abort;
4083         }
4084         spin_lock_init(&conf->device_lock);
4085         init_waitqueue_head(&conf->wait_for_stripe);
4086         init_waitqueue_head(&conf->wait_for_overlap);
4087         INIT_LIST_HEAD(&conf->handle_list);
4088         INIT_LIST_HEAD(&conf->delayed_list);
4089         INIT_LIST_HEAD(&conf->bitmap_list);
4090         INIT_LIST_HEAD(&conf->inactive_list);
4091         atomic_set(&conf->active_stripes, 0);
4092         atomic_set(&conf->preread_active_stripes, 0);
4093         atomic_set(&conf->active_aligned_reads, 0);
4094
4095         pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4096
4097         ITERATE_RDEV(mddev,rdev,tmp) {
4098                 raid_disk = rdev->raid_disk;
4099                 if (raid_disk >= conf->raid_disks
4100                     || raid_disk < 0)
4101                         continue;
4102                 disk = conf->disks + raid_disk;
4103
4104                 disk->rdev = rdev;
4105
4106                 if (test_bit(In_sync, &rdev->flags)) {
4107                         char b[BDEVNAME_SIZE];
4108                         printk(KERN_INFO "raid5: device %s operational as raid"
4109                                 " disk %d\n", bdevname(rdev->bdev,b),
4110                                 raid_disk);
4111                         working_disks++;
4112                 }
4113         }
4114
4115         /*
4116          * 0 for a fully functional array, 1 or 2 for a degraded array.
4117          */
4118         mddev->degraded = conf->raid_disks - working_disks;
4119         conf->mddev = mddev;
4120         conf->chunk_size = mddev->chunk_size;
4121         conf->level = mddev->level;
4122         if (conf->level == 6)
4123                 conf->max_degraded = 2;
4124         else
4125                 conf->max_degraded = 1;
4126         conf->algorithm = mddev->layout;
4127         conf->max_nr_stripes = NR_STRIPES;
4128         conf->expand_progress = mddev->reshape_position;
4129
4130         /* device size must be a multiple of chunk size */
4131         mddev->size &= ~(mddev->chunk_size/1024 -1);
4132         mddev->resync_max_sectors = mddev->size << 1;
4133
4134         if (conf->level == 6 && conf->raid_disks < 4) {
4135                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4136                        mdname(mddev), conf->raid_disks);
4137                 goto abort;
4138         }
4139         if (!conf->chunk_size || conf->chunk_size % 4) {
4140                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4141                         conf->chunk_size, mdname(mddev));
4142                 goto abort;
4143         }
4144         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4145                 printk(KERN_ERR 
4146                         "raid5: unsupported parity algorithm %d for %s\n",
4147                         conf->algorithm, mdname(mddev));
4148                 goto abort;
4149         }
4150         if (mddev->degraded > conf->max_degraded) {
4151                 printk(KERN_ERR "raid5: not enough operational devices for %s"
4152                         " (%d/%d failed)\n",
4153                         mdname(mddev), mddev->degraded, conf->raid_disks);
4154                 goto abort;
4155         }
4156
4157         if (mddev->degraded > 0 &&
4158             mddev->recovery_cp != MaxSector) {
4159                 if (mddev->ok_start_degraded)
4160                         printk(KERN_WARNING
4161                                "raid5: starting dirty degraded array: %s"
4162                                "- data corruption possible.\n",
4163                                mdname(mddev));
4164                 else {
4165                         printk(KERN_ERR
4166                                "raid5: cannot start dirty degraded array for %s\n",
4167                                mdname(mddev));
4168                         goto abort;
4169                 }
4170         }
4171
4172         {
4173                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4174                 if (!mddev->thread) {
4175                         printk(KERN_ERR 
4176                                 "raid5: couldn't allocate thread for %s\n",
4177                                 mdname(mddev));
4178                         goto abort;
4179                 }
4180         }
4181         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4182                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4183         if (grow_stripes(conf, conf->max_nr_stripes)) {
4184                 printk(KERN_ERR 
4185                         "raid5: couldn't allocate %dkB for buffers\n", memory);
4186                 shrink_stripes(conf);
4187                 md_unregister_thread(mddev->thread);
4188                 goto abort;
4189         } else
4190                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4191                         memory, mdname(mddev));
4192
4193         if (mddev->degraded == 0)
4194                 printk("raid5: raid level %d set %s active with %d out of %d"
4195                         " devices, algorithm %d\n", conf->level, mdname(mddev), 
4196                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4197                         conf->algorithm);
4198         else
4199                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4200                         " out of %d devices, algorithm %d\n", conf->level,
4201                         mdname(mddev), mddev->raid_disks - mddev->degraded,
4202                         mddev->raid_disks, conf->algorithm);
4203
4204         print_raid5_conf(conf);
4205
4206         if (conf->expand_progress != MaxSector) {
4207                 printk("...ok start reshape thread\n");
4208                 conf->expand_lo = conf->expand_progress;
4209                 atomic_set(&conf->reshape_stripes, 0);
4210                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4211                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4212                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4213                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4214                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4215                                                         "%s_reshape");
4216         }
4217
4218         /* read-ahead size must cover two whole stripes, which is
4219          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4220          */
4221         {
4222                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4223                 int stripe = data_disks *
4224                         (mddev->chunk_size / PAGE_SIZE);
4225                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4226                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4227         }
4228
4229         /* Ok, everything is just fine now */
4230         if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4231                 printk(KERN_WARNING
4232                        "raid5: failed to create sysfs attributes for %s\n",
4233                        mdname(mddev));
4234
4235         mddev->queue->unplug_fn = raid5_unplug_device;
4236         mddev->queue->backing_dev_info.congested_data = mddev;
4237         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4238
4239         mddev->array_size =  mddev->size * (conf->previous_raid_disks -
4240                                             conf->max_degraded);
4241
4242         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4243
4244         return 0;
4245 abort:
4246         if (conf) {
4247                 print_raid5_conf(conf);
4248                 safe_put_page(conf->spare_page);
4249                 kfree(conf->disks);
4250                 kfree(conf->stripe_hashtbl);
4251                 kfree(conf);
4252         }
4253         mddev->private = NULL;
4254         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4255         return -EIO;
4256 }
4257
4258
4259
4260 static int stop(mddev_t *mddev)
4261 {
4262         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4263
4264         md_unregister_thread(mddev->thread);
4265         mddev->thread = NULL;
4266         shrink_stripes(conf);
4267         kfree(conf->stripe_hashtbl);
4268         mddev->queue->backing_dev_info.congested_fn = NULL;
4269         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4270         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4271         kfree(conf->disks);
4272         kfree(conf);
4273         mddev->private = NULL;
4274         return 0;
4275 }
4276
4277 #ifdef DEBUG
4278 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4279 {
4280         int i;
4281
4282         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4283                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4284         seq_printf(seq, "sh %llu,  count %d.\n",
4285                    (unsigned long long)sh->sector, atomic_read(&sh->count));
4286         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4287         for (i = 0; i < sh->disks; i++) {
4288                 seq_printf(seq, "(cache%d: %p %ld) ",
4289                            i, sh->dev[i].page, sh->dev[i].flags);
4290         }
4291         seq_printf(seq, "\n");
4292 }
4293
4294 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4295 {
4296         struct stripe_head *sh;
4297         struct hlist_node *hn;
4298         int i;
4299
4300         spin_lock_irq(&conf->device_lock);
4301         for (i = 0; i < NR_HASH; i++) {
4302                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4303                         if (sh->raid_conf != conf)
4304                                 continue;
4305                         print_sh(seq, sh);
4306                 }
4307         }
4308         spin_unlock_irq(&conf->device_lock);
4309 }
4310 #endif
4311
4312 static void status (struct seq_file *seq, mddev_t *mddev)
4313 {
4314         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4315         int i;
4316
4317         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4318         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4319         for (i = 0; i < conf->raid_disks; i++)
4320                 seq_printf (seq, "%s",
4321                                conf->disks[i].rdev &&
4322                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4323         seq_printf (seq, "]");
4324 #ifdef DEBUG
4325         seq_printf (seq, "\n");
4326         printall(seq, conf);
4327 #endif
4328 }
4329
4330 static void print_raid5_conf (raid5_conf_t *conf)
4331 {
4332         int i;
4333         struct disk_info *tmp;
4334
4335         printk("RAID5 conf printout:\n");
4336         if (!conf) {
4337                 printk("(conf==NULL)\n");
4338                 return;
4339         }
4340         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4341                  conf->raid_disks - conf->mddev->degraded);
4342
4343         for (i = 0; i < conf->raid_disks; i++) {
4344                 char b[BDEVNAME_SIZE];
4345                 tmp = conf->disks + i;
4346                 if (tmp->rdev)
4347                 printk(" disk %d, o:%d, dev:%s\n",
4348                         i, !test_bit(Faulty, &tmp->rdev->flags),
4349                         bdevname(tmp->rdev->bdev,b));
4350         }
4351 }
4352
4353 static int raid5_spare_active(mddev_t *mddev)
4354 {
4355         int i;
4356         raid5_conf_t *conf = mddev->private;
4357         struct disk_info *tmp;
4358
4359         for (i = 0; i < conf->raid_disks; i++) {
4360                 tmp = conf->disks + i;
4361                 if (tmp->rdev
4362                     && !test_bit(Faulty, &tmp->rdev->flags)
4363                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4364                         unsigned long flags;
4365                         spin_lock_irqsave(&conf->device_lock, flags);
4366                         mddev->degraded--;
4367                         spin_unlock_irqrestore(&conf->device_lock, flags);
4368                 }
4369         }
4370         print_raid5_conf(conf);
4371         return 0;
4372 }
4373
4374 static int raid5_remove_disk(mddev_t *mddev, int number)
4375 {
4376         raid5_conf_t *conf = mddev->private;
4377         int err = 0;
4378         mdk_rdev_t *rdev;
4379         struct disk_info *p = conf->disks + number;
4380
4381         print_raid5_conf(conf);
4382         rdev = p->rdev;
4383         if (rdev) {
4384                 if (test_bit(In_sync, &rdev->flags) ||
4385                     atomic_read(&rdev->nr_pending)) {
4386                         err = -EBUSY;
4387                         goto abort;
4388                 }
4389                 p->rdev = NULL;
4390                 synchronize_rcu();
4391                 if (atomic_read(&rdev->nr_pending)) {
4392                         /* lost the race, try later */
4393                         err = -EBUSY;
4394                         p->rdev = rdev;
4395                 }
4396         }
4397 abort:
4398
4399         print_raid5_conf(conf);
4400         return err;
4401 }
4402
4403 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4404 {
4405         raid5_conf_t *conf = mddev->private;
4406         int found = 0;
4407         int disk;
4408         struct disk_info *p;
4409
4410         if (mddev->degraded > conf->max_degraded)
4411                 /* no point adding a device */
4412                 return 0;
4413
4414         /*
4415          * find the disk ... but prefer rdev->saved_raid_disk
4416          * if possible.
4417          */
4418         if (rdev->saved_raid_disk >= 0 &&
4419             conf->disks[rdev->saved_raid_disk].rdev == NULL)
4420                 disk = rdev->saved_raid_disk;
4421         else
4422                 disk = 0;
4423         for ( ; disk < conf->raid_disks; disk++)
4424                 if ((p=conf->disks + disk)->rdev == NULL) {
4425                         clear_bit(In_sync, &rdev->flags);
4426                         rdev->raid_disk = disk;
4427                         found = 1;
4428                         if (rdev->saved_raid_disk != disk)
4429                                 conf->fullsync = 1;
4430                         rcu_assign_pointer(p->rdev, rdev);
4431                         break;
4432                 }
4433         print_raid5_conf(conf);
4434         return found;
4435 }
4436
4437 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4438 {
4439         /* no resync is happening, and there is enough space
4440          * on all devices, so we can resize.
4441          * We need to make sure resync covers any new space.
4442          * If the array is shrinking we should possibly wait until
4443          * any io in the removed space completes, but it hardly seems
4444          * worth it.
4445          */
4446         raid5_conf_t *conf = mddev_to_conf(mddev);
4447
4448         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4449         mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4450         set_capacity(mddev->gendisk, mddev->array_size << 1);
4451         mddev->changed = 1;
4452         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
4453                 mddev->recovery_cp = mddev->size << 1;
4454                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4455         }
4456         mddev->size = sectors /2;
4457         mddev->resync_max_sectors = sectors;
4458         return 0;
4459 }
4460
4461 #ifdef CONFIG_MD_RAID5_RESHAPE
4462 static int raid5_check_reshape(mddev_t *mddev)
4463 {
4464         raid5_conf_t *conf = mddev_to_conf(mddev);
4465         int err;
4466
4467         if (mddev->delta_disks < 0 ||
4468             mddev->new_level != mddev->level)
4469                 return -EINVAL; /* Cannot shrink array or change level yet */
4470         if (mddev->delta_disks == 0)
4471                 return 0; /* nothing to do */
4472
4473         /* Can only proceed if there are plenty of stripe_heads.
4474          * We need a minimum of one full stripe,, and for sensible progress
4475          * it is best to have about 4 times that.
4476          * If we require 4 times, then the default 256 4K stripe_heads will
4477          * allow for chunk sizes up to 256K, which is probably OK.
4478          * If the chunk size is greater, user-space should request more
4479          * stripe_heads first.
4480          */
4481         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4482             (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4483                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
4484                        (mddev->chunk_size / STRIPE_SIZE)*4);
4485                 return -ENOSPC;
4486         }
4487
4488         err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4489         if (err)
4490                 return err;
4491
4492         if (mddev->degraded > conf->max_degraded)
4493                 return -EINVAL;
4494         /* looks like we might be able to manage this */
4495         return 0;
4496 }
4497
4498 static int raid5_start_reshape(mddev_t *mddev)
4499 {
4500         raid5_conf_t *conf = mddev_to_conf(mddev);
4501         mdk_rdev_t *rdev;
4502         struct list_head *rtmp;
4503         int spares = 0;
4504         int added_devices = 0;
4505         unsigned long flags;
4506
4507         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4508                 return -EBUSY;
4509
4510         ITERATE_RDEV(mddev, rdev, rtmp)
4511                 if (rdev->raid_disk < 0 &&
4512                     !test_bit(Faulty, &rdev->flags))
4513                         spares++;
4514
4515         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4516                 /* Not enough devices even to make a degraded array
4517                  * of that size
4518                  */
4519                 return -EINVAL;
4520
4521         atomic_set(&conf->reshape_stripes, 0);
4522         spin_lock_irq(&conf->device_lock);
4523         conf->previous_raid_disks = conf->raid_disks;
4524         conf->raid_disks += mddev->delta_disks;
4525         conf->expand_progress = 0;
4526         conf->expand_lo = 0;
4527         spin_unlock_irq(&conf->device_lock);
4528
4529         /* Add some new drives, as many as will fit.
4530          * We know there are enough to make the newly sized array work.
4531          */
4532         ITERATE_RDEV(mddev, rdev, rtmp)
4533                 if (rdev->raid_disk < 0 &&
4534                     !test_bit(Faulty, &rdev->flags)) {
4535                         if (raid5_add_disk(mddev, rdev)) {
4536                                 char nm[20];
4537                                 set_bit(In_sync, &rdev->flags);
4538                                 added_devices++;
4539                                 rdev->recovery_offset = 0;
4540                                 sprintf(nm, "rd%d", rdev->raid_disk);
4541                                 if (sysfs_create_link(&mddev->kobj,
4542                                                       &rdev->kobj, nm))
4543                                         printk(KERN_WARNING
4544                                                "raid5: failed to create "
4545                                                " link %s for %s\n",
4546                                                nm, mdname(mddev));
4547                         } else
4548                                 break;
4549                 }
4550
4551         spin_lock_irqsave(&conf->device_lock, flags);
4552         mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4553         spin_unlock_irqrestore(&conf->device_lock, flags);
4554         mddev->raid_disks = conf->raid_disks;
4555         mddev->reshape_position = 0;
4556         set_bit(MD_CHANGE_DEVS, &mddev->flags);
4557
4558         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4559         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4560         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4561         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4562         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4563                                                 "%s_reshape");
4564         if (!mddev->sync_thread) {
4565                 mddev->recovery = 0;
4566                 spin_lock_irq(&conf->device_lock);
4567                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4568                 conf->expand_progress = MaxSector;
4569                 spin_unlock_irq(&conf->device_lock);
4570                 return -EAGAIN;
4571         }
4572         md_wakeup_thread(mddev->sync_thread);
4573         md_new_event(mddev);
4574         return 0;
4575 }
4576 #endif
4577
4578 static void end_reshape(raid5_conf_t *conf)
4579 {
4580         struct block_device *bdev;
4581
4582         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4583                 conf->mddev->array_size = conf->mddev->size *
4584                         (conf->raid_disks - conf->max_degraded);
4585                 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4586                 conf->mddev->changed = 1;
4587
4588                 bdev = bdget_disk(conf->mddev->gendisk, 0);
4589                 if (bdev) {
4590                         mutex_lock(&bdev->bd_inode->i_mutex);
4591                         i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4592                         mutex_unlock(&bdev->bd_inode->i_mutex);
4593                         bdput(bdev);
4594                 }
4595                 spin_lock_irq(&conf->device_lock);
4596                 conf->expand_progress = MaxSector;
4597                 spin_unlock_irq(&conf->device_lock);
4598                 conf->mddev->reshape_position = MaxSector;
4599
4600                 /* read-ahead size must cover two whole stripes, which is
4601                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4602                  */
4603                 {
4604                         int data_disks = conf->previous_raid_disks - conf->max_degraded;
4605                         int stripe = data_disks *
4606                                 (conf->mddev->chunk_size / PAGE_SIZE);
4607                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4608                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4609                 }
4610         }
4611 }
4612
4613 static void raid5_quiesce(mddev_t *mddev, int state)
4614 {
4615         raid5_conf_t *conf = mddev_to_conf(mddev);
4616
4617         switch(state) {
4618         case 2: /* resume for a suspend */
4619                 wake_up(&conf->wait_for_overlap);
4620                 break;
4621
4622         case 1: /* stop all writes */
4623                 spin_lock_irq(&conf->device_lock);
4624                 conf->quiesce = 1;
4625                 wait_event_lock_irq(conf->wait_for_stripe,
4626                                     atomic_read(&conf->active_stripes) == 0 &&
4627                                     atomic_read(&conf->active_aligned_reads) == 0,
4628                                     conf->device_lock, /* nothing */);
4629                 spin_unlock_irq(&conf->device_lock);
4630                 break;
4631
4632         case 0: /* re-enable writes */
4633                 spin_lock_irq(&conf->device_lock);
4634                 conf->quiesce = 0;
4635                 wake_up(&conf->wait_for_stripe);
4636                 wake_up(&conf->wait_for_overlap);
4637                 spin_unlock_irq(&conf->device_lock);
4638                 break;
4639         }
4640 }
4641
4642 static struct mdk_personality raid6_personality =
4643 {
4644         .name           = "raid6",
4645         .level          = 6,
4646         .owner          = THIS_MODULE,
4647         .make_request   = make_request,
4648         .run            = run,
4649         .stop           = stop,
4650         .status         = status,
4651         .error_handler  = error,
4652         .hot_add_disk   = raid5_add_disk,
4653         .hot_remove_disk= raid5_remove_disk,
4654         .spare_active   = raid5_spare_active,
4655         .sync_request   = sync_request,
4656         .resize         = raid5_resize,
4657 #ifdef CONFIG_MD_RAID5_RESHAPE
4658         .check_reshape  = raid5_check_reshape,
4659         .start_reshape  = raid5_start_reshape,
4660 #endif
4661         .quiesce        = raid5_quiesce,
4662 };
4663 static struct mdk_personality raid5_personality =
4664 {
4665         .name           = "raid5",
4666         .level          = 5,
4667         .owner          = THIS_MODULE,
4668         .make_request   = make_request,
4669         .run            = run,
4670         .stop           = stop,
4671         .status         = status,
4672         .error_handler  = error,
4673         .hot_add_disk   = raid5_add_disk,
4674         .hot_remove_disk= raid5_remove_disk,
4675         .spare_active   = raid5_spare_active,
4676         .sync_request   = sync_request,
4677         .resize         = raid5_resize,
4678 #ifdef CONFIG_MD_RAID5_RESHAPE
4679         .check_reshape  = raid5_check_reshape,
4680         .start_reshape  = raid5_start_reshape,
4681 #endif
4682         .quiesce        = raid5_quiesce,
4683 };
4684
4685 static struct mdk_personality raid4_personality =
4686 {
4687         .name           = "raid4",
4688         .level          = 4,
4689         .owner          = THIS_MODULE,
4690         .make_request   = make_request,
4691         .run            = run,
4692         .stop           = stop,
4693         .status         = status,
4694         .error_handler  = error,
4695         .hot_add_disk   = raid5_add_disk,
4696         .hot_remove_disk= raid5_remove_disk,
4697         .spare_active   = raid5_spare_active,
4698         .sync_request   = sync_request,
4699         .resize         = raid5_resize,
4700 #ifdef CONFIG_MD_RAID5_RESHAPE
4701         .check_reshape  = raid5_check_reshape,
4702         .start_reshape  = raid5_start_reshape,
4703 #endif
4704         .quiesce        = raid5_quiesce,
4705 };
4706
4707 static int __init raid5_init(void)
4708 {
4709         int e;
4710
4711         e = raid6_select_algo();
4712         if ( e )
4713                 return e;
4714         register_md_personality(&raid6_personality);
4715         register_md_personality(&raid5_personality);
4716         register_md_personality(&raid4_personality);
4717         return 0;
4718 }
4719
4720 static void raid5_exit(void)
4721 {
4722         unregister_md_personality(&raid6_personality);
4723         unregister_md_personality(&raid5_personality);
4724         unregister_md_personality(&raid4_personality);
4725 }
4726
4727 module_init(raid5_init);
4728 module_exit(raid5_exit);
4729 MODULE_LICENSE("GPL");
4730 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4731 MODULE_ALIAS("md-raid5");
4732 MODULE_ALIAS("md-raid4");
4733 MODULE_ALIAS("md-level-5");
4734 MODULE_ALIAS("md-level-4");
4735 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4736 MODULE_ALIAS("md-raid6");
4737 MODULE_ALIAS("md-level-6");
4738
4739 /* This used to be two separate modules, they were: */
4740 MODULE_ALIAS("raid5");
4741 MODULE_ALIAS("raid6");