Merge branch 'for-2.6.26' of master.kernel.org:/pub/scm/linux/kernel/git/olof/pasemi
[linux-2.6] / 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         if (unlikely(sh->ops.count < 0)) {
380                 printk(KERN_ERR "pending: %#lx ops.pending: %#lx ops.ack: %#lx "
381                         "ops.complete: %#lx\n", pending, sh->ops.pending,
382                         sh->ops.ack, sh->ops.complete);
383                 BUG();
384         }
385
386         return pending;
387 }
388
389 static void
390 raid5_end_read_request(struct bio *bi, int error);
391 static void
392 raid5_end_write_request(struct bio *bi, int error);
393
394 static void ops_run_io(struct stripe_head *sh)
395 {
396         raid5_conf_t *conf = sh->raid_conf;
397         int i, disks = sh->disks;
398
399         might_sleep();
400
401         for (i = disks; i--; ) {
402                 int rw;
403                 struct bio *bi;
404                 mdk_rdev_t *rdev;
405                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
406                         rw = WRITE;
407                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
408                         rw = READ;
409                 else
410                         continue;
411
412                 bi = &sh->dev[i].req;
413
414                 bi->bi_rw = rw;
415                 if (rw == WRITE)
416                         bi->bi_end_io = raid5_end_write_request;
417                 else
418                         bi->bi_end_io = raid5_end_read_request;
419
420                 rcu_read_lock();
421                 rdev = rcu_dereference(conf->disks[i].rdev);
422                 if (rdev && test_bit(Faulty, &rdev->flags))
423                         rdev = NULL;
424                 if (rdev)
425                         atomic_inc(&rdev->nr_pending);
426                 rcu_read_unlock();
427
428                 if (rdev) {
429                         if (test_bit(STRIPE_SYNCING, &sh->state) ||
430                                 test_bit(STRIPE_EXPAND_SOURCE, &sh->state) ||
431                                 test_bit(STRIPE_EXPAND_READY, &sh->state))
432                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
433
434                         bi->bi_bdev = rdev->bdev;
435                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
436                                 __FUNCTION__, (unsigned long long)sh->sector,
437                                 bi->bi_rw, i);
438                         atomic_inc(&sh->count);
439                         bi->bi_sector = sh->sector + rdev->data_offset;
440                         bi->bi_flags = 1 << BIO_UPTODATE;
441                         bi->bi_vcnt = 1;
442                         bi->bi_max_vecs = 1;
443                         bi->bi_idx = 0;
444                         bi->bi_io_vec = &sh->dev[i].vec;
445                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
446                         bi->bi_io_vec[0].bv_offset = 0;
447                         bi->bi_size = STRIPE_SIZE;
448                         bi->bi_next = NULL;
449                         if (rw == WRITE &&
450                             test_bit(R5_ReWrite, &sh->dev[i].flags))
451                                 atomic_add(STRIPE_SECTORS,
452                                         &rdev->corrected_errors);
453                         generic_make_request(bi);
454                 } else {
455                         if (rw == WRITE)
456                                 set_bit(STRIPE_DEGRADED, &sh->state);
457                         pr_debug("skip op %ld on disc %d for sector %llu\n",
458                                 bi->bi_rw, i, (unsigned long long)sh->sector);
459                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
460                         set_bit(STRIPE_HANDLE, &sh->state);
461                 }
462         }
463 }
464
465 static struct dma_async_tx_descriptor *
466 async_copy_data(int frombio, struct bio *bio, struct page *page,
467         sector_t sector, struct dma_async_tx_descriptor *tx)
468 {
469         struct bio_vec *bvl;
470         struct page *bio_page;
471         int i;
472         int page_offset;
473
474         if (bio->bi_sector >= sector)
475                 page_offset = (signed)(bio->bi_sector - sector) * 512;
476         else
477                 page_offset = (signed)(sector - bio->bi_sector) * -512;
478         bio_for_each_segment(bvl, bio, i) {
479                 int len = bio_iovec_idx(bio, i)->bv_len;
480                 int clen;
481                 int b_offset = 0;
482
483                 if (page_offset < 0) {
484                         b_offset = -page_offset;
485                         page_offset += b_offset;
486                         len -= b_offset;
487                 }
488
489                 if (len > 0 && page_offset + len > STRIPE_SIZE)
490                         clen = STRIPE_SIZE - page_offset;
491                 else
492                         clen = len;
493
494                 if (clen > 0) {
495                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
496                         bio_page = bio_iovec_idx(bio, i)->bv_page;
497                         if (frombio)
498                                 tx = async_memcpy(page, bio_page, page_offset,
499                                         b_offset, clen,
500                                         ASYNC_TX_DEP_ACK,
501                                         tx, NULL, NULL);
502                         else
503                                 tx = async_memcpy(bio_page, page, b_offset,
504                                         page_offset, clen,
505                                         ASYNC_TX_DEP_ACK,
506                                         tx, NULL, NULL);
507                 }
508                 if (clen < len) /* hit end of page */
509                         break;
510                 page_offset +=  len;
511         }
512
513         return tx;
514 }
515
516 static void ops_complete_biofill(void *stripe_head_ref)
517 {
518         struct stripe_head *sh = stripe_head_ref;
519         struct bio *return_bi = NULL;
520         raid5_conf_t *conf = sh->raid_conf;
521         int i;
522
523         pr_debug("%s: stripe %llu\n", __FUNCTION__,
524                 (unsigned long long)sh->sector);
525
526         /* clear completed biofills */
527         for (i = sh->disks; i--; ) {
528                 struct r5dev *dev = &sh->dev[i];
529
530                 /* acknowledge completion of a biofill operation */
531                 /* and check if we need to reply to a read request,
532                  * new R5_Wantfill requests are held off until
533                  * !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending)
534                  */
535                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
536                         struct bio *rbi, *rbi2;
537
538                         /* The access to dev->read is outside of the
539                          * spin_lock_irq(&conf->device_lock), but is protected
540                          * by the STRIPE_OP_BIOFILL pending bit
541                          */
542                         BUG_ON(!dev->read);
543                         rbi = dev->read;
544                         dev->read = NULL;
545                         while (rbi && rbi->bi_sector <
546                                 dev->sector + STRIPE_SECTORS) {
547                                 rbi2 = r5_next_bio(rbi, dev->sector);
548                                 spin_lock_irq(&conf->device_lock);
549                                 if (--rbi->bi_phys_segments == 0) {
550                                         rbi->bi_next = return_bi;
551                                         return_bi = rbi;
552                                 }
553                                 spin_unlock_irq(&conf->device_lock);
554                                 rbi = rbi2;
555                         }
556                 }
557         }
558         set_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
559
560         return_io(return_bi);
561
562         set_bit(STRIPE_HANDLE, &sh->state);
563         release_stripe(sh);
564 }
565
566 static void ops_run_biofill(struct stripe_head *sh)
567 {
568         struct dma_async_tx_descriptor *tx = NULL;
569         raid5_conf_t *conf = sh->raid_conf;
570         int i;
571
572         pr_debug("%s: stripe %llu\n", __FUNCTION__,
573                 (unsigned long long)sh->sector);
574
575         for (i = sh->disks; i--; ) {
576                 struct r5dev *dev = &sh->dev[i];
577                 if (test_bit(R5_Wantfill, &dev->flags)) {
578                         struct bio *rbi;
579                         spin_lock_irq(&conf->device_lock);
580                         dev->read = rbi = dev->toread;
581                         dev->toread = NULL;
582                         spin_unlock_irq(&conf->device_lock);
583                         while (rbi && rbi->bi_sector <
584                                 dev->sector + STRIPE_SECTORS) {
585                                 tx = async_copy_data(0, rbi, dev->page,
586                                         dev->sector, tx);
587                                 rbi = r5_next_bio(rbi, dev->sector);
588                         }
589                 }
590         }
591
592         atomic_inc(&sh->count);
593         async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
594                 ops_complete_biofill, sh);
595 }
596
597 static void ops_complete_compute5(void *stripe_head_ref)
598 {
599         struct stripe_head *sh = stripe_head_ref;
600         int target = sh->ops.target;
601         struct r5dev *tgt = &sh->dev[target];
602
603         pr_debug("%s: stripe %llu\n", __FUNCTION__,
604                 (unsigned long long)sh->sector);
605
606         set_bit(R5_UPTODATE, &tgt->flags);
607         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
608         clear_bit(R5_Wantcompute, &tgt->flags);
609         set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
610         set_bit(STRIPE_HANDLE, &sh->state);
611         release_stripe(sh);
612 }
613
614 static struct dma_async_tx_descriptor *
615 ops_run_compute5(struct stripe_head *sh, unsigned long pending)
616 {
617         /* kernel stack size limits the total number of disks */
618         int disks = sh->disks;
619         struct page *xor_srcs[disks];
620         int target = sh->ops.target;
621         struct r5dev *tgt = &sh->dev[target];
622         struct page *xor_dest = tgt->page;
623         int count = 0;
624         struct dma_async_tx_descriptor *tx;
625         int i;
626
627         pr_debug("%s: stripe %llu block: %d\n",
628                 __FUNCTION__, (unsigned long long)sh->sector, target);
629         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
630
631         for (i = disks; i--; )
632                 if (i != target)
633                         xor_srcs[count++] = sh->dev[i].page;
634
635         atomic_inc(&sh->count);
636
637         if (unlikely(count == 1))
638                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
639                         0, NULL, ops_complete_compute5, sh);
640         else
641                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
642                         ASYNC_TX_XOR_ZERO_DST, NULL,
643                         ops_complete_compute5, sh);
644
645         /* ack now if postxor is not set to be run */
646         if (tx && !test_bit(STRIPE_OP_POSTXOR, &pending))
647                 async_tx_ack(tx);
648
649         return tx;
650 }
651
652 static void ops_complete_prexor(void *stripe_head_ref)
653 {
654         struct stripe_head *sh = stripe_head_ref;
655
656         pr_debug("%s: stripe %llu\n", __FUNCTION__,
657                 (unsigned long long)sh->sector);
658
659         set_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
660 }
661
662 static struct dma_async_tx_descriptor *
663 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
664 {
665         /* kernel stack size limits the total number of disks */
666         int disks = sh->disks;
667         struct page *xor_srcs[disks];
668         int count = 0, pd_idx = sh->pd_idx, i;
669
670         /* existing parity data subtracted */
671         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
672
673         pr_debug("%s: stripe %llu\n", __FUNCTION__,
674                 (unsigned long long)sh->sector);
675
676         for (i = disks; i--; ) {
677                 struct r5dev *dev = &sh->dev[i];
678                 /* Only process blocks that are known to be uptodate */
679                 if (dev->towrite && test_bit(R5_Wantprexor, &dev->flags))
680                         xor_srcs[count++] = dev->page;
681         }
682
683         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
684                 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
685                 ops_complete_prexor, sh);
686
687         return tx;
688 }
689
690 static struct dma_async_tx_descriptor *
691 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx,
692                  unsigned long pending)
693 {
694         int disks = sh->disks;
695         int pd_idx = sh->pd_idx, i;
696
697         /* check if prexor is active which means only process blocks
698          * that are part of a read-modify-write (Wantprexor)
699          */
700         int prexor = test_bit(STRIPE_OP_PREXOR, &pending);
701
702         pr_debug("%s: stripe %llu\n", __FUNCTION__,
703                 (unsigned long long)sh->sector);
704
705         for (i = disks; i--; ) {
706                 struct r5dev *dev = &sh->dev[i];
707                 struct bio *chosen;
708                 int towrite;
709
710                 towrite = 0;
711                 if (prexor) { /* rmw */
712                         if (dev->towrite &&
713                             test_bit(R5_Wantprexor, &dev->flags))
714                                 towrite = 1;
715                 } else { /* rcw */
716                         if (i != pd_idx && dev->towrite &&
717                                 test_bit(R5_LOCKED, &dev->flags))
718                                 towrite = 1;
719                 }
720
721                 if (towrite) {
722                         struct bio *wbi;
723
724                         spin_lock(&sh->lock);
725                         chosen = dev->towrite;
726                         dev->towrite = NULL;
727                         BUG_ON(dev->written);
728                         wbi = dev->written = chosen;
729                         spin_unlock(&sh->lock);
730
731                         while (wbi && wbi->bi_sector <
732                                 dev->sector + STRIPE_SECTORS) {
733                                 tx = async_copy_data(1, wbi, dev->page,
734                                         dev->sector, tx);
735                                 wbi = r5_next_bio(wbi, dev->sector);
736                         }
737                 }
738         }
739
740         return tx;
741 }
742
743 static void ops_complete_postxor(void *stripe_head_ref)
744 {
745         struct stripe_head *sh = stripe_head_ref;
746
747         pr_debug("%s: stripe %llu\n", __FUNCTION__,
748                 (unsigned long long)sh->sector);
749
750         set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
751         set_bit(STRIPE_HANDLE, &sh->state);
752         release_stripe(sh);
753 }
754
755 static void ops_complete_write(void *stripe_head_ref)
756 {
757         struct stripe_head *sh = stripe_head_ref;
758         int disks = sh->disks, i, pd_idx = sh->pd_idx;
759
760         pr_debug("%s: stripe %llu\n", __FUNCTION__,
761                 (unsigned long long)sh->sector);
762
763         for (i = disks; i--; ) {
764                 struct r5dev *dev = &sh->dev[i];
765                 if (dev->written || i == pd_idx)
766                         set_bit(R5_UPTODATE, &dev->flags);
767         }
768
769         set_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
770         set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
771
772         set_bit(STRIPE_HANDLE, &sh->state);
773         release_stripe(sh);
774 }
775
776 static void
777 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx,
778                 unsigned long pending)
779 {
780         /* kernel stack size limits the total number of disks */
781         int disks = sh->disks;
782         struct page *xor_srcs[disks];
783
784         int count = 0, pd_idx = sh->pd_idx, i;
785         struct page *xor_dest;
786         int prexor = test_bit(STRIPE_OP_PREXOR, &pending);
787         unsigned long flags;
788         dma_async_tx_callback callback;
789
790         pr_debug("%s: stripe %llu\n", __FUNCTION__,
791                 (unsigned long long)sh->sector);
792
793         /* check if prexor is active which means only process blocks
794          * that are part of a read-modify-write (written)
795          */
796         if (prexor) {
797                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
798                 for (i = disks; i--; ) {
799                         struct r5dev *dev = &sh->dev[i];
800                         if (dev->written)
801                                 xor_srcs[count++] = dev->page;
802                 }
803         } else {
804                 xor_dest = sh->dev[pd_idx].page;
805                 for (i = disks; i--; ) {
806                         struct r5dev *dev = &sh->dev[i];
807                         if (i != pd_idx)
808                                 xor_srcs[count++] = dev->page;
809                 }
810         }
811
812         /* check whether this postxor is part of a write */
813         callback = test_bit(STRIPE_OP_BIODRAIN, &pending) ?
814                 ops_complete_write : ops_complete_postxor;
815
816         /* 1/ if we prexor'd then the dest is reused as a source
817          * 2/ if we did not prexor then we are redoing the parity
818          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
819          * for the synchronous xor case
820          */
821         flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
822                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
823
824         atomic_inc(&sh->count);
825
826         if (unlikely(count == 1)) {
827                 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
828                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
829                         flags, tx, callback, sh);
830         } else
831                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
832                         flags, tx, callback, sh);
833 }
834
835 static void ops_complete_check(void *stripe_head_ref)
836 {
837         struct stripe_head *sh = stripe_head_ref;
838         int pd_idx = sh->pd_idx;
839
840         pr_debug("%s: stripe %llu\n", __FUNCTION__,
841                 (unsigned long long)sh->sector);
842
843         if (test_and_clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending) &&
844                 sh->ops.zero_sum_result == 0)
845                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
846
847         set_bit(STRIPE_OP_CHECK, &sh->ops.complete);
848         set_bit(STRIPE_HANDLE, &sh->state);
849         release_stripe(sh);
850 }
851
852 static void ops_run_check(struct stripe_head *sh)
853 {
854         /* kernel stack size limits the total number of disks */
855         int disks = sh->disks;
856         struct page *xor_srcs[disks];
857         struct dma_async_tx_descriptor *tx;
858
859         int count = 0, pd_idx = sh->pd_idx, i;
860         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
861
862         pr_debug("%s: stripe %llu\n", __FUNCTION__,
863                 (unsigned long long)sh->sector);
864
865         for (i = disks; i--; ) {
866                 struct r5dev *dev = &sh->dev[i];
867                 if (i != pd_idx)
868                         xor_srcs[count++] = dev->page;
869         }
870
871         tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
872                 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
873
874         if (tx)
875                 set_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
876         else
877                 clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
878
879         atomic_inc(&sh->count);
880         tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
881                 ops_complete_check, sh);
882 }
883
884 static void raid5_run_ops(struct stripe_head *sh, unsigned long pending)
885 {
886         int overlap_clear = 0, i, disks = sh->disks;
887         struct dma_async_tx_descriptor *tx = NULL;
888
889         if (test_bit(STRIPE_OP_BIOFILL, &pending)) {
890                 ops_run_biofill(sh);
891                 overlap_clear++;
892         }
893
894         if (test_bit(STRIPE_OP_COMPUTE_BLK, &pending))
895                 tx = ops_run_compute5(sh, pending);
896
897         if (test_bit(STRIPE_OP_PREXOR, &pending))
898                 tx = ops_run_prexor(sh, tx);
899
900         if (test_bit(STRIPE_OP_BIODRAIN, &pending)) {
901                 tx = ops_run_biodrain(sh, tx, pending);
902                 overlap_clear++;
903         }
904
905         if (test_bit(STRIPE_OP_POSTXOR, &pending))
906                 ops_run_postxor(sh, tx, pending);
907
908         if (test_bit(STRIPE_OP_CHECK, &pending))
909                 ops_run_check(sh);
910
911         if (test_bit(STRIPE_OP_IO, &pending))
912                 ops_run_io(sh);
913
914         if (overlap_clear)
915                 for (i = disks; i--; ) {
916                         struct r5dev *dev = &sh->dev[i];
917                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
918                                 wake_up(&sh->raid_conf->wait_for_overlap);
919                 }
920 }
921
922 static int grow_one_stripe(raid5_conf_t *conf)
923 {
924         struct stripe_head *sh;
925         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
926         if (!sh)
927                 return 0;
928         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
929         sh->raid_conf = conf;
930         spin_lock_init(&sh->lock);
931
932         if (grow_buffers(sh, conf->raid_disks)) {
933                 shrink_buffers(sh, conf->raid_disks);
934                 kmem_cache_free(conf->slab_cache, sh);
935                 return 0;
936         }
937         sh->disks = conf->raid_disks;
938         /* we just created an active stripe so... */
939         atomic_set(&sh->count, 1);
940         atomic_inc(&conf->active_stripes);
941         INIT_LIST_HEAD(&sh->lru);
942         release_stripe(sh);
943         return 1;
944 }
945
946 static int grow_stripes(raid5_conf_t *conf, int num)
947 {
948         struct kmem_cache *sc;
949         int devs = conf->raid_disks;
950
951         sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
952         sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
953         conf->active_name = 0;
954         sc = kmem_cache_create(conf->cache_name[conf->active_name],
955                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
956                                0, 0, NULL);
957         if (!sc)
958                 return 1;
959         conf->slab_cache = sc;
960         conf->pool_size = devs;
961         while (num--)
962                 if (!grow_one_stripe(conf))
963                         return 1;
964         return 0;
965 }
966
967 #ifdef CONFIG_MD_RAID5_RESHAPE
968 static int resize_stripes(raid5_conf_t *conf, int newsize)
969 {
970         /* Make all the stripes able to hold 'newsize' devices.
971          * New slots in each stripe get 'page' set to a new page.
972          *
973          * This happens in stages:
974          * 1/ create a new kmem_cache and allocate the required number of
975          *    stripe_heads.
976          * 2/ gather all the old stripe_heads and tranfer the pages across
977          *    to the new stripe_heads.  This will have the side effect of
978          *    freezing the array as once all stripe_heads have been collected,
979          *    no IO will be possible.  Old stripe heads are freed once their
980          *    pages have been transferred over, and the old kmem_cache is
981          *    freed when all stripes are done.
982          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
983          *    we simple return a failre status - no need to clean anything up.
984          * 4/ allocate new pages for the new slots in the new stripe_heads.
985          *    If this fails, we don't bother trying the shrink the
986          *    stripe_heads down again, we just leave them as they are.
987          *    As each stripe_head is processed the new one is released into
988          *    active service.
989          *
990          * Once step2 is started, we cannot afford to wait for a write,
991          * so we use GFP_NOIO allocations.
992          */
993         struct stripe_head *osh, *nsh;
994         LIST_HEAD(newstripes);
995         struct disk_info *ndisks;
996         int err = 0;
997         struct kmem_cache *sc;
998         int i;
999
1000         if (newsize <= conf->pool_size)
1001                 return 0; /* never bother to shrink */
1002
1003         md_allow_write(conf->mddev);
1004
1005         /* Step 1 */
1006         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1007                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1008                                0, 0, NULL);
1009         if (!sc)
1010                 return -ENOMEM;
1011
1012         for (i = conf->max_nr_stripes; i; i--) {
1013                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1014                 if (!nsh)
1015                         break;
1016
1017                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1018
1019                 nsh->raid_conf = conf;
1020                 spin_lock_init(&nsh->lock);
1021
1022                 list_add(&nsh->lru, &newstripes);
1023         }
1024         if (i) {
1025                 /* didn't get enough, give up */
1026                 while (!list_empty(&newstripes)) {
1027                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1028                         list_del(&nsh->lru);
1029                         kmem_cache_free(sc, nsh);
1030                 }
1031                 kmem_cache_destroy(sc);
1032                 return -ENOMEM;
1033         }
1034         /* Step 2 - Must use GFP_NOIO now.
1035          * OK, we have enough stripes, start collecting inactive
1036          * stripes and copying them over
1037          */
1038         list_for_each_entry(nsh, &newstripes, lru) {
1039                 spin_lock_irq(&conf->device_lock);
1040                 wait_event_lock_irq(conf->wait_for_stripe,
1041                                     !list_empty(&conf->inactive_list),
1042                                     conf->device_lock,
1043                                     unplug_slaves(conf->mddev)
1044                         );
1045                 osh = get_free_stripe(conf);
1046                 spin_unlock_irq(&conf->device_lock);
1047                 atomic_set(&nsh->count, 1);
1048                 for(i=0; i<conf->pool_size; i++)
1049                         nsh->dev[i].page = osh->dev[i].page;
1050                 for( ; i<newsize; i++)
1051                         nsh->dev[i].page = NULL;
1052                 kmem_cache_free(conf->slab_cache, osh);
1053         }
1054         kmem_cache_destroy(conf->slab_cache);
1055
1056         /* Step 3.
1057          * At this point, we are holding all the stripes so the array
1058          * is completely stalled, so now is a good time to resize
1059          * conf->disks.
1060          */
1061         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1062         if (ndisks) {
1063                 for (i=0; i<conf->raid_disks; i++)
1064                         ndisks[i] = conf->disks[i];
1065                 kfree(conf->disks);
1066                 conf->disks = ndisks;
1067         } else
1068                 err = -ENOMEM;
1069
1070         /* Step 4, return new stripes to service */
1071         while(!list_empty(&newstripes)) {
1072                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1073                 list_del_init(&nsh->lru);
1074                 for (i=conf->raid_disks; i < newsize; i++)
1075                         if (nsh->dev[i].page == NULL) {
1076                                 struct page *p = alloc_page(GFP_NOIO);
1077                                 nsh->dev[i].page = p;
1078                                 if (!p)
1079                                         err = -ENOMEM;
1080                         }
1081                 release_stripe(nsh);
1082         }
1083         /* critical section pass, GFP_NOIO no longer needed */
1084
1085         conf->slab_cache = sc;
1086         conf->active_name = 1-conf->active_name;
1087         conf->pool_size = newsize;
1088         return err;
1089 }
1090 #endif
1091
1092 static int drop_one_stripe(raid5_conf_t *conf)
1093 {
1094         struct stripe_head *sh;
1095
1096         spin_lock_irq(&conf->device_lock);
1097         sh = get_free_stripe(conf);
1098         spin_unlock_irq(&conf->device_lock);
1099         if (!sh)
1100                 return 0;
1101         BUG_ON(atomic_read(&sh->count));
1102         shrink_buffers(sh, conf->pool_size);
1103         kmem_cache_free(conf->slab_cache, sh);
1104         atomic_dec(&conf->active_stripes);
1105         return 1;
1106 }
1107
1108 static void shrink_stripes(raid5_conf_t *conf)
1109 {
1110         while (drop_one_stripe(conf))
1111                 ;
1112
1113         if (conf->slab_cache)
1114                 kmem_cache_destroy(conf->slab_cache);
1115         conf->slab_cache = NULL;
1116 }
1117
1118 static void raid5_end_read_request(struct bio * bi, int error)
1119 {
1120         struct stripe_head *sh = bi->bi_private;
1121         raid5_conf_t *conf = sh->raid_conf;
1122         int disks = sh->disks, i;
1123         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1124         char b[BDEVNAME_SIZE];
1125         mdk_rdev_t *rdev;
1126
1127
1128         for (i=0 ; i<disks; i++)
1129                 if (bi == &sh->dev[i].req)
1130                         break;
1131
1132         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1133                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1134                 uptodate);
1135         if (i == disks) {
1136                 BUG();
1137                 return;
1138         }
1139
1140         if (uptodate) {
1141                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1142                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1143                         rdev = conf->disks[i].rdev;
1144                         printk(KERN_INFO "raid5:%s: read error corrected (%lu sectors at %llu on %s)\n",
1145                                mdname(conf->mddev), STRIPE_SECTORS,
1146                                (unsigned long long)(sh->sector + rdev->data_offset),
1147                                bdevname(rdev->bdev, b));
1148                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1149                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1150                 }
1151                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1152                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1153         } else {
1154                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1155                 int retry = 0;
1156                 rdev = conf->disks[i].rdev;
1157
1158                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1159                 atomic_inc(&rdev->read_errors);
1160                 if (conf->mddev->degraded)
1161                         printk(KERN_WARNING "raid5:%s: read error not correctable (sector %llu on %s).\n",
1162                                mdname(conf->mddev),
1163                                (unsigned long long)(sh->sector + rdev->data_offset),
1164                                bdn);
1165                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1166                         /* Oh, no!!! */
1167                         printk(KERN_WARNING "raid5:%s: read error NOT corrected!! (sector %llu on %s).\n",
1168                                mdname(conf->mddev),
1169                                (unsigned long long)(sh->sector + rdev->data_offset),
1170                                bdn);
1171                 else if (atomic_read(&rdev->read_errors)
1172                          > conf->max_nr_stripes)
1173                         printk(KERN_WARNING
1174                                "raid5:%s: Too many read errors, failing device %s.\n",
1175                                mdname(conf->mddev), bdn);
1176                 else
1177                         retry = 1;
1178                 if (retry)
1179                         set_bit(R5_ReadError, &sh->dev[i].flags);
1180                 else {
1181                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1182                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1183                         md_error(conf->mddev, rdev);
1184                 }
1185         }
1186         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1187         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1188         set_bit(STRIPE_HANDLE, &sh->state);
1189         release_stripe(sh);
1190 }
1191
1192 static void raid5_end_write_request (struct bio *bi, int error)
1193 {
1194         struct stripe_head *sh = bi->bi_private;
1195         raid5_conf_t *conf = sh->raid_conf;
1196         int disks = sh->disks, i;
1197         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1198
1199         for (i=0 ; i<disks; i++)
1200                 if (bi == &sh->dev[i].req)
1201                         break;
1202
1203         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1204                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1205                 uptodate);
1206         if (i == disks) {
1207                 BUG();
1208                 return;
1209         }
1210
1211         if (!uptodate)
1212                 md_error(conf->mddev, conf->disks[i].rdev);
1213
1214         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1215         
1216         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1217         set_bit(STRIPE_HANDLE, &sh->state);
1218         release_stripe(sh);
1219 }
1220
1221
1222 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1223         
1224 static void raid5_build_block (struct stripe_head *sh, int i)
1225 {
1226         struct r5dev *dev = &sh->dev[i];
1227
1228         bio_init(&dev->req);
1229         dev->req.bi_io_vec = &dev->vec;
1230         dev->req.bi_vcnt++;
1231         dev->req.bi_max_vecs++;
1232         dev->vec.bv_page = dev->page;
1233         dev->vec.bv_len = STRIPE_SIZE;
1234         dev->vec.bv_offset = 0;
1235
1236         dev->req.bi_sector = sh->sector;
1237         dev->req.bi_private = sh;
1238
1239         dev->flags = 0;
1240         dev->sector = compute_blocknr(sh, i);
1241 }
1242
1243 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1244 {
1245         char b[BDEVNAME_SIZE];
1246         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1247         pr_debug("raid5: error called\n");
1248
1249         if (!test_bit(Faulty, &rdev->flags)) {
1250                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1251                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1252                         unsigned long flags;
1253                         spin_lock_irqsave(&conf->device_lock, flags);
1254                         mddev->degraded++;
1255                         spin_unlock_irqrestore(&conf->device_lock, flags);
1256                         /*
1257                          * if recovery was running, make sure it aborts.
1258                          */
1259                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
1260                 }
1261                 set_bit(Faulty, &rdev->flags);
1262                 printk (KERN_ALERT
1263                         "raid5: Disk failure on %s, disabling device."
1264                         " Operation continuing on %d devices\n",
1265                         bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1266         }
1267 }
1268
1269 /*
1270  * Input: a 'big' sector number,
1271  * Output: index of the data and parity disk, and the sector # in them.
1272  */
1273 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1274                         unsigned int data_disks, unsigned int * dd_idx,
1275                         unsigned int * pd_idx, raid5_conf_t *conf)
1276 {
1277         long stripe;
1278         unsigned long chunk_number;
1279         unsigned int chunk_offset;
1280         sector_t new_sector;
1281         int sectors_per_chunk = conf->chunk_size >> 9;
1282
1283         /* First compute the information on this sector */
1284
1285         /*
1286          * Compute the chunk number and the sector offset inside the chunk
1287          */
1288         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1289         chunk_number = r_sector;
1290         BUG_ON(r_sector != chunk_number);
1291
1292         /*
1293          * Compute the stripe number
1294          */
1295         stripe = chunk_number / data_disks;
1296
1297         /*
1298          * Compute the data disk and parity disk indexes inside the stripe
1299          */
1300         *dd_idx = chunk_number % data_disks;
1301
1302         /*
1303          * Select the parity disk based on the user selected algorithm.
1304          */
1305         switch(conf->level) {
1306         case 4:
1307                 *pd_idx = data_disks;
1308                 break;
1309         case 5:
1310                 switch (conf->algorithm) {
1311                 case ALGORITHM_LEFT_ASYMMETRIC:
1312                         *pd_idx = data_disks - stripe % raid_disks;
1313                         if (*dd_idx >= *pd_idx)
1314                                 (*dd_idx)++;
1315                         break;
1316                 case ALGORITHM_RIGHT_ASYMMETRIC:
1317                         *pd_idx = stripe % raid_disks;
1318                         if (*dd_idx >= *pd_idx)
1319                                 (*dd_idx)++;
1320                         break;
1321                 case ALGORITHM_LEFT_SYMMETRIC:
1322                         *pd_idx = data_disks - stripe % raid_disks;
1323                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1324                         break;
1325                 case ALGORITHM_RIGHT_SYMMETRIC:
1326                         *pd_idx = stripe % raid_disks;
1327                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1328                         break;
1329                 default:
1330                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1331                                 conf->algorithm);
1332                 }
1333                 break;
1334         case 6:
1335
1336                 /**** FIX THIS ****/
1337                 switch (conf->algorithm) {
1338                 case ALGORITHM_LEFT_ASYMMETRIC:
1339                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1340                         if (*pd_idx == raid_disks-1)
1341                                 (*dd_idx)++;    /* Q D D D P */
1342                         else if (*dd_idx >= *pd_idx)
1343                                 (*dd_idx) += 2; /* D D P Q D */
1344                         break;
1345                 case ALGORITHM_RIGHT_ASYMMETRIC:
1346                         *pd_idx = stripe % raid_disks;
1347                         if (*pd_idx == raid_disks-1)
1348                                 (*dd_idx)++;    /* Q D D D P */
1349                         else if (*dd_idx >= *pd_idx)
1350                                 (*dd_idx) += 2; /* D D P Q D */
1351                         break;
1352                 case ALGORITHM_LEFT_SYMMETRIC:
1353                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1354                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1355                         break;
1356                 case ALGORITHM_RIGHT_SYMMETRIC:
1357                         *pd_idx = stripe % raid_disks;
1358                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1359                         break;
1360                 default:
1361                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1362                                 conf->algorithm);
1363                 }
1364                 break;
1365         }
1366
1367         /*
1368          * Finally, compute the new sector number
1369          */
1370         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1371         return new_sector;
1372 }
1373
1374
1375 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1376 {
1377         raid5_conf_t *conf = sh->raid_conf;
1378         int raid_disks = sh->disks;
1379         int data_disks = raid_disks - conf->max_degraded;
1380         sector_t new_sector = sh->sector, check;
1381         int sectors_per_chunk = conf->chunk_size >> 9;
1382         sector_t stripe;
1383         int chunk_offset;
1384         int chunk_number, dummy1, dummy2, dd_idx = i;
1385         sector_t r_sector;
1386
1387
1388         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1389         stripe = new_sector;
1390         BUG_ON(new_sector != stripe);
1391
1392         if (i == sh->pd_idx)
1393                 return 0;
1394         switch(conf->level) {
1395         case 4: break;
1396         case 5:
1397                 switch (conf->algorithm) {
1398                 case ALGORITHM_LEFT_ASYMMETRIC:
1399                 case ALGORITHM_RIGHT_ASYMMETRIC:
1400                         if (i > sh->pd_idx)
1401                                 i--;
1402                         break;
1403                 case ALGORITHM_LEFT_SYMMETRIC:
1404                 case ALGORITHM_RIGHT_SYMMETRIC:
1405                         if (i < sh->pd_idx)
1406                                 i += raid_disks;
1407                         i -= (sh->pd_idx + 1);
1408                         break;
1409                 default:
1410                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1411                                conf->algorithm);
1412                 }
1413                 break;
1414         case 6:
1415                 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1416                         return 0; /* It is the Q disk */
1417                 switch (conf->algorithm) {
1418                 case ALGORITHM_LEFT_ASYMMETRIC:
1419                 case ALGORITHM_RIGHT_ASYMMETRIC:
1420                         if (sh->pd_idx == raid_disks-1)
1421                                 i--;    /* Q D D D P */
1422                         else if (i > sh->pd_idx)
1423                                 i -= 2; /* D D P Q D */
1424                         break;
1425                 case ALGORITHM_LEFT_SYMMETRIC:
1426                 case ALGORITHM_RIGHT_SYMMETRIC:
1427                         if (sh->pd_idx == raid_disks-1)
1428                                 i--; /* Q D D D P */
1429                         else {
1430                                 /* D D P Q D */
1431                                 if (i < sh->pd_idx)
1432                                         i += raid_disks;
1433                                 i -= (sh->pd_idx + 2);
1434                         }
1435                         break;
1436                 default:
1437                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1438                                 conf->algorithm);
1439                 }
1440                 break;
1441         }
1442
1443         chunk_number = stripe * data_disks + i;
1444         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1445
1446         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1447         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1448                 printk(KERN_ERR "compute_blocknr: map not correct\n");
1449                 return 0;
1450         }
1451         return r_sector;
1452 }
1453
1454
1455
1456 /*
1457  * Copy data between a page in the stripe cache, and one or more bion
1458  * The page could align with the middle of the bio, or there could be
1459  * several bion, each with several bio_vecs, which cover part of the page
1460  * Multiple bion are linked together on bi_next.  There may be extras
1461  * at the end of this list.  We ignore them.
1462  */
1463 static void copy_data(int frombio, struct bio *bio,
1464                      struct page *page,
1465                      sector_t sector)
1466 {
1467         char *pa = page_address(page);
1468         struct bio_vec *bvl;
1469         int i;
1470         int page_offset;
1471
1472         if (bio->bi_sector >= sector)
1473                 page_offset = (signed)(bio->bi_sector - sector) * 512;
1474         else
1475                 page_offset = (signed)(sector - bio->bi_sector) * -512;
1476         bio_for_each_segment(bvl, bio, i) {
1477                 int len = bio_iovec_idx(bio,i)->bv_len;
1478                 int clen;
1479                 int b_offset = 0;
1480
1481                 if (page_offset < 0) {
1482                         b_offset = -page_offset;
1483                         page_offset += b_offset;
1484                         len -= b_offset;
1485                 }
1486
1487                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1488                         clen = STRIPE_SIZE - page_offset;
1489                 else clen = len;
1490
1491                 if (clen > 0) {
1492                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1493                         if (frombio)
1494                                 memcpy(pa+page_offset, ba+b_offset, clen);
1495                         else
1496                                 memcpy(ba+b_offset, pa+page_offset, clen);
1497                         __bio_kunmap_atomic(ba, KM_USER0);
1498                 }
1499                 if (clen < len) /* hit end of page */
1500                         break;
1501                 page_offset +=  len;
1502         }
1503 }
1504
1505 #define check_xor()     do {                                              \
1506                                 if (count == MAX_XOR_BLOCKS) {            \
1507                                 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1508                                 count = 0;                                \
1509                            }                                              \
1510                         } while(0)
1511
1512 static void compute_parity6(struct stripe_head *sh, int method)
1513 {
1514         raid6_conf_t *conf = sh->raid_conf;
1515         int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1516         struct bio *chosen;
1517         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1518         void *ptrs[disks];
1519
1520         qd_idx = raid6_next_disk(pd_idx, disks);
1521         d0_idx = raid6_next_disk(qd_idx, disks);
1522
1523         pr_debug("compute_parity, stripe %llu, method %d\n",
1524                 (unsigned long long)sh->sector, method);
1525
1526         switch(method) {
1527         case READ_MODIFY_WRITE:
1528                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
1529         case RECONSTRUCT_WRITE:
1530                 for (i= disks; i-- ;)
1531                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1532                                 chosen = sh->dev[i].towrite;
1533                                 sh->dev[i].towrite = NULL;
1534
1535                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1536                                         wake_up(&conf->wait_for_overlap);
1537
1538                                 BUG_ON(sh->dev[i].written);
1539                                 sh->dev[i].written = chosen;
1540                         }
1541                 break;
1542         case CHECK_PARITY:
1543                 BUG();          /* Not implemented yet */
1544         }
1545
1546         for (i = disks; i--;)
1547                 if (sh->dev[i].written) {
1548                         sector_t sector = sh->dev[i].sector;
1549                         struct bio *wbi = sh->dev[i].written;
1550                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1551                                 copy_data(1, wbi, sh->dev[i].page, sector);
1552                                 wbi = r5_next_bio(wbi, sector);
1553                         }
1554
1555                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1556                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1557                 }
1558
1559 //      switch(method) {
1560 //      case RECONSTRUCT_WRITE:
1561 //      case CHECK_PARITY:
1562 //      case UPDATE_PARITY:
1563                 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1564                 /* FIX: Is this ordering of drives even remotely optimal? */
1565                 count = 0;
1566                 i = d0_idx;
1567                 do {
1568                         ptrs[count++] = page_address(sh->dev[i].page);
1569                         if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1570                                 printk("block %d/%d not uptodate on parity calc\n", i,count);
1571                         i = raid6_next_disk(i, disks);
1572                 } while ( i != d0_idx );
1573 //              break;
1574 //      }
1575
1576         raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1577
1578         switch(method) {
1579         case RECONSTRUCT_WRITE:
1580                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1581                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1582                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1583                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
1584                 break;
1585         case UPDATE_PARITY:
1586                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1587                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1588                 break;
1589         }
1590 }
1591
1592
1593 /* Compute one missing block */
1594 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1595 {
1596         int i, count, disks = sh->disks;
1597         void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1598         int pd_idx = sh->pd_idx;
1599         int qd_idx = raid6_next_disk(pd_idx, disks);
1600
1601         pr_debug("compute_block_1, stripe %llu, idx %d\n",
1602                 (unsigned long long)sh->sector, dd_idx);
1603
1604         if ( dd_idx == qd_idx ) {
1605                 /* We're actually computing the Q drive */
1606                 compute_parity6(sh, UPDATE_PARITY);
1607         } else {
1608                 dest = page_address(sh->dev[dd_idx].page);
1609                 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1610                 count = 0;
1611                 for (i = disks ; i--; ) {
1612                         if (i == dd_idx || i == qd_idx)
1613                                 continue;
1614                         p = page_address(sh->dev[i].page);
1615                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1616                                 ptr[count++] = p;
1617                         else
1618                                 printk("compute_block() %d, stripe %llu, %d"
1619                                        " not present\n", dd_idx,
1620                                        (unsigned long long)sh->sector, i);
1621
1622                         check_xor();
1623                 }
1624                 if (count)
1625                         xor_blocks(count, STRIPE_SIZE, dest, ptr);
1626                 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1627                 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1628         }
1629 }
1630
1631 /* Compute two missing blocks */
1632 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1633 {
1634         int i, count, disks = sh->disks;
1635         int pd_idx = sh->pd_idx;
1636         int qd_idx = raid6_next_disk(pd_idx, disks);
1637         int d0_idx = raid6_next_disk(qd_idx, disks);
1638         int faila, failb;
1639
1640         /* faila and failb are disk numbers relative to d0_idx */
1641         /* pd_idx become disks-2 and qd_idx become disks-1 */
1642         faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1643         failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1644
1645         BUG_ON(faila == failb);
1646         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1647
1648         pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1649                (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1650
1651         if ( failb == disks-1 ) {
1652                 /* Q disk is one of the missing disks */
1653                 if ( faila == disks-2 ) {
1654                         /* Missing P+Q, just recompute */
1655                         compute_parity6(sh, UPDATE_PARITY);
1656                         return;
1657                 } else {
1658                         /* We're missing D+Q; recompute D from P */
1659                         compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1660                         compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1661                         return;
1662                 }
1663         }
1664
1665         /* We're missing D+P or D+D; build pointer table */
1666         {
1667                 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1668                 void *ptrs[disks];
1669
1670                 count = 0;
1671                 i = d0_idx;
1672                 do {
1673                         ptrs[count++] = page_address(sh->dev[i].page);
1674                         i = raid6_next_disk(i, disks);
1675                         if (i != dd_idx1 && i != dd_idx2 &&
1676                             !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1677                                 printk("compute_2 with missing block %d/%d\n", count, i);
1678                 } while ( i != d0_idx );
1679
1680                 if ( failb == disks-2 ) {
1681                         /* We're missing D+P. */
1682                         raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1683                 } else {
1684                         /* We're missing D+D. */
1685                         raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1686                 }
1687
1688                 /* Both the above update both missing blocks */
1689                 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1690                 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1691         }
1692 }
1693
1694 static int
1695 handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
1696 {
1697         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1698         int locked = 0;
1699
1700         if (rcw) {
1701                 /* if we are not expanding this is a proper write request, and
1702                  * there will be bios with new data to be drained into the
1703                  * stripe cache
1704                  */
1705                 if (!expand) {
1706                         set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1707                         sh->ops.count++;
1708                 }
1709
1710                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1711                 sh->ops.count++;
1712
1713                 for (i = disks; i--; ) {
1714                         struct r5dev *dev = &sh->dev[i];
1715
1716                         if (dev->towrite) {
1717                                 set_bit(R5_LOCKED, &dev->flags);
1718                                 if (!expand)
1719                                         clear_bit(R5_UPTODATE, &dev->flags);
1720                                 locked++;
1721                         }
1722                 }
1723         } else {
1724                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1725                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1726
1727                 set_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
1728                 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1729                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1730
1731                 sh->ops.count += 3;
1732
1733                 for (i = disks; i--; ) {
1734                         struct r5dev *dev = &sh->dev[i];
1735                         if (i == pd_idx)
1736                                 continue;
1737
1738                         /* For a read-modify write there may be blocks that are
1739                          * locked for reading while others are ready to be
1740                          * written so we distinguish these blocks by the
1741                          * R5_Wantprexor bit
1742                          */
1743                         if (dev->towrite &&
1744                             (test_bit(R5_UPTODATE, &dev->flags) ||
1745                             test_bit(R5_Wantcompute, &dev->flags))) {
1746                                 set_bit(R5_Wantprexor, &dev->flags);
1747                                 set_bit(R5_LOCKED, &dev->flags);
1748                                 clear_bit(R5_UPTODATE, &dev->flags);
1749                                 locked++;
1750                         }
1751                 }
1752         }
1753
1754         /* keep the parity disk locked while asynchronous operations
1755          * are in flight
1756          */
1757         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1758         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1759         locked++;
1760
1761         pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1762                 __FUNCTION__, (unsigned long long)sh->sector,
1763                 locked, sh->ops.pending);
1764
1765         return locked;
1766 }
1767
1768 /*
1769  * Each stripe/dev can have one or more bion attached.
1770  * toread/towrite point to the first in a chain.
1771  * The bi_next chain must be in order.
1772  */
1773 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1774 {
1775         struct bio **bip;
1776         raid5_conf_t *conf = sh->raid_conf;
1777         int firstwrite=0;
1778
1779         pr_debug("adding bh b#%llu to stripe s#%llu\n",
1780                 (unsigned long long)bi->bi_sector,
1781                 (unsigned long long)sh->sector);
1782
1783
1784         spin_lock(&sh->lock);
1785         spin_lock_irq(&conf->device_lock);
1786         if (forwrite) {
1787                 bip = &sh->dev[dd_idx].towrite;
1788                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1789                         firstwrite = 1;
1790         } else
1791                 bip = &sh->dev[dd_idx].toread;
1792         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1793                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1794                         goto overlap;
1795                 bip = & (*bip)->bi_next;
1796         }
1797         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1798                 goto overlap;
1799
1800         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1801         if (*bip)
1802                 bi->bi_next = *bip;
1803         *bip = bi;
1804         bi->bi_phys_segments ++;
1805         spin_unlock_irq(&conf->device_lock);
1806         spin_unlock(&sh->lock);
1807
1808         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1809                 (unsigned long long)bi->bi_sector,
1810                 (unsigned long long)sh->sector, dd_idx);
1811
1812         if (conf->mddev->bitmap && firstwrite) {
1813                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1814                                   STRIPE_SECTORS, 0);
1815                 sh->bm_seq = conf->seq_flush+1;
1816                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1817         }
1818
1819         if (forwrite) {
1820                 /* check if page is covered */
1821                 sector_t sector = sh->dev[dd_idx].sector;
1822                 for (bi=sh->dev[dd_idx].towrite;
1823                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1824                              bi && bi->bi_sector <= sector;
1825                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1826                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1827                                 sector = bi->bi_sector + (bi->bi_size>>9);
1828                 }
1829                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1830                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1831         }
1832         return 1;
1833
1834  overlap:
1835         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1836         spin_unlock_irq(&conf->device_lock);
1837         spin_unlock(&sh->lock);
1838         return 0;
1839 }
1840
1841 static void end_reshape(raid5_conf_t *conf);
1842
1843 static int page_is_zero(struct page *p)
1844 {
1845         char *a = page_address(p);
1846         return ((*(u32*)a) == 0 &&
1847                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1848 }
1849
1850 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1851 {
1852         int sectors_per_chunk = conf->chunk_size >> 9;
1853         int pd_idx, dd_idx;
1854         int chunk_offset = sector_div(stripe, sectors_per_chunk);
1855
1856         raid5_compute_sector(stripe * (disks - conf->max_degraded)
1857                              *sectors_per_chunk + chunk_offset,
1858                              disks, disks - conf->max_degraded,
1859                              &dd_idx, &pd_idx, conf);
1860         return pd_idx;
1861 }
1862
1863 static void
1864 handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1865                                 struct stripe_head_state *s, int disks,
1866                                 struct bio **return_bi)
1867 {
1868         int i;
1869         for (i = disks; i--; ) {
1870                 struct bio *bi;
1871                 int bitmap_end = 0;
1872
1873                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1874                         mdk_rdev_t *rdev;
1875                         rcu_read_lock();
1876                         rdev = rcu_dereference(conf->disks[i].rdev);
1877                         if (rdev && test_bit(In_sync, &rdev->flags))
1878                                 /* multiple read failures in one stripe */
1879                                 md_error(conf->mddev, rdev);
1880                         rcu_read_unlock();
1881                 }
1882                 spin_lock_irq(&conf->device_lock);
1883                 /* fail all writes first */
1884                 bi = sh->dev[i].towrite;
1885                 sh->dev[i].towrite = NULL;
1886                 if (bi) {
1887                         s->to_write--;
1888                         bitmap_end = 1;
1889                 }
1890
1891                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1892                         wake_up(&conf->wait_for_overlap);
1893
1894                 while (bi && bi->bi_sector <
1895                         sh->dev[i].sector + STRIPE_SECTORS) {
1896                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1897                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1898                         if (--bi->bi_phys_segments == 0) {
1899                                 md_write_end(conf->mddev);
1900                                 bi->bi_next = *return_bi;
1901                                 *return_bi = bi;
1902                         }
1903                         bi = nextbi;
1904                 }
1905                 /* and fail all 'written' */
1906                 bi = sh->dev[i].written;
1907                 sh->dev[i].written = NULL;
1908                 if (bi) bitmap_end = 1;
1909                 while (bi && bi->bi_sector <
1910                        sh->dev[i].sector + STRIPE_SECTORS) {
1911                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1912                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1913                         if (--bi->bi_phys_segments == 0) {
1914                                 md_write_end(conf->mddev);
1915                                 bi->bi_next = *return_bi;
1916                                 *return_bi = bi;
1917                         }
1918                         bi = bi2;
1919                 }
1920
1921                 /* fail any reads if this device is non-operational and
1922                  * the data has not reached the cache yet.
1923                  */
1924                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
1925                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1926                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
1927                         bi = sh->dev[i].toread;
1928                         sh->dev[i].toread = NULL;
1929                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1930                                 wake_up(&conf->wait_for_overlap);
1931                         if (bi) s->to_read--;
1932                         while (bi && bi->bi_sector <
1933                                sh->dev[i].sector + STRIPE_SECTORS) {
1934                                 struct bio *nextbi =
1935                                         r5_next_bio(bi, sh->dev[i].sector);
1936                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1937                                 if (--bi->bi_phys_segments == 0) {
1938                                         bi->bi_next = *return_bi;
1939                                         *return_bi = bi;
1940                                 }
1941                                 bi = nextbi;
1942                         }
1943                 }
1944                 spin_unlock_irq(&conf->device_lock);
1945                 if (bitmap_end)
1946                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1947                                         STRIPE_SECTORS, 0, 0);
1948         }
1949
1950 }
1951
1952 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
1953  * to process
1954  */
1955 static int __handle_issuing_new_read_requests5(struct stripe_head *sh,
1956                         struct stripe_head_state *s, int disk_idx, int disks)
1957 {
1958         struct r5dev *dev = &sh->dev[disk_idx];
1959         struct r5dev *failed_dev = &sh->dev[s->failed_num];
1960
1961         /* don't schedule compute operations or reads on the parity block while
1962          * a check is in flight
1963          */
1964         if ((disk_idx == sh->pd_idx) &&
1965              test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
1966                 return ~0;
1967
1968         /* is the data in this block needed, and can we get it? */
1969         if (!test_bit(R5_LOCKED, &dev->flags) &&
1970             !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread ||
1971             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1972              s->syncing || s->expanding || (s->failed &&
1973              (failed_dev->toread || (failed_dev->towrite &&
1974              !test_bit(R5_OVERWRITE, &failed_dev->flags)
1975              ))))) {
1976                 /* 1/ We would like to get this block, possibly by computing it,
1977                  * but we might not be able to.
1978                  *
1979                  * 2/ Since parity check operations potentially make the parity
1980                  * block !uptodate it will need to be refreshed before any
1981                  * compute operations on data disks are scheduled.
1982                  *
1983                  * 3/ We hold off parity block re-reads until check operations
1984                  * have quiesced.
1985                  */
1986                 if ((s->uptodate == disks - 1) &&
1987                     !test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
1988                         set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
1989                         set_bit(R5_Wantcompute, &dev->flags);
1990                         sh->ops.target = disk_idx;
1991                         s->req_compute = 1;
1992                         sh->ops.count++;
1993                         /* Careful: from this point on 'uptodate' is in the eye
1994                          * of raid5_run_ops which services 'compute' operations
1995                          * before writes. R5_Wantcompute flags a block that will
1996                          * be R5_UPTODATE by the time it is needed for a
1997                          * subsequent operation.
1998                          */
1999                         s->uptodate++;
2000                         return 0; /* uptodate + compute == disks */
2001                 } else if ((s->uptodate < disks - 1) &&
2002                         test_bit(R5_Insync, &dev->flags)) {
2003                         /* Note: we hold off compute operations while checks are
2004                          * in flight, but we still prefer 'compute' over 'read'
2005                          * hence we only read if (uptodate < * disks-1)
2006                          */
2007                         set_bit(R5_LOCKED, &dev->flags);
2008                         set_bit(R5_Wantread, &dev->flags);
2009                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2010                                 sh->ops.count++;
2011                         s->locked++;
2012                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2013                                 s->syncing);
2014                 }
2015         }
2016
2017         return ~0;
2018 }
2019
2020 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2021                         struct stripe_head_state *s, int disks)
2022 {
2023         int i;
2024
2025         /* Clear completed compute operations.  Parity recovery
2026          * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled
2027          * later on in this routine
2028          */
2029         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2030                 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2031                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2032                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2033                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2034         }
2035
2036         /* look for blocks to read/compute, skip this if a compute
2037          * is already in flight, or if the stripe contents are in the
2038          * midst of changing due to a write
2039          */
2040         if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2041                 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2042                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2043                 for (i = disks; i--; )
2044                         if (__handle_issuing_new_read_requests5(
2045                                 sh, s, i, disks) == 0)
2046                                 break;
2047         }
2048         set_bit(STRIPE_HANDLE, &sh->state);
2049 }
2050
2051 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2052                         struct stripe_head_state *s, struct r6_state *r6s,
2053                         int disks)
2054 {
2055         int i;
2056         for (i = disks; i--; ) {
2057                 struct r5dev *dev = &sh->dev[i];
2058                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2059                     !test_bit(R5_UPTODATE, &dev->flags) &&
2060                     (dev->toread || (dev->towrite &&
2061                      !test_bit(R5_OVERWRITE, &dev->flags)) ||
2062                      s->syncing || s->expanding ||
2063                      (s->failed >= 1 &&
2064                       (sh->dev[r6s->failed_num[0]].toread ||
2065                        s->to_write)) ||
2066                      (s->failed >= 2 &&
2067                       (sh->dev[r6s->failed_num[1]].toread ||
2068                        s->to_write)))) {
2069                         /* we would like to get this block, possibly
2070                          * by computing it, but we might not be able to
2071                          */
2072                         if (s->uptodate == disks-1) {
2073                                 pr_debug("Computing stripe %llu block %d\n",
2074                                        (unsigned long long)sh->sector, i);
2075                                 compute_block_1(sh, i, 0);
2076                                 s->uptodate++;
2077                         } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2078                                 /* Computing 2-failure is *very* expensive; only
2079                                  * do it if failed >= 2
2080                                  */
2081                                 int other;
2082                                 for (other = disks; other--; ) {
2083                                         if (other == i)
2084                                                 continue;
2085                                         if (!test_bit(R5_UPTODATE,
2086                                               &sh->dev[other].flags))
2087                                                 break;
2088                                 }
2089                                 BUG_ON(other < 0);
2090                                 pr_debug("Computing stripe %llu blocks %d,%d\n",
2091                                        (unsigned long long)sh->sector,
2092                                        i, other);
2093                                 compute_block_2(sh, i, other);
2094                                 s->uptodate += 2;
2095                         } else if (test_bit(R5_Insync, &dev->flags)) {
2096                                 set_bit(R5_LOCKED, &dev->flags);
2097                                 set_bit(R5_Wantread, &dev->flags);
2098                                 s->locked++;
2099                                 pr_debug("Reading block %d (sync=%d)\n",
2100                                         i, s->syncing);
2101                         }
2102                 }
2103         }
2104         set_bit(STRIPE_HANDLE, &sh->state);
2105 }
2106
2107
2108 /* handle_completed_write_requests
2109  * any written block on an uptodate or failed drive can be returned.
2110  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2111  * never LOCKED, so we don't need to test 'failed' directly.
2112  */
2113 static void handle_completed_write_requests(raid5_conf_t *conf,
2114         struct stripe_head *sh, int disks, struct bio **return_bi)
2115 {
2116         int i;
2117         struct r5dev *dev;
2118
2119         for (i = disks; i--; )
2120                 if (sh->dev[i].written) {
2121                         dev = &sh->dev[i];
2122                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2123                                 test_bit(R5_UPTODATE, &dev->flags)) {
2124                                 /* We can return any write requests */
2125                                 struct bio *wbi, *wbi2;
2126                                 int bitmap_end = 0;
2127                                 pr_debug("Return write for disc %d\n", i);
2128                                 spin_lock_irq(&conf->device_lock);
2129                                 wbi = dev->written;
2130                                 dev->written = NULL;
2131                                 while (wbi && wbi->bi_sector <
2132                                         dev->sector + STRIPE_SECTORS) {
2133                                         wbi2 = r5_next_bio(wbi, dev->sector);
2134                                         if (--wbi->bi_phys_segments == 0) {
2135                                                 md_write_end(conf->mddev);
2136                                                 wbi->bi_next = *return_bi;
2137                                                 *return_bi = wbi;
2138                                         }
2139                                         wbi = wbi2;
2140                                 }
2141                                 if (dev->towrite == NULL)
2142                                         bitmap_end = 1;
2143                                 spin_unlock_irq(&conf->device_lock);
2144                                 if (bitmap_end)
2145                                         bitmap_endwrite(conf->mddev->bitmap,
2146                                                         sh->sector,
2147                                                         STRIPE_SECTORS,
2148                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2149                                                         0);
2150                         }
2151                 }
2152 }
2153
2154 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2155                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2156 {
2157         int rmw = 0, rcw = 0, i;
2158         for (i = disks; i--; ) {
2159                 /* would I have to read this buffer for read_modify_write */
2160                 struct r5dev *dev = &sh->dev[i];
2161                 if ((dev->towrite || i == sh->pd_idx) &&
2162                     !test_bit(R5_LOCKED, &dev->flags) &&
2163                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2164                       test_bit(R5_Wantcompute, &dev->flags))) {
2165                         if (test_bit(R5_Insync, &dev->flags))
2166                                 rmw++;
2167                         else
2168                                 rmw += 2*disks;  /* cannot read it */
2169                 }
2170                 /* Would I have to read this buffer for reconstruct_write */
2171                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2172                     !test_bit(R5_LOCKED, &dev->flags) &&
2173                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2174                     test_bit(R5_Wantcompute, &dev->flags))) {
2175                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2176                         else
2177                                 rcw += 2*disks;
2178                 }
2179         }
2180         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2181                 (unsigned long long)sh->sector, rmw, rcw);
2182         set_bit(STRIPE_HANDLE, &sh->state);
2183         if (rmw < rcw && rmw > 0)
2184                 /* prefer read-modify-write, but need to get some data */
2185                 for (i = disks; i--; ) {
2186                         struct r5dev *dev = &sh->dev[i];
2187                         if ((dev->towrite || i == sh->pd_idx) &&
2188                             !test_bit(R5_LOCKED, &dev->flags) &&
2189                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2190                             test_bit(R5_Wantcompute, &dev->flags)) &&
2191                             test_bit(R5_Insync, &dev->flags)) {
2192                                 if (
2193                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2194                                         pr_debug("Read_old block "
2195                                                 "%d for r-m-w\n", i);
2196                                         set_bit(R5_LOCKED, &dev->flags);
2197                                         set_bit(R5_Wantread, &dev->flags);
2198                                         if (!test_and_set_bit(
2199                                                 STRIPE_OP_IO, &sh->ops.pending))
2200                                                 sh->ops.count++;
2201                                         s->locked++;
2202                                 } else {
2203                                         set_bit(STRIPE_DELAYED, &sh->state);
2204                                         set_bit(STRIPE_HANDLE, &sh->state);
2205                                 }
2206                         }
2207                 }
2208         if (rcw <= rmw && rcw > 0)
2209                 /* want reconstruct write, but need to get some data */
2210                 for (i = disks; i--; ) {
2211                         struct r5dev *dev = &sh->dev[i];
2212                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2213                             i != sh->pd_idx &&
2214                             !test_bit(R5_LOCKED, &dev->flags) &&
2215                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2216                             test_bit(R5_Wantcompute, &dev->flags)) &&
2217                             test_bit(R5_Insync, &dev->flags)) {
2218                                 if (
2219                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2220                                         pr_debug("Read_old block "
2221                                                 "%d for Reconstruct\n", i);
2222                                         set_bit(R5_LOCKED, &dev->flags);
2223                                         set_bit(R5_Wantread, &dev->flags);
2224                                         if (!test_and_set_bit(
2225                                                 STRIPE_OP_IO, &sh->ops.pending))
2226                                                 sh->ops.count++;
2227                                         s->locked++;
2228                                 } else {
2229                                         set_bit(STRIPE_DELAYED, &sh->state);
2230                                         set_bit(STRIPE_HANDLE, &sh->state);
2231                                 }
2232                         }
2233                 }
2234         /* now if nothing is locked, and if we have enough data,
2235          * we can start a write request
2236          */
2237         /* since handle_stripe can be called at any time we need to handle the
2238          * case where a compute block operation has been submitted and then a
2239          * subsequent call wants to start a write request.  raid5_run_ops only
2240          * handles the case where compute block and postxor are requested
2241          * simultaneously.  If this is not the case then new writes need to be
2242          * held off until the compute completes.
2243          */
2244         if ((s->req_compute ||
2245             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) &&
2246                 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2247                 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2248                 s->locked += handle_write_operations5(sh, rcw == 0, 0);
2249 }
2250
2251 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2252                 struct stripe_head *sh, struct stripe_head_state *s,
2253                 struct r6_state *r6s, int disks)
2254 {
2255         int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2256         int qd_idx = r6s->qd_idx;
2257         for (i = disks; i--; ) {
2258                 struct r5dev *dev = &sh->dev[i];
2259                 /* Would I have to read this buffer for reconstruct_write */
2260                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2261                     && i != pd_idx && i != qd_idx
2262                     && (!test_bit(R5_LOCKED, &dev->flags)
2263                             ) &&
2264                     !test_bit(R5_UPTODATE, &dev->flags)) {
2265                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2266                         else {
2267                                 pr_debug("raid6: must_compute: "
2268                                         "disk %d flags=%#lx\n", i, dev->flags);
2269                                 must_compute++;
2270                         }
2271                 }
2272         }
2273         pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2274                (unsigned long long)sh->sector, rcw, must_compute);
2275         set_bit(STRIPE_HANDLE, &sh->state);
2276
2277         if (rcw > 0)
2278                 /* want reconstruct write, but need to get some data */
2279                 for (i = disks; i--; ) {
2280                         struct r5dev *dev = &sh->dev[i];
2281                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2282                             && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2283                             && !test_bit(R5_LOCKED, &dev->flags) &&
2284                             !test_bit(R5_UPTODATE, &dev->flags) &&
2285                             test_bit(R5_Insync, &dev->flags)) {
2286                                 if (
2287                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2288                                         pr_debug("Read_old stripe %llu "
2289                                                 "block %d for Reconstruct\n",
2290                                              (unsigned long long)sh->sector, i);
2291                                         set_bit(R5_LOCKED, &dev->flags);
2292                                         set_bit(R5_Wantread, &dev->flags);
2293                                         s->locked++;
2294                                 } else {
2295                                         pr_debug("Request delayed stripe %llu "
2296                                                 "block %d for Reconstruct\n",
2297                                              (unsigned long long)sh->sector, i);
2298                                         set_bit(STRIPE_DELAYED, &sh->state);
2299                                         set_bit(STRIPE_HANDLE, &sh->state);
2300                                 }
2301                         }
2302                 }
2303         /* now if nothing is locked, and if we have enough data, we can start a
2304          * write request
2305          */
2306         if (s->locked == 0 && rcw == 0 &&
2307             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2308                 if (must_compute > 0) {
2309                         /* We have failed blocks and need to compute them */
2310                         switch (s->failed) {
2311                         case 0:
2312                                 BUG();
2313                         case 1:
2314                                 compute_block_1(sh, r6s->failed_num[0], 0);
2315                                 break;
2316                         case 2:
2317                                 compute_block_2(sh, r6s->failed_num[0],
2318                                                 r6s->failed_num[1]);
2319                                 break;
2320                         default: /* This request should have been failed? */
2321                                 BUG();
2322                         }
2323                 }
2324
2325                 pr_debug("Computing parity for stripe %llu\n",
2326                         (unsigned long long)sh->sector);
2327                 compute_parity6(sh, RECONSTRUCT_WRITE);
2328                 /* now every locked buffer is ready to be written */
2329                 for (i = disks; i--; )
2330                         if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2331                                 pr_debug("Writing stripe %llu block %d\n",
2332                                        (unsigned long long)sh->sector, i);
2333                                 s->locked++;
2334                                 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2335                         }
2336                 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2337                 set_bit(STRIPE_INSYNC, &sh->state);
2338
2339                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2340                         atomic_dec(&conf->preread_active_stripes);
2341                         if (atomic_read(&conf->preread_active_stripes) <
2342                             IO_THRESHOLD)
2343                                 md_wakeup_thread(conf->mddev->thread);
2344                 }
2345         }
2346 }
2347
2348 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2349                                 struct stripe_head_state *s, int disks)
2350 {
2351         int canceled_check = 0;
2352
2353         set_bit(STRIPE_HANDLE, &sh->state);
2354
2355         /* complete a check operation */
2356         if (test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) {
2357             clear_bit(STRIPE_OP_CHECK, &sh->ops.ack);
2358             clear_bit(STRIPE_OP_CHECK, &sh->ops.pending);
2359                 if (s->failed == 0) {
2360                         if (sh->ops.zero_sum_result == 0)
2361                                 /* parity is correct (on disc,
2362                                  * not in buffer any more)
2363                                  */
2364                                 set_bit(STRIPE_INSYNC, &sh->state);
2365                         else {
2366                                 conf->mddev->resync_mismatches +=
2367                                         STRIPE_SECTORS;
2368                                 if (test_bit(
2369                                      MD_RECOVERY_CHECK, &conf->mddev->recovery))
2370                                         /* don't try to repair!! */
2371                                         set_bit(STRIPE_INSYNC, &sh->state);
2372                                 else {
2373                                         set_bit(STRIPE_OP_COMPUTE_BLK,
2374                                                 &sh->ops.pending);
2375                                         set_bit(STRIPE_OP_MOD_REPAIR_PD,
2376                                                 &sh->ops.pending);
2377                                         set_bit(R5_Wantcompute,
2378                                                 &sh->dev[sh->pd_idx].flags);
2379                                         sh->ops.target = sh->pd_idx;
2380                                         sh->ops.count++;
2381                                         s->uptodate++;
2382                                 }
2383                         }
2384                 } else
2385                         canceled_check = 1; /* STRIPE_INSYNC is not set */
2386         }
2387
2388         /* check if we can clear a parity disk reconstruct */
2389         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2390                 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2391
2392                 clear_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending);
2393                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2394                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2395                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2396         }
2397
2398         /* start a new check operation if there are no failures, the stripe is
2399          * not insync, and a repair is not in flight
2400          */
2401         if (s->failed == 0 &&
2402             !test_bit(STRIPE_INSYNC, &sh->state) &&
2403             !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2404                 if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2405                         BUG_ON(s->uptodate != disks);
2406                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2407                         sh->ops.count++;
2408                         s->uptodate--;
2409                 }
2410         }
2411
2412         /* Wait for check parity and compute block operations to complete
2413          * before write-back.  If a failure occurred while the check operation
2414          * was in flight we need to cycle this stripe through handle_stripe
2415          * since the parity block may not be uptodate
2416          */
2417         if (!canceled_check && !test_bit(STRIPE_INSYNC, &sh->state) &&
2418             !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) &&
2419             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) {
2420                 struct r5dev *dev;
2421                 /* either failed parity check, or recovery is happening */
2422                 if (s->failed == 0)
2423                         s->failed_num = sh->pd_idx;
2424                 dev = &sh->dev[s->failed_num];
2425                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2426                 BUG_ON(s->uptodate != disks);
2427
2428                 set_bit(R5_LOCKED, &dev->flags);
2429                 set_bit(R5_Wantwrite, &dev->flags);
2430                 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2431                         sh->ops.count++;
2432
2433                 clear_bit(STRIPE_DEGRADED, &sh->state);
2434                 s->locked++;
2435                 set_bit(STRIPE_INSYNC, &sh->state);
2436         }
2437 }
2438
2439
2440 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2441                                 struct stripe_head_state *s,
2442                                 struct r6_state *r6s, struct page *tmp_page,
2443                                 int disks)
2444 {
2445         int update_p = 0, update_q = 0;
2446         struct r5dev *dev;
2447         int pd_idx = sh->pd_idx;
2448         int qd_idx = r6s->qd_idx;
2449
2450         set_bit(STRIPE_HANDLE, &sh->state);
2451
2452         BUG_ON(s->failed > 2);
2453         BUG_ON(s->uptodate < disks);
2454         /* Want to check and possibly repair P and Q.
2455          * However there could be one 'failed' device, in which
2456          * case we can only check one of them, possibly using the
2457          * other to generate missing data
2458          */
2459
2460         /* If !tmp_page, we cannot do the calculations,
2461          * but as we have set STRIPE_HANDLE, we will soon be called
2462          * by stripe_handle with a tmp_page - just wait until then.
2463          */
2464         if (tmp_page) {
2465                 if (s->failed == r6s->q_failed) {
2466                         /* The only possible failed device holds 'Q', so it
2467                          * makes sense to check P (If anything else were failed,
2468                          * we would have used P to recreate it).
2469                          */
2470                         compute_block_1(sh, pd_idx, 1);
2471                         if (!page_is_zero(sh->dev[pd_idx].page)) {
2472                                 compute_block_1(sh, pd_idx, 0);
2473                                 update_p = 1;
2474                         }
2475                 }
2476                 if (!r6s->q_failed && s->failed < 2) {
2477                         /* q is not failed, and we didn't use it to generate
2478                          * anything, so it makes sense to check it
2479                          */
2480                         memcpy(page_address(tmp_page),
2481                                page_address(sh->dev[qd_idx].page),
2482                                STRIPE_SIZE);
2483                         compute_parity6(sh, UPDATE_PARITY);
2484                         if (memcmp(page_address(tmp_page),
2485                                    page_address(sh->dev[qd_idx].page),
2486                                    STRIPE_SIZE) != 0) {
2487                                 clear_bit(STRIPE_INSYNC, &sh->state);
2488                                 update_q = 1;
2489                         }
2490                 }
2491                 if (update_p || update_q) {
2492                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2493                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2494                                 /* don't try to repair!! */
2495                                 update_p = update_q = 0;
2496                 }
2497
2498                 /* now write out any block on a failed drive,
2499                  * or P or Q if they need it
2500                  */
2501
2502                 if (s->failed == 2) {
2503                         dev = &sh->dev[r6s->failed_num[1]];
2504                         s->locked++;
2505                         set_bit(R5_LOCKED, &dev->flags);
2506                         set_bit(R5_Wantwrite, &dev->flags);
2507                 }
2508                 if (s->failed >= 1) {
2509                         dev = &sh->dev[r6s->failed_num[0]];
2510                         s->locked++;
2511                         set_bit(R5_LOCKED, &dev->flags);
2512                         set_bit(R5_Wantwrite, &dev->flags);
2513                 }
2514
2515                 if (update_p) {
2516                         dev = &sh->dev[pd_idx];
2517                         s->locked++;
2518                         set_bit(R5_LOCKED, &dev->flags);
2519                         set_bit(R5_Wantwrite, &dev->flags);
2520                 }
2521                 if (update_q) {
2522                         dev = &sh->dev[qd_idx];
2523                         s->locked++;
2524                         set_bit(R5_LOCKED, &dev->flags);
2525                         set_bit(R5_Wantwrite, &dev->flags);
2526                 }
2527                 clear_bit(STRIPE_DEGRADED, &sh->state);
2528
2529                 set_bit(STRIPE_INSYNC, &sh->state);
2530         }
2531 }
2532
2533 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2534                                 struct r6_state *r6s)
2535 {
2536         int i;
2537
2538         /* We have read all the blocks in this stripe and now we need to
2539          * copy some of them into a target stripe for expand.
2540          */
2541         struct dma_async_tx_descriptor *tx = NULL;
2542         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2543         for (i = 0; i < sh->disks; i++)
2544                 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2545                         int dd_idx, pd_idx, j;
2546                         struct stripe_head *sh2;
2547
2548                         sector_t bn = compute_blocknr(sh, i);
2549                         sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2550                                                 conf->raid_disks -
2551                                                 conf->max_degraded, &dd_idx,
2552                                                 &pd_idx, conf);
2553                         sh2 = get_active_stripe(conf, s, conf->raid_disks,
2554                                                 pd_idx, 1);
2555                         if (sh2 == NULL)
2556                                 /* so far only the early blocks of this stripe
2557                                  * have been requested.  When later blocks
2558                                  * get requested, we will try again
2559                                  */
2560                                 continue;
2561                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2562                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2563                                 /* must have already done this block */
2564                                 release_stripe(sh2);
2565                                 continue;
2566                         }
2567
2568                         /* place all the copies on one channel */
2569                         tx = async_memcpy(sh2->dev[dd_idx].page,
2570                                 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2571                                 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2572
2573                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2574                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2575                         for (j = 0; j < conf->raid_disks; j++)
2576                                 if (j != sh2->pd_idx &&
2577                                     (!r6s || j != raid6_next_disk(sh2->pd_idx,
2578                                                                  sh2->disks)) &&
2579                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2580                                         break;
2581                         if (j == conf->raid_disks) {
2582                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2583                                 set_bit(STRIPE_HANDLE, &sh2->state);
2584                         }
2585                         release_stripe(sh2);
2586
2587                 }
2588         /* done submitting copies, wait for them to complete */
2589         if (tx) {
2590                 async_tx_ack(tx);
2591                 dma_wait_for_async_tx(tx);
2592         }
2593 }
2594
2595 /*
2596  * handle_stripe - do things to a stripe.
2597  *
2598  * We lock the stripe and then examine the state of various bits
2599  * to see what needs to be done.
2600  * Possible results:
2601  *    return some read request which now have data
2602  *    return some write requests which are safely on disc
2603  *    schedule a read on some buffers
2604  *    schedule a write of some buffers
2605  *    return confirmation of parity correctness
2606  *
2607  * buffers are taken off read_list or write_list, and bh_cache buffers
2608  * get BH_Lock set before the stripe lock is released.
2609  *
2610  */
2611
2612 static void handle_stripe5(struct stripe_head *sh)
2613 {
2614         raid5_conf_t *conf = sh->raid_conf;
2615         int disks = sh->disks, i;
2616         struct bio *return_bi = NULL;
2617         struct stripe_head_state s;
2618         struct r5dev *dev;
2619         unsigned long pending = 0;
2620
2621         memset(&s, 0, sizeof(s));
2622         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2623                 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2624                 atomic_read(&sh->count), sh->pd_idx,
2625                 sh->ops.pending, sh->ops.ack, sh->ops.complete);
2626
2627         spin_lock(&sh->lock);
2628         clear_bit(STRIPE_HANDLE, &sh->state);
2629         clear_bit(STRIPE_DELAYED, &sh->state);
2630
2631         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2632         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2633         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2634         /* Now to look around and see what can be done */
2635
2636         /* clean-up completed biofill operations */
2637         if (test_bit(STRIPE_OP_BIOFILL, &sh->ops.complete)) {
2638                 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending);
2639                 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack);
2640                 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
2641         }
2642
2643         rcu_read_lock();
2644         for (i=disks; i--; ) {
2645                 mdk_rdev_t *rdev;
2646                 struct r5dev *dev = &sh->dev[i];
2647                 clear_bit(R5_Insync, &dev->flags);
2648
2649                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2650                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2651                         dev->towrite, dev->written);
2652
2653                 /* maybe we can request a biofill operation
2654                  *
2655                  * new wantfill requests are only permitted while
2656                  * STRIPE_OP_BIOFILL is clear
2657                  */
2658                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2659                         !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2660                         set_bit(R5_Wantfill, &dev->flags);
2661
2662                 /* now count some things */
2663                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2664                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2665                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2666
2667                 if (test_bit(R5_Wantfill, &dev->flags))
2668                         s.to_fill++;
2669                 else if (dev->toread)
2670                         s.to_read++;
2671                 if (dev->towrite) {
2672                         s.to_write++;
2673                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2674                                 s.non_overwrite++;
2675                 }
2676                 if (dev->written)
2677                         s.written++;
2678                 rdev = rcu_dereference(conf->disks[i].rdev);
2679                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2680                         /* The ReadError flag will just be confusing now */
2681                         clear_bit(R5_ReadError, &dev->flags);
2682                         clear_bit(R5_ReWrite, &dev->flags);
2683                 }
2684                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2685                     || test_bit(R5_ReadError, &dev->flags)) {
2686                         s.failed++;
2687                         s.failed_num = i;
2688                 } else
2689                         set_bit(R5_Insync, &dev->flags);
2690         }
2691         rcu_read_unlock();
2692
2693         if (s.to_fill && !test_and_set_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2694                 sh->ops.count++;
2695
2696         pr_debug("locked=%d uptodate=%d to_read=%d"
2697                 " to_write=%d failed=%d failed_num=%d\n",
2698                 s.locked, s.uptodate, s.to_read, s.to_write,
2699                 s.failed, s.failed_num);
2700         /* check if the array has lost two devices and, if so, some requests might
2701          * need to be failed
2702          */
2703         if (s.failed > 1 && s.to_read+s.to_write+s.written)
2704                 handle_requests_to_failed_array(conf, sh, &s, disks,
2705                                                 &return_bi);
2706         if (s.failed > 1 && s.syncing) {
2707                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2708                 clear_bit(STRIPE_SYNCING, &sh->state);
2709                 s.syncing = 0;
2710         }
2711
2712         /* might be able to return some write requests if the parity block
2713          * is safe, or on a failed drive
2714          */
2715         dev = &sh->dev[sh->pd_idx];
2716         if ( s.written &&
2717              ((test_bit(R5_Insync, &dev->flags) &&
2718                !test_bit(R5_LOCKED, &dev->flags) &&
2719                test_bit(R5_UPTODATE, &dev->flags)) ||
2720                (s.failed == 1 && s.failed_num == sh->pd_idx)))
2721                 handle_completed_write_requests(conf, sh, disks, &return_bi);
2722
2723         /* Now we might consider reading some blocks, either to check/generate
2724          * parity, or to satisfy requests
2725          * or to load a block that is being partially written.
2726          */
2727         if (s.to_read || s.non_overwrite ||
2728             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding ||
2729             test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2730                 handle_issuing_new_read_requests5(sh, &s, disks);
2731
2732         /* Now we check to see if any write operations have recently
2733          * completed
2734          */
2735
2736         /* leave prexor set until postxor is done, allows us to distinguish
2737          * a rmw from a rcw during biodrain
2738          */
2739         if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2740                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2741
2742                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2743                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2744                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2745
2746                 for (i = disks; i--; )
2747                         clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2748         }
2749
2750         /* if only POSTXOR is set then this is an 'expand' postxor */
2751         if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2752                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2753
2754                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2755                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2756                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2757
2758                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2759                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2760                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2761
2762                 /* All the 'written' buffers and the parity block are ready to
2763                  * be written back to disk
2764                  */
2765                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2766                 for (i = disks; i--; ) {
2767                         dev = &sh->dev[i];
2768                         if (test_bit(R5_LOCKED, &dev->flags) &&
2769                                 (i == sh->pd_idx || dev->written)) {
2770                                 pr_debug("Writing block %d\n", i);
2771                                 set_bit(R5_Wantwrite, &dev->flags);
2772                                 if (!test_and_set_bit(
2773                                     STRIPE_OP_IO, &sh->ops.pending))
2774                                         sh->ops.count++;
2775                                 if (!test_bit(R5_Insync, &dev->flags) ||
2776                                     (i == sh->pd_idx && s.failed == 0))
2777                                         set_bit(STRIPE_INSYNC, &sh->state);
2778                         }
2779                 }
2780                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2781                         atomic_dec(&conf->preread_active_stripes);
2782                         if (atomic_read(&conf->preread_active_stripes) <
2783                                 IO_THRESHOLD)
2784                                 md_wakeup_thread(conf->mddev->thread);
2785                 }
2786         }
2787
2788         /* Now to consider new write requests and what else, if anything
2789          * should be read.  We do not handle new writes when:
2790          * 1/ A 'write' operation (copy+xor) is already in flight.
2791          * 2/ A 'check' operation is in flight, as it may clobber the parity
2792          *    block.
2793          */
2794         if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2795                           !test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2796                 handle_issuing_new_write_requests5(conf, sh, &s, disks);
2797
2798         /* maybe we need to check and possibly fix the parity for this stripe
2799          * Any reads will already have been scheduled, so we just see if enough
2800          * data is available.  The parity check is held off while parity
2801          * dependent operations are in flight.
2802          */
2803         if ((s.syncing && s.locked == 0 &&
2804              !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2805              !test_bit(STRIPE_INSYNC, &sh->state)) ||
2806               test_bit(STRIPE_OP_CHECK, &sh->ops.pending) ||
2807               test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending))
2808                 handle_parity_checks5(conf, sh, &s, disks);
2809
2810         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2811                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2812                 clear_bit(STRIPE_SYNCING, &sh->state);
2813         }
2814
2815         /* If the failed drive is just a ReadError, then we might need to progress
2816          * the repair/check process
2817          */
2818         if (s.failed == 1 && !conf->mddev->ro &&
2819             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2820             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2821             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2822                 ) {
2823                 dev = &sh->dev[s.failed_num];
2824                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2825                         set_bit(R5_Wantwrite, &dev->flags);
2826                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2827                                 sh->ops.count++;
2828                         set_bit(R5_ReWrite, &dev->flags);
2829                         set_bit(R5_LOCKED, &dev->flags);
2830                         s.locked++;
2831                 } else {
2832                         /* let's read it back */
2833                         set_bit(R5_Wantread, &dev->flags);
2834                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2835                                 sh->ops.count++;
2836                         set_bit(R5_LOCKED, &dev->flags);
2837                         s.locked++;
2838                 }
2839         }
2840
2841         /* Finish postxor operations initiated by the expansion
2842          * process
2843          */
2844         if (test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete) &&
2845                 !test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending)) {
2846
2847                 clear_bit(STRIPE_EXPANDING, &sh->state);
2848
2849                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2850                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2851                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2852
2853                 for (i = conf->raid_disks; i--; ) {
2854                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2855                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2856                                 sh->ops.count++;
2857                 }
2858         }
2859
2860         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2861                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2862                 /* Need to write out all blocks after computing parity */
2863                 sh->disks = conf->raid_disks;
2864                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2865                         conf->raid_disks);
2866                 s.locked += handle_write_operations5(sh, 1, 1);
2867         } else if (s.expanded &&
2868                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2869                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2870                 atomic_dec(&conf->reshape_stripes);
2871                 wake_up(&conf->wait_for_overlap);
2872                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2873         }
2874
2875         if (s.expanding && s.locked == 0 &&
2876             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2877                 handle_stripe_expansion(conf, sh, NULL);
2878
2879         if (sh->ops.count)
2880                 pending = get_stripe_work(sh);
2881
2882         spin_unlock(&sh->lock);
2883
2884         if (pending)
2885                 raid5_run_ops(sh, pending);
2886
2887         return_io(return_bi);
2888
2889 }
2890
2891 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2892 {
2893         raid6_conf_t *conf = sh->raid_conf;
2894         int disks = sh->disks;
2895         struct bio *return_bi = NULL;
2896         int i, pd_idx = sh->pd_idx;
2897         struct stripe_head_state s;
2898         struct r6_state r6s;
2899         struct r5dev *dev, *pdev, *qdev;
2900
2901         r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2902         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2903                 "pd_idx=%d, qd_idx=%d\n",
2904                (unsigned long long)sh->sector, sh->state,
2905                atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2906         memset(&s, 0, sizeof(s));
2907
2908         spin_lock(&sh->lock);
2909         clear_bit(STRIPE_HANDLE, &sh->state);
2910         clear_bit(STRIPE_DELAYED, &sh->state);
2911
2912         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2913         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2914         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2915         /* Now to look around and see what can be done */
2916
2917         rcu_read_lock();
2918         for (i=disks; i--; ) {
2919                 mdk_rdev_t *rdev;
2920                 dev = &sh->dev[i];
2921                 clear_bit(R5_Insync, &dev->flags);
2922
2923                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2924                         i, dev->flags, dev->toread, dev->towrite, dev->written);
2925                 /* maybe we can reply to a read */
2926                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2927                         struct bio *rbi, *rbi2;
2928                         pr_debug("Return read for disc %d\n", i);
2929                         spin_lock_irq(&conf->device_lock);
2930                         rbi = dev->toread;
2931                         dev->toread = NULL;
2932                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2933                                 wake_up(&conf->wait_for_overlap);
2934                         spin_unlock_irq(&conf->device_lock);
2935                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2936                                 copy_data(0, rbi, dev->page, dev->sector);
2937                                 rbi2 = r5_next_bio(rbi, dev->sector);
2938                                 spin_lock_irq(&conf->device_lock);
2939                                 if (--rbi->bi_phys_segments == 0) {
2940                                         rbi->bi_next = return_bi;
2941                                         return_bi = rbi;
2942                                 }
2943                                 spin_unlock_irq(&conf->device_lock);
2944                                 rbi = rbi2;
2945                         }
2946                 }
2947
2948                 /* now count some things */
2949                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2950                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2951
2952
2953                 if (dev->toread)
2954                         s.to_read++;
2955                 if (dev->towrite) {
2956                         s.to_write++;
2957                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2958                                 s.non_overwrite++;
2959                 }
2960                 if (dev->written)
2961                         s.written++;
2962                 rdev = rcu_dereference(conf->disks[i].rdev);
2963                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2964                         /* The ReadError flag will just be confusing now */
2965                         clear_bit(R5_ReadError, &dev->flags);
2966                         clear_bit(R5_ReWrite, &dev->flags);
2967                 }
2968                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2969                     || test_bit(R5_ReadError, &dev->flags)) {
2970                         if (s.failed < 2)
2971                                 r6s.failed_num[s.failed] = i;
2972                         s.failed++;
2973                 } else
2974                         set_bit(R5_Insync, &dev->flags);
2975         }
2976         rcu_read_unlock();
2977         pr_debug("locked=%d uptodate=%d to_read=%d"
2978                " to_write=%d failed=%d failed_num=%d,%d\n",
2979                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
2980                r6s.failed_num[0], r6s.failed_num[1]);
2981         /* check if the array has lost >2 devices and, if so, some requests
2982          * might need to be failed
2983          */
2984         if (s.failed > 2 && s.to_read+s.to_write+s.written)
2985                 handle_requests_to_failed_array(conf, sh, &s, disks,
2986                                                 &return_bi);
2987         if (s.failed > 2 && s.syncing) {
2988                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2989                 clear_bit(STRIPE_SYNCING, &sh->state);
2990                 s.syncing = 0;
2991         }
2992
2993         /*
2994          * might be able to return some write requests if the parity blocks
2995          * are safe, or on a failed drive
2996          */
2997         pdev = &sh->dev[pd_idx];
2998         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
2999                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3000         qdev = &sh->dev[r6s.qd_idx];
3001         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
3002                 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
3003
3004         if ( s.written &&
3005              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3006                              && !test_bit(R5_LOCKED, &pdev->flags)
3007                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3008              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3009                              && !test_bit(R5_LOCKED, &qdev->flags)
3010                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3011                 handle_completed_write_requests(conf, sh, disks, &return_bi);
3012
3013         /* Now we might consider reading some blocks, either to check/generate
3014          * parity, or to satisfy requests
3015          * or to load a block that is being partially written.
3016          */
3017         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3018             (s.syncing && (s.uptodate < disks)) || s.expanding)
3019                 handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
3020
3021         /* now to consider writing and what else, if anything should be read */
3022         if (s.to_write)
3023                 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
3024
3025         /* maybe we need to check and possibly fix the parity for this stripe
3026          * Any reads will already have been scheduled, so we just see if enough
3027          * data is available
3028          */
3029         if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3030                 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3031
3032         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3033                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3034                 clear_bit(STRIPE_SYNCING, &sh->state);
3035         }
3036
3037         /* If the failed drives are just a ReadError, then we might need
3038          * to progress the repair/check process
3039          */
3040         if (s.failed <= 2 && !conf->mddev->ro)
3041                 for (i = 0; i < s.failed; i++) {
3042                         dev = &sh->dev[r6s.failed_num[i]];
3043                         if (test_bit(R5_ReadError, &dev->flags)
3044                             && !test_bit(R5_LOCKED, &dev->flags)
3045                             && test_bit(R5_UPTODATE, &dev->flags)
3046                                 ) {
3047                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3048                                         set_bit(R5_Wantwrite, &dev->flags);
3049                                         set_bit(R5_ReWrite, &dev->flags);
3050                                         set_bit(R5_LOCKED, &dev->flags);
3051                                 } else {
3052                                         /* let's read it back */
3053                                         set_bit(R5_Wantread, &dev->flags);
3054                                         set_bit(R5_LOCKED, &dev->flags);
3055                                 }
3056                         }
3057                 }
3058
3059         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3060                 /* Need to write out all blocks after computing P&Q */
3061                 sh->disks = conf->raid_disks;
3062                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
3063                                              conf->raid_disks);
3064                 compute_parity6(sh, RECONSTRUCT_WRITE);
3065                 for (i = conf->raid_disks ; i-- ;  ) {
3066                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3067                         s.locked++;
3068                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3069                 }
3070                 clear_bit(STRIPE_EXPANDING, &sh->state);
3071         } else if (s.expanded) {
3072                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3073                 atomic_dec(&conf->reshape_stripes);
3074                 wake_up(&conf->wait_for_overlap);
3075                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3076         }
3077
3078         if (s.expanding && s.locked == 0 &&
3079             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
3080                 handle_stripe_expansion(conf, sh, &r6s);
3081
3082         spin_unlock(&sh->lock);
3083
3084         return_io(return_bi);
3085
3086         for (i=disks; i-- ;) {
3087                 int rw;
3088                 struct bio *bi;
3089                 mdk_rdev_t *rdev;
3090                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
3091                         rw = WRITE;
3092                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
3093                         rw = READ;
3094                 else
3095                         continue;
3096
3097                 bi = &sh->dev[i].req;
3098
3099                 bi->bi_rw = rw;
3100                 if (rw == WRITE)
3101                         bi->bi_end_io = raid5_end_write_request;
3102                 else
3103                         bi->bi_end_io = raid5_end_read_request;
3104
3105                 rcu_read_lock();
3106                 rdev = rcu_dereference(conf->disks[i].rdev);
3107                 if (rdev && test_bit(Faulty, &rdev->flags))
3108                         rdev = NULL;
3109                 if (rdev)
3110                         atomic_inc(&rdev->nr_pending);
3111                 rcu_read_unlock();
3112
3113                 if (rdev) {
3114                         if (s.syncing || s.expanding || s.expanded)
3115                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
3116
3117                         bi->bi_bdev = rdev->bdev;
3118                         pr_debug("for %llu schedule op %ld on disc %d\n",
3119                                 (unsigned long long)sh->sector, bi->bi_rw, i);
3120                         atomic_inc(&sh->count);
3121                         bi->bi_sector = sh->sector + rdev->data_offset;
3122                         bi->bi_flags = 1 << BIO_UPTODATE;
3123                         bi->bi_vcnt = 1;
3124                         bi->bi_max_vecs = 1;
3125                         bi->bi_idx = 0;
3126                         bi->bi_io_vec = &sh->dev[i].vec;
3127                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
3128                         bi->bi_io_vec[0].bv_offset = 0;
3129                         bi->bi_size = STRIPE_SIZE;
3130                         bi->bi_next = NULL;
3131                         if (rw == WRITE &&
3132                             test_bit(R5_ReWrite, &sh->dev[i].flags))
3133                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
3134                         generic_make_request(bi);
3135                 } else {
3136                         if (rw == WRITE)
3137                                 set_bit(STRIPE_DEGRADED, &sh->state);
3138                         pr_debug("skip op %ld on disc %d for sector %llu\n",
3139                                 bi->bi_rw, i, (unsigned long long)sh->sector);
3140                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
3141                         set_bit(STRIPE_HANDLE, &sh->state);
3142                 }
3143         }
3144 }
3145
3146 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3147 {
3148         if (sh->raid_conf->level == 6)
3149                 handle_stripe6(sh, tmp_page);
3150         else
3151                 handle_stripe5(sh);
3152 }
3153
3154
3155
3156 static void raid5_activate_delayed(raid5_conf_t *conf)
3157 {
3158         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3159                 while (!list_empty(&conf->delayed_list)) {
3160                         struct list_head *l = conf->delayed_list.next;
3161                         struct stripe_head *sh;
3162                         sh = list_entry(l, struct stripe_head, lru);
3163                         list_del_init(l);
3164                         clear_bit(STRIPE_DELAYED, &sh->state);
3165                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3166                                 atomic_inc(&conf->preread_active_stripes);
3167                         list_add_tail(&sh->lru, &conf->handle_list);
3168                 }
3169         } else
3170                 blk_plug_device(conf->mddev->queue);
3171 }
3172
3173 static void activate_bit_delay(raid5_conf_t *conf)
3174 {
3175         /* device_lock is held */
3176         struct list_head head;
3177         list_add(&head, &conf->bitmap_list);
3178         list_del_init(&conf->bitmap_list);
3179         while (!list_empty(&head)) {
3180                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3181                 list_del_init(&sh->lru);
3182                 atomic_inc(&sh->count);
3183                 __release_stripe(conf, sh);
3184         }
3185 }
3186
3187 static void unplug_slaves(mddev_t *mddev)
3188 {
3189         raid5_conf_t *conf = mddev_to_conf(mddev);
3190         int i;
3191
3192         rcu_read_lock();
3193         for (i=0; i<mddev->raid_disks; i++) {
3194                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3195                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3196                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3197
3198                         atomic_inc(&rdev->nr_pending);
3199                         rcu_read_unlock();
3200
3201                         blk_unplug(r_queue);
3202
3203                         rdev_dec_pending(rdev, mddev);
3204                         rcu_read_lock();
3205                 }
3206         }
3207         rcu_read_unlock();
3208 }
3209
3210 static void raid5_unplug_device(struct request_queue *q)
3211 {
3212         mddev_t *mddev = q->queuedata;
3213         raid5_conf_t *conf = mddev_to_conf(mddev);
3214         unsigned long flags;
3215
3216         spin_lock_irqsave(&conf->device_lock, flags);
3217
3218         if (blk_remove_plug(q)) {
3219                 conf->seq_flush++;
3220                 raid5_activate_delayed(conf);
3221         }
3222         md_wakeup_thread(mddev->thread);
3223
3224         spin_unlock_irqrestore(&conf->device_lock, flags);
3225
3226         unplug_slaves(mddev);
3227 }
3228
3229 static int raid5_congested(void *data, int bits)
3230 {
3231         mddev_t *mddev = data;
3232         raid5_conf_t *conf = mddev_to_conf(mddev);
3233
3234         /* No difference between reads and writes.  Just check
3235          * how busy the stripe_cache is
3236          */
3237         if (conf->inactive_blocked)
3238                 return 1;
3239         if (conf->quiesce)
3240                 return 1;
3241         if (list_empty_careful(&conf->inactive_list))
3242                 return 1;
3243
3244         return 0;
3245 }
3246
3247 /* We want read requests to align with chunks where possible,
3248  * but write requests don't need to.
3249  */
3250 static int raid5_mergeable_bvec(struct request_queue *q, struct bio *bio, struct bio_vec *biovec)
3251 {
3252         mddev_t *mddev = q->queuedata;
3253         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3254         int max;
3255         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3256         unsigned int bio_sectors = bio->bi_size >> 9;
3257
3258         if (bio_data_dir(bio) == WRITE)
3259                 return biovec->bv_len; /* always allow writes to be mergeable */
3260
3261         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3262         if (max < 0) max = 0;
3263         if (max <= biovec->bv_len && bio_sectors == 0)
3264                 return biovec->bv_len;
3265         else
3266                 return max;
3267 }
3268
3269
3270 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3271 {
3272         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3273         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3274         unsigned int bio_sectors = bio->bi_size >> 9;
3275
3276         return  chunk_sectors >=
3277                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3278 }
3279
3280 /*
3281  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3282  *  later sampled by raid5d.
3283  */
3284 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3285 {
3286         unsigned long flags;
3287
3288         spin_lock_irqsave(&conf->device_lock, flags);
3289
3290         bi->bi_next = conf->retry_read_aligned_list;
3291         conf->retry_read_aligned_list = bi;
3292
3293         spin_unlock_irqrestore(&conf->device_lock, flags);
3294         md_wakeup_thread(conf->mddev->thread);
3295 }
3296
3297
3298 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3299 {
3300         struct bio *bi;
3301
3302         bi = conf->retry_read_aligned;
3303         if (bi) {
3304                 conf->retry_read_aligned = NULL;
3305                 return bi;
3306         }
3307         bi = conf->retry_read_aligned_list;
3308         if(bi) {
3309                 conf->retry_read_aligned_list = bi->bi_next;
3310                 bi->bi_next = NULL;
3311                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3312                 bi->bi_hw_segments = 0; /* count of processed stripes */
3313         }
3314
3315         return bi;
3316 }
3317
3318
3319 /*
3320  *  The "raid5_align_endio" should check if the read succeeded and if it
3321  *  did, call bio_endio on the original bio (having bio_put the new bio
3322  *  first).
3323  *  If the read failed..
3324  */
3325 static void raid5_align_endio(struct bio *bi, int error)
3326 {
3327         struct bio* raid_bi  = bi->bi_private;
3328         mddev_t *mddev;
3329         raid5_conf_t *conf;
3330         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3331         mdk_rdev_t *rdev;
3332
3333         bio_put(bi);
3334
3335         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3336         conf = mddev_to_conf(mddev);
3337         rdev = (void*)raid_bi->bi_next;
3338         raid_bi->bi_next = NULL;
3339
3340         rdev_dec_pending(rdev, conf->mddev);
3341
3342         if (!error && uptodate) {
3343                 bio_endio(raid_bi, 0);
3344                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3345                         wake_up(&conf->wait_for_stripe);
3346                 return;
3347         }
3348
3349
3350         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3351
3352         add_bio_to_retry(raid_bi, conf);
3353 }
3354
3355 static int bio_fits_rdev(struct bio *bi)
3356 {
3357         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3358
3359         if ((bi->bi_size>>9) > q->max_sectors)
3360                 return 0;
3361         blk_recount_segments(q, bi);
3362         if (bi->bi_phys_segments > q->max_phys_segments ||
3363             bi->bi_hw_segments > q->max_hw_segments)
3364                 return 0;
3365
3366         if (q->merge_bvec_fn)
3367                 /* it's too hard to apply the merge_bvec_fn at this stage,
3368                  * just just give up
3369                  */
3370                 return 0;
3371
3372         return 1;
3373 }
3374
3375
3376 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3377 {
3378         mddev_t *mddev = q->queuedata;
3379         raid5_conf_t *conf = mddev_to_conf(mddev);
3380         const unsigned int raid_disks = conf->raid_disks;
3381         const unsigned int data_disks = raid_disks - conf->max_degraded;
3382         unsigned int dd_idx, pd_idx;
3383         struct bio* align_bi;
3384         mdk_rdev_t *rdev;
3385
3386         if (!in_chunk_boundary(mddev, raid_bio)) {
3387                 pr_debug("chunk_aligned_read : non aligned\n");
3388                 return 0;
3389         }
3390         /*
3391          * use bio_clone to make a copy of the bio
3392          */
3393         align_bi = bio_clone(raid_bio, GFP_NOIO);
3394         if (!align_bi)
3395                 return 0;
3396         /*
3397          *   set bi_end_io to a new function, and set bi_private to the
3398          *     original bio.
3399          */
3400         align_bi->bi_end_io  = raid5_align_endio;
3401         align_bi->bi_private = raid_bio;
3402         /*
3403          *      compute position
3404          */
3405         align_bi->bi_sector =  raid5_compute_sector(raid_bio->bi_sector,
3406                                         raid_disks,
3407                                         data_disks,
3408                                         &dd_idx,
3409                                         &pd_idx,
3410                                         conf);
3411
3412         rcu_read_lock();
3413         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3414         if (rdev && test_bit(In_sync, &rdev->flags)) {
3415                 atomic_inc(&rdev->nr_pending);
3416                 rcu_read_unlock();
3417                 raid_bio->bi_next = (void*)rdev;
3418                 align_bi->bi_bdev =  rdev->bdev;
3419                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3420                 align_bi->bi_sector += rdev->data_offset;
3421
3422                 if (!bio_fits_rdev(align_bi)) {
3423                         /* too big in some way */
3424                         bio_put(align_bi);
3425                         rdev_dec_pending(rdev, mddev);
3426                         return 0;
3427                 }
3428
3429                 spin_lock_irq(&conf->device_lock);
3430                 wait_event_lock_irq(conf->wait_for_stripe,
3431                                     conf->quiesce == 0,
3432                                     conf->device_lock, /* nothing */);
3433                 atomic_inc(&conf->active_aligned_reads);
3434                 spin_unlock_irq(&conf->device_lock);
3435
3436                 generic_make_request(align_bi);
3437                 return 1;
3438         } else {
3439                 rcu_read_unlock();
3440                 bio_put(align_bi);
3441                 return 0;
3442         }
3443 }
3444
3445
3446 static int make_request(struct request_queue *q, struct bio * bi)
3447 {
3448         mddev_t *mddev = q->queuedata;
3449         raid5_conf_t *conf = mddev_to_conf(mddev);
3450         unsigned int dd_idx, pd_idx;
3451         sector_t new_sector;
3452         sector_t logical_sector, last_sector;
3453         struct stripe_head *sh;
3454         const int rw = bio_data_dir(bi);
3455         int remaining;
3456
3457         if (unlikely(bio_barrier(bi))) {
3458                 bio_endio(bi, -EOPNOTSUPP);
3459                 return 0;
3460         }
3461
3462         md_write_start(mddev, bi);
3463
3464         disk_stat_inc(mddev->gendisk, ios[rw]);
3465         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3466
3467         if (rw == READ &&
3468              mddev->reshape_position == MaxSector &&
3469              chunk_aligned_read(q,bi))
3470                 return 0;
3471
3472         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3473         last_sector = bi->bi_sector + (bi->bi_size>>9);
3474         bi->bi_next = NULL;
3475         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3476
3477         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3478                 DEFINE_WAIT(w);
3479                 int disks, data_disks;
3480
3481         retry:
3482                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3483                 if (likely(conf->expand_progress == MaxSector))
3484                         disks = conf->raid_disks;
3485                 else {
3486                         /* spinlock is needed as expand_progress may be
3487                          * 64bit on a 32bit platform, and so it might be
3488                          * possible to see a half-updated value
3489                          * Ofcourse expand_progress could change after
3490                          * the lock is dropped, so once we get a reference
3491                          * to the stripe that we think it is, we will have
3492                          * to check again.
3493                          */
3494                         spin_lock_irq(&conf->device_lock);
3495                         disks = conf->raid_disks;
3496                         if (logical_sector >= conf->expand_progress)
3497                                 disks = conf->previous_raid_disks;
3498                         else {
3499                                 if (logical_sector >= conf->expand_lo) {
3500                                         spin_unlock_irq(&conf->device_lock);
3501                                         schedule();
3502                                         goto retry;
3503                                 }
3504                         }
3505                         spin_unlock_irq(&conf->device_lock);
3506                 }
3507                 data_disks = disks - conf->max_degraded;
3508
3509                 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3510                                                   &dd_idx, &pd_idx, conf);
3511                 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3512                         (unsigned long long)new_sector, 
3513                         (unsigned long long)logical_sector);
3514
3515                 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3516                 if (sh) {
3517                         if (unlikely(conf->expand_progress != MaxSector)) {
3518                                 /* expansion might have moved on while waiting for a
3519                                  * stripe, so we must do the range check again.
3520                                  * Expansion could still move past after this
3521                                  * test, but as we are holding a reference to
3522                                  * 'sh', we know that if that happens,
3523                                  *  STRIPE_EXPANDING will get set and the expansion
3524                                  * won't proceed until we finish with the stripe.
3525                                  */
3526                                 int must_retry = 0;
3527                                 spin_lock_irq(&conf->device_lock);
3528                                 if (logical_sector <  conf->expand_progress &&
3529                                     disks == conf->previous_raid_disks)
3530                                         /* mismatch, need to try again */
3531                                         must_retry = 1;
3532                                 spin_unlock_irq(&conf->device_lock);
3533                                 if (must_retry) {
3534                                         release_stripe(sh);
3535                                         goto retry;
3536                                 }
3537                         }
3538                         /* FIXME what if we get a false positive because these
3539                          * are being updated.
3540                          */
3541                         if (logical_sector >= mddev->suspend_lo &&
3542                             logical_sector < mddev->suspend_hi) {
3543                                 release_stripe(sh);
3544                                 schedule();
3545                                 goto retry;
3546                         }
3547
3548                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3549                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3550                                 /* Stripe is busy expanding or
3551                                  * add failed due to overlap.  Flush everything
3552                                  * and wait a while
3553                                  */
3554                                 raid5_unplug_device(mddev->queue);
3555                                 release_stripe(sh);
3556                                 schedule();
3557                                 goto retry;
3558                         }
3559                         finish_wait(&conf->wait_for_overlap, &w);
3560                         set_bit(STRIPE_HANDLE, &sh->state);
3561                         clear_bit(STRIPE_DELAYED, &sh->state);
3562                         release_stripe(sh);
3563                 } else {
3564                         /* cannot get stripe for read-ahead, just give-up */
3565                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3566                         finish_wait(&conf->wait_for_overlap, &w);
3567                         break;
3568                 }
3569                         
3570         }
3571         spin_lock_irq(&conf->device_lock);
3572         remaining = --bi->bi_phys_segments;
3573         spin_unlock_irq(&conf->device_lock);
3574         if (remaining == 0) {
3575
3576                 if ( rw == WRITE )
3577                         md_write_end(mddev);
3578
3579                 bi->bi_end_io(bi,
3580                               test_bit(BIO_UPTODATE, &bi->bi_flags)
3581                                 ? 0 : -EIO);
3582         }
3583         return 0;
3584 }
3585
3586 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3587 {
3588         /* reshaping is quite different to recovery/resync so it is
3589          * handled quite separately ... here.
3590          *
3591          * On each call to sync_request, we gather one chunk worth of
3592          * destination stripes and flag them as expanding.
3593          * Then we find all the source stripes and request reads.
3594          * As the reads complete, handle_stripe will copy the data
3595          * into the destination stripe and release that stripe.
3596          */
3597         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3598         struct stripe_head *sh;
3599         int pd_idx;
3600         sector_t first_sector, last_sector;
3601         int raid_disks = conf->previous_raid_disks;
3602         int data_disks = raid_disks - conf->max_degraded;
3603         int new_data_disks = conf->raid_disks - conf->max_degraded;
3604         int i;
3605         int dd_idx;
3606         sector_t writepos, safepos, gap;
3607
3608         if (sector_nr == 0 &&
3609             conf->expand_progress != 0) {
3610                 /* restarting in the middle, skip the initial sectors */
3611                 sector_nr = conf->expand_progress;
3612                 sector_div(sector_nr, new_data_disks);
3613                 *skipped = 1;
3614                 return sector_nr;
3615         }
3616
3617         /* we update the metadata when there is more than 3Meg
3618          * in the block range (that is rather arbitrary, should
3619          * probably be time based) or when the data about to be
3620          * copied would over-write the source of the data at
3621          * the front of the range.
3622          * i.e. one new_stripe forward from expand_progress new_maps
3623          * to after where expand_lo old_maps to
3624          */
3625         writepos = conf->expand_progress +
3626                 conf->chunk_size/512*(new_data_disks);
3627         sector_div(writepos, new_data_disks);
3628         safepos = conf->expand_lo;
3629         sector_div(safepos, data_disks);
3630         gap = conf->expand_progress - conf->expand_lo;
3631
3632         if (writepos >= safepos ||
3633             gap > (new_data_disks)*3000*2 /*3Meg*/) {
3634                 /* Cannot proceed until we've updated the superblock... */
3635                 wait_event(conf->wait_for_overlap,
3636                            atomic_read(&conf->reshape_stripes)==0);
3637                 mddev->reshape_position = conf->expand_progress;
3638                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3639                 md_wakeup_thread(mddev->thread);
3640                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3641                            kthread_should_stop());
3642                 spin_lock_irq(&conf->device_lock);
3643                 conf->expand_lo = mddev->reshape_position;
3644                 spin_unlock_irq(&conf->device_lock);
3645                 wake_up(&conf->wait_for_overlap);
3646         }
3647
3648         for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3649                 int j;
3650                 int skipped = 0;
3651                 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3652                 sh = get_active_stripe(conf, sector_nr+i,
3653                                        conf->raid_disks, pd_idx, 0);
3654                 set_bit(STRIPE_EXPANDING, &sh->state);
3655                 atomic_inc(&conf->reshape_stripes);
3656                 /* If any of this stripe is beyond the end of the old
3657                  * array, then we need to zero those blocks
3658                  */
3659                 for (j=sh->disks; j--;) {
3660                         sector_t s;
3661                         if (j == sh->pd_idx)
3662                                 continue;
3663                         if (conf->level == 6 &&
3664                             j == raid6_next_disk(sh->pd_idx, sh->disks))
3665                                 continue;
3666                         s = compute_blocknr(sh, j);
3667                         if (s < (mddev->array_size<<1)) {
3668                                 skipped = 1;
3669                                 continue;
3670                         }
3671                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3672                         set_bit(R5_Expanded, &sh->dev[j].flags);
3673                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
3674                 }
3675                 if (!skipped) {
3676                         set_bit(STRIPE_EXPAND_READY, &sh->state);
3677                         set_bit(STRIPE_HANDLE, &sh->state);
3678                 }
3679                 release_stripe(sh);
3680         }
3681         spin_lock_irq(&conf->device_lock);
3682         conf->expand_progress = (sector_nr + i) * new_data_disks;
3683         spin_unlock_irq(&conf->device_lock);
3684         /* Ok, those stripe are ready. We can start scheduling
3685          * reads on the source stripes.
3686          * The source stripes are determined by mapping the first and last
3687          * block on the destination stripes.
3688          */
3689         first_sector =
3690                 raid5_compute_sector(sector_nr*(new_data_disks),
3691                                      raid_disks, data_disks,
3692                                      &dd_idx, &pd_idx, conf);
3693         last_sector =
3694                 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3695                                      *(new_data_disks) -1,
3696                                      raid_disks, data_disks,
3697                                      &dd_idx, &pd_idx, conf);
3698         if (last_sector >= (mddev->size<<1))
3699                 last_sector = (mddev->size<<1)-1;
3700         while (first_sector <= last_sector) {
3701                 pd_idx = stripe_to_pdidx(first_sector, conf,
3702                                          conf->previous_raid_disks);
3703                 sh = get_active_stripe(conf, first_sector,
3704                                        conf->previous_raid_disks, pd_idx, 0);
3705                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3706                 set_bit(STRIPE_HANDLE, &sh->state);
3707                 release_stripe(sh);
3708                 first_sector += STRIPE_SECTORS;
3709         }
3710         /* If this takes us to the resync_max point where we have to pause,
3711          * then we need to write out the superblock.
3712          */
3713         sector_nr += conf->chunk_size>>9;
3714         if (sector_nr >= mddev->resync_max) {
3715                 /* Cannot proceed until we've updated the superblock... */
3716                 wait_event(conf->wait_for_overlap,
3717                            atomic_read(&conf->reshape_stripes) == 0);
3718                 mddev->reshape_position = conf->expand_progress;
3719                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3720                 md_wakeup_thread(mddev->thread);
3721                 wait_event(mddev->sb_wait,
3722                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3723                            || kthread_should_stop());
3724                 spin_lock_irq(&conf->device_lock);
3725                 conf->expand_lo = mddev->reshape_position;
3726                 spin_unlock_irq(&conf->device_lock);
3727                 wake_up(&conf->wait_for_overlap);
3728         }
3729         return conf->chunk_size>>9;
3730 }
3731
3732 /* FIXME go_faster isn't used */
3733 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3734 {
3735         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3736         struct stripe_head *sh;
3737         int pd_idx;
3738         int raid_disks = conf->raid_disks;
3739         sector_t max_sector = mddev->size << 1;
3740         int sync_blocks;
3741         int still_degraded = 0;
3742         int i;
3743
3744         if (sector_nr >= max_sector) {
3745                 /* just being told to finish up .. nothing much to do */
3746                 unplug_slaves(mddev);
3747                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3748                         end_reshape(conf);
3749                         return 0;
3750                 }
3751
3752                 if (mddev->curr_resync < max_sector) /* aborted */
3753                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3754                                         &sync_blocks, 1);
3755                 else /* completed sync */
3756                         conf->fullsync = 0;
3757                 bitmap_close_sync(mddev->bitmap);
3758
3759                 return 0;
3760         }
3761
3762         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3763                 return reshape_request(mddev, sector_nr, skipped);
3764
3765         /* No need to check resync_max as we never do more than one
3766          * stripe, and as resync_max will always be on a chunk boundary,
3767          * if the check in md_do_sync didn't fire, there is no chance
3768          * of overstepping resync_max here
3769          */
3770
3771         /* if there is too many failed drives and we are trying
3772          * to resync, then assert that we are finished, because there is
3773          * nothing we can do.
3774          */
3775         if (mddev->degraded >= conf->max_degraded &&
3776             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3777                 sector_t rv = (mddev->size << 1) - sector_nr;
3778                 *skipped = 1;
3779                 return rv;
3780         }
3781         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3782             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3783             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3784                 /* we can skip this block, and probably more */
3785                 sync_blocks /= STRIPE_SECTORS;
3786                 *skipped = 1;
3787                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3788         }
3789
3790
3791         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3792
3793         pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3794         sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3795         if (sh == NULL) {
3796                 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3797                 /* make sure we don't swamp the stripe cache if someone else
3798                  * is trying to get access
3799                  */
3800                 schedule_timeout_uninterruptible(1);
3801         }
3802         /* Need to check if array will still be degraded after recovery/resync
3803          * We don't need to check the 'failed' flag as when that gets set,
3804          * recovery aborts.
3805          */
3806         for (i=0; i<mddev->raid_disks; i++)
3807                 if (conf->disks[i].rdev == NULL)
3808                         still_degraded = 1;
3809
3810         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3811
3812         spin_lock(&sh->lock);
3813         set_bit(STRIPE_SYNCING, &sh->state);
3814         clear_bit(STRIPE_INSYNC, &sh->state);
3815         spin_unlock(&sh->lock);
3816
3817         handle_stripe(sh, NULL);
3818         release_stripe(sh);
3819
3820         return STRIPE_SECTORS;
3821 }
3822
3823 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3824 {
3825         /* We may not be able to submit a whole bio at once as there
3826          * may not be enough stripe_heads available.
3827          * We cannot pre-allocate enough stripe_heads as we may need
3828          * more than exist in the cache (if we allow ever large chunks).
3829          * So we do one stripe head at a time and record in
3830          * ->bi_hw_segments how many have been done.
3831          *
3832          * We *know* that this entire raid_bio is in one chunk, so
3833          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3834          */
3835         struct stripe_head *sh;
3836         int dd_idx, pd_idx;
3837         sector_t sector, logical_sector, last_sector;
3838         int scnt = 0;
3839         int remaining;
3840         int handled = 0;
3841
3842         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3843         sector = raid5_compute_sector(  logical_sector,
3844                                         conf->raid_disks,
3845                                         conf->raid_disks - conf->max_degraded,
3846                                         &dd_idx,
3847                                         &pd_idx,
3848                                         conf);
3849         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3850
3851         for (; logical_sector < last_sector;
3852              logical_sector += STRIPE_SECTORS,
3853                      sector += STRIPE_SECTORS,
3854                      scnt++) {
3855
3856                 if (scnt < raid_bio->bi_hw_segments)
3857                         /* already done this stripe */
3858                         continue;
3859
3860                 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3861
3862                 if (!sh) {
3863                         /* failed to get a stripe - must wait */
3864                         raid_bio->bi_hw_segments = scnt;
3865                         conf->retry_read_aligned = raid_bio;
3866                         return handled;
3867                 }
3868
3869                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3870                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3871                         release_stripe(sh);
3872                         raid_bio->bi_hw_segments = scnt;
3873                         conf->retry_read_aligned = raid_bio;
3874                         return handled;
3875                 }
3876
3877                 handle_stripe(sh, NULL);
3878                 release_stripe(sh);
3879                 handled++;
3880         }
3881         spin_lock_irq(&conf->device_lock);
3882         remaining = --raid_bio->bi_phys_segments;
3883         spin_unlock_irq(&conf->device_lock);
3884         if (remaining == 0) {
3885
3886                 raid_bio->bi_end_io(raid_bio,
3887                               test_bit(BIO_UPTODATE, &raid_bio->bi_flags)
3888                                 ? 0 : -EIO);
3889         }
3890         if (atomic_dec_and_test(&conf->active_aligned_reads))
3891                 wake_up(&conf->wait_for_stripe);
3892         return handled;
3893 }
3894
3895
3896
3897 /*
3898  * This is our raid5 kernel thread.
3899  *
3900  * We scan the hash table for stripes which can be handled now.
3901  * During the scan, completed stripes are saved for us by the interrupt
3902  * handler, so that they will not have to wait for our next wakeup.
3903  */
3904 static void raid5d(mddev_t *mddev)
3905 {
3906         struct stripe_head *sh;
3907         raid5_conf_t *conf = mddev_to_conf(mddev);
3908         int handled;
3909
3910         pr_debug("+++ raid5d active\n");
3911
3912         md_check_recovery(mddev);
3913
3914         handled = 0;
3915         spin_lock_irq(&conf->device_lock);
3916         while (1) {
3917                 struct list_head *first;
3918                 struct bio *bio;
3919
3920                 if (conf->seq_flush != conf->seq_write) {
3921                         int seq = conf->seq_flush;
3922                         spin_unlock_irq(&conf->device_lock);
3923                         bitmap_unplug(mddev->bitmap);
3924                         spin_lock_irq(&conf->device_lock);
3925                         conf->seq_write = seq;
3926                         activate_bit_delay(conf);
3927                 }
3928
3929                 while ((bio = remove_bio_from_retry(conf))) {
3930                         int ok;
3931                         spin_unlock_irq(&conf->device_lock);
3932                         ok = retry_aligned_read(conf, bio);
3933                         spin_lock_irq(&conf->device_lock);
3934                         if (!ok)
3935                                 break;
3936                         handled++;
3937                 }
3938
3939                 if (list_empty(&conf->handle_list)) {
3940                         async_tx_issue_pending_all();
3941                         break;
3942                 }
3943
3944                 first = conf->handle_list.next;
3945                 sh = list_entry(first, struct stripe_head, lru);
3946
3947                 list_del_init(first);
3948                 atomic_inc(&sh->count);
3949                 BUG_ON(atomic_read(&sh->count)!= 1);
3950                 spin_unlock_irq(&conf->device_lock);
3951                 
3952                 handled++;
3953                 handle_stripe(sh, conf->spare_page);
3954                 release_stripe(sh);
3955
3956                 spin_lock_irq(&conf->device_lock);
3957         }
3958         pr_debug("%d stripes handled\n", handled);
3959
3960         spin_unlock_irq(&conf->device_lock);
3961
3962         unplug_slaves(mddev);
3963
3964         pr_debug("--- raid5d inactive\n");
3965 }
3966
3967 static ssize_t
3968 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
3969 {
3970         raid5_conf_t *conf = mddev_to_conf(mddev);
3971         if (conf)
3972                 return sprintf(page, "%d\n", conf->max_nr_stripes);
3973         else
3974                 return 0;
3975 }
3976
3977 static ssize_t
3978 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
3979 {
3980         raid5_conf_t *conf = mddev_to_conf(mddev);
3981         char *end;
3982         int new;
3983         if (len >= PAGE_SIZE)
3984                 return -EINVAL;
3985         if (!conf)
3986                 return -ENODEV;
3987
3988         new = simple_strtoul(page, &end, 10);
3989         if (!*page || (*end && *end != '\n') )
3990                 return -EINVAL;
3991         if (new <= 16 || new > 32768)
3992                 return -EINVAL;
3993         while (new < conf->max_nr_stripes) {
3994                 if (drop_one_stripe(conf))
3995                         conf->max_nr_stripes--;
3996                 else
3997                         break;
3998         }
3999         md_allow_write(mddev);
4000         while (new > conf->max_nr_stripes) {
4001                 if (grow_one_stripe(conf))
4002                         conf->max_nr_stripes++;
4003                 else break;
4004         }
4005         return len;
4006 }
4007
4008 static struct md_sysfs_entry
4009 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4010                                 raid5_show_stripe_cache_size,
4011                                 raid5_store_stripe_cache_size);
4012
4013 static ssize_t
4014 stripe_cache_active_show(mddev_t *mddev, char *page)
4015 {
4016         raid5_conf_t *conf = mddev_to_conf(mddev);
4017         if (conf)
4018                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4019         else
4020                 return 0;
4021 }
4022
4023 static struct md_sysfs_entry
4024 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4025
4026 static struct attribute *raid5_attrs[] =  {
4027         &raid5_stripecache_size.attr,
4028         &raid5_stripecache_active.attr,
4029         NULL,
4030 };
4031 static struct attribute_group raid5_attrs_group = {
4032         .name = NULL,
4033         .attrs = raid5_attrs,
4034 };
4035
4036 static int run(mddev_t *mddev)
4037 {
4038         raid5_conf_t *conf;
4039         int raid_disk, memory;
4040         mdk_rdev_t *rdev;
4041         struct disk_info *disk;
4042         struct list_head *tmp;
4043         int working_disks = 0;
4044
4045         if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4046                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4047                        mdname(mddev), mddev->level);
4048                 return -EIO;
4049         }
4050
4051         if (mddev->reshape_position != MaxSector) {
4052                 /* Check that we can continue the reshape.
4053                  * Currently only disks can change, it must
4054                  * increase, and we must be past the point where
4055                  * a stripe over-writes itself
4056                  */
4057                 sector_t here_new, here_old;
4058                 int old_disks;
4059                 int max_degraded = (mddev->level == 5 ? 1 : 2);
4060
4061                 if (mddev->new_level != mddev->level ||
4062                     mddev->new_layout != mddev->layout ||
4063                     mddev->new_chunk != mddev->chunk_size) {
4064                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4065                                "required - aborting.\n",
4066                                mdname(mddev));
4067                         return -EINVAL;
4068                 }
4069                 if (mddev->delta_disks <= 0) {
4070                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4071                                "(reduce disks) required - aborting.\n",
4072                                mdname(mddev));
4073                         return -EINVAL;
4074                 }
4075                 old_disks = mddev->raid_disks - mddev->delta_disks;
4076                 /* reshape_position must be on a new-stripe boundary, and one
4077                  * further up in new geometry must map after here in old
4078                  * geometry.
4079                  */
4080                 here_new = mddev->reshape_position;
4081                 if (sector_div(here_new, (mddev->chunk_size>>9)*
4082                                (mddev->raid_disks - max_degraded))) {
4083                         printk(KERN_ERR "raid5: reshape_position not "
4084                                "on a stripe boundary\n");
4085                         return -EINVAL;
4086                 }
4087                 /* here_new is the stripe we will write to */
4088                 here_old = mddev->reshape_position;
4089                 sector_div(here_old, (mddev->chunk_size>>9)*
4090                            (old_disks-max_degraded));
4091                 /* here_old is the first stripe that we might need to read
4092                  * from */
4093                 if (here_new >= here_old) {
4094                         /* Reading from the same stripe as writing to - bad */
4095                         printk(KERN_ERR "raid5: reshape_position too early for "
4096                                "auto-recovery - aborting.\n");
4097                         return -EINVAL;
4098                 }
4099                 printk(KERN_INFO "raid5: reshape will continue\n");
4100                 /* OK, we should be able to continue; */
4101         }
4102
4103
4104         mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4105         if ((conf = mddev->private) == NULL)
4106                 goto abort;
4107         if (mddev->reshape_position == MaxSector) {
4108                 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4109         } else {
4110                 conf->raid_disks = mddev->raid_disks;
4111                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4112         }
4113
4114         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4115                               GFP_KERNEL);
4116         if (!conf->disks)
4117                 goto abort;
4118
4119         conf->mddev = mddev;
4120
4121         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4122                 goto abort;
4123
4124         if (mddev->level == 6) {
4125                 conf->spare_page = alloc_page(GFP_KERNEL);
4126                 if (!conf->spare_page)
4127                         goto abort;
4128         }
4129         spin_lock_init(&conf->device_lock);
4130         init_waitqueue_head(&conf->wait_for_stripe);
4131         init_waitqueue_head(&conf->wait_for_overlap);
4132         INIT_LIST_HEAD(&conf->handle_list);
4133         INIT_LIST_HEAD(&conf->delayed_list);
4134         INIT_LIST_HEAD(&conf->bitmap_list);
4135         INIT_LIST_HEAD(&conf->inactive_list);
4136         atomic_set(&conf->active_stripes, 0);
4137         atomic_set(&conf->preread_active_stripes, 0);
4138         atomic_set(&conf->active_aligned_reads, 0);
4139
4140         pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4141
4142         rdev_for_each(rdev, tmp, mddev) {
4143                 raid_disk = rdev->raid_disk;
4144                 if (raid_disk >= conf->raid_disks
4145                     || raid_disk < 0)
4146                         continue;
4147                 disk = conf->disks + raid_disk;
4148
4149                 disk->rdev = rdev;
4150
4151                 if (test_bit(In_sync, &rdev->flags)) {
4152                         char b[BDEVNAME_SIZE];
4153                         printk(KERN_INFO "raid5: device %s operational as raid"
4154                                 " disk %d\n", bdevname(rdev->bdev,b),
4155                                 raid_disk);
4156                         working_disks++;
4157                 }
4158         }
4159
4160         /*
4161          * 0 for a fully functional array, 1 or 2 for a degraded array.
4162          */
4163         mddev->degraded = conf->raid_disks - working_disks;
4164         conf->mddev = mddev;
4165         conf->chunk_size = mddev->chunk_size;
4166         conf->level = mddev->level;
4167         if (conf->level == 6)
4168                 conf->max_degraded = 2;
4169         else
4170                 conf->max_degraded = 1;
4171         conf->algorithm = mddev->layout;
4172         conf->max_nr_stripes = NR_STRIPES;
4173         conf->expand_progress = mddev->reshape_position;
4174
4175         /* device size must be a multiple of chunk size */
4176         mddev->size &= ~(mddev->chunk_size/1024 -1);
4177         mddev->resync_max_sectors = mddev->size << 1;
4178
4179         if (conf->level == 6 && conf->raid_disks < 4) {
4180                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4181                        mdname(mddev), conf->raid_disks);
4182                 goto abort;
4183         }
4184         if (!conf->chunk_size || conf->chunk_size % 4) {
4185                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4186                         conf->chunk_size, mdname(mddev));
4187                 goto abort;
4188         }
4189         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4190                 printk(KERN_ERR 
4191                         "raid5: unsupported parity algorithm %d for %s\n",
4192                         conf->algorithm, mdname(mddev));
4193                 goto abort;
4194         }
4195         if (mddev->degraded > conf->max_degraded) {
4196                 printk(KERN_ERR "raid5: not enough operational devices for %s"
4197                         " (%d/%d failed)\n",
4198                         mdname(mddev), mddev->degraded, conf->raid_disks);
4199                 goto abort;
4200         }
4201
4202         if (mddev->degraded > 0 &&
4203             mddev->recovery_cp != MaxSector) {
4204                 if (mddev->ok_start_degraded)
4205                         printk(KERN_WARNING
4206                                "raid5: starting dirty degraded array: %s"
4207                                "- data corruption possible.\n",
4208                                mdname(mddev));
4209                 else {
4210                         printk(KERN_ERR
4211                                "raid5: cannot start dirty degraded array for %s\n",
4212                                mdname(mddev));
4213                         goto abort;
4214                 }
4215         }
4216
4217         {
4218                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4219                 if (!mddev->thread) {
4220                         printk(KERN_ERR 
4221                                 "raid5: couldn't allocate thread for %s\n",
4222                                 mdname(mddev));
4223                         goto abort;
4224                 }
4225         }
4226         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4227                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4228         if (grow_stripes(conf, conf->max_nr_stripes)) {
4229                 printk(KERN_ERR 
4230                         "raid5: couldn't allocate %dkB for buffers\n", memory);
4231                 shrink_stripes(conf);
4232                 md_unregister_thread(mddev->thread);
4233                 goto abort;
4234         } else
4235                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4236                         memory, mdname(mddev));
4237
4238         if (mddev->degraded == 0)
4239                 printk("raid5: raid level %d set %s active with %d out of %d"
4240                         " devices, algorithm %d\n", conf->level, mdname(mddev), 
4241                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4242                         conf->algorithm);
4243         else
4244                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4245                         " out of %d devices, algorithm %d\n", conf->level,
4246                         mdname(mddev), mddev->raid_disks - mddev->degraded,
4247                         mddev->raid_disks, conf->algorithm);
4248
4249         print_raid5_conf(conf);
4250
4251         if (conf->expand_progress != MaxSector) {
4252                 printk("...ok start reshape thread\n");
4253                 conf->expand_lo = conf->expand_progress;
4254                 atomic_set(&conf->reshape_stripes, 0);
4255                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4256                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4257                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4258                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4259                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4260                                                         "%s_reshape");
4261         }
4262
4263         /* read-ahead size must cover two whole stripes, which is
4264          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4265          */
4266         {
4267                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4268                 int stripe = data_disks *
4269                         (mddev->chunk_size / PAGE_SIZE);
4270                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4271                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4272         }
4273
4274         /* Ok, everything is just fine now */
4275         if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4276                 printk(KERN_WARNING
4277                        "raid5: failed to create sysfs attributes for %s\n",
4278                        mdname(mddev));
4279
4280         mddev->queue->unplug_fn = raid5_unplug_device;
4281         mddev->queue->backing_dev_info.congested_data = mddev;
4282         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4283
4284         mddev->array_size =  mddev->size * (conf->previous_raid_disks -
4285                                             conf->max_degraded);
4286
4287         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4288
4289         return 0;
4290 abort:
4291         if (conf) {
4292                 print_raid5_conf(conf);
4293                 safe_put_page(conf->spare_page);
4294                 kfree(conf->disks);
4295                 kfree(conf->stripe_hashtbl);
4296                 kfree(conf);
4297         }
4298         mddev->private = NULL;
4299         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4300         return -EIO;
4301 }
4302
4303
4304
4305 static int stop(mddev_t *mddev)
4306 {
4307         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4308
4309         md_unregister_thread(mddev->thread);
4310         mddev->thread = NULL;
4311         shrink_stripes(conf);
4312         kfree(conf->stripe_hashtbl);
4313         mddev->queue->backing_dev_info.congested_fn = NULL;
4314         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4315         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4316         kfree(conf->disks);
4317         kfree(conf);
4318         mddev->private = NULL;
4319         return 0;
4320 }
4321
4322 #ifdef DEBUG
4323 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4324 {
4325         int i;
4326
4327         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4328                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4329         seq_printf(seq, "sh %llu,  count %d.\n",
4330                    (unsigned long long)sh->sector, atomic_read(&sh->count));
4331         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4332         for (i = 0; i < sh->disks; i++) {
4333                 seq_printf(seq, "(cache%d: %p %ld) ",
4334                            i, sh->dev[i].page, sh->dev[i].flags);
4335         }
4336         seq_printf(seq, "\n");
4337 }
4338
4339 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4340 {
4341         struct stripe_head *sh;
4342         struct hlist_node *hn;
4343         int i;
4344
4345         spin_lock_irq(&conf->device_lock);
4346         for (i = 0; i < NR_HASH; i++) {
4347                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4348                         if (sh->raid_conf != conf)
4349                                 continue;
4350                         print_sh(seq, sh);
4351                 }
4352         }
4353         spin_unlock_irq(&conf->device_lock);
4354 }
4355 #endif
4356
4357 static void status (struct seq_file *seq, mddev_t *mddev)
4358 {
4359         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4360         int i;
4361
4362         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4363         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4364         for (i = 0; i < conf->raid_disks; i++)
4365                 seq_printf (seq, "%s",
4366                                conf->disks[i].rdev &&
4367                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4368         seq_printf (seq, "]");
4369 #ifdef DEBUG
4370         seq_printf (seq, "\n");
4371         printall(seq, conf);
4372 #endif
4373 }
4374
4375 static void print_raid5_conf (raid5_conf_t *conf)
4376 {
4377         int i;
4378         struct disk_info *tmp;
4379
4380         printk("RAID5 conf printout:\n");
4381         if (!conf) {
4382                 printk("(conf==NULL)\n");
4383                 return;
4384         }
4385         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4386                  conf->raid_disks - conf->mddev->degraded);
4387
4388         for (i = 0; i < conf->raid_disks; i++) {
4389                 char b[BDEVNAME_SIZE];
4390                 tmp = conf->disks + i;
4391                 if (tmp->rdev)
4392                 printk(" disk %d, o:%d, dev:%s\n",
4393                         i, !test_bit(Faulty, &tmp->rdev->flags),
4394                         bdevname(tmp->rdev->bdev,b));
4395         }
4396 }
4397
4398 static int raid5_spare_active(mddev_t *mddev)
4399 {
4400         int i;
4401         raid5_conf_t *conf = mddev->private;
4402         struct disk_info *tmp;
4403
4404         for (i = 0; i < conf->raid_disks; i++) {
4405                 tmp = conf->disks + i;
4406                 if (tmp->rdev
4407                     && !test_bit(Faulty, &tmp->rdev->flags)
4408                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4409                         unsigned long flags;
4410                         spin_lock_irqsave(&conf->device_lock, flags);
4411                         mddev->degraded--;
4412                         spin_unlock_irqrestore(&conf->device_lock, flags);
4413                 }
4414         }
4415         print_raid5_conf(conf);
4416         return 0;
4417 }
4418
4419 static int raid5_remove_disk(mddev_t *mddev, int number)
4420 {
4421         raid5_conf_t *conf = mddev->private;
4422         int err = 0;
4423         mdk_rdev_t *rdev;
4424         struct disk_info *p = conf->disks + number;
4425
4426         print_raid5_conf(conf);
4427         rdev = p->rdev;
4428         if (rdev) {
4429                 if (test_bit(In_sync, &rdev->flags) ||
4430                     atomic_read(&rdev->nr_pending)) {
4431                         err = -EBUSY;
4432                         goto abort;
4433                 }
4434                 p->rdev = NULL;
4435                 synchronize_rcu();
4436                 if (atomic_read(&rdev->nr_pending)) {
4437                         /* lost the race, try later */
4438                         err = -EBUSY;
4439                         p->rdev = rdev;
4440                 }
4441         }
4442 abort:
4443
4444         print_raid5_conf(conf);
4445         return err;
4446 }
4447
4448 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4449 {
4450         raid5_conf_t *conf = mddev->private;
4451         int found = 0;
4452         int disk;
4453         struct disk_info *p;
4454
4455         if (mddev->degraded > conf->max_degraded)
4456                 /* no point adding a device */
4457                 return 0;
4458
4459         /*
4460          * find the disk ... but prefer rdev->saved_raid_disk
4461          * if possible.
4462          */
4463         if (rdev->saved_raid_disk >= 0 &&
4464             conf->disks[rdev->saved_raid_disk].rdev == NULL)
4465                 disk = rdev->saved_raid_disk;
4466         else
4467                 disk = 0;
4468         for ( ; disk < conf->raid_disks; disk++)
4469                 if ((p=conf->disks + disk)->rdev == NULL) {
4470                         clear_bit(In_sync, &rdev->flags);
4471                         rdev->raid_disk = disk;
4472                         found = 1;
4473                         if (rdev->saved_raid_disk != disk)
4474                                 conf->fullsync = 1;
4475                         rcu_assign_pointer(p->rdev, rdev);
4476                         break;
4477                 }
4478         print_raid5_conf(conf);
4479         return found;
4480 }
4481
4482 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4483 {
4484         /* no resync is happening, and there is enough space
4485          * on all devices, so we can resize.
4486          * We need to make sure resync covers any new space.
4487          * If the array is shrinking we should possibly wait until
4488          * any io in the removed space completes, but it hardly seems
4489          * worth it.
4490          */
4491         raid5_conf_t *conf = mddev_to_conf(mddev);
4492
4493         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4494         mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4495         set_capacity(mddev->gendisk, mddev->array_size << 1);
4496         mddev->changed = 1;
4497         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
4498                 mddev->recovery_cp = mddev->size << 1;
4499                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4500         }
4501         mddev->size = sectors /2;
4502         mddev->resync_max_sectors = sectors;
4503         return 0;
4504 }
4505
4506 #ifdef CONFIG_MD_RAID5_RESHAPE
4507 static int raid5_check_reshape(mddev_t *mddev)
4508 {
4509         raid5_conf_t *conf = mddev_to_conf(mddev);
4510         int err;
4511
4512         if (mddev->delta_disks < 0 ||
4513             mddev->new_level != mddev->level)
4514                 return -EINVAL; /* Cannot shrink array or change level yet */
4515         if (mddev->delta_disks == 0)
4516                 return 0; /* nothing to do */
4517
4518         /* Can only proceed if there are plenty of stripe_heads.
4519          * We need a minimum of one full stripe,, and for sensible progress
4520          * it is best to have about 4 times that.
4521          * If we require 4 times, then the default 256 4K stripe_heads will
4522          * allow for chunk sizes up to 256K, which is probably OK.
4523          * If the chunk size is greater, user-space should request more
4524          * stripe_heads first.
4525          */
4526         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4527             (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4528                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
4529                        (mddev->chunk_size / STRIPE_SIZE)*4);
4530                 return -ENOSPC;
4531         }
4532
4533         err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4534         if (err)
4535                 return err;
4536
4537         if (mddev->degraded > conf->max_degraded)
4538                 return -EINVAL;
4539         /* looks like we might be able to manage this */
4540         return 0;
4541 }
4542
4543 static int raid5_start_reshape(mddev_t *mddev)
4544 {
4545         raid5_conf_t *conf = mddev_to_conf(mddev);
4546         mdk_rdev_t *rdev;
4547         struct list_head *rtmp;
4548         int spares = 0;
4549         int added_devices = 0;
4550         unsigned long flags;
4551
4552         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4553                 return -EBUSY;
4554
4555         rdev_for_each(rdev, rtmp, mddev)
4556                 if (rdev->raid_disk < 0 &&
4557                     !test_bit(Faulty, &rdev->flags))
4558                         spares++;
4559
4560         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4561                 /* Not enough devices even to make a degraded array
4562                  * of that size
4563                  */
4564                 return -EINVAL;
4565
4566         atomic_set(&conf->reshape_stripes, 0);
4567         spin_lock_irq(&conf->device_lock);
4568         conf->previous_raid_disks = conf->raid_disks;
4569         conf->raid_disks += mddev->delta_disks;
4570         conf->expand_progress = 0;
4571         conf->expand_lo = 0;
4572         spin_unlock_irq(&conf->device_lock);
4573
4574         /* Add some new drives, as many as will fit.
4575          * We know there are enough to make the newly sized array work.
4576          */
4577         rdev_for_each(rdev, rtmp, mddev)
4578                 if (rdev->raid_disk < 0 &&
4579                     !test_bit(Faulty, &rdev->flags)) {
4580                         if (raid5_add_disk(mddev, rdev)) {
4581                                 char nm[20];
4582                                 set_bit(In_sync, &rdev->flags);
4583                                 added_devices++;
4584                                 rdev->recovery_offset = 0;
4585                                 sprintf(nm, "rd%d", rdev->raid_disk);
4586                                 if (sysfs_create_link(&mddev->kobj,
4587                                                       &rdev->kobj, nm))
4588                                         printk(KERN_WARNING
4589                                                "raid5: failed to create "
4590                                                " link %s for %s\n",
4591                                                nm, mdname(mddev));
4592                         } else
4593                                 break;
4594                 }
4595
4596         spin_lock_irqsave(&conf->device_lock, flags);
4597         mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4598         spin_unlock_irqrestore(&conf->device_lock, flags);
4599         mddev->raid_disks = conf->raid_disks;
4600         mddev->reshape_position = 0;
4601         set_bit(MD_CHANGE_DEVS, &mddev->flags);
4602
4603         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4604         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4605         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4606         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4607         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4608                                                 "%s_reshape");
4609         if (!mddev->sync_thread) {
4610                 mddev->recovery = 0;
4611                 spin_lock_irq(&conf->device_lock);
4612                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4613                 conf->expand_progress = MaxSector;
4614                 spin_unlock_irq(&conf->device_lock);
4615                 return -EAGAIN;
4616         }
4617         md_wakeup_thread(mddev->sync_thread);
4618         md_new_event(mddev);
4619         return 0;
4620 }
4621 #endif
4622
4623 static void end_reshape(raid5_conf_t *conf)
4624 {
4625         struct block_device *bdev;
4626
4627         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4628                 conf->mddev->array_size = conf->mddev->size *
4629                         (conf->raid_disks - conf->max_degraded);
4630                 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4631                 conf->mddev->changed = 1;
4632
4633                 bdev = bdget_disk(conf->mddev->gendisk, 0);
4634                 if (bdev) {
4635                         mutex_lock(&bdev->bd_inode->i_mutex);
4636                         i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4637                         mutex_unlock(&bdev->bd_inode->i_mutex);
4638                         bdput(bdev);
4639                 }
4640                 spin_lock_irq(&conf->device_lock);
4641                 conf->expand_progress = MaxSector;
4642                 spin_unlock_irq(&conf->device_lock);
4643                 conf->mddev->reshape_position = MaxSector;
4644
4645                 /* read-ahead size must cover two whole stripes, which is
4646                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4647                  */
4648                 {
4649                         int data_disks = conf->previous_raid_disks - conf->max_degraded;
4650                         int stripe = data_disks *
4651                                 (conf->mddev->chunk_size / PAGE_SIZE);
4652                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4653                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4654                 }
4655         }
4656 }
4657
4658 static void raid5_quiesce(mddev_t *mddev, int state)
4659 {
4660         raid5_conf_t *conf = mddev_to_conf(mddev);
4661
4662         switch(state) {
4663         case 2: /* resume for a suspend */
4664                 wake_up(&conf->wait_for_overlap);
4665                 break;
4666
4667         case 1: /* stop all writes */
4668                 spin_lock_irq(&conf->device_lock);
4669                 conf->quiesce = 1;
4670                 wait_event_lock_irq(conf->wait_for_stripe,
4671                                     atomic_read(&conf->active_stripes) == 0 &&
4672                                     atomic_read(&conf->active_aligned_reads) == 0,
4673                                     conf->device_lock, /* nothing */);
4674                 spin_unlock_irq(&conf->device_lock);
4675                 break;
4676
4677         case 0: /* re-enable writes */
4678                 spin_lock_irq(&conf->device_lock);
4679                 conf->quiesce = 0;
4680                 wake_up(&conf->wait_for_stripe);
4681                 wake_up(&conf->wait_for_overlap);
4682                 spin_unlock_irq(&conf->device_lock);
4683                 break;
4684         }
4685 }
4686
4687 static struct mdk_personality raid6_personality =
4688 {
4689         .name           = "raid6",
4690         .level          = 6,
4691         .owner          = THIS_MODULE,
4692         .make_request   = make_request,
4693         .run            = run,
4694         .stop           = stop,
4695         .status         = status,
4696         .error_handler  = error,
4697         .hot_add_disk   = raid5_add_disk,
4698         .hot_remove_disk= raid5_remove_disk,
4699         .spare_active   = raid5_spare_active,
4700         .sync_request   = sync_request,
4701         .resize         = raid5_resize,
4702 #ifdef CONFIG_MD_RAID5_RESHAPE
4703         .check_reshape  = raid5_check_reshape,
4704         .start_reshape  = raid5_start_reshape,
4705 #endif
4706         .quiesce        = raid5_quiesce,
4707 };
4708 static struct mdk_personality raid5_personality =
4709 {
4710         .name           = "raid5",
4711         .level          = 5,
4712         .owner          = THIS_MODULE,
4713         .make_request   = make_request,
4714         .run            = run,
4715         .stop           = stop,
4716         .status         = status,
4717         .error_handler  = error,
4718         .hot_add_disk   = raid5_add_disk,
4719         .hot_remove_disk= raid5_remove_disk,
4720         .spare_active   = raid5_spare_active,
4721         .sync_request   = sync_request,
4722         .resize         = raid5_resize,
4723 #ifdef CONFIG_MD_RAID5_RESHAPE
4724         .check_reshape  = raid5_check_reshape,
4725         .start_reshape  = raid5_start_reshape,
4726 #endif
4727         .quiesce        = raid5_quiesce,
4728 };
4729
4730 static struct mdk_personality raid4_personality =
4731 {
4732         .name           = "raid4",
4733         .level          = 4,
4734         .owner          = THIS_MODULE,
4735         .make_request   = make_request,
4736         .run            = run,
4737         .stop           = stop,
4738         .status         = status,
4739         .error_handler  = error,
4740         .hot_add_disk   = raid5_add_disk,
4741         .hot_remove_disk= raid5_remove_disk,
4742         .spare_active   = raid5_spare_active,
4743         .sync_request   = sync_request,
4744         .resize         = raid5_resize,
4745 #ifdef CONFIG_MD_RAID5_RESHAPE
4746         .check_reshape  = raid5_check_reshape,
4747         .start_reshape  = raid5_start_reshape,
4748 #endif
4749         .quiesce        = raid5_quiesce,
4750 };
4751
4752 static int __init raid5_init(void)
4753 {
4754         int e;
4755
4756         e = raid6_select_algo();
4757         if ( e )
4758                 return e;
4759         register_md_personality(&raid6_personality);
4760         register_md_personality(&raid5_personality);
4761         register_md_personality(&raid4_personality);
4762         return 0;
4763 }
4764
4765 static void raid5_exit(void)
4766 {
4767         unregister_md_personality(&raid6_personality);
4768         unregister_md_personality(&raid5_personality);
4769         unregister_md_personality(&raid4_personality);
4770 }
4771
4772 module_init(raid5_init);
4773 module_exit(raid5_exit);
4774 MODULE_LICENSE("GPL");
4775 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4776 MODULE_ALIAS("md-raid5");
4777 MODULE_ALIAS("md-raid4");
4778 MODULE_ALIAS("md-level-5");
4779 MODULE_ALIAS("md-level-4");
4780 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4781 MODULE_ALIAS("md-raid6");
4782 MODULE_ALIAS("md-level-6");
4783
4784 /* This used to be two separate modules, they were: */
4785 MODULE_ALIAS("raid5");
4786 MODULE_ALIAS("raid6");