pasemi_mac: flags as passed to spin_*_irqsave() should be unsigned long
[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                 int bytes = bi->bi_size;
112
113                 return_bi = bi->bi_next;
114                 bi->bi_next = NULL;
115                 bi->bi_size = 0;
116                 bi->bi_end_io(bi, bytes,
117                               test_bit(BIO_UPTODATE, &bi->bi_flags)
118                                 ? 0 : -EIO);
119                 bi = return_bi;
120         }
121 }
122
123 static void print_raid5_conf (raid5_conf_t *conf);
124
125 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
126 {
127         if (atomic_dec_and_test(&sh->count)) {
128                 BUG_ON(!list_empty(&sh->lru));
129                 BUG_ON(atomic_read(&conf->active_stripes)==0);
130                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
131                         if (test_bit(STRIPE_DELAYED, &sh->state)) {
132                                 list_add_tail(&sh->lru, &conf->delayed_list);
133                                 blk_plug_device(conf->mddev->queue);
134                         } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
135                                    sh->bm_seq - conf->seq_write > 0) {
136                                 list_add_tail(&sh->lru, &conf->bitmap_list);
137                                 blk_plug_device(conf->mddev->queue);
138                         } else {
139                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
140                                 list_add_tail(&sh->lru, &conf->handle_list);
141                         }
142                         md_wakeup_thread(conf->mddev->thread);
143                 } else {
144                         BUG_ON(sh->ops.pending);
145                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
146                                 atomic_dec(&conf->preread_active_stripes);
147                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
148                                         md_wakeup_thread(conf->mddev->thread);
149                         }
150                         atomic_dec(&conf->active_stripes);
151                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
152                                 list_add_tail(&sh->lru, &conf->inactive_list);
153                                 wake_up(&conf->wait_for_stripe);
154                                 if (conf->retry_read_aligned)
155                                         md_wakeup_thread(conf->mddev->thread);
156                         }
157                 }
158         }
159 }
160 static void release_stripe(struct stripe_head *sh)
161 {
162         raid5_conf_t *conf = sh->raid_conf;
163         unsigned long flags;
164
165         spin_lock_irqsave(&conf->device_lock, flags);
166         __release_stripe(conf, sh);
167         spin_unlock_irqrestore(&conf->device_lock, flags);
168 }
169
170 static inline void remove_hash(struct stripe_head *sh)
171 {
172         pr_debug("remove_hash(), stripe %llu\n",
173                 (unsigned long long)sh->sector);
174
175         hlist_del_init(&sh->hash);
176 }
177
178 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
179 {
180         struct hlist_head *hp = stripe_hash(conf, sh->sector);
181
182         pr_debug("insert_hash(), stripe %llu\n",
183                 (unsigned long long)sh->sector);
184
185         CHECK_DEVLOCK();
186         hlist_add_head(&sh->hash, hp);
187 }
188
189
190 /* find an idle stripe, make sure it is unhashed, and return it. */
191 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
192 {
193         struct stripe_head *sh = NULL;
194         struct list_head *first;
195
196         CHECK_DEVLOCK();
197         if (list_empty(&conf->inactive_list))
198                 goto out;
199         first = conf->inactive_list.next;
200         sh = list_entry(first, struct stripe_head, lru);
201         list_del_init(first);
202         remove_hash(sh);
203         atomic_inc(&conf->active_stripes);
204 out:
205         return sh;
206 }
207
208 static void shrink_buffers(struct stripe_head *sh, int num)
209 {
210         struct page *p;
211         int i;
212
213         for (i=0; i<num ; i++) {
214                 p = sh->dev[i].page;
215                 if (!p)
216                         continue;
217                 sh->dev[i].page = NULL;
218                 put_page(p);
219         }
220 }
221
222 static int grow_buffers(struct stripe_head *sh, int num)
223 {
224         int i;
225
226         for (i=0; i<num; i++) {
227                 struct page *page;
228
229                 if (!(page = alloc_page(GFP_KERNEL))) {
230                         return 1;
231                 }
232                 sh->dev[i].page = page;
233         }
234         return 0;
235 }
236
237 static void raid5_build_block (struct stripe_head *sh, int i);
238
239 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks)
240 {
241         raid5_conf_t *conf = sh->raid_conf;
242         int i;
243
244         BUG_ON(atomic_read(&sh->count) != 0);
245         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
246         BUG_ON(sh->ops.pending || sh->ops.ack || sh->ops.complete);
247
248         CHECK_DEVLOCK();
249         pr_debug("init_stripe called, stripe %llu\n",
250                 (unsigned long long)sh->sector);
251
252         remove_hash(sh);
253
254         sh->sector = sector;
255         sh->pd_idx = pd_idx;
256         sh->state = 0;
257
258         sh->disks = disks;
259
260         for (i = sh->disks; i--; ) {
261                 struct r5dev *dev = &sh->dev[i];
262
263                 if (dev->toread || dev->read || dev->towrite || dev->written ||
264                     test_bit(R5_LOCKED, &dev->flags)) {
265                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
266                                (unsigned long long)sh->sector, i, dev->toread,
267                                dev->read, dev->towrite, dev->written,
268                                test_bit(R5_LOCKED, &dev->flags));
269                         BUG();
270                 }
271                 dev->flags = 0;
272                 raid5_build_block(sh, i);
273         }
274         insert_hash(conf, sh);
275 }
276
277 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
278 {
279         struct stripe_head *sh;
280         struct hlist_node *hn;
281
282         CHECK_DEVLOCK();
283         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
284         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
285                 if (sh->sector == sector && sh->disks == disks)
286                         return sh;
287         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
288         return NULL;
289 }
290
291 static void unplug_slaves(mddev_t *mddev);
292 static void raid5_unplug_device(struct request_queue *q);
293
294 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks,
295                                              int pd_idx, int noblock)
296 {
297         struct stripe_head *sh;
298
299         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
300
301         spin_lock_irq(&conf->device_lock);
302
303         do {
304                 wait_event_lock_irq(conf->wait_for_stripe,
305                                     conf->quiesce == 0,
306                                     conf->device_lock, /* nothing */);
307                 sh = __find_stripe(conf, sector, disks);
308                 if (!sh) {
309                         if (!conf->inactive_blocked)
310                                 sh = get_free_stripe(conf);
311                         if (noblock && sh == NULL)
312                                 break;
313                         if (!sh) {
314                                 conf->inactive_blocked = 1;
315                                 wait_event_lock_irq(conf->wait_for_stripe,
316                                                     !list_empty(&conf->inactive_list) &&
317                                                     (atomic_read(&conf->active_stripes)
318                                                      < (conf->max_nr_stripes *3/4)
319                                                      || !conf->inactive_blocked),
320                                                     conf->device_lock,
321                                                     raid5_unplug_device(conf->mddev->queue)
322                                         );
323                                 conf->inactive_blocked = 0;
324                         } else
325                                 init_stripe(sh, sector, pd_idx, disks);
326                 } else {
327                         if (atomic_read(&sh->count)) {
328                           BUG_ON(!list_empty(&sh->lru));
329                         } else {
330                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
331                                         atomic_inc(&conf->active_stripes);
332                                 if (list_empty(&sh->lru) &&
333                                     !test_bit(STRIPE_EXPANDING, &sh->state))
334                                         BUG();
335                                 list_del_init(&sh->lru);
336                         }
337                 }
338         } while (sh == NULL);
339
340         if (sh)
341                 atomic_inc(&sh->count);
342
343         spin_unlock_irq(&conf->device_lock);
344         return sh;
345 }
346
347 /* test_and_ack_op() ensures that we only dequeue an operation once */
348 #define test_and_ack_op(op, pend) \
349 do {                                                    \
350         if (test_bit(op, &sh->ops.pending) &&           \
351                 !test_bit(op, &sh->ops.complete)) {     \
352                 if (test_and_set_bit(op, &sh->ops.ack)) \
353                         clear_bit(op, &pend);           \
354                 else                                    \
355                         ack++;                          \
356         } else                                          \
357                 clear_bit(op, &pend);                   \
358 } while (0)
359
360 /* find new work to run, do not resubmit work that is already
361  * in flight
362  */
363 static unsigned long get_stripe_work(struct stripe_head *sh)
364 {
365         unsigned long pending;
366         int ack = 0;
367
368         pending = sh->ops.pending;
369
370         test_and_ack_op(STRIPE_OP_BIOFILL, pending);
371         test_and_ack_op(STRIPE_OP_COMPUTE_BLK, pending);
372         test_and_ack_op(STRIPE_OP_PREXOR, pending);
373         test_and_ack_op(STRIPE_OP_BIODRAIN, pending);
374         test_and_ack_op(STRIPE_OP_POSTXOR, pending);
375         test_and_ack_op(STRIPE_OP_CHECK, pending);
376         if (test_and_clear_bit(STRIPE_OP_IO, &sh->ops.pending))
377                 ack++;
378
379         sh->ops.count -= ack;
380         BUG_ON(sh->ops.count < 0);
381
382         return pending;
383 }
384
385 static int
386 raid5_end_read_request(struct bio *bi, unsigned int bytes_done, int error);
387 static int
388 raid5_end_write_request (struct bio *bi, unsigned int bytes_done, int error);
389
390 static void ops_run_io(struct stripe_head *sh)
391 {
392         raid5_conf_t *conf = sh->raid_conf;
393         int i, disks = sh->disks;
394
395         might_sleep();
396
397         for (i = disks; i--; ) {
398                 int rw;
399                 struct bio *bi;
400                 mdk_rdev_t *rdev;
401                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
402                         rw = WRITE;
403                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
404                         rw = READ;
405                 else
406                         continue;
407
408                 bi = &sh->dev[i].req;
409
410                 bi->bi_rw = rw;
411                 if (rw == WRITE)
412                         bi->bi_end_io = raid5_end_write_request;
413                 else
414                         bi->bi_end_io = raid5_end_read_request;
415
416                 rcu_read_lock();
417                 rdev = rcu_dereference(conf->disks[i].rdev);
418                 if (rdev && test_bit(Faulty, &rdev->flags))
419                         rdev = NULL;
420                 if (rdev)
421                         atomic_inc(&rdev->nr_pending);
422                 rcu_read_unlock();
423
424                 if (rdev) {
425                         if (test_bit(STRIPE_SYNCING, &sh->state) ||
426                                 test_bit(STRIPE_EXPAND_SOURCE, &sh->state) ||
427                                 test_bit(STRIPE_EXPAND_READY, &sh->state))
428                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
429
430                         bi->bi_bdev = rdev->bdev;
431                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
432                                 __FUNCTION__, (unsigned long long)sh->sector,
433                                 bi->bi_rw, i);
434                         atomic_inc(&sh->count);
435                         bi->bi_sector = sh->sector + rdev->data_offset;
436                         bi->bi_flags = 1 << BIO_UPTODATE;
437                         bi->bi_vcnt = 1;
438                         bi->bi_max_vecs = 1;
439                         bi->bi_idx = 0;
440                         bi->bi_io_vec = &sh->dev[i].vec;
441                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
442                         bi->bi_io_vec[0].bv_offset = 0;
443                         bi->bi_size = STRIPE_SIZE;
444                         bi->bi_next = NULL;
445                         if (rw == WRITE &&
446                             test_bit(R5_ReWrite, &sh->dev[i].flags))
447                                 atomic_add(STRIPE_SECTORS,
448                                         &rdev->corrected_errors);
449                         generic_make_request(bi);
450                 } else {
451                         if (rw == WRITE)
452                                 set_bit(STRIPE_DEGRADED, &sh->state);
453                         pr_debug("skip op %ld on disc %d for sector %llu\n",
454                                 bi->bi_rw, i, (unsigned long long)sh->sector);
455                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
456                         set_bit(STRIPE_HANDLE, &sh->state);
457                 }
458         }
459 }
460
461 static struct dma_async_tx_descriptor *
462 async_copy_data(int frombio, struct bio *bio, struct page *page,
463         sector_t sector, struct dma_async_tx_descriptor *tx)
464 {
465         struct bio_vec *bvl;
466         struct page *bio_page;
467         int i;
468         int page_offset;
469
470         if (bio->bi_sector >= sector)
471                 page_offset = (signed)(bio->bi_sector - sector) * 512;
472         else
473                 page_offset = (signed)(sector - bio->bi_sector) * -512;
474         bio_for_each_segment(bvl, bio, i) {
475                 int len = bio_iovec_idx(bio, i)->bv_len;
476                 int clen;
477                 int b_offset = 0;
478
479                 if (page_offset < 0) {
480                         b_offset = -page_offset;
481                         page_offset += b_offset;
482                         len -= b_offset;
483                 }
484
485                 if (len > 0 && page_offset + len > STRIPE_SIZE)
486                         clen = STRIPE_SIZE - page_offset;
487                 else
488                         clen = len;
489
490                 if (clen > 0) {
491                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
492                         bio_page = bio_iovec_idx(bio, i)->bv_page;
493                         if (frombio)
494                                 tx = async_memcpy(page, bio_page, page_offset,
495                                         b_offset, clen,
496                                         ASYNC_TX_DEP_ACK,
497                                         tx, NULL, NULL);
498                         else
499                                 tx = async_memcpy(bio_page, page, b_offset,
500                                         page_offset, clen,
501                                         ASYNC_TX_DEP_ACK,
502                                         tx, NULL, NULL);
503                 }
504                 if (clen < len) /* hit end of page */
505                         break;
506                 page_offset +=  len;
507         }
508
509         return tx;
510 }
511
512 static void ops_complete_biofill(void *stripe_head_ref)
513 {
514         struct stripe_head *sh = stripe_head_ref;
515         struct bio *return_bi = NULL;
516         raid5_conf_t *conf = sh->raid_conf;
517         int i;
518
519         pr_debug("%s: stripe %llu\n", __FUNCTION__,
520                 (unsigned long long)sh->sector);
521
522         /* clear completed biofills */
523         for (i = sh->disks; i--; ) {
524                 struct r5dev *dev = &sh->dev[i];
525
526                 /* acknowledge completion of a biofill operation */
527                 /* and check if we need to reply to a read request,
528                  * new R5_Wantfill requests are held off until
529                  * !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending)
530                  */
531                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
532                         struct bio *rbi, *rbi2;
533
534                         /* The access to dev->read is outside of the
535                          * spin_lock_irq(&conf->device_lock), but is protected
536                          * by the STRIPE_OP_BIOFILL pending bit
537                          */
538                         BUG_ON(!dev->read);
539                         rbi = dev->read;
540                         dev->read = NULL;
541                         while (rbi && rbi->bi_sector <
542                                 dev->sector + STRIPE_SECTORS) {
543                                 rbi2 = r5_next_bio(rbi, dev->sector);
544                                 spin_lock_irq(&conf->device_lock);
545                                 if (--rbi->bi_phys_segments == 0) {
546                                         rbi->bi_next = return_bi;
547                                         return_bi = rbi;
548                                 }
549                                 spin_unlock_irq(&conf->device_lock);
550                                 rbi = rbi2;
551                         }
552                 }
553         }
554         clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack);
555         clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending);
556
557         return_io(return_bi);
558
559         set_bit(STRIPE_HANDLE, &sh->state);
560         release_stripe(sh);
561 }
562
563 static void ops_run_biofill(struct stripe_head *sh)
564 {
565         struct dma_async_tx_descriptor *tx = NULL;
566         raid5_conf_t *conf = sh->raid_conf;
567         int i;
568
569         pr_debug("%s: stripe %llu\n", __FUNCTION__,
570                 (unsigned long long)sh->sector);
571
572         for (i = sh->disks; i--; ) {
573                 struct r5dev *dev = &sh->dev[i];
574                 if (test_bit(R5_Wantfill, &dev->flags)) {
575                         struct bio *rbi;
576                         spin_lock_irq(&conf->device_lock);
577                         dev->read = rbi = dev->toread;
578                         dev->toread = NULL;
579                         spin_unlock_irq(&conf->device_lock);
580                         while (rbi && rbi->bi_sector <
581                                 dev->sector + STRIPE_SECTORS) {
582                                 tx = async_copy_data(0, rbi, dev->page,
583                                         dev->sector, tx);
584                                 rbi = r5_next_bio(rbi, dev->sector);
585                         }
586                 }
587         }
588
589         atomic_inc(&sh->count);
590         async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
591                 ops_complete_biofill, sh);
592 }
593
594 static void ops_complete_compute5(void *stripe_head_ref)
595 {
596         struct stripe_head *sh = stripe_head_ref;
597         int target = sh->ops.target;
598         struct r5dev *tgt = &sh->dev[target];
599
600         pr_debug("%s: stripe %llu\n", __FUNCTION__,
601                 (unsigned long long)sh->sector);
602
603         set_bit(R5_UPTODATE, &tgt->flags);
604         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
605         clear_bit(R5_Wantcompute, &tgt->flags);
606         set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
607         set_bit(STRIPE_HANDLE, &sh->state);
608         release_stripe(sh);
609 }
610
611 static struct dma_async_tx_descriptor *
612 ops_run_compute5(struct stripe_head *sh, unsigned long pending)
613 {
614         /* kernel stack size limits the total number of disks */
615         int disks = sh->disks;
616         struct page *xor_srcs[disks];
617         int target = sh->ops.target;
618         struct r5dev *tgt = &sh->dev[target];
619         struct page *xor_dest = tgt->page;
620         int count = 0;
621         struct dma_async_tx_descriptor *tx;
622         int i;
623
624         pr_debug("%s: stripe %llu block: %d\n",
625                 __FUNCTION__, (unsigned long long)sh->sector, target);
626         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
627
628         for (i = disks; i--; )
629                 if (i != target)
630                         xor_srcs[count++] = sh->dev[i].page;
631
632         atomic_inc(&sh->count);
633
634         if (unlikely(count == 1))
635                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
636                         0, NULL, ops_complete_compute5, sh);
637         else
638                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
639                         ASYNC_TX_XOR_ZERO_DST, NULL,
640                         ops_complete_compute5, sh);
641
642         /* ack now if postxor is not set to be run */
643         if (tx && !test_bit(STRIPE_OP_POSTXOR, &pending))
644                 async_tx_ack(tx);
645
646         return tx;
647 }
648
649 static void ops_complete_prexor(void *stripe_head_ref)
650 {
651         struct stripe_head *sh = stripe_head_ref;
652
653         pr_debug("%s: stripe %llu\n", __FUNCTION__,
654                 (unsigned long long)sh->sector);
655
656         set_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
657 }
658
659 static struct dma_async_tx_descriptor *
660 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
661 {
662         /* kernel stack size limits the total number of disks */
663         int disks = sh->disks;
664         struct page *xor_srcs[disks];
665         int count = 0, pd_idx = sh->pd_idx, i;
666
667         /* existing parity data subtracted */
668         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
669
670         pr_debug("%s: stripe %llu\n", __FUNCTION__,
671                 (unsigned long long)sh->sector);
672
673         for (i = disks; i--; ) {
674                 struct r5dev *dev = &sh->dev[i];
675                 /* Only process blocks that are known to be uptodate */
676                 if (dev->towrite && test_bit(R5_Wantprexor, &dev->flags))
677                         xor_srcs[count++] = dev->page;
678         }
679
680         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
681                 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
682                 ops_complete_prexor, sh);
683
684         return tx;
685 }
686
687 static struct dma_async_tx_descriptor *
688 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
689 {
690         int disks = sh->disks;
691         int pd_idx = sh->pd_idx, i;
692
693         /* check if prexor is active which means only process blocks
694          * that are part of a read-modify-write (Wantprexor)
695          */
696         int prexor = test_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
697
698         pr_debug("%s: stripe %llu\n", __FUNCTION__,
699                 (unsigned long long)sh->sector);
700
701         for (i = disks; i--; ) {
702                 struct r5dev *dev = &sh->dev[i];
703                 struct bio *chosen;
704                 int towrite;
705
706                 towrite = 0;
707                 if (prexor) { /* rmw */
708                         if (dev->towrite &&
709                             test_bit(R5_Wantprexor, &dev->flags))
710                                 towrite = 1;
711                 } else { /* rcw */
712                         if (i != pd_idx && dev->towrite &&
713                                 test_bit(R5_LOCKED, &dev->flags))
714                                 towrite = 1;
715                 }
716
717                 if (towrite) {
718                         struct bio *wbi;
719
720                         spin_lock(&sh->lock);
721                         chosen = dev->towrite;
722                         dev->towrite = NULL;
723                         BUG_ON(dev->written);
724                         wbi = dev->written = chosen;
725                         spin_unlock(&sh->lock);
726
727                         while (wbi && wbi->bi_sector <
728                                 dev->sector + STRIPE_SECTORS) {
729                                 tx = async_copy_data(1, wbi, dev->page,
730                                         dev->sector, tx);
731                                 wbi = r5_next_bio(wbi, dev->sector);
732                         }
733                 }
734         }
735
736         return tx;
737 }
738
739 static void ops_complete_postxor(void *stripe_head_ref)
740 {
741         struct stripe_head *sh = stripe_head_ref;
742
743         pr_debug("%s: stripe %llu\n", __FUNCTION__,
744                 (unsigned long long)sh->sector);
745
746         set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
747         set_bit(STRIPE_HANDLE, &sh->state);
748         release_stripe(sh);
749 }
750
751 static void ops_complete_write(void *stripe_head_ref)
752 {
753         struct stripe_head *sh = stripe_head_ref;
754         int disks = sh->disks, i, pd_idx = sh->pd_idx;
755
756         pr_debug("%s: stripe %llu\n", __FUNCTION__,
757                 (unsigned long long)sh->sector);
758
759         for (i = disks; i--; ) {
760                 struct r5dev *dev = &sh->dev[i];
761                 if (dev->written || i == pd_idx)
762                         set_bit(R5_UPTODATE, &dev->flags);
763         }
764
765         set_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
766         set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
767
768         set_bit(STRIPE_HANDLE, &sh->state);
769         release_stripe(sh);
770 }
771
772 static void
773 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
774 {
775         /* kernel stack size limits the total number of disks */
776         int disks = sh->disks;
777         struct page *xor_srcs[disks];
778
779         int count = 0, pd_idx = sh->pd_idx, i;
780         struct page *xor_dest;
781         int prexor = test_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
782         unsigned long flags;
783         dma_async_tx_callback callback;
784
785         pr_debug("%s: stripe %llu\n", __FUNCTION__,
786                 (unsigned long long)sh->sector);
787
788         /* check if prexor is active which means only process blocks
789          * that are part of a read-modify-write (written)
790          */
791         if (prexor) {
792                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
793                 for (i = disks; i--; ) {
794                         struct r5dev *dev = &sh->dev[i];
795                         if (dev->written)
796                                 xor_srcs[count++] = dev->page;
797                 }
798         } else {
799                 xor_dest = sh->dev[pd_idx].page;
800                 for (i = disks; i--; ) {
801                         struct r5dev *dev = &sh->dev[i];
802                         if (i != pd_idx)
803                                 xor_srcs[count++] = dev->page;
804                 }
805         }
806
807         /* check whether this postxor is part of a write */
808         callback = test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending) ?
809                 ops_complete_write : ops_complete_postxor;
810
811         /* 1/ if we prexor'd then the dest is reused as a source
812          * 2/ if we did not prexor then we are redoing the parity
813          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
814          * for the synchronous xor case
815          */
816         flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
817                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
818
819         atomic_inc(&sh->count);
820
821         if (unlikely(count == 1)) {
822                 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
823                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
824                         flags, tx, callback, sh);
825         } else
826                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
827                         flags, tx, callback, sh);
828 }
829
830 static void ops_complete_check(void *stripe_head_ref)
831 {
832         struct stripe_head *sh = stripe_head_ref;
833         int pd_idx = sh->pd_idx;
834
835         pr_debug("%s: stripe %llu\n", __FUNCTION__,
836                 (unsigned long long)sh->sector);
837
838         if (test_and_clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending) &&
839                 sh->ops.zero_sum_result == 0)
840                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
841
842         set_bit(STRIPE_OP_CHECK, &sh->ops.complete);
843         set_bit(STRIPE_HANDLE, &sh->state);
844         release_stripe(sh);
845 }
846
847 static void ops_run_check(struct stripe_head *sh)
848 {
849         /* kernel stack size limits the total number of disks */
850         int disks = sh->disks;
851         struct page *xor_srcs[disks];
852         struct dma_async_tx_descriptor *tx;
853
854         int count = 0, pd_idx = sh->pd_idx, i;
855         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
856
857         pr_debug("%s: stripe %llu\n", __FUNCTION__,
858                 (unsigned long long)sh->sector);
859
860         for (i = disks; i--; ) {
861                 struct r5dev *dev = &sh->dev[i];
862                 if (i != pd_idx)
863                         xor_srcs[count++] = dev->page;
864         }
865
866         tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
867                 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
868
869         if (tx)
870                 set_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
871         else
872                 clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
873
874         atomic_inc(&sh->count);
875         tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
876                 ops_complete_check, sh);
877 }
878
879 static void raid5_run_ops(struct stripe_head *sh, unsigned long pending)
880 {
881         int overlap_clear = 0, i, disks = sh->disks;
882         struct dma_async_tx_descriptor *tx = NULL;
883
884         if (test_bit(STRIPE_OP_BIOFILL, &pending)) {
885                 ops_run_biofill(sh);
886                 overlap_clear++;
887         }
888
889         if (test_bit(STRIPE_OP_COMPUTE_BLK, &pending))
890                 tx = ops_run_compute5(sh, pending);
891
892         if (test_bit(STRIPE_OP_PREXOR, &pending))
893                 tx = ops_run_prexor(sh, tx);
894
895         if (test_bit(STRIPE_OP_BIODRAIN, &pending)) {
896                 tx = ops_run_biodrain(sh, tx);
897                 overlap_clear++;
898         }
899
900         if (test_bit(STRIPE_OP_POSTXOR, &pending))
901                 ops_run_postxor(sh, tx);
902
903         if (test_bit(STRIPE_OP_CHECK, &pending))
904                 ops_run_check(sh);
905
906         if (test_bit(STRIPE_OP_IO, &pending))
907                 ops_run_io(sh);
908
909         if (overlap_clear)
910                 for (i = disks; i--; ) {
911                         struct r5dev *dev = &sh->dev[i];
912                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
913                                 wake_up(&sh->raid_conf->wait_for_overlap);
914                 }
915 }
916
917 static int grow_one_stripe(raid5_conf_t *conf)
918 {
919         struct stripe_head *sh;
920         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
921         if (!sh)
922                 return 0;
923         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
924         sh->raid_conf = conf;
925         spin_lock_init(&sh->lock);
926
927         if (grow_buffers(sh, conf->raid_disks)) {
928                 shrink_buffers(sh, conf->raid_disks);
929                 kmem_cache_free(conf->slab_cache, sh);
930                 return 0;
931         }
932         sh->disks = conf->raid_disks;
933         /* we just created an active stripe so... */
934         atomic_set(&sh->count, 1);
935         atomic_inc(&conf->active_stripes);
936         INIT_LIST_HEAD(&sh->lru);
937         release_stripe(sh);
938         return 1;
939 }
940
941 static int grow_stripes(raid5_conf_t *conf, int num)
942 {
943         struct kmem_cache *sc;
944         int devs = conf->raid_disks;
945
946         sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
947         sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
948         conf->active_name = 0;
949         sc = kmem_cache_create(conf->cache_name[conf->active_name],
950                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
951                                0, 0, NULL);
952         if (!sc)
953                 return 1;
954         conf->slab_cache = sc;
955         conf->pool_size = devs;
956         while (num--)
957                 if (!grow_one_stripe(conf))
958                         return 1;
959         return 0;
960 }
961
962 #ifdef CONFIG_MD_RAID5_RESHAPE
963 static int resize_stripes(raid5_conf_t *conf, int newsize)
964 {
965         /* Make all the stripes able to hold 'newsize' devices.
966          * New slots in each stripe get 'page' set to a new page.
967          *
968          * This happens in stages:
969          * 1/ create a new kmem_cache and allocate the required number of
970          *    stripe_heads.
971          * 2/ gather all the old stripe_heads and tranfer the pages across
972          *    to the new stripe_heads.  This will have the side effect of
973          *    freezing the array as once all stripe_heads have been collected,
974          *    no IO will be possible.  Old stripe heads are freed once their
975          *    pages have been transferred over, and the old kmem_cache is
976          *    freed when all stripes are done.
977          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
978          *    we simple return a failre status - no need to clean anything up.
979          * 4/ allocate new pages for the new slots in the new stripe_heads.
980          *    If this fails, we don't bother trying the shrink the
981          *    stripe_heads down again, we just leave them as they are.
982          *    As each stripe_head is processed the new one is released into
983          *    active service.
984          *
985          * Once step2 is started, we cannot afford to wait for a write,
986          * so we use GFP_NOIO allocations.
987          */
988         struct stripe_head *osh, *nsh;
989         LIST_HEAD(newstripes);
990         struct disk_info *ndisks;
991         int err = 0;
992         struct kmem_cache *sc;
993         int i;
994
995         if (newsize <= conf->pool_size)
996                 return 0; /* never bother to shrink */
997
998         md_allow_write(conf->mddev);
999
1000         /* Step 1 */
1001         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1002                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1003                                0, 0, NULL);
1004         if (!sc)
1005                 return -ENOMEM;
1006
1007         for (i = conf->max_nr_stripes; i; i--) {
1008                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1009                 if (!nsh)
1010                         break;
1011
1012                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1013
1014                 nsh->raid_conf = conf;
1015                 spin_lock_init(&nsh->lock);
1016
1017                 list_add(&nsh->lru, &newstripes);
1018         }
1019         if (i) {
1020                 /* didn't get enough, give up */
1021                 while (!list_empty(&newstripes)) {
1022                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1023                         list_del(&nsh->lru);
1024                         kmem_cache_free(sc, nsh);
1025                 }
1026                 kmem_cache_destroy(sc);
1027                 return -ENOMEM;
1028         }
1029         /* Step 2 - Must use GFP_NOIO now.
1030          * OK, we have enough stripes, start collecting inactive
1031          * stripes and copying them over
1032          */
1033         list_for_each_entry(nsh, &newstripes, lru) {
1034                 spin_lock_irq(&conf->device_lock);
1035                 wait_event_lock_irq(conf->wait_for_stripe,
1036                                     !list_empty(&conf->inactive_list),
1037                                     conf->device_lock,
1038                                     unplug_slaves(conf->mddev)
1039                         );
1040                 osh = get_free_stripe(conf);
1041                 spin_unlock_irq(&conf->device_lock);
1042                 atomic_set(&nsh->count, 1);
1043                 for(i=0; i<conf->pool_size; i++)
1044                         nsh->dev[i].page = osh->dev[i].page;
1045                 for( ; i<newsize; i++)
1046                         nsh->dev[i].page = NULL;
1047                 kmem_cache_free(conf->slab_cache, osh);
1048         }
1049         kmem_cache_destroy(conf->slab_cache);
1050
1051         /* Step 3.
1052          * At this point, we are holding all the stripes so the array
1053          * is completely stalled, so now is a good time to resize
1054          * conf->disks.
1055          */
1056         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1057         if (ndisks) {
1058                 for (i=0; i<conf->raid_disks; i++)
1059                         ndisks[i] = conf->disks[i];
1060                 kfree(conf->disks);
1061                 conf->disks = ndisks;
1062         } else
1063                 err = -ENOMEM;
1064
1065         /* Step 4, return new stripes to service */
1066         while(!list_empty(&newstripes)) {
1067                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1068                 list_del_init(&nsh->lru);
1069                 for (i=conf->raid_disks; i < newsize; i++)
1070                         if (nsh->dev[i].page == NULL) {
1071                                 struct page *p = alloc_page(GFP_NOIO);
1072                                 nsh->dev[i].page = p;
1073                                 if (!p)
1074                                         err = -ENOMEM;
1075                         }
1076                 release_stripe(nsh);
1077         }
1078         /* critical section pass, GFP_NOIO no longer needed */
1079
1080         conf->slab_cache = sc;
1081         conf->active_name = 1-conf->active_name;
1082         conf->pool_size = newsize;
1083         return err;
1084 }
1085 #endif
1086
1087 static int drop_one_stripe(raid5_conf_t *conf)
1088 {
1089         struct stripe_head *sh;
1090
1091         spin_lock_irq(&conf->device_lock);
1092         sh = get_free_stripe(conf);
1093         spin_unlock_irq(&conf->device_lock);
1094         if (!sh)
1095                 return 0;
1096         BUG_ON(atomic_read(&sh->count));
1097         shrink_buffers(sh, conf->pool_size);
1098         kmem_cache_free(conf->slab_cache, sh);
1099         atomic_dec(&conf->active_stripes);
1100         return 1;
1101 }
1102
1103 static void shrink_stripes(raid5_conf_t *conf)
1104 {
1105         while (drop_one_stripe(conf))
1106                 ;
1107
1108         if (conf->slab_cache)
1109                 kmem_cache_destroy(conf->slab_cache);
1110         conf->slab_cache = NULL;
1111 }
1112
1113 static int raid5_end_read_request(struct bio * bi, unsigned int bytes_done,
1114                                    int error)
1115 {
1116         struct stripe_head *sh = bi->bi_private;
1117         raid5_conf_t *conf = sh->raid_conf;
1118         int disks = sh->disks, i;
1119         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1120         char b[BDEVNAME_SIZE];
1121         mdk_rdev_t *rdev;
1122
1123         if (bi->bi_size)
1124                 return 1;
1125
1126         for (i=0 ; i<disks; i++)
1127                 if (bi == &sh->dev[i].req)
1128                         break;
1129
1130         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1131                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1132                 uptodate);
1133         if (i == disks) {
1134                 BUG();
1135                 return 0;
1136         }
1137
1138         if (uptodate) {
1139                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1140                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1141                         rdev = conf->disks[i].rdev;
1142                         printk(KERN_INFO "raid5:%s: read error corrected (%lu sectors at %llu on %s)\n",
1143                                mdname(conf->mddev), STRIPE_SECTORS,
1144                                (unsigned long long)sh->sector + rdev->data_offset,
1145                                bdevname(rdev->bdev, b));
1146                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1147                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1148                 }
1149                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1150                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1151         } else {
1152                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1153                 int retry = 0;
1154                 rdev = conf->disks[i].rdev;
1155
1156                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1157                 atomic_inc(&rdev->read_errors);
1158                 if (conf->mddev->degraded)
1159                         printk(KERN_WARNING "raid5:%s: read error not correctable (sector %llu on %s).\n",
1160                                mdname(conf->mddev),
1161                                (unsigned long long)sh->sector + rdev->data_offset,
1162                                bdn);
1163                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1164                         /* Oh, no!!! */
1165                         printk(KERN_WARNING "raid5:%s: read error NOT corrected!! (sector %llu on %s).\n",
1166                                mdname(conf->mddev),
1167                                (unsigned long long)sh->sector + rdev->data_offset,
1168                                bdn);
1169                 else if (atomic_read(&rdev->read_errors)
1170                          > conf->max_nr_stripes)
1171                         printk(KERN_WARNING
1172                                "raid5:%s: Too many read errors, failing device %s.\n",
1173                                mdname(conf->mddev), bdn);
1174                 else
1175                         retry = 1;
1176                 if (retry)
1177                         set_bit(R5_ReadError, &sh->dev[i].flags);
1178                 else {
1179                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1180                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1181                         md_error(conf->mddev, rdev);
1182                 }
1183         }
1184         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1185         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1186         set_bit(STRIPE_HANDLE, &sh->state);
1187         release_stripe(sh);
1188         return 0;
1189 }
1190
1191 static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
1192                                     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         if (bi->bi_size)
1200                 return 1;
1201
1202         for (i=0 ; i<disks; i++)
1203                 if (bi == &sh->dev[i].req)
1204                         break;
1205
1206         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1207                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1208                 uptodate);
1209         if (i == disks) {
1210                 BUG();
1211                 return 0;
1212         }
1213
1214         if (!uptodate)
1215                 md_error(conf->mddev, conf->disks[i].rdev);
1216
1217         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1218         
1219         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1220         set_bit(STRIPE_HANDLE, &sh->state);
1221         release_stripe(sh);
1222         return 0;
1223 }
1224
1225
1226 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1227         
1228 static void raid5_build_block (struct stripe_head *sh, int i)
1229 {
1230         struct r5dev *dev = &sh->dev[i];
1231
1232         bio_init(&dev->req);
1233         dev->req.bi_io_vec = &dev->vec;
1234         dev->req.bi_vcnt++;
1235         dev->req.bi_max_vecs++;
1236         dev->vec.bv_page = dev->page;
1237         dev->vec.bv_len = STRIPE_SIZE;
1238         dev->vec.bv_offset = 0;
1239
1240         dev->req.bi_sector = sh->sector;
1241         dev->req.bi_private = sh;
1242
1243         dev->flags = 0;
1244         dev->sector = compute_blocknr(sh, i);
1245 }
1246
1247 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1248 {
1249         char b[BDEVNAME_SIZE];
1250         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1251         pr_debug("raid5: error called\n");
1252
1253         if (!test_bit(Faulty, &rdev->flags)) {
1254                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1255                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1256                         unsigned long flags;
1257                         spin_lock_irqsave(&conf->device_lock, flags);
1258                         mddev->degraded++;
1259                         spin_unlock_irqrestore(&conf->device_lock, flags);
1260                         /*
1261                          * if recovery was running, make sure it aborts.
1262                          */
1263                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
1264                 }
1265                 set_bit(Faulty, &rdev->flags);
1266                 printk (KERN_ALERT
1267                         "raid5: Disk failure on %s, disabling device."
1268                         " Operation continuing on %d devices\n",
1269                         bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1270         }
1271 }
1272
1273 /*
1274  * Input: a 'big' sector number,
1275  * Output: index of the data and parity disk, and the sector # in them.
1276  */
1277 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1278                         unsigned int data_disks, unsigned int * dd_idx,
1279                         unsigned int * pd_idx, raid5_conf_t *conf)
1280 {
1281         long stripe;
1282         unsigned long chunk_number;
1283         unsigned int chunk_offset;
1284         sector_t new_sector;
1285         int sectors_per_chunk = conf->chunk_size >> 9;
1286
1287         /* First compute the information on this sector */
1288
1289         /*
1290          * Compute the chunk number and the sector offset inside the chunk
1291          */
1292         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1293         chunk_number = r_sector;
1294         BUG_ON(r_sector != chunk_number);
1295
1296         /*
1297          * Compute the stripe number
1298          */
1299         stripe = chunk_number / data_disks;
1300
1301         /*
1302          * Compute the data disk and parity disk indexes inside the stripe
1303          */
1304         *dd_idx = chunk_number % data_disks;
1305
1306         /*
1307          * Select the parity disk based on the user selected algorithm.
1308          */
1309         switch(conf->level) {
1310         case 4:
1311                 *pd_idx = data_disks;
1312                 break;
1313         case 5:
1314                 switch (conf->algorithm) {
1315                 case ALGORITHM_LEFT_ASYMMETRIC:
1316                         *pd_idx = data_disks - stripe % raid_disks;
1317                         if (*dd_idx >= *pd_idx)
1318                                 (*dd_idx)++;
1319                         break;
1320                 case ALGORITHM_RIGHT_ASYMMETRIC:
1321                         *pd_idx = stripe % raid_disks;
1322                         if (*dd_idx >= *pd_idx)
1323                                 (*dd_idx)++;
1324                         break;
1325                 case ALGORITHM_LEFT_SYMMETRIC:
1326                         *pd_idx = data_disks - stripe % raid_disks;
1327                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1328                         break;
1329                 case ALGORITHM_RIGHT_SYMMETRIC:
1330                         *pd_idx = stripe % raid_disks;
1331                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1332                         break;
1333                 default:
1334                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1335                                 conf->algorithm);
1336                 }
1337                 break;
1338         case 6:
1339
1340                 /**** FIX THIS ****/
1341                 switch (conf->algorithm) {
1342                 case ALGORITHM_LEFT_ASYMMETRIC:
1343                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1344                         if (*pd_idx == raid_disks-1)
1345                                 (*dd_idx)++;    /* Q D D D P */
1346                         else if (*dd_idx >= *pd_idx)
1347                                 (*dd_idx) += 2; /* D D P Q D */
1348                         break;
1349                 case ALGORITHM_RIGHT_ASYMMETRIC:
1350                         *pd_idx = stripe % raid_disks;
1351                         if (*pd_idx == raid_disks-1)
1352                                 (*dd_idx)++;    /* Q D D D P */
1353                         else if (*dd_idx >= *pd_idx)
1354                                 (*dd_idx) += 2; /* D D P Q D */
1355                         break;
1356                 case ALGORITHM_LEFT_SYMMETRIC:
1357                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1358                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1359                         break;
1360                 case ALGORITHM_RIGHT_SYMMETRIC:
1361                         *pd_idx = stripe % raid_disks;
1362                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1363                         break;
1364                 default:
1365                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1366                                 conf->algorithm);
1367                 }
1368                 break;
1369         }
1370
1371         /*
1372          * Finally, compute the new sector number
1373          */
1374         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1375         return new_sector;
1376 }
1377
1378
1379 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1380 {
1381         raid5_conf_t *conf = sh->raid_conf;
1382         int raid_disks = sh->disks;
1383         int data_disks = raid_disks - conf->max_degraded;
1384         sector_t new_sector = sh->sector, check;
1385         int sectors_per_chunk = conf->chunk_size >> 9;
1386         sector_t stripe;
1387         int chunk_offset;
1388         int chunk_number, dummy1, dummy2, dd_idx = i;
1389         sector_t r_sector;
1390
1391
1392         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1393         stripe = new_sector;
1394         BUG_ON(new_sector != stripe);
1395
1396         if (i == sh->pd_idx)
1397                 return 0;
1398         switch(conf->level) {
1399         case 4: break;
1400         case 5:
1401                 switch (conf->algorithm) {
1402                 case ALGORITHM_LEFT_ASYMMETRIC:
1403                 case ALGORITHM_RIGHT_ASYMMETRIC:
1404                         if (i > sh->pd_idx)
1405                                 i--;
1406                         break;
1407                 case ALGORITHM_LEFT_SYMMETRIC:
1408                 case ALGORITHM_RIGHT_SYMMETRIC:
1409                         if (i < sh->pd_idx)
1410                                 i += raid_disks;
1411                         i -= (sh->pd_idx + 1);
1412                         break;
1413                 default:
1414                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1415                                conf->algorithm);
1416                 }
1417                 break;
1418         case 6:
1419                 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1420                         return 0; /* It is the Q disk */
1421                 switch (conf->algorithm) {
1422                 case ALGORITHM_LEFT_ASYMMETRIC:
1423                 case ALGORITHM_RIGHT_ASYMMETRIC:
1424                         if (sh->pd_idx == raid_disks-1)
1425                                 i--;    /* Q D D D P */
1426                         else if (i > sh->pd_idx)
1427                                 i -= 2; /* D D P Q D */
1428                         break;
1429                 case ALGORITHM_LEFT_SYMMETRIC:
1430                 case ALGORITHM_RIGHT_SYMMETRIC:
1431                         if (sh->pd_idx == raid_disks-1)
1432                                 i--; /* Q D D D P */
1433                         else {
1434                                 /* D D P Q D */
1435                                 if (i < sh->pd_idx)
1436                                         i += raid_disks;
1437                                 i -= (sh->pd_idx + 2);
1438                         }
1439                         break;
1440                 default:
1441                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1442                                 conf->algorithm);
1443                 }
1444                 break;
1445         }
1446
1447         chunk_number = stripe * data_disks + i;
1448         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1449
1450         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1451         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1452                 printk(KERN_ERR "compute_blocknr: map not correct\n");
1453                 return 0;
1454         }
1455         return r_sector;
1456 }
1457
1458
1459
1460 /*
1461  * Copy data between a page in the stripe cache, and one or more bion
1462  * The page could align with the middle of the bio, or there could be
1463  * several bion, each with several bio_vecs, which cover part of the page
1464  * Multiple bion are linked together on bi_next.  There may be extras
1465  * at the end of this list.  We ignore them.
1466  */
1467 static void copy_data(int frombio, struct bio *bio,
1468                      struct page *page,
1469                      sector_t sector)
1470 {
1471         char *pa = page_address(page);
1472         struct bio_vec *bvl;
1473         int i;
1474         int page_offset;
1475
1476         if (bio->bi_sector >= sector)
1477                 page_offset = (signed)(bio->bi_sector - sector) * 512;
1478         else
1479                 page_offset = (signed)(sector - bio->bi_sector) * -512;
1480         bio_for_each_segment(bvl, bio, i) {
1481                 int len = bio_iovec_idx(bio,i)->bv_len;
1482                 int clen;
1483                 int b_offset = 0;
1484
1485                 if (page_offset < 0) {
1486                         b_offset = -page_offset;
1487                         page_offset += b_offset;
1488                         len -= b_offset;
1489                 }
1490
1491                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1492                         clen = STRIPE_SIZE - page_offset;
1493                 else clen = len;
1494
1495                 if (clen > 0) {
1496                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1497                         if (frombio)
1498                                 memcpy(pa+page_offset, ba+b_offset, clen);
1499                         else
1500                                 memcpy(ba+b_offset, pa+page_offset, clen);
1501                         __bio_kunmap_atomic(ba, KM_USER0);
1502                 }
1503                 if (clen < len) /* hit end of page */
1504                         break;
1505                 page_offset +=  len;
1506         }
1507 }
1508
1509 #define check_xor()     do {                                              \
1510                                 if (count == MAX_XOR_BLOCKS) {            \
1511                                 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1512                                 count = 0;                                \
1513                            }                                              \
1514                         } while(0)
1515
1516 static void compute_parity6(struct stripe_head *sh, int method)
1517 {
1518         raid6_conf_t *conf = sh->raid_conf;
1519         int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1520         struct bio *chosen;
1521         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1522         void *ptrs[disks];
1523
1524         qd_idx = raid6_next_disk(pd_idx, disks);
1525         d0_idx = raid6_next_disk(qd_idx, disks);
1526
1527         pr_debug("compute_parity, stripe %llu, method %d\n",
1528                 (unsigned long long)sh->sector, method);
1529
1530         switch(method) {
1531         case READ_MODIFY_WRITE:
1532                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
1533         case RECONSTRUCT_WRITE:
1534                 for (i= disks; i-- ;)
1535                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1536                                 chosen = sh->dev[i].towrite;
1537                                 sh->dev[i].towrite = NULL;
1538
1539                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1540                                         wake_up(&conf->wait_for_overlap);
1541
1542                                 BUG_ON(sh->dev[i].written);
1543                                 sh->dev[i].written = chosen;
1544                         }
1545                 break;
1546         case CHECK_PARITY:
1547                 BUG();          /* Not implemented yet */
1548         }
1549
1550         for (i = disks; i--;)
1551                 if (sh->dev[i].written) {
1552                         sector_t sector = sh->dev[i].sector;
1553                         struct bio *wbi = sh->dev[i].written;
1554                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1555                                 copy_data(1, wbi, sh->dev[i].page, sector);
1556                                 wbi = r5_next_bio(wbi, sector);
1557                         }
1558
1559                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1560                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1561                 }
1562
1563 //      switch(method) {
1564 //      case RECONSTRUCT_WRITE:
1565 //      case CHECK_PARITY:
1566 //      case UPDATE_PARITY:
1567                 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1568                 /* FIX: Is this ordering of drives even remotely optimal? */
1569                 count = 0;
1570                 i = d0_idx;
1571                 do {
1572                         ptrs[count++] = page_address(sh->dev[i].page);
1573                         if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1574                                 printk("block %d/%d not uptodate on parity calc\n", i,count);
1575                         i = raid6_next_disk(i, disks);
1576                 } while ( i != d0_idx );
1577 //              break;
1578 //      }
1579
1580         raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1581
1582         switch(method) {
1583         case RECONSTRUCT_WRITE:
1584                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1585                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1586                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1587                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
1588                 break;
1589         case UPDATE_PARITY:
1590                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1591                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1592                 break;
1593         }
1594 }
1595
1596
1597 /* Compute one missing block */
1598 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1599 {
1600         int i, count, disks = sh->disks;
1601         void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1602         int pd_idx = sh->pd_idx;
1603         int qd_idx = raid6_next_disk(pd_idx, disks);
1604
1605         pr_debug("compute_block_1, stripe %llu, idx %d\n",
1606                 (unsigned long long)sh->sector, dd_idx);
1607
1608         if ( dd_idx == qd_idx ) {
1609                 /* We're actually computing the Q drive */
1610                 compute_parity6(sh, UPDATE_PARITY);
1611         } else {
1612                 dest = page_address(sh->dev[dd_idx].page);
1613                 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1614                 count = 0;
1615                 for (i = disks ; i--; ) {
1616                         if (i == dd_idx || i == qd_idx)
1617                                 continue;
1618                         p = page_address(sh->dev[i].page);
1619                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1620                                 ptr[count++] = p;
1621                         else
1622                                 printk("compute_block() %d, stripe %llu, %d"
1623                                        " not present\n", dd_idx,
1624                                        (unsigned long long)sh->sector, i);
1625
1626                         check_xor();
1627                 }
1628                 if (count)
1629                         xor_blocks(count, STRIPE_SIZE, dest, ptr);
1630                 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1631                 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1632         }
1633 }
1634
1635 /* Compute two missing blocks */
1636 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1637 {
1638         int i, count, disks = sh->disks;
1639         int pd_idx = sh->pd_idx;
1640         int qd_idx = raid6_next_disk(pd_idx, disks);
1641         int d0_idx = raid6_next_disk(qd_idx, disks);
1642         int faila, failb;
1643
1644         /* faila and failb are disk numbers relative to d0_idx */
1645         /* pd_idx become disks-2 and qd_idx become disks-1 */
1646         faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1647         failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1648
1649         BUG_ON(faila == failb);
1650         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1651
1652         pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1653                (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1654
1655         if ( failb == disks-1 ) {
1656                 /* Q disk is one of the missing disks */
1657                 if ( faila == disks-2 ) {
1658                         /* Missing P+Q, just recompute */
1659                         compute_parity6(sh, UPDATE_PARITY);
1660                         return;
1661                 } else {
1662                         /* We're missing D+Q; recompute D from P */
1663                         compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1664                         compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1665                         return;
1666                 }
1667         }
1668
1669         /* We're missing D+P or D+D; build pointer table */
1670         {
1671                 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1672                 void *ptrs[disks];
1673
1674                 count = 0;
1675                 i = d0_idx;
1676                 do {
1677                         ptrs[count++] = page_address(sh->dev[i].page);
1678                         i = raid6_next_disk(i, disks);
1679                         if (i != dd_idx1 && i != dd_idx2 &&
1680                             !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1681                                 printk("compute_2 with missing block %d/%d\n", count, i);
1682                 } while ( i != d0_idx );
1683
1684                 if ( failb == disks-2 ) {
1685                         /* We're missing D+P. */
1686                         raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1687                 } else {
1688                         /* We're missing D+D. */
1689                         raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1690                 }
1691
1692                 /* Both the above update both missing blocks */
1693                 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1694                 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1695         }
1696 }
1697
1698 static int
1699 handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
1700 {
1701         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1702         int locked = 0;
1703
1704         if (rcw) {
1705                 /* if we are not expanding this is a proper write request, and
1706                  * there will be bios with new data to be drained into the
1707                  * stripe cache
1708                  */
1709                 if (!expand) {
1710                         set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1711                         sh->ops.count++;
1712                 }
1713
1714                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1715                 sh->ops.count++;
1716
1717                 for (i = disks; i--; ) {
1718                         struct r5dev *dev = &sh->dev[i];
1719
1720                         if (dev->towrite) {
1721                                 set_bit(R5_LOCKED, &dev->flags);
1722                                 if (!expand)
1723                                         clear_bit(R5_UPTODATE, &dev->flags);
1724                                 locked++;
1725                         }
1726                 }
1727         } else {
1728                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1729                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1730
1731                 set_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
1732                 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1733                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1734
1735                 sh->ops.count += 3;
1736
1737                 for (i = disks; i--; ) {
1738                         struct r5dev *dev = &sh->dev[i];
1739                         if (i == pd_idx)
1740                                 continue;
1741
1742                         /* For a read-modify write there may be blocks that are
1743                          * locked for reading while others are ready to be
1744                          * written so we distinguish these blocks by the
1745                          * R5_Wantprexor bit
1746                          */
1747                         if (dev->towrite &&
1748                             (test_bit(R5_UPTODATE, &dev->flags) ||
1749                             test_bit(R5_Wantcompute, &dev->flags))) {
1750                                 set_bit(R5_Wantprexor, &dev->flags);
1751                                 set_bit(R5_LOCKED, &dev->flags);
1752                                 clear_bit(R5_UPTODATE, &dev->flags);
1753                                 locked++;
1754                         }
1755                 }
1756         }
1757
1758         /* keep the parity disk locked while asynchronous operations
1759          * are in flight
1760          */
1761         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1762         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1763         locked++;
1764
1765         pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1766                 __FUNCTION__, (unsigned long long)sh->sector,
1767                 locked, sh->ops.pending);
1768
1769         return locked;
1770 }
1771
1772 /*
1773  * Each stripe/dev can have one or more bion attached.
1774  * toread/towrite point to the first in a chain.
1775  * The bi_next chain must be in order.
1776  */
1777 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1778 {
1779         struct bio **bip;
1780         raid5_conf_t *conf = sh->raid_conf;
1781         int firstwrite=0;
1782
1783         pr_debug("adding bh b#%llu to stripe s#%llu\n",
1784                 (unsigned long long)bi->bi_sector,
1785                 (unsigned long long)sh->sector);
1786
1787
1788         spin_lock(&sh->lock);
1789         spin_lock_irq(&conf->device_lock);
1790         if (forwrite) {
1791                 bip = &sh->dev[dd_idx].towrite;
1792                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1793                         firstwrite = 1;
1794         } else
1795                 bip = &sh->dev[dd_idx].toread;
1796         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1797                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1798                         goto overlap;
1799                 bip = & (*bip)->bi_next;
1800         }
1801         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1802                 goto overlap;
1803
1804         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1805         if (*bip)
1806                 bi->bi_next = *bip;
1807         *bip = bi;
1808         bi->bi_phys_segments ++;
1809         spin_unlock_irq(&conf->device_lock);
1810         spin_unlock(&sh->lock);
1811
1812         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1813                 (unsigned long long)bi->bi_sector,
1814                 (unsigned long long)sh->sector, dd_idx);
1815
1816         if (conf->mddev->bitmap && firstwrite) {
1817                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1818                                   STRIPE_SECTORS, 0);
1819                 sh->bm_seq = conf->seq_flush+1;
1820                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1821         }
1822
1823         if (forwrite) {
1824                 /* check if page is covered */
1825                 sector_t sector = sh->dev[dd_idx].sector;
1826                 for (bi=sh->dev[dd_idx].towrite;
1827                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1828                              bi && bi->bi_sector <= sector;
1829                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1830                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1831                                 sector = bi->bi_sector + (bi->bi_size>>9);
1832                 }
1833                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1834                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1835         }
1836         return 1;
1837
1838  overlap:
1839         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1840         spin_unlock_irq(&conf->device_lock);
1841         spin_unlock(&sh->lock);
1842         return 0;
1843 }
1844
1845 static void end_reshape(raid5_conf_t *conf);
1846
1847 static int page_is_zero(struct page *p)
1848 {
1849         char *a = page_address(p);
1850         return ((*(u32*)a) == 0 &&
1851                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1852 }
1853
1854 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1855 {
1856         int sectors_per_chunk = conf->chunk_size >> 9;
1857         int pd_idx, dd_idx;
1858         int chunk_offset = sector_div(stripe, sectors_per_chunk);
1859
1860         raid5_compute_sector(stripe * (disks - conf->max_degraded)
1861                              *sectors_per_chunk + chunk_offset,
1862                              disks, disks - conf->max_degraded,
1863                              &dd_idx, &pd_idx, conf);
1864         return pd_idx;
1865 }
1866
1867 static void
1868 handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1869                                 struct stripe_head_state *s, int disks,
1870                                 struct bio **return_bi)
1871 {
1872         int i;
1873         for (i = disks; i--; ) {
1874                 struct bio *bi;
1875                 int bitmap_end = 0;
1876
1877                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1878                         mdk_rdev_t *rdev;
1879                         rcu_read_lock();
1880                         rdev = rcu_dereference(conf->disks[i].rdev);
1881                         if (rdev && test_bit(In_sync, &rdev->flags))
1882                                 /* multiple read failures in one stripe */
1883                                 md_error(conf->mddev, rdev);
1884                         rcu_read_unlock();
1885                 }
1886                 spin_lock_irq(&conf->device_lock);
1887                 /* fail all writes first */
1888                 bi = sh->dev[i].towrite;
1889                 sh->dev[i].towrite = NULL;
1890                 if (bi) {
1891                         s->to_write--;
1892                         bitmap_end = 1;
1893                 }
1894
1895                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1896                         wake_up(&conf->wait_for_overlap);
1897
1898                 while (bi && bi->bi_sector <
1899                         sh->dev[i].sector + STRIPE_SECTORS) {
1900                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1901                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1902                         if (--bi->bi_phys_segments == 0) {
1903                                 md_write_end(conf->mddev);
1904                                 bi->bi_next = *return_bi;
1905                                 *return_bi = bi;
1906                         }
1907                         bi = nextbi;
1908                 }
1909                 /* and fail all 'written' */
1910                 bi = sh->dev[i].written;
1911                 sh->dev[i].written = NULL;
1912                 if (bi) bitmap_end = 1;
1913                 while (bi && bi->bi_sector <
1914                        sh->dev[i].sector + STRIPE_SECTORS) {
1915                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1916                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1917                         if (--bi->bi_phys_segments == 0) {
1918                                 md_write_end(conf->mddev);
1919                                 bi->bi_next = *return_bi;
1920                                 *return_bi = bi;
1921                         }
1922                         bi = bi2;
1923                 }
1924
1925                 /* fail any reads if this device is non-operational and
1926                  * the data has not reached the cache yet.
1927                  */
1928                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
1929                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1930                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
1931                         bi = sh->dev[i].toread;
1932                         sh->dev[i].toread = NULL;
1933                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1934                                 wake_up(&conf->wait_for_overlap);
1935                         if (bi) s->to_read--;
1936                         while (bi && bi->bi_sector <
1937                                sh->dev[i].sector + STRIPE_SECTORS) {
1938                                 struct bio *nextbi =
1939                                         r5_next_bio(bi, sh->dev[i].sector);
1940                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1941                                 if (--bi->bi_phys_segments == 0) {
1942                                         bi->bi_next = *return_bi;
1943                                         *return_bi = bi;
1944                                 }
1945                                 bi = nextbi;
1946                         }
1947                 }
1948                 spin_unlock_irq(&conf->device_lock);
1949                 if (bitmap_end)
1950                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1951                                         STRIPE_SECTORS, 0, 0);
1952         }
1953
1954 }
1955
1956 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
1957  * to process
1958  */
1959 static int __handle_issuing_new_read_requests5(struct stripe_head *sh,
1960                         struct stripe_head_state *s, int disk_idx, int disks)
1961 {
1962         struct r5dev *dev = &sh->dev[disk_idx];
1963         struct r5dev *failed_dev = &sh->dev[s->failed_num];
1964
1965         /* don't schedule compute operations or reads on the parity block while
1966          * a check is in flight
1967          */
1968         if ((disk_idx == sh->pd_idx) &&
1969              test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
1970                 return ~0;
1971
1972         /* is the data in this block needed, and can we get it? */
1973         if (!test_bit(R5_LOCKED, &dev->flags) &&
1974             !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread ||
1975             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1976              s->syncing || s->expanding || (s->failed &&
1977              (failed_dev->toread || (failed_dev->towrite &&
1978              !test_bit(R5_OVERWRITE, &failed_dev->flags)
1979              ))))) {
1980                 /* 1/ We would like to get this block, possibly by computing it,
1981                  * but we might not be able to.
1982                  *
1983                  * 2/ Since parity check operations potentially make the parity
1984                  * block !uptodate it will need to be refreshed before any
1985                  * compute operations on data disks are scheduled.
1986                  *
1987                  * 3/ We hold off parity block re-reads until check operations
1988                  * have quiesced.
1989                  */
1990                 if ((s->uptodate == disks - 1) &&
1991                     !test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
1992                         set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
1993                         set_bit(R5_Wantcompute, &dev->flags);
1994                         sh->ops.target = disk_idx;
1995                         s->req_compute = 1;
1996                         sh->ops.count++;
1997                         /* Careful: from this point on 'uptodate' is in the eye
1998                          * of raid5_run_ops which services 'compute' operations
1999                          * before writes. R5_Wantcompute flags a block that will
2000                          * be R5_UPTODATE by the time it is needed for a
2001                          * subsequent operation.
2002                          */
2003                         s->uptodate++;
2004                         return 0; /* uptodate + compute == disks */
2005                 } else if ((s->uptodate < disks - 1) &&
2006                         test_bit(R5_Insync, &dev->flags)) {
2007                         /* Note: we hold off compute operations while checks are
2008                          * in flight, but we still prefer 'compute' over 'read'
2009                          * hence we only read if (uptodate < * disks-1)
2010                          */
2011                         set_bit(R5_LOCKED, &dev->flags);
2012                         set_bit(R5_Wantread, &dev->flags);
2013                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2014                                 sh->ops.count++;
2015                         s->locked++;
2016                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2017                                 s->syncing);
2018                 }
2019         }
2020
2021         return ~0;
2022 }
2023
2024 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2025                         struct stripe_head_state *s, int disks)
2026 {
2027         int i;
2028
2029         /* Clear completed compute operations.  Parity recovery
2030          * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled
2031          * later on in this routine
2032          */
2033         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2034                 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2035                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2036                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2037                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2038         }
2039
2040         /* look for blocks to read/compute, skip this if a compute
2041          * is already in flight, or if the stripe contents are in the
2042          * midst of changing due to a write
2043          */
2044         if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2045                 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2046                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2047                 for (i = disks; i--; )
2048                         if (__handle_issuing_new_read_requests5(
2049                                 sh, s, i, disks) == 0)
2050                                 break;
2051         }
2052         set_bit(STRIPE_HANDLE, &sh->state);
2053 }
2054
2055 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2056                         struct stripe_head_state *s, struct r6_state *r6s,
2057                         int disks)
2058 {
2059         int i;
2060         for (i = disks; i--; ) {
2061                 struct r5dev *dev = &sh->dev[i];
2062                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2063                     !test_bit(R5_UPTODATE, &dev->flags) &&
2064                     (dev->toread || (dev->towrite &&
2065                      !test_bit(R5_OVERWRITE, &dev->flags)) ||
2066                      s->syncing || s->expanding ||
2067                      (s->failed >= 1 &&
2068                       (sh->dev[r6s->failed_num[0]].toread ||
2069                        s->to_write)) ||
2070                      (s->failed >= 2 &&
2071                       (sh->dev[r6s->failed_num[1]].toread ||
2072                        s->to_write)))) {
2073                         /* we would like to get this block, possibly
2074                          * by computing it, but we might not be able to
2075                          */
2076                         if (s->uptodate == disks-1) {
2077                                 pr_debug("Computing stripe %llu block %d\n",
2078                                        (unsigned long long)sh->sector, i);
2079                                 compute_block_1(sh, i, 0);
2080                                 s->uptodate++;
2081                         } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2082                                 /* Computing 2-failure is *very* expensive; only
2083                                  * do it if failed >= 2
2084                                  */
2085                                 int other;
2086                                 for (other = disks; other--; ) {
2087                                         if (other == i)
2088                                                 continue;
2089                                         if (!test_bit(R5_UPTODATE,
2090                                               &sh->dev[other].flags))
2091                                                 break;
2092                                 }
2093                                 BUG_ON(other < 0);
2094                                 pr_debug("Computing stripe %llu blocks %d,%d\n",
2095                                        (unsigned long long)sh->sector,
2096                                        i, other);
2097                                 compute_block_2(sh, i, other);
2098                                 s->uptodate += 2;
2099                         } else if (test_bit(R5_Insync, &dev->flags)) {
2100                                 set_bit(R5_LOCKED, &dev->flags);
2101                                 set_bit(R5_Wantread, &dev->flags);
2102                                 s->locked++;
2103                                 pr_debug("Reading block %d (sync=%d)\n",
2104                                         i, s->syncing);
2105                         }
2106                 }
2107         }
2108         set_bit(STRIPE_HANDLE, &sh->state);
2109 }
2110
2111
2112 /* handle_completed_write_requests
2113  * any written block on an uptodate or failed drive can be returned.
2114  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2115  * never LOCKED, so we don't need to test 'failed' directly.
2116  */
2117 static void handle_completed_write_requests(raid5_conf_t *conf,
2118         struct stripe_head *sh, int disks, struct bio **return_bi)
2119 {
2120         int i;
2121         struct r5dev *dev;
2122
2123         for (i = disks; i--; )
2124                 if (sh->dev[i].written) {
2125                         dev = &sh->dev[i];
2126                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2127                                 test_bit(R5_UPTODATE, &dev->flags)) {
2128                                 /* We can return any write requests */
2129                                 struct bio *wbi, *wbi2;
2130                                 int bitmap_end = 0;
2131                                 pr_debug("Return write for disc %d\n", i);
2132                                 spin_lock_irq(&conf->device_lock);
2133                                 wbi = dev->written;
2134                                 dev->written = NULL;
2135                                 while (wbi && wbi->bi_sector <
2136                                         dev->sector + STRIPE_SECTORS) {
2137                                         wbi2 = r5_next_bio(wbi, dev->sector);
2138                                         if (--wbi->bi_phys_segments == 0) {
2139                                                 md_write_end(conf->mddev);
2140                                                 wbi->bi_next = *return_bi;
2141                                                 *return_bi = wbi;
2142                                         }
2143                                         wbi = wbi2;
2144                                 }
2145                                 if (dev->towrite == NULL)
2146                                         bitmap_end = 1;
2147                                 spin_unlock_irq(&conf->device_lock);
2148                                 if (bitmap_end)
2149                                         bitmap_endwrite(conf->mddev->bitmap,
2150                                                         sh->sector,
2151                                                         STRIPE_SECTORS,
2152                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2153                                                         0);
2154                         }
2155                 }
2156 }
2157
2158 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2159                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2160 {
2161         int rmw = 0, rcw = 0, i;
2162         for (i = disks; i--; ) {
2163                 /* would I have to read this buffer for read_modify_write */
2164                 struct r5dev *dev = &sh->dev[i];
2165                 if ((dev->towrite || i == sh->pd_idx) &&
2166                     !test_bit(R5_LOCKED, &dev->flags) &&
2167                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2168                       test_bit(R5_Wantcompute, &dev->flags))) {
2169                         if (test_bit(R5_Insync, &dev->flags))
2170                                 rmw++;
2171                         else
2172                                 rmw += 2*disks;  /* cannot read it */
2173                 }
2174                 /* Would I have to read this buffer for reconstruct_write */
2175                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2176                     !test_bit(R5_LOCKED, &dev->flags) &&
2177                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2178                     test_bit(R5_Wantcompute, &dev->flags))) {
2179                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2180                         else
2181                                 rcw += 2*disks;
2182                 }
2183         }
2184         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2185                 (unsigned long long)sh->sector, rmw, rcw);
2186         set_bit(STRIPE_HANDLE, &sh->state);
2187         if (rmw < rcw && rmw > 0)
2188                 /* prefer read-modify-write, but need to get some data */
2189                 for (i = disks; i--; ) {
2190                         struct r5dev *dev = &sh->dev[i];
2191                         if ((dev->towrite || i == sh->pd_idx) &&
2192                             !test_bit(R5_LOCKED, &dev->flags) &&
2193                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2194                             test_bit(R5_Wantcompute, &dev->flags)) &&
2195                             test_bit(R5_Insync, &dev->flags)) {
2196                                 if (
2197                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2198                                         pr_debug("Read_old block "
2199                                                 "%d for r-m-w\n", i);
2200                                         set_bit(R5_LOCKED, &dev->flags);
2201                                         set_bit(R5_Wantread, &dev->flags);
2202                                         if (!test_and_set_bit(
2203                                                 STRIPE_OP_IO, &sh->ops.pending))
2204                                                 sh->ops.count++;
2205                                         s->locked++;
2206                                 } else {
2207                                         set_bit(STRIPE_DELAYED, &sh->state);
2208                                         set_bit(STRIPE_HANDLE, &sh->state);
2209                                 }
2210                         }
2211                 }
2212         if (rcw <= rmw && rcw > 0)
2213                 /* want reconstruct write, but need to get some data */
2214                 for (i = disks; i--; ) {
2215                         struct r5dev *dev = &sh->dev[i];
2216                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2217                             i != sh->pd_idx &&
2218                             !test_bit(R5_LOCKED, &dev->flags) &&
2219                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2220                             test_bit(R5_Wantcompute, &dev->flags)) &&
2221                             test_bit(R5_Insync, &dev->flags)) {
2222                                 if (
2223                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2224                                         pr_debug("Read_old block "
2225                                                 "%d for Reconstruct\n", i);
2226                                         set_bit(R5_LOCKED, &dev->flags);
2227                                         set_bit(R5_Wantread, &dev->flags);
2228                                         if (!test_and_set_bit(
2229                                                 STRIPE_OP_IO, &sh->ops.pending))
2230                                                 sh->ops.count++;
2231                                         s->locked++;
2232                                 } else {
2233                                         set_bit(STRIPE_DELAYED, &sh->state);
2234                                         set_bit(STRIPE_HANDLE, &sh->state);
2235                                 }
2236                         }
2237                 }
2238         /* now if nothing is locked, and if we have enough data,
2239          * we can start a write request
2240          */
2241         /* since handle_stripe can be called at any time we need to handle the
2242          * case where a compute block operation has been submitted and then a
2243          * subsequent call wants to start a write request.  raid5_run_ops only
2244          * handles the case where compute block and postxor are requested
2245          * simultaneously.  If this is not the case then new writes need to be
2246          * held off until the compute completes.
2247          */
2248         if ((s->req_compute ||
2249             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) &&
2250                 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2251                 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2252                 s->locked += handle_write_operations5(sh, rcw == 0, 0);
2253 }
2254
2255 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2256                 struct stripe_head *sh, struct stripe_head_state *s,
2257                 struct r6_state *r6s, int disks)
2258 {
2259         int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2260         int qd_idx = r6s->qd_idx;
2261         for (i = disks; i--; ) {
2262                 struct r5dev *dev = &sh->dev[i];
2263                 /* Would I have to read this buffer for reconstruct_write */
2264                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2265                     && i != pd_idx && i != qd_idx
2266                     && (!test_bit(R5_LOCKED, &dev->flags)
2267                             ) &&
2268                     !test_bit(R5_UPTODATE, &dev->flags)) {
2269                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2270                         else {
2271                                 pr_debug("raid6: must_compute: "
2272                                         "disk %d flags=%#lx\n", i, dev->flags);
2273                                 must_compute++;
2274                         }
2275                 }
2276         }
2277         pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2278                (unsigned long long)sh->sector, rcw, must_compute);
2279         set_bit(STRIPE_HANDLE, &sh->state);
2280
2281         if (rcw > 0)
2282                 /* want reconstruct write, but need to get some data */
2283                 for (i = disks; i--; ) {
2284                         struct r5dev *dev = &sh->dev[i];
2285                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2286                             && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2287                             && !test_bit(R5_LOCKED, &dev->flags) &&
2288                             !test_bit(R5_UPTODATE, &dev->flags) &&
2289                             test_bit(R5_Insync, &dev->flags)) {
2290                                 if (
2291                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2292                                         pr_debug("Read_old stripe %llu "
2293                                                 "block %d for Reconstruct\n",
2294                                              (unsigned long long)sh->sector, i);
2295                                         set_bit(R5_LOCKED, &dev->flags);
2296                                         set_bit(R5_Wantread, &dev->flags);
2297                                         s->locked++;
2298                                 } else {
2299                                         pr_debug("Request delayed stripe %llu "
2300                                                 "block %d for Reconstruct\n",
2301                                              (unsigned long long)sh->sector, i);
2302                                         set_bit(STRIPE_DELAYED, &sh->state);
2303                                         set_bit(STRIPE_HANDLE, &sh->state);
2304                                 }
2305                         }
2306                 }
2307         /* now if nothing is locked, and if we have enough data, we can start a
2308          * write request
2309          */
2310         if (s->locked == 0 && rcw == 0 &&
2311             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2312                 if (must_compute > 0) {
2313                         /* We have failed blocks and need to compute them */
2314                         switch (s->failed) {
2315                         case 0:
2316                                 BUG();
2317                         case 1:
2318                                 compute_block_1(sh, r6s->failed_num[0], 0);
2319                                 break;
2320                         case 2:
2321                                 compute_block_2(sh, r6s->failed_num[0],
2322                                                 r6s->failed_num[1]);
2323                                 break;
2324                         default: /* This request should have been failed? */
2325                                 BUG();
2326                         }
2327                 }
2328
2329                 pr_debug("Computing parity for stripe %llu\n",
2330                         (unsigned long long)sh->sector);
2331                 compute_parity6(sh, RECONSTRUCT_WRITE);
2332                 /* now every locked buffer is ready to be written */
2333                 for (i = disks; i--; )
2334                         if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2335                                 pr_debug("Writing stripe %llu block %d\n",
2336                                        (unsigned long long)sh->sector, i);
2337                                 s->locked++;
2338                                 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2339                         }
2340                 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2341                 set_bit(STRIPE_INSYNC, &sh->state);
2342
2343                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2344                         atomic_dec(&conf->preread_active_stripes);
2345                         if (atomic_read(&conf->preread_active_stripes) <
2346                             IO_THRESHOLD)
2347                                 md_wakeup_thread(conf->mddev->thread);
2348                 }
2349         }
2350 }
2351
2352 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2353                                 struct stripe_head_state *s, int disks)
2354 {
2355         set_bit(STRIPE_HANDLE, &sh->state);
2356         /* Take one of the following actions:
2357          * 1/ start a check parity operation if (uptodate == disks)
2358          * 2/ finish a check parity operation and act on the result
2359          * 3/ skip to the writeback section if we previously
2360          *    initiated a recovery operation
2361          */
2362         if (s->failed == 0 &&
2363             !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2364                 if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2365                         BUG_ON(s->uptodate != disks);
2366                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2367                         sh->ops.count++;
2368                         s->uptodate--;
2369                 } else if (
2370                        test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) {
2371                         clear_bit(STRIPE_OP_CHECK, &sh->ops.ack);
2372                         clear_bit(STRIPE_OP_CHECK, &sh->ops.pending);
2373
2374                         if (sh->ops.zero_sum_result == 0)
2375                                 /* parity is correct (on disc,
2376                                  * not in buffer any more)
2377                                  */
2378                                 set_bit(STRIPE_INSYNC, &sh->state);
2379                         else {
2380                                 conf->mddev->resync_mismatches +=
2381                                         STRIPE_SECTORS;
2382                                 if (test_bit(
2383                                      MD_RECOVERY_CHECK, &conf->mddev->recovery))
2384                                         /* don't try to repair!! */
2385                                         set_bit(STRIPE_INSYNC, &sh->state);
2386                                 else {
2387                                         set_bit(STRIPE_OP_COMPUTE_BLK,
2388                                                 &sh->ops.pending);
2389                                         set_bit(STRIPE_OP_MOD_REPAIR_PD,
2390                                                 &sh->ops.pending);
2391                                         set_bit(R5_Wantcompute,
2392                                                 &sh->dev[sh->pd_idx].flags);
2393                                         sh->ops.target = sh->pd_idx;
2394                                         sh->ops.count++;
2395                                         s->uptodate++;
2396                                 }
2397                         }
2398                 }
2399         }
2400
2401         /* check if we can clear a parity disk reconstruct */
2402         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2403                 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2404
2405                 clear_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending);
2406                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2407                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2408                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2409         }
2410
2411         /* Wait for check parity and compute block operations to complete
2412          * before write-back
2413          */
2414         if (!test_bit(STRIPE_INSYNC, &sh->state) &&
2415                 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) &&
2416                 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) {
2417                 struct r5dev *dev;
2418                 /* either failed parity check, or recovery is happening */
2419                 if (s->failed == 0)
2420                         s->failed_num = sh->pd_idx;
2421                 dev = &sh->dev[s->failed_num];
2422                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2423                 BUG_ON(s->uptodate != disks);
2424
2425                 set_bit(R5_LOCKED, &dev->flags);
2426                 set_bit(R5_Wantwrite, &dev->flags);
2427                 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2428                         sh->ops.count++;
2429
2430                 clear_bit(STRIPE_DEGRADED, &sh->state);
2431                 s->locked++;
2432                 set_bit(STRIPE_INSYNC, &sh->state);
2433         }
2434 }
2435
2436
2437 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2438                                 struct stripe_head_state *s,
2439                                 struct r6_state *r6s, struct page *tmp_page,
2440                                 int disks)
2441 {
2442         int update_p = 0, update_q = 0;
2443         struct r5dev *dev;
2444         int pd_idx = sh->pd_idx;
2445         int qd_idx = r6s->qd_idx;
2446
2447         set_bit(STRIPE_HANDLE, &sh->state);
2448
2449         BUG_ON(s->failed > 2);
2450         BUG_ON(s->uptodate < disks);
2451         /* Want to check and possibly repair P and Q.
2452          * However there could be one 'failed' device, in which
2453          * case we can only check one of them, possibly using the
2454          * other to generate missing data
2455          */
2456
2457         /* If !tmp_page, we cannot do the calculations,
2458          * but as we have set STRIPE_HANDLE, we will soon be called
2459          * by stripe_handle with a tmp_page - just wait until then.
2460          */
2461         if (tmp_page) {
2462                 if (s->failed == r6s->q_failed) {
2463                         /* The only possible failed device holds 'Q', so it
2464                          * makes sense to check P (If anything else were failed,
2465                          * we would have used P to recreate it).
2466                          */
2467                         compute_block_1(sh, pd_idx, 1);
2468                         if (!page_is_zero(sh->dev[pd_idx].page)) {
2469                                 compute_block_1(sh, pd_idx, 0);
2470                                 update_p = 1;
2471                         }
2472                 }
2473                 if (!r6s->q_failed && s->failed < 2) {
2474                         /* q is not failed, and we didn't use it to generate
2475                          * anything, so it makes sense to check it
2476                          */
2477                         memcpy(page_address(tmp_page),
2478                                page_address(sh->dev[qd_idx].page),
2479                                STRIPE_SIZE);
2480                         compute_parity6(sh, UPDATE_PARITY);
2481                         if (memcmp(page_address(tmp_page),
2482                                    page_address(sh->dev[qd_idx].page),
2483                                    STRIPE_SIZE) != 0) {
2484                                 clear_bit(STRIPE_INSYNC, &sh->state);
2485                                 update_q = 1;
2486                         }
2487                 }
2488                 if (update_p || update_q) {
2489                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2490                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2491                                 /* don't try to repair!! */
2492                                 update_p = update_q = 0;
2493                 }
2494
2495                 /* now write out any block on a failed drive,
2496                  * or P or Q if they need it
2497                  */
2498
2499                 if (s->failed == 2) {
2500                         dev = &sh->dev[r6s->failed_num[1]];
2501                         s->locked++;
2502                         set_bit(R5_LOCKED, &dev->flags);
2503                         set_bit(R5_Wantwrite, &dev->flags);
2504                 }
2505                 if (s->failed >= 1) {
2506                         dev = &sh->dev[r6s->failed_num[0]];
2507                         s->locked++;
2508                         set_bit(R5_LOCKED, &dev->flags);
2509                         set_bit(R5_Wantwrite, &dev->flags);
2510                 }
2511
2512                 if (update_p) {
2513                         dev = &sh->dev[pd_idx];
2514                         s->locked++;
2515                         set_bit(R5_LOCKED, &dev->flags);
2516                         set_bit(R5_Wantwrite, &dev->flags);
2517                 }
2518                 if (update_q) {
2519                         dev = &sh->dev[qd_idx];
2520                         s->locked++;
2521                         set_bit(R5_LOCKED, &dev->flags);
2522                         set_bit(R5_Wantwrite, &dev->flags);
2523                 }
2524                 clear_bit(STRIPE_DEGRADED, &sh->state);
2525
2526                 set_bit(STRIPE_INSYNC, &sh->state);
2527         }
2528 }
2529
2530 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2531                                 struct r6_state *r6s)
2532 {
2533         int i;
2534
2535         /* We have read all the blocks in this stripe and now we need to
2536          * copy some of them into a target stripe for expand.
2537          */
2538         struct dma_async_tx_descriptor *tx = NULL;
2539         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2540         for (i = 0; i < sh->disks; i++)
2541                 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2542                         int dd_idx, pd_idx, j;
2543                         struct stripe_head *sh2;
2544
2545                         sector_t bn = compute_blocknr(sh, i);
2546                         sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2547                                                 conf->raid_disks -
2548                                                 conf->max_degraded, &dd_idx,
2549                                                 &pd_idx, conf);
2550                         sh2 = get_active_stripe(conf, s, conf->raid_disks,
2551                                                 pd_idx, 1);
2552                         if (sh2 == NULL)
2553                                 /* so far only the early blocks of this stripe
2554                                  * have been requested.  When later blocks
2555                                  * get requested, we will try again
2556                                  */
2557                                 continue;
2558                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2559                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2560                                 /* must have already done this block */
2561                                 release_stripe(sh2);
2562                                 continue;
2563                         }
2564
2565                         /* place all the copies on one channel */
2566                         tx = async_memcpy(sh2->dev[dd_idx].page,
2567                                 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2568                                 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2569
2570                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2571                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2572                         for (j = 0; j < conf->raid_disks; j++)
2573                                 if (j != sh2->pd_idx &&
2574                                     (!r6s || j != raid6_next_disk(sh2->pd_idx,
2575                                                                  sh2->disks)) &&
2576                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2577                                         break;
2578                         if (j == conf->raid_disks) {
2579                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2580                                 set_bit(STRIPE_HANDLE, &sh2->state);
2581                         }
2582                         release_stripe(sh2);
2583
2584                 }
2585         /* done submitting copies, wait for them to complete */
2586         if (tx) {
2587                 async_tx_ack(tx);
2588                 dma_wait_for_async_tx(tx);
2589         }
2590 }
2591
2592 /*
2593  * handle_stripe - do things to a stripe.
2594  *
2595  * We lock the stripe and then examine the state of various bits
2596  * to see what needs to be done.
2597  * Possible results:
2598  *    return some read request which now have data
2599  *    return some write requests which are safely on disc
2600  *    schedule a read on some buffers
2601  *    schedule a write of some buffers
2602  *    return confirmation of parity correctness
2603  *
2604  * buffers are taken off read_list or write_list, and bh_cache buffers
2605  * get BH_Lock set before the stripe lock is released.
2606  *
2607  */
2608
2609 static void handle_stripe5(struct stripe_head *sh)
2610 {
2611         raid5_conf_t *conf = sh->raid_conf;
2612         int disks = sh->disks, i;
2613         struct bio *return_bi = NULL;
2614         struct stripe_head_state s;
2615         struct r5dev *dev;
2616         unsigned long pending = 0;
2617
2618         memset(&s, 0, sizeof(s));
2619         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2620                 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2621                 atomic_read(&sh->count), sh->pd_idx,
2622                 sh->ops.pending, sh->ops.ack, sh->ops.complete);
2623
2624         spin_lock(&sh->lock);
2625         clear_bit(STRIPE_HANDLE, &sh->state);
2626         clear_bit(STRIPE_DELAYED, &sh->state);
2627
2628         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2629         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2630         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2631         /* Now to look around and see what can be done */
2632
2633         rcu_read_lock();
2634         for (i=disks; i--; ) {
2635                 mdk_rdev_t *rdev;
2636                 struct r5dev *dev = &sh->dev[i];
2637                 clear_bit(R5_Insync, &dev->flags);
2638
2639                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2640                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2641                         dev->towrite, dev->written);
2642
2643                 /* maybe we can request a biofill operation
2644                  *
2645                  * new wantfill requests are only permitted while
2646                  * STRIPE_OP_BIOFILL is clear
2647                  */
2648                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2649                         !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2650                         set_bit(R5_Wantfill, &dev->flags);
2651
2652                 /* now count some things */
2653                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2654                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2655                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2656
2657                 if (test_bit(R5_Wantfill, &dev->flags))
2658                         s.to_fill++;
2659                 else if (dev->toread)
2660                         s.to_read++;
2661                 if (dev->towrite) {
2662                         s.to_write++;
2663                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2664                                 s.non_overwrite++;
2665                 }
2666                 if (dev->written)
2667                         s.written++;
2668                 rdev = rcu_dereference(conf->disks[i].rdev);
2669                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2670                         /* The ReadError flag will just be confusing now */
2671                         clear_bit(R5_ReadError, &dev->flags);
2672                         clear_bit(R5_ReWrite, &dev->flags);
2673                 }
2674                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2675                     || test_bit(R5_ReadError, &dev->flags)) {
2676                         s.failed++;
2677                         s.failed_num = i;
2678                 } else
2679                         set_bit(R5_Insync, &dev->flags);
2680         }
2681         rcu_read_unlock();
2682
2683         if (s.to_fill && !test_and_set_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2684                 sh->ops.count++;
2685
2686         pr_debug("locked=%d uptodate=%d to_read=%d"
2687                 " to_write=%d failed=%d failed_num=%d\n",
2688                 s.locked, s.uptodate, s.to_read, s.to_write,
2689                 s.failed, s.failed_num);
2690         /* check if the array has lost two devices and, if so, some requests might
2691          * need to be failed
2692          */
2693         if (s.failed > 1 && s.to_read+s.to_write+s.written)
2694                 handle_requests_to_failed_array(conf, sh, &s, disks,
2695                                                 &return_bi);
2696         if (s.failed > 1 && s.syncing) {
2697                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2698                 clear_bit(STRIPE_SYNCING, &sh->state);
2699                 s.syncing = 0;
2700         }
2701
2702         /* might be able to return some write requests if the parity block
2703          * is safe, or on a failed drive
2704          */
2705         dev = &sh->dev[sh->pd_idx];
2706         if ( s.written &&
2707              ((test_bit(R5_Insync, &dev->flags) &&
2708                !test_bit(R5_LOCKED, &dev->flags) &&
2709                test_bit(R5_UPTODATE, &dev->flags)) ||
2710                (s.failed == 1 && s.failed_num == sh->pd_idx)))
2711                 handle_completed_write_requests(conf, sh, disks, &return_bi);
2712
2713         /* Now we might consider reading some blocks, either to check/generate
2714          * parity, or to satisfy requests
2715          * or to load a block that is being partially written.
2716          */
2717         if (s.to_read || s.non_overwrite ||
2718             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding ||
2719             test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2720                 handle_issuing_new_read_requests5(sh, &s, disks);
2721
2722         /* Now we check to see if any write operations have recently
2723          * completed
2724          */
2725
2726         /* leave prexor set until postxor is done, allows us to distinguish
2727          * a rmw from a rcw during biodrain
2728          */
2729         if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2730                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2731
2732                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2733                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2734                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2735
2736                 for (i = disks; i--; )
2737                         clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2738         }
2739
2740         /* if only POSTXOR is set then this is an 'expand' postxor */
2741         if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2742                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2743
2744                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2745                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2746                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2747
2748                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2749                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2750                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2751
2752                 /* All the 'written' buffers and the parity block are ready to
2753                  * be written back to disk
2754                  */
2755                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2756                 for (i = disks; i--; ) {
2757                         dev = &sh->dev[i];
2758                         if (test_bit(R5_LOCKED, &dev->flags) &&
2759                                 (i == sh->pd_idx || dev->written)) {
2760                                 pr_debug("Writing block %d\n", i);
2761                                 set_bit(R5_Wantwrite, &dev->flags);
2762                                 if (!test_and_set_bit(
2763                                     STRIPE_OP_IO, &sh->ops.pending))
2764                                         sh->ops.count++;
2765                                 if (!test_bit(R5_Insync, &dev->flags) ||
2766                                     (i == sh->pd_idx && s.failed == 0))
2767                                         set_bit(STRIPE_INSYNC, &sh->state);
2768                         }
2769                 }
2770                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2771                         atomic_dec(&conf->preread_active_stripes);
2772                         if (atomic_read(&conf->preread_active_stripes) <
2773                                 IO_THRESHOLD)
2774                                 md_wakeup_thread(conf->mddev->thread);
2775                 }
2776         }
2777
2778         /* Now to consider new write requests and what else, if anything
2779          * should be read.  We do not handle new writes when:
2780          * 1/ A 'write' operation (copy+xor) is already in flight.
2781          * 2/ A 'check' operation is in flight, as it may clobber the parity
2782          *    block.
2783          */
2784         if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2785                           !test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2786                 handle_issuing_new_write_requests5(conf, sh, &s, disks);
2787
2788         /* maybe we need to check and possibly fix the parity for this stripe
2789          * Any reads will already have been scheduled, so we just see if enough
2790          * data is available.  The parity check is held off while parity
2791          * dependent operations are in flight.
2792          */
2793         if ((s.syncing && s.locked == 0 &&
2794              !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2795              !test_bit(STRIPE_INSYNC, &sh->state)) ||
2796               test_bit(STRIPE_OP_CHECK, &sh->ops.pending) ||
2797               test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending))
2798                 handle_parity_checks5(conf, sh, &s, disks);
2799
2800         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2801                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2802                 clear_bit(STRIPE_SYNCING, &sh->state);
2803         }
2804
2805         /* If the failed drive is just a ReadError, then we might need to progress
2806          * the repair/check process
2807          */
2808         if (s.failed == 1 && !conf->mddev->ro &&
2809             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2810             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2811             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2812                 ) {
2813                 dev = &sh->dev[s.failed_num];
2814                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2815                         set_bit(R5_Wantwrite, &dev->flags);
2816                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2817                                 sh->ops.count++;
2818                         set_bit(R5_ReWrite, &dev->flags);
2819                         set_bit(R5_LOCKED, &dev->flags);
2820                         s.locked++;
2821                 } else {
2822                         /* let's read it back */
2823                         set_bit(R5_Wantread, &dev->flags);
2824                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2825                                 sh->ops.count++;
2826                         set_bit(R5_LOCKED, &dev->flags);
2827                         s.locked++;
2828                 }
2829         }
2830
2831         /* Finish postxor operations initiated by the expansion
2832          * process
2833          */
2834         if (test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete) &&
2835                 !test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending)) {
2836
2837                 clear_bit(STRIPE_EXPANDING, &sh->state);
2838
2839                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2840                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2841                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2842
2843                 for (i = conf->raid_disks; i--; ) {
2844                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2845                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2846                                 sh->ops.count++;
2847                 }
2848         }
2849
2850         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2851                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2852                 /* Need to write out all blocks after computing parity */
2853                 sh->disks = conf->raid_disks;
2854                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2855                         conf->raid_disks);
2856                 s.locked += handle_write_operations5(sh, 1, 1);
2857         } else if (s.expanded &&
2858                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2859                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2860                 atomic_dec(&conf->reshape_stripes);
2861                 wake_up(&conf->wait_for_overlap);
2862                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2863         }
2864
2865         if (s.expanding && s.locked == 0)
2866                 handle_stripe_expansion(conf, sh, NULL);
2867
2868         if (sh->ops.count)
2869                 pending = get_stripe_work(sh);
2870
2871         spin_unlock(&sh->lock);
2872
2873         if (pending)
2874                 raid5_run_ops(sh, pending);
2875
2876         return_io(return_bi);
2877
2878 }
2879
2880 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2881 {
2882         raid6_conf_t *conf = sh->raid_conf;
2883         int disks = sh->disks;
2884         struct bio *return_bi = NULL;
2885         int i, pd_idx = sh->pd_idx;
2886         struct stripe_head_state s;
2887         struct r6_state r6s;
2888         struct r5dev *dev, *pdev, *qdev;
2889
2890         r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2891         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2892                 "pd_idx=%d, qd_idx=%d\n",
2893                (unsigned long long)sh->sector, sh->state,
2894                atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2895         memset(&s, 0, sizeof(s));
2896
2897         spin_lock(&sh->lock);
2898         clear_bit(STRIPE_HANDLE, &sh->state);
2899         clear_bit(STRIPE_DELAYED, &sh->state);
2900
2901         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2902         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2903         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2904         /* Now to look around and see what can be done */
2905
2906         rcu_read_lock();
2907         for (i=disks; i--; ) {
2908                 mdk_rdev_t *rdev;
2909                 dev = &sh->dev[i];
2910                 clear_bit(R5_Insync, &dev->flags);
2911
2912                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2913                         i, dev->flags, dev->toread, dev->towrite, dev->written);
2914                 /* maybe we can reply to a read */
2915                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2916                         struct bio *rbi, *rbi2;
2917                         pr_debug("Return read for disc %d\n", i);
2918                         spin_lock_irq(&conf->device_lock);
2919                         rbi = dev->toread;
2920                         dev->toread = NULL;
2921                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2922                                 wake_up(&conf->wait_for_overlap);
2923                         spin_unlock_irq(&conf->device_lock);
2924                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2925                                 copy_data(0, rbi, dev->page, dev->sector);
2926                                 rbi2 = r5_next_bio(rbi, dev->sector);
2927                                 spin_lock_irq(&conf->device_lock);
2928                                 if (--rbi->bi_phys_segments == 0) {
2929                                         rbi->bi_next = return_bi;
2930                                         return_bi = rbi;
2931                                 }
2932                                 spin_unlock_irq(&conf->device_lock);
2933                                 rbi = rbi2;
2934                         }
2935                 }
2936
2937                 /* now count some things */
2938                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2939                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2940
2941
2942                 if (dev->toread)
2943                         s.to_read++;
2944                 if (dev->towrite) {
2945                         s.to_write++;
2946                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2947                                 s.non_overwrite++;
2948                 }
2949                 if (dev->written)
2950                         s.written++;
2951                 rdev = rcu_dereference(conf->disks[i].rdev);
2952                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2953                         /* The ReadError flag will just be confusing now */
2954                         clear_bit(R5_ReadError, &dev->flags);
2955                         clear_bit(R5_ReWrite, &dev->flags);
2956                 }
2957                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2958                     || test_bit(R5_ReadError, &dev->flags)) {
2959                         if (s.failed < 2)
2960                                 r6s.failed_num[s.failed] = i;
2961                         s.failed++;
2962                 } else
2963                         set_bit(R5_Insync, &dev->flags);
2964         }
2965         rcu_read_unlock();
2966         pr_debug("locked=%d uptodate=%d to_read=%d"
2967                " to_write=%d failed=%d failed_num=%d,%d\n",
2968                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
2969                r6s.failed_num[0], r6s.failed_num[1]);
2970         /* check if the array has lost >2 devices and, if so, some requests
2971          * might need to be failed
2972          */
2973         if (s.failed > 2 && s.to_read+s.to_write+s.written)
2974                 handle_requests_to_failed_array(conf, sh, &s, disks,
2975                                                 &return_bi);
2976         if (s.failed > 2 && s.syncing) {
2977                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2978                 clear_bit(STRIPE_SYNCING, &sh->state);
2979                 s.syncing = 0;
2980         }
2981
2982         /*
2983          * might be able to return some write requests if the parity blocks
2984          * are safe, or on a failed drive
2985          */
2986         pdev = &sh->dev[pd_idx];
2987         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
2988                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
2989         qdev = &sh->dev[r6s.qd_idx];
2990         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
2991                 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
2992
2993         if ( s.written &&
2994              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
2995                              && !test_bit(R5_LOCKED, &pdev->flags)
2996                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
2997              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
2998                              && !test_bit(R5_LOCKED, &qdev->flags)
2999                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3000                 handle_completed_write_requests(conf, sh, disks, &return_bi);
3001
3002         /* Now we might consider reading some blocks, either to check/generate
3003          * parity, or to satisfy requests
3004          * or to load a block that is being partially written.
3005          */
3006         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3007             (s.syncing && (s.uptodate < disks)) || s.expanding)
3008                 handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
3009
3010         /* now to consider writing and what else, if anything should be read */
3011         if (s.to_write)
3012                 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
3013
3014         /* maybe we need to check and possibly fix the parity for this stripe
3015          * Any reads will already have been scheduled, so we just see if enough
3016          * data is available
3017          */
3018         if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3019                 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3020
3021         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3022                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3023                 clear_bit(STRIPE_SYNCING, &sh->state);
3024         }
3025
3026         /* If the failed drives are just a ReadError, then we might need
3027          * to progress the repair/check process
3028          */
3029         if (s.failed <= 2 && !conf->mddev->ro)
3030                 for (i = 0; i < s.failed; i++) {
3031                         dev = &sh->dev[r6s.failed_num[i]];
3032                         if (test_bit(R5_ReadError, &dev->flags)
3033                             && !test_bit(R5_LOCKED, &dev->flags)
3034                             && test_bit(R5_UPTODATE, &dev->flags)
3035                                 ) {
3036                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3037                                         set_bit(R5_Wantwrite, &dev->flags);
3038                                         set_bit(R5_ReWrite, &dev->flags);
3039                                         set_bit(R5_LOCKED, &dev->flags);
3040                                 } else {
3041                                         /* let's read it back */
3042                                         set_bit(R5_Wantread, &dev->flags);
3043                                         set_bit(R5_LOCKED, &dev->flags);
3044                                 }
3045                         }
3046                 }
3047
3048         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3049                 /* Need to write out all blocks after computing P&Q */
3050                 sh->disks = conf->raid_disks;
3051                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
3052                                              conf->raid_disks);
3053                 compute_parity6(sh, RECONSTRUCT_WRITE);
3054                 for (i = conf->raid_disks ; i-- ;  ) {
3055                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3056                         s.locked++;
3057                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3058                 }
3059                 clear_bit(STRIPE_EXPANDING, &sh->state);
3060         } else if (s.expanded) {
3061                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3062                 atomic_dec(&conf->reshape_stripes);
3063                 wake_up(&conf->wait_for_overlap);
3064                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3065         }
3066
3067         if (s.expanding && s.locked == 0)
3068                 handle_stripe_expansion(conf, sh, &r6s);
3069
3070         spin_unlock(&sh->lock);
3071
3072         return_io(return_bi);
3073
3074         for (i=disks; i-- ;) {
3075                 int rw;
3076                 struct bio *bi;
3077                 mdk_rdev_t *rdev;
3078                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
3079                         rw = WRITE;
3080                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
3081                         rw = READ;
3082                 else
3083                         continue;
3084
3085                 bi = &sh->dev[i].req;
3086
3087                 bi->bi_rw = rw;
3088                 if (rw == WRITE)
3089                         bi->bi_end_io = raid5_end_write_request;
3090                 else
3091                         bi->bi_end_io = raid5_end_read_request;
3092
3093                 rcu_read_lock();
3094                 rdev = rcu_dereference(conf->disks[i].rdev);
3095                 if (rdev && test_bit(Faulty, &rdev->flags))
3096                         rdev = NULL;
3097                 if (rdev)
3098                         atomic_inc(&rdev->nr_pending);
3099                 rcu_read_unlock();
3100
3101                 if (rdev) {
3102                         if (s.syncing || s.expanding || s.expanded)
3103                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
3104
3105                         bi->bi_bdev = rdev->bdev;
3106                         pr_debug("for %llu schedule op %ld on disc %d\n",
3107                                 (unsigned long long)sh->sector, bi->bi_rw, i);
3108                         atomic_inc(&sh->count);
3109                         bi->bi_sector = sh->sector + rdev->data_offset;
3110                         bi->bi_flags = 1 << BIO_UPTODATE;
3111                         bi->bi_vcnt = 1;
3112                         bi->bi_max_vecs = 1;
3113                         bi->bi_idx = 0;
3114                         bi->bi_io_vec = &sh->dev[i].vec;
3115                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
3116                         bi->bi_io_vec[0].bv_offset = 0;
3117                         bi->bi_size = STRIPE_SIZE;
3118                         bi->bi_next = NULL;
3119                         if (rw == WRITE &&
3120                             test_bit(R5_ReWrite, &sh->dev[i].flags))
3121                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
3122                         generic_make_request(bi);
3123                 } else {
3124                         if (rw == WRITE)
3125                                 set_bit(STRIPE_DEGRADED, &sh->state);
3126                         pr_debug("skip op %ld on disc %d for sector %llu\n",
3127                                 bi->bi_rw, i, (unsigned long long)sh->sector);
3128                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
3129                         set_bit(STRIPE_HANDLE, &sh->state);
3130                 }
3131         }
3132 }
3133
3134 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3135 {
3136         if (sh->raid_conf->level == 6)
3137                 handle_stripe6(sh, tmp_page);
3138         else
3139                 handle_stripe5(sh);
3140 }
3141
3142
3143
3144 static void raid5_activate_delayed(raid5_conf_t *conf)
3145 {
3146         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3147                 while (!list_empty(&conf->delayed_list)) {
3148                         struct list_head *l = conf->delayed_list.next;
3149                         struct stripe_head *sh;
3150                         sh = list_entry(l, struct stripe_head, lru);
3151                         list_del_init(l);
3152                         clear_bit(STRIPE_DELAYED, &sh->state);
3153                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3154                                 atomic_inc(&conf->preread_active_stripes);
3155                         list_add_tail(&sh->lru, &conf->handle_list);
3156                 }
3157         }
3158 }
3159
3160 static void activate_bit_delay(raid5_conf_t *conf)
3161 {
3162         /* device_lock is held */
3163         struct list_head head;
3164         list_add(&head, &conf->bitmap_list);
3165         list_del_init(&conf->bitmap_list);
3166         while (!list_empty(&head)) {
3167                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3168                 list_del_init(&sh->lru);
3169                 atomic_inc(&sh->count);
3170                 __release_stripe(conf, sh);
3171         }
3172 }
3173
3174 static void unplug_slaves(mddev_t *mddev)
3175 {
3176         raid5_conf_t *conf = mddev_to_conf(mddev);
3177         int i;
3178
3179         rcu_read_lock();
3180         for (i=0; i<mddev->raid_disks; i++) {
3181                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3182                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3183                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3184
3185                         atomic_inc(&rdev->nr_pending);
3186                         rcu_read_unlock();
3187
3188                         if (r_queue->unplug_fn)
3189                                 r_queue->unplug_fn(r_queue);
3190
3191                         rdev_dec_pending(rdev, mddev);
3192                         rcu_read_lock();
3193                 }
3194         }
3195         rcu_read_unlock();
3196 }
3197
3198 static void raid5_unplug_device(struct request_queue *q)
3199 {
3200         mddev_t *mddev = q->queuedata;
3201         raid5_conf_t *conf = mddev_to_conf(mddev);
3202         unsigned long flags;
3203
3204         spin_lock_irqsave(&conf->device_lock, flags);
3205
3206         if (blk_remove_plug(q)) {
3207                 conf->seq_flush++;
3208                 raid5_activate_delayed(conf);
3209         }
3210         md_wakeup_thread(mddev->thread);
3211
3212         spin_unlock_irqrestore(&conf->device_lock, flags);
3213
3214         unplug_slaves(mddev);
3215 }
3216
3217 static int raid5_issue_flush(struct request_queue *q, struct gendisk *disk,
3218                              sector_t *error_sector)
3219 {
3220         mddev_t *mddev = q->queuedata;
3221         raid5_conf_t *conf = mddev_to_conf(mddev);
3222         int i, ret = 0;
3223
3224         rcu_read_lock();
3225         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
3226                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3227                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
3228                         struct block_device *bdev = rdev->bdev;
3229                         struct request_queue *r_queue = bdev_get_queue(bdev);
3230
3231                         if (!r_queue->issue_flush_fn)
3232                                 ret = -EOPNOTSUPP;
3233                         else {
3234                                 atomic_inc(&rdev->nr_pending);
3235                                 rcu_read_unlock();
3236                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
3237                                                               error_sector);
3238                                 rdev_dec_pending(rdev, mddev);
3239                                 rcu_read_lock();
3240                         }
3241                 }
3242         }
3243         rcu_read_unlock();
3244         return ret;
3245 }
3246
3247 static int raid5_congested(void *data, int bits)
3248 {
3249         mddev_t *mddev = data;
3250         raid5_conf_t *conf = mddev_to_conf(mddev);
3251
3252         /* No difference between reads and writes.  Just check
3253          * how busy the stripe_cache is
3254          */
3255         if (conf->inactive_blocked)
3256                 return 1;
3257         if (conf->quiesce)
3258                 return 1;
3259         if (list_empty_careful(&conf->inactive_list))
3260                 return 1;
3261
3262         return 0;
3263 }
3264
3265 /* We want read requests to align with chunks where possible,
3266  * but write requests don't need to.
3267  */
3268 static int raid5_mergeable_bvec(struct request_queue *q, struct bio *bio, struct bio_vec *biovec)
3269 {
3270         mddev_t *mddev = q->queuedata;
3271         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3272         int max;
3273         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3274         unsigned int bio_sectors = bio->bi_size >> 9;
3275
3276         if (bio_data_dir(bio) == WRITE)
3277                 return biovec->bv_len; /* always allow writes to be mergeable */
3278
3279         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3280         if (max < 0) max = 0;
3281         if (max <= biovec->bv_len && bio_sectors == 0)
3282                 return biovec->bv_len;
3283         else
3284                 return max;
3285 }
3286
3287
3288 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3289 {
3290         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3291         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3292         unsigned int bio_sectors = bio->bi_size >> 9;
3293
3294         return  chunk_sectors >=
3295                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3296 }
3297
3298 /*
3299  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3300  *  later sampled by raid5d.
3301  */
3302 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3303 {
3304         unsigned long flags;
3305
3306         spin_lock_irqsave(&conf->device_lock, flags);
3307
3308         bi->bi_next = conf->retry_read_aligned_list;
3309         conf->retry_read_aligned_list = bi;
3310
3311         spin_unlock_irqrestore(&conf->device_lock, flags);
3312         md_wakeup_thread(conf->mddev->thread);
3313 }
3314
3315
3316 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3317 {
3318         struct bio *bi;
3319
3320         bi = conf->retry_read_aligned;
3321         if (bi) {
3322                 conf->retry_read_aligned = NULL;
3323                 return bi;
3324         }
3325         bi = conf->retry_read_aligned_list;
3326         if(bi) {
3327                 conf->retry_read_aligned_list = bi->bi_next;
3328                 bi->bi_next = NULL;
3329                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3330                 bi->bi_hw_segments = 0; /* count of processed stripes */
3331         }
3332
3333         return bi;
3334 }
3335
3336
3337 /*
3338  *  The "raid5_align_endio" should check if the read succeeded and if it
3339  *  did, call bio_endio on the original bio (having bio_put the new bio
3340  *  first).
3341  *  If the read failed..
3342  */
3343 static int raid5_align_endio(struct bio *bi, unsigned int bytes, int error)
3344 {
3345         struct bio* raid_bi  = bi->bi_private;
3346         mddev_t *mddev;
3347         raid5_conf_t *conf;
3348         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3349         mdk_rdev_t *rdev;
3350
3351         if (bi->bi_size)
3352                 return 1;
3353         bio_put(bi);
3354
3355         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3356         conf = mddev_to_conf(mddev);
3357         rdev = (void*)raid_bi->bi_next;
3358         raid_bi->bi_next = NULL;
3359
3360         rdev_dec_pending(rdev, conf->mddev);
3361
3362         if (!error && uptodate) {
3363                 bio_endio(raid_bi, bytes, 0);
3364                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3365                         wake_up(&conf->wait_for_stripe);
3366                 return 0;
3367         }
3368
3369
3370         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3371
3372         add_bio_to_retry(raid_bi, conf);
3373         return 0;
3374 }
3375
3376 static int bio_fits_rdev(struct bio *bi)
3377 {
3378         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3379
3380         if ((bi->bi_size>>9) > q->max_sectors)
3381                 return 0;
3382         blk_recount_segments(q, bi);
3383         if (bi->bi_phys_segments > q->max_phys_segments ||
3384             bi->bi_hw_segments > q->max_hw_segments)
3385                 return 0;
3386
3387         if (q->merge_bvec_fn)
3388                 /* it's too hard to apply the merge_bvec_fn at this stage,
3389                  * just just give up
3390                  */
3391                 return 0;
3392
3393         return 1;
3394 }
3395
3396
3397 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3398 {
3399         mddev_t *mddev = q->queuedata;
3400         raid5_conf_t *conf = mddev_to_conf(mddev);
3401         const unsigned int raid_disks = conf->raid_disks;
3402         const unsigned int data_disks = raid_disks - conf->max_degraded;
3403         unsigned int dd_idx, pd_idx;
3404         struct bio* align_bi;
3405         mdk_rdev_t *rdev;
3406
3407         if (!in_chunk_boundary(mddev, raid_bio)) {
3408                 pr_debug("chunk_aligned_read : non aligned\n");
3409                 return 0;
3410         }
3411         /*
3412          * use bio_clone to make a copy of the bio
3413          */
3414         align_bi = bio_clone(raid_bio, GFP_NOIO);
3415         if (!align_bi)
3416                 return 0;
3417         /*
3418          *   set bi_end_io to a new function, and set bi_private to the
3419          *     original bio.
3420          */
3421         align_bi->bi_end_io  = raid5_align_endio;
3422         align_bi->bi_private = raid_bio;
3423         /*
3424          *      compute position
3425          */
3426         align_bi->bi_sector =  raid5_compute_sector(raid_bio->bi_sector,
3427                                         raid_disks,
3428                                         data_disks,
3429                                         &dd_idx,
3430                                         &pd_idx,
3431                                         conf);
3432
3433         rcu_read_lock();
3434         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3435         if (rdev && test_bit(In_sync, &rdev->flags)) {
3436                 atomic_inc(&rdev->nr_pending);
3437                 rcu_read_unlock();
3438                 raid_bio->bi_next = (void*)rdev;
3439                 align_bi->bi_bdev =  rdev->bdev;
3440                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3441                 align_bi->bi_sector += rdev->data_offset;
3442
3443                 if (!bio_fits_rdev(align_bi)) {
3444                         /* too big in some way */
3445                         bio_put(align_bi);
3446                         rdev_dec_pending(rdev, mddev);
3447                         return 0;
3448                 }
3449
3450                 spin_lock_irq(&conf->device_lock);
3451                 wait_event_lock_irq(conf->wait_for_stripe,
3452                                     conf->quiesce == 0,
3453                                     conf->device_lock, /* nothing */);
3454                 atomic_inc(&conf->active_aligned_reads);
3455                 spin_unlock_irq(&conf->device_lock);
3456
3457                 generic_make_request(align_bi);
3458                 return 1;
3459         } else {
3460                 rcu_read_unlock();
3461                 bio_put(align_bi);
3462                 return 0;
3463         }
3464 }
3465
3466
3467 static int make_request(struct request_queue *q, struct bio * bi)
3468 {
3469         mddev_t *mddev = q->queuedata;
3470         raid5_conf_t *conf = mddev_to_conf(mddev);
3471         unsigned int dd_idx, pd_idx;
3472         sector_t new_sector;
3473         sector_t logical_sector, last_sector;
3474         struct stripe_head *sh;
3475         const int rw = bio_data_dir(bi);
3476         int remaining;
3477
3478         if (unlikely(bio_barrier(bi))) {
3479                 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
3480                 return 0;
3481         }
3482
3483         md_write_start(mddev, bi);
3484
3485         disk_stat_inc(mddev->gendisk, ios[rw]);
3486         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3487
3488         if (rw == READ &&
3489              mddev->reshape_position == MaxSector &&
3490              chunk_aligned_read(q,bi))
3491                 return 0;
3492
3493         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3494         last_sector = bi->bi_sector + (bi->bi_size>>9);
3495         bi->bi_next = NULL;
3496         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3497
3498         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3499                 DEFINE_WAIT(w);
3500                 int disks, data_disks;
3501
3502         retry:
3503                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3504                 if (likely(conf->expand_progress == MaxSector))
3505                         disks = conf->raid_disks;
3506                 else {
3507                         /* spinlock is needed as expand_progress may be
3508                          * 64bit on a 32bit platform, and so it might be
3509                          * possible to see a half-updated value
3510                          * Ofcourse expand_progress could change after
3511                          * the lock is dropped, so once we get a reference
3512                          * to the stripe that we think it is, we will have
3513                          * to check again.
3514                          */
3515                         spin_lock_irq(&conf->device_lock);
3516                         disks = conf->raid_disks;
3517                         if (logical_sector >= conf->expand_progress)
3518                                 disks = conf->previous_raid_disks;
3519                         else {
3520                                 if (logical_sector >= conf->expand_lo) {
3521                                         spin_unlock_irq(&conf->device_lock);
3522                                         schedule();
3523                                         goto retry;
3524                                 }
3525                         }
3526                         spin_unlock_irq(&conf->device_lock);
3527                 }
3528                 data_disks = disks - conf->max_degraded;
3529
3530                 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3531                                                   &dd_idx, &pd_idx, conf);
3532                 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3533                         (unsigned long long)new_sector, 
3534                         (unsigned long long)logical_sector);
3535
3536                 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3537                 if (sh) {
3538                         if (unlikely(conf->expand_progress != MaxSector)) {
3539                                 /* expansion might have moved on while waiting for a
3540                                  * stripe, so we must do the range check again.
3541                                  * Expansion could still move past after this
3542                                  * test, but as we are holding a reference to
3543                                  * 'sh', we know that if that happens,
3544                                  *  STRIPE_EXPANDING will get set and the expansion
3545                                  * won't proceed until we finish with the stripe.
3546                                  */
3547                                 int must_retry = 0;
3548                                 spin_lock_irq(&conf->device_lock);
3549                                 if (logical_sector <  conf->expand_progress &&
3550                                     disks == conf->previous_raid_disks)
3551                                         /* mismatch, need to try again */
3552                                         must_retry = 1;
3553                                 spin_unlock_irq(&conf->device_lock);
3554                                 if (must_retry) {
3555                                         release_stripe(sh);
3556                                         goto retry;
3557                                 }
3558                         }
3559                         /* FIXME what if we get a false positive because these
3560                          * are being updated.
3561                          */
3562                         if (logical_sector >= mddev->suspend_lo &&
3563                             logical_sector < mddev->suspend_hi) {
3564                                 release_stripe(sh);
3565                                 schedule();
3566                                 goto retry;
3567                         }
3568
3569                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3570                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3571                                 /* Stripe is busy expanding or
3572                                  * add failed due to overlap.  Flush everything
3573                                  * and wait a while
3574                                  */
3575                                 raid5_unplug_device(mddev->queue);
3576                                 release_stripe(sh);
3577                                 schedule();
3578                                 goto retry;
3579                         }
3580                         finish_wait(&conf->wait_for_overlap, &w);
3581                         handle_stripe(sh, NULL);
3582                         release_stripe(sh);
3583                 } else {
3584                         /* cannot get stripe for read-ahead, just give-up */
3585                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3586                         finish_wait(&conf->wait_for_overlap, &w);
3587                         break;
3588                 }
3589                         
3590         }
3591         spin_lock_irq(&conf->device_lock);
3592         remaining = --bi->bi_phys_segments;
3593         spin_unlock_irq(&conf->device_lock);
3594         if (remaining == 0) {
3595                 int bytes = bi->bi_size;
3596
3597                 if ( rw == WRITE )
3598                         md_write_end(mddev);
3599                 bi->bi_size = 0;
3600                 bi->bi_end_io(bi, bytes,
3601                               test_bit(BIO_UPTODATE, &bi->bi_flags)
3602                                 ? 0 : -EIO);
3603         }
3604         return 0;
3605 }
3606
3607 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3608 {
3609         /* reshaping is quite different to recovery/resync so it is
3610          * handled quite separately ... here.
3611          *
3612          * On each call to sync_request, we gather one chunk worth of
3613          * destination stripes and flag them as expanding.
3614          * Then we find all the source stripes and request reads.
3615          * As the reads complete, handle_stripe will copy the data
3616          * into the destination stripe and release that stripe.
3617          */
3618         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3619         struct stripe_head *sh;
3620         int pd_idx;
3621         sector_t first_sector, last_sector;
3622         int raid_disks = conf->previous_raid_disks;
3623         int data_disks = raid_disks - conf->max_degraded;
3624         int new_data_disks = conf->raid_disks - conf->max_degraded;
3625         int i;
3626         int dd_idx;
3627         sector_t writepos, safepos, gap;
3628
3629         if (sector_nr == 0 &&
3630             conf->expand_progress != 0) {
3631                 /* restarting in the middle, skip the initial sectors */
3632                 sector_nr = conf->expand_progress;
3633                 sector_div(sector_nr, new_data_disks);
3634                 *skipped = 1;
3635                 return sector_nr;
3636         }
3637
3638         /* we update the metadata when there is more than 3Meg
3639          * in the block range (that is rather arbitrary, should
3640          * probably be time based) or when the data about to be
3641          * copied would over-write the source of the data at
3642          * the front of the range.
3643          * i.e. one new_stripe forward from expand_progress new_maps
3644          * to after where expand_lo old_maps to
3645          */
3646         writepos = conf->expand_progress +
3647                 conf->chunk_size/512*(new_data_disks);
3648         sector_div(writepos, new_data_disks);
3649         safepos = conf->expand_lo;
3650         sector_div(safepos, data_disks);
3651         gap = conf->expand_progress - conf->expand_lo;
3652
3653         if (writepos >= safepos ||
3654             gap > (new_data_disks)*3000*2 /*3Meg*/) {
3655                 /* Cannot proceed until we've updated the superblock... */
3656                 wait_event(conf->wait_for_overlap,
3657                            atomic_read(&conf->reshape_stripes)==0);
3658                 mddev->reshape_position = conf->expand_progress;
3659                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3660                 md_wakeup_thread(mddev->thread);
3661                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3662                            kthread_should_stop());
3663                 spin_lock_irq(&conf->device_lock);
3664                 conf->expand_lo = mddev->reshape_position;
3665                 spin_unlock_irq(&conf->device_lock);
3666                 wake_up(&conf->wait_for_overlap);
3667         }
3668
3669         for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3670                 int j;
3671                 int skipped = 0;
3672                 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3673                 sh = get_active_stripe(conf, sector_nr+i,
3674                                        conf->raid_disks, pd_idx, 0);
3675                 set_bit(STRIPE_EXPANDING, &sh->state);
3676                 atomic_inc(&conf->reshape_stripes);
3677                 /* If any of this stripe is beyond the end of the old
3678                  * array, then we need to zero those blocks
3679                  */
3680                 for (j=sh->disks; j--;) {
3681                         sector_t s;
3682                         if (j == sh->pd_idx)
3683                                 continue;
3684                         if (conf->level == 6 &&
3685                             j == raid6_next_disk(sh->pd_idx, sh->disks))
3686                                 continue;
3687                         s = compute_blocknr(sh, j);
3688                         if (s < (mddev->array_size<<1)) {
3689                                 skipped = 1;
3690                                 continue;
3691                         }
3692                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3693                         set_bit(R5_Expanded, &sh->dev[j].flags);
3694                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
3695                 }
3696                 if (!skipped) {
3697                         set_bit(STRIPE_EXPAND_READY, &sh->state);
3698                         set_bit(STRIPE_HANDLE, &sh->state);
3699                 }
3700                 release_stripe(sh);
3701         }
3702         spin_lock_irq(&conf->device_lock);
3703         conf->expand_progress = (sector_nr + i) * new_data_disks;
3704         spin_unlock_irq(&conf->device_lock);
3705         /* Ok, those stripe are ready. We can start scheduling
3706          * reads on the source stripes.
3707          * The source stripes are determined by mapping the first and last
3708          * block on the destination stripes.
3709          */
3710         first_sector =
3711                 raid5_compute_sector(sector_nr*(new_data_disks),
3712                                      raid_disks, data_disks,
3713                                      &dd_idx, &pd_idx, conf);
3714         last_sector =
3715                 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3716                                      *(new_data_disks) -1,
3717                                      raid_disks, data_disks,
3718                                      &dd_idx, &pd_idx, conf);
3719         if (last_sector >= (mddev->size<<1))
3720                 last_sector = (mddev->size<<1)-1;
3721         while (first_sector <= last_sector) {
3722                 pd_idx = stripe_to_pdidx(first_sector, conf,
3723                                          conf->previous_raid_disks);
3724                 sh = get_active_stripe(conf, first_sector,
3725                                        conf->previous_raid_disks, pd_idx, 0);
3726                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3727                 set_bit(STRIPE_HANDLE, &sh->state);
3728                 release_stripe(sh);
3729                 first_sector += STRIPE_SECTORS;
3730         }
3731         return conf->chunk_size>>9;
3732 }
3733
3734 /* FIXME go_faster isn't used */
3735 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3736 {
3737         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3738         struct stripe_head *sh;
3739         int pd_idx;
3740         int raid_disks = conf->raid_disks;
3741         sector_t max_sector = mddev->size << 1;
3742         int sync_blocks;
3743         int still_degraded = 0;
3744         int i;
3745
3746         if (sector_nr >= max_sector) {
3747                 /* just being told to finish up .. nothing much to do */
3748                 unplug_slaves(mddev);
3749                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3750                         end_reshape(conf);
3751                         return 0;
3752                 }
3753
3754                 if (mddev->curr_resync < max_sector) /* aborted */
3755                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3756                                         &sync_blocks, 1);
3757                 else /* completed sync */
3758                         conf->fullsync = 0;
3759                 bitmap_close_sync(mddev->bitmap);
3760
3761                 return 0;
3762         }
3763
3764         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3765                 return reshape_request(mddev, sector_nr, skipped);
3766
3767         /* if there is too many failed drives and we are trying
3768          * to resync, then assert that we are finished, because there is
3769          * nothing we can do.
3770          */
3771         if (mddev->degraded >= conf->max_degraded &&
3772             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3773                 sector_t rv = (mddev->size << 1) - sector_nr;
3774                 *skipped = 1;
3775                 return rv;
3776         }
3777         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3778             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3779             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3780                 /* we can skip this block, and probably more */
3781                 sync_blocks /= STRIPE_SECTORS;
3782                 *skipped = 1;
3783                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3784         }
3785
3786         pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3787         sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3788         if (sh == NULL) {
3789                 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3790                 /* make sure we don't swamp the stripe cache if someone else
3791                  * is trying to get access
3792                  */
3793                 schedule_timeout_uninterruptible(1);
3794         }
3795         /* Need to check if array will still be degraded after recovery/resync
3796          * We don't need to check the 'failed' flag as when that gets set,
3797          * recovery aborts.
3798          */
3799         for (i=0; i<mddev->raid_disks; i++)
3800                 if (conf->disks[i].rdev == NULL)
3801                         still_degraded = 1;
3802
3803         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3804
3805         spin_lock(&sh->lock);
3806         set_bit(STRIPE_SYNCING, &sh->state);
3807         clear_bit(STRIPE_INSYNC, &sh->state);
3808         spin_unlock(&sh->lock);
3809
3810         handle_stripe(sh, NULL);
3811         release_stripe(sh);
3812
3813         return STRIPE_SECTORS;
3814 }
3815
3816 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3817 {
3818         /* We may not be able to submit a whole bio at once as there
3819          * may not be enough stripe_heads available.
3820          * We cannot pre-allocate enough stripe_heads as we may need
3821          * more than exist in the cache (if we allow ever large chunks).
3822          * So we do one stripe head at a time and record in
3823          * ->bi_hw_segments how many have been done.
3824          *
3825          * We *know* that this entire raid_bio is in one chunk, so
3826          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3827          */
3828         struct stripe_head *sh;
3829         int dd_idx, pd_idx;
3830         sector_t sector, logical_sector, last_sector;
3831         int scnt = 0;
3832         int remaining;
3833         int handled = 0;
3834
3835         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3836         sector = raid5_compute_sector(  logical_sector,
3837                                         conf->raid_disks,
3838                                         conf->raid_disks - conf->max_degraded,
3839                                         &dd_idx,
3840                                         &pd_idx,
3841                                         conf);
3842         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3843
3844         for (; logical_sector < last_sector;
3845              logical_sector += STRIPE_SECTORS,
3846                      sector += STRIPE_SECTORS,
3847                      scnt++) {
3848
3849                 if (scnt < raid_bio->bi_hw_segments)
3850                         /* already done this stripe */
3851                         continue;
3852
3853                 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3854
3855                 if (!sh) {
3856                         /* failed to get a stripe - must wait */
3857                         raid_bio->bi_hw_segments = scnt;
3858                         conf->retry_read_aligned = raid_bio;
3859                         return handled;
3860                 }
3861
3862                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3863                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3864                         release_stripe(sh);
3865                         raid_bio->bi_hw_segments = scnt;
3866                         conf->retry_read_aligned = raid_bio;
3867                         return handled;
3868                 }
3869
3870                 handle_stripe(sh, NULL);
3871                 release_stripe(sh);
3872                 handled++;
3873         }
3874         spin_lock_irq(&conf->device_lock);
3875         remaining = --raid_bio->bi_phys_segments;
3876         spin_unlock_irq(&conf->device_lock);
3877         if (remaining == 0) {
3878                 int bytes = raid_bio->bi_size;
3879
3880                 raid_bio->bi_size = 0;
3881                 raid_bio->bi_end_io(raid_bio, bytes,
3882                               test_bit(BIO_UPTODATE, &raid_bio->bi_flags)
3883                                 ? 0 : -EIO);
3884         }
3885         if (atomic_dec_and_test(&conf->active_aligned_reads))
3886                 wake_up(&conf->wait_for_stripe);
3887         return handled;
3888 }
3889
3890
3891
3892 /*
3893  * This is our raid5 kernel thread.
3894  *
3895  * We scan the hash table for stripes which can be handled now.
3896  * During the scan, completed stripes are saved for us by the interrupt
3897  * handler, so that they will not have to wait for our next wakeup.
3898  */
3899 static void raid5d (mddev_t *mddev)
3900 {
3901         struct stripe_head *sh;
3902         raid5_conf_t *conf = mddev_to_conf(mddev);
3903         int handled;
3904
3905         pr_debug("+++ raid5d active\n");
3906
3907         md_check_recovery(mddev);
3908
3909         handled = 0;
3910         spin_lock_irq(&conf->device_lock);
3911         while (1) {
3912                 struct list_head *first;
3913                 struct bio *bio;
3914
3915                 if (conf->seq_flush != conf->seq_write) {
3916                         int seq = conf->seq_flush;
3917                         spin_unlock_irq(&conf->device_lock);
3918                         bitmap_unplug(mddev->bitmap);
3919                         spin_lock_irq(&conf->device_lock);
3920                         conf->seq_write = seq;
3921                         activate_bit_delay(conf);
3922                 }
3923
3924                 if (list_empty(&conf->handle_list) &&
3925                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
3926                     !blk_queue_plugged(mddev->queue) &&
3927                     !list_empty(&conf->delayed_list))
3928                         raid5_activate_delayed(conf);
3929
3930                 while ((bio = remove_bio_from_retry(conf))) {
3931                         int ok;
3932                         spin_unlock_irq(&conf->device_lock);
3933                         ok = retry_aligned_read(conf, bio);
3934                         spin_lock_irq(&conf->device_lock);
3935                         if (!ok)
3936                                 break;
3937                         handled++;
3938                 }
3939
3940                 if (list_empty(&conf->handle_list)) {
3941                         async_tx_issue_pending_all();
3942                         break;
3943                 }
3944
3945                 first = conf->handle_list.next;
3946                 sh = list_entry(first, struct stripe_head, lru);
3947
3948                 list_del_init(first);
3949                 atomic_inc(&sh->count);
3950                 BUG_ON(atomic_read(&sh->count)!= 1);
3951                 spin_unlock_irq(&conf->device_lock);
3952                 
3953                 handled++;
3954                 handle_stripe(sh, conf->spare_page);
3955                 release_stripe(sh);
3956
3957                 spin_lock_irq(&conf->device_lock);
3958         }
3959         pr_debug("%d stripes handled\n", handled);
3960
3961         spin_unlock_irq(&conf->device_lock);
3962
3963         unplug_slaves(mddev);
3964
3965         pr_debug("--- raid5d inactive\n");
3966 }
3967
3968 static ssize_t
3969 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
3970 {
3971         raid5_conf_t *conf = mddev_to_conf(mddev);
3972         if (conf)
3973                 return sprintf(page, "%d\n", conf->max_nr_stripes);
3974         else
3975                 return 0;
3976 }
3977
3978 static ssize_t
3979 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
3980 {
3981         raid5_conf_t *conf = mddev_to_conf(mddev);
3982         char *end;
3983         int new;
3984         if (len >= PAGE_SIZE)
3985                 return -EINVAL;
3986         if (!conf)
3987                 return -ENODEV;
3988
3989         new = simple_strtoul(page, &end, 10);
3990         if (!*page || (*end && *end != '\n') )
3991                 return -EINVAL;
3992         if (new <= 16 || new > 32768)
3993                 return -EINVAL;
3994         while (new < conf->max_nr_stripes) {
3995                 if (drop_one_stripe(conf))
3996                         conf->max_nr_stripes--;
3997                 else
3998                         break;
3999         }
4000         md_allow_write(mddev);
4001         while (new > conf->max_nr_stripes) {
4002                 if (grow_one_stripe(conf))
4003                         conf->max_nr_stripes++;
4004                 else break;
4005         }
4006         return len;
4007 }
4008
4009 static struct md_sysfs_entry
4010 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4011                                 raid5_show_stripe_cache_size,
4012                                 raid5_store_stripe_cache_size);
4013
4014 static ssize_t
4015 stripe_cache_active_show(mddev_t *mddev, char *page)
4016 {
4017         raid5_conf_t *conf = mddev_to_conf(mddev);
4018         if (conf)
4019                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4020         else
4021                 return 0;
4022 }
4023
4024 static struct md_sysfs_entry
4025 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4026
4027 static struct attribute *raid5_attrs[] =  {
4028         &raid5_stripecache_size.attr,
4029         &raid5_stripecache_active.attr,
4030         NULL,
4031 };
4032 static struct attribute_group raid5_attrs_group = {
4033         .name = NULL,
4034         .attrs = raid5_attrs,
4035 };
4036
4037 static int run(mddev_t *mddev)
4038 {
4039         raid5_conf_t *conf;
4040         int raid_disk, memory;
4041         mdk_rdev_t *rdev;
4042         struct disk_info *disk;
4043         struct list_head *tmp;
4044         int working_disks = 0;
4045
4046         if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4047                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4048                        mdname(mddev), mddev->level);
4049                 return -EIO;
4050         }
4051
4052         if (mddev->reshape_position != MaxSector) {
4053                 /* Check that we can continue the reshape.
4054                  * Currently only disks can change, it must
4055                  * increase, and we must be past the point where
4056                  * a stripe over-writes itself
4057                  */
4058                 sector_t here_new, here_old;
4059                 int old_disks;
4060                 int max_degraded = (mddev->level == 5 ? 1 : 2);
4061
4062                 if (mddev->new_level != mddev->level ||
4063                     mddev->new_layout != mddev->layout ||
4064                     mddev->new_chunk != mddev->chunk_size) {
4065                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4066                                "required - aborting.\n",
4067                                mdname(mddev));
4068                         return -EINVAL;
4069                 }
4070                 if (mddev->delta_disks <= 0) {
4071                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4072                                "(reduce disks) required - aborting.\n",
4073                                mdname(mddev));
4074                         return -EINVAL;
4075                 }
4076                 old_disks = mddev->raid_disks - mddev->delta_disks;
4077                 /* reshape_position must be on a new-stripe boundary, and one
4078                  * further up in new geometry must map after here in old
4079                  * geometry.
4080                  */
4081                 here_new = mddev->reshape_position;
4082                 if (sector_div(here_new, (mddev->chunk_size>>9)*
4083                                (mddev->raid_disks - max_degraded))) {
4084                         printk(KERN_ERR "raid5: reshape_position not "
4085                                "on a stripe boundary\n");
4086                         return -EINVAL;
4087                 }
4088                 /* here_new is the stripe we will write to */
4089                 here_old = mddev->reshape_position;
4090                 sector_div(here_old, (mddev->chunk_size>>9)*
4091                            (old_disks-max_degraded));
4092                 /* here_old is the first stripe that we might need to read
4093                  * from */
4094                 if (here_new >= here_old) {
4095                         /* Reading from the same stripe as writing to - bad */
4096                         printk(KERN_ERR "raid5: reshape_position too early for "
4097                                "auto-recovery - aborting.\n");
4098                         return -EINVAL;
4099                 }
4100                 printk(KERN_INFO "raid5: reshape will continue\n");
4101                 /* OK, we should be able to continue; */
4102         }
4103
4104
4105         mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4106         if ((conf = mddev->private) == NULL)
4107                 goto abort;
4108         if (mddev->reshape_position == MaxSector) {
4109                 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4110         } else {
4111                 conf->raid_disks = mddev->raid_disks;
4112                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4113         }
4114
4115         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4116                               GFP_KERNEL);
4117         if (!conf->disks)
4118                 goto abort;
4119
4120         conf->mddev = mddev;
4121
4122         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4123                 goto abort;
4124
4125         if (mddev->level == 6) {
4126                 conf->spare_page = alloc_page(GFP_KERNEL);
4127                 if (!conf->spare_page)
4128                         goto abort;
4129         }
4130         spin_lock_init(&conf->device_lock);
4131         init_waitqueue_head(&conf->wait_for_stripe);
4132         init_waitqueue_head(&conf->wait_for_overlap);
4133         INIT_LIST_HEAD(&conf->handle_list);
4134         INIT_LIST_HEAD(&conf->delayed_list);
4135         INIT_LIST_HEAD(&conf->bitmap_list);
4136         INIT_LIST_HEAD(&conf->inactive_list);
4137         atomic_set(&conf->active_stripes, 0);
4138         atomic_set(&conf->preread_active_stripes, 0);
4139         atomic_set(&conf->active_aligned_reads, 0);
4140
4141         pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4142
4143         ITERATE_RDEV(mddev,rdev,tmp) {
4144                 raid_disk = rdev->raid_disk;
4145                 if (raid_disk >= conf->raid_disks
4146                     || raid_disk < 0)
4147                         continue;
4148                 disk = conf->disks + raid_disk;
4149
4150                 disk->rdev = rdev;
4151
4152                 if (test_bit(In_sync, &rdev->flags)) {
4153                         char b[BDEVNAME_SIZE];
4154                         printk(KERN_INFO "raid5: device %s operational as raid"
4155                                 " disk %d\n", bdevname(rdev->bdev,b),
4156                                 raid_disk);
4157                         working_disks++;
4158                 }
4159         }
4160
4161         /*
4162          * 0 for a fully functional array, 1 or 2 for a degraded array.
4163          */
4164         mddev->degraded = conf->raid_disks - working_disks;
4165         conf->mddev = mddev;
4166         conf->chunk_size = mddev->chunk_size;
4167         conf->level = mddev->level;
4168         if (conf->level == 6)
4169                 conf->max_degraded = 2;
4170         else
4171                 conf->max_degraded = 1;
4172         conf->algorithm = mddev->layout;
4173         conf->max_nr_stripes = NR_STRIPES;
4174         conf->expand_progress = mddev->reshape_position;
4175
4176         /* device size must be a multiple of chunk size */
4177         mddev->size &= ~(mddev->chunk_size/1024 -1);
4178         mddev->resync_max_sectors = mddev->size << 1;
4179
4180         if (conf->level == 6 && conf->raid_disks < 4) {
4181                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4182                        mdname(mddev), conf->raid_disks);
4183                 goto abort;
4184         }
4185         if (!conf->chunk_size || conf->chunk_size % 4) {
4186                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4187                         conf->chunk_size, mdname(mddev));
4188                 goto abort;
4189         }
4190         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4191                 printk(KERN_ERR 
4192                         "raid5: unsupported parity algorithm %d for %s\n",
4193                         conf->algorithm, mdname(mddev));
4194                 goto abort;
4195         }
4196         if (mddev->degraded > conf->max_degraded) {
4197                 printk(KERN_ERR "raid5: not enough operational devices for %s"
4198                         " (%d/%d failed)\n",
4199                         mdname(mddev), mddev->degraded, conf->raid_disks);
4200                 goto abort;
4201         }
4202
4203         if (mddev->degraded > 0 &&
4204             mddev->recovery_cp != MaxSector) {
4205                 if (mddev->ok_start_degraded)
4206                         printk(KERN_WARNING
4207                                "raid5: starting dirty degraded array: %s"
4208                                "- data corruption possible.\n",
4209                                mdname(mddev));
4210                 else {
4211                         printk(KERN_ERR
4212                                "raid5: cannot start dirty degraded array for %s\n",
4213                                mdname(mddev));
4214                         goto abort;
4215                 }
4216         }
4217
4218         {
4219                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4220                 if (!mddev->thread) {
4221                         printk(KERN_ERR 
4222                                 "raid5: couldn't allocate thread for %s\n",
4223                                 mdname(mddev));
4224                         goto abort;
4225                 }
4226         }
4227         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4228                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4229         if (grow_stripes(conf, conf->max_nr_stripes)) {
4230                 printk(KERN_ERR 
4231                         "raid5: couldn't allocate %dkB for buffers\n", memory);
4232                 shrink_stripes(conf);
4233                 md_unregister_thread(mddev->thread);
4234                 goto abort;
4235         } else
4236                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4237                         memory, mdname(mddev));
4238
4239         if (mddev->degraded == 0)
4240                 printk("raid5: raid level %d set %s active with %d out of %d"
4241                         " devices, algorithm %d\n", conf->level, mdname(mddev), 
4242                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4243                         conf->algorithm);
4244         else
4245                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4246                         " out of %d devices, algorithm %d\n", conf->level,
4247                         mdname(mddev), mddev->raid_disks - mddev->degraded,
4248                         mddev->raid_disks, conf->algorithm);
4249
4250         print_raid5_conf(conf);
4251
4252         if (conf->expand_progress != MaxSector) {
4253                 printk("...ok start reshape thread\n");
4254                 conf->expand_lo = conf->expand_progress;
4255                 atomic_set(&conf->reshape_stripes, 0);
4256                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4257                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4258                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4259                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4260                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4261                                                         "%s_reshape");
4262         }
4263
4264         /* read-ahead size must cover two whole stripes, which is
4265          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4266          */
4267         {
4268                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4269                 int stripe = data_disks *
4270                         (mddev->chunk_size / PAGE_SIZE);
4271                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4272                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4273         }
4274
4275         /* Ok, everything is just fine now */
4276         if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4277                 printk(KERN_WARNING
4278                        "raid5: failed to create sysfs attributes for %s\n",
4279                        mdname(mddev));
4280
4281         mddev->queue->unplug_fn = raid5_unplug_device;
4282         mddev->queue->issue_flush_fn = raid5_issue_flush;
4283         mddev->queue->backing_dev_info.congested_data = mddev;
4284         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4285
4286         mddev->array_size =  mddev->size * (conf->previous_raid_disks -
4287                                             conf->max_degraded);
4288
4289         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4290
4291         return 0;
4292 abort:
4293         if (conf) {
4294                 print_raid5_conf(conf);
4295                 safe_put_page(conf->spare_page);
4296                 kfree(conf->disks);
4297                 kfree(conf->stripe_hashtbl);
4298                 kfree(conf);
4299         }
4300         mddev->private = NULL;
4301         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4302         return -EIO;
4303 }
4304
4305
4306
4307 static int stop(mddev_t *mddev)
4308 {
4309         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4310
4311         md_unregister_thread(mddev->thread);
4312         mddev->thread = NULL;
4313         shrink_stripes(conf);
4314         kfree(conf->stripe_hashtbl);
4315         mddev->queue->backing_dev_info.congested_fn = NULL;
4316         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4317         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4318         kfree(conf->disks);
4319         kfree(conf);
4320         mddev->private = NULL;
4321         return 0;
4322 }
4323
4324 #ifdef DEBUG
4325 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4326 {
4327         int i;
4328
4329         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4330                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4331         seq_printf(seq, "sh %llu,  count %d.\n",
4332                    (unsigned long long)sh->sector, atomic_read(&sh->count));
4333         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4334         for (i = 0; i < sh->disks; i++) {
4335                 seq_printf(seq, "(cache%d: %p %ld) ",
4336                            i, sh->dev[i].page, sh->dev[i].flags);
4337         }
4338         seq_printf(seq, "\n");
4339 }
4340
4341 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4342 {
4343         struct stripe_head *sh;
4344         struct hlist_node *hn;
4345         int i;
4346
4347         spin_lock_irq(&conf->device_lock);
4348         for (i = 0; i < NR_HASH; i++) {
4349                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4350                         if (sh->raid_conf != conf)
4351                                 continue;
4352                         print_sh(seq, sh);
4353                 }
4354         }
4355         spin_unlock_irq(&conf->device_lock);
4356 }
4357 #endif
4358
4359 static void status (struct seq_file *seq, mddev_t *mddev)
4360 {
4361         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4362         int i;
4363
4364         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4365         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4366         for (i = 0; i < conf->raid_disks; i++)
4367                 seq_printf (seq, "%s",
4368                                conf->disks[i].rdev &&
4369                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4370         seq_printf (seq, "]");
4371 #ifdef DEBUG
4372         seq_printf (seq, "\n");
4373         printall(seq, conf);
4374 #endif
4375 }
4376
4377 static void print_raid5_conf (raid5_conf_t *conf)
4378 {
4379         int i;
4380         struct disk_info *tmp;
4381
4382         printk("RAID5 conf printout:\n");
4383         if (!conf) {
4384                 printk("(conf==NULL)\n");
4385                 return;
4386         }
4387         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4388                  conf->raid_disks - conf->mddev->degraded);
4389
4390         for (i = 0; i < conf->raid_disks; i++) {
4391                 char b[BDEVNAME_SIZE];
4392                 tmp = conf->disks + i;
4393                 if (tmp->rdev)
4394                 printk(" disk %d, o:%d, dev:%s\n",
4395                         i, !test_bit(Faulty, &tmp->rdev->flags),
4396                         bdevname(tmp->rdev->bdev,b));
4397         }
4398 }
4399
4400 static int raid5_spare_active(mddev_t *mddev)
4401 {
4402         int i;
4403         raid5_conf_t *conf = mddev->private;
4404         struct disk_info *tmp;
4405
4406         for (i = 0; i < conf->raid_disks; i++) {
4407                 tmp = conf->disks + i;
4408                 if (tmp->rdev
4409                     && !test_bit(Faulty, &tmp->rdev->flags)
4410                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4411                         unsigned long flags;
4412                         spin_lock_irqsave(&conf->device_lock, flags);
4413                         mddev->degraded--;
4414                         spin_unlock_irqrestore(&conf->device_lock, flags);
4415                 }
4416         }
4417         print_raid5_conf(conf);
4418         return 0;
4419 }
4420
4421 static int raid5_remove_disk(mddev_t *mddev, int number)
4422 {
4423         raid5_conf_t *conf = mddev->private;
4424         int err = 0;
4425         mdk_rdev_t *rdev;
4426         struct disk_info *p = conf->disks + number;
4427
4428         print_raid5_conf(conf);
4429         rdev = p->rdev;
4430         if (rdev) {
4431                 if (test_bit(In_sync, &rdev->flags) ||
4432                     atomic_read(&rdev->nr_pending)) {
4433                         err = -EBUSY;
4434                         goto abort;
4435                 }
4436                 p->rdev = NULL;
4437                 synchronize_rcu();
4438                 if (atomic_read(&rdev->nr_pending)) {
4439                         /* lost the race, try later */
4440                         err = -EBUSY;
4441                         p->rdev = rdev;
4442                 }
4443         }
4444 abort:
4445
4446         print_raid5_conf(conf);
4447         return err;
4448 }
4449
4450 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4451 {
4452         raid5_conf_t *conf = mddev->private;
4453         int found = 0;
4454         int disk;
4455         struct disk_info *p;
4456
4457         if (mddev->degraded > conf->max_degraded)
4458                 /* no point adding a device */
4459                 return 0;
4460
4461         /*
4462          * find the disk ... but prefer rdev->saved_raid_disk
4463          * if possible.
4464          */
4465         if (rdev->saved_raid_disk >= 0 &&
4466             conf->disks[rdev->saved_raid_disk].rdev == NULL)
4467                 disk = rdev->saved_raid_disk;
4468         else
4469                 disk = 0;
4470         for ( ; disk < conf->raid_disks; disk++)
4471                 if ((p=conf->disks + disk)->rdev == NULL) {
4472                         clear_bit(In_sync, &rdev->flags);
4473                         rdev->raid_disk = disk;
4474                         found = 1;
4475                         if (rdev->saved_raid_disk != disk)
4476                                 conf->fullsync = 1;
4477                         rcu_assign_pointer(p->rdev, rdev);
4478                         break;
4479                 }
4480         print_raid5_conf(conf);
4481         return found;
4482 }
4483
4484 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4485 {
4486         /* no resync is happening, and there is enough space
4487          * on all devices, so we can resize.
4488          * We need to make sure resync covers any new space.
4489          * If the array is shrinking we should possibly wait until
4490          * any io in the removed space completes, but it hardly seems
4491          * worth it.
4492          */
4493         raid5_conf_t *conf = mddev_to_conf(mddev);
4494
4495         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4496         mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4497         set_capacity(mddev->gendisk, mddev->array_size << 1);
4498         mddev->changed = 1;
4499         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
4500                 mddev->recovery_cp = mddev->size << 1;
4501                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4502         }
4503         mddev->size = sectors /2;
4504         mddev->resync_max_sectors = sectors;
4505         return 0;
4506 }
4507
4508 #ifdef CONFIG_MD_RAID5_RESHAPE
4509 static int raid5_check_reshape(mddev_t *mddev)
4510 {
4511         raid5_conf_t *conf = mddev_to_conf(mddev);
4512         int err;
4513
4514         if (mddev->delta_disks < 0 ||
4515             mddev->new_level != mddev->level)
4516                 return -EINVAL; /* Cannot shrink array or change level yet */
4517         if (mddev->delta_disks == 0)
4518                 return 0; /* nothing to do */
4519
4520         /* Can only proceed if there are plenty of stripe_heads.
4521          * We need a minimum of one full stripe,, and for sensible progress
4522          * it is best to have about 4 times that.
4523          * If we require 4 times, then the default 256 4K stripe_heads will
4524          * allow for chunk sizes up to 256K, which is probably OK.
4525          * If the chunk size is greater, user-space should request more
4526          * stripe_heads first.
4527          */
4528         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4529             (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4530                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
4531                        (mddev->chunk_size / STRIPE_SIZE)*4);
4532                 return -ENOSPC;
4533         }
4534
4535         err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4536         if (err)
4537                 return err;
4538
4539         if (mddev->degraded > conf->max_degraded)
4540                 return -EINVAL;
4541         /* looks like we might be able to manage this */
4542         return 0;
4543 }
4544
4545 static int raid5_start_reshape(mddev_t *mddev)
4546 {
4547         raid5_conf_t *conf = mddev_to_conf(mddev);
4548         mdk_rdev_t *rdev;
4549         struct list_head *rtmp;
4550         int spares = 0;
4551         int added_devices = 0;
4552         unsigned long flags;
4553
4554         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4555                 return -EBUSY;
4556
4557         ITERATE_RDEV(mddev, rdev, rtmp)
4558                 if (rdev->raid_disk < 0 &&
4559                     !test_bit(Faulty, &rdev->flags))
4560                         spares++;
4561
4562         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4563                 /* Not enough devices even to make a degraded array
4564                  * of that size
4565                  */
4566                 return -EINVAL;
4567
4568         atomic_set(&conf->reshape_stripes, 0);
4569         spin_lock_irq(&conf->device_lock);
4570         conf->previous_raid_disks = conf->raid_disks;
4571         conf->raid_disks += mddev->delta_disks;
4572         conf->expand_progress = 0;
4573         conf->expand_lo = 0;
4574         spin_unlock_irq(&conf->device_lock);
4575
4576         /* Add some new drives, as many as will fit.
4577          * We know there are enough to make the newly sized array work.
4578          */
4579         ITERATE_RDEV(mddev, rdev, rtmp)
4580                 if (rdev->raid_disk < 0 &&
4581                     !test_bit(Faulty, &rdev->flags)) {
4582                         if (raid5_add_disk(mddev, rdev)) {
4583                                 char nm[20];
4584                                 set_bit(In_sync, &rdev->flags);
4585                                 added_devices++;
4586                                 rdev->recovery_offset = 0;
4587                                 sprintf(nm, "rd%d", rdev->raid_disk);
4588                                 if (sysfs_create_link(&mddev->kobj,
4589                                                       &rdev->kobj, nm))
4590                                         printk(KERN_WARNING
4591                                                "raid5: failed to create "
4592                                                " link %s for %s\n",
4593                                                nm, mdname(mddev));
4594                         } else
4595                                 break;
4596                 }
4597
4598         spin_lock_irqsave(&conf->device_lock, flags);
4599         mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4600         spin_unlock_irqrestore(&conf->device_lock, flags);
4601         mddev->raid_disks = conf->raid_disks;
4602         mddev->reshape_position = 0;
4603         set_bit(MD_CHANGE_DEVS, &mddev->flags);
4604
4605         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4606         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4607         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4608         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4609         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4610                                                 "%s_reshape");
4611         if (!mddev->sync_thread) {
4612                 mddev->recovery = 0;
4613                 spin_lock_irq(&conf->device_lock);
4614                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4615                 conf->expand_progress = MaxSector;
4616                 spin_unlock_irq(&conf->device_lock);
4617                 return -EAGAIN;
4618         }
4619         md_wakeup_thread(mddev->sync_thread);
4620         md_new_event(mddev);
4621         return 0;
4622 }
4623 #endif
4624
4625 static void end_reshape(raid5_conf_t *conf)
4626 {
4627         struct block_device *bdev;
4628
4629         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4630                 conf->mddev->array_size = conf->mddev->size *
4631                         (conf->raid_disks - conf->max_degraded);
4632                 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4633                 conf->mddev->changed = 1;
4634
4635                 bdev = bdget_disk(conf->mddev->gendisk, 0);
4636                 if (bdev) {
4637                         mutex_lock(&bdev->bd_inode->i_mutex);
4638                         i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4639                         mutex_unlock(&bdev->bd_inode->i_mutex);
4640                         bdput(bdev);
4641                 }
4642                 spin_lock_irq(&conf->device_lock);
4643                 conf->expand_progress = MaxSector;
4644                 spin_unlock_irq(&conf->device_lock);
4645                 conf->mddev->reshape_position = MaxSector;
4646
4647                 /* read-ahead size must cover two whole stripes, which is
4648                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4649                  */
4650                 {
4651                         int data_disks = conf->previous_raid_disks - conf->max_degraded;
4652                         int stripe = data_disks *
4653                                 (conf->mddev->chunk_size / PAGE_SIZE);
4654                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4655                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4656                 }
4657         }
4658 }
4659
4660 static void raid5_quiesce(mddev_t *mddev, int state)
4661 {
4662         raid5_conf_t *conf = mddev_to_conf(mddev);
4663
4664         switch(state) {
4665         case 2: /* resume for a suspend */
4666                 wake_up(&conf->wait_for_overlap);
4667                 break;
4668
4669         case 1: /* stop all writes */
4670                 spin_lock_irq(&conf->device_lock);
4671                 conf->quiesce = 1;
4672                 wait_event_lock_irq(conf->wait_for_stripe,
4673                                     atomic_read(&conf->active_stripes) == 0 &&
4674                                     atomic_read(&conf->active_aligned_reads) == 0,
4675                                     conf->device_lock, /* nothing */);
4676                 spin_unlock_irq(&conf->device_lock);
4677                 break;
4678
4679         case 0: /* re-enable writes */
4680                 spin_lock_irq(&conf->device_lock);
4681                 conf->quiesce = 0;
4682                 wake_up(&conf->wait_for_stripe);
4683                 wake_up(&conf->wait_for_overlap);
4684                 spin_unlock_irq(&conf->device_lock);
4685                 break;
4686         }
4687 }
4688
4689 static struct mdk_personality raid6_personality =
4690 {
4691         .name           = "raid6",
4692         .level          = 6,
4693         .owner          = THIS_MODULE,
4694         .make_request   = make_request,
4695         .run            = run,
4696         .stop           = stop,
4697         .status         = status,
4698         .error_handler  = error,
4699         .hot_add_disk   = raid5_add_disk,
4700         .hot_remove_disk= raid5_remove_disk,
4701         .spare_active   = raid5_spare_active,
4702         .sync_request   = sync_request,
4703         .resize         = raid5_resize,
4704 #ifdef CONFIG_MD_RAID5_RESHAPE
4705         .check_reshape  = raid5_check_reshape,
4706         .start_reshape  = raid5_start_reshape,
4707 #endif
4708         .quiesce        = raid5_quiesce,
4709 };
4710 static struct mdk_personality raid5_personality =
4711 {
4712         .name           = "raid5",
4713         .level          = 5,
4714         .owner          = THIS_MODULE,
4715         .make_request   = make_request,
4716         .run            = run,
4717         .stop           = stop,
4718         .status         = status,
4719         .error_handler  = error,
4720         .hot_add_disk   = raid5_add_disk,
4721         .hot_remove_disk= raid5_remove_disk,
4722         .spare_active   = raid5_spare_active,
4723         .sync_request   = sync_request,
4724         .resize         = raid5_resize,
4725 #ifdef CONFIG_MD_RAID5_RESHAPE
4726         .check_reshape  = raid5_check_reshape,
4727         .start_reshape  = raid5_start_reshape,
4728 #endif
4729         .quiesce        = raid5_quiesce,
4730 };
4731
4732 static struct mdk_personality raid4_personality =
4733 {
4734         .name           = "raid4",
4735         .level          = 4,
4736         .owner          = THIS_MODULE,
4737         .make_request   = make_request,
4738         .run            = run,
4739         .stop           = stop,
4740         .status         = status,
4741         .error_handler  = error,
4742         .hot_add_disk   = raid5_add_disk,
4743         .hot_remove_disk= raid5_remove_disk,
4744         .spare_active   = raid5_spare_active,
4745         .sync_request   = sync_request,
4746         .resize         = raid5_resize,
4747 #ifdef CONFIG_MD_RAID5_RESHAPE
4748         .check_reshape  = raid5_check_reshape,
4749         .start_reshape  = raid5_start_reshape,
4750 #endif
4751         .quiesce        = raid5_quiesce,
4752 };
4753
4754 static int __init raid5_init(void)
4755 {
4756         int e;
4757
4758         e = raid6_select_algo();
4759         if ( e )
4760                 return e;
4761         register_md_personality(&raid6_personality);
4762         register_md_personality(&raid5_personality);
4763         register_md_personality(&raid4_personality);
4764         return 0;
4765 }
4766
4767 static void raid5_exit(void)
4768 {
4769         unregister_md_personality(&raid6_personality);
4770         unregister_md_personality(&raid5_personality);
4771         unregister_md_personality(&raid4_personality);
4772 }
4773
4774 module_init(raid5_init);
4775 module_exit(raid5_exit);
4776 MODULE_LICENSE("GPL");
4777 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4778 MODULE_ALIAS("md-raid5");
4779 MODULE_ALIAS("md-raid4");
4780 MODULE_ALIAS("md-level-5");
4781 MODULE_ALIAS("md-level-4");
4782 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4783 MODULE_ALIAS("md-raid6");
4784 MODULE_ALIAS("md-level-6");
4785
4786 /* This used to be two separate modules, they were: */
4787 MODULE_ALIAS("raid5");
4788 MODULE_ALIAS("raid6");