0b66afef2d82009be4f54c34880df588a8a22b61
[linux-2.6.git] / drivers / md / raid5.c
1 /*
2  * raid5.c : Multiple Devices driver for Linux
3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4  *         Copyright (C) 1999, 2000 Ingo Molnar
5  *         Copyright (C) 2002, 2003 H. Peter Anvin
6  *
7  * RAID-4/5/6 management functions.
8  * Thanks to Penguin Computing for making the RAID-6 development possible
9  * by donating a test server!
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 /*
22  * BITMAP UNPLUGGING:
23  *
24  * The sequencing for updating the bitmap reliably is a little
25  * subtle (and I got it wrong the first time) so it deserves some
26  * explanation.
27  *
28  * We group bitmap updates into batches.  Each batch has a number.
29  * We may write out several batches at once, but that isn't very important.
30  * conf->bm_write is the number of the last batch successfully written.
31  * conf->bm_flush is the number of the last batch that was closed to
32  *    new additions.
33  * When we discover that we will need to write to any block in a stripe
34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35  * the number of the batch it will be in. This is bm_flush+1.
36  * When we are ready to do a write, if that batch hasn't been written yet,
37  *   we plug the array and queue the stripe for later.
38  * When an unplug happens, we increment bm_flush, thus closing the current
39  *   batch.
40  * When we notice that bm_flush > bm_write, we write out all pending updates
41  * to the bitmap, and advance bm_write to where bm_flush was.
42  * This may occasionally write a bit out twice, but is sure never to
43  * miss any bits.
44  */
45
46 #include <linux/module.h>
47 #include <linux/slab.h>
48 #include <linux/highmem.h>
49 #include <linux/bitops.h>
50 #include <linux/kthread.h>
51 #include <asm/atomic.h>
52 #include "raid6.h"
53
54 #include <linux/raid/bitmap.h>
55 #include <linux/async_tx.h>
56
57 /*
58  * Stripe cache
59  */
60
61 #define NR_STRIPES              256
62 #define STRIPE_SIZE             PAGE_SIZE
63 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
64 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
65 #define IO_THRESHOLD            1
66 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
67 #define HASH_MASK               (NR_HASH - 1)
68
69 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
70
71 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
72  * order without overlap.  There may be several bio's per stripe+device, and
73  * a bio could span several devices.
74  * When walking this list for a particular stripe+device, we must never proceed
75  * beyond a bio that extends past this device, as the next bio might no longer
76  * be valid.
77  * This macro is used to determine the 'next' bio in the list, given the sector
78  * of the current stripe+device
79  */
80 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
81 /*
82  * The following can be used to debug the driver
83  */
84 #define RAID5_PARANOIA  1
85 #if RAID5_PARANOIA && defined(CONFIG_SMP)
86 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
87 #else
88 # define CHECK_DEVLOCK()
89 #endif
90
91 #ifdef DEBUG
92 #define inline
93 #define __inline__
94 #endif
95
96 #if !RAID6_USE_EMPTY_ZERO_PAGE
97 /* In .bss so it's zeroed */
98 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
99 #endif
100
101 static inline int raid6_next_disk(int disk, int raid_disks)
102 {
103         disk++;
104         return (disk < raid_disks) ? disk : 0;
105 }
106
107 static void return_io(struct bio *return_bi)
108 {
109         struct bio *bi = return_bi;
110         while (bi) {
111                 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(request_queue_t *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 | ASYNC_TX_KMAP_SRC,
497                                         tx, NULL, NULL);
498                         else
499                                 tx = async_memcpy(bio_page, page, b_offset,
500                                         page_offset, clen,
501                                         ASYNC_TX_DEP_ACK | ASYNC_TX_KMAP_DST,
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, more_to_read = 0;
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                 /* check if this stripe has new incoming reads */
526                 if (dev->toread)
527                         more_to_read++;
528
529                 /* acknowledge completion of a biofill operation */
530                 /* and check if we need to reply to a read request
531                 */
532                 if (test_bit(R5_Wantfill, &dev->flags) && !dev->toread) {
533                         struct bio *rbi, *rbi2;
534                         clear_bit(R5_Wantfill, &dev->flags);
535
536                         /* The access to dev->read is outside of the
537                          * spin_lock_irq(&conf->device_lock), but is protected
538                          * by the STRIPE_OP_BIOFILL pending bit
539                          */
540                         BUG_ON(!dev->read);
541                         rbi = dev->read;
542                         dev->read = NULL;
543                         while (rbi && rbi->bi_sector <
544                                 dev->sector + STRIPE_SECTORS) {
545                                 rbi2 = r5_next_bio(rbi, dev->sector);
546                                 spin_lock_irq(&conf->device_lock);
547                                 if (--rbi->bi_phys_segments == 0) {
548                                         rbi->bi_next = return_bi;
549                                         return_bi = rbi;
550                                 }
551                                 spin_unlock_irq(&conf->device_lock);
552                                 rbi = rbi2;
553                         }
554                 }
555         }
556         clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack);
557         clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending);
558
559         return_io(return_bi);
560
561         if (more_to_read)
562                 set_bit(STRIPE_HANDLE, &sh->state);
563         release_stripe(sh);
564 }
565
566 static void ops_run_biofill(struct stripe_head *sh)
567 {
568         struct dma_async_tx_descriptor *tx = NULL;
569         raid5_conf_t *conf = sh->raid_conf;
570         int i;
571
572         pr_debug("%s: stripe %llu\n", __FUNCTION__,
573                 (unsigned long long)sh->sector);
574
575         for (i = sh->disks; i--; ) {
576                 struct r5dev *dev = &sh->dev[i];
577                 if (test_bit(R5_Wantfill, &dev->flags)) {
578                         struct bio *rbi;
579                         spin_lock_irq(&conf->device_lock);
580                         dev->read = rbi = dev->toread;
581                         dev->toread = NULL;
582                         spin_unlock_irq(&conf->device_lock);
583                         while (rbi && rbi->bi_sector <
584                                 dev->sector + STRIPE_SECTORS) {
585                                 tx = async_copy_data(0, rbi, dev->page,
586                                         dev->sector, tx);
587                                 rbi = r5_next_bio(rbi, dev->sector);
588                         }
589                 }
590         }
591
592         atomic_inc(&sh->count);
593         async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
594                 ops_complete_biofill, sh);
595 }
596
597 static void ops_complete_compute5(void *stripe_head_ref)
598 {
599         struct stripe_head *sh = stripe_head_ref;
600         int target = sh->ops.target;
601         struct r5dev *tgt = &sh->dev[target];
602
603         pr_debug("%s: stripe %llu\n", __FUNCTION__,
604                 (unsigned long long)sh->sector);
605
606         set_bit(R5_UPTODATE, &tgt->flags);
607         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
608         clear_bit(R5_Wantcompute, &tgt->flags);
609         set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
610         set_bit(STRIPE_HANDLE, &sh->state);
611         release_stripe(sh);
612 }
613
614 static struct dma_async_tx_descriptor *
615 ops_run_compute5(struct stripe_head *sh, unsigned long pending)
616 {
617         /* kernel stack size limits the total number of disks */
618         int disks = sh->disks;
619         struct page *xor_srcs[disks];
620         int target = sh->ops.target;
621         struct r5dev *tgt = &sh->dev[target];
622         struct page *xor_dest = tgt->page;
623         int count = 0;
624         struct dma_async_tx_descriptor *tx;
625         int i;
626
627         pr_debug("%s: stripe %llu block: %d\n",
628                 __FUNCTION__, (unsigned long long)sh->sector, target);
629         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
630
631         for (i = disks; i--; )
632                 if (i != target)
633                         xor_srcs[count++] = sh->dev[i].page;
634
635         atomic_inc(&sh->count);
636
637         if (unlikely(count == 1))
638                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
639                         0, NULL, ops_complete_compute5, sh);
640         else
641                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
642                         ASYNC_TX_XOR_ZERO_DST, NULL,
643                         ops_complete_compute5, sh);
644
645         /* ack now if postxor is not set to be run */
646         if (tx && !test_bit(STRIPE_OP_POSTXOR, &pending))
647                 async_tx_ack(tx);
648
649         return tx;
650 }
651
652 static void ops_complete_prexor(void *stripe_head_ref)
653 {
654         struct stripe_head *sh = stripe_head_ref;
655
656         pr_debug("%s: stripe %llu\n", __FUNCTION__,
657                 (unsigned long long)sh->sector);
658
659         set_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
660 }
661
662 static struct dma_async_tx_descriptor *
663 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
664 {
665         /* kernel stack size limits the total number of disks */
666         int disks = sh->disks;
667         struct page *xor_srcs[disks];
668         int count = 0, pd_idx = sh->pd_idx, i;
669
670         /* existing parity data subtracted */
671         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
672
673         pr_debug("%s: stripe %llu\n", __FUNCTION__,
674                 (unsigned long long)sh->sector);
675
676         for (i = disks; i--; ) {
677                 struct r5dev *dev = &sh->dev[i];
678                 /* Only process blocks that are known to be uptodate */
679                 if (dev->towrite && test_bit(R5_Wantprexor, &dev->flags))
680                         xor_srcs[count++] = dev->page;
681         }
682
683         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
684                 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
685                 ops_complete_prexor, sh);
686
687         return tx;
688 }
689
690 static struct dma_async_tx_descriptor *
691 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
692 {
693         int disks = sh->disks;
694         int pd_idx = sh->pd_idx, i;
695
696         /* check if prexor is active which means only process blocks
697          * that are part of a read-modify-write (Wantprexor)
698          */
699         int prexor = test_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
700
701         pr_debug("%s: stripe %llu\n", __FUNCTION__,
702                 (unsigned long long)sh->sector);
703
704         for (i = disks; i--; ) {
705                 struct r5dev *dev = &sh->dev[i];
706                 struct bio *chosen;
707                 int towrite;
708
709                 towrite = 0;
710                 if (prexor) { /* rmw */
711                         if (dev->towrite &&
712                             test_bit(R5_Wantprexor, &dev->flags))
713                                 towrite = 1;
714                 } else { /* rcw */
715                         if (i != pd_idx && dev->towrite &&
716                                 test_bit(R5_LOCKED, &dev->flags))
717                                 towrite = 1;
718                 }
719
720                 if (towrite) {
721                         struct bio *wbi;
722
723                         spin_lock(&sh->lock);
724                         chosen = dev->towrite;
725                         dev->towrite = NULL;
726                         BUG_ON(dev->written);
727                         wbi = dev->written = chosen;
728                         spin_unlock(&sh->lock);
729
730                         while (wbi && wbi->bi_sector <
731                                 dev->sector + STRIPE_SECTORS) {
732                                 tx = async_copy_data(1, wbi, dev->page,
733                                         dev->sector, tx);
734                                 wbi = r5_next_bio(wbi, dev->sector);
735                         }
736                 }
737         }
738
739         return tx;
740 }
741
742 static void ops_complete_postxor(void *stripe_head_ref)
743 {
744         struct stripe_head *sh = stripe_head_ref;
745
746         pr_debug("%s: stripe %llu\n", __FUNCTION__,
747                 (unsigned long long)sh->sector);
748
749         set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
750         set_bit(STRIPE_HANDLE, &sh->state);
751         release_stripe(sh);
752 }
753
754 static void ops_complete_write(void *stripe_head_ref)
755 {
756         struct stripe_head *sh = stripe_head_ref;
757         int disks = sh->disks, i, pd_idx = sh->pd_idx;
758
759         pr_debug("%s: stripe %llu\n", __FUNCTION__,
760                 (unsigned long long)sh->sector);
761
762         for (i = disks; i--; ) {
763                 struct r5dev *dev = &sh->dev[i];
764                 if (dev->written || i == pd_idx)
765                         set_bit(R5_UPTODATE, &dev->flags);
766         }
767
768         set_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
769         set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
770
771         set_bit(STRIPE_HANDLE, &sh->state);
772         release_stripe(sh);
773 }
774
775 static void
776 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
777 {
778         /* kernel stack size limits the total number of disks */
779         int disks = sh->disks;
780         struct page *xor_srcs[disks];
781
782         int count = 0, pd_idx = sh->pd_idx, i;
783         struct page *xor_dest;
784         int prexor = test_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
785         unsigned long flags;
786         dma_async_tx_callback callback;
787
788         pr_debug("%s: stripe %llu\n", __FUNCTION__,
789                 (unsigned long long)sh->sector);
790
791         /* check if prexor is active which means only process blocks
792          * that are part of a read-modify-write (written)
793          */
794         if (prexor) {
795                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
796                 for (i = disks; i--; ) {
797                         struct r5dev *dev = &sh->dev[i];
798                         if (dev->written)
799                                 xor_srcs[count++] = dev->page;
800                 }
801         } else {
802                 xor_dest = sh->dev[pd_idx].page;
803                 for (i = disks; i--; ) {
804                         struct r5dev *dev = &sh->dev[i];
805                         if (i != pd_idx)
806                                 xor_srcs[count++] = dev->page;
807                 }
808         }
809
810         /* check whether this postxor is part of a write */
811         callback = test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending) ?
812                 ops_complete_write : ops_complete_postxor;
813
814         /* 1/ if we prexor'd then the dest is reused as a source
815          * 2/ if we did not prexor then we are redoing the parity
816          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
817          * for the synchronous xor case
818          */
819         flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
820                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
821
822         atomic_inc(&sh->count);
823
824         if (unlikely(count == 1)) {
825                 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
826                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
827                         flags, tx, callback, sh);
828         } else
829                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
830                         flags, tx, callback, sh);
831 }
832
833 static void ops_complete_check(void *stripe_head_ref)
834 {
835         struct stripe_head *sh = stripe_head_ref;
836         int pd_idx = sh->pd_idx;
837
838         pr_debug("%s: stripe %llu\n", __FUNCTION__,
839                 (unsigned long long)sh->sector);
840
841         if (test_and_clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending) &&
842                 sh->ops.zero_sum_result == 0)
843                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
844
845         set_bit(STRIPE_OP_CHECK, &sh->ops.complete);
846         set_bit(STRIPE_HANDLE, &sh->state);
847         release_stripe(sh);
848 }
849
850 static void ops_run_check(struct stripe_head *sh)
851 {
852         /* kernel stack size limits the total number of disks */
853         int disks = sh->disks;
854         struct page *xor_srcs[disks];
855         struct dma_async_tx_descriptor *tx;
856
857         int count = 0, pd_idx = sh->pd_idx, i;
858         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
859
860         pr_debug("%s: stripe %llu\n", __FUNCTION__,
861                 (unsigned long long)sh->sector);
862
863         for (i = disks; i--; ) {
864                 struct r5dev *dev = &sh->dev[i];
865                 if (i != pd_idx)
866                         xor_srcs[count++] = dev->page;
867         }
868
869         tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
870                 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
871
872         if (tx)
873                 set_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
874         else
875                 clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
876
877         atomic_inc(&sh->count);
878         tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
879                 ops_complete_check, sh);
880 }
881
882 static void raid5_run_ops(struct stripe_head *sh, unsigned long pending)
883 {
884         int overlap_clear = 0, i, disks = sh->disks;
885         struct dma_async_tx_descriptor *tx = NULL;
886
887         if (test_bit(STRIPE_OP_BIOFILL, &pending)) {
888                 ops_run_biofill(sh);
889                 overlap_clear++;
890         }
891
892         if (test_bit(STRIPE_OP_COMPUTE_BLK, &pending))
893                 tx = ops_run_compute5(sh, pending);
894
895         if (test_bit(STRIPE_OP_PREXOR, &pending))
896                 tx = ops_run_prexor(sh, tx);
897
898         if (test_bit(STRIPE_OP_BIODRAIN, &pending)) {
899                 tx = ops_run_biodrain(sh, tx);
900                 overlap_clear++;
901         }
902
903         if (test_bit(STRIPE_OP_POSTXOR, &pending))
904                 ops_run_postxor(sh, tx);
905
906         if (test_bit(STRIPE_OP_CHECK, &pending))
907                 ops_run_check(sh);
908
909         if (test_bit(STRIPE_OP_IO, &pending))
910                 ops_run_io(sh);
911
912         if (overlap_clear)
913                 for (i = disks; i--; ) {
914                         struct r5dev *dev = &sh->dev[i];
915                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
916                                 wake_up(&sh->raid_conf->wait_for_overlap);
917                 }
918 }
919
920 static int grow_one_stripe(raid5_conf_t *conf)
921 {
922         struct stripe_head *sh;
923         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
924         if (!sh)
925                 return 0;
926         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
927         sh->raid_conf = conf;
928         spin_lock_init(&sh->lock);
929
930         if (grow_buffers(sh, conf->raid_disks)) {
931                 shrink_buffers(sh, conf->raid_disks);
932                 kmem_cache_free(conf->slab_cache, sh);
933                 return 0;
934         }
935         sh->disks = conf->raid_disks;
936         /* we just created an active stripe so... */
937         atomic_set(&sh->count, 1);
938         atomic_inc(&conf->active_stripes);
939         INIT_LIST_HEAD(&sh->lru);
940         release_stripe(sh);
941         return 1;
942 }
943
944 static int grow_stripes(raid5_conf_t *conf, int num)
945 {
946         struct kmem_cache *sc;
947         int devs = conf->raid_disks;
948
949         sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
950         sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
951         conf->active_name = 0;
952         sc = kmem_cache_create(conf->cache_name[conf->active_name],
953                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
954                                0, 0, NULL, NULL);
955         if (!sc)
956                 return 1;
957         conf->slab_cache = sc;
958         conf->pool_size = devs;
959         while (num--)
960                 if (!grow_one_stripe(conf))
961                         return 1;
962         return 0;
963 }
964
965 #ifdef CONFIG_MD_RAID5_RESHAPE
966 static int resize_stripes(raid5_conf_t *conf, int newsize)
967 {
968         /* Make all the stripes able to hold 'newsize' devices.
969          * New slots in each stripe get 'page' set to a new page.
970          *
971          * This happens in stages:
972          * 1/ create a new kmem_cache and allocate the required number of
973          *    stripe_heads.
974          * 2/ gather all the old stripe_heads and tranfer the pages across
975          *    to the new stripe_heads.  This will have the side effect of
976          *    freezing the array as once all stripe_heads have been collected,
977          *    no IO will be possible.  Old stripe heads are freed once their
978          *    pages have been transferred over, and the old kmem_cache is
979          *    freed when all stripes are done.
980          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
981          *    we simple return a failre status - no need to clean anything up.
982          * 4/ allocate new pages for the new slots in the new stripe_heads.
983          *    If this fails, we don't bother trying the shrink the
984          *    stripe_heads down again, we just leave them as they are.
985          *    As each stripe_head is processed the new one is released into
986          *    active service.
987          *
988          * Once step2 is started, we cannot afford to wait for a write,
989          * so we use GFP_NOIO allocations.
990          */
991         struct stripe_head *osh, *nsh;
992         LIST_HEAD(newstripes);
993         struct disk_info *ndisks;
994         int err = 0;
995         struct kmem_cache *sc;
996         int i;
997
998         if (newsize <= conf->pool_size)
999                 return 0; /* never bother to shrink */
1000
1001         md_allow_write(conf->mddev);
1002
1003         /* Step 1 */
1004         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1005                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1006                                0, 0, NULL, NULL);
1007         if (!sc)
1008                 return -ENOMEM;
1009
1010         for (i = conf->max_nr_stripes; i; i--) {
1011                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1012                 if (!nsh)
1013                         break;
1014
1015                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1016
1017                 nsh->raid_conf = conf;
1018                 spin_lock_init(&nsh->lock);
1019
1020                 list_add(&nsh->lru, &newstripes);
1021         }
1022         if (i) {
1023                 /* didn't get enough, give up */
1024                 while (!list_empty(&newstripes)) {
1025                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1026                         list_del(&nsh->lru);
1027                         kmem_cache_free(sc, nsh);
1028                 }
1029                 kmem_cache_destroy(sc);
1030                 return -ENOMEM;
1031         }
1032         /* Step 2 - Must use GFP_NOIO now.
1033          * OK, we have enough stripes, start collecting inactive
1034          * stripes and copying them over
1035          */
1036         list_for_each_entry(nsh, &newstripes, lru) {
1037                 spin_lock_irq(&conf->device_lock);
1038                 wait_event_lock_irq(conf->wait_for_stripe,
1039                                     !list_empty(&conf->inactive_list),
1040                                     conf->device_lock,
1041                                     unplug_slaves(conf->mddev)
1042                         );
1043                 osh = get_free_stripe(conf);
1044                 spin_unlock_irq(&conf->device_lock);
1045                 atomic_set(&nsh->count, 1);
1046                 for(i=0; i<conf->pool_size; i++)
1047                         nsh->dev[i].page = osh->dev[i].page;
1048                 for( ; i<newsize; i++)
1049                         nsh->dev[i].page = NULL;
1050                 kmem_cache_free(conf->slab_cache, osh);
1051         }
1052         kmem_cache_destroy(conf->slab_cache);
1053
1054         /* Step 3.
1055          * At this point, we are holding all the stripes so the array
1056          * is completely stalled, so now is a good time to resize
1057          * conf->disks.
1058          */
1059         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1060         if (ndisks) {
1061                 for (i=0; i<conf->raid_disks; i++)
1062                         ndisks[i] = conf->disks[i];
1063                 kfree(conf->disks);
1064                 conf->disks = ndisks;
1065         } else
1066                 err = -ENOMEM;
1067
1068         /* Step 4, return new stripes to service */
1069         while(!list_empty(&newstripes)) {
1070                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1071                 list_del_init(&nsh->lru);
1072                 for (i=conf->raid_disks; i < newsize; i++)
1073                         if (nsh->dev[i].page == NULL) {
1074                                 struct page *p = alloc_page(GFP_NOIO);
1075                                 nsh->dev[i].page = p;
1076                                 if (!p)
1077                                         err = -ENOMEM;
1078                         }
1079                 release_stripe(nsh);
1080         }
1081         /* critical section pass, GFP_NOIO no longer needed */
1082
1083         conf->slab_cache = sc;
1084         conf->active_name = 1-conf->active_name;
1085         conf->pool_size = newsize;
1086         return err;
1087 }
1088 #endif
1089
1090 static int drop_one_stripe(raid5_conf_t *conf)
1091 {
1092         struct stripe_head *sh;
1093
1094         spin_lock_irq(&conf->device_lock);
1095         sh = get_free_stripe(conf);
1096         spin_unlock_irq(&conf->device_lock);
1097         if (!sh)
1098                 return 0;
1099         BUG_ON(atomic_read(&sh->count));
1100         shrink_buffers(sh, conf->pool_size);
1101         kmem_cache_free(conf->slab_cache, sh);
1102         atomic_dec(&conf->active_stripes);
1103         return 1;
1104 }
1105
1106 static void shrink_stripes(raid5_conf_t *conf)
1107 {
1108         while (drop_one_stripe(conf))
1109                 ;
1110
1111         if (conf->slab_cache)
1112                 kmem_cache_destroy(conf->slab_cache);
1113         conf->slab_cache = NULL;
1114 }
1115
1116 static int raid5_end_read_request(struct bio * bi, unsigned int bytes_done,
1117                                    int error)
1118 {
1119         struct stripe_head *sh = bi->bi_private;
1120         raid5_conf_t *conf = sh->raid_conf;
1121         int disks = sh->disks, i;
1122         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1123         char b[BDEVNAME_SIZE];
1124         mdk_rdev_t *rdev;
1125
1126         if (bi->bi_size)
1127                 return 1;
1128
1129         for (i=0 ; i<disks; i++)
1130                 if (bi == &sh->dev[i].req)
1131                         break;
1132
1133         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1134                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1135                 uptodate);
1136         if (i == disks) {
1137                 BUG();
1138                 return 0;
1139         }
1140
1141         if (uptodate) {
1142                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1143                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1144                         rdev = conf->disks[i].rdev;
1145                         printk(KERN_INFO "raid5:%s: read error corrected (%lu sectors at %llu on %s)\n",
1146                                mdname(conf->mddev), STRIPE_SECTORS,
1147                                (unsigned long long)sh->sector + rdev->data_offset,
1148                                bdevname(rdev->bdev, b));
1149                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1150                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1151                 }
1152                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1153                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1154         } else {
1155                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1156                 int retry = 0;
1157                 rdev = conf->disks[i].rdev;
1158
1159                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1160                 atomic_inc(&rdev->read_errors);
1161                 if (conf->mddev->degraded)
1162                         printk(KERN_WARNING "raid5:%s: read error not correctable (sector %llu on %s).\n",
1163                                mdname(conf->mddev),
1164                                (unsigned long long)sh->sector + rdev->data_offset,
1165                                bdn);
1166                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1167                         /* Oh, no!!! */
1168                         printk(KERN_WARNING "raid5:%s: read error NOT corrected!! (sector %llu on %s).\n",
1169                                mdname(conf->mddev),
1170                                (unsigned long long)sh->sector + rdev->data_offset,
1171                                bdn);
1172                 else if (atomic_read(&rdev->read_errors)
1173                          > conf->max_nr_stripes)
1174                         printk(KERN_WARNING
1175                                "raid5:%s: Too many read errors, failing device %s.\n",
1176                                mdname(conf->mddev), bdn);
1177                 else
1178                         retry = 1;
1179                 if (retry)
1180                         set_bit(R5_ReadError, &sh->dev[i].flags);
1181                 else {
1182                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1183                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1184                         md_error(conf->mddev, rdev);
1185                 }
1186         }
1187         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1188         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1189         set_bit(STRIPE_HANDLE, &sh->state);
1190         release_stripe(sh);
1191         return 0;
1192 }
1193
1194 static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
1195                                     int error)
1196 {
1197         struct stripe_head *sh = bi->bi_private;
1198         raid5_conf_t *conf = sh->raid_conf;
1199         int disks = sh->disks, i;
1200         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1201
1202         if (bi->bi_size)
1203                 return 1;
1204
1205         for (i=0 ; i<disks; i++)
1206                 if (bi == &sh->dev[i].req)
1207                         break;
1208
1209         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1210                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1211                 uptodate);
1212         if (i == disks) {
1213                 BUG();
1214                 return 0;
1215         }
1216
1217         if (!uptodate)
1218                 md_error(conf->mddev, conf->disks[i].rdev);
1219
1220         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1221         
1222         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1223         set_bit(STRIPE_HANDLE, &sh->state);
1224         release_stripe(sh);
1225         return 0;
1226 }
1227
1228
1229 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1230         
1231 static void raid5_build_block (struct stripe_head *sh, int i)
1232 {
1233         struct r5dev *dev = &sh->dev[i];
1234
1235         bio_init(&dev->req);
1236         dev->req.bi_io_vec = &dev->vec;
1237         dev->req.bi_vcnt++;
1238         dev->req.bi_max_vecs++;
1239         dev->vec.bv_page = dev->page;
1240         dev->vec.bv_len = STRIPE_SIZE;
1241         dev->vec.bv_offset = 0;
1242
1243         dev->req.bi_sector = sh->sector;
1244         dev->req.bi_private = sh;
1245
1246         dev->flags = 0;
1247         dev->sector = compute_blocknr(sh, i);
1248 }
1249
1250 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1251 {
1252         char b[BDEVNAME_SIZE];
1253         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1254         pr_debug("raid5: error called\n");
1255
1256         if (!test_bit(Faulty, &rdev->flags)) {
1257                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1258                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1259                         unsigned long flags;
1260                         spin_lock_irqsave(&conf->device_lock, flags);
1261                         mddev->degraded++;
1262                         spin_unlock_irqrestore(&conf->device_lock, flags);
1263                         /*
1264                          * if recovery was running, make sure it aborts.
1265                          */
1266                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
1267                 }
1268                 set_bit(Faulty, &rdev->flags);
1269                 printk (KERN_ALERT
1270                         "raid5: Disk failure on %s, disabling device."
1271                         " Operation continuing on %d devices\n",
1272                         bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1273         }
1274 }
1275
1276 /*
1277  * Input: a 'big' sector number,
1278  * Output: index of the data and parity disk, and the sector # in them.
1279  */
1280 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1281                         unsigned int data_disks, unsigned int * dd_idx,
1282                         unsigned int * pd_idx, raid5_conf_t *conf)
1283 {
1284         long stripe;
1285         unsigned long chunk_number;
1286         unsigned int chunk_offset;
1287         sector_t new_sector;
1288         int sectors_per_chunk = conf->chunk_size >> 9;
1289
1290         /* First compute the information on this sector */
1291
1292         /*
1293          * Compute the chunk number and the sector offset inside the chunk
1294          */
1295         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1296         chunk_number = r_sector;
1297         BUG_ON(r_sector != chunk_number);
1298
1299         /*
1300          * Compute the stripe number
1301          */
1302         stripe = chunk_number / data_disks;
1303
1304         /*
1305          * Compute the data disk and parity disk indexes inside the stripe
1306          */
1307         *dd_idx = chunk_number % data_disks;
1308
1309         /*
1310          * Select the parity disk based on the user selected algorithm.
1311          */
1312         switch(conf->level) {
1313         case 4:
1314                 *pd_idx = data_disks;
1315                 break;
1316         case 5:
1317                 switch (conf->algorithm) {
1318                 case ALGORITHM_LEFT_ASYMMETRIC:
1319                         *pd_idx = data_disks - stripe % raid_disks;
1320                         if (*dd_idx >= *pd_idx)
1321                                 (*dd_idx)++;
1322                         break;
1323                 case ALGORITHM_RIGHT_ASYMMETRIC:
1324                         *pd_idx = stripe % raid_disks;
1325                         if (*dd_idx >= *pd_idx)
1326                                 (*dd_idx)++;
1327                         break;
1328                 case ALGORITHM_LEFT_SYMMETRIC:
1329                         *pd_idx = data_disks - stripe % raid_disks;
1330                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1331                         break;
1332                 case ALGORITHM_RIGHT_SYMMETRIC:
1333                         *pd_idx = stripe % raid_disks;
1334                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1335                         break;
1336                 default:
1337                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1338                                 conf->algorithm);
1339                 }
1340                 break;
1341         case 6:
1342
1343                 /**** FIX THIS ****/
1344                 switch (conf->algorithm) {
1345                 case ALGORITHM_LEFT_ASYMMETRIC:
1346                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1347                         if (*pd_idx == raid_disks-1)
1348                                 (*dd_idx)++;    /* Q D D D P */
1349                         else if (*dd_idx >= *pd_idx)
1350                                 (*dd_idx) += 2; /* D D P Q D */
1351                         break;
1352                 case ALGORITHM_RIGHT_ASYMMETRIC:
1353                         *pd_idx = stripe % raid_disks;
1354                         if (*pd_idx == raid_disks-1)
1355                                 (*dd_idx)++;    /* Q D D D P */
1356                         else if (*dd_idx >= *pd_idx)
1357                                 (*dd_idx) += 2; /* D D P Q D */
1358                         break;
1359                 case ALGORITHM_LEFT_SYMMETRIC:
1360                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1361                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1362                         break;
1363                 case ALGORITHM_RIGHT_SYMMETRIC:
1364                         *pd_idx = stripe % raid_disks;
1365                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1366                         break;
1367                 default:
1368                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1369                                 conf->algorithm);
1370                 }
1371                 break;
1372         }
1373
1374         /*
1375          * Finally, compute the new sector number
1376          */
1377         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1378         return new_sector;
1379 }
1380
1381
1382 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1383 {
1384         raid5_conf_t *conf = sh->raid_conf;
1385         int raid_disks = sh->disks;
1386         int data_disks = raid_disks - conf->max_degraded;
1387         sector_t new_sector = sh->sector, check;
1388         int sectors_per_chunk = conf->chunk_size >> 9;
1389         sector_t stripe;
1390         int chunk_offset;
1391         int chunk_number, dummy1, dummy2, dd_idx = i;
1392         sector_t r_sector;
1393
1394
1395         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1396         stripe = new_sector;
1397         BUG_ON(new_sector != stripe);
1398
1399         if (i == sh->pd_idx)
1400                 return 0;
1401         switch(conf->level) {
1402         case 4: break;
1403         case 5:
1404                 switch (conf->algorithm) {
1405                 case ALGORITHM_LEFT_ASYMMETRIC:
1406                 case ALGORITHM_RIGHT_ASYMMETRIC:
1407                         if (i > sh->pd_idx)
1408                                 i--;
1409                         break;
1410                 case ALGORITHM_LEFT_SYMMETRIC:
1411                 case ALGORITHM_RIGHT_SYMMETRIC:
1412                         if (i < sh->pd_idx)
1413                                 i += raid_disks;
1414                         i -= (sh->pd_idx + 1);
1415                         break;
1416                 default:
1417                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1418                                conf->algorithm);
1419                 }
1420                 break;
1421         case 6:
1422                 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1423                         return 0; /* It is the Q disk */
1424                 switch (conf->algorithm) {
1425                 case ALGORITHM_LEFT_ASYMMETRIC:
1426                 case ALGORITHM_RIGHT_ASYMMETRIC:
1427                         if (sh->pd_idx == raid_disks-1)
1428                                 i--;    /* Q D D D P */
1429                         else if (i > sh->pd_idx)
1430                                 i -= 2; /* D D P Q D */
1431                         break;
1432                 case ALGORITHM_LEFT_SYMMETRIC:
1433                 case ALGORITHM_RIGHT_SYMMETRIC:
1434                         if (sh->pd_idx == raid_disks-1)
1435                                 i--; /* Q D D D P */
1436                         else {
1437                                 /* D D P Q D */
1438                                 if (i < sh->pd_idx)
1439                                         i += raid_disks;
1440                                 i -= (sh->pd_idx + 2);
1441                         }
1442                         break;
1443                 default:
1444                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1445                                 conf->algorithm);
1446                 }
1447                 break;
1448         }
1449
1450         chunk_number = stripe * data_disks + i;
1451         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1452
1453         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1454         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1455                 printk(KERN_ERR "compute_blocknr: map not correct\n");
1456                 return 0;
1457         }
1458         return r_sector;
1459 }
1460
1461
1462
1463 /*
1464  * Copy data between a page in the stripe cache, and one or more bion
1465  * The page could align with the middle of the bio, or there could be
1466  * several bion, each with several bio_vecs, which cover part of the page
1467  * Multiple bion are linked together on bi_next.  There may be extras
1468  * at the end of this list.  We ignore them.
1469  */
1470 static void copy_data(int frombio, struct bio *bio,
1471                      struct page *page,
1472                      sector_t sector)
1473 {
1474         char *pa = page_address(page);
1475         struct bio_vec *bvl;
1476         int i;
1477         int page_offset;
1478
1479         if (bio->bi_sector >= sector)
1480                 page_offset = (signed)(bio->bi_sector - sector) * 512;
1481         else
1482                 page_offset = (signed)(sector - bio->bi_sector) * -512;
1483         bio_for_each_segment(bvl, bio, i) {
1484                 int len = bio_iovec_idx(bio,i)->bv_len;
1485                 int clen;
1486                 int b_offset = 0;
1487
1488                 if (page_offset < 0) {
1489                         b_offset = -page_offset;
1490                         page_offset += b_offset;
1491                         len -= b_offset;
1492                 }
1493
1494                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1495                         clen = STRIPE_SIZE - page_offset;
1496                 else clen = len;
1497
1498                 if (clen > 0) {
1499                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1500                         if (frombio)
1501                                 memcpy(pa+page_offset, ba+b_offset, clen);
1502                         else
1503                                 memcpy(ba+b_offset, pa+page_offset, clen);
1504                         __bio_kunmap_atomic(ba, KM_USER0);
1505                 }
1506                 if (clen < len) /* hit end of page */
1507                         break;
1508                 page_offset +=  len;
1509         }
1510 }
1511
1512 #define check_xor()     do {                                              \
1513                                 if (count == MAX_XOR_BLOCKS) {            \
1514                                 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1515                                 count = 0;                                \
1516                            }                                              \
1517                         } while(0)
1518
1519 static void compute_parity6(struct stripe_head *sh, int method)
1520 {
1521         raid6_conf_t *conf = sh->raid_conf;
1522         int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1523         struct bio *chosen;
1524         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1525         void *ptrs[disks];
1526
1527         qd_idx = raid6_next_disk(pd_idx, disks);
1528         d0_idx = raid6_next_disk(qd_idx, disks);
1529
1530         pr_debug("compute_parity, stripe %llu, method %d\n",
1531                 (unsigned long long)sh->sector, method);
1532
1533         switch(method) {
1534         case READ_MODIFY_WRITE:
1535                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
1536         case RECONSTRUCT_WRITE:
1537                 for (i= disks; i-- ;)
1538                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1539                                 chosen = sh->dev[i].towrite;
1540                                 sh->dev[i].towrite = NULL;
1541
1542                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1543                                         wake_up(&conf->wait_for_overlap);
1544
1545                                 BUG_ON(sh->dev[i].written);
1546                                 sh->dev[i].written = chosen;
1547                         }
1548                 break;
1549         case CHECK_PARITY:
1550                 BUG();          /* Not implemented yet */
1551         }
1552
1553         for (i = disks; i--;)
1554                 if (sh->dev[i].written) {
1555                         sector_t sector = sh->dev[i].sector;
1556                         struct bio *wbi = sh->dev[i].written;
1557                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1558                                 copy_data(1, wbi, sh->dev[i].page, sector);
1559                                 wbi = r5_next_bio(wbi, sector);
1560                         }
1561
1562                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1563                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1564                 }
1565
1566 //      switch(method) {
1567 //      case RECONSTRUCT_WRITE:
1568 //      case CHECK_PARITY:
1569 //      case UPDATE_PARITY:
1570                 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1571                 /* FIX: Is this ordering of drives even remotely optimal? */
1572                 count = 0;
1573                 i = d0_idx;
1574                 do {
1575                         ptrs[count++] = page_address(sh->dev[i].page);
1576                         if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1577                                 printk("block %d/%d not uptodate on parity calc\n", i,count);
1578                         i = raid6_next_disk(i, disks);
1579                 } while ( i != d0_idx );
1580 //              break;
1581 //      }
1582
1583         raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1584
1585         switch(method) {
1586         case RECONSTRUCT_WRITE:
1587                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1588                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1589                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1590                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
1591                 break;
1592         case UPDATE_PARITY:
1593                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1594                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1595                 break;
1596         }
1597 }
1598
1599
1600 /* Compute one missing block */
1601 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1602 {
1603         int i, count, disks = sh->disks;
1604         void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1605         int pd_idx = sh->pd_idx;
1606         int qd_idx = raid6_next_disk(pd_idx, disks);
1607
1608         pr_debug("compute_block_1, stripe %llu, idx %d\n",
1609                 (unsigned long long)sh->sector, dd_idx);
1610
1611         if ( dd_idx == qd_idx ) {
1612                 /* We're actually computing the Q drive */
1613                 compute_parity6(sh, UPDATE_PARITY);
1614         } else {
1615                 dest = page_address(sh->dev[dd_idx].page);
1616                 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1617                 count = 0;
1618                 for (i = disks ; i--; ) {
1619                         if (i == dd_idx || i == qd_idx)
1620                                 continue;
1621                         p = page_address(sh->dev[i].page);
1622                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1623                                 ptr[count++] = p;
1624                         else
1625                                 printk("compute_block() %d, stripe %llu, %d"
1626                                        " not present\n", dd_idx,
1627                                        (unsigned long long)sh->sector, i);
1628
1629                         check_xor();
1630                 }
1631                 if (count)
1632                         xor_blocks(count, STRIPE_SIZE, dest, ptr);
1633                 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1634                 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1635         }
1636 }
1637
1638 /* Compute two missing blocks */
1639 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1640 {
1641         int i, count, disks = sh->disks;
1642         int pd_idx = sh->pd_idx;
1643         int qd_idx = raid6_next_disk(pd_idx, disks);
1644         int d0_idx = raid6_next_disk(qd_idx, disks);
1645         int faila, failb;
1646
1647         /* faila and failb are disk numbers relative to d0_idx */
1648         /* pd_idx become disks-2 and qd_idx become disks-1 */
1649         faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1650         failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1651
1652         BUG_ON(faila == failb);
1653         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1654
1655         pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1656                (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1657
1658         if ( failb == disks-1 ) {
1659                 /* Q disk is one of the missing disks */
1660                 if ( faila == disks-2 ) {
1661                         /* Missing P+Q, just recompute */
1662                         compute_parity6(sh, UPDATE_PARITY);
1663                         return;
1664                 } else {
1665                         /* We're missing D+Q; recompute D from P */
1666                         compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1667                         compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1668                         return;
1669                 }
1670         }
1671
1672         /* We're missing D+P or D+D; build pointer table */
1673         {
1674                 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1675                 void *ptrs[disks];
1676
1677                 count = 0;
1678                 i = d0_idx;
1679                 do {
1680                         ptrs[count++] = page_address(sh->dev[i].page);
1681                         i = raid6_next_disk(i, disks);
1682                         if (i != dd_idx1 && i != dd_idx2 &&
1683                             !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1684                                 printk("compute_2 with missing block %d/%d\n", count, i);
1685                 } while ( i != d0_idx );
1686
1687                 if ( failb == disks-2 ) {
1688                         /* We're missing D+P. */
1689                         raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1690                 } else {
1691                         /* We're missing D+D. */
1692                         raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1693                 }
1694
1695                 /* Both the above update both missing blocks */
1696                 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1697                 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1698         }
1699 }
1700
1701 static int
1702 handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
1703 {
1704         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1705         int locked = 0;
1706
1707         if (rcw) {
1708                 /* if we are not expanding this is a proper write request, and
1709                  * there will be bios with new data to be drained into the
1710                  * stripe cache
1711                  */
1712                 if (!expand) {
1713                         set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1714                         sh->ops.count++;
1715                 }
1716
1717                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1718                 sh->ops.count++;
1719
1720                 for (i = disks; i--; ) {
1721                         struct r5dev *dev = &sh->dev[i];
1722
1723                         if (dev->towrite) {
1724                                 set_bit(R5_LOCKED, &dev->flags);
1725                                 if (!expand)
1726                                         clear_bit(R5_UPTODATE, &dev->flags);
1727                                 locked++;
1728                         }
1729                 }
1730         } else {
1731                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1732                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1733
1734                 set_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
1735                 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1736                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1737
1738                 sh->ops.count += 3;
1739
1740                 for (i = disks; i--; ) {
1741                         struct r5dev *dev = &sh->dev[i];
1742                         if (i == pd_idx)
1743                                 continue;
1744
1745                         /* For a read-modify write there may be blocks that are
1746                          * locked for reading while others are ready to be
1747                          * written so we distinguish these blocks by the
1748                          * R5_Wantprexor bit
1749                          */
1750                         if (dev->towrite &&
1751                             (test_bit(R5_UPTODATE, &dev->flags) ||
1752                             test_bit(R5_Wantcompute, &dev->flags))) {
1753                                 set_bit(R5_Wantprexor, &dev->flags);
1754                                 set_bit(R5_LOCKED, &dev->flags);
1755                                 clear_bit(R5_UPTODATE, &dev->flags);
1756                                 locked++;
1757                         }
1758                 }
1759         }
1760
1761         /* keep the parity disk locked while asynchronous operations
1762          * are in flight
1763          */
1764         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1765         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1766         locked++;
1767
1768         pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1769                 __FUNCTION__, (unsigned long long)sh->sector,
1770                 locked, sh->ops.pending);
1771
1772         return locked;
1773 }
1774
1775 /*
1776  * Each stripe/dev can have one or more bion attached.
1777  * toread/towrite point to the first in a chain.
1778  * The bi_next chain must be in order.
1779  */
1780 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1781 {
1782         struct bio **bip;
1783         raid5_conf_t *conf = sh->raid_conf;
1784         int firstwrite=0;
1785
1786         pr_debug("adding bh b#%llu to stripe s#%llu\n",
1787                 (unsigned long long)bi->bi_sector,
1788                 (unsigned long long)sh->sector);
1789
1790
1791         spin_lock(&sh->lock);
1792         spin_lock_irq(&conf->device_lock);
1793         if (forwrite) {
1794                 bip = &sh->dev[dd_idx].towrite;
1795                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1796                         firstwrite = 1;
1797         } else
1798                 bip = &sh->dev[dd_idx].toread;
1799         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1800                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1801                         goto overlap;
1802                 bip = & (*bip)->bi_next;
1803         }
1804         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1805                 goto overlap;
1806
1807         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1808         if (*bip)
1809                 bi->bi_next = *bip;
1810         *bip = bi;
1811         bi->bi_phys_segments ++;
1812         spin_unlock_irq(&conf->device_lock);
1813         spin_unlock(&sh->lock);
1814
1815         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1816                 (unsigned long long)bi->bi_sector,
1817                 (unsigned long long)sh->sector, dd_idx);
1818
1819         if (conf->mddev->bitmap && firstwrite) {
1820                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1821                                   STRIPE_SECTORS, 0);
1822                 sh->bm_seq = conf->seq_flush+1;
1823                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1824         }
1825
1826         if (forwrite) {
1827                 /* check if page is covered */
1828                 sector_t sector = sh->dev[dd_idx].sector;
1829                 for (bi=sh->dev[dd_idx].towrite;
1830                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1831                              bi && bi->bi_sector <= sector;
1832                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1833                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1834                                 sector = bi->bi_sector + (bi->bi_size>>9);
1835                 }
1836                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1837                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1838         }
1839         return 1;
1840
1841  overlap:
1842         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1843         spin_unlock_irq(&conf->device_lock);
1844         spin_unlock(&sh->lock);
1845         return 0;
1846 }
1847
1848 static void end_reshape(raid5_conf_t *conf);
1849
1850 static int page_is_zero(struct page *p)
1851 {
1852         char *a = page_address(p);
1853         return ((*(u32*)a) == 0 &&
1854                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1855 }
1856
1857 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1858 {
1859         int sectors_per_chunk = conf->chunk_size >> 9;
1860         int pd_idx, dd_idx;
1861         int chunk_offset = sector_div(stripe, sectors_per_chunk);
1862
1863         raid5_compute_sector(stripe * (disks - conf->max_degraded)
1864                              *sectors_per_chunk + chunk_offset,
1865                              disks, disks - conf->max_degraded,
1866                              &dd_idx, &pd_idx, conf);
1867         return pd_idx;
1868 }
1869
1870 static void
1871 handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1872                                 struct stripe_head_state *s, int disks,
1873                                 struct bio **return_bi)
1874 {
1875         int i;
1876         for (i = disks; i--; ) {
1877                 struct bio *bi;
1878                 int bitmap_end = 0;
1879
1880                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1881                         mdk_rdev_t *rdev;
1882                         rcu_read_lock();
1883                         rdev = rcu_dereference(conf->disks[i].rdev);
1884                         if (rdev && test_bit(In_sync, &rdev->flags))
1885                                 /* multiple read failures in one stripe */
1886                                 md_error(conf->mddev, rdev);
1887                         rcu_read_unlock();
1888                 }
1889                 spin_lock_irq(&conf->device_lock);
1890                 /* fail all writes first */
1891                 bi = sh->dev[i].towrite;
1892                 sh->dev[i].towrite = NULL;
1893                 if (bi) {
1894                         s->to_write--;
1895                         bitmap_end = 1;
1896                 }
1897
1898                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1899                         wake_up(&conf->wait_for_overlap);
1900
1901                 while (bi && bi->bi_sector <
1902                         sh->dev[i].sector + STRIPE_SECTORS) {
1903                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1904                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1905                         if (--bi->bi_phys_segments == 0) {
1906                                 md_write_end(conf->mddev);
1907                                 bi->bi_next = *return_bi;
1908                                 *return_bi = bi;
1909                         }
1910                         bi = nextbi;
1911                 }
1912                 /* and fail all 'written' */
1913                 bi = sh->dev[i].written;
1914                 sh->dev[i].written = NULL;
1915                 if (bi) bitmap_end = 1;
1916                 while (bi && bi->bi_sector <
1917                        sh->dev[i].sector + STRIPE_SECTORS) {
1918                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1919                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1920                         if (--bi->bi_phys_segments == 0) {
1921                                 md_write_end(conf->mddev);
1922                                 bi->bi_next = *return_bi;
1923                                 *return_bi = bi;
1924                         }
1925                         bi = bi2;
1926                 }
1927
1928                 /* fail any reads if this device is non-operational and
1929                  * the data has not reached the cache yet.
1930                  */
1931                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
1932                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1933                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
1934                         bi = sh->dev[i].toread;
1935                         sh->dev[i].toread = NULL;
1936                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1937                                 wake_up(&conf->wait_for_overlap);
1938                         if (bi) s->to_read--;
1939                         while (bi && bi->bi_sector <
1940                                sh->dev[i].sector + STRIPE_SECTORS) {
1941                                 struct bio *nextbi =
1942                                         r5_next_bio(bi, sh->dev[i].sector);
1943                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1944                                 if (--bi->bi_phys_segments == 0) {
1945                                         bi->bi_next = *return_bi;
1946                                         *return_bi = bi;
1947                                 }
1948                                 bi = nextbi;
1949                         }
1950                 }
1951                 spin_unlock_irq(&conf->device_lock);
1952                 if (bitmap_end)
1953                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1954                                         STRIPE_SECTORS, 0, 0);
1955         }
1956
1957 }
1958
1959 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
1960  * to process
1961  */
1962 static int __handle_issuing_new_read_requests5(struct stripe_head *sh,
1963                         struct stripe_head_state *s, int disk_idx, int disks)
1964 {
1965         struct r5dev *dev = &sh->dev[disk_idx];
1966         struct r5dev *failed_dev = &sh->dev[s->failed_num];
1967
1968         /* don't schedule compute operations or reads on the parity block while
1969          * a check is in flight
1970          */
1971         if ((disk_idx == sh->pd_idx) &&
1972              test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
1973                 return ~0;
1974
1975         /* is the data in this block needed, and can we get it? */
1976         if (!test_bit(R5_LOCKED, &dev->flags) &&
1977             !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread ||
1978             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1979              s->syncing || s->expanding || (s->failed &&
1980              (failed_dev->toread || (failed_dev->towrite &&
1981              !test_bit(R5_OVERWRITE, &failed_dev->flags)
1982              ))))) {
1983                 /* 1/ We would like to get this block, possibly by computing it,
1984                  * but we might not be able to.
1985                  *
1986                  * 2/ Since parity check operations potentially make the parity
1987                  * block !uptodate it will need to be refreshed before any
1988                  * compute operations on data disks are scheduled.
1989                  *
1990                  * 3/ We hold off parity block re-reads until check operations
1991                  * have quiesced.
1992                  */
1993                 if ((s->uptodate == disks - 1) &&
1994                     !test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
1995                         set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
1996                         set_bit(R5_Wantcompute, &dev->flags);
1997                         sh->ops.target = disk_idx;
1998                         s->req_compute = 1;
1999                         sh->ops.count++;
2000                         /* Careful: from this point on 'uptodate' is in the eye
2001                          * of raid5_run_ops which services 'compute' operations
2002                          * before writes. R5_Wantcompute flags a block that will
2003                          * be R5_UPTODATE by the time it is needed for a
2004                          * subsequent operation.
2005                          */
2006                         s->uptodate++;
2007                         return 0; /* uptodate + compute == disks */
2008                 } else if ((s->uptodate < disks - 1) &&
2009                         test_bit(R5_Insync, &dev->flags)) {
2010                         /* Note: we hold off compute operations while checks are
2011                          * in flight, but we still prefer 'compute' over 'read'
2012                          * hence we only read if (uptodate < * disks-1)
2013                          */
2014                         set_bit(R5_LOCKED, &dev->flags);
2015                         set_bit(R5_Wantread, &dev->flags);
2016                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2017                                 sh->ops.count++;
2018                         s->locked++;
2019                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2020                                 s->syncing);
2021                 }
2022         }
2023
2024         return ~0;
2025 }
2026
2027 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2028                         struct stripe_head_state *s, int disks)
2029 {
2030         int i;
2031
2032         /* Clear completed compute operations.  Parity recovery
2033          * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled
2034          * later on in this routine
2035          */
2036         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2037                 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2038                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2039                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2040                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2041         }
2042
2043         /* look for blocks to read/compute, skip this if a compute
2044          * is already in flight, or if the stripe contents are in the
2045          * midst of changing due to a write
2046          */
2047         if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2048                 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2049                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2050                 for (i = disks; i--; )
2051                         if (__handle_issuing_new_read_requests5(
2052                                 sh, s, i, disks) == 0)
2053                                 break;
2054         }
2055         set_bit(STRIPE_HANDLE, &sh->state);
2056 }
2057
2058 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2059                         struct stripe_head_state *s, struct r6_state *r6s,
2060                         int disks)
2061 {
2062         int i;
2063         for (i = disks; i--; ) {
2064                 struct r5dev *dev = &sh->dev[i];
2065                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2066                     !test_bit(R5_UPTODATE, &dev->flags) &&
2067                     (dev->toread || (dev->towrite &&
2068                      !test_bit(R5_OVERWRITE, &dev->flags)) ||
2069                      s->syncing || s->expanding ||
2070                      (s->failed >= 1 &&
2071                       (sh->dev[r6s->failed_num[0]].toread ||
2072                        s->to_write)) ||
2073                      (s->failed >= 2 &&
2074                       (sh->dev[r6s->failed_num[1]].toread ||
2075                        s->to_write)))) {
2076                         /* we would like to get this block, possibly
2077                          * by computing it, but we might not be able to
2078                          */
2079                         if (s->uptodate == disks-1) {
2080                                 pr_debug("Computing stripe %llu block %d\n",
2081                                        (unsigned long long)sh->sector, i);
2082                                 compute_block_1(sh, i, 0);
2083                                 s->uptodate++;
2084                         } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2085                                 /* Computing 2-failure is *very* expensive; only
2086                                  * do it if failed >= 2
2087                                  */
2088                                 int other;
2089                                 for (other = disks; other--; ) {
2090                                         if (other == i)
2091                                                 continue;
2092                                         if (!test_bit(R5_UPTODATE,
2093                                               &sh->dev[other].flags))
2094                                                 break;
2095                                 }
2096                                 BUG_ON(other < 0);
2097                                 pr_debug("Computing stripe %llu blocks %d,%d\n",
2098                                        (unsigned long long)sh->sector,
2099                                        i, other);
2100                                 compute_block_2(sh, i, other);
2101                                 s->uptodate += 2;
2102                         } else if (test_bit(R5_Insync, &dev->flags)) {
2103                                 set_bit(R5_LOCKED, &dev->flags);
2104                                 set_bit(R5_Wantread, &dev->flags);
2105                                 s->locked++;
2106                                 pr_debug("Reading block %d (sync=%d)\n",
2107                                         i, s->syncing);
2108                         }
2109                 }
2110         }
2111         set_bit(STRIPE_HANDLE, &sh->state);
2112 }
2113
2114
2115 /* handle_completed_write_requests
2116  * any written block on an uptodate or failed drive can be returned.
2117  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2118  * never LOCKED, so we don't need to test 'failed' directly.
2119  */
2120 static void handle_completed_write_requests(raid5_conf_t *conf,
2121         struct stripe_head *sh, int disks, struct bio **return_bi)
2122 {
2123         int i;
2124         struct r5dev *dev;
2125
2126         for (i = disks; i--; )
2127                 if (sh->dev[i].written) {
2128                         dev = &sh->dev[i];
2129                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2130                                 test_bit(R5_UPTODATE, &dev->flags)) {
2131                                 /* We can return any write requests */
2132                                 struct bio *wbi, *wbi2;
2133                                 int bitmap_end = 0;
2134                                 pr_debug("Return write for disc %d\n", i);
2135                                 spin_lock_irq(&conf->device_lock);
2136                                 wbi = dev->written;
2137                                 dev->written = NULL;
2138                                 while (wbi && wbi->bi_sector <
2139                                         dev->sector + STRIPE_SECTORS) {
2140                                         wbi2 = r5_next_bio(wbi, dev->sector);
2141                                         if (--wbi->bi_phys_segments == 0) {
2142                                                 md_write_end(conf->mddev);
2143                                                 wbi->bi_next = *return_bi;
2144                                                 *return_bi = wbi;
2145                                         }
2146                                         wbi = wbi2;
2147                                 }
2148                                 if (dev->towrite == NULL)
2149                                         bitmap_end = 1;
2150                                 spin_unlock_irq(&conf->device_lock);
2151                                 if (bitmap_end)
2152                                         bitmap_endwrite(conf->mddev->bitmap,
2153                                                         sh->sector,
2154                                                         STRIPE_SECTORS,
2155                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2156                                                         0);
2157                         }
2158                 }
2159 }
2160
2161 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2162                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2163 {
2164         int rmw = 0, rcw = 0, i;
2165         for (i = disks; i--; ) {
2166                 /* would I have to read this buffer for read_modify_write */
2167                 struct r5dev *dev = &sh->dev[i];
2168                 if ((dev->towrite || i == sh->pd_idx) &&
2169                     !test_bit(R5_LOCKED, &dev->flags) &&
2170                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2171                       test_bit(R5_Wantcompute, &dev->flags))) {
2172                         if (test_bit(R5_Insync, &dev->flags))
2173                                 rmw++;
2174                         else
2175                                 rmw += 2*disks;  /* cannot read it */
2176                 }
2177                 /* Would I have to read this buffer for reconstruct_write */
2178                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2179                     !test_bit(R5_LOCKED, &dev->flags) &&
2180                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2181                     test_bit(R5_Wantcompute, &dev->flags))) {
2182                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2183                         else
2184                                 rcw += 2*disks;
2185                 }
2186         }
2187         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2188                 (unsigned long long)sh->sector, rmw, rcw);
2189         set_bit(STRIPE_HANDLE, &sh->state);
2190         if (rmw < rcw && rmw > 0)
2191                 /* prefer read-modify-write, but need to get some data */
2192                 for (i = disks; i--; ) {
2193                         struct r5dev *dev = &sh->dev[i];
2194                         if ((dev->towrite || i == sh->pd_idx) &&
2195                             !test_bit(R5_LOCKED, &dev->flags) &&
2196                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2197                             test_bit(R5_Wantcompute, &dev->flags)) &&
2198                             test_bit(R5_Insync, &dev->flags)) {
2199                                 if (
2200                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2201                                         pr_debug("Read_old block "
2202                                                 "%d for r-m-w\n", i);
2203                                         set_bit(R5_LOCKED, &dev->flags);
2204                                         set_bit(R5_Wantread, &dev->flags);
2205                                         if (!test_and_set_bit(
2206                                                 STRIPE_OP_IO, &sh->ops.pending))
2207                                                 sh->ops.count++;
2208                                         s->locked++;
2209                                 } else {
2210                                         set_bit(STRIPE_DELAYED, &sh->state);
2211                                         set_bit(STRIPE_HANDLE, &sh->state);
2212                                 }
2213                         }
2214                 }
2215         if (rcw <= rmw && rcw > 0)
2216                 /* want reconstruct write, but need to get some data */
2217                 for (i = disks; i--; ) {
2218                         struct r5dev *dev = &sh->dev[i];
2219                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2220                             i != sh->pd_idx &&
2221                             !test_bit(R5_LOCKED, &dev->flags) &&
2222                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2223                             test_bit(R5_Wantcompute, &dev->flags)) &&
2224                             test_bit(R5_Insync, &dev->flags)) {
2225                                 if (
2226                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2227                                         pr_debug("Read_old block "
2228                                                 "%d for Reconstruct\n", i);
2229                                         set_bit(R5_LOCKED, &dev->flags);
2230                                         set_bit(R5_Wantread, &dev->flags);
2231                                         if (!test_and_set_bit(
2232                                                 STRIPE_OP_IO, &sh->ops.pending))
2233                                                 sh->ops.count++;
2234                                         s->locked++;
2235                                 } else {
2236                                         set_bit(STRIPE_DELAYED, &sh->state);
2237                                         set_bit(STRIPE_HANDLE, &sh->state);
2238                                 }
2239                         }
2240                 }
2241         /* now if nothing is locked, and if we have enough data,
2242          * we can start a write request
2243          */
2244         /* since handle_stripe can be called at any time we need to handle the
2245          * case where a compute block operation has been submitted and then a
2246          * subsequent call wants to start a write request.  raid5_run_ops only
2247          * handles the case where compute block and postxor are requested
2248          * simultaneously.  If this is not the case then new writes need to be
2249          * held off until the compute completes.
2250          */
2251         if ((s->req_compute ||
2252             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) &&
2253                 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2254                 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2255                 s->locked += handle_write_operations5(sh, rcw == 0, 0);
2256 }
2257
2258 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2259                 struct stripe_head *sh, struct stripe_head_state *s,
2260                 struct r6_state *r6s, int disks)
2261 {
2262         int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2263         int qd_idx = r6s->qd_idx;
2264         for (i = disks; i--; ) {
2265                 struct r5dev *dev = &sh->dev[i];
2266                 /* Would I have to read this buffer for reconstruct_write */
2267                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2268                     && i != pd_idx && i != qd_idx
2269                     && (!test_bit(R5_LOCKED, &dev->flags)
2270                             ) &&
2271                     !test_bit(R5_UPTODATE, &dev->flags)) {
2272                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2273                         else {
2274                                 pr_debug("raid6: must_compute: "
2275                                         "disk %d flags=%#lx\n", i, dev->flags);
2276                                 must_compute++;
2277                         }
2278                 }
2279         }
2280         pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2281                (unsigned long long)sh->sector, rcw, must_compute);
2282         set_bit(STRIPE_HANDLE, &sh->state);
2283
2284         if (rcw > 0)
2285                 /* want reconstruct write, but need to get some data */
2286                 for (i = disks; i--; ) {
2287                         struct r5dev *dev = &sh->dev[i];
2288                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2289                             && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2290                             && !test_bit(R5_LOCKED, &dev->flags) &&
2291                             !test_bit(R5_UPTODATE, &dev->flags) &&
2292                             test_bit(R5_Insync, &dev->flags)) {
2293                                 if (
2294                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2295                                         pr_debug("Read_old stripe %llu "
2296                                                 "block %d for Reconstruct\n",
2297                                              (unsigned long long)sh->sector, i);
2298                                         set_bit(R5_LOCKED, &dev->flags);
2299                                         set_bit(R5_Wantread, &dev->flags);
2300                                         s->locked++;
2301                                 } else {
2302                                         pr_debug("Request delayed stripe %llu "
2303                                                 "block %d for Reconstruct\n",
2304                                              (unsigned long long)sh->sector, i);
2305                                         set_bit(STRIPE_DELAYED, &sh->state);
2306                                         set_bit(STRIPE_HANDLE, &sh->state);
2307                                 }
2308                         }
2309                 }
2310         /* now if nothing is locked, and if we have enough data, we can start a
2311          * write request
2312          */
2313         if (s->locked == 0 && rcw == 0 &&
2314             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2315                 if (must_compute > 0) {
2316                         /* We have failed blocks and need to compute them */
2317                         switch (s->failed) {
2318                         case 0:
2319                                 BUG();
2320                         case 1:
2321                                 compute_block_1(sh, r6s->failed_num[0], 0);
2322                                 break;
2323                         case 2:
2324                                 compute_block_2(sh, r6s->failed_num[0],
2325                                                 r6s->failed_num[1]);
2326                                 break;
2327                         default: /* This request should have been failed? */
2328                                 BUG();
2329                         }
2330                 }
2331
2332                 pr_debug("Computing parity for stripe %llu\n",
2333                         (unsigned long long)sh->sector);
2334                 compute_parity6(sh, RECONSTRUCT_WRITE);
2335                 /* now every locked buffer is ready to be written */
2336                 for (i = disks; i--; )
2337                         if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2338                                 pr_debug("Writing stripe %llu block %d\n",
2339                                        (unsigned long long)sh->sector, i);
2340                                 s->locked++;
2341                                 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2342                         }
2343                 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2344                 set_bit(STRIPE_INSYNC, &sh->state);
2345
2346                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2347                         atomic_dec(&conf->preread_active_stripes);
2348                         if (atomic_read(&conf->preread_active_stripes) <
2349                             IO_THRESHOLD)
2350                                 md_wakeup_thread(conf->mddev->thread);
2351                 }
2352         }
2353 }
2354
2355 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2356                                 struct stripe_head_state *s, int disks)
2357 {
2358         set_bit(STRIPE_HANDLE, &sh->state);
2359         /* Take one of the following actions:
2360          * 1/ start a check parity operation if (uptodate == disks)
2361          * 2/ finish a check parity operation and act on the result
2362          * 3/ skip to the writeback section if we previously
2363          *    initiated a recovery operation
2364          */
2365         if (s->failed == 0 &&
2366             !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2367                 if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2368                         BUG_ON(s->uptodate != disks);
2369                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2370                         sh->ops.count++;
2371                         s->uptodate--;
2372                 } else if (
2373                        test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) {
2374                         clear_bit(STRIPE_OP_CHECK, &sh->ops.ack);
2375                         clear_bit(STRIPE_OP_CHECK, &sh->ops.pending);
2376
2377                         if (sh->ops.zero_sum_result == 0)
2378                                 /* parity is correct (on disc,
2379                                  * not in buffer any more)
2380                                  */
2381                                 set_bit(STRIPE_INSYNC, &sh->state);
2382                         else {
2383                                 conf->mddev->resync_mismatches +=
2384                                         STRIPE_SECTORS;
2385                                 if (test_bit(
2386                                      MD_RECOVERY_CHECK, &conf->mddev->recovery))
2387                                         /* don't try to repair!! */
2388                                         set_bit(STRIPE_INSYNC, &sh->state);
2389                                 else {
2390                                         set_bit(STRIPE_OP_COMPUTE_BLK,
2391                                                 &sh->ops.pending);
2392                                         set_bit(STRIPE_OP_MOD_REPAIR_PD,
2393                                                 &sh->ops.pending);
2394                                         set_bit(R5_Wantcompute,
2395                                                 &sh->dev[sh->pd_idx].flags);
2396                                         sh->ops.target = sh->pd_idx;
2397                                         sh->ops.count++;
2398                                         s->uptodate++;
2399                                 }
2400                         }
2401                 }
2402         }
2403
2404         /* check if we can clear a parity disk reconstruct */
2405         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2406                 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2407
2408                 clear_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending);
2409                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2410                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2411                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2412         }
2413
2414         /* Wait for check parity and compute block operations to complete
2415          * before write-back
2416          */
2417         if (!test_bit(STRIPE_INSYNC, &sh->state) &&
2418                 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) &&
2419                 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) {
2420                 struct r5dev *dev;
2421                 /* either failed parity check, or recovery is happening */
2422                 if (s->failed == 0)
2423                         s->failed_num = sh->pd_idx;
2424                 dev = &sh->dev[s->failed_num];
2425                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2426                 BUG_ON(s->uptodate != disks);
2427
2428                 set_bit(R5_LOCKED, &dev->flags);
2429                 set_bit(R5_Wantwrite, &dev->flags);
2430                 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2431                         sh->ops.count++;
2432
2433                 clear_bit(STRIPE_DEGRADED, &sh->state);
2434                 s->locked++;
2435                 set_bit(STRIPE_INSYNC, &sh->state);
2436         }
2437 }
2438
2439
2440 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2441                                 struct stripe_head_state *s,
2442                                 struct r6_state *r6s, struct page *tmp_page,
2443                                 int disks)
2444 {
2445         int update_p = 0, update_q = 0;
2446         struct r5dev *dev;
2447         int pd_idx = sh->pd_idx;
2448         int qd_idx = r6s->qd_idx;
2449
2450         set_bit(STRIPE_HANDLE, &sh->state);
2451
2452         BUG_ON(s->failed > 2);
2453         BUG_ON(s->uptodate < disks);
2454         /* Want to check and possibly repair P and Q.
2455          * However there could be one 'failed' device, in which
2456          * case we can only check one of them, possibly using the
2457          * other to generate missing data
2458          */
2459
2460         /* If !tmp_page, we cannot do the calculations,
2461          * but as we have set STRIPE_HANDLE, we will soon be called
2462          * by stripe_handle with a tmp_page - just wait until then.
2463          */
2464         if (tmp_page) {
2465                 if (s->failed == r6s->q_failed) {
2466                         /* The only possible failed device holds 'Q', so it
2467                          * makes sense to check P (If anything else were failed,
2468                          * we would have used P to recreate it).
2469                          */
2470                         compute_block_1(sh, pd_idx, 1);
2471                         if (!page_is_zero(sh->dev[pd_idx].page)) {
2472                                 compute_block_1(sh, pd_idx, 0);
2473                                 update_p = 1;
2474                         }
2475                 }
2476                 if (!r6s->q_failed && s->failed < 2) {
2477                         /* q is not failed, and we didn't use it to generate
2478                          * anything, so it makes sense to check it
2479                          */
2480                         memcpy(page_address(tmp_page),
2481                                page_address(sh->dev[qd_idx].page),
2482                                STRIPE_SIZE);
2483                         compute_parity6(sh, UPDATE_PARITY);
2484                         if (memcmp(page_address(tmp_page),
2485                                    page_address(sh->dev[qd_idx].page),
2486                                    STRIPE_SIZE) != 0) {
2487                                 clear_bit(STRIPE_INSYNC, &sh->state);
2488                                 update_q = 1;
2489                         }
2490                 }
2491                 if (update_p || update_q) {
2492                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2493                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2494                                 /* don't try to repair!! */
2495                                 update_p = update_q = 0;
2496                 }
2497
2498                 /* now write out any block on a failed drive,
2499                  * or P or Q if they need it
2500                  */
2501
2502                 if (s->failed == 2) {
2503                         dev = &sh->dev[r6s->failed_num[1]];
2504                         s->locked++;
2505                         set_bit(R5_LOCKED, &dev->flags);
2506                         set_bit(R5_Wantwrite, &dev->flags);
2507                 }
2508                 if (s->failed >= 1) {
2509                         dev = &sh->dev[r6s->failed_num[0]];
2510                         s->locked++;
2511                         set_bit(R5_LOCKED, &dev->flags);
2512                         set_bit(R5_Wantwrite, &dev->flags);
2513                 }
2514
2515                 if (update_p) {
2516                         dev = &sh->dev[pd_idx];
2517                         s->locked++;
2518                         set_bit(R5_LOCKED, &dev->flags);
2519                         set_bit(R5_Wantwrite, &dev->flags);
2520                 }
2521                 if (update_q) {
2522                         dev = &sh->dev[qd_idx];
2523                         s->locked++;
2524                         set_bit(R5_LOCKED, &dev->flags);
2525                         set_bit(R5_Wantwrite, &dev->flags);
2526                 }
2527                 clear_bit(STRIPE_DEGRADED, &sh->state);
2528
2529                 set_bit(STRIPE_INSYNC, &sh->state);
2530         }
2531 }
2532
2533 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2534                                 struct r6_state *r6s)
2535 {
2536         int i;
2537
2538         /* We have read all the blocks in this stripe and now we need to
2539          * copy some of them into a target stripe for expand.
2540          */
2541         struct dma_async_tx_descriptor *tx = NULL;
2542         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2543         for (i = 0; i < sh->disks; i++)
2544                 if (i != sh->pd_idx && (r6s && i != r6s->qd_idx)) {
2545                         int dd_idx, pd_idx, j;
2546                         struct stripe_head *sh2;
2547
2548                         sector_t bn = compute_blocknr(sh, i);
2549                         sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2550                                                 conf->raid_disks -
2551                                                 conf->max_degraded, &dd_idx,
2552                                                 &pd_idx, conf);
2553                         sh2 = get_active_stripe(conf, s, conf->raid_disks,
2554                                                 pd_idx, 1);
2555                         if (sh2 == NULL)
2556                                 /* so far only the early blocks of this stripe
2557                                  * have been requested.  When later blocks
2558                                  * get requested, we will try again
2559                                  */
2560                                 continue;
2561                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2562                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2563                                 /* must have already done this block */
2564                                 release_stripe(sh2);
2565                                 continue;
2566                         }
2567
2568                         /* place all the copies on one channel */
2569                         tx = async_memcpy(sh2->dev[dd_idx].page,
2570                                 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2571                                 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2572
2573                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2574                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2575                         for (j = 0; j < conf->raid_disks; j++)
2576                                 if (j != sh2->pd_idx &&
2577                                     (r6s && j != r6s->qd_idx) &&
2578                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2579                                         break;
2580                         if (j == conf->raid_disks) {
2581                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2582                                 set_bit(STRIPE_HANDLE, &sh2->state);
2583                         }
2584                         release_stripe(sh2);
2585
2586                         /* done submitting copies, wait for them to complete */
2587                         if (i + 1 >= sh->disks) {
2588                                 async_tx_ack(tx);
2589                                 dma_wait_for_async_tx(tx);
2590                         }
2591                 }
2592 }
2593
2594 /*
2595  * handle_stripe - do things to a stripe.
2596  *
2597  * We lock the stripe and then examine the state of various bits
2598  * to see what needs to be done.
2599  * Possible results:
2600  *    return some read request which now have data
2601  *    return some write requests which are safely on disc
2602  *    schedule a read on some buffers
2603  *    schedule a write of some buffers
2604  *    return confirmation of parity correctness
2605  *
2606  * buffers are taken off read_list or write_list, and bh_cache buffers
2607  * get BH_Lock set before the stripe lock is released.
2608  *
2609  */
2610
2611 static void handle_stripe5(struct stripe_head *sh)
2612 {
2613         raid5_conf_t *conf = sh->raid_conf;
2614         int disks = sh->disks, i;
2615         struct bio *return_bi = NULL;
2616         struct stripe_head_state s;
2617         struct r5dev *dev;
2618         unsigned long pending = 0;
2619
2620         memset(&s, 0, sizeof(s));
2621         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2622                 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2623                 atomic_read(&sh->count), sh->pd_idx,
2624                 sh->ops.pending, sh->ops.ack, sh->ops.complete);
2625
2626         spin_lock(&sh->lock);
2627         clear_bit(STRIPE_HANDLE, &sh->state);
2628         clear_bit(STRIPE_DELAYED, &sh->state);
2629
2630         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2631         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2632         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2633         /* Now to look around and see what can be done */
2634
2635         rcu_read_lock();
2636         for (i=disks; i--; ) {
2637                 mdk_rdev_t *rdev;
2638                 struct r5dev *dev = &sh->dev[i];
2639                 clear_bit(R5_Insync, &dev->flags);
2640
2641                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2642                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2643                         dev->towrite, dev->written);
2644
2645                 /* maybe we can request a biofill operation
2646                  *
2647                  * new wantfill requests are only permitted while
2648                  * STRIPE_OP_BIOFILL is clear
2649                  */
2650                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2651                         !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2652                         set_bit(R5_Wantfill, &dev->flags);
2653
2654                 /* now count some things */
2655                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2656                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2657                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2658
2659                 if (test_bit(R5_Wantfill, &dev->flags))
2660                         s.to_fill++;
2661                 else if (dev->toread)
2662                         s.to_read++;
2663                 if (dev->towrite) {
2664                         s.to_write++;
2665                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2666                                 s.non_overwrite++;
2667                 }
2668                 if (dev->written)
2669                         s.written++;
2670                 rdev = rcu_dereference(conf->disks[i].rdev);
2671                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2672                         /* The ReadError flag will just be confusing now */
2673                         clear_bit(R5_ReadError, &dev->flags);
2674                         clear_bit(R5_ReWrite, &dev->flags);
2675                 }
2676                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2677                     || test_bit(R5_ReadError, &dev->flags)) {
2678                         s.failed++;
2679                         s.failed_num = i;
2680                 } else
2681                         set_bit(R5_Insync, &dev->flags);
2682         }
2683         rcu_read_unlock();
2684
2685         if (s.to_fill && !test_and_set_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2686                 sh->ops.count++;
2687
2688         pr_debug("locked=%d uptodate=%d to_read=%d"
2689                 " to_write=%d failed=%d failed_num=%d\n",
2690                 s.locked, s.uptodate, s.to_read, s.to_write,
2691                 s.failed, s.failed_num);
2692         /* check if the array has lost two devices and, if so, some requests might
2693          * need to be failed
2694          */
2695         if (s.failed > 1 && s.to_read+s.to_write+s.written)
2696                 handle_requests_to_failed_array(conf, sh, &s, disks,
2697                                                 &return_bi);
2698         if (s.failed > 1 && s.syncing) {
2699                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2700                 clear_bit(STRIPE_SYNCING, &sh->state);
2701                 s.syncing = 0;
2702         }
2703
2704         /* might be able to return some write requests if the parity block
2705          * is safe, or on a failed drive
2706          */
2707         dev = &sh->dev[sh->pd_idx];
2708         if ( s.written &&
2709              ((test_bit(R5_Insync, &dev->flags) &&
2710                !test_bit(R5_LOCKED, &dev->flags) &&
2711                test_bit(R5_UPTODATE, &dev->flags)) ||
2712                (s.failed == 1 && s.failed_num == sh->pd_idx)))
2713                 handle_completed_write_requests(conf, sh, disks, &return_bi);
2714
2715         /* Now we might consider reading some blocks, either to check/generate
2716          * parity, or to satisfy requests
2717          * or to load a block that is being partially written.
2718          */
2719         if (s.to_read || s.non_overwrite ||
2720             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding ||
2721             test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2722                 handle_issuing_new_read_requests5(sh, &s, disks);
2723
2724         /* Now we check to see if any write operations have recently
2725          * completed
2726          */
2727
2728         /* leave prexor set until postxor is done, allows us to distinguish
2729          * a rmw from a rcw during biodrain
2730          */
2731         if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2732                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2733
2734                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2735                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2736                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2737
2738                 for (i = disks; i--; )
2739                         clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2740         }
2741
2742         /* if only POSTXOR is set then this is an 'expand' postxor */
2743         if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2744                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2745
2746                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2747                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2748                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2749
2750                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2751                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2752                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2753
2754                 /* All the 'written' buffers and the parity block are ready to
2755                  * be written back to disk
2756                  */
2757                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2758                 for (i = disks; i--; ) {
2759                         dev = &sh->dev[i];
2760                         if (test_bit(R5_LOCKED, &dev->flags) &&
2761                                 (i == sh->pd_idx || dev->written)) {
2762                                 pr_debug("Writing block %d\n", i);
2763                                 set_bit(R5_Wantwrite, &dev->flags);
2764                                 if (!test_and_set_bit(
2765                                     STRIPE_OP_IO, &sh->ops.pending))
2766                                         sh->ops.count++;
2767                                 if (!test_bit(R5_Insync, &dev->flags) ||
2768                                     (i == sh->pd_idx && s.failed == 0))
2769                                         set_bit(STRIPE_INSYNC, &sh->state);
2770                         }
2771                 }
2772                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2773                         atomic_dec(&conf->preread_active_stripes);
2774                         if (atomic_read(&conf->preread_active_stripes) <
2775                                 IO_THRESHOLD)
2776                                 md_wakeup_thread(conf->mddev->thread);
2777                 }
2778         }
2779
2780         /* Now to consider new write requests and what else, if anything
2781          * should be read.  We do not handle new writes when:
2782          * 1/ A 'write' operation (copy+xor) is already in flight.
2783          * 2/ A 'check' operation is in flight, as it may clobber the parity
2784          *    block.
2785          */
2786         if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2787                           !test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2788                 handle_issuing_new_write_requests5(conf, sh, &s, disks);
2789
2790         /* maybe we need to check and possibly fix the parity for this stripe
2791          * Any reads will already have been scheduled, so we just see if enough
2792          * data is available.  The parity check is held off while parity
2793          * dependent operations are in flight.
2794          */
2795         if ((s.syncing && s.locked == 0 &&
2796              !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2797              !test_bit(STRIPE_INSYNC, &sh->state)) ||
2798               test_bit(STRIPE_OP_CHECK, &sh->ops.pending) ||
2799               test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending))
2800                 handle_parity_checks5(conf, sh, &s, disks);
2801
2802         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2803                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2804                 clear_bit(STRIPE_SYNCING, &sh->state);
2805         }
2806
2807         /* If the failed drive is just a ReadError, then we might need to progress
2808          * the repair/check process
2809          */
2810         if (s.failed == 1 && !conf->mddev->ro &&
2811             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2812             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2813             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2814                 ) {
2815                 dev = &sh->dev[s.failed_num];
2816                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2817                         set_bit(R5_Wantwrite, &dev->flags);
2818                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2819                                 sh->ops.count++;
2820                         set_bit(R5_ReWrite, &dev->flags);
2821                         set_bit(R5_LOCKED, &dev->flags);
2822                         s.locked++;
2823                 } else {
2824                         /* let's read it back */
2825                         set_bit(R5_Wantread, &dev->flags);
2826                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2827                                 sh->ops.count++;
2828                         set_bit(R5_LOCKED, &dev->flags);
2829                         s.locked++;
2830                 }
2831         }
2832
2833         /* Finish postxor operations initiated by the expansion
2834          * process
2835          */
2836         if (test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete) &&
2837                 !test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending)) {
2838
2839                 clear_bit(STRIPE_EXPANDING, &sh->state);
2840
2841                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2842                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2843                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2844
2845                 for (i = conf->raid_disks; i--; ) {
2846                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2847                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2848                                 sh->ops.count++;
2849                 }
2850         }
2851
2852         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2853                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2854                 /* Need to write out all blocks after computing parity */
2855                 sh->disks = conf->raid_disks;
2856                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2857                         conf->raid_disks);
2858                 s.locked += handle_write_operations5(sh, 0, 1);
2859         } else if (s.expanded &&
2860                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2861                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2862                 atomic_dec(&conf->reshape_stripes);
2863                 wake_up(&conf->wait_for_overlap);
2864                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2865         }
2866
2867         if (s.expanding && s.locked == 0)
2868                 handle_stripe_expansion(conf, sh, NULL);
2869
2870         if (sh->ops.count)
2871                 pending = get_stripe_work(sh);
2872
2873         spin_unlock(&sh->lock);
2874
2875         if (pending)
2876                 raid5_run_ops(sh, pending);
2877
2878         return_io(return_bi);
2879
2880 }
2881
2882 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2883 {
2884         raid6_conf_t *conf = sh->raid_conf;
2885         int disks = sh->disks;
2886         struct bio *return_bi = NULL;
2887         int i, pd_idx = sh->pd_idx;
2888         struct stripe_head_state s;
2889         struct r6_state r6s;
2890         struct r5dev *dev, *pdev, *qdev;
2891
2892         r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2893         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2894                 "pd_idx=%d, qd_idx=%d\n",
2895                (unsigned long long)sh->sector, sh->state,
2896                atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2897         memset(&s, 0, sizeof(s));
2898
2899         spin_lock(&sh->lock);
2900         clear_bit(STRIPE_HANDLE, &sh->state);
2901         clear_bit(STRIPE_DELAYED, &sh->state);
2902
2903         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2904         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2905         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2906         /* Now to look around and see what can be done */
2907
2908         rcu_read_lock();
2909         for (i=disks; i--; ) {
2910                 mdk_rdev_t *rdev;
2911                 dev = &sh->dev[i];
2912                 clear_bit(R5_Insync, &dev->flags);
2913
2914                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2915                         i, dev->flags, dev->toread, dev->towrite, dev->written);
2916                 /* maybe we can reply to a read */
2917                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2918                         struct bio *rbi, *rbi2;
2919                         pr_debug("Return read for disc %d\n", i);
2920                         spin_lock_irq(&conf->device_lock);
2921                         rbi = dev->toread;
2922                         dev->toread = NULL;
2923                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2924                                 wake_up(&conf->wait_for_overlap);
2925                         spin_unlock_irq(&conf->device_lock);
2926                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2927                                 copy_data(0, rbi, dev->page, dev->sector);
2928                                 rbi2 = r5_next_bio(rbi, dev->sector);
2929                                 spin_lock_irq(&conf->device_lock);
2930                                 if (--rbi->bi_phys_segments == 0) {
2931                                         rbi->bi_next = return_bi;
2932                                         return_bi = rbi;
2933                                 }
2934                                 spin_unlock_irq(&conf->device_lock);
2935                                 rbi = rbi2;
2936                         }
2937                 }
2938
2939                 /* now count some things */
2940                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2941                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2942
2943
2944                 if (dev->toread)
2945                         s.to_read++;
2946                 if (dev->towrite) {
2947                         s.to_write++;
2948                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2949                                 s.non_overwrite++;
2950                 }
2951                 if (dev->written)
2952                         s.written++;
2953                 rdev = rcu_dereference(conf->disks[i].rdev);
2954                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2955                         /* The ReadError flag will just be confusing now */
2956                         clear_bit(R5_ReadError, &dev->flags);
2957                         clear_bit(R5_ReWrite, &dev->flags);
2958                 }
2959                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2960                     || test_bit(R5_ReadError, &dev->flags)) {
2961                         if (s.failed < 2)
2962                                 r6s.failed_num[s.failed] = i;
2963                         s.failed++;
2964                 } else
2965                         set_bit(R5_Insync, &dev->flags);
2966         }
2967         rcu_read_unlock();
2968         pr_debug("locked=%d uptodate=%d to_read=%d"
2969                " to_write=%d failed=%d failed_num=%d,%d\n",
2970                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
2971                r6s.failed_num[0], r6s.failed_num[1]);
2972         /* check if the array has lost >2 devices and, if so, some requests
2973          * might need to be failed
2974          */
2975         if (s.failed > 2 && s.to_read+s.to_write+s.written)
2976                 handle_requests_to_failed_array(conf, sh, &s, disks,
2977                                                 &return_bi);
2978         if (s.failed > 2 && s.syncing) {
2979                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2980                 clear_bit(STRIPE_SYNCING, &sh->state);
2981                 s.syncing = 0;
2982         }
2983
2984         /*
2985          * might be able to return some write requests if the parity blocks
2986          * are safe, or on a failed drive
2987          */
2988         pdev = &sh->dev[pd_idx];
2989         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
2990                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
2991         qdev = &sh->dev[r6s.qd_idx];
2992         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
2993                 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
2994
2995         if ( s.written &&
2996              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
2997                              && !test_bit(R5_LOCKED, &pdev->flags)
2998                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
2999              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3000                              && !test_bit(R5_LOCKED, &qdev->flags)
3001                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3002                 handle_completed_write_requests(conf, sh, disks, &return_bi);
3003
3004         /* Now we might consider reading some blocks, either to check/generate
3005          * parity, or to satisfy requests
3006          * or to load a block that is being partially written.
3007          */
3008         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3009             (s.syncing && (s.uptodate < disks)) || s.expanding)
3010                 handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
3011
3012         /* now to consider writing and what else, if anything should be read */
3013         if (s.to_write)
3014                 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
3015
3016         /* maybe we need to check and possibly fix the parity for this stripe
3017          * Any reads will already have been scheduled, so we just see if enough
3018          * data is available
3019          */
3020         if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3021                 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3022
3023         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3024                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3025                 clear_bit(STRIPE_SYNCING, &sh->state);
3026         }
3027
3028         /* If the failed drives are just a ReadError, then we might need
3029          * to progress the repair/check process
3030          */
3031         if (s.failed <= 2 && !conf->mddev->ro)
3032                 for (i = 0; i < s.failed; i++) {
3033                         dev = &sh->dev[r6s.failed_num[i]];
3034                         if (test_bit(R5_ReadError, &dev->flags)
3035                             && !test_bit(R5_LOCKED, &dev->flags)
3036                             && test_bit(R5_UPTODATE, &dev->flags)
3037                                 ) {
3038                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3039                                         set_bit(R5_Wantwrite, &dev->flags);
3040                                         set_bit(R5_ReWrite, &dev->flags);
3041                                         set_bit(R5_LOCKED, &dev->flags);
3042                                 } else {
3043                                         /* let's read it back */
3044                                         set_bit(R5_Wantread, &dev->flags);
3045                                         set_bit(R5_LOCKED, &dev->flags);
3046                                 }
3047                         }
3048                 }
3049
3050         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3051                 /* Need to write out all blocks after computing P&Q */
3052                 sh->disks = conf->raid_disks;
3053                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
3054                                              conf->raid_disks);
3055                 compute_parity6(sh, RECONSTRUCT_WRITE);
3056                 for (i = conf->raid_disks ; i-- ;  ) {
3057                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3058                         s.locked++;
3059                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3060                 }
3061                 clear_bit(STRIPE_EXPANDING, &sh->state);
3062         } else if (s.expanded) {
3063                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3064                 atomic_dec(&conf->reshape_stripes);
3065                 wake_up(&conf->wait_for_overlap);
3066                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3067         }
3068
3069         if (s.expanding && s.locked == 0)
3070                 handle_stripe_expansion(conf, sh, &r6s);
3071
3072         spin_unlock(&sh->lock);
3073
3074         return_io(return_bi);
3075
3076         for (i=disks; i-- ;) {
3077                 int rw;
3078                 struct bio *bi;
3079                 mdk_rdev_t *rdev;
3080                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
3081                         rw = WRITE;
3082                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
3083                         rw = READ;
3084                 else
3085                         continue;
3086
3087                 bi = &sh->dev[i].req;
3088
3089                 bi->bi_rw = rw;
3090                 if (rw == WRITE)
3091                         bi->bi_end_io = raid5_end_write_request;
3092                 else
3093                         bi->bi_end_io = raid5_end_read_request;
3094
3095                 rcu_read_lock();
3096                 rdev = rcu_dereference(conf->disks[i].rdev);
3097                 if (rdev && test_bit(Faulty, &rdev->flags))
3098                         rdev = NULL;
3099                 if (rdev)
3100                         atomic_inc(&rdev->nr_pending);
3101                 rcu_read_unlock();
3102
3103                 if (rdev) {
3104                         if (s.syncing || s.expanding || s.expanded)
3105                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
3106
3107                         bi->bi_bdev = rdev->bdev;
3108                         pr_debug("for %llu schedule op %ld on disc %d\n",
3109                                 (unsigned long long)sh->sector, bi->bi_rw, i);
3110                         atomic_inc(&sh->count);
3111                         bi->bi_sector = sh->sector + rdev->data_offset;
3112                         bi->bi_flags = 1 << BIO_UPTODATE;
3113                         bi->bi_vcnt = 1;
3114                         bi->bi_max_vecs = 1;
3115                         bi->bi_idx = 0;
3116                         bi->bi_io_vec = &sh->dev[i].vec;
3117                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
3118                         bi->bi_io_vec[0].bv_offset = 0;
3119                         bi->bi_size = STRIPE_SIZE;
3120                         bi->bi_next = NULL;
3121                         if (rw == WRITE &&
3122                             test_bit(R5_ReWrite, &sh->dev[i].flags))
3123                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
3124                         generic_make_request(bi);
3125                 } else {
3126                         if (rw == WRITE)
3127                                 set_bit(STRIPE_DEGRADED, &sh->state);
3128                         pr_debug("skip op %ld on disc %d for sector %llu\n",
3129                                 bi->bi_rw, i, (unsigned long long)sh->sector);
3130                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
3131                         set_bit(STRIPE_HANDLE, &sh->state);
3132                 }
3133         }
3134 }
3135
3136 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3137 {
3138         if (sh->raid_conf->level == 6)
3139                 handle_stripe6(sh, tmp_page);
3140         else
3141                 handle_stripe5(sh);
3142 }
3143
3144
3145
3146 static void raid5_activate_delayed(raid5_conf_t *conf)
3147 {
3148         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3149                 while (!list_empty(&conf->delayed_list)) {
3150                         struct list_head *l = conf->delayed_list.next;
3151                         struct stripe_head *sh;
3152                         sh = list_entry(l, struct stripe_head, lru);
3153                         list_del_init(l);
3154                         clear_bit(STRIPE_DELAYED, &sh->state);
3155                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3156                                 atomic_inc(&conf->preread_active_stripes);
3157                         list_add_tail(&sh->lru, &conf->handle_list);
3158                 }
3159         }
3160 }
3161
3162 static void activate_bit_delay(raid5_conf_t *conf)
3163 {
3164         /* device_lock is held */
3165         struct list_head head;
3166         list_add(&head, &conf->bitmap_list);
3167         list_del_init(&conf->bitmap_list);
3168         while (!list_empty(&head)) {
3169                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3170                 list_del_init(&sh->lru);
3171                 atomic_inc(&sh->count);
3172                 __release_stripe(conf, sh);
3173         }
3174 }
3175
3176 static void unplug_slaves(mddev_t *mddev)
3177 {
3178         raid5_conf_t *conf = mddev_to_conf(mddev);
3179         int i;
3180
3181         rcu_read_lock();
3182         for (i=0; i<mddev->raid_disks; i++) {
3183                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3184                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3185                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
3186
3187                         atomic_inc(&rdev->nr_pending);
3188                         rcu_read_unlock();
3189
3190                         if (r_queue->unplug_fn)
3191                                 r_queue->unplug_fn(r_queue);
3192
3193                         rdev_dec_pending(rdev, mddev);
3194                         rcu_read_lock();
3195                 }
3196         }
3197         rcu_read_unlock();
3198 }
3199
3200 static void raid5_unplug_device(request_queue_t *q)
3201 {
3202         mddev_t *mddev = q->queuedata;
3203         raid5_conf_t *conf = mddev_to_conf(mddev);
3204         unsigned long flags;
3205
3206         spin_lock_irqsave(&conf->device_lock, flags);
3207
3208         if (blk_remove_plug(q)) {
3209                 conf->seq_flush++;
3210                 raid5_activate_delayed(conf);
3211         }
3212         md_wakeup_thread(mddev->thread);
3213
3214         spin_unlock_irqrestore(&conf->device_lock, flags);
3215
3216         unplug_slaves(mddev);
3217 }
3218
3219 static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
3220                              sector_t *error_sector)
3221 {
3222         mddev_t *mddev = q->queuedata;
3223         raid5_conf_t *conf = mddev_to_conf(mddev);
3224         int i, ret = 0;
3225
3226         rcu_read_lock();
3227         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
3228                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3229                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
3230                         struct block_device *bdev = rdev->bdev;
3231                         request_queue_t *r_queue = bdev_get_queue(bdev);
3232
3233                         if (!r_queue->issue_flush_fn)
3234                                 ret = -EOPNOTSUPP;
3235                         else {
3236                                 atomic_inc(&rdev->nr_pending);
3237                                 rcu_read_unlock();
3238                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
3239                                                               error_sector);
3240                                 rdev_dec_pending(rdev, mddev);
3241                                 rcu_read_lock();
3242                         }
3243                 }
3244         }
3245         rcu_read_unlock();
3246         return ret;
3247 }
3248
3249 static int raid5_congested(void *data, int bits)
3250 {
3251         mddev_t *mddev = data;
3252         raid5_conf_t *conf = mddev_to_conf(mddev);
3253
3254         /* No difference between reads and writes.  Just check
3255          * how busy the stripe_cache is
3256          */
3257         if (conf->inactive_blocked)
3258                 return 1;
3259         if (conf->quiesce)
3260                 return 1;
3261         if (list_empty_careful(&conf->inactive_list))
3262                 return 1;
3263
3264         return 0;
3265 }
3266
3267 /* We want read requests to align with chunks where possible,
3268  * but write requests don't need to.
3269  */
3270 static int raid5_mergeable_bvec(request_queue_t *q, struct bio *bio, struct bio_vec *biovec)
3271 {
3272         mddev_t *mddev = q->queuedata;
3273         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3274         int max;
3275         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3276         unsigned int bio_sectors = bio->bi_size >> 9;
3277
3278         if (bio_data_dir(bio) == WRITE)
3279                 return biovec->bv_len; /* always allow writes to be mergeable */
3280
3281         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3282         if (max < 0) max = 0;
3283         if (max <= biovec->bv_len && bio_sectors == 0)
3284                 return biovec->bv_len;
3285         else
3286                 return max;
3287 }
3288
3289
3290 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3291 {
3292         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3293         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3294         unsigned int bio_sectors = bio->bi_size >> 9;
3295
3296         return  chunk_sectors >=
3297                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3298 }
3299
3300 /*
3301  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3302  *  later sampled by raid5d.
3303  */
3304 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3305 {
3306         unsigned long flags;
3307
3308         spin_lock_irqsave(&conf->device_lock, flags);
3309
3310         bi->bi_next = conf->retry_read_aligned_list;
3311         conf->retry_read_aligned_list = bi;
3312
3313         spin_unlock_irqrestore(&conf->device_lock, flags);
3314         md_wakeup_thread(conf->mddev->thread);
3315 }
3316
3317
3318 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3319 {
3320         struct bio *bi;
3321
3322         bi = conf->retry_read_aligned;
3323         if (bi) {
3324                 conf->retry_read_aligned = NULL;
3325                 return bi;
3326         }
3327         bi = conf->retry_read_aligned_list;
3328         if(bi) {
3329                 conf->retry_read_aligned_list = bi->bi_next;
3330                 bi->bi_next = NULL;
3331                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3332                 bi->bi_hw_segments = 0; /* count of processed stripes */
3333         }
3334
3335         return bi;
3336 }
3337
3338
3339 /*
3340  *  The "raid5_align_endio" should check if the read succeeded and if it
3341  *  did, call bio_endio on the original bio (having bio_put the new bio
3342  *  first).
3343  *  If the read failed..
3344  */
3345 static int raid5_align_endio(struct bio *bi, unsigned int bytes, int error)
3346 {
3347         struct bio* raid_bi  = bi->bi_private;
3348         mddev_t *mddev;
3349         raid5_conf_t *conf;
3350         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3351         mdk_rdev_t *rdev;
3352
3353         if (bi->bi_size)
3354                 return 1;
3355         bio_put(bi);
3356
3357         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3358         conf = mddev_to_conf(mddev);
3359         rdev = (void*)raid_bi->bi_next;
3360         raid_bi->bi_next = NULL;
3361
3362         rdev_dec_pending(rdev, conf->mddev);
3363
3364         if (!error && uptodate) {
3365                 bio_endio(raid_bi, bytes, 0);
3366                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3367                         wake_up(&conf->wait_for_stripe);
3368                 return 0;
3369         }
3370
3371
3372         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3373
3374         add_bio_to_retry(raid_bi, conf);
3375         return 0;
3376 }
3377
3378 static int bio_fits_rdev(struct bio *bi)
3379 {
3380         request_queue_t *q = bdev_get_queue(bi->bi_bdev);
3381
3382         if ((bi->bi_size>>9) > q->max_sectors)
3383                 return 0;
3384         blk_recount_segments(q, bi);
3385         if (bi->bi_phys_segments > q->max_phys_segments ||
3386             bi->bi_hw_segments > q->max_hw_segments)
3387                 return 0;
3388
3389         if (q->merge_bvec_fn)
3390                 /* it's too hard to apply the merge_bvec_fn at this stage,
3391                  * just just give up
3392                  */
3393                 return 0;
3394
3395         return 1;
3396 }
3397
3398
3399 static int chunk_aligned_read(request_queue_t *q, struct bio * raid_bio)
3400 {
3401         mddev_t *mddev = q->queuedata;
3402         raid5_conf_t *conf = mddev_to_conf(mddev);
3403         const unsigned int raid_disks = conf->raid_disks;
3404         const unsigned int data_disks = raid_disks - conf->max_degraded;
3405         unsigned int dd_idx, pd_idx;
3406         struct bio* align_bi;
3407         mdk_rdev_t *rdev;
3408
3409         if (!in_chunk_boundary(mddev, raid_bio)) {
3410                 pr_debug("chunk_aligned_read : non aligned\n");
3411                 return 0;
3412         }
3413         /*
3414          * use bio_clone to make a copy of the bio
3415          */
3416         align_bi = bio_clone(raid_bio, GFP_NOIO);
3417         if (!align_bi)
3418                 return 0;
3419         /*
3420          *   set bi_end_io to a new function, and set bi_private to the
3421          *     original bio.
3422          */
3423         align_bi->bi_end_io  = raid5_align_endio;
3424         align_bi->bi_private = raid_bio;
3425         /*
3426          *      compute position
3427          */
3428         align_bi->bi_sector =  raid5_compute_sector(raid_bio->bi_sector,
3429                                         raid_disks,
3430                                         data_disks,
3431                                         &dd_idx,
3432                                         &pd_idx,
3433                                         conf);
3434
3435         rcu_read_lock();
3436         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3437         if (rdev && test_bit(In_sync, &rdev->flags)) {
3438                 atomic_inc(&rdev->nr_pending);
3439                 rcu_read_unlock();
3440                 raid_bio->bi_next = (void*)rdev;
3441                 align_bi->bi_bdev =  rdev->bdev;
3442                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3443                 align_bi->bi_sector += rdev->data_offset;
3444
3445                 if (!bio_fits_rdev(align_bi)) {
3446                         /* too big in some way */
3447                         bio_put(align_bi);
3448                         rdev_dec_pending(rdev, mddev);
3449                         return 0;
3450                 }
3451
3452                 spin_lock_irq(&conf->device_lock);
3453                 wait_event_lock_irq(conf->wait_for_stripe,
3454                                     conf->quiesce == 0,
3455                                     conf->device_lock, /* nothing */);
3456                 atomic_inc(&conf->active_aligned_reads);
3457                 spin_unlock_irq(&conf->device_lock);
3458
3459                 generic_make_request(align_bi);
3460                 return 1;
3461         } else {
3462                 rcu_read_unlock();
3463                 bio_put(align_bi);
3464                 return 0;
3465         }
3466 }
3467
3468
3469 static int make_request(request_queue_t *q, struct bio * bi)
3470 {
3471         mddev_t *mddev = q->queuedata;
3472         raid5_conf_t *conf = mddev_to_conf(mddev);
3473         unsigned int dd_idx, pd_idx;
3474         sector_t new_sector;
3475         sector_t logical_sector, last_sector;
3476         struct stripe_head *sh;
3477         const int rw = bio_data_dir(bi);
3478         int remaining;
3479
3480         if (unlikely(bio_barrier(bi))) {
3481                 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
3482                 return 0;
3483         }
3484
3485         md_write_start(mddev, bi);
3486
3487         disk_stat_inc(mddev->gendisk, ios[rw]);
3488         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3489
3490         if (rw == READ &&
3491              mddev->reshape_position == MaxSector &&
3492              chunk_aligned_read(q,bi))
3493                 return 0;
3494
3495         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3496         last_sector = bi->bi_sector + (bi->bi_size>>9);
3497         bi->bi_next = NULL;
3498         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3499
3500         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3501                 DEFINE_WAIT(w);
3502                 int disks, data_disks;
3503
3504         retry:
3505                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3506                 if (likely(conf->expand_progress == MaxSector))
3507                         disks = conf->raid_disks;
3508                 else {
3509                         /* spinlock is needed as expand_progress may be
3510                          * 64bit on a 32bit platform, and so it might be
3511                          * possible to see a half-updated value
3512                          * Ofcourse expand_progress could change after
3513                          * the lock is dropped, so once we get a reference
3514                          * to the stripe that we think it is, we will have
3515                          * to check again.
3516                          */
3517                         spin_lock_irq(&conf->device_lock);
3518                         disks = conf->raid_disks;
3519                         if (logical_sector >= conf->expand_progress)
3520                                 disks = conf->previous_raid_disks;
3521                         else {
3522                                 if (logical_sector >= conf->expand_lo) {
3523                                         spin_unlock_irq(&conf->device_lock);
3524                                         schedule();
3525                                         goto retry;
3526                                 }
3527                         }
3528                         spin_unlock_irq(&conf->device_lock);
3529                 }
3530                 data_disks = disks - conf->max_degraded;
3531
3532                 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3533                                                   &dd_idx, &pd_idx, conf);
3534                 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3535                         (unsigned long long)new_sector, 
3536                         (unsigned long long)logical_sector);
3537
3538                 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3539                 if (sh) {
3540                         if (unlikely(conf->expand_progress != MaxSector)) {
3541                                 /* expansion might have moved on while waiting for a
3542                                  * stripe, so we must do the range check again.
3543                                  * Expansion could still move past after this
3544                                  * test, but as we are holding a reference to
3545                                  * 'sh', we know that if that happens,
3546                                  *  STRIPE_EXPANDING will get set and the expansion
3547                                  * won't proceed until we finish with the stripe.
3548                                  */
3549                                 int must_retry = 0;
3550                                 spin_lock_irq(&conf->device_lock);
3551                                 if (logical_sector <  conf->expand_progress &&
3552                                     disks == conf->previous_raid_disks)
3553                                         /* mismatch, need to try again */
3554                                         must_retry = 1;
3555                                 spin_unlock_irq(&conf->device_lock);
3556                                 if (must_retry) {
3557                                         release_stripe(sh);
3558                                         goto retry;
3559                                 }
3560                         }
3561                         /* FIXME what if we get a false positive because these
3562                          * are being updated.
3563                          */
3564                         if (logical_sector >= mddev->suspend_lo &&
3565                             logical_sector < mddev->suspend_hi) {
3566                                 release_stripe(sh);
3567                                 schedule();
3568                                 goto retry;
3569                         }
3570
3571                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3572                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3573                                 /* Stripe is busy expanding or
3574                                  * add failed due to overlap.  Flush everything
3575                                  * and wait a while
3576                                  */
3577                                 raid5_unplug_device(mddev->queue);
3578                                 release_stripe(sh);
3579                                 schedule();
3580                                 goto retry;
3581                         }
3582                         finish_wait(&conf->wait_for_overlap, &w);
3583                         handle_stripe(sh, NULL);
3584                         release_stripe(sh);
3585                 } else {
3586                         /* cannot get stripe for read-ahead, just give-up */
3587                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3588                         finish_wait(&conf->wait_for_overlap, &w);
3589                         break;
3590                 }
3591                         
3592         }
3593         spin_lock_irq(&conf->device_lock);
3594         remaining = --bi->bi_phys_segments;
3595         spin_unlock_irq(&conf->device_lock);
3596         if (remaining == 0) {
3597                 int bytes = bi->bi_size;
3598
3599                 if ( rw == WRITE )
3600                         md_write_end(mddev);
3601                 bi->bi_size = 0;
3602                 bi->bi_end_io(bi, bytes,
3603                               test_bit(BIO_UPTODATE, &bi->bi_flags)
3604                                 ? 0 : -EIO);
3605         }
3606         return 0;
3607 }
3608
3609 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3610 {
3611         /* reshaping is quite different to recovery/resync so it is
3612          * handled quite separately ... here.
3613          *
3614          * On each call to sync_request, we gather one chunk worth of
3615          * destination stripes and flag them as expanding.
3616          * Then we find all the source stripes and request reads.
3617          * As the reads complete, handle_stripe will copy the data
3618          * into the destination stripe and release that stripe.
3619          */
3620         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3621         struct stripe_head *sh;
3622         int pd_idx;
3623         sector_t first_sector, last_sector;
3624         int raid_disks = conf->previous_raid_disks;
3625         int data_disks = raid_disks - conf->max_degraded;
3626         int new_data_disks = conf->raid_disks - conf->max_degraded;
3627         int i;
3628         int dd_idx;
3629         sector_t writepos, safepos, gap;
3630
3631         if (sector_nr == 0 &&
3632             conf->expand_progress != 0) {
3633                 /* restarting in the middle, skip the initial sectors */
3634                 sector_nr = conf->expand_progress;
3635                 sector_div(sector_nr, new_data_disks);
3636                 *skipped = 1;
3637                 return sector_nr;
3638         }
3639
3640         /* we update the metadata when there is more than 3Meg
3641          * in the block range (that is rather arbitrary, should
3642          * probably be time based) or when the data about to be
3643          * copied would over-write the source of the data at
3644          * the front of the range.
3645          * i.e. one new_stripe forward from expand_progress new_maps
3646          * to after where expand_lo old_maps to
3647          */
3648         writepos = conf->expand_progress +
3649                 conf->chunk_size/512*(new_data_disks);
3650         sector_div(writepos, new_data_disks);
3651         safepos = conf->expand_lo;
3652         sector_div(safepos, data_disks);
3653         gap = conf->expand_progress - conf->expand_lo;
3654
3655         if (writepos >= safepos ||
3656             gap > (new_data_disks)*3000*2 /*3Meg*/) {
3657                 /* Cannot proceed until we've updated the superblock... */
3658                 wait_event(conf->wait_for_overlap,
3659                            atomic_read(&conf->reshape_stripes)==0);
3660                 mddev->reshape_position = conf->expand_progress;
3661                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3662                 md_wakeup_thread(mddev->thread);
3663                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3664                            kthread_should_stop());
3665                 spin_lock_irq(&conf->device_lock);
3666                 conf->expand_lo = mddev->reshape_position;
3667                 spin_unlock_irq(&conf->device_lock);
3668                 wake_up(&conf->wait_for_overlap);
3669         }
3670
3671         for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3672                 int j;
3673                 int skipped = 0;
3674                 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3675                 sh = get_active_stripe(conf, sector_nr+i,
3676                                        conf->raid_disks, pd_idx, 0);
3677                 set_bit(STRIPE_EXPANDING, &sh->state);
3678                 atomic_inc(&conf->reshape_stripes);
3679                 /* If any of this stripe is beyond the end of the old
3680                  * array, then we need to zero those blocks
3681                  */
3682                 for (j=sh->disks; j--;) {
3683                         sector_t s;
3684                         if (j == sh->pd_idx)
3685                                 continue;
3686                         if (conf->level == 6 &&
3687                             j == raid6_next_disk(sh->pd_idx, sh->disks))
3688                                 continue;
3689                         s = compute_blocknr(sh, j);
3690                         if (s < (mddev->array_size<<1)) {
3691                                 skipped = 1;
3692                                 continue;
3693                         }
3694                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3695                         set_bit(R5_Expanded, &sh->dev[j].flags);
3696                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
3697                 }
3698                 if (!skipped) {
3699                         set_bit(STRIPE_EXPAND_READY, &sh->state);
3700                         set_bit(STRIPE_HANDLE, &sh->state);
3701                 }
3702                 release_stripe(sh);
3703         }
3704         spin_lock_irq(&conf->device_lock);
3705         conf->expand_progress = (sector_nr + i) * new_data_disks;
3706         spin_unlock_irq(&conf->device_lock);
3707         /* Ok, those stripe are ready. We can start scheduling
3708          * reads on the source stripes.
3709          * The source stripes are determined by mapping the first and last
3710          * block on the destination stripes.
3711          */
3712         first_sector =
3713                 raid5_compute_sector(sector_nr*(new_data_disks),
3714                                      raid_disks, data_disks,
3715                                      &dd_idx, &pd_idx, conf);
3716         last_sector =
3717                 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3718                                      *(new_data_disks) -1,
3719                                      raid_disks, data_disks,
3720                                      &dd_idx, &pd_idx, conf);
3721         if (last_sector >= (mddev->size<<1))
3722                 last_sector = (mddev->size<<1)-1;
3723         while (first_sector <= last_sector) {
3724                 pd_idx = stripe_to_pdidx(first_sector, conf,
3725                                          conf->previous_raid_disks);
3726                 sh = get_active_stripe(conf, first_sector,
3727                                        conf->previous_raid_disks, pd_idx, 0);
3728                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3729                 set_bit(STRIPE_HANDLE, &sh->state);
3730                 release_stripe(sh);
3731                 first_sector += STRIPE_SECTORS;
3732         }
3733         return conf->chunk_size>>9;
3734 }
3735
3736 /* FIXME go_faster isn't used */
3737 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3738 {
3739         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3740         struct stripe_head *sh;
3741         int pd_idx;
3742         int raid_disks = conf->raid_disks;
3743         sector_t max_sector = mddev->size << 1;
3744         int sync_blocks;
3745         int still_degraded = 0;
3746         int i;
3747
3748         if (sector_nr >= max_sector) {
3749                 /* just being told to finish up .. nothing much to do */
3750                 unplug_slaves(mddev);
3751                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3752                         end_reshape(conf);
3753                         return 0;
3754                 }
3755
3756                 if (mddev->curr_resync < max_sector) /* aborted */
3757                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3758                                         &sync_blocks, 1);
3759                 else /* completed sync */
3760                         conf->fullsync = 0;
3761                 bitmap_close_sync(mddev->bitmap);
3762
3763                 return 0;
3764         }
3765
3766         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3767                 return reshape_request(mddev, sector_nr, skipped);
3768
3769         /* if there is too many failed drives and we are trying
3770          * to resync, then assert that we are finished, because there is
3771          * nothing we can do.
3772          */
3773         if (mddev->degraded >= conf->max_degraded &&
3774             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3775                 sector_t rv = (mddev->size << 1) - sector_nr;
3776                 *skipped = 1;
3777                 return rv;
3778         }
3779         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3780             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3781             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3782                 /* we can skip this block, and probably more */
3783                 sync_blocks /= STRIPE_SECTORS;
3784                 *skipped = 1;
3785                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3786         }
3787
3788         pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3789         sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3790         if (sh == NULL) {
3791                 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3792                 /* make sure we don't swamp the stripe cache if someone else
3793                  * is trying to get access
3794                  */
3795                 schedule_timeout_uninterruptible(1);
3796         }
3797         /* Need to check if array will still be degraded after recovery/resync
3798          * We don't need to check the 'failed' flag as when that gets set,
3799          * recovery aborts.
3800          */
3801         for (i=0; i<mddev->raid_disks; i++)
3802                 if (conf->disks[i].rdev == NULL)
3803                         still_degraded = 1;
3804
3805         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3806
3807         spin_lock(&sh->lock);
3808         set_bit(STRIPE_SYNCING, &sh->state);
3809         clear_bit(STRIPE_INSYNC, &sh->state);
3810         spin_unlock(&sh->lock);
3811
3812         handle_stripe(sh, NULL);
3813         release_stripe(sh);
3814
3815         return STRIPE_SECTORS;
3816 }
3817
3818 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3819 {
3820         /* We may not be able to submit a whole bio at once as there
3821          * may not be enough stripe_heads available.
3822          * We cannot pre-allocate enough stripe_heads as we may need
3823          * more than exist in the cache (if we allow ever large chunks).
3824          * So we do one stripe head at a time and record in
3825          * ->bi_hw_segments how many have been done.
3826          *
3827          * We *know* that this entire raid_bio is in one chunk, so
3828          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3829          */
3830         struct stripe_head *sh;
3831         int dd_idx, pd_idx;
3832         sector_t sector, logical_sector, last_sector;
3833         int scnt = 0;
3834         int remaining;
3835         int handled = 0;
3836
3837         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3838         sector = raid5_compute_sector(  logical_sector,
3839                                         conf->raid_disks,
3840                                         conf->raid_disks - conf->max_degraded,
3841                                         &dd_idx,
3842                                         &pd_idx,
3843                                         conf);
3844         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3845
3846         for (; logical_sector < last_sector;
3847              logical_sector += STRIPE_SECTORS,
3848                      sector += STRIPE_SECTORS,
3849                      scnt++) {
3850
3851                 if (scnt < raid_bio->bi_hw_segments)
3852                         /* already done this stripe */
3853                         continue;
3854
3855                 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3856
3857                 if (!sh) {
3858                         /* failed to get a stripe - must wait */
3859                         raid_bio->bi_hw_segments = scnt;
3860                         conf->retry_read_aligned = raid_bio;
3861                         return handled;
3862                 }
3863
3864                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3865                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3866                         release_stripe(sh);
3867                         raid_bio->bi_hw_segments = scnt;
3868                         conf->retry_read_aligned = raid_bio;
3869                         return handled;
3870                 }
3871
3872                 handle_stripe(sh, NULL);
3873                 release_stripe(sh);
3874                 handled++;
3875         }
3876         spin_lock_irq(&conf->device_lock);
3877         remaining = --raid_bio->bi_phys_segments;
3878         spin_unlock_irq(&conf->device_lock);
3879         if (remaining == 0) {
3880                 int bytes = raid_bio->bi_size;
3881
3882                 raid_bio->bi_size = 0;
3883                 raid_bio->bi_end_io(raid_bio, bytes,
3884                               test_bit(BIO_UPTODATE, &raid_bio->bi_flags)
3885                                 ? 0 : -EIO);
3886         }
3887         if (atomic_dec_and_test(&conf->active_aligned_reads))
3888                 wake_up(&conf->wait_for_stripe);
3889         return handled;
3890 }
3891
3892
3893
3894 /*
3895  * This is our raid5 kernel thread.
3896  *
3897  * We scan the hash table for stripes which can be handled now.
3898  * During the scan, completed stripes are saved for us by the interrupt
3899  * handler, so that they will not have to wait for our next wakeup.
3900  */
3901 static void raid5d (mddev_t *mddev)
3902 {
3903         struct stripe_head *sh;
3904         raid5_conf_t *conf = mddev_to_conf(mddev);
3905         int handled;
3906
3907         pr_debug("+++ raid5d active\n");
3908
3909         md_check_recovery(mddev);
3910
3911         handled = 0;
3912         spin_lock_irq(&conf->device_lock);
3913         while (1) {
3914                 struct list_head *first;
3915                 struct bio *bio;
3916
3917                 if (conf->seq_flush != conf->seq_write) {
3918                         int seq = conf->seq_flush;
3919                         spin_unlock_irq(&conf->device_lock);
3920                         bitmap_unplug(mddev->bitmap);
3921                         spin_lock_irq(&conf->device_lock);
3922                         conf->seq_write = seq;
3923                         activate_bit_delay(conf);
3924                 }
3925
3926                 if (list_empty(&conf->handle_list) &&
3927                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
3928                     !blk_queue_plugged(mddev->queue) &&
3929                     !list_empty(&conf->delayed_list))
3930                         raid5_activate_delayed(conf);
3931
3932                 while ((bio = remove_bio_from_retry(conf))) {
3933                         int ok;
3934                         spin_unlock_irq(&conf->device_lock);
3935                         ok = retry_aligned_read(conf, bio);
3936                         spin_lock_irq(&conf->device_lock);
3937                         if (!ok)
3938                                 break;
3939                         handled++;
3940                 }
3941
3942                 if (list_empty(&conf->handle_list)) {
3943                         async_tx_issue_pending_all();
3944                         break;
3945                 }
3946
3947                 first = conf->handle_list.next;
3948                 sh = list_entry(first, struct stripe_head, lru);
3949
3950                 list_del_init(first);
3951                 atomic_inc(&sh->count);
3952                 BUG_ON(atomic_read(&sh->count)!= 1);
3953                 spin_unlock_irq(&conf->device_lock);
3954                 
3955                 handled++;
3956                 handle_stripe(sh, conf->spare_page);
3957                 release_stripe(sh);
3958
3959                 spin_lock_irq(&conf->device_lock);
3960         }
3961         pr_debug("%d stripes handled\n", handled);
3962
3963         spin_unlock_irq(&conf->device_lock);
3964
3965         unplug_slaves(mddev);
3966
3967         pr_debug("--- raid5d inactive\n");
3968 }
3969
3970 static ssize_t
3971 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
3972 {
3973         raid5_conf_t *conf = mddev_to_conf(mddev);
3974         if (conf)
3975                 return sprintf(page, "%d\n", conf->max_nr_stripes);
3976         else
3977                 return 0;
3978 }
3979
3980 static ssize_t
3981 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
3982 {
3983         raid5_conf_t *conf = mddev_to_conf(mddev);
3984         char *end;
3985         int new;
3986         if (len >= PAGE_SIZE)
3987                 return -EINVAL;
3988         if (!conf)
3989                 return -ENODEV;
3990
3991         new = simple_strtoul(page, &end, 10);
3992         if (!*page || (*end && *end != '\n') )
3993                 return -EINVAL;
3994         if (new <= 16 || new > 32768)
3995                 return -EINVAL;
3996         while (new < conf->max_nr_stripes) {
3997                 if (drop_one_stripe(conf))
3998                         conf->max_nr_stripes--;
3999                 else
4000                         break;
4001         }
4002         md_allow_write(mddev);
4003         while (new > conf->max_nr_stripes) {
4004                 if (grow_one_stripe(conf))
4005                         conf->max_nr_stripes++;
4006                 else break;
4007         }
4008         return len;
4009 }
4010
4011 static struct md_sysfs_entry
4012 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4013                                 raid5_show_stripe_cache_size,
4014                                 raid5_store_stripe_cache_size);
4015
4016 static ssize_t
4017 stripe_cache_active_show(mddev_t *mddev, char *page)
4018 {
4019         raid5_conf_t *conf = mddev_to_conf(mddev);
4020         if (conf)
4021                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4022         else
4023                 return 0;
4024 }
4025
4026 static struct md_sysfs_entry
4027 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4028
4029 static struct attribute *raid5_attrs[] =  {
4030         &raid5_stripecache_size.attr,
4031         &raid5_stripecache_active.attr,
4032         NULL,
4033 };
4034 static struct attribute_group raid5_attrs_group = {
4035         .name = NULL,
4036         .attrs = raid5_attrs,
4037 };
4038
4039 static int run(mddev_t *mddev)
4040 {
4041         raid5_conf_t *conf;
4042         int raid_disk, memory;
4043         mdk_rdev_t *rdev;
4044         struct disk_info *disk;
4045         struct list_head *tmp;
4046         int working_disks = 0;
4047
4048         if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4049                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4050                        mdname(mddev), mddev->level);
4051                 return -EIO;
4052         }
4053
4054         if (mddev->reshape_position != MaxSector) {
4055                 /* Check that we can continue the reshape.
4056                  * Currently only disks can change, it must
4057                  * increase, and we must be past the point where
4058                  * a stripe over-writes itself
4059                  */
4060                 sector_t here_new, here_old;
4061                 int old_disks;
4062                 int max_degraded = (mddev->level == 5 ? 1 : 2);
4063
4064                 if (mddev->new_level != mddev->level ||
4065                     mddev->new_layout != mddev->layout ||
4066                     mddev->new_chunk != mddev->chunk_size) {
4067                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4068                                "required - aborting.\n",
4069                                mdname(mddev));
4070                         return -EINVAL;
4071                 }
4072                 if (mddev->delta_disks <= 0) {
4073                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4074                                "(reduce disks) required - aborting.\n",
4075                                mdname(mddev));
4076                         return -EINVAL;
4077                 }
4078                 old_disks = mddev->raid_disks - mddev->delta_disks;
4079                 /* reshape_position must be on a new-stripe boundary, and one
4080                  * further up in new geometry must map after here in old
4081                  * geometry.
4082                  */
4083                 here_new = mddev->reshape_position;
4084                 if (sector_div(here_new, (mddev->chunk_size>>9)*
4085                                (mddev->raid_disks - max_degraded))) {
4086                         printk(KERN_ERR "raid5: reshape_position not "
4087                                "on a stripe boundary\n");
4088                         return -EINVAL;
4089                 }
4090                 /* here_new is the stripe we will write to */
4091                 here_old = mddev->reshape_position;
4092                 sector_div(here_old, (mddev->chunk_size>>9)*
4093                            (old_disks-max_degraded));
4094                 /* here_old is the first stripe that we might need to read
4095                  * from */
4096                 if (here_new >= here_old) {
4097                         /* Reading from the same stripe as writing to - bad */
4098                         printk(KERN_ERR "raid5: reshape_position too early for "
4099                                "auto-recovery - aborting.\n");
4100                         return -EINVAL;
4101                 }
4102                 printk(KERN_INFO "raid5: reshape will continue\n");
4103                 /* OK, we should be able to continue; */
4104         }
4105
4106
4107         mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4108         if ((conf = mddev->private) == NULL)
4109                 goto abort;
4110         if (mddev->reshape_position == MaxSector) {
4111                 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4112         } else {
4113                 conf->raid_disks = mddev->raid_disks;
4114                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4115         }
4116
4117         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4118                               GFP_KERNEL);
4119         if (!conf->disks)
4120                 goto abort;
4121
4122         conf->mddev = mddev;
4123
4124         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4125                 goto abort;
4126
4127         if (mddev->level == 6) {
4128                 conf->spare_page = alloc_page(GFP_KERNEL);
4129                 if (!conf->spare_page)
4130                         goto abort;
4131         }
4132         spin_lock_init(&conf->device_lock);
4133         init_waitqueue_head(&conf->wait_for_stripe);
4134         init_waitqueue_head(&conf->wait_for_overlap);
4135         INIT_LIST_HEAD(&conf->handle_list);
4136         INIT_LIST_HEAD(&conf->delayed_list);
4137         INIT_LIST_HEAD(&conf->bitmap_list);
4138         INIT_LIST_HEAD(&conf->inactive_list);
4139         atomic_set(&conf->active_stripes, 0);
4140         atomic_set(&conf->preread_active_stripes, 0);
4141         atomic_set(&conf->active_aligned_reads, 0);
4142
4143         pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4144
4145         ITERATE_RDEV(mddev,rdev,tmp) {
4146                 raid_disk = rdev->raid_disk;
4147                 if (raid_disk >= conf->raid_disks
4148                     || raid_disk < 0)
4149                         continue;
4150                 disk = conf->disks + raid_disk;
4151
4152                 disk->rdev = rdev;
4153
4154                 if (test_bit(In_sync, &rdev->flags)) {
4155                         char b[BDEVNAME_SIZE];
4156                         printk(KERN_INFO "raid5: device %s operational as raid"
4157                                 " disk %d\n", bdevname(rdev->bdev,b),
4158                                 raid_disk);
4159                         working_disks++;
4160                 }
4161         }
4162
4163         /*
4164          * 0 for a fully functional array, 1 or 2 for a degraded array.
4165          */
4166         mddev->degraded = conf->raid_disks - working_disks;
4167         conf->mddev = mddev;
4168         conf->chunk_size = mddev->chunk_size;
4169         conf->level = mddev->level;
4170         if (conf->level == 6)
4171                 conf->max_degraded = 2;
4172         else
4173                 conf->max_degraded = 1;
4174         conf->algorithm = mddev->layout;
4175         conf->max_nr_stripes = NR_STRIPES;
4176         conf->expand_progress = mddev->reshape_position;
4177
4178         /* device size must be a multiple of chunk size */
4179         mddev->size &= ~(mddev->chunk_size/1024 -1);
4180         mddev->resync_max_sectors = mddev->size << 1;
4181
4182         if (conf->level == 6 && conf->raid_disks < 4) {
4183                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4184                        mdname(mddev), conf->raid_disks);
4185                 goto abort;
4186         }
4187         if (!conf->chunk_size || conf->chunk_size % 4) {
4188                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4189                         conf->chunk_size, mdname(mddev));
4190                 goto abort;
4191         }
4192         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4193                 printk(KERN_ERR 
4194                         "raid5: unsupported parity algorithm %d for %s\n",
4195                         conf->algorithm, mdname(mddev));
4196                 goto abort;
4197         }
4198         if (mddev->degraded > conf->max_degraded) {
4199                 printk(KERN_ERR "raid5: not enough operational devices for %s"
4200                         " (%d/%d failed)\n",
4201                         mdname(mddev), mddev->degraded, conf->raid_disks);
4202                 goto abort;
4203         }
4204
4205         if (mddev->degraded > 0 &&
4206             mddev->recovery_cp != MaxSector) {
4207                 if (mddev->ok_start_degraded)
4208                         printk(KERN_WARNING
4209                                "raid5: starting dirty degraded array: %s"
4210                                "- data corruption possible.\n",
4211                                mdname(mddev));
4212                 else {
4213                         printk(KERN_ERR
4214                                "raid5: cannot start dirty degraded array for %s\n",
4215                                mdname(mddev));
4216                         goto abort;
4217                 }
4218         }
4219
4220         {
4221                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4222                 if (!mddev->thread) {
4223                         printk(KERN_ERR 
4224                                 "raid5: couldn't allocate thread for %s\n",
4225                                 mdname(mddev));
4226                         goto abort;
4227                 }
4228         }
4229         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4230                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4231         if (grow_stripes(conf, conf->max_nr_stripes)) {
4232                 printk(KERN_ERR 
4233                         "raid5: couldn't allocate %dkB for buffers\n", memory);
4234                 shrink_stripes(conf);
4235                 md_unregister_thread(mddev->thread);
4236                 goto abort;
4237         } else
4238                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4239                         memory, mdname(mddev));
4240
4241         if (mddev->degraded == 0)
4242                 printk("raid5: raid level %d set %s active with %d out of %d"
4243                         " devices, algorithm %d\n", conf->level, mdname(mddev), 
4244                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4245                         conf->algorithm);
4246         else
4247                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4248                         " out of %d devices, algorithm %d\n", conf->level,
4249                         mdname(mddev), mddev->raid_disks - mddev->degraded,
4250                         mddev->raid_disks, conf->algorithm);
4251
4252         print_raid5_conf(conf);
4253
4254         if (conf->expand_progress != MaxSector) {
4255                 printk("...ok start reshape thread\n");
4256                 conf->expand_lo = conf->expand_progress;
4257                 atomic_set(&conf->reshape_stripes, 0);
4258                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4259                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4260                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4261                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4262                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4263                                                         "%s_reshape");
4264         }
4265
4266         /* read-ahead size must cover two whole stripes, which is
4267          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4268          */
4269         {
4270                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4271                 int stripe = data_disks *
4272                         (mddev->chunk_size / PAGE_SIZE);
4273                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4274                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4275         }
4276
4277         /* Ok, everything is just fine now */
4278         if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4279                 printk(KERN_WARNING
4280                        "raid5: failed to create sysfs attributes for %s\n",
4281                        mdname(mddev));
4282
4283         mddev->queue->unplug_fn = raid5_unplug_device;
4284         mddev->queue->issue_flush_fn = raid5_issue_flush;
4285         mddev->queue->backing_dev_info.congested_data = mddev;
4286         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4287
4288         mddev->array_size =  mddev->size * (conf->previous_raid_disks -
4289                                             conf->max_degraded);
4290
4291         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4292
4293         return 0;
4294 abort:
4295         if (conf) {
4296                 print_raid5_conf(conf);
4297                 safe_put_page(conf->spare_page);
4298                 kfree(conf->disks);
4299                 kfree(conf->stripe_hashtbl);
4300                 kfree(conf);
4301         }
4302         mddev->private = NULL;
4303         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4304         return -EIO;
4305 }
4306
4307
4308
4309 static int stop(mddev_t *mddev)
4310 {
4311         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4312
4313         md_unregister_thread(mddev->thread);
4314         mddev->thread = NULL;
4315         shrink_stripes(conf);
4316         kfree(conf->stripe_hashtbl);
4317         mddev->queue->backing_dev_info.congested_fn = NULL;
4318         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4319         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4320         kfree(conf->disks);
4321         kfree(conf);
4322         mddev->private = NULL;
4323         return 0;
4324 }
4325
4326 #ifdef DEBUG
4327 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4328 {
4329         int i;
4330
4331         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4332                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4333         seq_printf(seq, "sh %llu,  count %d.\n",
4334                    (unsigned long long)sh->sector, atomic_read(&sh->count));
4335         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4336         for (i = 0; i < sh->disks; i++) {
4337                 seq_printf(seq, "(cache%d: %p %ld) ",
4338                            i, sh->dev[i].page, sh->dev[i].flags);
4339         }
4340         seq_printf(seq, "\n");
4341 }
4342
4343 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4344 {
4345         struct stripe_head *sh;
4346         struct hlist_node *hn;
4347         int i;
4348
4349         spin_lock_irq(&conf->device_lock);
4350         for (i = 0; i < NR_HASH; i++) {
4351                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4352                         if (sh->raid_conf != conf)
4353                                 continue;
4354                         print_sh(seq, sh);
4355                 }
4356         }
4357         spin_unlock_irq(&conf->device_lock);
4358 }
4359 #endif
4360
4361 static void status (struct seq_file *seq, mddev_t *mddev)
4362 {
4363         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4364         int i;
4365
4366         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4367         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4368         for (i = 0; i < conf->raid_disks; i++)
4369                 seq_printf (seq, "%s",
4370                                conf->disks[i].rdev &&
4371                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4372         seq_printf (seq, "]");
4373 #ifdef DEBUG
4374         seq_printf (seq, "\n");
4375         printall(seq, conf);
4376 #endif
4377 }
4378
4379 static void print_raid5_conf (raid5_conf_t *conf)
4380 {
4381         int i;
4382         struct disk_info *tmp;
4383
4384         printk("RAID5 conf printout:\n");
4385         if (!conf) {
4386                 printk("(conf==NULL)\n");
4387                 return;
4388         }
4389         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4390                  conf->raid_disks - conf->mddev->degraded);
4391
4392         for (i = 0; i < conf->raid_disks; i++) {
4393                 char b[BDEVNAME_SIZE];
4394                 tmp = conf->disks + i;
4395                 if (tmp->rdev)
4396                 printk(" disk %d, o:%d, dev:%s\n",
4397                         i, !test_bit(Faulty, &tmp->rdev->flags),
4398                         bdevname(tmp->rdev->bdev,b));
4399         }
4400 }
4401
4402 static int raid5_spare_active(mddev_t *mddev)
4403 {
4404         int i;
4405         raid5_conf_t *conf = mddev->private;
4406         struct disk_info *tmp;
4407
4408         for (i = 0; i < conf->raid_disks; i++) {
4409                 tmp = conf->disks + i;
4410                 if (tmp->rdev
4411                     && !test_bit(Faulty, &tmp->rdev->flags)
4412                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4413                         unsigned long flags;
4414                         spin_lock_irqsave(&conf->device_lock, flags);
4415                         mddev->degraded--;
4416                         spin_unlock_irqrestore(&conf->device_lock, flags);
4417                 }
4418         }
4419         print_raid5_conf(conf);
4420         return 0;
4421 }
4422
4423 static int raid5_remove_disk(mddev_t *mddev, int number)
4424 {
4425         raid5_conf_t *conf = mddev->private;
4426         int err = 0;
4427         mdk_rdev_t *rdev;
4428         struct disk_info *p = conf->disks + number;
4429
4430         print_raid5_conf(conf);
4431         rdev = p->rdev;
4432         if (rdev) {
4433                 if (test_bit(In_sync, &rdev->flags) ||
4434                     atomic_read(&rdev->nr_pending)) {
4435                         err = -EBUSY;
4436                         goto abort;
4437                 }
4438                 p->rdev = NULL;
4439                 synchronize_rcu();
4440                 if (atomic_read(&rdev->nr_pending)) {
4441                         /* lost the race, try later */
4442                         err = -EBUSY;
4443                         p->rdev = rdev;
4444                 }
4445         }
4446 abort:
4447
4448         print_raid5_conf(conf);
4449         return err;
4450 }
4451
4452 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4453 {
4454         raid5_conf_t *conf = mddev->private;
4455         int found = 0;
4456         int disk;
4457         struct disk_info *p;
4458
4459         if (mddev->degraded > conf->max_degraded)
4460                 /* no point adding a device */
4461                 return 0;
4462
4463         /*
4464          * find the disk ... but prefer rdev->saved_raid_disk
4465          * if possible.
4466          */
4467         if (rdev->saved_raid_disk >= 0 &&
4468             conf->disks[rdev->saved_raid_disk].rdev == NULL)
4469                 disk = rdev->saved_raid_disk;
4470         else
4471                 disk = 0;
4472         for ( ; disk < conf->raid_disks; disk++)
4473                 if ((p=conf->disks + disk)->rdev == NULL) {
4474                         clear_bit(In_sync, &rdev->flags);
4475                         rdev->raid_disk = disk;
4476                         found = 1;
4477                         if (rdev->saved_raid_disk != disk)
4478                                 conf->fullsync = 1;
4479                         rcu_assign_pointer(p->rdev, rdev);
4480                         break;
4481                 }
4482         print_raid5_conf(conf);
4483         return found;
4484 }
4485
4486 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4487 {
4488         /* no resync is happening, and there is enough space
4489          * on all devices, so we can resize.
4490          * We need to make sure resync covers any new space.
4491          * If the array is shrinking we should possibly wait until
4492          * any io in the removed space completes, but it hardly seems
4493          * worth it.
4494          */
4495         raid5_conf_t *conf = mddev_to_conf(mddev);
4496
4497         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4498         mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4499         set_capacity(mddev->gendisk, mddev->array_size << 1);
4500         mddev->changed = 1;
4501         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
4502                 mddev->recovery_cp = mddev->size << 1;
4503                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4504         }
4505         mddev->size = sectors /2;
4506         mddev->resync_max_sectors = sectors;
4507         return 0;
4508 }
4509
4510 #ifdef CONFIG_MD_RAID5_RESHAPE
4511 static int raid5_check_reshape(mddev_t *mddev)
4512 {
4513         raid5_conf_t *conf = mddev_to_conf(mddev);
4514         int err;
4515
4516         if (mddev->delta_disks < 0 ||
4517             mddev->new_level != mddev->level)
4518                 return -EINVAL; /* Cannot shrink array or change level yet */
4519         if (mddev->delta_disks == 0)
4520                 return 0; /* nothing to do */
4521
4522         /* Can only proceed if there are plenty of stripe_heads.
4523          * We need a minimum of one full stripe,, and for sensible progress
4524          * it is best to have about 4 times that.
4525          * If we require 4 times, then the default 256 4K stripe_heads will
4526          * allow for chunk sizes up to 256K, which is probably OK.
4527          * If the chunk size is greater, user-space should request more
4528          * stripe_heads first.
4529          */
4530         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4531             (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4532                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
4533                        (mddev->chunk_size / STRIPE_SIZE)*4);
4534                 return -ENOSPC;
4535         }
4536
4537         err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4538         if (err)
4539                 return err;
4540
4541         if (mddev->degraded > conf->max_degraded)
4542                 return -EINVAL;
4543         /* looks like we might be able to manage this */
4544         return 0;
4545 }
4546
4547 static int raid5_start_reshape(mddev_t *mddev)
4548 {
4549         raid5_conf_t *conf = mddev_to_conf(mddev);
4550         mdk_rdev_t *rdev;
4551         struct list_head *rtmp;
4552         int spares = 0;
4553         int added_devices = 0;
4554         unsigned long flags;
4555
4556         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4557                 return -EBUSY;
4558
4559         ITERATE_RDEV(mddev, rdev, rtmp)
4560                 if (rdev->raid_disk < 0 &&
4561                     !test_bit(Faulty, &rdev->flags))
4562                         spares++;
4563
4564         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4565                 /* Not enough devices even to make a degraded array
4566                  * of that size
4567                  */
4568                 return -EINVAL;
4569
4570         atomic_set(&conf->reshape_stripes, 0);
4571         spin_lock_irq(&conf->device_lock);
4572         conf->previous_raid_disks = conf->raid_disks;
4573         conf->raid_disks += mddev->delta_disks;
4574         conf->expand_progress = 0;
4575         conf->expand_lo = 0;
4576         spin_unlock_irq(&conf->device_lock);
4577
4578         /* Add some new drives, as many as will fit.
4579          * We know there are enough to make the newly sized array work.
4580          */
4581         ITERATE_RDEV(mddev, rdev, rtmp)
4582                 if (rdev->raid_disk < 0 &&
4583                     !test_bit(Faulty, &rdev->flags)) {
4584                         if (raid5_add_disk(mddev, rdev)) {
4585                                 char nm[20];
4586                                 set_bit(In_sync, &rdev->flags);
4587                                 added_devices++;
4588                                 rdev->recovery_offset = 0;
4589                                 sprintf(nm, "rd%d", rdev->raid_disk);
4590                                 if (sysfs_create_link(&mddev->kobj,
4591                                                       &rdev->kobj, nm))
4592                                         printk(KERN_WARNING
4593                                                "raid5: failed to create "
4594                                                " link %s for %s\n",
4595                                                nm, mdname(mddev));
4596                         } else
4597                                 break;
4598                 }
4599
4600         spin_lock_irqsave(&conf->device_lock, flags);
4601         mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4602         spin_unlock_irqrestore(&conf->device_lock, flags);
4603         mddev->raid_disks = conf->raid_disks;
4604         mddev->reshape_position = 0;
4605         set_bit(MD_CHANGE_DEVS, &mddev->flags);
4606
4607         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4608         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4609         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4610         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4611         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4612                                                 "%s_reshape");
4613         if (!mddev->sync_thread) {
4614                 mddev->recovery = 0;
4615                 spin_lock_irq(&conf->device_lock);
4616                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4617                 conf->expand_progress = MaxSector;
4618                 spin_unlock_irq(&conf->device_lock);
4619                 return -EAGAIN;
4620         }
4621         md_wakeup_thread(mddev->sync_thread);
4622         md_new_event(mddev);
4623         return 0;
4624 }
4625 #endif
4626
4627 static void end_reshape(raid5_conf_t *conf)
4628 {
4629         struct block_device *bdev;
4630
4631         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4632                 conf->mddev->array_size = conf->mddev->size *
4633                         (conf->raid_disks - conf->max_degraded);
4634                 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4635                 conf->mddev->changed = 1;
4636
4637                 bdev = bdget_disk(conf->mddev->gendisk, 0);
4638                 if (bdev) {
4639                         mutex_lock(&bdev->bd_inode->i_mutex);
4640                         i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4641                         mutex_unlock(&bdev->bd_inode->i_mutex);
4642                         bdput(bdev);
4643                 }
4644                 spin_lock_irq(&conf->device_lock);
4645                 conf->expand_progress = MaxSector;
4646                 spin_unlock_irq(&conf->device_lock);
4647                 conf->mddev->reshape_position = MaxSector;
4648
4649                 /* read-ahead size must cover two whole stripes, which is
4650                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4651                  */
4652                 {
4653                         int data_disks = conf->previous_raid_disks - conf->max_degraded;
4654                         int stripe = data_disks *
4655                                 (conf->mddev->chunk_size / PAGE_SIZE);
4656                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4657                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4658                 }
4659         }
4660 }
4661
4662 static void raid5_quiesce(mddev_t *mddev, int state)
4663 {
4664         raid5_conf_t *conf = mddev_to_conf(mddev);
4665
4666         switch(state) {
4667         case 2: /* resume for a suspend */
4668                 wake_up(&conf->wait_for_overlap);
4669                 break;
4670
4671         case 1: /* stop all writes */
4672                 spin_lock_irq(&conf->device_lock);
4673                 conf->quiesce = 1;
4674                 wait_event_lock_irq(conf->wait_for_stripe,
4675                                     atomic_read(&conf->active_stripes) == 0 &&
4676                                     atomic_read(&conf->active_aligned_reads) == 0,
4677                                     conf->device_lock, /* nothing */);
4678                 spin_unlock_irq(&conf->device_lock);
4679                 break;
4680
4681         case 0: /* re-enable writes */
4682                 spin_lock_irq(&conf->device_lock);
4683                 conf->quiesce = 0;
4684                 wake_up(&conf->wait_for_stripe);
4685                 wake_up(&conf->wait_for_overlap);
4686                 spin_unlock_irq(&conf->device_lock);
4687                 break;
4688         }
4689 }
4690
4691 static struct mdk_personality raid6_personality =
4692 {
4693         .name           = "raid6",
4694         .level          = 6,
4695         .owner          = THIS_MODULE,
4696         .make_request   = make_request,
4697         .run            = run,
4698         .stop           = stop,
4699         .status         = status,
4700         .error_handler  = error,
4701         .hot_add_disk   = raid5_add_disk,
4702         .hot_remove_disk= raid5_remove_disk,
4703         .spare_active   = raid5_spare_active,
4704         .sync_request   = sync_request,
4705         .resize         = raid5_resize,
4706 #ifdef CONFIG_MD_RAID5_RESHAPE
4707         .check_reshape  = raid5_check_reshape,
4708         .start_reshape  = raid5_start_reshape,
4709 #endif
4710         .quiesce        = raid5_quiesce,
4711 };
4712 static struct mdk_personality raid5_personality =
4713 {
4714         .name           = "raid5",
4715         .level          = 5,
4716         .owner          = THIS_MODULE,
4717         .make_request   = make_request,
4718         .run            = run,
4719         .stop           = stop,
4720         .status         = status,
4721         .error_handler  = error,
4722         .hot_add_disk   = raid5_add_disk,
4723         .hot_remove_disk= raid5_remove_disk,
4724         .spare_active   = raid5_spare_active,
4725         .sync_request   = sync_request,
4726         .resize         = raid5_resize,
4727 #ifdef CONFIG_MD_RAID5_RESHAPE
4728         .check_reshape  = raid5_check_reshape,
4729         .start_reshape  = raid5_start_reshape,
4730 #endif
4731         .quiesce        = raid5_quiesce,
4732 };
4733
4734 static struct mdk_personality raid4_personality =
4735 {
4736         .name           = "raid4",
4737         .level          = 4,
4738         .owner          = THIS_MODULE,
4739         .make_request   = make_request,
4740         .run            = run,
4741         .stop           = stop,
4742         .status         = status,
4743         .error_handler  = error,
4744         .hot_add_disk   = raid5_add_disk,
4745         .hot_remove_disk= raid5_remove_disk,
4746         .spare_active   = raid5_spare_active,
4747         .sync_request   = sync_request,
4748         .resize         = raid5_resize,
4749 #ifdef CONFIG_MD_RAID5_RESHAPE
4750         .check_reshape  = raid5_check_reshape,
4751         .start_reshape  = raid5_start_reshape,
4752 #endif
4753         .quiesce        = raid5_quiesce,
4754 };
4755
4756 static int __init raid5_init(void)
4757 {
4758         int e;
4759
4760         e = raid6_select_algo();
4761         if ( e )
4762                 return e;
4763         register_md_personality(&raid6_personality);
4764         register_md_personality(&raid5_personality);
4765         register_md_personality(&raid4_personality);
4766         return 0;
4767 }
4768
4769 static void raid5_exit(void)
4770 {
4771         unregister_md_personality(&raid6_personality);
4772         unregister_md_personality(&raid5_personality);
4773         unregister_md_personality(&raid4_personality);
4774 }
4775
4776 module_init(raid5_init);
4777 module_exit(raid5_exit);
4778 MODULE_LICENSE("GPL");
4779 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4780 MODULE_ALIAS("md-raid5");
4781 MODULE_ALIAS("md-raid4");
4782 MODULE_ALIAS("md-level-5");
4783 MODULE_ALIAS("md-level-4");
4784 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4785 MODULE_ALIAS("md-raid6");
4786 MODULE_ALIAS("md-level-6");
4787
4788 /* This used to be two separate modules, they were: */
4789 MODULE_ALIAS("raid5");
4790 MODULE_ALIAS("raid6");