]> err.no Git - linux-2.6/blob - drivers/md/raid5.c
Merge branch 'for-2.6.27' of git://linux-nfs.org/~bfields/linux
[linux-2.6] / drivers / md / raid5.c
1 /*
2  * raid5.c : Multiple Devices driver for Linux
3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4  *         Copyright (C) 1999, 2000 Ingo Molnar
5  *         Copyright (C) 2002, 2003 H. Peter Anvin
6  *
7  * RAID-4/5/6 management functions.
8  * Thanks to Penguin Computing for making the RAID-6 development possible
9  * by donating a test server!
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 /*
22  * BITMAP UNPLUGGING:
23  *
24  * The sequencing for updating the bitmap reliably is a little
25  * subtle (and I got it wrong the first time) so it deserves some
26  * explanation.
27  *
28  * We group bitmap updates into batches.  Each batch has a number.
29  * We may write out several batches at once, but that isn't very important.
30  * conf->bm_write is the number of the last batch successfully written.
31  * conf->bm_flush is the number of the last batch that was closed to
32  *    new additions.
33  * When we discover that we will need to write to any block in a stripe
34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35  * the number of the batch it will be in. This is bm_flush+1.
36  * When we are ready to do a write, if that batch hasn't been written yet,
37  *   we plug the array and queue the stripe for later.
38  * When an unplug happens, we increment bm_flush, thus closing the current
39  *   batch.
40  * When we notice that bm_flush > bm_write, we write out all pending updates
41  * to the bitmap, and advance bm_write to where bm_flush was.
42  * This may occasionally write a bit out twice, but is sure never to
43  * miss any bits.
44  */
45
46 #include <linux/module.h>
47 #include <linux/slab.h>
48 #include <linux/highmem.h>
49 #include <linux/bitops.h>
50 #include <linux/kthread.h>
51 #include <asm/atomic.h>
52 #include "raid6.h"
53
54 #include <linux/raid/bitmap.h>
55 #include <linux/async_tx.h>
56
57 /*
58  * Stripe cache
59  */
60
61 #define NR_STRIPES              256
62 #define STRIPE_SIZE             PAGE_SIZE
63 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
64 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
65 #define IO_THRESHOLD            1
66 #define BYPASS_THRESHOLD        1
67 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
68 #define HASH_MASK               (NR_HASH - 1)
69
70 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
71
72 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
73  * order without overlap.  There may be several bio's per stripe+device, and
74  * a bio could span several devices.
75  * When walking this list for a particular stripe+device, we must never proceed
76  * beyond a bio that extends past this device, as the next bio might no longer
77  * be valid.
78  * This macro is used to determine the 'next' bio in the list, given the sector
79  * of the current stripe+device
80  */
81 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
82 /*
83  * The following can be used to debug the driver
84  */
85 #define RAID5_PARANOIA  1
86 #if RAID5_PARANOIA && defined(CONFIG_SMP)
87 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
88 #else
89 # define CHECK_DEVLOCK()
90 #endif
91
92 #ifdef DEBUG
93 #define inline
94 #define __inline__
95 #endif
96
97 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
98
99 #if !RAID6_USE_EMPTY_ZERO_PAGE
100 /* In .bss so it's zeroed */
101 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
102 #endif
103
104 static inline int raid6_next_disk(int disk, int raid_disks)
105 {
106         disk++;
107         return (disk < raid_disks) ? disk : 0;
108 }
109
110 static void return_io(struct bio *return_bi)
111 {
112         struct bio *bi = return_bi;
113         while (bi) {
114
115                 return_bi = bi->bi_next;
116                 bi->bi_next = NULL;
117                 bi->bi_size = 0;
118                 bi->bi_end_io(bi,
119                               test_bit(BIO_UPTODATE, &bi->bi_flags)
120                                 ? 0 : -EIO);
121                 bi = return_bi;
122         }
123 }
124
125 static void print_raid5_conf (raid5_conf_t *conf);
126
127 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
128 {
129         if (atomic_dec_and_test(&sh->count)) {
130                 BUG_ON(!list_empty(&sh->lru));
131                 BUG_ON(atomic_read(&conf->active_stripes)==0);
132                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
133                         if (test_bit(STRIPE_DELAYED, &sh->state)) {
134                                 list_add_tail(&sh->lru, &conf->delayed_list);
135                                 blk_plug_device(conf->mddev->queue);
136                         } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
137                                    sh->bm_seq - conf->seq_write > 0) {
138                                 list_add_tail(&sh->lru, &conf->bitmap_list);
139                                 blk_plug_device(conf->mddev->queue);
140                         } else {
141                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
142                                 list_add_tail(&sh->lru, &conf->handle_list);
143                         }
144                         md_wakeup_thread(conf->mddev->thread);
145                 } else {
146                         BUG_ON(sh->ops.pending);
147                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
148                                 atomic_dec(&conf->preread_active_stripes);
149                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
150                                         md_wakeup_thread(conf->mddev->thread);
151                         }
152                         atomic_dec(&conf->active_stripes);
153                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
154                                 list_add_tail(&sh->lru, &conf->inactive_list);
155                                 wake_up(&conf->wait_for_stripe);
156                                 if (conf->retry_read_aligned)
157                                         md_wakeup_thread(conf->mddev->thread);
158                         }
159                 }
160         }
161 }
162 static void release_stripe(struct stripe_head *sh)
163 {
164         raid5_conf_t *conf = sh->raid_conf;
165         unsigned long flags;
166
167         spin_lock_irqsave(&conf->device_lock, flags);
168         __release_stripe(conf, sh);
169         spin_unlock_irqrestore(&conf->device_lock, flags);
170 }
171
172 static inline void remove_hash(struct stripe_head *sh)
173 {
174         pr_debug("remove_hash(), stripe %llu\n",
175                 (unsigned long long)sh->sector);
176
177         hlist_del_init(&sh->hash);
178 }
179
180 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
181 {
182         struct hlist_head *hp = stripe_hash(conf, sh->sector);
183
184         pr_debug("insert_hash(), stripe %llu\n",
185                 (unsigned long long)sh->sector);
186
187         CHECK_DEVLOCK();
188         hlist_add_head(&sh->hash, hp);
189 }
190
191
192 /* find an idle stripe, make sure it is unhashed, and return it. */
193 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
194 {
195         struct stripe_head *sh = NULL;
196         struct list_head *first;
197
198         CHECK_DEVLOCK();
199         if (list_empty(&conf->inactive_list))
200                 goto out;
201         first = conf->inactive_list.next;
202         sh = list_entry(first, struct stripe_head, lru);
203         list_del_init(first);
204         remove_hash(sh);
205         atomic_inc(&conf->active_stripes);
206 out:
207         return sh;
208 }
209
210 static void shrink_buffers(struct stripe_head *sh, int num)
211 {
212         struct page *p;
213         int i;
214
215         for (i=0; i<num ; i++) {
216                 p = sh->dev[i].page;
217                 if (!p)
218                         continue;
219                 sh->dev[i].page = NULL;
220                 put_page(p);
221         }
222 }
223
224 static int grow_buffers(struct stripe_head *sh, int num)
225 {
226         int i;
227
228         for (i=0; i<num; i++) {
229                 struct page *page;
230
231                 if (!(page = alloc_page(GFP_KERNEL))) {
232                         return 1;
233                 }
234                 sh->dev[i].page = page;
235         }
236         return 0;
237 }
238
239 static void raid5_build_block (struct stripe_head *sh, int i);
240
241 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks)
242 {
243         raid5_conf_t *conf = sh->raid_conf;
244         int i;
245
246         BUG_ON(atomic_read(&sh->count) != 0);
247         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
248         BUG_ON(sh->ops.pending || sh->ops.ack || sh->ops.complete);
249
250         CHECK_DEVLOCK();
251         pr_debug("init_stripe called, stripe %llu\n",
252                 (unsigned long long)sh->sector);
253
254         remove_hash(sh);
255
256         sh->sector = sector;
257         sh->pd_idx = pd_idx;
258         sh->state = 0;
259
260         sh->disks = disks;
261
262         for (i = sh->disks; i--; ) {
263                 struct r5dev *dev = &sh->dev[i];
264
265                 if (dev->toread || dev->read || dev->towrite || dev->written ||
266                     test_bit(R5_LOCKED, &dev->flags)) {
267                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
268                                (unsigned long long)sh->sector, i, dev->toread,
269                                dev->read, dev->towrite, dev->written,
270                                test_bit(R5_LOCKED, &dev->flags));
271                         BUG();
272                 }
273                 dev->flags = 0;
274                 raid5_build_block(sh, i);
275         }
276         insert_hash(conf, sh);
277 }
278
279 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
280 {
281         struct stripe_head *sh;
282         struct hlist_node *hn;
283
284         CHECK_DEVLOCK();
285         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
286         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
287                 if (sh->sector == sector && sh->disks == disks)
288                         return sh;
289         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
290         return NULL;
291 }
292
293 static void unplug_slaves(mddev_t *mddev);
294 static void raid5_unplug_device(struct request_queue *q);
295
296 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks,
297                                              int pd_idx, int noblock)
298 {
299         struct stripe_head *sh;
300
301         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
302
303         spin_lock_irq(&conf->device_lock);
304
305         do {
306                 wait_event_lock_irq(conf->wait_for_stripe,
307                                     conf->quiesce == 0,
308                                     conf->device_lock, /* nothing */);
309                 sh = __find_stripe(conf, sector, disks);
310                 if (!sh) {
311                         if (!conf->inactive_blocked)
312                                 sh = get_free_stripe(conf);
313                         if (noblock && sh == NULL)
314                                 break;
315                         if (!sh) {
316                                 conf->inactive_blocked = 1;
317                                 wait_event_lock_irq(conf->wait_for_stripe,
318                                                     !list_empty(&conf->inactive_list) &&
319                                                     (atomic_read(&conf->active_stripes)
320                                                      < (conf->max_nr_stripes *3/4)
321                                                      || !conf->inactive_blocked),
322                                                     conf->device_lock,
323                                                     raid5_unplug_device(conf->mddev->queue)
324                                         );
325                                 conf->inactive_blocked = 0;
326                         } else
327                                 init_stripe(sh, sector, pd_idx, disks);
328                 } else {
329                         if (atomic_read(&sh->count)) {
330                           BUG_ON(!list_empty(&sh->lru));
331                         } else {
332                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
333                                         atomic_inc(&conf->active_stripes);
334                                 if (list_empty(&sh->lru) &&
335                                     !test_bit(STRIPE_EXPANDING, &sh->state))
336                                         BUG();
337                                 list_del_init(&sh->lru);
338                         }
339                 }
340         } while (sh == NULL);
341
342         if (sh)
343                 atomic_inc(&sh->count);
344
345         spin_unlock_irq(&conf->device_lock);
346         return sh;
347 }
348
349 /* test_and_ack_op() ensures that we only dequeue an operation once */
350 #define test_and_ack_op(op, pend) \
351 do {                                                    \
352         if (test_bit(op, &sh->ops.pending) &&           \
353                 !test_bit(op, &sh->ops.complete)) {     \
354                 if (test_and_set_bit(op, &sh->ops.ack)) \
355                         clear_bit(op, &pend);           \
356                 else                                    \
357                         ack++;                          \
358         } else                                          \
359                 clear_bit(op, &pend);                   \
360 } while (0)
361
362 /* find new work to run, do not resubmit work that is already
363  * in flight
364  */
365 static unsigned long get_stripe_work(struct stripe_head *sh)
366 {
367         unsigned long pending;
368         int ack = 0;
369
370         pending = sh->ops.pending;
371
372         test_and_ack_op(STRIPE_OP_BIOFILL, pending);
373         test_and_ack_op(STRIPE_OP_COMPUTE_BLK, pending);
374         test_and_ack_op(STRIPE_OP_PREXOR, pending);
375         test_and_ack_op(STRIPE_OP_BIODRAIN, pending);
376         test_and_ack_op(STRIPE_OP_POSTXOR, pending);
377         test_and_ack_op(STRIPE_OP_CHECK, pending);
378         if (test_and_clear_bit(STRIPE_OP_IO, &sh->ops.pending))
379                 ack++;
380
381         sh->ops.count -= ack;
382         if (unlikely(sh->ops.count < 0)) {
383                 printk(KERN_ERR "pending: %#lx ops.pending: %#lx ops.ack: %#lx "
384                         "ops.complete: %#lx\n", pending, sh->ops.pending,
385                         sh->ops.ack, sh->ops.complete);
386                 BUG();
387         }
388
389         return pending;
390 }
391
392 static void
393 raid5_end_read_request(struct bio *bi, int error);
394 static void
395 raid5_end_write_request(struct bio *bi, int error);
396
397 static void ops_run_io(struct stripe_head *sh)
398 {
399         raid5_conf_t *conf = sh->raid_conf;
400         int i, disks = sh->disks;
401
402         might_sleep();
403
404         set_bit(STRIPE_IO_STARTED, &sh->state);
405         for (i = disks; i--; ) {
406                 int rw;
407                 struct bio *bi;
408                 mdk_rdev_t *rdev;
409                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
410                         rw = WRITE;
411                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
412                         rw = READ;
413                 else
414                         continue;
415
416                 bi = &sh->dev[i].req;
417
418                 bi->bi_rw = rw;
419                 if (rw == WRITE)
420                         bi->bi_end_io = raid5_end_write_request;
421                 else
422                         bi->bi_end_io = raid5_end_read_request;
423
424                 rcu_read_lock();
425                 rdev = rcu_dereference(conf->disks[i].rdev);
426                 if (rdev && test_bit(Faulty, &rdev->flags))
427                         rdev = NULL;
428                 if (rdev)
429                         atomic_inc(&rdev->nr_pending);
430                 rcu_read_unlock();
431
432                 if (rdev) {
433                         if (test_bit(STRIPE_SYNCING, &sh->state) ||
434                                 test_bit(STRIPE_EXPAND_SOURCE, &sh->state) ||
435                                 test_bit(STRIPE_EXPAND_READY, &sh->state))
436                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
437
438                         bi->bi_bdev = rdev->bdev;
439                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
440                                 __func__, (unsigned long long)sh->sector,
441                                 bi->bi_rw, i);
442                         atomic_inc(&sh->count);
443                         bi->bi_sector = sh->sector + rdev->data_offset;
444                         bi->bi_flags = 1 << BIO_UPTODATE;
445                         bi->bi_vcnt = 1;
446                         bi->bi_max_vecs = 1;
447                         bi->bi_idx = 0;
448                         bi->bi_io_vec = &sh->dev[i].vec;
449                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
450                         bi->bi_io_vec[0].bv_offset = 0;
451                         bi->bi_size = STRIPE_SIZE;
452                         bi->bi_next = NULL;
453                         if (rw == WRITE &&
454                             test_bit(R5_ReWrite, &sh->dev[i].flags))
455                                 atomic_add(STRIPE_SECTORS,
456                                         &rdev->corrected_errors);
457                         generic_make_request(bi);
458                 } else {
459                         if (rw == WRITE)
460                                 set_bit(STRIPE_DEGRADED, &sh->state);
461                         pr_debug("skip op %ld on disc %d for sector %llu\n",
462                                 bi->bi_rw, i, (unsigned long long)sh->sector);
463                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
464                         set_bit(STRIPE_HANDLE, &sh->state);
465                 }
466         }
467 }
468
469 static struct dma_async_tx_descriptor *
470 async_copy_data(int frombio, struct bio *bio, struct page *page,
471         sector_t sector, struct dma_async_tx_descriptor *tx)
472 {
473         struct bio_vec *bvl;
474         struct page *bio_page;
475         int i;
476         int page_offset;
477
478         if (bio->bi_sector >= sector)
479                 page_offset = (signed)(bio->bi_sector - sector) * 512;
480         else
481                 page_offset = (signed)(sector - bio->bi_sector) * -512;
482         bio_for_each_segment(bvl, bio, i) {
483                 int len = bio_iovec_idx(bio, i)->bv_len;
484                 int clen;
485                 int b_offset = 0;
486
487                 if (page_offset < 0) {
488                         b_offset = -page_offset;
489                         page_offset += b_offset;
490                         len -= b_offset;
491                 }
492
493                 if (len > 0 && page_offset + len > STRIPE_SIZE)
494                         clen = STRIPE_SIZE - page_offset;
495                 else
496                         clen = len;
497
498                 if (clen > 0) {
499                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
500                         bio_page = bio_iovec_idx(bio, i)->bv_page;
501                         if (frombio)
502                                 tx = async_memcpy(page, bio_page, page_offset,
503                                         b_offset, clen,
504                                         ASYNC_TX_DEP_ACK,
505                                         tx, NULL, NULL);
506                         else
507                                 tx = async_memcpy(bio_page, page, b_offset,
508                                         page_offset, clen,
509                                         ASYNC_TX_DEP_ACK,
510                                         tx, NULL, NULL);
511                 }
512                 if (clen < len) /* hit end of page */
513                         break;
514                 page_offset +=  len;
515         }
516
517         return tx;
518 }
519
520 static void ops_complete_biofill(void *stripe_head_ref)
521 {
522         struct stripe_head *sh = stripe_head_ref;
523         struct bio *return_bi = NULL;
524         raid5_conf_t *conf = sh->raid_conf;
525         int i;
526
527         pr_debug("%s: stripe %llu\n", __func__,
528                 (unsigned long long)sh->sector);
529
530         /* clear completed biofills */
531         for (i = sh->disks; i--; ) {
532                 struct r5dev *dev = &sh->dev[i];
533
534                 /* acknowledge completion of a biofill operation */
535                 /* and check if we need to reply to a read request,
536                  * new R5_Wantfill requests are held off until
537                  * !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending)
538                  */
539                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
540                         struct bio *rbi, *rbi2;
541
542                         /* The access to dev->read is outside of the
543                          * spin_lock_irq(&conf->device_lock), but is protected
544                          * by the STRIPE_OP_BIOFILL pending bit
545                          */
546                         BUG_ON(!dev->read);
547                         rbi = dev->read;
548                         dev->read = NULL;
549                         while (rbi && rbi->bi_sector <
550                                 dev->sector + STRIPE_SECTORS) {
551                                 rbi2 = r5_next_bio(rbi, dev->sector);
552                                 spin_lock_irq(&conf->device_lock);
553                                 if (--rbi->bi_phys_segments == 0) {
554                                         rbi->bi_next = return_bi;
555                                         return_bi = rbi;
556                                 }
557                                 spin_unlock_irq(&conf->device_lock);
558                                 rbi = rbi2;
559                         }
560                 }
561         }
562         set_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
563
564         return_io(return_bi);
565
566         set_bit(STRIPE_HANDLE, &sh->state);
567         release_stripe(sh);
568 }
569
570 static void ops_run_biofill(struct stripe_head *sh)
571 {
572         struct dma_async_tx_descriptor *tx = NULL;
573         raid5_conf_t *conf = sh->raid_conf;
574         int i;
575
576         pr_debug("%s: stripe %llu\n", __func__,
577                 (unsigned long long)sh->sector);
578
579         for (i = sh->disks; i--; ) {
580                 struct r5dev *dev = &sh->dev[i];
581                 if (test_bit(R5_Wantfill, &dev->flags)) {
582                         struct bio *rbi;
583                         spin_lock_irq(&conf->device_lock);
584                         dev->read = rbi = dev->toread;
585                         dev->toread = NULL;
586                         spin_unlock_irq(&conf->device_lock);
587                         while (rbi && rbi->bi_sector <
588                                 dev->sector + STRIPE_SECTORS) {
589                                 tx = async_copy_data(0, rbi, dev->page,
590                                         dev->sector, tx);
591                                 rbi = r5_next_bio(rbi, dev->sector);
592                         }
593                 }
594         }
595
596         atomic_inc(&sh->count);
597         async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
598                 ops_complete_biofill, sh);
599 }
600
601 static void ops_complete_compute5(void *stripe_head_ref)
602 {
603         struct stripe_head *sh = stripe_head_ref;
604         int target = sh->ops.target;
605         struct r5dev *tgt = &sh->dev[target];
606
607         pr_debug("%s: stripe %llu\n", __func__,
608                 (unsigned long long)sh->sector);
609
610         set_bit(R5_UPTODATE, &tgt->flags);
611         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
612         clear_bit(R5_Wantcompute, &tgt->flags);
613         set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
614         set_bit(STRIPE_HANDLE, &sh->state);
615         release_stripe(sh);
616 }
617
618 static struct dma_async_tx_descriptor *
619 ops_run_compute5(struct stripe_head *sh, unsigned long pending)
620 {
621         /* kernel stack size limits the total number of disks */
622         int disks = sh->disks;
623         struct page *xor_srcs[disks];
624         int target = sh->ops.target;
625         struct r5dev *tgt = &sh->dev[target];
626         struct page *xor_dest = tgt->page;
627         int count = 0;
628         struct dma_async_tx_descriptor *tx;
629         int i;
630
631         pr_debug("%s: stripe %llu block: %d\n",
632                 __func__, (unsigned long long)sh->sector, target);
633         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
634
635         for (i = disks; i--; )
636                 if (i != target)
637                         xor_srcs[count++] = sh->dev[i].page;
638
639         atomic_inc(&sh->count);
640
641         if (unlikely(count == 1))
642                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
643                         0, NULL, ops_complete_compute5, sh);
644         else
645                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
646                         ASYNC_TX_XOR_ZERO_DST, NULL,
647                         ops_complete_compute5, sh);
648
649         /* ack now if postxor is not set to be run */
650         if (tx && !test_bit(STRIPE_OP_POSTXOR, &pending))
651                 async_tx_ack(tx);
652
653         return tx;
654 }
655
656 static void ops_complete_prexor(void *stripe_head_ref)
657 {
658         struct stripe_head *sh = stripe_head_ref;
659
660         pr_debug("%s: stripe %llu\n", __func__,
661                 (unsigned long long)sh->sector);
662
663         set_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
664 }
665
666 static struct dma_async_tx_descriptor *
667 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
668 {
669         /* kernel stack size limits the total number of disks */
670         int disks = sh->disks;
671         struct page *xor_srcs[disks];
672         int count = 0, pd_idx = sh->pd_idx, i;
673
674         /* existing parity data subtracted */
675         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
676
677         pr_debug("%s: stripe %llu\n", __func__,
678                 (unsigned long long)sh->sector);
679
680         for (i = disks; i--; ) {
681                 struct r5dev *dev = &sh->dev[i];
682                 /* Only process blocks that are known to be uptodate */
683                 if (dev->towrite && test_bit(R5_Wantprexor, &dev->flags))
684                         xor_srcs[count++] = dev->page;
685         }
686
687         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
688                 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
689                 ops_complete_prexor, sh);
690
691         return tx;
692 }
693
694 static struct dma_async_tx_descriptor *
695 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx,
696                  unsigned long pending)
697 {
698         int disks = sh->disks;
699         int pd_idx = sh->pd_idx, i;
700
701         /* check if prexor is active which means only process blocks
702          * that are part of a read-modify-write (Wantprexor)
703          */
704         int prexor = test_bit(STRIPE_OP_PREXOR, &pending);
705
706         pr_debug("%s: stripe %llu\n", __func__,
707                 (unsigned long long)sh->sector);
708
709         for (i = disks; i--; ) {
710                 struct r5dev *dev = &sh->dev[i];
711                 struct bio *chosen;
712                 int towrite;
713
714                 towrite = 0;
715                 if (prexor) { /* rmw */
716                         if (dev->towrite &&
717                             test_bit(R5_Wantprexor, &dev->flags))
718                                 towrite = 1;
719                 } else { /* rcw */
720                         if (i != pd_idx && dev->towrite &&
721                                 test_bit(R5_LOCKED, &dev->flags))
722                                 towrite = 1;
723                 }
724
725                 if (towrite) {
726                         struct bio *wbi;
727
728                         spin_lock(&sh->lock);
729                         chosen = dev->towrite;
730                         dev->towrite = NULL;
731                         BUG_ON(dev->written);
732                         wbi = dev->written = chosen;
733                         spin_unlock(&sh->lock);
734
735                         while (wbi && wbi->bi_sector <
736                                 dev->sector + STRIPE_SECTORS) {
737                                 tx = async_copy_data(1, wbi, dev->page,
738                                         dev->sector, tx);
739                                 wbi = r5_next_bio(wbi, dev->sector);
740                         }
741                 }
742         }
743
744         return tx;
745 }
746
747 static void ops_complete_postxor(void *stripe_head_ref)
748 {
749         struct stripe_head *sh = stripe_head_ref;
750
751         pr_debug("%s: stripe %llu\n", __func__,
752                 (unsigned long long)sh->sector);
753
754         set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
755         set_bit(STRIPE_HANDLE, &sh->state);
756         release_stripe(sh);
757 }
758
759 static void ops_complete_write(void *stripe_head_ref)
760 {
761         struct stripe_head *sh = stripe_head_ref;
762         int disks = sh->disks, i, pd_idx = sh->pd_idx;
763
764         pr_debug("%s: stripe %llu\n", __func__,
765                 (unsigned long long)sh->sector);
766
767         for (i = disks; i--; ) {
768                 struct r5dev *dev = &sh->dev[i];
769                 if (dev->written || i == pd_idx)
770                         set_bit(R5_UPTODATE, &dev->flags);
771         }
772
773         set_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
774         set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
775
776         set_bit(STRIPE_HANDLE, &sh->state);
777         release_stripe(sh);
778 }
779
780 static void
781 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx,
782                 unsigned long pending)
783 {
784         /* kernel stack size limits the total number of disks */
785         int disks = sh->disks;
786         struct page *xor_srcs[disks];
787
788         int count = 0, pd_idx = sh->pd_idx, i;
789         struct page *xor_dest;
790         int prexor = test_bit(STRIPE_OP_PREXOR, &pending);
791         unsigned long flags;
792         dma_async_tx_callback callback;
793
794         pr_debug("%s: stripe %llu\n", __func__,
795                 (unsigned long long)sh->sector);
796
797         /* check if prexor is active which means only process blocks
798          * that are part of a read-modify-write (written)
799          */
800         if (prexor) {
801                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
802                 for (i = disks; i--; ) {
803                         struct r5dev *dev = &sh->dev[i];
804                         if (dev->written)
805                                 xor_srcs[count++] = dev->page;
806                 }
807         } else {
808                 xor_dest = sh->dev[pd_idx].page;
809                 for (i = disks; i--; ) {
810                         struct r5dev *dev = &sh->dev[i];
811                         if (i != pd_idx)
812                                 xor_srcs[count++] = dev->page;
813                 }
814         }
815
816         /* check whether this postxor is part of a write */
817         callback = test_bit(STRIPE_OP_BIODRAIN, &pending) ?
818                 ops_complete_write : ops_complete_postxor;
819
820         /* 1/ if we prexor'd then the dest is reused as a source
821          * 2/ if we did not prexor then we are redoing the parity
822          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
823          * for the synchronous xor case
824          */
825         flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
826                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
827
828         atomic_inc(&sh->count);
829
830         if (unlikely(count == 1)) {
831                 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
832                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
833                         flags, tx, callback, sh);
834         } else
835                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
836                         flags, tx, callback, sh);
837 }
838
839 static void ops_complete_check(void *stripe_head_ref)
840 {
841         struct stripe_head *sh = stripe_head_ref;
842         int pd_idx = sh->pd_idx;
843
844         pr_debug("%s: stripe %llu\n", __func__,
845                 (unsigned long long)sh->sector);
846
847         if (test_and_clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending) &&
848                 sh->ops.zero_sum_result == 0)
849                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
850
851         set_bit(STRIPE_OP_CHECK, &sh->ops.complete);
852         set_bit(STRIPE_HANDLE, &sh->state);
853         release_stripe(sh);
854 }
855
856 static void ops_run_check(struct stripe_head *sh)
857 {
858         /* kernel stack size limits the total number of disks */
859         int disks = sh->disks;
860         struct page *xor_srcs[disks];
861         struct dma_async_tx_descriptor *tx;
862
863         int count = 0, pd_idx = sh->pd_idx, i;
864         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
865
866         pr_debug("%s: stripe %llu\n", __func__,
867                 (unsigned long long)sh->sector);
868
869         for (i = disks; i--; ) {
870                 struct r5dev *dev = &sh->dev[i];
871                 if (i != pd_idx)
872                         xor_srcs[count++] = dev->page;
873         }
874
875         tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
876                 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
877
878         if (tx)
879                 set_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
880         else
881                 clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
882
883         atomic_inc(&sh->count);
884         tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
885                 ops_complete_check, sh);
886 }
887
888 static void raid5_run_ops(struct stripe_head *sh, unsigned long pending)
889 {
890         int overlap_clear = 0, i, disks = sh->disks;
891         struct dma_async_tx_descriptor *tx = NULL;
892
893         if (test_bit(STRIPE_OP_BIOFILL, &pending)) {
894                 ops_run_biofill(sh);
895                 overlap_clear++;
896         }
897
898         if (test_bit(STRIPE_OP_COMPUTE_BLK, &pending))
899                 tx = ops_run_compute5(sh, pending);
900
901         if (test_bit(STRIPE_OP_PREXOR, &pending))
902                 tx = ops_run_prexor(sh, tx);
903
904         if (test_bit(STRIPE_OP_BIODRAIN, &pending)) {
905                 tx = ops_run_biodrain(sh, tx, pending);
906                 overlap_clear++;
907         }
908
909         if (test_bit(STRIPE_OP_POSTXOR, &pending))
910                 ops_run_postxor(sh, tx, pending);
911
912         if (test_bit(STRIPE_OP_CHECK, &pending))
913                 ops_run_check(sh);
914
915         if (test_bit(STRIPE_OP_IO, &pending))
916                 ops_run_io(sh);
917
918         if (overlap_clear)
919                 for (i = disks; i--; ) {
920                         struct r5dev *dev = &sh->dev[i];
921                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
922                                 wake_up(&sh->raid_conf->wait_for_overlap);
923                 }
924 }
925
926 static int grow_one_stripe(raid5_conf_t *conf)
927 {
928         struct stripe_head *sh;
929         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
930         if (!sh)
931                 return 0;
932         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
933         sh->raid_conf = conf;
934         spin_lock_init(&sh->lock);
935
936         if (grow_buffers(sh, conf->raid_disks)) {
937                 shrink_buffers(sh, conf->raid_disks);
938                 kmem_cache_free(conf->slab_cache, sh);
939                 return 0;
940         }
941         sh->disks = conf->raid_disks;
942         /* we just created an active stripe so... */
943         atomic_set(&sh->count, 1);
944         atomic_inc(&conf->active_stripes);
945         INIT_LIST_HEAD(&sh->lru);
946         release_stripe(sh);
947         return 1;
948 }
949
950 static int grow_stripes(raid5_conf_t *conf, int num)
951 {
952         struct kmem_cache *sc;
953         int devs = conf->raid_disks;
954
955         sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
956         sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
957         conf->active_name = 0;
958         sc = kmem_cache_create(conf->cache_name[conf->active_name],
959                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
960                                0, 0, NULL);
961         if (!sc)
962                 return 1;
963         conf->slab_cache = sc;
964         conf->pool_size = devs;
965         while (num--)
966                 if (!grow_one_stripe(conf))
967                         return 1;
968         return 0;
969 }
970
971 #ifdef CONFIG_MD_RAID5_RESHAPE
972 static int resize_stripes(raid5_conf_t *conf, int newsize)
973 {
974         /* Make all the stripes able to hold 'newsize' devices.
975          * New slots in each stripe get 'page' set to a new page.
976          *
977          * This happens in stages:
978          * 1/ create a new kmem_cache and allocate the required number of
979          *    stripe_heads.
980          * 2/ gather all the old stripe_heads and tranfer the pages across
981          *    to the new stripe_heads.  This will have the side effect of
982          *    freezing the array as once all stripe_heads have been collected,
983          *    no IO will be possible.  Old stripe heads are freed once their
984          *    pages have been transferred over, and the old kmem_cache is
985          *    freed when all stripes are done.
986          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
987          *    we simple return a failre status - no need to clean anything up.
988          * 4/ allocate new pages for the new slots in the new stripe_heads.
989          *    If this fails, we don't bother trying the shrink the
990          *    stripe_heads down again, we just leave them as they are.
991          *    As each stripe_head is processed the new one is released into
992          *    active service.
993          *
994          * Once step2 is started, we cannot afford to wait for a write,
995          * so we use GFP_NOIO allocations.
996          */
997         struct stripe_head *osh, *nsh;
998         LIST_HEAD(newstripes);
999         struct disk_info *ndisks;
1000         int err = 0;
1001         struct kmem_cache *sc;
1002         int i;
1003
1004         if (newsize <= conf->pool_size)
1005                 return 0; /* never bother to shrink */
1006
1007         md_allow_write(conf->mddev);
1008
1009         /* Step 1 */
1010         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1011                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1012                                0, 0, NULL);
1013         if (!sc)
1014                 return -ENOMEM;
1015
1016         for (i = conf->max_nr_stripes; i; i--) {
1017                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1018                 if (!nsh)
1019                         break;
1020
1021                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1022
1023                 nsh->raid_conf = conf;
1024                 spin_lock_init(&nsh->lock);
1025
1026                 list_add(&nsh->lru, &newstripes);
1027         }
1028         if (i) {
1029                 /* didn't get enough, give up */
1030                 while (!list_empty(&newstripes)) {
1031                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1032                         list_del(&nsh->lru);
1033                         kmem_cache_free(sc, nsh);
1034                 }
1035                 kmem_cache_destroy(sc);
1036                 return -ENOMEM;
1037         }
1038         /* Step 2 - Must use GFP_NOIO now.
1039          * OK, we have enough stripes, start collecting inactive
1040          * stripes and copying them over
1041          */
1042         list_for_each_entry(nsh, &newstripes, lru) {
1043                 spin_lock_irq(&conf->device_lock);
1044                 wait_event_lock_irq(conf->wait_for_stripe,
1045                                     !list_empty(&conf->inactive_list),
1046                                     conf->device_lock,
1047                                     unplug_slaves(conf->mddev)
1048                         );
1049                 osh = get_free_stripe(conf);
1050                 spin_unlock_irq(&conf->device_lock);
1051                 atomic_set(&nsh->count, 1);
1052                 for(i=0; i<conf->pool_size; i++)
1053                         nsh->dev[i].page = osh->dev[i].page;
1054                 for( ; i<newsize; i++)
1055                         nsh->dev[i].page = NULL;
1056                 kmem_cache_free(conf->slab_cache, osh);
1057         }
1058         kmem_cache_destroy(conf->slab_cache);
1059
1060         /* Step 3.
1061          * At this point, we are holding all the stripes so the array
1062          * is completely stalled, so now is a good time to resize
1063          * conf->disks.
1064          */
1065         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1066         if (ndisks) {
1067                 for (i=0; i<conf->raid_disks; i++)
1068                         ndisks[i] = conf->disks[i];
1069                 kfree(conf->disks);
1070                 conf->disks = ndisks;
1071         } else
1072                 err = -ENOMEM;
1073
1074         /* Step 4, return new stripes to service */
1075         while(!list_empty(&newstripes)) {
1076                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1077                 list_del_init(&nsh->lru);
1078                 for (i=conf->raid_disks; i < newsize; i++)
1079                         if (nsh->dev[i].page == NULL) {
1080                                 struct page *p = alloc_page(GFP_NOIO);
1081                                 nsh->dev[i].page = p;
1082                                 if (!p)
1083                                         err = -ENOMEM;
1084                         }
1085                 release_stripe(nsh);
1086         }
1087         /* critical section pass, GFP_NOIO no longer needed */
1088
1089         conf->slab_cache = sc;
1090         conf->active_name = 1-conf->active_name;
1091         conf->pool_size = newsize;
1092         return err;
1093 }
1094 #endif
1095
1096 static int drop_one_stripe(raid5_conf_t *conf)
1097 {
1098         struct stripe_head *sh;
1099
1100         spin_lock_irq(&conf->device_lock);
1101         sh = get_free_stripe(conf);
1102         spin_unlock_irq(&conf->device_lock);
1103         if (!sh)
1104                 return 0;
1105         BUG_ON(atomic_read(&sh->count));
1106         shrink_buffers(sh, conf->pool_size);
1107         kmem_cache_free(conf->slab_cache, sh);
1108         atomic_dec(&conf->active_stripes);
1109         return 1;
1110 }
1111
1112 static void shrink_stripes(raid5_conf_t *conf)
1113 {
1114         while (drop_one_stripe(conf))
1115                 ;
1116
1117         if (conf->slab_cache)
1118                 kmem_cache_destroy(conf->slab_cache);
1119         conf->slab_cache = NULL;
1120 }
1121
1122 static void raid5_end_read_request(struct bio * bi, int error)
1123 {
1124         struct stripe_head *sh = bi->bi_private;
1125         raid5_conf_t *conf = sh->raid_conf;
1126         int disks = sh->disks, i;
1127         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1128         char b[BDEVNAME_SIZE];
1129         mdk_rdev_t *rdev;
1130
1131
1132         for (i=0 ; i<disks; i++)
1133                 if (bi == &sh->dev[i].req)
1134                         break;
1135
1136         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1137                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1138                 uptodate);
1139         if (i == disks) {
1140                 BUG();
1141                 return;
1142         }
1143
1144         if (uptodate) {
1145                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1146                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1147                         rdev = conf->disks[i].rdev;
1148                         printk_rl(KERN_INFO "raid5:%s: read error corrected"
1149                                   " (%lu sectors at %llu on %s)\n",
1150                                   mdname(conf->mddev), STRIPE_SECTORS,
1151                                   (unsigned long long)(sh->sector
1152                                                        + rdev->data_offset),
1153                                   bdevname(rdev->bdev, b));
1154                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1155                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1156                 }
1157                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1158                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1159         } else {
1160                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1161                 int retry = 0;
1162                 rdev = conf->disks[i].rdev;
1163
1164                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1165                 atomic_inc(&rdev->read_errors);
1166                 if (conf->mddev->degraded)
1167                         printk_rl(KERN_WARNING
1168                                   "raid5:%s: read error not correctable "
1169                                   "(sector %llu on %s).\n",
1170                                   mdname(conf->mddev),
1171                                   (unsigned long long)(sh->sector
1172                                                        + rdev->data_offset),
1173                                   bdn);
1174                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1175                         /* Oh, no!!! */
1176                         printk_rl(KERN_WARNING
1177                                   "raid5:%s: read error NOT corrected!! "
1178                                   "(sector %llu on %s).\n",
1179                                   mdname(conf->mddev),
1180                                   (unsigned long long)(sh->sector
1181                                                        + rdev->data_offset),
1182                                   bdn);
1183                 else if (atomic_read(&rdev->read_errors)
1184                          > conf->max_nr_stripes)
1185                         printk(KERN_WARNING
1186                                "raid5:%s: Too many read errors, failing device %s.\n",
1187                                mdname(conf->mddev), bdn);
1188                 else
1189                         retry = 1;
1190                 if (retry)
1191                         set_bit(R5_ReadError, &sh->dev[i].flags);
1192                 else {
1193                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1194                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1195                         md_error(conf->mddev, rdev);
1196                 }
1197         }
1198         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1199         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1200         set_bit(STRIPE_HANDLE, &sh->state);
1201         release_stripe(sh);
1202 }
1203
1204 static void raid5_end_write_request (struct bio *bi, int error)
1205 {
1206         struct stripe_head *sh = bi->bi_private;
1207         raid5_conf_t *conf = sh->raid_conf;
1208         int disks = sh->disks, i;
1209         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1210
1211         for (i=0 ; i<disks; i++)
1212                 if (bi == &sh->dev[i].req)
1213                         break;
1214
1215         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1216                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1217                 uptodate);
1218         if (i == disks) {
1219                 BUG();
1220                 return;
1221         }
1222
1223         if (!uptodate)
1224                 md_error(conf->mddev, conf->disks[i].rdev);
1225
1226         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1227         
1228         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1229         set_bit(STRIPE_HANDLE, &sh->state);
1230         release_stripe(sh);
1231 }
1232
1233
1234 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1235         
1236 static void raid5_build_block (struct stripe_head *sh, int i)
1237 {
1238         struct r5dev *dev = &sh->dev[i];
1239
1240         bio_init(&dev->req);
1241         dev->req.bi_io_vec = &dev->vec;
1242         dev->req.bi_vcnt++;
1243         dev->req.bi_max_vecs++;
1244         dev->vec.bv_page = dev->page;
1245         dev->vec.bv_len = STRIPE_SIZE;
1246         dev->vec.bv_offset = 0;
1247
1248         dev->req.bi_sector = sh->sector;
1249         dev->req.bi_private = sh;
1250
1251         dev->flags = 0;
1252         dev->sector = compute_blocknr(sh, i);
1253 }
1254
1255 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1256 {
1257         char b[BDEVNAME_SIZE];
1258         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1259         pr_debug("raid5: error called\n");
1260
1261         if (!test_bit(Faulty, &rdev->flags)) {
1262                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1263                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1264                         unsigned long flags;
1265                         spin_lock_irqsave(&conf->device_lock, flags);
1266                         mddev->degraded++;
1267                         spin_unlock_irqrestore(&conf->device_lock, flags);
1268                         /*
1269                          * if recovery was running, make sure it aborts.
1270                          */
1271                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1272                 }
1273                 set_bit(Faulty, &rdev->flags);
1274                 printk (KERN_ALERT
1275                         "raid5: Disk failure on %s, disabling device.\n"
1276                         "raid5: Operation continuing on %d devices.\n",
1277                         bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1278         }
1279 }
1280
1281 /*
1282  * Input: a 'big' sector number,
1283  * Output: index of the data and parity disk, and the sector # in them.
1284  */
1285 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1286                         unsigned int data_disks, unsigned int * dd_idx,
1287                         unsigned int * pd_idx, raid5_conf_t *conf)
1288 {
1289         long stripe;
1290         unsigned long chunk_number;
1291         unsigned int chunk_offset;
1292         sector_t new_sector;
1293         int sectors_per_chunk = conf->chunk_size >> 9;
1294
1295         /* First compute the information on this sector */
1296
1297         /*
1298          * Compute the chunk number and the sector offset inside the chunk
1299          */
1300         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1301         chunk_number = r_sector;
1302         BUG_ON(r_sector != chunk_number);
1303
1304         /*
1305          * Compute the stripe number
1306          */
1307         stripe = chunk_number / data_disks;
1308
1309         /*
1310          * Compute the data disk and parity disk indexes inside the stripe
1311          */
1312         *dd_idx = chunk_number % data_disks;
1313
1314         /*
1315          * Select the parity disk based on the user selected algorithm.
1316          */
1317         switch(conf->level) {
1318         case 4:
1319                 *pd_idx = data_disks;
1320                 break;
1321         case 5:
1322                 switch (conf->algorithm) {
1323                 case ALGORITHM_LEFT_ASYMMETRIC:
1324                         *pd_idx = data_disks - stripe % raid_disks;
1325                         if (*dd_idx >= *pd_idx)
1326                                 (*dd_idx)++;
1327                         break;
1328                 case ALGORITHM_RIGHT_ASYMMETRIC:
1329                         *pd_idx = stripe % raid_disks;
1330                         if (*dd_idx >= *pd_idx)
1331                                 (*dd_idx)++;
1332                         break;
1333                 case ALGORITHM_LEFT_SYMMETRIC:
1334                         *pd_idx = data_disks - stripe % raid_disks;
1335                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1336                         break;
1337                 case ALGORITHM_RIGHT_SYMMETRIC:
1338                         *pd_idx = stripe % raid_disks;
1339                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1340                         break;
1341                 default:
1342                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1343                                 conf->algorithm);
1344                 }
1345                 break;
1346         case 6:
1347
1348                 /**** FIX THIS ****/
1349                 switch (conf->algorithm) {
1350                 case ALGORITHM_LEFT_ASYMMETRIC:
1351                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1352                         if (*pd_idx == raid_disks-1)
1353                                 (*dd_idx)++;    /* Q D D D P */
1354                         else if (*dd_idx >= *pd_idx)
1355                                 (*dd_idx) += 2; /* D D P Q D */
1356                         break;
1357                 case ALGORITHM_RIGHT_ASYMMETRIC:
1358                         *pd_idx = stripe % raid_disks;
1359                         if (*pd_idx == raid_disks-1)
1360                                 (*dd_idx)++;    /* Q D D D P */
1361                         else if (*dd_idx >= *pd_idx)
1362                                 (*dd_idx) += 2; /* D D P Q D */
1363                         break;
1364                 case ALGORITHM_LEFT_SYMMETRIC:
1365                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1366                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1367                         break;
1368                 case ALGORITHM_RIGHT_SYMMETRIC:
1369                         *pd_idx = stripe % raid_disks;
1370                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1371                         break;
1372                 default:
1373                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1374                                 conf->algorithm);
1375                 }
1376                 break;
1377         }
1378
1379         /*
1380          * Finally, compute the new sector number
1381          */
1382         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1383         return new_sector;
1384 }
1385
1386
1387 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1388 {
1389         raid5_conf_t *conf = sh->raid_conf;
1390         int raid_disks = sh->disks;
1391         int data_disks = raid_disks - conf->max_degraded;
1392         sector_t new_sector = sh->sector, check;
1393         int sectors_per_chunk = conf->chunk_size >> 9;
1394         sector_t stripe;
1395         int chunk_offset;
1396         int chunk_number, dummy1, dummy2, dd_idx = i;
1397         sector_t r_sector;
1398
1399
1400         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1401         stripe = new_sector;
1402         BUG_ON(new_sector != stripe);
1403
1404         if (i == sh->pd_idx)
1405                 return 0;
1406         switch(conf->level) {
1407         case 4: break;
1408         case 5:
1409                 switch (conf->algorithm) {
1410                 case ALGORITHM_LEFT_ASYMMETRIC:
1411                 case ALGORITHM_RIGHT_ASYMMETRIC:
1412                         if (i > sh->pd_idx)
1413                                 i--;
1414                         break;
1415                 case ALGORITHM_LEFT_SYMMETRIC:
1416                 case ALGORITHM_RIGHT_SYMMETRIC:
1417                         if (i < sh->pd_idx)
1418                                 i += raid_disks;
1419                         i -= (sh->pd_idx + 1);
1420                         break;
1421                 default:
1422                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1423                                conf->algorithm);
1424                 }
1425                 break;
1426         case 6:
1427                 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1428                         return 0; /* It is the Q disk */
1429                 switch (conf->algorithm) {
1430                 case ALGORITHM_LEFT_ASYMMETRIC:
1431                 case ALGORITHM_RIGHT_ASYMMETRIC:
1432                         if (sh->pd_idx == raid_disks-1)
1433                                 i--;    /* Q D D D P */
1434                         else if (i > sh->pd_idx)
1435                                 i -= 2; /* D D P Q D */
1436                         break;
1437                 case ALGORITHM_LEFT_SYMMETRIC:
1438                 case ALGORITHM_RIGHT_SYMMETRIC:
1439                         if (sh->pd_idx == raid_disks-1)
1440                                 i--; /* Q D D D P */
1441                         else {
1442                                 /* D D P Q D */
1443                                 if (i < sh->pd_idx)
1444                                         i += raid_disks;
1445                                 i -= (sh->pd_idx + 2);
1446                         }
1447                         break;
1448                 default:
1449                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1450                                 conf->algorithm);
1451                 }
1452                 break;
1453         }
1454
1455         chunk_number = stripe * data_disks + i;
1456         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1457
1458         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1459         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1460                 printk(KERN_ERR "compute_blocknr: map not correct\n");
1461                 return 0;
1462         }
1463         return r_sector;
1464 }
1465
1466
1467
1468 /*
1469  * Copy data between a page in the stripe cache, and one or more bion
1470  * The page could align with the middle of the bio, or there could be
1471  * several bion, each with several bio_vecs, which cover part of the page
1472  * Multiple bion are linked together on bi_next.  There may be extras
1473  * at the end of this list.  We ignore them.
1474  */
1475 static void copy_data(int frombio, struct bio *bio,
1476                      struct page *page,
1477                      sector_t sector)
1478 {
1479         char *pa = page_address(page);
1480         struct bio_vec *bvl;
1481         int i;
1482         int page_offset;
1483
1484         if (bio->bi_sector >= sector)
1485                 page_offset = (signed)(bio->bi_sector - sector) * 512;
1486         else
1487                 page_offset = (signed)(sector - bio->bi_sector) * -512;
1488         bio_for_each_segment(bvl, bio, i) {
1489                 int len = bio_iovec_idx(bio,i)->bv_len;
1490                 int clen;
1491                 int b_offset = 0;
1492
1493                 if (page_offset < 0) {
1494                         b_offset = -page_offset;
1495                         page_offset += b_offset;
1496                         len -= b_offset;
1497                 }
1498
1499                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1500                         clen = STRIPE_SIZE - page_offset;
1501                 else clen = len;
1502
1503                 if (clen > 0) {
1504                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1505                         if (frombio)
1506                                 memcpy(pa+page_offset, ba+b_offset, clen);
1507                         else
1508                                 memcpy(ba+b_offset, pa+page_offset, clen);
1509                         __bio_kunmap_atomic(ba, KM_USER0);
1510                 }
1511                 if (clen < len) /* hit end of page */
1512                         break;
1513                 page_offset +=  len;
1514         }
1515 }
1516
1517 #define check_xor()     do {                                              \
1518                                 if (count == MAX_XOR_BLOCKS) {            \
1519                                 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1520                                 count = 0;                                \
1521                            }                                              \
1522                         } while(0)
1523
1524 static void compute_parity6(struct stripe_head *sh, int method)
1525 {
1526         raid6_conf_t *conf = sh->raid_conf;
1527         int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1528         struct bio *chosen;
1529         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1530         void *ptrs[disks];
1531
1532         qd_idx = raid6_next_disk(pd_idx, disks);
1533         d0_idx = raid6_next_disk(qd_idx, disks);
1534
1535         pr_debug("compute_parity, stripe %llu, method %d\n",
1536                 (unsigned long long)sh->sector, method);
1537
1538         switch(method) {
1539         case READ_MODIFY_WRITE:
1540                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
1541         case RECONSTRUCT_WRITE:
1542                 for (i= disks; i-- ;)
1543                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1544                                 chosen = sh->dev[i].towrite;
1545                                 sh->dev[i].towrite = NULL;
1546
1547                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1548                                         wake_up(&conf->wait_for_overlap);
1549
1550                                 BUG_ON(sh->dev[i].written);
1551                                 sh->dev[i].written = chosen;
1552                         }
1553                 break;
1554         case CHECK_PARITY:
1555                 BUG();          /* Not implemented yet */
1556         }
1557
1558         for (i = disks; i--;)
1559                 if (sh->dev[i].written) {
1560                         sector_t sector = sh->dev[i].sector;
1561                         struct bio *wbi = sh->dev[i].written;
1562                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1563                                 copy_data(1, wbi, sh->dev[i].page, sector);
1564                                 wbi = r5_next_bio(wbi, sector);
1565                         }
1566
1567                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1568                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1569                 }
1570
1571 //      switch(method) {
1572 //      case RECONSTRUCT_WRITE:
1573 //      case CHECK_PARITY:
1574 //      case UPDATE_PARITY:
1575                 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1576                 /* FIX: Is this ordering of drives even remotely optimal? */
1577                 count = 0;
1578                 i = d0_idx;
1579                 do {
1580                         ptrs[count++] = page_address(sh->dev[i].page);
1581                         if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1582                                 printk("block %d/%d not uptodate on parity calc\n", i,count);
1583                         i = raid6_next_disk(i, disks);
1584                 } while ( i != d0_idx );
1585 //              break;
1586 //      }
1587
1588         raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1589
1590         switch(method) {
1591         case RECONSTRUCT_WRITE:
1592                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1593                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1594                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1595                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
1596                 break;
1597         case UPDATE_PARITY:
1598                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1599                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1600                 break;
1601         }
1602 }
1603
1604
1605 /* Compute one missing block */
1606 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1607 {
1608         int i, count, disks = sh->disks;
1609         void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1610         int pd_idx = sh->pd_idx;
1611         int qd_idx = raid6_next_disk(pd_idx, disks);
1612
1613         pr_debug("compute_block_1, stripe %llu, idx %d\n",
1614                 (unsigned long long)sh->sector, dd_idx);
1615
1616         if ( dd_idx == qd_idx ) {
1617                 /* We're actually computing the Q drive */
1618                 compute_parity6(sh, UPDATE_PARITY);
1619         } else {
1620                 dest = page_address(sh->dev[dd_idx].page);
1621                 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1622                 count = 0;
1623                 for (i = disks ; i--; ) {
1624                         if (i == dd_idx || i == qd_idx)
1625                                 continue;
1626                         p = page_address(sh->dev[i].page);
1627                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1628                                 ptr[count++] = p;
1629                         else
1630                                 printk("compute_block() %d, stripe %llu, %d"
1631                                        " not present\n", dd_idx,
1632                                        (unsigned long long)sh->sector, i);
1633
1634                         check_xor();
1635                 }
1636                 if (count)
1637                         xor_blocks(count, STRIPE_SIZE, dest, ptr);
1638                 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1639                 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1640         }
1641 }
1642
1643 /* Compute two missing blocks */
1644 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1645 {
1646         int i, count, disks = sh->disks;
1647         int pd_idx = sh->pd_idx;
1648         int qd_idx = raid6_next_disk(pd_idx, disks);
1649         int d0_idx = raid6_next_disk(qd_idx, disks);
1650         int faila, failb;
1651
1652         /* faila and failb are disk numbers relative to d0_idx */
1653         /* pd_idx become disks-2 and qd_idx become disks-1 */
1654         faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1655         failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1656
1657         BUG_ON(faila == failb);
1658         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1659
1660         pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1661                (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1662
1663         if ( failb == disks-1 ) {
1664                 /* Q disk is one of the missing disks */
1665                 if ( faila == disks-2 ) {
1666                         /* Missing P+Q, just recompute */
1667                         compute_parity6(sh, UPDATE_PARITY);
1668                         return;
1669                 } else {
1670                         /* We're missing D+Q; recompute D from P */
1671                         compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1672                         compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1673                         return;
1674                 }
1675         }
1676
1677         /* We're missing D+P or D+D; build pointer table */
1678         {
1679                 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1680                 void *ptrs[disks];
1681
1682                 count = 0;
1683                 i = d0_idx;
1684                 do {
1685                         ptrs[count++] = page_address(sh->dev[i].page);
1686                         i = raid6_next_disk(i, disks);
1687                         if (i != dd_idx1 && i != dd_idx2 &&
1688                             !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1689                                 printk("compute_2 with missing block %d/%d\n", count, i);
1690                 } while ( i != d0_idx );
1691
1692                 if ( failb == disks-2 ) {
1693                         /* We're missing D+P. */
1694                         raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1695                 } else {
1696                         /* We're missing D+D. */
1697                         raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1698                 }
1699
1700                 /* Both the above update both missing blocks */
1701                 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1702                 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1703         }
1704 }
1705
1706 static int
1707 handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
1708 {
1709         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1710         int locked = 0;
1711
1712         if (rcw) {
1713                 /* if we are not expanding this is a proper write request, and
1714                  * there will be bios with new data to be drained into the
1715                  * stripe cache
1716                  */
1717                 if (!expand) {
1718                         set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1719                         sh->ops.count++;
1720                 }
1721
1722                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1723                 sh->ops.count++;
1724
1725                 for (i = disks; i--; ) {
1726                         struct r5dev *dev = &sh->dev[i];
1727
1728                         if (dev->towrite) {
1729                                 set_bit(R5_LOCKED, &dev->flags);
1730                                 if (!expand)
1731                                         clear_bit(R5_UPTODATE, &dev->flags);
1732                                 locked++;
1733                         }
1734                 }
1735                 if (locked + 1 == disks)
1736                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1737                                 atomic_inc(&sh->raid_conf->pending_full_writes);
1738         } else {
1739                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1740                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1741
1742                 set_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
1743                 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1744                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1745
1746                 sh->ops.count += 3;
1747
1748                 for (i = disks; i--; ) {
1749                         struct r5dev *dev = &sh->dev[i];
1750                         if (i == pd_idx)
1751                                 continue;
1752
1753                         /* For a read-modify write there may be blocks that are
1754                          * locked for reading while others are ready to be
1755                          * written so we distinguish these blocks by the
1756                          * R5_Wantprexor bit
1757                          */
1758                         if (dev->towrite &&
1759                             (test_bit(R5_UPTODATE, &dev->flags) ||
1760                             test_bit(R5_Wantcompute, &dev->flags))) {
1761                                 set_bit(R5_Wantprexor, &dev->flags);
1762                                 set_bit(R5_LOCKED, &dev->flags);
1763                                 clear_bit(R5_UPTODATE, &dev->flags);
1764                                 locked++;
1765                         }
1766                 }
1767         }
1768
1769         /* keep the parity disk locked while asynchronous operations
1770          * are in flight
1771          */
1772         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1773         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1774         locked++;
1775
1776         pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1777                 __func__, (unsigned long long)sh->sector,
1778                 locked, sh->ops.pending);
1779
1780         return locked;
1781 }
1782
1783 /*
1784  * Each stripe/dev can have one or more bion attached.
1785  * toread/towrite point to the first in a chain.
1786  * The bi_next chain must be in order.
1787  */
1788 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1789 {
1790         struct bio **bip;
1791         raid5_conf_t *conf = sh->raid_conf;
1792         int firstwrite=0;
1793
1794         pr_debug("adding bh b#%llu to stripe s#%llu\n",
1795                 (unsigned long long)bi->bi_sector,
1796                 (unsigned long long)sh->sector);
1797
1798
1799         spin_lock(&sh->lock);
1800         spin_lock_irq(&conf->device_lock);
1801         if (forwrite) {
1802                 bip = &sh->dev[dd_idx].towrite;
1803                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1804                         firstwrite = 1;
1805         } else
1806                 bip = &sh->dev[dd_idx].toread;
1807         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1808                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1809                         goto overlap;
1810                 bip = & (*bip)->bi_next;
1811         }
1812         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1813                 goto overlap;
1814
1815         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1816         if (*bip)
1817                 bi->bi_next = *bip;
1818         *bip = bi;
1819         bi->bi_phys_segments ++;
1820         spin_unlock_irq(&conf->device_lock);
1821         spin_unlock(&sh->lock);
1822
1823         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1824                 (unsigned long long)bi->bi_sector,
1825                 (unsigned long long)sh->sector, dd_idx);
1826
1827         if (conf->mddev->bitmap && firstwrite) {
1828                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1829                                   STRIPE_SECTORS, 0);
1830                 sh->bm_seq = conf->seq_flush+1;
1831                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1832         }
1833
1834         if (forwrite) {
1835                 /* check if page is covered */
1836                 sector_t sector = sh->dev[dd_idx].sector;
1837                 for (bi=sh->dev[dd_idx].towrite;
1838                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1839                              bi && bi->bi_sector <= sector;
1840                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1841                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1842                                 sector = bi->bi_sector + (bi->bi_size>>9);
1843                 }
1844                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1845                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1846         }
1847         return 1;
1848
1849  overlap:
1850         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1851         spin_unlock_irq(&conf->device_lock);
1852         spin_unlock(&sh->lock);
1853         return 0;
1854 }
1855
1856 static void end_reshape(raid5_conf_t *conf);
1857
1858 static int page_is_zero(struct page *p)
1859 {
1860         char *a = page_address(p);
1861         return ((*(u32*)a) == 0 &&
1862                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1863 }
1864
1865 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1866 {
1867         int sectors_per_chunk = conf->chunk_size >> 9;
1868         int pd_idx, dd_idx;
1869         int chunk_offset = sector_div(stripe, sectors_per_chunk);
1870
1871         raid5_compute_sector(stripe * (disks - conf->max_degraded)
1872                              *sectors_per_chunk + chunk_offset,
1873                              disks, disks - conf->max_degraded,
1874                              &dd_idx, &pd_idx, conf);
1875         return pd_idx;
1876 }
1877
1878 static void
1879 handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1880                                 struct stripe_head_state *s, int disks,
1881                                 struct bio **return_bi)
1882 {
1883         int i;
1884         for (i = disks; i--; ) {
1885                 struct bio *bi;
1886                 int bitmap_end = 0;
1887
1888                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1889                         mdk_rdev_t *rdev;
1890                         rcu_read_lock();
1891                         rdev = rcu_dereference(conf->disks[i].rdev);
1892                         if (rdev && test_bit(In_sync, &rdev->flags))
1893                                 /* multiple read failures in one stripe */
1894                                 md_error(conf->mddev, rdev);
1895                         rcu_read_unlock();
1896                 }
1897                 spin_lock_irq(&conf->device_lock);
1898                 /* fail all writes first */
1899                 bi = sh->dev[i].towrite;
1900                 sh->dev[i].towrite = NULL;
1901                 if (bi) {
1902                         s->to_write--;
1903                         bitmap_end = 1;
1904                 }
1905
1906                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1907                         wake_up(&conf->wait_for_overlap);
1908
1909                 while (bi && bi->bi_sector <
1910                         sh->dev[i].sector + STRIPE_SECTORS) {
1911                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1912                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1913                         if (--bi->bi_phys_segments == 0) {
1914                                 md_write_end(conf->mddev);
1915                                 bi->bi_next = *return_bi;
1916                                 *return_bi = bi;
1917                         }
1918                         bi = nextbi;
1919                 }
1920                 /* and fail all 'written' */
1921                 bi = sh->dev[i].written;
1922                 sh->dev[i].written = NULL;
1923                 if (bi) bitmap_end = 1;
1924                 while (bi && bi->bi_sector <
1925                        sh->dev[i].sector + STRIPE_SECTORS) {
1926                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1927                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1928                         if (--bi->bi_phys_segments == 0) {
1929                                 md_write_end(conf->mddev);
1930                                 bi->bi_next = *return_bi;
1931                                 *return_bi = bi;
1932                         }
1933                         bi = bi2;
1934                 }
1935
1936                 /* fail any reads if this device is non-operational and
1937                  * the data has not reached the cache yet.
1938                  */
1939                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
1940                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1941                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
1942                         bi = sh->dev[i].toread;
1943                         sh->dev[i].toread = NULL;
1944                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1945                                 wake_up(&conf->wait_for_overlap);
1946                         if (bi) s->to_read--;
1947                         while (bi && bi->bi_sector <
1948                                sh->dev[i].sector + STRIPE_SECTORS) {
1949                                 struct bio *nextbi =
1950                                         r5_next_bio(bi, sh->dev[i].sector);
1951                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1952                                 if (--bi->bi_phys_segments == 0) {
1953                                         bi->bi_next = *return_bi;
1954                                         *return_bi = bi;
1955                                 }
1956                                 bi = nextbi;
1957                         }
1958                 }
1959                 spin_unlock_irq(&conf->device_lock);
1960                 if (bitmap_end)
1961                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1962                                         STRIPE_SECTORS, 0, 0);
1963         }
1964
1965         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
1966                 if (atomic_dec_and_test(&conf->pending_full_writes))
1967                         md_wakeup_thread(conf->mddev->thread);
1968 }
1969
1970 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
1971  * to process
1972  */
1973 static int __handle_issuing_new_read_requests5(struct stripe_head *sh,
1974                         struct stripe_head_state *s, int disk_idx, int disks)
1975 {
1976         struct r5dev *dev = &sh->dev[disk_idx];
1977         struct r5dev *failed_dev = &sh->dev[s->failed_num];
1978
1979         /* don't schedule compute operations or reads on the parity block while
1980          * a check is in flight
1981          */
1982         if ((disk_idx == sh->pd_idx) &&
1983              test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
1984                 return ~0;
1985
1986         /* is the data in this block needed, and can we get it? */
1987         if (!test_bit(R5_LOCKED, &dev->flags) &&
1988             !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread ||
1989             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1990              s->syncing || s->expanding || (s->failed &&
1991              (failed_dev->toread || (failed_dev->towrite &&
1992              !test_bit(R5_OVERWRITE, &failed_dev->flags)
1993              ))))) {
1994                 /* 1/ We would like to get this block, possibly by computing it,
1995                  * but we might not be able to.
1996                  *
1997                  * 2/ Since parity check operations potentially make the parity
1998                  * block !uptodate it will need to be refreshed before any
1999                  * compute operations on data disks are scheduled.
2000                  *
2001                  * 3/ We hold off parity block re-reads until check operations
2002                  * have quiesced.
2003                  */
2004                 if ((s->uptodate == disks - 1) &&
2005                     (s->failed && disk_idx == s->failed_num) &&
2006                     !test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2007                         set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2008                         set_bit(R5_Wantcompute, &dev->flags);
2009                         sh->ops.target = disk_idx;
2010                         s->req_compute = 1;
2011                         sh->ops.count++;
2012                         /* Careful: from this point on 'uptodate' is in the eye
2013                          * of raid5_run_ops which services 'compute' operations
2014                          * before writes. R5_Wantcompute flags a block that will
2015                          * be R5_UPTODATE by the time it is needed for a
2016                          * subsequent operation.
2017                          */
2018                         s->uptodate++;
2019                         return 0; /* uptodate + compute == disks */
2020                 } else if (test_bit(R5_Insync, &dev->flags)) {
2021                         set_bit(R5_LOCKED, &dev->flags);
2022                         set_bit(R5_Wantread, &dev->flags);
2023                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2024                                 sh->ops.count++;
2025                         s->locked++;
2026                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2027                                 s->syncing);
2028                 }
2029         }
2030
2031         return ~0;
2032 }
2033
2034 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2035                         struct stripe_head_state *s, int disks)
2036 {
2037         int i;
2038
2039         /* Clear completed compute operations.  Parity recovery
2040          * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled
2041          * later on in this routine
2042          */
2043         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2044                 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2045                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2046                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2047                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2048         }
2049
2050         /* look for blocks to read/compute, skip this if a compute
2051          * is already in flight, or if the stripe contents are in the
2052          * midst of changing due to a write
2053          */
2054         if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2055                 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2056                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2057                 for (i = disks; i--; )
2058                         if (__handle_issuing_new_read_requests5(
2059                                 sh, s, i, disks) == 0)
2060                                 break;
2061         }
2062         set_bit(STRIPE_HANDLE, &sh->state);
2063 }
2064
2065 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2066                         struct stripe_head_state *s, struct r6_state *r6s,
2067                         int disks)
2068 {
2069         int i;
2070         for (i = disks; i--; ) {
2071                 struct r5dev *dev = &sh->dev[i];
2072                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2073                     !test_bit(R5_UPTODATE, &dev->flags) &&
2074                     (dev->toread || (dev->towrite &&
2075                      !test_bit(R5_OVERWRITE, &dev->flags)) ||
2076                      s->syncing || s->expanding ||
2077                      (s->failed >= 1 &&
2078                       (sh->dev[r6s->failed_num[0]].toread ||
2079                        s->to_write)) ||
2080                      (s->failed >= 2 &&
2081                       (sh->dev[r6s->failed_num[1]].toread ||
2082                        s->to_write)))) {
2083                         /* we would like to get this block, possibly
2084                          * by computing it, but we might not be able to
2085                          */
2086                         if ((s->uptodate == disks - 1) &&
2087                             (s->failed && (i == r6s->failed_num[0] ||
2088                                            i == r6s->failed_num[1]))) {
2089                                 pr_debug("Computing stripe %llu block %d\n",
2090                                        (unsigned long long)sh->sector, i);
2091                                 compute_block_1(sh, i, 0);
2092                                 s->uptodate++;
2093                         } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2094                                 /* Computing 2-failure is *very* expensive; only
2095                                  * do it if failed >= 2
2096                                  */
2097                                 int other;
2098                                 for (other = disks; other--; ) {
2099                                         if (other == i)
2100                                                 continue;
2101                                         if (!test_bit(R5_UPTODATE,
2102                                               &sh->dev[other].flags))
2103                                                 break;
2104                                 }
2105                                 BUG_ON(other < 0);
2106                                 pr_debug("Computing stripe %llu blocks %d,%d\n",
2107                                        (unsigned long long)sh->sector,
2108                                        i, other);
2109                                 compute_block_2(sh, i, other);
2110                                 s->uptodate += 2;
2111                         } else if (test_bit(R5_Insync, &dev->flags)) {
2112                                 set_bit(R5_LOCKED, &dev->flags);
2113                                 set_bit(R5_Wantread, &dev->flags);
2114                                 s->locked++;
2115                                 pr_debug("Reading block %d (sync=%d)\n",
2116                                         i, s->syncing);
2117                         }
2118                 }
2119         }
2120         set_bit(STRIPE_HANDLE, &sh->state);
2121 }
2122
2123
2124 /* handle_completed_write_requests
2125  * any written block on an uptodate or failed drive can be returned.
2126  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2127  * never LOCKED, so we don't need to test 'failed' directly.
2128  */
2129 static void handle_completed_write_requests(raid5_conf_t *conf,
2130         struct stripe_head *sh, int disks, struct bio **return_bi)
2131 {
2132         int i;
2133         struct r5dev *dev;
2134
2135         for (i = disks; i--; )
2136                 if (sh->dev[i].written) {
2137                         dev = &sh->dev[i];
2138                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2139                                 test_bit(R5_UPTODATE, &dev->flags)) {
2140                                 /* We can return any write requests */
2141                                 struct bio *wbi, *wbi2;
2142                                 int bitmap_end = 0;
2143                                 pr_debug("Return write for disc %d\n", i);
2144                                 spin_lock_irq(&conf->device_lock);
2145                                 wbi = dev->written;
2146                                 dev->written = NULL;
2147                                 while (wbi && wbi->bi_sector <
2148                                         dev->sector + STRIPE_SECTORS) {
2149                                         wbi2 = r5_next_bio(wbi, dev->sector);
2150                                         if (--wbi->bi_phys_segments == 0) {
2151                                                 md_write_end(conf->mddev);
2152                                                 wbi->bi_next = *return_bi;
2153                                                 *return_bi = wbi;
2154                                         }
2155                                         wbi = wbi2;
2156                                 }
2157                                 if (dev->towrite == NULL)
2158                                         bitmap_end = 1;
2159                                 spin_unlock_irq(&conf->device_lock);
2160                                 if (bitmap_end)
2161                                         bitmap_endwrite(conf->mddev->bitmap,
2162                                                         sh->sector,
2163                                                         STRIPE_SECTORS,
2164                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2165                                                         0);
2166                         }
2167                 }
2168
2169         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2170                 if (atomic_dec_and_test(&conf->pending_full_writes))
2171                         md_wakeup_thread(conf->mddev->thread);
2172 }
2173
2174 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2175                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2176 {
2177         int rmw = 0, rcw = 0, i;
2178         for (i = disks; i--; ) {
2179                 /* would I have to read this buffer for read_modify_write */
2180                 struct r5dev *dev = &sh->dev[i];
2181                 if ((dev->towrite || i == sh->pd_idx) &&
2182                     !test_bit(R5_LOCKED, &dev->flags) &&
2183                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2184                       test_bit(R5_Wantcompute, &dev->flags))) {
2185                         if (test_bit(R5_Insync, &dev->flags))
2186                                 rmw++;
2187                         else
2188                                 rmw += 2*disks;  /* cannot read it */
2189                 }
2190                 /* Would I have to read this buffer for reconstruct_write */
2191                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2192                     !test_bit(R5_LOCKED, &dev->flags) &&
2193                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2194                     test_bit(R5_Wantcompute, &dev->flags))) {
2195                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2196                         else
2197                                 rcw += 2*disks;
2198                 }
2199         }
2200         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2201                 (unsigned long long)sh->sector, rmw, rcw);
2202         set_bit(STRIPE_HANDLE, &sh->state);
2203         if (rmw < rcw && rmw > 0)
2204                 /* prefer read-modify-write, but need to get some data */
2205                 for (i = disks; i--; ) {
2206                         struct r5dev *dev = &sh->dev[i];
2207                         if ((dev->towrite || i == sh->pd_idx) &&
2208                             !test_bit(R5_LOCKED, &dev->flags) &&
2209                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2210                             test_bit(R5_Wantcompute, &dev->flags)) &&
2211                             test_bit(R5_Insync, &dev->flags)) {
2212                                 if (
2213                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2214                                         pr_debug("Read_old block "
2215                                                 "%d for r-m-w\n", i);
2216                                         set_bit(R5_LOCKED, &dev->flags);
2217                                         set_bit(R5_Wantread, &dev->flags);
2218                                         if (!test_and_set_bit(
2219                                                 STRIPE_OP_IO, &sh->ops.pending))
2220                                                 sh->ops.count++;
2221                                         s->locked++;
2222                                 } else {
2223                                         set_bit(STRIPE_DELAYED, &sh->state);
2224                                         set_bit(STRIPE_HANDLE, &sh->state);
2225                                 }
2226                         }
2227                 }
2228         if (rcw <= rmw && rcw > 0)
2229                 /* want reconstruct write, but need to get some data */
2230                 for (i = disks; i--; ) {
2231                         struct r5dev *dev = &sh->dev[i];
2232                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2233                             i != sh->pd_idx &&
2234                             !test_bit(R5_LOCKED, &dev->flags) &&
2235                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2236                             test_bit(R5_Wantcompute, &dev->flags)) &&
2237                             test_bit(R5_Insync, &dev->flags)) {
2238                                 if (
2239                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2240                                         pr_debug("Read_old block "
2241                                                 "%d for Reconstruct\n", i);
2242                                         set_bit(R5_LOCKED, &dev->flags);
2243                                         set_bit(R5_Wantread, &dev->flags);
2244                                         if (!test_and_set_bit(
2245                                                 STRIPE_OP_IO, &sh->ops.pending))
2246                                                 sh->ops.count++;
2247                                         s->locked++;
2248                                 } else {
2249                                         set_bit(STRIPE_DELAYED, &sh->state);
2250                                         set_bit(STRIPE_HANDLE, &sh->state);
2251                                 }
2252                         }
2253                 }
2254         /* now if nothing is locked, and if we have enough data,
2255          * we can start a write request
2256          */
2257         /* since handle_stripe can be called at any time we need to handle the
2258          * case where a compute block operation has been submitted and then a
2259          * subsequent call wants to start a write request.  raid5_run_ops only
2260          * handles the case where compute block and postxor are requested
2261          * simultaneously.  If this is not the case then new writes need to be
2262          * held off until the compute completes.
2263          */
2264         if ((s->req_compute ||
2265             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) &&
2266                 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2267                 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2268                 s->locked += handle_write_operations5(sh, rcw == 0, 0);
2269 }
2270
2271 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2272                 struct stripe_head *sh, struct stripe_head_state *s,
2273                 struct r6_state *r6s, int disks)
2274 {
2275         int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2276         int qd_idx = r6s->qd_idx;
2277         for (i = disks; i--; ) {
2278                 struct r5dev *dev = &sh->dev[i];
2279                 /* Would I have to read this buffer for reconstruct_write */
2280                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2281                     && i != pd_idx && i != qd_idx
2282                     && (!test_bit(R5_LOCKED, &dev->flags)
2283                             ) &&
2284                     !test_bit(R5_UPTODATE, &dev->flags)) {
2285                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2286                         else {
2287                                 pr_debug("raid6: must_compute: "
2288                                         "disk %d flags=%#lx\n", i, dev->flags);
2289                                 must_compute++;
2290                         }
2291                 }
2292         }
2293         pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2294                (unsigned long long)sh->sector, rcw, must_compute);
2295         set_bit(STRIPE_HANDLE, &sh->state);
2296
2297         if (rcw > 0)
2298                 /* want reconstruct write, but need to get some data */
2299                 for (i = disks; i--; ) {
2300                         struct r5dev *dev = &sh->dev[i];
2301                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2302                             && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2303                             && !test_bit(R5_LOCKED, &dev->flags) &&
2304                             !test_bit(R5_UPTODATE, &dev->flags) &&
2305                             test_bit(R5_Insync, &dev->flags)) {
2306                                 if (
2307                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2308                                         pr_debug("Read_old stripe %llu "
2309                                                 "block %d for Reconstruct\n",
2310                                              (unsigned long long)sh->sector, i);
2311                                         set_bit(R5_LOCKED, &dev->flags);
2312                                         set_bit(R5_Wantread, &dev->flags);
2313                                         s->locked++;
2314                                 } else {
2315                                         pr_debug("Request delayed stripe %llu "
2316                                                 "block %d for Reconstruct\n",
2317                                              (unsigned long long)sh->sector, i);
2318                                         set_bit(STRIPE_DELAYED, &sh->state);
2319                                         set_bit(STRIPE_HANDLE, &sh->state);
2320                                 }
2321                         }
2322                 }
2323         /* now if nothing is locked, and if we have enough data, we can start a
2324          * write request
2325          */
2326         if (s->locked == 0 && rcw == 0 &&
2327             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2328                 if (must_compute > 0) {
2329                         /* We have failed blocks and need to compute them */
2330                         switch (s->failed) {
2331                         case 0:
2332                                 BUG();
2333                         case 1:
2334                                 compute_block_1(sh, r6s->failed_num[0], 0);
2335                                 break;
2336                         case 2:
2337                                 compute_block_2(sh, r6s->failed_num[0],
2338                                                 r6s->failed_num[1]);
2339                                 break;
2340                         default: /* This request should have been failed? */
2341                                 BUG();
2342                         }
2343                 }
2344
2345                 pr_debug("Computing parity for stripe %llu\n",
2346                         (unsigned long long)sh->sector);
2347                 compute_parity6(sh, RECONSTRUCT_WRITE);
2348                 /* now every locked buffer is ready to be written */
2349                 for (i = disks; i--; )
2350                         if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2351                                 pr_debug("Writing stripe %llu block %d\n",
2352                                        (unsigned long long)sh->sector, i);
2353                                 s->locked++;
2354                                 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2355                         }
2356                 if (s->locked == disks)
2357                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2358                                 atomic_inc(&conf->pending_full_writes);
2359                 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2360                 set_bit(STRIPE_INSYNC, &sh->state);
2361
2362                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2363                         atomic_dec(&conf->preread_active_stripes);
2364                         if (atomic_read(&conf->preread_active_stripes) <
2365                             IO_THRESHOLD)
2366                                 md_wakeup_thread(conf->mddev->thread);
2367                 }
2368         }
2369 }
2370
2371 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2372                                 struct stripe_head_state *s, int disks)
2373 {
2374         int canceled_check = 0;
2375
2376         set_bit(STRIPE_HANDLE, &sh->state);
2377
2378         /* complete a check operation */
2379         if (test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) {
2380                 clear_bit(STRIPE_OP_CHECK, &sh->ops.ack);
2381                 clear_bit(STRIPE_OP_CHECK, &sh->ops.pending);
2382                 if (s->failed == 0) {
2383                         if (sh->ops.zero_sum_result == 0)
2384                                 /* parity is correct (on disc,
2385                                  * not in buffer any more)
2386                                  */
2387                                 set_bit(STRIPE_INSYNC, &sh->state);
2388                         else {
2389                                 conf->mddev->resync_mismatches +=
2390                                         STRIPE_SECTORS;
2391                                 if (test_bit(
2392                                      MD_RECOVERY_CHECK, &conf->mddev->recovery))
2393                                         /* don't try to repair!! */
2394                                         set_bit(STRIPE_INSYNC, &sh->state);
2395                                 else {
2396                                         set_bit(STRIPE_OP_COMPUTE_BLK,
2397                                                 &sh->ops.pending);
2398                                         set_bit(STRIPE_OP_MOD_REPAIR_PD,
2399                                                 &sh->ops.pending);
2400                                         set_bit(R5_Wantcompute,
2401                                                 &sh->dev[sh->pd_idx].flags);
2402                                         sh->ops.target = sh->pd_idx;
2403                                         sh->ops.count++;
2404                                         s->uptodate++;
2405                                 }
2406                         }
2407                 } else
2408                         canceled_check = 1; /* STRIPE_INSYNC is not set */
2409         }
2410
2411         /* start a new check operation if there are no failures, the stripe is
2412          * not insync, and a repair is not in flight
2413          */
2414         if (s->failed == 0 &&
2415             !test_bit(STRIPE_INSYNC, &sh->state) &&
2416             !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2417                 if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2418                         BUG_ON(s->uptodate != disks);
2419                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2420                         sh->ops.count++;
2421                         s->uptodate--;
2422                 }
2423         }
2424
2425         /* check if we can clear a parity disk reconstruct */
2426         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2427             test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2428
2429                 clear_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending);
2430                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2431                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2432                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2433         }
2434
2435
2436         /* Wait for check parity and compute block operations to complete
2437          * before write-back.  If a failure occurred while the check operation
2438          * was in flight we need to cycle this stripe through handle_stripe
2439          * since the parity block may not be uptodate
2440          */
2441         if (!canceled_check && !test_bit(STRIPE_INSYNC, &sh->state) &&
2442             !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) &&
2443             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) {
2444                 struct r5dev *dev;
2445                 /* either failed parity check, or recovery is happening */
2446                 if (s->failed == 0)
2447                         s->failed_num = sh->pd_idx;
2448                 dev = &sh->dev[s->failed_num];
2449                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2450                 BUG_ON(s->uptodate != disks);
2451
2452                 set_bit(R5_LOCKED, &dev->flags);
2453                 set_bit(R5_Wantwrite, &dev->flags);
2454                 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2455                         sh->ops.count++;
2456
2457                 clear_bit(STRIPE_DEGRADED, &sh->state);
2458                 s->locked++;
2459                 set_bit(STRIPE_INSYNC, &sh->state);
2460         }
2461 }
2462
2463
2464 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2465                                 struct stripe_head_state *s,
2466                                 struct r6_state *r6s, struct page *tmp_page,
2467                                 int disks)
2468 {
2469         int update_p = 0, update_q = 0;
2470         struct r5dev *dev;
2471         int pd_idx = sh->pd_idx;
2472         int qd_idx = r6s->qd_idx;
2473
2474         set_bit(STRIPE_HANDLE, &sh->state);
2475
2476         BUG_ON(s->failed > 2);
2477         BUG_ON(s->uptodate < disks);
2478         /* Want to check and possibly repair P and Q.
2479          * However there could be one 'failed' device, in which
2480          * case we can only check one of them, possibly using the
2481          * other to generate missing data
2482          */
2483
2484         /* If !tmp_page, we cannot do the calculations,
2485          * but as we have set STRIPE_HANDLE, we will soon be called
2486          * by stripe_handle with a tmp_page - just wait until then.
2487          */
2488         if (tmp_page) {
2489                 if (s->failed == r6s->q_failed) {
2490                         /* The only possible failed device holds 'Q', so it
2491                          * makes sense to check P (If anything else were failed,
2492                          * we would have used P to recreate it).
2493                          */
2494                         compute_block_1(sh, pd_idx, 1);
2495                         if (!page_is_zero(sh->dev[pd_idx].page)) {
2496                                 compute_block_1(sh, pd_idx, 0);
2497                                 update_p = 1;
2498                         }
2499                 }
2500                 if (!r6s->q_failed && s->failed < 2) {
2501                         /* q is not failed, and we didn't use it to generate
2502                          * anything, so it makes sense to check it
2503                          */
2504                         memcpy(page_address(tmp_page),
2505                                page_address(sh->dev[qd_idx].page),
2506                                STRIPE_SIZE);
2507                         compute_parity6(sh, UPDATE_PARITY);
2508                         if (memcmp(page_address(tmp_page),
2509                                    page_address(sh->dev[qd_idx].page),
2510                                    STRIPE_SIZE) != 0) {
2511                                 clear_bit(STRIPE_INSYNC, &sh->state);
2512                                 update_q = 1;
2513                         }
2514                 }
2515                 if (update_p || update_q) {
2516                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2517                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2518                                 /* don't try to repair!! */
2519                                 update_p = update_q = 0;
2520                 }
2521
2522                 /* now write out any block on a failed drive,
2523                  * or P or Q if they need it
2524                  */
2525
2526                 if (s->failed == 2) {
2527                         dev = &sh->dev[r6s->failed_num[1]];
2528                         s->locked++;
2529                         set_bit(R5_LOCKED, &dev->flags);
2530                         set_bit(R5_Wantwrite, &dev->flags);
2531                 }
2532                 if (s->failed >= 1) {
2533                         dev = &sh->dev[r6s->failed_num[0]];
2534                         s->locked++;
2535                         set_bit(R5_LOCKED, &dev->flags);
2536                         set_bit(R5_Wantwrite, &dev->flags);
2537                 }
2538
2539                 if (update_p) {
2540                         dev = &sh->dev[pd_idx];
2541                         s->locked++;
2542                         set_bit(R5_LOCKED, &dev->flags);
2543                         set_bit(R5_Wantwrite, &dev->flags);
2544                 }
2545                 if (update_q) {
2546                         dev = &sh->dev[qd_idx];
2547                         s->locked++;
2548                         set_bit(R5_LOCKED, &dev->flags);
2549                         set_bit(R5_Wantwrite, &dev->flags);
2550                 }
2551                 clear_bit(STRIPE_DEGRADED, &sh->state);
2552
2553                 set_bit(STRIPE_INSYNC, &sh->state);
2554         }
2555 }
2556
2557 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2558                                 struct r6_state *r6s)
2559 {
2560         int i;
2561
2562         /* We have read all the blocks in this stripe and now we need to
2563          * copy some of them into a target stripe for expand.
2564          */
2565         struct dma_async_tx_descriptor *tx = NULL;
2566         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2567         for (i = 0; i < sh->disks; i++)
2568                 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2569                         int dd_idx, pd_idx, j;
2570                         struct stripe_head *sh2;
2571
2572                         sector_t bn = compute_blocknr(sh, i);
2573                         sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2574                                                 conf->raid_disks -
2575                                                 conf->max_degraded, &dd_idx,
2576                                                 &pd_idx, conf);
2577                         sh2 = get_active_stripe(conf, s, conf->raid_disks,
2578                                                 pd_idx, 1);
2579                         if (sh2 == NULL)
2580                                 /* so far only the early blocks of this stripe
2581                                  * have been requested.  When later blocks
2582                                  * get requested, we will try again
2583                                  */
2584                                 continue;
2585                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2586                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2587                                 /* must have already done this block */
2588                                 release_stripe(sh2);
2589                                 continue;
2590                         }
2591
2592                         /* place all the copies on one channel */
2593                         tx = async_memcpy(sh2->dev[dd_idx].page,
2594                                 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2595                                 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2596
2597                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2598                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2599                         for (j = 0; j < conf->raid_disks; j++)
2600                                 if (j != sh2->pd_idx &&
2601                                     (!r6s || j != raid6_next_disk(sh2->pd_idx,
2602                                                                  sh2->disks)) &&
2603                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2604                                         break;
2605                         if (j == conf->raid_disks) {
2606                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2607                                 set_bit(STRIPE_HANDLE, &sh2->state);
2608                         }
2609                         release_stripe(sh2);
2610
2611                 }
2612         /* done submitting copies, wait for them to complete */
2613         if (tx) {
2614                 async_tx_ack(tx);
2615                 dma_wait_for_async_tx(tx);
2616         }
2617 }
2618
2619
2620 /*
2621  * handle_stripe - do things to a stripe.
2622  *
2623  * We lock the stripe and then examine the state of various bits
2624  * to see what needs to be done.
2625  * Possible results:
2626  *    return some read request which now have data
2627  *    return some write requests which are safely on disc
2628  *    schedule a read on some buffers
2629  *    schedule a write of some buffers
2630  *    return confirmation of parity correctness
2631  *
2632  * buffers are taken off read_list or write_list, and bh_cache buffers
2633  * get BH_Lock set before the stripe lock is released.
2634  *
2635  */
2636
2637 static void handle_stripe5(struct stripe_head *sh)
2638 {
2639         raid5_conf_t *conf = sh->raid_conf;
2640         int disks = sh->disks, i;
2641         struct bio *return_bi = NULL;
2642         struct stripe_head_state s;
2643         struct r5dev *dev;
2644         unsigned long pending = 0;
2645         mdk_rdev_t *blocked_rdev = NULL;
2646         int prexor;
2647
2648         memset(&s, 0, sizeof(s));
2649         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2650                 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2651                 atomic_read(&sh->count), sh->pd_idx,
2652                 sh->ops.pending, sh->ops.ack, sh->ops.complete);
2653
2654         spin_lock(&sh->lock);
2655         clear_bit(STRIPE_HANDLE, &sh->state);
2656         clear_bit(STRIPE_DELAYED, &sh->state);
2657
2658         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2659         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2660         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2661         /* Now to look around and see what can be done */
2662
2663         /* clean-up completed biofill operations */
2664         if (test_bit(STRIPE_OP_BIOFILL, &sh->ops.complete)) {
2665                 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending);
2666                 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack);
2667                 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
2668         }
2669
2670         rcu_read_lock();
2671         for (i=disks; i--; ) {
2672                 mdk_rdev_t *rdev;
2673                 struct r5dev *dev = &sh->dev[i];
2674                 clear_bit(R5_Insync, &dev->flags);
2675
2676                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2677                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2678                         dev->towrite, dev->written);
2679
2680                 /* maybe we can request a biofill operation
2681                  *
2682                  * new wantfill requests are only permitted while
2683                  * STRIPE_OP_BIOFILL is clear
2684                  */
2685                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2686                         !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2687                         set_bit(R5_Wantfill, &dev->flags);
2688
2689                 /* now count some things */
2690                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2691                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2692                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2693
2694                 if (test_bit(R5_Wantfill, &dev->flags))
2695                         s.to_fill++;
2696                 else if (dev->toread)
2697                         s.to_read++;
2698                 if (dev->towrite) {
2699                         s.to_write++;
2700                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2701                                 s.non_overwrite++;
2702                 }
2703                 if (dev->written)
2704                         s.written++;
2705                 rdev = rcu_dereference(conf->disks[i].rdev);
2706                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2707                         blocked_rdev = rdev;
2708                         atomic_inc(&rdev->nr_pending);
2709                         break;
2710                 }
2711                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2712                         /* The ReadError flag will just be confusing now */
2713                         clear_bit(R5_ReadError, &dev->flags);
2714                         clear_bit(R5_ReWrite, &dev->flags);
2715                 }
2716                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2717                     || test_bit(R5_ReadError, &dev->flags)) {
2718                         s.failed++;
2719                         s.failed_num = i;
2720                 } else
2721                         set_bit(R5_Insync, &dev->flags);
2722         }
2723         rcu_read_unlock();
2724
2725         if (unlikely(blocked_rdev)) {
2726                 set_bit(STRIPE_HANDLE, &sh->state);
2727                 goto unlock;
2728         }
2729
2730         if (s.to_fill && !test_and_set_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2731                 sh->ops.count++;
2732
2733         pr_debug("locked=%d uptodate=%d to_read=%d"
2734                 " to_write=%d failed=%d failed_num=%d\n",
2735                 s.locked, s.uptodate, s.to_read, s.to_write,
2736                 s.failed, s.failed_num);
2737         /* check if the array has lost two devices and, if so, some requests might
2738          * need to be failed
2739          */
2740         if (s.failed > 1 && s.to_read+s.to_write+s.written)
2741                 handle_requests_to_failed_array(conf, sh, &s, disks,
2742                                                 &return_bi);
2743         if (s.failed > 1 && s.syncing) {
2744                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2745                 clear_bit(STRIPE_SYNCING, &sh->state);
2746                 s.syncing = 0;
2747         }
2748
2749         /* might be able to return some write requests if the parity block
2750          * is safe, or on a failed drive
2751          */
2752         dev = &sh->dev[sh->pd_idx];
2753         if ( s.written &&
2754              ((test_bit(R5_Insync, &dev->flags) &&
2755                !test_bit(R5_LOCKED, &dev->flags) &&
2756                test_bit(R5_UPTODATE, &dev->flags)) ||
2757                (s.failed == 1 && s.failed_num == sh->pd_idx)))
2758                 handle_completed_write_requests(conf, sh, disks, &return_bi);
2759
2760         /* Now we might consider reading some blocks, either to check/generate
2761          * parity, or to satisfy requests
2762          * or to load a block that is being partially written.
2763          */
2764         if (s.to_read || s.non_overwrite ||
2765             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding ||
2766             test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2767                 handle_issuing_new_read_requests5(sh, &s, disks);
2768
2769         /* Now we check to see if any write operations have recently
2770          * completed
2771          */
2772
2773         /* leave prexor set until postxor is done, allows us to distinguish
2774          * a rmw from a rcw during biodrain
2775          */
2776         prexor = 0;
2777         if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2778                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2779
2780                 prexor = 1;
2781                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2782                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2783                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2784
2785                 for (i = disks; i--; )
2786                         clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2787         }
2788
2789         /* if only POSTXOR is set then this is an 'expand' postxor */
2790         if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2791                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2792
2793                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2794                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2795                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2796
2797                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2798                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2799                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2800
2801                 /* All the 'written' buffers and the parity block are ready to
2802                  * be written back to disk
2803                  */
2804                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2805                 for (i = disks; i--; ) {
2806                         dev = &sh->dev[i];
2807                         if (test_bit(R5_LOCKED, &dev->flags) &&
2808                                 (i == sh->pd_idx || dev->written)) {
2809                                 pr_debug("Writing block %d\n", i);
2810                                 set_bit(R5_Wantwrite, &dev->flags);
2811                                 if (!test_and_set_bit(
2812                                     STRIPE_OP_IO, &sh->ops.pending))
2813                                         sh->ops.count++;
2814                                 if (prexor)
2815                                         continue;
2816                                 if (!test_bit(R5_Insync, &dev->flags) ||
2817                                     (i == sh->pd_idx && s.failed == 0))
2818                                         set_bit(STRIPE_INSYNC, &sh->state);
2819                         }
2820                 }
2821                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2822                         atomic_dec(&conf->preread_active_stripes);
2823                         if (atomic_read(&conf->preread_active_stripes) <
2824                                 IO_THRESHOLD)
2825                                 md_wakeup_thread(conf->mddev->thread);
2826                 }
2827         }
2828
2829         /* Now to consider new write requests and what else, if anything
2830          * should be read.  We do not handle new writes when:
2831          * 1/ A 'write' operation (copy+xor) is already in flight.
2832          * 2/ A 'check' operation is in flight, as it may clobber the parity
2833          *    block.
2834          */
2835         if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2836                           !test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2837                 handle_issuing_new_write_requests5(conf, sh, &s, disks);
2838
2839         /* maybe we need to check and possibly fix the parity for this stripe
2840          * Any reads will already have been scheduled, so we just see if enough
2841          * data is available.  The parity check is held off while parity
2842          * dependent operations are in flight.
2843          */
2844         if ((s.syncing && s.locked == 0 &&
2845              !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2846              !test_bit(STRIPE_INSYNC, &sh->state)) ||
2847               test_bit(STRIPE_OP_CHECK, &sh->ops.pending) ||
2848               test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending))
2849                 handle_parity_checks5(conf, sh, &s, disks);
2850
2851         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2852                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2853                 clear_bit(STRIPE_SYNCING, &sh->state);
2854         }
2855
2856         /* If the failed drive is just a ReadError, then we might need to progress
2857          * the repair/check process
2858          */
2859         if (s.failed == 1 && !conf->mddev->ro &&
2860             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2861             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2862             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2863                 ) {
2864                 dev = &sh->dev[s.failed_num];
2865                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2866                         set_bit(R5_Wantwrite, &dev->flags);
2867                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2868                                 sh->ops.count++;
2869                         set_bit(R5_ReWrite, &dev->flags);
2870                         set_bit(R5_LOCKED, &dev->flags);
2871                         s.locked++;
2872                 } else {
2873                         /* let's read it back */
2874                         set_bit(R5_Wantread, &dev->flags);
2875                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2876                                 sh->ops.count++;
2877                         set_bit(R5_LOCKED, &dev->flags);
2878                         s.locked++;
2879                 }
2880         }
2881
2882         /* Finish postxor operations initiated by the expansion
2883          * process
2884          */
2885         if (test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete) &&
2886                 !test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending)) {
2887
2888                 clear_bit(STRIPE_EXPANDING, &sh->state);
2889
2890                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2891                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2892                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2893
2894                 for (i = conf->raid_disks; i--; ) {
2895                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2896                         set_bit(R5_LOCKED, &dev->flags);
2897                         s.locked++;
2898                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2899                                 sh->ops.count++;
2900                 }
2901         }
2902
2903         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2904                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2905                 /* Need to write out all blocks after computing parity */
2906                 sh->disks = conf->raid_disks;
2907                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2908                         conf->raid_disks);
2909                 s.locked += handle_write_operations5(sh, 1, 1);
2910         } else if (s.expanded &&
2911                    s.locked == 0 &&
2912                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2913                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2914                 atomic_dec(&conf->reshape_stripes);
2915                 wake_up(&conf->wait_for_overlap);
2916                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2917         }
2918
2919         if (s.expanding && s.locked == 0 &&
2920             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2921                 handle_stripe_expansion(conf, sh, NULL);
2922
2923         if (sh->ops.count)
2924                 pending = get_stripe_work(sh);
2925
2926  unlock:
2927         spin_unlock(&sh->lock);
2928
2929         /* wait for this device to become unblocked */
2930         if (unlikely(blocked_rdev))
2931                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
2932
2933         if (pending)
2934                 raid5_run_ops(sh, pending);
2935
2936         return_io(return_bi);
2937
2938 }
2939
2940 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2941 {
2942         raid6_conf_t *conf = sh->raid_conf;
2943         int disks = sh->disks;
2944         struct bio *return_bi = NULL;
2945         int i, pd_idx = sh->pd_idx;
2946         struct stripe_head_state s;
2947         struct r6_state r6s;
2948         struct r5dev *dev, *pdev, *qdev;
2949         mdk_rdev_t *blocked_rdev = NULL;
2950
2951         r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2952         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2953                 "pd_idx=%d, qd_idx=%d\n",
2954                (unsigned long long)sh->sector, sh->state,
2955                atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2956         memset(&s, 0, sizeof(s));
2957
2958         spin_lock(&sh->lock);
2959         clear_bit(STRIPE_HANDLE, &sh->state);
2960         clear_bit(STRIPE_DELAYED, &sh->state);
2961
2962         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2963         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2964         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2965         /* Now to look around and see what can be done */
2966
2967         rcu_read_lock();
2968         for (i=disks; i--; ) {
2969                 mdk_rdev_t *rdev;
2970                 dev = &sh->dev[i];
2971                 clear_bit(R5_Insync, &dev->flags);
2972
2973                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2974                         i, dev->flags, dev->toread, dev->towrite, dev->written);
2975                 /* maybe we can reply to a read */
2976                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2977                         struct bio *rbi, *rbi2;
2978                         pr_debug("Return read for disc %d\n", i);
2979                         spin_lock_irq(&conf->device_lock);
2980                         rbi = dev->toread;
2981                         dev->toread = NULL;
2982                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2983                                 wake_up(&conf->wait_for_overlap);
2984                         spin_unlock_irq(&conf->device_lock);
2985                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2986                                 copy_data(0, rbi, dev->page, dev->sector);
2987                                 rbi2 = r5_next_bio(rbi, dev->sector);
2988                                 spin_lock_irq(&conf->device_lock);
2989                                 if (--rbi->bi_phys_segments == 0) {
2990                                         rbi->bi_next = return_bi;
2991                                         return_bi = rbi;
2992                                 }
2993                                 spin_unlock_irq(&conf->device_lock);
2994                                 rbi = rbi2;
2995                         }
2996                 }
2997
2998                 /* now count some things */
2999                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3000                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3001
3002
3003                 if (dev->toread)
3004                         s.to_read++;
3005                 if (dev->towrite) {
3006                         s.to_write++;
3007                         if (!test_bit(R5_OVERWRITE, &dev->flags))
3008                                 s.non_overwrite++;
3009                 }
3010                 if (dev->written)
3011                         s.written++;
3012                 rdev = rcu_dereference(conf->disks[i].rdev);
3013                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3014                         blocked_rdev = rdev;
3015                         atomic_inc(&rdev->nr_pending);
3016                         break;
3017                 }
3018                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3019                         /* The ReadError flag will just be confusing now */
3020                         clear_bit(R5_ReadError, &dev->flags);
3021                         clear_bit(R5_ReWrite, &dev->flags);
3022                 }
3023                 if (!rdev || !test_bit(In_sync, &rdev->flags)
3024                     || test_bit(R5_ReadError, &dev->flags)) {
3025                         if (s.failed < 2)
3026                                 r6s.failed_num[s.failed] = i;
3027                         s.failed++;
3028                 } else
3029                         set_bit(R5_Insync, &dev->flags);
3030         }
3031         rcu_read_unlock();
3032
3033         if (unlikely(blocked_rdev)) {
3034                 set_bit(STRIPE_HANDLE, &sh->state);
3035                 goto unlock;
3036         }
3037         pr_debug("locked=%d uptodate=%d to_read=%d"
3038                " to_write=%d failed=%d failed_num=%d,%d\n",
3039                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3040                r6s.failed_num[0], r6s.failed_num[1]);
3041         /* check if the array has lost >2 devices and, if so, some requests
3042          * might need to be failed
3043          */
3044         if (s.failed > 2 && s.to_read+s.to_write+s.written)
3045                 handle_requests_to_failed_array(conf, sh, &s, disks,
3046                                                 &return_bi);
3047         if (s.failed > 2 && s.syncing) {
3048                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3049                 clear_bit(STRIPE_SYNCING, &sh->state);
3050                 s.syncing = 0;
3051         }
3052
3053         /*
3054          * might be able to return some write requests if the parity blocks
3055          * are safe, or on a failed drive
3056          */
3057         pdev = &sh->dev[pd_idx];
3058         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3059                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3060         qdev = &sh->dev[r6s.qd_idx];
3061         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
3062                 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
3063
3064         if ( s.written &&
3065              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3066                              && !test_bit(R5_LOCKED, &pdev->flags)
3067                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3068              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3069                              && !test_bit(R5_LOCKED, &qdev->flags)
3070                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3071                 handle_completed_write_requests(conf, sh, disks, &return_bi);
3072
3073         /* Now we might consider reading some blocks, either to check/generate
3074          * parity, or to satisfy requests
3075          * or to load a block that is being partially written.
3076          */
3077         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3078             (s.syncing && (s.uptodate < disks)) || s.expanding)
3079                 handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
3080
3081         /* now to consider writing and what else, if anything should be read */
3082         if (s.to_write)
3083                 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
3084
3085         /* maybe we need to check and possibly fix the parity for this stripe
3086          * Any reads will already have been scheduled, so we just see if enough
3087          * data is available
3088          */
3089         if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3090                 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3091
3092         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3093                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3094                 clear_bit(STRIPE_SYNCING, &sh->state);
3095         }
3096
3097         /* If the failed drives are just a ReadError, then we might need
3098          * to progress the repair/check process
3099          */
3100         if (s.failed <= 2 && !conf->mddev->ro)
3101                 for (i = 0; i < s.failed; i++) {
3102                         dev = &sh->dev[r6s.failed_num[i]];
3103                         if (test_bit(R5_ReadError, &dev->flags)
3104                             && !test_bit(R5_LOCKED, &dev->flags)
3105                             && test_bit(R5_UPTODATE, &dev->flags)
3106                                 ) {
3107                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3108                                         set_bit(R5_Wantwrite, &dev->flags);
3109                                         set_bit(R5_ReWrite, &dev->flags);
3110                                         set_bit(R5_LOCKED, &dev->flags);
3111                                 } else {
3112                                         /* let's read it back */
3113                                         set_bit(R5_Wantread, &dev->flags);
3114                                         set_bit(R5_LOCKED, &dev->flags);
3115                                 }
3116                         }
3117                 }
3118
3119         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3120                 /* Need to write out all blocks after computing P&Q */
3121                 sh->disks = conf->raid_disks;
3122                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
3123                                              conf->raid_disks);
3124                 compute_parity6(sh, RECONSTRUCT_WRITE);
3125                 for (i = conf->raid_disks ; i-- ;  ) {
3126                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3127                         s.locked++;
3128                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3129                 }
3130                 clear_bit(STRIPE_EXPANDING, &sh->state);
3131         } else if (s.expanded) {
3132                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3133                 atomic_dec(&conf->reshape_stripes);
3134                 wake_up(&conf->wait_for_overlap);
3135                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3136         }
3137
3138         if (s.expanding && s.locked == 0 &&
3139             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
3140                 handle_stripe_expansion(conf, sh, &r6s);
3141
3142  unlock:
3143         spin_unlock(&sh->lock);
3144
3145         /* wait for this device to become unblocked */
3146         if (unlikely(blocked_rdev))
3147                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3148
3149         return_io(return_bi);
3150
3151         for (i=disks; i-- ;) {
3152                 int rw;
3153                 struct bio *bi;
3154                 mdk_rdev_t *rdev;
3155                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
3156                         rw = WRITE;
3157                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
3158                         rw = READ;
3159                 else
3160                         continue;
3161
3162                 set_bit(STRIPE_IO_STARTED, &sh->state);
3163
3164                 bi = &sh->dev[i].req;
3165
3166                 bi->bi_rw = rw;
3167                 if (rw == WRITE)
3168                         bi->bi_end_io = raid5_end_write_request;
3169                 else
3170                         bi->bi_end_io = raid5_end_read_request;
3171
3172                 rcu_read_lock();
3173                 rdev = rcu_dereference(conf->disks[i].rdev);
3174                 if (rdev && test_bit(Faulty, &rdev->flags))
3175                         rdev = NULL;
3176                 if (rdev)
3177                         atomic_inc(&rdev->nr_pending);
3178                 rcu_read_unlock();
3179
3180                 if (rdev) {
3181                         if (s.syncing || s.expanding || s.expanded)
3182                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
3183
3184                         bi->bi_bdev = rdev->bdev;
3185                         pr_debug("for %llu schedule op %ld on disc %d\n",
3186                                 (unsigned long long)sh->sector, bi->bi_rw, i);
3187                         atomic_inc(&sh->count);
3188                         bi->bi_sector = sh->sector + rdev->data_offset;
3189                         bi->bi_flags = 1 << BIO_UPTODATE;
3190                         bi->bi_vcnt = 1;
3191                         bi->bi_max_vecs = 1;
3192                         bi->bi_idx = 0;
3193                         bi->bi_io_vec = &sh->dev[i].vec;
3194                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
3195                         bi->bi_io_vec[0].bv_offset = 0;
3196                         bi->bi_size = STRIPE_SIZE;
3197                         bi->bi_next = NULL;
3198                         if (rw == WRITE &&
3199                             test_bit(R5_ReWrite, &sh->dev[i].flags))
3200                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
3201                         generic_make_request(bi);
3202                 } else {
3203                         if (rw == WRITE)
3204                                 set_bit(STRIPE_DEGRADED, &sh->state);
3205                         pr_debug("skip op %ld on disc %d for sector %llu\n",
3206                                 bi->bi_rw, i, (unsigned long long)sh->sector);
3207                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
3208                         set_bit(STRIPE_HANDLE, &sh->state);
3209                 }
3210         }
3211 }
3212
3213 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3214 {
3215         if (sh->raid_conf->level == 6)
3216                 handle_stripe6(sh, tmp_page);
3217         else
3218                 handle_stripe5(sh);
3219 }
3220
3221
3222
3223 static void raid5_activate_delayed(raid5_conf_t *conf)
3224 {
3225         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3226                 while (!list_empty(&conf->delayed_list)) {
3227                         struct list_head *l = conf->delayed_list.next;
3228                         struct stripe_head *sh;
3229                         sh = list_entry(l, struct stripe_head, lru);
3230                         list_del_init(l);
3231                         clear_bit(STRIPE_DELAYED, &sh->state);
3232                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3233                                 atomic_inc(&conf->preread_active_stripes);
3234                         list_add_tail(&sh->lru, &conf->hold_list);
3235                 }
3236         } else
3237                 blk_plug_device(conf->mddev->queue);
3238 }
3239
3240 static void activate_bit_delay(raid5_conf_t *conf)
3241 {
3242         /* device_lock is held */
3243         struct list_head head;
3244         list_add(&head, &conf->bitmap_list);
3245         list_del_init(&conf->bitmap_list);
3246         while (!list_empty(&head)) {
3247                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3248                 list_del_init(&sh->lru);
3249                 atomic_inc(&sh->count);
3250                 __release_stripe(conf, sh);
3251         }
3252 }
3253
3254 static void unplug_slaves(mddev_t *mddev)
3255 {
3256         raid5_conf_t *conf = mddev_to_conf(mddev);
3257         int i;
3258
3259         rcu_read_lock();
3260         for (i=0; i<mddev->raid_disks; i++) {
3261                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3262                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3263                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3264
3265                         atomic_inc(&rdev->nr_pending);
3266                         rcu_read_unlock();
3267
3268                         blk_unplug(r_queue);
3269
3270                         rdev_dec_pending(rdev, mddev);
3271                         rcu_read_lock();
3272                 }
3273         }
3274         rcu_read_unlock();
3275 }
3276
3277 static void raid5_unplug_device(struct request_queue *q)
3278 {
3279         mddev_t *mddev = q->queuedata;
3280         raid5_conf_t *conf = mddev_to_conf(mddev);
3281         unsigned long flags;
3282
3283         spin_lock_irqsave(&conf->device_lock, flags);
3284
3285         if (blk_remove_plug(q)) {
3286                 conf->seq_flush++;
3287                 raid5_activate_delayed(conf);
3288         }
3289         md_wakeup_thread(mddev->thread);
3290
3291         spin_unlock_irqrestore(&conf->device_lock, flags);
3292
3293         unplug_slaves(mddev);
3294 }
3295
3296 static int raid5_congested(void *data, int bits)
3297 {
3298         mddev_t *mddev = data;
3299         raid5_conf_t *conf = mddev_to_conf(mddev);
3300
3301         /* No difference between reads and writes.  Just check
3302          * how busy the stripe_cache is
3303          */
3304         if (conf->inactive_blocked)
3305                 return 1;
3306         if (conf->quiesce)
3307                 return 1;
3308         if (list_empty_careful(&conf->inactive_list))
3309                 return 1;
3310
3311         return 0;
3312 }
3313
3314 /* We want read requests to align with chunks where possible,
3315  * but write requests don't need to.
3316  */
3317 static int raid5_mergeable_bvec(struct request_queue *q,
3318                                 struct bvec_merge_data *bvm,
3319                                 struct bio_vec *biovec)
3320 {
3321         mddev_t *mddev = q->queuedata;
3322         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3323         int max;
3324         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3325         unsigned int bio_sectors = bvm->bi_size >> 9;
3326
3327         if ((bvm->bi_rw & 1) == WRITE)
3328                 return biovec->bv_len; /* always allow writes to be mergeable */
3329
3330         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3331         if (max < 0) max = 0;
3332         if (max <= biovec->bv_len && bio_sectors == 0)
3333                 return biovec->bv_len;
3334         else
3335                 return max;
3336 }
3337
3338
3339 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3340 {
3341         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3342         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3343         unsigned int bio_sectors = bio->bi_size >> 9;
3344
3345         return  chunk_sectors >=
3346                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3347 }
3348
3349 /*
3350  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3351  *  later sampled by raid5d.
3352  */
3353 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3354 {
3355         unsigned long flags;
3356
3357         spin_lock_irqsave(&conf->device_lock, flags);
3358
3359         bi->bi_next = conf->retry_read_aligned_list;
3360         conf->retry_read_aligned_list = bi;
3361
3362         spin_unlock_irqrestore(&conf->device_lock, flags);
3363         md_wakeup_thread(conf->mddev->thread);
3364 }
3365
3366
3367 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3368 {
3369         struct bio *bi;
3370
3371         bi = conf->retry_read_aligned;
3372         if (bi) {
3373                 conf->retry_read_aligned = NULL;
3374                 return bi;
3375         }
3376         bi = conf->retry_read_aligned_list;
3377         if(bi) {
3378                 conf->retry_read_aligned_list = bi->bi_next;
3379                 bi->bi_next = NULL;
3380                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3381                 bi->bi_hw_segments = 0; /* count of processed stripes */
3382         }
3383
3384         return bi;
3385 }
3386
3387
3388 /*
3389  *  The "raid5_align_endio" should check if the read succeeded and if it
3390  *  did, call bio_endio on the original bio (having bio_put the new bio
3391  *  first).
3392  *  If the read failed..
3393  */
3394 static void raid5_align_endio(struct bio *bi, int error)
3395 {
3396         struct bio* raid_bi  = bi->bi_private;
3397         mddev_t *mddev;
3398         raid5_conf_t *conf;
3399         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3400         mdk_rdev_t *rdev;
3401
3402         bio_put(bi);
3403
3404         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3405         conf = mddev_to_conf(mddev);
3406         rdev = (void*)raid_bi->bi_next;
3407         raid_bi->bi_next = NULL;
3408
3409         rdev_dec_pending(rdev, conf->mddev);
3410
3411         if (!error && uptodate) {
3412                 bio_endio(raid_bi, 0);
3413                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3414                         wake_up(&conf->wait_for_stripe);
3415                 return;
3416         }
3417
3418
3419         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3420
3421         add_bio_to_retry(raid_bi, conf);
3422 }
3423
3424 static int bio_fits_rdev(struct bio *bi)
3425 {
3426         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3427
3428         if ((bi->bi_size>>9) > q->max_sectors)
3429                 return 0;
3430         blk_recount_segments(q, bi);
3431         if (bi->bi_phys_segments > q->max_phys_segments ||
3432             bi->bi_hw_segments > q->max_hw_segments)
3433                 return 0;
3434
3435         if (q->merge_bvec_fn)
3436                 /* it's too hard to apply the merge_bvec_fn at this stage,
3437                  * just just give up
3438                  */
3439                 return 0;
3440
3441         return 1;
3442 }
3443
3444
3445 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3446 {
3447         mddev_t *mddev = q->queuedata;
3448         raid5_conf_t *conf = mddev_to_conf(mddev);
3449         const unsigned int raid_disks = conf->raid_disks;
3450         const unsigned int data_disks = raid_disks - conf->max_degraded;
3451         unsigned int dd_idx, pd_idx;
3452         struct bio* align_bi;
3453         mdk_rdev_t *rdev;
3454
3455         if (!in_chunk_boundary(mddev, raid_bio)) {
3456                 pr_debug("chunk_aligned_read : non aligned\n");
3457                 return 0;
3458         }
3459         /*
3460          * use bio_clone to make a copy of the bio
3461          */
3462         align_bi = bio_clone(raid_bio, GFP_NOIO);
3463         if (!align_bi)
3464                 return 0;
3465         /*
3466          *   set bi_end_io to a new function, and set bi_private to the
3467          *     original bio.
3468          */
3469         align_bi->bi_end_io  = raid5_align_endio;
3470         align_bi->bi_private = raid_bio;
3471         /*
3472          *      compute position
3473          */
3474         align_bi->bi_sector =  raid5_compute_sector(raid_bio->bi_sector,
3475                                         raid_disks,
3476                                         data_disks,
3477                                         &dd_idx,
3478                                         &pd_idx,
3479                                         conf);
3480
3481         rcu_read_lock();
3482         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3483         if (rdev && test_bit(In_sync, &rdev->flags)) {
3484                 atomic_inc(&rdev->nr_pending);
3485                 rcu_read_unlock();
3486                 raid_bio->bi_next = (void*)rdev;
3487                 align_bi->bi_bdev =  rdev->bdev;
3488                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3489                 align_bi->bi_sector += rdev->data_offset;
3490
3491                 if (!bio_fits_rdev(align_bi)) {
3492                         /* too big in some way */
3493                         bio_put(align_bi);
3494                         rdev_dec_pending(rdev, mddev);
3495                         return 0;
3496                 }
3497
3498                 spin_lock_irq(&conf->device_lock);
3499                 wait_event_lock_irq(conf->wait_for_stripe,
3500                                     conf->quiesce == 0,
3501                                     conf->device_lock, /* nothing */);
3502                 atomic_inc(&conf->active_aligned_reads);
3503                 spin_unlock_irq(&conf->device_lock);
3504
3505                 generic_make_request(align_bi);
3506                 return 1;
3507         } else {
3508                 rcu_read_unlock();
3509                 bio_put(align_bi);
3510                 return 0;
3511         }
3512 }
3513
3514 /* __get_priority_stripe - get the next stripe to process
3515  *
3516  * Full stripe writes are allowed to pass preread active stripes up until
3517  * the bypass_threshold is exceeded.  In general the bypass_count
3518  * increments when the handle_list is handled before the hold_list; however, it
3519  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3520  * stripe with in flight i/o.  The bypass_count will be reset when the
3521  * head of the hold_list has changed, i.e. the head was promoted to the
3522  * handle_list.
3523  */
3524 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3525 {
3526         struct stripe_head *sh;
3527
3528         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3529                   __func__,
3530                   list_empty(&conf->handle_list) ? "empty" : "busy",
3531                   list_empty(&conf->hold_list) ? "empty" : "busy",
3532                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
3533
3534         if (!list_empty(&conf->handle_list)) {
3535                 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3536
3537                 if (list_empty(&conf->hold_list))
3538                         conf->bypass_count = 0;
3539                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3540                         if (conf->hold_list.next == conf->last_hold)
3541                                 conf->bypass_count++;
3542                         else {
3543                                 conf->last_hold = conf->hold_list.next;
3544                                 conf->bypass_count -= conf->bypass_threshold;
3545                                 if (conf->bypass_count < 0)
3546                                         conf->bypass_count = 0;
3547                         }
3548                 }
3549         } else if (!list_empty(&conf->hold_list) &&
3550                    ((conf->bypass_threshold &&
3551                      conf->bypass_count > conf->bypass_threshold) ||
3552                     atomic_read(&conf->pending_full_writes) == 0)) {
3553                 sh = list_entry(conf->hold_list.next,
3554                                 typeof(*sh), lru);
3555                 conf->bypass_count -= conf->bypass_threshold;
3556                 if (conf->bypass_count < 0)
3557                         conf->bypass_count = 0;
3558         } else
3559                 return NULL;
3560
3561         list_del_init(&sh->lru);
3562         atomic_inc(&sh->count);
3563         BUG_ON(atomic_read(&sh->count) != 1);
3564         return sh;
3565 }
3566
3567 static int make_request(struct request_queue *q, struct bio * bi)
3568 {
3569         mddev_t *mddev = q->queuedata;
3570         raid5_conf_t *conf = mddev_to_conf(mddev);
3571         unsigned int dd_idx, pd_idx;
3572         sector_t new_sector;
3573         sector_t logical_sector, last_sector;
3574         struct stripe_head *sh;
3575         const int rw = bio_data_dir(bi);
3576         int remaining;
3577
3578         if (unlikely(bio_barrier(bi))) {
3579                 bio_endio(bi, -EOPNOTSUPP);
3580                 return 0;
3581         }
3582
3583         md_write_start(mddev, bi);
3584
3585         disk_stat_inc(mddev->gendisk, ios[rw]);
3586         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3587
3588         if (rw == READ &&
3589              mddev->reshape_position == MaxSector &&
3590              chunk_aligned_read(q,bi))
3591                 return 0;
3592
3593         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3594         last_sector = bi->bi_sector + (bi->bi_size>>9);
3595         bi->bi_next = NULL;
3596         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3597
3598         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3599                 DEFINE_WAIT(w);
3600                 int disks, data_disks;
3601
3602         retry:
3603                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3604                 if (likely(conf->expand_progress == MaxSector))
3605                         disks = conf->raid_disks;
3606                 else {
3607                         /* spinlock is needed as expand_progress may be
3608                          * 64bit on a 32bit platform, and so it might be
3609                          * possible to see a half-updated value
3610                          * Ofcourse expand_progress could change after
3611                          * the lock is dropped, so once we get a reference
3612                          * to the stripe that we think it is, we will have
3613                          * to check again.
3614                          */
3615                         spin_lock_irq(&conf->device_lock);
3616                         disks = conf->raid_disks;
3617                         if (logical_sector >= conf->expand_progress)
3618                                 disks = conf->previous_raid_disks;
3619                         else {
3620                                 if (logical_sector >= conf->expand_lo) {
3621                                         spin_unlock_irq(&conf->device_lock);
3622                                         schedule();
3623                                         goto retry;
3624                                 }
3625                         }
3626                         spin_unlock_irq(&conf->device_lock);
3627                 }
3628                 data_disks = disks - conf->max_degraded;
3629
3630                 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3631                                                   &dd_idx, &pd_idx, conf);
3632                 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3633                         (unsigned long long)new_sector, 
3634                         (unsigned long long)logical_sector);
3635
3636                 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3637                 if (sh) {
3638                         if (unlikely(conf->expand_progress != MaxSector)) {
3639                                 /* expansion might have moved on while waiting for a
3640                                  * stripe, so we must do the range check again.
3641                                  * Expansion could still move past after this
3642                                  * test, but as we are holding a reference to
3643                                  * 'sh', we know that if that happens,
3644                                  *  STRIPE_EXPANDING will get set and the expansion
3645                                  * won't proceed until we finish with the stripe.
3646                                  */
3647                                 int must_retry = 0;
3648                                 spin_lock_irq(&conf->device_lock);
3649                                 if (logical_sector <  conf->expand_progress &&
3650                                     disks == conf->previous_raid_disks)
3651                                         /* mismatch, need to try again */
3652                                         must_retry = 1;
3653                                 spin_unlock_irq(&conf->device_lock);
3654                                 if (must_retry) {
3655                                         release_stripe(sh);
3656                                         goto retry;
3657                                 }
3658                         }
3659                         /* FIXME what if we get a false positive because these
3660                          * are being updated.
3661                          */
3662                         if (logical_sector >= mddev->suspend_lo &&
3663                             logical_sector < mddev->suspend_hi) {
3664                                 release_stripe(sh);
3665                                 schedule();
3666                                 goto retry;
3667                         }
3668
3669                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3670                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3671                                 /* Stripe is busy expanding or
3672                                  * add failed due to overlap.  Flush everything
3673                                  * and wait a while
3674                                  */
3675                                 raid5_unplug_device(mddev->queue);
3676                                 release_stripe(sh);
3677                                 schedule();
3678                                 goto retry;
3679                         }
3680                         finish_wait(&conf->wait_for_overlap, &w);
3681                         set_bit(STRIPE_HANDLE, &sh->state);
3682                         clear_bit(STRIPE_DELAYED, &sh->state);
3683                         release_stripe(sh);
3684                 } else {
3685                         /* cannot get stripe for read-ahead, just give-up */
3686                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3687                         finish_wait(&conf->wait_for_overlap, &w);
3688                         break;
3689                 }
3690                         
3691         }
3692         spin_lock_irq(&conf->device_lock);
3693         remaining = --bi->bi_phys_segments;
3694         spin_unlock_irq(&conf->device_lock);
3695         if (remaining == 0) {
3696
3697                 if ( rw == WRITE )
3698                         md_write_end(mddev);
3699
3700                 bi->bi_end_io(bi,
3701                               test_bit(BIO_UPTODATE, &bi->bi_flags)
3702                                 ? 0 : -EIO);
3703         }
3704         return 0;
3705 }
3706
3707 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3708 {
3709         /* reshaping is quite different to recovery/resync so it is
3710          * handled quite separately ... here.
3711          *
3712          * On each call to sync_request, we gather one chunk worth of
3713          * destination stripes and flag them as expanding.
3714          * Then we find all the source stripes and request reads.
3715          * As the reads complete, handle_stripe will copy the data
3716          * into the destination stripe and release that stripe.
3717          */
3718         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3719         struct stripe_head *sh;
3720         int pd_idx;
3721         sector_t first_sector, last_sector;
3722         int raid_disks = conf->previous_raid_disks;
3723         int data_disks = raid_disks - conf->max_degraded;
3724         int new_data_disks = conf->raid_disks - conf->max_degraded;
3725         int i;
3726         int dd_idx;
3727         sector_t writepos, safepos, gap;
3728
3729         if (sector_nr == 0 &&
3730             conf->expand_progress != 0) {
3731                 /* restarting in the middle, skip the initial sectors */
3732                 sector_nr = conf->expand_progress;
3733                 sector_div(sector_nr, new_data_disks);
3734                 *skipped = 1;
3735                 return sector_nr;
3736         }
3737
3738         /* we update the metadata when there is more than 3Meg
3739          * in the block range (that is rather arbitrary, should
3740          * probably be time based) or when the data about to be
3741          * copied would over-write the source of the data at
3742          * the front of the range.
3743          * i.e. one new_stripe forward from expand_progress new_maps
3744          * to after where expand_lo old_maps to
3745          */
3746         writepos = conf->expand_progress +
3747                 conf->chunk_size/512*(new_data_disks);
3748         sector_div(writepos, new_data_disks);
3749         safepos = conf->expand_lo;
3750         sector_div(safepos, data_disks);
3751         gap = conf->expand_progress - conf->expand_lo;
3752
3753         if (writepos >= safepos ||
3754             gap > (new_data_disks)*3000*2 /*3Meg*/) {
3755                 /* Cannot proceed until we've updated the superblock... */
3756                 wait_event(conf->wait_for_overlap,
3757                            atomic_read(&conf->reshape_stripes)==0);
3758                 mddev->reshape_position = conf->expand_progress;
3759                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3760                 md_wakeup_thread(mddev->thread);
3761                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3762                            kthread_should_stop());
3763                 spin_lock_irq(&conf->device_lock);
3764                 conf->expand_lo = mddev->reshape_position;
3765                 spin_unlock_irq(&conf->device_lock);
3766                 wake_up(&conf->wait_for_overlap);
3767         }
3768
3769         for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3770                 int j;
3771                 int skipped = 0;
3772                 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3773                 sh = get_active_stripe(conf, sector_nr+i,
3774                                        conf->raid_disks, pd_idx, 0);
3775                 set_bit(STRIPE_EXPANDING, &sh->state);
3776                 atomic_inc(&conf->reshape_stripes);
3777                 /* If any of this stripe is beyond the end of the old
3778                  * array, then we need to zero those blocks
3779                  */
3780                 for (j=sh->disks; j--;) {
3781                         sector_t s;
3782                         if (j == sh->pd_idx)
3783                                 continue;
3784                         if (conf->level == 6 &&
3785                             j == raid6_next_disk(sh->pd_idx, sh->disks))
3786                                 continue;
3787                         s = compute_blocknr(sh, j);
3788                         if (s < (mddev->array_size<<1)) {
3789                                 skipped = 1;
3790                                 continue;
3791                         }
3792                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3793                         set_bit(R5_Expanded, &sh->dev[j].flags);
3794                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
3795                 }
3796                 if (!skipped) {
3797                         set_bit(STRIPE_EXPAND_READY, &sh->state);
3798                         set_bit(STRIPE_HANDLE, &sh->state);
3799                 }
3800                 release_stripe(sh);
3801         }
3802         spin_lock_irq(&conf->device_lock);
3803         conf->expand_progress = (sector_nr + i) * new_data_disks;
3804         spin_unlock_irq(&conf->device_lock);
3805         /* Ok, those stripe are ready. We can start scheduling
3806          * reads on the source stripes.
3807          * The source stripes are determined by mapping the first and last
3808          * block on the destination stripes.
3809          */
3810         first_sector =
3811                 raid5_compute_sector(sector_nr*(new_data_disks),
3812                                      raid_disks, data_disks,
3813                                      &dd_idx, &pd_idx, conf);
3814         last_sector =
3815                 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3816                                      *(new_data_disks) -1,
3817                                      raid_disks, data_disks,
3818                                      &dd_idx, &pd_idx, conf);
3819         if (last_sector >= (mddev->size<<1))
3820                 last_sector = (mddev->size<<1)-1;
3821         while (first_sector <= last_sector) {
3822                 pd_idx = stripe_to_pdidx(first_sector, conf,
3823                                          conf->previous_raid_disks);
3824                 sh = get_active_stripe(conf, first_sector,
3825                                        conf->previous_raid_disks, pd_idx, 0);
3826                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3827                 set_bit(STRIPE_HANDLE, &sh->state);
3828                 release_stripe(sh);
3829                 first_sector += STRIPE_SECTORS;
3830         }
3831         /* If this takes us to the resync_max point where we have to pause,
3832          * then we need to write out the superblock.
3833          */
3834         sector_nr += conf->chunk_size>>9;
3835         if (sector_nr >= mddev->resync_max) {
3836                 /* Cannot proceed until we've updated the superblock... */
3837                 wait_event(conf->wait_for_overlap,
3838                            atomic_read(&conf->reshape_stripes) == 0);
3839                 mddev->reshape_position = conf->expand_progress;
3840                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3841                 md_wakeup_thread(mddev->thread);
3842                 wait_event(mddev->sb_wait,
3843                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3844                            || kthread_should_stop());
3845                 spin_lock_irq(&conf->device_lock);
3846                 conf->expand_lo = mddev->reshape_position;
3847                 spin_unlock_irq(&conf->device_lock);
3848                 wake_up(&conf->wait_for_overlap);
3849         }
3850         return conf->chunk_size>>9;
3851 }
3852
3853 /* FIXME go_faster isn't used */
3854 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3855 {
3856         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3857         struct stripe_head *sh;
3858         int pd_idx;
3859         int raid_disks = conf->raid_disks;
3860         sector_t max_sector = mddev->size << 1;
3861         int sync_blocks;
3862         int still_degraded = 0;
3863         int i;
3864
3865         if (sector_nr >= max_sector) {
3866                 /* just being told to finish up .. nothing much to do */
3867                 unplug_slaves(mddev);
3868                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3869                         end_reshape(conf);
3870                         return 0;
3871                 }
3872
3873                 if (mddev->curr_resync < max_sector) /* aborted */
3874                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3875                                         &sync_blocks, 1);
3876                 else /* completed sync */
3877                         conf->fullsync = 0;
3878                 bitmap_close_sync(mddev->bitmap);
3879
3880                 return 0;
3881         }
3882
3883         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3884                 return reshape_request(mddev, sector_nr, skipped);
3885
3886         /* No need to check resync_max as we never do more than one
3887          * stripe, and as resync_max will always be on a chunk boundary,
3888          * if the check in md_do_sync didn't fire, there is no chance
3889          * of overstepping resync_max here
3890          */
3891
3892         /* if there is too many failed drives and we are trying
3893          * to resync, then assert that we are finished, because there is
3894          * nothing we can do.
3895          */
3896         if (mddev->degraded >= conf->max_degraded &&
3897             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3898                 sector_t rv = (mddev->size << 1) - sector_nr;
3899                 *skipped = 1;
3900                 return rv;
3901         }
3902         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3903             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3904             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3905                 /* we can skip this block, and probably more */
3906                 sync_blocks /= STRIPE_SECTORS;
3907                 *skipped = 1;
3908                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3909         }
3910
3911
3912         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3913
3914         pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3915         sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3916         if (sh == NULL) {
3917                 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3918                 /* make sure we don't swamp the stripe cache if someone else
3919                  * is trying to get access
3920                  */
3921                 schedule_timeout_uninterruptible(1);
3922         }
3923         /* Need to check if array will still be degraded after recovery/resync
3924          * We don't need to check the 'failed' flag as when that gets set,
3925          * recovery aborts.
3926          */
3927         for (i=0; i<mddev->raid_disks; i++)
3928                 if (conf->disks[i].rdev == NULL)
3929                         still_degraded = 1;
3930
3931         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3932
3933         spin_lock(&sh->lock);
3934         set_bit(STRIPE_SYNCING, &sh->state);
3935         clear_bit(STRIPE_INSYNC, &sh->state);
3936         spin_unlock(&sh->lock);
3937
3938         handle_stripe(sh, NULL);
3939         release_stripe(sh);
3940
3941         return STRIPE_SECTORS;
3942 }
3943
3944 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3945 {
3946         /* We may not be able to submit a whole bio at once as there
3947          * may not be enough stripe_heads available.
3948          * We cannot pre-allocate enough stripe_heads as we may need
3949          * more than exist in the cache (if we allow ever large chunks).
3950          * So we do one stripe head at a time and record in
3951          * ->bi_hw_segments how many have been done.
3952          *
3953          * We *know* that this entire raid_bio is in one chunk, so
3954          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3955          */
3956         struct stripe_head *sh;
3957         int dd_idx, pd_idx;
3958         sector_t sector, logical_sector, last_sector;
3959         int scnt = 0;
3960         int remaining;
3961         int handled = 0;
3962
3963         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3964         sector = raid5_compute_sector(  logical_sector,
3965                                         conf->raid_disks,
3966                                         conf->raid_disks - conf->max_degraded,
3967                                         &dd_idx,
3968                                         &pd_idx,
3969                                         conf);
3970         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3971
3972         for (; logical_sector < last_sector;
3973              logical_sector += STRIPE_SECTORS,
3974                      sector += STRIPE_SECTORS,
3975                      scnt++) {
3976
3977                 if (scnt < raid_bio->bi_hw_segments)
3978                         /* already done this stripe */
3979                         continue;
3980
3981                 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3982
3983                 if (!sh) {
3984                         /* failed to get a stripe - must wait */
3985                         raid_bio->bi_hw_segments = scnt;
3986                         conf->retry_read_aligned = raid_bio;
3987                         return handled;
3988                 }
3989
3990                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3991                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3992                         release_stripe(sh);
3993                         raid_bio->bi_hw_segments = scnt;
3994                         conf->retry_read_aligned = raid_bio;
3995                         return handled;
3996                 }
3997
3998                 handle_stripe(sh, NULL);
3999                 release_stripe(sh);
4000                 handled++;
4001         }
4002         spin_lock_irq(&conf->device_lock);
4003         remaining = --raid_bio->bi_phys_segments;
4004         spin_unlock_irq(&conf->device_lock);
4005         if (remaining == 0) {
4006
4007                 raid_bio->bi_end_io(raid_bio,
4008                               test_bit(BIO_UPTODATE, &raid_bio->bi_flags)
4009                                 ? 0 : -EIO);
4010         }
4011         if (atomic_dec_and_test(&conf->active_aligned_reads))
4012                 wake_up(&conf->wait_for_stripe);
4013         return handled;
4014 }
4015
4016
4017
4018 /*
4019  * This is our raid5 kernel thread.
4020  *
4021  * We scan the hash table for stripes which can be handled now.
4022  * During the scan, completed stripes are saved for us by the interrupt
4023  * handler, so that they will not have to wait for our next wakeup.
4024  */
4025 static void raid5d(mddev_t *mddev)
4026 {
4027         struct stripe_head *sh;
4028         raid5_conf_t *conf = mddev_to_conf(mddev);
4029         int handled;
4030
4031         pr_debug("+++ raid5d active\n");
4032
4033         md_check_recovery(mddev);
4034
4035         handled = 0;
4036         spin_lock_irq(&conf->device_lock);
4037         while (1) {
4038                 struct bio *bio;
4039
4040                 if (conf->seq_flush != conf->seq_write) {
4041                         int seq = conf->seq_flush;
4042                         spin_unlock_irq(&conf->device_lock);
4043                         bitmap_unplug(mddev->bitmap);
4044                         spin_lock_irq(&conf->device_lock);
4045                         conf->seq_write = seq;
4046                         activate_bit_delay(conf);
4047                 }
4048
4049                 while ((bio = remove_bio_from_retry(conf))) {
4050                         int ok;
4051                         spin_unlock_irq(&conf->device_lock);
4052                         ok = retry_aligned_read(conf, bio);
4053                         spin_lock_irq(&conf->device_lock);
4054                         if (!ok)
4055                                 break;
4056                         handled++;
4057                 }
4058
4059                 sh = __get_priority_stripe(conf);
4060
4061                 if (!sh) {
4062                         async_tx_issue_pending_all();
4063                         break;
4064                 }
4065                 spin_unlock_irq(&conf->device_lock);
4066                 
4067                 handled++;
4068                 handle_stripe(sh, conf->spare_page);
4069                 release_stripe(sh);
4070
4071                 spin_lock_irq(&conf->device_lock);
4072         }
4073         pr_debug("%d stripes handled\n", handled);
4074
4075         spin_unlock_irq(&conf->device_lock);
4076
4077         unplug_slaves(mddev);
4078
4079         pr_debug("--- raid5d inactive\n");
4080 }
4081
4082 static ssize_t
4083 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4084 {
4085         raid5_conf_t *conf = mddev_to_conf(mddev);
4086         if (conf)
4087                 return sprintf(page, "%d\n", conf->max_nr_stripes);
4088         else
4089                 return 0;
4090 }
4091
4092 static ssize_t
4093 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4094 {
4095         raid5_conf_t *conf = mddev_to_conf(mddev);
4096         unsigned long new;
4097         if (len >= PAGE_SIZE)
4098                 return -EINVAL;
4099         if (!conf)
4100                 return -ENODEV;
4101
4102         if (strict_strtoul(page, 10, &new))
4103                 return -EINVAL;
4104         if (new <= 16 || new > 32768)
4105                 return -EINVAL;
4106         while (new < conf->max_nr_stripes) {
4107                 if (drop_one_stripe(conf))
4108                         conf->max_nr_stripes--;
4109                 else
4110                         break;
4111         }
4112         md_allow_write(mddev);
4113         while (new > conf->max_nr_stripes) {
4114                 if (grow_one_stripe(conf))
4115                         conf->max_nr_stripes++;
4116                 else break;
4117         }
4118         return len;
4119 }
4120
4121 static struct md_sysfs_entry
4122 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4123                                 raid5_show_stripe_cache_size,
4124                                 raid5_store_stripe_cache_size);
4125
4126 static ssize_t
4127 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4128 {
4129         raid5_conf_t *conf = mddev_to_conf(mddev);
4130         if (conf)
4131                 return sprintf(page, "%d\n", conf->bypass_threshold);
4132         else
4133                 return 0;
4134 }
4135
4136 static ssize_t
4137 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4138 {
4139         raid5_conf_t *conf = mddev_to_conf(mddev);
4140         unsigned long new;
4141         if (len >= PAGE_SIZE)
4142                 return -EINVAL;
4143         if (!conf)
4144                 return -ENODEV;
4145
4146         if (strict_strtoul(page, 10, &new))
4147                 return -EINVAL;
4148         if (new > conf->max_nr_stripes)
4149                 return -EINVAL;
4150         conf->bypass_threshold = new;
4151         return len;
4152 }
4153
4154 static struct md_sysfs_entry
4155 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4156                                         S_IRUGO | S_IWUSR,
4157                                         raid5_show_preread_threshold,
4158                                         raid5_store_preread_threshold);
4159
4160 static ssize_t
4161 stripe_cache_active_show(mddev_t *mddev, char *page)
4162 {
4163         raid5_conf_t *conf = mddev_to_conf(mddev);
4164         if (conf)
4165                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4166         else
4167                 return 0;
4168 }
4169
4170 static struct md_sysfs_entry
4171 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4172
4173 static struct attribute *raid5_attrs[] =  {
4174         &raid5_stripecache_size.attr,
4175         &raid5_stripecache_active.attr,
4176         &raid5_preread_bypass_threshold.attr,
4177         NULL,
4178 };
4179 static struct attribute_group raid5_attrs_group = {
4180         .name = NULL,
4181         .attrs = raid5_attrs,
4182 };
4183
4184 static int run(mddev_t *mddev)
4185 {
4186         raid5_conf_t *conf;
4187         int raid_disk, memory;
4188         mdk_rdev_t *rdev;
4189         struct disk_info *disk;
4190         struct list_head *tmp;
4191         int working_disks = 0;
4192
4193         if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4194                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4195                        mdname(mddev), mddev->level);
4196                 return -EIO;
4197         }
4198
4199         if (mddev->reshape_position != MaxSector) {
4200                 /* Check that we can continue the reshape.
4201                  * Currently only disks can change, it must
4202                  * increase, and we must be past the point where
4203                  * a stripe over-writes itself
4204                  */
4205                 sector_t here_new, here_old;
4206                 int old_disks;
4207                 int max_degraded = (mddev->level == 5 ? 1 : 2);
4208
4209                 if (mddev->new_level != mddev->level ||
4210                     mddev->new_layout != mddev->layout ||
4211                     mddev->new_chunk != mddev->chunk_size) {
4212                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4213                                "required - aborting.\n",
4214                                mdname(mddev));
4215                         return -EINVAL;
4216                 }
4217                 if (mddev->delta_disks <= 0) {
4218                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4219                                "(reduce disks) required - aborting.\n",
4220                                mdname(mddev));
4221                         return -EINVAL;
4222                 }
4223                 old_disks = mddev->raid_disks - mddev->delta_disks;
4224                 /* reshape_position must be on a new-stripe boundary, and one
4225                  * further up in new geometry must map after here in old
4226                  * geometry.
4227                  */
4228                 here_new = mddev->reshape_position;
4229                 if (sector_div(here_new, (mddev->chunk_size>>9)*
4230                                (mddev->raid_disks - max_degraded))) {
4231                         printk(KERN_ERR "raid5: reshape_position not "
4232                                "on a stripe boundary\n");
4233                         return -EINVAL;
4234                 }
4235                 /* here_new is the stripe we will write to */
4236                 here_old = mddev->reshape_position;
4237                 sector_div(here_old, (mddev->chunk_size>>9)*
4238                            (old_disks-max_degraded));
4239                 /* here_old is the first stripe that we might need to read
4240                  * from */
4241                 if (here_new >= here_old) {
4242                         /* Reading from the same stripe as writing to - bad */
4243                         printk(KERN_ERR "raid5: reshape_position too early for "
4244                                "auto-recovery - aborting.\n");
4245                         return -EINVAL;
4246                 }
4247                 printk(KERN_INFO "raid5: reshape will continue\n");
4248                 /* OK, we should be able to continue; */
4249         }
4250
4251
4252         mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4253         if ((conf = mddev->private) == NULL)
4254                 goto abort;
4255         if (mddev->reshape_position == MaxSector) {
4256                 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4257         } else {
4258                 conf->raid_disks = mddev->raid_disks;
4259                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4260         }
4261
4262         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4263                               GFP_KERNEL);
4264         if (!conf->disks)
4265                 goto abort;
4266
4267         conf->mddev = mddev;
4268
4269         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4270                 goto abort;
4271
4272         if (mddev->level == 6) {
4273                 conf->spare_page = alloc_page(GFP_KERNEL);
4274                 if (!conf->spare_page)
4275                         goto abort;
4276         }
4277         spin_lock_init(&conf->device_lock);
4278         mddev->queue->queue_lock = &conf->device_lock;
4279         init_waitqueue_head(&conf->wait_for_stripe);
4280         init_waitqueue_head(&conf->wait_for_overlap);
4281         INIT_LIST_HEAD(&conf->handle_list);
4282         INIT_LIST_HEAD(&conf->hold_list);
4283         INIT_LIST_HEAD(&conf->delayed_list);
4284         INIT_LIST_HEAD(&conf->bitmap_list);
4285         INIT_LIST_HEAD(&conf->inactive_list);
4286         atomic_set(&conf->active_stripes, 0);
4287         atomic_set(&conf->preread_active_stripes, 0);
4288         atomic_set(&conf->active_aligned_reads, 0);
4289         conf->bypass_threshold = BYPASS_THRESHOLD;
4290
4291         pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4292
4293         rdev_for_each(rdev, tmp, mddev) {
4294                 raid_disk = rdev->raid_disk;
4295                 if (raid_disk >= conf->raid_disks
4296                     || raid_disk < 0)
4297                         continue;
4298                 disk = conf->disks + raid_disk;
4299
4300                 disk->rdev = rdev;
4301
4302                 if (test_bit(In_sync, &rdev->flags)) {
4303                         char b[BDEVNAME_SIZE];
4304                         printk(KERN_INFO "raid5: device %s operational as raid"
4305                                 " disk %d\n", bdevname(rdev->bdev,b),
4306                                 raid_disk);
4307                         working_disks++;
4308                 } else
4309                         /* Cannot rely on bitmap to complete recovery */
4310                         conf->fullsync = 1;
4311         }
4312
4313         /*
4314          * 0 for a fully functional array, 1 or 2 for a degraded array.
4315          */
4316         mddev->degraded = conf->raid_disks - working_disks;
4317         conf->mddev = mddev;
4318         conf->chunk_size = mddev->chunk_size;
4319         conf->level = mddev->level;
4320         if (conf->level == 6)
4321                 conf->max_degraded = 2;
4322         else
4323                 conf->max_degraded = 1;
4324         conf->algorithm = mddev->layout;
4325         conf->max_nr_stripes = NR_STRIPES;
4326         conf->expand_progress = mddev->reshape_position;
4327
4328         /* device size must be a multiple of chunk size */
4329         mddev->size &= ~(mddev->chunk_size/1024 -1);
4330         mddev->resync_max_sectors = mddev->size << 1;
4331
4332         if (conf->level == 6 && conf->raid_disks < 4) {
4333                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4334                        mdname(mddev), conf->raid_disks);
4335                 goto abort;
4336         }
4337         if (!conf->chunk_size || conf->chunk_size % 4) {
4338                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4339                         conf->chunk_size, mdname(mddev));
4340                 goto abort;
4341         }
4342         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4343                 printk(KERN_ERR 
4344                         "raid5: unsupported parity algorithm %d for %s\n",
4345                         conf->algorithm, mdname(mddev));
4346                 goto abort;
4347         }
4348         if (mddev->degraded > conf->max_degraded) {
4349                 printk(KERN_ERR "raid5: not enough operational devices for %s"
4350                         " (%d/%d failed)\n",
4351                         mdname(mddev), mddev->degraded, conf->raid_disks);
4352                 goto abort;
4353         }
4354
4355         if (mddev->degraded > 0 &&
4356             mddev->recovery_cp != MaxSector) {
4357                 if (mddev->ok_start_degraded)
4358                         printk(KERN_WARNING
4359                                "raid5: starting dirty degraded array: %s"
4360                                "- data corruption possible.\n",
4361                                mdname(mddev));
4362                 else {
4363                         printk(KERN_ERR
4364                                "raid5: cannot start dirty degraded array for %s\n",
4365                                mdname(mddev));
4366                         goto abort;
4367                 }
4368         }
4369
4370         {
4371                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4372                 if (!mddev->thread) {
4373                         printk(KERN_ERR 
4374                                 "raid5: couldn't allocate thread for %s\n",
4375                                 mdname(mddev));
4376                         goto abort;
4377                 }
4378         }
4379         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4380                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4381         if (grow_stripes(conf, conf->max_nr_stripes)) {
4382                 printk(KERN_ERR 
4383                         "raid5: couldn't allocate %dkB for buffers\n", memory);
4384                 shrink_stripes(conf);
4385                 md_unregister_thread(mddev->thread);
4386                 goto abort;
4387         } else
4388                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4389                         memory, mdname(mddev));
4390
4391         if (mddev->degraded == 0)
4392                 printk("raid5: raid level %d set %s active with %d out of %d"
4393                         " devices, algorithm %d\n", conf->level, mdname(mddev), 
4394                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4395                         conf->algorithm);
4396         else
4397                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4398                         " out of %d devices, algorithm %d\n", conf->level,
4399                         mdname(mddev), mddev->raid_disks - mddev->degraded,
4400                         mddev->raid_disks, conf->algorithm);
4401
4402         print_raid5_conf(conf);
4403
4404         if (conf->expand_progress != MaxSector) {
4405                 printk("...ok start reshape thread\n");
4406                 conf->expand_lo = conf->expand_progress;
4407                 atomic_set(&conf->reshape_stripes, 0);
4408                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4409                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4410                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4411                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4412                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4413                                                         "%s_reshape");
4414         }
4415
4416         /* read-ahead size must cover two whole stripes, which is
4417          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4418          */
4419         {
4420                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4421                 int stripe = data_disks *
4422                         (mddev->chunk_size / PAGE_SIZE);
4423                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4424                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4425         }
4426
4427         /* Ok, everything is just fine now */
4428         if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4429                 printk(KERN_WARNING
4430                        "raid5: failed to create sysfs attributes for %s\n",
4431                        mdname(mddev));
4432
4433         mddev->queue->unplug_fn = raid5_unplug_device;
4434         mddev->queue->backing_dev_info.congested_data = mddev;
4435         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4436
4437         mddev->array_size =  mddev->size * (conf->previous_raid_disks -
4438                                             conf->max_degraded);
4439
4440         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4441
4442         return 0;
4443 abort:
4444         if (conf) {
4445                 print_raid5_conf(conf);
4446                 safe_put_page(conf->spare_page);
4447                 kfree(conf->disks);
4448                 kfree(conf->stripe_hashtbl);
4449                 kfree(conf);
4450         }
4451         mddev->private = NULL;
4452         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4453         return -EIO;
4454 }
4455
4456
4457
4458 static int stop(mddev_t *mddev)
4459 {
4460         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4461
4462         md_unregister_thread(mddev->thread);
4463         mddev->thread = NULL;
4464         shrink_stripes(conf);
4465         kfree(conf->stripe_hashtbl);
4466         mddev->queue->backing_dev_info.congested_fn = NULL;
4467         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4468         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4469         kfree(conf->disks);
4470         kfree(conf);
4471         mddev->private = NULL;
4472         return 0;
4473 }
4474
4475 #ifdef DEBUG
4476 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4477 {
4478         int i;
4479
4480         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4481                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4482         seq_printf(seq, "sh %llu,  count %d.\n",
4483                    (unsigned long long)sh->sector, atomic_read(&sh->count));
4484         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4485         for (i = 0; i < sh->disks; i++) {
4486                 seq_printf(seq, "(cache%d: %p %ld) ",
4487                            i, sh->dev[i].page, sh->dev[i].flags);
4488         }
4489         seq_printf(seq, "\n");
4490 }
4491
4492 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4493 {
4494         struct stripe_head *sh;
4495         struct hlist_node *hn;
4496         int i;
4497
4498         spin_lock_irq(&conf->device_lock);
4499         for (i = 0; i < NR_HASH; i++) {
4500                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4501                         if (sh->raid_conf != conf)
4502                                 continue;
4503                         print_sh(seq, sh);
4504                 }
4505         }
4506         spin_unlock_irq(&conf->device_lock);
4507 }
4508 #endif
4509
4510 static void status (struct seq_file *seq, mddev_t *mddev)
4511 {
4512         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4513         int i;
4514
4515         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4516         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4517         for (i = 0; i < conf->raid_disks; i++)
4518                 seq_printf (seq, "%s",
4519                                conf->disks[i].rdev &&
4520                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4521         seq_printf (seq, "]");
4522 #ifdef DEBUG
4523         seq_printf (seq, "\n");
4524         printall(seq, conf);
4525 #endif
4526 }
4527
4528 static void print_raid5_conf (raid5_conf_t *conf)
4529 {
4530         int i;
4531         struct disk_info *tmp;
4532
4533         printk("RAID5 conf printout:\n");
4534         if (!conf) {
4535                 printk("(conf==NULL)\n");
4536                 return;
4537         }
4538         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4539                  conf->raid_disks - conf->mddev->degraded);
4540
4541         for (i = 0; i < conf->raid_disks; i++) {
4542                 char b[BDEVNAME_SIZE];
4543                 tmp = conf->disks + i;
4544                 if (tmp->rdev)
4545                 printk(" disk %d, o:%d, dev:%s\n",
4546                         i, !test_bit(Faulty, &tmp->rdev->flags),
4547                         bdevname(tmp->rdev->bdev,b));
4548         }
4549 }
4550
4551 static int raid5_spare_active(mddev_t *mddev)
4552 {
4553         int i;
4554         raid5_conf_t *conf = mddev->private;
4555         struct disk_info *tmp;
4556
4557         for (i = 0; i < conf->raid_disks; i++) {
4558                 tmp = conf->disks + i;
4559                 if (tmp->rdev
4560                     && !test_bit(Faulty, &tmp->rdev->flags)
4561                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4562                         unsigned long flags;
4563                         spin_lock_irqsave(&conf->device_lock, flags);
4564                         mddev->degraded--;
4565                         spin_unlock_irqrestore(&conf->device_lock, flags);
4566                 }
4567         }
4568         print_raid5_conf(conf);
4569         return 0;
4570 }
4571
4572 static int raid5_remove_disk(mddev_t *mddev, int number)
4573 {
4574         raid5_conf_t *conf = mddev->private;
4575         int err = 0;
4576         mdk_rdev_t *rdev;
4577         struct disk_info *p = conf->disks + number;
4578
4579         print_raid5_conf(conf);
4580         rdev = p->rdev;
4581         if (rdev) {
4582                 if (test_bit(In_sync, &rdev->flags) ||
4583                     atomic_read(&rdev->nr_pending)) {
4584                         err = -EBUSY;
4585                         goto abort;
4586                 }
4587                 /* Only remove non-faulty devices if recovery
4588                  * isn't possible.
4589                  */
4590                 if (!test_bit(Faulty, &rdev->flags) &&
4591                     mddev->degraded <= conf->max_degraded) {
4592                         err = -EBUSY;
4593                         goto abort;
4594                 }
4595                 p->rdev = NULL;
4596                 synchronize_rcu();
4597                 if (atomic_read(&rdev->nr_pending)) {
4598                         /* lost the race, try later */
4599                         err = -EBUSY;
4600                         p->rdev = rdev;
4601                 }
4602         }
4603 abort:
4604
4605         print_raid5_conf(conf);
4606         return err;
4607 }
4608
4609 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4610 {
4611         raid5_conf_t *conf = mddev->private;
4612         int found = 0;
4613         int disk;
4614         struct disk_info *p;
4615
4616         if (mddev->degraded > conf->max_degraded)
4617                 /* no point adding a device */
4618                 return 0;
4619
4620         /*
4621          * find the disk ... but prefer rdev->saved_raid_disk
4622          * if possible.
4623          */
4624         if (rdev->saved_raid_disk >= 0 &&
4625             conf->disks[rdev->saved_raid_disk].rdev == NULL)
4626                 disk = rdev->saved_raid_disk;
4627         else
4628                 disk = 0;
4629         for ( ; disk < conf->raid_disks; disk++)
4630                 if ((p=conf->disks + disk)->rdev == NULL) {
4631                         clear_bit(In_sync, &rdev->flags);
4632                         rdev->raid_disk = disk;
4633                         found = 1;
4634                         if (rdev->saved_raid_disk != disk)
4635                                 conf->fullsync = 1;
4636                         rcu_assign_pointer(p->rdev, rdev);
4637                         break;
4638                 }
4639         print_raid5_conf(conf);
4640         return found;
4641 }
4642
4643 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4644 {
4645         /* no resync is happening, and there is enough space
4646          * on all devices, so we can resize.
4647          * We need to make sure resync covers any new space.
4648          * If the array is shrinking we should possibly wait until
4649          * any io in the removed space completes, but it hardly seems
4650          * worth it.
4651          */
4652         raid5_conf_t *conf = mddev_to_conf(mddev);
4653
4654         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4655         mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4656         set_capacity(mddev->gendisk, mddev->array_size << 1);
4657         mddev->changed = 1;
4658         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
4659                 mddev->recovery_cp = mddev->size << 1;
4660                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4661         }
4662         mddev->size = sectors /2;
4663         mddev->resync_max_sectors = sectors;
4664         return 0;
4665 }
4666
4667 #ifdef CONFIG_MD_RAID5_RESHAPE
4668 static int raid5_check_reshape(mddev_t *mddev)
4669 {
4670         raid5_conf_t *conf = mddev_to_conf(mddev);
4671         int err;
4672
4673         if (mddev->delta_disks < 0 ||
4674             mddev->new_level != mddev->level)
4675                 return -EINVAL; /* Cannot shrink array or change level yet */
4676         if (mddev->delta_disks == 0)
4677                 return 0; /* nothing to do */
4678
4679         /* Can only proceed if there are plenty of stripe_heads.
4680          * We need a minimum of one full stripe,, and for sensible progress
4681          * it is best to have about 4 times that.
4682          * If we require 4 times, then the default 256 4K stripe_heads will
4683          * allow for chunk sizes up to 256K, which is probably OK.
4684          * If the chunk size is greater, user-space should request more
4685          * stripe_heads first.
4686          */
4687         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4688             (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4689                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
4690                        (mddev->chunk_size / STRIPE_SIZE)*4);
4691                 return -ENOSPC;
4692         }
4693
4694         err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4695         if (err)
4696                 return err;
4697
4698         if (mddev->degraded > conf->max_degraded)
4699                 return -EINVAL;
4700         /* looks like we might be able to manage this */
4701         return 0;
4702 }
4703
4704 static int raid5_start_reshape(mddev_t *mddev)
4705 {
4706         raid5_conf_t *conf = mddev_to_conf(mddev);
4707         mdk_rdev_t *rdev;
4708         struct list_head *rtmp;
4709         int spares = 0;
4710         int added_devices = 0;
4711         unsigned long flags;
4712
4713         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4714                 return -EBUSY;
4715
4716         rdev_for_each(rdev, rtmp, mddev)
4717                 if (rdev->raid_disk < 0 &&
4718                     !test_bit(Faulty, &rdev->flags))
4719                         spares++;
4720
4721         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4722                 /* Not enough devices even to make a degraded array
4723                  * of that size
4724                  */
4725                 return -EINVAL;
4726
4727         atomic_set(&conf->reshape_stripes, 0);
4728         spin_lock_irq(&conf->device_lock);
4729         conf->previous_raid_disks = conf->raid_disks;
4730         conf->raid_disks += mddev->delta_disks;
4731         conf->expand_progress = 0;
4732         conf->expand_lo = 0;
4733         spin_unlock_irq(&conf->device_lock);
4734
4735         /* Add some new drives, as many as will fit.
4736          * We know there are enough to make the newly sized array work.
4737          */
4738         rdev_for_each(rdev, rtmp, mddev)
4739                 if (rdev->raid_disk < 0 &&
4740                     !test_bit(Faulty, &rdev->flags)) {
4741                         if (raid5_add_disk(mddev, rdev)) {
4742                                 char nm[20];
4743                                 set_bit(In_sync, &rdev->flags);
4744                                 added_devices++;
4745                                 rdev->recovery_offset = 0;
4746                                 sprintf(nm, "rd%d", rdev->raid_disk);
4747                                 if (sysfs_create_link(&mddev->kobj,
4748                                                       &rdev->kobj, nm))
4749                                         printk(KERN_WARNING
4750                                                "raid5: failed to create "
4751                                                " link %s for %s\n",
4752                                                nm, mdname(mddev));
4753                         } else
4754                                 break;
4755                 }
4756
4757         spin_lock_irqsave(&conf->device_lock, flags);
4758         mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4759         spin_unlock_irqrestore(&conf->device_lock, flags);
4760         mddev->raid_disks = conf->raid_disks;
4761         mddev->reshape_position = 0;
4762         set_bit(MD_CHANGE_DEVS, &mddev->flags);
4763
4764         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4765         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4766         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4767         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4768         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4769                                                 "%s_reshape");
4770         if (!mddev->sync_thread) {
4771                 mddev->recovery = 0;
4772                 spin_lock_irq(&conf->device_lock);
4773                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4774                 conf->expand_progress = MaxSector;
4775                 spin_unlock_irq(&conf->device_lock);
4776                 return -EAGAIN;
4777         }
4778         md_wakeup_thread(mddev->sync_thread);
4779         md_new_event(mddev);
4780         return 0;
4781 }
4782 #endif
4783
4784 static void end_reshape(raid5_conf_t *conf)
4785 {
4786         struct block_device *bdev;
4787
4788         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4789                 conf->mddev->array_size = conf->mddev->size *
4790                         (conf->raid_disks - conf->max_degraded);
4791                 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4792                 conf->mddev->changed = 1;
4793
4794                 bdev = bdget_disk(conf->mddev->gendisk, 0);
4795                 if (bdev) {
4796                         mutex_lock(&bdev->bd_inode->i_mutex);
4797                         i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4798                         mutex_unlock(&bdev->bd_inode->i_mutex);
4799                         bdput(bdev);
4800                 }
4801                 spin_lock_irq(&conf->device_lock);
4802                 conf->expand_progress = MaxSector;
4803                 spin_unlock_irq(&conf->device_lock);
4804                 conf->mddev->reshape_position = MaxSector;
4805
4806                 /* read-ahead size must cover two whole stripes, which is
4807                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4808                  */
4809                 {
4810                         int data_disks = conf->previous_raid_disks - conf->max_degraded;
4811                         int stripe = data_disks *
4812                                 (conf->mddev->chunk_size / PAGE_SIZE);
4813                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4814                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4815                 }
4816         }
4817 }
4818
4819 static void raid5_quiesce(mddev_t *mddev, int state)
4820 {
4821         raid5_conf_t *conf = mddev_to_conf(mddev);
4822
4823         switch(state) {
4824         case 2: /* resume for a suspend */
4825                 wake_up(&conf->wait_for_overlap);
4826                 break;
4827
4828         case 1: /* stop all writes */
4829                 spin_lock_irq(&conf->device_lock);
4830                 conf->quiesce = 1;
4831                 wait_event_lock_irq(conf->wait_for_stripe,
4832                                     atomic_read(&conf->active_stripes) == 0 &&
4833                                     atomic_read(&conf->active_aligned_reads) == 0,
4834                                     conf->device_lock, /* nothing */);
4835                 spin_unlock_irq(&conf->device_lock);
4836                 break;
4837
4838         case 0: /* re-enable writes */
4839                 spin_lock_irq(&conf->device_lock);
4840                 conf->quiesce = 0;
4841                 wake_up(&conf->wait_for_stripe);
4842                 wake_up(&conf->wait_for_overlap);
4843                 spin_unlock_irq(&conf->device_lock);
4844                 break;
4845         }
4846 }
4847
4848 static struct mdk_personality raid6_personality =
4849 {
4850         .name           = "raid6",
4851         .level          = 6,
4852         .owner          = THIS_MODULE,
4853         .make_request   = make_request,
4854         .run            = run,
4855         .stop           = stop,
4856         .status         = status,
4857         .error_handler  = error,
4858         .hot_add_disk   = raid5_add_disk,
4859         .hot_remove_disk= raid5_remove_disk,
4860         .spare_active   = raid5_spare_active,
4861         .sync_request   = sync_request,
4862         .resize         = raid5_resize,
4863 #ifdef CONFIG_MD_RAID5_RESHAPE
4864         .check_reshape  = raid5_check_reshape,
4865         .start_reshape  = raid5_start_reshape,
4866 #endif
4867         .quiesce        = raid5_quiesce,
4868 };
4869 static struct mdk_personality raid5_personality =
4870 {
4871         .name           = "raid5",
4872         .level          = 5,
4873         .owner          = THIS_MODULE,
4874         .make_request   = make_request,
4875         .run            = run,
4876         .stop           = stop,
4877         .status         = status,
4878         .error_handler  = error,
4879         .hot_add_disk   = raid5_add_disk,
4880         .hot_remove_disk= raid5_remove_disk,
4881         .spare_active   = raid5_spare_active,
4882         .sync_request   = sync_request,
4883         .resize         = raid5_resize,
4884 #ifdef CONFIG_MD_RAID5_RESHAPE
4885         .check_reshape  = raid5_check_reshape,
4886         .start_reshape  = raid5_start_reshape,
4887 #endif
4888         .quiesce        = raid5_quiesce,
4889 };
4890
4891 static struct mdk_personality raid4_personality =
4892 {
4893         .name           = "raid4",
4894         .level          = 4,
4895         .owner          = THIS_MODULE,
4896         .make_request   = make_request,
4897         .run            = run,
4898         .stop           = stop,
4899         .status         = status,
4900         .error_handler  = error,
4901         .hot_add_disk   = raid5_add_disk,
4902         .hot_remove_disk= raid5_remove_disk,
4903         .spare_active   = raid5_spare_active,
4904         .sync_request   = sync_request,
4905         .resize         = raid5_resize,
4906 #ifdef CONFIG_MD_RAID5_RESHAPE
4907         .check_reshape  = raid5_check_reshape,
4908         .start_reshape  = raid5_start_reshape,
4909 #endif
4910         .quiesce        = raid5_quiesce,
4911 };
4912
4913 static int __init raid5_init(void)
4914 {
4915         int e;
4916
4917         e = raid6_select_algo();
4918         if ( e )
4919                 return e;
4920         register_md_personality(&raid6_personality);
4921         register_md_personality(&raid5_personality);
4922         register_md_personality(&raid4_personality);
4923         return 0;
4924 }
4925
4926 static void raid5_exit(void)
4927 {
4928         unregister_md_personality(&raid6_personality);
4929         unregister_md_personality(&raid5_personality);
4930         unregister_md_personality(&raid4_personality);
4931 }
4932
4933 module_init(raid5_init);
4934 module_exit(raid5_exit);
4935 MODULE_LICENSE("GPL");
4936 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4937 MODULE_ALIAS("md-raid5");
4938 MODULE_ALIAS("md-raid4");
4939 MODULE_ALIAS("md-level-5");
4940 MODULE_ALIAS("md-level-4");
4941 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4942 MODULE_ALIAS("md-raid6");
4943 MODULE_ALIAS("md-level-6");
4944
4945 /* This used to be two separate modules, they were: */
4946 MODULE_ALIAS("raid5");
4947 MODULE_ALIAS("raid6");