Add bvec_merge_data to handle stacked devices and ->merge_bvec()
[linux-2.6.git] / drivers / md / raid5.c
1 /*
2  * raid5.c : Multiple Devices driver for Linux
3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4  *         Copyright (C) 1999, 2000 Ingo Molnar
5  *         Copyright (C) 2002, 2003 H. Peter Anvin
6  *
7  * RAID-4/5/6 management functions.
8  * Thanks to Penguin Computing for making the RAID-6 development possible
9  * by donating a test server!
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 /*
22  * BITMAP UNPLUGGING:
23  *
24  * The sequencing for updating the bitmap reliably is a little
25  * subtle (and I got it wrong the first time) so it deserves some
26  * explanation.
27  *
28  * We group bitmap updates into batches.  Each batch has a number.
29  * We may write out several batches at once, but that isn't very important.
30  * conf->bm_write is the number of the last batch successfully written.
31  * conf->bm_flush is the number of the last batch that was closed to
32  *    new additions.
33  * When we discover that we will need to write to any block in a stripe
34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35  * the number of the batch it will be in. This is bm_flush+1.
36  * When we are ready to do a write, if that batch hasn't been written yet,
37  *   we plug the array and queue the stripe for later.
38  * When an unplug happens, we increment bm_flush, thus closing the current
39  *   batch.
40  * When we notice that bm_flush > bm_write, we write out all pending updates
41  * to the bitmap, and advance bm_write to where bm_flush was.
42  * This may occasionally write a bit out twice, but is sure never to
43  * miss any bits.
44  */
45
46 #include <linux/module.h>
47 #include <linux/slab.h>
48 #include <linux/highmem.h>
49 #include <linux/bitops.h>
50 #include <linux/kthread.h>
51 #include <asm/atomic.h>
52 #include "raid6.h"
53
54 #include <linux/raid/bitmap.h>
55 #include <linux/async_tx.h>
56
57 /*
58  * Stripe cache
59  */
60
61 #define NR_STRIPES              256
62 #define STRIPE_SIZE             PAGE_SIZE
63 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
64 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
65 #define IO_THRESHOLD            1
66 #define 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 ((s->uptodate < disks - 1) &&
2021                         test_bit(R5_Insync, &dev->flags)) {
2022                         /* Note: we hold off compute operations while checks are
2023                          * in flight, but we still prefer 'compute' over 'read'
2024                          * hence we only read if (uptodate < * disks-1)
2025                          */
2026                         set_bit(R5_LOCKED, &dev->flags);
2027                         set_bit(R5_Wantread, &dev->flags);
2028                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2029                                 sh->ops.count++;
2030                         s->locked++;
2031                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2032                                 s->syncing);
2033                 }
2034         }
2035
2036         return ~0;
2037 }
2038
2039 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2040                         struct stripe_head_state *s, int disks)
2041 {
2042         int i;
2043
2044         /* Clear completed compute operations.  Parity recovery
2045          * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled
2046          * later on in this routine
2047          */
2048         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2049                 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2050                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2051                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2052                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2053         }
2054
2055         /* look for blocks to read/compute, skip this if a compute
2056          * is already in flight, or if the stripe contents are in the
2057          * midst of changing due to a write
2058          */
2059         if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2060                 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2061                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2062                 for (i = disks; i--; )
2063                         if (__handle_issuing_new_read_requests5(
2064                                 sh, s, i, disks) == 0)
2065                                 break;
2066         }
2067         set_bit(STRIPE_HANDLE, &sh->state);
2068 }
2069
2070 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2071                         struct stripe_head_state *s, struct r6_state *r6s,
2072                         int disks)
2073 {
2074         int i;
2075         for (i = disks; i--; ) {
2076                 struct r5dev *dev = &sh->dev[i];
2077                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2078                     !test_bit(R5_UPTODATE, &dev->flags) &&
2079                     (dev->toread || (dev->towrite &&
2080                      !test_bit(R5_OVERWRITE, &dev->flags)) ||
2081                      s->syncing || s->expanding ||
2082                      (s->failed >= 1 &&
2083                       (sh->dev[r6s->failed_num[0]].toread ||
2084                        s->to_write)) ||
2085                      (s->failed >= 2 &&
2086                       (sh->dev[r6s->failed_num[1]].toread ||
2087                        s->to_write)))) {
2088                         /* we would like to get this block, possibly
2089                          * by computing it, but we might not be able to
2090                          */
2091                         if ((s->uptodate == disks - 1) &&
2092                             (s->failed && (i == r6s->failed_num[0] ||
2093                                            i == r6s->failed_num[1]))) {
2094                                 pr_debug("Computing stripe %llu block %d\n",
2095                                        (unsigned long long)sh->sector, i);
2096                                 compute_block_1(sh, i, 0);
2097                                 s->uptodate++;
2098                         } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2099                                 /* Computing 2-failure is *very* expensive; only
2100                                  * do it if failed >= 2
2101                                  */
2102                                 int other;
2103                                 for (other = disks; other--; ) {
2104                                         if (other == i)
2105                                                 continue;
2106                                         if (!test_bit(R5_UPTODATE,
2107                                               &sh->dev[other].flags))
2108                                                 break;
2109                                 }
2110                                 BUG_ON(other < 0);
2111                                 pr_debug("Computing stripe %llu blocks %d,%d\n",
2112                                        (unsigned long long)sh->sector,
2113                                        i, other);
2114                                 compute_block_2(sh, i, other);
2115                                 s->uptodate += 2;
2116                         } else if (test_bit(R5_Insync, &dev->flags)) {
2117                                 set_bit(R5_LOCKED, &dev->flags);
2118                                 set_bit(R5_Wantread, &dev->flags);
2119                                 s->locked++;
2120                                 pr_debug("Reading block %d (sync=%d)\n",
2121                                         i, s->syncing);
2122                         }
2123                 }
2124         }
2125         set_bit(STRIPE_HANDLE, &sh->state);
2126 }
2127
2128
2129 /* handle_completed_write_requests
2130  * any written block on an uptodate or failed drive can be returned.
2131  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2132  * never LOCKED, so we don't need to test 'failed' directly.
2133  */
2134 static void handle_completed_write_requests(raid5_conf_t *conf,
2135         struct stripe_head *sh, int disks, struct bio **return_bi)
2136 {
2137         int i;
2138         struct r5dev *dev;
2139
2140         for (i = disks; i--; )
2141                 if (sh->dev[i].written) {
2142                         dev = &sh->dev[i];
2143                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2144                                 test_bit(R5_UPTODATE, &dev->flags)) {
2145                                 /* We can return any write requests */
2146                                 struct bio *wbi, *wbi2;
2147                                 int bitmap_end = 0;
2148                                 pr_debug("Return write for disc %d\n", i);
2149                                 spin_lock_irq(&conf->device_lock);
2150                                 wbi = dev->written;
2151                                 dev->written = NULL;
2152                                 while (wbi && wbi->bi_sector <
2153                                         dev->sector + STRIPE_SECTORS) {
2154                                         wbi2 = r5_next_bio(wbi, dev->sector);
2155                                         if (--wbi->bi_phys_segments == 0) {
2156                                                 md_write_end(conf->mddev);
2157                                                 wbi->bi_next = *return_bi;
2158                                                 *return_bi = wbi;
2159                                         }
2160                                         wbi = wbi2;
2161                                 }
2162                                 if (dev->towrite == NULL)
2163                                         bitmap_end = 1;
2164                                 spin_unlock_irq(&conf->device_lock);
2165                                 if (bitmap_end)
2166                                         bitmap_endwrite(conf->mddev->bitmap,
2167                                                         sh->sector,
2168                                                         STRIPE_SECTORS,
2169                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2170                                                         0);
2171                         }
2172                 }
2173
2174         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2175                 if (atomic_dec_and_test(&conf->pending_full_writes))
2176                         md_wakeup_thread(conf->mddev->thread);
2177 }
2178
2179 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2180                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2181 {
2182         int rmw = 0, rcw = 0, i;
2183         for (i = disks; i--; ) {
2184                 /* would I have to read this buffer for read_modify_write */
2185                 struct r5dev *dev = &sh->dev[i];
2186                 if ((dev->towrite || i == sh->pd_idx) &&
2187                     !test_bit(R5_LOCKED, &dev->flags) &&
2188                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2189                       test_bit(R5_Wantcompute, &dev->flags))) {
2190                         if (test_bit(R5_Insync, &dev->flags))
2191                                 rmw++;
2192                         else
2193                                 rmw += 2*disks;  /* cannot read it */
2194                 }
2195                 /* Would I have to read this buffer for reconstruct_write */
2196                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2197                     !test_bit(R5_LOCKED, &dev->flags) &&
2198                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2199                     test_bit(R5_Wantcompute, &dev->flags))) {
2200                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2201                         else
2202                                 rcw += 2*disks;
2203                 }
2204         }
2205         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2206                 (unsigned long long)sh->sector, rmw, rcw);
2207         set_bit(STRIPE_HANDLE, &sh->state);
2208         if (rmw < rcw && rmw > 0)
2209                 /* prefer read-modify-write, but need to get some data */
2210                 for (i = disks; i--; ) {
2211                         struct r5dev *dev = &sh->dev[i];
2212                         if ((dev->towrite || i == sh->pd_idx) &&
2213                             !test_bit(R5_LOCKED, &dev->flags) &&
2214                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2215                             test_bit(R5_Wantcompute, &dev->flags)) &&
2216                             test_bit(R5_Insync, &dev->flags)) {
2217                                 if (
2218                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2219                                         pr_debug("Read_old block "
2220                                                 "%d for r-m-w\n", i);
2221                                         set_bit(R5_LOCKED, &dev->flags);
2222                                         set_bit(R5_Wantread, &dev->flags);
2223                                         if (!test_and_set_bit(
2224                                                 STRIPE_OP_IO, &sh->ops.pending))
2225                                                 sh->ops.count++;
2226                                         s->locked++;
2227                                 } else {
2228                                         set_bit(STRIPE_DELAYED, &sh->state);
2229                                         set_bit(STRIPE_HANDLE, &sh->state);
2230                                 }
2231                         }
2232                 }
2233         if (rcw <= rmw && rcw > 0)
2234                 /* want reconstruct write, but need to get some data */
2235                 for (i = disks; i--; ) {
2236                         struct r5dev *dev = &sh->dev[i];
2237                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2238                             i != sh->pd_idx &&
2239                             !test_bit(R5_LOCKED, &dev->flags) &&
2240                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2241                             test_bit(R5_Wantcompute, &dev->flags)) &&
2242                             test_bit(R5_Insync, &dev->flags)) {
2243                                 if (
2244                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2245                                         pr_debug("Read_old block "
2246                                                 "%d for Reconstruct\n", i);
2247                                         set_bit(R5_LOCKED, &dev->flags);
2248                                         set_bit(R5_Wantread, &dev->flags);
2249                                         if (!test_and_set_bit(
2250                                                 STRIPE_OP_IO, &sh->ops.pending))
2251                                                 sh->ops.count++;
2252                                         s->locked++;
2253                                 } else {
2254                                         set_bit(STRIPE_DELAYED, &sh->state);
2255                                         set_bit(STRIPE_HANDLE, &sh->state);
2256                                 }
2257                         }
2258                 }
2259         /* now if nothing is locked, and if we have enough data,
2260          * we can start a write request
2261          */
2262         /* since handle_stripe can be called at any time we need to handle the
2263          * case where a compute block operation has been submitted and then a
2264          * subsequent call wants to start a write request.  raid5_run_ops only
2265          * handles the case where compute block and postxor are requested
2266          * simultaneously.  If this is not the case then new writes need to be
2267          * held off until the compute completes.
2268          */
2269         if ((s->req_compute ||
2270             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) &&
2271                 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2272                 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2273                 s->locked += handle_write_operations5(sh, rcw == 0, 0);
2274 }
2275
2276 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2277                 struct stripe_head *sh, struct stripe_head_state *s,
2278                 struct r6_state *r6s, int disks)
2279 {
2280         int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2281         int qd_idx = r6s->qd_idx;
2282         for (i = disks; i--; ) {
2283                 struct r5dev *dev = &sh->dev[i];
2284                 /* Would I have to read this buffer for reconstruct_write */
2285                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2286                     && i != pd_idx && i != qd_idx
2287                     && (!test_bit(R5_LOCKED, &dev->flags)
2288                             ) &&
2289                     !test_bit(R5_UPTODATE, &dev->flags)) {
2290                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2291                         else {
2292                                 pr_debug("raid6: must_compute: "
2293                                         "disk %d flags=%#lx\n", i, dev->flags);
2294                                 must_compute++;
2295                         }
2296                 }
2297         }
2298         pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2299                (unsigned long long)sh->sector, rcw, must_compute);
2300         set_bit(STRIPE_HANDLE, &sh->state);
2301
2302         if (rcw > 0)
2303                 /* want reconstruct write, but need to get some data */
2304                 for (i = disks; i--; ) {
2305                         struct r5dev *dev = &sh->dev[i];
2306                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2307                             && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2308                             && !test_bit(R5_LOCKED, &dev->flags) &&
2309                             !test_bit(R5_UPTODATE, &dev->flags) &&
2310                             test_bit(R5_Insync, &dev->flags)) {
2311                                 if (
2312                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2313                                         pr_debug("Read_old stripe %llu "
2314                                                 "block %d for Reconstruct\n",
2315                                              (unsigned long long)sh->sector, i);
2316                                         set_bit(R5_LOCKED, &dev->flags);
2317                                         set_bit(R5_Wantread, &dev->flags);
2318                                         s->locked++;
2319                                 } else {
2320                                         pr_debug("Request delayed stripe %llu "
2321                                                 "block %d for Reconstruct\n",
2322                                              (unsigned long long)sh->sector, i);
2323                                         set_bit(STRIPE_DELAYED, &sh->state);
2324                                         set_bit(STRIPE_HANDLE, &sh->state);
2325                                 }
2326                         }
2327                 }
2328         /* now if nothing is locked, and if we have enough data, we can start a
2329          * write request
2330          */
2331         if (s->locked == 0 && rcw == 0 &&
2332             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2333                 if (must_compute > 0) {
2334                         /* We have failed blocks and need to compute them */
2335                         switch (s->failed) {
2336                         case 0:
2337                                 BUG();
2338                         case 1:
2339                                 compute_block_1(sh, r6s->failed_num[0], 0);
2340                                 break;
2341                         case 2:
2342                                 compute_block_2(sh, r6s->failed_num[0],
2343                                                 r6s->failed_num[1]);
2344                                 break;
2345                         default: /* This request should have been failed? */
2346                                 BUG();
2347                         }
2348                 }
2349
2350                 pr_debug("Computing parity for stripe %llu\n",
2351                         (unsigned long long)sh->sector);
2352                 compute_parity6(sh, RECONSTRUCT_WRITE);
2353                 /* now every locked buffer is ready to be written */
2354                 for (i = disks; i--; )
2355                         if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2356                                 pr_debug("Writing stripe %llu block %d\n",
2357                                        (unsigned long long)sh->sector, i);
2358                                 s->locked++;
2359                                 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2360                         }
2361                 if (s->locked == disks)
2362                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2363                                 atomic_inc(&conf->pending_full_writes);
2364                 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2365                 set_bit(STRIPE_INSYNC, &sh->state);
2366
2367                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2368                         atomic_dec(&conf->preread_active_stripes);
2369                         if (atomic_read(&conf->preread_active_stripes) <
2370                             IO_THRESHOLD)
2371                                 md_wakeup_thread(conf->mddev->thread);
2372                 }
2373         }
2374 }
2375
2376 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2377                                 struct stripe_head_state *s, int disks)
2378 {
2379         int canceled_check = 0;
2380
2381         set_bit(STRIPE_HANDLE, &sh->state);
2382
2383         /* complete a check operation */
2384         if (test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) {
2385                 clear_bit(STRIPE_OP_CHECK, &sh->ops.ack);
2386                 clear_bit(STRIPE_OP_CHECK, &sh->ops.pending);
2387                 if (s->failed == 0) {
2388                         if (sh->ops.zero_sum_result == 0)
2389                                 /* parity is correct (on disc,
2390                                  * not in buffer any more)
2391                                  */
2392                                 set_bit(STRIPE_INSYNC, &sh->state);
2393                         else {
2394                                 conf->mddev->resync_mismatches +=
2395                                         STRIPE_SECTORS;
2396                                 if (test_bit(
2397                                      MD_RECOVERY_CHECK, &conf->mddev->recovery))
2398                                         /* don't try to repair!! */
2399                                         set_bit(STRIPE_INSYNC, &sh->state);
2400                                 else {
2401                                         set_bit(STRIPE_OP_COMPUTE_BLK,
2402                                                 &sh->ops.pending);
2403                                         set_bit(STRIPE_OP_MOD_REPAIR_PD,
2404                                                 &sh->ops.pending);
2405                                         set_bit(R5_Wantcompute,
2406                                                 &sh->dev[sh->pd_idx].flags);
2407                                         sh->ops.target = sh->pd_idx;
2408                                         sh->ops.count++;
2409                                         s->uptodate++;
2410                                 }
2411                         }
2412                 } else
2413                         canceled_check = 1; /* STRIPE_INSYNC is not set */
2414         }
2415
2416         /* start a new check operation if there are no failures, the stripe is
2417          * not insync, and a repair is not in flight
2418          */
2419         if (s->failed == 0 &&
2420             !test_bit(STRIPE_INSYNC, &sh->state) &&
2421             !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2422                 if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2423                         BUG_ON(s->uptodate != disks);
2424                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2425                         sh->ops.count++;
2426                         s->uptodate--;
2427                 }
2428         }
2429
2430         /* check if we can clear a parity disk reconstruct */
2431         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2432             test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2433
2434                 clear_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending);
2435                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2436                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2437                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2438         }
2439
2440
2441         /* Wait for check parity and compute block operations to complete
2442          * before write-back.  If a failure occurred while the check operation
2443          * was in flight we need to cycle this stripe through handle_stripe
2444          * since the parity block may not be uptodate
2445          */
2446         if (!canceled_check && !test_bit(STRIPE_INSYNC, &sh->state) &&
2447             !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) &&
2448             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) {
2449                 struct r5dev *dev;
2450                 /* either failed parity check, or recovery is happening */
2451                 if (s->failed == 0)
2452                         s->failed_num = sh->pd_idx;
2453                 dev = &sh->dev[s->failed_num];
2454                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2455                 BUG_ON(s->uptodate != disks);
2456
2457                 set_bit(R5_LOCKED, &dev->flags);
2458                 set_bit(R5_Wantwrite, &dev->flags);
2459                 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2460                         sh->ops.count++;
2461
2462                 clear_bit(STRIPE_DEGRADED, &sh->state);
2463                 s->locked++;
2464                 set_bit(STRIPE_INSYNC, &sh->state);
2465         }
2466 }
2467
2468
2469 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2470                                 struct stripe_head_state *s,
2471                                 struct r6_state *r6s, struct page *tmp_page,
2472                                 int disks)
2473 {
2474         int update_p = 0, update_q = 0;
2475         struct r5dev *dev;
2476         int pd_idx = sh->pd_idx;
2477         int qd_idx = r6s->qd_idx;
2478
2479         set_bit(STRIPE_HANDLE, &sh->state);
2480
2481         BUG_ON(s->failed > 2);
2482         BUG_ON(s->uptodate < disks);
2483         /* Want to check and possibly repair P and Q.
2484          * However there could be one 'failed' device, in which
2485          * case we can only check one of them, possibly using the
2486          * other to generate missing data
2487          */
2488
2489         /* If !tmp_page, we cannot do the calculations,
2490          * but as we have set STRIPE_HANDLE, we will soon be called
2491          * by stripe_handle with a tmp_page - just wait until then.
2492          */
2493         if (tmp_page) {
2494                 if (s->failed == r6s->q_failed) {
2495                         /* The only possible failed device holds 'Q', so it
2496                          * makes sense to check P (If anything else were failed,
2497                          * we would have used P to recreate it).
2498                          */
2499                         compute_block_1(sh, pd_idx, 1);
2500                         if (!page_is_zero(sh->dev[pd_idx].page)) {
2501                                 compute_block_1(sh, pd_idx, 0);
2502                                 update_p = 1;
2503                         }
2504                 }
2505                 if (!r6s->q_failed && s->failed < 2) {
2506                         /* q is not failed, and we didn't use it to generate
2507                          * anything, so it makes sense to check it
2508                          */
2509                         memcpy(page_address(tmp_page),
2510                                page_address(sh->dev[qd_idx].page),
2511                                STRIPE_SIZE);
2512                         compute_parity6(sh, UPDATE_PARITY);
2513                         if (memcmp(page_address(tmp_page),
2514                                    page_address(sh->dev[qd_idx].page),
2515                                    STRIPE_SIZE) != 0) {
2516                                 clear_bit(STRIPE_INSYNC, &sh->state);
2517                                 update_q = 1;
2518                         }
2519                 }
2520                 if (update_p || update_q) {
2521                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2522                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2523                                 /* don't try to repair!! */
2524                                 update_p = update_q = 0;
2525                 }
2526
2527                 /* now write out any block on a failed drive,
2528                  * or P or Q if they need it
2529                  */
2530
2531                 if (s->failed == 2) {
2532                         dev = &sh->dev[r6s->failed_num[1]];
2533                         s->locked++;
2534                         set_bit(R5_LOCKED, &dev->flags);
2535                         set_bit(R5_Wantwrite, &dev->flags);
2536                 }
2537                 if (s->failed >= 1) {
2538                         dev = &sh->dev[r6s->failed_num[0]];
2539                         s->locked++;
2540                         set_bit(R5_LOCKED, &dev->flags);
2541                         set_bit(R5_Wantwrite, &dev->flags);
2542                 }
2543
2544                 if (update_p) {
2545                         dev = &sh->dev[pd_idx];
2546                         s->locked++;
2547                         set_bit(R5_LOCKED, &dev->flags);
2548                         set_bit(R5_Wantwrite, &dev->flags);
2549                 }
2550                 if (update_q) {
2551                         dev = &sh->dev[qd_idx];
2552                         s->locked++;
2553                         set_bit(R5_LOCKED, &dev->flags);
2554                         set_bit(R5_Wantwrite, &dev->flags);
2555                 }
2556                 clear_bit(STRIPE_DEGRADED, &sh->state);
2557
2558                 set_bit(STRIPE_INSYNC, &sh->state);
2559         }
2560 }
2561
2562 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2563                                 struct r6_state *r6s)
2564 {
2565         int i;
2566
2567         /* We have read all the blocks in this stripe and now we need to
2568          * copy some of them into a target stripe for expand.
2569          */
2570         struct dma_async_tx_descriptor *tx = NULL;
2571         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2572         for (i = 0; i < sh->disks; i++)
2573                 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2574                         int dd_idx, pd_idx, j;
2575                         struct stripe_head *sh2;
2576
2577                         sector_t bn = compute_blocknr(sh, i);
2578                         sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2579                                                 conf->raid_disks -
2580                                                 conf->max_degraded, &dd_idx,
2581                                                 &pd_idx, conf);
2582                         sh2 = get_active_stripe(conf, s, conf->raid_disks,
2583                                                 pd_idx, 1);
2584                         if (sh2 == NULL)
2585                                 /* so far only the early blocks of this stripe
2586                                  * have been requested.  When later blocks
2587                                  * get requested, we will try again
2588                                  */
2589                                 continue;
2590                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2591                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2592                                 /* must have already done this block */
2593                                 release_stripe(sh2);
2594                                 continue;
2595                         }
2596
2597                         /* place all the copies on one channel */
2598                         tx = async_memcpy(sh2->dev[dd_idx].page,
2599                                 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2600                                 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2601
2602                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2603                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2604                         for (j = 0; j < conf->raid_disks; j++)
2605                                 if (j != sh2->pd_idx &&
2606                                     (!r6s || j != raid6_next_disk(sh2->pd_idx,
2607                                                                  sh2->disks)) &&
2608                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2609                                         break;
2610                         if (j == conf->raid_disks) {
2611                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2612                                 set_bit(STRIPE_HANDLE, &sh2->state);
2613                         }
2614                         release_stripe(sh2);
2615
2616                 }
2617         /* done submitting copies, wait for them to complete */
2618         if (tx) {
2619                 async_tx_ack(tx);
2620                 dma_wait_for_async_tx(tx);
2621         }
2622 }
2623
2624
2625 /*
2626  * handle_stripe - do things to a stripe.
2627  *
2628  * We lock the stripe and then examine the state of various bits
2629  * to see what needs to be done.
2630  * Possible results:
2631  *    return some read request which now have data
2632  *    return some write requests which are safely on disc
2633  *    schedule a read on some buffers
2634  *    schedule a write of some buffers
2635  *    return confirmation of parity correctness
2636  *
2637  * buffers are taken off read_list or write_list, and bh_cache buffers
2638  * get BH_Lock set before the stripe lock is released.
2639  *
2640  */
2641
2642 static void handle_stripe5(struct stripe_head *sh)
2643 {
2644         raid5_conf_t *conf = sh->raid_conf;
2645         int disks = sh->disks, i;
2646         struct bio *return_bi = NULL;
2647         struct stripe_head_state s;
2648         struct r5dev *dev;
2649         unsigned long pending = 0;
2650         mdk_rdev_t *blocked_rdev = NULL;
2651         int prexor;
2652
2653         memset(&s, 0, sizeof(s));
2654         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2655                 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2656                 atomic_read(&sh->count), sh->pd_idx,
2657                 sh->ops.pending, sh->ops.ack, sh->ops.complete);
2658
2659         spin_lock(&sh->lock);
2660         clear_bit(STRIPE_HANDLE, &sh->state);
2661         clear_bit(STRIPE_DELAYED, &sh->state);
2662
2663         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2664         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2665         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2666         /* Now to look around and see what can be done */
2667
2668         /* clean-up completed biofill operations */
2669         if (test_bit(STRIPE_OP_BIOFILL, &sh->ops.complete)) {
2670                 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending);
2671                 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack);
2672                 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
2673         }
2674
2675         rcu_read_lock();
2676         for (i=disks; i--; ) {
2677                 mdk_rdev_t *rdev;
2678                 struct r5dev *dev = &sh->dev[i];
2679                 clear_bit(R5_Insync, &dev->flags);
2680
2681                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2682                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2683                         dev->towrite, dev->written);
2684
2685                 /* maybe we can request a biofill operation
2686                  *
2687                  * new wantfill requests are only permitted while
2688                  * STRIPE_OP_BIOFILL is clear
2689                  */
2690                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2691                         !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2692                         set_bit(R5_Wantfill, &dev->flags);
2693
2694                 /* now count some things */
2695                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2696                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2697                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2698
2699                 if (test_bit(R5_Wantfill, &dev->flags))
2700                         s.to_fill++;
2701                 else if (dev->toread)
2702                         s.to_read++;
2703                 if (dev->towrite) {
2704                         s.to_write++;
2705                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2706                                 s.non_overwrite++;
2707                 }
2708                 if (dev->written)
2709                         s.written++;
2710                 rdev = rcu_dereference(conf->disks[i].rdev);
2711                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2712                         blocked_rdev = rdev;
2713                         atomic_inc(&rdev->nr_pending);
2714                         break;
2715                 }
2716                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2717                         /* The ReadError flag will just be confusing now */
2718                         clear_bit(R5_ReadError, &dev->flags);
2719                         clear_bit(R5_ReWrite, &dev->flags);
2720                 }
2721                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2722                     || test_bit(R5_ReadError, &dev->flags)) {
2723                         s.failed++;
2724                         s.failed_num = i;
2725                 } else
2726                         set_bit(R5_Insync, &dev->flags);
2727         }
2728         rcu_read_unlock();
2729
2730         if (unlikely(blocked_rdev)) {
2731                 set_bit(STRIPE_HANDLE, &sh->state);
2732                 goto unlock;
2733         }
2734
2735         if (s.to_fill && !test_and_set_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2736                 sh->ops.count++;
2737
2738         pr_debug("locked=%d uptodate=%d to_read=%d"
2739                 " to_write=%d failed=%d failed_num=%d\n",
2740                 s.locked, s.uptodate, s.to_read, s.to_write,
2741                 s.failed, s.failed_num);
2742         /* check if the array has lost two devices and, if so, some requests might
2743          * need to be failed
2744          */
2745         if (s.failed > 1 && s.to_read+s.to_write+s.written)
2746                 handle_requests_to_failed_array(conf, sh, &s, disks,
2747                                                 &return_bi);
2748         if (s.failed > 1 && s.syncing) {
2749                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2750                 clear_bit(STRIPE_SYNCING, &sh->state);
2751                 s.syncing = 0;
2752         }
2753
2754         /* might be able to return some write requests if the parity block
2755          * is safe, or on a failed drive
2756          */
2757         dev = &sh->dev[sh->pd_idx];
2758         if ( s.written &&
2759              ((test_bit(R5_Insync, &dev->flags) &&
2760                !test_bit(R5_LOCKED, &dev->flags) &&
2761                test_bit(R5_UPTODATE, &dev->flags)) ||
2762                (s.failed == 1 && s.failed_num == sh->pd_idx)))
2763                 handle_completed_write_requests(conf, sh, disks, &return_bi);
2764
2765         /* Now we might consider reading some blocks, either to check/generate
2766          * parity, or to satisfy requests
2767          * or to load a block that is being partially written.
2768          */
2769         if (s.to_read || s.non_overwrite ||
2770             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding ||
2771             test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2772                 handle_issuing_new_read_requests5(sh, &s, disks);
2773
2774         /* Now we check to see if any write operations have recently
2775          * completed
2776          */
2777
2778         /* leave prexor set until postxor is done, allows us to distinguish
2779          * a rmw from a rcw during biodrain
2780          */
2781         prexor = 0;
2782         if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2783                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2784
2785                 prexor = 1;
2786                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2787                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2788                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2789
2790                 for (i = disks; i--; )
2791                         clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2792         }
2793
2794         /* if only POSTXOR is set then this is an 'expand' postxor */
2795         if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2796                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2797
2798                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2799                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2800                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2801
2802                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2803                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2804                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2805
2806                 /* All the 'written' buffers and the parity block are ready to
2807                  * be written back to disk
2808                  */
2809                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2810                 for (i = disks; i--; ) {
2811                         dev = &sh->dev[i];
2812                         if (test_bit(R5_LOCKED, &dev->flags) &&
2813                                 (i == sh->pd_idx || dev->written)) {
2814                                 pr_debug("Writing block %d\n", i);
2815                                 set_bit(R5_Wantwrite, &dev->flags);
2816                                 if (!test_and_set_bit(
2817                                     STRIPE_OP_IO, &sh->ops.pending))
2818                                         sh->ops.count++;
2819                                 if (prexor)
2820                                         continue;
2821                                 if (!test_bit(R5_Insync, &dev->flags) ||
2822                                     (i == sh->pd_idx && s.failed == 0))
2823                                         set_bit(STRIPE_INSYNC, &sh->state);
2824                         }
2825                 }
2826                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2827                         atomic_dec(&conf->preread_active_stripes);
2828                         if (atomic_read(&conf->preread_active_stripes) <
2829                                 IO_THRESHOLD)
2830                                 md_wakeup_thread(conf->mddev->thread);
2831                 }
2832         }
2833
2834         /* Now to consider new write requests and what else, if anything
2835          * should be read.  We do not handle new writes when:
2836          * 1/ A 'write' operation (copy+xor) is already in flight.
2837          * 2/ A 'check' operation is in flight, as it may clobber the parity
2838          *    block.
2839          */
2840         if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2841                           !test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2842                 handle_issuing_new_write_requests5(conf, sh, &s, disks);
2843
2844         /* maybe we need to check and possibly fix the parity for this stripe
2845          * Any reads will already have been scheduled, so we just see if enough
2846          * data is available.  The parity check is held off while parity
2847          * dependent operations are in flight.
2848          */
2849         if ((s.syncing && s.locked == 0 &&
2850              !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2851              !test_bit(STRIPE_INSYNC, &sh->state)) ||
2852               test_bit(STRIPE_OP_CHECK, &sh->ops.pending) ||
2853               test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending))
2854                 handle_parity_checks5(conf, sh, &s, disks);
2855
2856         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2857                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2858                 clear_bit(STRIPE_SYNCING, &sh->state);
2859         }
2860
2861         /* If the failed drive is just a ReadError, then we might need to progress
2862          * the repair/check process
2863          */
2864         if (s.failed == 1 && !conf->mddev->ro &&
2865             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2866             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2867             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2868                 ) {
2869                 dev = &sh->dev[s.failed_num];
2870                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2871                         set_bit(R5_Wantwrite, &dev->flags);
2872                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2873                                 sh->ops.count++;
2874                         set_bit(R5_ReWrite, &dev->flags);
2875                         set_bit(R5_LOCKED, &dev->flags);
2876                         s.locked++;
2877                 } else {
2878                         /* let's read it back */
2879                         set_bit(R5_Wantread, &dev->flags);
2880                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2881                                 sh->ops.count++;
2882                         set_bit(R5_LOCKED, &dev->flags);
2883                         s.locked++;
2884                 }
2885         }
2886
2887         /* Finish postxor operations initiated by the expansion
2888          * process
2889          */
2890         if (test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete) &&
2891                 !test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending)) {
2892
2893                 clear_bit(STRIPE_EXPANDING, &sh->state);
2894
2895                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2896                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2897                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2898
2899                 for (i = conf->raid_disks; i--; ) {
2900                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2901                         set_bit(R5_LOCKED, &dev->flags);
2902                         s.locked++;
2903                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2904                                 sh->ops.count++;
2905                 }
2906         }
2907
2908         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2909                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2910                 /* Need to write out all blocks after computing parity */
2911                 sh->disks = conf->raid_disks;
2912                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2913                         conf->raid_disks);
2914                 s.locked += handle_write_operations5(sh, 1, 1);
2915         } else if (s.expanded &&
2916                    s.locked == 0 &&
2917                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2918                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2919                 atomic_dec(&conf->reshape_stripes);
2920                 wake_up(&conf->wait_for_overlap);
2921                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2922         }
2923
2924         if (s.expanding && s.locked == 0 &&
2925             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2926                 handle_stripe_expansion(conf, sh, NULL);
2927
2928         if (sh->ops.count)
2929                 pending = get_stripe_work(sh);
2930
2931  unlock:
2932         spin_unlock(&sh->lock);
2933
2934         /* wait for this device to become unblocked */
2935         if (unlikely(blocked_rdev))
2936                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
2937
2938         if (pending)
2939                 raid5_run_ops(sh, pending);
2940
2941         return_io(return_bi);
2942
2943 }
2944
2945 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2946 {
2947         raid6_conf_t *conf = sh->raid_conf;
2948         int disks = sh->disks;
2949         struct bio *return_bi = NULL;
2950         int i, pd_idx = sh->pd_idx;
2951         struct stripe_head_state s;
2952         struct r6_state r6s;
2953         struct r5dev *dev, *pdev, *qdev;
2954         mdk_rdev_t *blocked_rdev = NULL;
2955
2956         r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2957         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2958                 "pd_idx=%d, qd_idx=%d\n",
2959                (unsigned long long)sh->sector, sh->state,
2960                atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2961         memset(&s, 0, sizeof(s));
2962
2963         spin_lock(&sh->lock);
2964         clear_bit(STRIPE_HANDLE, &sh->state);
2965         clear_bit(STRIPE_DELAYED, &sh->state);
2966
2967         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2968         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2969         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2970         /* Now to look around and see what can be done */
2971
2972         rcu_read_lock();
2973         for (i=disks; i--; ) {
2974                 mdk_rdev_t *rdev;
2975                 dev = &sh->dev[i];
2976                 clear_bit(R5_Insync, &dev->flags);
2977
2978                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2979                         i, dev->flags, dev->toread, dev->towrite, dev->written);
2980                 /* maybe we can reply to a read */
2981                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2982                         struct bio *rbi, *rbi2;
2983                         pr_debug("Return read for disc %d\n", i);
2984                         spin_lock_irq(&conf->device_lock);
2985                         rbi = dev->toread;
2986                         dev->toread = NULL;
2987                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2988                                 wake_up(&conf->wait_for_overlap);
2989                         spin_unlock_irq(&conf->device_lock);
2990                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2991                                 copy_data(0, rbi, dev->page, dev->sector);
2992                                 rbi2 = r5_next_bio(rbi, dev->sector);
2993                                 spin_lock_irq(&conf->device_lock);
2994                                 if (--rbi->bi_phys_segments == 0) {
2995                                         rbi->bi_next = return_bi;
2996                                         return_bi = rbi;
2997                                 }
2998                                 spin_unlock_irq(&conf->device_lock);
2999                                 rbi = rbi2;
3000                         }
3001                 }
3002
3003                 /* now count some things */
3004                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3005                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3006
3007
3008                 if (dev->toread)
3009                         s.to_read++;
3010                 if (dev->towrite) {
3011                         s.to_write++;
3012                         if (!test_bit(R5_OVERWRITE, &dev->flags))
3013                                 s.non_overwrite++;
3014                 }
3015                 if (dev->written)
3016                         s.written++;
3017                 rdev = rcu_dereference(conf->disks[i].rdev);
3018                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3019                         blocked_rdev = rdev;
3020                         atomic_inc(&rdev->nr_pending);
3021                         break;
3022                 }
3023                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3024                         /* The ReadError flag will just be confusing now */
3025                         clear_bit(R5_ReadError, &dev->flags);
3026                         clear_bit(R5_ReWrite, &dev->flags);
3027                 }
3028                 if (!rdev || !test_bit(In_sync, &rdev->flags)
3029                     || test_bit(R5_ReadError, &dev->flags)) {
3030                         if (s.failed < 2)
3031                                 r6s.failed_num[s.failed] = i;
3032                         s.failed++;
3033                 } else
3034                         set_bit(R5_Insync, &dev->flags);
3035         }
3036         rcu_read_unlock();
3037
3038         if (unlikely(blocked_rdev)) {
3039                 set_bit(STRIPE_HANDLE, &sh->state);
3040                 goto unlock;
3041         }
3042         pr_debug("locked=%d uptodate=%d to_read=%d"
3043                " to_write=%d failed=%d failed_num=%d,%d\n",
3044                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3045                r6s.failed_num[0], r6s.failed_num[1]);
3046         /* check if the array has lost >2 devices and, if so, some requests
3047          * might need to be failed
3048          */
3049         if (s.failed > 2 && s.to_read+s.to_write+s.written)
3050                 handle_requests_to_failed_array(conf, sh, &s, disks,
3051                                                 &return_bi);
3052         if (s.failed > 2 && s.syncing) {
3053                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3054                 clear_bit(STRIPE_SYNCING, &sh->state);
3055                 s.syncing = 0;
3056         }
3057
3058         /*
3059          * might be able to return some write requests if the parity blocks
3060          * are safe, or on a failed drive
3061          */
3062         pdev = &sh->dev[pd_idx];
3063         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3064                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3065         qdev = &sh->dev[r6s.qd_idx];
3066         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
3067                 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
3068
3069         if ( s.written &&
3070              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3071                              && !test_bit(R5_LOCKED, &pdev->flags)
3072                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3073              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3074                              && !test_bit(R5_LOCKED, &qdev->flags)
3075                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3076                 handle_completed_write_requests(conf, sh, disks, &return_bi);
3077
3078         /* Now we might consider reading some blocks, either to check/generate
3079          * parity, or to satisfy requests
3080          * or to load a block that is being partially written.
3081          */
3082         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3083             (s.syncing && (s.uptodate < disks)) || s.expanding)
3084                 handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
3085
3086         /* now to consider writing and what else, if anything should be read */
3087         if (s.to_write)
3088                 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
3089
3090         /* maybe we need to check and possibly fix the parity for this stripe
3091          * Any reads will already have been scheduled, so we just see if enough
3092          * data is available
3093          */
3094         if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3095                 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3096
3097         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3098                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3099                 clear_bit(STRIPE_SYNCING, &sh->state);
3100         }
3101
3102         /* If the failed drives are just a ReadError, then we might need
3103          * to progress the repair/check process
3104          */
3105         if (s.failed <= 2 && !conf->mddev->ro)
3106                 for (i = 0; i < s.failed; i++) {
3107                         dev = &sh->dev[r6s.failed_num[i]];
3108                         if (test_bit(R5_ReadError, &dev->flags)
3109                             && !test_bit(R5_LOCKED, &dev->flags)
3110                             && test_bit(R5_UPTODATE, &dev->flags)
3111                                 ) {
3112                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3113                                         set_bit(R5_Wantwrite, &dev->flags);
3114                                         set_bit(R5_ReWrite, &dev->flags);
3115                                         set_bit(R5_LOCKED, &dev->flags);
3116                                 } else {
3117                                         /* let's read it back */
3118                                         set_bit(R5_Wantread, &dev->flags);
3119                                         set_bit(R5_LOCKED, &dev->flags);
3120                                 }
3121                         }
3122                 }
3123
3124         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3125                 /* Need to write out all blocks after computing P&Q */
3126                 sh->disks = conf->raid_disks;
3127                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
3128                                              conf->raid_disks);
3129                 compute_parity6(sh, RECONSTRUCT_WRITE);
3130                 for (i = conf->raid_disks ; i-- ;  ) {
3131                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3132                         s.locked++;
3133                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3134                 }
3135                 clear_bit(STRIPE_EXPANDING, &sh->state);
3136         } else if (s.expanded) {
3137                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3138                 atomic_dec(&conf->reshape_stripes);
3139                 wake_up(&conf->wait_for_overlap);
3140                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3141         }
3142
3143         if (s.expanding && s.locked == 0 &&
3144             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
3145                 handle_stripe_expansion(conf, sh, &r6s);
3146
3147  unlock:
3148         spin_unlock(&sh->lock);
3149
3150         /* wait for this device to become unblocked */
3151         if (unlikely(blocked_rdev))
3152                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3153
3154         return_io(return_bi);
3155
3156         for (i=disks; i-- ;) {
3157                 int rw;
3158                 struct bio *bi;
3159                 mdk_rdev_t *rdev;
3160                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
3161                         rw = WRITE;
3162                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
3163                         rw = READ;
3164                 else
3165                         continue;
3166
3167                 set_bit(STRIPE_IO_STARTED, &sh->state);
3168
3169                 bi = &sh->dev[i].req;
3170
3171                 bi->bi_rw = rw;
3172                 if (rw == WRITE)
3173                         bi->bi_end_io = raid5_end_write_request;
3174                 else
3175                         bi->bi_end_io = raid5_end_read_request;
3176
3177                 rcu_read_lock();
3178                 rdev = rcu_dereference(conf->disks[i].rdev);
3179                 if (rdev && test_bit(Faulty, &rdev->flags))
3180                         rdev = NULL;
3181                 if (rdev)
3182                         atomic_inc(&rdev->nr_pending);
3183                 rcu_read_unlock();
3184
3185                 if (rdev) {
3186                         if (s.syncing || s.expanding || s.expanded)
3187                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
3188
3189                         bi->bi_bdev = rdev->bdev;
3190                         pr_debug("for %llu schedule op %ld on disc %d\n",
3191                                 (unsigned long long)sh->sector, bi->bi_rw, i);
3192                         atomic_inc(&sh->count);
3193                         bi->bi_sector = sh->sector + rdev->data_offset;
3194                         bi->bi_flags = 1 << BIO_UPTODATE;
3195                         bi->bi_vcnt = 1;
3196                         bi->bi_max_vecs = 1;
3197                         bi->bi_idx = 0;
3198                         bi->bi_io_vec = &sh->dev[i].vec;
3199                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
3200                         bi->bi_io_vec[0].bv_offset = 0;
3201                         bi->bi_size = STRIPE_SIZE;
3202                         bi->bi_next = NULL;
3203                         if (rw == WRITE &&
3204                             test_bit(R5_ReWrite, &sh->dev[i].flags))
3205                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
3206                         generic_make_request(bi);
3207                 } else {
3208                         if (rw == WRITE)
3209                                 set_bit(STRIPE_DEGRADED, &sh->state);
3210                         pr_debug("skip op %ld on disc %d for sector %llu\n",
3211                                 bi->bi_rw, i, (unsigned long long)sh->sector);
3212                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
3213                         set_bit(STRIPE_HANDLE, &sh->state);
3214                 }
3215         }
3216 }
3217
3218 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3219 {
3220         if (sh->raid_conf->level == 6)
3221                 handle_stripe6(sh, tmp_page);
3222         else
3223                 handle_stripe5(sh);
3224 }
3225
3226
3227
3228 static void raid5_activate_delayed(raid5_conf_t *conf)
3229 {
3230         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3231                 while (!list_empty(&conf->delayed_list)) {
3232                         struct list_head *l = conf->delayed_list.next;
3233                         struct stripe_head *sh;
3234                         sh = list_entry(l, struct stripe_head, lru);
3235                         list_del_init(l);
3236                         clear_bit(STRIPE_DELAYED, &sh->state);
3237                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3238                                 atomic_inc(&conf->preread_active_stripes);
3239                         list_add_tail(&sh->lru, &conf->hold_list);
3240                 }
3241         } else
3242                 blk_plug_device(conf->mddev->queue);
3243 }
3244
3245 static void activate_bit_delay(raid5_conf_t *conf)
3246 {
3247         /* device_lock is held */
3248         struct list_head head;
3249         list_add(&head, &conf->bitmap_list);
3250         list_del_init(&conf->bitmap_list);
3251         while (!list_empty(&head)) {
3252                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3253                 list_del_init(&sh->lru);
3254                 atomic_inc(&sh->count);
3255                 __release_stripe(conf, sh);
3256         }
3257 }
3258
3259 static void unplug_slaves(mddev_t *mddev)
3260 {
3261         raid5_conf_t *conf = mddev_to_conf(mddev);
3262         int i;
3263
3264         rcu_read_lock();
3265         for (i=0; i<mddev->raid_disks; i++) {
3266                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3267                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3268                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3269
3270                         atomic_inc(&rdev->nr_pending);
3271                         rcu_read_unlock();
3272
3273                         blk_unplug(r_queue);
3274
3275                         rdev_dec_pending(rdev, mddev);
3276                         rcu_read_lock();
3277                 }
3278         }
3279         rcu_read_unlock();
3280 }
3281
3282 static void raid5_unplug_device(struct request_queue *q)
3283 {
3284         mddev_t *mddev = q->queuedata;
3285         raid5_conf_t *conf = mddev_to_conf(mddev);
3286         unsigned long flags;
3287
3288         spin_lock_irqsave(&conf->device_lock, flags);
3289
3290         if (blk_remove_plug(q)) {
3291                 conf->seq_flush++;
3292                 raid5_activate_delayed(conf);
3293         }
3294         md_wakeup_thread(mddev->thread);
3295
3296         spin_unlock_irqrestore(&conf->device_lock, flags);
3297
3298         unplug_slaves(mddev);
3299 }
3300
3301 static int raid5_congested(void *data, int bits)
3302 {
3303         mddev_t *mddev = data;
3304         raid5_conf_t *conf = mddev_to_conf(mddev);
3305
3306         /* No difference between reads and writes.  Just check
3307          * how busy the stripe_cache is
3308          */
3309         if (conf->inactive_blocked)
3310                 return 1;
3311         if (conf->quiesce)
3312                 return 1;
3313         if (list_empty_careful(&conf->inactive_list))
3314                 return 1;
3315
3316         return 0;
3317 }
3318
3319 /* We want read requests to align with chunks where possible,
3320  * but write requests don't need to.
3321  */
3322 static int raid5_mergeable_bvec(struct request_queue *q,
3323                                 struct bvec_merge_data *bvm,
3324                                 struct bio_vec *biovec)
3325 {
3326         mddev_t *mddev = q->queuedata;
3327         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3328         int max;
3329         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3330         unsigned int bio_sectors = bvm->bi_size >> 9;
3331
3332         if ((bvm->bi_rw & 1) == WRITE)
3333                 return biovec->bv_len; /* always allow writes to be mergeable */
3334
3335         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3336         if (max < 0) max = 0;
3337         if (max <= biovec->bv_len && bio_sectors == 0)
3338                 return biovec->bv_len;
3339         else
3340                 return max;
3341 }
3342
3343
3344 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3345 {
3346         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3347         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3348         unsigned int bio_sectors = bio->bi_size >> 9;
3349
3350         return  chunk_sectors >=
3351                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3352 }
3353
3354 /*
3355  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3356  *  later sampled by raid5d.
3357  */
3358 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3359 {
3360         unsigned long flags;
3361
3362         spin_lock_irqsave(&conf->device_lock, flags);
3363
3364         bi->bi_next = conf->retry_read_aligned_list;
3365         conf->retry_read_aligned_list = bi;
3366
3367         spin_unlock_irqrestore(&conf->device_lock, flags);
3368         md_wakeup_thread(conf->mddev->thread);
3369 }
3370
3371
3372 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3373 {
3374         struct bio *bi;
3375
3376         bi = conf->retry_read_aligned;
3377         if (bi) {
3378                 conf->retry_read_aligned = NULL;
3379                 return bi;
3380         }
3381         bi = conf->retry_read_aligned_list;
3382         if(bi) {
3383                 conf->retry_read_aligned_list = bi->bi_next;
3384                 bi->bi_next = NULL;
3385                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3386                 bi->bi_hw_segments = 0; /* count of processed stripes */
3387         }
3388
3389         return bi;
3390 }
3391
3392
3393 /*
3394  *  The "raid5_align_endio" should check if the read succeeded and if it
3395  *  did, call bio_endio on the original bio (having bio_put the new bio
3396  *  first).
3397  *  If the read failed..
3398  */
3399 static void raid5_align_endio(struct bio *bi, int error)
3400 {
3401         struct bio* raid_bi  = bi->bi_private;
3402         mddev_t *mddev;
3403         raid5_conf_t *conf;
3404         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3405         mdk_rdev_t *rdev;
3406
3407         bio_put(bi);
3408
3409         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3410         conf = mddev_to_conf(mddev);
3411         rdev = (void*)raid_bi->bi_next;
3412         raid_bi->bi_next = NULL;
3413
3414         rdev_dec_pending(rdev, conf->mddev);
3415
3416         if (!error && uptodate) {
3417                 bio_endio(raid_bi, 0);
3418                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3419                         wake_up(&conf->wait_for_stripe);
3420                 return;
3421         }
3422
3423
3424         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3425
3426         add_bio_to_retry(raid_bi, conf);
3427 }
3428
3429 static int bio_fits_rdev(struct bio *bi)
3430 {
3431         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3432
3433         if ((bi->bi_size>>9) > q->max_sectors)
3434                 return 0;
3435         blk_recount_segments(q, bi);
3436         if (bi->bi_phys_segments > q->max_phys_segments ||
3437             bi->bi_hw_segments > q->max_hw_segments)
3438                 return 0;
3439
3440         if (q->merge_bvec_fn)
3441                 /* it's too hard to apply the merge_bvec_fn at this stage,
3442                  * just just give up
3443                  */
3444                 return 0;
3445
3446         return 1;
3447 }
3448
3449
3450 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3451 {
3452         mddev_t *mddev = q->queuedata;
3453         raid5_conf_t *conf = mddev_to_conf(mddev);
3454         const unsigned int raid_disks = conf->raid_disks;
3455         const unsigned int data_disks = raid_disks - conf->max_degraded;
3456         unsigned int dd_idx, pd_idx;
3457         struct bio* align_bi;
3458         mdk_rdev_t *rdev;
3459
3460         if (!in_chunk_boundary(mddev, raid_bio)) {
3461                 pr_debug("chunk_aligned_read : non aligned\n");
3462                 return 0;
3463         }
3464         /*
3465          * use bio_clone to make a copy of the bio
3466          */
3467         align_bi = bio_clone(raid_bio, GFP_NOIO);
3468         if (!align_bi)
3469                 return 0;
3470         /*
3471          *   set bi_end_io to a new function, and set bi_private to the
3472          *     original bio.
3473          */
3474         align_bi->bi_end_io  = raid5_align_endio;
3475         align_bi->bi_private = raid_bio;
3476         /*
3477          *      compute position
3478          */
3479         align_bi->bi_sector =  raid5_compute_sector(raid_bio->bi_sector,
3480                                         raid_disks,
3481                                         data_disks,
3482                                         &dd_idx,
3483                                         &pd_idx,
3484                                         conf);
3485
3486         rcu_read_lock();
3487         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3488         if (rdev && test_bit(In_sync, &rdev->flags)) {
3489                 atomic_inc(&rdev->nr_pending);
3490                 rcu_read_unlock();
3491                 raid_bio->bi_next = (void*)rdev;
3492                 align_bi->bi_bdev =  rdev->bdev;
3493                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3494                 align_bi->bi_sector += rdev->data_offset;
3495
3496                 if (!bio_fits_rdev(align_bi)) {
3497                         /* too big in some way */
3498                         bio_put(align_bi);
3499                         rdev_dec_pending(rdev, mddev);
3500                         return 0;
3501                 }
3502
3503                 spin_lock_irq(&conf->device_lock);
3504                 wait_event_lock_irq(conf->wait_for_stripe,
3505                                     conf->quiesce == 0,
3506                                     conf->device_lock, /* nothing */);
3507                 atomic_inc(&conf->active_aligned_reads);
3508                 spin_unlock_irq(&conf->device_lock);
3509
3510                 generic_make_request(align_bi);
3511                 return 1;
3512         } else {
3513                 rcu_read_unlock();
3514                 bio_put(align_bi);
3515                 return 0;
3516         }
3517 }
3518
3519 /* __get_priority_stripe - get the next stripe to process
3520  *
3521  * Full stripe writes are allowed to pass preread active stripes up until
3522  * the bypass_threshold is exceeded.  In general the bypass_count
3523  * increments when the handle_list is handled before the hold_list; however, it
3524  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3525  * stripe with in flight i/o.  The bypass_count will be reset when the
3526  * head of the hold_list has changed, i.e. the head was promoted to the
3527  * handle_list.
3528  */
3529 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3530 {
3531         struct stripe_head *sh;
3532
3533         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3534                   __func__,
3535                   list_empty(&conf->handle_list) ? "empty" : "busy",
3536                   list_empty(&conf->hold_list) ? "empty" : "busy",
3537                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
3538
3539         if (!list_empty(&conf->handle_list)) {
3540                 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3541
3542                 if (list_empty(&conf->hold_list))
3543                         conf->bypass_count = 0;
3544                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3545                         if (conf->hold_list.next == conf->last_hold)
3546                                 conf->bypass_count++;
3547                         else {
3548                                 conf->last_hold = conf->hold_list.next;
3549                                 conf->bypass_count -= conf->bypass_threshold;
3550                                 if (conf->bypass_count < 0)
3551                                         conf->bypass_count = 0;
3552                         }
3553                 }
3554         } else if (!list_empty(&conf->hold_list) &&
3555                    ((conf->bypass_threshold &&
3556                      conf->bypass_count > conf->bypass_threshold) ||
3557                     atomic_read(&conf->pending_full_writes) == 0)) {
3558                 sh = list_entry(conf->hold_list.next,
3559                                 typeof(*sh), lru);
3560                 conf->bypass_count -= conf->bypass_threshold;
3561                 if (conf->bypass_count < 0)
3562                         conf->bypass_count = 0;
3563         } else
3564                 return NULL;
3565
3566         list_del_init(&sh->lru);
3567         atomic_inc(&sh->count);
3568         BUG_ON(atomic_read(&sh->count) != 1);
3569         return sh;
3570 }
3571
3572 static int make_request(struct request_queue *q, struct bio * bi)
3573 {
3574         mddev_t *mddev = q->queuedata;
3575         raid5_conf_t *conf = mddev_to_conf(mddev);
3576         unsigned int dd_idx, pd_idx;
3577         sector_t new_sector;
3578         sector_t logical_sector, last_sector;
3579         struct stripe_head *sh;
3580         const int rw = bio_data_dir(bi);
3581         int remaining;
3582
3583         if (unlikely(bio_barrier(bi))) {
3584                 bio_endio(bi, -EOPNOTSUPP);
3585                 return 0;
3586         }
3587
3588         md_write_start(mddev, bi);
3589
3590         disk_stat_inc(mddev->gendisk, ios[rw]);
3591         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3592
3593         if (rw == READ &&
3594              mddev->reshape_position == MaxSector &&
3595              chunk_aligned_read(q,bi))
3596                 return 0;
3597
3598         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3599         last_sector = bi->bi_sector + (bi->bi_size>>9);
3600         bi->bi_next = NULL;
3601         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3602
3603         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3604                 DEFINE_WAIT(w);
3605                 int disks, data_disks;
3606
3607         retry:
3608                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3609                 if (likely(conf->expand_progress == MaxSector))
3610                         disks = conf->raid_disks;
3611                 else {
3612                         /* spinlock is needed as expand_progress may be
3613                          * 64bit on a 32bit platform, and so it might be
3614                          * possible to see a half-updated value
3615                          * Ofcourse expand_progress could change after
3616                          * the lock is dropped, so once we get a reference
3617                          * to the stripe that we think it is, we will have
3618                          * to check again.
3619                          */
3620                         spin_lock_irq(&conf->device_lock);
3621                         disks = conf->raid_disks;
3622                         if (logical_sector >= conf->expand_progress)
3623                                 disks = conf->previous_raid_disks;
3624                         else {
3625                                 if (logical_sector >= conf->expand_lo) {
3626                                         spin_unlock_irq(&conf->device_lock);
3627                                         schedule();
3628                                         goto retry;
3629                                 }
3630                         }
3631                         spin_unlock_irq(&conf->device_lock);
3632                 }
3633                 data_disks = disks - conf->max_degraded;
3634
3635                 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3636                                                   &dd_idx, &pd_idx, conf);
3637                 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3638                         (unsigned long long)new_sector, 
3639                         (unsigned long long)logical_sector);
3640
3641                 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3642                 if (sh) {
3643                         if (unlikely(conf->expand_progress != MaxSector)) {
3644                                 /* expansion might have moved on while waiting for a
3645                                  * stripe, so we must do the range check again.
3646                                  * Expansion could still move past after this
3647                                  * test, but as we are holding a reference to
3648                                  * 'sh', we know that if that happens,
3649                                  *  STRIPE_EXPANDING will get set and the expansion
3650                                  * won't proceed until we finish with the stripe.
3651                                  */
3652                                 int must_retry = 0;
3653                                 spin_lock_irq(&conf->device_lock);
3654                                 if (logical_sector <  conf->expand_progress &&
3655                                     disks == conf->previous_raid_disks)
3656                                         /* mismatch, need to try again */
3657                                         must_retry = 1;
3658                                 spin_unlock_irq(&conf->device_lock);
3659                                 if (must_retry) {
3660                                         release_stripe(sh);
3661                                         goto retry;
3662                                 }
3663                         }
3664                         /* FIXME what if we get a false positive because these
3665                          * are being updated.
3666                          */
3667                         if (logical_sector >= mddev->suspend_lo &&
3668                             logical_sector < mddev->suspend_hi) {
3669                                 release_stripe(sh);
3670                                 schedule();
3671                                 goto retry;
3672                         }
3673
3674                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3675                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3676                                 /* Stripe is busy expanding or
3677                                  * add failed due to overlap.  Flush everything
3678                                  * and wait a while
3679                                  */
3680                                 raid5_unplug_device(mddev->queue);
3681                                 release_stripe(sh);
3682                                 schedule();
3683                                 goto retry;
3684                         }
3685                         finish_wait(&conf->wait_for_overlap, &w);
3686                         set_bit(STRIPE_HANDLE, &sh->state);
3687                         clear_bit(STRIPE_DELAYED, &sh->state);
3688                         release_stripe(sh);
3689                 } else {
3690                         /* cannot get stripe for read-ahead, just give-up */
3691                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3692                         finish_wait(&conf->wait_for_overlap, &w);
3693                         break;
3694                 }
3695                         
3696         }
3697         spin_lock_irq(&conf->device_lock);
3698         remaining = --bi->bi_phys_segments;
3699         spin_unlock_irq(&conf->device_lock);
3700         if (remaining == 0) {
3701
3702                 if ( rw == WRITE )
3703                         md_write_end(mddev);
3704
3705                 bi->bi_end_io(bi,
3706                               test_bit(BIO_UPTODATE, &bi->bi_flags)
3707                                 ? 0 : -EIO);
3708         }
3709         return 0;
3710 }
3711
3712 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3713 {
3714         /* reshaping is quite different to recovery/resync so it is
3715          * handled quite separately ... here.
3716          *
3717          * On each call to sync_request, we gather one chunk worth of
3718          * destination stripes and flag them as expanding.
3719          * Then we find all the source stripes and request reads.
3720          * As the reads complete, handle_stripe will copy the data
3721          * into the destination stripe and release that stripe.
3722          */
3723         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3724         struct stripe_head *sh;
3725         int pd_idx;
3726         sector_t first_sector, last_sector;
3727         int raid_disks = conf->previous_raid_disks;
3728         int data_disks = raid_disks - conf->max_degraded;
3729         int new_data_disks = conf->raid_disks - conf->max_degraded;
3730         int i;
3731         int dd_idx;
3732         sector_t writepos, safepos, gap;
3733
3734         if (sector_nr == 0 &&
3735             conf->expand_progress != 0) {
3736                 /* restarting in the middle, skip the initial sectors */
3737                 sector_nr = conf->expand_progress;
3738                 sector_div(sector_nr, new_data_disks);
3739                 *skipped = 1;
3740                 return sector_nr;
3741         }
3742
3743         /* we update the metadata when there is more than 3Meg
3744          * in the block range (that is rather arbitrary, should
3745          * probably be time based) or when the data about to be
3746          * copied would over-write the source of the data at
3747          * the front of the range.
3748          * i.e. one new_stripe forward from expand_progress new_maps
3749          * to after where expand_lo old_maps to
3750          */
3751         writepos = conf->expand_progress +
3752                 conf->chunk_size/512*(new_data_disks);
3753         sector_div(writepos, new_data_disks);
3754         safepos = conf->expand_lo;
3755         sector_div(safepos, data_disks);
3756         gap = conf->expand_progress - conf->expand_lo;
3757
3758         if (writepos >= safepos ||
3759             gap > (new_data_disks)*3000*2 /*3Meg*/) {
3760                 /* Cannot proceed until we've updated the superblock... */
3761                 wait_event(conf->wait_for_overlap,
3762                            atomic_read(&conf->reshape_stripes)==0);
3763                 mddev->reshape_position = conf->expand_progress;
3764                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3765                 md_wakeup_thread(mddev->thread);
3766                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3767                            kthread_should_stop());
3768                 spin_lock_irq(&conf->device_lock);
3769                 conf->expand_lo = mddev->reshape_position;
3770                 spin_unlock_irq(&conf->device_lock);
3771                 wake_up(&conf->wait_for_overlap);
3772         }
3773
3774         for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3775                 int j;
3776                 int skipped = 0;
3777                 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3778                 sh = get_active_stripe(conf, sector_nr+i,
3779                                        conf->raid_disks, pd_idx, 0);
3780                 set_bit(STRIPE_EXPANDING, &sh->state);
3781                 atomic_inc(&conf->reshape_stripes);
3782                 /* If any of this stripe is beyond the end of the old
3783                  * array, then we need to zero those blocks
3784                  */
3785                 for (j=sh->disks; j--;) {
3786                         sector_t s;
3787                         if (j == sh->pd_idx)
3788                                 continue;
3789                         if (conf->level == 6 &&
3790                             j == raid6_next_disk(sh->pd_idx, sh->disks))
3791                                 continue;
3792                         s = compute_blocknr(sh, j);
3793                         if (s < (mddev->array_size<<1)) {
3794                                 skipped = 1;
3795                                 continue;
3796                         }
3797                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3798                         set_bit(R5_Expanded, &sh->dev[j].flags);
3799                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
3800                 }
3801                 if (!skipped) {
3802                         set_bit(STRIPE_EXPAND_READY, &sh->state);
3803                         set_bit(STRIPE_HANDLE, &sh->state);
3804                 }
3805                 release_stripe(sh);
3806         }
3807         spin_lock_irq(&conf->device_lock);
3808         conf->expand_progress = (sector_nr + i) * new_data_disks;
3809         spin_unlock_irq(&conf->device_lock);
3810         /* Ok, those stripe are ready. We can start scheduling
3811          * reads on the source stripes.
3812          * The source stripes are determined by mapping the first and last
3813          * block on the destination stripes.
3814          */
3815         first_sector =
3816                 raid5_compute_sector(sector_nr*(new_data_disks),
3817                                      raid_disks, data_disks,
3818                                      &dd_idx, &pd_idx, conf);
3819         last_sector =
3820                 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3821                                      *(new_data_disks) -1,
3822                                      raid_disks, data_disks,
3823                                      &dd_idx, &pd_idx, conf);
3824         if (last_sector >= (mddev->size<<1))
3825                 last_sector = (mddev->size<<1)-1;
3826         while (first_sector <= last_sector) {
3827                 pd_idx = stripe_to_pdidx(first_sector, conf,
3828                                          conf->previous_raid_disks);
3829                 sh = get_active_stripe(conf, first_sector,
3830                                        conf->previous_raid_disks, pd_idx, 0);
3831                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3832                 set_bit(STRIPE_HANDLE, &sh->state);
3833                 release_stripe(sh);
3834                 first_sector += STRIPE_SECTORS;
3835         }
3836         /* If this takes us to the resync_max point where we have to pause,
3837          * then we need to write out the superblock.
3838          */
3839         sector_nr += conf->chunk_size>>9;
3840         if (sector_nr >= mddev->resync_max) {
3841                 /* Cannot proceed until we've updated the superblock... */
3842                 wait_event(conf->wait_for_overlap,
3843                            atomic_read(&conf->reshape_stripes) == 0);
3844                 mddev->reshape_position = conf->expand_progress;
3845                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3846                 md_wakeup_thread(mddev->thread);
3847                 wait_event(mddev->sb_wait,
3848                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3849                            || kthread_should_stop());
3850                 spin_lock_irq(&conf->device_lock);
3851                 conf->expand_lo = mddev->reshape_position;
3852                 spin_unlock_irq(&conf->device_lock);
3853                 wake_up(&conf->wait_for_overlap);
3854         }
3855         return conf->chunk_size>>9;
3856 }
3857
3858 /* FIXME go_faster isn't used */
3859 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3860 {
3861         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3862         struct stripe_head *sh;
3863         int pd_idx;
3864         int raid_disks = conf->raid_disks;
3865         sector_t max_sector = mddev->size << 1;
3866         int sync_blocks;
3867         int still_degraded = 0;
3868         int i;
3869
3870         if (sector_nr >= max_sector) {
3871                 /* just being told to finish up .. nothing much to do */
3872                 unplug_slaves(mddev);
3873                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3874                         end_reshape(conf);
3875                         return 0;
3876                 }
3877
3878                 if (mddev->curr_resync < max_sector) /* aborted */
3879                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3880                                         &sync_blocks, 1);
3881                 else /* completed sync */
3882                         conf->fullsync = 0;
3883                 bitmap_close_sync(mddev->bitmap);
3884
3885                 return 0;
3886         }
3887
3888         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3889                 return reshape_request(mddev, sector_nr, skipped);
3890
3891         /* No need to check resync_max as we never do more than one
3892          * stripe, and as resync_max will always be on a chunk boundary,
3893          * if the check in md_do_sync didn't fire, there is no chance
3894          * of overstepping resync_max here
3895          */
3896
3897         /* if there is too many failed drives and we are trying
3898          * to resync, then assert that we are finished, because there is
3899          * nothing we can do.
3900          */
3901         if (mddev->degraded >= conf->max_degraded &&
3902             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3903                 sector_t rv = (mddev->size << 1) - sector_nr;
3904                 *skipped = 1;
3905                 return rv;
3906         }
3907         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3908             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3909             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3910                 /* we can skip this block, and probably more */
3911                 sync_blocks /= STRIPE_SECTORS;
3912                 *skipped = 1;
3913                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3914         }
3915
3916
3917         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3918
3919         pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3920         sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3921         if (sh == NULL) {
3922                 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3923                 /* make sure we don't swamp the stripe cache if someone else
3924                  * is trying to get access
3925                  */
3926                 schedule_timeout_uninterruptible(1);
3927         }
3928         /* Need to check if array will still be degraded after recovery/resync
3929          * We don't need to check the 'failed' flag as when that gets set,
3930          * recovery aborts.
3931          */
3932         for (i=0; i<mddev->raid_disks; i++)
3933                 if (conf->disks[i].rdev == NULL)
3934                         still_degraded = 1;
3935
3936         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3937
3938         spin_lock(&sh->lock);
3939         set_bit(STRIPE_SYNCING, &sh->state);
3940         clear_bit(STRIPE_INSYNC, &sh->state);
3941         spin_unlock(&sh->lock);
3942
3943         handle_stripe(sh, NULL);
3944         release_stripe(sh);
3945
3946         return STRIPE_SECTORS;
3947 }
3948
3949 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3950 {
3951         /* We may not be able to submit a whole bio at once as there
3952          * may not be enough stripe_heads available.
3953          * We cannot pre-allocate enough stripe_heads as we may need
3954          * more than exist in the cache (if we allow ever large chunks).
3955          * So we do one stripe head at a time and record in
3956          * ->bi_hw_segments how many have been done.
3957          *
3958          * We *know* that this entire raid_bio is in one chunk, so
3959          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3960          */
3961         struct stripe_head *sh;
3962         int dd_idx, pd_idx;
3963         sector_t sector, logical_sector, last_sector;
3964         int scnt = 0;
3965         int remaining;
3966         int handled = 0;
3967
3968         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3969         sector = raid5_compute_sector(  logical_sector,
3970                                         conf->raid_disks,
3971                                         conf->raid_disks - conf->max_degraded,
3972                                         &dd_idx,
3973                                         &pd_idx,
3974                                         conf);
3975         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3976
3977         for (; logical_sector < last_sector;
3978              logical_sector += STRIPE_SECTORS,
3979                      sector += STRIPE_SECTORS,
3980                      scnt++) {
3981
3982                 if (scnt < raid_bio->bi_hw_segments)
3983                         /* already done this stripe */
3984                         continue;
3985
3986                 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3987
3988                 if (!sh) {
3989                         /* failed to get a stripe - must wait */
3990                         raid_bio->bi_hw_segments = scnt;
3991                         conf->retry_read_aligned = raid_bio;
3992                         return handled;
3993                 }
3994
3995                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3996                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3997                         release_stripe(sh);
3998                         raid_bio->bi_hw_segments = scnt;
3999                         conf->retry_read_aligned = raid_bio;
4000                         return handled;
4001                 }
4002
4003                 handle_stripe(sh, NULL);
4004                 release_stripe(sh);
4005                 handled++;
4006         }
4007         spin_lock_irq(&conf->device_lock);
4008         remaining = --raid_bio->bi_phys_segments;
4009         spin_unlock_irq(&conf->device_lock);
4010         if (remaining == 0) {
4011
4012                 raid_bio->bi_end_io(raid_bio,
4013                               test_bit(BIO_UPTODATE, &raid_bio->bi_flags)
4014                                 ? 0 : -EIO);
4015         }
4016         if (atomic_dec_and_test(&conf->active_aligned_reads))
4017                 wake_up(&conf->wait_for_stripe);
4018         return handled;
4019 }
4020
4021
4022
4023 /*
4024  * This is our raid5 kernel thread.
4025  *
4026  * We scan the hash table for stripes which can be handled now.
4027  * During the scan, completed stripes are saved for us by the interrupt
4028  * handler, so that they will not have to wait for our next wakeup.
4029  */
4030 static void raid5d(mddev_t *mddev)
4031 {
4032         struct stripe_head *sh;
4033         raid5_conf_t *conf = mddev_to_conf(mddev);
4034         int handled;
4035
4036         pr_debug("+++ raid5d active\n");
4037
4038         md_check_recovery(mddev);
4039
4040         handled = 0;
4041         spin_lock_irq(&conf->device_lock);
4042         while (1) {
4043                 struct bio *bio;
4044
4045                 if (conf->seq_flush != conf->seq_write) {
4046                         int seq = conf->seq_flush;
4047                         spin_unlock_irq(&conf->device_lock);
4048                         bitmap_unplug(mddev->bitmap);
4049                         spin_lock_irq(&conf->device_lock);
4050                         conf->seq_write = seq;
4051                         activate_bit_delay(conf);
4052                 }
4053
4054                 while ((bio = remove_bio_from_retry(conf))) {
4055                         int ok;
4056                         spin_unlock_irq(&conf->device_lock);
4057                         ok = retry_aligned_read(conf, bio);
4058                         spin_lock_irq(&conf->device_lock);
4059                         if (!ok)
4060                                 break;
4061                         handled++;
4062                 }
4063
4064                 sh = __get_priority_stripe(conf);
4065
4066                 if (!sh) {
4067                         async_tx_issue_pending_all();
4068                         break;
4069                 }
4070                 spin_unlock_irq(&conf->device_lock);
4071                 
4072                 handled++;
4073                 handle_stripe(sh, conf->spare_page);
4074                 release_stripe(sh);
4075
4076                 spin_lock_irq(&conf->device_lock);
4077         }
4078         pr_debug("%d stripes handled\n", handled);
4079
4080         spin_unlock_irq(&conf->device_lock);
4081
4082         unplug_slaves(mddev);
4083
4084         pr_debug("--- raid5d inactive\n");
4085 }
4086
4087 static ssize_t
4088 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4089 {
4090         raid5_conf_t *conf = mddev_to_conf(mddev);
4091         if (conf)
4092                 return sprintf(page, "%d\n", conf->max_nr_stripes);
4093         else
4094                 return 0;
4095 }
4096
4097 static ssize_t
4098 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4099 {
4100         raid5_conf_t *conf = mddev_to_conf(mddev);
4101         unsigned long new;
4102         if (len >= PAGE_SIZE)
4103                 return -EINVAL;
4104         if (!conf)
4105                 return -ENODEV;
4106
4107         if (strict_strtoul(page, 10, &new))
4108                 return -EINVAL;
4109         if (new <= 16 || new > 32768)
4110                 return -EINVAL;
4111         while (new < conf->max_nr_stripes) {
4112                 if (drop_one_stripe(conf))
4113                         conf->max_nr_stripes--;
4114                 else
4115                         break;
4116         }
4117         md_allow_write(mddev);
4118         while (new > conf->max_nr_stripes) {
4119                 if (grow_one_stripe(conf))
4120                         conf->max_nr_stripes++;
4121                 else break;
4122         }
4123         return len;
4124 }
4125
4126 static struct md_sysfs_entry
4127 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4128                                 raid5_show_stripe_cache_size,
4129                                 raid5_store_stripe_cache_size);
4130
4131 static ssize_t
4132 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4133 {
4134         raid5_conf_t *conf = mddev_to_conf(mddev);
4135         if (conf)
4136                 return sprintf(page, "%d\n", conf->bypass_threshold);
4137         else
4138                 return 0;
4139 }
4140
4141 static ssize_t
4142 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4143 {
4144         raid5_conf_t *conf = mddev_to_conf(mddev);
4145         unsigned long new;
4146         if (len >= PAGE_SIZE)
4147                 return -EINVAL;
4148         if (!conf)
4149                 return -ENODEV;
4150
4151         if (strict_strtoul(page, 10, &new))
4152                 return -EINVAL;
4153         if (new > conf->max_nr_stripes)
4154                 return -EINVAL;
4155         conf->bypass_threshold = new;
4156         return len;
4157 }
4158
4159 static struct md_sysfs_entry
4160 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4161                                         S_IRUGO | S_IWUSR,
4162                                         raid5_show_preread_threshold,
4163                                         raid5_store_preread_threshold);
4164
4165 static ssize_t
4166 stripe_cache_active_show(mddev_t *mddev, char *page)
4167 {
4168         raid5_conf_t *conf = mddev_to_conf(mddev);
4169         if (conf)
4170                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4171         else
4172                 return 0;
4173 }
4174
4175 static struct md_sysfs_entry
4176 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4177
4178 static struct attribute *raid5_attrs[] =  {
4179         &raid5_stripecache_size.attr,
4180         &raid5_stripecache_active.attr,
4181         &raid5_preread_bypass_threshold.attr,
4182         NULL,
4183 };
4184 static struct attribute_group raid5_attrs_group = {
4185         .name = NULL,
4186         .attrs = raid5_attrs,
4187 };
4188
4189 static int run(mddev_t *mddev)
4190 {
4191         raid5_conf_t *conf;
4192         int raid_disk, memory;
4193         mdk_rdev_t *rdev;
4194         struct disk_info *disk;
4195         struct list_head *tmp;
4196         int working_disks = 0;
4197
4198         if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4199                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4200                        mdname(mddev), mddev->level);
4201                 return -EIO;
4202         }
4203
4204         if (mddev->reshape_position != MaxSector) {
4205                 /* Check that we can continue the reshape.
4206                  * Currently only disks can change, it must
4207                  * increase, and we must be past the point where
4208                  * a stripe over-writes itself
4209                  */
4210                 sector_t here_new, here_old;
4211                 int old_disks;
4212                 int max_degraded = (mddev->level == 5 ? 1 : 2);
4213
4214                 if (mddev->new_level != mddev->level ||
4215                     mddev->new_layout != mddev->layout ||
4216                     mddev->new_chunk != mddev->chunk_size) {
4217                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4218                                "required - aborting.\n",
4219                                mdname(mddev));
4220                         return -EINVAL;
4221                 }
4222                 if (mddev->delta_disks <= 0) {
4223                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4224                                "(reduce disks) required - aborting.\n",
4225                                mdname(mddev));
4226                         return -EINVAL;
4227                 }
4228                 old_disks = mddev->raid_disks - mddev->delta_disks;
4229                 /* reshape_position must be on a new-stripe boundary, and one
4230                  * further up in new geometry must map after here in old
4231                  * geometry.
4232                  */
4233                 here_new = mddev->reshape_position;
4234                 if (sector_div(here_new, (mddev->chunk_size>>9)*
4235                                (mddev->raid_disks - max_degraded))) {
4236                         printk(KERN_ERR "raid5: reshape_position not "
4237                                "on a stripe boundary\n");
4238                         return -EINVAL;
4239                 }
4240                 /* here_new is the stripe we will write to */
4241                 here_old = mddev->reshape_position;
4242                 sector_div(here_old, (mddev->chunk_size>>9)*
4243                            (old_disks-max_degraded));
4244                 /* here_old is the first stripe that we might need to read
4245                  * from */
4246                 if (here_new >= here_old) {
4247                         /* Reading from the same stripe as writing to - bad */
4248                         printk(KERN_ERR "raid5: reshape_position too early for "
4249                                "auto-recovery - aborting.\n");
4250                         return -EINVAL;
4251                 }
4252                 printk(KERN_INFO "raid5: reshape will continue\n");
4253                 /* OK, we should be able to continue; */
4254         }
4255
4256
4257         mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4258         if ((conf = mddev->private) == NULL)
4259                 goto abort;
4260         if (mddev->reshape_position == MaxSector) {
4261                 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4262         } else {
4263                 conf->raid_disks = mddev->raid_disks;
4264                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4265         }
4266
4267         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4268                               GFP_KERNEL);
4269         if (!conf->disks)
4270                 goto abort;
4271
4272         conf->mddev = mddev;
4273
4274         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4275                 goto abort;
4276
4277         if (mddev->level == 6) {
4278                 conf->spare_page = alloc_page(GFP_KERNEL);
4279                 if (!conf->spare_page)
4280                         goto abort;
4281         }
4282         spin_lock_init(&conf->device_lock);
4283         mddev->queue->queue_lock = &conf->device_lock;
4284         init_waitqueue_head(&conf->wait_for_stripe);
4285         init_waitqueue_head(&conf->wait_for_overlap);
4286         INIT_LIST_HEAD(&conf->handle_list);
4287         INIT_LIST_HEAD(&conf->hold_list);
4288         INIT_LIST_HEAD(&conf->delayed_list);
4289         INIT_LIST_HEAD(&conf->bitmap_list);
4290         INIT_LIST_HEAD(&conf->inactive_list);
4291         atomic_set(&conf->active_stripes, 0);
4292         atomic_set(&conf->preread_active_stripes, 0);
4293         atomic_set(&conf->active_aligned_reads, 0);
4294         conf->bypass_threshold = BYPASS_THRESHOLD;
4295
4296         pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4297
4298         rdev_for_each(rdev, tmp, mddev) {
4299                 raid_disk = rdev->raid_disk;
4300                 if (raid_disk >= conf->raid_disks
4301                     || raid_disk < 0)
4302                         continue;
4303                 disk = conf->disks + raid_disk;
4304
4305                 disk->rdev = rdev;
4306
4307                 if (test_bit(In_sync, &rdev->flags)) {
4308                         char b[BDEVNAME_SIZE];
4309                         printk(KERN_INFO "raid5: device %s operational as raid"
4310                                 " disk %d\n", bdevname(rdev->bdev,b),
4311                                 raid_disk);
4312                         working_disks++;
4313                 } else
4314                         /* Cannot rely on bitmap to complete recovery */
4315                         conf->fullsync = 1;
4316         }
4317
4318         /*
4319          * 0 for a fully functional array, 1 or 2 for a degraded array.
4320          */
4321         mddev->degraded = conf->raid_disks - working_disks;
4322         conf->mddev = mddev;
4323         conf->chunk_size = mddev->chunk_size;
4324         conf->level = mddev->level;
4325         if (conf->level == 6)
4326                 conf->max_degraded = 2;
4327         else
4328                 conf->max_degraded = 1;
4329         conf->algorithm = mddev->layout;
4330         conf->max_nr_stripes = NR_STRIPES;
4331         conf->expand_progress = mddev->reshape_position;
4332
4333         /* device size must be a multiple of chunk size */
4334         mddev->size &= ~(mddev->chunk_size/1024 -1);
4335         mddev->resync_max_sectors = mddev->size << 1;
4336
4337         if (conf->level == 6 && conf->raid_disks < 4) {
4338                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4339                        mdname(mddev), conf->raid_disks);
4340                 goto abort;
4341         }
4342         if (!conf->chunk_size || conf->chunk_size % 4) {
4343                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4344                         conf->chunk_size, mdname(mddev));
4345                 goto abort;
4346         }
4347         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4348                 printk(KERN_ERR 
4349                         "raid5: unsupported parity algorithm %d for %s\n",
4350                         conf->algorithm, mdname(mddev));
4351                 goto abort;
4352         }
4353         if (mddev->degraded > conf->max_degraded) {
4354                 printk(KERN_ERR "raid5: not enough operational devices for %s"
4355                         " (%d/%d failed)\n",
4356                         mdname(mddev), mddev->degraded, conf->raid_disks);
4357                 goto abort;
4358         }
4359
4360         if (mddev->degraded > 0 &&
4361             mddev->recovery_cp != MaxSector) {
4362                 if (mddev->ok_start_degraded)
4363                         printk(KERN_WARNING
4364                                "raid5: starting dirty degraded array: %s"
4365                                "- data corruption possible.\n",
4366                                mdname(mddev));
4367                 else {
4368                         printk(KERN_ERR
4369                                "raid5: cannot start dirty degraded array for %s\n",
4370                                mdname(mddev));
4371                         goto abort;
4372                 }
4373         }
4374
4375         {
4376                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4377                 if (!mddev->thread) {
4378                         printk(KERN_ERR 
4379                                 "raid5: couldn't allocate thread for %s\n",
4380                                 mdname(mddev));
4381                         goto abort;
4382                 }
4383         }
4384         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4385                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4386         if (grow_stripes(conf, conf->max_nr_stripes)) {
4387                 printk(KERN_ERR 
4388                         "raid5: couldn't allocate %dkB for buffers\n", memory);
4389                 shrink_stripes(conf);
4390                 md_unregister_thread(mddev->thread);
4391                 goto abort;
4392         } else
4393                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4394                         memory, mdname(mddev));
4395
4396         if (mddev->degraded == 0)
4397                 printk("raid5: raid level %d set %s active with %d out of %d"
4398                         " devices, algorithm %d\n", conf->level, mdname(mddev), 
4399                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4400                         conf->algorithm);
4401         else
4402                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4403                         " out of %d devices, algorithm %d\n", conf->level,
4404                         mdname(mddev), mddev->raid_disks - mddev->degraded,
4405                         mddev->raid_disks, conf->algorithm);
4406
4407         print_raid5_conf(conf);
4408
4409         if (conf->expand_progress != MaxSector) {
4410                 printk("...ok start reshape thread\n");
4411                 conf->expand_lo = conf->expand_progress;
4412                 atomic_set(&conf->reshape_stripes, 0);
4413                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4414                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4415                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4416                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4417                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4418                                                         "%s_reshape");
4419         }
4420
4421         /* read-ahead size must cover two whole stripes, which is
4422          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4423          */
4424         {
4425                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4426                 int stripe = data_disks *
4427                         (mddev->chunk_size / PAGE_SIZE);
4428                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4429                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4430         }
4431
4432         /* Ok, everything is just fine now */
4433         if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4434                 printk(KERN_WARNING
4435                        "raid5: failed to create sysfs attributes for %s\n",
4436                        mdname(mddev));
4437
4438         mddev->queue->unplug_fn = raid5_unplug_device;
4439         mddev->queue->backing_dev_info.congested_data = mddev;
4440         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4441
4442         mddev->array_size =  mddev->size * (conf->previous_raid_disks -
4443                                             conf->max_degraded);
4444
4445         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4446
4447         return 0;
4448 abort:
4449         if (conf) {
4450                 print_raid5_conf(conf);
4451                 safe_put_page(conf->spare_page);
4452                 kfree(conf->disks);
4453                 kfree(conf->stripe_hashtbl);
4454                 kfree(conf);
4455         }
4456         mddev->private = NULL;
4457         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4458         return -EIO;
4459 }
4460
4461
4462
4463 static int stop(mddev_t *mddev)
4464 {
4465         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4466
4467         md_unregister_thread(mddev->thread);
4468         mddev->thread = NULL;
4469         shrink_stripes(conf);
4470         kfree(conf->stripe_hashtbl);
4471         mddev->queue->backing_dev_info.congested_fn = NULL;
4472         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4473         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4474         kfree(conf->disks);
4475         kfree(conf);
4476         mddev->private = NULL;
4477         return 0;
4478 }
4479
4480 #ifdef DEBUG
4481 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4482 {
4483         int i;
4484
4485         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4486                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4487         seq_printf(seq, "sh %llu,  count %d.\n",
4488                    (unsigned long long)sh->sector, atomic_read(&sh->count));
4489         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4490         for (i = 0; i < sh->disks; i++) {
4491                 seq_printf(seq, "(cache%d: %p %ld) ",
4492                            i, sh->dev[i].page, sh->dev[i].flags);
4493         }
4494         seq_printf(seq, "\n");
4495 }
4496
4497 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4498 {
4499         struct stripe_head *sh;
4500         struct hlist_node *hn;
4501         int i;
4502
4503         spin_lock_irq(&conf->device_lock);
4504         for (i = 0; i < NR_HASH; i++) {
4505                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4506                         if (sh->raid_conf != conf)
4507                                 continue;
4508                         print_sh(seq, sh);
4509                 }
4510         }
4511         spin_unlock_irq(&conf->device_lock);
4512 }
4513 #endif
4514
4515 static void status (struct seq_file *seq, mddev_t *mddev)
4516 {
4517         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4518         int i;
4519
4520         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4521         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4522         for (i = 0; i < conf->raid_disks; i++)
4523                 seq_printf (seq, "%s",
4524                                conf->disks[i].rdev &&
4525                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4526         seq_printf (seq, "]");
4527 #ifdef DEBUG
4528         seq_printf (seq, "\n");
4529         printall(seq, conf);
4530 #endif
4531 }
4532
4533 static void print_raid5_conf (raid5_conf_t *conf)
4534 {
4535         int i;
4536         struct disk_info *tmp;
4537
4538         printk("RAID5 conf printout:\n");
4539         if (!conf) {
4540                 printk("(conf==NULL)\n");
4541                 return;
4542         }
4543         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4544                  conf->raid_disks - conf->mddev->degraded);
4545
4546         for (i = 0; i < conf->raid_disks; i++) {
4547                 char b[BDEVNAME_SIZE];
4548                 tmp = conf->disks + i;
4549                 if (tmp->rdev)
4550                 printk(" disk %d, o:%d, dev:%s\n",
4551                         i, !test_bit(Faulty, &tmp->rdev->flags),
4552                         bdevname(tmp->rdev->bdev,b));
4553         }
4554 }
4555
4556 static int raid5_spare_active(mddev_t *mddev)
4557 {
4558         int i;
4559         raid5_conf_t *conf = mddev->private;
4560         struct disk_info *tmp;
4561
4562         for (i = 0; i < conf->raid_disks; i++) {
4563                 tmp = conf->disks + i;
4564                 if (tmp->rdev
4565                     && !test_bit(Faulty, &tmp->rdev->flags)
4566                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4567                         unsigned long flags;
4568                         spin_lock_irqsave(&conf->device_lock, flags);
4569                         mddev->degraded--;
4570                         spin_unlock_irqrestore(&conf->device_lock, flags);
4571                 }
4572         }
4573         print_raid5_conf(conf);
4574         return 0;
4575 }
4576
4577 static int raid5_remove_disk(mddev_t *mddev, int number)
4578 {
4579         raid5_conf_t *conf = mddev->private;
4580         int err = 0;
4581         mdk_rdev_t *rdev;
4582         struct disk_info *p = conf->disks + number;
4583
4584         print_raid5_conf(conf);
4585         rdev = p->rdev;
4586         if (rdev) {
4587                 if (test_bit(In_sync, &rdev->flags) ||
4588                     atomic_read(&rdev->nr_pending)) {
4589                         err = -EBUSY;
4590                         goto abort;
4591                 }
4592                 /* Only remove non-faulty devices if recovery
4593                  * isn't possible.
4594                  */
4595                 if (!test_bit(Faulty, &rdev->flags) &&
4596                     mddev->degraded <= conf->max_degraded) {
4597                         err = -EBUSY;
4598                         goto abort;
4599                 }
4600                 p->rdev = NULL;
4601                 synchronize_rcu();
4602                 if (atomic_read(&rdev->nr_pending)) {
4603                         /* lost the race, try later */
4604                         err = -EBUSY;
4605                         p->rdev = rdev;
4606                 }
4607         }
4608 abort:
4609
4610         print_raid5_conf(conf);
4611         return err;
4612 }
4613
4614 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4615 {
4616         raid5_conf_t *conf = mddev->private;
4617         int found = 0;
4618         int disk;
4619         struct disk_info *p;
4620
4621         if (mddev->degraded > conf->max_degraded)
4622                 /* no point adding a device */
4623                 return 0;
4624
4625         /*
4626          * find the disk ... but prefer rdev->saved_raid_disk
4627          * if possible.
4628          */
4629         if (rdev->saved_raid_disk >= 0 &&
4630             conf->disks[rdev->saved_raid_disk].rdev == NULL)
4631                 disk = rdev->saved_raid_disk;
4632         else
4633                 disk = 0;
4634         for ( ; disk < conf->raid_disks; disk++)
4635                 if ((p=conf->disks + disk)->rdev == NULL) {
4636                         clear_bit(In_sync, &rdev->flags);
4637                         rdev->raid_disk = disk;
4638                         found = 1;
4639                         if (rdev->saved_raid_disk != disk)
4640                                 conf->fullsync = 1;
4641                         rcu_assign_pointer(p->rdev, rdev);
4642                         break;
4643                 }
4644         print_raid5_conf(conf);
4645         return found;
4646 }
4647
4648 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4649 {
4650         /* no resync is happening, and there is enough space
4651          * on all devices, so we can resize.
4652          * We need to make sure resync covers any new space.
4653          * If the array is shrinking we should possibly wait until
4654          * any io in the removed space completes, but it hardly seems
4655          * worth it.
4656          */
4657         raid5_conf_t *conf = mddev_to_conf(mddev);
4658
4659         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4660         mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4661         set_capacity(mddev->gendisk, mddev->array_size << 1);
4662         mddev->changed = 1;
4663         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
4664                 mddev->recovery_cp = mddev->size << 1;
4665                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4666         }
4667         mddev->size = sectors /2;
4668         mddev->resync_max_sectors = sectors;
4669         return 0;
4670 }
4671
4672 #ifdef CONFIG_MD_RAID5_RESHAPE
4673 static int raid5_check_reshape(mddev_t *mddev)
4674 {
4675         raid5_conf_t *conf = mddev_to_conf(mddev);
4676         int err;
4677
4678         if (mddev->delta_disks < 0 ||
4679             mddev->new_level != mddev->level)
4680                 return -EINVAL; /* Cannot shrink array or change level yet */
4681         if (mddev->delta_disks == 0)
4682                 return 0; /* nothing to do */
4683
4684         /* Can only proceed if there are plenty of stripe_heads.
4685          * We need a minimum of one full stripe,, and for sensible progress
4686          * it is best to have about 4 times that.
4687          * If we require 4 times, then the default 256 4K stripe_heads will
4688          * allow for chunk sizes up to 256K, which is probably OK.
4689          * If the chunk size is greater, user-space should request more
4690          * stripe_heads first.
4691          */
4692         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4693             (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4694                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
4695                        (mddev->chunk_size / STRIPE_SIZE)*4);
4696                 return -ENOSPC;
4697         }
4698
4699         err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4700         if (err)
4701                 return err;
4702
4703         if (mddev->degraded > conf->max_degraded)
4704                 return -EINVAL;
4705         /* looks like we might be able to manage this */
4706         return 0;
4707 }
4708
4709 static int raid5_start_reshape(mddev_t *mddev)
4710 {
4711         raid5_conf_t *conf = mddev_to_conf(mddev);
4712         mdk_rdev_t *rdev;
4713         struct list_head *rtmp;
4714         int spares = 0;
4715         int added_devices = 0;
4716         unsigned long flags;
4717
4718         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4719                 return -EBUSY;
4720
4721         rdev_for_each(rdev, rtmp, mddev)
4722                 if (rdev->raid_disk < 0 &&
4723                     !test_bit(Faulty, &rdev->flags))
4724                         spares++;
4725
4726         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4727                 /* Not enough devices even to make a degraded array
4728                  * of that size
4729                  */
4730                 return -EINVAL;
4731
4732         atomic_set(&conf->reshape_stripes, 0);
4733         spin_lock_irq(&conf->device_lock);
4734         conf->previous_raid_disks = conf->raid_disks;
4735         conf->raid_disks += mddev->delta_disks;
4736         conf->expand_progress = 0;
4737         conf->expand_lo = 0;
4738         spin_unlock_irq(&conf->device_lock);
4739
4740         /* Add some new drives, as many as will fit.
4741          * We know there are enough to make the newly sized array work.
4742          */
4743         rdev_for_each(rdev, rtmp, mddev)
4744                 if (rdev->raid_disk < 0 &&
4745                     !test_bit(Faulty, &rdev->flags)) {
4746                         if (raid5_add_disk(mddev, rdev)) {
4747                                 char nm[20];
4748                                 set_bit(In_sync, &rdev->flags);
4749                                 added_devices++;
4750                                 rdev->recovery_offset = 0;
4751                                 sprintf(nm, "rd%d", rdev->raid_disk);
4752                                 if (sysfs_create_link(&mddev->kobj,
4753                                                       &rdev->kobj, nm))
4754                                         printk(KERN_WARNING
4755                                                "raid5: failed to create "
4756                                                " link %s for %s\n",
4757                                                nm, mdname(mddev));
4758                         } else
4759                                 break;
4760                 }
4761
4762         spin_lock_irqsave(&conf->device_lock, flags);
4763         mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4764         spin_unlock_irqrestore(&conf->device_lock, flags);
4765         mddev->raid_disks = conf->raid_disks;
4766         mddev->reshape_position = 0;
4767         set_bit(MD_CHANGE_DEVS, &mddev->flags);
4768
4769         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4770         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4771         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4772         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4773         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4774                                                 "%s_reshape");
4775         if (!mddev->sync_thread) {
4776                 mddev->recovery = 0;
4777                 spin_lock_irq(&conf->device_lock);
4778                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4779                 conf->expand_progress = MaxSector;
4780                 spin_unlock_irq(&conf->device_lock);
4781                 return -EAGAIN;
4782         }
4783         md_wakeup_thread(mddev->sync_thread);
4784         md_new_event(mddev);
4785         return 0;
4786 }
4787 #endif
4788
4789 static void end_reshape(raid5_conf_t *conf)
4790 {
4791         struct block_device *bdev;
4792
4793         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4794                 conf->mddev->array_size = conf->mddev->size *
4795                         (conf->raid_disks - conf->max_degraded);
4796                 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4797                 conf->mddev->changed = 1;
4798
4799                 bdev = bdget_disk(conf->mddev->gendisk, 0);
4800                 if (bdev) {
4801                         mutex_lock(&bdev->bd_inode->i_mutex);
4802                         i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4803                         mutex_unlock(&bdev->bd_inode->i_mutex);
4804                         bdput(bdev);
4805                 }
4806                 spin_lock_irq(&conf->device_lock);
4807                 conf->expand_progress = MaxSector;
4808                 spin_unlock_irq(&conf->device_lock);
4809                 conf->mddev->reshape_position = MaxSector;
4810
4811                 /* read-ahead size must cover two whole stripes, which is
4812                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4813                  */
4814                 {
4815                         int data_disks = conf->previous_raid_disks - conf->max_degraded;
4816                         int stripe = data_disks *
4817                                 (conf->mddev->chunk_size / PAGE_SIZE);
4818                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4819                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4820                 }
4821         }
4822 }
4823
4824 static void raid5_quiesce(mddev_t *mddev, int state)
4825 {
4826         raid5_conf_t *conf = mddev_to_conf(mddev);
4827
4828         switch(state) {
4829         case 2: /* resume for a suspend */
4830                 wake_up(&conf->wait_for_overlap);
4831                 break;
4832
4833         case 1: /* stop all writes */
4834                 spin_lock_irq(&conf->device_lock);
4835                 conf->quiesce = 1;
4836                 wait_event_lock_irq(conf->wait_for_stripe,
4837                                     atomic_read(&conf->active_stripes) == 0 &&
4838                                     atomic_read(&conf->active_aligned_reads) == 0,
4839                                     conf->device_lock, /* nothing */);
4840                 spin_unlock_irq(&conf->device_lock);
4841                 break;
4842
4843         case 0: /* re-enable writes */
4844                 spin_lock_irq(&conf->device_lock);
4845                 conf->quiesce = 0;
4846                 wake_up(&conf->wait_for_stripe);
4847                 wake_up(&conf->wait_for_overlap);
4848                 spin_unlock_irq(&conf->device_lock);
4849                 break;
4850         }
4851 }
4852
4853 static struct mdk_personality raid6_personality =
4854 {
4855         .name           = "raid6",
4856         .level          = 6,
4857         .owner          = THIS_MODULE,
4858         .make_request   = make_request,
4859         .run            = run,
4860         .stop           = stop,
4861         .status         = status,
4862         .error_handler  = error,
4863         .hot_add_disk   = raid5_add_disk,
4864         .hot_remove_disk= raid5_remove_disk,
4865         .spare_active   = raid5_spare_active,
4866         .sync_request   = sync_request,
4867         .resize         = raid5_resize,
4868 #ifdef CONFIG_MD_RAID5_RESHAPE
4869         .check_reshape  = raid5_check_reshape,
4870         .start_reshape  = raid5_start_reshape,
4871 #endif
4872         .quiesce        = raid5_quiesce,
4873 };
4874 static struct mdk_personality raid5_personality =
4875 {
4876         .name           = "raid5",
4877         .level          = 5,
4878         .owner          = THIS_MODULE,
4879         .make_request   = make_request,
4880         .run            = run,
4881         .stop           = stop,
4882         .status         = status,
4883         .error_handler  = error,
4884         .hot_add_disk   = raid5_add_disk,
4885         .hot_remove_disk= raid5_remove_disk,
4886         .spare_active   = raid5_spare_active,
4887         .sync_request   = sync_request,
4888         .resize         = raid5_resize,
4889 #ifdef CONFIG_MD_RAID5_RESHAPE
4890         .check_reshape  = raid5_check_reshape,
4891         .start_reshape  = raid5_start_reshape,
4892 #endif
4893         .quiesce        = raid5_quiesce,
4894 };
4895
4896 static struct mdk_personality raid4_personality =
4897 {
4898         .name           = "raid4",
4899         .level          = 4,
4900         .owner          = THIS_MODULE,
4901         .make_request   = make_request,
4902         .run            = run,
4903         .stop           = stop,
4904         .status         = status,
4905         .error_handler  = error,
4906         .hot_add_disk   = raid5_add_disk,
4907         .hot_remove_disk= raid5_remove_disk,
4908         .spare_active   = raid5_spare_active,
4909         .sync_request   = sync_request,
4910         .resize         = raid5_resize,
4911 #ifdef CONFIG_MD_RAID5_RESHAPE
4912         .check_reshape  = raid5_check_reshape,
4913         .start_reshape  = raid5_start_reshape,
4914 #endif
4915         .quiesce        = raid5_quiesce,
4916 };
4917
4918 static int __init raid5_init(void)
4919 {
4920         int e;
4921
4922         e = raid6_select_algo();
4923         if ( e )
4924                 return e;
4925         register_md_personality(&raid6_personality);
4926         register_md_personality(&raid5_personality);
4927         register_md_personality(&raid4_personality);
4928         return 0;
4929 }
4930
4931 static void raid5_exit(void)
4932 {
4933         unregister_md_personality(&raid6_personality);
4934         unregister_md_personality(&raid5_personality);
4935         unregister_md_personality(&raid4_personality);
4936 }
4937
4938 module_init(raid5_init);
4939 module_exit(raid5_exit);
4940 MODULE_LICENSE("GPL");
4941 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4942 MODULE_ALIAS("md-raid5");
4943 MODULE_ALIAS("md-raid4");
4944 MODULE_ALIAS("md-level-5");
4945 MODULE_ALIAS("md-level-4");
4946 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4947 MODULE_ALIAS("md-raid6");
4948 MODULE_ALIAS("md-level-6");
4949
4950 /* This used to be two separate modules, they were: */
4951 MODULE_ALIAS("raid5");
4952 MODULE_ALIAS("raid6");