md: common infrastructure for running operations with raid5_run_ops
[linux-2.6.git] / drivers / md / raid5.c
1 /*
2  * raid5.c : Multiple Devices driver for Linux
3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4  *         Copyright (C) 1999, 2000 Ingo Molnar
5  *         Copyright (C) 2002, 2003 H. Peter Anvin
6  *
7  * RAID-4/5/6 management functions.
8  * Thanks to Penguin Computing for making the RAID-6 development possible
9  * by donating a test server!
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 /*
22  * BITMAP UNPLUGGING:
23  *
24  * The sequencing for updating the bitmap reliably is a little
25  * subtle (and I got it wrong the first time) so it deserves some
26  * explanation.
27  *
28  * We group bitmap updates into batches.  Each batch has a number.
29  * We may write out several batches at once, but that isn't very important.
30  * conf->bm_write is the number of the last batch successfully written.
31  * conf->bm_flush is the number of the last batch that was closed to
32  *    new additions.
33  * When we discover that we will need to write to any block in a stripe
34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35  * the number of the batch it will be in. This is bm_flush+1.
36  * When we are ready to do a write, if that batch hasn't been written yet,
37  *   we plug the array and queue the stripe for later.
38  * When an unplug happens, we increment bm_flush, thus closing the current
39  *   batch.
40  * When we notice that bm_flush > bm_write, we write out all pending updates
41  * to the bitmap, and advance bm_write to where bm_flush was.
42  * This may occasionally write a bit out twice, but is sure never to
43  * miss any bits.
44  */
45
46 #include <linux/module.h>
47 #include <linux/slab.h>
48 #include <linux/highmem.h>
49 #include <linux/bitops.h>
50 #include <linux/kthread.h>
51 #include <asm/atomic.h>
52 #include "raid6.h"
53
54 #include <linux/raid/bitmap.h>
55 #include <linux/async_tx.h>
56
57 /*
58  * Stripe cache
59  */
60
61 #define NR_STRIPES              256
62 #define STRIPE_SIZE             PAGE_SIZE
63 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
64 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
65 #define IO_THRESHOLD            1
66 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
67 #define HASH_MASK               (NR_HASH - 1)
68
69 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
70
71 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
72  * order without overlap.  There may be several bio's per stripe+device, and
73  * a bio could span several devices.
74  * When walking this list for a particular stripe+device, we must never proceed
75  * beyond a bio that extends past this device, as the next bio might no longer
76  * be valid.
77  * This macro is used to determine the 'next' bio in the list, given the sector
78  * of the current stripe+device
79  */
80 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
81 /*
82  * The following can be used to debug the driver
83  */
84 #define RAID5_PARANOIA  1
85 #if RAID5_PARANOIA && defined(CONFIG_SMP)
86 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
87 #else
88 # define CHECK_DEVLOCK()
89 #endif
90
91 #ifdef DEBUG
92 #define inline
93 #define __inline__
94 #endif
95
96 #if !RAID6_USE_EMPTY_ZERO_PAGE
97 /* In .bss so it's zeroed */
98 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
99 #endif
100
101 static inline int raid6_next_disk(int disk, int raid_disks)
102 {
103         disk++;
104         return (disk < raid_disks) ? disk : 0;
105 }
106
107 static void return_io(struct bio *return_bi)
108 {
109         struct bio *bi = return_bi;
110         while (bi) {
111                 int bytes = bi->bi_size;
112
113                 return_bi = bi->bi_next;
114                 bi->bi_next = NULL;
115                 bi->bi_size = 0;
116                 bi->bi_end_io(bi, bytes,
117                               test_bit(BIO_UPTODATE, &bi->bi_flags)
118                                 ? 0 : -EIO);
119                 bi = return_bi;
120         }
121 }
122
123 static void print_raid5_conf (raid5_conf_t *conf);
124
125 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
126 {
127         if (atomic_dec_and_test(&sh->count)) {
128                 BUG_ON(!list_empty(&sh->lru));
129                 BUG_ON(atomic_read(&conf->active_stripes)==0);
130                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
131                         if (test_bit(STRIPE_DELAYED, &sh->state)) {
132                                 list_add_tail(&sh->lru, &conf->delayed_list);
133                                 blk_plug_device(conf->mddev->queue);
134                         } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
135                                    sh->bm_seq - conf->seq_write > 0) {
136                                 list_add_tail(&sh->lru, &conf->bitmap_list);
137                                 blk_plug_device(conf->mddev->queue);
138                         } else {
139                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
140                                 list_add_tail(&sh->lru, &conf->handle_list);
141                         }
142                         md_wakeup_thread(conf->mddev->thread);
143                 } else {
144                         BUG_ON(sh->ops.pending);
145                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
146                                 atomic_dec(&conf->preread_active_stripes);
147                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
148                                         md_wakeup_thread(conf->mddev->thread);
149                         }
150                         atomic_dec(&conf->active_stripes);
151                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
152                                 list_add_tail(&sh->lru, &conf->inactive_list);
153                                 wake_up(&conf->wait_for_stripe);
154                                 if (conf->retry_read_aligned)
155                                         md_wakeup_thread(conf->mddev->thread);
156                         }
157                 }
158         }
159 }
160 static void release_stripe(struct stripe_head *sh)
161 {
162         raid5_conf_t *conf = sh->raid_conf;
163         unsigned long flags;
164
165         spin_lock_irqsave(&conf->device_lock, flags);
166         __release_stripe(conf, sh);
167         spin_unlock_irqrestore(&conf->device_lock, flags);
168 }
169
170 static inline void remove_hash(struct stripe_head *sh)
171 {
172         pr_debug("remove_hash(), stripe %llu\n",
173                 (unsigned long long)sh->sector);
174
175         hlist_del_init(&sh->hash);
176 }
177
178 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
179 {
180         struct hlist_head *hp = stripe_hash(conf, sh->sector);
181
182         pr_debug("insert_hash(), stripe %llu\n",
183                 (unsigned long long)sh->sector);
184
185         CHECK_DEVLOCK();
186         hlist_add_head(&sh->hash, hp);
187 }
188
189
190 /* find an idle stripe, make sure it is unhashed, and return it. */
191 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
192 {
193         struct stripe_head *sh = NULL;
194         struct list_head *first;
195
196         CHECK_DEVLOCK();
197         if (list_empty(&conf->inactive_list))
198                 goto out;
199         first = conf->inactive_list.next;
200         sh = list_entry(first, struct stripe_head, lru);
201         list_del_init(first);
202         remove_hash(sh);
203         atomic_inc(&conf->active_stripes);
204 out:
205         return sh;
206 }
207
208 static void shrink_buffers(struct stripe_head *sh, int num)
209 {
210         struct page *p;
211         int i;
212
213         for (i=0; i<num ; i++) {
214                 p = sh->dev[i].page;
215                 if (!p)
216                         continue;
217                 sh->dev[i].page = NULL;
218                 put_page(p);
219         }
220 }
221
222 static int grow_buffers(struct stripe_head *sh, int num)
223 {
224         int i;
225
226         for (i=0; i<num; i++) {
227                 struct page *page;
228
229                 if (!(page = alloc_page(GFP_KERNEL))) {
230                         return 1;
231                 }
232                 sh->dev[i].page = page;
233         }
234         return 0;
235 }
236
237 static void raid5_build_block (struct stripe_head *sh, int i);
238
239 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks)
240 {
241         raid5_conf_t *conf = sh->raid_conf;
242         int i;
243
244         BUG_ON(atomic_read(&sh->count) != 0);
245         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
246         BUG_ON(sh->ops.pending || sh->ops.ack || sh->ops.complete);
247
248         CHECK_DEVLOCK();
249         pr_debug("init_stripe called, stripe %llu\n",
250                 (unsigned long long)sh->sector);
251
252         remove_hash(sh);
253
254         sh->sector = sector;
255         sh->pd_idx = pd_idx;
256         sh->state = 0;
257
258         sh->disks = disks;
259
260         for (i = sh->disks; i--; ) {
261                 struct r5dev *dev = &sh->dev[i];
262
263                 if (dev->toread || dev->read || dev->towrite || dev->written ||
264                     test_bit(R5_LOCKED, &dev->flags)) {
265                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
266                                (unsigned long long)sh->sector, i, dev->toread,
267                                dev->read, dev->towrite, dev->written,
268                                test_bit(R5_LOCKED, &dev->flags));
269                         BUG();
270                 }
271                 dev->flags = 0;
272                 raid5_build_block(sh, i);
273         }
274         insert_hash(conf, sh);
275 }
276
277 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
278 {
279         struct stripe_head *sh;
280         struct hlist_node *hn;
281
282         CHECK_DEVLOCK();
283         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
284         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
285                 if (sh->sector == sector && sh->disks == disks)
286                         return sh;
287         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
288         return NULL;
289 }
290
291 static void unplug_slaves(mddev_t *mddev);
292 static void raid5_unplug_device(request_queue_t *q);
293
294 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks,
295                                              int pd_idx, int noblock)
296 {
297         struct stripe_head *sh;
298
299         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
300
301         spin_lock_irq(&conf->device_lock);
302
303         do {
304                 wait_event_lock_irq(conf->wait_for_stripe,
305                                     conf->quiesce == 0,
306                                     conf->device_lock, /* nothing */);
307                 sh = __find_stripe(conf, sector, disks);
308                 if (!sh) {
309                         if (!conf->inactive_blocked)
310                                 sh = get_free_stripe(conf);
311                         if (noblock && sh == NULL)
312                                 break;
313                         if (!sh) {
314                                 conf->inactive_blocked = 1;
315                                 wait_event_lock_irq(conf->wait_for_stripe,
316                                                     !list_empty(&conf->inactive_list) &&
317                                                     (atomic_read(&conf->active_stripes)
318                                                      < (conf->max_nr_stripes *3/4)
319                                                      || !conf->inactive_blocked),
320                                                     conf->device_lock,
321                                                     raid5_unplug_device(conf->mddev->queue)
322                                         );
323                                 conf->inactive_blocked = 0;
324                         } else
325                                 init_stripe(sh, sector, pd_idx, disks);
326                 } else {
327                         if (atomic_read(&sh->count)) {
328                           BUG_ON(!list_empty(&sh->lru));
329                         } else {
330                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
331                                         atomic_inc(&conf->active_stripes);
332                                 if (list_empty(&sh->lru) &&
333                                     !test_bit(STRIPE_EXPANDING, &sh->state))
334                                         BUG();
335                                 list_del_init(&sh->lru);
336                         }
337                 }
338         } while (sh == NULL);
339
340         if (sh)
341                 atomic_inc(&sh->count);
342
343         spin_unlock_irq(&conf->device_lock);
344         return sh;
345 }
346
347 /* test_and_ack_op() ensures that we only dequeue an operation once */
348 #define test_and_ack_op(op, pend) \
349 do {                                                    \
350         if (test_bit(op, &sh->ops.pending) &&           \
351                 !test_bit(op, &sh->ops.complete)) {     \
352                 if (test_and_set_bit(op, &sh->ops.ack)) \
353                         clear_bit(op, &pend);           \
354                 else                                    \
355                         ack++;                          \
356         } else                                          \
357                 clear_bit(op, &pend);                   \
358 } while (0)
359
360 /* find new work to run, do not resubmit work that is already
361  * in flight
362  */
363 static unsigned long get_stripe_work(struct stripe_head *sh)
364 {
365         unsigned long pending;
366         int ack = 0;
367
368         pending = sh->ops.pending;
369
370         test_and_ack_op(STRIPE_OP_BIOFILL, pending);
371         test_and_ack_op(STRIPE_OP_COMPUTE_BLK, pending);
372         test_and_ack_op(STRIPE_OP_PREXOR, pending);
373         test_and_ack_op(STRIPE_OP_BIODRAIN, pending);
374         test_and_ack_op(STRIPE_OP_POSTXOR, pending);
375         test_and_ack_op(STRIPE_OP_CHECK, pending);
376         if (test_and_clear_bit(STRIPE_OP_IO, &sh->ops.pending))
377                 ack++;
378
379         sh->ops.count -= ack;
380         BUG_ON(sh->ops.count < 0);
381
382         return pending;
383 }
384
385 static int
386 raid5_end_read_request(struct bio *bi, unsigned int bytes_done, int error);
387 static int
388 raid5_end_write_request (struct bio *bi, unsigned int bytes_done, int error);
389
390 static void ops_run_io(struct stripe_head *sh)
391 {
392         raid5_conf_t *conf = sh->raid_conf;
393         int i, disks = sh->disks;
394
395         might_sleep();
396
397         for (i = disks; i--; ) {
398                 int rw;
399                 struct bio *bi;
400                 mdk_rdev_t *rdev;
401                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
402                         rw = WRITE;
403                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
404                         rw = READ;
405                 else
406                         continue;
407
408                 bi = &sh->dev[i].req;
409
410                 bi->bi_rw = rw;
411                 if (rw == WRITE)
412                         bi->bi_end_io = raid5_end_write_request;
413                 else
414                         bi->bi_end_io = raid5_end_read_request;
415
416                 rcu_read_lock();
417                 rdev = rcu_dereference(conf->disks[i].rdev);
418                 if (rdev && test_bit(Faulty, &rdev->flags))
419                         rdev = NULL;
420                 if (rdev)
421                         atomic_inc(&rdev->nr_pending);
422                 rcu_read_unlock();
423
424                 if (rdev) {
425                         if (test_bit(STRIPE_SYNCING, &sh->state) ||
426                                 test_bit(STRIPE_EXPAND_SOURCE, &sh->state) ||
427                                 test_bit(STRIPE_EXPAND_READY, &sh->state))
428                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
429
430                         bi->bi_bdev = rdev->bdev;
431                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
432                                 __FUNCTION__, (unsigned long long)sh->sector,
433                                 bi->bi_rw, i);
434                         atomic_inc(&sh->count);
435                         bi->bi_sector = sh->sector + rdev->data_offset;
436                         bi->bi_flags = 1 << BIO_UPTODATE;
437                         bi->bi_vcnt = 1;
438                         bi->bi_max_vecs = 1;
439                         bi->bi_idx = 0;
440                         bi->bi_io_vec = &sh->dev[i].vec;
441                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
442                         bi->bi_io_vec[0].bv_offset = 0;
443                         bi->bi_size = STRIPE_SIZE;
444                         bi->bi_next = NULL;
445                         if (rw == WRITE &&
446                             test_bit(R5_ReWrite, &sh->dev[i].flags))
447                                 atomic_add(STRIPE_SECTORS,
448                                         &rdev->corrected_errors);
449                         generic_make_request(bi);
450                 } else {
451                         if (rw == WRITE)
452                                 set_bit(STRIPE_DEGRADED, &sh->state);
453                         pr_debug("skip op %ld on disc %d for sector %llu\n",
454                                 bi->bi_rw, i, (unsigned long long)sh->sector);
455                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
456                         set_bit(STRIPE_HANDLE, &sh->state);
457                 }
458         }
459 }
460
461 static struct dma_async_tx_descriptor *
462 async_copy_data(int frombio, struct bio *bio, struct page *page,
463         sector_t sector, struct dma_async_tx_descriptor *tx)
464 {
465         struct bio_vec *bvl;
466         struct page *bio_page;
467         int i;
468         int page_offset;
469
470         if (bio->bi_sector >= sector)
471                 page_offset = (signed)(bio->bi_sector - sector) * 512;
472         else
473                 page_offset = (signed)(sector - bio->bi_sector) * -512;
474         bio_for_each_segment(bvl, bio, i) {
475                 int len = bio_iovec_idx(bio, i)->bv_len;
476                 int clen;
477                 int b_offset = 0;
478
479                 if (page_offset < 0) {
480                         b_offset = -page_offset;
481                         page_offset += b_offset;
482                         len -= b_offset;
483                 }
484
485                 if (len > 0 && page_offset + len > STRIPE_SIZE)
486                         clen = STRIPE_SIZE - page_offset;
487                 else
488                         clen = len;
489
490                 if (clen > 0) {
491                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
492                         bio_page = bio_iovec_idx(bio, i)->bv_page;
493                         if (frombio)
494                                 tx = async_memcpy(page, bio_page, page_offset,
495                                         b_offset, clen,
496                                         ASYNC_TX_DEP_ACK | ASYNC_TX_KMAP_SRC,
497                                         tx, NULL, NULL);
498                         else
499                                 tx = async_memcpy(bio_page, page, b_offset,
500                                         page_offset, clen,
501                                         ASYNC_TX_DEP_ACK | ASYNC_TX_KMAP_DST,
502                                         tx, NULL, NULL);
503                 }
504                 if (clen < len) /* hit end of page */
505                         break;
506                 page_offset +=  len;
507         }
508
509         return tx;
510 }
511
512 static void ops_complete_biofill(void *stripe_head_ref)
513 {
514         struct stripe_head *sh = stripe_head_ref;
515         struct bio *return_bi = NULL;
516         raid5_conf_t *conf = sh->raid_conf;
517         int i, more_to_read = 0;
518
519         pr_debug("%s: stripe %llu\n", __FUNCTION__,
520                 (unsigned long long)sh->sector);
521
522         /* clear completed biofills */
523         for (i = sh->disks; i--; ) {
524                 struct r5dev *dev = &sh->dev[i];
525                 /* check if this stripe has new incoming reads */
526                 if (dev->toread)
527                         more_to_read++;
528
529                 /* acknowledge completion of a biofill operation */
530                 /* and check if we need to reply to a read request
531                 */
532                 if (test_bit(R5_Wantfill, &dev->flags) && !dev->toread) {
533                         struct bio *rbi, *rbi2;
534                         clear_bit(R5_Wantfill, &dev->flags);
535
536                         /* The access to dev->read is outside of the
537                          * spin_lock_irq(&conf->device_lock), but is protected
538                          * by the STRIPE_OP_BIOFILL pending bit
539                          */
540                         BUG_ON(!dev->read);
541                         rbi = dev->read;
542                         dev->read = NULL;
543                         while (rbi && rbi->bi_sector <
544                                 dev->sector + STRIPE_SECTORS) {
545                                 rbi2 = r5_next_bio(rbi, dev->sector);
546                                 spin_lock_irq(&conf->device_lock);
547                                 if (--rbi->bi_phys_segments == 0) {
548                                         rbi->bi_next = return_bi;
549                                         return_bi = rbi;
550                                 }
551                                 spin_unlock_irq(&conf->device_lock);
552                                 rbi = rbi2;
553                         }
554                 }
555         }
556         clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack);
557         clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending);
558
559         return_io(return_bi);
560
561         if (more_to_read)
562                 set_bit(STRIPE_HANDLE, &sh->state);
563         release_stripe(sh);
564 }
565
566 static void ops_run_biofill(struct stripe_head *sh)
567 {
568         struct dma_async_tx_descriptor *tx = NULL;
569         raid5_conf_t *conf = sh->raid_conf;
570         int i;
571
572         pr_debug("%s: stripe %llu\n", __FUNCTION__,
573                 (unsigned long long)sh->sector);
574
575         for (i = sh->disks; i--; ) {
576                 struct r5dev *dev = &sh->dev[i];
577                 if (test_bit(R5_Wantfill, &dev->flags)) {
578                         struct bio *rbi;
579                         spin_lock_irq(&conf->device_lock);
580                         dev->read = rbi = dev->toread;
581                         dev->toread = NULL;
582                         spin_unlock_irq(&conf->device_lock);
583                         while (rbi && rbi->bi_sector <
584                                 dev->sector + STRIPE_SECTORS) {
585                                 tx = async_copy_data(0, rbi, dev->page,
586                                         dev->sector, tx);
587                                 rbi = r5_next_bio(rbi, dev->sector);
588                         }
589                 }
590         }
591
592         atomic_inc(&sh->count);
593         async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
594                 ops_complete_biofill, sh);
595 }
596
597 static void ops_complete_compute5(void *stripe_head_ref)
598 {
599         struct stripe_head *sh = stripe_head_ref;
600         int target = sh->ops.target;
601         struct r5dev *tgt = &sh->dev[target];
602
603         pr_debug("%s: stripe %llu\n", __FUNCTION__,
604                 (unsigned long long)sh->sector);
605
606         set_bit(R5_UPTODATE, &tgt->flags);
607         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
608         clear_bit(R5_Wantcompute, &tgt->flags);
609         set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
610         set_bit(STRIPE_HANDLE, &sh->state);
611         release_stripe(sh);
612 }
613
614 static struct dma_async_tx_descriptor *
615 ops_run_compute5(struct stripe_head *sh, unsigned long pending)
616 {
617         /* kernel stack size limits the total number of disks */
618         int disks = sh->disks;
619         struct page *xor_srcs[disks];
620         int target = sh->ops.target;
621         struct r5dev *tgt = &sh->dev[target];
622         struct page *xor_dest = tgt->page;
623         int count = 0;
624         struct dma_async_tx_descriptor *tx;
625         int i;
626
627         pr_debug("%s: stripe %llu block: %d\n",
628                 __FUNCTION__, (unsigned long long)sh->sector, target);
629         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
630
631         for (i = disks; i--; )
632                 if (i != target)
633                         xor_srcs[count++] = sh->dev[i].page;
634
635         atomic_inc(&sh->count);
636
637         if (unlikely(count == 1))
638                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
639                         0, NULL, ops_complete_compute5, sh);
640         else
641                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
642                         ASYNC_TX_XOR_ZERO_DST, NULL,
643                         ops_complete_compute5, sh);
644
645         /* ack now if postxor is not set to be run */
646         if (tx && !test_bit(STRIPE_OP_POSTXOR, &pending))
647                 async_tx_ack(tx);
648
649         return tx;
650 }
651
652 static void ops_complete_prexor(void *stripe_head_ref)
653 {
654         struct stripe_head *sh = stripe_head_ref;
655
656         pr_debug("%s: stripe %llu\n", __FUNCTION__,
657                 (unsigned long long)sh->sector);
658
659         set_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
660 }
661
662 static struct dma_async_tx_descriptor *
663 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
664 {
665         /* kernel stack size limits the total number of disks */
666         int disks = sh->disks;
667         struct page *xor_srcs[disks];
668         int count = 0, pd_idx = sh->pd_idx, i;
669
670         /* existing parity data subtracted */
671         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
672
673         pr_debug("%s: stripe %llu\n", __FUNCTION__,
674                 (unsigned long long)sh->sector);
675
676         for (i = disks; i--; ) {
677                 struct r5dev *dev = &sh->dev[i];
678                 /* Only process blocks that are known to be uptodate */
679                 if (dev->towrite && test_bit(R5_Wantprexor, &dev->flags))
680                         xor_srcs[count++] = dev->page;
681         }
682
683         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
684                 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
685                 ops_complete_prexor, sh);
686
687         return tx;
688 }
689
690 static struct dma_async_tx_descriptor *
691 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
692 {
693         int disks = sh->disks;
694         int pd_idx = sh->pd_idx, i;
695
696         /* check if prexor is active which means only process blocks
697          * that are part of a read-modify-write (Wantprexor)
698          */
699         int prexor = test_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
700
701         pr_debug("%s: stripe %llu\n", __FUNCTION__,
702                 (unsigned long long)sh->sector);
703
704         for (i = disks; i--; ) {
705                 struct r5dev *dev = &sh->dev[i];
706                 struct bio *chosen;
707                 int towrite;
708
709                 towrite = 0;
710                 if (prexor) { /* rmw */
711                         if (dev->towrite &&
712                             test_bit(R5_Wantprexor, &dev->flags))
713                                 towrite = 1;
714                 } else { /* rcw */
715                         if (i != pd_idx && dev->towrite &&
716                                 test_bit(R5_LOCKED, &dev->flags))
717                                 towrite = 1;
718                 }
719
720                 if (towrite) {
721                         struct bio *wbi;
722
723                         spin_lock(&sh->lock);
724                         chosen = dev->towrite;
725                         dev->towrite = NULL;
726                         BUG_ON(dev->written);
727                         wbi = dev->written = chosen;
728                         spin_unlock(&sh->lock);
729
730                         while (wbi && wbi->bi_sector <
731                                 dev->sector + STRIPE_SECTORS) {
732                                 tx = async_copy_data(1, wbi, dev->page,
733                                         dev->sector, tx);
734                                 wbi = r5_next_bio(wbi, dev->sector);
735                         }
736                 }
737         }
738
739         return tx;
740 }
741
742 static void ops_complete_postxor(void *stripe_head_ref)
743 {
744         struct stripe_head *sh = stripe_head_ref;
745
746         pr_debug("%s: stripe %llu\n", __FUNCTION__,
747                 (unsigned long long)sh->sector);
748
749         set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
750         set_bit(STRIPE_HANDLE, &sh->state);
751         release_stripe(sh);
752 }
753
754 static void ops_complete_write(void *stripe_head_ref)
755 {
756         struct stripe_head *sh = stripe_head_ref;
757         int disks = sh->disks, i, pd_idx = sh->pd_idx;
758
759         pr_debug("%s: stripe %llu\n", __FUNCTION__,
760                 (unsigned long long)sh->sector);
761
762         for (i = disks; i--; ) {
763                 struct r5dev *dev = &sh->dev[i];
764                 if (dev->written || i == pd_idx)
765                         set_bit(R5_UPTODATE, &dev->flags);
766         }
767
768         set_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
769         set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
770
771         set_bit(STRIPE_HANDLE, &sh->state);
772         release_stripe(sh);
773 }
774
775 static void
776 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
777 {
778         /* kernel stack size limits the total number of disks */
779         int disks = sh->disks;
780         struct page *xor_srcs[disks];
781
782         int count = 0, pd_idx = sh->pd_idx, i;
783         struct page *xor_dest;
784         int prexor = test_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
785         unsigned long flags;
786         dma_async_tx_callback callback;
787
788         pr_debug("%s: stripe %llu\n", __FUNCTION__,
789                 (unsigned long long)sh->sector);
790
791         /* check if prexor is active which means only process blocks
792          * that are part of a read-modify-write (written)
793          */
794         if (prexor) {
795                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
796                 for (i = disks; i--; ) {
797                         struct r5dev *dev = &sh->dev[i];
798                         if (dev->written)
799                                 xor_srcs[count++] = dev->page;
800                 }
801         } else {
802                 xor_dest = sh->dev[pd_idx].page;
803                 for (i = disks; i--; ) {
804                         struct r5dev *dev = &sh->dev[i];
805                         if (i != pd_idx)
806                                 xor_srcs[count++] = dev->page;
807                 }
808         }
809
810         /* check whether this postxor is part of a write */
811         callback = test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending) ?
812                 ops_complete_write : ops_complete_postxor;
813
814         /* 1/ if we prexor'd then the dest is reused as a source
815          * 2/ if we did not prexor then we are redoing the parity
816          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
817          * for the synchronous xor case
818          */
819         flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
820                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
821
822         atomic_inc(&sh->count);
823
824         if (unlikely(count == 1)) {
825                 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
826                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
827                         flags, tx, callback, sh);
828         } else
829                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
830                         flags, tx, callback, sh);
831 }
832
833 static void ops_complete_check(void *stripe_head_ref)
834 {
835         struct stripe_head *sh = stripe_head_ref;
836         int pd_idx = sh->pd_idx;
837
838         pr_debug("%s: stripe %llu\n", __FUNCTION__,
839                 (unsigned long long)sh->sector);
840
841         if (test_and_clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending) &&
842                 sh->ops.zero_sum_result == 0)
843                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
844
845         set_bit(STRIPE_OP_CHECK, &sh->ops.complete);
846         set_bit(STRIPE_HANDLE, &sh->state);
847         release_stripe(sh);
848 }
849
850 static void ops_run_check(struct stripe_head *sh)
851 {
852         /* kernel stack size limits the total number of disks */
853         int disks = sh->disks;
854         struct page *xor_srcs[disks];
855         struct dma_async_tx_descriptor *tx;
856
857         int count = 0, pd_idx = sh->pd_idx, i;
858         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
859
860         pr_debug("%s: stripe %llu\n", __FUNCTION__,
861                 (unsigned long long)sh->sector);
862
863         for (i = disks; i--; ) {
864                 struct r5dev *dev = &sh->dev[i];
865                 if (i != pd_idx)
866                         xor_srcs[count++] = dev->page;
867         }
868
869         tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
870                 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
871
872         if (tx)
873                 set_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
874         else
875                 clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
876
877         atomic_inc(&sh->count);
878         tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
879                 ops_complete_check, sh);
880 }
881
882 static void raid5_run_ops(struct stripe_head *sh, unsigned long pending)
883 {
884         int overlap_clear = 0, i, disks = sh->disks;
885         struct dma_async_tx_descriptor *tx = NULL;
886
887         if (test_bit(STRIPE_OP_BIOFILL, &pending)) {
888                 ops_run_biofill(sh);
889                 overlap_clear++;
890         }
891
892         if (test_bit(STRIPE_OP_COMPUTE_BLK, &pending))
893                 tx = ops_run_compute5(sh, pending);
894
895         if (test_bit(STRIPE_OP_PREXOR, &pending))
896                 tx = ops_run_prexor(sh, tx);
897
898         if (test_bit(STRIPE_OP_BIODRAIN, &pending)) {
899                 tx = ops_run_biodrain(sh, tx);
900                 overlap_clear++;
901         }
902
903         if (test_bit(STRIPE_OP_POSTXOR, &pending))
904                 ops_run_postxor(sh, tx);
905
906         if (test_bit(STRIPE_OP_CHECK, &pending))
907                 ops_run_check(sh);
908
909         if (test_bit(STRIPE_OP_IO, &pending))
910                 ops_run_io(sh);
911
912         if (overlap_clear)
913                 for (i = disks; i--; ) {
914                         struct r5dev *dev = &sh->dev[i];
915                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
916                                 wake_up(&sh->raid_conf->wait_for_overlap);
917                 }
918 }
919
920 static int grow_one_stripe(raid5_conf_t *conf)
921 {
922         struct stripe_head *sh;
923         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
924         if (!sh)
925                 return 0;
926         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
927         sh->raid_conf = conf;
928         spin_lock_init(&sh->lock);
929
930         if (grow_buffers(sh, conf->raid_disks)) {
931                 shrink_buffers(sh, conf->raid_disks);
932                 kmem_cache_free(conf->slab_cache, sh);
933                 return 0;
934         }
935         sh->disks = conf->raid_disks;
936         /* we just created an active stripe so... */
937         atomic_set(&sh->count, 1);
938         atomic_inc(&conf->active_stripes);
939         INIT_LIST_HEAD(&sh->lru);
940         release_stripe(sh);
941         return 1;
942 }
943
944 static int grow_stripes(raid5_conf_t *conf, int num)
945 {
946         struct kmem_cache *sc;
947         int devs = conf->raid_disks;
948
949         sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
950         sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
951         conf->active_name = 0;
952         sc = kmem_cache_create(conf->cache_name[conf->active_name],
953                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
954                                0, 0, NULL, NULL);
955         if (!sc)
956                 return 1;
957         conf->slab_cache = sc;
958         conf->pool_size = devs;
959         while (num--)
960                 if (!grow_one_stripe(conf))
961                         return 1;
962         return 0;
963 }
964
965 #ifdef CONFIG_MD_RAID5_RESHAPE
966 static int resize_stripes(raid5_conf_t *conf, int newsize)
967 {
968         /* Make all the stripes able to hold 'newsize' devices.
969          * New slots in each stripe get 'page' set to a new page.
970          *
971          * This happens in stages:
972          * 1/ create a new kmem_cache and allocate the required number of
973          *    stripe_heads.
974          * 2/ gather all the old stripe_heads and tranfer the pages across
975          *    to the new stripe_heads.  This will have the side effect of
976          *    freezing the array as once all stripe_heads have been collected,
977          *    no IO will be possible.  Old stripe heads are freed once their
978          *    pages have been transferred over, and the old kmem_cache is
979          *    freed when all stripes are done.
980          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
981          *    we simple return a failre status - no need to clean anything up.
982          * 4/ allocate new pages for the new slots in the new stripe_heads.
983          *    If this fails, we don't bother trying the shrink the
984          *    stripe_heads down again, we just leave them as they are.
985          *    As each stripe_head is processed the new one is released into
986          *    active service.
987          *
988          * Once step2 is started, we cannot afford to wait for a write,
989          * so we use GFP_NOIO allocations.
990          */
991         struct stripe_head *osh, *nsh;
992         LIST_HEAD(newstripes);
993         struct disk_info *ndisks;
994         int err = 0;
995         struct kmem_cache *sc;
996         int i;
997
998         if (newsize <= conf->pool_size)
999                 return 0; /* never bother to shrink */
1000
1001         md_allow_write(conf->mddev);
1002
1003         /* Step 1 */
1004         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1005                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1006                                0, 0, NULL, NULL);
1007         if (!sc)
1008                 return -ENOMEM;
1009
1010         for (i = conf->max_nr_stripes; i; i--) {
1011                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1012                 if (!nsh)
1013                         break;
1014
1015                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1016
1017                 nsh->raid_conf = conf;
1018                 spin_lock_init(&nsh->lock);
1019
1020                 list_add(&nsh->lru, &newstripes);
1021         }
1022         if (i) {
1023                 /* didn't get enough, give up */
1024                 while (!list_empty(&newstripes)) {
1025                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1026                         list_del(&nsh->lru);
1027                         kmem_cache_free(sc, nsh);
1028                 }
1029                 kmem_cache_destroy(sc);
1030                 return -ENOMEM;
1031         }
1032         /* Step 2 - Must use GFP_NOIO now.
1033          * OK, we have enough stripes, start collecting inactive
1034          * stripes and copying them over
1035          */
1036         list_for_each_entry(nsh, &newstripes, lru) {
1037                 spin_lock_irq(&conf->device_lock);
1038                 wait_event_lock_irq(conf->wait_for_stripe,
1039                                     !list_empty(&conf->inactive_list),
1040                                     conf->device_lock,
1041                                     unplug_slaves(conf->mddev)
1042                         );
1043                 osh = get_free_stripe(conf);
1044                 spin_unlock_irq(&conf->device_lock);
1045                 atomic_set(&nsh->count, 1);
1046                 for(i=0; i<conf->pool_size; i++)
1047                         nsh->dev[i].page = osh->dev[i].page;
1048                 for( ; i<newsize; i++)
1049                         nsh->dev[i].page = NULL;
1050                 kmem_cache_free(conf->slab_cache, osh);
1051         }
1052         kmem_cache_destroy(conf->slab_cache);
1053
1054         /* Step 3.
1055          * At this point, we are holding all the stripes so the array
1056          * is completely stalled, so now is a good time to resize
1057          * conf->disks.
1058          */
1059         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1060         if (ndisks) {
1061                 for (i=0; i<conf->raid_disks; i++)
1062                         ndisks[i] = conf->disks[i];
1063                 kfree(conf->disks);
1064                 conf->disks = ndisks;
1065         } else
1066                 err = -ENOMEM;
1067
1068         /* Step 4, return new stripes to service */
1069         while(!list_empty(&newstripes)) {
1070                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1071                 list_del_init(&nsh->lru);
1072                 for (i=conf->raid_disks; i < newsize; i++)
1073                         if (nsh->dev[i].page == NULL) {
1074                                 struct page *p = alloc_page(GFP_NOIO);
1075                                 nsh->dev[i].page = p;
1076                                 if (!p)
1077                                         err = -ENOMEM;
1078                         }
1079                 release_stripe(nsh);
1080         }
1081         /* critical section pass, GFP_NOIO no longer needed */
1082
1083         conf->slab_cache = sc;
1084         conf->active_name = 1-conf->active_name;
1085         conf->pool_size = newsize;
1086         return err;
1087 }
1088 #endif
1089
1090 static int drop_one_stripe(raid5_conf_t *conf)
1091 {
1092         struct stripe_head *sh;
1093
1094         spin_lock_irq(&conf->device_lock);
1095         sh = get_free_stripe(conf);
1096         spin_unlock_irq(&conf->device_lock);
1097         if (!sh)
1098                 return 0;
1099         BUG_ON(atomic_read(&sh->count));
1100         shrink_buffers(sh, conf->pool_size);
1101         kmem_cache_free(conf->slab_cache, sh);
1102         atomic_dec(&conf->active_stripes);
1103         return 1;
1104 }
1105
1106 static void shrink_stripes(raid5_conf_t *conf)
1107 {
1108         while (drop_one_stripe(conf))
1109                 ;
1110
1111         if (conf->slab_cache)
1112                 kmem_cache_destroy(conf->slab_cache);
1113         conf->slab_cache = NULL;
1114 }
1115
1116 static int raid5_end_read_request(struct bio * bi, unsigned int bytes_done,
1117                                    int error)
1118 {
1119         struct stripe_head *sh = bi->bi_private;
1120         raid5_conf_t *conf = sh->raid_conf;
1121         int disks = sh->disks, i;
1122         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1123         char b[BDEVNAME_SIZE];
1124         mdk_rdev_t *rdev;
1125
1126         if (bi->bi_size)
1127                 return 1;
1128
1129         for (i=0 ; i<disks; i++)
1130                 if (bi == &sh->dev[i].req)
1131                         break;
1132
1133         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1134                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1135                 uptodate);
1136         if (i == disks) {
1137                 BUG();
1138                 return 0;
1139         }
1140
1141         if (uptodate) {
1142                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1143                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1144                         rdev = conf->disks[i].rdev;
1145                         printk(KERN_INFO "raid5:%s: read error corrected (%lu sectors at %llu on %s)\n",
1146                                mdname(conf->mddev), STRIPE_SECTORS,
1147                                (unsigned long long)sh->sector + rdev->data_offset,
1148                                bdevname(rdev->bdev, b));
1149                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1150                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1151                 }
1152                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1153                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1154         } else {
1155                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1156                 int retry = 0;
1157                 rdev = conf->disks[i].rdev;
1158
1159                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1160                 atomic_inc(&rdev->read_errors);
1161                 if (conf->mddev->degraded)
1162                         printk(KERN_WARNING "raid5:%s: read error not correctable (sector %llu on %s).\n",
1163                                mdname(conf->mddev),
1164                                (unsigned long long)sh->sector + rdev->data_offset,
1165                                bdn);
1166                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1167                         /* Oh, no!!! */
1168                         printk(KERN_WARNING "raid5:%s: read error NOT corrected!! (sector %llu on %s).\n",
1169                                mdname(conf->mddev),
1170                                (unsigned long long)sh->sector + rdev->data_offset,
1171                                bdn);
1172                 else if (atomic_read(&rdev->read_errors)
1173                          > conf->max_nr_stripes)
1174                         printk(KERN_WARNING
1175                                "raid5:%s: Too many read errors, failing device %s.\n",
1176                                mdname(conf->mddev), bdn);
1177                 else
1178                         retry = 1;
1179                 if (retry)
1180                         set_bit(R5_ReadError, &sh->dev[i].flags);
1181                 else {
1182                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1183                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1184                         md_error(conf->mddev, rdev);
1185                 }
1186         }
1187         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1188         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1189         set_bit(STRIPE_HANDLE, &sh->state);
1190         release_stripe(sh);
1191         return 0;
1192 }
1193
1194 static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
1195                                     int error)
1196 {
1197         struct stripe_head *sh = bi->bi_private;
1198         raid5_conf_t *conf = sh->raid_conf;
1199         int disks = sh->disks, i;
1200         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1201
1202         if (bi->bi_size)
1203                 return 1;
1204
1205         for (i=0 ; i<disks; i++)
1206                 if (bi == &sh->dev[i].req)
1207                         break;
1208
1209         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1210                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1211                 uptodate);
1212         if (i == disks) {
1213                 BUG();
1214                 return 0;
1215         }
1216
1217         if (!uptodate)
1218                 md_error(conf->mddev, conf->disks[i].rdev);
1219
1220         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1221         
1222         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1223         set_bit(STRIPE_HANDLE, &sh->state);
1224         release_stripe(sh);
1225         return 0;
1226 }
1227
1228
1229 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1230         
1231 static void raid5_build_block (struct stripe_head *sh, int i)
1232 {
1233         struct r5dev *dev = &sh->dev[i];
1234
1235         bio_init(&dev->req);
1236         dev->req.bi_io_vec = &dev->vec;
1237         dev->req.bi_vcnt++;
1238         dev->req.bi_max_vecs++;
1239         dev->vec.bv_page = dev->page;
1240         dev->vec.bv_len = STRIPE_SIZE;
1241         dev->vec.bv_offset = 0;
1242
1243         dev->req.bi_sector = sh->sector;
1244         dev->req.bi_private = sh;
1245
1246         dev->flags = 0;
1247         dev->sector = compute_blocknr(sh, i);
1248 }
1249
1250 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1251 {
1252         char b[BDEVNAME_SIZE];
1253         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1254         pr_debug("raid5: error called\n");
1255
1256         if (!test_bit(Faulty, &rdev->flags)) {
1257                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1258                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1259                         unsigned long flags;
1260                         spin_lock_irqsave(&conf->device_lock, flags);
1261                         mddev->degraded++;
1262                         spin_unlock_irqrestore(&conf->device_lock, flags);
1263                         /*
1264                          * if recovery was running, make sure it aborts.
1265                          */
1266                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
1267                 }
1268                 set_bit(Faulty, &rdev->flags);
1269                 printk (KERN_ALERT
1270                         "raid5: Disk failure on %s, disabling device."
1271                         " Operation continuing on %d devices\n",
1272                         bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1273         }
1274 }
1275
1276 /*
1277  * Input: a 'big' sector number,
1278  * Output: index of the data and parity disk, and the sector # in them.
1279  */
1280 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1281                         unsigned int data_disks, unsigned int * dd_idx,
1282                         unsigned int * pd_idx, raid5_conf_t *conf)
1283 {
1284         long stripe;
1285         unsigned long chunk_number;
1286         unsigned int chunk_offset;
1287         sector_t new_sector;
1288         int sectors_per_chunk = conf->chunk_size >> 9;
1289
1290         /* First compute the information on this sector */
1291
1292         /*
1293          * Compute the chunk number and the sector offset inside the chunk
1294          */
1295         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1296         chunk_number = r_sector;
1297         BUG_ON(r_sector != chunk_number);
1298
1299         /*
1300          * Compute the stripe number
1301          */
1302         stripe = chunk_number / data_disks;
1303
1304         /*
1305          * Compute the data disk and parity disk indexes inside the stripe
1306          */
1307         *dd_idx = chunk_number % data_disks;
1308
1309         /*
1310          * Select the parity disk based on the user selected algorithm.
1311          */
1312         switch(conf->level) {
1313         case 4:
1314                 *pd_idx = data_disks;
1315                 break;
1316         case 5:
1317                 switch (conf->algorithm) {
1318                 case ALGORITHM_LEFT_ASYMMETRIC:
1319                         *pd_idx = data_disks - stripe % raid_disks;
1320                         if (*dd_idx >= *pd_idx)
1321                                 (*dd_idx)++;
1322                         break;
1323                 case ALGORITHM_RIGHT_ASYMMETRIC:
1324                         *pd_idx = stripe % raid_disks;
1325                         if (*dd_idx >= *pd_idx)
1326                                 (*dd_idx)++;
1327                         break;
1328                 case ALGORITHM_LEFT_SYMMETRIC:
1329                         *pd_idx = data_disks - stripe % raid_disks;
1330                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1331                         break;
1332                 case ALGORITHM_RIGHT_SYMMETRIC:
1333                         *pd_idx = stripe % raid_disks;
1334                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1335                         break;
1336                 default:
1337                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1338                                 conf->algorithm);
1339                 }
1340                 break;
1341         case 6:
1342
1343                 /**** FIX THIS ****/
1344                 switch (conf->algorithm) {
1345                 case ALGORITHM_LEFT_ASYMMETRIC:
1346                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1347                         if (*pd_idx == raid_disks-1)
1348                                 (*dd_idx)++;    /* Q D D D P */
1349                         else if (*dd_idx >= *pd_idx)
1350                                 (*dd_idx) += 2; /* D D P Q D */
1351                         break;
1352                 case ALGORITHM_RIGHT_ASYMMETRIC:
1353                         *pd_idx = stripe % raid_disks;
1354                         if (*pd_idx == raid_disks-1)
1355                                 (*dd_idx)++;    /* Q D D D P */
1356                         else if (*dd_idx >= *pd_idx)
1357                                 (*dd_idx) += 2; /* D D P Q D */
1358                         break;
1359                 case ALGORITHM_LEFT_SYMMETRIC:
1360                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1361                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1362                         break;
1363                 case ALGORITHM_RIGHT_SYMMETRIC:
1364                         *pd_idx = stripe % raid_disks;
1365                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1366                         break;
1367                 default:
1368                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1369                                 conf->algorithm);
1370                 }
1371                 break;
1372         }
1373
1374         /*
1375          * Finally, compute the new sector number
1376          */
1377         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1378         return new_sector;
1379 }
1380
1381
1382 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1383 {
1384         raid5_conf_t *conf = sh->raid_conf;
1385         int raid_disks = sh->disks;
1386         int data_disks = raid_disks - conf->max_degraded;
1387         sector_t new_sector = sh->sector, check;
1388         int sectors_per_chunk = conf->chunk_size >> 9;
1389         sector_t stripe;
1390         int chunk_offset;
1391         int chunk_number, dummy1, dummy2, dd_idx = i;
1392         sector_t r_sector;
1393
1394
1395         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1396         stripe = new_sector;
1397         BUG_ON(new_sector != stripe);
1398
1399         if (i == sh->pd_idx)
1400                 return 0;
1401         switch(conf->level) {
1402         case 4: break;
1403         case 5:
1404                 switch (conf->algorithm) {
1405                 case ALGORITHM_LEFT_ASYMMETRIC:
1406                 case ALGORITHM_RIGHT_ASYMMETRIC:
1407                         if (i > sh->pd_idx)
1408                                 i--;
1409                         break;
1410                 case ALGORITHM_LEFT_SYMMETRIC:
1411                 case ALGORITHM_RIGHT_SYMMETRIC:
1412                         if (i < sh->pd_idx)
1413                                 i += raid_disks;
1414                         i -= (sh->pd_idx + 1);
1415                         break;
1416                 default:
1417                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1418                                conf->algorithm);
1419                 }
1420                 break;
1421         case 6:
1422                 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1423                         return 0; /* It is the Q disk */
1424                 switch (conf->algorithm) {
1425                 case ALGORITHM_LEFT_ASYMMETRIC:
1426                 case ALGORITHM_RIGHT_ASYMMETRIC:
1427                         if (sh->pd_idx == raid_disks-1)
1428                                 i--;    /* Q D D D P */
1429                         else if (i > sh->pd_idx)
1430                                 i -= 2; /* D D P Q D */
1431                         break;
1432                 case ALGORITHM_LEFT_SYMMETRIC:
1433                 case ALGORITHM_RIGHT_SYMMETRIC:
1434                         if (sh->pd_idx == raid_disks-1)
1435                                 i--; /* Q D D D P */
1436                         else {
1437                                 /* D D P Q D */
1438                                 if (i < sh->pd_idx)
1439                                         i += raid_disks;
1440                                 i -= (sh->pd_idx + 2);
1441                         }
1442                         break;
1443                 default:
1444                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1445                                 conf->algorithm);
1446                 }
1447                 break;
1448         }
1449
1450         chunk_number = stripe * data_disks + i;
1451         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1452
1453         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1454         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1455                 printk(KERN_ERR "compute_blocknr: map not correct\n");
1456                 return 0;
1457         }
1458         return r_sector;
1459 }
1460
1461
1462
1463 /*
1464  * Copy data between a page in the stripe cache, and one or more bion
1465  * The page could align with the middle of the bio, or there could be
1466  * several bion, each with several bio_vecs, which cover part of the page
1467  * Multiple bion are linked together on bi_next.  There may be extras
1468  * at the end of this list.  We ignore them.
1469  */
1470 static void copy_data(int frombio, struct bio *bio,
1471                      struct page *page,
1472                      sector_t sector)
1473 {
1474         char *pa = page_address(page);
1475         struct bio_vec *bvl;
1476         int i;
1477         int page_offset;
1478
1479         if (bio->bi_sector >= sector)
1480                 page_offset = (signed)(bio->bi_sector - sector) * 512;
1481         else
1482                 page_offset = (signed)(sector - bio->bi_sector) * -512;
1483         bio_for_each_segment(bvl, bio, i) {
1484                 int len = bio_iovec_idx(bio,i)->bv_len;
1485                 int clen;
1486                 int b_offset = 0;
1487
1488                 if (page_offset < 0) {
1489                         b_offset = -page_offset;
1490                         page_offset += b_offset;
1491                         len -= b_offset;
1492                 }
1493
1494                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1495                         clen = STRIPE_SIZE - page_offset;
1496                 else clen = len;
1497
1498                 if (clen > 0) {
1499                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1500                         if (frombio)
1501                                 memcpy(pa+page_offset, ba+b_offset, clen);
1502                         else
1503                                 memcpy(ba+b_offset, pa+page_offset, clen);
1504                         __bio_kunmap_atomic(ba, KM_USER0);
1505                 }
1506                 if (clen < len) /* hit end of page */
1507                         break;
1508                 page_offset +=  len;
1509         }
1510 }
1511
1512 #define check_xor()     do {                                              \
1513                                 if (count == MAX_XOR_BLOCKS) {            \
1514                                 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1515                                 count = 0;                                \
1516                            }                                              \
1517                         } while(0)
1518
1519
1520 static void compute_block(struct stripe_head *sh, int dd_idx)
1521 {
1522         int i, count, disks = sh->disks;
1523         void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1524
1525         pr_debug("compute_block, stripe %llu, idx %d\n",
1526                 (unsigned long long)sh->sector, dd_idx);
1527
1528         dest = page_address(sh->dev[dd_idx].page);
1529         memset(dest, 0, STRIPE_SIZE);
1530         count = 0;
1531         for (i = disks ; i--; ) {
1532                 if (i == dd_idx)
1533                         continue;
1534                 p = page_address(sh->dev[i].page);
1535                 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1536                         ptr[count++] = p;
1537                 else
1538                         printk(KERN_ERR "compute_block() %d, stripe %llu, %d"
1539                                 " not present\n", dd_idx,
1540                                 (unsigned long long)sh->sector, i);
1541
1542                 check_xor();
1543         }
1544         if (count)
1545                 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1546         set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1547 }
1548
1549 static void compute_parity5(struct stripe_head *sh, int method)
1550 {
1551         raid5_conf_t *conf = sh->raid_conf;
1552         int i, pd_idx = sh->pd_idx, disks = sh->disks, count;
1553         void *ptr[MAX_XOR_BLOCKS], *dest;
1554         struct bio *chosen;
1555
1556         pr_debug("compute_parity5, stripe %llu, method %d\n",
1557                 (unsigned long long)sh->sector, method);
1558
1559         count = 0;
1560         dest = page_address(sh->dev[pd_idx].page);
1561         switch(method) {
1562         case READ_MODIFY_WRITE:
1563                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags));
1564                 for (i=disks ; i-- ;) {
1565                         if (i==pd_idx)
1566                                 continue;
1567                         if (sh->dev[i].towrite &&
1568                             test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
1569                                 ptr[count++] = page_address(sh->dev[i].page);
1570                                 chosen = sh->dev[i].towrite;
1571                                 sh->dev[i].towrite = NULL;
1572
1573                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1574                                         wake_up(&conf->wait_for_overlap);
1575
1576                                 BUG_ON(sh->dev[i].written);
1577                                 sh->dev[i].written = chosen;
1578                                 check_xor();
1579                         }
1580                 }
1581                 break;
1582         case RECONSTRUCT_WRITE:
1583                 memset(dest, 0, STRIPE_SIZE);
1584                 for (i= disks; i-- ;)
1585                         if (i!=pd_idx && sh->dev[i].towrite) {
1586                                 chosen = sh->dev[i].towrite;
1587                                 sh->dev[i].towrite = NULL;
1588
1589                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1590                                         wake_up(&conf->wait_for_overlap);
1591
1592                                 BUG_ON(sh->dev[i].written);
1593                                 sh->dev[i].written = chosen;
1594                         }
1595                 break;
1596         case CHECK_PARITY:
1597                 break;
1598         }
1599         if (count) {
1600                 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1601                 count = 0;
1602         }
1603         
1604         for (i = disks; i--;)
1605                 if (sh->dev[i].written) {
1606                         sector_t sector = sh->dev[i].sector;
1607                         struct bio *wbi = sh->dev[i].written;
1608                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1609                                 copy_data(1, wbi, sh->dev[i].page, sector);
1610                                 wbi = r5_next_bio(wbi, sector);
1611                         }
1612
1613                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1614                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1615                 }
1616
1617         switch(method) {
1618         case RECONSTRUCT_WRITE:
1619         case CHECK_PARITY:
1620                 for (i=disks; i--;)
1621                         if (i != pd_idx) {
1622                                 ptr[count++] = page_address(sh->dev[i].page);
1623                                 check_xor();
1624                         }
1625                 break;
1626         case READ_MODIFY_WRITE:
1627                 for (i = disks; i--;)
1628                         if (sh->dev[i].written) {
1629                                 ptr[count++] = page_address(sh->dev[i].page);
1630                                 check_xor();
1631                         }
1632         }
1633         if (count)
1634                 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1635
1636         if (method != CHECK_PARITY) {
1637                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1638                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1639         } else
1640                 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1641 }
1642
1643 static void compute_parity6(struct stripe_head *sh, int method)
1644 {
1645         raid6_conf_t *conf = sh->raid_conf;
1646         int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1647         struct bio *chosen;
1648         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1649         void *ptrs[disks];
1650
1651         qd_idx = raid6_next_disk(pd_idx, disks);
1652         d0_idx = raid6_next_disk(qd_idx, disks);
1653
1654         pr_debug("compute_parity, stripe %llu, method %d\n",
1655                 (unsigned long long)sh->sector, method);
1656
1657         switch(method) {
1658         case READ_MODIFY_WRITE:
1659                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
1660         case RECONSTRUCT_WRITE:
1661                 for (i= disks; i-- ;)
1662                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1663                                 chosen = sh->dev[i].towrite;
1664                                 sh->dev[i].towrite = NULL;
1665
1666                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1667                                         wake_up(&conf->wait_for_overlap);
1668
1669                                 BUG_ON(sh->dev[i].written);
1670                                 sh->dev[i].written = chosen;
1671                         }
1672                 break;
1673         case CHECK_PARITY:
1674                 BUG();          /* Not implemented yet */
1675         }
1676
1677         for (i = disks; i--;)
1678                 if (sh->dev[i].written) {
1679                         sector_t sector = sh->dev[i].sector;
1680                         struct bio *wbi = sh->dev[i].written;
1681                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1682                                 copy_data(1, wbi, sh->dev[i].page, sector);
1683                                 wbi = r5_next_bio(wbi, sector);
1684                         }
1685
1686                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1687                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1688                 }
1689
1690 //      switch(method) {
1691 //      case RECONSTRUCT_WRITE:
1692 //      case CHECK_PARITY:
1693 //      case UPDATE_PARITY:
1694                 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1695                 /* FIX: Is this ordering of drives even remotely optimal? */
1696                 count = 0;
1697                 i = d0_idx;
1698                 do {
1699                         ptrs[count++] = page_address(sh->dev[i].page);
1700                         if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1701                                 printk("block %d/%d not uptodate on parity calc\n", i,count);
1702                         i = raid6_next_disk(i, disks);
1703                 } while ( i != d0_idx );
1704 //              break;
1705 //      }
1706
1707         raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1708
1709         switch(method) {
1710         case RECONSTRUCT_WRITE:
1711                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1712                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1713                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1714                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
1715                 break;
1716         case UPDATE_PARITY:
1717                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1718                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1719                 break;
1720         }
1721 }
1722
1723
1724 /* Compute one missing block */
1725 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1726 {
1727         int i, count, disks = sh->disks;
1728         void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1729         int pd_idx = sh->pd_idx;
1730         int qd_idx = raid6_next_disk(pd_idx, disks);
1731
1732         pr_debug("compute_block_1, stripe %llu, idx %d\n",
1733                 (unsigned long long)sh->sector, dd_idx);
1734
1735         if ( dd_idx == qd_idx ) {
1736                 /* We're actually computing the Q drive */
1737                 compute_parity6(sh, UPDATE_PARITY);
1738         } else {
1739                 dest = page_address(sh->dev[dd_idx].page);
1740                 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1741                 count = 0;
1742                 for (i = disks ; i--; ) {
1743                         if (i == dd_idx || i == qd_idx)
1744                                 continue;
1745                         p = page_address(sh->dev[i].page);
1746                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1747                                 ptr[count++] = p;
1748                         else
1749                                 printk("compute_block() %d, stripe %llu, %d"
1750                                        " not present\n", dd_idx,
1751                                        (unsigned long long)sh->sector, i);
1752
1753                         check_xor();
1754                 }
1755                 if (count)
1756                         xor_blocks(count, STRIPE_SIZE, dest, ptr);
1757                 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1758                 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1759         }
1760 }
1761
1762 /* Compute two missing blocks */
1763 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1764 {
1765         int i, count, disks = sh->disks;
1766         int pd_idx = sh->pd_idx;
1767         int qd_idx = raid6_next_disk(pd_idx, disks);
1768         int d0_idx = raid6_next_disk(qd_idx, disks);
1769         int faila, failb;
1770
1771         /* faila and failb are disk numbers relative to d0_idx */
1772         /* pd_idx become disks-2 and qd_idx become disks-1 */
1773         faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1774         failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1775
1776         BUG_ON(faila == failb);
1777         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1778
1779         pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1780                (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1781
1782         if ( failb == disks-1 ) {
1783                 /* Q disk is one of the missing disks */
1784                 if ( faila == disks-2 ) {
1785                         /* Missing P+Q, just recompute */
1786                         compute_parity6(sh, UPDATE_PARITY);
1787                         return;
1788                 } else {
1789                         /* We're missing D+Q; recompute D from P */
1790                         compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1791                         compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1792                         return;
1793                 }
1794         }
1795
1796         /* We're missing D+P or D+D; build pointer table */
1797         {
1798                 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1799                 void *ptrs[disks];
1800
1801                 count = 0;
1802                 i = d0_idx;
1803                 do {
1804                         ptrs[count++] = page_address(sh->dev[i].page);
1805                         i = raid6_next_disk(i, disks);
1806                         if (i != dd_idx1 && i != dd_idx2 &&
1807                             !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1808                                 printk("compute_2 with missing block %d/%d\n", count, i);
1809                 } while ( i != d0_idx );
1810
1811                 if ( failb == disks-2 ) {
1812                         /* We're missing D+P. */
1813                         raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1814                 } else {
1815                         /* We're missing D+D. */
1816                         raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1817                 }
1818
1819                 /* Both the above update both missing blocks */
1820                 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1821                 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1822         }
1823 }
1824
1825
1826
1827 /*
1828  * Each stripe/dev can have one or more bion attached.
1829  * toread/towrite point to the first in a chain.
1830  * The bi_next chain must be in order.
1831  */
1832 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1833 {
1834         struct bio **bip;
1835         raid5_conf_t *conf = sh->raid_conf;
1836         int firstwrite=0;
1837
1838         pr_debug("adding bh b#%llu to stripe s#%llu\n",
1839                 (unsigned long long)bi->bi_sector,
1840                 (unsigned long long)sh->sector);
1841
1842
1843         spin_lock(&sh->lock);
1844         spin_lock_irq(&conf->device_lock);
1845         if (forwrite) {
1846                 bip = &sh->dev[dd_idx].towrite;
1847                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1848                         firstwrite = 1;
1849         } else
1850                 bip = &sh->dev[dd_idx].toread;
1851         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1852                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1853                         goto overlap;
1854                 bip = & (*bip)->bi_next;
1855         }
1856         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1857                 goto overlap;
1858
1859         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1860         if (*bip)
1861                 bi->bi_next = *bip;
1862         *bip = bi;
1863         bi->bi_phys_segments ++;
1864         spin_unlock_irq(&conf->device_lock);
1865         spin_unlock(&sh->lock);
1866
1867         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1868                 (unsigned long long)bi->bi_sector,
1869                 (unsigned long long)sh->sector, dd_idx);
1870
1871         if (conf->mddev->bitmap && firstwrite) {
1872                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1873                                   STRIPE_SECTORS, 0);
1874                 sh->bm_seq = conf->seq_flush+1;
1875                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1876         }
1877
1878         if (forwrite) {
1879                 /* check if page is covered */
1880                 sector_t sector = sh->dev[dd_idx].sector;
1881                 for (bi=sh->dev[dd_idx].towrite;
1882                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1883                              bi && bi->bi_sector <= sector;
1884                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1885                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1886                                 sector = bi->bi_sector + (bi->bi_size>>9);
1887                 }
1888                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1889                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1890         }
1891         return 1;
1892
1893  overlap:
1894         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1895         spin_unlock_irq(&conf->device_lock);
1896         spin_unlock(&sh->lock);
1897         return 0;
1898 }
1899
1900 static void end_reshape(raid5_conf_t *conf);
1901
1902 static int page_is_zero(struct page *p)
1903 {
1904         char *a = page_address(p);
1905         return ((*(u32*)a) == 0 &&
1906                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1907 }
1908
1909 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1910 {
1911         int sectors_per_chunk = conf->chunk_size >> 9;
1912         int pd_idx, dd_idx;
1913         int chunk_offset = sector_div(stripe, sectors_per_chunk);
1914
1915         raid5_compute_sector(stripe * (disks - conf->max_degraded)
1916                              *sectors_per_chunk + chunk_offset,
1917                              disks, disks - conf->max_degraded,
1918                              &dd_idx, &pd_idx, conf);
1919         return pd_idx;
1920 }
1921
1922 static void
1923 handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1924                                 struct stripe_head_state *s, int disks,
1925                                 struct bio **return_bi)
1926 {
1927         int i;
1928         for (i = disks; i--; ) {
1929                 struct bio *bi;
1930                 int bitmap_end = 0;
1931
1932                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1933                         mdk_rdev_t *rdev;
1934                         rcu_read_lock();
1935                         rdev = rcu_dereference(conf->disks[i].rdev);
1936                         if (rdev && test_bit(In_sync, &rdev->flags))
1937                                 /* multiple read failures in one stripe */
1938                                 md_error(conf->mddev, rdev);
1939                         rcu_read_unlock();
1940                 }
1941                 spin_lock_irq(&conf->device_lock);
1942                 /* fail all writes first */
1943                 bi = sh->dev[i].towrite;
1944                 sh->dev[i].towrite = NULL;
1945                 if (bi) {
1946                         s->to_write--;
1947                         bitmap_end = 1;
1948                 }
1949
1950                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1951                         wake_up(&conf->wait_for_overlap);
1952
1953                 while (bi && bi->bi_sector <
1954                         sh->dev[i].sector + STRIPE_SECTORS) {
1955                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1956                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1957                         if (--bi->bi_phys_segments == 0) {
1958                                 md_write_end(conf->mddev);
1959                                 bi->bi_next = *return_bi;
1960                                 *return_bi = bi;
1961                         }
1962                         bi = nextbi;
1963                 }
1964                 /* and fail all 'written' */
1965                 bi = sh->dev[i].written;
1966                 sh->dev[i].written = NULL;
1967                 if (bi) bitmap_end = 1;
1968                 while (bi && bi->bi_sector <
1969                        sh->dev[i].sector + STRIPE_SECTORS) {
1970                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1971                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1972                         if (--bi->bi_phys_segments == 0) {
1973                                 md_write_end(conf->mddev);
1974                                 bi->bi_next = *return_bi;
1975                                 *return_bi = bi;
1976                         }
1977                         bi = bi2;
1978                 }
1979
1980                 /* fail any reads if this device is non-operational */
1981                 if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1982                     test_bit(R5_ReadError, &sh->dev[i].flags)) {
1983                         bi = sh->dev[i].toread;
1984                         sh->dev[i].toread = NULL;
1985                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1986                                 wake_up(&conf->wait_for_overlap);
1987                         if (bi) s->to_read--;
1988                         while (bi && bi->bi_sector <
1989                                sh->dev[i].sector + STRIPE_SECTORS) {
1990                                 struct bio *nextbi =
1991                                         r5_next_bio(bi, sh->dev[i].sector);
1992                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1993                                 if (--bi->bi_phys_segments == 0) {
1994                                         bi->bi_next = *return_bi;
1995                                         *return_bi = bi;
1996                                 }
1997                                 bi = nextbi;
1998                         }
1999                 }
2000                 spin_unlock_irq(&conf->device_lock);
2001                 if (bitmap_end)
2002                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2003                                         STRIPE_SECTORS, 0, 0);
2004         }
2005
2006 }
2007
2008 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2009                         struct stripe_head_state *s, int disks)
2010 {
2011         int i;
2012         for (i = disks; i--; ) {
2013                 struct r5dev *dev = &sh->dev[i];
2014                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2015                     !test_bit(R5_UPTODATE, &dev->flags) &&
2016                     (dev->toread ||
2017                      (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2018                      s->syncing || s->expanding ||
2019                      (s->failed && (sh->dev[s->failed_num].toread ||
2020                         (sh->dev[s->failed_num].towrite &&
2021                         !test_bit(R5_OVERWRITE, &sh->dev[s->failed_num].flags))
2022                       )))) {
2023                         /* we would like to get this block, possibly
2024                          * by computing it, but we might not be able to
2025                          */
2026                         if (s->uptodate == disks-1) {
2027                                 pr_debug("Computing block %d\n", i);
2028                                 compute_block(sh, i);
2029                                 s->uptodate++;
2030                         } else if (test_bit(R5_Insync, &dev->flags)) {
2031                                 set_bit(R5_LOCKED, &dev->flags);
2032                                 set_bit(R5_Wantread, &dev->flags);
2033                                 s->locked++;
2034                                 pr_debug("Reading block %d (sync=%d)\n",
2035                                         i, s->syncing);
2036                         }
2037                 }
2038         }
2039         set_bit(STRIPE_HANDLE, &sh->state);
2040 }
2041
2042 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2043                         struct stripe_head_state *s, struct r6_state *r6s,
2044                         int disks)
2045 {
2046         int i;
2047         for (i = disks; i--; ) {
2048                 struct r5dev *dev = &sh->dev[i];
2049                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2050                     !test_bit(R5_UPTODATE, &dev->flags) &&
2051                     (dev->toread || (dev->towrite &&
2052                      !test_bit(R5_OVERWRITE, &dev->flags)) ||
2053                      s->syncing || s->expanding ||
2054                      (s->failed >= 1 &&
2055                       (sh->dev[r6s->failed_num[0]].toread ||
2056                        s->to_write)) ||
2057                      (s->failed >= 2 &&
2058                       (sh->dev[r6s->failed_num[1]].toread ||
2059                        s->to_write)))) {
2060                         /* we would like to get this block, possibly
2061                          * by computing it, but we might not be able to
2062                          */
2063                         if (s->uptodate == disks-1) {
2064                                 pr_debug("Computing stripe %llu block %d\n",
2065                                        (unsigned long long)sh->sector, i);
2066                                 compute_block_1(sh, i, 0);
2067                                 s->uptodate++;
2068                         } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2069                                 /* Computing 2-failure is *very* expensive; only
2070                                  * do it if failed >= 2
2071                                  */
2072                                 int other;
2073                                 for (other = disks; other--; ) {
2074                                         if (other == i)
2075                                                 continue;
2076                                         if (!test_bit(R5_UPTODATE,
2077                                               &sh->dev[other].flags))
2078                                                 break;
2079                                 }
2080                                 BUG_ON(other < 0);
2081                                 pr_debug("Computing stripe %llu blocks %d,%d\n",
2082                                        (unsigned long long)sh->sector,
2083                                        i, other);
2084                                 compute_block_2(sh, i, other);
2085                                 s->uptodate += 2;
2086                         } else if (test_bit(R5_Insync, &dev->flags)) {
2087                                 set_bit(R5_LOCKED, &dev->flags);
2088                                 set_bit(R5_Wantread, &dev->flags);
2089                                 s->locked++;
2090                                 pr_debug("Reading block %d (sync=%d)\n",
2091                                         i, s->syncing);
2092                         }
2093                 }
2094         }
2095         set_bit(STRIPE_HANDLE, &sh->state);
2096 }
2097
2098
2099 /* handle_completed_write_requests
2100  * any written block on an uptodate or failed drive can be returned.
2101  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2102  * never LOCKED, so we don't need to test 'failed' directly.
2103  */
2104 static void handle_completed_write_requests(raid5_conf_t *conf,
2105         struct stripe_head *sh, int disks, struct bio **return_bi)
2106 {
2107         int i;
2108         struct r5dev *dev;
2109
2110         for (i = disks; i--; )
2111                 if (sh->dev[i].written) {
2112                         dev = &sh->dev[i];
2113                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2114                                 test_bit(R5_UPTODATE, &dev->flags)) {
2115                                 /* We can return any write requests */
2116                                 struct bio *wbi, *wbi2;
2117                                 int bitmap_end = 0;
2118                                 pr_debug("Return write for disc %d\n", i);
2119                                 spin_lock_irq(&conf->device_lock);
2120                                 wbi = dev->written;
2121                                 dev->written = NULL;
2122                                 while (wbi && wbi->bi_sector <
2123                                         dev->sector + STRIPE_SECTORS) {
2124                                         wbi2 = r5_next_bio(wbi, dev->sector);
2125                                         if (--wbi->bi_phys_segments == 0) {
2126                                                 md_write_end(conf->mddev);
2127                                                 wbi->bi_next = *return_bi;
2128                                                 *return_bi = wbi;
2129                                         }
2130                                         wbi = wbi2;
2131                                 }
2132                                 if (dev->towrite == NULL)
2133                                         bitmap_end = 1;
2134                                 spin_unlock_irq(&conf->device_lock);
2135                                 if (bitmap_end)
2136                                         bitmap_endwrite(conf->mddev->bitmap,
2137                                                         sh->sector,
2138                                                         STRIPE_SECTORS,
2139                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2140                                                         0);
2141                         }
2142                 }
2143 }
2144
2145 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2146                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2147 {
2148         int rmw = 0, rcw = 0, i;
2149         for (i = disks; i--; ) {
2150                 /* would I have to read this buffer for read_modify_write */
2151                 struct r5dev *dev = &sh->dev[i];
2152                 if ((dev->towrite || i == sh->pd_idx) &&
2153                     !test_bit(R5_LOCKED, &dev->flags) &&
2154                     !test_bit(R5_UPTODATE, &dev->flags)) {
2155                         if (test_bit(R5_Insync, &dev->flags))
2156                                 rmw++;
2157                         else
2158                                 rmw += 2*disks;  /* cannot read it */
2159                 }
2160                 /* Would I have to read this buffer for reconstruct_write */
2161                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2162                     !test_bit(R5_LOCKED, &dev->flags) &&
2163                     !test_bit(R5_UPTODATE, &dev->flags)) {
2164                         if (test_bit(R5_Insync, &dev->flags))
2165                                 rcw++;
2166                         else
2167                                 rcw += 2*disks;
2168                 }
2169         }
2170         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2171                 (unsigned long long)sh->sector, rmw, rcw);
2172         set_bit(STRIPE_HANDLE, &sh->state);
2173         if (rmw < rcw && rmw > 0)
2174                 /* prefer read-modify-write, but need to get some data */
2175                 for (i = disks; i--; ) {
2176                         struct r5dev *dev = &sh->dev[i];
2177                         if ((dev->towrite || i == sh->pd_idx) &&
2178                             !test_bit(R5_LOCKED, &dev->flags) &&
2179                             !test_bit(R5_UPTODATE, &dev->flags) &&
2180                             test_bit(R5_Insync, &dev->flags)) {
2181                                 if (
2182                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2183                                         pr_debug("Read_old block "
2184                                                 "%d for r-m-w\n", i);
2185                                         set_bit(R5_LOCKED, &dev->flags);
2186                                         set_bit(R5_Wantread, &dev->flags);
2187                                         s->locked++;
2188                                 } else {
2189                                         set_bit(STRIPE_DELAYED, &sh->state);
2190                                         set_bit(STRIPE_HANDLE, &sh->state);
2191                                 }
2192                         }
2193                 }
2194         if (rcw <= rmw && rcw > 0)
2195                 /* want reconstruct write, but need to get some data */
2196                 for (i = disks; i--; ) {
2197                         struct r5dev *dev = &sh->dev[i];
2198                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2199                             i != sh->pd_idx &&
2200                             !test_bit(R5_LOCKED, &dev->flags) &&
2201                             !test_bit(R5_UPTODATE, &dev->flags) &&
2202                             test_bit(R5_Insync, &dev->flags)) {
2203                                 if (
2204                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2205                                         pr_debug("Read_old block "
2206                                                 "%d for Reconstruct\n", i);
2207                                         set_bit(R5_LOCKED, &dev->flags);
2208                                         set_bit(R5_Wantread, &dev->flags);
2209                                         s->locked++;
2210                                 } else {
2211                                         set_bit(STRIPE_DELAYED, &sh->state);
2212                                         set_bit(STRIPE_HANDLE, &sh->state);
2213                                 }
2214                         }
2215                 }
2216         /* now if nothing is locked, and if we have enough data,
2217          * we can start a write request
2218          */
2219         if (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2220             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2221                 pr_debug("Computing parity...\n");
2222                 compute_parity5(sh, rcw == 0 ?
2223                         RECONSTRUCT_WRITE : READ_MODIFY_WRITE);
2224                 /* now every locked buffer is ready to be written */
2225                 for (i = disks; i--; )
2226                         if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2227                                 pr_debug("Writing block %d\n", i);
2228                                 s->locked++;
2229                                 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2230                                 if (!test_bit(R5_Insync, &sh->dev[i].flags)
2231                                     || (i == sh->pd_idx && s->failed == 0))
2232                                         set_bit(STRIPE_INSYNC, &sh->state);
2233                         }
2234                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2235                         atomic_dec(&conf->preread_active_stripes);
2236                         if (atomic_read(&conf->preread_active_stripes) <
2237                             IO_THRESHOLD)
2238                                 md_wakeup_thread(conf->mddev->thread);
2239                 }
2240         }
2241 }
2242
2243 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2244                 struct stripe_head *sh, struct stripe_head_state *s,
2245                 struct r6_state *r6s, int disks)
2246 {
2247         int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2248         int qd_idx = r6s->qd_idx;
2249         for (i = disks; i--; ) {
2250                 struct r5dev *dev = &sh->dev[i];
2251                 /* Would I have to read this buffer for reconstruct_write */
2252                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2253                     && i != pd_idx && i != qd_idx
2254                     && (!test_bit(R5_LOCKED, &dev->flags)
2255                             ) &&
2256                     !test_bit(R5_UPTODATE, &dev->flags)) {
2257                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2258                         else {
2259                                 pr_debug("raid6: must_compute: "
2260                                         "disk %d flags=%#lx\n", i, dev->flags);
2261                                 must_compute++;
2262                         }
2263                 }
2264         }
2265         pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2266                (unsigned long long)sh->sector, rcw, must_compute);
2267         set_bit(STRIPE_HANDLE, &sh->state);
2268
2269         if (rcw > 0)
2270                 /* want reconstruct write, but need to get some data */
2271                 for (i = disks; i--; ) {
2272                         struct r5dev *dev = &sh->dev[i];
2273                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2274                             && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2275                             && !test_bit(R5_LOCKED, &dev->flags) &&
2276                             !test_bit(R5_UPTODATE, &dev->flags) &&
2277                             test_bit(R5_Insync, &dev->flags)) {
2278                                 if (
2279                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2280                                         pr_debug("Read_old stripe %llu "
2281                                                 "block %d for Reconstruct\n",
2282                                              (unsigned long long)sh->sector, i);
2283                                         set_bit(R5_LOCKED, &dev->flags);
2284                                         set_bit(R5_Wantread, &dev->flags);
2285                                         s->locked++;
2286                                 } else {
2287                                         pr_debug("Request delayed stripe %llu "
2288                                                 "block %d for Reconstruct\n",
2289                                              (unsigned long long)sh->sector, i);
2290                                         set_bit(STRIPE_DELAYED, &sh->state);
2291                                         set_bit(STRIPE_HANDLE, &sh->state);
2292                                 }
2293                         }
2294                 }
2295         /* now if nothing is locked, and if we have enough data, we can start a
2296          * write request
2297          */
2298         if (s->locked == 0 && rcw == 0 &&
2299             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2300                 if (must_compute > 0) {
2301                         /* We have failed blocks and need to compute them */
2302                         switch (s->failed) {
2303                         case 0:
2304                                 BUG();
2305                         case 1:
2306                                 compute_block_1(sh, r6s->failed_num[0], 0);
2307                                 break;
2308                         case 2:
2309                                 compute_block_2(sh, r6s->failed_num[0],
2310                                                 r6s->failed_num[1]);
2311                                 break;
2312                         default: /* This request should have been failed? */
2313                                 BUG();
2314                         }
2315                 }
2316
2317                 pr_debug("Computing parity for stripe %llu\n",
2318                         (unsigned long long)sh->sector);
2319                 compute_parity6(sh, RECONSTRUCT_WRITE);
2320                 /* now every locked buffer is ready to be written */
2321                 for (i = disks; i--; )
2322                         if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2323                                 pr_debug("Writing stripe %llu block %d\n",
2324                                        (unsigned long long)sh->sector, i);
2325                                 s->locked++;
2326                                 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2327                         }
2328                 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2329                 set_bit(STRIPE_INSYNC, &sh->state);
2330
2331                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2332                         atomic_dec(&conf->preread_active_stripes);
2333                         if (atomic_read(&conf->preread_active_stripes) <
2334                             IO_THRESHOLD)
2335                                 md_wakeup_thread(conf->mddev->thread);
2336                 }
2337         }
2338 }
2339
2340 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2341                                 struct stripe_head_state *s, int disks)
2342 {
2343         set_bit(STRIPE_HANDLE, &sh->state);
2344         if (s->failed == 0) {
2345                 BUG_ON(s->uptodate != disks);
2346                 compute_parity5(sh, CHECK_PARITY);
2347                 s->uptodate--;
2348                 if (page_is_zero(sh->dev[sh->pd_idx].page)) {
2349                         /* parity is correct (on disc, not in buffer any more)
2350                          */
2351                         set_bit(STRIPE_INSYNC, &sh->state);
2352                 } else {
2353                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2354                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2355                                 /* don't try to repair!! */
2356                                 set_bit(STRIPE_INSYNC, &sh->state);
2357                         else {
2358                                 compute_block(sh, sh->pd_idx);
2359                                 s->uptodate++;
2360                         }
2361                 }
2362         }
2363         if (!test_bit(STRIPE_INSYNC, &sh->state)) {
2364                 struct r5dev *dev;
2365                 /* either failed parity check, or recovery is happening */
2366                 if (s->failed == 0)
2367                         s->failed_num = sh->pd_idx;
2368                 dev = &sh->dev[s->failed_num];
2369                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2370                 BUG_ON(s->uptodate != disks);
2371
2372                 set_bit(R5_LOCKED, &dev->flags);
2373                 set_bit(R5_Wantwrite, &dev->flags);
2374                 clear_bit(STRIPE_DEGRADED, &sh->state);
2375                 s->locked++;
2376                 set_bit(STRIPE_INSYNC, &sh->state);
2377         }
2378 }
2379
2380
2381 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2382                                 struct stripe_head_state *s,
2383                                 struct r6_state *r6s, struct page *tmp_page,
2384                                 int disks)
2385 {
2386         int update_p = 0, update_q = 0;
2387         struct r5dev *dev;
2388         int pd_idx = sh->pd_idx;
2389         int qd_idx = r6s->qd_idx;
2390
2391         set_bit(STRIPE_HANDLE, &sh->state);
2392
2393         BUG_ON(s->failed > 2);
2394         BUG_ON(s->uptodate < disks);
2395         /* Want to check and possibly repair P and Q.
2396          * However there could be one 'failed' device, in which
2397          * case we can only check one of them, possibly using the
2398          * other to generate missing data
2399          */
2400
2401         /* If !tmp_page, we cannot do the calculations,
2402          * but as we have set STRIPE_HANDLE, we will soon be called
2403          * by stripe_handle with a tmp_page - just wait until then.
2404          */
2405         if (tmp_page) {
2406                 if (s->failed == r6s->q_failed) {
2407                         /* The only possible failed device holds 'Q', so it
2408                          * makes sense to check P (If anything else were failed,
2409                          * we would have used P to recreate it).
2410                          */
2411                         compute_block_1(sh, pd_idx, 1);
2412                         if (!page_is_zero(sh->dev[pd_idx].page)) {
2413                                 compute_block_1(sh, pd_idx, 0);
2414                                 update_p = 1;
2415                         }
2416                 }
2417                 if (!r6s->q_failed && s->failed < 2) {
2418                         /* q is not failed, and we didn't use it to generate
2419                          * anything, so it makes sense to check it
2420                          */
2421                         memcpy(page_address(tmp_page),
2422                                page_address(sh->dev[qd_idx].page),
2423                                STRIPE_SIZE);
2424                         compute_parity6(sh, UPDATE_PARITY);
2425                         if (memcmp(page_address(tmp_page),
2426                                    page_address(sh->dev[qd_idx].page),
2427                                    STRIPE_SIZE) != 0) {
2428                                 clear_bit(STRIPE_INSYNC, &sh->state);
2429                                 update_q = 1;
2430                         }
2431                 }
2432                 if (update_p || update_q) {
2433                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2434                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2435                                 /* don't try to repair!! */
2436                                 update_p = update_q = 0;
2437                 }
2438
2439                 /* now write out any block on a failed drive,
2440                  * or P or Q if they need it
2441                  */
2442
2443                 if (s->failed == 2) {
2444                         dev = &sh->dev[r6s->failed_num[1]];
2445                         s->locked++;
2446                         set_bit(R5_LOCKED, &dev->flags);
2447                         set_bit(R5_Wantwrite, &dev->flags);
2448                 }
2449                 if (s->failed >= 1) {
2450                         dev = &sh->dev[r6s->failed_num[0]];
2451                         s->locked++;
2452                         set_bit(R5_LOCKED, &dev->flags);
2453                         set_bit(R5_Wantwrite, &dev->flags);
2454                 }
2455
2456                 if (update_p) {
2457                         dev = &sh->dev[pd_idx];
2458                         s->locked++;
2459                         set_bit(R5_LOCKED, &dev->flags);
2460                         set_bit(R5_Wantwrite, &dev->flags);
2461                 }
2462                 if (update_q) {
2463                         dev = &sh->dev[qd_idx];
2464                         s->locked++;
2465                         set_bit(R5_LOCKED, &dev->flags);
2466                         set_bit(R5_Wantwrite, &dev->flags);
2467                 }
2468                 clear_bit(STRIPE_DEGRADED, &sh->state);
2469
2470                 set_bit(STRIPE_INSYNC, &sh->state);
2471         }
2472 }
2473
2474 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2475                                 struct r6_state *r6s)
2476 {
2477         int i;
2478
2479         /* We have read all the blocks in this stripe and now we need to
2480          * copy some of them into a target stripe for expand.
2481          */
2482         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2483         for (i = 0; i < sh->disks; i++)
2484                 if (i != sh->pd_idx && (r6s && i != r6s->qd_idx)) {
2485                         int dd_idx, pd_idx, j;
2486                         struct stripe_head *sh2;
2487
2488                         sector_t bn = compute_blocknr(sh, i);
2489                         sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2490                                                 conf->raid_disks -
2491                                                 conf->max_degraded, &dd_idx,
2492                                                 &pd_idx, conf);
2493                         sh2 = get_active_stripe(conf, s, conf->raid_disks,
2494                                                 pd_idx, 1);
2495                         if (sh2 == NULL)
2496                                 /* so far only the early blocks of this stripe
2497                                  * have been requested.  When later blocks
2498                                  * get requested, we will try again
2499                                  */
2500                                 continue;
2501                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2502                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2503                                 /* must have already done this block */
2504                                 release_stripe(sh2);
2505                                 continue;
2506                         }
2507                         memcpy(page_address(sh2->dev[dd_idx].page),
2508                                page_address(sh->dev[i].page),
2509                                STRIPE_SIZE);
2510                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2511                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2512                         for (j = 0; j < conf->raid_disks; j++)
2513                                 if (j != sh2->pd_idx &&
2514                                     (r6s && j != r6s->qd_idx) &&
2515                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2516                                         break;
2517                         if (j == conf->raid_disks) {
2518                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2519                                 set_bit(STRIPE_HANDLE, &sh2->state);
2520                         }
2521                         release_stripe(sh2);
2522                 }
2523 }
2524
2525 /*
2526  * handle_stripe - do things to a stripe.
2527  *
2528  * We lock the stripe and then examine the state of various bits
2529  * to see what needs to be done.
2530  * Possible results:
2531  *    return some read request which now have data
2532  *    return some write requests which are safely on disc
2533  *    schedule a read on some buffers
2534  *    schedule a write of some buffers
2535  *    return confirmation of parity correctness
2536  *
2537  * buffers are taken off read_list or write_list, and bh_cache buffers
2538  * get BH_Lock set before the stripe lock is released.
2539  *
2540  */
2541
2542 static void handle_stripe5(struct stripe_head *sh)
2543 {
2544         raid5_conf_t *conf = sh->raid_conf;
2545         int disks = sh->disks, i;
2546         struct bio *return_bi = NULL;
2547         struct stripe_head_state s;
2548         struct r5dev *dev;
2549         unsigned long pending = 0;
2550
2551         memset(&s, 0, sizeof(s));
2552         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2553                 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2554                 atomic_read(&sh->count), sh->pd_idx,
2555                 sh->ops.pending, sh->ops.ack, sh->ops.complete);
2556
2557         spin_lock(&sh->lock);
2558         clear_bit(STRIPE_HANDLE, &sh->state);
2559         clear_bit(STRIPE_DELAYED, &sh->state);
2560
2561         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2562         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2563         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2564         /* Now to look around and see what can be done */
2565
2566         rcu_read_lock();
2567         for (i=disks; i--; ) {
2568                 mdk_rdev_t *rdev;
2569                 struct r5dev *dev = &sh->dev[i];
2570                 clear_bit(R5_Insync, &dev->flags);
2571
2572                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2573                         i, dev->flags, dev->toread, dev->towrite, dev->written);
2574                 /* maybe we can reply to a read */
2575                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2576                         struct bio *rbi, *rbi2;
2577                         pr_debug("Return read for disc %d\n", i);
2578                         spin_lock_irq(&conf->device_lock);
2579                         rbi = dev->toread;
2580                         dev->toread = NULL;
2581                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2582                                 wake_up(&conf->wait_for_overlap);
2583                         spin_unlock_irq(&conf->device_lock);
2584                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2585                                 copy_data(0, rbi, dev->page, dev->sector);
2586                                 rbi2 = r5_next_bio(rbi, dev->sector);
2587                                 spin_lock_irq(&conf->device_lock);
2588                                 if (--rbi->bi_phys_segments == 0) {
2589                                         rbi->bi_next = return_bi;
2590                                         return_bi = rbi;
2591                                 }
2592                                 spin_unlock_irq(&conf->device_lock);
2593                                 rbi = rbi2;
2594                         }
2595                 }
2596
2597                 /* now count some things */
2598                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2599                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2600
2601                 if (dev->toread)
2602                         s.to_read++;
2603                 if (dev->towrite) {
2604                         s.to_write++;
2605                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2606                                 s.non_overwrite++;
2607                 }
2608                 if (dev->written)
2609                         s.written++;
2610                 rdev = rcu_dereference(conf->disks[i].rdev);
2611                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2612                         /* The ReadError flag will just be confusing now */
2613                         clear_bit(R5_ReadError, &dev->flags);
2614                         clear_bit(R5_ReWrite, &dev->flags);
2615                 }
2616                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2617                     || test_bit(R5_ReadError, &dev->flags)) {
2618                         s.failed++;
2619                         s.failed_num = i;
2620                 } else
2621                         set_bit(R5_Insync, &dev->flags);
2622         }
2623         rcu_read_unlock();
2624         pr_debug("locked=%d uptodate=%d to_read=%d"
2625                 " to_write=%d failed=%d failed_num=%d\n",
2626                 s.locked, s.uptodate, s.to_read, s.to_write,
2627                 s.failed, s.failed_num);
2628         /* check if the array has lost two devices and, if so, some requests might
2629          * need to be failed
2630          */
2631         if (s.failed > 1 && s.to_read+s.to_write+s.written)
2632                 handle_requests_to_failed_array(conf, sh, &s, disks,
2633                                                 &return_bi);
2634         if (s.failed > 1 && s.syncing) {
2635                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2636                 clear_bit(STRIPE_SYNCING, &sh->state);
2637                 s.syncing = 0;
2638         }
2639
2640         /* might be able to return some write requests if the parity block
2641          * is safe, or on a failed drive
2642          */
2643         dev = &sh->dev[sh->pd_idx];
2644         if ( s.written &&
2645              ((test_bit(R5_Insync, &dev->flags) &&
2646                !test_bit(R5_LOCKED, &dev->flags) &&
2647                test_bit(R5_UPTODATE, &dev->flags)) ||
2648                (s.failed == 1 && s.failed_num == sh->pd_idx)))
2649                 handle_completed_write_requests(conf, sh, disks, &return_bi);
2650
2651         /* Now we might consider reading some blocks, either to check/generate
2652          * parity, or to satisfy requests
2653          * or to load a block that is being partially written.
2654          */
2655         if (s.to_read || s.non_overwrite ||
2656                 (s.syncing && (s.uptodate < disks)) || s.expanding)
2657                 handle_issuing_new_read_requests5(sh, &s, disks);
2658
2659         /* now to consider writing and what else, if anything should be read */
2660         if (s.to_write)
2661                 handle_issuing_new_write_requests5(conf, sh, &s, disks);
2662
2663         /* maybe we need to check and possibly fix the parity for this stripe
2664          * Any reads will already have been scheduled, so we just see if enough data
2665          * is available
2666          */
2667         if (s.syncing && s.locked == 0 &&
2668             !test_bit(STRIPE_INSYNC, &sh->state))
2669                 handle_parity_checks5(conf, sh, &s, disks);
2670         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2671                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2672                 clear_bit(STRIPE_SYNCING, &sh->state);
2673         }
2674
2675         /* If the failed drive is just a ReadError, then we might need to progress
2676          * the repair/check process
2677          */
2678         if (s.failed == 1 && !conf->mddev->ro &&
2679             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2680             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2681             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2682                 ) {
2683                 dev = &sh->dev[s.failed_num];
2684                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2685                         set_bit(R5_Wantwrite, &dev->flags);
2686                         set_bit(R5_ReWrite, &dev->flags);
2687                         set_bit(R5_LOCKED, &dev->flags);
2688                         s.locked++;
2689                 } else {
2690                         /* let's read it back */
2691                         set_bit(R5_Wantread, &dev->flags);
2692                         set_bit(R5_LOCKED, &dev->flags);
2693                         s.locked++;
2694                 }
2695         }
2696
2697         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
2698                 /* Need to write out all blocks after computing parity */
2699                 sh->disks = conf->raid_disks;
2700                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf, conf->raid_disks);
2701                 compute_parity5(sh, RECONSTRUCT_WRITE);
2702                 for (i = conf->raid_disks; i--; ) {
2703                         set_bit(R5_LOCKED, &sh->dev[i].flags);
2704                         s.locked++;
2705                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2706                 }
2707                 clear_bit(STRIPE_EXPANDING, &sh->state);
2708         } else if (s.expanded) {
2709                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2710                 atomic_dec(&conf->reshape_stripes);
2711                 wake_up(&conf->wait_for_overlap);
2712                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2713         }
2714
2715         if (s.expanding && s.locked == 0)
2716                 handle_stripe_expansion(conf, sh, NULL);
2717
2718         if (sh->ops.count)
2719                 pending = get_stripe_work(sh);
2720
2721         spin_unlock(&sh->lock);
2722
2723         if (pending)
2724                 raid5_run_ops(sh, pending);
2725
2726         return_io(return_bi);
2727
2728         for (i=disks; i-- ;) {
2729                 int rw;
2730                 struct bio *bi;
2731                 mdk_rdev_t *rdev;
2732                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
2733                         rw = WRITE;
2734                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
2735                         rw = READ;
2736                 else
2737                         continue;
2738  
2739                 bi = &sh->dev[i].req;
2740  
2741                 bi->bi_rw = rw;
2742                 if (rw == WRITE)
2743                         bi->bi_end_io = raid5_end_write_request;
2744                 else
2745                         bi->bi_end_io = raid5_end_read_request;
2746  
2747                 rcu_read_lock();
2748                 rdev = rcu_dereference(conf->disks[i].rdev);
2749                 if (rdev && test_bit(Faulty, &rdev->flags))
2750                         rdev = NULL;
2751                 if (rdev)
2752                         atomic_inc(&rdev->nr_pending);
2753                 rcu_read_unlock();
2754  
2755                 if (rdev) {
2756                         if (s.syncing || s.expanding || s.expanded)
2757                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
2758
2759                         bi->bi_bdev = rdev->bdev;
2760                         pr_debug("for %llu schedule op %ld on disc %d\n",
2761                                 (unsigned long long)sh->sector, bi->bi_rw, i);
2762                         atomic_inc(&sh->count);
2763                         bi->bi_sector = sh->sector + rdev->data_offset;
2764                         bi->bi_flags = 1 << BIO_UPTODATE;
2765                         bi->bi_vcnt = 1;        
2766                         bi->bi_max_vecs = 1;
2767                         bi->bi_idx = 0;
2768                         bi->bi_io_vec = &sh->dev[i].vec;
2769                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
2770                         bi->bi_io_vec[0].bv_offset = 0;
2771                         bi->bi_size = STRIPE_SIZE;
2772                         bi->bi_next = NULL;
2773                         if (rw == WRITE &&
2774                             test_bit(R5_ReWrite, &sh->dev[i].flags))
2775                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2776                         generic_make_request(bi);
2777                 } else {
2778                         if (rw == WRITE)
2779                                 set_bit(STRIPE_DEGRADED, &sh->state);
2780                         pr_debug("skip op %ld on disc %d for sector %llu\n",
2781                                 bi->bi_rw, i, (unsigned long long)sh->sector);
2782                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2783                         set_bit(STRIPE_HANDLE, &sh->state);
2784                 }
2785         }
2786 }
2787
2788 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2789 {
2790         raid6_conf_t *conf = sh->raid_conf;
2791         int disks = sh->disks;
2792         struct bio *return_bi = NULL;
2793         int i, pd_idx = sh->pd_idx;
2794         struct stripe_head_state s;
2795         struct r6_state r6s;
2796         struct r5dev *dev, *pdev, *qdev;
2797
2798         r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2799         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2800                 "pd_idx=%d, qd_idx=%d\n",
2801                (unsigned long long)sh->sector, sh->state,
2802                atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2803         memset(&s, 0, sizeof(s));
2804
2805         spin_lock(&sh->lock);
2806         clear_bit(STRIPE_HANDLE, &sh->state);
2807         clear_bit(STRIPE_DELAYED, &sh->state);
2808
2809         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2810         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2811         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2812         /* Now to look around and see what can be done */
2813
2814         rcu_read_lock();
2815         for (i=disks; i--; ) {
2816                 mdk_rdev_t *rdev;
2817                 dev = &sh->dev[i];
2818                 clear_bit(R5_Insync, &dev->flags);
2819
2820                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2821                         i, dev->flags, dev->toread, dev->towrite, dev->written);
2822                 /* maybe we can reply to a read */
2823                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2824                         struct bio *rbi, *rbi2;
2825                         pr_debug("Return read for disc %d\n", i);
2826                         spin_lock_irq(&conf->device_lock);
2827                         rbi = dev->toread;
2828                         dev->toread = NULL;
2829                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2830                                 wake_up(&conf->wait_for_overlap);
2831                         spin_unlock_irq(&conf->device_lock);
2832                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2833                                 copy_data(0, rbi, dev->page, dev->sector);
2834                                 rbi2 = r5_next_bio(rbi, dev->sector);
2835                                 spin_lock_irq(&conf->device_lock);
2836                                 if (--rbi->bi_phys_segments == 0) {
2837                                         rbi->bi_next = return_bi;
2838                                         return_bi = rbi;
2839                                 }
2840                                 spin_unlock_irq(&conf->device_lock);
2841                                 rbi = rbi2;
2842                         }
2843                 }
2844
2845                 /* now count some things */
2846                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2847                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2848
2849
2850                 if (dev->toread)
2851                         s.to_read++;
2852                 if (dev->towrite) {
2853                         s.to_write++;
2854                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2855                                 s.non_overwrite++;
2856                 }
2857                 if (dev->written)
2858                         s.written++;
2859                 rdev = rcu_dereference(conf->disks[i].rdev);
2860                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2861                         /* The ReadError flag will just be confusing now */
2862                         clear_bit(R5_ReadError, &dev->flags);
2863                         clear_bit(R5_ReWrite, &dev->flags);
2864                 }
2865                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2866                     || test_bit(R5_ReadError, &dev->flags)) {
2867                         if (s.failed < 2)
2868                                 r6s.failed_num[s.failed] = i;
2869                         s.failed++;
2870                 } else
2871                         set_bit(R5_Insync, &dev->flags);
2872         }
2873         rcu_read_unlock();
2874         pr_debug("locked=%d uptodate=%d to_read=%d"
2875                " to_write=%d failed=%d failed_num=%d,%d\n",
2876                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
2877                r6s.failed_num[0], r6s.failed_num[1]);
2878         /* check if the array has lost >2 devices and, if so, some requests
2879          * might need to be failed
2880          */
2881         if (s.failed > 2 && s.to_read+s.to_write+s.written)
2882                 handle_requests_to_failed_array(conf, sh, &s, disks,
2883                                                 &return_bi);
2884         if (s.failed > 2 && s.syncing) {
2885                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2886                 clear_bit(STRIPE_SYNCING, &sh->state);
2887                 s.syncing = 0;
2888         }
2889
2890         /*
2891          * might be able to return some write requests if the parity blocks
2892          * are safe, or on a failed drive
2893          */
2894         pdev = &sh->dev[pd_idx];
2895         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
2896                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
2897         qdev = &sh->dev[r6s.qd_idx];
2898         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
2899                 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
2900
2901         if ( s.written &&
2902              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
2903                              && !test_bit(R5_LOCKED, &pdev->flags)
2904                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
2905              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
2906                              && !test_bit(R5_LOCKED, &qdev->flags)
2907                              && test_bit(R5_UPTODATE, &qdev->flags)))))
2908                 handle_completed_write_requests(conf, sh, disks, &return_bi);
2909
2910         /* Now we might consider reading some blocks, either to check/generate
2911          * parity, or to satisfy requests
2912          * or to load a block that is being partially written.
2913          */
2914         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
2915             (s.syncing && (s.uptodate < disks)) || s.expanding)
2916                 handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
2917
2918         /* now to consider writing and what else, if anything should be read */
2919         if (s.to_write)
2920                 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
2921
2922         /* maybe we need to check and possibly fix the parity for this stripe
2923          * Any reads will already have been scheduled, so we just see if enough
2924          * data is available
2925          */
2926         if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
2927                 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
2928
2929         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2930                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2931                 clear_bit(STRIPE_SYNCING, &sh->state);
2932         }
2933
2934         /* If the failed drives are just a ReadError, then we might need
2935          * to progress the repair/check process
2936          */
2937         if (s.failed <= 2 && !conf->mddev->ro)
2938                 for (i = 0; i < s.failed; i++) {
2939                         dev = &sh->dev[r6s.failed_num[i]];
2940                         if (test_bit(R5_ReadError, &dev->flags)
2941                             && !test_bit(R5_LOCKED, &dev->flags)
2942                             && test_bit(R5_UPTODATE, &dev->flags)
2943                                 ) {
2944                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2945                                         set_bit(R5_Wantwrite, &dev->flags);
2946                                         set_bit(R5_ReWrite, &dev->flags);
2947                                         set_bit(R5_LOCKED, &dev->flags);
2948                                 } else {
2949                                         /* let's read it back */
2950                                         set_bit(R5_Wantread, &dev->flags);
2951                                         set_bit(R5_LOCKED, &dev->flags);
2952                                 }
2953                         }
2954                 }
2955
2956         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
2957                 /* Need to write out all blocks after computing P&Q */
2958                 sh->disks = conf->raid_disks;
2959                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2960                                              conf->raid_disks);
2961                 compute_parity6(sh, RECONSTRUCT_WRITE);
2962                 for (i = conf->raid_disks ; i-- ;  ) {
2963                         set_bit(R5_LOCKED, &sh->dev[i].flags);
2964                         s.locked++;
2965                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2966                 }
2967                 clear_bit(STRIPE_EXPANDING, &sh->state);
2968         } else if (s.expanded) {
2969                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2970                 atomic_dec(&conf->reshape_stripes);
2971                 wake_up(&conf->wait_for_overlap);
2972                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2973         }
2974
2975         if (s.expanding && s.locked == 0)
2976                 handle_stripe_expansion(conf, sh, &r6s);
2977
2978         spin_unlock(&sh->lock);
2979
2980         return_io(return_bi);
2981
2982         for (i=disks; i-- ;) {
2983                 int rw;
2984                 struct bio *bi;
2985                 mdk_rdev_t *rdev;
2986                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
2987                         rw = WRITE;
2988                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
2989                         rw = READ;
2990                 else
2991                         continue;
2992
2993                 bi = &sh->dev[i].req;
2994
2995                 bi->bi_rw = rw;
2996                 if (rw == WRITE)
2997                         bi->bi_end_io = raid5_end_write_request;
2998                 else
2999                         bi->bi_end_io = raid5_end_read_request;
3000
3001                 rcu_read_lock();
3002                 rdev = rcu_dereference(conf->disks[i].rdev);
3003                 if (rdev && test_bit(Faulty, &rdev->flags))
3004                         rdev = NULL;
3005                 if (rdev)
3006                         atomic_inc(&rdev->nr_pending);
3007                 rcu_read_unlock();
3008
3009                 if (rdev) {
3010                         if (s.syncing || s.expanding || s.expanded)
3011                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
3012
3013                         bi->bi_bdev = rdev->bdev;
3014                         pr_debug("for %llu schedule op %ld on disc %d\n",
3015                                 (unsigned long long)sh->sector, bi->bi_rw, i);
3016                         atomic_inc(&sh->count);
3017                         bi->bi_sector = sh->sector + rdev->data_offset;
3018                         bi->bi_flags = 1 << BIO_UPTODATE;
3019                         bi->bi_vcnt = 1;
3020                         bi->bi_max_vecs = 1;
3021                         bi->bi_idx = 0;
3022                         bi->bi_io_vec = &sh->dev[i].vec;
3023                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
3024                         bi->bi_io_vec[0].bv_offset = 0;
3025                         bi->bi_size = STRIPE_SIZE;
3026                         bi->bi_next = NULL;
3027                         if (rw == WRITE &&
3028                             test_bit(R5_ReWrite, &sh->dev[i].flags))
3029                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
3030                         generic_make_request(bi);
3031                 } else {
3032                         if (rw == WRITE)
3033                                 set_bit(STRIPE_DEGRADED, &sh->state);
3034                         pr_debug("skip op %ld on disc %d for sector %llu\n",
3035                                 bi->bi_rw, i, (unsigned long long)sh->sector);
3036                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
3037                         set_bit(STRIPE_HANDLE, &sh->state);
3038                 }
3039         }
3040 }
3041
3042 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3043 {
3044         if (sh->raid_conf->level == 6)
3045                 handle_stripe6(sh, tmp_page);
3046         else
3047                 handle_stripe5(sh);
3048 }
3049
3050
3051
3052 static void raid5_activate_delayed(raid5_conf_t *conf)
3053 {
3054         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3055                 while (!list_empty(&conf->delayed_list)) {
3056                         struct list_head *l = conf->delayed_list.next;
3057                         struct stripe_head *sh;
3058                         sh = list_entry(l, struct stripe_head, lru);
3059                         list_del_init(l);
3060                         clear_bit(STRIPE_DELAYED, &sh->state);
3061                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3062                                 atomic_inc(&conf->preread_active_stripes);
3063                         list_add_tail(&sh->lru, &conf->handle_list);
3064                 }
3065         }
3066 }
3067
3068 static void activate_bit_delay(raid5_conf_t *conf)
3069 {
3070         /* device_lock is held */
3071         struct list_head head;
3072         list_add(&head, &conf->bitmap_list);
3073         list_del_init(&conf->bitmap_list);
3074         while (!list_empty(&head)) {
3075                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3076                 list_del_init(&sh->lru);
3077                 atomic_inc(&sh->count);
3078                 __release_stripe(conf, sh);
3079         }
3080 }
3081
3082 static void unplug_slaves(mddev_t *mddev)
3083 {
3084         raid5_conf_t *conf = mddev_to_conf(mddev);
3085         int i;
3086
3087         rcu_read_lock();
3088         for (i=0; i<mddev->raid_disks; i++) {
3089                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3090                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3091                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
3092
3093                         atomic_inc(&rdev->nr_pending);
3094                         rcu_read_unlock();
3095
3096                         if (r_queue->unplug_fn)
3097                                 r_queue->unplug_fn(r_queue);
3098
3099                         rdev_dec_pending(rdev, mddev);
3100                         rcu_read_lock();
3101                 }
3102         }
3103         rcu_read_unlock();
3104 }
3105
3106 static void raid5_unplug_device(request_queue_t *q)
3107 {
3108         mddev_t *mddev = q->queuedata;
3109         raid5_conf_t *conf = mddev_to_conf(mddev);
3110         unsigned long flags;
3111
3112         spin_lock_irqsave(&conf->device_lock, flags);
3113
3114         if (blk_remove_plug(q)) {
3115                 conf->seq_flush++;
3116                 raid5_activate_delayed(conf);
3117         }
3118         md_wakeup_thread(mddev->thread);
3119
3120         spin_unlock_irqrestore(&conf->device_lock, flags);
3121
3122         unplug_slaves(mddev);
3123 }
3124
3125 static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
3126                              sector_t *error_sector)
3127 {
3128         mddev_t *mddev = q->queuedata;
3129         raid5_conf_t *conf = mddev_to_conf(mddev);
3130         int i, ret = 0;
3131
3132         rcu_read_lock();
3133         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
3134                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3135                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
3136                         struct block_device *bdev = rdev->bdev;
3137                         request_queue_t *r_queue = bdev_get_queue(bdev);
3138
3139                         if (!r_queue->issue_flush_fn)
3140                                 ret = -EOPNOTSUPP;
3141                         else {
3142                                 atomic_inc(&rdev->nr_pending);
3143                                 rcu_read_unlock();
3144                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
3145                                                               error_sector);
3146                                 rdev_dec_pending(rdev, mddev);
3147                                 rcu_read_lock();
3148                         }
3149                 }
3150         }
3151         rcu_read_unlock();
3152         return ret;
3153 }
3154
3155 static int raid5_congested(void *data, int bits)
3156 {
3157         mddev_t *mddev = data;
3158         raid5_conf_t *conf = mddev_to_conf(mddev);
3159
3160         /* No difference between reads and writes.  Just check
3161          * how busy the stripe_cache is
3162          */
3163         if (conf->inactive_blocked)
3164                 return 1;
3165         if (conf->quiesce)
3166                 return 1;
3167         if (list_empty_careful(&conf->inactive_list))
3168                 return 1;
3169
3170         return 0;
3171 }
3172
3173 /* We want read requests to align with chunks where possible,
3174  * but write requests don't need to.
3175  */
3176 static int raid5_mergeable_bvec(request_queue_t *q, struct bio *bio, struct bio_vec *biovec)
3177 {
3178         mddev_t *mddev = q->queuedata;
3179         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3180         int max;
3181         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3182         unsigned int bio_sectors = bio->bi_size >> 9;
3183
3184         if (bio_data_dir(bio) == WRITE)
3185                 return biovec->bv_len; /* always allow writes to be mergeable */
3186
3187         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3188         if (max < 0) max = 0;
3189         if (max <= biovec->bv_len && bio_sectors == 0)
3190                 return biovec->bv_len;
3191         else
3192                 return max;
3193 }
3194
3195
3196 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3197 {
3198         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3199         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3200         unsigned int bio_sectors = bio->bi_size >> 9;
3201
3202         return  chunk_sectors >=
3203                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3204 }
3205
3206 /*
3207  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3208  *  later sampled by raid5d.
3209  */
3210 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3211 {
3212         unsigned long flags;
3213
3214         spin_lock_irqsave(&conf->device_lock, flags);
3215
3216         bi->bi_next = conf->retry_read_aligned_list;
3217         conf->retry_read_aligned_list = bi;
3218
3219         spin_unlock_irqrestore(&conf->device_lock, flags);
3220         md_wakeup_thread(conf->mddev->thread);
3221 }
3222
3223
3224 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3225 {
3226         struct bio *bi;
3227
3228         bi = conf->retry_read_aligned;
3229         if (bi) {
3230                 conf->retry_read_aligned = NULL;
3231                 return bi;
3232         }
3233         bi = conf->retry_read_aligned_list;
3234         if(bi) {
3235                 conf->retry_read_aligned_list = bi->bi_next;
3236                 bi->bi_next = NULL;
3237                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3238                 bi->bi_hw_segments = 0; /* count of processed stripes */
3239         }
3240
3241         return bi;
3242 }
3243
3244
3245 /*
3246  *  The "raid5_align_endio" should check if the read succeeded and if it
3247  *  did, call bio_endio on the original bio (having bio_put the new bio
3248  *  first).
3249  *  If the read failed..
3250  */
3251 static int raid5_align_endio(struct bio *bi, unsigned int bytes, int error)
3252 {
3253         struct bio* raid_bi  = bi->bi_private;
3254         mddev_t *mddev;
3255         raid5_conf_t *conf;
3256         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3257         mdk_rdev_t *rdev;
3258
3259         if (bi->bi_size)
3260                 return 1;
3261         bio_put(bi);
3262
3263         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3264         conf = mddev_to_conf(mddev);
3265         rdev = (void*)raid_bi->bi_next;
3266         raid_bi->bi_next = NULL;
3267
3268         rdev_dec_pending(rdev, conf->mddev);
3269
3270         if (!error && uptodate) {
3271                 bio_endio(raid_bi, bytes, 0);
3272                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3273                         wake_up(&conf->wait_for_stripe);
3274                 return 0;
3275         }
3276
3277
3278         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3279
3280         add_bio_to_retry(raid_bi, conf);
3281         return 0;
3282 }
3283
3284 static int bio_fits_rdev(struct bio *bi)
3285 {
3286         request_queue_t *q = bdev_get_queue(bi->bi_bdev);
3287
3288         if ((bi->bi_size>>9) > q->max_sectors)
3289                 return 0;
3290         blk_recount_segments(q, bi);
3291         if (bi->bi_phys_segments > q->max_phys_segments ||
3292             bi->bi_hw_segments > q->max_hw_segments)
3293                 return 0;
3294
3295         if (q->merge_bvec_fn)
3296                 /* it's too hard to apply the merge_bvec_fn at this stage,
3297                  * just just give up
3298                  */
3299                 return 0;
3300
3301         return 1;
3302 }
3303
3304
3305 static int chunk_aligned_read(request_queue_t *q, struct bio * raid_bio)
3306 {
3307         mddev_t *mddev = q->queuedata;
3308         raid5_conf_t *conf = mddev_to_conf(mddev);
3309         const unsigned int raid_disks = conf->raid_disks;
3310         const unsigned int data_disks = raid_disks - conf->max_degraded;
3311         unsigned int dd_idx, pd_idx;
3312         struct bio* align_bi;
3313         mdk_rdev_t *rdev;
3314
3315         if (!in_chunk_boundary(mddev, raid_bio)) {
3316                 pr_debug("chunk_aligned_read : non aligned\n");
3317                 return 0;
3318         }
3319         /*
3320          * use bio_clone to make a copy of the bio
3321          */
3322         align_bi = bio_clone(raid_bio, GFP_NOIO);
3323         if (!align_bi)
3324                 return 0;
3325         /*
3326          *   set bi_end_io to a new function, and set bi_private to the
3327          *     original bio.
3328          */
3329         align_bi->bi_end_io  = raid5_align_endio;
3330         align_bi->bi_private = raid_bio;
3331         /*
3332          *      compute position
3333          */
3334         align_bi->bi_sector =  raid5_compute_sector(raid_bio->bi_sector,
3335                                         raid_disks,
3336                                         data_disks,
3337                                         &dd_idx,
3338                                         &pd_idx,
3339                                         conf);
3340
3341         rcu_read_lock();
3342         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3343         if (rdev && test_bit(In_sync, &rdev->flags)) {
3344                 atomic_inc(&rdev->nr_pending);
3345                 rcu_read_unlock();
3346                 raid_bio->bi_next = (void*)rdev;
3347                 align_bi->bi_bdev =  rdev->bdev;
3348                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3349                 align_bi->bi_sector += rdev->data_offset;
3350
3351                 if (!bio_fits_rdev(align_bi)) {
3352                         /* too big in some way */
3353                         bio_put(align_bi);
3354                         rdev_dec_pending(rdev, mddev);
3355                         return 0;
3356                 }
3357
3358                 spin_lock_irq(&conf->device_lock);
3359                 wait_event_lock_irq(conf->wait_for_stripe,
3360                                     conf->quiesce == 0,
3361                                     conf->device_lock, /* nothing */);
3362                 atomic_inc(&conf->active_aligned_reads);
3363                 spin_unlock_irq(&conf->device_lock);
3364
3365                 generic_make_request(align_bi);
3366                 return 1;
3367         } else {
3368                 rcu_read_unlock();
3369                 bio_put(align_bi);
3370                 return 0;
3371         }
3372 }
3373
3374
3375 static int make_request(request_queue_t *q, struct bio * bi)
3376 {
3377         mddev_t *mddev = q->queuedata;
3378         raid5_conf_t *conf = mddev_to_conf(mddev);
3379         unsigned int dd_idx, pd_idx;
3380         sector_t new_sector;
3381         sector_t logical_sector, last_sector;
3382         struct stripe_head *sh;
3383         const int rw = bio_data_dir(bi);
3384         int remaining;
3385
3386         if (unlikely(bio_barrier(bi))) {
3387                 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
3388                 return 0;
3389         }
3390
3391         md_write_start(mddev, bi);
3392
3393         disk_stat_inc(mddev->gendisk, ios[rw]);
3394         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3395
3396         if (rw == READ &&
3397              mddev->reshape_position == MaxSector &&
3398              chunk_aligned_read(q,bi))
3399                 return 0;
3400
3401         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3402         last_sector = bi->bi_sector + (bi->bi_size>>9);
3403         bi->bi_next = NULL;
3404         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3405
3406         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3407                 DEFINE_WAIT(w);
3408                 int disks, data_disks;
3409
3410         retry:
3411                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3412                 if (likely(conf->expand_progress == MaxSector))
3413                         disks = conf->raid_disks;
3414                 else {
3415                         /* spinlock is needed as expand_progress may be
3416                          * 64bit on a 32bit platform, and so it might be
3417                          * possible to see a half-updated value
3418                          * Ofcourse expand_progress could change after
3419                          * the lock is dropped, so once we get a reference
3420                          * to the stripe that we think it is, we will have
3421                          * to check again.
3422                          */
3423                         spin_lock_irq(&conf->device_lock);
3424                         disks = conf->raid_disks;
3425                         if (logical_sector >= conf->expand_progress)
3426                                 disks = conf->previous_raid_disks;
3427                         else {
3428                                 if (logical_sector >= conf->expand_lo) {
3429                                         spin_unlock_irq(&conf->device_lock);
3430                                         schedule();
3431                                         goto retry;
3432                                 }
3433                         }
3434                         spin_unlock_irq(&conf->device_lock);
3435                 }
3436                 data_disks = disks - conf->max_degraded;
3437
3438                 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3439                                                   &dd_idx, &pd_idx, conf);
3440                 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3441                         (unsigned long long)new_sector, 
3442                         (unsigned long long)logical_sector);
3443
3444                 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3445                 if (sh) {
3446                         if (unlikely(conf->expand_progress != MaxSector)) {
3447                                 /* expansion might have moved on while waiting for a
3448                                  * stripe, so we must do the range check again.
3449                                  * Expansion could still move past after this
3450                                  * test, but as we are holding a reference to
3451                                  * 'sh', we know that if that happens,
3452                                  *  STRIPE_EXPANDING will get set and the expansion
3453                                  * won't proceed until we finish with the stripe.
3454                                  */
3455                                 int must_retry = 0;
3456                                 spin_lock_irq(&conf->device_lock);
3457                                 if (logical_sector <  conf->expand_progress &&
3458                                     disks == conf->previous_raid_disks)
3459                                         /* mismatch, need to try again */
3460                                         must_retry = 1;
3461                                 spin_unlock_irq(&conf->device_lock);
3462                                 if (must_retry) {
3463                                         release_stripe(sh);
3464                                         goto retry;
3465                                 }
3466                         }
3467                         /* FIXME what if we get a false positive because these
3468                          * are being updated.
3469                          */
3470                         if (logical_sector >= mddev->suspend_lo &&
3471                             logical_sector < mddev->suspend_hi) {
3472                                 release_stripe(sh);
3473                                 schedule();
3474                                 goto retry;
3475                         }
3476
3477                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3478                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3479                                 /* Stripe is busy expanding or
3480                                  * add failed due to overlap.  Flush everything
3481                                  * and wait a while
3482                                  */
3483                                 raid5_unplug_device(mddev->queue);
3484                                 release_stripe(sh);
3485                                 schedule();
3486                                 goto retry;
3487                         }
3488                         finish_wait(&conf->wait_for_overlap, &w);
3489                         handle_stripe(sh, NULL);
3490                         release_stripe(sh);
3491                 } else {
3492                         /* cannot get stripe for read-ahead, just give-up */
3493                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3494                         finish_wait(&conf->wait_for_overlap, &w);
3495                         break;
3496                 }
3497                         
3498         }
3499         spin_lock_irq(&conf->device_lock);
3500         remaining = --bi->bi_phys_segments;
3501         spin_unlock_irq(&conf->device_lock);
3502         if (remaining == 0) {
3503                 int bytes = bi->bi_size;
3504
3505                 if ( rw == WRITE )
3506                         md_write_end(mddev);
3507                 bi->bi_size = 0;
3508                 bi->bi_end_io(bi, bytes,
3509                               test_bit(BIO_UPTODATE, &bi->bi_flags)
3510                                 ? 0 : -EIO);
3511         }
3512         return 0;
3513 }
3514
3515 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3516 {
3517         /* reshaping is quite different to recovery/resync so it is
3518          * handled quite separately ... here.
3519          *
3520          * On each call to sync_request, we gather one chunk worth of
3521          * destination stripes and flag them as expanding.
3522          * Then we find all the source stripes and request reads.
3523          * As the reads complete, handle_stripe will copy the data
3524          * into the destination stripe and release that stripe.
3525          */
3526         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3527         struct stripe_head *sh;
3528         int pd_idx;
3529         sector_t first_sector, last_sector;
3530         int raid_disks = conf->previous_raid_disks;
3531         int data_disks = raid_disks - conf->max_degraded;
3532         int new_data_disks = conf->raid_disks - conf->max_degraded;
3533         int i;
3534         int dd_idx;
3535         sector_t writepos, safepos, gap;
3536
3537         if (sector_nr == 0 &&
3538             conf->expand_progress != 0) {
3539                 /* restarting in the middle, skip the initial sectors */
3540                 sector_nr = conf->expand_progress;
3541                 sector_div(sector_nr, new_data_disks);
3542                 *skipped = 1;
3543                 return sector_nr;
3544         }
3545
3546         /* we update the metadata when there is more than 3Meg
3547          * in the block range (that is rather arbitrary, should
3548          * probably be time based) or when the data about to be
3549          * copied would over-write the source of the data at
3550          * the front of the range.
3551          * i.e. one new_stripe forward from expand_progress new_maps
3552          * to after where expand_lo old_maps to
3553          */
3554         writepos = conf->expand_progress +
3555                 conf->chunk_size/512*(new_data_disks);
3556         sector_div(writepos, new_data_disks);
3557         safepos = conf->expand_lo;
3558         sector_div(safepos, data_disks);
3559         gap = conf->expand_progress - conf->expand_lo;
3560
3561         if (writepos >= safepos ||
3562             gap > (new_data_disks)*3000*2 /*3Meg*/) {
3563                 /* Cannot proceed until we've updated the superblock... */
3564                 wait_event(conf->wait_for_overlap,
3565                            atomic_read(&conf->reshape_stripes)==0);
3566                 mddev->reshape_position = conf->expand_progress;
3567                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3568                 md_wakeup_thread(mddev->thread);
3569                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3570                            kthread_should_stop());
3571                 spin_lock_irq(&conf->device_lock);
3572                 conf->expand_lo = mddev->reshape_position;
3573                 spin_unlock_irq(&conf->device_lock);
3574                 wake_up(&conf->wait_for_overlap);
3575         }
3576
3577         for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3578                 int j;
3579                 int skipped = 0;
3580                 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3581                 sh = get_active_stripe(conf, sector_nr+i,
3582                                        conf->raid_disks, pd_idx, 0);
3583                 set_bit(STRIPE_EXPANDING, &sh->state);
3584                 atomic_inc(&conf->reshape_stripes);
3585                 /* If any of this stripe is beyond the end of the old
3586                  * array, then we need to zero those blocks
3587                  */
3588                 for (j=sh->disks; j--;) {
3589                         sector_t s;
3590                         if (j == sh->pd_idx)
3591                                 continue;
3592                         if (conf->level == 6 &&
3593                             j == raid6_next_disk(sh->pd_idx, sh->disks))
3594                                 continue;
3595                         s = compute_blocknr(sh, j);
3596                         if (s < (mddev->array_size<<1)) {
3597                                 skipped = 1;
3598                                 continue;
3599                         }
3600                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3601                         set_bit(R5_Expanded, &sh->dev[j].flags);
3602                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
3603                 }
3604                 if (!skipped) {
3605                         set_bit(STRIPE_EXPAND_READY, &sh->state);
3606                         set_bit(STRIPE_HANDLE, &sh->state);
3607                 }
3608                 release_stripe(sh);
3609         }
3610         spin_lock_irq(&conf->device_lock);
3611         conf->expand_progress = (sector_nr + i) * new_data_disks;
3612         spin_unlock_irq(&conf->device_lock);
3613         /* Ok, those stripe are ready. We can start scheduling
3614          * reads on the source stripes.
3615          * The source stripes are determined by mapping the first and last
3616          * block on the destination stripes.
3617          */
3618         first_sector =
3619                 raid5_compute_sector(sector_nr*(new_data_disks),
3620                                      raid_disks, data_disks,
3621                                      &dd_idx, &pd_idx, conf);
3622         last_sector =
3623                 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3624                                      *(new_data_disks) -1,
3625                                      raid_disks, data_disks,
3626                                      &dd_idx, &pd_idx, conf);
3627         if (last_sector >= (mddev->size<<1))
3628                 last_sector = (mddev->size<<1)-1;
3629         while (first_sector <= last_sector) {
3630                 pd_idx = stripe_to_pdidx(first_sector, conf,
3631                                          conf->previous_raid_disks);
3632                 sh = get_active_stripe(conf, first_sector,
3633                                        conf->previous_raid_disks, pd_idx, 0);
3634                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3635                 set_bit(STRIPE_HANDLE, &sh->state);
3636                 release_stripe(sh);
3637                 first_sector += STRIPE_SECTORS;
3638         }
3639         return conf->chunk_size>>9;
3640 }
3641
3642 /* FIXME go_faster isn't used */
3643 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3644 {
3645         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3646         struct stripe_head *sh;
3647         int pd_idx;
3648         int raid_disks = conf->raid_disks;
3649         sector_t max_sector = mddev->size << 1;
3650         int sync_blocks;
3651         int still_degraded = 0;
3652         int i;
3653
3654         if (sector_nr >= max_sector) {
3655                 /* just being told to finish up .. nothing much to do */
3656                 unplug_slaves(mddev);
3657                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3658                         end_reshape(conf);
3659                         return 0;
3660                 }
3661
3662                 if (mddev->curr_resync < max_sector) /* aborted */
3663                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3664                                         &sync_blocks, 1);
3665                 else /* completed sync */
3666                         conf->fullsync = 0;
3667                 bitmap_close_sync(mddev->bitmap);
3668
3669                 return 0;
3670         }
3671
3672         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3673                 return reshape_request(mddev, sector_nr, skipped);
3674
3675         /* if there is too many failed drives and we are trying
3676          * to resync, then assert that we are finished, because there is
3677          * nothing we can do.
3678          */
3679         if (mddev->degraded >= conf->max_degraded &&
3680             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3681                 sector_t rv = (mddev->size << 1) - sector_nr;
3682                 *skipped = 1;
3683                 return rv;
3684         }
3685         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3686             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3687             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3688                 /* we can skip this block, and probably more */
3689                 sync_blocks /= STRIPE_SECTORS;
3690                 *skipped = 1;
3691                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3692         }
3693
3694         pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3695         sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3696         if (sh == NULL) {
3697                 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3698                 /* make sure we don't swamp the stripe cache if someone else
3699                  * is trying to get access
3700                  */
3701                 schedule_timeout_uninterruptible(1);
3702         }
3703         /* Need to check if array will still be degraded after recovery/resync
3704          * We don't need to check the 'failed' flag as when that gets set,
3705          * recovery aborts.
3706          */
3707         for (i=0; i<mddev->raid_disks; i++)
3708                 if (conf->disks[i].rdev == NULL)
3709                         still_degraded = 1;
3710
3711         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3712
3713         spin_lock(&sh->lock);
3714         set_bit(STRIPE_SYNCING, &sh->state);
3715         clear_bit(STRIPE_INSYNC, &sh->state);
3716         spin_unlock(&sh->lock);
3717
3718         handle_stripe(sh, NULL);
3719         release_stripe(sh);
3720
3721         return STRIPE_SECTORS;
3722 }
3723
3724 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3725 {
3726         /* We may not be able to submit a whole bio at once as there
3727          * may not be enough stripe_heads available.
3728          * We cannot pre-allocate enough stripe_heads as we may need
3729          * more than exist in the cache (if we allow ever large chunks).
3730          * So we do one stripe head at a time and record in
3731          * ->bi_hw_segments how many have been done.
3732          *
3733          * We *know* that this entire raid_bio is in one chunk, so
3734          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3735          */
3736         struct stripe_head *sh;
3737         int dd_idx, pd_idx;
3738         sector_t sector, logical_sector, last_sector;
3739         int scnt = 0;
3740         int remaining;
3741         int handled = 0;
3742
3743         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3744         sector = raid5_compute_sector(  logical_sector,
3745                                         conf->raid_disks,
3746                                         conf->raid_disks - conf->max_degraded,
3747                                         &dd_idx,
3748                                         &pd_idx,
3749                                         conf);
3750         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3751
3752         for (; logical_sector < last_sector;
3753              logical_sector += STRIPE_SECTORS,
3754                      sector += STRIPE_SECTORS,
3755                      scnt++) {
3756
3757                 if (scnt < raid_bio->bi_hw_segments)
3758                         /* already done this stripe */
3759                         continue;
3760
3761                 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3762
3763                 if (!sh) {
3764                         /* failed to get a stripe - must wait */
3765                         raid_bio->bi_hw_segments = scnt;
3766                         conf->retry_read_aligned = raid_bio;
3767                         return handled;
3768                 }
3769
3770                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3771                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3772                         release_stripe(sh);
3773                         raid_bio->bi_hw_segments = scnt;
3774                         conf->retry_read_aligned = raid_bio;
3775                         return handled;
3776                 }
3777
3778                 handle_stripe(sh, NULL);
3779                 release_stripe(sh);
3780                 handled++;
3781         }
3782         spin_lock_irq(&conf->device_lock);
3783         remaining = --raid_bio->bi_phys_segments;
3784         spin_unlock_irq(&conf->device_lock);
3785         if (remaining == 0) {
3786                 int bytes = raid_bio->bi_size;
3787
3788                 raid_bio->bi_size = 0;
3789                 raid_bio->bi_end_io(raid_bio, bytes,
3790                               test_bit(BIO_UPTODATE, &raid_bio->bi_flags)
3791                                 ? 0 : -EIO);
3792         }
3793         if (atomic_dec_and_test(&conf->active_aligned_reads))
3794                 wake_up(&conf->wait_for_stripe);
3795         return handled;
3796 }
3797
3798
3799
3800 /*
3801  * This is our raid5 kernel thread.
3802  *
3803  * We scan the hash table for stripes which can be handled now.
3804  * During the scan, completed stripes are saved for us by the interrupt
3805  * handler, so that they will not have to wait for our next wakeup.
3806  */
3807 static void raid5d (mddev_t *mddev)
3808 {
3809         struct stripe_head *sh;
3810         raid5_conf_t *conf = mddev_to_conf(mddev);
3811         int handled;
3812
3813         pr_debug("+++ raid5d active\n");
3814
3815         md_check_recovery(mddev);
3816
3817         handled = 0;
3818         spin_lock_irq(&conf->device_lock);
3819         while (1) {
3820                 struct list_head *first;
3821                 struct bio *bio;
3822
3823                 if (conf->seq_flush != conf->seq_write) {
3824                         int seq = conf->seq_flush;
3825                         spin_unlock_irq(&conf->device_lock);
3826                         bitmap_unplug(mddev->bitmap);
3827                         spin_lock_irq(&conf->device_lock);
3828                         conf->seq_write = seq;
3829                         activate_bit_delay(conf);
3830                 }
3831
3832                 if (list_empty(&conf->handle_list) &&
3833                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
3834                     !blk_queue_plugged(mddev->queue) &&
3835                     !list_empty(&conf->delayed_list))
3836                         raid5_activate_delayed(conf);
3837
3838                 while ((bio = remove_bio_from_retry(conf))) {
3839                         int ok;
3840                         spin_unlock_irq(&conf->device_lock);
3841                         ok = retry_aligned_read(conf, bio);
3842                         spin_lock_irq(&conf->device_lock);
3843                         if (!ok)
3844                                 break;
3845                         handled++;
3846                 }
3847
3848                 if (list_empty(&conf->handle_list)) {
3849                         async_tx_issue_pending_all();
3850                         break;
3851                 }
3852
3853                 first = conf->handle_list.next;
3854                 sh = list_entry(first, struct stripe_head, lru);
3855
3856                 list_del_init(first);
3857                 atomic_inc(&sh->count);
3858                 BUG_ON(atomic_read(&sh->count)!= 1);
3859                 spin_unlock_irq(&conf->device_lock);
3860                 
3861                 handled++;
3862                 handle_stripe(sh, conf->spare_page);
3863                 release_stripe(sh);
3864
3865                 spin_lock_irq(&conf->device_lock);
3866         }
3867         pr_debug("%d stripes handled\n", handled);
3868
3869         spin_unlock_irq(&conf->device_lock);
3870
3871         unplug_slaves(mddev);
3872
3873         pr_debug("--- raid5d inactive\n");
3874 }
3875
3876 static ssize_t
3877 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
3878 {
3879         raid5_conf_t *conf = mddev_to_conf(mddev);
3880         if (conf)
3881                 return sprintf(page, "%d\n", conf->max_nr_stripes);
3882         else
3883                 return 0;
3884 }
3885
3886 static ssize_t
3887 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
3888 {
3889         raid5_conf_t *conf = mddev_to_conf(mddev);
3890         char *end;
3891         int new;
3892         if (len >= PAGE_SIZE)
3893                 return -EINVAL;
3894         if (!conf)
3895                 return -ENODEV;
3896
3897         new = simple_strtoul(page, &end, 10);
3898         if (!*page || (*end && *end != '\n') )
3899                 return -EINVAL;
3900         if (new <= 16 || new > 32768)
3901                 return -EINVAL;
3902         while (new < conf->max_nr_stripes) {
3903                 if (drop_one_stripe(conf))
3904                         conf->max_nr_stripes--;
3905                 else
3906                         break;
3907         }
3908         md_allow_write(mddev);
3909         while (new > conf->max_nr_stripes) {
3910                 if (grow_one_stripe(conf))
3911                         conf->max_nr_stripes++;
3912                 else break;
3913         }
3914         return len;
3915 }
3916
3917 static struct md_sysfs_entry
3918 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
3919                                 raid5_show_stripe_cache_size,
3920                                 raid5_store_stripe_cache_size);
3921
3922 static ssize_t
3923 stripe_cache_active_show(mddev_t *mddev, char *page)
3924 {
3925         raid5_conf_t *conf = mddev_to_conf(mddev);
3926         if (conf)
3927                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
3928         else
3929                 return 0;
3930 }
3931
3932 static struct md_sysfs_entry
3933 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
3934
3935 static struct attribute *raid5_attrs[] =  {
3936         &raid5_stripecache_size.attr,
3937         &raid5_stripecache_active.attr,
3938         NULL,
3939 };
3940 static struct attribute_group raid5_attrs_group = {
3941         .name = NULL,
3942         .attrs = raid5_attrs,
3943 };
3944
3945 static int run(mddev_t *mddev)
3946 {
3947         raid5_conf_t *conf;
3948         int raid_disk, memory;
3949         mdk_rdev_t *rdev;
3950         struct disk_info *disk;
3951         struct list_head *tmp;
3952         int working_disks = 0;
3953
3954         if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
3955                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
3956                        mdname(mddev), mddev->level);
3957                 return -EIO;
3958         }
3959
3960         if (mddev->reshape_position != MaxSector) {
3961                 /* Check that we can continue the reshape.
3962                  * Currently only disks can change, it must
3963                  * increase, and we must be past the point where
3964                  * a stripe over-writes itself
3965                  */
3966                 sector_t here_new, here_old;
3967                 int old_disks;
3968                 int max_degraded = (mddev->level == 5 ? 1 : 2);
3969
3970                 if (mddev->new_level != mddev->level ||
3971                     mddev->new_layout != mddev->layout ||
3972                     mddev->new_chunk != mddev->chunk_size) {
3973                         printk(KERN_ERR "raid5: %s: unsupported reshape "
3974                                "required - aborting.\n",
3975                                mdname(mddev));
3976                         return -EINVAL;
3977                 }
3978                 if (mddev->delta_disks <= 0) {
3979                         printk(KERN_ERR "raid5: %s: unsupported reshape "
3980                                "(reduce disks) required - aborting.\n",
3981                                mdname(mddev));
3982                         return -EINVAL;
3983                 }
3984                 old_disks = mddev->raid_disks - mddev->delta_disks;
3985                 /* reshape_position must be on a new-stripe boundary, and one
3986                  * further up in new geometry must map after here in old
3987                  * geometry.
3988                  */
3989                 here_new = mddev->reshape_position;
3990                 if (sector_div(here_new, (mddev->chunk_size>>9)*
3991                                (mddev->raid_disks - max_degraded))) {
3992                         printk(KERN_ERR "raid5: reshape_position not "
3993                                "on a stripe boundary\n");
3994                         return -EINVAL;
3995                 }
3996                 /* here_new is the stripe we will write to */
3997                 here_old = mddev->reshape_position;
3998                 sector_div(here_old, (mddev->chunk_size>>9)*
3999                            (old_disks-max_degraded));
4000                 /* here_old is the first stripe that we might need to read
4001                  * from */
4002                 if (here_new >= here_old) {
4003                         /* Reading from the same stripe as writing to - bad */
4004                         printk(KERN_ERR "raid5: reshape_position too early for "
4005                                "auto-recovery - aborting.\n");
4006                         return -EINVAL;
4007                 }
4008                 printk(KERN_INFO "raid5: reshape will continue\n");
4009                 /* OK, we should be able to continue; */
4010         }
4011
4012
4013         mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4014         if ((conf = mddev->private) == NULL)
4015                 goto abort;
4016         if (mddev->reshape_position == MaxSector) {
4017                 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4018         } else {
4019                 conf->raid_disks = mddev->raid_disks;
4020                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4021         }
4022
4023         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4024                               GFP_KERNEL);
4025         if (!conf->disks)
4026                 goto abort;
4027
4028         conf->mddev = mddev;
4029
4030         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4031                 goto abort;
4032
4033         if (mddev->level == 6) {
4034                 conf->spare_page = alloc_page(GFP_KERNEL);
4035                 if (!conf->spare_page)
4036                         goto abort;
4037         }
4038         spin_lock_init(&conf->device_lock);
4039         init_waitqueue_head(&conf->wait_for_stripe);
4040         init_waitqueue_head(&conf->wait_for_overlap);
4041         INIT_LIST_HEAD(&conf->handle_list);
4042         INIT_LIST_HEAD(&conf->delayed_list);
4043         INIT_LIST_HEAD(&conf->bitmap_list);
4044         INIT_LIST_HEAD(&conf->inactive_list);
4045         atomic_set(&conf->active_stripes, 0);
4046         atomic_set(&conf->preread_active_stripes, 0);
4047         atomic_set(&conf->active_aligned_reads, 0);
4048
4049         pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4050
4051         ITERATE_RDEV(mddev,rdev,tmp) {
4052                 raid_disk = rdev->raid_disk;
4053                 if (raid_disk >= conf->raid_disks
4054                     || raid_disk < 0)
4055                         continue;
4056                 disk = conf->disks + raid_disk;
4057
4058                 disk->rdev = rdev;
4059
4060                 if (test_bit(In_sync, &rdev->flags)) {
4061                         char b[BDEVNAME_SIZE];
4062                         printk(KERN_INFO "raid5: device %s operational as raid"
4063                                 " disk %d\n", bdevname(rdev->bdev,b),
4064                                 raid_disk);
4065                         working_disks++;
4066                 }
4067         }
4068
4069         /*
4070          * 0 for a fully functional array, 1 or 2 for a degraded array.
4071          */
4072         mddev->degraded = conf->raid_disks - working_disks;
4073         conf->mddev = mddev;
4074         conf->chunk_size = mddev->chunk_size;
4075         conf->level = mddev->level;
4076         if (conf->level == 6)
4077                 conf->max_degraded = 2;
4078         else
4079                 conf->max_degraded = 1;
4080         conf->algorithm = mddev->layout;
4081         conf->max_nr_stripes = NR_STRIPES;
4082         conf->expand_progress = mddev->reshape_position;
4083
4084         /* device size must be a multiple of chunk size */
4085         mddev->size &= ~(mddev->chunk_size/1024 -1);
4086         mddev->resync_max_sectors = mddev->size << 1;
4087
4088         if (conf->level == 6 && conf->raid_disks < 4) {
4089                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4090                        mdname(mddev), conf->raid_disks);
4091                 goto abort;
4092         }
4093         if (!conf->chunk_size || conf->chunk_size % 4) {
4094                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4095                         conf->chunk_size, mdname(mddev));
4096                 goto abort;
4097         }
4098         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4099                 printk(KERN_ERR 
4100                         "raid5: unsupported parity algorithm %d for %s\n",
4101                         conf->algorithm, mdname(mddev));
4102                 goto abort;
4103         }
4104         if (mddev->degraded > conf->max_degraded) {
4105                 printk(KERN_ERR "raid5: not enough operational devices for %s"
4106                         " (%d/%d failed)\n",
4107                         mdname(mddev), mddev->degraded, conf->raid_disks);
4108                 goto abort;
4109         }
4110
4111         if (mddev->degraded > 0 &&
4112             mddev->recovery_cp != MaxSector) {
4113                 if (mddev->ok_start_degraded)
4114                         printk(KERN_WARNING
4115                                "raid5: starting dirty degraded array: %s"
4116                                "- data corruption possible.\n",
4117                                mdname(mddev));
4118                 else {
4119                         printk(KERN_ERR
4120                                "raid5: cannot start dirty degraded array for %s\n",
4121                                mdname(mddev));
4122                         goto abort;
4123                 }
4124         }
4125
4126         {
4127                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4128                 if (!mddev->thread) {
4129                         printk(KERN_ERR 
4130                                 "raid5: couldn't allocate thread for %s\n",
4131                                 mdname(mddev));
4132                         goto abort;
4133                 }
4134         }
4135         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4136                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4137         if (grow_stripes(conf, conf->max_nr_stripes)) {
4138                 printk(KERN_ERR 
4139                         "raid5: couldn't allocate %dkB for buffers\n", memory);
4140                 shrink_stripes(conf);
4141                 md_unregister_thread(mddev->thread);
4142                 goto abort;
4143         } else
4144                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4145                         memory, mdname(mddev));
4146
4147         if (mddev->degraded == 0)
4148                 printk("raid5: raid level %d set %s active with %d out of %d"
4149                         " devices, algorithm %d\n", conf->level, mdname(mddev), 
4150                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4151                         conf->algorithm);
4152         else
4153                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4154                         " out of %d devices, algorithm %d\n", conf->level,
4155                         mdname(mddev), mddev->raid_disks - mddev->degraded,
4156                         mddev->raid_disks, conf->algorithm);
4157
4158         print_raid5_conf(conf);
4159
4160         if (conf->expand_progress != MaxSector) {
4161                 printk("...ok start reshape thread\n");
4162                 conf->expand_lo = conf->expand_progress;
4163                 atomic_set(&conf->reshape_stripes, 0);
4164                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4165                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4166                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4167                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4168                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4169                                                         "%s_reshape");
4170         }
4171
4172         /* read-ahead size must cover two whole stripes, which is
4173          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4174          */
4175         {
4176                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4177                 int stripe = data_disks *
4178                         (mddev->chunk_size / PAGE_SIZE);
4179                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4180                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4181         }
4182
4183         /* Ok, everything is just fine now */
4184         if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4185                 printk(KERN_WARNING
4186                        "raid5: failed to create sysfs attributes for %s\n",
4187                        mdname(mddev));
4188
4189         mddev->queue->unplug_fn = raid5_unplug_device;
4190         mddev->queue->issue_flush_fn = raid5_issue_flush;
4191         mddev->queue->backing_dev_info.congested_data = mddev;
4192         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4193
4194         mddev->array_size =  mddev->size * (conf->previous_raid_disks -
4195                                             conf->max_degraded);
4196
4197         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4198
4199         return 0;
4200 abort:
4201         if (conf) {
4202                 print_raid5_conf(conf);
4203                 safe_put_page(conf->spare_page);
4204                 kfree(conf->disks);
4205                 kfree(conf->stripe_hashtbl);
4206                 kfree(conf);
4207         }
4208         mddev->private = NULL;
4209         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4210         return -EIO;
4211 }
4212
4213
4214
4215 static int stop(mddev_t *mddev)
4216 {
4217         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4218
4219         md_unregister_thread(mddev->thread);
4220         mddev->thread = NULL;
4221         shrink_stripes(conf);
4222         kfree(conf->stripe_hashtbl);
4223         mddev->queue->backing_dev_info.congested_fn = NULL;
4224         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4225         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4226         kfree(conf->disks);
4227         kfree(conf);
4228         mddev->private = NULL;
4229         return 0;
4230 }
4231
4232 #ifdef DEBUG
4233 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4234 {
4235         int i;
4236
4237         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4238                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4239         seq_printf(seq, "sh %llu,  count %d.\n",
4240                    (unsigned long long)sh->sector, atomic_read(&sh->count));
4241         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4242         for (i = 0; i < sh->disks; i++) {
4243                 seq_printf(seq, "(cache%d: %p %ld) ",
4244                            i, sh->dev[i].page, sh->dev[i].flags);
4245         }
4246         seq_printf(seq, "\n");
4247 }
4248
4249 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4250 {
4251         struct stripe_head *sh;
4252         struct hlist_node *hn;
4253         int i;
4254
4255         spin_lock_irq(&conf->device_lock);
4256         for (i = 0; i < NR_HASH; i++) {
4257                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4258                         if (sh->raid_conf != conf)
4259                                 continue;
4260                         print_sh(seq, sh);
4261                 }
4262         }
4263         spin_unlock_irq(&conf->device_lock);
4264 }
4265 #endif
4266
4267 static void status (struct seq_file *seq, mddev_t *mddev)
4268 {
4269         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4270         int i;
4271
4272         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4273         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4274         for (i = 0; i < conf->raid_disks; i++)
4275                 seq_printf (seq, "%s",
4276                                conf->disks[i].rdev &&
4277                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4278         seq_printf (seq, "]");
4279 #ifdef DEBUG
4280         seq_printf (seq, "\n");
4281         printall(seq, conf);
4282 #endif
4283 }
4284
4285 static void print_raid5_conf (raid5_conf_t *conf)
4286 {
4287         int i;
4288         struct disk_info *tmp;
4289
4290         printk("RAID5 conf printout:\n");
4291         if (!conf) {
4292                 printk("(conf==NULL)\n");
4293                 return;
4294         }
4295         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4296                  conf->raid_disks - conf->mddev->degraded);
4297
4298         for (i = 0; i < conf->raid_disks; i++) {
4299                 char b[BDEVNAME_SIZE];
4300                 tmp = conf->disks + i;
4301                 if (tmp->rdev)
4302                 printk(" disk %d, o:%d, dev:%s\n",
4303                         i, !test_bit(Faulty, &tmp->rdev->flags),
4304                         bdevname(tmp->rdev->bdev,b));
4305         }
4306 }
4307
4308 static int raid5_spare_active(mddev_t *mddev)
4309 {
4310         int i;
4311         raid5_conf_t *conf = mddev->private;
4312         struct disk_info *tmp;
4313
4314         for (i = 0; i < conf->raid_disks; i++) {
4315                 tmp = conf->disks + i;
4316                 if (tmp->rdev
4317                     && !test_bit(Faulty, &tmp->rdev->flags)
4318                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4319                         unsigned long flags;
4320                         spin_lock_irqsave(&conf->device_lock, flags);
4321                         mddev->degraded--;
4322                         spin_unlock_irqrestore(&conf->device_lock, flags);
4323                 }
4324         }
4325         print_raid5_conf(conf);
4326         return 0;
4327 }
4328
4329 static int raid5_remove_disk(mddev_t *mddev, int number)
4330 {
4331         raid5_conf_t *conf = mddev->private;
4332         int err = 0;
4333         mdk_rdev_t *rdev;
4334         struct disk_info *p = conf->disks + number;
4335
4336         print_raid5_conf(conf);
4337         rdev = p->rdev;
4338         if (rdev) {
4339                 if (test_bit(In_sync, &rdev->flags) ||
4340                     atomic_read(&rdev->nr_pending)) {
4341                         err = -EBUSY;
4342                         goto abort;
4343                 }
4344                 p->rdev = NULL;
4345                 synchronize_rcu();
4346                 if (atomic_read(&rdev->nr_pending)) {
4347                         /* lost the race, try later */
4348                         err = -EBUSY;
4349                         p->rdev = rdev;
4350                 }
4351         }
4352 abort:
4353
4354         print_raid5_conf(conf);
4355         return err;
4356 }
4357
4358 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4359 {
4360         raid5_conf_t *conf = mddev->private;
4361         int found = 0;
4362         int disk;
4363         struct disk_info *p;
4364
4365         if (mddev->degraded > conf->max_degraded)
4366                 /* no point adding a device */
4367                 return 0;
4368
4369         /*
4370          * find the disk ... but prefer rdev->saved_raid_disk
4371          * if possible.
4372          */
4373         if (rdev->saved_raid_disk >= 0 &&
4374             conf->disks[rdev->saved_raid_disk].rdev == NULL)
4375                 disk = rdev->saved_raid_disk;
4376         else
4377                 disk = 0;
4378         for ( ; disk < conf->raid_disks; disk++)
4379                 if ((p=conf->disks + disk)->rdev == NULL) {
4380                         clear_bit(In_sync, &rdev->flags);
4381                         rdev->raid_disk = disk;
4382                         found = 1;
4383                         if (rdev->saved_raid_disk != disk)
4384                                 conf->fullsync = 1;
4385                         rcu_assign_pointer(p->rdev, rdev);
4386                         break;
4387                 }
4388         print_raid5_conf(conf);
4389         return found;
4390 }
4391
4392 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4393 {
4394         /* no resync is happening, and there is enough space
4395          * on all devices, so we can resize.
4396          * We need to make sure resync covers any new space.
4397          * If the array is shrinking we should possibly wait until
4398          * any io in the removed space completes, but it hardly seems
4399          * worth it.
4400          */
4401         raid5_conf_t *conf = mddev_to_conf(mddev);
4402
4403         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4404         mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4405         set_capacity(mddev->gendisk, mddev->array_size << 1);
4406         mddev->changed = 1;
4407         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
4408                 mddev->recovery_cp = mddev->size << 1;
4409                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4410         }
4411         mddev->size = sectors /2;
4412         mddev->resync_max_sectors = sectors;
4413         return 0;
4414 }
4415
4416 #ifdef CONFIG_MD_RAID5_RESHAPE
4417 static int raid5_check_reshape(mddev_t *mddev)
4418 {
4419         raid5_conf_t *conf = mddev_to_conf(mddev);
4420         int err;
4421
4422         if (mddev->delta_disks < 0 ||
4423             mddev->new_level != mddev->level)
4424                 return -EINVAL; /* Cannot shrink array or change level yet */
4425         if (mddev->delta_disks == 0)
4426                 return 0; /* nothing to do */
4427
4428         /* Can only proceed if there are plenty of stripe_heads.
4429          * We need a minimum of one full stripe,, and for sensible progress
4430          * it is best to have about 4 times that.
4431          * If we require 4 times, then the default 256 4K stripe_heads will
4432          * allow for chunk sizes up to 256K, which is probably OK.
4433          * If the chunk size is greater, user-space should request more
4434          * stripe_heads first.
4435          */
4436         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4437             (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4438                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
4439                        (mddev->chunk_size / STRIPE_SIZE)*4);
4440                 return -ENOSPC;
4441         }
4442
4443         err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4444         if (err)
4445                 return err;
4446
4447         if (mddev->degraded > conf->max_degraded)
4448                 return -EINVAL;
4449         /* looks like we might be able to manage this */
4450         return 0;
4451 }
4452
4453 static int raid5_start_reshape(mddev_t *mddev)
4454 {
4455         raid5_conf_t *conf = mddev_to_conf(mddev);
4456         mdk_rdev_t *rdev;
4457         struct list_head *rtmp;
4458         int spares = 0;
4459         int added_devices = 0;
4460         unsigned long flags;
4461
4462         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4463                 return -EBUSY;
4464
4465         ITERATE_RDEV(mddev, rdev, rtmp)
4466                 if (rdev->raid_disk < 0 &&
4467                     !test_bit(Faulty, &rdev->flags))
4468                         spares++;
4469
4470         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4471                 /* Not enough devices even to make a degraded array
4472                  * of that size
4473                  */
4474                 return -EINVAL;
4475
4476         atomic_set(&conf->reshape_stripes, 0);
4477         spin_lock_irq(&conf->device_lock);
4478         conf->previous_raid_disks = conf->raid_disks;
4479         conf->raid_disks += mddev->delta_disks;
4480         conf->expand_progress = 0;
4481         conf->expand_lo = 0;
4482         spin_unlock_irq(&conf->device_lock);
4483
4484         /* Add some new drives, as many as will fit.
4485          * We know there are enough to make the newly sized array work.
4486          */
4487         ITERATE_RDEV(mddev, rdev, rtmp)
4488                 if (rdev->raid_disk < 0 &&
4489                     !test_bit(Faulty, &rdev->flags)) {
4490                         if (raid5_add_disk(mddev, rdev)) {
4491                                 char nm[20];
4492                                 set_bit(In_sync, &rdev->flags);
4493                                 added_devices++;
4494                                 rdev->recovery_offset = 0;
4495                                 sprintf(nm, "rd%d", rdev->raid_disk);
4496                                 if (sysfs_create_link(&mddev->kobj,
4497                                                       &rdev->kobj, nm))
4498                                         printk(KERN_WARNING
4499                                                "raid5: failed to create "
4500                                                " link %s for %s\n",
4501                                                nm, mdname(mddev));
4502                         } else
4503                                 break;
4504                 }
4505
4506         spin_lock_irqsave(&conf->device_lock, flags);
4507         mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4508         spin_unlock_irqrestore(&conf->device_lock, flags);
4509         mddev->raid_disks = conf->raid_disks;
4510         mddev->reshape_position = 0;
4511         set_bit(MD_CHANGE_DEVS, &mddev->flags);
4512
4513         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4514         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4515         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4516         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4517         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4518                                                 "%s_reshape");
4519         if (!mddev->sync_thread) {
4520                 mddev->recovery = 0;
4521                 spin_lock_irq(&conf->device_lock);
4522                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4523                 conf->expand_progress = MaxSector;
4524                 spin_unlock_irq(&conf->device_lock);
4525                 return -EAGAIN;
4526         }
4527         md_wakeup_thread(mddev->sync_thread);
4528         md_new_event(mddev);
4529         return 0;
4530 }
4531 #endif
4532
4533 static void end_reshape(raid5_conf_t *conf)
4534 {
4535         struct block_device *bdev;
4536
4537         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4538                 conf->mddev->array_size = conf->mddev->size *
4539                         (conf->raid_disks - conf->max_degraded);
4540                 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4541                 conf->mddev->changed = 1;
4542
4543                 bdev = bdget_disk(conf->mddev->gendisk, 0);
4544                 if (bdev) {
4545                         mutex_lock(&bdev->bd_inode->i_mutex);
4546                         i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4547                         mutex_unlock(&bdev->bd_inode->i_mutex);
4548                         bdput(bdev);
4549                 }
4550                 spin_lock_irq(&conf->device_lock);
4551                 conf->expand_progress = MaxSector;
4552                 spin_unlock_irq(&conf->device_lock);
4553                 conf->mddev->reshape_position = MaxSector;
4554
4555                 /* read-ahead size must cover two whole stripes, which is
4556                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4557                  */
4558                 {
4559                         int data_disks = conf->previous_raid_disks - conf->max_degraded;
4560                         int stripe = data_disks *
4561                                 (conf->mddev->chunk_size / PAGE_SIZE);
4562                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4563                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4564                 }
4565         }
4566 }
4567
4568 static void raid5_quiesce(mddev_t *mddev, int state)
4569 {
4570         raid5_conf_t *conf = mddev_to_conf(mddev);
4571
4572         switch(state) {
4573         case 2: /* resume for a suspend */
4574                 wake_up(&conf->wait_for_overlap);
4575                 break;
4576
4577         case 1: /* stop all writes */
4578                 spin_lock_irq(&conf->device_lock);
4579                 conf->quiesce = 1;
4580                 wait_event_lock_irq(conf->wait_for_stripe,
4581                                     atomic_read(&conf->active_stripes) == 0 &&
4582                                     atomic_read(&conf->active_aligned_reads) == 0,
4583                                     conf->device_lock, /* nothing */);
4584                 spin_unlock_irq(&conf->device_lock);
4585                 break;
4586
4587         case 0: /* re-enable writes */
4588                 spin_lock_irq(&conf->device_lock);
4589                 conf->quiesce = 0;
4590                 wake_up(&conf->wait_for_stripe);
4591                 wake_up(&conf->wait_for_overlap);
4592                 spin_unlock_irq(&conf->device_lock);
4593                 break;
4594         }
4595 }
4596
4597 static struct mdk_personality raid6_personality =
4598 {
4599         .name           = "raid6",
4600         .level          = 6,
4601         .owner          = THIS_MODULE,
4602         .make_request   = make_request,
4603         .run            = run,
4604         .stop           = stop,
4605         .status         = status,
4606         .error_handler  = error,
4607         .hot_add_disk   = raid5_add_disk,
4608         .hot_remove_disk= raid5_remove_disk,
4609         .spare_active   = raid5_spare_active,
4610         .sync_request   = sync_request,
4611         .resize         = raid5_resize,
4612 #ifdef CONFIG_MD_RAID5_RESHAPE
4613         .check_reshape  = raid5_check_reshape,
4614         .start_reshape  = raid5_start_reshape,
4615 #endif
4616         .quiesce        = raid5_quiesce,
4617 };
4618 static struct mdk_personality raid5_personality =
4619 {
4620         .name           = "raid5",
4621         .level          = 5,
4622         .owner          = THIS_MODULE,
4623         .make_request   = make_request,
4624         .run            = run,
4625         .stop           = stop,
4626         .status         = status,
4627         .error_handler  = error,
4628         .hot_add_disk   = raid5_add_disk,
4629         .hot_remove_disk= raid5_remove_disk,
4630         .spare_active   = raid5_spare_active,
4631         .sync_request   = sync_request,
4632         .resize         = raid5_resize,
4633 #ifdef CONFIG_MD_RAID5_RESHAPE
4634         .check_reshape  = raid5_check_reshape,
4635         .start_reshape  = raid5_start_reshape,
4636 #endif
4637         .quiesce        = raid5_quiesce,
4638 };
4639
4640 static struct mdk_personality raid4_personality =
4641 {
4642         .name           = "raid4",
4643         .level          = 4,
4644         .owner          = THIS_MODULE,
4645         .make_request   = make_request,
4646         .run            = run,
4647         .stop           = stop,
4648         .status         = status,
4649         .error_handler  = error,
4650         .hot_add_disk   = raid5_add_disk,
4651         .hot_remove_disk= raid5_remove_disk,
4652         .spare_active   = raid5_spare_active,
4653         .sync_request   = sync_request,
4654         .resize         = raid5_resize,
4655 #ifdef CONFIG_MD_RAID5_RESHAPE
4656         .check_reshape  = raid5_check_reshape,
4657         .start_reshape  = raid5_start_reshape,
4658 #endif
4659         .quiesce        = raid5_quiesce,
4660 };
4661
4662 static int __init raid5_init(void)
4663 {
4664         int e;
4665
4666         e = raid6_select_algo();
4667         if ( e )
4668                 return e;
4669         register_md_personality(&raid6_personality);
4670         register_md_personality(&raid5_personality);
4671         register_md_personality(&raid4_personality);
4672         return 0;
4673 }
4674
4675 static void raid5_exit(void)
4676 {
4677         unregister_md_personality(&raid6_personality);
4678         unregister_md_personality(&raid5_personality);
4679         unregister_md_personality(&raid4_personality);
4680 }
4681
4682 module_init(raid5_init);
4683 module_exit(raid5_exit);
4684 MODULE_LICENSE("GPL");
4685 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4686 MODULE_ALIAS("md-raid5");
4687 MODULE_ALIAS("md-raid4");
4688 MODULE_ALIAS("md-level-5");
4689 MODULE_ALIAS("md-level-4");
4690 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4691 MODULE_ALIAS("md-raid6");
4692 MODULE_ALIAS("md-level-6");
4693
4694 /* This used to be two separate modules, they were: */
4695 MODULE_ALIAS("raid5");
4696 MODULE_ALIAS("raid6");