md: handle_stripe5 - add request/completion logic for async read 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 static int
1826 handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
1827 {
1828         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1829         int locked = 0;
1830
1831         if (rcw) {
1832                 /* if we are not expanding this is a proper write request, and
1833                  * there will be bios with new data to be drained into the
1834                  * stripe cache
1835                  */
1836                 if (!expand) {
1837                         set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1838                         sh->ops.count++;
1839                 }
1840
1841                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1842                 sh->ops.count++;
1843
1844                 for (i = disks; i--; ) {
1845                         struct r5dev *dev = &sh->dev[i];
1846
1847                         if (dev->towrite) {
1848                                 set_bit(R5_LOCKED, &dev->flags);
1849                                 if (!expand)
1850                                         clear_bit(R5_UPTODATE, &dev->flags);
1851                                 locked++;
1852                         }
1853                 }
1854         } else {
1855                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1856                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1857
1858                 set_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
1859                 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1860                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1861
1862                 sh->ops.count += 3;
1863
1864                 for (i = disks; i--; ) {
1865                         struct r5dev *dev = &sh->dev[i];
1866                         if (i == pd_idx)
1867                                 continue;
1868
1869                         /* For a read-modify write there may be blocks that are
1870                          * locked for reading while others are ready to be
1871                          * written so we distinguish these blocks by the
1872                          * R5_Wantprexor bit
1873                          */
1874                         if (dev->towrite &&
1875                             (test_bit(R5_UPTODATE, &dev->flags) ||
1876                             test_bit(R5_Wantcompute, &dev->flags))) {
1877                                 set_bit(R5_Wantprexor, &dev->flags);
1878                                 set_bit(R5_LOCKED, &dev->flags);
1879                                 clear_bit(R5_UPTODATE, &dev->flags);
1880                                 locked++;
1881                         }
1882                 }
1883         }
1884
1885         /* keep the parity disk locked while asynchronous operations
1886          * are in flight
1887          */
1888         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1889         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1890         locked++;
1891
1892         pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1893                 __FUNCTION__, (unsigned long long)sh->sector,
1894                 locked, sh->ops.pending);
1895
1896         return locked;
1897 }
1898
1899 /*
1900  * Each stripe/dev can have one or more bion attached.
1901  * toread/towrite point to the first in a chain.
1902  * The bi_next chain must be in order.
1903  */
1904 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1905 {
1906         struct bio **bip;
1907         raid5_conf_t *conf = sh->raid_conf;
1908         int firstwrite=0;
1909
1910         pr_debug("adding bh b#%llu to stripe s#%llu\n",
1911                 (unsigned long long)bi->bi_sector,
1912                 (unsigned long long)sh->sector);
1913
1914
1915         spin_lock(&sh->lock);
1916         spin_lock_irq(&conf->device_lock);
1917         if (forwrite) {
1918                 bip = &sh->dev[dd_idx].towrite;
1919                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1920                         firstwrite = 1;
1921         } else
1922                 bip = &sh->dev[dd_idx].toread;
1923         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1924                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1925                         goto overlap;
1926                 bip = & (*bip)->bi_next;
1927         }
1928         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1929                 goto overlap;
1930
1931         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1932         if (*bip)
1933                 bi->bi_next = *bip;
1934         *bip = bi;
1935         bi->bi_phys_segments ++;
1936         spin_unlock_irq(&conf->device_lock);
1937         spin_unlock(&sh->lock);
1938
1939         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1940                 (unsigned long long)bi->bi_sector,
1941                 (unsigned long long)sh->sector, dd_idx);
1942
1943         if (conf->mddev->bitmap && firstwrite) {
1944                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1945                                   STRIPE_SECTORS, 0);
1946                 sh->bm_seq = conf->seq_flush+1;
1947                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1948         }
1949
1950         if (forwrite) {
1951                 /* check if page is covered */
1952                 sector_t sector = sh->dev[dd_idx].sector;
1953                 for (bi=sh->dev[dd_idx].towrite;
1954                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1955                              bi && bi->bi_sector <= sector;
1956                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1957                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1958                                 sector = bi->bi_sector + (bi->bi_size>>9);
1959                 }
1960                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1961                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1962         }
1963         return 1;
1964
1965  overlap:
1966         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1967         spin_unlock_irq(&conf->device_lock);
1968         spin_unlock(&sh->lock);
1969         return 0;
1970 }
1971
1972 static void end_reshape(raid5_conf_t *conf);
1973
1974 static int page_is_zero(struct page *p)
1975 {
1976         char *a = page_address(p);
1977         return ((*(u32*)a) == 0 &&
1978                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1979 }
1980
1981 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1982 {
1983         int sectors_per_chunk = conf->chunk_size >> 9;
1984         int pd_idx, dd_idx;
1985         int chunk_offset = sector_div(stripe, sectors_per_chunk);
1986
1987         raid5_compute_sector(stripe * (disks - conf->max_degraded)
1988                              *sectors_per_chunk + chunk_offset,
1989                              disks, disks - conf->max_degraded,
1990                              &dd_idx, &pd_idx, conf);
1991         return pd_idx;
1992 }
1993
1994 static void
1995 handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1996                                 struct stripe_head_state *s, int disks,
1997                                 struct bio **return_bi)
1998 {
1999         int i;
2000         for (i = disks; i--; ) {
2001                 struct bio *bi;
2002                 int bitmap_end = 0;
2003
2004                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2005                         mdk_rdev_t *rdev;
2006                         rcu_read_lock();
2007                         rdev = rcu_dereference(conf->disks[i].rdev);
2008                         if (rdev && test_bit(In_sync, &rdev->flags))
2009                                 /* multiple read failures in one stripe */
2010                                 md_error(conf->mddev, rdev);
2011                         rcu_read_unlock();
2012                 }
2013                 spin_lock_irq(&conf->device_lock);
2014                 /* fail all writes first */
2015                 bi = sh->dev[i].towrite;
2016                 sh->dev[i].towrite = NULL;
2017                 if (bi) {
2018                         s->to_write--;
2019                         bitmap_end = 1;
2020                 }
2021
2022                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2023                         wake_up(&conf->wait_for_overlap);
2024
2025                 while (bi && bi->bi_sector <
2026                         sh->dev[i].sector + STRIPE_SECTORS) {
2027                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2028                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2029                         if (--bi->bi_phys_segments == 0) {
2030                                 md_write_end(conf->mddev);
2031                                 bi->bi_next = *return_bi;
2032                                 *return_bi = bi;
2033                         }
2034                         bi = nextbi;
2035                 }
2036                 /* and fail all 'written' */
2037                 bi = sh->dev[i].written;
2038                 sh->dev[i].written = NULL;
2039                 if (bi) bitmap_end = 1;
2040                 while (bi && bi->bi_sector <
2041                        sh->dev[i].sector + STRIPE_SECTORS) {
2042                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2043                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2044                         if (--bi->bi_phys_segments == 0) {
2045                                 md_write_end(conf->mddev);
2046                                 bi->bi_next = *return_bi;
2047                                 *return_bi = bi;
2048                         }
2049                         bi = bi2;
2050                 }
2051
2052                 /* fail any reads if this device is non-operational and
2053                  * the data has not reached the cache yet.
2054                  */
2055                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2056                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2057                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
2058                         bi = sh->dev[i].toread;
2059                         sh->dev[i].toread = NULL;
2060                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2061                                 wake_up(&conf->wait_for_overlap);
2062                         if (bi) s->to_read--;
2063                         while (bi && bi->bi_sector <
2064                                sh->dev[i].sector + STRIPE_SECTORS) {
2065                                 struct bio *nextbi =
2066                                         r5_next_bio(bi, sh->dev[i].sector);
2067                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2068                                 if (--bi->bi_phys_segments == 0) {
2069                                         bi->bi_next = *return_bi;
2070                                         *return_bi = bi;
2071                                 }
2072                                 bi = nextbi;
2073                         }
2074                 }
2075                 spin_unlock_irq(&conf->device_lock);
2076                 if (bitmap_end)
2077                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2078                                         STRIPE_SECTORS, 0, 0);
2079         }
2080
2081 }
2082
2083 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
2084  * to process
2085  */
2086 static int __handle_issuing_new_read_requests5(struct stripe_head *sh,
2087                         struct stripe_head_state *s, int disk_idx, int disks)
2088 {
2089         struct r5dev *dev = &sh->dev[disk_idx];
2090         struct r5dev *failed_dev = &sh->dev[s->failed_num];
2091
2092         /* don't schedule compute operations or reads on the parity block while
2093          * a check is in flight
2094          */
2095         if ((disk_idx == sh->pd_idx) &&
2096              test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2097                 return ~0;
2098
2099         /* is the data in this block needed, and can we get it? */
2100         if (!test_bit(R5_LOCKED, &dev->flags) &&
2101             !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread ||
2102             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2103              s->syncing || s->expanding || (s->failed &&
2104              (failed_dev->toread || (failed_dev->towrite &&
2105              !test_bit(R5_OVERWRITE, &failed_dev->flags)
2106              ))))) {
2107                 /* 1/ We would like to get this block, possibly by computing it,
2108                  * but we might not be able to.
2109                  *
2110                  * 2/ Since parity check operations potentially make the parity
2111                  * block !uptodate it will need to be refreshed before any
2112                  * compute operations on data disks are scheduled.
2113                  *
2114                  * 3/ We hold off parity block re-reads until check operations
2115                  * have quiesced.
2116                  */
2117                 if ((s->uptodate == disks - 1) &&
2118                     !test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2119                         set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2120                         set_bit(R5_Wantcompute, &dev->flags);
2121                         sh->ops.target = disk_idx;
2122                         s->req_compute = 1;
2123                         sh->ops.count++;
2124                         /* Careful: from this point on 'uptodate' is in the eye
2125                          * of raid5_run_ops which services 'compute' operations
2126                          * before writes. R5_Wantcompute flags a block that will
2127                          * be R5_UPTODATE by the time it is needed for a
2128                          * subsequent operation.
2129                          */
2130                         s->uptodate++;
2131                         return 0; /* uptodate + compute == disks */
2132                 } else if ((s->uptodate < disks - 1) &&
2133                         test_bit(R5_Insync, &dev->flags)) {
2134                         /* Note: we hold off compute operations while checks are
2135                          * in flight, but we still prefer 'compute' over 'read'
2136                          * hence we only read if (uptodate < * disks-1)
2137                          */
2138                         set_bit(R5_LOCKED, &dev->flags);
2139                         set_bit(R5_Wantread, &dev->flags);
2140                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2141                                 sh->ops.count++;
2142                         s->locked++;
2143                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2144                                 s->syncing);
2145                 }
2146         }
2147
2148         return ~0;
2149 }
2150
2151 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2152                         struct stripe_head_state *s, int disks)
2153 {
2154         int i;
2155
2156         /* Clear completed compute operations.  Parity recovery
2157          * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled
2158          * later on in this routine
2159          */
2160         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2161                 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2162                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2163                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2164                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2165         }
2166
2167         /* look for blocks to read/compute, skip this if a compute
2168          * is already in flight, or if the stripe contents are in the
2169          * midst of changing due to a write
2170          */
2171         if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2172                 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2173                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2174                 for (i = disks; i--; )
2175                         if (__handle_issuing_new_read_requests5(
2176                                 sh, s, i, disks) == 0)
2177                                 break;
2178         }
2179         set_bit(STRIPE_HANDLE, &sh->state);
2180 }
2181
2182 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2183                         struct stripe_head_state *s, struct r6_state *r6s,
2184                         int disks)
2185 {
2186         int i;
2187         for (i = disks; i--; ) {
2188                 struct r5dev *dev = &sh->dev[i];
2189                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2190                     !test_bit(R5_UPTODATE, &dev->flags) &&
2191                     (dev->toread || (dev->towrite &&
2192                      !test_bit(R5_OVERWRITE, &dev->flags)) ||
2193                      s->syncing || s->expanding ||
2194                      (s->failed >= 1 &&
2195                       (sh->dev[r6s->failed_num[0]].toread ||
2196                        s->to_write)) ||
2197                      (s->failed >= 2 &&
2198                       (sh->dev[r6s->failed_num[1]].toread ||
2199                        s->to_write)))) {
2200                         /* we would like to get this block, possibly
2201                          * by computing it, but we might not be able to
2202                          */
2203                         if (s->uptodate == disks-1) {
2204                                 pr_debug("Computing stripe %llu block %d\n",
2205                                        (unsigned long long)sh->sector, i);
2206                                 compute_block_1(sh, i, 0);
2207                                 s->uptodate++;
2208                         } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2209                                 /* Computing 2-failure is *very* expensive; only
2210                                  * do it if failed >= 2
2211                                  */
2212                                 int other;
2213                                 for (other = disks; other--; ) {
2214                                         if (other == i)
2215                                                 continue;
2216                                         if (!test_bit(R5_UPTODATE,
2217                                               &sh->dev[other].flags))
2218                                                 break;
2219                                 }
2220                                 BUG_ON(other < 0);
2221                                 pr_debug("Computing stripe %llu blocks %d,%d\n",
2222                                        (unsigned long long)sh->sector,
2223                                        i, other);
2224                                 compute_block_2(sh, i, other);
2225                                 s->uptodate += 2;
2226                         } else if (test_bit(R5_Insync, &dev->flags)) {
2227                                 set_bit(R5_LOCKED, &dev->flags);
2228                                 set_bit(R5_Wantread, &dev->flags);
2229                                 s->locked++;
2230                                 pr_debug("Reading block %d (sync=%d)\n",
2231                                         i, s->syncing);
2232                         }
2233                 }
2234         }
2235         set_bit(STRIPE_HANDLE, &sh->state);
2236 }
2237
2238
2239 /* handle_completed_write_requests
2240  * any written block on an uptodate or failed drive can be returned.
2241  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2242  * never LOCKED, so we don't need to test 'failed' directly.
2243  */
2244 static void handle_completed_write_requests(raid5_conf_t *conf,
2245         struct stripe_head *sh, int disks, struct bio **return_bi)
2246 {
2247         int i;
2248         struct r5dev *dev;
2249
2250         for (i = disks; i--; )
2251                 if (sh->dev[i].written) {
2252                         dev = &sh->dev[i];
2253                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2254                                 test_bit(R5_UPTODATE, &dev->flags)) {
2255                                 /* We can return any write requests */
2256                                 struct bio *wbi, *wbi2;
2257                                 int bitmap_end = 0;
2258                                 pr_debug("Return write for disc %d\n", i);
2259                                 spin_lock_irq(&conf->device_lock);
2260                                 wbi = dev->written;
2261                                 dev->written = NULL;
2262                                 while (wbi && wbi->bi_sector <
2263                                         dev->sector + STRIPE_SECTORS) {
2264                                         wbi2 = r5_next_bio(wbi, dev->sector);
2265                                         if (--wbi->bi_phys_segments == 0) {
2266                                                 md_write_end(conf->mddev);
2267                                                 wbi->bi_next = *return_bi;
2268                                                 *return_bi = wbi;
2269                                         }
2270                                         wbi = wbi2;
2271                                 }
2272                                 if (dev->towrite == NULL)
2273                                         bitmap_end = 1;
2274                                 spin_unlock_irq(&conf->device_lock);
2275                                 if (bitmap_end)
2276                                         bitmap_endwrite(conf->mddev->bitmap,
2277                                                         sh->sector,
2278                                                         STRIPE_SECTORS,
2279                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2280                                                         0);
2281                         }
2282                 }
2283 }
2284
2285 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2286                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2287 {
2288         int rmw = 0, rcw = 0, i;
2289         for (i = disks; i--; ) {
2290                 /* would I have to read this buffer for read_modify_write */
2291                 struct r5dev *dev = &sh->dev[i];
2292                 if ((dev->towrite || i == sh->pd_idx) &&
2293                     !test_bit(R5_LOCKED, &dev->flags) &&
2294                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2295                       test_bit(R5_Wantcompute, &dev->flags))) {
2296                         if (test_bit(R5_Insync, &dev->flags))
2297                                 rmw++;
2298                         else
2299                                 rmw += 2*disks;  /* cannot read it */
2300                 }
2301                 /* Would I have to read this buffer for reconstruct_write */
2302                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2303                     !test_bit(R5_LOCKED, &dev->flags) &&
2304                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2305                     test_bit(R5_Wantcompute, &dev->flags))) {
2306                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2307                         else
2308                                 rcw += 2*disks;
2309                 }
2310         }
2311         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2312                 (unsigned long long)sh->sector, rmw, rcw);
2313         set_bit(STRIPE_HANDLE, &sh->state);
2314         if (rmw < rcw && rmw > 0)
2315                 /* prefer read-modify-write, but need to get some data */
2316                 for (i = disks; i--; ) {
2317                         struct r5dev *dev = &sh->dev[i];
2318                         if ((dev->towrite || i == sh->pd_idx) &&
2319                             !test_bit(R5_LOCKED, &dev->flags) &&
2320                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2321                             test_bit(R5_Wantcompute, &dev->flags)) &&
2322                             test_bit(R5_Insync, &dev->flags)) {
2323                                 if (
2324                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2325                                         pr_debug("Read_old block "
2326                                                 "%d for r-m-w\n", i);
2327                                         set_bit(R5_LOCKED, &dev->flags);
2328                                         set_bit(R5_Wantread, &dev->flags);
2329                                         s->locked++;
2330                                 } else {
2331                                         set_bit(STRIPE_DELAYED, &sh->state);
2332                                         set_bit(STRIPE_HANDLE, &sh->state);
2333                                 }
2334                         }
2335                 }
2336         if (rcw <= rmw && rcw > 0)
2337                 /* want reconstruct write, but need to get some data */
2338                 for (i = disks; i--; ) {
2339                         struct r5dev *dev = &sh->dev[i];
2340                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2341                             i != sh->pd_idx &&
2342                             !test_bit(R5_LOCKED, &dev->flags) &&
2343                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2344                             test_bit(R5_Wantcompute, &dev->flags)) &&
2345                             test_bit(R5_Insync, &dev->flags)) {
2346                                 if (
2347                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2348                                         pr_debug("Read_old block "
2349                                                 "%d for Reconstruct\n", i);
2350                                         set_bit(R5_LOCKED, &dev->flags);
2351                                         set_bit(R5_Wantread, &dev->flags);
2352                                         s->locked++;
2353                                 } else {
2354                                         set_bit(STRIPE_DELAYED, &sh->state);
2355                                         set_bit(STRIPE_HANDLE, &sh->state);
2356                                 }
2357                         }
2358                 }
2359         /* now if nothing is locked, and if we have enough data,
2360          * we can start a write request
2361          */
2362         /* since handle_stripe can be called at any time we need to handle the
2363          * case where a compute block operation has been submitted and then a
2364          * subsequent call wants to start a write request.  raid5_run_ops only
2365          * handles the case where compute block and postxor are requested
2366          * simultaneously.  If this is not the case then new writes need to be
2367          * held off until the compute completes.
2368          */
2369         if ((s->req_compute ||
2370             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) &&
2371                 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2372                 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2373                 s->locked += handle_write_operations5(sh, rcw == 0, 0);
2374 }
2375
2376 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2377                 struct stripe_head *sh, struct stripe_head_state *s,
2378                 struct r6_state *r6s, int disks)
2379 {
2380         int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2381         int qd_idx = r6s->qd_idx;
2382         for (i = disks; i--; ) {
2383                 struct r5dev *dev = &sh->dev[i];
2384                 /* Would I have to read this buffer for reconstruct_write */
2385                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2386                     && i != pd_idx && i != qd_idx
2387                     && (!test_bit(R5_LOCKED, &dev->flags)
2388                             ) &&
2389                     !test_bit(R5_UPTODATE, &dev->flags)) {
2390                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2391                         else {
2392                                 pr_debug("raid6: must_compute: "
2393                                         "disk %d flags=%#lx\n", i, dev->flags);
2394                                 must_compute++;
2395                         }
2396                 }
2397         }
2398         pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2399                (unsigned long long)sh->sector, rcw, must_compute);
2400         set_bit(STRIPE_HANDLE, &sh->state);
2401
2402         if (rcw > 0)
2403                 /* want reconstruct write, but need to get some data */
2404                 for (i = disks; i--; ) {
2405                         struct r5dev *dev = &sh->dev[i];
2406                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2407                             && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2408                             && !test_bit(R5_LOCKED, &dev->flags) &&
2409                             !test_bit(R5_UPTODATE, &dev->flags) &&
2410                             test_bit(R5_Insync, &dev->flags)) {
2411                                 if (
2412                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2413                                         pr_debug("Read_old stripe %llu "
2414                                                 "block %d for Reconstruct\n",
2415                                              (unsigned long long)sh->sector, i);
2416                                         set_bit(R5_LOCKED, &dev->flags);
2417                                         set_bit(R5_Wantread, &dev->flags);
2418                                         s->locked++;
2419                                 } else {
2420                                         pr_debug("Request delayed stripe %llu "
2421                                                 "block %d for Reconstruct\n",
2422                                              (unsigned long long)sh->sector, i);
2423                                         set_bit(STRIPE_DELAYED, &sh->state);
2424                                         set_bit(STRIPE_HANDLE, &sh->state);
2425                                 }
2426                         }
2427                 }
2428         /* now if nothing is locked, and if we have enough data, we can start a
2429          * write request
2430          */
2431         if (s->locked == 0 && rcw == 0 &&
2432             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2433                 if (must_compute > 0) {
2434                         /* We have failed blocks and need to compute them */
2435                         switch (s->failed) {
2436                         case 0:
2437                                 BUG();
2438                         case 1:
2439                                 compute_block_1(sh, r6s->failed_num[0], 0);
2440                                 break;
2441                         case 2:
2442                                 compute_block_2(sh, r6s->failed_num[0],
2443                                                 r6s->failed_num[1]);
2444                                 break;
2445                         default: /* This request should have been failed? */
2446                                 BUG();
2447                         }
2448                 }
2449
2450                 pr_debug("Computing parity for stripe %llu\n",
2451                         (unsigned long long)sh->sector);
2452                 compute_parity6(sh, RECONSTRUCT_WRITE);
2453                 /* now every locked buffer is ready to be written */
2454                 for (i = disks; i--; )
2455                         if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2456                                 pr_debug("Writing stripe %llu block %d\n",
2457                                        (unsigned long long)sh->sector, i);
2458                                 s->locked++;
2459                                 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2460                         }
2461                 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2462                 set_bit(STRIPE_INSYNC, &sh->state);
2463
2464                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2465                         atomic_dec(&conf->preread_active_stripes);
2466                         if (atomic_read(&conf->preread_active_stripes) <
2467                             IO_THRESHOLD)
2468                                 md_wakeup_thread(conf->mddev->thread);
2469                 }
2470         }
2471 }
2472
2473 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2474                                 struct stripe_head_state *s, int disks)
2475 {
2476         set_bit(STRIPE_HANDLE, &sh->state);
2477         /* Take one of the following actions:
2478          * 1/ start a check parity operation if (uptodate == disks)
2479          * 2/ finish a check parity operation and act on the result
2480          * 3/ skip to the writeback section if we previously
2481          *    initiated a recovery operation
2482          */
2483         if (s->failed == 0 &&
2484             !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2485                 if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2486                         BUG_ON(s->uptodate != disks);
2487                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2488                         sh->ops.count++;
2489                         s->uptodate--;
2490                 } else if (
2491                        test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) {
2492                         clear_bit(STRIPE_OP_CHECK, &sh->ops.ack);
2493                         clear_bit(STRIPE_OP_CHECK, &sh->ops.pending);
2494
2495                         if (sh->ops.zero_sum_result == 0)
2496                                 /* parity is correct (on disc,
2497                                  * not in buffer any more)
2498                                  */
2499                                 set_bit(STRIPE_INSYNC, &sh->state);
2500                         else {
2501                                 conf->mddev->resync_mismatches +=
2502                                         STRIPE_SECTORS;
2503                                 if (test_bit(
2504                                      MD_RECOVERY_CHECK, &conf->mddev->recovery))
2505                                         /* don't try to repair!! */
2506                                         set_bit(STRIPE_INSYNC, &sh->state);
2507                                 else {
2508                                         set_bit(STRIPE_OP_COMPUTE_BLK,
2509                                                 &sh->ops.pending);
2510                                         set_bit(STRIPE_OP_MOD_REPAIR_PD,
2511                                                 &sh->ops.pending);
2512                                         set_bit(R5_Wantcompute,
2513                                                 &sh->dev[sh->pd_idx].flags);
2514                                         sh->ops.target = sh->pd_idx;
2515                                         sh->ops.count++;
2516                                         s->uptodate++;
2517                                 }
2518                         }
2519                 }
2520         }
2521
2522         /* check if we can clear a parity disk reconstruct */
2523         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2524                 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2525
2526                 clear_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending);
2527                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2528                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2529                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2530         }
2531
2532         /* Wait for check parity and compute block operations to complete
2533          * before write-back
2534          */
2535         if (!test_bit(STRIPE_INSYNC, &sh->state) &&
2536                 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) &&
2537                 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) {
2538                 struct r5dev *dev;
2539                 /* either failed parity check, or recovery is happening */
2540                 if (s->failed == 0)
2541                         s->failed_num = sh->pd_idx;
2542                 dev = &sh->dev[s->failed_num];
2543                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2544                 BUG_ON(s->uptodate != disks);
2545
2546                 set_bit(R5_LOCKED, &dev->flags);
2547                 set_bit(R5_Wantwrite, &dev->flags);
2548                 clear_bit(STRIPE_DEGRADED, &sh->state);
2549                 s->locked++;
2550                 set_bit(STRIPE_INSYNC, &sh->state);
2551         }
2552 }
2553
2554
2555 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2556                                 struct stripe_head_state *s,
2557                                 struct r6_state *r6s, struct page *tmp_page,
2558                                 int disks)
2559 {
2560         int update_p = 0, update_q = 0;
2561         struct r5dev *dev;
2562         int pd_idx = sh->pd_idx;
2563         int qd_idx = r6s->qd_idx;
2564
2565         set_bit(STRIPE_HANDLE, &sh->state);
2566
2567         BUG_ON(s->failed > 2);
2568         BUG_ON(s->uptodate < disks);
2569         /* Want to check and possibly repair P and Q.
2570          * However there could be one 'failed' device, in which
2571          * case we can only check one of them, possibly using the
2572          * other to generate missing data
2573          */
2574
2575         /* If !tmp_page, we cannot do the calculations,
2576          * but as we have set STRIPE_HANDLE, we will soon be called
2577          * by stripe_handle with a tmp_page - just wait until then.
2578          */
2579         if (tmp_page) {
2580                 if (s->failed == r6s->q_failed) {
2581                         /* The only possible failed device holds 'Q', so it
2582                          * makes sense to check P (If anything else were failed,
2583                          * we would have used P to recreate it).
2584                          */
2585                         compute_block_1(sh, pd_idx, 1);
2586                         if (!page_is_zero(sh->dev[pd_idx].page)) {
2587                                 compute_block_1(sh, pd_idx, 0);
2588                                 update_p = 1;
2589                         }
2590                 }
2591                 if (!r6s->q_failed && s->failed < 2) {
2592                         /* q is not failed, and we didn't use it to generate
2593                          * anything, so it makes sense to check it
2594                          */
2595                         memcpy(page_address(tmp_page),
2596                                page_address(sh->dev[qd_idx].page),
2597                                STRIPE_SIZE);
2598                         compute_parity6(sh, UPDATE_PARITY);
2599                         if (memcmp(page_address(tmp_page),
2600                                    page_address(sh->dev[qd_idx].page),
2601                                    STRIPE_SIZE) != 0) {
2602                                 clear_bit(STRIPE_INSYNC, &sh->state);
2603                                 update_q = 1;
2604                         }
2605                 }
2606                 if (update_p || update_q) {
2607                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2608                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2609                                 /* don't try to repair!! */
2610                                 update_p = update_q = 0;
2611                 }
2612
2613                 /* now write out any block on a failed drive,
2614                  * or P or Q if they need it
2615                  */
2616
2617                 if (s->failed == 2) {
2618                         dev = &sh->dev[r6s->failed_num[1]];
2619                         s->locked++;
2620                         set_bit(R5_LOCKED, &dev->flags);
2621                         set_bit(R5_Wantwrite, &dev->flags);
2622                 }
2623                 if (s->failed >= 1) {
2624                         dev = &sh->dev[r6s->failed_num[0]];
2625                         s->locked++;
2626                         set_bit(R5_LOCKED, &dev->flags);
2627                         set_bit(R5_Wantwrite, &dev->flags);
2628                 }
2629
2630                 if (update_p) {
2631                         dev = &sh->dev[pd_idx];
2632                         s->locked++;
2633                         set_bit(R5_LOCKED, &dev->flags);
2634                         set_bit(R5_Wantwrite, &dev->flags);
2635                 }
2636                 if (update_q) {
2637                         dev = &sh->dev[qd_idx];
2638                         s->locked++;
2639                         set_bit(R5_LOCKED, &dev->flags);
2640                         set_bit(R5_Wantwrite, &dev->flags);
2641                 }
2642                 clear_bit(STRIPE_DEGRADED, &sh->state);
2643
2644                 set_bit(STRIPE_INSYNC, &sh->state);
2645         }
2646 }
2647
2648 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2649                                 struct r6_state *r6s)
2650 {
2651         int i;
2652
2653         /* We have read all the blocks in this stripe and now we need to
2654          * copy some of them into a target stripe for expand.
2655          */
2656         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2657         for (i = 0; i < sh->disks; i++)
2658                 if (i != sh->pd_idx && (r6s && i != r6s->qd_idx)) {
2659                         int dd_idx, pd_idx, j;
2660                         struct stripe_head *sh2;
2661
2662                         sector_t bn = compute_blocknr(sh, i);
2663                         sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2664                                                 conf->raid_disks -
2665                                                 conf->max_degraded, &dd_idx,
2666                                                 &pd_idx, conf);
2667                         sh2 = get_active_stripe(conf, s, conf->raid_disks,
2668                                                 pd_idx, 1);
2669                         if (sh2 == NULL)
2670                                 /* so far only the early blocks of this stripe
2671                                  * have been requested.  When later blocks
2672                                  * get requested, we will try again
2673                                  */
2674                                 continue;
2675                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2676                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2677                                 /* must have already done this block */
2678                                 release_stripe(sh2);
2679                                 continue;
2680                         }
2681                         memcpy(page_address(sh2->dev[dd_idx].page),
2682                                page_address(sh->dev[i].page),
2683                                STRIPE_SIZE);
2684                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2685                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2686                         for (j = 0; j < conf->raid_disks; j++)
2687                                 if (j != sh2->pd_idx &&
2688                                     (r6s && j != r6s->qd_idx) &&
2689                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2690                                         break;
2691                         if (j == conf->raid_disks) {
2692                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2693                                 set_bit(STRIPE_HANDLE, &sh2->state);
2694                         }
2695                         release_stripe(sh2);
2696                 }
2697 }
2698
2699 /*
2700  * handle_stripe - do things to a stripe.
2701  *
2702  * We lock the stripe and then examine the state of various bits
2703  * to see what needs to be done.
2704  * Possible results:
2705  *    return some read request which now have data
2706  *    return some write requests which are safely on disc
2707  *    schedule a read on some buffers
2708  *    schedule a write of some buffers
2709  *    return confirmation of parity correctness
2710  *
2711  * buffers are taken off read_list or write_list, and bh_cache buffers
2712  * get BH_Lock set before the stripe lock is released.
2713  *
2714  */
2715
2716 static void handle_stripe5(struct stripe_head *sh)
2717 {
2718         raid5_conf_t *conf = sh->raid_conf;
2719         int disks = sh->disks, i;
2720         struct bio *return_bi = NULL;
2721         struct stripe_head_state s;
2722         struct r5dev *dev;
2723         unsigned long pending = 0;
2724
2725         memset(&s, 0, sizeof(s));
2726         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2727                 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2728                 atomic_read(&sh->count), sh->pd_idx,
2729                 sh->ops.pending, sh->ops.ack, sh->ops.complete);
2730
2731         spin_lock(&sh->lock);
2732         clear_bit(STRIPE_HANDLE, &sh->state);
2733         clear_bit(STRIPE_DELAYED, &sh->state);
2734
2735         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2736         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2737         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2738         /* Now to look around and see what can be done */
2739
2740         rcu_read_lock();
2741         for (i=disks; i--; ) {
2742                 mdk_rdev_t *rdev;
2743                 struct r5dev *dev = &sh->dev[i];
2744                 clear_bit(R5_Insync, &dev->flags);
2745
2746                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2747                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2748                         dev->towrite, dev->written);
2749
2750                 /* maybe we can request a biofill operation
2751                  *
2752                  * new wantfill requests are only permitted while
2753                  * STRIPE_OP_BIOFILL is clear
2754                  */
2755                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2756                         !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2757                         set_bit(R5_Wantfill, &dev->flags);
2758
2759                 /* now count some things */
2760                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2761                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2762                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2763
2764                 if (test_bit(R5_Wantfill, &dev->flags))
2765                         s.to_fill++;
2766                 else if (dev->toread)
2767                         s.to_read++;
2768                 if (dev->towrite) {
2769                         s.to_write++;
2770                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2771                                 s.non_overwrite++;
2772                 }
2773                 if (dev->written)
2774                         s.written++;
2775                 rdev = rcu_dereference(conf->disks[i].rdev);
2776                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2777                         /* The ReadError flag will just be confusing now */
2778                         clear_bit(R5_ReadError, &dev->flags);
2779                         clear_bit(R5_ReWrite, &dev->flags);
2780                 }
2781                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2782                     || test_bit(R5_ReadError, &dev->flags)) {
2783                         s.failed++;
2784                         s.failed_num = i;
2785                 } else
2786                         set_bit(R5_Insync, &dev->flags);
2787         }
2788         rcu_read_unlock();
2789
2790         if (s.to_fill && !test_and_set_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2791                 sh->ops.count++;
2792
2793         pr_debug("locked=%d uptodate=%d to_read=%d"
2794                 " to_write=%d failed=%d failed_num=%d\n",
2795                 s.locked, s.uptodate, s.to_read, s.to_write,
2796                 s.failed, s.failed_num);
2797         /* check if the array has lost two devices and, if so, some requests might
2798          * need to be failed
2799          */
2800         if (s.failed > 1 && s.to_read+s.to_write+s.written)
2801                 handle_requests_to_failed_array(conf, sh, &s, disks,
2802                                                 &return_bi);
2803         if (s.failed > 1 && s.syncing) {
2804                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2805                 clear_bit(STRIPE_SYNCING, &sh->state);
2806                 s.syncing = 0;
2807         }
2808
2809         /* might be able to return some write requests if the parity block
2810          * is safe, or on a failed drive
2811          */
2812         dev = &sh->dev[sh->pd_idx];
2813         if ( s.written &&
2814              ((test_bit(R5_Insync, &dev->flags) &&
2815                !test_bit(R5_LOCKED, &dev->flags) &&
2816                test_bit(R5_UPTODATE, &dev->flags)) ||
2817                (s.failed == 1 && s.failed_num == sh->pd_idx)))
2818                 handle_completed_write_requests(conf, sh, disks, &return_bi);
2819
2820         /* Now we might consider reading some blocks, either to check/generate
2821          * parity, or to satisfy requests
2822          * or to load a block that is being partially written.
2823          */
2824         if (s.to_read || s.non_overwrite ||
2825             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding ||
2826             test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2827                 handle_issuing_new_read_requests5(sh, &s, disks);
2828
2829         /* Now we check to see if any write operations have recently
2830          * completed
2831          */
2832
2833         /* leave prexor set until postxor is done, allows us to distinguish
2834          * a rmw from a rcw during biodrain
2835          */
2836         if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2837                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2838
2839                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2840                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2841                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2842
2843                 for (i = disks; i--; )
2844                         clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2845         }
2846
2847         /* if only POSTXOR is set then this is an 'expand' postxor */
2848         if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2849                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2850
2851                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2852                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2853                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2854
2855                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2856                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2857                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2858
2859                 /* All the 'written' buffers and the parity block are ready to
2860                  * be written back to disk
2861                  */
2862                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2863                 for (i = disks; i--; ) {
2864                         dev = &sh->dev[i];
2865                         if (test_bit(R5_LOCKED, &dev->flags) &&
2866                                 (i == sh->pd_idx || dev->written)) {
2867                                 pr_debug("Writing block %d\n", i);
2868                                 set_bit(R5_Wantwrite, &dev->flags);
2869                                 if (!test_and_set_bit(
2870                                     STRIPE_OP_IO, &sh->ops.pending))
2871                                         sh->ops.count++;
2872                                 if (!test_bit(R5_Insync, &dev->flags) ||
2873                                     (i == sh->pd_idx && s.failed == 0))
2874                                         set_bit(STRIPE_INSYNC, &sh->state);
2875                         }
2876                 }
2877                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2878                         atomic_dec(&conf->preread_active_stripes);
2879                         if (atomic_read(&conf->preread_active_stripes) <
2880                                 IO_THRESHOLD)
2881                                 md_wakeup_thread(conf->mddev->thread);
2882                 }
2883         }
2884
2885         /* Now to consider new write requests and what else, if anything
2886          * should be read.  We do not handle new writes when:
2887          * 1/ A 'write' operation (copy+xor) is already in flight.
2888          * 2/ A 'check' operation is in flight, as it may clobber the parity
2889          *    block.
2890          */
2891         if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2892                           !test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2893                 handle_issuing_new_write_requests5(conf, sh, &s, disks);
2894
2895         /* maybe we need to check and possibly fix the parity for this stripe
2896          * Any reads will already have been scheduled, so we just see if enough
2897          * data is available.  The parity check is held off while parity
2898          * dependent operations are in flight.
2899          */
2900         if ((s.syncing && s.locked == 0 &&
2901              !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2902              !test_bit(STRIPE_INSYNC, &sh->state)) ||
2903               test_bit(STRIPE_OP_CHECK, &sh->ops.pending) ||
2904               test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending))
2905                 handle_parity_checks5(conf, sh, &s, disks);
2906
2907         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2908                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2909                 clear_bit(STRIPE_SYNCING, &sh->state);
2910         }
2911
2912         /* If the failed drive is just a ReadError, then we might need to progress
2913          * the repair/check process
2914          */
2915         if (s.failed == 1 && !conf->mddev->ro &&
2916             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2917             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2918             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2919                 ) {
2920                 dev = &sh->dev[s.failed_num];
2921                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2922                         set_bit(R5_Wantwrite, &dev->flags);
2923                         set_bit(R5_ReWrite, &dev->flags);
2924                         set_bit(R5_LOCKED, &dev->flags);
2925                         s.locked++;
2926                 } else {
2927                         /* let's read it back */
2928                         set_bit(R5_Wantread, &dev->flags);
2929                         set_bit(R5_LOCKED, &dev->flags);
2930                         s.locked++;
2931                 }
2932         }
2933
2934         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
2935                 /* Need to write out all blocks after computing parity */
2936                 sh->disks = conf->raid_disks;
2937                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf, conf->raid_disks);
2938                 compute_parity5(sh, RECONSTRUCT_WRITE);
2939                 for (i = conf->raid_disks; i--; ) {
2940                         set_bit(R5_LOCKED, &sh->dev[i].flags);
2941                         s.locked++;
2942                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2943                 }
2944                 clear_bit(STRIPE_EXPANDING, &sh->state);
2945         } else if (s.expanded) {
2946                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2947                 atomic_dec(&conf->reshape_stripes);
2948                 wake_up(&conf->wait_for_overlap);
2949                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2950         }
2951
2952         if (s.expanding && s.locked == 0)
2953                 handle_stripe_expansion(conf, sh, NULL);
2954
2955         if (sh->ops.count)
2956                 pending = get_stripe_work(sh);
2957
2958         spin_unlock(&sh->lock);
2959
2960         if (pending)
2961                 raid5_run_ops(sh, pending);
2962
2963         return_io(return_bi);
2964
2965         for (i=disks; i-- ;) {
2966                 int rw;
2967                 struct bio *bi;
2968                 mdk_rdev_t *rdev;
2969                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
2970                         rw = WRITE;
2971                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
2972                         rw = READ;
2973                 else
2974                         continue;
2975  
2976                 bi = &sh->dev[i].req;
2977  
2978                 bi->bi_rw = rw;
2979                 if (rw == WRITE)
2980                         bi->bi_end_io = raid5_end_write_request;
2981                 else
2982                         bi->bi_end_io = raid5_end_read_request;
2983  
2984                 rcu_read_lock();
2985                 rdev = rcu_dereference(conf->disks[i].rdev);
2986                 if (rdev && test_bit(Faulty, &rdev->flags))
2987                         rdev = NULL;
2988                 if (rdev)
2989                         atomic_inc(&rdev->nr_pending);
2990                 rcu_read_unlock();
2991  
2992                 if (rdev) {
2993                         if (s.syncing || s.expanding || s.expanded)
2994                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
2995
2996                         bi->bi_bdev = rdev->bdev;
2997                         pr_debug("for %llu schedule op %ld on disc %d\n",
2998                                 (unsigned long long)sh->sector, bi->bi_rw, i);
2999                         atomic_inc(&sh->count);
3000                         bi->bi_sector = sh->sector + rdev->data_offset;
3001                         bi->bi_flags = 1 << BIO_UPTODATE;
3002                         bi->bi_vcnt = 1;        
3003                         bi->bi_max_vecs = 1;
3004                         bi->bi_idx = 0;
3005                         bi->bi_io_vec = &sh->dev[i].vec;
3006                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
3007                         bi->bi_io_vec[0].bv_offset = 0;
3008                         bi->bi_size = STRIPE_SIZE;
3009                         bi->bi_next = NULL;
3010                         if (rw == WRITE &&
3011                             test_bit(R5_ReWrite, &sh->dev[i].flags))
3012                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
3013                         generic_make_request(bi);
3014                 } else {
3015                         if (rw == WRITE)
3016                                 set_bit(STRIPE_DEGRADED, &sh->state);
3017                         pr_debug("skip op %ld on disc %d for sector %llu\n",
3018                                 bi->bi_rw, i, (unsigned long long)sh->sector);
3019                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
3020                         set_bit(STRIPE_HANDLE, &sh->state);
3021                 }
3022         }
3023 }
3024
3025 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
3026 {
3027         raid6_conf_t *conf = sh->raid_conf;
3028         int disks = sh->disks;
3029         struct bio *return_bi = NULL;
3030         int i, pd_idx = sh->pd_idx;
3031         struct stripe_head_state s;
3032         struct r6_state r6s;
3033         struct r5dev *dev, *pdev, *qdev;
3034
3035         r6s.qd_idx = raid6_next_disk(pd_idx, disks);
3036         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3037                 "pd_idx=%d, qd_idx=%d\n",
3038                (unsigned long long)sh->sector, sh->state,
3039                atomic_read(&sh->count), pd_idx, r6s.qd_idx);
3040         memset(&s, 0, sizeof(s));
3041
3042         spin_lock(&sh->lock);
3043         clear_bit(STRIPE_HANDLE, &sh->state);
3044         clear_bit(STRIPE_DELAYED, &sh->state);
3045
3046         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3047         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3048         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3049         /* Now to look around and see what can be done */
3050
3051         rcu_read_lock();
3052         for (i=disks; i--; ) {
3053                 mdk_rdev_t *rdev;
3054                 dev = &sh->dev[i];
3055                 clear_bit(R5_Insync, &dev->flags);
3056
3057                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3058                         i, dev->flags, dev->toread, dev->towrite, dev->written);
3059                 /* maybe we can reply to a read */
3060                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
3061                         struct bio *rbi, *rbi2;
3062                         pr_debug("Return read for disc %d\n", i);
3063                         spin_lock_irq(&conf->device_lock);
3064                         rbi = dev->toread;
3065                         dev->toread = NULL;
3066                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
3067                                 wake_up(&conf->wait_for_overlap);
3068                         spin_unlock_irq(&conf->device_lock);
3069                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
3070                                 copy_data(0, rbi, dev->page, dev->sector);
3071                                 rbi2 = r5_next_bio(rbi, dev->sector);
3072                                 spin_lock_irq(&conf->device_lock);
3073                                 if (--rbi->bi_phys_segments == 0) {
3074                                         rbi->bi_next = return_bi;
3075                                         return_bi = rbi;
3076                                 }
3077                                 spin_unlock_irq(&conf->device_lock);
3078                                 rbi = rbi2;
3079                         }
3080                 }
3081
3082                 /* now count some things */
3083                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3084                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3085
3086
3087                 if (dev->toread)
3088                         s.to_read++;
3089                 if (dev->towrite) {
3090                         s.to_write++;
3091                         if (!test_bit(R5_OVERWRITE, &dev->flags))
3092                                 s.non_overwrite++;
3093                 }
3094                 if (dev->written)
3095                         s.written++;
3096                 rdev = rcu_dereference(conf->disks[i].rdev);
3097                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3098                         /* The ReadError flag will just be confusing now */
3099                         clear_bit(R5_ReadError, &dev->flags);
3100                         clear_bit(R5_ReWrite, &dev->flags);
3101                 }
3102                 if (!rdev || !test_bit(In_sync, &rdev->flags)
3103                     || test_bit(R5_ReadError, &dev->flags)) {
3104                         if (s.failed < 2)
3105                                 r6s.failed_num[s.failed] = i;
3106                         s.failed++;
3107                 } else
3108                         set_bit(R5_Insync, &dev->flags);
3109         }
3110         rcu_read_unlock();
3111         pr_debug("locked=%d uptodate=%d to_read=%d"
3112                " to_write=%d failed=%d failed_num=%d,%d\n",
3113                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3114                r6s.failed_num[0], r6s.failed_num[1]);
3115         /* check if the array has lost >2 devices and, if so, some requests
3116          * might need to be failed
3117          */
3118         if (s.failed > 2 && s.to_read+s.to_write+s.written)
3119                 handle_requests_to_failed_array(conf, sh, &s, disks,
3120                                                 &return_bi);
3121         if (s.failed > 2 && s.syncing) {
3122                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3123                 clear_bit(STRIPE_SYNCING, &sh->state);
3124                 s.syncing = 0;
3125         }
3126
3127         /*
3128          * might be able to return some write requests if the parity blocks
3129          * are safe, or on a failed drive
3130          */
3131         pdev = &sh->dev[pd_idx];
3132         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3133                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3134         qdev = &sh->dev[r6s.qd_idx];
3135         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
3136                 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
3137
3138         if ( s.written &&
3139              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3140                              && !test_bit(R5_LOCKED, &pdev->flags)
3141                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3142              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3143                              && !test_bit(R5_LOCKED, &qdev->flags)
3144                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3145                 handle_completed_write_requests(conf, sh, disks, &return_bi);
3146
3147         /* Now we might consider reading some blocks, either to check/generate
3148          * parity, or to satisfy requests
3149          * or to load a block that is being partially written.
3150          */
3151         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3152             (s.syncing && (s.uptodate < disks)) || s.expanding)
3153                 handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
3154
3155         /* now to consider writing and what else, if anything should be read */
3156         if (s.to_write)
3157                 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
3158
3159         /* maybe we need to check and possibly fix the parity for this stripe
3160          * Any reads will already have been scheduled, so we just see if enough
3161          * data is available
3162          */
3163         if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3164                 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3165
3166         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3167                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3168                 clear_bit(STRIPE_SYNCING, &sh->state);
3169         }
3170
3171         /* If the failed drives are just a ReadError, then we might need
3172          * to progress the repair/check process
3173          */
3174         if (s.failed <= 2 && !conf->mddev->ro)
3175                 for (i = 0; i < s.failed; i++) {
3176                         dev = &sh->dev[r6s.failed_num[i]];
3177                         if (test_bit(R5_ReadError, &dev->flags)
3178                             && !test_bit(R5_LOCKED, &dev->flags)
3179                             && test_bit(R5_UPTODATE, &dev->flags)
3180                                 ) {
3181                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3182                                         set_bit(R5_Wantwrite, &dev->flags);
3183                                         set_bit(R5_ReWrite, &dev->flags);
3184                                         set_bit(R5_LOCKED, &dev->flags);
3185                                 } else {
3186                                         /* let's read it back */
3187                                         set_bit(R5_Wantread, &dev->flags);
3188                                         set_bit(R5_LOCKED, &dev->flags);
3189                                 }
3190                         }
3191                 }
3192
3193         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3194                 /* Need to write out all blocks after computing P&Q */
3195                 sh->disks = conf->raid_disks;
3196                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
3197                                              conf->raid_disks);
3198                 compute_parity6(sh, RECONSTRUCT_WRITE);
3199                 for (i = conf->raid_disks ; i-- ;  ) {
3200                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3201                         s.locked++;
3202                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3203                 }
3204                 clear_bit(STRIPE_EXPANDING, &sh->state);
3205         } else if (s.expanded) {
3206                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3207                 atomic_dec(&conf->reshape_stripes);
3208                 wake_up(&conf->wait_for_overlap);
3209                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3210         }
3211
3212         if (s.expanding && s.locked == 0)
3213                 handle_stripe_expansion(conf, sh, &r6s);
3214
3215         spin_unlock(&sh->lock);
3216
3217         return_io(return_bi);
3218
3219         for (i=disks; i-- ;) {
3220                 int rw;
3221                 struct bio *bi;
3222                 mdk_rdev_t *rdev;
3223                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
3224                         rw = WRITE;
3225                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
3226                         rw = READ;
3227                 else
3228                         continue;
3229
3230                 bi = &sh->dev[i].req;
3231
3232                 bi->bi_rw = rw;
3233                 if (rw == WRITE)
3234                         bi->bi_end_io = raid5_end_write_request;
3235                 else
3236                         bi->bi_end_io = raid5_end_read_request;
3237
3238                 rcu_read_lock();
3239                 rdev = rcu_dereference(conf->disks[i].rdev);
3240                 if (rdev && test_bit(Faulty, &rdev->flags))
3241                         rdev = NULL;
3242                 if (rdev)
3243                         atomic_inc(&rdev->nr_pending);
3244                 rcu_read_unlock();
3245
3246                 if (rdev) {
3247                         if (s.syncing || s.expanding || s.expanded)
3248                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
3249
3250                         bi->bi_bdev = rdev->bdev;
3251                         pr_debug("for %llu schedule op %ld on disc %d\n",
3252                                 (unsigned long long)sh->sector, bi->bi_rw, i);
3253                         atomic_inc(&sh->count);
3254                         bi->bi_sector = sh->sector + rdev->data_offset;
3255                         bi->bi_flags = 1 << BIO_UPTODATE;
3256                         bi->bi_vcnt = 1;
3257                         bi->bi_max_vecs = 1;
3258                         bi->bi_idx = 0;
3259                         bi->bi_io_vec = &sh->dev[i].vec;
3260                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
3261                         bi->bi_io_vec[0].bv_offset = 0;
3262                         bi->bi_size = STRIPE_SIZE;
3263                         bi->bi_next = NULL;
3264                         if (rw == WRITE &&
3265                             test_bit(R5_ReWrite, &sh->dev[i].flags))
3266                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
3267                         generic_make_request(bi);
3268                 } else {
3269                         if (rw == WRITE)
3270                                 set_bit(STRIPE_DEGRADED, &sh->state);
3271                         pr_debug("skip op %ld on disc %d for sector %llu\n",
3272                                 bi->bi_rw, i, (unsigned long long)sh->sector);
3273                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
3274                         set_bit(STRIPE_HANDLE, &sh->state);
3275                 }
3276         }
3277 }
3278
3279 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3280 {
3281         if (sh->raid_conf->level == 6)
3282                 handle_stripe6(sh, tmp_page);
3283         else
3284                 handle_stripe5(sh);
3285 }
3286
3287
3288
3289 static void raid5_activate_delayed(raid5_conf_t *conf)
3290 {
3291         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3292                 while (!list_empty(&conf->delayed_list)) {
3293                         struct list_head *l = conf->delayed_list.next;
3294                         struct stripe_head *sh;
3295                         sh = list_entry(l, struct stripe_head, lru);
3296                         list_del_init(l);
3297                         clear_bit(STRIPE_DELAYED, &sh->state);
3298                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3299                                 atomic_inc(&conf->preread_active_stripes);
3300                         list_add_tail(&sh->lru, &conf->handle_list);
3301                 }
3302         }
3303 }
3304
3305 static void activate_bit_delay(raid5_conf_t *conf)
3306 {
3307         /* device_lock is held */
3308         struct list_head head;
3309         list_add(&head, &conf->bitmap_list);
3310         list_del_init(&conf->bitmap_list);
3311         while (!list_empty(&head)) {
3312                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3313                 list_del_init(&sh->lru);
3314                 atomic_inc(&sh->count);
3315                 __release_stripe(conf, sh);
3316         }
3317 }
3318
3319 static void unplug_slaves(mddev_t *mddev)
3320 {
3321         raid5_conf_t *conf = mddev_to_conf(mddev);
3322         int i;
3323
3324         rcu_read_lock();
3325         for (i=0; i<mddev->raid_disks; i++) {
3326                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3327                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3328                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
3329
3330                         atomic_inc(&rdev->nr_pending);
3331                         rcu_read_unlock();
3332
3333                         if (r_queue->unplug_fn)
3334                                 r_queue->unplug_fn(r_queue);
3335
3336                         rdev_dec_pending(rdev, mddev);
3337                         rcu_read_lock();
3338                 }
3339         }
3340         rcu_read_unlock();
3341 }
3342
3343 static void raid5_unplug_device(request_queue_t *q)
3344 {
3345         mddev_t *mddev = q->queuedata;
3346         raid5_conf_t *conf = mddev_to_conf(mddev);
3347         unsigned long flags;
3348
3349         spin_lock_irqsave(&conf->device_lock, flags);
3350
3351         if (blk_remove_plug(q)) {
3352                 conf->seq_flush++;
3353                 raid5_activate_delayed(conf);
3354         }
3355         md_wakeup_thread(mddev->thread);
3356
3357         spin_unlock_irqrestore(&conf->device_lock, flags);
3358
3359         unplug_slaves(mddev);
3360 }
3361
3362 static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
3363                              sector_t *error_sector)
3364 {
3365         mddev_t *mddev = q->queuedata;
3366         raid5_conf_t *conf = mddev_to_conf(mddev);
3367         int i, ret = 0;
3368
3369         rcu_read_lock();
3370         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
3371                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3372                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
3373                         struct block_device *bdev = rdev->bdev;
3374                         request_queue_t *r_queue = bdev_get_queue(bdev);
3375
3376                         if (!r_queue->issue_flush_fn)
3377                                 ret = -EOPNOTSUPP;
3378                         else {
3379                                 atomic_inc(&rdev->nr_pending);
3380                                 rcu_read_unlock();
3381                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
3382                                                               error_sector);
3383                                 rdev_dec_pending(rdev, mddev);
3384                                 rcu_read_lock();
3385                         }
3386                 }
3387         }
3388         rcu_read_unlock();
3389         return ret;
3390 }
3391
3392 static int raid5_congested(void *data, int bits)
3393 {
3394         mddev_t *mddev = data;
3395         raid5_conf_t *conf = mddev_to_conf(mddev);
3396
3397         /* No difference between reads and writes.  Just check
3398          * how busy the stripe_cache is
3399          */
3400         if (conf->inactive_blocked)
3401                 return 1;
3402         if (conf->quiesce)
3403                 return 1;
3404         if (list_empty_careful(&conf->inactive_list))
3405                 return 1;
3406
3407         return 0;
3408 }
3409
3410 /* We want read requests to align with chunks where possible,
3411  * but write requests don't need to.
3412  */
3413 static int raid5_mergeable_bvec(request_queue_t *q, struct bio *bio, struct bio_vec *biovec)
3414 {
3415         mddev_t *mddev = q->queuedata;
3416         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3417         int max;
3418         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3419         unsigned int bio_sectors = bio->bi_size >> 9;
3420
3421         if (bio_data_dir(bio) == WRITE)
3422                 return biovec->bv_len; /* always allow writes to be mergeable */
3423
3424         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3425         if (max < 0) max = 0;
3426         if (max <= biovec->bv_len && bio_sectors == 0)
3427                 return biovec->bv_len;
3428         else
3429                 return max;
3430 }
3431
3432
3433 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3434 {
3435         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3436         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3437         unsigned int bio_sectors = bio->bi_size >> 9;
3438
3439         return  chunk_sectors >=
3440                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3441 }
3442
3443 /*
3444  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3445  *  later sampled by raid5d.
3446  */
3447 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3448 {
3449         unsigned long flags;
3450
3451         spin_lock_irqsave(&conf->device_lock, flags);
3452
3453         bi->bi_next = conf->retry_read_aligned_list;
3454         conf->retry_read_aligned_list = bi;
3455
3456         spin_unlock_irqrestore(&conf->device_lock, flags);
3457         md_wakeup_thread(conf->mddev->thread);
3458 }
3459
3460
3461 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3462 {
3463         struct bio *bi;
3464
3465         bi = conf->retry_read_aligned;
3466         if (bi) {
3467                 conf->retry_read_aligned = NULL;
3468                 return bi;
3469         }
3470         bi = conf->retry_read_aligned_list;
3471         if(bi) {
3472                 conf->retry_read_aligned_list = bi->bi_next;
3473                 bi->bi_next = NULL;
3474                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3475                 bi->bi_hw_segments = 0; /* count of processed stripes */
3476         }
3477
3478         return bi;
3479 }
3480
3481
3482 /*
3483  *  The "raid5_align_endio" should check if the read succeeded and if it
3484  *  did, call bio_endio on the original bio (having bio_put the new bio
3485  *  first).
3486  *  If the read failed..
3487  */
3488 static int raid5_align_endio(struct bio *bi, unsigned int bytes, int error)
3489 {
3490         struct bio* raid_bi  = bi->bi_private;
3491         mddev_t *mddev;
3492         raid5_conf_t *conf;
3493         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3494         mdk_rdev_t *rdev;
3495
3496         if (bi->bi_size)
3497                 return 1;
3498         bio_put(bi);
3499
3500         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3501         conf = mddev_to_conf(mddev);
3502         rdev = (void*)raid_bi->bi_next;
3503         raid_bi->bi_next = NULL;
3504
3505         rdev_dec_pending(rdev, conf->mddev);
3506
3507         if (!error && uptodate) {
3508                 bio_endio(raid_bi, bytes, 0);
3509                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3510                         wake_up(&conf->wait_for_stripe);
3511                 return 0;
3512         }
3513
3514
3515         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3516
3517         add_bio_to_retry(raid_bi, conf);
3518         return 0;
3519 }
3520
3521 static int bio_fits_rdev(struct bio *bi)
3522 {
3523         request_queue_t *q = bdev_get_queue(bi->bi_bdev);
3524
3525         if ((bi->bi_size>>9) > q->max_sectors)
3526                 return 0;
3527         blk_recount_segments(q, bi);
3528         if (bi->bi_phys_segments > q->max_phys_segments ||
3529             bi->bi_hw_segments > q->max_hw_segments)
3530                 return 0;
3531
3532         if (q->merge_bvec_fn)
3533                 /* it's too hard to apply the merge_bvec_fn at this stage,
3534                  * just just give up
3535                  */
3536                 return 0;
3537
3538         return 1;
3539 }
3540
3541
3542 static int chunk_aligned_read(request_queue_t *q, struct bio * raid_bio)
3543 {
3544         mddev_t *mddev = q->queuedata;
3545         raid5_conf_t *conf = mddev_to_conf(mddev);
3546         const unsigned int raid_disks = conf->raid_disks;
3547         const unsigned int data_disks = raid_disks - conf->max_degraded;
3548         unsigned int dd_idx, pd_idx;
3549         struct bio* align_bi;
3550         mdk_rdev_t *rdev;
3551
3552         if (!in_chunk_boundary(mddev, raid_bio)) {
3553                 pr_debug("chunk_aligned_read : non aligned\n");
3554                 return 0;
3555         }
3556         /*
3557          * use bio_clone to make a copy of the bio
3558          */
3559         align_bi = bio_clone(raid_bio, GFP_NOIO);
3560         if (!align_bi)
3561                 return 0;
3562         /*
3563          *   set bi_end_io to a new function, and set bi_private to the
3564          *     original bio.
3565          */
3566         align_bi->bi_end_io  = raid5_align_endio;
3567         align_bi->bi_private = raid_bio;
3568         /*
3569          *      compute position
3570          */
3571         align_bi->bi_sector =  raid5_compute_sector(raid_bio->bi_sector,
3572                                         raid_disks,
3573                                         data_disks,
3574                                         &dd_idx,
3575                                         &pd_idx,
3576                                         conf);
3577
3578         rcu_read_lock();
3579         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3580         if (rdev && test_bit(In_sync, &rdev->flags)) {
3581                 atomic_inc(&rdev->nr_pending);
3582                 rcu_read_unlock();
3583                 raid_bio->bi_next = (void*)rdev;
3584                 align_bi->bi_bdev =  rdev->bdev;
3585                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3586                 align_bi->bi_sector += rdev->data_offset;
3587
3588                 if (!bio_fits_rdev(align_bi)) {
3589                         /* too big in some way */
3590                         bio_put(align_bi);
3591                         rdev_dec_pending(rdev, mddev);
3592                         return 0;
3593                 }
3594
3595                 spin_lock_irq(&conf->device_lock);
3596                 wait_event_lock_irq(conf->wait_for_stripe,
3597                                     conf->quiesce == 0,
3598                                     conf->device_lock, /* nothing */);
3599                 atomic_inc(&conf->active_aligned_reads);
3600                 spin_unlock_irq(&conf->device_lock);
3601
3602                 generic_make_request(align_bi);
3603                 return 1;
3604         } else {
3605                 rcu_read_unlock();
3606                 bio_put(align_bi);
3607                 return 0;
3608         }
3609 }
3610
3611
3612 static int make_request(request_queue_t *q, struct bio * bi)
3613 {
3614         mddev_t *mddev = q->queuedata;
3615         raid5_conf_t *conf = mddev_to_conf(mddev);
3616         unsigned int dd_idx, pd_idx;
3617         sector_t new_sector;
3618         sector_t logical_sector, last_sector;
3619         struct stripe_head *sh;
3620         const int rw = bio_data_dir(bi);
3621         int remaining;
3622
3623         if (unlikely(bio_barrier(bi))) {
3624                 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
3625                 return 0;
3626         }
3627
3628         md_write_start(mddev, bi);
3629
3630         disk_stat_inc(mddev->gendisk, ios[rw]);
3631         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3632
3633         if (rw == READ &&
3634              mddev->reshape_position == MaxSector &&
3635              chunk_aligned_read(q,bi))
3636                 return 0;
3637
3638         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3639         last_sector = bi->bi_sector + (bi->bi_size>>9);
3640         bi->bi_next = NULL;
3641         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3642
3643         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3644                 DEFINE_WAIT(w);
3645                 int disks, data_disks;
3646
3647         retry:
3648                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3649                 if (likely(conf->expand_progress == MaxSector))
3650                         disks = conf->raid_disks;
3651                 else {
3652                         /* spinlock is needed as expand_progress may be
3653                          * 64bit on a 32bit platform, and so it might be
3654                          * possible to see a half-updated value
3655                          * Ofcourse expand_progress could change after
3656                          * the lock is dropped, so once we get a reference
3657                          * to the stripe that we think it is, we will have
3658                          * to check again.
3659                          */
3660                         spin_lock_irq(&conf->device_lock);
3661                         disks = conf->raid_disks;
3662                         if (logical_sector >= conf->expand_progress)
3663                                 disks = conf->previous_raid_disks;
3664                         else {
3665                                 if (logical_sector >= conf->expand_lo) {
3666                                         spin_unlock_irq(&conf->device_lock);
3667                                         schedule();
3668                                         goto retry;
3669                                 }
3670                         }
3671                         spin_unlock_irq(&conf->device_lock);
3672                 }
3673                 data_disks = disks - conf->max_degraded;
3674
3675                 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3676                                                   &dd_idx, &pd_idx, conf);
3677                 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3678                         (unsigned long long)new_sector, 
3679                         (unsigned long long)logical_sector);
3680
3681                 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3682                 if (sh) {
3683                         if (unlikely(conf->expand_progress != MaxSector)) {
3684                                 /* expansion might have moved on while waiting for a
3685                                  * stripe, so we must do the range check again.
3686                                  * Expansion could still move past after this
3687                                  * test, but as we are holding a reference to
3688                                  * 'sh', we know that if that happens,
3689                                  *  STRIPE_EXPANDING will get set and the expansion
3690                                  * won't proceed until we finish with the stripe.
3691                                  */
3692                                 int must_retry = 0;
3693                                 spin_lock_irq(&conf->device_lock);
3694                                 if (logical_sector <  conf->expand_progress &&
3695                                     disks == conf->previous_raid_disks)
3696                                         /* mismatch, need to try again */
3697                                         must_retry = 1;
3698                                 spin_unlock_irq(&conf->device_lock);
3699                                 if (must_retry) {
3700                                         release_stripe(sh);
3701                                         goto retry;
3702                                 }
3703                         }
3704                         /* FIXME what if we get a false positive because these
3705                          * are being updated.
3706                          */
3707                         if (logical_sector >= mddev->suspend_lo &&
3708                             logical_sector < mddev->suspend_hi) {
3709                                 release_stripe(sh);
3710                                 schedule();
3711                                 goto retry;
3712                         }
3713
3714                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3715                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3716                                 /* Stripe is busy expanding or
3717                                  * add failed due to overlap.  Flush everything
3718                                  * and wait a while
3719                                  */
3720                                 raid5_unplug_device(mddev->queue);
3721                                 release_stripe(sh);
3722                                 schedule();
3723                                 goto retry;
3724                         }
3725                         finish_wait(&conf->wait_for_overlap, &w);
3726                         handle_stripe(sh, NULL);
3727                         release_stripe(sh);
3728                 } else {
3729                         /* cannot get stripe for read-ahead, just give-up */
3730                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3731                         finish_wait(&conf->wait_for_overlap, &w);
3732                         break;
3733                 }
3734                         
3735         }
3736         spin_lock_irq(&conf->device_lock);
3737         remaining = --bi->bi_phys_segments;
3738         spin_unlock_irq(&conf->device_lock);
3739         if (remaining == 0) {
3740                 int bytes = bi->bi_size;
3741
3742                 if ( rw == WRITE )
3743                         md_write_end(mddev);
3744                 bi->bi_size = 0;
3745                 bi->bi_end_io(bi, bytes,
3746                               test_bit(BIO_UPTODATE, &bi->bi_flags)
3747                                 ? 0 : -EIO);
3748         }
3749         return 0;
3750 }
3751
3752 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3753 {
3754         /* reshaping is quite different to recovery/resync so it is
3755          * handled quite separately ... here.
3756          *
3757          * On each call to sync_request, we gather one chunk worth of
3758          * destination stripes and flag them as expanding.
3759          * Then we find all the source stripes and request reads.
3760          * As the reads complete, handle_stripe will copy the data
3761          * into the destination stripe and release that stripe.
3762          */
3763         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3764         struct stripe_head *sh;
3765         int pd_idx;
3766         sector_t first_sector, last_sector;
3767         int raid_disks = conf->previous_raid_disks;
3768         int data_disks = raid_disks - conf->max_degraded;
3769         int new_data_disks = conf->raid_disks - conf->max_degraded;
3770         int i;
3771         int dd_idx;
3772         sector_t writepos, safepos, gap;
3773
3774         if (sector_nr == 0 &&
3775             conf->expand_progress != 0) {
3776                 /* restarting in the middle, skip the initial sectors */
3777                 sector_nr = conf->expand_progress;
3778                 sector_div(sector_nr, new_data_disks);
3779                 *skipped = 1;
3780                 return sector_nr;
3781         }
3782
3783         /* we update the metadata when there is more than 3Meg
3784          * in the block range (that is rather arbitrary, should
3785          * probably be time based) or when the data about to be
3786          * copied would over-write the source of the data at
3787          * the front of the range.
3788          * i.e. one new_stripe forward from expand_progress new_maps
3789          * to after where expand_lo old_maps to
3790          */
3791         writepos = conf->expand_progress +
3792                 conf->chunk_size/512*(new_data_disks);
3793         sector_div(writepos, new_data_disks);
3794         safepos = conf->expand_lo;
3795         sector_div(safepos, data_disks);
3796         gap = conf->expand_progress - conf->expand_lo;
3797
3798         if (writepos >= safepos ||
3799             gap > (new_data_disks)*3000*2 /*3Meg*/) {
3800                 /* Cannot proceed until we've updated the superblock... */
3801                 wait_event(conf->wait_for_overlap,
3802                            atomic_read(&conf->reshape_stripes)==0);
3803                 mddev->reshape_position = conf->expand_progress;
3804                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3805                 md_wakeup_thread(mddev->thread);
3806                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3807                            kthread_should_stop());
3808                 spin_lock_irq(&conf->device_lock);
3809                 conf->expand_lo = mddev->reshape_position;
3810                 spin_unlock_irq(&conf->device_lock);
3811                 wake_up(&conf->wait_for_overlap);
3812         }
3813
3814         for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3815                 int j;
3816                 int skipped = 0;
3817                 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3818                 sh = get_active_stripe(conf, sector_nr+i,
3819                                        conf->raid_disks, pd_idx, 0);
3820                 set_bit(STRIPE_EXPANDING, &sh->state);
3821                 atomic_inc(&conf->reshape_stripes);
3822                 /* If any of this stripe is beyond the end of the old
3823                  * array, then we need to zero those blocks
3824                  */
3825                 for (j=sh->disks; j--;) {
3826                         sector_t s;
3827                         if (j == sh->pd_idx)
3828                                 continue;
3829                         if (conf->level == 6 &&
3830                             j == raid6_next_disk(sh->pd_idx, sh->disks))
3831                                 continue;
3832                         s = compute_blocknr(sh, j);
3833                         if (s < (mddev->array_size<<1)) {
3834                                 skipped = 1;
3835                                 continue;
3836                         }
3837                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3838                         set_bit(R5_Expanded, &sh->dev[j].flags);
3839                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
3840                 }
3841                 if (!skipped) {
3842                         set_bit(STRIPE_EXPAND_READY, &sh->state);
3843                         set_bit(STRIPE_HANDLE, &sh->state);
3844                 }
3845                 release_stripe(sh);
3846         }
3847         spin_lock_irq(&conf->device_lock);
3848         conf->expand_progress = (sector_nr + i) * new_data_disks;
3849         spin_unlock_irq(&conf->device_lock);
3850         /* Ok, those stripe are ready. We can start scheduling
3851          * reads on the source stripes.
3852          * The source stripes are determined by mapping the first and last
3853          * block on the destination stripes.
3854          */
3855         first_sector =
3856                 raid5_compute_sector(sector_nr*(new_data_disks),
3857                                      raid_disks, data_disks,
3858                                      &dd_idx, &pd_idx, conf);
3859         last_sector =
3860                 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3861                                      *(new_data_disks) -1,
3862                                      raid_disks, data_disks,
3863                                      &dd_idx, &pd_idx, conf);
3864         if (last_sector >= (mddev->size<<1))
3865                 last_sector = (mddev->size<<1)-1;
3866         while (first_sector <= last_sector) {
3867                 pd_idx = stripe_to_pdidx(first_sector, conf,
3868                                          conf->previous_raid_disks);
3869                 sh = get_active_stripe(conf, first_sector,
3870                                        conf->previous_raid_disks, pd_idx, 0);
3871                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3872                 set_bit(STRIPE_HANDLE, &sh->state);
3873                 release_stripe(sh);
3874                 first_sector += STRIPE_SECTORS;
3875         }
3876         return conf->chunk_size>>9;
3877 }
3878
3879 /* FIXME go_faster isn't used */
3880 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3881 {
3882         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3883         struct stripe_head *sh;
3884         int pd_idx;
3885         int raid_disks = conf->raid_disks;
3886         sector_t max_sector = mddev->size << 1;
3887         int sync_blocks;
3888         int still_degraded = 0;
3889         int i;
3890
3891         if (sector_nr >= max_sector) {
3892                 /* just being told to finish up .. nothing much to do */
3893                 unplug_slaves(mddev);
3894                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3895                         end_reshape(conf);
3896                         return 0;
3897                 }
3898
3899                 if (mddev->curr_resync < max_sector) /* aborted */
3900                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3901                                         &sync_blocks, 1);
3902                 else /* completed sync */
3903                         conf->fullsync = 0;
3904                 bitmap_close_sync(mddev->bitmap);
3905
3906                 return 0;
3907         }
3908
3909         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3910                 return reshape_request(mddev, sector_nr, skipped);
3911
3912         /* if there is too many failed drives and we are trying
3913          * to resync, then assert that we are finished, because there is
3914          * nothing we can do.
3915          */
3916         if (mddev->degraded >= conf->max_degraded &&
3917             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3918                 sector_t rv = (mddev->size << 1) - sector_nr;
3919                 *skipped = 1;
3920                 return rv;
3921         }
3922         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3923             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3924             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3925                 /* we can skip this block, and probably more */
3926                 sync_blocks /= STRIPE_SECTORS;
3927                 *skipped = 1;
3928                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3929         }
3930
3931         pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3932         sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3933         if (sh == NULL) {
3934                 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3935                 /* make sure we don't swamp the stripe cache if someone else
3936                  * is trying to get access
3937                  */
3938                 schedule_timeout_uninterruptible(1);
3939         }
3940         /* Need to check if array will still be degraded after recovery/resync
3941          * We don't need to check the 'failed' flag as when that gets set,
3942          * recovery aborts.
3943          */
3944         for (i=0; i<mddev->raid_disks; i++)
3945                 if (conf->disks[i].rdev == NULL)
3946                         still_degraded = 1;
3947
3948         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3949
3950         spin_lock(&sh->lock);
3951         set_bit(STRIPE_SYNCING, &sh->state);
3952         clear_bit(STRIPE_INSYNC, &sh->state);
3953         spin_unlock(&sh->lock);
3954
3955         handle_stripe(sh, NULL);
3956         release_stripe(sh);
3957
3958         return STRIPE_SECTORS;
3959 }
3960
3961 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3962 {
3963         /* We may not be able to submit a whole bio at once as there
3964          * may not be enough stripe_heads available.
3965          * We cannot pre-allocate enough stripe_heads as we may need
3966          * more than exist in the cache (if we allow ever large chunks).
3967          * So we do one stripe head at a time and record in
3968          * ->bi_hw_segments how many have been done.
3969          *
3970          * We *know* that this entire raid_bio is in one chunk, so
3971          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3972          */
3973         struct stripe_head *sh;
3974         int dd_idx, pd_idx;
3975         sector_t sector, logical_sector, last_sector;
3976         int scnt = 0;
3977         int remaining;
3978         int handled = 0;
3979
3980         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3981         sector = raid5_compute_sector(  logical_sector,
3982                                         conf->raid_disks,
3983                                         conf->raid_disks - conf->max_degraded,
3984                                         &dd_idx,
3985                                         &pd_idx,
3986                                         conf);
3987         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3988
3989         for (; logical_sector < last_sector;
3990              logical_sector += STRIPE_SECTORS,
3991                      sector += STRIPE_SECTORS,
3992                      scnt++) {
3993
3994                 if (scnt < raid_bio->bi_hw_segments)
3995                         /* already done this stripe */
3996                         continue;
3997
3998                 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3999
4000                 if (!sh) {
4001                         /* failed to get a stripe - must wait */
4002                         raid_bio->bi_hw_segments = scnt;
4003                         conf->retry_read_aligned = raid_bio;
4004                         return handled;
4005                 }
4006
4007                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4008                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4009                         release_stripe(sh);
4010                         raid_bio->bi_hw_segments = scnt;
4011                         conf->retry_read_aligned = raid_bio;
4012                         return handled;
4013                 }
4014
4015                 handle_stripe(sh, NULL);
4016                 release_stripe(sh);
4017                 handled++;
4018         }
4019         spin_lock_irq(&conf->device_lock);
4020         remaining = --raid_bio->bi_phys_segments;
4021         spin_unlock_irq(&conf->device_lock);
4022         if (remaining == 0) {
4023                 int bytes = raid_bio->bi_size;
4024
4025                 raid_bio->bi_size = 0;
4026                 raid_bio->bi_end_io(raid_bio, bytes,
4027                               test_bit(BIO_UPTODATE, &raid_bio->bi_flags)
4028                                 ? 0 : -EIO);
4029         }
4030         if (atomic_dec_and_test(&conf->active_aligned_reads))
4031                 wake_up(&conf->wait_for_stripe);
4032         return handled;
4033 }
4034
4035
4036
4037 /*
4038  * This is our raid5 kernel thread.
4039  *
4040  * We scan the hash table for stripes which can be handled now.
4041  * During the scan, completed stripes are saved for us by the interrupt
4042  * handler, so that they will not have to wait for our next wakeup.
4043  */
4044 static void raid5d (mddev_t *mddev)
4045 {
4046         struct stripe_head *sh;
4047         raid5_conf_t *conf = mddev_to_conf(mddev);
4048         int handled;
4049
4050         pr_debug("+++ raid5d active\n");
4051
4052         md_check_recovery(mddev);
4053
4054         handled = 0;
4055         spin_lock_irq(&conf->device_lock);
4056         while (1) {
4057                 struct list_head *first;
4058                 struct bio *bio;
4059
4060                 if (conf->seq_flush != conf->seq_write) {
4061                         int seq = conf->seq_flush;
4062                         spin_unlock_irq(&conf->device_lock);
4063                         bitmap_unplug(mddev->bitmap);
4064                         spin_lock_irq(&conf->device_lock);
4065                         conf->seq_write = seq;
4066                         activate_bit_delay(conf);
4067                 }
4068
4069                 if (list_empty(&conf->handle_list) &&
4070                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
4071                     !blk_queue_plugged(mddev->queue) &&
4072                     !list_empty(&conf->delayed_list))
4073                         raid5_activate_delayed(conf);
4074
4075                 while ((bio = remove_bio_from_retry(conf))) {
4076                         int ok;
4077                         spin_unlock_irq(&conf->device_lock);
4078                         ok = retry_aligned_read(conf, bio);
4079                         spin_lock_irq(&conf->device_lock);
4080                         if (!ok)
4081                                 break;
4082                         handled++;
4083                 }
4084
4085                 if (list_empty(&conf->handle_list)) {
4086                         async_tx_issue_pending_all();
4087                         break;
4088                 }
4089
4090                 first = conf->handle_list.next;
4091                 sh = list_entry(first, struct stripe_head, lru);
4092
4093                 list_del_init(first);
4094                 atomic_inc(&sh->count);
4095                 BUG_ON(atomic_read(&sh->count)!= 1);
4096                 spin_unlock_irq(&conf->device_lock);
4097                 
4098                 handled++;
4099                 handle_stripe(sh, conf->spare_page);
4100                 release_stripe(sh);
4101
4102                 spin_lock_irq(&conf->device_lock);
4103         }
4104         pr_debug("%d stripes handled\n", handled);
4105
4106         spin_unlock_irq(&conf->device_lock);
4107
4108         unplug_slaves(mddev);
4109
4110         pr_debug("--- raid5d inactive\n");
4111 }
4112
4113 static ssize_t
4114 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4115 {
4116         raid5_conf_t *conf = mddev_to_conf(mddev);
4117         if (conf)
4118                 return sprintf(page, "%d\n", conf->max_nr_stripes);
4119         else
4120                 return 0;
4121 }
4122
4123 static ssize_t
4124 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4125 {
4126         raid5_conf_t *conf = mddev_to_conf(mddev);
4127         char *end;
4128         int new;
4129         if (len >= PAGE_SIZE)
4130                 return -EINVAL;
4131         if (!conf)
4132                 return -ENODEV;
4133
4134         new = simple_strtoul(page, &end, 10);
4135         if (!*page || (*end && *end != '\n') )
4136                 return -EINVAL;
4137         if (new <= 16 || new > 32768)
4138                 return -EINVAL;
4139         while (new < conf->max_nr_stripes) {
4140                 if (drop_one_stripe(conf))
4141                         conf->max_nr_stripes--;
4142                 else
4143                         break;
4144         }
4145         md_allow_write(mddev);
4146         while (new > conf->max_nr_stripes) {
4147                 if (grow_one_stripe(conf))
4148                         conf->max_nr_stripes++;
4149                 else break;
4150         }
4151         return len;
4152 }
4153
4154 static struct md_sysfs_entry
4155 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4156                                 raid5_show_stripe_cache_size,
4157                                 raid5_store_stripe_cache_size);
4158
4159 static ssize_t
4160 stripe_cache_active_show(mddev_t *mddev, char *page)
4161 {
4162         raid5_conf_t *conf = mddev_to_conf(mddev);
4163         if (conf)
4164                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4165         else
4166                 return 0;
4167 }
4168
4169 static struct md_sysfs_entry
4170 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4171
4172 static struct attribute *raid5_attrs[] =  {
4173         &raid5_stripecache_size.attr,
4174         &raid5_stripecache_active.attr,
4175         NULL,
4176 };
4177 static struct attribute_group raid5_attrs_group = {
4178         .name = NULL,
4179         .attrs = raid5_attrs,
4180 };
4181
4182 static int run(mddev_t *mddev)
4183 {
4184         raid5_conf_t *conf;
4185         int raid_disk, memory;
4186         mdk_rdev_t *rdev;
4187         struct disk_info *disk;
4188         struct list_head *tmp;
4189         int working_disks = 0;
4190
4191         if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4192                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4193                        mdname(mddev), mddev->level);
4194                 return -EIO;
4195         }
4196
4197         if (mddev->reshape_position != MaxSector) {
4198                 /* Check that we can continue the reshape.
4199                  * Currently only disks can change, it must
4200                  * increase, and we must be past the point where
4201                  * a stripe over-writes itself
4202                  */
4203                 sector_t here_new, here_old;
4204                 int old_disks;
4205                 int max_degraded = (mddev->level == 5 ? 1 : 2);
4206
4207                 if (mddev->new_level != mddev->level ||
4208                     mddev->new_layout != mddev->layout ||
4209                     mddev->new_chunk != mddev->chunk_size) {
4210                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4211                                "required - aborting.\n",
4212                                mdname(mddev));
4213                         return -EINVAL;
4214                 }
4215                 if (mddev->delta_disks <= 0) {
4216                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4217                                "(reduce disks) required - aborting.\n",
4218                                mdname(mddev));
4219                         return -EINVAL;
4220                 }
4221                 old_disks = mddev->raid_disks - mddev->delta_disks;
4222                 /* reshape_position must be on a new-stripe boundary, and one
4223                  * further up in new geometry must map after here in old
4224                  * geometry.
4225                  */
4226                 here_new = mddev->reshape_position;
4227                 if (sector_div(here_new, (mddev->chunk_size>>9)*
4228                                (mddev->raid_disks - max_degraded))) {
4229                         printk(KERN_ERR "raid5: reshape_position not "
4230                                "on a stripe boundary\n");
4231                         return -EINVAL;
4232                 }
4233                 /* here_new is the stripe we will write to */
4234                 here_old = mddev->reshape_position;
4235                 sector_div(here_old, (mddev->chunk_size>>9)*
4236                            (old_disks-max_degraded));
4237                 /* here_old is the first stripe that we might need to read
4238                  * from */
4239                 if (here_new >= here_old) {
4240                         /* Reading from the same stripe as writing to - bad */
4241                         printk(KERN_ERR "raid5: reshape_position too early for "
4242                                "auto-recovery - aborting.\n");
4243                         return -EINVAL;
4244                 }
4245                 printk(KERN_INFO "raid5: reshape will continue\n");
4246                 /* OK, we should be able to continue; */
4247         }
4248
4249
4250         mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4251         if ((conf = mddev->private) == NULL)
4252                 goto abort;
4253         if (mddev->reshape_position == MaxSector) {
4254                 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4255         } else {
4256                 conf->raid_disks = mddev->raid_disks;
4257                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4258         }
4259
4260         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4261                               GFP_KERNEL);
4262         if (!conf->disks)
4263                 goto abort;
4264
4265         conf->mddev = mddev;
4266
4267         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4268                 goto abort;
4269
4270         if (mddev->level == 6) {
4271                 conf->spare_page = alloc_page(GFP_KERNEL);
4272                 if (!conf->spare_page)
4273                         goto abort;
4274         }
4275         spin_lock_init(&conf->device_lock);
4276         init_waitqueue_head(&conf->wait_for_stripe);
4277         init_waitqueue_head(&conf->wait_for_overlap);
4278         INIT_LIST_HEAD(&conf->handle_list);
4279         INIT_LIST_HEAD(&conf->delayed_list);
4280         INIT_LIST_HEAD(&conf->bitmap_list);
4281         INIT_LIST_HEAD(&conf->inactive_list);
4282         atomic_set(&conf->active_stripes, 0);
4283         atomic_set(&conf->preread_active_stripes, 0);
4284         atomic_set(&conf->active_aligned_reads, 0);
4285
4286         pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4287
4288         ITERATE_RDEV(mddev,rdev,tmp) {
4289                 raid_disk = rdev->raid_disk;
4290                 if (raid_disk >= conf->raid_disks
4291                     || raid_disk < 0)
4292                         continue;
4293                 disk = conf->disks + raid_disk;
4294
4295                 disk->rdev = rdev;
4296
4297                 if (test_bit(In_sync, &rdev->flags)) {
4298                         char b[BDEVNAME_SIZE];
4299                         printk(KERN_INFO "raid5: device %s operational as raid"
4300                                 " disk %d\n", bdevname(rdev->bdev,b),
4301                                 raid_disk);
4302                         working_disks++;
4303                 }
4304         }
4305
4306         /*
4307          * 0 for a fully functional array, 1 or 2 for a degraded array.
4308          */
4309         mddev->degraded = conf->raid_disks - working_disks;
4310         conf->mddev = mddev;
4311         conf->chunk_size = mddev->chunk_size;
4312         conf->level = mddev->level;
4313         if (conf->level == 6)
4314                 conf->max_degraded = 2;
4315         else
4316                 conf->max_degraded = 1;
4317         conf->algorithm = mddev->layout;
4318         conf->max_nr_stripes = NR_STRIPES;
4319         conf->expand_progress = mddev->reshape_position;
4320
4321         /* device size must be a multiple of chunk size */
4322         mddev->size &= ~(mddev->chunk_size/1024 -1);
4323         mddev->resync_max_sectors = mddev->size << 1;
4324
4325         if (conf->level == 6 && conf->raid_disks < 4) {
4326                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4327                        mdname(mddev), conf->raid_disks);
4328                 goto abort;
4329         }
4330         if (!conf->chunk_size || conf->chunk_size % 4) {
4331                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4332                         conf->chunk_size, mdname(mddev));
4333                 goto abort;
4334         }
4335         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4336                 printk(KERN_ERR 
4337                         "raid5: unsupported parity algorithm %d for %s\n",
4338                         conf->algorithm, mdname(mddev));
4339                 goto abort;
4340         }
4341         if (mddev->degraded > conf->max_degraded) {
4342                 printk(KERN_ERR "raid5: not enough operational devices for %s"
4343                         " (%d/%d failed)\n",
4344                         mdname(mddev), mddev->degraded, conf->raid_disks);
4345                 goto abort;
4346         }
4347
4348         if (mddev->degraded > 0 &&
4349             mddev->recovery_cp != MaxSector) {
4350                 if (mddev->ok_start_degraded)
4351                         printk(KERN_WARNING
4352                                "raid5: starting dirty degraded array: %s"
4353                                "- data corruption possible.\n",
4354                                mdname(mddev));
4355                 else {
4356                         printk(KERN_ERR
4357                                "raid5: cannot start dirty degraded array for %s\n",
4358                                mdname(mddev));
4359                         goto abort;
4360                 }
4361         }
4362
4363         {
4364                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4365                 if (!mddev->thread) {
4366                         printk(KERN_ERR 
4367                                 "raid5: couldn't allocate thread for %s\n",
4368                                 mdname(mddev));
4369                         goto abort;
4370                 }
4371         }
4372         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4373                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4374         if (grow_stripes(conf, conf->max_nr_stripes)) {
4375                 printk(KERN_ERR 
4376                         "raid5: couldn't allocate %dkB for buffers\n", memory);
4377                 shrink_stripes(conf);
4378                 md_unregister_thread(mddev->thread);
4379                 goto abort;
4380         } else
4381                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4382                         memory, mdname(mddev));
4383
4384         if (mddev->degraded == 0)
4385                 printk("raid5: raid level %d set %s active with %d out of %d"
4386                         " devices, algorithm %d\n", conf->level, mdname(mddev), 
4387                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4388                         conf->algorithm);
4389         else
4390                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4391                         " out of %d devices, algorithm %d\n", conf->level,
4392                         mdname(mddev), mddev->raid_disks - mddev->degraded,
4393                         mddev->raid_disks, conf->algorithm);
4394
4395         print_raid5_conf(conf);
4396
4397         if (conf->expand_progress != MaxSector) {
4398                 printk("...ok start reshape thread\n");
4399                 conf->expand_lo = conf->expand_progress;
4400                 atomic_set(&conf->reshape_stripes, 0);
4401                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4402                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4403                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4404                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4405                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4406                                                         "%s_reshape");
4407         }
4408
4409         /* read-ahead size must cover two whole stripes, which is
4410          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4411          */
4412         {
4413                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4414                 int stripe = data_disks *
4415                         (mddev->chunk_size / PAGE_SIZE);
4416                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4417                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4418         }
4419
4420         /* Ok, everything is just fine now */
4421         if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4422                 printk(KERN_WARNING
4423                        "raid5: failed to create sysfs attributes for %s\n",
4424                        mdname(mddev));
4425
4426         mddev->queue->unplug_fn = raid5_unplug_device;
4427         mddev->queue->issue_flush_fn = raid5_issue_flush;
4428         mddev->queue->backing_dev_info.congested_data = mddev;
4429         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4430
4431         mddev->array_size =  mddev->size * (conf->previous_raid_disks -
4432                                             conf->max_degraded);
4433
4434         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4435
4436         return 0;
4437 abort:
4438         if (conf) {
4439                 print_raid5_conf(conf);
4440                 safe_put_page(conf->spare_page);
4441                 kfree(conf->disks);
4442                 kfree(conf->stripe_hashtbl);
4443                 kfree(conf);
4444         }
4445         mddev->private = NULL;
4446         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4447         return -EIO;
4448 }
4449
4450
4451
4452 static int stop(mddev_t *mddev)
4453 {
4454         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4455
4456         md_unregister_thread(mddev->thread);
4457         mddev->thread = NULL;
4458         shrink_stripes(conf);
4459         kfree(conf->stripe_hashtbl);
4460         mddev->queue->backing_dev_info.congested_fn = NULL;
4461         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4462         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4463         kfree(conf->disks);
4464         kfree(conf);
4465         mddev->private = NULL;
4466         return 0;
4467 }
4468
4469 #ifdef DEBUG
4470 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4471 {
4472         int i;
4473
4474         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4475                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4476         seq_printf(seq, "sh %llu,  count %d.\n",
4477                    (unsigned long long)sh->sector, atomic_read(&sh->count));
4478         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4479         for (i = 0; i < sh->disks; i++) {
4480                 seq_printf(seq, "(cache%d: %p %ld) ",
4481                            i, sh->dev[i].page, sh->dev[i].flags);
4482         }
4483         seq_printf(seq, "\n");
4484 }
4485
4486 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4487 {
4488         struct stripe_head *sh;
4489         struct hlist_node *hn;
4490         int i;
4491
4492         spin_lock_irq(&conf->device_lock);
4493         for (i = 0; i < NR_HASH; i++) {
4494                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4495                         if (sh->raid_conf != conf)
4496                                 continue;
4497                         print_sh(seq, sh);
4498                 }
4499         }
4500         spin_unlock_irq(&conf->device_lock);
4501 }
4502 #endif
4503
4504 static void status (struct seq_file *seq, mddev_t *mddev)
4505 {
4506         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4507         int i;
4508
4509         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4510         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4511         for (i = 0; i < conf->raid_disks; i++)
4512                 seq_printf (seq, "%s",
4513                                conf->disks[i].rdev &&
4514                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4515         seq_printf (seq, "]");
4516 #ifdef DEBUG
4517         seq_printf (seq, "\n");
4518         printall(seq, conf);
4519 #endif
4520 }
4521
4522 static void print_raid5_conf (raid5_conf_t *conf)
4523 {
4524         int i;
4525         struct disk_info *tmp;
4526
4527         printk("RAID5 conf printout:\n");
4528         if (!conf) {
4529                 printk("(conf==NULL)\n");
4530                 return;
4531         }
4532         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4533                  conf->raid_disks - conf->mddev->degraded);
4534
4535         for (i = 0; i < conf->raid_disks; i++) {
4536                 char b[BDEVNAME_SIZE];
4537                 tmp = conf->disks + i;
4538                 if (tmp->rdev)
4539                 printk(" disk %d, o:%d, dev:%s\n",
4540                         i, !test_bit(Faulty, &tmp->rdev->flags),
4541                         bdevname(tmp->rdev->bdev,b));
4542         }
4543 }
4544
4545 static int raid5_spare_active(mddev_t *mddev)
4546 {
4547         int i;
4548         raid5_conf_t *conf = mddev->private;
4549         struct disk_info *tmp;
4550
4551         for (i = 0; i < conf->raid_disks; i++) {
4552                 tmp = conf->disks + i;
4553                 if (tmp->rdev
4554                     && !test_bit(Faulty, &tmp->rdev->flags)
4555                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4556                         unsigned long flags;
4557                         spin_lock_irqsave(&conf->device_lock, flags);
4558                         mddev->degraded--;
4559                         spin_unlock_irqrestore(&conf->device_lock, flags);
4560                 }
4561         }
4562         print_raid5_conf(conf);
4563         return 0;
4564 }
4565
4566 static int raid5_remove_disk(mddev_t *mddev, int number)
4567 {
4568         raid5_conf_t *conf = mddev->private;
4569         int err = 0;
4570         mdk_rdev_t *rdev;
4571         struct disk_info *p = conf->disks + number;
4572
4573         print_raid5_conf(conf);
4574         rdev = p->rdev;
4575         if (rdev) {
4576                 if (test_bit(In_sync, &rdev->flags) ||
4577                     atomic_read(&rdev->nr_pending)) {
4578                         err = -EBUSY;
4579                         goto abort;
4580                 }
4581                 p->rdev = NULL;
4582                 synchronize_rcu();
4583                 if (atomic_read(&rdev->nr_pending)) {
4584                         /* lost the race, try later */
4585                         err = -EBUSY;
4586                         p->rdev = rdev;
4587                 }
4588         }
4589 abort:
4590
4591         print_raid5_conf(conf);
4592         return err;
4593 }
4594
4595 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4596 {
4597         raid5_conf_t *conf = mddev->private;
4598         int found = 0;
4599         int disk;
4600         struct disk_info *p;
4601
4602         if (mddev->degraded > conf->max_degraded)
4603                 /* no point adding a device */
4604                 return 0;
4605
4606         /*
4607          * find the disk ... but prefer rdev->saved_raid_disk
4608          * if possible.
4609          */
4610         if (rdev->saved_raid_disk >= 0 &&
4611             conf->disks[rdev->saved_raid_disk].rdev == NULL)
4612                 disk = rdev->saved_raid_disk;
4613         else
4614                 disk = 0;
4615         for ( ; disk < conf->raid_disks; disk++)
4616                 if ((p=conf->disks + disk)->rdev == NULL) {
4617                         clear_bit(In_sync, &rdev->flags);
4618                         rdev->raid_disk = disk;
4619                         found = 1;
4620                         if (rdev->saved_raid_disk != disk)
4621                                 conf->fullsync = 1;
4622                         rcu_assign_pointer(p->rdev, rdev);
4623                         break;
4624                 }
4625         print_raid5_conf(conf);
4626         return found;
4627 }
4628
4629 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4630 {
4631         /* no resync is happening, and there is enough space
4632          * on all devices, so we can resize.
4633          * We need to make sure resync covers any new space.
4634          * If the array is shrinking we should possibly wait until
4635          * any io in the removed space completes, but it hardly seems
4636          * worth it.
4637          */
4638         raid5_conf_t *conf = mddev_to_conf(mddev);
4639
4640         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4641         mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4642         set_capacity(mddev->gendisk, mddev->array_size << 1);
4643         mddev->changed = 1;
4644         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
4645                 mddev->recovery_cp = mddev->size << 1;
4646                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4647         }
4648         mddev->size = sectors /2;
4649         mddev->resync_max_sectors = sectors;
4650         return 0;
4651 }
4652
4653 #ifdef CONFIG_MD_RAID5_RESHAPE
4654 static int raid5_check_reshape(mddev_t *mddev)
4655 {
4656         raid5_conf_t *conf = mddev_to_conf(mddev);
4657         int err;
4658
4659         if (mddev->delta_disks < 0 ||
4660             mddev->new_level != mddev->level)
4661                 return -EINVAL; /* Cannot shrink array or change level yet */
4662         if (mddev->delta_disks == 0)
4663                 return 0; /* nothing to do */
4664
4665         /* Can only proceed if there are plenty of stripe_heads.
4666          * We need a minimum of one full stripe,, and for sensible progress
4667          * it is best to have about 4 times that.
4668          * If we require 4 times, then the default 256 4K stripe_heads will
4669          * allow for chunk sizes up to 256K, which is probably OK.
4670          * If the chunk size is greater, user-space should request more
4671          * stripe_heads first.
4672          */
4673         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4674             (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4675                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
4676                        (mddev->chunk_size / STRIPE_SIZE)*4);
4677                 return -ENOSPC;
4678         }
4679
4680         err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4681         if (err)
4682                 return err;
4683
4684         if (mddev->degraded > conf->max_degraded)
4685                 return -EINVAL;
4686         /* looks like we might be able to manage this */
4687         return 0;
4688 }
4689
4690 static int raid5_start_reshape(mddev_t *mddev)
4691 {
4692         raid5_conf_t *conf = mddev_to_conf(mddev);
4693         mdk_rdev_t *rdev;
4694         struct list_head *rtmp;
4695         int spares = 0;
4696         int added_devices = 0;
4697         unsigned long flags;
4698
4699         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4700                 return -EBUSY;
4701
4702         ITERATE_RDEV(mddev, rdev, rtmp)
4703                 if (rdev->raid_disk < 0 &&
4704                     !test_bit(Faulty, &rdev->flags))
4705                         spares++;
4706
4707         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4708                 /* Not enough devices even to make a degraded array
4709                  * of that size
4710                  */
4711                 return -EINVAL;
4712
4713         atomic_set(&conf->reshape_stripes, 0);
4714         spin_lock_irq(&conf->device_lock);
4715         conf->previous_raid_disks = conf->raid_disks;
4716         conf->raid_disks += mddev->delta_disks;
4717         conf->expand_progress = 0;
4718         conf->expand_lo = 0;
4719         spin_unlock_irq(&conf->device_lock);
4720
4721         /* Add some new drives, as many as will fit.
4722          * We know there are enough to make the newly sized array work.
4723          */
4724         ITERATE_RDEV(mddev, rdev, rtmp)
4725                 if (rdev->raid_disk < 0 &&
4726                     !test_bit(Faulty, &rdev->flags)) {
4727                         if (raid5_add_disk(mddev, rdev)) {
4728                                 char nm[20];
4729                                 set_bit(In_sync, &rdev->flags);
4730                                 added_devices++;
4731                                 rdev->recovery_offset = 0;
4732                                 sprintf(nm, "rd%d", rdev->raid_disk);
4733                                 if (sysfs_create_link(&mddev->kobj,
4734                                                       &rdev->kobj, nm))
4735                                         printk(KERN_WARNING
4736                                                "raid5: failed to create "
4737                                                " link %s for %s\n",
4738                                                nm, mdname(mddev));
4739                         } else
4740                                 break;
4741                 }
4742
4743         spin_lock_irqsave(&conf->device_lock, flags);
4744         mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4745         spin_unlock_irqrestore(&conf->device_lock, flags);
4746         mddev->raid_disks = conf->raid_disks;
4747         mddev->reshape_position = 0;
4748         set_bit(MD_CHANGE_DEVS, &mddev->flags);
4749
4750         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4751         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4752         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4753         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4754         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4755                                                 "%s_reshape");
4756         if (!mddev->sync_thread) {
4757                 mddev->recovery = 0;
4758                 spin_lock_irq(&conf->device_lock);
4759                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4760                 conf->expand_progress = MaxSector;
4761                 spin_unlock_irq(&conf->device_lock);
4762                 return -EAGAIN;
4763         }
4764         md_wakeup_thread(mddev->sync_thread);
4765         md_new_event(mddev);
4766         return 0;
4767 }
4768 #endif
4769
4770 static void end_reshape(raid5_conf_t *conf)
4771 {
4772         struct block_device *bdev;
4773
4774         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4775                 conf->mddev->array_size = conf->mddev->size *
4776                         (conf->raid_disks - conf->max_degraded);
4777                 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4778                 conf->mddev->changed = 1;
4779
4780                 bdev = bdget_disk(conf->mddev->gendisk, 0);
4781                 if (bdev) {
4782                         mutex_lock(&bdev->bd_inode->i_mutex);
4783                         i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4784                         mutex_unlock(&bdev->bd_inode->i_mutex);
4785                         bdput(bdev);
4786                 }
4787                 spin_lock_irq(&conf->device_lock);
4788                 conf->expand_progress = MaxSector;
4789                 spin_unlock_irq(&conf->device_lock);
4790                 conf->mddev->reshape_position = MaxSector;
4791
4792                 /* read-ahead size must cover two whole stripes, which is
4793                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4794                  */
4795                 {
4796                         int data_disks = conf->previous_raid_disks - conf->max_degraded;
4797                         int stripe = data_disks *
4798                                 (conf->mddev->chunk_size / PAGE_SIZE);
4799                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4800                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4801                 }
4802         }
4803 }
4804
4805 static void raid5_quiesce(mddev_t *mddev, int state)
4806 {
4807         raid5_conf_t *conf = mddev_to_conf(mddev);
4808
4809         switch(state) {
4810         case 2: /* resume for a suspend */
4811                 wake_up(&conf->wait_for_overlap);
4812                 break;
4813
4814         case 1: /* stop all writes */
4815                 spin_lock_irq(&conf->device_lock);
4816                 conf->quiesce = 1;
4817                 wait_event_lock_irq(conf->wait_for_stripe,
4818                                     atomic_read(&conf->active_stripes) == 0 &&
4819                                     atomic_read(&conf->active_aligned_reads) == 0,
4820                                     conf->device_lock, /* nothing */);
4821                 spin_unlock_irq(&conf->device_lock);
4822                 break;
4823
4824         case 0: /* re-enable writes */
4825                 spin_lock_irq(&conf->device_lock);
4826                 conf->quiesce = 0;
4827                 wake_up(&conf->wait_for_stripe);
4828                 wake_up(&conf->wait_for_overlap);
4829                 spin_unlock_irq(&conf->device_lock);
4830                 break;
4831         }
4832 }
4833
4834 static struct mdk_personality raid6_personality =
4835 {
4836         .name           = "raid6",
4837         .level          = 6,
4838         .owner          = THIS_MODULE,
4839         .make_request   = make_request,
4840         .run            = run,
4841         .stop           = stop,
4842         .status         = status,
4843         .error_handler  = error,
4844         .hot_add_disk   = raid5_add_disk,
4845         .hot_remove_disk= raid5_remove_disk,
4846         .spare_active   = raid5_spare_active,
4847         .sync_request   = sync_request,
4848         .resize         = raid5_resize,
4849 #ifdef CONFIG_MD_RAID5_RESHAPE
4850         .check_reshape  = raid5_check_reshape,
4851         .start_reshape  = raid5_start_reshape,
4852 #endif
4853         .quiesce        = raid5_quiesce,
4854 };
4855 static struct mdk_personality raid5_personality =
4856 {
4857         .name           = "raid5",
4858         .level          = 5,
4859         .owner          = THIS_MODULE,
4860         .make_request   = make_request,
4861         .run            = run,
4862         .stop           = stop,
4863         .status         = status,
4864         .error_handler  = error,
4865         .hot_add_disk   = raid5_add_disk,
4866         .hot_remove_disk= raid5_remove_disk,
4867         .spare_active   = raid5_spare_active,
4868         .sync_request   = sync_request,
4869         .resize         = raid5_resize,
4870 #ifdef CONFIG_MD_RAID5_RESHAPE
4871         .check_reshape  = raid5_check_reshape,
4872         .start_reshape  = raid5_start_reshape,
4873 #endif
4874         .quiesce        = raid5_quiesce,
4875 };
4876
4877 static struct mdk_personality raid4_personality =
4878 {
4879         .name           = "raid4",
4880         .level          = 4,
4881         .owner          = THIS_MODULE,
4882         .make_request   = make_request,
4883         .run            = run,
4884         .stop           = stop,
4885         .status         = status,
4886         .error_handler  = error,
4887         .hot_add_disk   = raid5_add_disk,
4888         .hot_remove_disk= raid5_remove_disk,
4889         .spare_active   = raid5_spare_active,
4890         .sync_request   = sync_request,
4891         .resize         = raid5_resize,
4892 #ifdef CONFIG_MD_RAID5_RESHAPE
4893         .check_reshape  = raid5_check_reshape,
4894         .start_reshape  = raid5_start_reshape,
4895 #endif
4896         .quiesce        = raid5_quiesce,
4897 };
4898
4899 static int __init raid5_init(void)
4900 {
4901         int e;
4902
4903         e = raid6_select_algo();
4904         if ( e )
4905                 return e;
4906         register_md_personality(&raid6_personality);
4907         register_md_personality(&raid5_personality);
4908         register_md_personality(&raid4_personality);
4909         return 0;
4910 }
4911
4912 static void raid5_exit(void)
4913 {
4914         unregister_md_personality(&raid6_personality);
4915         unregister_md_personality(&raid5_personality);
4916         unregister_md_personality(&raid4_personality);
4917 }
4918
4919 module_init(raid5_init);
4920 module_exit(raid5_exit);
4921 MODULE_LICENSE("GPL");
4922 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4923 MODULE_ALIAS("md-raid5");
4924 MODULE_ALIAS("md-raid4");
4925 MODULE_ALIAS("md-level-5");
4926 MODULE_ALIAS("md-level-4");
4927 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4928 MODULE_ALIAS("md-raid6");
4929 MODULE_ALIAS("md-level-6");
4930
4931 /* This used to be two separate modules, they were: */
4932 MODULE_ALIAS("raid5");
4933 MODULE_ALIAS("raid6");