968dacaced6de669181bba2d1e1902fbea3ffa8f
[linux-2.6.git] / drivers / md / raid5.c
1 /*
2  * raid5.c : Multiple Devices driver for Linux
3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4  *         Copyright (C) 1999, 2000 Ingo Molnar
5  *         Copyright (C) 2002, 2003 H. Peter Anvin
6  *
7  * RAID-4/5/6 management functions.
8  * Thanks to Penguin Computing for making the RAID-6 development possible
9  * by donating a test server!
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 /*
22  * BITMAP UNPLUGGING:
23  *
24  * The sequencing for updating the bitmap reliably is a little
25  * subtle (and I got it wrong the first time) so it deserves some
26  * explanation.
27  *
28  * We group bitmap updates into batches.  Each batch has a number.
29  * We may write out several batches at once, but that isn't very important.
30  * conf->bm_write is the number of the last batch successfully written.
31  * conf->bm_flush is the number of the last batch that was closed to
32  *    new additions.
33  * When we discover that we will need to write to any block in a stripe
34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35  * the number of the batch it will be in. This is bm_flush+1.
36  * When we are ready to do a write, if that batch hasn't been written yet,
37  *   we plug the array and queue the stripe for later.
38  * When an unplug happens, we increment bm_flush, thus closing the current
39  *   batch.
40  * When we notice that bm_flush > bm_write, we write out all pending updates
41  * to the bitmap, and advance bm_write to where bm_flush was.
42  * This may occasionally write a bit out twice, but is sure never to
43  * miss any bits.
44  */
45
46 #include <linux/module.h>
47 #include <linux/slab.h>
48 #include <linux/highmem.h>
49 #include <linux/bitops.h>
50 #include <linux/kthread.h>
51 #include <asm/atomic.h>
52 #include "raid6.h"
53
54 #include <linux/raid/bitmap.h>
55 #include <linux/async_tx.h>
56
57 /*
58  * Stripe cache
59  */
60
61 #define NR_STRIPES              256
62 #define STRIPE_SIZE             PAGE_SIZE
63 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
64 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
65 #define IO_THRESHOLD            1
66 #define BYPASS_THRESHOLD        1
67 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
68 #define HASH_MASK               (NR_HASH - 1)
69
70 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
71
72 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
73  * order without overlap.  There may be several bio's per stripe+device, and
74  * a bio could span several devices.
75  * When walking this list for a particular stripe+device, we must never proceed
76  * beyond a bio that extends past this device, as the next bio might no longer
77  * be valid.
78  * This macro is used to determine the 'next' bio in the list, given the sector
79  * of the current stripe+device
80  */
81 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
82 /*
83  * The following can be used to debug the driver
84  */
85 #define RAID5_PARANOIA  1
86 #if RAID5_PARANOIA && defined(CONFIG_SMP)
87 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
88 #else
89 # define CHECK_DEVLOCK()
90 #endif
91
92 #ifdef DEBUG
93 #define inline
94 #define __inline__
95 #endif
96
97 #if !RAID6_USE_EMPTY_ZERO_PAGE
98 /* In .bss so it's zeroed */
99 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
100 #endif
101
102 static inline int raid6_next_disk(int disk, int raid_disks)
103 {
104         disk++;
105         return (disk < raid_disks) ? disk : 0;
106 }
107
108 static void return_io(struct bio *return_bi)
109 {
110         struct bio *bi = return_bi;
111         while (bi) {
112
113                 return_bi = bi->bi_next;
114                 bi->bi_next = NULL;
115                 bi->bi_size = 0;
116                 bi->bi_end_io(bi,
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(struct request_queue *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         if (unlikely(sh->ops.count < 0)) {
381                 printk(KERN_ERR "pending: %#lx ops.pending: %#lx ops.ack: %#lx "
382                         "ops.complete: %#lx\n", pending, sh->ops.pending,
383                         sh->ops.ack, sh->ops.complete);
384                 BUG();
385         }
386
387         return pending;
388 }
389
390 static void
391 raid5_end_read_request(struct bio *bi, int error);
392 static void
393 raid5_end_write_request(struct bio *bi, int error);
394
395 static void ops_run_io(struct stripe_head *sh)
396 {
397         raid5_conf_t *conf = sh->raid_conf;
398         int i, disks = sh->disks;
399
400         might_sleep();
401
402         set_bit(STRIPE_IO_STARTED, &sh->state);
403         for (i = disks; i--; ) {
404                 int rw;
405                 struct bio *bi;
406                 mdk_rdev_t *rdev;
407                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
408                         rw = WRITE;
409                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
410                         rw = READ;
411                 else
412                         continue;
413
414                 bi = &sh->dev[i].req;
415
416                 bi->bi_rw = rw;
417                 if (rw == WRITE)
418                         bi->bi_end_io = raid5_end_write_request;
419                 else
420                         bi->bi_end_io = raid5_end_read_request;
421
422                 rcu_read_lock();
423                 rdev = rcu_dereference(conf->disks[i].rdev);
424                 if (rdev && test_bit(Faulty, &rdev->flags))
425                         rdev = NULL;
426                 if (rdev)
427                         atomic_inc(&rdev->nr_pending);
428                 rcu_read_unlock();
429
430                 if (rdev) {
431                         if (test_bit(STRIPE_SYNCING, &sh->state) ||
432                                 test_bit(STRIPE_EXPAND_SOURCE, &sh->state) ||
433                                 test_bit(STRIPE_EXPAND_READY, &sh->state))
434                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
435
436                         bi->bi_bdev = rdev->bdev;
437                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
438                                 __func__, (unsigned long long)sh->sector,
439                                 bi->bi_rw, i);
440                         atomic_inc(&sh->count);
441                         bi->bi_sector = sh->sector + rdev->data_offset;
442                         bi->bi_flags = 1 << BIO_UPTODATE;
443                         bi->bi_vcnt = 1;
444                         bi->bi_max_vecs = 1;
445                         bi->bi_idx = 0;
446                         bi->bi_io_vec = &sh->dev[i].vec;
447                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
448                         bi->bi_io_vec[0].bv_offset = 0;
449                         bi->bi_size = STRIPE_SIZE;
450                         bi->bi_next = NULL;
451                         if (rw == WRITE &&
452                             test_bit(R5_ReWrite, &sh->dev[i].flags))
453                                 atomic_add(STRIPE_SECTORS,
454                                         &rdev->corrected_errors);
455                         generic_make_request(bi);
456                 } else {
457                         if (rw == WRITE)
458                                 set_bit(STRIPE_DEGRADED, &sh->state);
459                         pr_debug("skip op %ld on disc %d for sector %llu\n",
460                                 bi->bi_rw, i, (unsigned long long)sh->sector);
461                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
462                         set_bit(STRIPE_HANDLE, &sh->state);
463                 }
464         }
465 }
466
467 static struct dma_async_tx_descriptor *
468 async_copy_data(int frombio, struct bio *bio, struct page *page,
469         sector_t sector, struct dma_async_tx_descriptor *tx)
470 {
471         struct bio_vec *bvl;
472         struct page *bio_page;
473         int i;
474         int page_offset;
475
476         if (bio->bi_sector >= sector)
477                 page_offset = (signed)(bio->bi_sector - sector) * 512;
478         else
479                 page_offset = (signed)(sector - bio->bi_sector) * -512;
480         bio_for_each_segment(bvl, bio, i) {
481                 int len = bio_iovec_idx(bio, i)->bv_len;
482                 int clen;
483                 int b_offset = 0;
484
485                 if (page_offset < 0) {
486                         b_offset = -page_offset;
487                         page_offset += b_offset;
488                         len -= b_offset;
489                 }
490
491                 if (len > 0 && page_offset + len > STRIPE_SIZE)
492                         clen = STRIPE_SIZE - page_offset;
493                 else
494                         clen = len;
495
496                 if (clen > 0) {
497                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
498                         bio_page = bio_iovec_idx(bio, i)->bv_page;
499                         if (frombio)
500                                 tx = async_memcpy(page, bio_page, page_offset,
501                                         b_offset, clen,
502                                         ASYNC_TX_DEP_ACK,
503                                         tx, NULL, NULL);
504                         else
505                                 tx = async_memcpy(bio_page, page, b_offset,
506                                         page_offset, clen,
507                                         ASYNC_TX_DEP_ACK,
508                                         tx, NULL, NULL);
509                 }
510                 if (clen < len) /* hit end of page */
511                         break;
512                 page_offset +=  len;
513         }
514
515         return tx;
516 }
517
518 static void ops_complete_biofill(void *stripe_head_ref)
519 {
520         struct stripe_head *sh = stripe_head_ref;
521         struct bio *return_bi = NULL;
522         raid5_conf_t *conf = sh->raid_conf;
523         int i;
524
525         pr_debug("%s: stripe %llu\n", __func__,
526                 (unsigned long long)sh->sector);
527
528         /* clear completed biofills */
529         for (i = sh->disks; i--; ) {
530                 struct r5dev *dev = &sh->dev[i];
531
532                 /* acknowledge completion of a biofill operation */
533                 /* and check if we need to reply to a read request,
534                  * new R5_Wantfill requests are held off until
535                  * !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending)
536                  */
537                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
538                         struct bio *rbi, *rbi2;
539
540                         /* The access to dev->read is outside of the
541                          * spin_lock_irq(&conf->device_lock), but is protected
542                          * by the STRIPE_OP_BIOFILL pending bit
543                          */
544                         BUG_ON(!dev->read);
545                         rbi = dev->read;
546                         dev->read = NULL;
547                         while (rbi && rbi->bi_sector <
548                                 dev->sector + STRIPE_SECTORS) {
549                                 rbi2 = r5_next_bio(rbi, dev->sector);
550                                 spin_lock_irq(&conf->device_lock);
551                                 if (--rbi->bi_phys_segments == 0) {
552                                         rbi->bi_next = return_bi;
553                                         return_bi = rbi;
554                                 }
555                                 spin_unlock_irq(&conf->device_lock);
556                                 rbi = rbi2;
557                         }
558                 }
559         }
560         set_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
561
562         return_io(return_bi);
563
564         set_bit(STRIPE_HANDLE, &sh->state);
565         release_stripe(sh);
566 }
567
568 static void ops_run_biofill(struct stripe_head *sh)
569 {
570         struct dma_async_tx_descriptor *tx = NULL;
571         raid5_conf_t *conf = sh->raid_conf;
572         int i;
573
574         pr_debug("%s: stripe %llu\n", __func__,
575                 (unsigned long long)sh->sector);
576
577         for (i = sh->disks; i--; ) {
578                 struct r5dev *dev = &sh->dev[i];
579                 if (test_bit(R5_Wantfill, &dev->flags)) {
580                         struct bio *rbi;
581                         spin_lock_irq(&conf->device_lock);
582                         dev->read = rbi = dev->toread;
583                         dev->toread = NULL;
584                         spin_unlock_irq(&conf->device_lock);
585                         while (rbi && rbi->bi_sector <
586                                 dev->sector + STRIPE_SECTORS) {
587                                 tx = async_copy_data(0, rbi, dev->page,
588                                         dev->sector, tx);
589                                 rbi = r5_next_bio(rbi, dev->sector);
590                         }
591                 }
592         }
593
594         atomic_inc(&sh->count);
595         async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
596                 ops_complete_biofill, sh);
597 }
598
599 static void ops_complete_compute5(void *stripe_head_ref)
600 {
601         struct stripe_head *sh = stripe_head_ref;
602         int target = sh->ops.target;
603         struct r5dev *tgt = &sh->dev[target];
604
605         pr_debug("%s: stripe %llu\n", __func__,
606                 (unsigned long long)sh->sector);
607
608         set_bit(R5_UPTODATE, &tgt->flags);
609         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
610         clear_bit(R5_Wantcompute, &tgt->flags);
611         set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
612         set_bit(STRIPE_HANDLE, &sh->state);
613         release_stripe(sh);
614 }
615
616 static struct dma_async_tx_descriptor *
617 ops_run_compute5(struct stripe_head *sh, unsigned long pending)
618 {
619         /* kernel stack size limits the total number of disks */
620         int disks = sh->disks;
621         struct page *xor_srcs[disks];
622         int target = sh->ops.target;
623         struct r5dev *tgt = &sh->dev[target];
624         struct page *xor_dest = tgt->page;
625         int count = 0;
626         struct dma_async_tx_descriptor *tx;
627         int i;
628
629         pr_debug("%s: stripe %llu block: %d\n",
630                 __func__, (unsigned long long)sh->sector, target);
631         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
632
633         for (i = disks; i--; )
634                 if (i != target)
635                         xor_srcs[count++] = sh->dev[i].page;
636
637         atomic_inc(&sh->count);
638
639         if (unlikely(count == 1))
640                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
641                         0, NULL, ops_complete_compute5, sh);
642         else
643                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
644                         ASYNC_TX_XOR_ZERO_DST, NULL,
645                         ops_complete_compute5, sh);
646
647         /* ack now if postxor is not set to be run */
648         if (tx && !test_bit(STRIPE_OP_POSTXOR, &pending))
649                 async_tx_ack(tx);
650
651         return tx;
652 }
653
654 static void ops_complete_prexor(void *stripe_head_ref)
655 {
656         struct stripe_head *sh = stripe_head_ref;
657
658         pr_debug("%s: stripe %llu\n", __func__,
659                 (unsigned long long)sh->sector);
660
661         set_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
662 }
663
664 static struct dma_async_tx_descriptor *
665 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
666 {
667         /* kernel stack size limits the total number of disks */
668         int disks = sh->disks;
669         struct page *xor_srcs[disks];
670         int count = 0, pd_idx = sh->pd_idx, i;
671
672         /* existing parity data subtracted */
673         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
674
675         pr_debug("%s: stripe %llu\n", __func__,
676                 (unsigned long long)sh->sector);
677
678         for (i = disks; i--; ) {
679                 struct r5dev *dev = &sh->dev[i];
680                 /* Only process blocks that are known to be uptodate */
681                 if (dev->towrite && test_bit(R5_Wantprexor, &dev->flags))
682                         xor_srcs[count++] = dev->page;
683         }
684
685         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
686                 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
687                 ops_complete_prexor, sh);
688
689         return tx;
690 }
691
692 static struct dma_async_tx_descriptor *
693 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx,
694                  unsigned long pending)
695 {
696         int disks = sh->disks;
697         int pd_idx = sh->pd_idx, i;
698
699         /* check if prexor is active which means only process blocks
700          * that are part of a read-modify-write (Wantprexor)
701          */
702         int prexor = test_bit(STRIPE_OP_PREXOR, &pending);
703
704         pr_debug("%s: stripe %llu\n", __func__,
705                 (unsigned long long)sh->sector);
706
707         for (i = disks; i--; ) {
708                 struct r5dev *dev = &sh->dev[i];
709                 struct bio *chosen;
710                 int towrite;
711
712                 towrite = 0;
713                 if (prexor) { /* rmw */
714                         if (dev->towrite &&
715                             test_bit(R5_Wantprexor, &dev->flags))
716                                 towrite = 1;
717                 } else { /* rcw */
718                         if (i != pd_idx && dev->towrite &&
719                                 test_bit(R5_LOCKED, &dev->flags))
720                                 towrite = 1;
721                 }
722
723                 if (towrite) {
724                         struct bio *wbi;
725
726                         spin_lock(&sh->lock);
727                         chosen = dev->towrite;
728                         dev->towrite = NULL;
729                         BUG_ON(dev->written);
730                         wbi = dev->written = chosen;
731                         spin_unlock(&sh->lock);
732
733                         while (wbi && wbi->bi_sector <
734                                 dev->sector + STRIPE_SECTORS) {
735                                 tx = async_copy_data(1, wbi, dev->page,
736                                         dev->sector, tx);
737                                 wbi = r5_next_bio(wbi, dev->sector);
738                         }
739                 }
740         }
741
742         return tx;
743 }
744
745 static void ops_complete_postxor(void *stripe_head_ref)
746 {
747         struct stripe_head *sh = stripe_head_ref;
748
749         pr_debug("%s: stripe %llu\n", __func__,
750                 (unsigned long long)sh->sector);
751
752         set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
753         set_bit(STRIPE_HANDLE, &sh->state);
754         release_stripe(sh);
755 }
756
757 static void ops_complete_write(void *stripe_head_ref)
758 {
759         struct stripe_head *sh = stripe_head_ref;
760         int disks = sh->disks, i, pd_idx = sh->pd_idx;
761
762         pr_debug("%s: stripe %llu\n", __func__,
763                 (unsigned long long)sh->sector);
764
765         for (i = disks; i--; ) {
766                 struct r5dev *dev = &sh->dev[i];
767                 if (dev->written || i == pd_idx)
768                         set_bit(R5_UPTODATE, &dev->flags);
769         }
770
771         set_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
772         set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
773
774         set_bit(STRIPE_HANDLE, &sh->state);
775         release_stripe(sh);
776 }
777
778 static void
779 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx,
780                 unsigned long pending)
781 {
782         /* kernel stack size limits the total number of disks */
783         int disks = sh->disks;
784         struct page *xor_srcs[disks];
785
786         int count = 0, pd_idx = sh->pd_idx, i;
787         struct page *xor_dest;
788         int prexor = test_bit(STRIPE_OP_PREXOR, &pending);
789         unsigned long flags;
790         dma_async_tx_callback callback;
791
792         pr_debug("%s: stripe %llu\n", __func__,
793                 (unsigned long long)sh->sector);
794
795         /* check if prexor is active which means only process blocks
796          * that are part of a read-modify-write (written)
797          */
798         if (prexor) {
799                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
800                 for (i = disks; i--; ) {
801                         struct r5dev *dev = &sh->dev[i];
802                         if (dev->written)
803                                 xor_srcs[count++] = dev->page;
804                 }
805         } else {
806                 xor_dest = sh->dev[pd_idx].page;
807                 for (i = disks; i--; ) {
808                         struct r5dev *dev = &sh->dev[i];
809                         if (i != pd_idx)
810                                 xor_srcs[count++] = dev->page;
811                 }
812         }
813
814         /* check whether this postxor is part of a write */
815         callback = test_bit(STRIPE_OP_BIODRAIN, &pending) ?
816                 ops_complete_write : ops_complete_postxor;
817
818         /* 1/ if we prexor'd then the dest is reused as a source
819          * 2/ if we did not prexor then we are redoing the parity
820          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
821          * for the synchronous xor case
822          */
823         flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
824                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
825
826         atomic_inc(&sh->count);
827
828         if (unlikely(count == 1)) {
829                 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
830                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
831                         flags, tx, callback, sh);
832         } else
833                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
834                         flags, tx, callback, sh);
835 }
836
837 static void ops_complete_check(void *stripe_head_ref)
838 {
839         struct stripe_head *sh = stripe_head_ref;
840         int pd_idx = sh->pd_idx;
841
842         pr_debug("%s: stripe %llu\n", __func__,
843                 (unsigned long long)sh->sector);
844
845         if (test_and_clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending) &&
846                 sh->ops.zero_sum_result == 0)
847                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
848
849         set_bit(STRIPE_OP_CHECK, &sh->ops.complete);
850         set_bit(STRIPE_HANDLE, &sh->state);
851         release_stripe(sh);
852 }
853
854 static void ops_run_check(struct stripe_head *sh)
855 {
856         /* kernel stack size limits the total number of disks */
857         int disks = sh->disks;
858         struct page *xor_srcs[disks];
859         struct dma_async_tx_descriptor *tx;
860
861         int count = 0, pd_idx = sh->pd_idx, i;
862         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
863
864         pr_debug("%s: stripe %llu\n", __func__,
865                 (unsigned long long)sh->sector);
866
867         for (i = disks; i--; ) {
868                 struct r5dev *dev = &sh->dev[i];
869                 if (i != pd_idx)
870                         xor_srcs[count++] = dev->page;
871         }
872
873         tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
874                 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
875
876         if (tx)
877                 set_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
878         else
879                 clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
880
881         atomic_inc(&sh->count);
882         tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
883                 ops_complete_check, sh);
884 }
885
886 static void raid5_run_ops(struct stripe_head *sh, unsigned long pending)
887 {
888         int overlap_clear = 0, i, disks = sh->disks;
889         struct dma_async_tx_descriptor *tx = NULL;
890
891         if (test_bit(STRIPE_OP_BIOFILL, &pending)) {
892                 ops_run_biofill(sh);
893                 overlap_clear++;
894         }
895
896         if (test_bit(STRIPE_OP_COMPUTE_BLK, &pending))
897                 tx = ops_run_compute5(sh, pending);
898
899         if (test_bit(STRIPE_OP_PREXOR, &pending))
900                 tx = ops_run_prexor(sh, tx);
901
902         if (test_bit(STRIPE_OP_BIODRAIN, &pending)) {
903                 tx = ops_run_biodrain(sh, tx, pending);
904                 overlap_clear++;
905         }
906
907         if (test_bit(STRIPE_OP_POSTXOR, &pending))
908                 ops_run_postxor(sh, tx, pending);
909
910         if (test_bit(STRIPE_OP_CHECK, &pending))
911                 ops_run_check(sh);
912
913         if (test_bit(STRIPE_OP_IO, &pending))
914                 ops_run_io(sh);
915
916         if (overlap_clear)
917                 for (i = disks; i--; ) {
918                         struct r5dev *dev = &sh->dev[i];
919                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
920                                 wake_up(&sh->raid_conf->wait_for_overlap);
921                 }
922 }
923
924 static int grow_one_stripe(raid5_conf_t *conf)
925 {
926         struct stripe_head *sh;
927         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
928         if (!sh)
929                 return 0;
930         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
931         sh->raid_conf = conf;
932         spin_lock_init(&sh->lock);
933
934         if (grow_buffers(sh, conf->raid_disks)) {
935                 shrink_buffers(sh, conf->raid_disks);
936                 kmem_cache_free(conf->slab_cache, sh);
937                 return 0;
938         }
939         sh->disks = conf->raid_disks;
940         /* we just created an active stripe so... */
941         atomic_set(&sh->count, 1);
942         atomic_inc(&conf->active_stripes);
943         INIT_LIST_HEAD(&sh->lru);
944         release_stripe(sh);
945         return 1;
946 }
947
948 static int grow_stripes(raid5_conf_t *conf, int num)
949 {
950         struct kmem_cache *sc;
951         int devs = conf->raid_disks;
952
953         sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
954         sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
955         conf->active_name = 0;
956         sc = kmem_cache_create(conf->cache_name[conf->active_name],
957                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
958                                0, 0, NULL);
959         if (!sc)
960                 return 1;
961         conf->slab_cache = sc;
962         conf->pool_size = devs;
963         while (num--)
964                 if (!grow_one_stripe(conf))
965                         return 1;
966         return 0;
967 }
968
969 #ifdef CONFIG_MD_RAID5_RESHAPE
970 static int resize_stripes(raid5_conf_t *conf, int newsize)
971 {
972         /* Make all the stripes able to hold 'newsize' devices.
973          * New slots in each stripe get 'page' set to a new page.
974          *
975          * This happens in stages:
976          * 1/ create a new kmem_cache and allocate the required number of
977          *    stripe_heads.
978          * 2/ gather all the old stripe_heads and tranfer the pages across
979          *    to the new stripe_heads.  This will have the side effect of
980          *    freezing the array as once all stripe_heads have been collected,
981          *    no IO will be possible.  Old stripe heads are freed once their
982          *    pages have been transferred over, and the old kmem_cache is
983          *    freed when all stripes are done.
984          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
985          *    we simple return a failre status - no need to clean anything up.
986          * 4/ allocate new pages for the new slots in the new stripe_heads.
987          *    If this fails, we don't bother trying the shrink the
988          *    stripe_heads down again, we just leave them as they are.
989          *    As each stripe_head is processed the new one is released into
990          *    active service.
991          *
992          * Once step2 is started, we cannot afford to wait for a write,
993          * so we use GFP_NOIO allocations.
994          */
995         struct stripe_head *osh, *nsh;
996         LIST_HEAD(newstripes);
997         struct disk_info *ndisks;
998         int err = 0;
999         struct kmem_cache *sc;
1000         int i;
1001
1002         if (newsize <= conf->pool_size)
1003                 return 0; /* never bother to shrink */
1004
1005         md_allow_write(conf->mddev);
1006
1007         /* Step 1 */
1008         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1009                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1010                                0, 0, NULL);
1011         if (!sc)
1012                 return -ENOMEM;
1013
1014         for (i = conf->max_nr_stripes; i; i--) {
1015                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1016                 if (!nsh)
1017                         break;
1018
1019                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1020
1021                 nsh->raid_conf = conf;
1022                 spin_lock_init(&nsh->lock);
1023
1024                 list_add(&nsh->lru, &newstripes);
1025         }
1026         if (i) {
1027                 /* didn't get enough, give up */
1028                 while (!list_empty(&newstripes)) {
1029                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1030                         list_del(&nsh->lru);
1031                         kmem_cache_free(sc, nsh);
1032                 }
1033                 kmem_cache_destroy(sc);
1034                 return -ENOMEM;
1035         }
1036         /* Step 2 - Must use GFP_NOIO now.
1037          * OK, we have enough stripes, start collecting inactive
1038          * stripes and copying them over
1039          */
1040         list_for_each_entry(nsh, &newstripes, lru) {
1041                 spin_lock_irq(&conf->device_lock);
1042                 wait_event_lock_irq(conf->wait_for_stripe,
1043                                     !list_empty(&conf->inactive_list),
1044                                     conf->device_lock,
1045                                     unplug_slaves(conf->mddev)
1046                         );
1047                 osh = get_free_stripe(conf);
1048                 spin_unlock_irq(&conf->device_lock);
1049                 atomic_set(&nsh->count, 1);
1050                 for(i=0; i<conf->pool_size; i++)
1051                         nsh->dev[i].page = osh->dev[i].page;
1052                 for( ; i<newsize; i++)
1053                         nsh->dev[i].page = NULL;
1054                 kmem_cache_free(conf->slab_cache, osh);
1055         }
1056         kmem_cache_destroy(conf->slab_cache);
1057
1058         /* Step 3.
1059          * At this point, we are holding all the stripes so the array
1060          * is completely stalled, so now is a good time to resize
1061          * conf->disks.
1062          */
1063         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1064         if (ndisks) {
1065                 for (i=0; i<conf->raid_disks; i++)
1066                         ndisks[i] = conf->disks[i];
1067                 kfree(conf->disks);
1068                 conf->disks = ndisks;
1069         } else
1070                 err = -ENOMEM;
1071
1072         /* Step 4, return new stripes to service */
1073         while(!list_empty(&newstripes)) {
1074                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1075                 list_del_init(&nsh->lru);
1076                 for (i=conf->raid_disks; i < newsize; i++)
1077                         if (nsh->dev[i].page == NULL) {
1078                                 struct page *p = alloc_page(GFP_NOIO);
1079                                 nsh->dev[i].page = p;
1080                                 if (!p)
1081                                         err = -ENOMEM;
1082                         }
1083                 release_stripe(nsh);
1084         }
1085         /* critical section pass, GFP_NOIO no longer needed */
1086
1087         conf->slab_cache = sc;
1088         conf->active_name = 1-conf->active_name;
1089         conf->pool_size = newsize;
1090         return err;
1091 }
1092 #endif
1093
1094 static int drop_one_stripe(raid5_conf_t *conf)
1095 {
1096         struct stripe_head *sh;
1097
1098         spin_lock_irq(&conf->device_lock);
1099         sh = get_free_stripe(conf);
1100         spin_unlock_irq(&conf->device_lock);
1101         if (!sh)
1102                 return 0;
1103         BUG_ON(atomic_read(&sh->count));
1104         shrink_buffers(sh, conf->pool_size);
1105         kmem_cache_free(conf->slab_cache, sh);
1106         atomic_dec(&conf->active_stripes);
1107         return 1;
1108 }
1109
1110 static void shrink_stripes(raid5_conf_t *conf)
1111 {
1112         while (drop_one_stripe(conf))
1113                 ;
1114
1115         if (conf->slab_cache)
1116                 kmem_cache_destroy(conf->slab_cache);
1117         conf->slab_cache = NULL;
1118 }
1119
1120 static void raid5_end_read_request(struct bio * bi, int error)
1121 {
1122         struct stripe_head *sh = bi->bi_private;
1123         raid5_conf_t *conf = sh->raid_conf;
1124         int disks = sh->disks, i;
1125         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1126         char b[BDEVNAME_SIZE];
1127         mdk_rdev_t *rdev;
1128
1129
1130         for (i=0 ; i<disks; i++)
1131                 if (bi == &sh->dev[i].req)
1132                         break;
1133
1134         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1135                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1136                 uptodate);
1137         if (i == disks) {
1138                 BUG();
1139                 return;
1140         }
1141
1142         if (uptodate) {
1143                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1144                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1145                         rdev = conf->disks[i].rdev;
1146                         printk(KERN_INFO "raid5:%s: read error corrected (%lu sectors at %llu on %s)\n",
1147                                mdname(conf->mddev), STRIPE_SECTORS,
1148                                (unsigned long long)(sh->sector + rdev->data_offset),
1149                                bdevname(rdev->bdev, b));
1150                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1151                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1152                 }
1153                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1154                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1155         } else {
1156                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1157                 int retry = 0;
1158                 rdev = conf->disks[i].rdev;
1159
1160                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1161                 atomic_inc(&rdev->read_errors);
1162                 if (conf->mddev->degraded)
1163                         printk(KERN_WARNING "raid5:%s: read error not correctable (sector %llu on %s).\n",
1164                                mdname(conf->mddev),
1165                                (unsigned long long)(sh->sector + rdev->data_offset),
1166                                bdn);
1167                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1168                         /* Oh, no!!! */
1169                         printk(KERN_WARNING "raid5:%s: read error NOT corrected!! (sector %llu on %s).\n",
1170                                mdname(conf->mddev),
1171                                (unsigned long long)(sh->sector + rdev->data_offset),
1172                                bdn);
1173                 else if (atomic_read(&rdev->read_errors)
1174                          > conf->max_nr_stripes)
1175                         printk(KERN_WARNING
1176                                "raid5:%s: Too many read errors, failing device %s.\n",
1177                                mdname(conf->mddev), bdn);
1178                 else
1179                         retry = 1;
1180                 if (retry)
1181                         set_bit(R5_ReadError, &sh->dev[i].flags);
1182                 else {
1183                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1184                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1185                         md_error(conf->mddev, rdev);
1186                 }
1187         }
1188         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1189         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1190         set_bit(STRIPE_HANDLE, &sh->state);
1191         release_stripe(sh);
1192 }
1193
1194 static void raid5_end_write_request (struct bio *bi, int error)
1195 {
1196         struct stripe_head *sh = bi->bi_private;
1197         raid5_conf_t *conf = sh->raid_conf;
1198         int disks = sh->disks, i;
1199         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1200
1201         for (i=0 ; i<disks; i++)
1202                 if (bi == &sh->dev[i].req)
1203                         break;
1204
1205         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1206                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1207                 uptodate);
1208         if (i == disks) {
1209                 BUG();
1210                 return;
1211         }
1212
1213         if (!uptodate)
1214                 md_error(conf->mddev, conf->disks[i].rdev);
1215
1216         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1217         
1218         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1219         set_bit(STRIPE_HANDLE, &sh->state);
1220         release_stripe(sh);
1221 }
1222
1223
1224 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1225         
1226 static void raid5_build_block (struct stripe_head *sh, int i)
1227 {
1228         struct r5dev *dev = &sh->dev[i];
1229
1230         bio_init(&dev->req);
1231         dev->req.bi_io_vec = &dev->vec;
1232         dev->req.bi_vcnt++;
1233         dev->req.bi_max_vecs++;
1234         dev->vec.bv_page = dev->page;
1235         dev->vec.bv_len = STRIPE_SIZE;
1236         dev->vec.bv_offset = 0;
1237
1238         dev->req.bi_sector = sh->sector;
1239         dev->req.bi_private = sh;
1240
1241         dev->flags = 0;
1242         dev->sector = compute_blocknr(sh, i);
1243 }
1244
1245 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1246 {
1247         char b[BDEVNAME_SIZE];
1248         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1249         pr_debug("raid5: error called\n");
1250
1251         if (!test_bit(Faulty, &rdev->flags)) {
1252                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1253                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1254                         unsigned long flags;
1255                         spin_lock_irqsave(&conf->device_lock, flags);
1256                         mddev->degraded++;
1257                         spin_unlock_irqrestore(&conf->device_lock, flags);
1258                         /*
1259                          * if recovery was running, make sure it aborts.
1260                          */
1261                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
1262                 }
1263                 set_bit(Faulty, &rdev->flags);
1264                 printk (KERN_ALERT
1265                         "raid5: Disk failure on %s, disabling device.\n"
1266                         "raid5: Operation continuing on %d devices.\n",
1267                         bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1268         }
1269 }
1270
1271 /*
1272  * Input: a 'big' sector number,
1273  * Output: index of the data and parity disk, and the sector # in them.
1274  */
1275 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1276                         unsigned int data_disks, unsigned int * dd_idx,
1277                         unsigned int * pd_idx, raid5_conf_t *conf)
1278 {
1279         long stripe;
1280         unsigned long chunk_number;
1281         unsigned int chunk_offset;
1282         sector_t new_sector;
1283         int sectors_per_chunk = conf->chunk_size >> 9;
1284
1285         /* First compute the information on this sector */
1286
1287         /*
1288          * Compute the chunk number and the sector offset inside the chunk
1289          */
1290         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1291         chunk_number = r_sector;
1292         BUG_ON(r_sector != chunk_number);
1293
1294         /*
1295          * Compute the stripe number
1296          */
1297         stripe = chunk_number / data_disks;
1298
1299         /*
1300          * Compute the data disk and parity disk indexes inside the stripe
1301          */
1302         *dd_idx = chunk_number % data_disks;
1303
1304         /*
1305          * Select the parity disk based on the user selected algorithm.
1306          */
1307         switch(conf->level) {
1308         case 4:
1309                 *pd_idx = data_disks;
1310                 break;
1311         case 5:
1312                 switch (conf->algorithm) {
1313                 case ALGORITHM_LEFT_ASYMMETRIC:
1314                         *pd_idx = data_disks - stripe % raid_disks;
1315                         if (*dd_idx >= *pd_idx)
1316                                 (*dd_idx)++;
1317                         break;
1318                 case ALGORITHM_RIGHT_ASYMMETRIC:
1319                         *pd_idx = stripe % raid_disks;
1320                         if (*dd_idx >= *pd_idx)
1321                                 (*dd_idx)++;
1322                         break;
1323                 case ALGORITHM_LEFT_SYMMETRIC:
1324                         *pd_idx = data_disks - stripe % raid_disks;
1325                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1326                         break;
1327                 case ALGORITHM_RIGHT_SYMMETRIC:
1328                         *pd_idx = stripe % raid_disks;
1329                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1330                         break;
1331                 default:
1332                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1333                                 conf->algorithm);
1334                 }
1335                 break;
1336         case 6:
1337
1338                 /**** FIX THIS ****/
1339                 switch (conf->algorithm) {
1340                 case ALGORITHM_LEFT_ASYMMETRIC:
1341                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1342                         if (*pd_idx == raid_disks-1)
1343                                 (*dd_idx)++;    /* Q D D D P */
1344                         else if (*dd_idx >= *pd_idx)
1345                                 (*dd_idx) += 2; /* D D P Q D */
1346                         break;
1347                 case ALGORITHM_RIGHT_ASYMMETRIC:
1348                         *pd_idx = stripe % raid_disks;
1349                         if (*pd_idx == raid_disks-1)
1350                                 (*dd_idx)++;    /* Q D D D P */
1351                         else if (*dd_idx >= *pd_idx)
1352                                 (*dd_idx) += 2; /* D D P Q D */
1353                         break;
1354                 case ALGORITHM_LEFT_SYMMETRIC:
1355                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1356                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1357                         break;
1358                 case ALGORITHM_RIGHT_SYMMETRIC:
1359                         *pd_idx = stripe % raid_disks;
1360                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1361                         break;
1362                 default:
1363                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1364                                 conf->algorithm);
1365                 }
1366                 break;
1367         }
1368
1369         /*
1370          * Finally, compute the new sector number
1371          */
1372         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1373         return new_sector;
1374 }
1375
1376
1377 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1378 {
1379         raid5_conf_t *conf = sh->raid_conf;
1380         int raid_disks = sh->disks;
1381         int data_disks = raid_disks - conf->max_degraded;
1382         sector_t new_sector = sh->sector, check;
1383         int sectors_per_chunk = conf->chunk_size >> 9;
1384         sector_t stripe;
1385         int chunk_offset;
1386         int chunk_number, dummy1, dummy2, dd_idx = i;
1387         sector_t r_sector;
1388
1389
1390         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1391         stripe = new_sector;
1392         BUG_ON(new_sector != stripe);
1393
1394         if (i == sh->pd_idx)
1395                 return 0;
1396         switch(conf->level) {
1397         case 4: break;
1398         case 5:
1399                 switch (conf->algorithm) {
1400                 case ALGORITHM_LEFT_ASYMMETRIC:
1401                 case ALGORITHM_RIGHT_ASYMMETRIC:
1402                         if (i > sh->pd_idx)
1403                                 i--;
1404                         break;
1405                 case ALGORITHM_LEFT_SYMMETRIC:
1406                 case ALGORITHM_RIGHT_SYMMETRIC:
1407                         if (i < sh->pd_idx)
1408                                 i += raid_disks;
1409                         i -= (sh->pd_idx + 1);
1410                         break;
1411                 default:
1412                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1413                                conf->algorithm);
1414                 }
1415                 break;
1416         case 6:
1417                 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1418                         return 0; /* It is the Q disk */
1419                 switch (conf->algorithm) {
1420                 case ALGORITHM_LEFT_ASYMMETRIC:
1421                 case ALGORITHM_RIGHT_ASYMMETRIC:
1422                         if (sh->pd_idx == raid_disks-1)
1423                                 i--;    /* Q D D D P */
1424                         else if (i > sh->pd_idx)
1425                                 i -= 2; /* D D P Q D */
1426                         break;
1427                 case ALGORITHM_LEFT_SYMMETRIC:
1428                 case ALGORITHM_RIGHT_SYMMETRIC:
1429                         if (sh->pd_idx == raid_disks-1)
1430                                 i--; /* Q D D D P */
1431                         else {
1432                                 /* D D P Q D */
1433                                 if (i < sh->pd_idx)
1434                                         i += raid_disks;
1435                                 i -= (sh->pd_idx + 2);
1436                         }
1437                         break;
1438                 default:
1439                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1440                                 conf->algorithm);
1441                 }
1442                 break;
1443         }
1444
1445         chunk_number = stripe * data_disks + i;
1446         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1447
1448         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1449         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1450                 printk(KERN_ERR "compute_blocknr: map not correct\n");
1451                 return 0;
1452         }
1453         return r_sector;
1454 }
1455
1456
1457
1458 /*
1459  * Copy data between a page in the stripe cache, and one or more bion
1460  * The page could align with the middle of the bio, or there could be
1461  * several bion, each with several bio_vecs, which cover part of the page
1462  * Multiple bion are linked together on bi_next.  There may be extras
1463  * at the end of this list.  We ignore them.
1464  */
1465 static void copy_data(int frombio, struct bio *bio,
1466                      struct page *page,
1467                      sector_t sector)
1468 {
1469         char *pa = page_address(page);
1470         struct bio_vec *bvl;
1471         int i;
1472         int page_offset;
1473
1474         if (bio->bi_sector >= sector)
1475                 page_offset = (signed)(bio->bi_sector - sector) * 512;
1476         else
1477                 page_offset = (signed)(sector - bio->bi_sector) * -512;
1478         bio_for_each_segment(bvl, bio, i) {
1479                 int len = bio_iovec_idx(bio,i)->bv_len;
1480                 int clen;
1481                 int b_offset = 0;
1482
1483                 if (page_offset < 0) {
1484                         b_offset = -page_offset;
1485                         page_offset += b_offset;
1486                         len -= b_offset;
1487                 }
1488
1489                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1490                         clen = STRIPE_SIZE - page_offset;
1491                 else clen = len;
1492
1493                 if (clen > 0) {
1494                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1495                         if (frombio)
1496                                 memcpy(pa+page_offset, ba+b_offset, clen);
1497                         else
1498                                 memcpy(ba+b_offset, pa+page_offset, clen);
1499                         __bio_kunmap_atomic(ba, KM_USER0);
1500                 }
1501                 if (clen < len) /* hit end of page */
1502                         break;
1503                 page_offset +=  len;
1504         }
1505 }
1506
1507 #define check_xor()     do {                                              \
1508                                 if (count == MAX_XOR_BLOCKS) {            \
1509                                 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1510                                 count = 0;                                \
1511                            }                                              \
1512                         } while(0)
1513
1514 static void compute_parity6(struct stripe_head *sh, int method)
1515 {
1516         raid6_conf_t *conf = sh->raid_conf;
1517         int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1518         struct bio *chosen;
1519         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1520         void *ptrs[disks];
1521
1522         qd_idx = raid6_next_disk(pd_idx, disks);
1523         d0_idx = raid6_next_disk(qd_idx, disks);
1524
1525         pr_debug("compute_parity, stripe %llu, method %d\n",
1526                 (unsigned long long)sh->sector, method);
1527
1528         switch(method) {
1529         case READ_MODIFY_WRITE:
1530                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
1531         case RECONSTRUCT_WRITE:
1532                 for (i= disks; i-- ;)
1533                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1534                                 chosen = sh->dev[i].towrite;
1535                                 sh->dev[i].towrite = NULL;
1536
1537                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1538                                         wake_up(&conf->wait_for_overlap);
1539
1540                                 BUG_ON(sh->dev[i].written);
1541                                 sh->dev[i].written = chosen;
1542                         }
1543                 break;
1544         case CHECK_PARITY:
1545                 BUG();          /* Not implemented yet */
1546         }
1547
1548         for (i = disks; i--;)
1549                 if (sh->dev[i].written) {
1550                         sector_t sector = sh->dev[i].sector;
1551                         struct bio *wbi = sh->dev[i].written;
1552                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1553                                 copy_data(1, wbi, sh->dev[i].page, sector);
1554                                 wbi = r5_next_bio(wbi, sector);
1555                         }
1556
1557                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1558                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1559                 }
1560
1561 //      switch(method) {
1562 //      case RECONSTRUCT_WRITE:
1563 //      case CHECK_PARITY:
1564 //      case UPDATE_PARITY:
1565                 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1566                 /* FIX: Is this ordering of drives even remotely optimal? */
1567                 count = 0;
1568                 i = d0_idx;
1569                 do {
1570                         ptrs[count++] = page_address(sh->dev[i].page);
1571                         if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1572                                 printk("block %d/%d not uptodate on parity calc\n", i,count);
1573                         i = raid6_next_disk(i, disks);
1574                 } while ( i != d0_idx );
1575 //              break;
1576 //      }
1577
1578         raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1579
1580         switch(method) {
1581         case RECONSTRUCT_WRITE:
1582                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1583                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1584                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1585                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
1586                 break;
1587         case UPDATE_PARITY:
1588                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1589                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1590                 break;
1591         }
1592 }
1593
1594
1595 /* Compute one missing block */
1596 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1597 {
1598         int i, count, disks = sh->disks;
1599         void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1600         int pd_idx = sh->pd_idx;
1601         int qd_idx = raid6_next_disk(pd_idx, disks);
1602
1603         pr_debug("compute_block_1, stripe %llu, idx %d\n",
1604                 (unsigned long long)sh->sector, dd_idx);
1605
1606         if ( dd_idx == qd_idx ) {
1607                 /* We're actually computing the Q drive */
1608                 compute_parity6(sh, UPDATE_PARITY);
1609         } else {
1610                 dest = page_address(sh->dev[dd_idx].page);
1611                 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1612                 count = 0;
1613                 for (i = disks ; i--; ) {
1614                         if (i == dd_idx || i == qd_idx)
1615                                 continue;
1616                         p = page_address(sh->dev[i].page);
1617                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1618                                 ptr[count++] = p;
1619                         else
1620                                 printk("compute_block() %d, stripe %llu, %d"
1621                                        " not present\n", dd_idx,
1622                                        (unsigned long long)sh->sector, i);
1623
1624                         check_xor();
1625                 }
1626                 if (count)
1627                         xor_blocks(count, STRIPE_SIZE, dest, ptr);
1628                 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1629                 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1630         }
1631 }
1632
1633 /* Compute two missing blocks */
1634 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1635 {
1636         int i, count, disks = sh->disks;
1637         int pd_idx = sh->pd_idx;
1638         int qd_idx = raid6_next_disk(pd_idx, disks);
1639         int d0_idx = raid6_next_disk(qd_idx, disks);
1640         int faila, failb;
1641
1642         /* faila and failb are disk numbers relative to d0_idx */
1643         /* pd_idx become disks-2 and qd_idx become disks-1 */
1644         faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1645         failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1646
1647         BUG_ON(faila == failb);
1648         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1649
1650         pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1651                (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1652
1653         if ( failb == disks-1 ) {
1654                 /* Q disk is one of the missing disks */
1655                 if ( faila == disks-2 ) {
1656                         /* Missing P+Q, just recompute */
1657                         compute_parity6(sh, UPDATE_PARITY);
1658                         return;
1659                 } else {
1660                         /* We're missing D+Q; recompute D from P */
1661                         compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1662                         compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1663                         return;
1664                 }
1665         }
1666
1667         /* We're missing D+P or D+D; build pointer table */
1668         {
1669                 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1670                 void *ptrs[disks];
1671
1672                 count = 0;
1673                 i = d0_idx;
1674                 do {
1675                         ptrs[count++] = page_address(sh->dev[i].page);
1676                         i = raid6_next_disk(i, disks);
1677                         if (i != dd_idx1 && i != dd_idx2 &&
1678                             !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1679                                 printk("compute_2 with missing block %d/%d\n", count, i);
1680                 } while ( i != d0_idx );
1681
1682                 if ( failb == disks-2 ) {
1683                         /* We're missing D+P. */
1684                         raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1685                 } else {
1686                         /* We're missing D+D. */
1687                         raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1688                 }
1689
1690                 /* Both the above update both missing blocks */
1691                 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1692                 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1693         }
1694 }
1695
1696 static int
1697 handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
1698 {
1699         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1700         int locked = 0;
1701
1702         if (rcw) {
1703                 /* if we are not expanding this is a proper write request, and
1704                  * there will be bios with new data to be drained into the
1705                  * stripe cache
1706                  */
1707                 if (!expand) {
1708                         set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1709                         sh->ops.count++;
1710                 }
1711
1712                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1713                 sh->ops.count++;
1714
1715                 for (i = disks; i--; ) {
1716                         struct r5dev *dev = &sh->dev[i];
1717
1718                         if (dev->towrite) {
1719                                 set_bit(R5_LOCKED, &dev->flags);
1720                                 if (!expand)
1721                                         clear_bit(R5_UPTODATE, &dev->flags);
1722                                 locked++;
1723                         }
1724                 }
1725                 if (locked + 1 == disks)
1726                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1727                                 atomic_inc(&sh->raid_conf->pending_full_writes);
1728         } else {
1729                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1730                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1731
1732                 set_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
1733                 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1734                 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1735
1736                 sh->ops.count += 3;
1737
1738                 for (i = disks; i--; ) {
1739                         struct r5dev *dev = &sh->dev[i];
1740                         if (i == pd_idx)
1741                                 continue;
1742
1743                         /* For a read-modify write there may be blocks that are
1744                          * locked for reading while others are ready to be
1745                          * written so we distinguish these blocks by the
1746                          * R5_Wantprexor bit
1747                          */
1748                         if (dev->towrite &&
1749                             (test_bit(R5_UPTODATE, &dev->flags) ||
1750                             test_bit(R5_Wantcompute, &dev->flags))) {
1751                                 set_bit(R5_Wantprexor, &dev->flags);
1752                                 set_bit(R5_LOCKED, &dev->flags);
1753                                 clear_bit(R5_UPTODATE, &dev->flags);
1754                                 locked++;
1755                         }
1756                 }
1757         }
1758
1759         /* keep the parity disk locked while asynchronous operations
1760          * are in flight
1761          */
1762         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1763         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1764         locked++;
1765
1766         pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1767                 __func__, (unsigned long long)sh->sector,
1768                 locked, sh->ops.pending);
1769
1770         return locked;
1771 }
1772
1773 /*
1774  * Each stripe/dev can have one or more bion attached.
1775  * toread/towrite point to the first in a chain.
1776  * The bi_next chain must be in order.
1777  */
1778 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1779 {
1780         struct bio **bip;
1781         raid5_conf_t *conf = sh->raid_conf;
1782         int firstwrite=0;
1783
1784         pr_debug("adding bh b#%llu to stripe s#%llu\n",
1785                 (unsigned long long)bi->bi_sector,
1786                 (unsigned long long)sh->sector);
1787
1788
1789         spin_lock(&sh->lock);
1790         spin_lock_irq(&conf->device_lock);
1791         if (forwrite) {
1792                 bip = &sh->dev[dd_idx].towrite;
1793                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1794                         firstwrite = 1;
1795         } else
1796                 bip = &sh->dev[dd_idx].toread;
1797         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1798                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1799                         goto overlap;
1800                 bip = & (*bip)->bi_next;
1801         }
1802         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1803                 goto overlap;
1804
1805         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1806         if (*bip)
1807                 bi->bi_next = *bip;
1808         *bip = bi;
1809         bi->bi_phys_segments ++;
1810         spin_unlock_irq(&conf->device_lock);
1811         spin_unlock(&sh->lock);
1812
1813         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1814                 (unsigned long long)bi->bi_sector,
1815                 (unsigned long long)sh->sector, dd_idx);
1816
1817         if (conf->mddev->bitmap && firstwrite) {
1818                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1819                                   STRIPE_SECTORS, 0);
1820                 sh->bm_seq = conf->seq_flush+1;
1821                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1822         }
1823
1824         if (forwrite) {
1825                 /* check if page is covered */
1826                 sector_t sector = sh->dev[dd_idx].sector;
1827                 for (bi=sh->dev[dd_idx].towrite;
1828                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1829                              bi && bi->bi_sector <= sector;
1830                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1831                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1832                                 sector = bi->bi_sector + (bi->bi_size>>9);
1833                 }
1834                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1835                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1836         }
1837         return 1;
1838
1839  overlap:
1840         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1841         spin_unlock_irq(&conf->device_lock);
1842         spin_unlock(&sh->lock);
1843         return 0;
1844 }
1845
1846 static void end_reshape(raid5_conf_t *conf);
1847
1848 static int page_is_zero(struct page *p)
1849 {
1850         char *a = page_address(p);
1851         return ((*(u32*)a) == 0 &&
1852                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1853 }
1854
1855 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1856 {
1857         int sectors_per_chunk = conf->chunk_size >> 9;
1858         int pd_idx, dd_idx;
1859         int chunk_offset = sector_div(stripe, sectors_per_chunk);
1860
1861         raid5_compute_sector(stripe * (disks - conf->max_degraded)
1862                              *sectors_per_chunk + chunk_offset,
1863                              disks, disks - conf->max_degraded,
1864                              &dd_idx, &pd_idx, conf);
1865         return pd_idx;
1866 }
1867
1868 static void
1869 handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1870                                 struct stripe_head_state *s, int disks,
1871                                 struct bio **return_bi)
1872 {
1873         int i;
1874         for (i = disks; i--; ) {
1875                 struct bio *bi;
1876                 int bitmap_end = 0;
1877
1878                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1879                         mdk_rdev_t *rdev;
1880                         rcu_read_lock();
1881                         rdev = rcu_dereference(conf->disks[i].rdev);
1882                         if (rdev && test_bit(In_sync, &rdev->flags))
1883                                 /* multiple read failures in one stripe */
1884                                 md_error(conf->mddev, rdev);
1885                         rcu_read_unlock();
1886                 }
1887                 spin_lock_irq(&conf->device_lock);
1888                 /* fail all writes first */
1889                 bi = sh->dev[i].towrite;
1890                 sh->dev[i].towrite = NULL;
1891                 if (bi) {
1892                         s->to_write--;
1893                         bitmap_end = 1;
1894                 }
1895
1896                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1897                         wake_up(&conf->wait_for_overlap);
1898
1899                 while (bi && bi->bi_sector <
1900                         sh->dev[i].sector + STRIPE_SECTORS) {
1901                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1902                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1903                         if (--bi->bi_phys_segments == 0) {
1904                                 md_write_end(conf->mddev);
1905                                 bi->bi_next = *return_bi;
1906                                 *return_bi = bi;
1907                         }
1908                         bi = nextbi;
1909                 }
1910                 /* and fail all 'written' */
1911                 bi = sh->dev[i].written;
1912                 sh->dev[i].written = NULL;
1913                 if (bi) bitmap_end = 1;
1914                 while (bi && bi->bi_sector <
1915                        sh->dev[i].sector + STRIPE_SECTORS) {
1916                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1917                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1918                         if (--bi->bi_phys_segments == 0) {
1919                                 md_write_end(conf->mddev);
1920                                 bi->bi_next = *return_bi;
1921                                 *return_bi = bi;
1922                         }
1923                         bi = bi2;
1924                 }
1925
1926                 /* fail any reads if this device is non-operational and
1927                  * the data has not reached the cache yet.
1928                  */
1929                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
1930                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1931                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
1932                         bi = sh->dev[i].toread;
1933                         sh->dev[i].toread = NULL;
1934                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1935                                 wake_up(&conf->wait_for_overlap);
1936                         if (bi) s->to_read--;
1937                         while (bi && bi->bi_sector <
1938                                sh->dev[i].sector + STRIPE_SECTORS) {
1939                                 struct bio *nextbi =
1940                                         r5_next_bio(bi, sh->dev[i].sector);
1941                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1942                                 if (--bi->bi_phys_segments == 0) {
1943                                         bi->bi_next = *return_bi;
1944                                         *return_bi = bi;
1945                                 }
1946                                 bi = nextbi;
1947                         }
1948                 }
1949                 spin_unlock_irq(&conf->device_lock);
1950                 if (bitmap_end)
1951                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1952                                         STRIPE_SECTORS, 0, 0);
1953         }
1954
1955         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
1956                 if (atomic_dec_and_test(&conf->pending_full_writes))
1957                         md_wakeup_thread(conf->mddev->thread);
1958 }
1959
1960 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
1961  * to process
1962  */
1963 static int __handle_issuing_new_read_requests5(struct stripe_head *sh,
1964                         struct stripe_head_state *s, int disk_idx, int disks)
1965 {
1966         struct r5dev *dev = &sh->dev[disk_idx];
1967         struct r5dev *failed_dev = &sh->dev[s->failed_num];
1968
1969         /* don't schedule compute operations or reads on the parity block while
1970          * a check is in flight
1971          */
1972         if ((disk_idx == sh->pd_idx) &&
1973              test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
1974                 return ~0;
1975
1976         /* is the data in this block needed, and can we get it? */
1977         if (!test_bit(R5_LOCKED, &dev->flags) &&
1978             !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread ||
1979             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1980              s->syncing || s->expanding || (s->failed &&
1981              (failed_dev->toread || (failed_dev->towrite &&
1982              !test_bit(R5_OVERWRITE, &failed_dev->flags)
1983              ))))) {
1984                 /* 1/ We would like to get this block, possibly by computing it,
1985                  * but we might not be able to.
1986                  *
1987                  * 2/ Since parity check operations potentially make the parity
1988                  * block !uptodate it will need to be refreshed before any
1989                  * compute operations on data disks are scheduled.
1990                  *
1991                  * 3/ We hold off parity block re-reads until check operations
1992                  * have quiesced.
1993                  */
1994                 if ((s->uptodate == disks - 1) &&
1995                     !test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
1996                         set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
1997                         set_bit(R5_Wantcompute, &dev->flags);
1998                         sh->ops.target = disk_idx;
1999                         s->req_compute = 1;
2000                         sh->ops.count++;
2001                         /* Careful: from this point on 'uptodate' is in the eye
2002                          * of raid5_run_ops which services 'compute' operations
2003                          * before writes. R5_Wantcompute flags a block that will
2004                          * be R5_UPTODATE by the time it is needed for a
2005                          * subsequent operation.
2006                          */
2007                         s->uptodate++;
2008                         return 0; /* uptodate + compute == disks */
2009                 } else if ((s->uptodate < disks - 1) &&
2010                         test_bit(R5_Insync, &dev->flags)) {
2011                         /* Note: we hold off compute operations while checks are
2012                          * in flight, but we still prefer 'compute' over 'read'
2013                          * hence we only read if (uptodate < * disks-1)
2014                          */
2015                         set_bit(R5_LOCKED, &dev->flags);
2016                         set_bit(R5_Wantread, &dev->flags);
2017                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2018                                 sh->ops.count++;
2019                         s->locked++;
2020                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2021                                 s->syncing);
2022                 }
2023         }
2024
2025         return ~0;
2026 }
2027
2028 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2029                         struct stripe_head_state *s, int disks)
2030 {
2031         int i;
2032
2033         /* Clear completed compute operations.  Parity recovery
2034          * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled
2035          * later on in this routine
2036          */
2037         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2038                 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2039                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2040                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2041                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2042         }
2043
2044         /* look for blocks to read/compute, skip this if a compute
2045          * is already in flight, or if the stripe contents are in the
2046          * midst of changing due to a write
2047          */
2048         if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2049                 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2050                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2051                 for (i = disks; i--; )
2052                         if (__handle_issuing_new_read_requests5(
2053                                 sh, s, i, disks) == 0)
2054                                 break;
2055         }
2056         set_bit(STRIPE_HANDLE, &sh->state);
2057 }
2058
2059 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2060                         struct stripe_head_state *s, struct r6_state *r6s,
2061                         int disks)
2062 {
2063         int i;
2064         for (i = disks; i--; ) {
2065                 struct r5dev *dev = &sh->dev[i];
2066                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2067                     !test_bit(R5_UPTODATE, &dev->flags) &&
2068                     (dev->toread || (dev->towrite &&
2069                      !test_bit(R5_OVERWRITE, &dev->flags)) ||
2070                      s->syncing || s->expanding ||
2071                      (s->failed >= 1 &&
2072                       (sh->dev[r6s->failed_num[0]].toread ||
2073                        s->to_write)) ||
2074                      (s->failed >= 2 &&
2075                       (sh->dev[r6s->failed_num[1]].toread ||
2076                        s->to_write)))) {
2077                         /* we would like to get this block, possibly
2078                          * by computing it, but we might not be able to
2079                          */
2080                         if (s->uptodate == disks-1) {
2081                                 pr_debug("Computing stripe %llu block %d\n",
2082                                        (unsigned long long)sh->sector, i);
2083                                 compute_block_1(sh, i, 0);
2084                                 s->uptodate++;
2085                         } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2086                                 /* Computing 2-failure is *very* expensive; only
2087                                  * do it if failed >= 2
2088                                  */
2089                                 int other;
2090                                 for (other = disks; other--; ) {
2091                                         if (other == i)
2092                                                 continue;
2093                                         if (!test_bit(R5_UPTODATE,
2094                                               &sh->dev[other].flags))
2095                                                 break;
2096                                 }
2097                                 BUG_ON(other < 0);
2098                                 pr_debug("Computing stripe %llu blocks %d,%d\n",
2099                                        (unsigned long long)sh->sector,
2100                                        i, other);
2101                                 compute_block_2(sh, i, other);
2102                                 s->uptodate += 2;
2103                         } else if (test_bit(R5_Insync, &dev->flags)) {
2104                                 set_bit(R5_LOCKED, &dev->flags);
2105                                 set_bit(R5_Wantread, &dev->flags);
2106                                 s->locked++;
2107                                 pr_debug("Reading block %d (sync=%d)\n",
2108                                         i, s->syncing);
2109                         }
2110                 }
2111         }
2112         set_bit(STRIPE_HANDLE, &sh->state);
2113 }
2114
2115
2116 /* handle_completed_write_requests
2117  * any written block on an uptodate or failed drive can be returned.
2118  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2119  * never LOCKED, so we don't need to test 'failed' directly.
2120  */
2121 static void handle_completed_write_requests(raid5_conf_t *conf,
2122         struct stripe_head *sh, int disks, struct bio **return_bi)
2123 {
2124         int i;
2125         struct r5dev *dev;
2126
2127         for (i = disks; i--; )
2128                 if (sh->dev[i].written) {
2129                         dev = &sh->dev[i];
2130                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2131                                 test_bit(R5_UPTODATE, &dev->flags)) {
2132                                 /* We can return any write requests */
2133                                 struct bio *wbi, *wbi2;
2134                                 int bitmap_end = 0;
2135                                 pr_debug("Return write for disc %d\n", i);
2136                                 spin_lock_irq(&conf->device_lock);
2137                                 wbi = dev->written;
2138                                 dev->written = NULL;
2139                                 while (wbi && wbi->bi_sector <
2140                                         dev->sector + STRIPE_SECTORS) {
2141                                         wbi2 = r5_next_bio(wbi, dev->sector);
2142                                         if (--wbi->bi_phys_segments == 0) {
2143                                                 md_write_end(conf->mddev);
2144                                                 wbi->bi_next = *return_bi;
2145                                                 *return_bi = wbi;
2146                                         }
2147                                         wbi = wbi2;
2148                                 }
2149                                 if (dev->towrite == NULL)
2150                                         bitmap_end = 1;
2151                                 spin_unlock_irq(&conf->device_lock);
2152                                 if (bitmap_end)
2153                                         bitmap_endwrite(conf->mddev->bitmap,
2154                                                         sh->sector,
2155                                                         STRIPE_SECTORS,
2156                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2157                                                         0);
2158                         }
2159                 }
2160
2161         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2162                 if (atomic_dec_and_test(&conf->pending_full_writes))
2163                         md_wakeup_thread(conf->mddev->thread);
2164 }
2165
2166 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2167                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2168 {
2169         int rmw = 0, rcw = 0, i;
2170         for (i = disks; i--; ) {
2171                 /* would I have to read this buffer for read_modify_write */
2172                 struct r5dev *dev = &sh->dev[i];
2173                 if ((dev->towrite || i == sh->pd_idx) &&
2174                     !test_bit(R5_LOCKED, &dev->flags) &&
2175                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2176                       test_bit(R5_Wantcompute, &dev->flags))) {
2177                         if (test_bit(R5_Insync, &dev->flags))
2178                                 rmw++;
2179                         else
2180                                 rmw += 2*disks;  /* cannot read it */
2181                 }
2182                 /* Would I have to read this buffer for reconstruct_write */
2183                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2184                     !test_bit(R5_LOCKED, &dev->flags) &&
2185                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2186                     test_bit(R5_Wantcompute, &dev->flags))) {
2187                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2188                         else
2189                                 rcw += 2*disks;
2190                 }
2191         }
2192         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2193                 (unsigned long long)sh->sector, rmw, rcw);
2194         set_bit(STRIPE_HANDLE, &sh->state);
2195         if (rmw < rcw && rmw > 0)
2196                 /* prefer read-modify-write, but need to get some data */
2197                 for (i = disks; i--; ) {
2198                         struct r5dev *dev = &sh->dev[i];
2199                         if ((dev->towrite || i == sh->pd_idx) &&
2200                             !test_bit(R5_LOCKED, &dev->flags) &&
2201                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2202                             test_bit(R5_Wantcompute, &dev->flags)) &&
2203                             test_bit(R5_Insync, &dev->flags)) {
2204                                 if (
2205                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2206                                         pr_debug("Read_old block "
2207                                                 "%d for r-m-w\n", i);
2208                                         set_bit(R5_LOCKED, &dev->flags);
2209                                         set_bit(R5_Wantread, &dev->flags);
2210                                         if (!test_and_set_bit(
2211                                                 STRIPE_OP_IO, &sh->ops.pending))
2212                                                 sh->ops.count++;
2213                                         s->locked++;
2214                                 } else {
2215                                         set_bit(STRIPE_DELAYED, &sh->state);
2216                                         set_bit(STRIPE_HANDLE, &sh->state);
2217                                 }
2218                         }
2219                 }
2220         if (rcw <= rmw && rcw > 0)
2221                 /* want reconstruct write, but need to get some data */
2222                 for (i = disks; i--; ) {
2223                         struct r5dev *dev = &sh->dev[i];
2224                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2225                             i != sh->pd_idx &&
2226                             !test_bit(R5_LOCKED, &dev->flags) &&
2227                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2228                             test_bit(R5_Wantcompute, &dev->flags)) &&
2229                             test_bit(R5_Insync, &dev->flags)) {
2230                                 if (
2231                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2232                                         pr_debug("Read_old block "
2233                                                 "%d for Reconstruct\n", i);
2234                                         set_bit(R5_LOCKED, &dev->flags);
2235                                         set_bit(R5_Wantread, &dev->flags);
2236                                         if (!test_and_set_bit(
2237                                                 STRIPE_OP_IO, &sh->ops.pending))
2238                                                 sh->ops.count++;
2239                                         s->locked++;
2240                                 } else {
2241                                         set_bit(STRIPE_DELAYED, &sh->state);
2242                                         set_bit(STRIPE_HANDLE, &sh->state);
2243                                 }
2244                         }
2245                 }
2246         /* now if nothing is locked, and if we have enough data,
2247          * we can start a write request
2248          */
2249         /* since handle_stripe can be called at any time we need to handle the
2250          * case where a compute block operation has been submitted and then a
2251          * subsequent call wants to start a write request.  raid5_run_ops only
2252          * handles the case where compute block and postxor are requested
2253          * simultaneously.  If this is not the case then new writes need to be
2254          * held off until the compute completes.
2255          */
2256         if ((s->req_compute ||
2257             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) &&
2258                 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2259                 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2260                 s->locked += handle_write_operations5(sh, rcw == 0, 0);
2261 }
2262
2263 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2264                 struct stripe_head *sh, struct stripe_head_state *s,
2265                 struct r6_state *r6s, int disks)
2266 {
2267         int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2268         int qd_idx = r6s->qd_idx;
2269         for (i = disks; i--; ) {
2270                 struct r5dev *dev = &sh->dev[i];
2271                 /* Would I have to read this buffer for reconstruct_write */
2272                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2273                     && i != pd_idx && i != qd_idx
2274                     && (!test_bit(R5_LOCKED, &dev->flags)
2275                             ) &&
2276                     !test_bit(R5_UPTODATE, &dev->flags)) {
2277                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2278                         else {
2279                                 pr_debug("raid6: must_compute: "
2280                                         "disk %d flags=%#lx\n", i, dev->flags);
2281                                 must_compute++;
2282                         }
2283                 }
2284         }
2285         pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2286                (unsigned long long)sh->sector, rcw, must_compute);
2287         set_bit(STRIPE_HANDLE, &sh->state);
2288
2289         if (rcw > 0)
2290                 /* want reconstruct write, but need to get some data */
2291                 for (i = disks; i--; ) {
2292                         struct r5dev *dev = &sh->dev[i];
2293                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2294                             && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2295                             && !test_bit(R5_LOCKED, &dev->flags) &&
2296                             !test_bit(R5_UPTODATE, &dev->flags) &&
2297                             test_bit(R5_Insync, &dev->flags)) {
2298                                 if (
2299                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2300                                         pr_debug("Read_old stripe %llu "
2301                                                 "block %d for Reconstruct\n",
2302                                              (unsigned long long)sh->sector, i);
2303                                         set_bit(R5_LOCKED, &dev->flags);
2304                                         set_bit(R5_Wantread, &dev->flags);
2305                                         s->locked++;
2306                                 } else {
2307                                         pr_debug("Request delayed stripe %llu "
2308                                                 "block %d for Reconstruct\n",
2309                                              (unsigned long long)sh->sector, i);
2310                                         set_bit(STRIPE_DELAYED, &sh->state);
2311                                         set_bit(STRIPE_HANDLE, &sh->state);
2312                                 }
2313                         }
2314                 }
2315         /* now if nothing is locked, and if we have enough data, we can start a
2316          * write request
2317          */
2318         if (s->locked == 0 && rcw == 0 &&
2319             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2320                 if (must_compute > 0) {
2321                         /* We have failed blocks and need to compute them */
2322                         switch (s->failed) {
2323                         case 0:
2324                                 BUG();
2325                         case 1:
2326                                 compute_block_1(sh, r6s->failed_num[0], 0);
2327                                 break;
2328                         case 2:
2329                                 compute_block_2(sh, r6s->failed_num[0],
2330                                                 r6s->failed_num[1]);
2331                                 break;
2332                         default: /* This request should have been failed? */
2333                                 BUG();
2334                         }
2335                 }
2336
2337                 pr_debug("Computing parity for stripe %llu\n",
2338                         (unsigned long long)sh->sector);
2339                 compute_parity6(sh, RECONSTRUCT_WRITE);
2340                 /* now every locked buffer is ready to be written */
2341                 for (i = disks; i--; )
2342                         if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2343                                 pr_debug("Writing stripe %llu block %d\n",
2344                                        (unsigned long long)sh->sector, i);
2345                                 s->locked++;
2346                                 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2347                         }
2348                 if (s->locked == disks)
2349                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2350                                 atomic_inc(&conf->pending_full_writes);
2351                 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2352                 set_bit(STRIPE_INSYNC, &sh->state);
2353
2354                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2355                         atomic_dec(&conf->preread_active_stripes);
2356                         if (atomic_read(&conf->preread_active_stripes) <
2357                             IO_THRESHOLD)
2358                                 md_wakeup_thread(conf->mddev->thread);
2359                 }
2360         }
2361 }
2362
2363 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2364                                 struct stripe_head_state *s, int disks)
2365 {
2366         int canceled_check = 0;
2367
2368         set_bit(STRIPE_HANDLE, &sh->state);
2369
2370         /* complete a check operation */
2371         if (test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) {
2372             clear_bit(STRIPE_OP_CHECK, &sh->ops.ack);
2373             clear_bit(STRIPE_OP_CHECK, &sh->ops.pending);
2374                 if (s->failed == 0) {
2375                         if (sh->ops.zero_sum_result == 0)
2376                                 /* parity is correct (on disc,
2377                                  * not in buffer any more)
2378                                  */
2379                                 set_bit(STRIPE_INSYNC, &sh->state);
2380                         else {
2381                                 conf->mddev->resync_mismatches +=
2382                                         STRIPE_SECTORS;
2383                                 if (test_bit(
2384                                      MD_RECOVERY_CHECK, &conf->mddev->recovery))
2385                                         /* don't try to repair!! */
2386                                         set_bit(STRIPE_INSYNC, &sh->state);
2387                                 else {
2388                                         set_bit(STRIPE_OP_COMPUTE_BLK,
2389                                                 &sh->ops.pending);
2390                                         set_bit(STRIPE_OP_MOD_REPAIR_PD,
2391                                                 &sh->ops.pending);
2392                                         set_bit(R5_Wantcompute,
2393                                                 &sh->dev[sh->pd_idx].flags);
2394                                         sh->ops.target = sh->pd_idx;
2395                                         sh->ops.count++;
2396                                         s->uptodate++;
2397                                 }
2398                         }
2399                 } else
2400                         canceled_check = 1; /* STRIPE_INSYNC is not set */
2401         }
2402
2403         /* check if we can clear a parity disk reconstruct */
2404         if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2405                 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2406
2407                 clear_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending);
2408                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2409                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2410                 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2411         }
2412
2413         /* start a new check operation if there are no failures, the stripe is
2414          * not insync, and a repair is not in flight
2415          */
2416         if (s->failed == 0 &&
2417             !test_bit(STRIPE_INSYNC, &sh->state) &&
2418             !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2419                 if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2420                         BUG_ON(s->uptodate != disks);
2421                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2422                         sh->ops.count++;
2423                         s->uptodate--;
2424                 }
2425         }
2426
2427         /* Wait for check parity and compute block operations to complete
2428          * before write-back.  If a failure occurred while the check operation
2429          * was in flight we need to cycle this stripe through handle_stripe
2430          * since the parity block may not be uptodate
2431          */
2432         if (!canceled_check && !test_bit(STRIPE_INSYNC, &sh->state) &&
2433             !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) &&
2434             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) {
2435                 struct r5dev *dev;
2436                 /* either failed parity check, or recovery is happening */
2437                 if (s->failed == 0)
2438                         s->failed_num = sh->pd_idx;
2439                 dev = &sh->dev[s->failed_num];
2440                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2441                 BUG_ON(s->uptodate != disks);
2442
2443                 set_bit(R5_LOCKED, &dev->flags);
2444                 set_bit(R5_Wantwrite, &dev->flags);
2445                 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2446                         sh->ops.count++;
2447
2448                 clear_bit(STRIPE_DEGRADED, &sh->state);
2449                 s->locked++;
2450                 set_bit(STRIPE_INSYNC, &sh->state);
2451         }
2452 }
2453
2454
2455 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2456                                 struct stripe_head_state *s,
2457                                 struct r6_state *r6s, struct page *tmp_page,
2458                                 int disks)
2459 {
2460         int update_p = 0, update_q = 0;
2461         struct r5dev *dev;
2462         int pd_idx = sh->pd_idx;
2463         int qd_idx = r6s->qd_idx;
2464
2465         set_bit(STRIPE_HANDLE, &sh->state);
2466
2467         BUG_ON(s->failed > 2);
2468         BUG_ON(s->uptodate < disks);
2469         /* Want to check and possibly repair P and Q.
2470          * However there could be one 'failed' device, in which
2471          * case we can only check one of them, possibly using the
2472          * other to generate missing data
2473          */
2474
2475         /* If !tmp_page, we cannot do the calculations,
2476          * but as we have set STRIPE_HANDLE, we will soon be called
2477          * by stripe_handle with a tmp_page - just wait until then.
2478          */
2479         if (tmp_page) {
2480                 if (s->failed == r6s->q_failed) {
2481                         /* The only possible failed device holds 'Q', so it
2482                          * makes sense to check P (If anything else were failed,
2483                          * we would have used P to recreate it).
2484                          */
2485                         compute_block_1(sh, pd_idx, 1);
2486                         if (!page_is_zero(sh->dev[pd_idx].page)) {
2487                                 compute_block_1(sh, pd_idx, 0);
2488                                 update_p = 1;
2489                         }
2490                 }
2491                 if (!r6s->q_failed && s->failed < 2) {
2492                         /* q is not failed, and we didn't use it to generate
2493                          * anything, so it makes sense to check it
2494                          */
2495                         memcpy(page_address(tmp_page),
2496                                page_address(sh->dev[qd_idx].page),
2497                                STRIPE_SIZE);
2498                         compute_parity6(sh, UPDATE_PARITY);
2499                         if (memcmp(page_address(tmp_page),
2500                                    page_address(sh->dev[qd_idx].page),
2501                                    STRIPE_SIZE) != 0) {
2502                                 clear_bit(STRIPE_INSYNC, &sh->state);
2503                                 update_q = 1;
2504                         }
2505                 }
2506                 if (update_p || update_q) {
2507                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2508                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2509                                 /* don't try to repair!! */
2510                                 update_p = update_q = 0;
2511                 }
2512
2513                 /* now write out any block on a failed drive,
2514                  * or P or Q if they need it
2515                  */
2516
2517                 if (s->failed == 2) {
2518                         dev = &sh->dev[r6s->failed_num[1]];
2519                         s->locked++;
2520                         set_bit(R5_LOCKED, &dev->flags);
2521                         set_bit(R5_Wantwrite, &dev->flags);
2522                 }
2523                 if (s->failed >= 1) {
2524                         dev = &sh->dev[r6s->failed_num[0]];
2525                         s->locked++;
2526                         set_bit(R5_LOCKED, &dev->flags);
2527                         set_bit(R5_Wantwrite, &dev->flags);
2528                 }
2529
2530                 if (update_p) {
2531                         dev = &sh->dev[pd_idx];
2532                         s->locked++;
2533                         set_bit(R5_LOCKED, &dev->flags);
2534                         set_bit(R5_Wantwrite, &dev->flags);
2535                 }
2536                 if (update_q) {
2537                         dev = &sh->dev[qd_idx];
2538                         s->locked++;
2539                         set_bit(R5_LOCKED, &dev->flags);
2540                         set_bit(R5_Wantwrite, &dev->flags);
2541                 }
2542                 clear_bit(STRIPE_DEGRADED, &sh->state);
2543
2544                 set_bit(STRIPE_INSYNC, &sh->state);
2545         }
2546 }
2547
2548 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2549                                 struct r6_state *r6s)
2550 {
2551         int i;
2552
2553         /* We have read all the blocks in this stripe and now we need to
2554          * copy some of them into a target stripe for expand.
2555          */
2556         struct dma_async_tx_descriptor *tx = NULL;
2557         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2558         for (i = 0; i < sh->disks; i++)
2559                 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2560                         int dd_idx, pd_idx, j;
2561                         struct stripe_head *sh2;
2562
2563                         sector_t bn = compute_blocknr(sh, i);
2564                         sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2565                                                 conf->raid_disks -
2566                                                 conf->max_degraded, &dd_idx,
2567                                                 &pd_idx, conf);
2568                         sh2 = get_active_stripe(conf, s, conf->raid_disks,
2569                                                 pd_idx, 1);
2570                         if (sh2 == NULL)
2571                                 /* so far only the early blocks of this stripe
2572                                  * have been requested.  When later blocks
2573                                  * get requested, we will try again
2574                                  */
2575                                 continue;
2576                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2577                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2578                                 /* must have already done this block */
2579                                 release_stripe(sh2);
2580                                 continue;
2581                         }
2582
2583                         /* place all the copies on one channel */
2584                         tx = async_memcpy(sh2->dev[dd_idx].page,
2585                                 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2586                                 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2587
2588                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2589                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2590                         for (j = 0; j < conf->raid_disks; j++)
2591                                 if (j != sh2->pd_idx &&
2592                                     (!r6s || j != raid6_next_disk(sh2->pd_idx,
2593                                                                  sh2->disks)) &&
2594                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2595                                         break;
2596                         if (j == conf->raid_disks) {
2597                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2598                                 set_bit(STRIPE_HANDLE, &sh2->state);
2599                         }
2600                         release_stripe(sh2);
2601
2602                 }
2603         /* done submitting copies, wait for them to complete */
2604         if (tx) {
2605                 async_tx_ack(tx);
2606                 dma_wait_for_async_tx(tx);
2607         }
2608 }
2609
2610 /*
2611  * handle_stripe - do things to a stripe.
2612  *
2613  * We lock the stripe and then examine the state of various bits
2614  * to see what needs to be done.
2615  * Possible results:
2616  *    return some read request which now have data
2617  *    return some write requests which are safely on disc
2618  *    schedule a read on some buffers
2619  *    schedule a write of some buffers
2620  *    return confirmation of parity correctness
2621  *
2622  * buffers are taken off read_list or write_list, and bh_cache buffers
2623  * get BH_Lock set before the stripe lock is released.
2624  *
2625  */
2626
2627 static void handle_stripe5(struct stripe_head *sh)
2628 {
2629         raid5_conf_t *conf = sh->raid_conf;
2630         int disks = sh->disks, i;
2631         struct bio *return_bi = NULL;
2632         struct stripe_head_state s;
2633         struct r5dev *dev;
2634         unsigned long pending = 0;
2635
2636         memset(&s, 0, sizeof(s));
2637         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2638                 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2639                 atomic_read(&sh->count), sh->pd_idx,
2640                 sh->ops.pending, sh->ops.ack, sh->ops.complete);
2641
2642         spin_lock(&sh->lock);
2643         clear_bit(STRIPE_HANDLE, &sh->state);
2644         clear_bit(STRIPE_DELAYED, &sh->state);
2645
2646         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2647         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2648         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2649         /* Now to look around and see what can be done */
2650
2651         /* clean-up completed biofill operations */
2652         if (test_bit(STRIPE_OP_BIOFILL, &sh->ops.complete)) {
2653                 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending);
2654                 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack);
2655                 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
2656         }
2657
2658         rcu_read_lock();
2659         for (i=disks; i--; ) {
2660                 mdk_rdev_t *rdev;
2661                 struct r5dev *dev = &sh->dev[i];
2662                 clear_bit(R5_Insync, &dev->flags);
2663
2664                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2665                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2666                         dev->towrite, dev->written);
2667
2668                 /* maybe we can request a biofill operation
2669                  *
2670                  * new wantfill requests are only permitted while
2671                  * STRIPE_OP_BIOFILL is clear
2672                  */
2673                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2674                         !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2675                         set_bit(R5_Wantfill, &dev->flags);
2676
2677                 /* now count some things */
2678                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2679                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2680                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2681
2682                 if (test_bit(R5_Wantfill, &dev->flags))
2683                         s.to_fill++;
2684                 else if (dev->toread)
2685                         s.to_read++;
2686                 if (dev->towrite) {
2687                         s.to_write++;
2688                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2689                                 s.non_overwrite++;
2690                 }
2691                 if (dev->written)
2692                         s.written++;
2693                 rdev = rcu_dereference(conf->disks[i].rdev);
2694                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2695                         /* The ReadError flag will just be confusing now */
2696                         clear_bit(R5_ReadError, &dev->flags);
2697                         clear_bit(R5_ReWrite, &dev->flags);
2698                 }
2699                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2700                     || test_bit(R5_ReadError, &dev->flags)) {
2701                         s.failed++;
2702                         s.failed_num = i;
2703                 } else
2704                         set_bit(R5_Insync, &dev->flags);
2705         }
2706         rcu_read_unlock();
2707
2708         if (s.to_fill && !test_and_set_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2709                 sh->ops.count++;
2710
2711         pr_debug("locked=%d uptodate=%d to_read=%d"
2712                 " to_write=%d failed=%d failed_num=%d\n",
2713                 s.locked, s.uptodate, s.to_read, s.to_write,
2714                 s.failed, s.failed_num);
2715         /* check if the array has lost two devices and, if so, some requests might
2716          * need to be failed
2717          */
2718         if (s.failed > 1 && s.to_read+s.to_write+s.written)
2719                 handle_requests_to_failed_array(conf, sh, &s, disks,
2720                                                 &return_bi);
2721         if (s.failed > 1 && s.syncing) {
2722                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2723                 clear_bit(STRIPE_SYNCING, &sh->state);
2724                 s.syncing = 0;
2725         }
2726
2727         /* might be able to return some write requests if the parity block
2728          * is safe, or on a failed drive
2729          */
2730         dev = &sh->dev[sh->pd_idx];
2731         if ( s.written &&
2732              ((test_bit(R5_Insync, &dev->flags) &&
2733                !test_bit(R5_LOCKED, &dev->flags) &&
2734                test_bit(R5_UPTODATE, &dev->flags)) ||
2735                (s.failed == 1 && s.failed_num == sh->pd_idx)))
2736                 handle_completed_write_requests(conf, sh, disks, &return_bi);
2737
2738         /* Now we might consider reading some blocks, either to check/generate
2739          * parity, or to satisfy requests
2740          * or to load a block that is being partially written.
2741          */
2742         if (s.to_read || s.non_overwrite ||
2743             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding ||
2744             test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2745                 handle_issuing_new_read_requests5(sh, &s, disks);
2746
2747         /* Now we check to see if any write operations have recently
2748          * completed
2749          */
2750
2751         /* leave prexor set until postxor is done, allows us to distinguish
2752          * a rmw from a rcw during biodrain
2753          */
2754         if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2755                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2756
2757                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2758                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2759                 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2760
2761                 for (i = disks; i--; )
2762                         clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2763         }
2764
2765         /* if only POSTXOR is set then this is an 'expand' postxor */
2766         if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2767                 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2768
2769                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2770                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2771                 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2772
2773                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2774                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2775                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2776
2777                 /* All the 'written' buffers and the parity block are ready to
2778                  * be written back to disk
2779                  */
2780                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2781                 for (i = disks; i--; ) {
2782                         dev = &sh->dev[i];
2783                         if (test_bit(R5_LOCKED, &dev->flags) &&
2784                                 (i == sh->pd_idx || dev->written)) {
2785                                 pr_debug("Writing block %d\n", i);
2786                                 set_bit(R5_Wantwrite, &dev->flags);
2787                                 if (!test_and_set_bit(
2788                                     STRIPE_OP_IO, &sh->ops.pending))
2789                                         sh->ops.count++;
2790                                 if (!test_bit(R5_Insync, &dev->flags) ||
2791                                     (i == sh->pd_idx && s.failed == 0))
2792                                         set_bit(STRIPE_INSYNC, &sh->state);
2793                         }
2794                 }
2795                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2796                         atomic_dec(&conf->preread_active_stripes);
2797                         if (atomic_read(&conf->preread_active_stripes) <
2798                                 IO_THRESHOLD)
2799                                 md_wakeup_thread(conf->mddev->thread);
2800                 }
2801         }
2802
2803         /* Now to consider new write requests and what else, if anything
2804          * should be read.  We do not handle new writes when:
2805          * 1/ A 'write' operation (copy+xor) is already in flight.
2806          * 2/ A 'check' operation is in flight, as it may clobber the parity
2807          *    block.
2808          */
2809         if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2810                           !test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2811                 handle_issuing_new_write_requests5(conf, sh, &s, disks);
2812
2813         /* maybe we need to check and possibly fix the parity for this stripe
2814          * Any reads will already have been scheduled, so we just see if enough
2815          * data is available.  The parity check is held off while parity
2816          * dependent operations are in flight.
2817          */
2818         if ((s.syncing && s.locked == 0 &&
2819              !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2820              !test_bit(STRIPE_INSYNC, &sh->state)) ||
2821               test_bit(STRIPE_OP_CHECK, &sh->ops.pending) ||
2822               test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending))
2823                 handle_parity_checks5(conf, sh, &s, disks);
2824
2825         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2826                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2827                 clear_bit(STRIPE_SYNCING, &sh->state);
2828         }
2829
2830         /* If the failed drive is just a ReadError, then we might need to progress
2831          * the repair/check process
2832          */
2833         if (s.failed == 1 && !conf->mddev->ro &&
2834             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2835             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2836             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2837                 ) {
2838                 dev = &sh->dev[s.failed_num];
2839                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2840                         set_bit(R5_Wantwrite, &dev->flags);
2841                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2842                                 sh->ops.count++;
2843                         set_bit(R5_ReWrite, &dev->flags);
2844                         set_bit(R5_LOCKED, &dev->flags);
2845                         s.locked++;
2846                 } else {
2847                         /* let's read it back */
2848                         set_bit(R5_Wantread, &dev->flags);
2849                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2850                                 sh->ops.count++;
2851                         set_bit(R5_LOCKED, &dev->flags);
2852                         s.locked++;
2853                 }
2854         }
2855
2856         /* Finish postxor operations initiated by the expansion
2857          * process
2858          */
2859         if (test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete) &&
2860                 !test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending)) {
2861
2862                 clear_bit(STRIPE_EXPANDING, &sh->state);
2863
2864                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2865                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2866                 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2867
2868                 for (i = conf->raid_disks; i--; ) {
2869                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2870                         if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2871                                 sh->ops.count++;
2872                 }
2873         }
2874
2875         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2876                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2877                 /* Need to write out all blocks after computing parity */
2878                 sh->disks = conf->raid_disks;
2879                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2880                         conf->raid_disks);
2881                 s.locked += handle_write_operations5(sh, 1, 1);
2882         } else if (s.expanded &&
2883                 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2884                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2885                 atomic_dec(&conf->reshape_stripes);
2886                 wake_up(&conf->wait_for_overlap);
2887                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2888         }
2889
2890         if (s.expanding && s.locked == 0 &&
2891             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2892                 handle_stripe_expansion(conf, sh, NULL);
2893
2894         if (sh->ops.count)
2895                 pending = get_stripe_work(sh);
2896
2897         spin_unlock(&sh->lock);
2898
2899         if (pending)
2900                 raid5_run_ops(sh, pending);
2901
2902         return_io(return_bi);
2903
2904 }
2905
2906 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2907 {
2908         raid6_conf_t *conf = sh->raid_conf;
2909         int disks = sh->disks;
2910         struct bio *return_bi = NULL;
2911         int i, pd_idx = sh->pd_idx;
2912         struct stripe_head_state s;
2913         struct r6_state r6s;
2914         struct r5dev *dev, *pdev, *qdev;
2915
2916         r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2917         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2918                 "pd_idx=%d, qd_idx=%d\n",
2919                (unsigned long long)sh->sector, sh->state,
2920                atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2921         memset(&s, 0, sizeof(s));
2922
2923         spin_lock(&sh->lock);
2924         clear_bit(STRIPE_HANDLE, &sh->state);
2925         clear_bit(STRIPE_DELAYED, &sh->state);
2926
2927         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2928         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2929         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2930         /* Now to look around and see what can be done */
2931
2932         rcu_read_lock();
2933         for (i=disks; i--; ) {
2934                 mdk_rdev_t *rdev;
2935                 dev = &sh->dev[i];
2936                 clear_bit(R5_Insync, &dev->flags);
2937
2938                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2939                         i, dev->flags, dev->toread, dev->towrite, dev->written);
2940                 /* maybe we can reply to a read */
2941                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2942                         struct bio *rbi, *rbi2;
2943                         pr_debug("Return read for disc %d\n", i);
2944                         spin_lock_irq(&conf->device_lock);
2945                         rbi = dev->toread;
2946                         dev->toread = NULL;
2947                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2948                                 wake_up(&conf->wait_for_overlap);
2949                         spin_unlock_irq(&conf->device_lock);
2950                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2951                                 copy_data(0, rbi, dev->page, dev->sector);
2952                                 rbi2 = r5_next_bio(rbi, dev->sector);
2953                                 spin_lock_irq(&conf->device_lock);
2954                                 if (--rbi->bi_phys_segments == 0) {
2955                                         rbi->bi_next = return_bi;
2956                                         return_bi = rbi;
2957                                 }
2958                                 spin_unlock_irq(&conf->device_lock);
2959                                 rbi = rbi2;
2960                         }
2961                 }
2962
2963                 /* now count some things */
2964                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2965                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2966
2967
2968                 if (dev->toread)
2969                         s.to_read++;
2970                 if (dev->towrite) {
2971                         s.to_write++;
2972                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2973                                 s.non_overwrite++;
2974                 }
2975                 if (dev->written)
2976                         s.written++;
2977                 rdev = rcu_dereference(conf->disks[i].rdev);
2978                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2979                         /* The ReadError flag will just be confusing now */
2980                         clear_bit(R5_ReadError, &dev->flags);
2981                         clear_bit(R5_ReWrite, &dev->flags);
2982                 }
2983                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2984                     || test_bit(R5_ReadError, &dev->flags)) {
2985                         if (s.failed < 2)
2986                                 r6s.failed_num[s.failed] = i;
2987                         s.failed++;
2988                 } else
2989                         set_bit(R5_Insync, &dev->flags);
2990         }
2991         rcu_read_unlock();
2992         pr_debug("locked=%d uptodate=%d to_read=%d"
2993                " to_write=%d failed=%d failed_num=%d,%d\n",
2994                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
2995                r6s.failed_num[0], r6s.failed_num[1]);
2996         /* check if the array has lost >2 devices and, if so, some requests
2997          * might need to be failed
2998          */
2999         if (s.failed > 2 && s.to_read+s.to_write+s.written)
3000                 handle_requests_to_failed_array(conf, sh, &s, disks,
3001                                                 &return_bi);
3002         if (s.failed > 2 && s.syncing) {
3003                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3004                 clear_bit(STRIPE_SYNCING, &sh->state);
3005                 s.syncing = 0;
3006         }
3007
3008         /*
3009          * might be able to return some write requests if the parity blocks
3010          * are safe, or on a failed drive
3011          */
3012         pdev = &sh->dev[pd_idx];
3013         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3014                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3015         qdev = &sh->dev[r6s.qd_idx];
3016         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
3017                 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
3018
3019         if ( s.written &&
3020              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3021                              && !test_bit(R5_LOCKED, &pdev->flags)
3022                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3023              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3024                              && !test_bit(R5_LOCKED, &qdev->flags)
3025                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3026                 handle_completed_write_requests(conf, sh, disks, &return_bi);
3027
3028         /* Now we might consider reading some blocks, either to check/generate
3029          * parity, or to satisfy requests
3030          * or to load a block that is being partially written.
3031          */
3032         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3033             (s.syncing && (s.uptodate < disks)) || s.expanding)
3034                 handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
3035
3036         /* now to consider writing and what else, if anything should be read */
3037         if (s.to_write)
3038                 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
3039
3040         /* maybe we need to check and possibly fix the parity for this stripe
3041          * Any reads will already have been scheduled, so we just see if enough
3042          * data is available
3043          */
3044         if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3045                 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3046
3047         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3048                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3049                 clear_bit(STRIPE_SYNCING, &sh->state);
3050         }
3051
3052         /* If the failed drives are just a ReadError, then we might need
3053          * to progress the repair/check process
3054          */
3055         if (s.failed <= 2 && !conf->mddev->ro)
3056                 for (i = 0; i < s.failed; i++) {
3057                         dev = &sh->dev[r6s.failed_num[i]];
3058                         if (test_bit(R5_ReadError, &dev->flags)
3059                             && !test_bit(R5_LOCKED, &dev->flags)
3060                             && test_bit(R5_UPTODATE, &dev->flags)
3061                                 ) {
3062                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3063                                         set_bit(R5_Wantwrite, &dev->flags);
3064                                         set_bit(R5_ReWrite, &dev->flags);
3065                                         set_bit(R5_LOCKED, &dev->flags);
3066                                 } else {
3067                                         /* let's read it back */
3068                                         set_bit(R5_Wantread, &dev->flags);
3069                                         set_bit(R5_LOCKED, &dev->flags);
3070                                 }
3071                         }
3072                 }
3073
3074         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3075                 /* Need to write out all blocks after computing P&Q */
3076                 sh->disks = conf->raid_disks;
3077                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
3078                                              conf->raid_disks);
3079                 compute_parity6(sh, RECONSTRUCT_WRITE);
3080                 for (i = conf->raid_disks ; i-- ;  ) {
3081                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3082                         s.locked++;
3083                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3084                 }
3085                 clear_bit(STRIPE_EXPANDING, &sh->state);
3086         } else if (s.expanded) {
3087                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3088                 atomic_dec(&conf->reshape_stripes);
3089                 wake_up(&conf->wait_for_overlap);
3090                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3091         }
3092
3093         if (s.expanding && s.locked == 0 &&
3094             !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
3095                 handle_stripe_expansion(conf, sh, &r6s);
3096
3097         spin_unlock(&sh->lock);
3098
3099         return_io(return_bi);
3100
3101         for (i=disks; i-- ;) {
3102                 int rw;
3103                 struct bio *bi;
3104                 mdk_rdev_t *rdev;
3105                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
3106                         rw = WRITE;
3107                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
3108                         rw = READ;
3109                 else
3110                         continue;
3111
3112                 set_bit(STRIPE_IO_STARTED, &sh->state);
3113
3114                 bi = &sh->dev[i].req;
3115
3116                 bi->bi_rw = rw;
3117                 if (rw == WRITE)
3118                         bi->bi_end_io = raid5_end_write_request;
3119                 else
3120                         bi->bi_end_io = raid5_end_read_request;
3121
3122                 rcu_read_lock();
3123                 rdev = rcu_dereference(conf->disks[i].rdev);
3124                 if (rdev && test_bit(Faulty, &rdev->flags))
3125                         rdev = NULL;
3126                 if (rdev)
3127                         atomic_inc(&rdev->nr_pending);
3128                 rcu_read_unlock();
3129
3130                 if (rdev) {
3131                         if (s.syncing || s.expanding || s.expanded)
3132                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
3133
3134                         bi->bi_bdev = rdev->bdev;
3135                         pr_debug("for %llu schedule op %ld on disc %d\n",
3136                                 (unsigned long long)sh->sector, bi->bi_rw, i);
3137                         atomic_inc(&sh->count);
3138                         bi->bi_sector = sh->sector + rdev->data_offset;
3139                         bi->bi_flags = 1 << BIO_UPTODATE;
3140                         bi->bi_vcnt = 1;
3141                         bi->bi_max_vecs = 1;
3142                         bi->bi_idx = 0;
3143                         bi->bi_io_vec = &sh->dev[i].vec;
3144                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
3145                         bi->bi_io_vec[0].bv_offset = 0;
3146                         bi->bi_size = STRIPE_SIZE;
3147                         bi->bi_next = NULL;
3148                         if (rw == WRITE &&
3149                             test_bit(R5_ReWrite, &sh->dev[i].flags))
3150                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
3151                         generic_make_request(bi);
3152                 } else {
3153                         if (rw == WRITE)
3154                                 set_bit(STRIPE_DEGRADED, &sh->state);
3155                         pr_debug("skip op %ld on disc %d for sector %llu\n",
3156                                 bi->bi_rw, i, (unsigned long long)sh->sector);
3157                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
3158                         set_bit(STRIPE_HANDLE, &sh->state);
3159                 }
3160         }
3161 }
3162
3163 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3164 {
3165         if (sh->raid_conf->level == 6)
3166                 handle_stripe6(sh, tmp_page);
3167         else
3168                 handle_stripe5(sh);
3169 }
3170
3171
3172
3173 static void raid5_activate_delayed(raid5_conf_t *conf)
3174 {
3175         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3176                 while (!list_empty(&conf->delayed_list)) {
3177                         struct list_head *l = conf->delayed_list.next;
3178                         struct stripe_head *sh;
3179                         sh = list_entry(l, struct stripe_head, lru);
3180                         list_del_init(l);
3181                         clear_bit(STRIPE_DELAYED, &sh->state);
3182                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3183                                 atomic_inc(&conf->preread_active_stripes);
3184                         list_add_tail(&sh->lru, &conf->hold_list);
3185                 }
3186         } else
3187                 blk_plug_device(conf->mddev->queue);
3188 }
3189
3190 static void activate_bit_delay(raid5_conf_t *conf)
3191 {
3192         /* device_lock is held */
3193         struct list_head head;
3194         list_add(&head, &conf->bitmap_list);
3195         list_del_init(&conf->bitmap_list);
3196         while (!list_empty(&head)) {
3197                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3198                 list_del_init(&sh->lru);
3199                 atomic_inc(&sh->count);
3200                 __release_stripe(conf, sh);
3201         }
3202 }
3203
3204 static void unplug_slaves(mddev_t *mddev)
3205 {
3206         raid5_conf_t *conf = mddev_to_conf(mddev);
3207         int i;
3208
3209         rcu_read_lock();
3210         for (i=0; i<mddev->raid_disks; i++) {
3211                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3212                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3213                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3214
3215                         atomic_inc(&rdev->nr_pending);
3216                         rcu_read_unlock();
3217
3218                         blk_unplug(r_queue);
3219
3220                         rdev_dec_pending(rdev, mddev);
3221                         rcu_read_lock();
3222                 }
3223         }
3224         rcu_read_unlock();
3225 }
3226
3227 static void raid5_unplug_device(struct request_queue *q)
3228 {
3229         mddev_t *mddev = q->queuedata;
3230         raid5_conf_t *conf = mddev_to_conf(mddev);
3231         unsigned long flags;
3232
3233         spin_lock_irqsave(&conf->device_lock, flags);
3234
3235         if (blk_remove_plug(q)) {
3236                 conf->seq_flush++;
3237                 raid5_activate_delayed(conf);
3238         }
3239         md_wakeup_thread(mddev->thread);
3240
3241         spin_unlock_irqrestore(&conf->device_lock, flags);
3242
3243         unplug_slaves(mddev);
3244 }
3245
3246 static int raid5_congested(void *data, int bits)
3247 {
3248         mddev_t *mddev = data;
3249         raid5_conf_t *conf = mddev_to_conf(mddev);
3250
3251         /* No difference between reads and writes.  Just check
3252          * how busy the stripe_cache is
3253          */
3254         if (conf->inactive_blocked)
3255                 return 1;
3256         if (conf->quiesce)
3257                 return 1;
3258         if (list_empty_careful(&conf->inactive_list))
3259                 return 1;
3260
3261         return 0;
3262 }
3263
3264 /* We want read requests to align with chunks where possible,
3265  * but write requests don't need to.
3266  */
3267 static int raid5_mergeable_bvec(struct request_queue *q, struct bio *bio, struct bio_vec *biovec)
3268 {
3269         mddev_t *mddev = q->queuedata;
3270         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3271         int max;
3272         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3273         unsigned int bio_sectors = bio->bi_size >> 9;
3274
3275         if (bio_data_dir(bio) == WRITE)
3276                 return biovec->bv_len; /* always allow writes to be mergeable */
3277
3278         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3279         if (max < 0) max = 0;
3280         if (max <= biovec->bv_len && bio_sectors == 0)
3281                 return biovec->bv_len;
3282         else
3283                 return max;
3284 }
3285
3286
3287 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3288 {
3289         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3290         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3291         unsigned int bio_sectors = bio->bi_size >> 9;
3292
3293         return  chunk_sectors >=
3294                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3295 }
3296
3297 /*
3298  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3299  *  later sampled by raid5d.
3300  */
3301 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3302 {
3303         unsigned long flags;
3304
3305         spin_lock_irqsave(&conf->device_lock, flags);
3306
3307         bi->bi_next = conf->retry_read_aligned_list;
3308         conf->retry_read_aligned_list = bi;
3309
3310         spin_unlock_irqrestore(&conf->device_lock, flags);
3311         md_wakeup_thread(conf->mddev->thread);
3312 }
3313
3314
3315 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3316 {
3317         struct bio *bi;
3318
3319         bi = conf->retry_read_aligned;
3320         if (bi) {
3321                 conf->retry_read_aligned = NULL;
3322                 return bi;
3323         }
3324         bi = conf->retry_read_aligned_list;
3325         if(bi) {
3326                 conf->retry_read_aligned_list = bi->bi_next;
3327                 bi->bi_next = NULL;
3328                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3329                 bi->bi_hw_segments = 0; /* count of processed stripes */
3330         }
3331
3332         return bi;
3333 }
3334
3335
3336 /*
3337  *  The "raid5_align_endio" should check if the read succeeded and if it
3338  *  did, call bio_endio on the original bio (having bio_put the new bio
3339  *  first).
3340  *  If the read failed..
3341  */
3342 static void raid5_align_endio(struct bio *bi, int error)
3343 {
3344         struct bio* raid_bi  = bi->bi_private;
3345         mddev_t *mddev;
3346         raid5_conf_t *conf;
3347         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3348         mdk_rdev_t *rdev;
3349
3350         bio_put(bi);
3351
3352         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3353         conf = mddev_to_conf(mddev);
3354         rdev = (void*)raid_bi->bi_next;
3355         raid_bi->bi_next = NULL;
3356
3357         rdev_dec_pending(rdev, conf->mddev);
3358
3359         if (!error && uptodate) {
3360                 bio_endio(raid_bi, 0);
3361                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3362                         wake_up(&conf->wait_for_stripe);
3363                 return;
3364         }
3365
3366
3367         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3368
3369         add_bio_to_retry(raid_bi, conf);
3370 }
3371
3372 static int bio_fits_rdev(struct bio *bi)
3373 {
3374         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3375
3376         if ((bi->bi_size>>9) > q->max_sectors)
3377                 return 0;
3378         blk_recount_segments(q, bi);
3379         if (bi->bi_phys_segments > q->max_phys_segments ||
3380             bi->bi_hw_segments > q->max_hw_segments)
3381                 return 0;
3382
3383         if (q->merge_bvec_fn)
3384                 /* it's too hard to apply the merge_bvec_fn at this stage,
3385                  * just just give up
3386                  */
3387                 return 0;
3388
3389         return 1;
3390 }
3391
3392
3393 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3394 {
3395         mddev_t *mddev = q->queuedata;
3396         raid5_conf_t *conf = mddev_to_conf(mddev);
3397         const unsigned int raid_disks = conf->raid_disks;
3398         const unsigned int data_disks = raid_disks - conf->max_degraded;
3399         unsigned int dd_idx, pd_idx;
3400         struct bio* align_bi;
3401         mdk_rdev_t *rdev;
3402
3403         if (!in_chunk_boundary(mddev, raid_bio)) {
3404                 pr_debug("chunk_aligned_read : non aligned\n");
3405                 return 0;
3406         }
3407         /*
3408          * use bio_clone to make a copy of the bio
3409          */
3410         align_bi = bio_clone(raid_bio, GFP_NOIO);
3411         if (!align_bi)
3412                 return 0;
3413         /*
3414          *   set bi_end_io to a new function, and set bi_private to the
3415          *     original bio.
3416          */
3417         align_bi->bi_end_io  = raid5_align_endio;
3418         align_bi->bi_private = raid_bio;
3419         /*
3420          *      compute position
3421          */
3422         align_bi->bi_sector =  raid5_compute_sector(raid_bio->bi_sector,
3423                                         raid_disks,
3424                                         data_disks,
3425                                         &dd_idx,
3426                                         &pd_idx,
3427                                         conf);
3428
3429         rcu_read_lock();
3430         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3431         if (rdev && test_bit(In_sync, &rdev->flags)) {
3432                 atomic_inc(&rdev->nr_pending);
3433                 rcu_read_unlock();
3434                 raid_bio->bi_next = (void*)rdev;
3435                 align_bi->bi_bdev =  rdev->bdev;
3436                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3437                 align_bi->bi_sector += rdev->data_offset;
3438
3439                 if (!bio_fits_rdev(align_bi)) {
3440                         /* too big in some way */
3441                         bio_put(align_bi);
3442                         rdev_dec_pending(rdev, mddev);
3443                         return 0;
3444                 }
3445
3446                 spin_lock_irq(&conf->device_lock);
3447                 wait_event_lock_irq(conf->wait_for_stripe,
3448                                     conf->quiesce == 0,
3449                                     conf->device_lock, /* nothing */);
3450                 atomic_inc(&conf->active_aligned_reads);
3451                 spin_unlock_irq(&conf->device_lock);
3452
3453                 generic_make_request(align_bi);
3454                 return 1;
3455         } else {
3456                 rcu_read_unlock();
3457                 bio_put(align_bi);
3458                 return 0;
3459         }
3460 }
3461
3462 /* __get_priority_stripe - get the next stripe to process
3463  *
3464  * Full stripe writes are allowed to pass preread active stripes up until
3465  * the bypass_threshold is exceeded.  In general the bypass_count
3466  * increments when the handle_list is handled before the hold_list; however, it
3467  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3468  * stripe with in flight i/o.  The bypass_count will be reset when the
3469  * head of the hold_list has changed, i.e. the head was promoted to the
3470  * handle_list.
3471  */
3472 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3473 {
3474         struct stripe_head *sh;
3475
3476         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3477                   __func__,
3478                   list_empty(&conf->handle_list) ? "empty" : "busy",
3479                   list_empty(&conf->hold_list) ? "empty" : "busy",
3480                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
3481
3482         if (!list_empty(&conf->handle_list)) {
3483                 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3484
3485                 if (list_empty(&conf->hold_list))
3486                         conf->bypass_count = 0;
3487                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3488                         if (conf->hold_list.next == conf->last_hold)
3489                                 conf->bypass_count++;
3490                         else {
3491                                 conf->last_hold = conf->hold_list.next;
3492                                 conf->bypass_count -= conf->bypass_threshold;
3493                                 if (conf->bypass_count < 0)
3494                                         conf->bypass_count = 0;
3495                         }
3496                 }
3497         } else if (!list_empty(&conf->hold_list) &&
3498                    ((conf->bypass_threshold &&
3499                      conf->bypass_count > conf->bypass_threshold) ||
3500                     atomic_read(&conf->pending_full_writes) == 0)) {
3501                 sh = list_entry(conf->hold_list.next,
3502                                 typeof(*sh), lru);
3503                 conf->bypass_count -= conf->bypass_threshold;
3504                 if (conf->bypass_count < 0)
3505                         conf->bypass_count = 0;
3506         } else
3507                 return NULL;
3508
3509         list_del_init(&sh->lru);
3510         atomic_inc(&sh->count);
3511         BUG_ON(atomic_read(&sh->count) != 1);
3512         return sh;
3513 }
3514
3515 static int make_request(struct request_queue *q, struct bio * bi)
3516 {
3517         mddev_t *mddev = q->queuedata;
3518         raid5_conf_t *conf = mddev_to_conf(mddev);
3519         unsigned int dd_idx, pd_idx;
3520         sector_t new_sector;
3521         sector_t logical_sector, last_sector;
3522         struct stripe_head *sh;
3523         const int rw = bio_data_dir(bi);
3524         int remaining;
3525
3526         if (unlikely(bio_barrier(bi))) {
3527                 bio_endio(bi, -EOPNOTSUPP);
3528                 return 0;
3529         }
3530
3531         md_write_start(mddev, bi);
3532
3533         disk_stat_inc(mddev->gendisk, ios[rw]);
3534         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3535
3536         if (rw == READ &&
3537              mddev->reshape_position == MaxSector &&
3538              chunk_aligned_read(q,bi))
3539                 return 0;
3540
3541         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3542         last_sector = bi->bi_sector + (bi->bi_size>>9);
3543         bi->bi_next = NULL;
3544         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3545
3546         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3547                 DEFINE_WAIT(w);
3548                 int disks, data_disks;
3549
3550         retry:
3551                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3552                 if (likely(conf->expand_progress == MaxSector))
3553                         disks = conf->raid_disks;
3554                 else {
3555                         /* spinlock is needed as expand_progress may be
3556                          * 64bit on a 32bit platform, and so it might be
3557                          * possible to see a half-updated value
3558                          * Ofcourse expand_progress could change after
3559                          * the lock is dropped, so once we get a reference
3560                          * to the stripe that we think it is, we will have
3561                          * to check again.
3562                          */
3563                         spin_lock_irq(&conf->device_lock);
3564                         disks = conf->raid_disks;
3565                         if (logical_sector >= conf->expand_progress)
3566                                 disks = conf->previous_raid_disks;
3567                         else {
3568                                 if (logical_sector >= conf->expand_lo) {
3569                                         spin_unlock_irq(&conf->device_lock);
3570                                         schedule();
3571                                         goto retry;
3572                                 }
3573                         }
3574                         spin_unlock_irq(&conf->device_lock);
3575                 }
3576                 data_disks = disks - conf->max_degraded;
3577
3578                 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3579                                                   &dd_idx, &pd_idx, conf);
3580                 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3581                         (unsigned long long)new_sector, 
3582                         (unsigned long long)logical_sector);
3583
3584                 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3585                 if (sh) {
3586                         if (unlikely(conf->expand_progress != MaxSector)) {
3587                                 /* expansion might have moved on while waiting for a
3588                                  * stripe, so we must do the range check again.
3589                                  * Expansion could still move past after this
3590                                  * test, but as we are holding a reference to
3591                                  * 'sh', we know that if that happens,
3592                                  *  STRIPE_EXPANDING will get set and the expansion
3593                                  * won't proceed until we finish with the stripe.
3594                                  */
3595                                 int must_retry = 0;
3596                                 spin_lock_irq(&conf->device_lock);
3597                                 if (logical_sector <  conf->expand_progress &&
3598                                     disks == conf->previous_raid_disks)
3599                                         /* mismatch, need to try again */
3600                                         must_retry = 1;
3601                                 spin_unlock_irq(&conf->device_lock);
3602                                 if (must_retry) {
3603                                         release_stripe(sh);
3604                                         goto retry;
3605                                 }
3606                         }
3607                         /* FIXME what if we get a false positive because these
3608                          * are being updated.
3609                          */
3610                         if (logical_sector >= mddev->suspend_lo &&
3611                             logical_sector < mddev->suspend_hi) {
3612                                 release_stripe(sh);
3613                                 schedule();
3614                                 goto retry;
3615                         }
3616
3617                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3618                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3619                                 /* Stripe is busy expanding or
3620                                  * add failed due to overlap.  Flush everything
3621                                  * and wait a while
3622                                  */
3623                                 raid5_unplug_device(mddev->queue);
3624                                 release_stripe(sh);
3625                                 schedule();
3626                                 goto retry;
3627                         }
3628                         finish_wait(&conf->wait_for_overlap, &w);
3629                         set_bit(STRIPE_HANDLE, &sh->state);
3630                         clear_bit(STRIPE_DELAYED, &sh->state);
3631                         release_stripe(sh);
3632                 } else {
3633                         /* cannot get stripe for read-ahead, just give-up */
3634                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3635                         finish_wait(&conf->wait_for_overlap, &w);
3636                         break;
3637                 }
3638                         
3639         }
3640         spin_lock_irq(&conf->device_lock);
3641         remaining = --bi->bi_phys_segments;
3642         spin_unlock_irq(&conf->device_lock);
3643         if (remaining == 0) {
3644
3645                 if ( rw == WRITE )
3646                         md_write_end(mddev);
3647
3648                 bi->bi_end_io(bi,
3649                               test_bit(BIO_UPTODATE, &bi->bi_flags)
3650                                 ? 0 : -EIO);
3651         }
3652         return 0;
3653 }
3654
3655 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3656 {
3657         /* reshaping is quite different to recovery/resync so it is
3658          * handled quite separately ... here.
3659          *
3660          * On each call to sync_request, we gather one chunk worth of
3661          * destination stripes and flag them as expanding.
3662          * Then we find all the source stripes and request reads.
3663          * As the reads complete, handle_stripe will copy the data
3664          * into the destination stripe and release that stripe.
3665          */
3666         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3667         struct stripe_head *sh;
3668         int pd_idx;
3669         sector_t first_sector, last_sector;
3670         int raid_disks = conf->previous_raid_disks;
3671         int data_disks = raid_disks - conf->max_degraded;
3672         int new_data_disks = conf->raid_disks - conf->max_degraded;
3673         int i;
3674         int dd_idx;
3675         sector_t writepos, safepos, gap;
3676
3677         if (sector_nr == 0 &&
3678             conf->expand_progress != 0) {
3679                 /* restarting in the middle, skip the initial sectors */
3680                 sector_nr = conf->expand_progress;
3681                 sector_div(sector_nr, new_data_disks);
3682                 *skipped = 1;
3683                 return sector_nr;
3684         }
3685
3686         /* we update the metadata when there is more than 3Meg
3687          * in the block range (that is rather arbitrary, should
3688          * probably be time based) or when the data about to be
3689          * copied would over-write the source of the data at
3690          * the front of the range.
3691          * i.e. one new_stripe forward from expand_progress new_maps
3692          * to after where expand_lo old_maps to
3693          */
3694         writepos = conf->expand_progress +
3695                 conf->chunk_size/512*(new_data_disks);
3696         sector_div(writepos, new_data_disks);
3697         safepos = conf->expand_lo;
3698         sector_div(safepos, data_disks);
3699         gap = conf->expand_progress - conf->expand_lo;
3700
3701         if (writepos >= safepos ||
3702             gap > (new_data_disks)*3000*2 /*3Meg*/) {
3703                 /* Cannot proceed until we've updated the superblock... */
3704                 wait_event(conf->wait_for_overlap,
3705                            atomic_read(&conf->reshape_stripes)==0);
3706                 mddev->reshape_position = conf->expand_progress;
3707                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3708                 md_wakeup_thread(mddev->thread);
3709                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3710                            kthread_should_stop());
3711                 spin_lock_irq(&conf->device_lock);
3712                 conf->expand_lo = mddev->reshape_position;
3713                 spin_unlock_irq(&conf->device_lock);
3714                 wake_up(&conf->wait_for_overlap);
3715         }
3716
3717         for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3718                 int j;
3719                 int skipped = 0;
3720                 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3721                 sh = get_active_stripe(conf, sector_nr+i,
3722                                        conf->raid_disks, pd_idx, 0);
3723                 set_bit(STRIPE_EXPANDING, &sh->state);
3724                 atomic_inc(&conf->reshape_stripes);
3725                 /* If any of this stripe is beyond the end of the old
3726                  * array, then we need to zero those blocks
3727                  */
3728                 for (j=sh->disks; j--;) {
3729                         sector_t s;
3730                         if (j == sh->pd_idx)
3731                                 continue;
3732                         if (conf->level == 6 &&
3733                             j == raid6_next_disk(sh->pd_idx, sh->disks))
3734                                 continue;
3735                         s = compute_blocknr(sh, j);
3736                         if (s < (mddev->array_size<<1)) {
3737                                 skipped = 1;
3738                                 continue;
3739                         }
3740                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3741                         set_bit(R5_Expanded, &sh->dev[j].flags);
3742                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
3743                 }
3744                 if (!skipped) {
3745                         set_bit(STRIPE_EXPAND_READY, &sh->state);
3746                         set_bit(STRIPE_HANDLE, &sh->state);
3747                 }
3748                 release_stripe(sh);
3749         }
3750         spin_lock_irq(&conf->device_lock);
3751         conf->expand_progress = (sector_nr + i) * new_data_disks;
3752         spin_unlock_irq(&conf->device_lock);
3753         /* Ok, those stripe are ready. We can start scheduling
3754          * reads on the source stripes.
3755          * The source stripes are determined by mapping the first and last
3756          * block on the destination stripes.
3757          */
3758         first_sector =
3759                 raid5_compute_sector(sector_nr*(new_data_disks),
3760                                      raid_disks, data_disks,
3761                                      &dd_idx, &pd_idx, conf);
3762         last_sector =
3763                 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3764                                      *(new_data_disks) -1,
3765                                      raid_disks, data_disks,
3766                                      &dd_idx, &pd_idx, conf);
3767         if (last_sector >= (mddev->size<<1))
3768                 last_sector = (mddev->size<<1)-1;
3769         while (first_sector <= last_sector) {
3770                 pd_idx = stripe_to_pdidx(first_sector, conf,
3771                                          conf->previous_raid_disks);
3772                 sh = get_active_stripe(conf, first_sector,
3773                                        conf->previous_raid_disks, pd_idx, 0);
3774                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3775                 set_bit(STRIPE_HANDLE, &sh->state);
3776                 release_stripe(sh);
3777                 first_sector += STRIPE_SECTORS;
3778         }
3779         /* If this takes us to the resync_max point where we have to pause,
3780          * then we need to write out the superblock.
3781          */
3782         sector_nr += conf->chunk_size>>9;
3783         if (sector_nr >= mddev->resync_max) {
3784                 /* Cannot proceed until we've updated the superblock... */
3785                 wait_event(conf->wait_for_overlap,
3786                            atomic_read(&conf->reshape_stripes) == 0);
3787                 mddev->reshape_position = conf->expand_progress;
3788                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3789                 md_wakeup_thread(mddev->thread);
3790                 wait_event(mddev->sb_wait,
3791                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3792                            || kthread_should_stop());
3793                 spin_lock_irq(&conf->device_lock);
3794                 conf->expand_lo = mddev->reshape_position;
3795                 spin_unlock_irq(&conf->device_lock);
3796                 wake_up(&conf->wait_for_overlap);
3797         }
3798         return conf->chunk_size>>9;
3799 }
3800
3801 /* FIXME go_faster isn't used */
3802 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3803 {
3804         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3805         struct stripe_head *sh;
3806         int pd_idx;
3807         int raid_disks = conf->raid_disks;
3808         sector_t max_sector = mddev->size << 1;
3809         int sync_blocks;
3810         int still_degraded = 0;
3811         int i;
3812
3813         if (sector_nr >= max_sector) {
3814                 /* just being told to finish up .. nothing much to do */
3815                 unplug_slaves(mddev);
3816                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3817                         end_reshape(conf);
3818                         return 0;
3819                 }
3820
3821                 if (mddev->curr_resync < max_sector) /* aborted */
3822                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3823                                         &sync_blocks, 1);
3824                 else /* completed sync */
3825                         conf->fullsync = 0;
3826                 bitmap_close_sync(mddev->bitmap);
3827
3828                 return 0;
3829         }
3830
3831         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3832                 return reshape_request(mddev, sector_nr, skipped);
3833
3834         /* No need to check resync_max as we never do more than one
3835          * stripe, and as resync_max will always be on a chunk boundary,
3836          * if the check in md_do_sync didn't fire, there is no chance
3837          * of overstepping resync_max here
3838          */
3839
3840         /* if there is too many failed drives and we are trying
3841          * to resync, then assert that we are finished, because there is
3842          * nothing we can do.
3843          */
3844         if (mddev->degraded >= conf->max_degraded &&
3845             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3846                 sector_t rv = (mddev->size << 1) - sector_nr;
3847                 *skipped = 1;
3848                 return rv;
3849         }
3850         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3851             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3852             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3853                 /* we can skip this block, and probably more */
3854                 sync_blocks /= STRIPE_SECTORS;
3855                 *skipped = 1;
3856                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3857         }
3858
3859
3860         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3861
3862         pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3863         sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3864         if (sh == NULL) {
3865                 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3866                 /* make sure we don't swamp the stripe cache if someone else
3867                  * is trying to get access
3868                  */
3869                 schedule_timeout_uninterruptible(1);
3870         }
3871         /* Need to check if array will still be degraded after recovery/resync
3872          * We don't need to check the 'failed' flag as when that gets set,
3873          * recovery aborts.
3874          */
3875         for (i=0; i<mddev->raid_disks; i++)
3876                 if (conf->disks[i].rdev == NULL)
3877                         still_degraded = 1;
3878
3879         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3880
3881         spin_lock(&sh->lock);
3882         set_bit(STRIPE_SYNCING, &sh->state);
3883         clear_bit(STRIPE_INSYNC, &sh->state);
3884         spin_unlock(&sh->lock);
3885
3886         handle_stripe(sh, NULL);
3887         release_stripe(sh);
3888
3889         return STRIPE_SECTORS;
3890 }
3891
3892 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3893 {
3894         /* We may not be able to submit a whole bio at once as there
3895          * may not be enough stripe_heads available.
3896          * We cannot pre-allocate enough stripe_heads as we may need
3897          * more than exist in the cache (if we allow ever large chunks).
3898          * So we do one stripe head at a time and record in
3899          * ->bi_hw_segments how many have been done.
3900          *
3901          * We *know* that this entire raid_bio is in one chunk, so
3902          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3903          */
3904         struct stripe_head *sh;
3905         int dd_idx, pd_idx;
3906         sector_t sector, logical_sector, last_sector;
3907         int scnt = 0;
3908         int remaining;
3909         int handled = 0;
3910
3911         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3912         sector = raid5_compute_sector(  logical_sector,
3913                                         conf->raid_disks,
3914                                         conf->raid_disks - conf->max_degraded,
3915                                         &dd_idx,
3916                                         &pd_idx,
3917                                         conf);
3918         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3919
3920         for (; logical_sector < last_sector;
3921              logical_sector += STRIPE_SECTORS,
3922                      sector += STRIPE_SECTORS,
3923                      scnt++) {
3924
3925                 if (scnt < raid_bio->bi_hw_segments)
3926                         /* already done this stripe */
3927                         continue;
3928
3929                 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3930
3931                 if (!sh) {
3932                         /* failed to get a stripe - must wait */
3933                         raid_bio->bi_hw_segments = scnt;
3934                         conf->retry_read_aligned = raid_bio;
3935                         return handled;
3936                 }
3937
3938                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3939                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3940                         release_stripe(sh);
3941                         raid_bio->bi_hw_segments = scnt;
3942                         conf->retry_read_aligned = raid_bio;
3943                         return handled;
3944                 }
3945
3946                 handle_stripe(sh, NULL);
3947                 release_stripe(sh);
3948                 handled++;
3949         }
3950         spin_lock_irq(&conf->device_lock);
3951         remaining = --raid_bio->bi_phys_segments;
3952         spin_unlock_irq(&conf->device_lock);
3953         if (remaining == 0) {
3954
3955                 raid_bio->bi_end_io(raid_bio,
3956                               test_bit(BIO_UPTODATE, &raid_bio->bi_flags)
3957                                 ? 0 : -EIO);
3958         }
3959         if (atomic_dec_and_test(&conf->active_aligned_reads))
3960                 wake_up(&conf->wait_for_stripe);
3961         return handled;
3962 }
3963
3964
3965
3966 /*
3967  * This is our raid5 kernel thread.
3968  *
3969  * We scan the hash table for stripes which can be handled now.
3970  * During the scan, completed stripes are saved for us by the interrupt
3971  * handler, so that they will not have to wait for our next wakeup.
3972  */
3973 static void raid5d(mddev_t *mddev)
3974 {
3975         struct stripe_head *sh;
3976         raid5_conf_t *conf = mddev_to_conf(mddev);
3977         int handled;
3978
3979         pr_debug("+++ raid5d active\n");
3980
3981         md_check_recovery(mddev);
3982
3983         handled = 0;
3984         spin_lock_irq(&conf->device_lock);
3985         while (1) {
3986                 struct bio *bio;
3987
3988                 if (conf->seq_flush != conf->seq_write) {
3989                         int seq = conf->seq_flush;
3990                         spin_unlock_irq(&conf->device_lock);
3991                         bitmap_unplug(mddev->bitmap);
3992                         spin_lock_irq(&conf->device_lock);
3993                         conf->seq_write = seq;
3994                         activate_bit_delay(conf);
3995                 }
3996
3997                 while ((bio = remove_bio_from_retry(conf))) {
3998                         int ok;
3999                         spin_unlock_irq(&conf->device_lock);
4000                         ok = retry_aligned_read(conf, bio);
4001                         spin_lock_irq(&conf->device_lock);
4002                         if (!ok)
4003                                 break;
4004                         handled++;
4005                 }
4006
4007                 sh = __get_priority_stripe(conf);
4008
4009                 if (!sh) {
4010                         async_tx_issue_pending_all();
4011                         break;
4012                 }
4013                 spin_unlock_irq(&conf->device_lock);
4014                 
4015                 handled++;
4016                 handle_stripe(sh, conf->spare_page);
4017                 release_stripe(sh);
4018
4019                 spin_lock_irq(&conf->device_lock);
4020         }
4021         pr_debug("%d stripes handled\n", handled);
4022
4023         spin_unlock_irq(&conf->device_lock);
4024
4025         unplug_slaves(mddev);
4026
4027         pr_debug("--- raid5d inactive\n");
4028 }
4029
4030 static ssize_t
4031 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4032 {
4033         raid5_conf_t *conf = mddev_to_conf(mddev);
4034         if (conf)
4035                 return sprintf(page, "%d\n", conf->max_nr_stripes);
4036         else
4037                 return 0;
4038 }
4039
4040 static ssize_t
4041 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4042 {
4043         raid5_conf_t *conf = mddev_to_conf(mddev);
4044         unsigned long new;
4045         if (len >= PAGE_SIZE)
4046                 return -EINVAL;
4047         if (!conf)
4048                 return -ENODEV;
4049
4050         if (strict_strtoul(page, 10, &new))
4051                 return -EINVAL;
4052         if (new <= 16 || new > 32768)
4053                 return -EINVAL;
4054         while (new < conf->max_nr_stripes) {
4055                 if (drop_one_stripe(conf))
4056                         conf->max_nr_stripes--;
4057                 else
4058                         break;
4059         }
4060         md_allow_write(mddev);
4061         while (new > conf->max_nr_stripes) {
4062                 if (grow_one_stripe(conf))
4063                         conf->max_nr_stripes++;
4064                 else break;
4065         }
4066         return len;
4067 }
4068
4069 static struct md_sysfs_entry
4070 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4071                                 raid5_show_stripe_cache_size,
4072                                 raid5_store_stripe_cache_size);
4073
4074 static ssize_t
4075 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4076 {
4077         raid5_conf_t *conf = mddev_to_conf(mddev);
4078         if (conf)
4079                 return sprintf(page, "%d\n", conf->bypass_threshold);
4080         else
4081                 return 0;
4082 }
4083
4084 static ssize_t
4085 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4086 {
4087         raid5_conf_t *conf = mddev_to_conf(mddev);
4088         unsigned long new;
4089         if (len >= PAGE_SIZE)
4090                 return -EINVAL;
4091         if (!conf)
4092                 return -ENODEV;
4093
4094         if (strict_strtoul(page, 10, &new))
4095                 return -EINVAL;
4096         if (new > conf->max_nr_stripes)
4097                 return -EINVAL;
4098         conf->bypass_threshold = new;
4099         return len;
4100 }
4101
4102 static struct md_sysfs_entry
4103 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4104                                         S_IRUGO | S_IWUSR,
4105                                         raid5_show_preread_threshold,
4106                                         raid5_store_preread_threshold);
4107
4108 static ssize_t
4109 stripe_cache_active_show(mddev_t *mddev, char *page)
4110 {
4111         raid5_conf_t *conf = mddev_to_conf(mddev);
4112         if (conf)
4113                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4114         else
4115                 return 0;
4116 }
4117
4118 static struct md_sysfs_entry
4119 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4120
4121 static struct attribute *raid5_attrs[] =  {
4122         &raid5_stripecache_size.attr,
4123         &raid5_stripecache_active.attr,
4124         &raid5_preread_bypass_threshold.attr,
4125         NULL,
4126 };
4127 static struct attribute_group raid5_attrs_group = {
4128         .name = NULL,
4129         .attrs = raid5_attrs,
4130 };
4131
4132 static int run(mddev_t *mddev)
4133 {
4134         raid5_conf_t *conf;
4135         int raid_disk, memory;
4136         mdk_rdev_t *rdev;
4137         struct disk_info *disk;
4138         struct list_head *tmp;
4139         int working_disks = 0;
4140
4141         if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4142                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4143                        mdname(mddev), mddev->level);
4144                 return -EIO;
4145         }
4146
4147         if (mddev->reshape_position != MaxSector) {
4148                 /* Check that we can continue the reshape.
4149                  * Currently only disks can change, it must
4150                  * increase, and we must be past the point where
4151                  * a stripe over-writes itself
4152                  */
4153                 sector_t here_new, here_old;
4154                 int old_disks;
4155                 int max_degraded = (mddev->level == 5 ? 1 : 2);
4156
4157                 if (mddev->new_level != mddev->level ||
4158                     mddev->new_layout != mddev->layout ||
4159                     mddev->new_chunk != mddev->chunk_size) {
4160                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4161                                "required - aborting.\n",
4162                                mdname(mddev));
4163                         return -EINVAL;
4164                 }
4165                 if (mddev->delta_disks <= 0) {
4166                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4167                                "(reduce disks) required - aborting.\n",
4168                                mdname(mddev));
4169                         return -EINVAL;
4170                 }
4171                 old_disks = mddev->raid_disks - mddev->delta_disks;
4172                 /* reshape_position must be on a new-stripe boundary, and one
4173                  * further up in new geometry must map after here in old
4174                  * geometry.
4175                  */
4176                 here_new = mddev->reshape_position;
4177                 if (sector_div(here_new, (mddev->chunk_size>>9)*
4178                                (mddev->raid_disks - max_degraded))) {
4179                         printk(KERN_ERR "raid5: reshape_position not "
4180                                "on a stripe boundary\n");
4181                         return -EINVAL;
4182                 }
4183                 /* here_new is the stripe we will write to */
4184                 here_old = mddev->reshape_position;
4185                 sector_div(here_old, (mddev->chunk_size>>9)*
4186                            (old_disks-max_degraded));
4187                 /* here_old is the first stripe that we might need to read
4188                  * from */
4189                 if (here_new >= here_old) {
4190                         /* Reading from the same stripe as writing to - bad */
4191                         printk(KERN_ERR "raid5: reshape_position too early for "
4192                                "auto-recovery - aborting.\n");
4193                         return -EINVAL;
4194                 }
4195                 printk(KERN_INFO "raid5: reshape will continue\n");
4196                 /* OK, we should be able to continue; */
4197         }
4198
4199
4200         mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4201         if ((conf = mddev->private) == NULL)
4202                 goto abort;
4203         if (mddev->reshape_position == MaxSector) {
4204                 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4205         } else {
4206                 conf->raid_disks = mddev->raid_disks;
4207                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4208         }
4209
4210         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4211                               GFP_KERNEL);
4212         if (!conf->disks)
4213                 goto abort;
4214
4215         conf->mddev = mddev;
4216
4217         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4218                 goto abort;
4219
4220         if (mddev->level == 6) {
4221                 conf->spare_page = alloc_page(GFP_KERNEL);
4222                 if (!conf->spare_page)
4223                         goto abort;
4224         }
4225         spin_lock_init(&conf->device_lock);
4226         init_waitqueue_head(&conf->wait_for_stripe);
4227         init_waitqueue_head(&conf->wait_for_overlap);
4228         INIT_LIST_HEAD(&conf->handle_list);
4229         INIT_LIST_HEAD(&conf->hold_list);
4230         INIT_LIST_HEAD(&conf->delayed_list);
4231         INIT_LIST_HEAD(&conf->bitmap_list);
4232         INIT_LIST_HEAD(&conf->inactive_list);
4233         atomic_set(&conf->active_stripes, 0);
4234         atomic_set(&conf->preread_active_stripes, 0);
4235         atomic_set(&conf->active_aligned_reads, 0);
4236         conf->bypass_threshold = BYPASS_THRESHOLD;
4237
4238         pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4239
4240         rdev_for_each(rdev, tmp, mddev) {
4241                 raid_disk = rdev->raid_disk;
4242                 if (raid_disk >= conf->raid_disks
4243                     || raid_disk < 0)
4244                         continue;
4245                 disk = conf->disks + raid_disk;
4246
4247                 disk->rdev = rdev;
4248
4249                 if (test_bit(In_sync, &rdev->flags)) {
4250                         char b[BDEVNAME_SIZE];
4251                         printk(KERN_INFO "raid5: device %s operational as raid"
4252                                 " disk %d\n", bdevname(rdev->bdev,b),
4253                                 raid_disk);
4254                         working_disks++;
4255                 }
4256         }
4257
4258         /*
4259          * 0 for a fully functional array, 1 or 2 for a degraded array.
4260          */
4261         mddev->degraded = conf->raid_disks - working_disks;
4262         conf->mddev = mddev;
4263         conf->chunk_size = mddev->chunk_size;
4264         conf->level = mddev->level;
4265         if (conf->level == 6)
4266                 conf->max_degraded = 2;
4267         else
4268                 conf->max_degraded = 1;
4269         conf->algorithm = mddev->layout;
4270         conf->max_nr_stripes = NR_STRIPES;
4271         conf->expand_progress = mddev->reshape_position;
4272
4273         /* device size must be a multiple of chunk size */
4274         mddev->size &= ~(mddev->chunk_size/1024 -1);
4275         mddev->resync_max_sectors = mddev->size << 1;
4276
4277         if (conf->level == 6 && conf->raid_disks < 4) {
4278                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4279                        mdname(mddev), conf->raid_disks);
4280                 goto abort;
4281         }
4282         if (!conf->chunk_size || conf->chunk_size % 4) {
4283                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4284                         conf->chunk_size, mdname(mddev));
4285                 goto abort;
4286         }
4287         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4288                 printk(KERN_ERR 
4289                         "raid5: unsupported parity algorithm %d for %s\n",
4290                         conf->algorithm, mdname(mddev));
4291                 goto abort;
4292         }
4293         if (mddev->degraded > conf->max_degraded) {
4294                 printk(KERN_ERR "raid5: not enough operational devices for %s"
4295                         " (%d/%d failed)\n",
4296                         mdname(mddev), mddev->degraded, conf->raid_disks);
4297                 goto abort;
4298         }
4299
4300         if (mddev->degraded > 0 &&
4301             mddev->recovery_cp != MaxSector) {
4302                 if (mddev->ok_start_degraded)
4303                         printk(KERN_WARNING
4304                                "raid5: starting dirty degraded array: %s"
4305                                "- data corruption possible.\n",
4306                                mdname(mddev));
4307                 else {
4308                         printk(KERN_ERR
4309                                "raid5: cannot start dirty degraded array for %s\n",
4310                                mdname(mddev));
4311                         goto abort;
4312                 }
4313         }
4314
4315         {
4316                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4317                 if (!mddev->thread) {
4318                         printk(KERN_ERR 
4319                                 "raid5: couldn't allocate thread for %s\n",
4320                                 mdname(mddev));
4321                         goto abort;
4322                 }
4323         }
4324         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4325                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4326         if (grow_stripes(conf, conf->max_nr_stripes)) {
4327                 printk(KERN_ERR 
4328                         "raid5: couldn't allocate %dkB for buffers\n", memory);
4329                 shrink_stripes(conf);
4330                 md_unregister_thread(mddev->thread);
4331                 goto abort;
4332         } else
4333                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4334                         memory, mdname(mddev));
4335
4336         if (mddev->degraded == 0)
4337                 printk("raid5: raid level %d set %s active with %d out of %d"
4338                         " devices, algorithm %d\n", conf->level, mdname(mddev), 
4339                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4340                         conf->algorithm);
4341         else
4342                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4343                         " out of %d devices, algorithm %d\n", conf->level,
4344                         mdname(mddev), mddev->raid_disks - mddev->degraded,
4345                         mddev->raid_disks, conf->algorithm);
4346
4347         print_raid5_conf(conf);
4348
4349         if (conf->expand_progress != MaxSector) {
4350                 printk("...ok start reshape thread\n");
4351                 conf->expand_lo = conf->expand_progress;
4352                 atomic_set(&conf->reshape_stripes, 0);
4353                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4354                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4355                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4356                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4357                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4358                                                         "%s_reshape");
4359         }
4360
4361         /* read-ahead size must cover two whole stripes, which is
4362          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4363          */
4364         {
4365                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4366                 int stripe = data_disks *
4367                         (mddev->chunk_size / PAGE_SIZE);
4368                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4369                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4370         }
4371
4372         /* Ok, everything is just fine now */
4373         if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4374                 printk(KERN_WARNING
4375                        "raid5: failed to create sysfs attributes for %s\n",
4376                        mdname(mddev));
4377
4378         mddev->queue->unplug_fn = raid5_unplug_device;
4379         mddev->queue->backing_dev_info.congested_data = mddev;
4380         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4381
4382         mddev->array_size =  mddev->size * (conf->previous_raid_disks -
4383                                             conf->max_degraded);
4384
4385         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4386
4387         return 0;
4388 abort:
4389         if (conf) {
4390                 print_raid5_conf(conf);
4391                 safe_put_page(conf->spare_page);
4392                 kfree(conf->disks);
4393                 kfree(conf->stripe_hashtbl);
4394                 kfree(conf);
4395         }
4396         mddev->private = NULL;
4397         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4398         return -EIO;
4399 }
4400
4401
4402
4403 static int stop(mddev_t *mddev)
4404 {
4405         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4406
4407         md_unregister_thread(mddev->thread);
4408         mddev->thread = NULL;
4409         shrink_stripes(conf);
4410         kfree(conf->stripe_hashtbl);
4411         mddev->queue->backing_dev_info.congested_fn = NULL;
4412         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4413         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4414         kfree(conf->disks);
4415         kfree(conf);
4416         mddev->private = NULL;
4417         return 0;
4418 }
4419
4420 #ifdef DEBUG
4421 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4422 {
4423         int i;
4424
4425         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4426                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4427         seq_printf(seq, "sh %llu,  count %d.\n",
4428                    (unsigned long long)sh->sector, atomic_read(&sh->count));
4429         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4430         for (i = 0; i < sh->disks; i++) {
4431                 seq_printf(seq, "(cache%d: %p %ld) ",
4432                            i, sh->dev[i].page, sh->dev[i].flags);
4433         }
4434         seq_printf(seq, "\n");
4435 }
4436
4437 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4438 {
4439         struct stripe_head *sh;
4440         struct hlist_node *hn;
4441         int i;
4442
4443         spin_lock_irq(&conf->device_lock);
4444         for (i = 0; i < NR_HASH; i++) {
4445                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4446                         if (sh->raid_conf != conf)
4447                                 continue;
4448                         print_sh(seq, sh);
4449                 }
4450         }
4451         spin_unlock_irq(&conf->device_lock);
4452 }
4453 #endif
4454
4455 static void status (struct seq_file *seq, mddev_t *mddev)
4456 {
4457         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4458         int i;
4459
4460         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4461         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4462         for (i = 0; i < conf->raid_disks; i++)
4463                 seq_printf (seq, "%s",
4464                                conf->disks[i].rdev &&
4465                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4466         seq_printf (seq, "]");
4467 #ifdef DEBUG
4468         seq_printf (seq, "\n");
4469         printall(seq, conf);
4470 #endif
4471 }
4472
4473 static void print_raid5_conf (raid5_conf_t *conf)
4474 {
4475         int i;
4476         struct disk_info *tmp;
4477
4478         printk("RAID5 conf printout:\n");
4479         if (!conf) {
4480                 printk("(conf==NULL)\n");
4481                 return;
4482         }
4483         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4484                  conf->raid_disks - conf->mddev->degraded);
4485
4486         for (i = 0; i < conf->raid_disks; i++) {
4487                 char b[BDEVNAME_SIZE];
4488                 tmp = conf->disks + i;
4489                 if (tmp->rdev)
4490                 printk(" disk %d, o:%d, dev:%s\n",
4491                         i, !test_bit(Faulty, &tmp->rdev->flags),
4492                         bdevname(tmp->rdev->bdev,b));
4493         }
4494 }
4495
4496 static int raid5_spare_active(mddev_t *mddev)
4497 {
4498         int i;
4499         raid5_conf_t *conf = mddev->private;
4500         struct disk_info *tmp;
4501
4502         for (i = 0; i < conf->raid_disks; i++) {
4503                 tmp = conf->disks + i;
4504                 if (tmp->rdev
4505                     && !test_bit(Faulty, &tmp->rdev->flags)
4506                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4507                         unsigned long flags;
4508                         spin_lock_irqsave(&conf->device_lock, flags);
4509                         mddev->degraded--;
4510                         spin_unlock_irqrestore(&conf->device_lock, flags);
4511                 }
4512         }
4513         print_raid5_conf(conf);
4514         return 0;
4515 }
4516
4517 static int raid5_remove_disk(mddev_t *mddev, int number)
4518 {
4519         raid5_conf_t *conf = mddev->private;
4520         int err = 0;
4521         mdk_rdev_t *rdev;
4522         struct disk_info *p = conf->disks + number;
4523
4524         print_raid5_conf(conf);
4525         rdev = p->rdev;
4526         if (rdev) {
4527                 if (test_bit(In_sync, &rdev->flags) ||
4528                     atomic_read(&rdev->nr_pending)) {
4529                         err = -EBUSY;
4530                         goto abort;
4531                 }
4532                 p->rdev = NULL;
4533                 synchronize_rcu();
4534                 if (atomic_read(&rdev->nr_pending)) {
4535                         /* lost the race, try later */
4536                         err = -EBUSY;
4537                         p->rdev = rdev;
4538                 }
4539         }
4540 abort:
4541
4542         print_raid5_conf(conf);
4543         return err;
4544 }
4545
4546 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4547 {
4548         raid5_conf_t *conf = mddev->private;
4549         int found = 0;
4550         int disk;
4551         struct disk_info *p;
4552
4553         if (mddev->degraded > conf->max_degraded)
4554                 /* no point adding a device */
4555                 return 0;
4556
4557         /*
4558          * find the disk ... but prefer rdev->saved_raid_disk
4559          * if possible.
4560          */
4561         if (rdev->saved_raid_disk >= 0 &&
4562             conf->disks[rdev->saved_raid_disk].rdev == NULL)
4563                 disk = rdev->saved_raid_disk;
4564         else
4565                 disk = 0;
4566         for ( ; disk < conf->raid_disks; disk++)
4567                 if ((p=conf->disks + disk)->rdev == NULL) {
4568                         clear_bit(In_sync, &rdev->flags);
4569                         rdev->raid_disk = disk;
4570                         found = 1;
4571                         if (rdev->saved_raid_disk != disk)
4572                                 conf->fullsync = 1;
4573                         rcu_assign_pointer(p->rdev, rdev);
4574                         break;
4575                 }
4576         print_raid5_conf(conf);
4577         return found;
4578 }
4579
4580 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4581 {
4582         /* no resync is happening, and there is enough space
4583          * on all devices, so we can resize.
4584          * We need to make sure resync covers any new space.
4585          * If the array is shrinking we should possibly wait until
4586          * any io in the removed space completes, but it hardly seems
4587          * worth it.
4588          */
4589         raid5_conf_t *conf = mddev_to_conf(mddev);
4590
4591         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4592         mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4593         set_capacity(mddev->gendisk, mddev->array_size << 1);
4594         mddev->changed = 1;
4595         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
4596                 mddev->recovery_cp = mddev->size << 1;
4597                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4598         }
4599         mddev->size = sectors /2;
4600         mddev->resync_max_sectors = sectors;
4601         return 0;
4602 }
4603
4604 #ifdef CONFIG_MD_RAID5_RESHAPE
4605 static int raid5_check_reshape(mddev_t *mddev)
4606 {
4607         raid5_conf_t *conf = mddev_to_conf(mddev);
4608         int err;
4609
4610         if (mddev->delta_disks < 0 ||
4611             mddev->new_level != mddev->level)
4612                 return -EINVAL; /* Cannot shrink array or change level yet */
4613         if (mddev->delta_disks == 0)
4614                 return 0; /* nothing to do */
4615
4616         /* Can only proceed if there are plenty of stripe_heads.
4617          * We need a minimum of one full stripe,, and for sensible progress
4618          * it is best to have about 4 times that.
4619          * If we require 4 times, then the default 256 4K stripe_heads will
4620          * allow for chunk sizes up to 256K, which is probably OK.
4621          * If the chunk size is greater, user-space should request more
4622          * stripe_heads first.
4623          */
4624         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4625             (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4626                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
4627                        (mddev->chunk_size / STRIPE_SIZE)*4);
4628                 return -ENOSPC;
4629         }
4630
4631         err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4632         if (err)
4633                 return err;
4634
4635         if (mddev->degraded > conf->max_degraded)
4636                 return -EINVAL;
4637         /* looks like we might be able to manage this */
4638         return 0;
4639 }
4640
4641 static int raid5_start_reshape(mddev_t *mddev)
4642 {
4643         raid5_conf_t *conf = mddev_to_conf(mddev);
4644         mdk_rdev_t *rdev;
4645         struct list_head *rtmp;
4646         int spares = 0;
4647         int added_devices = 0;
4648         unsigned long flags;
4649
4650         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4651                 return -EBUSY;
4652
4653         rdev_for_each(rdev, rtmp, mddev)
4654                 if (rdev->raid_disk < 0 &&
4655                     !test_bit(Faulty, &rdev->flags))
4656                         spares++;
4657
4658         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4659                 /* Not enough devices even to make a degraded array
4660                  * of that size
4661                  */
4662                 return -EINVAL;
4663
4664         atomic_set(&conf->reshape_stripes, 0);
4665         spin_lock_irq(&conf->device_lock);
4666         conf->previous_raid_disks = conf->raid_disks;
4667         conf->raid_disks += mddev->delta_disks;
4668         conf->expand_progress = 0;
4669         conf->expand_lo = 0;
4670         spin_unlock_irq(&conf->device_lock);
4671
4672         /* Add some new drives, as many as will fit.
4673          * We know there are enough to make the newly sized array work.
4674          */
4675         rdev_for_each(rdev, rtmp, mddev)
4676                 if (rdev->raid_disk < 0 &&
4677                     !test_bit(Faulty, &rdev->flags)) {
4678                         if (raid5_add_disk(mddev, rdev)) {
4679                                 char nm[20];
4680                                 set_bit(In_sync, &rdev->flags);
4681                                 added_devices++;
4682                                 rdev->recovery_offset = 0;
4683                                 sprintf(nm, "rd%d", rdev->raid_disk);
4684                                 if (sysfs_create_link(&mddev->kobj,
4685                                                       &rdev->kobj, nm))
4686                                         printk(KERN_WARNING
4687                                                "raid5: failed to create "
4688                                                " link %s for %s\n",
4689                                                nm, mdname(mddev));
4690                         } else
4691                                 break;
4692                 }
4693
4694         spin_lock_irqsave(&conf->device_lock, flags);
4695         mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4696         spin_unlock_irqrestore(&conf->device_lock, flags);
4697         mddev->raid_disks = conf->raid_disks;
4698         mddev->reshape_position = 0;
4699         set_bit(MD_CHANGE_DEVS, &mddev->flags);
4700
4701         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4702         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4703         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4704         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4705         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4706                                                 "%s_reshape");
4707         if (!mddev->sync_thread) {
4708                 mddev->recovery = 0;
4709                 spin_lock_irq(&conf->device_lock);
4710                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4711                 conf->expand_progress = MaxSector;
4712                 spin_unlock_irq(&conf->device_lock);
4713                 return -EAGAIN;
4714         }
4715         md_wakeup_thread(mddev->sync_thread);
4716         md_new_event(mddev);
4717         return 0;
4718 }
4719 #endif
4720
4721 static void end_reshape(raid5_conf_t *conf)
4722 {
4723         struct block_device *bdev;
4724
4725         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4726                 conf->mddev->array_size = conf->mddev->size *
4727                         (conf->raid_disks - conf->max_degraded);
4728                 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4729                 conf->mddev->changed = 1;
4730
4731                 bdev = bdget_disk(conf->mddev->gendisk, 0);
4732                 if (bdev) {
4733                         mutex_lock(&bdev->bd_inode->i_mutex);
4734                         i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4735                         mutex_unlock(&bdev->bd_inode->i_mutex);
4736                         bdput(bdev);
4737                 }
4738                 spin_lock_irq(&conf->device_lock);
4739                 conf->expand_progress = MaxSector;
4740                 spin_unlock_irq(&conf->device_lock);
4741                 conf->mddev->reshape_position = MaxSector;
4742
4743                 /* read-ahead size must cover two whole stripes, which is
4744                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4745                  */
4746                 {
4747                         int data_disks = conf->previous_raid_disks - conf->max_degraded;
4748                         int stripe = data_disks *
4749                                 (conf->mddev->chunk_size / PAGE_SIZE);
4750                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4751                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4752                 }
4753         }
4754 }
4755
4756 static void raid5_quiesce(mddev_t *mddev, int state)
4757 {
4758         raid5_conf_t *conf = mddev_to_conf(mddev);
4759
4760         switch(state) {
4761         case 2: /* resume for a suspend */
4762                 wake_up(&conf->wait_for_overlap);
4763                 break;
4764
4765         case 1: /* stop all writes */
4766                 spin_lock_irq(&conf->device_lock);
4767                 conf->quiesce = 1;
4768                 wait_event_lock_irq(conf->wait_for_stripe,
4769                                     atomic_read(&conf->active_stripes) == 0 &&
4770                                     atomic_read(&conf->active_aligned_reads) == 0,
4771                                     conf->device_lock, /* nothing */);
4772                 spin_unlock_irq(&conf->device_lock);
4773                 break;
4774
4775         case 0: /* re-enable writes */
4776                 spin_lock_irq(&conf->device_lock);
4777                 conf->quiesce = 0;
4778                 wake_up(&conf->wait_for_stripe);
4779                 wake_up(&conf->wait_for_overlap);
4780                 spin_unlock_irq(&conf->device_lock);
4781                 break;
4782         }
4783 }
4784
4785 static struct mdk_personality raid6_personality =
4786 {
4787         .name           = "raid6",
4788         .level          = 6,
4789         .owner          = THIS_MODULE,
4790         .make_request   = make_request,
4791         .run            = run,
4792         .stop           = stop,
4793         .status         = status,
4794         .error_handler  = error,
4795         .hot_add_disk   = raid5_add_disk,
4796         .hot_remove_disk= raid5_remove_disk,
4797         .spare_active   = raid5_spare_active,
4798         .sync_request   = sync_request,
4799         .resize         = raid5_resize,
4800 #ifdef CONFIG_MD_RAID5_RESHAPE
4801         .check_reshape  = raid5_check_reshape,
4802         .start_reshape  = raid5_start_reshape,
4803 #endif
4804         .quiesce        = raid5_quiesce,
4805 };
4806 static struct mdk_personality raid5_personality =
4807 {
4808         .name           = "raid5",
4809         .level          = 5,
4810         .owner          = THIS_MODULE,
4811         .make_request   = make_request,
4812         .run            = run,
4813         .stop           = stop,
4814         .status         = status,
4815         .error_handler  = error,
4816         .hot_add_disk   = raid5_add_disk,
4817         .hot_remove_disk= raid5_remove_disk,
4818         .spare_active   = raid5_spare_active,
4819         .sync_request   = sync_request,
4820         .resize         = raid5_resize,
4821 #ifdef CONFIG_MD_RAID5_RESHAPE
4822         .check_reshape  = raid5_check_reshape,
4823         .start_reshape  = raid5_start_reshape,
4824 #endif
4825         .quiesce        = raid5_quiesce,
4826 };
4827
4828 static struct mdk_personality raid4_personality =
4829 {
4830         .name           = "raid4",
4831         .level          = 4,
4832         .owner          = THIS_MODULE,
4833         .make_request   = make_request,
4834         .run            = run,
4835         .stop           = stop,
4836         .status         = status,
4837         .error_handler  = error,
4838         .hot_add_disk   = raid5_add_disk,
4839         .hot_remove_disk= raid5_remove_disk,
4840         .spare_active   = raid5_spare_active,
4841         .sync_request   = sync_request,
4842         .resize         = raid5_resize,
4843 #ifdef CONFIG_MD_RAID5_RESHAPE
4844         .check_reshape  = raid5_check_reshape,
4845         .start_reshape  = raid5_start_reshape,
4846 #endif
4847         .quiesce        = raid5_quiesce,
4848 };
4849
4850 static int __init raid5_init(void)
4851 {
4852         int e;
4853
4854         e = raid6_select_algo();
4855         if ( e )
4856                 return e;
4857         register_md_personality(&raid6_personality);
4858         register_md_personality(&raid5_personality);
4859         register_md_personality(&raid4_personality);
4860         return 0;
4861 }
4862
4863 static void raid5_exit(void)
4864 {
4865         unregister_md_personality(&raid6_personality);
4866         unregister_md_personality(&raid5_personality);
4867         unregister_md_personality(&raid4_personality);
4868 }
4869
4870 module_init(raid5_init);
4871 module_exit(raid5_exit);
4872 MODULE_LICENSE("GPL");
4873 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4874 MODULE_ALIAS("md-raid5");
4875 MODULE_ALIAS("md-raid4");
4876 MODULE_ALIAS("md-level-5");
4877 MODULE_ALIAS("md-level-4");
4878 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4879 MODULE_ALIAS("md-raid6");
4880 MODULE_ALIAS("md-level-6");
4881
4882 /* This used to be two separate modules, they were: */
4883 MODULE_ALIAS("raid5");
4884 MODULE_ALIAS("raid6");