?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Linux v2.6.18-rc5
[pohmelfs.git] / drivers / md / raid5.c
blob450066007160ca04d1a8956abe870bc1b363210a
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
56 /*
57 * Stripe cache
58 */
59
60 #define NR_STRIPES 256
61 #define STRIPE_SIZE PAGE_SIZE
62 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
63 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
64 #define IO_THRESHOLD 1
65 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
66 #define HASH_MASK (NR_HASH - 1)
67
68 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
69
70 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
71 * order without overlap. There may be several bio's per stripe+device, and
72 * a bio could span several devices.
73 * When walking this list for a particular stripe+device, we must never proceed
74 * beyond a bio that extends past this device, as the next bio might no longer
75 * be valid.
76 * This macro is used to determine the 'next' bio in the list, given the sector
77 * of the current stripe+device
78 */
79 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
80 /*
81 * The following can be used to debug the driver
82 */
83 #define RAID5_DEBUG 0
84 #define RAID5_PARANOIA 1
85 #if RAID5_PARANOIA && defined(CONFIG_SMP)
86 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
87 #else
88 # define CHECK_DEVLOCK()
89 #endif
90
91 #define PRINTK(x...) ((void)(RAID5_DEBUG && printk(x)))
92 #if RAID5_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 static void print_raid5_conf (raid5_conf_t *conf);
108
109 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
110 {
111 if (atomic_dec_and_test(&sh->count)) {
112 BUG_ON(!list_empty(&sh->lru));
113 BUG_ON(atomic_read(&conf->active_stripes)==0);
114 if (test_bit(STRIPE_HANDLE, &sh->state)) {
115 if (test_bit(STRIPE_DELAYED, &sh->state)) {
116 list_add_tail(&sh->lru, &conf->delayed_list);
117 blk_plug_device(conf->mddev->queue);
118 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
119 sh->bm_seq - conf->seq_write > 0) {
120 list_add_tail(&sh->lru, &conf->bitmap_list);
121 blk_plug_device(conf->mddev->queue);
122 } else {
123 clear_bit(STRIPE_BIT_DELAY, &sh->state);
124 list_add_tail(&sh->lru, &conf->handle_list);
125 }
126 md_wakeup_thread(conf->mddev->thread);
127 } else {
128 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
129 atomic_dec(&conf->preread_active_stripes);
130 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
131 md_wakeup_thread(conf->mddev->thread);
132 }
133 atomic_dec(&conf->active_stripes);
134 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
135 list_add_tail(&sh->lru, &conf->inactive_list);
136 wake_up(&conf->wait_for_stripe);
137 }
138 }
139 }
140 }
141 static void release_stripe(struct stripe_head *sh)
142 {
143 raid5_conf_t *conf = sh->raid_conf;
144 unsigned long flags;
145
146 spin_lock_irqsave(&conf->device_lock, flags);
147 __release_stripe(conf, sh);
148 spin_unlock_irqrestore(&conf->device_lock, flags);
149 }
150
151 static inline void remove_hash(struct stripe_head *sh)
152 {
153 PRINTK("remove_hash(), stripe %llu\n", (unsigned long long)sh->sector);
154
155 hlist_del_init(&sh->hash);
156 }
157
158 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
159 {
160 struct hlist_head *hp = stripe_hash(conf, sh->sector);
161
162 PRINTK("insert_hash(), stripe %llu\n", (unsigned long long)sh->sector);
163
164 CHECK_DEVLOCK();
165 hlist_add_head(&sh->hash, hp);
166 }
167
168
169 /* find an idle stripe, make sure it is unhashed, and return it. */
170 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
171 {
172 struct stripe_head *sh = NULL;
173 struct list_head *first;
174
175 CHECK_DEVLOCK();
176 if (list_empty(&conf->inactive_list))
177 goto out;
178 first = conf->inactive_list.next;
179 sh = list_entry(first, struct stripe_head, lru);
180 list_del_init(first);
181 remove_hash(sh);
182 atomic_inc(&conf->active_stripes);
183 out:
184 return sh;
185 }
186
187 static void shrink_buffers(struct stripe_head *sh, int num)
188 {
189 struct page *p;
190 int i;
191
192 for (i=0; i<num ; i++) {
193 p = sh->dev[i].page;
194 if (!p)
195 continue;
196 sh->dev[i].page = NULL;
197 put_page(p);
198 }
199 }
200
201 static int grow_buffers(struct stripe_head *sh, int num)
202 {
203 int i;
204
205 for (i=0; i<num; i++) {
206 struct page *page;
207
208 if (!(page = alloc_page(GFP_KERNEL))) {
209 return 1;
210 }
211 sh->dev[i].page = page;
212 }
213 return 0;
214 }
215
216 static void raid5_build_block (struct stripe_head *sh, int i);
217
218 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks)
219 {
220 raid5_conf_t *conf = sh->raid_conf;
221 int i;
222
223 BUG_ON(atomic_read(&sh->count) != 0);
224 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
225
226 CHECK_DEVLOCK();
227 PRINTK("init_stripe called, stripe %llu\n",
228 (unsigned long long)sh->sector);
229
230 remove_hash(sh);
231
232 sh->sector = sector;
233 sh->pd_idx = pd_idx;
234 sh->state = 0;
235
236 sh->disks = disks;
237
238 for (i = sh->disks; i--; ) {
239 struct r5dev *dev = &sh->dev[i];
240
241 if (dev->toread || dev->towrite || dev->written ||
242 test_bit(R5_LOCKED, &dev->flags)) {
243 printk("sector=%llx i=%d %p %p %p %d\n",
244 (unsigned long long)sh->sector, i, dev->toread,
245 dev->towrite, dev->written,
246 test_bit(R5_LOCKED, &dev->flags));
247 BUG();
248 }
249 dev->flags = 0;
250 raid5_build_block(sh, i);
251 }
252 insert_hash(conf, sh);
253 }
254
255 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
256 {
257 struct stripe_head *sh;
258 struct hlist_node *hn;
259
260 CHECK_DEVLOCK();
261 PRINTK("__find_stripe, sector %llu\n", (unsigned long long)sector);
262 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
263 if (sh->sector == sector && sh->disks == disks)
264 return sh;
265 PRINTK("__stripe %llu not in cache\n", (unsigned long long)sector);
266 return NULL;
267 }
268
269 static void unplug_slaves(mddev_t *mddev);
270 static void raid5_unplug_device(request_queue_t *q);
271
272 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks,
273 int pd_idx, int noblock)
274 {
275 struct stripe_head *sh;
276
277 PRINTK("get_stripe, sector %llu\n", (unsigned long long)sector);
278
279 spin_lock_irq(&conf->device_lock);
280
281 do {
282 wait_event_lock_irq(conf->wait_for_stripe,
283 conf->quiesce == 0,
284 conf->device_lock, /* nothing */);
285 sh = __find_stripe(conf, sector, disks);
286 if (!sh) {
287 if (!conf->inactive_blocked)
288 sh = get_free_stripe(conf);
289 if (noblock && sh == NULL)
290 break;
291 if (!sh) {
292 conf->inactive_blocked = 1;
293 wait_event_lock_irq(conf->wait_for_stripe,
294 !list_empty(&conf->inactive_list) &&
295 (atomic_read(&conf->active_stripes)
296 < (conf->max_nr_stripes *3/4)
297 || !conf->inactive_blocked),
298 conf->device_lock,
299 raid5_unplug_device(conf->mddev->queue)
300 );
301 conf->inactive_blocked = 0;
302 } else
303 init_stripe(sh, sector, pd_idx, disks);
304 } else {
305 if (atomic_read(&sh->count)) {
306 BUG_ON(!list_empty(&sh->lru));
307 } else {
308 if (!test_bit(STRIPE_HANDLE, &sh->state))
309 atomic_inc(&conf->active_stripes);
310 if (list_empty(&sh->lru) &&
311 !test_bit(STRIPE_EXPANDING, &sh->state))
312 BUG();
313 list_del_init(&sh->lru);
314 }
315 }
316 } while (sh == NULL);
317
318 if (sh)
319 atomic_inc(&sh->count);
320
321 spin_unlock_irq(&conf->device_lock);
322 return sh;
323 }
324
325 static int grow_one_stripe(raid5_conf_t *conf)
326 {
327 struct stripe_head *sh;
328 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
329 if (!sh)
330 return 0;
331 memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
332 sh->raid_conf = conf;
333 spin_lock_init(&sh->lock);
334
335 if (grow_buffers(sh, conf->raid_disks)) {
336 shrink_buffers(sh, conf->raid_disks);
337 kmem_cache_free(conf->slab_cache, sh);
338 return 0;
339 }
340 sh->disks = conf->raid_disks;
341 /* we just created an active stripe so... */
342 atomic_set(&sh->count, 1);
343 atomic_inc(&conf->active_stripes);
344 INIT_LIST_HEAD(&sh->lru);
345 release_stripe(sh);
346 return 1;
347 }
348
349 static int grow_stripes(raid5_conf_t *conf, int num)
350 {
351 kmem_cache_t *sc;
352 int devs = conf->raid_disks;
353
354 sprintf(conf->cache_name[0], "raid5/%s", mdname(conf->mddev));
355 sprintf(conf->cache_name[1], "raid5/%s-alt", mdname(conf->mddev));
356 conf->active_name = 0;
357 sc = kmem_cache_create(conf->cache_name[conf->active_name],
358 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
359 0, 0, NULL, NULL);
360 if (!sc)
361 return 1;
362 conf->slab_cache = sc;
363 conf->pool_size = devs;
364 while (num--)
365 if (!grow_one_stripe(conf))
366 return 1;
367 return 0;
368 }
369
370 #ifdef CONFIG_MD_RAID5_RESHAPE
371 static int resize_stripes(raid5_conf_t *conf, int newsize)
372 {
373 /* Make all the stripes able to hold 'newsize' devices.
374 * New slots in each stripe get 'page' set to a new page.
375 *
376 * This happens in stages:
377 * 1/ create a new kmem_cache and allocate the required number of
378 * stripe_heads.
379 * 2/ gather all the old stripe_heads and tranfer the pages across
380 * to the new stripe_heads. This will have the side effect of
381 * freezing the array as once all stripe_heads have been collected,
382 * no IO will be possible. Old stripe heads are freed once their
383 * pages have been transferred over, and the old kmem_cache is
384 * freed when all stripes are done.
385 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
386 * we simple return a failre status - no need to clean anything up.
387 * 4/ allocate new pages for the new slots in the new stripe_heads.
388 * If this fails, we don't bother trying the shrink the
389 * stripe_heads down again, we just leave them as they are.
390 * As each stripe_head is processed the new one is released into
391 * active service.
392 *
393 * Once step2 is started, we cannot afford to wait for a write,
394 * so we use GFP_NOIO allocations.
395 */
396 struct stripe_head *osh, *nsh;
397 LIST_HEAD(newstripes);
398 struct disk_info *ndisks;
399 int err = 0;
400 kmem_cache_t *sc;
401 int i;
402
403 if (newsize <= conf->pool_size)
404 return 0; /* never bother to shrink */
405
406 /* Step 1 */
407 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
408 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
409 0, 0, NULL, NULL);
410 if (!sc)
411 return -ENOMEM;
412
413 for (i = conf->max_nr_stripes; i; i--) {
414 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
415 if (!nsh)
416 break;
417
418 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
419
420 nsh->raid_conf = conf;
421 spin_lock_init(&nsh->lock);
422
423 list_add(&nsh->lru, &newstripes);
424 }
425 if (i) {
426 /* didn't get enough, give up */
427 while (!list_empty(&newstripes)) {
428 nsh = list_entry(newstripes.next, struct stripe_head, lru);
429 list_del(&nsh->lru);
430 kmem_cache_free(sc, nsh);
431 }
432 kmem_cache_destroy(sc);
433 return -ENOMEM;
434 }
435 /* Step 2 - Must use GFP_NOIO now.
436 * OK, we have enough stripes, start collecting inactive
437 * stripes and copying them over
438 */
439 list_for_each_entry(nsh, &newstripes, lru) {
440 spin_lock_irq(&conf->device_lock);
441 wait_event_lock_irq(conf->wait_for_stripe,
442 !list_empty(&conf->inactive_list),
443 conf->device_lock,
444 unplug_slaves(conf->mddev)
445 );
446 osh = get_free_stripe(conf);
447 spin_unlock_irq(&conf->device_lock);
448 atomic_set(&nsh->count, 1);
449 for(i=0; i<conf->pool_size; i++)
450 nsh->dev[i].page = osh->dev[i].page;
451 for( ; i<newsize; i++)
452 nsh->dev[i].page = NULL;
453 kmem_cache_free(conf->slab_cache, osh);
454 }
455 kmem_cache_destroy(conf->slab_cache);
456
457 /* Step 3.
458 * At this point, we are holding all the stripes so the array
459 * is completely stalled, so now is a good time to resize
460 * conf->disks.
461 */
462 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
463 if (ndisks) {
464 for (i=0; i<conf->raid_disks; i++)
465 ndisks[i] = conf->disks[i];
466 kfree(conf->disks);
467 conf->disks = ndisks;
468 } else
469 err = -ENOMEM;
470
471 /* Step 4, return new stripes to service */
472 while(!list_empty(&newstripes)) {
473 nsh = list_entry(newstripes.next, struct stripe_head, lru);
474 list_del_init(&nsh->lru);
475 for (i=conf->raid_disks; i < newsize; i++)
476 if (nsh->dev[i].page == NULL) {
477 struct page *p = alloc_page(GFP_NOIO);
478 nsh->dev[i].page = p;
479 if (!p)
480 err = -ENOMEM;
481 }
482 release_stripe(nsh);
483 }
484 /* critical section pass, GFP_NOIO no longer needed */
485
486 conf->slab_cache = sc;
487 conf->active_name = 1-conf->active_name;
488 conf->pool_size = newsize;
489 return err;
490 }
491 #endif
492
493 static int drop_one_stripe(raid5_conf_t *conf)
494 {
495 struct stripe_head *sh;
496
497 spin_lock_irq(&conf->device_lock);
498 sh = get_free_stripe(conf);
499 spin_unlock_irq(&conf->device_lock);
500 if (!sh)
501 return 0;
502 BUG_ON(atomic_read(&sh->count));
503 shrink_buffers(sh, conf->pool_size);
504 kmem_cache_free(conf->slab_cache, sh);
505 atomic_dec(&conf->active_stripes);
506 return 1;
507 }
508
509 static void shrink_stripes(raid5_conf_t *conf)
510 {
511 while (drop_one_stripe(conf))
512 ;
513
514 if (conf->slab_cache)
515 kmem_cache_destroy(conf->slab_cache);
516 conf->slab_cache = NULL;
517 }
518
519 static int raid5_end_read_request(struct bio * bi, unsigned int bytes_done,
520 int error)
521 {
522 struct stripe_head *sh = bi->bi_private;
523 raid5_conf_t *conf = sh->raid_conf;
524 int disks = sh->disks, i;
525 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
526 char b[BDEVNAME_SIZE];
527 mdk_rdev_t *rdev;
528
529 if (bi->bi_size)
530 return 1;
531
532 for (i=0 ; i<disks; i++)
533 if (bi == &sh->dev[i].req)
534 break;
535
536 PRINTK("end_read_request %llu/%d, count: %d, uptodate %d.\n",
537 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
538 uptodate);
539 if (i == disks) {
540 BUG();
541 return 0;
542 }
543
544 if (uptodate) {
545 #if 0
546 struct bio *bio;
547 unsigned long flags;
548 spin_lock_irqsave(&conf->device_lock, flags);
549 /* we can return a buffer if we bypassed the cache or
550 * if the top buffer is not in highmem. If there are
551 * multiple buffers, leave the extra work to
552 * handle_stripe
553 */
554 buffer = sh->bh_read[i];
555 if (buffer &&
556 (!PageHighMem(buffer->b_page)
557 || buffer->b_page == bh->b_page )
558 ) {
559 sh->bh_read[i] = buffer->b_reqnext;
560 buffer->b_reqnext = NULL;
561 } else
562 buffer = NULL;
563 spin_unlock_irqrestore(&conf->device_lock, flags);
564 if (sh->bh_page[i]==bh->b_page)
565 set_buffer_uptodate(bh);
566 if (buffer) {
567 if (buffer->b_page != bh->b_page)
568 memcpy(buffer->b_data, bh->b_data, bh->b_size);
569 buffer->b_end_io(buffer, 1);
570 }
571 #else
572 set_bit(R5_UPTODATE, &sh->dev[i].flags);
573 #endif
574 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
575 rdev = conf->disks[i].rdev;
576 printk(KERN_INFO "raid5:%s: read error corrected (%lu sectors at %llu on %s)\n",
577 mdname(conf->mddev), STRIPE_SECTORS,
578 (unsigned long long)sh->sector + rdev->data_offset,
579 bdevname(rdev->bdev, b));
580 clear_bit(R5_ReadError, &sh->dev[i].flags);
581 clear_bit(R5_ReWrite, &sh->dev[i].flags);
582 }
583 if (atomic_read(&conf->disks[i].rdev->read_errors))
584 atomic_set(&conf->disks[i].rdev->read_errors, 0);
585 } else {
586 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
587 int retry = 0;
588 rdev = conf->disks[i].rdev;
589
590 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
591 atomic_inc(&rdev->read_errors);
592 if (conf->mddev->degraded)
593 printk(KERN_WARNING "raid5:%s: read error not correctable (sector %llu on %s).\n",
594 mdname(conf->mddev),
595 (unsigned long long)sh->sector + rdev->data_offset,
596 bdn);
597 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
598 /* Oh, no!!! */
599 printk(KERN_WARNING "raid5:%s: read error NOT corrected!! (sector %llu on %s).\n",
600 mdname(conf->mddev),
601 (unsigned long long)sh->sector + rdev->data_offset,
602 bdn);
603 else if (atomic_read(&rdev->read_errors)
604 > conf->max_nr_stripes)
605 printk(KERN_WARNING
606 "raid5:%s: Too many read errors, failing device %s.\n",
607 mdname(conf->mddev), bdn);
608 else
609 retry = 1;
610 if (retry)
611 set_bit(R5_ReadError, &sh->dev[i].flags);
612 else {
613 clear_bit(R5_ReadError, &sh->dev[i].flags);
614 clear_bit(R5_ReWrite, &sh->dev[i].flags);
615 md_error(conf->mddev, rdev);
616 }
617 }
618 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
619 #if 0
620 /* must restore b_page before unlocking buffer... */
621 if (sh->bh_page[i] != bh->b_page) {
622 bh->b_page = sh->bh_page[i];
623 bh->b_data = page_address(bh->b_page);
624 clear_buffer_uptodate(bh);
625 }
626 #endif
627 clear_bit(R5_LOCKED, &sh->dev[i].flags);
628 set_bit(STRIPE_HANDLE, &sh->state);
629 release_stripe(sh);
630 return 0;
631 }
632
633 static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
634 int error)
635 {
636 struct stripe_head *sh = bi->bi_private;
637 raid5_conf_t *conf = sh->raid_conf;
638 int disks = sh->disks, i;
639 unsigned long flags;
640 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
641
642 if (bi->bi_size)
643 return 1;
644
645 for (i=0 ; i<disks; i++)
646 if (bi == &sh->dev[i].req)
647 break;
648
649 PRINTK("end_write_request %llu/%d, count %d, uptodate: %d.\n",
650 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
651 uptodate);
652 if (i == disks) {
653 BUG();
654 return 0;
655 }
656
657 spin_lock_irqsave(&conf->device_lock, flags);
658 if (!uptodate)
659 md_error(conf->mddev, conf->disks[i].rdev);
660
661 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
662
663 clear_bit(R5_LOCKED, &sh->dev[i].flags);
664 set_bit(STRIPE_HANDLE, &sh->state);
665 __release_stripe(conf, sh);
666 spin_unlock_irqrestore(&conf->device_lock, flags);
667 return 0;
668 }
669
670
671 static sector_t compute_blocknr(struct stripe_head *sh, int i);
672
673 static void raid5_build_block (struct stripe_head *sh, int i)
674 {
675 struct r5dev *dev = &sh->dev[i];
676
677 bio_init(&dev->req);
678 dev->req.bi_io_vec = &dev->vec;
679 dev->req.bi_vcnt++;
680 dev->req.bi_max_vecs++;
681 dev->vec.bv_page = dev->page;
682 dev->vec.bv_len = STRIPE_SIZE;
683 dev->vec.bv_offset = 0;
684
685 dev->req.bi_sector = sh->sector;
686 dev->req.bi_private = sh;
687
688 dev->flags = 0;
689 dev->sector = compute_blocknr(sh, i);
690 }
691
692 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
693 {
694 char b[BDEVNAME_SIZE];
695 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
696 PRINTK("raid5: error called\n");
697
698 if (!test_bit(Faulty, &rdev->flags)) {
699 mddev->sb_dirty = 1;
700 if (test_bit(In_sync, &rdev->flags)) {
701 conf->working_disks--;
702 mddev->degraded++;
703 conf->failed_disks++;
704 clear_bit(In_sync, &rdev->flags);
705 /*
706 * if recovery was running, make sure it aborts.
707 */
708 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
709 }
710 set_bit(Faulty, &rdev->flags);
711 printk (KERN_ALERT
712 "raid5: Disk failure on %s, disabling device."
713 " Operation continuing on %d devices\n",
714 bdevname(rdev->bdev,b), conf->working_disks);
715 }
716 }
717
718 /*
719 * Input: a 'big' sector number,
720 * Output: index of the data and parity disk, and the sector # in them.
721 */
722 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
723 unsigned int data_disks, unsigned int * dd_idx,
724 unsigned int * pd_idx, raid5_conf_t *conf)
725 {
726 long stripe;
727 unsigned long chunk_number;
728 unsigned int chunk_offset;
729 sector_t new_sector;
730 int sectors_per_chunk = conf->chunk_size >> 9;
731
732 /* First compute the information on this sector */
733
734 /*
735 * Compute the chunk number and the sector offset inside the chunk
736 */
737 chunk_offset = sector_div(r_sector, sectors_per_chunk);
738 chunk_number = r_sector;
739 BUG_ON(r_sector != chunk_number);
740
741 /*
742 * Compute the stripe number
743 */
744 stripe = chunk_number / data_disks;
745
746 /*
747 * Compute the data disk and parity disk indexes inside the stripe
748 */
749 *dd_idx = chunk_number % data_disks;
750
751 /*
752 * Select the parity disk based on the user selected algorithm.
753 */
754 switch(conf->level) {
755 case 4:
756 *pd_idx = data_disks;
757 break;
758 case 5:
759 switch (conf->algorithm) {
760 case ALGORITHM_LEFT_ASYMMETRIC:
761 *pd_idx = data_disks - stripe % raid_disks;
762 if (*dd_idx >= *pd_idx)
763 (*dd_idx)++;
764 break;
765 case ALGORITHM_RIGHT_ASYMMETRIC:
766 *pd_idx = stripe % raid_disks;
767 if (*dd_idx >= *pd_idx)
768 (*dd_idx)++;
769 break;
770 case ALGORITHM_LEFT_SYMMETRIC:
771 *pd_idx = data_disks - stripe % raid_disks;
772 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
773 break;
774 case ALGORITHM_RIGHT_SYMMETRIC:
775 *pd_idx = stripe % raid_disks;
776 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
777 break;
778 default:
779 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
780 conf->algorithm);
781 }
782 break;
783 case 6:
784
785 /**** FIX THIS ****/
786 switch (conf->algorithm) {
787 case ALGORITHM_LEFT_ASYMMETRIC:
788 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
789 if (*pd_idx == raid_disks-1)
790 (*dd_idx)++; /* Q D D D P */
791 else if (*dd_idx >= *pd_idx)
792 (*dd_idx) += 2; /* D D P Q D */
793 break;
794 case ALGORITHM_RIGHT_ASYMMETRIC:
795 *pd_idx = stripe % raid_disks;
796 if (*pd_idx == raid_disks-1)
797 (*dd_idx)++; /* Q D D D P */
798 else if (*dd_idx >= *pd_idx)
799 (*dd_idx) += 2; /* D D P Q D */
800 break;
801 case ALGORITHM_LEFT_SYMMETRIC:
802 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
803 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
804 break;
805 case ALGORITHM_RIGHT_SYMMETRIC:
806 *pd_idx = stripe % raid_disks;
807 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
808 break;
809 default:
810 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
811 conf->algorithm);
812 }
813 break;
814 }
815
816 /*
817 * Finally, compute the new sector number
818 */
819 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
820 return new_sector;
821 }
822
823
824 static sector_t compute_blocknr(struct stripe_head *sh, int i)
825 {
826 raid5_conf_t *conf = sh->raid_conf;
827 int raid_disks = sh->disks, data_disks = raid_disks - 1;
828 sector_t new_sector = sh->sector, check;
829 int sectors_per_chunk = conf->chunk_size >> 9;
830 sector_t stripe;
831 int chunk_offset;
832 int chunk_number, dummy1, dummy2, dd_idx = i;
833 sector_t r_sector;
834
835
836 chunk_offset = sector_div(new_sector, sectors_per_chunk);
837 stripe = new_sector;
838 BUG_ON(new_sector != stripe);
839
840 if (i == sh->pd_idx)
841 return 0;
842 switch(conf->level) {
843 case 4: break;
844 case 5:
845 switch (conf->algorithm) {
846 case ALGORITHM_LEFT_ASYMMETRIC:
847 case ALGORITHM_RIGHT_ASYMMETRIC:
848 if (i > sh->pd_idx)
849 i--;
850 break;
851 case ALGORITHM_LEFT_SYMMETRIC:
852 case ALGORITHM_RIGHT_SYMMETRIC:
853 if (i < sh->pd_idx)
854 i += raid_disks;
855 i -= (sh->pd_idx + 1);
856 break;
857 default:
858 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
859 conf->algorithm);
860 }
861 break;
862 case 6:
863 data_disks = raid_disks - 2;
864 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
865 return 0; /* It is the Q disk */
866 switch (conf->algorithm) {
867 case ALGORITHM_LEFT_ASYMMETRIC:
868 case ALGORITHM_RIGHT_ASYMMETRIC:
869 if (sh->pd_idx == raid_disks-1)
870 i--; /* Q D D D P */
871 else if (i > sh->pd_idx)
872 i -= 2; /* D D P Q D */
873 break;
874 case ALGORITHM_LEFT_SYMMETRIC:
875 case ALGORITHM_RIGHT_SYMMETRIC:
876 if (sh->pd_idx == raid_disks-1)
877 i--; /* Q D D D P */
878 else {
879 /* D D P Q D */
880 if (i < sh->pd_idx)
881 i += raid_disks;
882 i -= (sh->pd_idx + 2);
883 }
884 break;
885 default:
886 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
887 conf->algorithm);
888 }
889 break;
890 }
891
892 chunk_number = stripe * data_disks + i;
893 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
894
895 check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
896 if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
897 printk(KERN_ERR "compute_blocknr: map not correct\n");
898 return 0;
899 }
900 return r_sector;
901 }
902
903
904
905 /*
906 * Copy data between a page in the stripe cache, and one or more bion
907 * The page could align with the middle of the bio, or there could be
908 * several bion, each with several bio_vecs, which cover part of the page
909 * Multiple bion are linked together on bi_next. There may be extras
910 * at the end of this list. We ignore them.
911 */
912 static void copy_data(int frombio, struct bio *bio,
913 struct page *page,
914 sector_t sector)
915 {
916 char *pa = page_address(page);
917 struct bio_vec *bvl;
918 int i;
919 int page_offset;
920
921 if (bio->bi_sector >= sector)
922 page_offset = (signed)(bio->bi_sector - sector) * 512;
923 else
924 page_offset = (signed)(sector - bio->bi_sector) * -512;
925 bio_for_each_segment(bvl, bio, i) {
926 int len = bio_iovec_idx(bio,i)->bv_len;
927 int clen;
928 int b_offset = 0;
929
930 if (page_offset < 0) {
931 b_offset = -page_offset;
932 page_offset += b_offset;
933 len -= b_offset;
934 }
935
936 if (len > 0 && page_offset + len > STRIPE_SIZE)
937 clen = STRIPE_SIZE - page_offset;
938 else clen = len;
939
940 if (clen > 0) {
941 char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
942 if (frombio)
943 memcpy(pa+page_offset, ba+b_offset, clen);
944 else
945 memcpy(ba+b_offset, pa+page_offset, clen);
946 __bio_kunmap_atomic(ba, KM_USER0);
947 }
948 if (clen < len) /* hit end of page */
949 break;
950 page_offset += len;
951 }
952 }
953
954 #define check_xor() do { \
955 if (count == MAX_XOR_BLOCKS) { \
956 xor_block(count, STRIPE_SIZE, ptr); \
957 count = 1; \
958 } \
959 } while(0)
960
961
962 static void compute_block(struct stripe_head *sh, int dd_idx)
963 {
964 int i, count, disks = sh->disks;
965 void *ptr[MAX_XOR_BLOCKS], *p;
966
967 PRINTK("compute_block, stripe %llu, idx %d\n",
968 (unsigned long long)sh->sector, dd_idx);
969
970 ptr[0] = page_address(sh->dev[dd_idx].page);
971 memset(ptr[0], 0, STRIPE_SIZE);
972 count = 1;
973 for (i = disks ; i--; ) {
974 if (i == dd_idx)
975 continue;
976 p = page_address(sh->dev[i].page);
977 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
978 ptr[count++] = p;
979 else
980 printk(KERN_ERR "compute_block() %d, stripe %llu, %d"
981 " not present\n", dd_idx,
982 (unsigned long long)sh->sector, i);
983
984 check_xor();
985 }
986 if (count != 1)
987 xor_block(count, STRIPE_SIZE, ptr);
988 set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
989 }
990
991 static void compute_parity5(struct stripe_head *sh, int method)
992 {
993 raid5_conf_t *conf = sh->raid_conf;
994 int i, pd_idx = sh->pd_idx, disks = sh->disks, count;
995 void *ptr[MAX_XOR_BLOCKS];
996 struct bio *chosen;
997
998 PRINTK("compute_parity5, stripe %llu, method %d\n",
999 (unsigned long long)sh->sector, method);
1000
1001 count = 1;
1002 ptr[0] = page_address(sh->dev[pd_idx].page);
1003 switch(method) {
1004 case READ_MODIFY_WRITE:
1005 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags));
1006 for (i=disks ; i-- ;) {
1007 if (i==pd_idx)
1008 continue;
1009 if (sh->dev[i].towrite &&
1010 test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
1011 ptr[count++] = page_address(sh->dev[i].page);
1012 chosen = sh->dev[i].towrite;
1013 sh->dev[i].towrite = NULL;
1014
1015 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1016 wake_up(&conf->wait_for_overlap);
1017
1018 BUG_ON(sh->dev[i].written);
1019 sh->dev[i].written = chosen;
1020 check_xor();
1021 }
1022 }
1023 break;
1024 case RECONSTRUCT_WRITE:
1025 memset(ptr[0], 0, STRIPE_SIZE);
1026 for (i= disks; i-- ;)
1027 if (i!=pd_idx && sh->dev[i].towrite) {
1028 chosen = sh->dev[i].towrite;
1029 sh->dev[i].towrite = NULL;
1030
1031 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1032 wake_up(&conf->wait_for_overlap);
1033
1034 BUG_ON(sh->dev[i].written);
1035 sh->dev[i].written = chosen;
1036 }
1037 break;
1038 case CHECK_PARITY:
1039 break;
1040 }
1041 if (count>1) {
1042 xor_block(count, STRIPE_SIZE, ptr);
1043 count = 1;
1044 }
1045
1046 for (i = disks; i--;)
1047 if (sh->dev[i].written) {
1048 sector_t sector = sh->dev[i].sector;
1049 struct bio *wbi = sh->dev[i].written;
1050 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1051 copy_data(1, wbi, sh->dev[i].page, sector);
1052 wbi = r5_next_bio(wbi, sector);
1053 }
1054
1055 set_bit(R5_LOCKED, &sh->dev[i].flags);
1056 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1057 }
1058
1059 switch(method) {
1060 case RECONSTRUCT_WRITE:
1061 case CHECK_PARITY:
1062 for (i=disks; i--;)
1063 if (i != pd_idx) {
1064 ptr[count++] = page_address(sh->dev[i].page);
1065 check_xor();
1066 }
1067 break;
1068 case READ_MODIFY_WRITE:
1069 for (i = disks; i--;)
1070 if (sh->dev[i].written) {
1071 ptr[count++] = page_address(sh->dev[i].page);
1072 check_xor();
1073 }
1074 }
1075 if (count != 1)
1076 xor_block(count, STRIPE_SIZE, ptr);
1077
1078 if (method != CHECK_PARITY) {
1079 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1080 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1081 } else
1082 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1083 }
1084
1085 static void compute_parity6(struct stripe_head *sh, int method)
1086 {
1087 raid6_conf_t *conf = sh->raid_conf;
1088 int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = conf->raid_disks, count;
1089 struct bio *chosen;
1090 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1091 void *ptrs[disks];
1092
1093 qd_idx = raid6_next_disk(pd_idx, disks);
1094 d0_idx = raid6_next_disk(qd_idx, disks);
1095
1096 PRINTK("compute_parity, stripe %llu, method %d\n",
1097 (unsigned long long)sh->sector, method);
1098
1099 switch(method) {
1100 case READ_MODIFY_WRITE:
1101 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1102 case RECONSTRUCT_WRITE:
1103 for (i= disks; i-- ;)
1104 if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1105 chosen = sh->dev[i].towrite;
1106 sh->dev[i].towrite = NULL;
1107
1108 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1109 wake_up(&conf->wait_for_overlap);
1110
1111 if (sh->dev[i].written) BUG();
1112 sh->dev[i].written = chosen;
1113 }
1114 break;
1115 case CHECK_PARITY:
1116 BUG(); /* Not implemented yet */
1117 }
1118
1119 for (i = disks; i--;)
1120 if (sh->dev[i].written) {
1121 sector_t sector = sh->dev[i].sector;
1122 struct bio *wbi = sh->dev[i].written;
1123 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1124 copy_data(1, wbi, sh->dev[i].page, sector);
1125 wbi = r5_next_bio(wbi, sector);
1126 }
1127
1128 set_bit(R5_LOCKED, &sh->dev[i].flags);
1129 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1130 }
1131
1132 // switch(method) {
1133 // case RECONSTRUCT_WRITE:
1134 // case CHECK_PARITY:
1135 // case UPDATE_PARITY:
1136 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1137 /* FIX: Is this ordering of drives even remotely optimal? */
1138 count = 0;
1139 i = d0_idx;
1140 do {
1141 ptrs[count++] = page_address(sh->dev[i].page);
1142 if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1143 printk("block %d/%d not uptodate on parity calc\n", i,count);
1144 i = raid6_next_disk(i, disks);
1145 } while ( i != d0_idx );
1146 // break;
1147 // }
1148
1149 raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1150
1151 switch(method) {
1152 case RECONSTRUCT_WRITE:
1153 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1154 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1155 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1156 set_bit(R5_LOCKED, &sh->dev[qd_idx].flags);
1157 break;
1158 case UPDATE_PARITY:
1159 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1160 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1161 break;
1162 }
1163 }
1164
1165
1166 /* Compute one missing block */
1167 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1168 {
1169 raid6_conf_t *conf = sh->raid_conf;
1170 int i, count, disks = conf->raid_disks;
1171 void *ptr[MAX_XOR_BLOCKS], *p;
1172 int pd_idx = sh->pd_idx;
1173 int qd_idx = raid6_next_disk(pd_idx, disks);
1174
1175 PRINTK("compute_block_1, stripe %llu, idx %d\n",
1176 (unsigned long long)sh->sector, dd_idx);
1177
1178 if ( dd_idx == qd_idx ) {
1179 /* We're actually computing the Q drive */
1180 compute_parity6(sh, UPDATE_PARITY);
1181 } else {
1182 ptr[0] = page_address(sh->dev[dd_idx].page);
1183 if (!nozero) memset(ptr[0], 0, STRIPE_SIZE);
1184 count = 1;
1185 for (i = disks ; i--; ) {
1186 if (i == dd_idx || i == qd_idx)
1187 continue;
1188 p = page_address(sh->dev[i].page);
1189 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1190 ptr[count++] = p;
1191 else
1192 printk("compute_block() %d, stripe %llu, %d"
1193 " not present\n", dd_idx,
1194 (unsigned long long)sh->sector, i);
1195
1196 check_xor();
1197 }
1198 if (count != 1)
1199 xor_block(count, STRIPE_SIZE, ptr);
1200 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1201 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1202 }
1203 }
1204
1205 /* Compute two missing blocks */
1206 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1207 {
1208 raid6_conf_t *conf = sh->raid_conf;
1209 int i, count, disks = conf->raid_disks;
1210 int pd_idx = sh->pd_idx;
1211 int qd_idx = raid6_next_disk(pd_idx, disks);
1212 int d0_idx = raid6_next_disk(qd_idx, disks);
1213 int faila, failb;
1214
1215 /* faila and failb are disk numbers relative to d0_idx */
1216 /* pd_idx become disks-2 and qd_idx become disks-1 */
1217 faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1218 failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1219
1220 BUG_ON(faila == failb);
1221 if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1222
1223 PRINTK("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1224 (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1225
1226 if ( failb == disks-1 ) {
1227 /* Q disk is one of the missing disks */
1228 if ( faila == disks-2 ) {
1229 /* Missing P+Q, just recompute */
1230 compute_parity6(sh, UPDATE_PARITY);
1231 return;
1232 } else {
1233 /* We're missing D+Q; recompute D from P */
1234 compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1235 compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1236 return;
1237 }
1238 }
1239
1240 /* We're missing D+P or D+D; build pointer table */
1241 {
1242 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1243 void *ptrs[disks];
1244
1245 count = 0;
1246 i = d0_idx;
1247 do {
1248 ptrs[count++] = page_address(sh->dev[i].page);
1249 i = raid6_next_disk(i, disks);
1250 if (i != dd_idx1 && i != dd_idx2 &&
1251 !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1252 printk("compute_2 with missing block %d/%d\n", count, i);
1253 } while ( i != d0_idx );
1254
1255 if ( failb == disks-2 ) {
1256 /* We're missing D+P. */
1257 raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1258 } else {
1259 /* We're missing D+D. */
1260 raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1261 }
1262
1263 /* Both the above update both missing blocks */
1264 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1265 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1266 }
1267 }
1268
1269
1270
1271 /*
1272 * Each stripe/dev can have one or more bion attached.
1273 * toread/towrite point to the first in a chain.
1274 * The bi_next chain must be in order.
1275 */
1276 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1277 {
1278 struct bio **bip;
1279 raid5_conf_t *conf = sh->raid_conf;
1280 int firstwrite=0;
1281
1282 PRINTK("adding bh b#%llu to stripe s#%llu\n",
1283 (unsigned long long)bi->bi_sector,
1284 (unsigned long long)sh->sector);
1285
1286
1287 spin_lock(&sh->lock);
1288 spin_lock_irq(&conf->device_lock);
1289 if (forwrite) {
1290 bip = &sh->dev[dd_idx].towrite;
1291 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1292 firstwrite = 1;
1293 } else
1294 bip = &sh->dev[dd_idx].toread;
1295 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1296 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1297 goto overlap;
1298 bip = & (*bip)->bi_next;
1299 }
1300 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1301 goto overlap;
1302
1303 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1304 if (*bip)
1305 bi->bi_next = *bip;
1306 *bip = bi;
1307 bi->bi_phys_segments ++;
1308 spin_unlock_irq(&conf->device_lock);
1309 spin_unlock(&sh->lock);
1310
1311 PRINTK("added bi b#%llu to stripe s#%llu, disk %d.\n",
1312 (unsigned long long)bi->bi_sector,
1313 (unsigned long long)sh->sector, dd_idx);
1314
1315 if (conf->mddev->bitmap && firstwrite) {
1316 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1317 STRIPE_SECTORS, 0);
1318 sh->bm_seq = conf->seq_flush+1;
1319 set_bit(STRIPE_BIT_DELAY, &sh->state);
1320 }
1321
1322 if (forwrite) {
1323 /* check if page is covered */
1324 sector_t sector = sh->dev[dd_idx].sector;
1325 for (bi=sh->dev[dd_idx].towrite;
1326 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1327 bi && bi->bi_sector <= sector;
1328 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1329 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1330 sector = bi->bi_sector + (bi->bi_size>>9);
1331 }
1332 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1333 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1334 }
1335 return 1;
1336
1337 overlap:
1338 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1339 spin_unlock_irq(&conf->device_lock);
1340 spin_unlock(&sh->lock);
1341 return 0;
1342 }
1343
1344 static void end_reshape(raid5_conf_t *conf);
1345
1346 static int page_is_zero(struct page *p)
1347 {
1348 char *a = page_address(p);
1349 return ((*(u32*)a) == 0 &&
1350 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1351 }
1352
1353 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1354 {
1355 int sectors_per_chunk = conf->chunk_size >> 9;
1356 sector_t x = stripe;
1357 int pd_idx, dd_idx;
1358 int chunk_offset = sector_div(x, sectors_per_chunk);
1359 stripe = x;
1360 raid5_compute_sector(stripe*(disks-1)*sectors_per_chunk
1361 + chunk_offset, disks, disks-1, &dd_idx, &pd_idx, conf);
1362 return pd_idx;
1363 }
1364
1365
1366 /*
1367 * handle_stripe - do things to a stripe.
1368 *
1369 * We lock the stripe and then examine the state of various bits
1370 * to see what needs to be done.
1371 * Possible results:
1372 * return some read request which now have data
1373 * return some write requests which are safely on disc
1374 * schedule a read on some buffers
1375 * schedule a write of some buffers
1376 * return confirmation of parity correctness
1377 *
1378 * Parity calculations are done inside the stripe lock
1379 * buffers are taken off read_list or write_list, and bh_cache buffers
1380 * get BH_Lock set before the stripe lock is released.
1381 *
1382 */
1383
1384 static void handle_stripe5(struct stripe_head *sh)
1385 {
1386 raid5_conf_t *conf = sh->raid_conf;
1387 int disks = sh->disks;
1388 struct bio *return_bi= NULL;
1389 struct bio *bi;
1390 int i;
1391 int syncing, expanding, expanded;
1392 int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
1393 int non_overwrite = 0;
1394 int failed_num=0;
1395 struct r5dev *dev;
1396
1397 PRINTK("handling stripe %llu, cnt=%d, pd_idx=%d\n",
1398 (unsigned long long)sh->sector, atomic_read(&sh->count),
1399 sh->pd_idx);
1400
1401 spin_lock(&sh->lock);
1402 clear_bit(STRIPE_HANDLE, &sh->state);
1403 clear_bit(STRIPE_DELAYED, &sh->state);
1404
1405 syncing = test_bit(STRIPE_SYNCING, &sh->state);
1406 expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
1407 expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
1408 /* Now to look around and see what can be done */
1409
1410 rcu_read_lock();
1411 for (i=disks; i--; ) {
1412 mdk_rdev_t *rdev;
1413 dev = &sh->dev[i];
1414 clear_bit(R5_Insync, &dev->flags);
1415
1416 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
1417 i, dev->flags, dev->toread, dev->towrite, dev->written);
1418 /* maybe we can reply to a read */
1419 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
1420 struct bio *rbi, *rbi2;
1421 PRINTK("Return read for disc %d\n", i);
1422 spin_lock_irq(&conf->device_lock);
1423 rbi = dev->toread;
1424 dev->toread = NULL;
1425 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1426 wake_up(&conf->wait_for_overlap);
1427 spin_unlock_irq(&conf->device_lock);
1428 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1429 copy_data(0, rbi, dev->page, dev->sector);
1430 rbi2 = r5_next_bio(rbi, dev->sector);
1431 spin_lock_irq(&conf->device_lock);
1432 if (--rbi->bi_phys_segments == 0) {
1433 rbi->bi_next = return_bi;
1434 return_bi = rbi;
1435 }
1436 spin_unlock_irq(&conf->device_lock);
1437 rbi = rbi2;
1438 }
1439 }
1440
1441 /* now count some things */
1442 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
1443 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
1444
1445
1446 if (dev->toread) to_read++;
1447 if (dev->towrite) {
1448 to_write++;
1449 if (!test_bit(R5_OVERWRITE, &dev->flags))
1450 non_overwrite++;
1451 }
1452 if (dev->written) written++;
1453 rdev = rcu_dereference(conf->disks[i].rdev);
1454 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
1455 /* The ReadError flag will just be confusing now */
1456 clear_bit(R5_ReadError, &dev->flags);
1457 clear_bit(R5_ReWrite, &dev->flags);
1458 }
1459 if (!rdev || !test_bit(In_sync, &rdev->flags)
1460 || test_bit(R5_ReadError, &dev->flags)) {
1461 failed++;
1462 failed_num = i;
1463 } else
1464 set_bit(R5_Insync, &dev->flags);
1465 }
1466 rcu_read_unlock();
1467 PRINTK("locked=%d uptodate=%d to_read=%d"
1468 " to_write=%d failed=%d failed_num=%d\n",
1469 locked, uptodate, to_read, to_write, failed, failed_num);
1470 /* check if the array has lost two devices and, if so, some requests might
1471 * need to be failed
1472 */
1473 if (failed > 1 && to_read+to_write+written) {
1474 for (i=disks; i--; ) {
1475 int bitmap_end = 0;
1476
1477 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1478 mdk_rdev_t *rdev;
1479 rcu_read_lock();
1480 rdev = rcu_dereference(conf->disks[i].rdev);
1481 if (rdev && test_bit(In_sync, &rdev->flags))
1482 /* multiple read failures in one stripe */
1483 md_error(conf->mddev, rdev);
1484 rcu_read_unlock();
1485 }
1486
1487 spin_lock_irq(&conf->device_lock);
1488 /* fail all writes first */
1489 bi = sh->dev[i].towrite;
1490 sh->dev[i].towrite = NULL;
1491 if (bi) { to_write--; bitmap_end = 1; }
1492
1493 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1494 wake_up(&conf->wait_for_overlap);
1495
1496 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1497 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1498 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1499 if (--bi->bi_phys_segments == 0) {
1500 md_write_end(conf->mddev);
1501 bi->bi_next = return_bi;
1502 return_bi = bi;
1503 }
1504 bi = nextbi;
1505 }
1506 /* and fail all 'written' */
1507 bi = sh->dev[i].written;
1508 sh->dev[i].written = NULL;
1509 if (bi) bitmap_end = 1;
1510 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
1511 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1512 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1513 if (--bi->bi_phys_segments == 0) {
1514 md_write_end(conf->mddev);
1515 bi->bi_next = return_bi;
1516 return_bi = bi;
1517 }
1518 bi = bi2;
1519 }
1520
1521 /* fail any reads if this device is non-operational */
1522 if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1523 test_bit(R5_ReadError, &sh->dev[i].flags)) {
1524 bi = sh->dev[i].toread;
1525 sh->dev[i].toread = NULL;
1526 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1527 wake_up(&conf->wait_for_overlap);
1528 if (bi) to_read--;
1529 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1530 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1531 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1532 if (--bi->bi_phys_segments == 0) {
1533 bi->bi_next = return_bi;
1534 return_bi = bi;
1535 }
1536 bi = nextbi;
1537 }
1538 }
1539 spin_unlock_irq(&conf->device_lock);
1540 if (bitmap_end)
1541 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1542 STRIPE_SECTORS, 0, 0);
1543 }
1544 }
1545 if (failed > 1 && syncing) {
1546 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
1547 clear_bit(STRIPE_SYNCING, &sh->state);
1548 syncing = 0;
1549 }
1550
1551 /* might be able to return some write requests if the parity block
1552 * is safe, or on a failed drive
1553 */
1554 dev = &sh->dev[sh->pd_idx];
1555 if ( written &&
1556 ( (test_bit(R5_Insync, &dev->flags) && !test_bit(R5_LOCKED, &dev->flags) &&
1557 test_bit(R5_UPTODATE, &dev->flags))
1558 || (failed == 1 && failed_num == sh->pd_idx))
1559 ) {
1560 /* any written block on an uptodate or failed drive can be returned.
1561 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
1562 * never LOCKED, so we don't need to test 'failed' directly.
1563 */
1564 for (i=disks; i--; )
1565 if (sh->dev[i].written) {
1566 dev = &sh->dev[i];
1567 if (!test_bit(R5_LOCKED, &dev->flags) &&
1568 test_bit(R5_UPTODATE, &dev->flags) ) {
1569 /* We can return any write requests */
1570 struct bio *wbi, *wbi2;
1571 int bitmap_end = 0;
1572 PRINTK("Return write for disc %d\n", i);
1573 spin_lock_irq(&conf->device_lock);
1574 wbi = dev->written;
1575 dev->written = NULL;
1576 while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1577 wbi2 = r5_next_bio(wbi, dev->sector);
1578 if (--wbi->bi_phys_segments == 0) {
1579 md_write_end(conf->mddev);
1580 wbi->bi_next = return_bi;
1581 return_bi = wbi;
1582 }
1583 wbi = wbi2;
1584 }
1585 if (dev->towrite == NULL)
1586 bitmap_end = 1;
1587 spin_unlock_irq(&conf->device_lock);
1588 if (bitmap_end)
1589 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1590 STRIPE_SECTORS,
1591 !test_bit(STRIPE_DEGRADED, &sh->state), 0);
1592 }
1593 }
1594 }
1595
1596 /* Now we might consider reading some blocks, either to check/generate
1597 * parity, or to satisfy requests
1598 * or to load a block that is being partially written.
1599 */
1600 if (to_read || non_overwrite || (syncing && (uptodate < disks)) || expanding) {
1601 for (i=disks; i--;) {
1602 dev = &sh->dev[i];
1603 if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1604 (dev->toread ||
1605 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1606 syncing ||
1607 expanding ||
1608 (failed && (sh->dev[failed_num].toread ||
1609 (sh->dev[failed_num].towrite && !test_bit(R5_OVERWRITE, &sh->dev[failed_num].flags))))
1610 )
1611 ) {
1612 /* we would like to get this block, possibly
1613 * by computing it, but we might not be able to
1614 */
1615 if (uptodate == disks-1) {
1616 PRINTK("Computing block %d\n", i);
1617 compute_block(sh, i);
1618 uptodate++;
1619 } else if (test_bit(R5_Insync, &dev->flags)) {
1620 set_bit(R5_LOCKED, &dev->flags);
1621 set_bit(R5_Wantread, &dev->flags);
1622 #if 0
1623 /* if I am just reading this block and we don't have
1624 a failed drive, or any pending writes then sidestep the cache */
1625 if (sh->bh_read[i] && !sh->bh_read[i]->b_reqnext &&
1626 ! syncing && !failed && !to_write) {
1627 sh->bh_cache[i]->b_page = sh->bh_read[i]->b_page;
1628 sh->bh_cache[i]->b_data = sh->bh_read[i]->b_data;
1629 }
1630 #endif
1631 locked++;
1632 PRINTK("Reading block %d (sync=%d)\n",
1633 i, syncing);
1634 }
1635 }
1636 }
1637 set_bit(STRIPE_HANDLE, &sh->state);
1638 }
1639
1640 /* now to consider writing and what else, if anything should be read */
1641 if (to_write) {
1642 int rmw=0, rcw=0;
1643 for (i=disks ; i--;) {
1644 /* would I have to read this buffer for read_modify_write */
1645 dev = &sh->dev[i];
1646 if ((dev->towrite || i == sh->pd_idx) &&
1647 (!test_bit(R5_LOCKED, &dev->flags)
1648 #if 0
1649 || sh->bh_page[i]!=bh->b_page
1650 #endif
1651 ) &&
1652 !test_bit(R5_UPTODATE, &dev->flags)) {
1653 if (test_bit(R5_Insync, &dev->flags)
1654 /* && !(!mddev->insync && i == sh->pd_idx) */
1655 )
1656 rmw++;
1657 else rmw += 2*disks; /* cannot read it */
1658 }
1659 /* Would I have to read this buffer for reconstruct_write */
1660 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1661 (!test_bit(R5_LOCKED, &dev->flags)
1662 #if 0
1663 || sh->bh_page[i] != bh->b_page
1664 #endif
1665 ) &&
1666 !test_bit(R5_UPTODATE, &dev->flags)) {
1667 if (test_bit(R5_Insync, &dev->flags)) rcw++;
1668 else rcw += 2*disks;
1669 }
1670 }
1671 PRINTK("for sector %llu, rmw=%d rcw=%d\n",
1672 (unsigned long long)sh->sector, rmw, rcw);
1673 set_bit(STRIPE_HANDLE, &sh->state);
1674 if (rmw < rcw && rmw > 0)
1675 /* prefer read-modify-write, but need to get some data */
1676 for (i=disks; i--;) {
1677 dev = &sh->dev[i];
1678 if ((dev->towrite || i == sh->pd_idx) &&
1679 !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1680 test_bit(R5_Insync, &dev->flags)) {
1681 if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1682 {
1683 PRINTK("Read_old block %d for r-m-w\n", i);
1684 set_bit(R5_LOCKED, &dev->flags);
1685 set_bit(R5_Wantread, &dev->flags);
1686 locked++;
1687 } else {
1688 set_bit(STRIPE_DELAYED, &sh->state);
1689 set_bit(STRIPE_HANDLE, &sh->state);
1690 }
1691 }
1692 }
1693 if (rcw <= rmw && rcw > 0)
1694 /* want reconstruct write, but need to get some data */
1695 for (i=disks; i--;) {
1696 dev = &sh->dev[i];
1697 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1698 !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1699 test_bit(R5_Insync, &dev->flags)) {
1700 if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1701 {
1702 PRINTK("Read_old block %d for Reconstruct\n", i);
1703 set_bit(R5_LOCKED, &dev->flags);
1704 set_bit(R5_Wantread, &dev->flags);
1705 locked++;
1706 } else {
1707 set_bit(STRIPE_DELAYED, &sh->state);
1708 set_bit(STRIPE_HANDLE, &sh->state);
1709 }
1710 }
1711 }
1712 /* now if nothing is locked, and if we have enough data, we can start a write request */
1713 if (locked == 0 && (rcw == 0 ||rmw == 0) &&
1714 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
1715 PRINTK("Computing parity...\n");
1716 compute_parity5(sh, rcw==0 ? RECONSTRUCT_WRITE : READ_MODIFY_WRITE);
1717 /* now every locked buffer is ready to be written */
1718 for (i=disks; i--;)
1719 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
1720 PRINTK("Writing block %d\n", i);
1721 locked++;
1722 set_bit(R5_Wantwrite, &sh->dev[i].flags);
1723 if (!test_bit(R5_Insync, &sh->dev[i].flags)
1724 || (i==sh->pd_idx && failed == 0))
1725 set_bit(STRIPE_INSYNC, &sh->state);
1726 }
1727 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
1728 atomic_dec(&conf->preread_active_stripes);
1729 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
1730 md_wakeup_thread(conf->mddev->thread);
1731 }
1732 }
1733 }
1734
1735 /* maybe we need to check and possibly fix the parity for this stripe
1736 * Any reads will already have been scheduled, so we just see if enough data
1737 * is available
1738 */
1739 if (syncing && locked == 0 &&
1740 !test_bit(STRIPE_INSYNC, &sh->state)) {
1741 set_bit(STRIPE_HANDLE, &sh->state);
1742 if (failed == 0) {
1743 BUG_ON(uptodate != disks);
1744 compute_parity5(sh, CHECK_PARITY);
1745 uptodate--;
1746 if (page_is_zero(sh->dev[sh->pd_idx].page)) {
1747 /* parity is correct (on disc, not in buffer any more) */
1748 set_bit(STRIPE_INSYNC, &sh->state);
1749 } else {
1750 conf->mddev->resync_mismatches += STRIPE_SECTORS;
1751 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
1752 /* don't try to repair!! */
1753 set_bit(STRIPE_INSYNC, &sh->state);
1754 else {
1755 compute_block(sh, sh->pd_idx);
1756 uptodate++;
1757 }
1758 }
1759 }
1760 if (!test_bit(STRIPE_INSYNC, &sh->state)) {
1761 /* either failed parity check, or recovery is happening */
1762 if (failed==0)
1763 failed_num = sh->pd_idx;
1764 dev = &sh->dev[failed_num];
1765 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
1766 BUG_ON(uptodate != disks);
1767
1768 set_bit(R5_LOCKED, &dev->flags);
1769 set_bit(R5_Wantwrite, &dev->flags);
1770 clear_bit(STRIPE_DEGRADED, &sh->state);
1771 locked++;
1772 set_bit(STRIPE_INSYNC, &sh->state);
1773 }
1774 }
1775 if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
1776 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
1777 clear_bit(STRIPE_SYNCING, &sh->state);
1778 }
1779
1780 /* If the failed drive is just a ReadError, then we might need to progress
1781 * the repair/check process
1782 */
1783 if (failed == 1 && ! conf->mddev->ro &&
1784 test_bit(R5_ReadError, &sh->dev[failed_num].flags)
1785 && !test_bit(R5_LOCKED, &sh->dev[failed_num].flags)
1786 && test_bit(R5_UPTODATE, &sh->dev[failed_num].flags)
1787 ) {
1788 dev = &sh->dev[failed_num];
1789 if (!test_bit(R5_ReWrite, &dev->flags)) {
1790 set_bit(R5_Wantwrite, &dev->flags);
1791 set_bit(R5_ReWrite, &dev->flags);
1792 set_bit(R5_LOCKED, &dev->flags);
1793 locked++;
1794 } else {
1795 /* let's read it back */
1796 set_bit(R5_Wantread, &dev->flags);
1797 set_bit(R5_LOCKED, &dev->flags);
1798 locked++;
1799 }
1800 }
1801
1802 if (expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
1803 /* Need to write out all blocks after computing parity */
1804 sh->disks = conf->raid_disks;
1805 sh->pd_idx = stripe_to_pdidx(sh->sector, conf, conf->raid_disks);
1806 compute_parity5(sh, RECONSTRUCT_WRITE);
1807 for (i= conf->raid_disks; i--;) {
1808 set_bit(R5_LOCKED, &sh->dev[i].flags);
1809 locked++;
1810 set_bit(R5_Wantwrite, &sh->dev[i].flags);
1811 }
1812 clear_bit(STRIPE_EXPANDING, &sh->state);
1813 } else if (expanded) {
1814 clear_bit(STRIPE_EXPAND_READY, &sh->state);
1815 atomic_dec(&conf->reshape_stripes);
1816 wake_up(&conf->wait_for_overlap);
1817 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
1818 }
1819
1820 if (expanding && locked == 0) {
1821 /* We have read all the blocks in this stripe and now we need to
1822 * copy some of them into a target stripe for expand.
1823 */
1824 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
1825 for (i=0; i< sh->disks; i++)
1826 if (i != sh->pd_idx) {
1827 int dd_idx, pd_idx, j;
1828 struct stripe_head *sh2;
1829
1830 sector_t bn = compute_blocknr(sh, i);
1831 sector_t s = raid5_compute_sector(bn, conf->raid_disks,
1832 conf->raid_disks-1,
1833 &dd_idx, &pd_idx, conf);
1834 sh2 = get_active_stripe(conf, s, conf->raid_disks, pd_idx, 1);
1835 if (sh2 == NULL)
1836 /* so far only the early blocks of this stripe
1837 * have been requested. When later blocks
1838 * get requested, we will try again
1839 */
1840 continue;
1841 if(!test_bit(STRIPE_EXPANDING, &sh2->state) ||
1842 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
1843 /* must have already done this block */
1844 release_stripe(sh2);
1845 continue;
1846 }
1847 memcpy(page_address(sh2->dev[dd_idx].page),
1848 page_address(sh->dev[i].page),
1849 STRIPE_SIZE);
1850 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
1851 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
1852 for (j=0; j<conf->raid_disks; j++)
1853 if (j != sh2->pd_idx &&
1854 !test_bit(R5_Expanded, &sh2->dev[j].flags))
1855 break;
1856 if (j == conf->raid_disks) {
1857 set_bit(STRIPE_EXPAND_READY, &sh2->state);
1858 set_bit(STRIPE_HANDLE, &sh2->state);
1859 }
1860 release_stripe(sh2);
1861 }
1862 }
1863
1864 spin_unlock(&sh->lock);
1865
1866 while ((bi=return_bi)) {
1867 int bytes = bi->bi_size;
1868
1869 return_bi = bi->bi_next;
1870 bi->bi_next = NULL;
1871 bi->bi_size = 0;
1872 bi->bi_end_io(bi, bytes, 0);
1873 }
1874 for (i=disks; i-- ;) {
1875 int rw;
1876 struct bio *bi;
1877 mdk_rdev_t *rdev;
1878 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
1879 rw = 1;
1880 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1881 rw = 0;
1882 else
1883 continue;
1884
1885 bi = &sh->dev[i].req;
1886
1887 bi->bi_rw = rw;
1888 if (rw)
1889 bi->bi_end_io = raid5_end_write_request;
1890 else
1891 bi->bi_end_io = raid5_end_read_request;
1892
1893 rcu_read_lock();
1894 rdev = rcu_dereference(conf->disks[i].rdev);
1895 if (rdev && test_bit(Faulty, &rdev->flags))
1896 rdev = NULL;
1897 if (rdev)
1898 atomic_inc(&rdev->nr_pending);
1899 rcu_read_unlock();
1900
1901 if (rdev) {
1902 if (syncing || expanding || expanded)
1903 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1904
1905 bi->bi_bdev = rdev->bdev;
1906 PRINTK("for %llu schedule op %ld on disc %d\n",
1907 (unsigned long long)sh->sector, bi->bi_rw, i);
1908 atomic_inc(&sh->count);
1909 bi->bi_sector = sh->sector + rdev->data_offset;
1910 bi->bi_flags = 1 << BIO_UPTODATE;
1911 bi->bi_vcnt = 1;
1912 bi->bi_max_vecs = 1;
1913 bi->bi_idx = 0;
1914 bi->bi_io_vec = &sh->dev[i].vec;
1915 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1916 bi->bi_io_vec[0].bv_offset = 0;
1917 bi->bi_size = STRIPE_SIZE;
1918 bi->bi_next = NULL;
1919 if (rw == WRITE &&
1920 test_bit(R5_ReWrite, &sh->dev[i].flags))
1921 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1922 generic_make_request(bi);
1923 } else {
1924 if (rw == 1)
1925 set_bit(STRIPE_DEGRADED, &sh->state);
1926 PRINTK("skip op %ld on disc %d for sector %llu\n",
1927 bi->bi_rw, i, (unsigned long long)sh->sector);
1928 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1929 set_bit(STRIPE_HANDLE, &sh->state);
1930 }
1931 }
1932 }
1933
1934 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
1935 {
1936 raid6_conf_t *conf = sh->raid_conf;
1937 int disks = conf->raid_disks;
1938 struct bio *return_bi= NULL;
1939 struct bio *bi;
1940 int i;
1941 int syncing;
1942 int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
1943 int non_overwrite = 0;
1944 int failed_num[2] = {0, 0};
1945 struct r5dev *dev, *pdev, *qdev;
1946 int pd_idx = sh->pd_idx;
1947 int qd_idx = raid6_next_disk(pd_idx, disks);
1948 int p_failed, q_failed;
1949
1950 PRINTK("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d, qd_idx=%d\n",
1951 (unsigned long long)sh->sector, sh->state, atomic_read(&sh->count),
1952 pd_idx, qd_idx);
1953
1954 spin_lock(&sh->lock);
1955 clear_bit(STRIPE_HANDLE, &sh->state);
1956 clear_bit(STRIPE_DELAYED, &sh->state);
1957
1958 syncing = test_bit(STRIPE_SYNCING, &sh->state);
1959 /* Now to look around and see what can be done */
1960
1961 rcu_read_lock();
1962 for (i=disks; i--; ) {
1963 mdk_rdev_t *rdev;
1964 dev = &sh->dev[i];
1965 clear_bit(R5_Insync, &dev->flags);
1966
1967 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
1968 i, dev->flags, dev->toread, dev->towrite, dev->written);
1969 /* maybe we can reply to a read */
1970 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
1971 struct bio *rbi, *rbi2;
1972 PRINTK("Return read for disc %d\n", i);
1973 spin_lock_irq(&conf->device_lock);
1974 rbi = dev->toread;
1975 dev->toread = NULL;
1976 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1977 wake_up(&conf->wait_for_overlap);
1978 spin_unlock_irq(&conf->device_lock);
1979 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1980 copy_data(0, rbi, dev->page, dev->sector);
1981 rbi2 = r5_next_bio(rbi, dev->sector);
1982 spin_lock_irq(&conf->device_lock);
1983 if (--rbi->bi_phys_segments == 0) {
1984 rbi->bi_next = return_bi;
1985 return_bi = rbi;
1986 }
1987 spin_unlock_irq(&conf->device_lock);
1988 rbi = rbi2;
1989 }
1990 }
1991
1992 /* now count some things */
1993 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
1994 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
1995
1996
1997 if (dev->toread) to_read++;
1998 if (dev->towrite) {
1999 to_write++;
2000 if (!test_bit(R5_OVERWRITE, &dev->flags))
2001 non_overwrite++;
2002 }
2003 if (dev->written) written++;
2004 rdev = rcu_dereference(conf->disks[i].rdev);
2005 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2006 /* The ReadError flag will just be confusing now */
2007 clear_bit(R5_ReadError, &dev->flags);
2008 clear_bit(R5_ReWrite, &dev->flags);
2009 }
2010 if (!rdev || !test_bit(In_sync, &rdev->flags)
2011 || test_bit(R5_ReadError, &dev->flags)) {
2012 if ( failed < 2 )
2013 failed_num[failed] = i;
2014 failed++;
2015 } else
2016 set_bit(R5_Insync, &dev->flags);
2017 }
2018 rcu_read_unlock();
2019 PRINTK("locked=%d uptodate=%d to_read=%d"
2020 " to_write=%d failed=%d failed_num=%d,%d\n",
2021 locked, uptodate, to_read, to_write, failed,
2022 failed_num[0], failed_num[1]);
2023 /* check if the array has lost >2 devices and, if so, some requests might
2024 * need to be failed
2025 */
2026 if (failed > 2 && to_read+to_write+written) {
2027 for (i=disks; i--; ) {
2028 int bitmap_end = 0;
2029
2030 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2031 mdk_rdev_t *rdev;
2032 rcu_read_lock();
2033 rdev = rcu_dereference(conf->disks[i].rdev);
2034 if (rdev && test_bit(In_sync, &rdev->flags))
2035 /* multiple read failures in one stripe */
2036 md_error(conf->mddev, rdev);
2037 rcu_read_unlock();
2038 }
2039
2040 spin_lock_irq(&conf->device_lock);
2041 /* fail all writes first */
2042 bi = sh->dev[i].towrite;
2043 sh->dev[i].towrite = NULL;
2044 if (bi) { to_write--; bitmap_end = 1; }
2045
2046 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2047 wake_up(&conf->wait_for_overlap);
2048
2049 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
2050 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2051 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2052 if (--bi->bi_phys_segments == 0) {
2053 md_write_end(conf->mddev);
2054 bi->bi_next = return_bi;
2055 return_bi = bi;
2056 }
2057 bi = nextbi;
2058 }
2059 /* and fail all 'written' */
2060 bi = sh->dev[i].written;
2061 sh->dev[i].written = NULL;
2062 if (bi) bitmap_end = 1;
2063 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
2064 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2065 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2066 if (--bi->bi_phys_segments == 0) {
2067 md_write_end(conf->mddev);
2068 bi->bi_next = return_bi;
2069 return_bi = bi;
2070 }
2071 bi = bi2;
2072 }
2073
2074 /* fail any reads if this device is non-operational */
2075 if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2076 test_bit(R5_ReadError, &sh->dev[i].flags)) {
2077 bi = sh->dev[i].toread;
2078 sh->dev[i].toread = NULL;
2079 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2080 wake_up(&conf->wait_for_overlap);
2081 if (bi) to_read--;
2082 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
2083 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2084 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2085 if (--bi->bi_phys_segments == 0) {
2086 bi->bi_next = return_bi;
2087 return_bi = bi;
2088 }
2089 bi = nextbi;
2090 }
2091 }
2092 spin_unlock_irq(&conf->device_lock);
2093 if (bitmap_end)
2094 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2095 STRIPE_SECTORS, 0, 0);
2096 }
2097 }
2098 if (failed > 2 && syncing) {
2099 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2100 clear_bit(STRIPE_SYNCING, &sh->state);
2101 syncing = 0;
2102 }
2103
2104 /*
2105 * might be able to return some write requests if the parity blocks
2106 * are safe, or on a failed drive
2107 */
2108 pdev = &sh->dev[pd_idx];
2109 p_failed = (failed >= 1 && failed_num[0] == pd_idx)
2110 || (failed >= 2 && failed_num[1] == pd_idx);
2111 qdev = &sh->dev[qd_idx];
2112 q_failed = (failed >= 1 && failed_num[0] == qd_idx)
2113 || (failed >= 2 && failed_num[1] == qd_idx);
2114
2115 if ( written &&
2116 ( p_failed || ((test_bit(R5_Insync, &pdev->flags)
2117 && !test_bit(R5_LOCKED, &pdev->flags)
2118 && test_bit(R5_UPTODATE, &pdev->flags))) ) &&
2119 ( q_failed || ((test_bit(R5_Insync, &qdev->flags)
2120 && !test_bit(R5_LOCKED, &qdev->flags)
2121 && test_bit(R5_UPTODATE, &qdev->flags))) ) ) {
2122 /* any written block on an uptodate or failed drive can be
2123 * returned. Note that if we 'wrote' to a failed drive,
2124 * it will be UPTODATE, but never LOCKED, so we don't need
2125 * to test 'failed' directly.
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 int bitmap_end = 0;
2134 struct bio *wbi, *wbi2;
2135 PRINTK("Return write for stripe %llu disc %d\n",
2136 (unsigned long long)sh->sector, i);
2137 spin_lock_irq(&conf->device_lock);
2138 wbi = dev->written;
2139 dev->written = NULL;
2140 while (wbi && wbi->bi_sector < 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, sh->sector,
2154 STRIPE_SECTORS,
2155 !test_bit(STRIPE_DEGRADED, &sh->state), 0);
2156 }
2157 }
2158 }
2159
2160 /* Now we might consider reading some blocks, either to check/generate
2161 * parity, or to satisfy requests
2162 * or to load a block that is being partially written.
2163 */
2164 if (to_read || non_overwrite || (to_write && failed) || (syncing && (uptodate < disks))) {
2165 for (i=disks; i--;) {
2166 dev = &sh->dev[i];
2167 if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
2168 (dev->toread ||
2169 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2170 syncing ||
2171 (failed >= 1 && (sh->dev[failed_num[0]].toread || to_write)) ||
2172 (failed >= 2 && (sh->dev[failed_num[1]].toread || to_write))
2173 )
2174 ) {
2175 /* we would like to get this block, possibly
2176 * by computing it, but we might not be able to
2177 */
2178 if (uptodate == disks-1) {
2179 PRINTK("Computing stripe %llu block %d\n",
2180 (unsigned long long)sh->sector, i);
2181 compute_block_1(sh, i, 0);
2182 uptodate++;
2183 } else if ( uptodate == disks-2 && failed >= 2 ) {
2184 /* Computing 2-failure is *very* expensive; only do it if failed >= 2 */
2185 int other;
2186 for (other=disks; other--;) {
2187 if ( other == i )
2188 continue;
2189 if ( !test_bit(R5_UPTODATE, &sh->dev[other].flags) )
2190 break;
2191 }
2192 BUG_ON(other < 0);
2193 PRINTK("Computing stripe %llu blocks %d,%d\n",
2194 (unsigned long long)sh->sector, i, other);
2195 compute_block_2(sh, i, other);
2196 uptodate += 2;
2197 } else if (test_bit(R5_Insync, &dev->flags)) {
2198 set_bit(R5_LOCKED, &dev->flags);
2199 set_bit(R5_Wantread, &dev->flags);
2200 #if 0
2201 /* if I am just reading this block and we don't have
2202 a failed drive, or any pending writes then sidestep the cache */
2203 if (sh->bh_read[i] && !sh->bh_read[i]->b_reqnext &&
2204 ! syncing && !failed && !to_write) {
2205 sh->bh_cache[i]->b_page = sh->bh_read[i]->b_page;
2206 sh->bh_cache[i]->b_data = sh->bh_read[i]->b_data;
2207 }
2208 #endif
2209 locked++;
2210 PRINTK("Reading block %d (sync=%d)\n",
2211 i, syncing);
2212 }
2213 }
2214 }
2215 set_bit(STRIPE_HANDLE, &sh->state);
2216 }
2217
2218 /* now to consider writing and what else, if anything should be read */
2219 if (to_write) {
2220 int rcw=0, must_compute=0;
2221 for (i=disks ; i--;) {
2222 dev = &sh->dev[i];
2223 /* Would I have to read this buffer for reconstruct_write */
2224 if (!test_bit(R5_OVERWRITE, &dev->flags)
2225 && i != pd_idx && i != qd_idx
2226 && (!test_bit(R5_LOCKED, &dev->flags)
2227 #if 0
2228 || sh->bh_page[i] != bh->b_page
2229 #endif
2230 ) &&
2231 !test_bit(R5_UPTODATE, &dev->flags)) {
2232 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2233 else {
2234 PRINTK("raid6: must_compute: disk %d flags=%#lx\n", i, dev->flags);
2235 must_compute++;
2236 }
2237 }
2238 }
2239 PRINTK("for sector %llu, rcw=%d, must_compute=%d\n",
2240 (unsigned long long)sh->sector, rcw, must_compute);
2241 set_bit(STRIPE_HANDLE, &sh->state);
2242
2243 if (rcw > 0)
2244 /* want reconstruct write, but need to get some data */
2245 for (i=disks; i--;) {
2246 dev = &sh->dev[i];
2247 if (!test_bit(R5_OVERWRITE, &dev->flags)
2248 && !(failed == 0 && (i == pd_idx || i == qd_idx))
2249 && !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
2250 test_bit(R5_Insync, &dev->flags)) {
2251 if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
2252 {
2253 PRINTK("Read_old stripe %llu block %d for Reconstruct\n",
2254 (unsigned long long)sh->sector, i);
2255 set_bit(R5_LOCKED, &dev->flags);
2256 set_bit(R5_Wantread, &dev->flags);
2257 locked++;
2258 } else {
2259 PRINTK("Request delayed stripe %llu block %d for Reconstruct\n",
2260 (unsigned long long)sh->sector, i);
2261 set_bit(STRIPE_DELAYED, &sh->state);
2262 set_bit(STRIPE_HANDLE, &sh->state);
2263 }
2264 }
2265 }
2266 /* now if nothing is locked, and if we have enough data, we can start a write request */
2267 if (locked == 0 && rcw == 0 &&
2268 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2269 if ( must_compute > 0 ) {
2270 /* We have failed blocks and need to compute them */
2271 switch ( failed ) {
2272 case 0: BUG();
2273 case 1: compute_block_1(sh, failed_num[0], 0); break;
2274 case 2: compute_block_2(sh, failed_num[0], failed_num[1]); break;
2275 default: BUG(); /* This request should have been failed? */
2276 }
2277 }
2278
2279 PRINTK("Computing parity for stripe %llu\n", (unsigned long long)sh->sector);
2280 compute_parity6(sh, RECONSTRUCT_WRITE);
2281 /* now every locked buffer is ready to be written */
2282 for (i=disks; i--;)
2283 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2284 PRINTK("Writing stripe %llu block %d\n",
2285 (unsigned long long)sh->sector, i);
2286 locked++;
2287 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2288 }
2289 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2290 set_bit(STRIPE_INSYNC, &sh->state);
2291
2292 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2293 atomic_dec(&conf->preread_active_stripes);
2294 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
2295 md_wakeup_thread(conf->mddev->thread);
2296 }
2297 }
2298 }
2299
2300 /* maybe we need to check and possibly fix the parity for this stripe
2301 * Any reads will already have been scheduled, so we just see if enough data
2302 * is available
2303 */
2304 if (syncing && locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state)) {
2305 int update_p = 0, update_q = 0;
2306 struct r5dev *dev;
2307
2308 set_bit(STRIPE_HANDLE, &sh->state);
2309
2310 BUG_ON(failed>2);
2311 BUG_ON(uptodate < disks);
2312 /* Want to check and possibly repair P and Q.
2313 * However there could be one 'failed' device, in which
2314 * case we can only check one of them, possibly using the
2315 * other to generate missing data
2316 */
2317
2318 /* If !tmp_page, we cannot do the calculations,
2319 * but as we have set STRIPE_HANDLE, we will soon be called
2320 * by stripe_handle with a tmp_page - just wait until then.
2321 */
2322 if (tmp_page) {
2323 if (failed == q_failed) {
2324 /* The only possible failed device holds 'Q', so it makes
2325 * sense to check P (If anything else were failed, we would
2326 * have used P to recreate it).
2327 */
2328 compute_block_1(sh, pd_idx, 1);
2329 if (!page_is_zero(sh->dev[pd_idx].page)) {
2330 compute_block_1(sh,pd_idx,0);
2331 update_p = 1;
2332 }
2333 }
2334 if (!q_failed && failed < 2) {
2335 /* q is not failed, and we didn't use it to generate
2336 * anything, so it makes sense to check it
2337 */
2338 memcpy(page_address(tmp_page),
2339 page_address(sh->dev[qd_idx].page),
2340 STRIPE_SIZE);
2341 compute_parity6(sh, UPDATE_PARITY);
2342 if (memcmp(page_address(tmp_page),
2343 page_address(sh->dev[qd_idx].page),
2344 STRIPE_SIZE)!= 0) {
2345 clear_bit(STRIPE_INSYNC, &sh->state);
2346 update_q = 1;
2347 }
2348 }
2349 if (update_p || update_q) {
2350 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2351 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2352 /* don't try to repair!! */
2353 update_p = update_q = 0;
2354 }
2355
2356 /* now write out any block on a failed drive,
2357 * or P or Q if they need it
2358 */
2359
2360 if (failed == 2) {
2361 dev = &sh->dev[failed_num[1]];
2362 locked++;
2363 set_bit(R5_LOCKED, &dev->flags);
2364 set_bit(R5_Wantwrite, &dev->flags);
2365 }
2366 if (failed >= 1) {
2367 dev = &sh->dev[failed_num[0]];
2368 locked++;
2369 set_bit(R5_LOCKED, &dev->flags);
2370 set_bit(R5_Wantwrite, &dev->flags);
2371 }
2372
2373 if (update_p) {
2374 dev = &sh->dev[pd_idx];
2375 locked ++;
2376 set_bit(R5_LOCKED, &dev->flags);
2377 set_bit(R5_Wantwrite, &dev->flags);
2378 }
2379 if (update_q) {
2380 dev = &sh->dev[qd_idx];
2381 locked++;
2382 set_bit(R5_LOCKED, &dev->flags);
2383 set_bit(R5_Wantwrite, &dev->flags);
2384 }
2385 clear_bit(STRIPE_DEGRADED, &sh->state);
2386
2387 set_bit(STRIPE_INSYNC, &sh->state);
2388 }
2389 }
2390
2391 if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2392 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2393 clear_bit(STRIPE_SYNCING, &sh->state);
2394 }
2395
2396 /* If the failed drives are just a ReadError, then we might need
2397 * to progress the repair/check process
2398 */
2399 if (failed <= 2 && ! conf->mddev->ro)
2400 for (i=0; i<failed;i++) {
2401 dev = &sh->dev[failed_num[i]];
2402 if (test_bit(R5_ReadError, &dev->flags)
2403 && !test_bit(R5_LOCKED, &dev->flags)
2404 && test_bit(R5_UPTODATE, &dev->flags)
2405 ) {
2406 if (!test_bit(R5_ReWrite, &dev->flags)) {
2407 set_bit(R5_Wantwrite, &dev->flags);
2408 set_bit(R5_ReWrite, &dev->flags);
2409 set_bit(R5_LOCKED, &dev->flags);
2410 } else {
2411 /* let's read it back */
2412 set_bit(R5_Wantread, &dev->flags);
2413 set_bit(R5_LOCKED, &dev->flags);
2414 }
2415 }
2416 }
2417 spin_unlock(&sh->lock);
2418
2419 while ((bi=return_bi)) {
2420 int bytes = bi->bi_size;
2421
2422 return_bi = bi->bi_next;
2423 bi->bi_next = NULL;
2424 bi->bi_size = 0;
2425 bi->bi_end_io(bi, bytes, 0);
2426 }
2427 for (i=disks; i-- ;) {
2428 int rw;
2429 struct bio *bi;
2430 mdk_rdev_t *rdev;
2431 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
2432 rw = 1;
2433 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
2434 rw = 0;
2435 else
2436 continue;
2437
2438 bi = &sh->dev[i].req;
2439
2440 bi->bi_rw = rw;
2441 if (rw)
2442 bi->bi_end_io = raid5_end_write_request;
2443 else
2444 bi->bi_end_io = raid5_end_read_request;
2445
2446 rcu_read_lock();
2447 rdev = rcu_dereference(conf->disks[i].rdev);
2448 if (rdev && test_bit(Faulty, &rdev->flags))
2449 rdev = NULL;
2450 if (rdev)
2451 atomic_inc(&rdev->nr_pending);
2452 rcu_read_unlock();
2453
2454 if (rdev) {
2455 if (syncing)
2456 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
2457
2458 bi->bi_bdev = rdev->bdev;
2459 PRINTK("for %llu schedule op %ld on disc %d\n",
2460 (unsigned long long)sh->sector, bi->bi_rw, i);
2461 atomic_inc(&sh->count);
2462 bi->bi_sector = sh->sector + rdev->data_offset;
2463 bi->bi_flags = 1 << BIO_UPTODATE;
2464 bi->bi_vcnt = 1;
2465 bi->bi_max_vecs = 1;
2466 bi->bi_idx = 0;
2467 bi->bi_io_vec = &sh->dev[i].vec;
2468 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
2469 bi->bi_io_vec[0].bv_offset = 0;
2470 bi->bi_size = STRIPE_SIZE;
2471 bi->bi_next = NULL;
2472 if (rw == WRITE &&
2473 test_bit(R5_ReWrite, &sh->dev[i].flags))
2474 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2475 generic_make_request(bi);
2476 } else {
2477 if (rw == 1)
2478 set_bit(STRIPE_DEGRADED, &sh->state);
2479 PRINTK("skip op %ld on disc %d for sector %llu\n",
2480 bi->bi_rw, i, (unsigned long long)sh->sector);
2481 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2482 set_bit(STRIPE_HANDLE, &sh->state);
2483 }
2484 }
2485 }
2486
2487 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
2488 {
2489 if (sh->raid_conf->level == 6)
2490 handle_stripe6(sh, tmp_page);
2491 else
2492 handle_stripe5(sh);
2493 }
2494
2495
2496
2497 static void raid5_activate_delayed(raid5_conf_t *conf)
2498 {
2499 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
2500 while (!list_empty(&conf->delayed_list)) {
2501 struct list_head *l = conf->delayed_list.next;
2502 struct stripe_head *sh;
2503 sh = list_entry(l, struct stripe_head, lru);
2504 list_del_init(l);
2505 clear_bit(STRIPE_DELAYED, &sh->state);
2506 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
2507 atomic_inc(&conf->preread_active_stripes);
2508 list_add_tail(&sh->lru, &conf->handle_list);
2509 }
2510 }
2511 }
2512
2513 static void activate_bit_delay(raid5_conf_t *conf)
2514 {
2515 /* device_lock is held */
2516 struct list_head head;
2517 list_add(&head, &conf->bitmap_list);
2518 list_del_init(&conf->bitmap_list);
2519 while (!list_empty(&head)) {
2520 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
2521 list_del_init(&sh->lru);
2522 atomic_inc(&sh->count);
2523 __release_stripe(conf, sh);
2524 }
2525 }
2526
2527 static void unplug_slaves(mddev_t *mddev)
2528 {
2529 raid5_conf_t *conf = mddev_to_conf(mddev);
2530 int i;
2531
2532 rcu_read_lock();
2533 for (i=0; i<mddev->raid_disks; i++) {
2534 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
2535 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
2536 request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
2537
2538 atomic_inc(&rdev->nr_pending);
2539 rcu_read_unlock();
2540
2541 if (r_queue->unplug_fn)
2542 r_queue->unplug_fn(r_queue);
2543
2544 rdev_dec_pending(rdev, mddev);
2545 rcu_read_lock();
2546 }
2547 }
2548 rcu_read_unlock();
2549 }
2550
2551 static void raid5_unplug_device(request_queue_t *q)
2552 {
2553 mddev_t *mddev = q->queuedata;
2554 raid5_conf_t *conf = mddev_to_conf(mddev);
2555 unsigned long flags;
2556
2557 spin_lock_irqsave(&conf->device_lock, flags);
2558
2559 if (blk_remove_plug(q)) {
2560 conf->seq_flush++;
2561 raid5_activate_delayed(conf);
2562 }
2563 md_wakeup_thread(mddev->thread);
2564
2565 spin_unlock_irqrestore(&conf->device_lock, flags);
2566
2567 unplug_slaves(mddev);
2568 }
2569
2570 static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
2571 sector_t *error_sector)
2572 {
2573 mddev_t *mddev = q->queuedata;
2574 raid5_conf_t *conf = mddev_to_conf(mddev);
2575 int i, ret = 0;
2576
2577 rcu_read_lock();
2578 for (i=0; i<mddev->raid_disks && ret == 0; i++) {
2579 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
2580 if (rdev && !test_bit(Faulty, &rdev->flags)) {
2581 struct block_device *bdev = rdev->bdev;
2582 request_queue_t *r_queue = bdev_get_queue(bdev);
2583
2584 if (!r_queue->issue_flush_fn)
2585 ret = -EOPNOTSUPP;
2586 else {
2587 atomic_inc(&rdev->nr_pending);
2588 rcu_read_unlock();
2589 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
2590 error_sector);
2591 rdev_dec_pending(rdev, mddev);
2592 rcu_read_lock();
2593 }
2594 }
2595 }
2596 rcu_read_unlock();
2597 return ret;
2598 }
2599
2600 static int make_request(request_queue_t *q, struct bio * bi)
2601 {
2602 mddev_t *mddev = q->queuedata;
2603 raid5_conf_t *conf = mddev_to_conf(mddev);
2604 unsigned int dd_idx, pd_idx;
2605 sector_t new_sector;
2606 sector_t logical_sector, last_sector;
2607 struct stripe_head *sh;
2608 const int rw = bio_data_dir(bi);
2609 int remaining;
2610
2611 if (unlikely(bio_barrier(bi))) {
2612 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
2613 return 0;
2614 }
2615
2616 md_write_start(mddev, bi);
2617
2618 disk_stat_inc(mddev->gendisk, ios[rw]);
2619 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
2620
2621 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
2622 last_sector = bi->bi_sector + (bi->bi_size>>9);
2623 bi->bi_next = NULL;
2624 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
2625
2626 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
2627 DEFINE_WAIT(w);
2628 int disks, data_disks;
2629
2630 retry:
2631 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
2632 if (likely(conf->expand_progress == MaxSector))
2633 disks = conf->raid_disks;
2634 else {
2635 /* spinlock is needed as expand_progress may be
2636 * 64bit on a 32bit platform, and so it might be
2637 * possible to see a half-updated value
2638 * Ofcourse expand_progress could change after
2639 * the lock is dropped, so once we get a reference
2640 * to the stripe that we think it is, we will have
2641 * to check again.
2642 */
2643 spin_lock_irq(&conf->device_lock);
2644 disks = conf->raid_disks;
2645 if (logical_sector >= conf->expand_progress)
2646 disks = conf->previous_raid_disks;
2647 else {
2648 if (logical_sector >= conf->expand_lo) {
2649 spin_unlock_irq(&conf->device_lock);
2650 schedule();
2651 goto retry;
2652 }
2653 }
2654 spin_unlock_irq(&conf->device_lock);
2655 }
2656 data_disks = disks - conf->max_degraded;
2657
2658 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
2659 &dd_idx, &pd_idx, conf);
2660 PRINTK("raid5: make_request, sector %llu logical %llu\n",
2661 (unsigned long long)new_sector,
2662 (unsigned long long)logical_sector);
2663
2664 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
2665 if (sh) {
2666 if (unlikely(conf->expand_progress != MaxSector)) {
2667 /* expansion might have moved on while waiting for a
2668 * stripe, so we must do the range check again.
2669 * Expansion could still move past after this
2670 * test, but as we are holding a reference to
2671 * 'sh', we know that if that happens,
2672 * STRIPE_EXPANDING will get set and the expansion
2673 * won't proceed until we finish with the stripe.
2674 */
2675 int must_retry = 0;
2676 spin_lock_irq(&conf->device_lock);
2677 if (logical_sector < conf->expand_progress &&
2678 disks == conf->previous_raid_disks)
2679 /* mismatch, need to try again */
2680 must_retry = 1;
2681 spin_unlock_irq(&conf->device_lock);
2682 if (must_retry) {
2683 release_stripe(sh);
2684 goto retry;
2685 }
2686 }
2687 /* FIXME what if we get a false positive because these
2688 * are being updated.
2689 */
2690 if (logical_sector >= mddev->suspend_lo &&
2691 logical_sector < mddev->suspend_hi) {
2692 release_stripe(sh);
2693 schedule();
2694 goto retry;
2695 }
2696
2697 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
2698 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
2699 /* Stripe is busy expanding or
2700 * add failed due to overlap. Flush everything
2701 * and wait a while
2702 */
2703 raid5_unplug_device(mddev->queue);
2704 release_stripe(sh);
2705 schedule();
2706 goto retry;
2707 }
2708 finish_wait(&conf->wait_for_overlap, &w);
2709 handle_stripe(sh, NULL);
2710 release_stripe(sh);
2711 } else {
2712 /* cannot get stripe for read-ahead, just give-up */
2713 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2714 finish_wait(&conf->wait_for_overlap, &w);
2715 break;
2716 }
2717
2718 }
2719 spin_lock_irq(&conf->device_lock);
2720 remaining = --bi->bi_phys_segments;
2721 spin_unlock_irq(&conf->device_lock);
2722 if (remaining == 0) {
2723 int bytes = bi->bi_size;
2724
2725 if ( rw == WRITE )
2726 md_write_end(mddev);
2727 bi->bi_size = 0;
2728 bi->bi_end_io(bi, bytes, 0);
2729 }
2730 return 0;
2731 }
2732
2733 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
2734 {
2735 /* reshaping is quite different to recovery/resync so it is
2736 * handled quite separately ... here.
2737 *
2738 * On each call to sync_request, we gather one chunk worth of
2739 * destination stripes and flag them as expanding.
2740 * Then we find all the source stripes and request reads.
2741 * As the reads complete, handle_stripe will copy the data
2742 * into the destination stripe and release that stripe.
2743 */
2744 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
2745 struct stripe_head *sh;
2746 int pd_idx;
2747 sector_t first_sector, last_sector;
2748 int raid_disks;
2749 int data_disks;
2750 int i;
2751 int dd_idx;
2752 sector_t writepos, safepos, gap;
2753
2754 if (sector_nr == 0 &&
2755 conf->expand_progress != 0) {
2756 /* restarting in the middle, skip the initial sectors */
2757 sector_nr = conf->expand_progress;
2758 sector_div(sector_nr, conf->raid_disks-1);
2759 *skipped = 1;
2760 return sector_nr;
2761 }
2762
2763 /* we update the metadata when there is more than 3Meg
2764 * in the block range (that is rather arbitrary, should
2765 * probably be time based) or when the data about to be
2766 * copied would over-write the source of the data at
2767 * the front of the range.
2768 * i.e. one new_stripe forward from expand_progress new_maps
2769 * to after where expand_lo old_maps to
2770 */
2771 writepos = conf->expand_progress +
2772 conf->chunk_size/512*(conf->raid_disks-1);
2773 sector_div(writepos, conf->raid_disks-1);
2774 safepos = conf->expand_lo;
2775 sector_div(safepos, conf->previous_raid_disks-1);
2776 gap = conf->expand_progress - conf->expand_lo;
2777
2778 if (writepos >= safepos ||
2779 gap > (conf->raid_disks-1)*3000*2 /*3Meg*/) {
2780 /* Cannot proceed until we've updated the superblock... */
2781 wait_event(conf->wait_for_overlap,
2782 atomic_read(&conf->reshape_stripes)==0);
2783 mddev->reshape_position = conf->expand_progress;
2784 mddev->sb_dirty = 1;
2785 md_wakeup_thread(mddev->thread);
2786 wait_event(mddev->sb_wait, mddev->sb_dirty == 0 ||
2787 kthread_should_stop());
2788 spin_lock_irq(&conf->device_lock);
2789 conf->expand_lo = mddev->reshape_position;
2790 spin_unlock_irq(&conf->device_lock);
2791 wake_up(&conf->wait_for_overlap);
2792 }
2793
2794 for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
2795 int j;
2796 int skipped = 0;
2797 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
2798 sh = get_active_stripe(conf, sector_nr+i,
2799 conf->raid_disks, pd_idx, 0);
2800 set_bit(STRIPE_EXPANDING, &sh->state);
2801 atomic_inc(&conf->reshape_stripes);
2802 /* If any of this stripe is beyond the end of the old
2803 * array, then we need to zero those blocks
2804 */
2805 for (j=sh->disks; j--;) {
2806 sector_t s;
2807 if (j == sh->pd_idx)
2808 continue;
2809 s = compute_blocknr(sh, j);
2810 if (s < (mddev->array_size<<1)) {
2811 skipped = 1;
2812 continue;
2813 }
2814 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
2815 set_bit(R5_Expanded, &sh->dev[j].flags);
2816 set_bit(R5_UPTODATE, &sh->dev[j].flags);
2817 }
2818 if (!skipped) {
2819 set_bit(STRIPE_EXPAND_READY, &sh->state);
2820 set_bit(STRIPE_HANDLE, &sh->state);
2821 }
2822 release_stripe(sh);
2823 }
2824 spin_lock_irq(&conf->device_lock);
2825 conf->expand_progress = (sector_nr + i)*(conf->raid_disks-1);
2826 spin_unlock_irq(&conf->device_lock);
2827 /* Ok, those stripe are ready. We can start scheduling
2828 * reads on the source stripes.
2829 * The source stripes are determined by mapping the first and last
2830 * block on the destination stripes.
2831 */
2832 raid_disks = conf->previous_raid_disks;
2833 data_disks = raid_disks - 1;
2834 first_sector =
2835 raid5_compute_sector(sector_nr*(conf->raid_disks-1),
2836 raid_disks, data_disks,
2837 &dd_idx, &pd_idx, conf);
2838 last_sector =
2839 raid5_compute_sector((sector_nr+conf->chunk_size/512)
2840 *(conf->raid_disks-1) -1,
2841 raid_disks, data_disks,
2842 &dd_idx, &pd_idx, conf);
2843 if (last_sector >= (mddev->size<<1))
2844 last_sector = (mddev->size<<1)-1;
2845 while (first_sector <= last_sector) {
2846 pd_idx = stripe_to_pdidx(first_sector, conf, conf->previous_raid_disks);
2847 sh = get_active_stripe(conf, first_sector,
2848 conf->previous_raid_disks, pd_idx, 0);
2849 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2850 set_bit(STRIPE_HANDLE, &sh->state);
2851 release_stripe(sh);
2852 first_sector += STRIPE_SECTORS;
2853 }
2854 return conf->chunk_size>>9;
2855 }
2856
2857 /* FIXME go_faster isn't used */
2858 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
2859 {
2860 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
2861 struct stripe_head *sh;
2862 int pd_idx;
2863 int raid_disks = conf->raid_disks;
2864 sector_t max_sector = mddev->size << 1;
2865 int sync_blocks;
2866 int still_degraded = 0;
2867 int i;
2868
2869 if (sector_nr >= max_sector) {
2870 /* just being told to finish up .. nothing much to do */
2871 unplug_slaves(mddev);
2872 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2873 end_reshape(conf);
2874 return 0;
2875 }
2876
2877 if (mddev->curr_resync < max_sector) /* aborted */
2878 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2879 &sync_blocks, 1);
2880 else /* completed sync */
2881 conf->fullsync = 0;
2882 bitmap_close_sync(mddev->bitmap);
2883
2884 return 0;
2885 }
2886
2887 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2888 return reshape_request(mddev, sector_nr, skipped);
2889
2890 /* if there is too many failed drives and we are trying
2891 * to resync, then assert that we are finished, because there is
2892 * nothing we can do.
2893 */
2894 if (mddev->degraded >= conf->max_degraded &&
2895 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2896 sector_t rv = (mddev->size << 1) - sector_nr;
2897 *skipped = 1;
2898 return rv;
2899 }
2900 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2901 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2902 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
2903 /* we can skip this block, and probably more */
2904 sync_blocks /= STRIPE_SECTORS;
2905 *skipped = 1;
2906 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
2907 }
2908
2909 pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
2910 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
2911 if (sh == NULL) {
2912 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
2913 /* make sure we don't swamp the stripe cache if someone else
2914 * is trying to get access
2915 */
2916 schedule_timeout_uninterruptible(1);
2917 }
2918 /* Need to check if array will still be degraded after recovery/resync
2919 * We don't need to check the 'failed' flag as when that gets set,
2920 * recovery aborts.
2921 */
2922 for (i=0; i<mddev->raid_disks; i++)
2923 if (conf->disks[i].rdev == NULL)
2924 still_degraded = 1;
2925
2926 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
2927
2928 spin_lock(&sh->lock);
2929 set_bit(STRIPE_SYNCING, &sh->state);
2930 clear_bit(STRIPE_INSYNC, &sh->state);
2931 spin_unlock(&sh->lock);
2932
2933 handle_stripe(sh, NULL);
2934 release_stripe(sh);
2935
2936 return STRIPE_SECTORS;
2937 }
2938
2939 /*
2940 * This is our raid5 kernel thread.
2941 *
2942 * We scan the hash table for stripes which can be handled now.
2943 * During the scan, completed stripes are saved for us by the interrupt
2944 * handler, so that they will not have to wait for our next wakeup.
2945 */
2946 static void raid5d (mddev_t *mddev)
2947 {
2948 struct stripe_head *sh;
2949 raid5_conf_t *conf = mddev_to_conf(mddev);
2950 int handled;
2951
2952 PRINTK("+++ raid5d active\n");
2953
2954 md_check_recovery(mddev);
2955
2956 handled = 0;
2957 spin_lock_irq(&conf->device_lock);
2958 while (1) {
2959 struct list_head *first;
2960
2961 if (conf->seq_flush != conf->seq_write) {
2962 int seq = conf->seq_flush;
2963 spin_unlock_irq(&conf->device_lock);
2964 bitmap_unplug(mddev->bitmap);
2965 spin_lock_irq(&conf->device_lock);
2966 conf->seq_write = seq;
2967 activate_bit_delay(conf);
2968 }
2969
2970 if (list_empty(&conf->handle_list) &&
2971 atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
2972 !blk_queue_plugged(mddev->queue) &&
2973 !list_empty(&conf->delayed_list))
2974 raid5_activate_delayed(conf);
2975
2976 if (list_empty(&conf->handle_list))
2977 break;
2978
2979 first = conf->handle_list.next;
2980 sh = list_entry(first, struct stripe_head, lru);
2981
2982 list_del_init(first);
2983 atomic_inc(&sh->count);
2984 BUG_ON(atomic_read(&sh->count)!= 1);
2985 spin_unlock_irq(&conf->device_lock);
2986
2987 handled++;
2988 handle_stripe(sh, conf->spare_page);
2989 release_stripe(sh);
2990
2991 spin_lock_irq(&conf->device_lock);
2992 }
2993 PRINTK("%d stripes handled\n", handled);
2994
2995 spin_unlock_irq(&conf->device_lock);
2996
2997 unplug_slaves(mddev);
2998
2999 PRINTK("--- raid5d inactive\n");
3000 }
3001
3002 static ssize_t
3003 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
3004 {
3005 raid5_conf_t *conf = mddev_to_conf(mddev);
3006 if (conf)
3007 return sprintf(page, "%d\n", conf->max_nr_stripes);
3008 else
3009 return 0;
3010 }
3011
3012 static ssize_t
3013 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
3014 {
3015 raid5_conf_t *conf = mddev_to_conf(mddev);
3016 char *end;
3017 int new;
3018 if (len >= PAGE_SIZE)
3019 return -EINVAL;
3020 if (!conf)
3021 return -ENODEV;
3022
3023 new = simple_strtoul(page, &end, 10);
3024 if (!*page || (*end && *end != '\n') )
3025 return -EINVAL;
3026 if (new <= 16 || new > 32768)
3027 return -EINVAL;
3028 while (new < conf->max_nr_stripes) {
3029 if (drop_one_stripe(conf))
3030 conf->max_nr_stripes--;
3031 else
3032 break;
3033 }
3034 while (new > conf->max_nr_stripes) {
3035 if (grow_one_stripe(conf))
3036 conf->max_nr_stripes++;
3037 else break;
3038 }
3039 return len;
3040 }
3041
3042 static struct md_sysfs_entry
3043 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
3044 raid5_show_stripe_cache_size,
3045 raid5_store_stripe_cache_size);
3046
3047 static ssize_t
3048 stripe_cache_active_show(mddev_t *mddev, char *page)
3049 {
3050 raid5_conf_t *conf = mddev_to_conf(mddev);
3051 if (conf)
3052 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
3053 else
3054 return 0;
3055 }
3056
3057 static struct md_sysfs_entry
3058 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
3059
3060 static struct attribute *raid5_attrs[] = {
3061 &raid5_stripecache_size.attr,
3062 &raid5_stripecache_active.attr,
3063 NULL,
3064 };
3065 static struct attribute_group raid5_attrs_group = {
3066 .name = NULL,
3067 .attrs = raid5_attrs,
3068 };
3069
3070 static int run(mddev_t *mddev)
3071 {
3072 raid5_conf_t *conf;
3073 int raid_disk, memory;
3074 mdk_rdev_t *rdev;
3075 struct disk_info *disk;
3076 struct list_head *tmp;
3077
3078 if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
3079 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
3080 mdname(mddev), mddev->level);
3081 return -EIO;
3082 }
3083
3084 if (mddev->reshape_position != MaxSector) {
3085 /* Check that we can continue the reshape.
3086 * Currently only disks can change, it must
3087 * increase, and we must be past the point where
3088 * a stripe over-writes itself
3089 */
3090 sector_t here_new, here_old;
3091 int old_disks;
3092
3093 if (mddev->new_level != mddev->level ||
3094 mddev->new_layout != mddev->layout ||
3095 mddev->new_chunk != mddev->chunk_size) {
3096 printk(KERN_ERR "raid5: %s: unsupported reshape required - aborting.\n",
3097 mdname(mddev));
3098 return -EINVAL;
3099 }
3100 if (mddev->delta_disks <= 0) {
3101 printk(KERN_ERR "raid5: %s: unsupported reshape (reduce disks) required - aborting.\n",
3102 mdname(mddev));
3103 return -EINVAL;
3104 }
3105 old_disks = mddev->raid_disks - mddev->delta_disks;
3106 /* reshape_position must be on a new-stripe boundary, and one
3107 * further up in new geometry must map after here in old geometry.
3108 */
3109 here_new = mddev->reshape_position;
3110 if (sector_div(here_new, (mddev->chunk_size>>9)*(mddev->raid_disks-1))) {
3111 printk(KERN_ERR "raid5: reshape_position not on a stripe boundary\n");
3112 return -EINVAL;
3113 }
3114 /* here_new is the stripe we will write to */
3115 here_old = mddev->reshape_position;
3116 sector_div(here_old, (mddev->chunk_size>>9)*(old_disks-1));
3117 /* here_old is the first stripe that we might need to read from */
3118 if (here_new >= here_old) {
3119 /* Reading from the same stripe as writing to - bad */
3120 printk(KERN_ERR "raid5: reshape_position too early for auto-recovery - aborting.\n");
3121 return -EINVAL;
3122 }
3123 printk(KERN_INFO "raid5: reshape will continue\n");
3124 /* OK, we should be able to continue; */
3125 }
3126
3127
3128 mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
3129 if ((conf = mddev->private) == NULL)
3130 goto abort;
3131 if (mddev->reshape_position == MaxSector) {
3132 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
3133 } else {
3134 conf->raid_disks = mddev->raid_disks;
3135 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
3136 }
3137
3138 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
3139 GFP_KERNEL);
3140 if (!conf->disks)
3141 goto abort;
3142
3143 conf->mddev = mddev;
3144
3145 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
3146 goto abort;
3147
3148 if (mddev->level == 6) {
3149 conf->spare_page = alloc_page(GFP_KERNEL);
3150 if (!conf->spare_page)
3151 goto abort;
3152 }
3153 spin_lock_init(&conf->device_lock);
3154 init_waitqueue_head(&conf->wait_for_stripe);
3155 init_waitqueue_head(&conf->wait_for_overlap);
3156 INIT_LIST_HEAD(&conf->handle_list);
3157 INIT_LIST_HEAD(&conf->delayed_list);
3158 INIT_LIST_HEAD(&conf->bitmap_list);
3159 INIT_LIST_HEAD(&conf->inactive_list);
3160 atomic_set(&conf->active_stripes, 0);
3161 atomic_set(&conf->preread_active_stripes, 0);
3162
3163 PRINTK("raid5: run(%s) called.\n", mdname(mddev));
3164
3165 ITERATE_RDEV(mddev,rdev,tmp) {
3166 raid_disk = rdev->raid_disk;
3167 if (raid_disk >= conf->raid_disks
3168 || raid_disk < 0)
3169 continue;
3170 disk = conf->disks + raid_disk;
3171
3172 disk->rdev = rdev;
3173
3174 if (test_bit(In_sync, &rdev->flags)) {
3175 char b[BDEVNAME_SIZE];
3176 printk(KERN_INFO "raid5: device %s operational as raid"
3177 " disk %d\n", bdevname(rdev->bdev,b),
3178 raid_disk);
3179 conf->working_disks++;
3180 }
3181 }
3182
3183 /*
3184 * 0 for a fully functional array, 1 or 2 for a degraded array.
3185 */
3186 mddev->degraded = conf->failed_disks = conf->raid_disks - conf->working_disks;
3187 conf->mddev = mddev;
3188 conf->chunk_size = mddev->chunk_size;
3189 conf->level = mddev->level;
3190 if (conf->level == 6)
3191 conf->max_degraded = 2;
3192 else
3193 conf->max_degraded = 1;
3194 conf->algorithm = mddev->layout;
3195 conf->max_nr_stripes = NR_STRIPES;
3196 conf->expand_progress = mddev->reshape_position;
3197
3198 /* device size must be a multiple of chunk size */
3199 mddev->size &= ~(mddev->chunk_size/1024 -1);
3200 mddev->resync_max_sectors = mddev->size << 1;
3201
3202 if (conf->level == 6 && conf->raid_disks < 4) {
3203 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
3204 mdname(mddev), conf->raid_disks);
3205 goto abort;
3206 }
3207 if (!conf->chunk_size || conf->chunk_size % 4) {
3208 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
3209 conf->chunk_size, mdname(mddev));
3210 goto abort;
3211 }
3212 if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
3213 printk(KERN_ERR
3214 "raid5: unsupported parity algorithm %d for %s\n",
3215 conf->algorithm, mdname(mddev));
3216 goto abort;
3217 }
3218 if (mddev->degraded > conf->max_degraded) {
3219 printk(KERN_ERR "raid5: not enough operational devices for %s"
3220 " (%d/%d failed)\n",
3221 mdname(mddev), conf->failed_disks, conf->raid_disks);
3222 goto abort;
3223 }
3224
3225 if (mddev->degraded > 0 &&
3226 mddev->recovery_cp != MaxSector) {
3227 if (mddev->ok_start_degraded)
3228 printk(KERN_WARNING
3229 "raid5: starting dirty degraded array: %s"
3230 "- data corruption possible.\n",
3231 mdname(mddev));
3232 else {
3233 printk(KERN_ERR
3234 "raid5: cannot start dirty degraded array for %s\n",
3235 mdname(mddev));
3236 goto abort;
3237 }
3238 }
3239
3240 {
3241 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
3242 if (!mddev->thread) {
3243 printk(KERN_ERR
3244 "raid5: couldn't allocate thread for %s\n",
3245 mdname(mddev));
3246 goto abort;
3247 }
3248 }
3249 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
3250 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
3251 if (grow_stripes(conf, conf->max_nr_stripes)) {
3252 printk(KERN_ERR
3253 "raid5: couldn't allocate %dkB for buffers\n", memory);
3254 shrink_stripes(conf);
3255 md_unregister_thread(mddev->thread);
3256 goto abort;
3257 } else
3258 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
3259 memory, mdname(mddev));
3260
3261 if (mddev->degraded == 0)
3262 printk("raid5: raid level %d set %s active with %d out of %d"
3263 " devices, algorithm %d\n", conf->level, mdname(mddev),
3264 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
3265 conf->algorithm);
3266 else
3267 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
3268 " out of %d devices, algorithm %d\n", conf->level,
3269 mdname(mddev), mddev->raid_disks - mddev->degraded,
3270 mddev->raid_disks, conf->algorithm);
3271
3272 print_raid5_conf(conf);
3273
3274 if (conf->expand_progress != MaxSector) {
3275 printk("...ok start reshape thread\n");
3276 conf->expand_lo = conf->expand_progress;
3277 atomic_set(&conf->reshape_stripes, 0);
3278 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3279 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3280 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3281 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3282 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3283 "%s_reshape");
3284 }
3285
3286 /* read-ahead size must cover two whole stripes, which is
3287 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
3288 */
3289 {
3290 int data_disks = conf->previous_raid_disks - conf->max_degraded;
3291 int stripe = data_disks *
3292 (mddev->chunk_size / PAGE_SIZE);
3293 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3294 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3295 }
3296
3297 /* Ok, everything is just fine now */
3298 sysfs_create_group(&mddev->kobj, &raid5_attrs_group);
3299
3300 mddev->queue->unplug_fn = raid5_unplug_device;
3301 mddev->queue->issue_flush_fn = raid5_issue_flush;
3302 mddev->array_size = mddev->size * (conf->previous_raid_disks -
3303 conf->max_degraded);
3304
3305 return 0;
3306 abort:
3307 if (conf) {
3308 print_raid5_conf(conf);
3309 safe_put_page(conf->spare_page);
3310 kfree(conf->disks);
3311 kfree(conf->stripe_hashtbl);
3312 kfree(conf);
3313 }
3314 mddev->private = NULL;
3315 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
3316 return -EIO;
3317 }
3318
3319
3320
3321 static int stop(mddev_t *mddev)
3322 {
3323 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3324
3325 md_unregister_thread(mddev->thread);
3326 mddev->thread = NULL;
3327 shrink_stripes(conf);
3328 kfree(conf->stripe_hashtbl);
3329 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
3330 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
3331 kfree(conf->disks);
3332 kfree(conf);
3333 mddev->private = NULL;
3334 return 0;
3335 }
3336
3337 #if RAID5_DEBUG
3338 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
3339 {
3340 int i;
3341
3342 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
3343 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
3344 seq_printf(seq, "sh %llu, count %d.\n",
3345 (unsigned long long)sh->sector, atomic_read(&sh->count));
3346 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
3347 for (i = 0; i < sh->disks; i++) {
3348 seq_printf(seq, "(cache%d: %p %ld) ",
3349 i, sh->dev[i].page, sh->dev[i].flags);
3350 }
3351 seq_printf(seq, "\n");
3352 }
3353
3354 static void printall (struct seq_file *seq, raid5_conf_t *conf)
3355 {
3356 struct stripe_head *sh;
3357 struct hlist_node *hn;
3358 int i;
3359
3360 spin_lock_irq(&conf->device_lock);
3361 for (i = 0; i < NR_HASH; i++) {
3362 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
3363 if (sh->raid_conf != conf)
3364 continue;
3365 print_sh(seq, sh);
3366 }
3367 }
3368 spin_unlock_irq(&conf->device_lock);
3369 }
3370 #endif
3371
3372 static void status (struct seq_file *seq, mddev_t *mddev)
3373 {
3374 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3375 int i;
3376
3377 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
3378 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->working_disks);
3379 for (i = 0; i < conf->raid_disks; i++)
3380 seq_printf (seq, "%s",
3381 conf->disks[i].rdev &&
3382 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
3383 seq_printf (seq, "]");
3384 #if RAID5_DEBUG
3385 seq_printf (seq, "\n");
3386 printall(seq, conf);
3387 #endif
3388 }
3389
3390 static void print_raid5_conf (raid5_conf_t *conf)
3391 {
3392 int i;
3393 struct disk_info *tmp;
3394
3395 printk("RAID5 conf printout:\n");
3396 if (!conf) {
3397 printk("(conf==NULL)\n");
3398 return;
3399 }
3400 printk(" --- rd:%d wd:%d fd:%d\n", conf->raid_disks,
3401 conf->working_disks, conf->failed_disks);
3402
3403 for (i = 0; i < conf->raid_disks; i++) {
3404 char b[BDEVNAME_SIZE];
3405 tmp = conf->disks + i;
3406 if (tmp->rdev)
3407 printk(" disk %d, o:%d, dev:%s\n",
3408 i, !test_bit(Faulty, &tmp->rdev->flags),
3409 bdevname(tmp->rdev->bdev,b));
3410 }
3411 }
3412
3413 static int raid5_spare_active(mddev_t *mddev)
3414 {
3415 int i;
3416 raid5_conf_t *conf = mddev->private;
3417 struct disk_info *tmp;
3418
3419 for (i = 0; i < conf->raid_disks; i++) {
3420 tmp = conf->disks + i;
3421 if (tmp->rdev
3422 && !test_bit(Faulty, &tmp->rdev->flags)
3423 && !test_bit(In_sync, &tmp->rdev->flags)) {
3424 mddev->degraded--;
3425 conf->failed_disks--;
3426 conf->working_disks++;
3427 set_bit(In_sync, &tmp->rdev->flags);
3428 }
3429 }
3430 print_raid5_conf(conf);
3431 return 0;
3432 }
3433
3434 static int raid5_remove_disk(mddev_t *mddev, int number)
3435 {
3436 raid5_conf_t *conf = mddev->private;
3437 int err = 0;
3438 mdk_rdev_t *rdev;
3439 struct disk_info *p = conf->disks + number;
3440
3441 print_raid5_conf(conf);
3442 rdev = p->rdev;
3443 if (rdev) {
3444 if (test_bit(In_sync, &rdev->flags) ||
3445 atomic_read(&rdev->nr_pending)) {
3446 err = -EBUSY;
3447 goto abort;
3448 }
3449 p->rdev = NULL;
3450 synchronize_rcu();
3451 if (atomic_read(&rdev->nr_pending)) {
3452 /* lost the race, try later */
3453 err = -EBUSY;
3454 p->rdev = rdev;
3455 }
3456 }
3457 abort:
3458
3459 print_raid5_conf(conf);
3460 return err;
3461 }
3462
3463 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
3464 {
3465 raid5_conf_t *conf = mddev->private;
3466 int found = 0;
3467 int disk;
3468 struct disk_info *p;
3469
3470 if (mddev->degraded > conf->max_degraded)
3471 /* no point adding a device */
3472 return 0;
3473
3474 /*
3475 * find the disk ... but prefer rdev->saved_raid_disk
3476 * if possible.
3477 */
3478 if (rdev->saved_raid_disk >= 0 &&
3479 conf->disks[rdev->saved_raid_disk].rdev == NULL)
3480 disk = rdev->saved_raid_disk;
3481 else
3482 disk = 0;
3483 for ( ; disk < conf->raid_disks; disk++)
3484 if ((p=conf->disks + disk)->rdev == NULL) {
3485 clear_bit(In_sync, &rdev->flags);
3486 rdev->raid_disk = disk;
3487 found = 1;
3488 if (rdev->saved_raid_disk != disk)
3489 conf->fullsync = 1;
3490 rcu_assign_pointer(p->rdev, rdev);
3491 break;
3492 }
3493 print_raid5_conf(conf);
3494 return found;
3495 }
3496
3497 static int raid5_resize(mddev_t *mddev, sector_t sectors)
3498 {
3499 /* no resync is happening, and there is enough space
3500 * on all devices, so we can resize.
3501 * We need to make sure resync covers any new space.
3502 * If the array is shrinking we should possibly wait until
3503 * any io in the removed space completes, but it hardly seems
3504 * worth it.
3505 */
3506 raid5_conf_t *conf = mddev_to_conf(mddev);
3507
3508 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
3509 mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
3510 set_capacity(mddev->gendisk, mddev->array_size << 1);
3511 mddev->changed = 1;
3512 if (sectors/2 > mddev->size && mddev->recovery_cp == MaxSector) {
3513 mddev->recovery_cp = mddev->size << 1;
3514 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3515 }
3516 mddev->size = sectors /2;
3517 mddev->resync_max_sectors = sectors;
3518 return 0;
3519 }
3520
3521 #ifdef CONFIG_MD_RAID5_RESHAPE
3522 static int raid5_check_reshape(mddev_t *mddev)
3523 {
3524 raid5_conf_t *conf = mddev_to_conf(mddev);
3525 int err;
3526
3527 if (mddev->delta_disks < 0 ||
3528 mddev->new_level != mddev->level)
3529 return -EINVAL; /* Cannot shrink array or change level yet */
3530 if (mddev->delta_disks == 0)
3531 return 0; /* nothing to do */
3532
3533 /* Can only proceed if there are plenty of stripe_heads.
3534 * We need a minimum of one full stripe,, and for sensible progress
3535 * it is best to have about 4 times that.
3536 * If we require 4 times, then the default 256 4K stripe_heads will
3537 * allow for chunk sizes up to 256K, which is probably OK.
3538 * If the chunk size is greater, user-space should request more
3539 * stripe_heads first.
3540 */
3541 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
3542 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
3543 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
3544 (mddev->chunk_size / STRIPE_SIZE)*4);
3545 return -ENOSPC;
3546 }
3547
3548 err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
3549 if (err)
3550 return err;
3551
3552 /* looks like we might be able to manage this */
3553 return 0;
3554 }
3555
3556 static int raid5_start_reshape(mddev_t *mddev)
3557 {
3558 raid5_conf_t *conf = mddev_to_conf(mddev);
3559 mdk_rdev_t *rdev;
3560 struct list_head *rtmp;
3561 int spares = 0;
3562 int added_devices = 0;
3563
3564 if (mddev->degraded ||
3565 test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
3566 return -EBUSY;
3567
3568 ITERATE_RDEV(mddev, rdev, rtmp)
3569 if (rdev->raid_disk < 0 &&
3570 !test_bit(Faulty, &rdev->flags))
3571 spares++;
3572
3573 if (spares < mddev->delta_disks-1)
3574 /* Not enough devices even to make a degraded array
3575 * of that size
3576 */
3577 return -EINVAL;
3578
3579 atomic_set(&conf->reshape_stripes, 0);
3580 spin_lock_irq(&conf->device_lock);
3581 conf->previous_raid_disks = conf->raid_disks;
3582 conf->raid_disks += mddev->delta_disks;
3583 conf->expand_progress = 0;
3584 conf->expand_lo = 0;
3585 spin_unlock_irq(&conf->device_lock);
3586
3587 /* Add some new drives, as many as will fit.
3588 * We know there are enough to make the newly sized array work.
3589 */
3590 ITERATE_RDEV(mddev, rdev, rtmp)
3591 if (rdev->raid_disk < 0 &&
3592 !test_bit(Faulty, &rdev->flags)) {
3593 if (raid5_add_disk(mddev, rdev)) {
3594 char nm[20];
3595 set_bit(In_sync, &rdev->flags);
3596 conf->working_disks++;
3597 added_devices++;
3598 rdev->recovery_offset = 0;
3599 sprintf(nm, "rd%d", rdev->raid_disk);
3600 sysfs_create_link(&mddev->kobj, &rdev->kobj, nm);
3601 } else
3602 break;
3603 }
3604
3605 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
3606 mddev->raid_disks = conf->raid_disks;
3607 mddev->reshape_position = 0;
3608 mddev->sb_dirty = 1;
3609
3610 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3611 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3612 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3613 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3614 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3615 "%s_reshape");
3616 if (!mddev->sync_thread) {
3617 mddev->recovery = 0;
3618 spin_lock_irq(&conf->device_lock);
3619 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
3620 conf->expand_progress = MaxSector;
3621 spin_unlock_irq(&conf->device_lock);
3622 return -EAGAIN;
3623 }
3624 md_wakeup_thread(mddev->sync_thread);
3625 md_new_event(mddev);
3626 return 0;
3627 }
3628 #endif
3629
3630 static void end_reshape(raid5_conf_t *conf)
3631 {
3632 struct block_device *bdev;
3633
3634 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
3635 conf->mddev->array_size = conf->mddev->size * (conf->raid_disks-1);
3636 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
3637 conf->mddev->changed = 1;
3638
3639 bdev = bdget_disk(conf->mddev->gendisk, 0);
3640 if (bdev) {
3641 mutex_lock(&bdev->bd_inode->i_mutex);
3642 i_size_write(bdev->bd_inode, conf->mddev->array_size << 10);
3643 mutex_unlock(&bdev->bd_inode->i_mutex);
3644 bdput(bdev);
3645 }
3646 spin_lock_irq(&conf->device_lock);
3647 conf->expand_progress = MaxSector;
3648 spin_unlock_irq(&conf->device_lock);
3649 conf->mddev->reshape_position = MaxSector;
3650
3651 /* read-ahead size must cover two whole stripes, which is
3652 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
3653 */
3654 {
3655 int data_disks = conf->previous_raid_disks - conf->max_degraded;
3656 int stripe = data_disks *
3657 (conf->mddev->chunk_size / PAGE_SIZE);
3658 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3659 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3660 }
3661 }
3662 }
3663
3664 static void raid5_quiesce(mddev_t *mddev, int state)
3665 {
3666 raid5_conf_t *conf = mddev_to_conf(mddev);
3667
3668 switch(state) {
3669 case 2: /* resume for a suspend */
3670 wake_up(&conf->wait_for_overlap);
3671 break;
3672
3673 case 1: /* stop all writes */
3674 spin_lock_irq(&conf->device_lock);
3675 conf->quiesce = 1;
3676 wait_event_lock_irq(conf->wait_for_stripe,
3677 atomic_read(&conf->active_stripes) == 0,
3678 conf->device_lock, /* nothing */);
3679 spin_unlock_irq(&conf->device_lock);
3680 break;
3681
3682 case 0: /* re-enable writes */
3683 spin_lock_irq(&conf->device_lock);
3684 conf->quiesce = 0;
3685 wake_up(&conf->wait_for_stripe);
3686 wake_up(&conf->wait_for_overlap);
3687 spin_unlock_irq(&conf->device_lock);
3688 break;
3689 }
3690 }
3691
3692 static struct mdk_personality raid6_personality =
3693 {
3694 .name = "raid6",
3695 .level = 6,
3696 .owner = THIS_MODULE,
3697 .make_request = make_request,
3698 .run = run,
3699 .stop = stop,
3700 .status = status,
3701 .error_handler = error,
3702 .hot_add_disk = raid5_add_disk,
3703 .hot_remove_disk= raid5_remove_disk,
3704 .spare_active = raid5_spare_active,
3705 .sync_request = sync_request,
3706 .resize = raid5_resize,
3707 .quiesce = raid5_quiesce,
3708 };
3709 static struct mdk_personality raid5_personality =
3710 {
3711 .name = "raid5",
3712 .level = 5,
3713 .owner = THIS_MODULE,
3714 .make_request = make_request,
3715 .run = run,
3716 .stop = stop,
3717 .status = status,
3718 .error_handler = error,
3719 .hot_add_disk = raid5_add_disk,
3720 .hot_remove_disk= raid5_remove_disk,
3721 .spare_active = raid5_spare_active,
3722 .sync_request = sync_request,
3723 .resize = raid5_resize,
3724 #ifdef CONFIG_MD_RAID5_RESHAPE
3725 .check_reshape = raid5_check_reshape,
3726 .start_reshape = raid5_start_reshape,
3727 #endif
3728 .quiesce = raid5_quiesce,
3729 };
3730
3731 static struct mdk_personality raid4_personality =
3732 {
3733 .name = "raid4",
3734 .level = 4,
3735 .owner = THIS_MODULE,
3736 .make_request = make_request,
3737 .run = run,
3738 .stop = stop,
3739 .status = status,
3740 .error_handler = error,
3741 .hot_add_disk = raid5_add_disk,
3742 .hot_remove_disk= raid5_remove_disk,
3743 .spare_active = raid5_spare_active,
3744 .sync_request = sync_request,
3745 .resize = raid5_resize,
3746 .quiesce = raid5_quiesce,
3747 };
3748
3749 static int __init raid5_init(void)
3750 {
3751 int e;
3752
3753 e = raid6_select_algo();
3754 if ( e )
3755 return e;
3756 register_md_personality(&raid6_personality);
3757 register_md_personality(&raid5_personality);
3758 register_md_personality(&raid4_personality);
3759 return 0;
3760 }
3761
3762 static void raid5_exit(void)
3763 {
3764 unregister_md_personality(&raid6_personality);
3765 unregister_md_personality(&raid5_personality);
3766 unregister_md_personality(&raid4_personality);
3767 }
3768
3769 module_init(raid5_init);
3770 module_exit(raid5_exit);
3771 MODULE_LICENSE("GPL");
3772 MODULE_ALIAS("md-personality-4"); /* RAID5 */
3773 MODULE_ALIAS("md-raid5");
3774 MODULE_ALIAS("md-raid4");
3775 MODULE_ALIAS("md-level-5");
3776 MODULE_ALIAS("md-level-4");
3777 MODULE_ALIAS("md-personality-8"); /* RAID6 */
3778 MODULE_ALIAS("md-raid6");
3779 MODULE_ALIAS("md-level-6");
3780
3781 /* This used to be two separate modules, they were: */
3782 MODULE_ALIAS("raid5");
3783 MODULE_ALIAS("raid6");