blob: 77610b98d4e0e82bd7bf999c94650eeb9e63689d [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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 *
6 * RAID-5 management functions.
7 *
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2, or (at your option)
11 * any later version.
12 *
13 * You should have received a copy of the GNU General Public License
14 * (for example /usr/src/linux/COPYING); if not, write to the Free
15 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
16 */
17
18
19#include <linux/config.h>
20#include <linux/module.h>
21#include <linux/slab.h>
22#include <linux/raid/raid5.h>
23#include <linux/highmem.h>
24#include <linux/bitops.h>
25#include <asm/atomic.h>
26
NeilBrown72626682005-09-09 16:23:54 -070027#include <linux/raid/bitmap.h>
28
Linus Torvalds1da177e2005-04-16 15:20:36 -070029/*
30 * Stripe cache
31 */
32
33#define NR_STRIPES 256
34#define STRIPE_SIZE PAGE_SIZE
35#define STRIPE_SHIFT (PAGE_SHIFT - 9)
36#define STRIPE_SECTORS (STRIPE_SIZE>>9)
37#define IO_THRESHOLD 1
38#define HASH_PAGES 1
39#define HASH_PAGES_ORDER 0
40#define NR_HASH (HASH_PAGES * PAGE_SIZE / sizeof(struct stripe_head *))
41#define HASH_MASK (NR_HASH - 1)
42
43#define stripe_hash(conf, sect) ((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK])
44
45/* bio's attached to a stripe+device for I/O are linked together in bi_sector
46 * order without overlap. There may be several bio's per stripe+device, and
47 * a bio could span several devices.
48 * When walking this list for a particular stripe+device, we must never proceed
49 * beyond a bio that extends past this device, as the next bio might no longer
50 * be valid.
51 * This macro is used to determine the 'next' bio in the list, given the sector
52 * of the current stripe+device
53 */
54#define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
55/*
56 * The following can be used to debug the driver
57 */
58#define RAID5_DEBUG 0
59#define RAID5_PARANOIA 1
60#if RAID5_PARANOIA && defined(CONFIG_SMP)
61# define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
62#else
63# define CHECK_DEVLOCK()
64#endif
65
66#define PRINTK(x...) ((void)(RAID5_DEBUG && printk(x)))
67#if RAID5_DEBUG
68#define inline
69#define __inline__
70#endif
71
72static void print_raid5_conf (raid5_conf_t *conf);
73
74static inline void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
75{
76 if (atomic_dec_and_test(&sh->count)) {
77 if (!list_empty(&sh->lru))
78 BUG();
79 if (atomic_read(&conf->active_stripes)==0)
80 BUG();
81 if (test_bit(STRIPE_HANDLE, &sh->state)) {
82 if (test_bit(STRIPE_DELAYED, &sh->state))
83 list_add_tail(&sh->lru, &conf->delayed_list);
NeilBrown72626682005-09-09 16:23:54 -070084 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
85 conf->seq_write == sh->bm_seq)
86 list_add_tail(&sh->lru, &conf->bitmap_list);
87 else {
88 clear_bit(STRIPE_BIT_DELAY, &sh->state);
Linus Torvalds1da177e2005-04-16 15:20:36 -070089 list_add_tail(&sh->lru, &conf->handle_list);
NeilBrown72626682005-09-09 16:23:54 -070090 }
Linus Torvalds1da177e2005-04-16 15:20:36 -070091 md_wakeup_thread(conf->mddev->thread);
92 } else {
93 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
94 atomic_dec(&conf->preread_active_stripes);
95 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
96 md_wakeup_thread(conf->mddev->thread);
97 }
98 list_add_tail(&sh->lru, &conf->inactive_list);
99 atomic_dec(&conf->active_stripes);
100 if (!conf->inactive_blocked ||
101 atomic_read(&conf->active_stripes) < (NR_STRIPES*3/4))
102 wake_up(&conf->wait_for_stripe);
103 }
104 }
105}
106static void release_stripe(struct stripe_head *sh)
107{
108 raid5_conf_t *conf = sh->raid_conf;
109 unsigned long flags;
110
111 spin_lock_irqsave(&conf->device_lock, flags);
112 __release_stripe(conf, sh);
113 spin_unlock_irqrestore(&conf->device_lock, flags);
114}
115
116static void remove_hash(struct stripe_head *sh)
117{
118 PRINTK("remove_hash(), stripe %llu\n", (unsigned long long)sh->sector);
119
120 if (sh->hash_pprev) {
121 if (sh->hash_next)
122 sh->hash_next->hash_pprev = sh->hash_pprev;
123 *sh->hash_pprev = sh->hash_next;
124 sh->hash_pprev = NULL;
125 }
126}
127
128static __inline__ void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
129{
130 struct stripe_head **shp = &stripe_hash(conf, sh->sector);
131
132 PRINTK("insert_hash(), stripe %llu\n", (unsigned long long)sh->sector);
133
134 CHECK_DEVLOCK();
135 if ((sh->hash_next = *shp) != NULL)
136 (*shp)->hash_pprev = &sh->hash_next;
137 *shp = sh;
138 sh->hash_pprev = shp;
139}
140
141
142/* find an idle stripe, make sure it is unhashed, and return it. */
143static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
144{
145 struct stripe_head *sh = NULL;
146 struct list_head *first;
147
148 CHECK_DEVLOCK();
149 if (list_empty(&conf->inactive_list))
150 goto out;
151 first = conf->inactive_list.next;
152 sh = list_entry(first, struct stripe_head, lru);
153 list_del_init(first);
154 remove_hash(sh);
155 atomic_inc(&conf->active_stripes);
156out:
157 return sh;
158}
159
160static void shrink_buffers(struct stripe_head *sh, int num)
161{
162 struct page *p;
163 int i;
164
165 for (i=0; i<num ; i++) {
166 p = sh->dev[i].page;
167 if (!p)
168 continue;
169 sh->dev[i].page = NULL;
170 page_cache_release(p);
171 }
172}
173
174static int grow_buffers(struct stripe_head *sh, int num)
175{
176 int i;
177
178 for (i=0; i<num; i++) {
179 struct page *page;
180
181 if (!(page = alloc_page(GFP_KERNEL))) {
182 return 1;
183 }
184 sh->dev[i].page = page;
185 }
186 return 0;
187}
188
189static void raid5_build_block (struct stripe_head *sh, int i);
190
191static inline void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx)
192{
193 raid5_conf_t *conf = sh->raid_conf;
194 int disks = conf->raid_disks, i;
195
196 if (atomic_read(&sh->count) != 0)
197 BUG();
198 if (test_bit(STRIPE_HANDLE, &sh->state))
199 BUG();
200
201 CHECK_DEVLOCK();
202 PRINTK("init_stripe called, stripe %llu\n",
203 (unsigned long long)sh->sector);
204
205 remove_hash(sh);
206
207 sh->sector = sector;
208 sh->pd_idx = pd_idx;
209 sh->state = 0;
210
211 for (i=disks; i--; ) {
212 struct r5dev *dev = &sh->dev[i];
213
214 if (dev->toread || dev->towrite || dev->written ||
215 test_bit(R5_LOCKED, &dev->flags)) {
216 printk("sector=%llx i=%d %p %p %p %d\n",
217 (unsigned long long)sh->sector, i, dev->toread,
218 dev->towrite, dev->written,
219 test_bit(R5_LOCKED, &dev->flags));
220 BUG();
221 }
222 dev->flags = 0;
223 raid5_build_block(sh, i);
224 }
225 insert_hash(conf, sh);
226}
227
228static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector)
229{
230 struct stripe_head *sh;
231
232 CHECK_DEVLOCK();
233 PRINTK("__find_stripe, sector %llu\n", (unsigned long long)sector);
234 for (sh = stripe_hash(conf, sector); sh; sh = sh->hash_next)
235 if (sh->sector == sector)
236 return sh;
237 PRINTK("__stripe %llu not in cache\n", (unsigned long long)sector);
238 return NULL;
239}
240
241static void unplug_slaves(mddev_t *mddev);
242static void raid5_unplug_device(request_queue_t *q);
243
244static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector,
245 int pd_idx, int noblock)
246{
247 struct stripe_head *sh;
248
249 PRINTK("get_stripe, sector %llu\n", (unsigned long long)sector);
250
251 spin_lock_irq(&conf->device_lock);
252
253 do {
NeilBrown72626682005-09-09 16:23:54 -0700254 wait_event_lock_irq(conf->wait_for_stripe,
255 conf->quiesce == 0,
256 conf->device_lock, /* nothing */);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257 sh = __find_stripe(conf, sector);
258 if (!sh) {
259 if (!conf->inactive_blocked)
260 sh = get_free_stripe(conf);
261 if (noblock && sh == NULL)
262 break;
263 if (!sh) {
264 conf->inactive_blocked = 1;
265 wait_event_lock_irq(conf->wait_for_stripe,
266 !list_empty(&conf->inactive_list) &&
267 (atomic_read(&conf->active_stripes) < (NR_STRIPES *3/4)
268 || !conf->inactive_blocked),
269 conf->device_lock,
270 unplug_slaves(conf->mddev);
271 );
272 conf->inactive_blocked = 0;
273 } else
274 init_stripe(sh, sector, pd_idx);
275 } else {
276 if (atomic_read(&sh->count)) {
277 if (!list_empty(&sh->lru))
278 BUG();
279 } else {
280 if (!test_bit(STRIPE_HANDLE, &sh->state))
281 atomic_inc(&conf->active_stripes);
282 if (list_empty(&sh->lru))
283 BUG();
284 list_del_init(&sh->lru);
285 }
286 }
287 } while (sh == NULL);
288
289 if (sh)
290 atomic_inc(&sh->count);
291
292 spin_unlock_irq(&conf->device_lock);
293 return sh;
294}
295
NeilBrown3f294f42005-11-08 21:39:25 -0800296static int grow_one_stripe(raid5_conf_t *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297{
298 struct stripe_head *sh;
NeilBrown3f294f42005-11-08 21:39:25 -0800299 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
300 if (!sh)
301 return 0;
302 memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
303 sh->raid_conf = conf;
304 spin_lock_init(&sh->lock);
305
306 if (grow_buffers(sh, conf->raid_disks)) {
307 shrink_buffers(sh, conf->raid_disks);
308 kmem_cache_free(conf->slab_cache, sh);
309 return 0;
310 }
311 /* we just created an active stripe so... */
312 atomic_set(&sh->count, 1);
313 atomic_inc(&conf->active_stripes);
314 INIT_LIST_HEAD(&sh->lru);
315 release_stripe(sh);
316 return 1;
317}
318
319static int grow_stripes(raid5_conf_t *conf, int num)
320{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321 kmem_cache_t *sc;
322 int devs = conf->raid_disks;
323
324 sprintf(conf->cache_name, "raid5/%s", mdname(conf->mddev));
325
326 sc = kmem_cache_create(conf->cache_name,
327 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
328 0, 0, NULL, NULL);
329 if (!sc)
330 return 1;
331 conf->slab_cache = sc;
332 while (num--) {
NeilBrown3f294f42005-11-08 21:39:25 -0800333 if (!grow_one_stripe(conf))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335 }
336 return 0;
337}
338
NeilBrown3f294f42005-11-08 21:39:25 -0800339static int drop_one_stripe(raid5_conf_t *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340{
341 struct stripe_head *sh;
342
NeilBrown3f294f42005-11-08 21:39:25 -0800343 spin_lock_irq(&conf->device_lock);
344 sh = get_free_stripe(conf);
345 spin_unlock_irq(&conf->device_lock);
346 if (!sh)
347 return 0;
348 if (atomic_read(&sh->count))
349 BUG();
350 shrink_buffers(sh, conf->raid_disks);
351 kmem_cache_free(conf->slab_cache, sh);
352 atomic_dec(&conf->active_stripes);
353 return 1;
354}
355
356static void shrink_stripes(raid5_conf_t *conf)
357{
358 while (drop_one_stripe(conf))
359 ;
360
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361 kmem_cache_destroy(conf->slab_cache);
362 conf->slab_cache = NULL;
363}
364
NeilBrown4e5314b2005-11-08 21:39:22 -0800365static int raid5_end_read_request(struct bio * bi, unsigned int bytes_done,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366 int error)
367{
368 struct stripe_head *sh = bi->bi_private;
369 raid5_conf_t *conf = sh->raid_conf;
370 int disks = conf->raid_disks, i;
371 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
372
373 if (bi->bi_size)
374 return 1;
375
376 for (i=0 ; i<disks; i++)
377 if (bi == &sh->dev[i].req)
378 break;
379
380 PRINTK("end_read_request %llu/%d, count: %d, uptodate %d.\n",
381 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
382 uptodate);
383 if (i == disks) {
384 BUG();
385 return 0;
386 }
387
388 if (uptodate) {
389#if 0
390 struct bio *bio;
391 unsigned long flags;
392 spin_lock_irqsave(&conf->device_lock, flags);
393 /* we can return a buffer if we bypassed the cache or
394 * if the top buffer is not in highmem. If there are
395 * multiple buffers, leave the extra work to
396 * handle_stripe
397 */
398 buffer = sh->bh_read[i];
399 if (buffer &&
400 (!PageHighMem(buffer->b_page)
401 || buffer->b_page == bh->b_page )
402 ) {
403 sh->bh_read[i] = buffer->b_reqnext;
404 buffer->b_reqnext = NULL;
405 } else
406 buffer = NULL;
407 spin_unlock_irqrestore(&conf->device_lock, flags);
408 if (sh->bh_page[i]==bh->b_page)
409 set_buffer_uptodate(bh);
410 if (buffer) {
411 if (buffer->b_page != bh->b_page)
412 memcpy(buffer->b_data, bh->b_data, bh->b_size);
413 buffer->b_end_io(buffer, 1);
414 }
415#else
416 set_bit(R5_UPTODATE, &sh->dev[i].flags);
NeilBrown4e5314b2005-11-08 21:39:22 -0800417#endif
418 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
419 printk("R5: read error corrected!!\n");
420 clear_bit(R5_ReadError, &sh->dev[i].flags);
421 clear_bit(R5_ReWrite, &sh->dev[i].flags);
422 }
NeilBrownba22dcb2005-11-08 21:39:31 -0800423 if (atomic_read(&conf->disks[i].rdev->read_errors))
424 atomic_set(&conf->disks[i].rdev->read_errors, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425 } else {
NeilBrownba22dcb2005-11-08 21:39:31 -0800426 int retry = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
NeilBrownba22dcb2005-11-08 21:39:31 -0800428 atomic_inc(&conf->disks[i].rdev->read_errors);
429 if (conf->mddev->degraded)
NeilBrown4e5314b2005-11-08 21:39:22 -0800430 printk("R5: read error not correctable.\n");
NeilBrownba22dcb2005-11-08 21:39:31 -0800431 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
NeilBrown4e5314b2005-11-08 21:39:22 -0800432 /* Oh, no!!! */
433 printk("R5: read error NOT corrected!!\n");
NeilBrownba22dcb2005-11-08 21:39:31 -0800434 else if (atomic_read(&conf->disks[i].rdev->read_errors)
435 > conf->max_nr_stripes)
436 printk("raid5: Too many read errors, failing device.\n");
437 else
438 retry = 1;
439 if (retry)
440 set_bit(R5_ReadError, &sh->dev[i].flags);
441 else {
NeilBrown4e5314b2005-11-08 21:39:22 -0800442 clear_bit(R5_ReadError, &sh->dev[i].flags);
443 clear_bit(R5_ReWrite, &sh->dev[i].flags);
444 md_error(conf->mddev, conf->disks[i].rdev);
NeilBrownba22dcb2005-11-08 21:39:31 -0800445 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 }
447 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
448#if 0
449 /* must restore b_page before unlocking buffer... */
450 if (sh->bh_page[i] != bh->b_page) {
451 bh->b_page = sh->bh_page[i];
452 bh->b_data = page_address(bh->b_page);
453 clear_buffer_uptodate(bh);
454 }
455#endif
456 clear_bit(R5_LOCKED, &sh->dev[i].flags);
457 set_bit(STRIPE_HANDLE, &sh->state);
458 release_stripe(sh);
459 return 0;
460}
461
462static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
463 int error)
464{
465 struct stripe_head *sh = bi->bi_private;
466 raid5_conf_t *conf = sh->raid_conf;
467 int disks = conf->raid_disks, i;
468 unsigned long flags;
469 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
470
471 if (bi->bi_size)
472 return 1;
473
474 for (i=0 ; i<disks; i++)
475 if (bi == &sh->dev[i].req)
476 break;
477
478 PRINTK("end_write_request %llu/%d, count %d, uptodate: %d.\n",
479 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
480 uptodate);
481 if (i == disks) {
482 BUG();
483 return 0;
484 }
485
486 spin_lock_irqsave(&conf->device_lock, flags);
487 if (!uptodate)
488 md_error(conf->mddev, conf->disks[i].rdev);
489
490 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
491
492 clear_bit(R5_LOCKED, &sh->dev[i].flags);
493 set_bit(STRIPE_HANDLE, &sh->state);
494 __release_stripe(conf, sh);
495 spin_unlock_irqrestore(&conf->device_lock, flags);
496 return 0;
497}
498
499
500static sector_t compute_blocknr(struct stripe_head *sh, int i);
501
502static void raid5_build_block (struct stripe_head *sh, int i)
503{
504 struct r5dev *dev = &sh->dev[i];
505
506 bio_init(&dev->req);
507 dev->req.bi_io_vec = &dev->vec;
508 dev->req.bi_vcnt++;
509 dev->req.bi_max_vecs++;
510 dev->vec.bv_page = dev->page;
511 dev->vec.bv_len = STRIPE_SIZE;
512 dev->vec.bv_offset = 0;
513
514 dev->req.bi_sector = sh->sector;
515 dev->req.bi_private = sh;
516
517 dev->flags = 0;
518 if (i != sh->pd_idx)
519 dev->sector = compute_blocknr(sh, i);
520}
521
522static void error(mddev_t *mddev, mdk_rdev_t *rdev)
523{
524 char b[BDEVNAME_SIZE];
525 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
526 PRINTK("raid5: error called\n");
527
528 if (!rdev->faulty) {
529 mddev->sb_dirty = 1;
530 if (rdev->in_sync) {
531 conf->working_disks--;
532 mddev->degraded++;
533 conf->failed_disks++;
534 rdev->in_sync = 0;
535 /*
536 * if recovery was running, make sure it aborts.
537 */
538 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
539 }
540 rdev->faulty = 1;
541 printk (KERN_ALERT
542 "raid5: Disk failure on %s, disabling device."
543 " Operation continuing on %d devices\n",
544 bdevname(rdev->bdev,b), conf->working_disks);
545 }
546}
547
548/*
549 * Input: a 'big' sector number,
550 * Output: index of the data and parity disk, and the sector # in them.
551 */
552static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
553 unsigned int data_disks, unsigned int * dd_idx,
554 unsigned int * pd_idx, raid5_conf_t *conf)
555{
556 long stripe;
557 unsigned long chunk_number;
558 unsigned int chunk_offset;
559 sector_t new_sector;
560 int sectors_per_chunk = conf->chunk_size >> 9;
561
562 /* First compute the information on this sector */
563
564 /*
565 * Compute the chunk number and the sector offset inside the chunk
566 */
567 chunk_offset = sector_div(r_sector, sectors_per_chunk);
568 chunk_number = r_sector;
569 BUG_ON(r_sector != chunk_number);
570
571 /*
572 * Compute the stripe number
573 */
574 stripe = chunk_number / data_disks;
575
576 /*
577 * Compute the data disk and parity disk indexes inside the stripe
578 */
579 *dd_idx = chunk_number % data_disks;
580
581 /*
582 * Select the parity disk based on the user selected algorithm.
583 */
584 if (conf->level == 4)
585 *pd_idx = data_disks;
586 else switch (conf->algorithm) {
587 case ALGORITHM_LEFT_ASYMMETRIC:
588 *pd_idx = data_disks - stripe % raid_disks;
589 if (*dd_idx >= *pd_idx)
590 (*dd_idx)++;
591 break;
592 case ALGORITHM_RIGHT_ASYMMETRIC:
593 *pd_idx = stripe % raid_disks;
594 if (*dd_idx >= *pd_idx)
595 (*dd_idx)++;
596 break;
597 case ALGORITHM_LEFT_SYMMETRIC:
598 *pd_idx = data_disks - stripe % raid_disks;
599 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
600 break;
601 case ALGORITHM_RIGHT_SYMMETRIC:
602 *pd_idx = stripe % raid_disks;
603 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
604 break;
605 default:
606 printk("raid5: unsupported algorithm %d\n",
607 conf->algorithm);
608 }
609
610 /*
611 * Finally, compute the new sector number
612 */
613 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
614 return new_sector;
615}
616
617
618static sector_t compute_blocknr(struct stripe_head *sh, int i)
619{
620 raid5_conf_t *conf = sh->raid_conf;
621 int raid_disks = conf->raid_disks, data_disks = raid_disks - 1;
622 sector_t new_sector = sh->sector, check;
623 int sectors_per_chunk = conf->chunk_size >> 9;
624 sector_t stripe;
625 int chunk_offset;
626 int chunk_number, dummy1, dummy2, dd_idx = i;
627 sector_t r_sector;
628
629 chunk_offset = sector_div(new_sector, sectors_per_chunk);
630 stripe = new_sector;
631 BUG_ON(new_sector != stripe);
632
633
634 switch (conf->algorithm) {
635 case ALGORITHM_LEFT_ASYMMETRIC:
636 case ALGORITHM_RIGHT_ASYMMETRIC:
637 if (i > sh->pd_idx)
638 i--;
639 break;
640 case ALGORITHM_LEFT_SYMMETRIC:
641 case ALGORITHM_RIGHT_SYMMETRIC:
642 if (i < sh->pd_idx)
643 i += raid_disks;
644 i -= (sh->pd_idx + 1);
645 break;
646 default:
647 printk("raid5: unsupported algorithm %d\n",
648 conf->algorithm);
649 }
650
651 chunk_number = stripe * data_disks + i;
652 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
653
654 check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
655 if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
656 printk("compute_blocknr: map not correct\n");
657 return 0;
658 }
659 return r_sector;
660}
661
662
663
664/*
665 * Copy data between a page in the stripe cache, and a bio.
666 * There are no alignment or size guarantees between the page or the
667 * bio except that there is some overlap.
668 * All iovecs in the bio must be considered.
669 */
670static void copy_data(int frombio, struct bio *bio,
671 struct page *page,
672 sector_t sector)
673{
674 char *pa = page_address(page);
675 struct bio_vec *bvl;
676 int i;
677 int page_offset;
678
679 if (bio->bi_sector >= sector)
680 page_offset = (signed)(bio->bi_sector - sector) * 512;
681 else
682 page_offset = (signed)(sector - bio->bi_sector) * -512;
683 bio_for_each_segment(bvl, bio, i) {
684 int len = bio_iovec_idx(bio,i)->bv_len;
685 int clen;
686 int b_offset = 0;
687
688 if (page_offset < 0) {
689 b_offset = -page_offset;
690 page_offset += b_offset;
691 len -= b_offset;
692 }
693
694 if (len > 0 && page_offset + len > STRIPE_SIZE)
695 clen = STRIPE_SIZE - page_offset;
696 else clen = len;
697
698 if (clen > 0) {
699 char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
700 if (frombio)
701 memcpy(pa+page_offset, ba+b_offset, clen);
702 else
703 memcpy(ba+b_offset, pa+page_offset, clen);
704 __bio_kunmap_atomic(ba, KM_USER0);
705 }
706 if (clen < len) /* hit end of page */
707 break;
708 page_offset += len;
709 }
710}
711
712#define check_xor() do { \
713 if (count == MAX_XOR_BLOCKS) { \
714 xor_block(count, STRIPE_SIZE, ptr); \
715 count = 1; \
716 } \
717 } while(0)
718
719
720static void compute_block(struct stripe_head *sh, int dd_idx)
721{
722 raid5_conf_t *conf = sh->raid_conf;
723 int i, count, disks = conf->raid_disks;
724 void *ptr[MAX_XOR_BLOCKS], *p;
725
726 PRINTK("compute_block, stripe %llu, idx %d\n",
727 (unsigned long long)sh->sector, dd_idx);
728
729 ptr[0] = page_address(sh->dev[dd_idx].page);
730 memset(ptr[0], 0, STRIPE_SIZE);
731 count = 1;
732 for (i = disks ; i--; ) {
733 if (i == dd_idx)
734 continue;
735 p = page_address(sh->dev[i].page);
736 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
737 ptr[count++] = p;
738 else
739 printk("compute_block() %d, stripe %llu, %d"
740 " not present\n", dd_idx,
741 (unsigned long long)sh->sector, i);
742
743 check_xor();
744 }
745 if (count != 1)
746 xor_block(count, STRIPE_SIZE, ptr);
747 set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
748}
749
750static void compute_parity(struct stripe_head *sh, int method)
751{
752 raid5_conf_t *conf = sh->raid_conf;
753 int i, pd_idx = sh->pd_idx, disks = conf->raid_disks, count;
754 void *ptr[MAX_XOR_BLOCKS];
755 struct bio *chosen;
756
757 PRINTK("compute_parity, stripe %llu, method %d\n",
758 (unsigned long long)sh->sector, method);
759
760 count = 1;
761 ptr[0] = page_address(sh->dev[pd_idx].page);
762 switch(method) {
763 case READ_MODIFY_WRITE:
764 if (!test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags))
765 BUG();
766 for (i=disks ; i-- ;) {
767 if (i==pd_idx)
768 continue;
769 if (sh->dev[i].towrite &&
770 test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
771 ptr[count++] = page_address(sh->dev[i].page);
772 chosen = sh->dev[i].towrite;
773 sh->dev[i].towrite = NULL;
774
775 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
776 wake_up(&conf->wait_for_overlap);
777
778 if (sh->dev[i].written) BUG();
779 sh->dev[i].written = chosen;
780 check_xor();
781 }
782 }
783 break;
784 case RECONSTRUCT_WRITE:
785 memset(ptr[0], 0, STRIPE_SIZE);
786 for (i= disks; i-- ;)
787 if (i!=pd_idx && sh->dev[i].towrite) {
788 chosen = sh->dev[i].towrite;
789 sh->dev[i].towrite = NULL;
790
791 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
792 wake_up(&conf->wait_for_overlap);
793
794 if (sh->dev[i].written) BUG();
795 sh->dev[i].written = chosen;
796 }
797 break;
798 case CHECK_PARITY:
799 break;
800 }
801 if (count>1) {
802 xor_block(count, STRIPE_SIZE, ptr);
803 count = 1;
804 }
805
806 for (i = disks; i--;)
807 if (sh->dev[i].written) {
808 sector_t sector = sh->dev[i].sector;
809 struct bio *wbi = sh->dev[i].written;
810 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
811 copy_data(1, wbi, sh->dev[i].page, sector);
812 wbi = r5_next_bio(wbi, sector);
813 }
814
815 set_bit(R5_LOCKED, &sh->dev[i].flags);
816 set_bit(R5_UPTODATE, &sh->dev[i].flags);
817 }
818
819 switch(method) {
820 case RECONSTRUCT_WRITE:
821 case CHECK_PARITY:
822 for (i=disks; i--;)
823 if (i != pd_idx) {
824 ptr[count++] = page_address(sh->dev[i].page);
825 check_xor();
826 }
827 break;
828 case READ_MODIFY_WRITE:
829 for (i = disks; i--;)
830 if (sh->dev[i].written) {
831 ptr[count++] = page_address(sh->dev[i].page);
832 check_xor();
833 }
834 }
835 if (count != 1)
836 xor_block(count, STRIPE_SIZE, ptr);
837
838 if (method != CHECK_PARITY) {
839 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
840 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
841 } else
842 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
843}
844
845/*
846 * Each stripe/dev can have one or more bion attached.
847 * toread/towrite point to the first in a chain.
848 * The bi_next chain must be in order.
849 */
850static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
851{
852 struct bio **bip;
853 raid5_conf_t *conf = sh->raid_conf;
NeilBrown72626682005-09-09 16:23:54 -0700854 int firstwrite=0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855
856 PRINTK("adding bh b#%llu to stripe s#%llu\n",
857 (unsigned long long)bi->bi_sector,
858 (unsigned long long)sh->sector);
859
860
861 spin_lock(&sh->lock);
862 spin_lock_irq(&conf->device_lock);
NeilBrown72626682005-09-09 16:23:54 -0700863 if (forwrite) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864 bip = &sh->dev[dd_idx].towrite;
NeilBrown72626682005-09-09 16:23:54 -0700865 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
866 firstwrite = 1;
867 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -0700868 bip = &sh->dev[dd_idx].toread;
869 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
870 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
871 goto overlap;
872 bip = & (*bip)->bi_next;
873 }
874 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
875 goto overlap;
876
877 if (*bip && bi->bi_next && (*bip) != bi->bi_next)
878 BUG();
879 if (*bip)
880 bi->bi_next = *bip;
881 *bip = bi;
882 bi->bi_phys_segments ++;
883 spin_unlock_irq(&conf->device_lock);
884 spin_unlock(&sh->lock);
885
886 PRINTK("added bi b#%llu to stripe s#%llu, disk %d.\n",
887 (unsigned long long)bi->bi_sector,
888 (unsigned long long)sh->sector, dd_idx);
889
NeilBrown72626682005-09-09 16:23:54 -0700890 if (conf->mddev->bitmap && firstwrite) {
891 sh->bm_seq = conf->seq_write;
892 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
893 STRIPE_SECTORS, 0);
894 set_bit(STRIPE_BIT_DELAY, &sh->state);
895 }
896
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 if (forwrite) {
898 /* check if page is covered */
899 sector_t sector = sh->dev[dd_idx].sector;
900 for (bi=sh->dev[dd_idx].towrite;
901 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
902 bi && bi->bi_sector <= sector;
903 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
904 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
905 sector = bi->bi_sector + (bi->bi_size>>9);
906 }
907 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
908 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
909 }
910 return 1;
911
912 overlap:
913 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
914 spin_unlock_irq(&conf->device_lock);
915 spin_unlock(&sh->lock);
916 return 0;
917}
918
919
920/*
921 * handle_stripe - do things to a stripe.
922 *
923 * We lock the stripe and then examine the state of various bits
924 * to see what needs to be done.
925 * Possible results:
926 * return some read request which now have data
927 * return some write requests which are safely on disc
928 * schedule a read on some buffers
929 * schedule a write of some buffers
930 * return confirmation of parity correctness
931 *
932 * Parity calculations are done inside the stripe lock
933 * buffers are taken off read_list or write_list, and bh_cache buffers
934 * get BH_Lock set before the stripe lock is released.
935 *
936 */
937
938static void handle_stripe(struct stripe_head *sh)
939{
940 raid5_conf_t *conf = sh->raid_conf;
941 int disks = conf->raid_disks;
942 struct bio *return_bi= NULL;
943 struct bio *bi;
944 int i;
945 int syncing;
946 int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
947 int non_overwrite = 0;
948 int failed_num=0;
949 struct r5dev *dev;
950
951 PRINTK("handling stripe %llu, cnt=%d, pd_idx=%d\n",
952 (unsigned long long)sh->sector, atomic_read(&sh->count),
953 sh->pd_idx);
954
955 spin_lock(&sh->lock);
956 clear_bit(STRIPE_HANDLE, &sh->state);
957 clear_bit(STRIPE_DELAYED, &sh->state);
958
959 syncing = test_bit(STRIPE_SYNCING, &sh->state);
960 /* Now to look around and see what can be done */
961
962 for (i=disks; i--; ) {
963 mdk_rdev_t *rdev;
964 dev = &sh->dev[i];
965 clear_bit(R5_Insync, &dev->flags);
966 clear_bit(R5_Syncio, &dev->flags);
967
968 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
969 i, dev->flags, dev->toread, dev->towrite, dev->written);
970 /* maybe we can reply to a read */
971 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
972 struct bio *rbi, *rbi2;
973 PRINTK("Return read for disc %d\n", i);
974 spin_lock_irq(&conf->device_lock);
975 rbi = dev->toread;
976 dev->toread = NULL;
977 if (test_and_clear_bit(R5_Overlap, &dev->flags))
978 wake_up(&conf->wait_for_overlap);
979 spin_unlock_irq(&conf->device_lock);
980 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
981 copy_data(0, rbi, dev->page, dev->sector);
982 rbi2 = r5_next_bio(rbi, dev->sector);
983 spin_lock_irq(&conf->device_lock);
984 if (--rbi->bi_phys_segments == 0) {
985 rbi->bi_next = return_bi;
986 return_bi = rbi;
987 }
988 spin_unlock_irq(&conf->device_lock);
989 rbi = rbi2;
990 }
991 }
992
993 /* now count some things */
994 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
995 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
996
997
998 if (dev->toread) to_read++;
999 if (dev->towrite) {
1000 to_write++;
1001 if (!test_bit(R5_OVERWRITE, &dev->flags))
1002 non_overwrite++;
1003 }
1004 if (dev->written) written++;
1005 rdev = conf->disks[i].rdev; /* FIXME, should I be looking rdev */
1006 if (!rdev || !rdev->in_sync) {
NeilBrown4e5314b2005-11-08 21:39:22 -08001007 /* The ReadError flag wil just be confusing now */
1008 clear_bit(R5_ReadError, &dev->flags);
1009 clear_bit(R5_ReWrite, &dev->flags);
1010 }
1011 if (!rdev || !rdev->in_sync
1012 || test_bit(R5_ReadError, &dev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001013 failed++;
1014 failed_num = i;
1015 } else
1016 set_bit(R5_Insync, &dev->flags);
1017 }
1018 PRINTK("locked=%d uptodate=%d to_read=%d"
1019 " to_write=%d failed=%d failed_num=%d\n",
1020 locked, uptodate, to_read, to_write, failed, failed_num);
1021 /* check if the array has lost two devices and, if so, some requests might
1022 * need to be failed
1023 */
1024 if (failed > 1 && to_read+to_write+written) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001025 for (i=disks; i--; ) {
NeilBrown72626682005-09-09 16:23:54 -07001026 int bitmap_end = 0;
NeilBrown4e5314b2005-11-08 21:39:22 -08001027
1028 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1029 mdk_rdev_t *rdev = conf->disks[i].rdev;
1030 if (rdev && rdev->in_sync)
1031 /* multiple read failures in one stripe */
1032 md_error(conf->mddev, rdev);
1033 }
1034
NeilBrown72626682005-09-09 16:23:54 -07001035 spin_lock_irq(&conf->device_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036 /* fail all writes first */
1037 bi = sh->dev[i].towrite;
1038 sh->dev[i].towrite = NULL;
NeilBrown72626682005-09-09 16:23:54 -07001039 if (bi) { to_write--; bitmap_end = 1; }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040
1041 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1042 wake_up(&conf->wait_for_overlap);
1043
1044 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1045 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1046 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1047 if (--bi->bi_phys_segments == 0) {
1048 md_write_end(conf->mddev);
1049 bi->bi_next = return_bi;
1050 return_bi = bi;
1051 }
1052 bi = nextbi;
1053 }
1054 /* and fail all 'written' */
1055 bi = sh->dev[i].written;
1056 sh->dev[i].written = NULL;
NeilBrown72626682005-09-09 16:23:54 -07001057 if (bi) bitmap_end = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001058 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
1059 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1060 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1061 if (--bi->bi_phys_segments == 0) {
1062 md_write_end(conf->mddev);
1063 bi->bi_next = return_bi;
1064 return_bi = bi;
1065 }
1066 bi = bi2;
1067 }
1068
1069 /* fail any reads if this device is non-operational */
NeilBrown4e5314b2005-11-08 21:39:22 -08001070 if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1071 test_bit(R5_ReadError, &sh->dev[i].flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072 bi = sh->dev[i].toread;
1073 sh->dev[i].toread = NULL;
1074 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1075 wake_up(&conf->wait_for_overlap);
1076 if (bi) to_read--;
1077 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1078 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1079 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1080 if (--bi->bi_phys_segments == 0) {
1081 bi->bi_next = return_bi;
1082 return_bi = bi;
1083 }
1084 bi = nextbi;
1085 }
1086 }
NeilBrown72626682005-09-09 16:23:54 -07001087 spin_unlock_irq(&conf->device_lock);
1088 if (bitmap_end)
1089 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1090 STRIPE_SECTORS, 0, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092 }
1093 if (failed > 1 && syncing) {
1094 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
1095 clear_bit(STRIPE_SYNCING, &sh->state);
1096 syncing = 0;
1097 }
1098
1099 /* might be able to return some write requests if the parity block
1100 * is safe, or on a failed drive
1101 */
1102 dev = &sh->dev[sh->pd_idx];
1103 if ( written &&
1104 ( (test_bit(R5_Insync, &dev->flags) && !test_bit(R5_LOCKED, &dev->flags) &&
1105 test_bit(R5_UPTODATE, &dev->flags))
1106 || (failed == 1 && failed_num == sh->pd_idx))
1107 ) {
1108 /* any written block on an uptodate or failed drive can be returned.
1109 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
1110 * never LOCKED, so we don't need to test 'failed' directly.
1111 */
1112 for (i=disks; i--; )
1113 if (sh->dev[i].written) {
1114 dev = &sh->dev[i];
1115 if (!test_bit(R5_LOCKED, &dev->flags) &&
1116 test_bit(R5_UPTODATE, &dev->flags) ) {
1117 /* We can return any write requests */
1118 struct bio *wbi, *wbi2;
NeilBrown72626682005-09-09 16:23:54 -07001119 int bitmap_end = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120 PRINTK("Return write for disc %d\n", i);
1121 spin_lock_irq(&conf->device_lock);
1122 wbi = dev->written;
1123 dev->written = NULL;
1124 while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1125 wbi2 = r5_next_bio(wbi, dev->sector);
1126 if (--wbi->bi_phys_segments == 0) {
1127 md_write_end(conf->mddev);
1128 wbi->bi_next = return_bi;
1129 return_bi = wbi;
1130 }
1131 wbi = wbi2;
1132 }
NeilBrown72626682005-09-09 16:23:54 -07001133 if (dev->towrite == NULL)
1134 bitmap_end = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135 spin_unlock_irq(&conf->device_lock);
NeilBrown72626682005-09-09 16:23:54 -07001136 if (bitmap_end)
1137 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1138 STRIPE_SECTORS,
1139 !test_bit(STRIPE_DEGRADED, &sh->state), 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140 }
1141 }
1142 }
1143
1144 /* Now we might consider reading some blocks, either to check/generate
1145 * parity, or to satisfy requests
1146 * or to load a block that is being partially written.
1147 */
1148 if (to_read || non_overwrite || (syncing && (uptodate < disks))) {
1149 for (i=disks; i--;) {
1150 dev = &sh->dev[i];
1151 if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1152 (dev->toread ||
1153 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1154 syncing ||
1155 (failed && (sh->dev[failed_num].toread ||
1156 (sh->dev[failed_num].towrite && !test_bit(R5_OVERWRITE, &sh->dev[failed_num].flags))))
1157 )
1158 ) {
1159 /* we would like to get this block, possibly
1160 * by computing it, but we might not be able to
1161 */
1162 if (uptodate == disks-1) {
1163 PRINTK("Computing block %d\n", i);
1164 compute_block(sh, i);
1165 uptodate++;
1166 } else if (test_bit(R5_Insync, &dev->flags)) {
1167 set_bit(R5_LOCKED, &dev->flags);
1168 set_bit(R5_Wantread, &dev->flags);
1169#if 0
1170 /* if I am just reading this block and we don't have
1171 a failed drive, or any pending writes then sidestep the cache */
1172 if (sh->bh_read[i] && !sh->bh_read[i]->b_reqnext &&
1173 ! syncing && !failed && !to_write) {
1174 sh->bh_cache[i]->b_page = sh->bh_read[i]->b_page;
1175 sh->bh_cache[i]->b_data = sh->bh_read[i]->b_data;
1176 }
1177#endif
1178 locked++;
1179 PRINTK("Reading block %d (sync=%d)\n",
1180 i, syncing);
1181 if (syncing)
1182 md_sync_acct(conf->disks[i].rdev->bdev,
1183 STRIPE_SECTORS);
1184 }
1185 }
1186 }
1187 set_bit(STRIPE_HANDLE, &sh->state);
1188 }
1189
1190 /* now to consider writing and what else, if anything should be read */
1191 if (to_write) {
1192 int rmw=0, rcw=0;
1193 for (i=disks ; i--;) {
1194 /* would I have to read this buffer for read_modify_write */
1195 dev = &sh->dev[i];
1196 if ((dev->towrite || i == sh->pd_idx) &&
1197 (!test_bit(R5_LOCKED, &dev->flags)
1198#if 0
1199|| sh->bh_page[i]!=bh->b_page
1200#endif
1201 ) &&
1202 !test_bit(R5_UPTODATE, &dev->flags)) {
1203 if (test_bit(R5_Insync, &dev->flags)
1204/* && !(!mddev->insync && i == sh->pd_idx) */
1205 )
1206 rmw++;
1207 else rmw += 2*disks; /* cannot read it */
1208 }
1209 /* Would I have to read this buffer for reconstruct_write */
1210 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1211 (!test_bit(R5_LOCKED, &dev->flags)
1212#if 0
1213|| sh->bh_page[i] != bh->b_page
1214#endif
1215 ) &&
1216 !test_bit(R5_UPTODATE, &dev->flags)) {
1217 if (test_bit(R5_Insync, &dev->flags)) rcw++;
1218 else rcw += 2*disks;
1219 }
1220 }
1221 PRINTK("for sector %llu, rmw=%d rcw=%d\n",
1222 (unsigned long long)sh->sector, rmw, rcw);
1223 set_bit(STRIPE_HANDLE, &sh->state);
1224 if (rmw < rcw && rmw > 0)
1225 /* prefer read-modify-write, but need to get some data */
1226 for (i=disks; i--;) {
1227 dev = &sh->dev[i];
1228 if ((dev->towrite || i == sh->pd_idx) &&
1229 !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1230 test_bit(R5_Insync, &dev->flags)) {
1231 if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1232 {
1233 PRINTK("Read_old block %d for r-m-w\n", i);
1234 set_bit(R5_LOCKED, &dev->flags);
1235 set_bit(R5_Wantread, &dev->flags);
1236 locked++;
1237 } else {
1238 set_bit(STRIPE_DELAYED, &sh->state);
1239 set_bit(STRIPE_HANDLE, &sh->state);
1240 }
1241 }
1242 }
1243 if (rcw <= rmw && rcw > 0)
1244 /* want reconstruct write, but need to get some data */
1245 for (i=disks; i--;) {
1246 dev = &sh->dev[i];
1247 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1248 !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1249 test_bit(R5_Insync, &dev->flags)) {
1250 if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1251 {
1252 PRINTK("Read_old block %d for Reconstruct\n", i);
1253 set_bit(R5_LOCKED, &dev->flags);
1254 set_bit(R5_Wantread, &dev->flags);
1255 locked++;
1256 } else {
1257 set_bit(STRIPE_DELAYED, &sh->state);
1258 set_bit(STRIPE_HANDLE, &sh->state);
1259 }
1260 }
1261 }
1262 /* now if nothing is locked, and if we have enough data, we can start a write request */
NeilBrown72626682005-09-09 16:23:54 -07001263 if (locked == 0 && (rcw == 0 ||rmw == 0) &&
1264 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265 PRINTK("Computing parity...\n");
1266 compute_parity(sh, rcw==0 ? RECONSTRUCT_WRITE : READ_MODIFY_WRITE);
1267 /* now every locked buffer is ready to be written */
1268 for (i=disks; i--;)
1269 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
1270 PRINTK("Writing block %d\n", i);
1271 locked++;
1272 set_bit(R5_Wantwrite, &sh->dev[i].flags);
1273 if (!test_bit(R5_Insync, &sh->dev[i].flags)
1274 || (i==sh->pd_idx && failed == 0))
1275 set_bit(STRIPE_INSYNC, &sh->state);
1276 }
1277 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
1278 atomic_dec(&conf->preread_active_stripes);
1279 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
1280 md_wakeup_thread(conf->mddev->thread);
1281 }
1282 }
1283 }
1284
1285 /* maybe we need to check and possibly fix the parity for this stripe
1286 * Any reads will already have been scheduled, so we just see if enough data
1287 * is available
1288 */
1289 if (syncing && locked == 0 &&
1290 !test_bit(STRIPE_INSYNC, &sh->state) && failed <= 1) {
1291 set_bit(STRIPE_HANDLE, &sh->state);
1292 if (failed == 0) {
1293 char *pagea;
1294 if (uptodate != disks)
1295 BUG();
1296 compute_parity(sh, CHECK_PARITY);
1297 uptodate--;
1298 pagea = page_address(sh->dev[sh->pd_idx].page);
1299 if ((*(u32*)pagea) == 0 &&
1300 !memcmp(pagea, pagea+4, STRIPE_SIZE-4)) {
1301 /* parity is correct (on disc, not in buffer any more) */
1302 set_bit(STRIPE_INSYNC, &sh->state);
NeilBrown9d888832005-11-08 21:39:26 -08001303 } else {
1304 conf->mddev->resync_mismatches += STRIPE_SECTORS;
1305 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
1306 /* don't try to repair!! */
1307 set_bit(STRIPE_INSYNC, &sh->state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308 }
1309 }
1310 if (!test_bit(STRIPE_INSYNC, &sh->state)) {
1311 if (failed==0)
1312 failed_num = sh->pd_idx;
1313 /* should be able to compute the missing block and write it to spare */
1314 if (!test_bit(R5_UPTODATE, &sh->dev[failed_num].flags)) {
1315 if (uptodate+1 != disks)
1316 BUG();
1317 compute_block(sh, failed_num);
1318 uptodate++;
1319 }
1320 if (uptodate != disks)
1321 BUG();
1322 dev = &sh->dev[failed_num];
1323 set_bit(R5_LOCKED, &dev->flags);
1324 set_bit(R5_Wantwrite, &dev->flags);
NeilBrown72626682005-09-09 16:23:54 -07001325 clear_bit(STRIPE_DEGRADED, &sh->state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001326 locked++;
1327 set_bit(STRIPE_INSYNC, &sh->state);
1328 set_bit(R5_Syncio, &dev->flags);
1329 }
1330 }
1331 if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
1332 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
1333 clear_bit(STRIPE_SYNCING, &sh->state);
1334 }
NeilBrown4e5314b2005-11-08 21:39:22 -08001335
1336 /* If the failed drive is just a ReadError, then we might need to progress
1337 * the repair/check process
1338 */
NeilBrownba22dcb2005-11-08 21:39:31 -08001339 if (failed == 1 && ! conf->mddev->ro &&
1340 test_bit(R5_ReadError, &sh->dev[failed_num].flags)
NeilBrown4e5314b2005-11-08 21:39:22 -08001341 && !test_bit(R5_LOCKED, &sh->dev[failed_num].flags)
1342 && test_bit(R5_UPTODATE, &sh->dev[failed_num].flags)
1343 ) {
1344 dev = &sh->dev[failed_num];
1345 if (!test_bit(R5_ReWrite, &dev->flags)) {
1346 set_bit(R5_Wantwrite, &dev->flags);
1347 set_bit(R5_ReWrite, &dev->flags);
1348 set_bit(R5_LOCKED, &dev->flags);
1349 } else {
1350 /* let's read it back */
1351 set_bit(R5_Wantread, &dev->flags);
1352 set_bit(R5_LOCKED, &dev->flags);
1353 }
1354 }
1355
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356 spin_unlock(&sh->lock);
1357
1358 while ((bi=return_bi)) {
1359 int bytes = bi->bi_size;
1360
1361 return_bi = bi->bi_next;
1362 bi->bi_next = NULL;
1363 bi->bi_size = 0;
1364 bi->bi_end_io(bi, bytes, 0);
1365 }
1366 for (i=disks; i-- ;) {
1367 int rw;
1368 struct bio *bi;
1369 mdk_rdev_t *rdev;
1370 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
1371 rw = 1;
1372 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1373 rw = 0;
1374 else
1375 continue;
1376
1377 bi = &sh->dev[i].req;
1378
1379 bi->bi_rw = rw;
1380 if (rw)
1381 bi->bi_end_io = raid5_end_write_request;
1382 else
1383 bi->bi_end_io = raid5_end_read_request;
1384
1385 rcu_read_lock();
Suzanne Woodd6065f72005-11-08 21:39:27 -08001386 rdev = rcu_dereference(conf->disks[i].rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387 if (rdev && rdev->faulty)
1388 rdev = NULL;
1389 if (rdev)
1390 atomic_inc(&rdev->nr_pending);
1391 rcu_read_unlock();
1392
1393 if (rdev) {
1394 if (test_bit(R5_Syncio, &sh->dev[i].flags))
1395 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1396
1397 bi->bi_bdev = rdev->bdev;
1398 PRINTK("for %llu schedule op %ld on disc %d\n",
1399 (unsigned long long)sh->sector, bi->bi_rw, i);
1400 atomic_inc(&sh->count);
1401 bi->bi_sector = sh->sector + rdev->data_offset;
1402 bi->bi_flags = 1 << BIO_UPTODATE;
1403 bi->bi_vcnt = 1;
1404 bi->bi_max_vecs = 1;
1405 bi->bi_idx = 0;
1406 bi->bi_io_vec = &sh->dev[i].vec;
1407 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1408 bi->bi_io_vec[0].bv_offset = 0;
1409 bi->bi_size = STRIPE_SIZE;
1410 bi->bi_next = NULL;
1411 generic_make_request(bi);
1412 } else {
NeilBrown72626682005-09-09 16:23:54 -07001413 if (rw == 1)
1414 set_bit(STRIPE_DEGRADED, &sh->state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415 PRINTK("skip op %ld on disc %d for sector %llu\n",
1416 bi->bi_rw, i, (unsigned long long)sh->sector);
1417 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1418 set_bit(STRIPE_HANDLE, &sh->state);
1419 }
1420 }
1421}
1422
1423static inline void raid5_activate_delayed(raid5_conf_t *conf)
1424{
1425 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
1426 while (!list_empty(&conf->delayed_list)) {
1427 struct list_head *l = conf->delayed_list.next;
1428 struct stripe_head *sh;
1429 sh = list_entry(l, struct stripe_head, lru);
1430 list_del_init(l);
1431 clear_bit(STRIPE_DELAYED, &sh->state);
1432 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1433 atomic_inc(&conf->preread_active_stripes);
1434 list_add_tail(&sh->lru, &conf->handle_list);
1435 }
1436 }
1437}
1438
NeilBrown72626682005-09-09 16:23:54 -07001439static inline void activate_bit_delay(raid5_conf_t *conf)
1440{
1441 /* device_lock is held */
1442 struct list_head head;
1443 list_add(&head, &conf->bitmap_list);
1444 list_del_init(&conf->bitmap_list);
1445 while (!list_empty(&head)) {
1446 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
1447 list_del_init(&sh->lru);
1448 atomic_inc(&sh->count);
1449 __release_stripe(conf, sh);
1450 }
1451}
1452
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453static void unplug_slaves(mddev_t *mddev)
1454{
1455 raid5_conf_t *conf = mddev_to_conf(mddev);
1456 int i;
1457
1458 rcu_read_lock();
1459 for (i=0; i<mddev->raid_disks; i++) {
Suzanne Woodd6065f72005-11-08 21:39:27 -08001460 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001461 if (rdev && !rdev->faulty && atomic_read(&rdev->nr_pending)) {
1462 request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
1463
1464 atomic_inc(&rdev->nr_pending);
1465 rcu_read_unlock();
1466
1467 if (r_queue->unplug_fn)
1468 r_queue->unplug_fn(r_queue);
1469
1470 rdev_dec_pending(rdev, mddev);
1471 rcu_read_lock();
1472 }
1473 }
1474 rcu_read_unlock();
1475}
1476
1477static void raid5_unplug_device(request_queue_t *q)
1478{
1479 mddev_t *mddev = q->queuedata;
1480 raid5_conf_t *conf = mddev_to_conf(mddev);
1481 unsigned long flags;
1482
1483 spin_lock_irqsave(&conf->device_lock, flags);
1484
NeilBrown72626682005-09-09 16:23:54 -07001485 if (blk_remove_plug(q)) {
1486 conf->seq_flush++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487 raid5_activate_delayed(conf);
NeilBrown72626682005-09-09 16:23:54 -07001488 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489 md_wakeup_thread(mddev->thread);
1490
1491 spin_unlock_irqrestore(&conf->device_lock, flags);
1492
1493 unplug_slaves(mddev);
1494}
1495
1496static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
1497 sector_t *error_sector)
1498{
1499 mddev_t *mddev = q->queuedata;
1500 raid5_conf_t *conf = mddev_to_conf(mddev);
1501 int i, ret = 0;
1502
1503 rcu_read_lock();
1504 for (i=0; i<mddev->raid_disks && ret == 0; i++) {
Suzanne Woodd6065f72005-11-08 21:39:27 -08001505 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506 if (rdev && !rdev->faulty) {
1507 struct block_device *bdev = rdev->bdev;
1508 request_queue_t *r_queue = bdev_get_queue(bdev);
1509
1510 if (!r_queue->issue_flush_fn)
1511 ret = -EOPNOTSUPP;
1512 else {
1513 atomic_inc(&rdev->nr_pending);
1514 rcu_read_unlock();
1515 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
1516 error_sector);
1517 rdev_dec_pending(rdev, mddev);
1518 rcu_read_lock();
1519 }
1520 }
1521 }
1522 rcu_read_unlock();
1523 return ret;
1524}
1525
1526static inline void raid5_plug_device(raid5_conf_t *conf)
1527{
1528 spin_lock_irq(&conf->device_lock);
1529 blk_plug_device(conf->mddev->queue);
1530 spin_unlock_irq(&conf->device_lock);
1531}
1532
1533static int make_request (request_queue_t *q, struct bio * bi)
1534{
1535 mddev_t *mddev = q->queuedata;
1536 raid5_conf_t *conf = mddev_to_conf(mddev);
1537 const unsigned int raid_disks = conf->raid_disks;
1538 const unsigned int data_disks = raid_disks - 1;
1539 unsigned int dd_idx, pd_idx;
1540 sector_t new_sector;
1541 sector_t logical_sector, last_sector;
1542 struct stripe_head *sh;
Jens Axboea3623572005-11-01 09:26:16 +01001543 const int rw = bio_data_dir(bi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544
NeilBrowne5dcdd82005-09-09 16:23:41 -07001545 if (unlikely(bio_barrier(bi))) {
1546 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
1547 return 0;
1548 }
1549
NeilBrown3d310eb2005-06-21 17:17:26 -07001550 md_write_start(mddev, bi);
NeilBrown06d91a52005-06-21 17:17:12 -07001551
Jens Axboea3623572005-11-01 09:26:16 +01001552 disk_stat_inc(mddev->gendisk, ios[rw]);
1553 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001554
1555 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
1556 last_sector = bi->bi_sector + (bi->bi_size>>9);
1557 bi->bi_next = NULL;
1558 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
NeilBrown06d91a52005-06-21 17:17:12 -07001559
Linus Torvalds1da177e2005-04-16 15:20:36 -07001560 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
1561 DEFINE_WAIT(w);
1562
1563 new_sector = raid5_compute_sector(logical_sector,
1564 raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1565
1566 PRINTK("raid5: make_request, sector %llu logical %llu\n",
1567 (unsigned long long)new_sector,
1568 (unsigned long long)logical_sector);
1569
1570 retry:
1571 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
1572 sh = get_active_stripe(conf, new_sector, pd_idx, (bi->bi_rw&RWA_MASK));
1573 if (sh) {
1574 if (!add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
1575 /* Add failed due to overlap. Flush everything
1576 * and wait a while
1577 */
1578 raid5_unplug_device(mddev->queue);
1579 release_stripe(sh);
1580 schedule();
1581 goto retry;
1582 }
1583 finish_wait(&conf->wait_for_overlap, &w);
1584 raid5_plug_device(conf);
1585 handle_stripe(sh);
1586 release_stripe(sh);
1587
1588 } else {
1589 /* cannot get stripe for read-ahead, just give-up */
1590 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1591 finish_wait(&conf->wait_for_overlap, &w);
1592 break;
1593 }
1594
1595 }
1596 spin_lock_irq(&conf->device_lock);
1597 if (--bi->bi_phys_segments == 0) {
1598 int bytes = bi->bi_size;
1599
1600 if ( bio_data_dir(bi) == WRITE )
1601 md_write_end(mddev);
1602 bi->bi_size = 0;
1603 bi->bi_end_io(bi, bytes, 0);
1604 }
1605 spin_unlock_irq(&conf->device_lock);
1606 return 0;
1607}
1608
1609/* FIXME go_faster isn't used */
NeilBrown57afd892005-06-21 17:17:13 -07001610static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611{
1612 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1613 struct stripe_head *sh;
1614 int sectors_per_chunk = conf->chunk_size >> 9;
1615 sector_t x;
1616 unsigned long stripe;
1617 int chunk_offset;
1618 int dd_idx, pd_idx;
1619 sector_t first_sector;
1620 int raid_disks = conf->raid_disks;
1621 int data_disks = raid_disks-1;
NeilBrown72626682005-09-09 16:23:54 -07001622 sector_t max_sector = mddev->size << 1;
1623 int sync_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624
NeilBrown72626682005-09-09 16:23:54 -07001625 if (sector_nr >= max_sector) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626 /* just being told to finish up .. nothing much to do */
1627 unplug_slaves(mddev);
NeilBrown72626682005-09-09 16:23:54 -07001628
1629 if (mddev->curr_resync < max_sector) /* aborted */
1630 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1631 &sync_blocks, 1);
1632 else /* compelted sync */
1633 conf->fullsync = 0;
1634 bitmap_close_sync(mddev->bitmap);
1635
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636 return 0;
1637 }
1638 /* if there is 1 or more failed drives and we are trying
1639 * to resync, then assert that we are finished, because there is
1640 * nothing we can do.
1641 */
1642 if (mddev->degraded >= 1 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
NeilBrown57afd892005-06-21 17:17:13 -07001643 sector_t rv = (mddev->size << 1) - sector_nr;
1644 *skipped = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645 return rv;
1646 }
NeilBrown72626682005-09-09 16:23:54 -07001647 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1648 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
1649 /* we can skip this block, and probably more */
1650 sync_blocks /= STRIPE_SECTORS;
1651 *skipped = 1;
1652 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
1653 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654
1655 x = sector_nr;
1656 chunk_offset = sector_div(x, sectors_per_chunk);
1657 stripe = x;
1658 BUG_ON(x != stripe);
1659
1660 first_sector = raid5_compute_sector((sector_t)stripe*data_disks*sectors_per_chunk
1661 + chunk_offset, raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1662 sh = get_active_stripe(conf, sector_nr, pd_idx, 1);
1663 if (sh == NULL) {
1664 sh = get_active_stripe(conf, sector_nr, pd_idx, 0);
1665 /* make sure we don't swamp the stripe cache if someone else
1666 * is trying to get access
1667 */
Nishanth Aravamudan66c006a2005-11-07 01:01:17 -08001668 schedule_timeout_uninterruptible(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001669 }
NeilBrown72626682005-09-09 16:23:54 -07001670 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671 spin_lock(&sh->lock);
1672 set_bit(STRIPE_SYNCING, &sh->state);
1673 clear_bit(STRIPE_INSYNC, &sh->state);
1674 spin_unlock(&sh->lock);
1675
1676 handle_stripe(sh);
1677 release_stripe(sh);
1678
1679 return STRIPE_SECTORS;
1680}
1681
1682/*
1683 * This is our raid5 kernel thread.
1684 *
1685 * We scan the hash table for stripes which can be handled now.
1686 * During the scan, completed stripes are saved for us by the interrupt
1687 * handler, so that they will not have to wait for our next wakeup.
1688 */
1689static void raid5d (mddev_t *mddev)
1690{
1691 struct stripe_head *sh;
1692 raid5_conf_t *conf = mddev_to_conf(mddev);
1693 int handled;
1694
1695 PRINTK("+++ raid5d active\n");
1696
1697 md_check_recovery(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001698
1699 handled = 0;
1700 spin_lock_irq(&conf->device_lock);
1701 while (1) {
1702 struct list_head *first;
1703
NeilBrown72626682005-09-09 16:23:54 -07001704 if (conf->seq_flush - conf->seq_write > 0) {
1705 int seq = conf->seq_flush;
1706 bitmap_unplug(mddev->bitmap);
1707 conf->seq_write = seq;
1708 activate_bit_delay(conf);
1709 }
1710
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711 if (list_empty(&conf->handle_list) &&
1712 atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
1713 !blk_queue_plugged(mddev->queue) &&
1714 !list_empty(&conf->delayed_list))
1715 raid5_activate_delayed(conf);
1716
1717 if (list_empty(&conf->handle_list))
1718 break;
1719
1720 first = conf->handle_list.next;
1721 sh = list_entry(first, struct stripe_head, lru);
1722
1723 list_del_init(first);
1724 atomic_inc(&sh->count);
1725 if (atomic_read(&sh->count)!= 1)
1726 BUG();
1727 spin_unlock_irq(&conf->device_lock);
1728
1729 handled++;
1730 handle_stripe(sh);
1731 release_stripe(sh);
1732
1733 spin_lock_irq(&conf->device_lock);
1734 }
1735 PRINTK("%d stripes handled\n", handled);
1736
1737 spin_unlock_irq(&conf->device_lock);
1738
1739 unplug_slaves(mddev);
1740
1741 PRINTK("--- raid5d inactive\n");
1742}
1743
NeilBrown3f294f42005-11-08 21:39:25 -08001744static ssize_t
NeilBrown007583c2005-11-08 21:39:30 -08001745raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
NeilBrown3f294f42005-11-08 21:39:25 -08001746{
NeilBrown007583c2005-11-08 21:39:30 -08001747 raid5_conf_t *conf = mddev_to_conf(mddev);
NeilBrown3f294f42005-11-08 21:39:25 -08001748 return sprintf(page, "%d\n", conf->max_nr_stripes);
1749}
1750
1751static ssize_t
NeilBrown007583c2005-11-08 21:39:30 -08001752raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
NeilBrown3f294f42005-11-08 21:39:25 -08001753{
NeilBrown007583c2005-11-08 21:39:30 -08001754 raid5_conf_t *conf = mddev_to_conf(mddev);
NeilBrown3f294f42005-11-08 21:39:25 -08001755 char *end;
1756 int new;
1757 if (len >= PAGE_SIZE)
1758 return -EINVAL;
1759
1760 new = simple_strtoul(page, &end, 10);
1761 if (!*page || (*end && *end != '\n') )
1762 return -EINVAL;
1763 if (new <= 16 || new > 32768)
1764 return -EINVAL;
1765 while (new < conf->max_nr_stripes) {
1766 if (drop_one_stripe(conf))
1767 conf->max_nr_stripes--;
1768 else
1769 break;
1770 }
1771 while (new > conf->max_nr_stripes) {
1772 if (grow_one_stripe(conf))
1773 conf->max_nr_stripes++;
1774 else break;
1775 }
1776 return len;
1777}
NeilBrown007583c2005-11-08 21:39:30 -08001778
1779static struct md_sysfs_entry raid5_stripecache_size = {
NeilBrown3f294f42005-11-08 21:39:25 -08001780 .attr = {.name = "stripe_cache_size", .mode = S_IRUGO | S_IWUSR },
1781 .show = raid5_show_stripe_cache_size,
1782 .store = raid5_store_stripe_cache_size,
1783};
1784
1785static ssize_t
NeilBrown007583c2005-11-08 21:39:30 -08001786raid5_show_stripe_cache_active(mddev_t *mddev, char *page)
NeilBrown3f294f42005-11-08 21:39:25 -08001787{
NeilBrown007583c2005-11-08 21:39:30 -08001788 raid5_conf_t *conf = mddev_to_conf(mddev);
NeilBrown3f294f42005-11-08 21:39:25 -08001789 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
1790}
1791
NeilBrown007583c2005-11-08 21:39:30 -08001792static struct md_sysfs_entry raid5_stripecache_active = {
NeilBrown3f294f42005-11-08 21:39:25 -08001793 .attr = {.name = "stripe_cache_active", .mode = S_IRUGO},
1794 .show = raid5_show_stripe_cache_active,
1795};
1796
NeilBrown007583c2005-11-08 21:39:30 -08001797static struct attribute *raid5_attrs[] = {
NeilBrown3f294f42005-11-08 21:39:25 -08001798 &raid5_stripecache_size.attr,
1799 &raid5_stripecache_active.attr,
1800 NULL,
1801};
NeilBrown007583c2005-11-08 21:39:30 -08001802static struct attribute_group raid5_attrs_group = {
1803 .name = NULL,
1804 .attrs = raid5_attrs,
NeilBrown3f294f42005-11-08 21:39:25 -08001805};
1806
NeilBrown72626682005-09-09 16:23:54 -07001807static int run(mddev_t *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001808{
1809 raid5_conf_t *conf;
1810 int raid_disk, memory;
1811 mdk_rdev_t *rdev;
1812 struct disk_info *disk;
1813 struct list_head *tmp;
1814
1815 if (mddev->level != 5 && mddev->level != 4) {
1816 printk("raid5: %s: raid level not set to 4/5 (%d)\n", mdname(mddev), mddev->level);
1817 return -EIO;
1818 }
1819
1820 mddev->private = kmalloc (sizeof (raid5_conf_t)
1821 + mddev->raid_disks * sizeof(struct disk_info),
1822 GFP_KERNEL);
1823 if ((conf = mddev->private) == NULL)
1824 goto abort;
1825 memset (conf, 0, sizeof (*conf) + mddev->raid_disks * sizeof(struct disk_info) );
1826 conf->mddev = mddev;
1827
1828 if ((conf->stripe_hashtbl = (struct stripe_head **) __get_free_pages(GFP_ATOMIC, HASH_PAGES_ORDER)) == NULL)
1829 goto abort;
1830 memset(conf->stripe_hashtbl, 0, HASH_PAGES * PAGE_SIZE);
1831
1832 spin_lock_init(&conf->device_lock);
1833 init_waitqueue_head(&conf->wait_for_stripe);
1834 init_waitqueue_head(&conf->wait_for_overlap);
1835 INIT_LIST_HEAD(&conf->handle_list);
1836 INIT_LIST_HEAD(&conf->delayed_list);
NeilBrown72626682005-09-09 16:23:54 -07001837 INIT_LIST_HEAD(&conf->bitmap_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838 INIT_LIST_HEAD(&conf->inactive_list);
1839 atomic_set(&conf->active_stripes, 0);
1840 atomic_set(&conf->preread_active_stripes, 0);
1841
Linus Torvalds1da177e2005-04-16 15:20:36 -07001842 PRINTK("raid5: run(%s) called.\n", mdname(mddev));
1843
1844 ITERATE_RDEV(mddev,rdev,tmp) {
1845 raid_disk = rdev->raid_disk;
1846 if (raid_disk >= mddev->raid_disks
1847 || raid_disk < 0)
1848 continue;
1849 disk = conf->disks + raid_disk;
1850
1851 disk->rdev = rdev;
1852
1853 if (rdev->in_sync) {
1854 char b[BDEVNAME_SIZE];
1855 printk(KERN_INFO "raid5: device %s operational as raid"
1856 " disk %d\n", bdevname(rdev->bdev,b),
1857 raid_disk);
1858 conf->working_disks++;
1859 }
1860 }
1861
1862 conf->raid_disks = mddev->raid_disks;
1863 /*
1864 * 0 for a fully functional array, 1 for a degraded array.
1865 */
1866 mddev->degraded = conf->failed_disks = conf->raid_disks - conf->working_disks;
1867 conf->mddev = mddev;
1868 conf->chunk_size = mddev->chunk_size;
1869 conf->level = mddev->level;
1870 conf->algorithm = mddev->layout;
1871 conf->max_nr_stripes = NR_STRIPES;
1872
1873 /* device size must be a multiple of chunk size */
1874 mddev->size &= ~(mddev->chunk_size/1024 -1);
NeilBrownb1581562005-07-31 22:34:50 -07001875 mddev->resync_max_sectors = mddev->size << 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876
1877 if (!conf->chunk_size || conf->chunk_size % 4) {
1878 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
1879 conf->chunk_size, mdname(mddev));
1880 goto abort;
1881 }
1882 if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
1883 printk(KERN_ERR
1884 "raid5: unsupported parity algorithm %d for %s\n",
1885 conf->algorithm, mdname(mddev));
1886 goto abort;
1887 }
1888 if (mddev->degraded > 1) {
1889 printk(KERN_ERR "raid5: not enough operational devices for %s"
1890 " (%d/%d failed)\n",
1891 mdname(mddev), conf->failed_disks, conf->raid_disks);
1892 goto abort;
1893 }
1894
1895 if (mddev->degraded == 1 &&
1896 mddev->recovery_cp != MaxSector) {
1897 printk(KERN_ERR
1898 "raid5: cannot start dirty degraded array for %s\n",
1899 mdname(mddev));
1900 goto abort;
1901 }
1902
1903 {
1904 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
1905 if (!mddev->thread) {
1906 printk(KERN_ERR
1907 "raid5: couldn't allocate thread for %s\n",
1908 mdname(mddev));
1909 goto abort;
1910 }
1911 }
1912memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
1913 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
1914 if (grow_stripes(conf, conf->max_nr_stripes)) {
1915 printk(KERN_ERR
1916 "raid5: couldn't allocate %dkB for buffers\n", memory);
1917 shrink_stripes(conf);
1918 md_unregister_thread(mddev->thread);
1919 goto abort;
1920 } else
1921 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
1922 memory, mdname(mddev));
1923
1924 if (mddev->degraded == 0)
1925 printk("raid5: raid level %d set %s active with %d out of %d"
1926 " devices, algorithm %d\n", conf->level, mdname(mddev),
1927 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
1928 conf->algorithm);
1929 else
1930 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
1931 " out of %d devices, algorithm %d\n", conf->level,
1932 mdname(mddev), mddev->raid_disks - mddev->degraded,
1933 mddev->raid_disks, conf->algorithm);
1934
1935 print_raid5_conf(conf);
1936
1937 /* read-ahead size must cover two whole stripes, which is
1938 * 2 * (n-1) * chunksize where 'n' is the number of raid devices
1939 */
1940 {
1941 int stripe = (mddev->raid_disks-1) * mddev->chunk_size
1942 / PAGE_CACHE_SIZE;
1943 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
1944 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
1945 }
1946
1947 /* Ok, everything is just fine now */
NeilBrown007583c2005-11-08 21:39:30 -08001948 sysfs_create_group(&mddev->kobj, &raid5_attrs_group);
NeilBrown7a5febe2005-05-16 21:53:16 -07001949
NeilBrown72626682005-09-09 16:23:54 -07001950 if (mddev->bitmap)
1951 mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
1952
NeilBrown7a5febe2005-05-16 21:53:16 -07001953 mddev->queue->unplug_fn = raid5_unplug_device;
1954 mddev->queue->issue_flush_fn = raid5_issue_flush;
1955
Linus Torvalds1da177e2005-04-16 15:20:36 -07001956 mddev->array_size = mddev->size * (mddev->raid_disks - 1);
1957 return 0;
1958abort:
1959 if (conf) {
1960 print_raid5_conf(conf);
1961 if (conf->stripe_hashtbl)
1962 free_pages((unsigned long) conf->stripe_hashtbl,
1963 HASH_PAGES_ORDER);
1964 kfree(conf);
1965 }
1966 mddev->private = NULL;
1967 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
1968 return -EIO;
1969}
1970
1971
1972
NeilBrown3f294f42005-11-08 21:39:25 -08001973static int stop(mddev_t *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001974{
1975 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1976
1977 md_unregister_thread(mddev->thread);
1978 mddev->thread = NULL;
1979 shrink_stripes(conf);
1980 free_pages((unsigned long) conf->stripe_hashtbl, HASH_PAGES_ORDER);
1981 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
NeilBrown007583c2005-11-08 21:39:30 -08001982 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001983 mddev->private = NULL;
1984 return 0;
1985}
1986
1987#if RAID5_DEBUG
1988static void print_sh (struct stripe_head *sh)
1989{
1990 int i;
1991
1992 printk("sh %llu, pd_idx %d, state %ld.\n",
1993 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
1994 printk("sh %llu, count %d.\n",
1995 (unsigned long long)sh->sector, atomic_read(&sh->count));
1996 printk("sh %llu, ", (unsigned long long)sh->sector);
1997 for (i = 0; i < sh->raid_conf->raid_disks; i++) {
1998 printk("(cache%d: %p %ld) ",
1999 i, sh->dev[i].page, sh->dev[i].flags);
2000 }
2001 printk("\n");
2002}
2003
2004static void printall (raid5_conf_t *conf)
2005{
2006 struct stripe_head *sh;
2007 int i;
2008
2009 spin_lock_irq(&conf->device_lock);
2010 for (i = 0; i < NR_HASH; i++) {
2011 sh = conf->stripe_hashtbl[i];
2012 for (; sh; sh = sh->hash_next) {
2013 if (sh->raid_conf != conf)
2014 continue;
2015 print_sh(sh);
2016 }
2017 }
2018 spin_unlock_irq(&conf->device_lock);
2019}
2020#endif
2021
2022static void status (struct seq_file *seq, mddev_t *mddev)
2023{
2024 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
2025 int i;
2026
2027 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
2028 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->working_disks);
2029 for (i = 0; i < conf->raid_disks; i++)
2030 seq_printf (seq, "%s",
2031 conf->disks[i].rdev &&
2032 conf->disks[i].rdev->in_sync ? "U" : "_");
2033 seq_printf (seq, "]");
2034#if RAID5_DEBUG
2035#define D(x) \
2036 seq_printf (seq, "<"#x":%d>", atomic_read(&conf->x))
2037 printall(conf);
2038#endif
2039}
2040
2041static void print_raid5_conf (raid5_conf_t *conf)
2042{
2043 int i;
2044 struct disk_info *tmp;
2045
2046 printk("RAID5 conf printout:\n");
2047 if (!conf) {
2048 printk("(conf==NULL)\n");
2049 return;
2050 }
2051 printk(" --- rd:%d wd:%d fd:%d\n", conf->raid_disks,
2052 conf->working_disks, conf->failed_disks);
2053
2054 for (i = 0; i < conf->raid_disks; i++) {
2055 char b[BDEVNAME_SIZE];
2056 tmp = conf->disks + i;
2057 if (tmp->rdev)
2058 printk(" disk %d, o:%d, dev:%s\n",
2059 i, !tmp->rdev->faulty,
2060 bdevname(tmp->rdev->bdev,b));
2061 }
2062}
2063
2064static int raid5_spare_active(mddev_t *mddev)
2065{
2066 int i;
2067 raid5_conf_t *conf = mddev->private;
2068 struct disk_info *tmp;
2069
2070 for (i = 0; i < conf->raid_disks; i++) {
2071 tmp = conf->disks + i;
2072 if (tmp->rdev
2073 && !tmp->rdev->faulty
2074 && !tmp->rdev->in_sync) {
2075 mddev->degraded--;
2076 conf->failed_disks--;
2077 conf->working_disks++;
2078 tmp->rdev->in_sync = 1;
2079 }
2080 }
2081 print_raid5_conf(conf);
2082 return 0;
2083}
2084
2085static int raid5_remove_disk(mddev_t *mddev, int number)
2086{
2087 raid5_conf_t *conf = mddev->private;
2088 int err = 0;
2089 mdk_rdev_t *rdev;
2090 struct disk_info *p = conf->disks + number;
2091
2092 print_raid5_conf(conf);
2093 rdev = p->rdev;
2094 if (rdev) {
2095 if (rdev->in_sync ||
2096 atomic_read(&rdev->nr_pending)) {
2097 err = -EBUSY;
2098 goto abort;
2099 }
2100 p->rdev = NULL;
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002101 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002102 if (atomic_read(&rdev->nr_pending)) {
2103 /* lost the race, try later */
2104 err = -EBUSY;
2105 p->rdev = rdev;
2106 }
2107 }
2108abort:
2109
2110 print_raid5_conf(conf);
2111 return err;
2112}
2113
2114static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
2115{
2116 raid5_conf_t *conf = mddev->private;
2117 int found = 0;
2118 int disk;
2119 struct disk_info *p;
2120
2121 if (mddev->degraded > 1)
2122 /* no point adding a device */
2123 return 0;
2124
2125 /*
2126 * find the disk ...
2127 */
2128 for (disk=0; disk < mddev->raid_disks; disk++)
2129 if ((p=conf->disks + disk)->rdev == NULL) {
2130 rdev->in_sync = 0;
2131 rdev->raid_disk = disk;
2132 found = 1;
NeilBrown72626682005-09-09 16:23:54 -07002133 if (rdev->saved_raid_disk != disk)
2134 conf->fullsync = 1;
Suzanne Woodd6065f72005-11-08 21:39:27 -08002135 rcu_assign_pointer(p->rdev, rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136 break;
2137 }
2138 print_raid5_conf(conf);
2139 return found;
2140}
2141
2142static int raid5_resize(mddev_t *mddev, sector_t sectors)
2143{
2144 /* no resync is happening, and there is enough space
2145 * on all devices, so we can resize.
2146 * We need to make sure resync covers any new space.
2147 * If the array is shrinking we should possibly wait until
2148 * any io in the removed space completes, but it hardly seems
2149 * worth it.
2150 */
2151 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
2152 mddev->array_size = (sectors * (mddev->raid_disks-1))>>1;
2153 set_capacity(mddev->gendisk, mddev->array_size << 1);
2154 mddev->changed = 1;
2155 if (sectors/2 > mddev->size && mddev->recovery_cp == MaxSector) {
2156 mddev->recovery_cp = mddev->size << 1;
2157 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2158 }
2159 mddev->size = sectors /2;
NeilBrown4b5c7ae2005-07-27 11:43:28 -07002160 mddev->resync_max_sectors = sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002161 return 0;
2162}
2163
NeilBrown72626682005-09-09 16:23:54 -07002164static void raid5_quiesce(mddev_t *mddev, int state)
2165{
2166 raid5_conf_t *conf = mddev_to_conf(mddev);
2167
2168 switch(state) {
2169 case 1: /* stop all writes */
2170 spin_lock_irq(&conf->device_lock);
2171 conf->quiesce = 1;
2172 wait_event_lock_irq(conf->wait_for_stripe,
2173 atomic_read(&conf->active_stripes) == 0,
2174 conf->device_lock, /* nothing */);
2175 spin_unlock_irq(&conf->device_lock);
2176 break;
2177
2178 case 0: /* re-enable writes */
2179 spin_lock_irq(&conf->device_lock);
2180 conf->quiesce = 0;
2181 wake_up(&conf->wait_for_stripe);
2182 spin_unlock_irq(&conf->device_lock);
2183 break;
2184 }
2185 if (mddev->thread) {
2186 if (mddev->bitmap)
2187 mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2188 else
2189 mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2190 md_wakeup_thread(mddev->thread);
2191 }
2192}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002193static mdk_personality_t raid5_personality=
2194{
2195 .name = "raid5",
2196 .owner = THIS_MODULE,
2197 .make_request = make_request,
2198 .run = run,
2199 .stop = stop,
2200 .status = status,
2201 .error_handler = error,
2202 .hot_add_disk = raid5_add_disk,
2203 .hot_remove_disk= raid5_remove_disk,
2204 .spare_active = raid5_spare_active,
2205 .sync_request = sync_request,
2206 .resize = raid5_resize,
NeilBrown72626682005-09-09 16:23:54 -07002207 .quiesce = raid5_quiesce,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208};
2209
2210static int __init raid5_init (void)
2211{
2212 return register_md_personality (RAID5, &raid5_personality);
2213}
2214
2215static void raid5_exit (void)
2216{
2217 unregister_md_personality (RAID5);
2218}
2219
2220module_init(raid5_init);
2221module_exit(raid5_exit);
2222MODULE_LICENSE("GPL");
2223MODULE_ALIAS("md-personality-4"); /* RAID5 */