[PATCH] md: Fix several raid1 bugs which cause a memory leak
[linux-2.6] / drivers / md / raid0.c
1 /*
2    raid0.c : Multiple Devices driver for Linux
3              Copyright (C) 1994-96 Marc ZYNGIER
4              <zyngier@ufr-info-p7.ibp.fr> or
5              <maz@gloups.fdn.fr>
6              Copyright (C) 1999, 2000 Ingo Molnar, Red Hat
7
8
9    RAID-0 management functions.
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 #include <linux/module.h>
22 #include <linux/raid/raid0.h>
23
24 #define MAJOR_NR MD_MAJOR
25 #define MD_DRIVER
26 #define MD_PERSONALITY
27
28 static void raid0_unplug(request_queue_t *q)
29 {
30         mddev_t *mddev = q->queuedata;
31         raid0_conf_t *conf = mddev_to_conf(mddev);
32         mdk_rdev_t **devlist = conf->strip_zone[0].dev;
33         int i;
34
35         for (i=0; i<mddev->raid_disks; i++) {
36                 request_queue_t *r_queue = bdev_get_queue(devlist[i]->bdev);
37
38                 if (r_queue->unplug_fn)
39                         r_queue->unplug_fn(r_queue);
40         }
41 }
42
43 static int raid0_issue_flush(request_queue_t *q, struct gendisk *disk,
44                              sector_t *error_sector)
45 {
46         mddev_t *mddev = q->queuedata;
47         raid0_conf_t *conf = mddev_to_conf(mddev);
48         mdk_rdev_t **devlist = conf->strip_zone[0].dev;
49         int i, ret = 0;
50
51         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
52                 struct block_device *bdev = devlist[i]->bdev;
53                 request_queue_t *r_queue = bdev_get_queue(bdev);
54
55                 if (!r_queue->issue_flush_fn)
56                         ret = -EOPNOTSUPP;
57                 else
58                         ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk, error_sector);
59         }
60         return ret;
61 }
62
63
64 static int create_strip_zones (mddev_t *mddev)
65 {
66         int i, c, j;
67         sector_t current_offset, curr_zone_offset;
68         sector_t min_spacing;
69         raid0_conf_t *conf = mddev_to_conf(mddev);
70         mdk_rdev_t *smallest, *rdev1, *rdev2, *rdev;
71         struct list_head *tmp1, *tmp2;
72         struct strip_zone *zone;
73         int cnt;
74         char b[BDEVNAME_SIZE];
75  
76         /*
77          * The number of 'same size groups'
78          */
79         conf->nr_strip_zones = 0;
80  
81         ITERATE_RDEV(mddev,rdev1,tmp1) {
82                 printk("raid0: looking at %s\n",
83                         bdevname(rdev1->bdev,b));
84                 c = 0;
85                 ITERATE_RDEV(mddev,rdev2,tmp2) {
86                         printk("raid0:   comparing %s(%llu)",
87                                bdevname(rdev1->bdev,b),
88                                (unsigned long long)rdev1->size);
89                         printk(" with %s(%llu)\n",
90                                bdevname(rdev2->bdev,b),
91                                (unsigned long long)rdev2->size);
92                         if (rdev2 == rdev1) {
93                                 printk("raid0:   END\n");
94                                 break;
95                         }
96                         if (rdev2->size == rdev1->size)
97                         {
98                                 /*
99                                  * Not unique, don't count it as a new
100                                  * group
101                                  */
102                                 printk("raid0:   EQUAL\n");
103                                 c = 1;
104                                 break;
105                         }
106                         printk("raid0:   NOT EQUAL\n");
107                 }
108                 if (!c) {
109                         printk("raid0:   ==> UNIQUE\n");
110                         conf->nr_strip_zones++;
111                         printk("raid0: %d zones\n", conf->nr_strip_zones);
112                 }
113         }
114         printk("raid0: FINAL %d zones\n", conf->nr_strip_zones);
115
116         conf->strip_zone = kzalloc(sizeof(struct strip_zone)*
117                                 conf->nr_strip_zones, GFP_KERNEL);
118         if (!conf->strip_zone)
119                 return 1;
120         conf->devlist = kzalloc(sizeof(mdk_rdev_t*)*
121                                 conf->nr_strip_zones*mddev->raid_disks,
122                                 GFP_KERNEL);
123         if (!conf->devlist)
124                 return 1;
125
126         /* The first zone must contain all devices, so here we check that
127          * there is a proper alignment of slots to devices and find them all
128          */
129         zone = &conf->strip_zone[0];
130         cnt = 0;
131         smallest = NULL;
132         zone->dev = conf->devlist;
133         ITERATE_RDEV(mddev, rdev1, tmp1) {
134                 int j = rdev1->raid_disk;
135
136                 if (j < 0 || j >= mddev->raid_disks) {
137                         printk("raid0: bad disk number %d - aborting!\n", j);
138                         goto abort;
139                 }
140                 if (zone->dev[j]) {
141                         printk("raid0: multiple devices for %d - aborting!\n",
142                                 j);
143                         goto abort;
144                 }
145                 zone->dev[j] = rdev1;
146
147                 blk_queue_stack_limits(mddev->queue,
148                                        rdev1->bdev->bd_disk->queue);
149                 /* as we don't honour merge_bvec_fn, we must never risk
150                  * violating it, so limit ->max_sector to one PAGE, as
151                  * a one page request is never in violation.
152                  */
153
154                 if (rdev1->bdev->bd_disk->queue->merge_bvec_fn &&
155                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
156                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
157
158                 if (!smallest || (rdev1->size <smallest->size))
159                         smallest = rdev1;
160                 cnt++;
161         }
162         if (cnt != mddev->raid_disks) {
163                 printk("raid0: too few disks (%d of %d) - aborting!\n",
164                         cnt, mddev->raid_disks);
165                 goto abort;
166         }
167         zone->nb_dev = cnt;
168         zone->size = smallest->size * cnt;
169         zone->zone_offset = 0;
170
171         current_offset = smallest->size;
172         curr_zone_offset = zone->size;
173
174         /* now do the other zones */
175         for (i = 1; i < conf->nr_strip_zones; i++)
176         {
177                 zone = conf->strip_zone + i;
178                 zone->dev = conf->strip_zone[i-1].dev + mddev->raid_disks;
179
180                 printk("raid0: zone %d\n", i);
181                 zone->dev_offset = current_offset;
182                 smallest = NULL;
183                 c = 0;
184
185                 for (j=0; j<cnt; j++) {
186                         char b[BDEVNAME_SIZE];
187                         rdev = conf->strip_zone[0].dev[j];
188                         printk("raid0: checking %s ...", bdevname(rdev->bdev,b));
189                         if (rdev->size > current_offset)
190                         {
191                                 printk(" contained as device %d\n", c);
192                                 zone->dev[c] = rdev;
193                                 c++;
194                                 if (!smallest || (rdev->size <smallest->size)) {
195                                         smallest = rdev;
196                                         printk("  (%llu) is smallest!.\n", 
197                                                 (unsigned long long)rdev->size);
198                                 }
199                         } else
200                                 printk(" nope.\n");
201                 }
202
203                 zone->nb_dev = c;
204                 zone->size = (smallest->size - current_offset) * c;
205                 printk("raid0: zone->nb_dev: %d, size: %llu\n",
206                         zone->nb_dev, (unsigned long long)zone->size);
207
208                 zone->zone_offset = curr_zone_offset;
209                 curr_zone_offset += zone->size;
210
211                 current_offset = smallest->size;
212                 printk("raid0: current zone offset: %llu\n",
213                         (unsigned long long)current_offset);
214         }
215
216         /* Now find appropriate hash spacing.
217          * We want a number which causes most hash entries to cover
218          * at most two strips, but the hash table must be at most
219          * 1 PAGE.  We choose the smallest strip, or contiguous collection
220          * of strips, that has big enough size.  We never consider the last
221          * strip though as it's size has no bearing on the efficacy of the hash
222          * table.
223          */
224         conf->hash_spacing = curr_zone_offset;
225         min_spacing = curr_zone_offset;
226         sector_div(min_spacing, PAGE_SIZE/sizeof(struct strip_zone*));
227         for (i=0; i < conf->nr_strip_zones-1; i++) {
228                 sector_t sz = 0;
229                 for (j=i; j<conf->nr_strip_zones-1 &&
230                              sz < min_spacing ; j++)
231                         sz += conf->strip_zone[j].size;
232                 if (sz >= min_spacing && sz < conf->hash_spacing)
233                         conf->hash_spacing = sz;
234         }
235
236         mddev->queue->unplug_fn = raid0_unplug;
237
238         mddev->queue->issue_flush_fn = raid0_issue_flush;
239
240         printk("raid0: done.\n");
241         return 0;
242  abort:
243         return 1;
244 }
245
246 /**
247  *      raid0_mergeable_bvec -- tell bio layer if a two requests can be merged
248  *      @q: request queue
249  *      @bio: the buffer head that's been built up so far
250  *      @biovec: the request that could be merged to it.
251  *
252  *      Return amount of bytes we can accept at this offset
253  */
254 static int raid0_mergeable_bvec(request_queue_t *q, struct bio *bio, struct bio_vec *biovec)
255 {
256         mddev_t *mddev = q->queuedata;
257         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
258         int max;
259         unsigned int chunk_sectors = mddev->chunk_size >> 9;
260         unsigned int bio_sectors = bio->bi_size >> 9;
261
262         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
263         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
264         if (max <= biovec->bv_len && bio_sectors == 0)
265                 return biovec->bv_len;
266         else 
267                 return max;
268 }
269
270 static int raid0_run (mddev_t *mddev)
271 {
272         unsigned  cur=0, i=0, nb_zone;
273         s64 size;
274         raid0_conf_t *conf;
275         mdk_rdev_t *rdev;
276         struct list_head *tmp;
277
278         if (mddev->chunk_size == 0) {
279                 printk(KERN_ERR "md/raid0: non-zero chunk size required.\n");
280                 return -EINVAL;
281         }
282         printk(KERN_INFO "%s: setting max_sectors to %d, segment boundary to %d\n",
283                mdname(mddev),
284                mddev->chunk_size >> 9,
285                (mddev->chunk_size>>1)-1);
286         blk_queue_max_sectors(mddev->queue, mddev->chunk_size >> 9);
287         blk_queue_segment_boundary(mddev->queue, (mddev->chunk_size>>1) - 1);
288
289         conf = kmalloc(sizeof (raid0_conf_t), GFP_KERNEL);
290         if (!conf)
291                 goto out;
292         mddev->private = (void *)conf;
293  
294         conf->strip_zone = NULL;
295         conf->devlist = NULL;
296         if (create_strip_zones (mddev)) 
297                 goto out_free_conf;
298
299         /* calculate array device size */
300         mddev->array_size = 0;
301         ITERATE_RDEV(mddev,rdev,tmp)
302                 mddev->array_size += rdev->size;
303
304         printk("raid0 : md_size is %llu blocks.\n", 
305                 (unsigned long long)mddev->array_size);
306         printk("raid0 : conf->hash_spacing is %llu blocks.\n",
307                 (unsigned long long)conf->hash_spacing);
308         {
309                 sector_t s = mddev->array_size;
310                 sector_t space = conf->hash_spacing;
311                 int round;
312                 conf->preshift = 0;
313                 if (sizeof(sector_t) > sizeof(u32)) {
314                         /*shift down space and s so that sector_div will work */
315                         while (space > (sector_t) (~(u32)0)) {
316                                 s >>= 1;
317                                 space >>= 1;
318                                 s += 1; /* force round-up */
319                                 conf->preshift++;
320                         }
321                 }
322                 round = sector_div(s, (u32)space) ? 1 : 0;
323                 nb_zone = s + round;
324         }
325         printk("raid0 : nb_zone is %d.\n", nb_zone);
326
327         printk("raid0 : Allocating %Zd bytes for hash.\n",
328                                 nb_zone*sizeof(struct strip_zone*));
329         conf->hash_table = kmalloc (sizeof (struct strip_zone *)*nb_zone, GFP_KERNEL);
330         if (!conf->hash_table)
331                 goto out_free_conf;
332         size = conf->strip_zone[cur].size;
333
334         for (i=0; i< nb_zone; i++) {
335                 conf->hash_table[i] = conf->strip_zone + cur;
336                 while (size <= conf->hash_spacing) {
337                         cur++;
338                         size += conf->strip_zone[cur].size;
339                 }
340                 size -= conf->hash_spacing;
341         }
342         if (conf->preshift) {
343                 conf->hash_spacing >>= conf->preshift;
344                 /* round hash_spacing up so when we divide by it, we
345                  * err on the side of too-low, which is safest
346                  */
347                 conf->hash_spacing++;
348         }
349
350         /* calculate the max read-ahead size.
351          * For read-ahead of large files to be effective, we need to
352          * readahead at least twice a whole stripe. i.e. number of devices
353          * multiplied by chunk size times 2.
354          * If an individual device has an ra_pages greater than the
355          * chunk size, then we will not drive that device as hard as it
356          * wants.  We consider this a configuration error: a larger
357          * chunksize should be used in that case.
358          */
359         {
360                 int stripe = mddev->raid_disks * mddev->chunk_size / PAGE_SIZE;
361                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
362                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
363         }
364
365
366         blk_queue_merge_bvec(mddev->queue, raid0_mergeable_bvec);
367         return 0;
368
369 out_free_conf:
370         kfree(conf->strip_zone);
371         kfree(conf->devlist);
372         kfree(conf);
373         mddev->private = NULL;
374 out:
375         return -ENOMEM;
376 }
377
378 static int raid0_stop (mddev_t *mddev)
379 {
380         raid0_conf_t *conf = mddev_to_conf(mddev);
381
382         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
383         kfree(conf->hash_table);
384         conf->hash_table = NULL;
385         kfree(conf->strip_zone);
386         conf->strip_zone = NULL;
387         kfree(conf);
388         mddev->private = NULL;
389
390         return 0;
391 }
392
393 static int raid0_make_request (request_queue_t *q, struct bio *bio)
394 {
395         mddev_t *mddev = q->queuedata;
396         unsigned int sect_in_chunk, chunksize_bits,  chunk_size, chunk_sects;
397         raid0_conf_t *conf = mddev_to_conf(mddev);
398         struct strip_zone *zone;
399         mdk_rdev_t *tmp_dev;
400         unsigned long chunk;
401         sector_t block, rsect;
402         const int rw = bio_data_dir(bio);
403
404         if (unlikely(bio_barrier(bio))) {
405                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
406                 return 0;
407         }
408
409         disk_stat_inc(mddev->gendisk, ios[rw]);
410         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
411
412         chunk_size = mddev->chunk_size >> 10;
413         chunk_sects = mddev->chunk_size >> 9;
414         chunksize_bits = ffz(~chunk_size);
415         block = bio->bi_sector >> 1;
416         
417
418         if (unlikely(chunk_sects < (bio->bi_sector & (chunk_sects - 1)) + (bio->bi_size >> 9))) {
419                 struct bio_pair *bp;
420                 /* Sanity check -- queue functions should prevent this happening */
421                 if (bio->bi_vcnt != 1 ||
422                     bio->bi_idx != 0)
423                         goto bad_map;
424                 /* This is a one page bio that upper layers
425                  * refuse to split for us, so we need to split it.
426                  */
427                 bp = bio_split(bio, bio_split_pool, chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
428                 if (raid0_make_request(q, &bp->bio1))
429                         generic_make_request(&bp->bio1);
430                 if (raid0_make_request(q, &bp->bio2))
431                         generic_make_request(&bp->bio2);
432
433                 bio_pair_release(bp);
434                 return 0;
435         }
436  
437
438         {
439                 sector_t x = block >> conf->preshift;
440                 sector_div(x, (u32)conf->hash_spacing);
441                 zone = conf->hash_table[x];
442         }
443  
444         while (block >= (zone->zone_offset + zone->size)) 
445                 zone++;
446     
447         sect_in_chunk = bio->bi_sector & ((chunk_size<<1) -1);
448
449
450         {
451                 sector_t x =  (block - zone->zone_offset) >> chunksize_bits;
452
453                 sector_div(x, zone->nb_dev);
454                 chunk = x;
455                 BUG_ON(x != (sector_t)chunk);
456
457                 x = block >> chunksize_bits;
458                 tmp_dev = zone->dev[sector_div(x, zone->nb_dev)];
459         }
460         rsect = (((chunk << chunksize_bits) + zone->dev_offset)<<1)
461                 + sect_in_chunk;
462  
463         bio->bi_bdev = tmp_dev->bdev;
464         bio->bi_sector = rsect + tmp_dev->data_offset;
465
466         /*
467          * Let the main block layer submit the IO and resolve recursion:
468          */
469         return 1;
470
471 bad_map:
472         printk("raid0_make_request bug: can't convert block across chunks"
473                 " or bigger than %dk %llu %d\n", chunk_size, 
474                 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
475
476         bio_io_error(bio, bio->bi_size);
477         return 0;
478 }
479                            
480 static void raid0_status (struct seq_file *seq, mddev_t *mddev)
481 {
482 #undef MD_DEBUG
483 #ifdef MD_DEBUG
484         int j, k, h;
485         char b[BDEVNAME_SIZE];
486         raid0_conf_t *conf = mddev_to_conf(mddev);
487   
488         h = 0;
489         for (j = 0; j < conf->nr_strip_zones; j++) {
490                 seq_printf(seq, "      z%d", j);
491                 if (conf->hash_table[h] == conf->strip_zone+j)
492                         seq_printf("(h%d)", h++);
493                 seq_printf(seq, "=[");
494                 for (k = 0; k < conf->strip_zone[j].nb_dev; k++)
495                         seq_printf (seq, "%s/", bdevname(
496                                 conf->strip_zone[j].dev[k]->bdev,b));
497
498                 seq_printf (seq, "] zo=%d do=%d s=%d\n",
499                                 conf->strip_zone[j].zone_offset,
500                                 conf->strip_zone[j].dev_offset,
501                                 conf->strip_zone[j].size);
502         }
503 #endif
504         seq_printf(seq, " %dk chunks", mddev->chunk_size/1024);
505         return;
506 }
507
508 static struct mdk_personality raid0_personality=
509 {
510         .name           = "raid0",
511         .level          = 0,
512         .owner          = THIS_MODULE,
513         .make_request   = raid0_make_request,
514         .run            = raid0_run,
515         .stop           = raid0_stop,
516         .status         = raid0_status,
517 };
518
519 static int __init raid0_init (void)
520 {
521         return register_md_personality (&raid0_personality);
522 }
523
524 static void raid0_exit (void)
525 {
526         unregister_md_personality (&raid0_personality);
527 }
528
529 module_init(raid0_init);
530 module_exit(raid0_exit);
531 MODULE_LICENSE("GPL");
532 MODULE_ALIAS("md-personality-2"); /* RAID0 */
533 MODULE_ALIAS("md-raid0");
534 MODULE_ALIAS("md-level-0");