Merge git://git.kernel.org/pub/scm/linux/kernel/git/sfrench/cifs-2.6
[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 static int raid0_congested(void *data, int bits)
64 {
65         mddev_t *mddev = data;
66         raid0_conf_t *conf = mddev_to_conf(mddev);
67         mdk_rdev_t **devlist = conf->strip_zone[0].dev;
68         int i, ret = 0;
69
70         for (i = 0; i < mddev->raid_disks && !ret ; i++) {
71                 request_queue_t *q = bdev_get_queue(devlist[i]->bdev);
72
73                 ret |= bdi_congested(&q->backing_dev_info, bits);
74         }
75         return ret;
76 }
77
78
79 static int create_strip_zones (mddev_t *mddev)
80 {
81         int i, c, j;
82         sector_t current_offset, curr_zone_offset;
83         sector_t min_spacing;
84         raid0_conf_t *conf = mddev_to_conf(mddev);
85         mdk_rdev_t *smallest, *rdev1, *rdev2, *rdev;
86         struct list_head *tmp1, *tmp2;
87         struct strip_zone *zone;
88         int cnt;
89         char b[BDEVNAME_SIZE];
90  
91         /*
92          * The number of 'same size groups'
93          */
94         conf->nr_strip_zones = 0;
95  
96         ITERATE_RDEV(mddev,rdev1,tmp1) {
97                 printk("raid0: looking at %s\n",
98                         bdevname(rdev1->bdev,b));
99                 c = 0;
100                 ITERATE_RDEV(mddev,rdev2,tmp2) {
101                         printk("raid0:   comparing %s(%llu)",
102                                bdevname(rdev1->bdev,b),
103                                (unsigned long long)rdev1->size);
104                         printk(" with %s(%llu)\n",
105                                bdevname(rdev2->bdev,b),
106                                (unsigned long long)rdev2->size);
107                         if (rdev2 == rdev1) {
108                                 printk("raid0:   END\n");
109                                 break;
110                         }
111                         if (rdev2->size == rdev1->size)
112                         {
113                                 /*
114                                  * Not unique, don't count it as a new
115                                  * group
116                                  */
117                                 printk("raid0:   EQUAL\n");
118                                 c = 1;
119                                 break;
120                         }
121                         printk("raid0:   NOT EQUAL\n");
122                 }
123                 if (!c) {
124                         printk("raid0:   ==> UNIQUE\n");
125                         conf->nr_strip_zones++;
126                         printk("raid0: %d zones\n", conf->nr_strip_zones);
127                 }
128         }
129         printk("raid0: FINAL %d zones\n", conf->nr_strip_zones);
130
131         conf->strip_zone = kzalloc(sizeof(struct strip_zone)*
132                                 conf->nr_strip_zones, GFP_KERNEL);
133         if (!conf->strip_zone)
134                 return 1;
135         conf->devlist = kzalloc(sizeof(mdk_rdev_t*)*
136                                 conf->nr_strip_zones*mddev->raid_disks,
137                                 GFP_KERNEL);
138         if (!conf->devlist)
139                 return 1;
140
141         /* The first zone must contain all devices, so here we check that
142          * there is a proper alignment of slots to devices and find them all
143          */
144         zone = &conf->strip_zone[0];
145         cnt = 0;
146         smallest = NULL;
147         zone->dev = conf->devlist;
148         ITERATE_RDEV(mddev, rdev1, tmp1) {
149                 int j = rdev1->raid_disk;
150
151                 if (j < 0 || j >= mddev->raid_disks) {
152                         printk("raid0: bad disk number %d - aborting!\n", j);
153                         goto abort;
154                 }
155                 if (zone->dev[j]) {
156                         printk("raid0: multiple devices for %d - aborting!\n",
157                                 j);
158                         goto abort;
159                 }
160                 zone->dev[j] = rdev1;
161
162                 blk_queue_stack_limits(mddev->queue,
163                                        rdev1->bdev->bd_disk->queue);
164                 /* as we don't honour merge_bvec_fn, we must never risk
165                  * violating it, so limit ->max_sector to one PAGE, as
166                  * a one page request is never in violation.
167                  */
168
169                 if (rdev1->bdev->bd_disk->queue->merge_bvec_fn &&
170                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
171                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
172
173                 if (!smallest || (rdev1->size <smallest->size))
174                         smallest = rdev1;
175                 cnt++;
176         }
177         if (cnt != mddev->raid_disks) {
178                 printk("raid0: too few disks (%d of %d) - aborting!\n",
179                         cnt, mddev->raid_disks);
180                 goto abort;
181         }
182         zone->nb_dev = cnt;
183         zone->size = smallest->size * cnt;
184         zone->zone_offset = 0;
185
186         current_offset = smallest->size;
187         curr_zone_offset = zone->size;
188
189         /* now do the other zones */
190         for (i = 1; i < conf->nr_strip_zones; i++)
191         {
192                 zone = conf->strip_zone + i;
193                 zone->dev = conf->strip_zone[i-1].dev + mddev->raid_disks;
194
195                 printk("raid0: zone %d\n", i);
196                 zone->dev_offset = current_offset;
197                 smallest = NULL;
198                 c = 0;
199
200                 for (j=0; j<cnt; j++) {
201                         char b[BDEVNAME_SIZE];
202                         rdev = conf->strip_zone[0].dev[j];
203                         printk("raid0: checking %s ...", bdevname(rdev->bdev,b));
204                         if (rdev->size > current_offset)
205                         {
206                                 printk(" contained as device %d\n", c);
207                                 zone->dev[c] = rdev;
208                                 c++;
209                                 if (!smallest || (rdev->size <smallest->size)) {
210                                         smallest = rdev;
211                                         printk("  (%llu) is smallest!.\n", 
212                                                 (unsigned long long)rdev->size);
213                                 }
214                         } else
215                                 printk(" nope.\n");
216                 }
217
218                 zone->nb_dev = c;
219                 zone->size = (smallest->size - current_offset) * c;
220                 printk("raid0: zone->nb_dev: %d, size: %llu\n",
221                         zone->nb_dev, (unsigned long long)zone->size);
222
223                 zone->zone_offset = curr_zone_offset;
224                 curr_zone_offset += zone->size;
225
226                 current_offset = smallest->size;
227                 printk("raid0: current zone offset: %llu\n",
228                         (unsigned long long)current_offset);
229         }
230
231         /* Now find appropriate hash spacing.
232          * We want a number which causes most hash entries to cover
233          * at most two strips, but the hash table must be at most
234          * 1 PAGE.  We choose the smallest strip, or contiguous collection
235          * of strips, that has big enough size.  We never consider the last
236          * strip though as it's size has no bearing on the efficacy of the hash
237          * table.
238          */
239         conf->hash_spacing = curr_zone_offset;
240         min_spacing = curr_zone_offset;
241         sector_div(min_spacing, PAGE_SIZE/sizeof(struct strip_zone*));
242         for (i=0; i < conf->nr_strip_zones-1; i++) {
243                 sector_t sz = 0;
244                 for (j=i; j<conf->nr_strip_zones-1 &&
245                              sz < min_spacing ; j++)
246                         sz += conf->strip_zone[j].size;
247                 if (sz >= min_spacing && sz < conf->hash_spacing)
248                         conf->hash_spacing = sz;
249         }
250
251         mddev->queue->unplug_fn = raid0_unplug;
252
253         mddev->queue->issue_flush_fn = raid0_issue_flush;
254         mddev->queue->backing_dev_info.congested_fn = raid0_congested;
255         mddev->queue->backing_dev_info.congested_data = mddev;
256
257         printk("raid0: done.\n");
258         return 0;
259  abort:
260         return 1;
261 }
262
263 /**
264  *      raid0_mergeable_bvec -- tell bio layer if a two requests can be merged
265  *      @q: request queue
266  *      @bio: the buffer head that's been built up so far
267  *      @biovec: the request that could be merged to it.
268  *
269  *      Return amount of bytes we can accept at this offset
270  */
271 static int raid0_mergeable_bvec(request_queue_t *q, struct bio *bio, struct bio_vec *biovec)
272 {
273         mddev_t *mddev = q->queuedata;
274         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
275         int max;
276         unsigned int chunk_sectors = mddev->chunk_size >> 9;
277         unsigned int bio_sectors = bio->bi_size >> 9;
278
279         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
280         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
281         if (max <= biovec->bv_len && bio_sectors == 0)
282                 return biovec->bv_len;
283         else 
284                 return max;
285 }
286
287 static int raid0_run (mddev_t *mddev)
288 {
289         unsigned  cur=0, i=0, nb_zone;
290         s64 size;
291         raid0_conf_t *conf;
292         mdk_rdev_t *rdev;
293         struct list_head *tmp;
294
295         if (mddev->chunk_size == 0) {
296                 printk(KERN_ERR "md/raid0: non-zero chunk size required.\n");
297                 return -EINVAL;
298         }
299         printk(KERN_INFO "%s: setting max_sectors to %d, segment boundary to %d\n",
300                mdname(mddev),
301                mddev->chunk_size >> 9,
302                (mddev->chunk_size>>1)-1);
303         blk_queue_max_sectors(mddev->queue, mddev->chunk_size >> 9);
304         blk_queue_segment_boundary(mddev->queue, (mddev->chunk_size>>1) - 1);
305
306         conf = kmalloc(sizeof (raid0_conf_t), GFP_KERNEL);
307         if (!conf)
308                 goto out;
309         mddev->private = (void *)conf;
310  
311         conf->strip_zone = NULL;
312         conf->devlist = NULL;
313         if (create_strip_zones (mddev)) 
314                 goto out_free_conf;
315
316         /* calculate array device size */
317         mddev->array_size = 0;
318         ITERATE_RDEV(mddev,rdev,tmp)
319                 mddev->array_size += rdev->size;
320
321         printk("raid0 : md_size is %llu blocks.\n", 
322                 (unsigned long long)mddev->array_size);
323         printk("raid0 : conf->hash_spacing is %llu blocks.\n",
324                 (unsigned long long)conf->hash_spacing);
325         {
326                 sector_t s = mddev->array_size;
327                 sector_t space = conf->hash_spacing;
328                 int round;
329                 conf->preshift = 0;
330                 if (sizeof(sector_t) > sizeof(u32)) {
331                         /*shift down space and s so that sector_div will work */
332                         while (space > (sector_t) (~(u32)0)) {
333                                 s >>= 1;
334                                 space >>= 1;
335                                 s += 1; /* force round-up */
336                                 conf->preshift++;
337                         }
338                 }
339                 round = sector_div(s, (u32)space) ? 1 : 0;
340                 nb_zone = s + round;
341         }
342         printk("raid0 : nb_zone is %d.\n", nb_zone);
343
344         printk("raid0 : Allocating %Zd bytes for hash.\n",
345                                 nb_zone*sizeof(struct strip_zone*));
346         conf->hash_table = kmalloc (sizeof (struct strip_zone *)*nb_zone, GFP_KERNEL);
347         if (!conf->hash_table)
348                 goto out_free_conf;
349         size = conf->strip_zone[cur].size;
350
351         conf->hash_table[0] = conf->strip_zone + cur;
352         for (i=1; i< nb_zone; i++) {
353                 while (size <= conf->hash_spacing) {
354                         cur++;
355                         size += conf->strip_zone[cur].size;
356                 }
357                 size -= conf->hash_spacing;
358                 conf->hash_table[i] = conf->strip_zone + cur;
359         }
360         if (conf->preshift) {
361                 conf->hash_spacing >>= conf->preshift;
362                 /* round hash_spacing up so when we divide by it, we
363                  * err on the side of too-low, which is safest
364                  */
365                 conf->hash_spacing++;
366         }
367
368         /* calculate the max read-ahead size.
369          * For read-ahead of large files to be effective, we need to
370          * readahead at least twice a whole stripe. i.e. number of devices
371          * multiplied by chunk size times 2.
372          * If an individual device has an ra_pages greater than the
373          * chunk size, then we will not drive that device as hard as it
374          * wants.  We consider this a configuration error: a larger
375          * chunksize should be used in that case.
376          */
377         {
378                 int stripe = mddev->raid_disks * mddev->chunk_size / PAGE_SIZE;
379                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
380                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
381         }
382
383
384         blk_queue_merge_bvec(mddev->queue, raid0_mergeable_bvec);
385         return 0;
386
387 out_free_conf:
388         kfree(conf->strip_zone);
389         kfree(conf->devlist);
390         kfree(conf);
391         mddev->private = NULL;
392 out:
393         return -ENOMEM;
394 }
395
396 static int raid0_stop (mddev_t *mddev)
397 {
398         raid0_conf_t *conf = mddev_to_conf(mddev);
399
400         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
401         kfree(conf->hash_table);
402         conf->hash_table = NULL;
403         kfree(conf->strip_zone);
404         conf->strip_zone = NULL;
405         kfree(conf);
406         mddev->private = NULL;
407
408         return 0;
409 }
410
411 static int raid0_make_request (request_queue_t *q, struct bio *bio)
412 {
413         mddev_t *mddev = q->queuedata;
414         unsigned int sect_in_chunk, chunksize_bits,  chunk_size, chunk_sects;
415         raid0_conf_t *conf = mddev_to_conf(mddev);
416         struct strip_zone *zone;
417         mdk_rdev_t *tmp_dev;
418         sector_t chunk;
419         sector_t block, rsect;
420         const int rw = bio_data_dir(bio);
421
422         if (unlikely(bio_barrier(bio))) {
423                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
424                 return 0;
425         }
426
427         disk_stat_inc(mddev->gendisk, ios[rw]);
428         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
429
430         chunk_size = mddev->chunk_size >> 10;
431         chunk_sects = mddev->chunk_size >> 9;
432         chunksize_bits = ffz(~chunk_size);
433         block = bio->bi_sector >> 1;
434         
435
436         if (unlikely(chunk_sects < (bio->bi_sector & (chunk_sects - 1)) + (bio->bi_size >> 9))) {
437                 struct bio_pair *bp;
438                 /* Sanity check -- queue functions should prevent this happening */
439                 if (bio->bi_vcnt != 1 ||
440                     bio->bi_idx != 0)
441                         goto bad_map;
442                 /* This is a one page bio that upper layers
443                  * refuse to split for us, so we need to split it.
444                  */
445                 bp = bio_split(bio, bio_split_pool, chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
446                 if (raid0_make_request(q, &bp->bio1))
447                         generic_make_request(&bp->bio1);
448                 if (raid0_make_request(q, &bp->bio2))
449                         generic_make_request(&bp->bio2);
450
451                 bio_pair_release(bp);
452                 return 0;
453         }
454  
455
456         {
457                 sector_t x = block >> conf->preshift;
458                 sector_div(x, (u32)conf->hash_spacing);
459                 zone = conf->hash_table[x];
460         }
461  
462         while (block >= (zone->zone_offset + zone->size)) 
463                 zone++;
464     
465         sect_in_chunk = bio->bi_sector & ((chunk_size<<1) -1);
466
467
468         {
469                 sector_t x =  (block - zone->zone_offset) >> chunksize_bits;
470
471                 sector_div(x, zone->nb_dev);
472                 chunk = x;
473
474                 x = block >> chunksize_bits;
475                 tmp_dev = zone->dev[sector_div(x, zone->nb_dev)];
476         }
477         rsect = (((chunk << chunksize_bits) + zone->dev_offset)<<1)
478                 + sect_in_chunk;
479  
480         bio->bi_bdev = tmp_dev->bdev;
481         bio->bi_sector = rsect + tmp_dev->data_offset;
482
483         /*
484          * Let the main block layer submit the IO and resolve recursion:
485          */
486         return 1;
487
488 bad_map:
489         printk("raid0_make_request bug: can't convert block across chunks"
490                 " or bigger than %dk %llu %d\n", chunk_size, 
491                 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
492
493         bio_io_error(bio, bio->bi_size);
494         return 0;
495 }
496                            
497 static void raid0_status (struct seq_file *seq, mddev_t *mddev)
498 {
499 #undef MD_DEBUG
500 #ifdef MD_DEBUG
501         int j, k, h;
502         char b[BDEVNAME_SIZE];
503         raid0_conf_t *conf = mddev_to_conf(mddev);
504   
505         h = 0;
506         for (j = 0; j < conf->nr_strip_zones; j++) {
507                 seq_printf(seq, "      z%d", j);
508                 if (conf->hash_table[h] == conf->strip_zone+j)
509                         seq_printf("(h%d)", h++);
510                 seq_printf(seq, "=[");
511                 for (k = 0; k < conf->strip_zone[j].nb_dev; k++)
512                         seq_printf (seq, "%s/", bdevname(
513                                 conf->strip_zone[j].dev[k]->bdev,b));
514
515                 seq_printf (seq, "] zo=%d do=%d s=%d\n",
516                                 conf->strip_zone[j].zone_offset,
517                                 conf->strip_zone[j].dev_offset,
518                                 conf->strip_zone[j].size);
519         }
520 #endif
521         seq_printf(seq, " %dk chunks", mddev->chunk_size/1024);
522         return;
523 }
524
525 static struct mdk_personality raid0_personality=
526 {
527         .name           = "raid0",
528         .level          = 0,
529         .owner          = THIS_MODULE,
530         .make_request   = raid0_make_request,
531         .run            = raid0_run,
532         .stop           = raid0_stop,
533         .status         = raid0_status,
534 };
535
536 static int __init raid0_init (void)
537 {
538         return register_md_personality (&raid0_personality);
539 }
540
541 static void raid0_exit (void)
542 {
543         unregister_md_personality (&raid0_personality);
544 }
545
546 module_init(raid0_init);
547 module_exit(raid0_exit);
548 MODULE_LICENSE("GPL");
549 MODULE_ALIAS("md-personality-2"); /* RAID0 */
550 MODULE_ALIAS("md-raid0");
551 MODULE_ALIAS("md-level-0");