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