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; 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                         break;
  58                 }
  59
  60                 ret =r_queue->issue_flush_fn(r_queue, bdev->bd_disk, error_sector);
  61                 if (ret)
  62                         break;
  63         }
  64         return ret;
  65 }
  66
  67
  68 static int create_strip_zones (mddev_t *mddev)
  69 {
  70         int i, c, j;
  71         sector_t current_offset, curr_zone_offset;
  72         sector_t min_spacing;
  73         raid0_conf_t *conf = mddev_to_conf(mddev);
  74         mdk_rdev_t *smallest, *rdev1, *rdev2, *rdev;
  75         struct list_head *tmp1, *tmp2;
  76         struct strip_zone *zone;
  77         int cnt;
  78         char b[BDEVNAME_SIZE];
  79
  80         /*
  81          * The number of 'same size groups'
  82          */
  83         conf->nr_strip_zones = 0;
  84
  85         ITERATE_RDEV(mddev,rdev1,tmp1) {
  86                 printk("raid0: looking at %s\n",
  87                         bdevname(rdev1->bdev,b));
  88                 c = 0;
  89                 ITERATE_RDEV(mddev,rdev2,tmp2) {
  90                         printk("raid0:   comparing %s(%llu)",
  91                                bdevname(rdev1->bdev,b),
  92                                (unsigned long long)rdev1->size);
  93                         printk(" with %s(%llu)\n",
  94                                bdevname(rdev2->bdev,b),
  95                                (unsigned long long)rdev2->size);
  96                         if (rdev2 == rdev1) {
  97                                 printk("raid0:   END\n");
  98                                 break;
  99                         }
 100                         if (rdev2->size == rdev1->size)
 101                         {
 102                                 /*
 103                                  * Not unique, don't count it as a new
 104                                  * group
 105                                  */
 106                                 printk("raid0:   EQUAL\n");
 107                                 c = 1;
 108                                 break;
 109                         }
 110                         printk("raid0:   NOT EQUAL\n");
 111                 }
 112                 if (!c) {
 113                         printk("raid0:   ==> UNIQUE\n");
 114                         conf->nr_strip_zones++;
 115                         printk("raid0: %d zones\n", conf->nr_strip_zones);
 116                 }
 117         }
 118         printk("raid0: FINAL %d zones\n", conf->nr_strip_zones);
 119
 120         conf->strip_zone = kmalloc(sizeof(struct strip_zone)*
 121                                 conf->nr_strip_zones, GFP_KERNEL);
 122         if (!conf->strip_zone)
 123                 return 1;
 124         conf->devlist = kmalloc(sizeof(mdk_rdev_t*)*
 125                                 conf->nr_strip_zones*mddev->raid_disks,
 126                                 GFP_KERNEL);
 127         if (!conf->devlist)
 128                 return 1;
 129
 130         memset(conf->strip_zone, 0,sizeof(struct strip_zone)*
 131                                    conf->nr_strip_zones);
 132         memset(conf->devlist, 0,
 133                sizeof(mdk_rdev_t*) * conf->nr_strip_zones * mddev->raid_disks);
 134
 135         /* The first zone must contain all devices, so here we check that
 136          * there is a proper alignment of slots to devices and find them all
 137          */
 138         zone = &conf->strip_zone[0];
 139         cnt = 0;
 140         smallest = NULL;
 141         zone->dev = conf->devlist;
 142         ITERATE_RDEV(mddev, rdev1, tmp1) {
 143                 int j = rdev1->raid_disk;
 144
 145                 if (j < 0 || j >= mddev->raid_disks) {
 146                         printk("raid0: bad disk number %d - aborting!\n", j);
 147                         goto abort;
 148                 }
 149                 if (zone->dev[j]) {
 150                         printk("raid0: multiple devices for %d - aborting!\n",
 151                                 j);
 152                         goto abort;
 153                 }
 154                 zone->dev[j] = rdev1;
 155
 156                 blk_queue_stack_limits(mddev->queue,
 157                                        rdev1->bdev->bd_disk->queue);
 158                 /* as we don't honour merge_bvec_fn, we must never risk
 159                  * violating it, so limit ->max_sector to one PAGE, as
 160                  * a one page request is never in violation.
 161                  */
 162
 163                 if (rdev1->bdev->bd_disk->queue->merge_bvec_fn &&
 164                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
 165                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
 166
 167                 if (!smallest || (rdev1->size <smallest->size))
 168                         smallest = rdev1;
 169                 cnt++;
 170         }
 171         if (cnt != mddev->raid_disks) {
 172                 printk("raid0: too few disks (%d of %d) - aborting!\n",
 173                         cnt, mddev->raid_disks);
 174                 goto abort;
 175         }
 176         zone->nb_dev = cnt;
 177         zone->size = smallest->size * cnt;
 178         zone->zone_offset = 0;
 179
 180         current_offset = smallest->size;
 181         curr_zone_offset = zone->size;
 182
 183         /* now do the other zones */
 184         for (i = 1; i < conf->nr_strip_zones; i++)
 185         {
 186                 zone = conf->strip_zone + i;
 187                 zone->dev = conf->strip_zone[i-1].dev + mddev->raid_disks;
 188
 189                 printk("raid0: zone %d\n", i);
 190                 zone->dev_offset = current_offset;
 191                 smallest = NULL;
 192                 c = 0;
 193
 194                 for (j=0; j<cnt; j++) {
 195                         char b[BDEVNAME_SIZE];
 196                         rdev = conf->strip_zone[0].dev[j];
 197                         printk("raid0: checking %s ...", bdevname(rdev->bdev,b));
 198                         if (rdev->size > current_offset)
 199                         {
 200                                 printk(" contained as device %d\n", c);
 201                                 zone->dev[c] = rdev;
 202                                 c++;
 203                                 if (!smallest || (rdev->size <smallest->size)) {
 204                                         smallest = rdev;
 205                                         printk("  (%llu) is smallest!.\n",
 206                                                 (unsigned long long)rdev->size);
 207                                 }
 208                         } else
 209                                 printk(" nope.\n");
 210                 }
 211
 212                 zone->nb_dev = c;
 213                 zone->size = (smallest->size - current_offset) * c;
 214                 printk("raid0: zone->nb_dev: %d, size: %llu\n",
 215                         zone->nb_dev, (unsigned long long)zone->size);
 216
 217                 zone->zone_offset = curr_zone_offset;
 218                 curr_zone_offset += zone->size;
 219
 220                 current_offset = smallest->size;
 221                 printk("raid0: current zone offset: %llu\n",
 222                         (unsigned long long)current_offset);
 223         }
 224
 225         /* Now find appropriate hash spacing.
 226          * We want a number which causes most hash entries to cover
 227          * at most two strips, but the hash table must be at most
 228          * 1 PAGE.  We choose the smallest strip, or contiguous collection
 229          * of strips, that has big enough size.  We never consider the last
 230          * strip though as it's size has no bearing on the efficacy of the hash
 231          * table.
 232          */
 233         conf->hash_spacing = curr_zone_offset;
 234         min_spacing = curr_zone_offset;
 235         sector_div(min_spacing, PAGE_SIZE/sizeof(struct strip_zone*));
 236         for (i=0; i < conf->nr_strip_zones-1; i++) {
 237                 sector_t sz = 0;
 238                 for (j=i; j<conf->nr_strip_zones-1 &&
 239                              sz < min_spacing ; j++)
 240                         sz += conf->strip_zone[j].size;
 241                 if (sz >= min_spacing && sz < conf->hash_spacing)
 242                         conf->hash_spacing = sz;
 243         }
 244
 245         mddev->queue->unplug_fn = raid0_unplug;
 246
 247         mddev->queue->issue_flush_fn = raid0_issue_flush;
 248
 249         printk("raid0: done.\n");
 250         return 0;
 251  abort:
 252         return 1;
 253 }
 254
 255 /**
 256  *      raid0_mergeable_bvec -- tell bio layer if a two requests can be merged
 257  *      @q: request queue
 258  *      @bio: the buffer head that's been built up so far
 259  *      @biovec: the request that could be merged to it.
 260  *
 261  *      Return amount of bytes we can accept at this offset
 262  */
 263 static int raid0_mergeable_bvec(request_queue_t *q, struct bio *bio, struct bio_vec *biovec)
 264 {
 265         mddev_t *mddev = q->queuedata;
 266         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
 267         int max;
 268         unsigned int chunk_sectors = mddev->chunk_size >> 9;
 269         unsigned int bio_sectors = bio->bi_size >> 9;
 270
 271         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
 272         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
 273         if (max <= biovec->bv_len && bio_sectors == 0)
 274                 return biovec->bv_len;
 275         else
 276                 return max;
 277 }
 278
 279 static int raid0_run (mddev_t *mddev)
 280 {
 281         unsigned  cur=0, i=0, nb_zone;
 282         s64 size;
 283         raid0_conf_t *conf;
 284         mdk_rdev_t *rdev;
 285         struct list_head *tmp;
 286
 287         printk("%s: setting max_sectors to %d, segment boundary to %d\n",
 288                mdname(mddev),
 289                mddev->chunk_size >> 9,
 290                (mddev->chunk_size>>1)-1);
 291         blk_queue_max_sectors(mddev->queue, mddev->chunk_size >> 9);
 292         blk_queue_segment_boundary(mddev->queue, (mddev->chunk_size>>1) - 1);
 293
 294         conf = kmalloc(sizeof (raid0_conf_t), GFP_KERNEL);
 295         if (!conf)
 296                 goto out;
 297         mddev->private = (void *)conf;
 298
 299         conf->strip_zone = NULL;
 300         conf->devlist = NULL;
 301         if (create_strip_zones (mddev))
 302                 goto out_free_conf;
 303
 304         /* calculate array device size */
 305         mddev->array_size = 0;
 306         ITERATE_RDEV(mddev,rdev,tmp)
 307                 mddev->array_size += rdev->size;
 308
 309         printk("raid0 : md_size is %llu blocks.\n",
 310                 (unsigned long long)mddev->array_size);
 311         printk("raid0 : conf->hash_spacing is %llu blocks.\n",
 312                 (unsigned long long)conf->hash_spacing);
 313         {
 314 #if __GNUC__ < 3
 315                 volatile
 316 #endif
 317                 sector_t s = mddev->array_size;
 318                 sector_t space = conf->hash_spacing;
 319                 int round;
 320                 conf->preshift = 0;
 321                 if (sizeof(sector_t) > sizeof(unsigned long)) {
 322                         /*shift down space and s so that sector_div will work */
 323                         while (space > (sector_t) (~(unsigned long)0)) {
 324                                 s >>= 1;
 325                                 space >>= 1;
 326                                 s += 1; /* force round-up */
 327                                 conf->preshift++;
 328                         }
 329                 }
 330                 round = sector_div(s, (unsigned long)space) ? 1 : 0;
 331                 nb_zone = s + round;
 332         }
 333         printk("raid0 : nb_zone is %d.\n", nb_zone);
 334
 335         printk("raid0 : Allocating %Zd bytes for hash.\n",
 336                                 nb_zone*sizeof(struct strip_zone*));
 337         conf->hash_table = kmalloc (sizeof (struct strip_zone *)*nb_zone, GFP_KERNEL);
 338         if (!conf->hash_table)
 339                 goto out_free_conf;
 340         size = conf->strip_zone[cur].size;
 341
 342         for (i=0; i< nb_zone; i++) {
 343                 conf->hash_table[i] = conf->strip_zone + cur;
 344                 while (size <= conf->hash_spacing) {
 345                         cur++;
 346                         size += conf->strip_zone[cur].size;
 347                 }
 348                 size -= conf->hash_spacing;
 349         }
 350         if (conf->preshift) {
 351                 conf->hash_spacing >>= conf->preshift;
 352                 /* round hash_spacing up so when we divide by it, we
 353                  * err on the side of too-low, which is safest
 354                  */
 355                 conf->hash_spacing++;
 356         }
 357
 358         /* calculate the max read-ahead size.
 359          * For read-ahead of large files to be effective, we need to
 360          * readahead at least twice a whole stripe. i.e. number of devices
 361          * multiplied by chunk size times 2.
 362          * If an individual device has an ra_pages greater than the
 363          * chunk size, then we will not drive that device as hard as it
 364          * wants.  We consider this a configuration error: a larger
 365          * chunksize should be used in that case.
 366          */
 367         {
 368                 int stripe = mddev->raid_disks * mddev->chunk_size / PAGE_CACHE_SIZE;
 369                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
 370                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
 371         }
 372
 373
 374         blk_queue_merge_bvec(mddev->queue, raid0_mergeable_bvec);
 375         return 0;
 376
 377 out_free_conf:
 378         if (conf->strip_zone)
 379                 kfree(conf->strip_zone);
 380         if (conf->devlist)
 381                 kfree (conf->devlist);
 382         kfree(conf);
 383         mddev->private = NULL;
 384 out:
 385         return 1;
 386 }
 387
 388 static int raid0_stop (mddev_t *mddev)
 389 {
 390         raid0_conf_t *conf = mddev_to_conf(mddev);
 391
 392         kfree (conf->hash_table);
 393         conf->hash_table = NULL;
 394         kfree (conf->strip_zone);
 395         conf->strip_zone = NULL;
 396         kfree (conf);
 397         mddev->private = NULL;
 398
 399         return 0;
 400 }
 401
 402 static int raid0_make_request (request_queue_t *q, struct bio *bio)
 403 {
 404         mddev_t *mddev = q->queuedata;
 405         unsigned int sect_in_chunk, chunksize_bits,  chunk_size, chunk_sects;
 406         raid0_conf_t *conf = mddev_to_conf(mddev);
 407         struct strip_zone *zone;
 408         mdk_rdev_t *tmp_dev;
 409         unsigned long chunk;
 410         sector_t block, rsect;
 411
 412         if (bio_data_dir(bio)==WRITE) {
 413                 disk_stat_inc(mddev->gendisk, writes);
 414                 disk_stat_add(mddev->gendisk, write_sectors, bio_sectors(bio));
 415         } else {
 416                 disk_stat_inc(mddev->gendisk, reads);
 417                 disk_stat_add(mddev->gendisk, read_sectors, bio_sectors(bio));
 418         }
 419
 420         chunk_size = mddev->chunk_size >> 10;
 421         chunk_sects = mddev->chunk_size >> 9;
 422         chunksize_bits = ffz(~chunk_size);
 423         block = bio->bi_sector >> 1;
 424
 425
 426         if (unlikely(chunk_sects < (bio->bi_sector & (chunk_sects - 1)) + (bio->bi_size >> 9))) {
 427                 struct bio_pair *bp;
 428                 /* Sanity check -- queue functions should prevent this happening */
 429                 if (bio->bi_vcnt != 1 ||
 430                     bio->bi_idx != 0)
 431                         goto bad_map;
 432                 /* This is a one page bio that upper layers
 433                  * refuse to split for us, so we need to split it.
 434                  */
 435                 bp = bio_split(bio, bio_split_pool, chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
 436                 if (raid0_make_request(q, &bp->bio1))
 437                         generic_make_request(&bp->bio1);
 438                 if (raid0_make_request(q, &bp->bio2))
 439                         generic_make_request(&bp->bio2);
 440
 441                 bio_pair_release(bp);
 442                 return 0;
 443         }
 444
 445
 446         {
 447 #if __GNUC__ < 3
 448                 volatile
 449 #endif
 450                 sector_t x = block >> conf->preshift;
 451                 sector_div(x, (unsigned long)conf->hash_spacing);
 452                 zone = conf->hash_table[x];
 453         }
 454
 455         while (block >= (zone->zone_offset + zone->size))
 456                 zone++;
 457
 458         sect_in_chunk = bio->bi_sector & ((chunk_size<<1) -1);
 459
 460
 461         {
 462                 sector_t x =  (block - zone->zone_offset) >> chunksize_bits;
 463
 464                 sector_div(x, zone->nb_dev);
 465                 chunk = x;
 466                 BUG_ON(x != (sector_t)chunk);
 467
 468                 x = block >> chunksize_bits;
 469                 tmp_dev = zone->dev[sector_div(x, zone->nb_dev)];
 470         }
 471         rsect = (((chunk << chunksize_bits) + zone->dev_offset)<<1)
 472                 + sect_in_chunk;
 473
 474         bio->bi_bdev = tmp_dev->bdev;
 475         bio->bi_sector = rsect + tmp_dev->data_offset;
 476
 477         /*
 478          * Let the main block layer submit the IO and resolve recursion:
 479          */
 480         return 1;
 481
 482 bad_map:
 483         printk("raid0_make_request bug: can't convert block across chunks"
 484                 " or bigger than %dk %llu %d\n", chunk_size,
 485                 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 486
 487         bio_io_error(bio, bio->bi_size);
 488         return 0;
 489 }
 490
 491 static void raid0_status (struct seq_file *seq, mddev_t *mddev)
 492 {
 493 #undef MD_DEBUG
 494 #ifdef MD_DEBUG
 495         int j, k, h;
 496         char b[BDEVNAME_SIZE];
 497         raid0_conf_t *conf = mddev_to_conf(mddev);
 498
 499         h = 0;
 500         for (j = 0; j < conf->nr_strip_zones; j++) {
 501                 seq_printf(seq, "      z%d", j);
 502                 if (conf->hash_table[h] == conf->strip_zone+j)
 503                         seq_printf("(h%d)", h++);
 504                 seq_printf(seq, "=[");
 505                 for (k = 0; k < conf->strip_zone[j].nb_dev; k++)
 506                         seq_printf (seq, "%s/", bdevname(
 507                                 conf->strip_zone[j].dev[k]->bdev,b));
 508
 509                 seq_printf (seq, "] zo=%d do=%d s=%d\n",
 510                                 conf->strip_zone[j].zone_offset,
 511                                 conf->strip_zone[j].dev_offset,
 512                                 conf->strip_zone[j].size);
 513         }
 514 #endif
 515         seq_printf(seq, " %dk chunks", mddev->chunk_size/1024);
 516         return;
 517 }
 518
 519 static mdk_personality_t raid0_personality=
 520 {
 521         .name           = "raid0",
 522         .owner          = THIS_MODULE,
 523         .make_request   = raid0_make_request,
 524         .run            = raid0_run,
 525         .stop           = raid0_stop,
 526         .status         = raid0_status,
 527 };
 528
 529 static int __init raid0_init (void)
 530 {
 531         return register_md_personality (RAID0, &raid0_personality);
 532 }
 533
 534 static void raid0_exit (void)
 535 {
 536         unregister_md_personality (RAID0);
 537 }
 538
 539 module_init(raid0_init);
 540 module_exit(raid0_exit);
 541 MODULE_LICENSE("GPL");
 542 MODULE_ALIAS("md-personality-2"); /* RAID0 */