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/raid/raid0.h>
  22
  23 static void raid0_unplug(struct request_queue *q)
  24 {
  25         mddev_t *mddev = q->queuedata;
  26         raid0_conf_t *conf = mddev_to_conf(mddev);
  27         mdk_rdev_t **devlist = conf->strip_zone[0].dev;
  28         int i;
  29
  30         for (i=0; i<mddev->raid_disks; i++) {
  31                 struct request_queue *r_queue = bdev_get_queue(devlist[i]->bdev);
  32
  33                 blk_unplug(r_queue);
  34         }
  35 }
  36
  37 static int raid0_congested(void *data, int bits)
  38 {
  39         mddev_t *mddev = data;
  40         raid0_conf_t *conf = mddev_to_conf(mddev);
  41         mdk_rdev_t **devlist = conf->strip_zone[0].dev;
  42         int i, ret = 0;
  43
  44         for (i = 0; i < mddev->raid_disks && !ret ; i++) {
  45                 struct request_queue *q = bdev_get_queue(devlist[i]->bdev);
  46
  47                 ret |= bdi_congested(&q->backing_dev_info, bits);
  48         }
  49         return ret;
  50 }
  51
  52
  53 static int create_strip_zones (mddev_t *mddev)
  54 {
  55         int i, c, j;
  56         sector_t current_start, curr_zone_start;
  57         sector_t min_spacing;
  58         raid0_conf_t *conf = mddev_to_conf(mddev);
  59         mdk_rdev_t *smallest, *rdev1, *rdev2, *rdev;
  60         struct list_head *tmp1, *tmp2;
  61         struct strip_zone *zone;
  62         int cnt;
  63         char b[BDEVNAME_SIZE];
  64
  65         /*
  66          * The number of 'same size groups'
  67          */
  68         conf->nr_strip_zones = 0;
  69
  70         rdev_for_each(rdev1, tmp1, mddev) {
  71                 printk(KERN_INFO "raid0: looking at %s\n",
  72                         bdevname(rdev1->bdev,b));
  73                 c = 0;
  74                 rdev_for_each(rdev2, tmp2, mddev) {
  75                         printk(KERN_INFO "raid0:   comparing %s(%llu)",
  76                                bdevname(rdev1->bdev,b),
  77                                (unsigned long long)rdev1->size);
  78                         printk(KERN_INFO " with %s(%llu)\n",
  79                                bdevname(rdev2->bdev,b),
  80                                (unsigned long long)rdev2->size);
  81                         if (rdev2 == rdev1) {
  82                                 printk(KERN_INFO "raid0:   END\n");
  83                                 break;
  84                         }
  85                         if (rdev2->size == rdev1->size)
  86                         {
  87                                 /*
  88                                  * Not unique, don't count it as a new
  89                                  * group
  90                                  */
  91                                 printk(KERN_INFO "raid0:   EQUAL\n");
  92                                 c = 1;
  93                                 break;
  94                         }
  95                         printk(KERN_INFO "raid0:   NOT EQUAL\n");
  96                 }
  97                 if (!c) {
  98                         printk(KERN_INFO "raid0:   ==> UNIQUE\n");
  99                         conf->nr_strip_zones++;
 100                         printk(KERN_INFO "raid0: %d zones\n",
 101                                 conf->nr_strip_zones);
 102                 }
 103         }
 104         printk(KERN_INFO "raid0: FINAL %d zones\n", conf->nr_strip_zones);
 105
 106         conf->strip_zone = kzalloc(sizeof(struct strip_zone)*
 107                                 conf->nr_strip_zones, GFP_KERNEL);
 108         if (!conf->strip_zone)
 109                 return 1;
 110         conf->devlist = kzalloc(sizeof(mdk_rdev_t*)*
 111                                 conf->nr_strip_zones*mddev->raid_disks,
 112                                 GFP_KERNEL);
 113         if (!conf->devlist)
 114                 return 1;
 115
 116         /* The first zone must contain all devices, so here we check that
 117          * there is a proper alignment of slots to devices and find them all
 118          */
 119         zone = &conf->strip_zone[0];
 120         cnt = 0;
 121         smallest = NULL;
 122         zone->dev = conf->devlist;
 123         rdev_for_each(rdev1, tmp1, mddev) {
 124                 int j = rdev1->raid_disk;
 125
 126                 if (j < 0 || j >= mddev->raid_disks) {
 127                         printk(KERN_ERR "raid0: bad disk number %d - "
 128                                 "aborting!\n", j);
 129                         goto abort;
 130                 }
 131                 if (zone->dev[j]) {
 132                         printk(KERN_ERR "raid0: multiple devices for %d - "
 133                                 "aborting!\n", j);
 134                         goto abort;
 135                 }
 136                 zone->dev[j] = rdev1;
 137
 138                 blk_queue_stack_limits(mddev->queue,
 139                                        rdev1->bdev->bd_disk->queue);
 140                 /* as we don't honour merge_bvec_fn, we must never risk
 141                  * violating it, so limit ->max_sector to one PAGE, as
 142                  * a one page request is never in violation.
 143                  */
 144
 145                 if (rdev1->bdev->bd_disk->queue->merge_bvec_fn &&
 146                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
 147                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
 148
 149                 if (!smallest || (rdev1->size <smallest->size))
 150                         smallest = rdev1;
 151                 cnt++;
 152         }
 153         if (cnt != mddev->raid_disks) {
 154                 printk(KERN_ERR "raid0: too few disks (%d of %d) - "
 155                         "aborting!\n", cnt, mddev->raid_disks);
 156                 goto abort;
 157         }
 158         zone->nb_dev = cnt;
 159         zone->sectors = smallest->size * cnt * 2;
 160         zone->zone_start = 0;
 161
 162         current_start = smallest->size * 2;
 163         curr_zone_start = zone->sectors;
 164
 165         /* now do the other zones */
 166         for (i = 1; i < conf->nr_strip_zones; i++)
 167         {
 168                 zone = conf->strip_zone + i;
 169                 zone->dev = conf->strip_zone[i-1].dev + mddev->raid_disks;
 170
 171                 printk(KERN_INFO "raid0: zone %d\n", i);
 172                 zone->dev_start = current_start;
 173                 smallest = NULL;
 174                 c = 0;
 175
 176                 for (j=0; j<cnt; j++) {
 177                         char b[BDEVNAME_SIZE];
 178                         rdev = conf->strip_zone[0].dev[j];
 179                         printk(KERN_INFO "raid0: checking %s ...",
 180                                 bdevname(rdev->bdev, b));
 181                         if (rdev->size > current_start / 2) {
 182                                 printk(KERN_INFO " contained as device %d\n",
 183                                         c);
 184                                 zone->dev[c] = rdev;
 185                                 c++;
 186                                 if (!smallest || (rdev->size <smallest->size)) {
 187                                         smallest = rdev;
 188                                         printk(KERN_INFO "  (%llu) is smallest!.\n",
 189                                                 (unsigned long long)rdev->size);
 190                                 }
 191                         } else
 192                                 printk(KERN_INFO " nope.\n");
 193                 }
 194
 195                 zone->nb_dev = c;
 196                 zone->sectors = (smallest->size * 2 - current_start) * c;
 197                 printk(KERN_INFO "raid0: zone->nb_dev: %d, sectors: %llu\n",
 198                         zone->nb_dev, (unsigned long long)zone->sectors);
 199
 200                 zone->zone_start = curr_zone_start;
 201                 curr_zone_start += zone->sectors;
 202
 203                 current_start = smallest->size * 2;
 204                 printk(KERN_INFO "raid0: current zone start: %llu\n",
 205                         (unsigned long long)current_start);
 206         }
 207
 208         /* Now find appropriate hash spacing.
 209          * We want a number which causes most hash entries to cover
 210          * at most two strips, but the hash table must be at most
 211          * 1 PAGE.  We choose the smallest strip, or contiguous collection
 212          * of strips, that has big enough size.  We never consider the last
 213          * strip though as it's size has no bearing on the efficacy of the hash
 214          * table.
 215          */
 216         conf->spacing = curr_zone_start;
 217         min_spacing = curr_zone_start;
 218         sector_div(min_spacing, PAGE_SIZE/sizeof(struct strip_zone*));
 219         for (i=0; i < conf->nr_strip_zones-1; i++) {
 220                 sector_t s = 0;
 221                 for (j = i; j < conf->nr_strip_zones - 1 &&
 222                                 s < min_spacing; j++)
 223                         s += conf->strip_zone[j].sectors;
 224                 if (s >= min_spacing && s < conf->spacing)
 225                         conf->spacing = s;
 226         }
 227
 228         mddev->queue->unplug_fn = raid0_unplug;
 229
 230         mddev->queue->backing_dev_info.congested_fn = raid0_congested;
 231         mddev->queue->backing_dev_info.congested_data = mddev;
 232
 233         printk(KERN_INFO "raid0: done.\n");
 234         return 0;
 235  abort:
 236         return 1;
 237 }
 238
 239 /**
 240  *      raid0_mergeable_bvec -- tell bio layer if a two requests can be merged
 241  *      @q: request queue
 242  *      @bvm: properties of new bio
 243  *      @biovec: the request that could be merged to it.
 244  *
 245  *      Return amount of bytes we can accept at this offset
 246  */
 247 static int raid0_mergeable_bvec(struct request_queue *q,
 248                                 struct bvec_merge_data *bvm,
 249                                 struct bio_vec *biovec)
 250 {
 251         mddev_t *mddev = q->queuedata;
 252         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 253         int max;
 254         unsigned int chunk_sectors = mddev->chunk_size >> 9;
 255         unsigned int bio_sectors = bvm->bi_size >> 9;
 256
 257         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
 258         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
 259         if (max <= biovec->bv_len && bio_sectors == 0)
 260                 return biovec->bv_len;
 261         else
 262                 return max;
 263 }
 264
 265 static int raid0_run (mddev_t *mddev)
 266 {
 267         unsigned  cur=0, i=0, nb_zone;
 268         s64 sectors;
 269         raid0_conf_t *conf;
 270         mdk_rdev_t *rdev;
 271         struct list_head *tmp;
 272
 273         if (mddev->chunk_size == 0) {
 274                 printk(KERN_ERR "md/raid0: non-zero chunk size required.\n");
 275                 return -EINVAL;
 276         }
 277         printk(KERN_INFO "%s: setting max_sectors to %d, segment boundary to %d\n",
 278                mdname(mddev),
 279                mddev->chunk_size >> 9,
 280                (mddev->chunk_size>>1)-1);
 281         blk_queue_max_sectors(mddev->queue, mddev->chunk_size >> 9);
 282         blk_queue_segment_boundary(mddev->queue, (mddev->chunk_size>>1) - 1);
 283         mddev->queue->queue_lock = &mddev->queue->__queue_lock;
 284
 285         conf = kmalloc(sizeof (raid0_conf_t), GFP_KERNEL);
 286         if (!conf)
 287                 goto out;
 288         mddev->private = (void *)conf;
 289
 290         conf->strip_zone = NULL;
 291         conf->devlist = NULL;
 292         if (create_strip_zones (mddev))
 293                 goto out_free_conf;
 294
 295         /* calculate array device size */
 296         mddev->array_sectors = 0;
 297         rdev_for_each(rdev, tmp, mddev)
 298                 mddev->array_sectors += rdev->size * 2;
 299
 300         printk(KERN_INFO "raid0 : md_size is %llu sectors.\n",
 301                 (unsigned long long)mddev->array_sectors);
 302         printk(KERN_INFO "raid0 : conf->spacing is %llu sectors.\n",
 303                 (unsigned long long)conf->spacing);
 304         {
 305                 sector_t s = mddev->array_sectors;
 306                 sector_t space = conf->spacing;
 307                 int round;
 308                 conf->sector_shift = 0;
 309                 if (sizeof(sector_t) > sizeof(u32)) {
 310                         /*shift down space and s so that sector_div will work */
 311                         while (space > (sector_t) (~(u32)0)) {
 312                                 s >>= 1;
 313                                 space >>= 1;
 314                                 s += 1; /* force round-up */
 315                                 conf->sector_shift++;
 316                         }
 317                 }
 318                 round = sector_div(s, (u32)space) ? 1 : 0;
 319                 nb_zone = s + round;
 320         }
 321         printk(KERN_INFO "raid0 : nb_zone is %d.\n", nb_zone);
 322
 323         printk(KERN_INFO "raid0 : Allocating %zu bytes for hash.\n",
 324                                 nb_zone*sizeof(struct strip_zone*));
 325         conf->hash_table = kmalloc (sizeof (struct strip_zone *)*nb_zone, GFP_KERNEL);
 326         if (!conf->hash_table)
 327                 goto out_free_conf;
 328         sectors = conf->strip_zone[cur].sectors;
 329
 330         conf->hash_table[0] = conf->strip_zone + cur;
 331         for (i=1; i< nb_zone; i++) {
 332                 while (sectors <= conf->spacing) {
 333                         cur++;
 334                         sectors += conf->strip_zone[cur].sectors;
 335                 }
 336                 sectors -= conf->spacing;
 337                 conf->hash_table[i] = conf->strip_zone + cur;
 338         }
 339         if (conf->sector_shift) {
 340                 conf->spacing >>= conf->sector_shift;
 341                 /* round spacing up so when we divide by it, we
 342                  * err on the side of too-low, which is safest
 343                  */
 344                 conf->spacing++;
 345         }
 346
 347         /* calculate the max read-ahead size.
 348          * For read-ahead of large files to be effective, we need to
 349          * readahead at least twice a whole stripe. i.e. number of devices
 350          * multiplied by chunk size times 2.
 351          * If an individual device has an ra_pages greater than the
 352          * chunk size, then we will not drive that device as hard as it
 353          * wants.  We consider this a configuration error: a larger
 354          * chunksize should be used in that case.
 355          */
 356         {
 357                 int stripe = mddev->raid_disks * mddev->chunk_size / PAGE_SIZE;
 358                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
 359                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
 360         }
 361
 362
 363         blk_queue_merge_bvec(mddev->queue, raid0_mergeable_bvec);
 364         return 0;
 365
 366 out_free_conf:
 367         kfree(conf->strip_zone);
 368         kfree(conf->devlist);
 369         kfree(conf);
 370         mddev->private = NULL;
 371 out:
 372         return -ENOMEM;
 373 }
 374
 375 static int raid0_stop (mddev_t *mddev)
 376 {
 377         raid0_conf_t *conf = mddev_to_conf(mddev);
 378
 379         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
 380         kfree(conf->hash_table);
 381         conf->hash_table = NULL;
 382         kfree(conf->strip_zone);
 383         conf->strip_zone = NULL;
 384         kfree(conf);
 385         mddev->private = NULL;
 386
 387         return 0;
 388 }
 389
 390 static int raid0_make_request (struct request_queue *q, struct bio *bio)
 391 {
 392         mddev_t *mddev = q->queuedata;
 393         unsigned int sect_in_chunk, chunksect_bits, chunk_sects;
 394         raid0_conf_t *conf = mddev_to_conf(mddev);
 395         struct strip_zone *zone;
 396         mdk_rdev_t *tmp_dev;
 397         sector_t chunk;
 398         sector_t sector, rsect;
 399         const int rw = bio_data_dir(bio);
 400         int cpu;
 401
 402         if (unlikely(bio_barrier(bio))) {
 403                 bio_endio(bio, -EOPNOTSUPP);
 404                 return 0;
 405         }
 406
 407         cpu = part_stat_lock();
 408         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
 409         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
 410                       bio_sectors(bio));
 411         part_stat_unlock();
 412
 413         chunk_sects = mddev->chunk_size >> 9;
 414         chunksect_bits = ffz(~chunk_sects);
 415         sector = bio->bi_sector;
 416
 417         if (unlikely(chunk_sects < (bio->bi_sector & (chunk_sects - 1)) + (bio->bi_size >> 9))) {
 418                 struct bio_pair *bp;
 419                 /* Sanity check -- queue functions should prevent this happening */
 420                 if (bio->bi_vcnt != 1 ||
 421                     bio->bi_idx != 0)
 422                         goto bad_map;
 423                 /* This is a one page bio that upper layers
 424                  * refuse to split for us, so we need to split it.
 425                  */
 426                 bp = bio_split(bio, chunk_sects - (bio->bi_sector & (chunk_sects - 1)));
 427                 if (raid0_make_request(q, &bp->bio1))
 428                         generic_make_request(&bp->bio1);
 429                 if (raid0_make_request(q, &bp->bio2))
 430                         generic_make_request(&bp->bio2);
 431
 432                 bio_pair_release(bp);
 433                 return 0;
 434         }
 435
 436
 437         {
 438                 sector_t x = sector >> conf->sector_shift;
 439                 sector_div(x, (u32)conf->spacing);
 440                 zone = conf->hash_table[x];
 441         }
 442
 443         while (sector >= zone->zone_start + zone->sectors)
 444                 zone++;
 445
 446         sect_in_chunk = bio->bi_sector & (chunk_sects - 1);
 447
 448
 449         {
 450                 sector_t x = (sector - zone->zone_start) >> chunksect_bits;
 451
 452                 sector_div(x, zone->nb_dev);
 453                 chunk = x;
 454
 455                 x = sector >> chunksect_bits;
 456                 tmp_dev = zone->dev[sector_div(x, zone->nb_dev)];
 457         }
 458         rsect = (chunk << chunksect_bits) + zone->dev_start + sect_in_chunk;
 459
 460         bio->bi_bdev = tmp_dev->bdev;
 461         bio->bi_sector = rsect + tmp_dev->data_offset;
 462
 463         /*
 464          * Let the main block layer submit the IO and resolve recursion:
 465          */
 466         return 1;
 467
 468 bad_map:
 469         printk("raid0_make_request bug: can't convert block across chunks"
 470                 " or bigger than %dk %llu %d\n", chunk_sects / 2,
 471                 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 472
 473         bio_io_error(bio);
 474         return 0;
 475 }
 476
 477 static void raid0_status (struct seq_file *seq, mddev_t *mddev)
 478 {
 479 #undef MD_DEBUG
 480 #ifdef MD_DEBUG
 481         int j, k, h;
 482         char b[BDEVNAME_SIZE];
 483         raid0_conf_t *conf = mddev_to_conf(mddev);
 484
 485         h = 0;
 486         for (j = 0; j < conf->nr_strip_zones; j++) {
 487                 seq_printf(seq, "      z%d", j);
 488                 if (conf->hash_table[h] == conf->strip_zone+j)
 489                         seq_printf(seq, "(h%d)", h++);
 490                 seq_printf(seq, "=[");
 491                 for (k = 0; k < conf->strip_zone[j].nb_dev; k++)
 492                         seq_printf(seq, "%s/", bdevname(
 493                                 conf->strip_zone[j].dev[k]->bdev,b));
 494
 495                 seq_printf(seq, "] zs=%d ds=%d s=%d\n",
 496                                 conf->strip_zone[j].zone_start,
 497                                 conf->strip_zone[j].dev_start,
 498                                 conf->strip_zone[j].sectors);
 499         }
 500 #endif
 501         seq_printf(seq, " %dk chunks", mddev->chunk_size/1024);
 502         return;
 503 }
 504
 505 static struct mdk_personality raid0_personality=
 506 {
 507         .name           = "raid0",
 508         .level          = 0,
 509         .owner          = THIS_MODULE,
 510         .make_request   = raid0_make_request,
 511         .run            = raid0_run,
 512         .stop           = raid0_stop,
 513         .status         = raid0_status,
 514 };
 515
 516 static int __init raid0_init (void)
 517 {
 518         return register_md_personality (&raid0_personality);
 519 }
 520
 521 static void raid0_exit (void)
 522 {
 523         unregister_md_personality (&raid0_personality);
 524 }
 525
 526 module_init(raid0_init);
 527 module_exit(raid0_exit);
 528 MODULE_LICENSE("GPL");
 529 MODULE_ALIAS("md-personality-2"); /* RAID0 */
 530 MODULE_ALIAS("md-raid0");
 531 MODULE_ALIAS("md-level-0");