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