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