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