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/blkdev.h>
  22 #include <linux/seq_file.h>
  23 #include "md.h"
  24 #include "raid0.h"
  25
  26 static void raid0_unplug(struct request_queue *q)
  27 {
  28         mddev_t *mddev = q->queuedata;
  29         raid0_conf_t *conf = mddev->private;
  30         mdk_rdev_t **devlist = conf->devlist;
  31         int i;
  32
  33         for (i=0; i<mddev->raid_disks; i++) {
  34                 struct request_queue *r_queue = bdev_get_queue(devlist[i]->bdev);
  35
  36                 blk_unplug(r_queue);
  37         }
  38 }
  39
  40 static int raid0_congested(void *data, int bits)
  41 {
  42         mddev_t *mddev = data;
  43         raid0_conf_t *conf = mddev->private;
  44         mdk_rdev_t **devlist = conf->devlist;
  45         int i, ret = 0;
  46
  47         if (mddev_congested(mddev, bits))
  48                 return 1;
  49
  50         for (i = 0; i < mddev->raid_disks && !ret ; i++) {
  51                 struct request_queue *q = bdev_get_queue(devlist[i]->bdev);
  52
  53                 ret |= bdi_congested(&q->backing_dev_info, bits);
  54         }
  55         return ret;
  56 }
  57
  58 /*
  59  * inform the user of the raid configuration
  60 */
  61 static void dump_zones(mddev_t *mddev)
  62 {
  63         int j, k, h;
  64         sector_t zone_size = 0;
  65         sector_t zone_start = 0;
  66         char b[BDEVNAME_SIZE];
  67         raid0_conf_t *conf = mddev->private;
  68         printk(KERN_INFO "******* %s configuration *********\n",
  69                 mdname(mddev));
  70         h = 0;
  71         for (j = 0; j < conf->nr_strip_zones; j++) {
  72                 printk(KERN_INFO "zone%d=[", j);
  73                 for (k = 0; k < conf->strip_zone[j].nb_dev; k++)
  74                         printk("%s/",
  75                         bdevname(conf->devlist[j*mddev->raid_disks
  76                                                 + k]->bdev, b));
  77                 printk("]\n");
  78
  79                 zone_size  = conf->strip_zone[j].zone_end - zone_start;
  80                 printk(KERN_INFO "        zone offset=%llukb "
  81                                 "device offset=%llukb size=%llukb\n",
  82                         (unsigned long long)zone_start>>1,
  83                         (unsigned long long)conf->strip_zone[j].dev_start>>1,
  84                         (unsigned long long)zone_size>>1);
  85                 zone_start = conf->strip_zone[j].zone_end;
  86         }
  87         printk(KERN_INFO "**********************************\n\n");
  88 }
  89
  90 static int create_strip_zones(mddev_t *mddev)
  91 {
  92         int i, c, err;
  93         sector_t curr_zone_end, sectors;
  94         mdk_rdev_t *smallest, *rdev1, *rdev2, *rdev, **dev;
  95         struct strip_zone *zone;
  96         int cnt;
  97         char b[BDEVNAME_SIZE];
  98         raid0_conf_t *conf = kzalloc(sizeof(*conf), GFP_KERNEL);
  99
 100         if (!conf)
 101                 return -ENOMEM;
 102         list_for_each_entry(rdev1, &mddev->disks, same_set) {
 103                 printk(KERN_INFO "raid0: looking at %s\n",
 104                         bdevname(rdev1->bdev,b));
 105                 c = 0;
 106
 107                 /* round size to chunk_size */
 108                 sectors = rdev1->sectors;
 109                 sector_div(sectors, mddev->chunk_sectors);
 110                 rdev1->sectors = sectors * mddev->chunk_sectors;
 111
 112                 list_for_each_entry(rdev2, &mddev->disks, same_set) {
 113                         printk(KERN_INFO "raid0:   comparing %s(%llu)",
 114                                bdevname(rdev1->bdev,b),
 115                                (unsigned long long)rdev1->sectors);
 116                         printk(KERN_INFO " with %s(%llu)\n",
 117                                bdevname(rdev2->bdev,b),
 118                                (unsigned long long)rdev2->sectors);
 119                         if (rdev2 == rdev1) {
 120                                 printk(KERN_INFO "raid0:   END\n");
 121                                 break;
 122                         }
 123                         if (rdev2->sectors == rdev1->sectors) {
 124                                 /*
 125                                  * Not unique, don't count it as a new
 126                                  * group
 127                                  */
 128                                 printk(KERN_INFO "raid0:   EQUAL\n");
 129                                 c = 1;
 130                                 break;
 131                         }
 132                         printk(KERN_INFO "raid0:   NOT EQUAL\n");
 133                 }
 134                 if (!c) {
 135                         printk(KERN_INFO "raid0:   ==> UNIQUE\n");
 136                         conf->nr_strip_zones++;
 137                         printk(KERN_INFO "raid0: %d zones\n",
 138                                 conf->nr_strip_zones);
 139                 }
 140         }
 141         printk(KERN_INFO "raid0: FINAL %d zones\n", conf->nr_strip_zones);
 142         err = -ENOMEM;
 143         conf->strip_zone = kzalloc(sizeof(struct strip_zone)*
 144                                 conf->nr_strip_zones, GFP_KERNEL);
 145         if (!conf->strip_zone)
 146                 goto abort;
 147         conf->devlist = kzalloc(sizeof(mdk_rdev_t*)*
 148                                 conf->nr_strip_zones*mddev->raid_disks,
 149                                 GFP_KERNEL);
 150         if (!conf->devlist)
 151                 goto abort;
 152
 153         /* The first zone must contain all devices, so here we check that
 154          * there is a proper alignment of slots to devices and find them all
 155          */
 156         zone = &conf->strip_zone[0];
 157         cnt = 0;
 158         smallest = NULL;
 159         dev = conf->devlist;
 160         err = -EINVAL;
 161         list_for_each_entry(rdev1, &mddev->disks, same_set) {
 162                 int j = rdev1->raid_disk;
 163
 164                 if (j < 0 || j >= mddev->raid_disks) {
 165                         printk(KERN_ERR "raid0: bad disk number %d - "
 166                                 "aborting!\n", j);
 167                         goto abort;
 168                 }
 169                 if (dev[j]) {
 170                         printk(KERN_ERR "raid0: multiple devices for %d - "
 171                                 "aborting!\n", j);
 172                         goto abort;
 173                 }
 174                 dev[j] = rdev1;
 175
 176                 disk_stack_limits(mddev->gendisk, rdev1->bdev,
 177                                   rdev1->data_offset << 9);
 178                 /* as we don't honour merge_bvec_fn, we must never risk
 179                  * violating it, so limit ->max_sector to one PAGE, as
 180                  * a one page request is never in violation.
 181                  */
 182
 183                 if (rdev1->bdev->bd_disk->queue->merge_bvec_fn &&
 184                     queue_max_sectors(mddev->queue) > (PAGE_SIZE>>9))
 185                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
 186
 187                 if (!smallest || (rdev1->sectors < smallest->sectors))
 188                         smallest = rdev1;
 189                 cnt++;
 190         }
 191         if (cnt != mddev->raid_disks) {
 192                 printk(KERN_ERR "raid0: too few disks (%d of %d) - "
 193                         "aborting!\n", cnt, mddev->raid_disks);
 194                 goto abort;
 195         }
 196         zone->nb_dev = cnt;
 197         zone->zone_end = smallest->sectors * cnt;
 198
 199         curr_zone_end = zone->zone_end;
 200
 201         /* now do the other zones */
 202         for (i = 1; i < conf->nr_strip_zones; i++)
 203         {
 204                 int j;
 205
 206                 zone = conf->strip_zone + i;
 207                 dev = conf->devlist + i * mddev->raid_disks;
 208
 209                 printk(KERN_INFO "raid0: zone %d\n", i);
 210                 zone->dev_start = smallest->sectors;
 211                 smallest = NULL;
 212                 c = 0;
 213
 214                 for (j=0; j<cnt; j++) {
 215                         rdev = conf->devlist[j];
 216                         printk(KERN_INFO "raid0: checking %s ...",
 217                                 bdevname(rdev->bdev, b));
 218                         if (rdev->sectors <= zone->dev_start) {
 219                                 printk(KERN_INFO " nope.\n");
 220                                 continue;
 221                         }
 222                         printk(KERN_INFO " contained as device %d\n", c);
 223                         dev[c] = rdev;
 224                         c++;
 225                         if (!smallest || rdev->sectors < smallest->sectors) {
 226                                 smallest = rdev;
 227                                 printk(KERN_INFO "  (%llu) is smallest!.\n",
 228                                         (unsigned long long)rdev->sectors);
 229                         }
 230                 }
 231
 232                 zone->nb_dev = c;
 233                 sectors = (smallest->sectors - zone->dev_start) * c;
 234                 printk(KERN_INFO "raid0: zone->nb_dev: %d, sectors: %llu\n",
 235                         zone->nb_dev, (unsigned long long)sectors);
 236
 237                 curr_zone_end += sectors;
 238                 zone->zone_end = curr_zone_end;
 239
 240                 printk(KERN_INFO "raid0: current zone start: %llu\n",
 241                         (unsigned long long)smallest->sectors);
 242         }
 243         mddev->queue->unplug_fn = raid0_unplug;
 244         mddev->queue->backing_dev_info.congested_fn = raid0_congested;
 245         mddev->queue->backing_dev_info.congested_data = mddev;
 246
 247         /*
 248          * now since we have the hard sector sizes, we can make sure
 249          * chunk size is a multiple of that sector size
 250          */
 251         if ((mddev->chunk_sectors << 9) % queue_logical_block_size(mddev->queue)) {
 252                 printk(KERN_ERR "%s chunk_size of %d not valid\n",
 253                        mdname(mddev),
 254                        mddev->chunk_sectors << 9);
 255                 goto abort;
 256         }
 257
 258         blk_queue_io_min(mddev->queue, mddev->chunk_sectors << 9);
 259         blk_queue_io_opt(mddev->queue,
 260                          (mddev->chunk_sectors << 9) * mddev->raid_disks);
 261
 262         printk(KERN_INFO "raid0: done.\n");
 263         mddev->private = conf;
 264         return 0;
 265 abort:
 266         kfree(conf->strip_zone);
 267         kfree(conf->devlist);
 268         kfree(conf);
 269         mddev->private = NULL;
 270         return err;
 271 }
 272
 273 /**
 274  *      raid0_mergeable_bvec -- tell bio layer if a two requests can be merged
 275  *      @q: request queue
 276  *      @bvm: properties of new bio
 277  *      @biovec: the request that could be merged to it.
 278  *
 279  *      Return amount of bytes we can accept at this offset
 280  */
 281 static int raid0_mergeable_bvec(struct request_queue *q,
 282                                 struct bvec_merge_data *bvm,
 283                                 struct bio_vec *biovec)
 284 {
 285         mddev_t *mddev = q->queuedata;
 286         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 287         int max;
 288         unsigned int chunk_sectors = mddev->chunk_sectors;
 289         unsigned int bio_sectors = bvm->bi_size >> 9;
 290
 291         if (is_power_of_2(chunk_sectors))
 292                 max =  (chunk_sectors - ((sector & (chunk_sectors-1))
 293                                                 + bio_sectors)) << 9;
 294         else
 295                 max =  (chunk_sectors - (sector_div(sector, chunk_sectors)
 296                                                 + bio_sectors)) << 9;
 297         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
 298         if (max <= biovec->bv_len && bio_sectors == 0)
 299                 return biovec->bv_len;
 300         else
 301                 return max;
 302 }
 303
 304 static sector_t raid0_size(mddev_t *mddev, sector_t sectors, int raid_disks)
 305 {
 306         sector_t array_sectors = 0;
 307         mdk_rdev_t *rdev;
 308
 309         WARN_ONCE(sectors || raid_disks,
 310                   "%s does not support generic reshape\n", __func__);
 311
 312         list_for_each_entry(rdev, &mddev->disks, same_set)
 313                 array_sectors += rdev->sectors;
 314
 315         return array_sectors;
 316 }
 317
 318 static int raid0_run(mddev_t *mddev)
 319 {
 320         int ret;
 321
 322         if (mddev->chunk_sectors == 0) {
 323                 printk(KERN_ERR "md/raid0: chunk size must be set.\n");
 324                 return -EINVAL;
 325         }
 326         if (md_check_no_bitmap(mddev))
 327                 return -EINVAL;
 328         blk_queue_max_sectors(mddev->queue, mddev->chunk_sectors);
 329         mddev->queue->queue_lock = &mddev->queue->__queue_lock;
 330
 331         ret = create_strip_zones(mddev);
 332         if (ret < 0)
 333                 return ret;
 334
 335         /* calculate array device size */
 336         md_set_array_sectors(mddev, raid0_size(mddev, 0, 0));
 337
 338         printk(KERN_INFO "raid0 : md_size is %llu sectors.\n",
 339                 (unsigned long long)mddev->array_sectors);
 340         /* calculate the max read-ahead size.
 341          * For read-ahead of large files to be effective, we need to
 342          * readahead at least twice a whole stripe. i.e. number of devices
 343          * multiplied by chunk size times 2.
 344          * If an individual device has an ra_pages greater than the
 345          * chunk size, then we will not drive that device as hard as it
 346          * wants.  We consider this a configuration error: a larger
 347          * chunksize should be used in that case.
 348          */
 349         {
 350                 int stripe = mddev->raid_disks *
 351                         (mddev->chunk_sectors << 9) / PAGE_SIZE;
 352                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
 353                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
 354         }
 355
 356         blk_queue_merge_bvec(mddev->queue, raid0_mergeable_bvec);
 357         dump_zones(mddev);
 358         md_integrity_register(mddev);
 359         return 0;
 360 }
 361
 362 static int raid0_stop(mddev_t *mddev)
 363 {
 364         raid0_conf_t *conf = mddev->private;
 365
 366         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
 367         kfree(conf->strip_zone);
 368         kfree(conf->devlist);
 369         kfree(conf);
 370         mddev->private = NULL;
 371         return 0;
 372 }
 373
 374 /* Find the zone which holds a particular offset
 375  * Update *sectorp to be an offset in that zone
 376  */
 377 static struct strip_zone *find_zone(struct raid0_private_data *conf,
 378                                     sector_t *sectorp)
 379 {
 380         int i;
 381         struct strip_zone *z = conf->strip_zone;
 382         sector_t sector = *sectorp;
 383
 384         for (i = 0; i < conf->nr_strip_zones; i++)
 385                 if (sector < z[i].zone_end) {
 386                         if (i)
 387                                 *sectorp = sector - z[i-1].zone_end;
 388                         return z + i;
 389                 }
 390         BUG();
 391 }
 392
 393 /*
 394  * remaps the bio to the target device. we separate two flows.
 395  * power 2 flow and a general flow for the sake of perfromance
 396 */
 397 static mdk_rdev_t *map_sector(mddev_t *mddev, struct strip_zone *zone,
 398                                 sector_t sector, sector_t *sector_offset)
 399 {
 400         unsigned int sect_in_chunk;
 401         sector_t chunk;
 402         raid0_conf_t *conf = mddev->private;
 403         unsigned int chunk_sects = mddev->chunk_sectors;
 404
 405         if (is_power_of_2(chunk_sects)) {
 406                 int chunksect_bits = ffz(~chunk_sects);
 407                 /* find the sector offset inside the chunk */
 408                 sect_in_chunk  = sector & (chunk_sects - 1);
 409                 sector >>= chunksect_bits;
 410                 /* chunk in zone */
 411                 chunk = *sector_offset;
 412                 /* quotient is the chunk in real device*/
 413                 sector_div(chunk, zone->nb_dev << chunksect_bits);
 414         } else{
 415                 sect_in_chunk = sector_div(sector, chunk_sects);
 416                 chunk = *sector_offset;
 417                 sector_div(chunk, chunk_sects * zone->nb_dev);
 418         }
 419         /*
 420         *  position the bio over the real device
 421         *  real sector = chunk in device + starting of zone
 422         *       + the position in the chunk
 423         */
 424         *sector_offset = (chunk * chunk_sects) + sect_in_chunk;
 425         return conf->devlist[(zone - conf->strip_zone)*mddev->raid_disks
 426                              + sector_div(sector, zone->nb_dev)];
 427 }
 428
 429 /*
 430  * Is io distribute over 1 or more chunks ?
 431 */
 432 static inline int is_io_in_chunk_boundary(mddev_t *mddev,
 433                         unsigned int chunk_sects, struct bio *bio)
 434 {
 435         if (likely(is_power_of_2(chunk_sects))) {
 436                 return chunk_sects >= ((bio->bi_sector & (chunk_sects-1))
 437                                         + (bio->bi_size >> 9));
 438         } else{
 439                 sector_t sector = bio->bi_sector;
 440                 return chunk_sects >= (sector_div(sector, chunk_sects)
 441                                                 + (bio->bi_size >> 9));
 442         }
 443 }
 444
 445 static int raid0_make_request(struct request_queue *q, struct bio *bio)
 446 {
 447         mddev_t *mddev = q->queuedata;
 448         unsigned int chunk_sects;
 449         sector_t sector_offset;
 450         struct strip_zone *zone;
 451         mdk_rdev_t *tmp_dev;
 452         const int rw = bio_data_dir(bio);
 453         int cpu;
 454
 455         if (unlikely(bio_rw_flagged(bio, BIO_RW_BARRIER))) {
 456                 md_barrier_request(mddev, bio);
 457                 return 0;
 458         }
 459
 460         cpu = part_stat_lock();
 461         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
 462         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
 463                       bio_sectors(bio));
 464         part_stat_unlock();
 465
 466         chunk_sects = mddev->chunk_sectors;
 467         if (unlikely(!is_io_in_chunk_boundary(mddev, chunk_sects, bio))) {
 468                 sector_t sector = bio->bi_sector;
 469                 struct bio_pair *bp;
 470                 /* Sanity check -- queue functions should prevent this happening */
 471                 if (bio->bi_vcnt != 1 ||
 472                     bio->bi_idx != 0)
 473                         goto bad_map;
 474                 /* This is a one page bio that upper layers
 475                  * refuse to split for us, so we need to split it.
 476                  */
 477                 if (likely(is_power_of_2(chunk_sects)))
 478                         bp = bio_split(bio, chunk_sects - (sector &
 479                                                            (chunk_sects-1)));
 480                 else
 481                         bp = bio_split(bio, chunk_sects -
 482                                        sector_div(sector, chunk_sects));
 483                 if (raid0_make_request(q, &bp->bio1))
 484                         generic_make_request(&bp->bio1);
 485                 if (raid0_make_request(q, &bp->bio2))
 486                         generic_make_request(&bp->bio2);
 487
 488                 bio_pair_release(bp);
 489                 return 0;
 490         }
 491
 492         sector_offset = bio->bi_sector;
 493         zone =  find_zone(mddev->private, &sector_offset);
 494         tmp_dev = map_sector(mddev, zone, bio->bi_sector,
 495                              &sector_offset);
 496         bio->bi_bdev = tmp_dev->bdev;
 497         bio->bi_sector = sector_offset + zone->dev_start +
 498                 tmp_dev->data_offset;
 499         /*
 500          * Let the main block layer submit the IO and resolve recursion:
 501          */
 502         return 1;
 503
 504 bad_map:
 505         printk("raid0_make_request bug: can't convert block across chunks"
 506                 " or bigger than %dk %llu %d\n", chunk_sects / 2,
 507                 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 508
 509         bio_io_error(bio);
 510         return 0;
 511 }
 512
 513 static void raid0_status(struct seq_file *seq, mddev_t *mddev)
 514 {
 515 #undef MD_DEBUG
 516 #ifdef MD_DEBUG
 517         int j, k, h;
 518         char b[BDEVNAME_SIZE];
 519         raid0_conf_t *conf = mddev->private;
 520
 521         sector_t zone_size;
 522         sector_t zone_start = 0;
 523         h = 0;
 524
 525         for (j = 0; j < conf->nr_strip_zones; j++) {
 526                 seq_printf(seq, "      z%d", j);
 527                 seq_printf(seq, "=[");
 528                 for (k = 0; k < conf->strip_zone[j].nb_dev; k++)
 529                         seq_printf(seq, "%s/", bdevname(
 530                                 conf->devlist[j*mddev->raid_disks + k]
 531                                                 ->bdev, b));
 532
 533                 zone_size  = conf->strip_zone[j].zone_end - zone_start;
 534                 seq_printf(seq, "] ze=%lld ds=%lld s=%lld\n",
 535                         (unsigned long long)zone_start>>1,
 536                         (unsigned long long)conf->strip_zone[j].dev_start>>1,
 537                         (unsigned long long)zone_size>>1);
 538                 zone_start = conf->strip_zone[j].zone_end;
 539         }
 540 #endif
 541         seq_printf(seq, " %dk chunks", mddev->chunk_sectors / 2);
 542         return;
 543 }
 544
 545 static struct mdk_personality raid0_personality=
 546 {
 547         .name           = "raid0",
 548         .level          = 0,
 549         .owner          = THIS_MODULE,
 550         .make_request   = raid0_make_request,
 551         .run            = raid0_run,
 552         .stop           = raid0_stop,
 553         .status         = raid0_status,
 554         .size           = raid0_size,
 555 };
 556
 557 static int __init raid0_init (void)
 558 {
 559         return register_md_personality (&raid0_personality);
 560 }
 561
 562 static void raid0_exit (void)
 563 {
 564         unregister_md_personality (&raid0_personality);
 565 }
 566
 567 module_init(raid0_init);
 568 module_exit(raid0_exit);
 569 MODULE_LICENSE("GPL");
 570 MODULE_DESCRIPTION("RAID0 (striping) personality for MD");
 571 MODULE_ALIAS("md-personality-2"); /* RAID0 */
 572 MODULE_ALIAS("md-raid0");
 573 MODULE_ALIAS("md-level-0");