drivers/md/dm-log.c

   1 /*
   2  * Copyright (C) 2003 Sistina Software
   3  *
   4  * This file is released under the LGPL.
   5  */
   6
   7 #include <linux/init.h>
   8 #include <linux/slab.h>
   9 #include <linux/module.h>
  10 #include <linux/vmalloc.h>
  11
  12 #include "dm-log.h"
  13 #include "dm-io.h"
  14
  15 static LIST_HEAD(_log_types);
  16 static spinlock_t _lock = SPIN_LOCK_UNLOCKED;
  17
  18 int dm_register_dirty_log_type(struct dirty_log_type *type)
  19 {
  20         if (!try_module_get(type->module))
  21                 return -EINVAL;
  22
  23         spin_lock(&_lock);
  24         type->use_count = 0;
  25         list_add(&type->list, &_log_types);
  26         spin_unlock(&_lock);
  27
  28         return 0;
  29 }
  30
  31 int dm_unregister_dirty_log_type(struct dirty_log_type *type)
  32 {
  33         spin_lock(&_lock);
  34
  35         if (type->use_count)
  36                 DMWARN("Attempt to unregister a log type that is still in use");
  37         else {
  38                 list_del(&type->list);
  39                 module_put(type->module);
  40         }
  41
  42         spin_unlock(&_lock);
  43
  44         return 0;
  45 }
  46
  47 static struct dirty_log_type *get_type(const char *type_name)
  48 {
  49         struct dirty_log_type *type;
  50
  51         spin_lock(&_lock);
  52         list_for_each_entry (type, &_log_types, list)
  53                 if (!strcmp(type_name, type->name)) {
  54                         type->use_count++;
  55                         spin_unlock(&_lock);
  56                         return type;
  57                 }
  58
  59         spin_unlock(&_lock);
  60         return NULL;
  61 }
  62
  63 static void put_type(struct dirty_log_type *type)
  64 {
  65         spin_lock(&_lock);
  66         type->use_count--;
  67         spin_unlock(&_lock);
  68 }
  69
  70 struct dirty_log *dm_create_dirty_log(const char *type_name, struct dm_target *ti,
  71                                       unsigned int argc, char **argv)
  72 {
  73         struct dirty_log_type *type;
  74         struct dirty_log *log;
  75
  76         log = kmalloc(sizeof(*log), GFP_KERNEL);
  77         if (!log)
  78                 return NULL;
  79
  80         type = get_type(type_name);
  81         if (!type) {
  82                 kfree(log);
  83                 return NULL;
  84         }
  85
  86         log->type = type;
  87         if (type->ctr(log, ti, argc, argv)) {
  88                 kfree(log);
  89                 put_type(type);
  90                 return NULL;
  91         }
  92
  93         return log;
  94 }
  95
  96 void dm_destroy_dirty_log(struct dirty_log *log)
  97 {
  98         log->type->dtr(log);
  99         put_type(log->type);
 100         kfree(log);
 101 }
 102
 103 /*-----------------------------------------------------------------
 104  * Persistent and core logs share a lot of their implementation.
 105  * FIXME: need a reload method to be called from a resume
 106  *---------------------------------------------------------------*/
 107 /*
 108  * Magic for persistent mirrors: "MiRr"
 109  */
 110 #define MIRROR_MAGIC 0x4D695272
 111
 112 /*
 113  * The on-disk version of the metadata.
 114  */
 115 #define MIRROR_DISK_VERSION 1
 116 #define LOG_OFFSET 2
 117
 118 struct log_header {
 119         uint32_t magic;
 120
 121         /*
 122          * Simple, incrementing version. no backward
 123          * compatibility.
 124          */
 125         uint32_t version;
 126         sector_t nr_regions;
 127 };
 128
 129 struct log_c {
 130         struct dm_target *ti;
 131         int touched;
 132         sector_t region_size;
 133         unsigned int region_count;
 134         region_t sync_count;
 135
 136         unsigned bitset_uint32_count;
 137         uint32_t *clean_bits;
 138         uint32_t *sync_bits;
 139         uint32_t *recovering_bits;      /* FIXME: this seems excessive */
 140
 141         int sync_search;
 142
 143         /*
 144          * Disk log fields
 145          */
 146         struct dm_dev *log_dev;
 147         struct log_header header;
 148
 149         struct io_region header_location;
 150         struct log_header *disk_header;
 151
 152         struct io_region bits_location;
 153         uint32_t *disk_bits;
 154 };
 155
 156 /*
 157  * The touched member needs to be updated every time we access
 158  * one of the bitsets.
 159  */
 160 static  inline int log_test_bit(uint32_t *bs, unsigned bit)
 161 {
 162         return test_bit(bit, (unsigned long *) bs) ? 1 : 0;
 163 }
 164
 165 static inline void log_set_bit(struct log_c *l,
 166                                uint32_t *bs, unsigned bit)
 167 {
 168         set_bit(bit, (unsigned long *) bs);
 169         l->touched = 1;
 170 }
 171
 172 static inline void log_clear_bit(struct log_c *l,
 173                                  uint32_t *bs, unsigned bit)
 174 {
 175         clear_bit(bit, (unsigned long *) bs);
 176         l->touched = 1;
 177 }
 178
 179 /*----------------------------------------------------------------
 180  * Header IO
 181  *--------------------------------------------------------------*/
 182 static void header_to_disk(struct log_header *core, struct log_header *disk)
 183 {
 184         disk->magic = cpu_to_le32(core->magic);
 185         disk->version = cpu_to_le32(core->version);
 186         disk->nr_regions = cpu_to_le64(core->nr_regions);
 187 }
 188
 189 static void header_from_disk(struct log_header *core, struct log_header *disk)
 190 {
 191         core->magic = le32_to_cpu(disk->magic);
 192         core->version = le32_to_cpu(disk->version);
 193         core->nr_regions = le64_to_cpu(disk->nr_regions);
 194 }
 195
 196 static int read_header(struct log_c *log)
 197 {
 198         int r;
 199         unsigned long ebits;
 200
 201         r = dm_io_sync_vm(1, &log->header_location, READ,
 202                           log->disk_header, &ebits);
 203         if (r)
 204                 return r;
 205
 206         header_from_disk(&log->header, log->disk_header);
 207
 208         if (log->header.magic != MIRROR_MAGIC) {
 209                 log->header.magic = MIRROR_MAGIC;
 210                 log->header.version = MIRROR_DISK_VERSION;
 211                 log->header.nr_regions = 0;
 212         }
 213
 214         if (log->header.version != MIRROR_DISK_VERSION) {
 215                 DMWARN("incompatible disk log version");
 216                 return -EINVAL;
 217         }
 218
 219         return 0;
 220 }
 221
 222 static inline int write_header(struct log_c *log)
 223 {
 224         unsigned long ebits;
 225
 226         header_to_disk(&log->header, log->disk_header);
 227         return dm_io_sync_vm(1, &log->header_location, WRITE,
 228                              log->disk_header, &ebits);
 229 }
 230
 231 /*----------------------------------------------------------------
 232  * Bits IO
 233  *--------------------------------------------------------------*/
 234 static inline void bits_to_core(uint32_t *core, uint32_t *disk, unsigned count)
 235 {
 236         unsigned i;
 237
 238         for (i = 0; i < count; i++)
 239                 core[i] = le32_to_cpu(disk[i]);
 240 }
 241
 242 static inline void bits_to_disk(uint32_t *core, uint32_t *disk, unsigned count)
 243 {
 244         unsigned i;
 245
 246         /* copy across the clean/dirty bitset */
 247         for (i = 0; i < count; i++)
 248                 disk[i] = cpu_to_le32(core[i]);
 249 }
 250
 251 static int read_bits(struct log_c *log)
 252 {
 253         int r;
 254         unsigned long ebits;
 255
 256         r = dm_io_sync_vm(1, &log->bits_location, READ,
 257                           log->disk_bits, &ebits);
 258         if (r)
 259                 return r;
 260
 261         bits_to_core(log->clean_bits, log->disk_bits,
 262                      log->bitset_uint32_count);
 263         return 0;
 264 }
 265
 266 static int write_bits(struct log_c *log)
 267 {
 268         unsigned long ebits;
 269         bits_to_disk(log->clean_bits, log->disk_bits,
 270                      log->bitset_uint32_count);
 271         return dm_io_sync_vm(1, &log->bits_location, WRITE,
 272                              log->disk_bits, &ebits);
 273 }
 274
 275 /*----------------------------------------------------------------
 276  * constructor/destructor
 277  *--------------------------------------------------------------*/
 278 #define BYTE_SHIFT 3
 279 static int core_ctr(struct dirty_log *log, struct dm_target *ti,
 280                     unsigned int argc, char **argv)
 281 {
 282         struct log_c *lc;
 283         sector_t region_size;
 284         unsigned int region_count;
 285         size_t bitset_size;
 286
 287         if (argc != 1) {
 288                 DMWARN("wrong number of arguments to log_c");
 289                 return -EINVAL;
 290         }
 291
 292         if (sscanf(argv[0], SECTOR_FORMAT, &region_size) != 1) {
 293                 DMWARN("invalid region size string");
 294                 return -EINVAL;
 295         }
 296
 297         region_count = dm_div_up(ti->len, region_size);
 298
 299         lc = kmalloc(sizeof(*lc), GFP_KERNEL);
 300         if (!lc) {
 301                 DMWARN("couldn't allocate core log");
 302                 return -ENOMEM;
 303         }
 304
 305         lc->ti = ti;
 306         lc->touched = 0;
 307         lc->region_size = region_size;
 308         lc->region_count = region_count;
 309
 310         /*
 311          * Work out how many words we need to hold the bitset.
 312          */
 313         bitset_size = dm_round_up(region_count,
 314                                   sizeof(*lc->clean_bits) << BYTE_SHIFT);
 315         bitset_size >>= BYTE_SHIFT;
 316
 317         lc->bitset_uint32_count = bitset_size / 4;
 318         lc->clean_bits = vmalloc(bitset_size);
 319         if (!lc->clean_bits) {
 320                 DMWARN("couldn't allocate clean bitset");
 321                 kfree(lc);
 322                 return -ENOMEM;
 323         }
 324         memset(lc->clean_bits, -1, bitset_size);
 325
 326         lc->sync_bits = vmalloc(bitset_size);
 327         if (!lc->sync_bits) {
 328                 DMWARN("couldn't allocate sync bitset");
 329                 vfree(lc->clean_bits);
 330                 kfree(lc);
 331                 return -ENOMEM;
 332         }
 333         memset(lc->sync_bits, 0, bitset_size);
 334         lc->sync_count = 0;
 335
 336         lc->recovering_bits = vmalloc(bitset_size);
 337         if (!lc->recovering_bits) {
 338                 DMWARN("couldn't allocate sync bitset");
 339                 vfree(lc->sync_bits);
 340                 vfree(lc->clean_bits);
 341                 kfree(lc);
 342                 return -ENOMEM;
 343         }
 344         memset(lc->recovering_bits, 0, bitset_size);
 345         lc->sync_search = 0;
 346         log->context = lc;
 347         return 0;
 348 }
 349
 350 static void core_dtr(struct dirty_log *log)
 351 {
 352         struct log_c *lc = (struct log_c *) log->context;
 353         vfree(lc->clean_bits);
 354         vfree(lc->sync_bits);
 355         vfree(lc->recovering_bits);
 356         kfree(lc);
 357 }
 358
 359 static int disk_ctr(struct dirty_log *log, struct dm_target *ti,
 360                     unsigned int argc, char **argv)
 361 {
 362         int r;
 363         size_t size;
 364         struct log_c *lc;
 365         struct dm_dev *dev;
 366
 367         if (argc != 2) {
 368                 DMWARN("wrong number of arguments to log_d");
 369                 return -EINVAL;
 370         }
 371
 372         r = dm_get_device(ti, argv[0], 0, 0 /* FIXME */,
 373                           FMODE_READ | FMODE_WRITE, &dev);
 374         if (r)
 375                 return r;
 376
 377         r = core_ctr(log, ti, argc - 1, argv + 1);
 378         if (r) {
 379                 dm_put_device(ti, dev);
 380                 return r;
 381         }
 382
 383         lc = (struct log_c *) log->context;
 384         lc->log_dev = dev;
 385
 386         /* setup the disk header fields */
 387         lc->header_location.bdev = lc->log_dev->bdev;
 388         lc->header_location.sector = 0;
 389         lc->header_location.count = 1;
 390
 391         /*
 392          * We can't read less than this amount, even though we'll
 393          * not be using most of this space.
 394          */
 395         lc->disk_header = vmalloc(1 << SECTOR_SHIFT);
 396         if (!lc->disk_header)
 397                 goto bad;
 398
 399         /* setup the disk bitset fields */
 400         lc->bits_location.bdev = lc->log_dev->bdev;
 401         lc->bits_location.sector = LOG_OFFSET;
 402
 403         size = dm_round_up(lc->bitset_uint32_count * sizeof(uint32_t),
 404                            1 << SECTOR_SHIFT);
 405         lc->bits_location.count = size >> SECTOR_SHIFT;
 406         lc->disk_bits = vmalloc(size);
 407         if (!lc->disk_bits) {
 408                 vfree(lc->disk_header);
 409                 goto bad;
 410         }
 411         return 0;
 412
 413  bad:
 414         dm_put_device(ti, lc->log_dev);
 415         core_dtr(log);
 416         return -ENOMEM;
 417 }
 418
 419 static void disk_dtr(struct dirty_log *log)
 420 {
 421         struct log_c *lc = (struct log_c *) log->context;
 422         dm_put_device(lc->ti, lc->log_dev);
 423         vfree(lc->disk_header);
 424         vfree(lc->disk_bits);
 425         core_dtr(log);
 426 }
 427
 428 static int count_bits32(uint32_t *addr, unsigned size)
 429 {
 430         int count = 0, i;
 431
 432         for (i = 0; i < size; i++) {
 433                 count += hweight32(*(addr+i));
 434         }
 435         return count;
 436 }
 437
 438 static int disk_resume(struct dirty_log *log)
 439 {
 440         int r;
 441         unsigned i;
 442         struct log_c *lc = (struct log_c *) log->context;
 443         size_t size = lc->bitset_uint32_count * sizeof(uint32_t);
 444
 445         /* read the disk header */
 446         r = read_header(lc);
 447         if (r)
 448                 return r;
 449
 450         /* read the bits */
 451         r = read_bits(lc);
 452         if (r)
 453                 return r;
 454
 455         /* zero any new bits if the mirror has grown */
 456         for (i = lc->header.nr_regions; i < lc->region_count; i++)
 457                 /* FIXME: amazingly inefficient */
 458                 log_clear_bit(lc, lc->clean_bits, i);
 459
 460         /* copy clean across to sync */
 461         memcpy(lc->sync_bits, lc->clean_bits, size);
 462         lc->sync_count = count_bits32(lc->clean_bits, lc->bitset_uint32_count);
 463
 464         /* write the bits */
 465         r = write_bits(lc);
 466         if (r)
 467                 return r;
 468
 469         /* set the correct number of regions in the header */
 470         lc->header.nr_regions = lc->region_count;
 471
 472         /* write the new header */
 473         return write_header(lc);
 474 }
 475
 476 static sector_t core_get_region_size(struct dirty_log *log)
 477 {
 478         struct log_c *lc = (struct log_c *) log->context;
 479         return lc->region_size;
 480 }
 481
 482 static int core_is_clean(struct dirty_log *log, region_t region)
 483 {
 484         struct log_c *lc = (struct log_c *) log->context;
 485         return log_test_bit(lc->clean_bits, region);
 486 }
 487
 488 static int core_in_sync(struct dirty_log *log, region_t region, int block)
 489 {
 490         struct log_c *lc = (struct log_c *) log->context;
 491         return log_test_bit(lc->sync_bits, region);
 492 }
 493
 494 static int core_flush(struct dirty_log *log)
 495 {
 496         /* no op */
 497         return 0;
 498 }
 499
 500 static int disk_flush(struct dirty_log *log)
 501 {
 502         int r;
 503         struct log_c *lc = (struct log_c *) log->context;
 504
 505         /* only write if the log has changed */
 506         if (!lc->touched)
 507                 return 0;
 508
 509         r = write_bits(lc);
 510         if (!r)
 511                 lc->touched = 0;
 512
 513         return r;
 514 }
 515
 516 static void core_mark_region(struct dirty_log *log, region_t region)
 517 {
 518         struct log_c *lc = (struct log_c *) log->context;
 519         log_clear_bit(lc, lc->clean_bits, region);
 520 }
 521
 522 static void core_clear_region(struct dirty_log *log, region_t region)
 523 {
 524         struct log_c *lc = (struct log_c *) log->context;
 525         log_set_bit(lc, lc->clean_bits, region);
 526 }
 527
 528 static int core_get_resync_work(struct dirty_log *log, region_t *region)
 529 {
 530         struct log_c *lc = (struct log_c *) log->context;
 531
 532         if (lc->sync_search >= lc->region_count)
 533                 return 0;
 534
 535         do {
 536                 *region = find_next_zero_bit((unsigned long *) lc->sync_bits,
 537                                              lc->region_count,
 538                                              lc->sync_search);
 539                 lc->sync_search = *region + 1;
 540
 541                 if (*region == lc->region_count)
 542                         return 0;
 543
 544         } while (log_test_bit(lc->recovering_bits, *region));
 545
 546         log_set_bit(lc, lc->recovering_bits, *region);
 547         return 1;
 548 }
 549
 550 static void core_complete_resync_work(struct dirty_log *log, region_t region,
 551                                       int success)
 552 {
 553         struct log_c *lc = (struct log_c *) log->context;
 554
 555         log_clear_bit(lc, lc->recovering_bits, region);
 556         if (success) {
 557                 log_set_bit(lc, lc->sync_bits, region);
 558                 lc->sync_count++;
 559         }
 560 }
 561
 562 static region_t core_get_sync_count(struct dirty_log *log)
 563 {
 564         struct log_c *lc = (struct log_c *) log->context;
 565
 566         return lc->sync_count;
 567 }
 568
 569 static struct dirty_log_type _core_type = {
 570         .name = "core",
 571         .module = THIS_MODULE,
 572         .ctr = core_ctr,
 573         .dtr = core_dtr,
 574         .get_region_size = core_get_region_size,
 575         .is_clean = core_is_clean,
 576         .in_sync = core_in_sync,
 577         .flush = core_flush,
 578         .mark_region = core_mark_region,
 579         .clear_region = core_clear_region,
 580         .get_resync_work = core_get_resync_work,
 581         .complete_resync_work = core_complete_resync_work,
 582         .get_sync_count = core_get_sync_count
 583 };
 584
 585 static struct dirty_log_type _disk_type = {
 586         .name = "disk",
 587         .module = THIS_MODULE,
 588         .ctr = disk_ctr,
 589         .dtr = disk_dtr,
 590         .suspend = disk_flush,
 591         .resume = disk_resume,
 592         .get_region_size = core_get_region_size,
 593         .is_clean = core_is_clean,
 594         .in_sync = core_in_sync,
 595         .flush = disk_flush,
 596         .mark_region = core_mark_region,
 597         .clear_region = core_clear_region,
 598         .get_resync_work = core_get_resync_work,
 599         .complete_resync_work = core_complete_resync_work,
 600         .get_sync_count = core_get_sync_count
 601 };
 602
 603 int __init dm_dirty_log_init(void)
 604 {
 605         int r;
 606
 607         r = dm_register_dirty_log_type(&_core_type);
 608         if (r)
 609                 DMWARN("couldn't register core log");
 610
 611         r = dm_register_dirty_log_type(&_disk_type);
 612         if (r) {
 613                 DMWARN("couldn't register disk type");
 614                 dm_unregister_dirty_log_type(&_core_type);
 615         }
 616
 617         return r;
 618 }
 619
 620 void dm_dirty_log_exit(void)
 621 {
 622         dm_unregister_dirty_log_type(&_disk_type);
 623         dm_unregister_dirty_log_type(&_core_type);
 624 }
 625
 626 EXPORT_SYMBOL(dm_register_dirty_log_type);
 627 EXPORT_SYMBOL(dm_unregister_dirty_log_type);
 628 EXPORT_SYMBOL(dm_create_dirty_log);
 629 EXPORT_SYMBOL(dm_destroy_dirty_log);