extent_io.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #define _XOPEN_SOURCE 600
  19 #define __USE_XOPEN2K
  20 #include <stdio.h>
  21 #include <stdlib.h>
  22 #include <sys/types.h>
  23 #include <sys/stat.h>
  24 #include <fcntl.h>
  25 #include <unistd.h>
  26 #include "kerncompat.h"
  27 #include "extent_io.h"
  28 #include "list.h"
  29
  30 u64 cache_max = 1024 * 1024 * 32;
  31
  32 void extent_io_tree_init(struct extent_io_tree *tree)
  33 {
  34         cache_tree_init(&tree->state);
  35         cache_tree_init(&tree->cache);
  36         INIT_LIST_HEAD(&tree->lru);
  37         tree->cache_size = 0;
  38 }
  39
  40 static struct extent_state *alloc_extent_state(void)
  41 {
  42         struct extent_state *state;
  43
  44         state = malloc(sizeof(*state));
  45         if (!state)
  46                 return NULL;
  47         state->refs = 1;
  48         state->state = 0;
  49         state->private = 0;
  50         return state;
  51 }
  52
  53 static void free_extent_state(struct extent_state *state)
  54 {
  55         state->refs--;
  56         BUG_ON(state->refs < 0);
  57         if (state->refs == 0)
  58                 free(state);
  59 }
  60
  61 void extent_io_tree_cleanup(struct extent_io_tree *tree)
  62 {
  63         struct extent_state *es;
  64         struct extent_buffer *eb;
  65         struct cache_extent *cache;
  66
  67         while(!list_empty(&tree->lru)) {
  68                 eb = list_entry(tree->lru.next, struct extent_buffer, lru);
  69                 if (eb->refs != 1) {
  70                         fprintf(stderr, "extent buffer leak: "
  71                                 "start %llu len %u\n",
  72                                 (unsigned long long)eb->start, eb->len);
  73                         eb->refs = 1;
  74                 }
  75                 free_extent_buffer(eb);
  76         }
  77         while (1) {
  78                 cache = find_first_cache_extent(&tree->state, 0);
  79                 if (!cache)
  80                         break;
  81                 es = container_of(cache, struct extent_state, cache_node);
  82                 remove_cache_extent(&tree->state, &es->cache_node);
  83                 free_extent_state(es);
  84         }
  85 }
  86
  87 static inline void update_extent_state(struct extent_state *state)
  88 {
  89         state->cache_node.start = state->start;
  90         state->cache_node.size = state->end + 1 - state->start;
  91 }
  92
  93 /*
  94  * Utility function to look for merge candidates inside a given range.
  95  * Any extents with matching state are merged together into a single
  96  * extent in the tree. Extents with EXTENT_IO in their state field are
  97  * not merged
  98  */
  99 static int merge_state(struct extent_io_tree *tree,
 100                        struct extent_state *state)
 101 {
 102         struct extent_state *other;
 103         struct cache_extent *other_node;
 104
 105         if (state->state & EXTENT_IOBITS)
 106                 return 0;
 107
 108         other_node = prev_cache_extent(&state->cache_node);
 109         if (other_node) {
 110                 other = container_of(other_node, struct extent_state,
 111                                      cache_node);
 112                 if (other->end == state->start - 1 &&
 113                     other->state == state->state) {
 114                         state->start = other->start;
 115                         update_extent_state(state);
 116                         remove_cache_extent(&tree->state, &other->cache_node);
 117                         free_extent_state(other);
 118                 }
 119         }
 120         other_node = next_cache_extent(&state->cache_node);
 121         if (other_node) {
 122                 other = container_of(other_node, struct extent_state,
 123                                      cache_node);
 124                 if (other->start == state->end + 1 &&
 125                     other->state == state->state) {
 126                         other->start = state->start;
 127                         update_extent_state(other);
 128                         remove_cache_extent(&tree->state, &state->cache_node);
 129                         free_extent_state(state);
 130                 }
 131         }
 132         return 0;
 133 }
 134
 135 /*
 136  * insert an extent_state struct into the tree.  'bits' are set on the
 137  * struct before it is inserted.
 138  */
 139 static int insert_state(struct extent_io_tree *tree,
 140                         struct extent_state *state, u64 start, u64 end,
 141                         int bits)
 142 {
 143         int ret;
 144
 145         BUG_ON(end < start);
 146         state->state |= bits;
 147         state->start = start;
 148         state->end = end;
 149         update_extent_state(state);
 150         ret = insert_existing_cache_extent(&tree->state, &state->cache_node);
 151         BUG_ON(ret);
 152         merge_state(tree, state);
 153         return 0;
 154 }
 155
 156 /*
 157  * split a given extent state struct in two, inserting the preallocated
 158  * struct 'prealloc' as the newly created second half.  'split' indicates an
 159  * offset inside 'orig' where it should be split.
 160  */
 161 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 162                        struct extent_state *prealloc, u64 split)
 163 {
 164         int ret;
 165         prealloc->start = orig->start;
 166         prealloc->end = split - 1;
 167         prealloc->state = orig->state;
 168         update_extent_state(prealloc);
 169         orig->start = split;
 170         update_extent_state(orig);
 171         ret = insert_existing_cache_extent(&tree->state,
 172                                            &prealloc->cache_node);
 173         BUG_ON(ret);
 174         return 0;
 175 }
 176
 177 /*
 178  * clear some bits on a range in the tree.
 179  */
 180 static int clear_state_bit(struct extent_io_tree *tree,
 181                             struct extent_state *state, int bits)
 182 {
 183         int ret = state->state & bits;
 184
 185         state->state &= ~bits;
 186         if (state->state == 0) {
 187                 remove_cache_extent(&tree->state, &state->cache_node);
 188                 free_extent_state(state);
 189         } else {
 190                 merge_state(tree, state);
 191         }
 192         return ret;
 193 }
 194
 195 /*
 196  * set some bits on a range in the tree.
 197  */
 198 int clear_extent_bits(struct extent_io_tree *tree, u64 start,
 199                       u64 end, int bits, gfp_t mask)
 200 {
 201         struct extent_state *state;
 202         struct extent_state *prealloc = NULL;
 203         struct cache_extent *node;
 204         int err;
 205         int set = 0;
 206
 207 again:
 208         prealloc = alloc_extent_state();
 209         if (!prealloc)
 210                 return -ENOMEM;
 211
 212         /*
 213          * this search will find the extents that end after
 214          * our range starts
 215          */
 216         node = find_first_cache_extent(&tree->state, start);
 217         if (!node)
 218                 goto out;
 219         state = container_of(node, struct extent_state, cache_node);
 220         if (state->start > end)
 221                 goto out;
 222
 223         /*
 224          *     | ---- desired range ---- |
 225          *  | state | or
 226          *  | ------------- state -------------- |
 227          *
 228          * We need to split the extent we found, and may flip
 229          * bits on second half.
 230          *
 231          * If the extent we found extends past our range, we
 232          * just split and search again.  It'll get split again
 233          * the next time though.
 234          *
 235          * If the extent we found is inside our range, we clear
 236          * the desired bit on it.
 237          */
 238         if (state->start < start) {
 239                 err = split_state(tree, state, prealloc, start);
 240                 BUG_ON(err == -EEXIST);
 241                 prealloc = NULL;
 242                 if (err)
 243                         goto out;
 244                 if (state->end <= end) {
 245                         start = state->end + 1;
 246                         set |= clear_state_bit(tree, state, bits);
 247                 } else {
 248                         start = state->start;
 249                 }
 250                 goto search_again;
 251         }
 252         /*
 253          * | ---- desired range ---- |
 254          *                        | state |
 255          * We need to split the extent, and clear the bit
 256          * on the first half
 257          */
 258         if (state->start <= end && state->end > end) {
 259                 err = split_state(tree, state, prealloc, end + 1);
 260                 BUG_ON(err == -EEXIST);
 261
 262                 set |= clear_state_bit(tree, prealloc, bits);
 263                 prealloc = NULL;
 264                 goto out;
 265         }
 266
 267         start = state->end + 1;
 268         set |= clear_state_bit(tree, state, bits);
 269         goto search_again;
 270 out:
 271         if (prealloc)
 272                 free_extent_state(prealloc);
 273         return set;
 274
 275 search_again:
 276         if (start > end)
 277                 goto out;
 278         goto again;
 279 }
 280
 281 /*
 282  * set some bits on a range in the tree.
 283  */
 284 int set_extent_bits(struct extent_io_tree *tree, u64 start,
 285                     u64 end, int bits, gfp_t mask)
 286 {
 287         struct extent_state *state;
 288         struct extent_state *prealloc = NULL;
 289         struct cache_extent *node;
 290         int err = 0;
 291         int set;
 292         u64 last_start;
 293         u64 last_end;
 294 again:
 295         prealloc = alloc_extent_state();
 296         if (!prealloc)
 297                 return -ENOMEM;
 298
 299         /*
 300          * this search will find the extents that end after
 301          * our range starts
 302          */
 303         node = find_first_cache_extent(&tree->state, start);
 304         if (!node) {
 305                 err = insert_state(tree, prealloc, start, end, bits);
 306                 BUG_ON(err == -EEXIST);
 307                 prealloc = NULL;
 308                 goto out;
 309         }
 310
 311         state = container_of(node, struct extent_state, cache_node);
 312         last_start = state->start;
 313         last_end = state->end;
 314
 315         /*
 316          * | ---- desired range ---- |
 317          * | state |
 318          *
 319          * Just lock what we found and keep going
 320          */
 321         if (state->start == start && state->end <= end) {
 322                 set = state->state & bits;
 323                 state->state |= bits;
 324                 start = state->end + 1;
 325                 merge_state(tree, state);
 326                 goto search_again;
 327         }
 328         /*
 329          *     | ---- desired range ---- |
 330          * | state |
 331          *   or
 332          * | ------------- state -------------- |
 333          *
 334          * We need to split the extent we found, and may flip bits on
 335          * second half.
 336          *
 337          * If the extent we found extends past our
 338          * range, we just split and search again.  It'll get split
 339          * again the next time though.
 340          *
 341          * If the extent we found is inside our range, we set the
 342          * desired bit on it.
 343          */
 344         if (state->start < start) {
 345                 set = state->state & bits;
 346                 err = split_state(tree, state, prealloc, start);
 347                 BUG_ON(err == -EEXIST);
 348                 prealloc = NULL;
 349                 if (err)
 350                         goto out;
 351                 if (state->end <= end) {
 352                         state->state |= bits;
 353                         start = state->end + 1;
 354                         merge_state(tree, state);
 355                 } else {
 356                         start = state->start;
 357                 }
 358                 goto search_again;
 359         }
 360         /*
 361          * | ---- desired range ---- |
 362          *     | state | or               | state |
 363          *
 364          * There's a hole, we need to insert something in it and
 365          * ignore the extent we found.
 366          */
 367         if (state->start > start) {
 368                 u64 this_end;
 369                 if (end < last_start)
 370                         this_end = end;
 371                 else
 372                         this_end = last_start -1;
 373                 err = insert_state(tree, prealloc, start, this_end,
 374                                 bits);
 375                 BUG_ON(err == -EEXIST);
 376                 prealloc = NULL;
 377                 if (err)
 378                         goto out;
 379                 start = this_end + 1;
 380                 goto search_again;
 381         }
 382         /*
 383          * | ---- desired range ---- |
 384          * | ---------- state ---------- |
 385          * We need to split the extent, and set the bit
 386          * on the first half
 387          */
 388         set = state->state & bits;
 389         err = split_state(tree, state, prealloc, end + 1);
 390         BUG_ON(err == -EEXIST);
 391
 392         state->state |= bits;
 393         merge_state(tree, prealloc);
 394         prealloc = NULL;
 395 out:
 396         if (prealloc)
 397                 free_extent_state(prealloc);
 398         return err;
 399 search_again:
 400         if (start > end)
 401                 goto out;
 402         goto again;
 403 }
 404
 405 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 406                      gfp_t mask)
 407 {
 408         return set_extent_bits(tree, start, end, EXTENT_DIRTY, mask);
 409 }
 410
 411 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 412                        gfp_t mask)
 413 {
 414         return clear_extent_bits(tree, start, end, EXTENT_DIRTY, mask);
 415 }
 416
 417 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
 418                           u64 *start_ret, u64 *end_ret, int bits)
 419 {
 420         struct cache_extent *node;
 421         struct extent_state *state;
 422         int ret = 1;
 423
 424         /*
 425          * this search will find all the extents that end after
 426          * our range starts.
 427          */
 428         node = find_first_cache_extent(&tree->state, start);
 429         if (!node)
 430                 goto out;
 431
 432         while(1) {
 433                 state = container_of(node, struct extent_state, cache_node);
 434                 if (state->end >= start && (state->state & bits)) {
 435                         *start_ret = state->start;
 436                         *end_ret = state->end;
 437                         ret = 0;
 438                         break;
 439                 }
 440                 node = next_cache_extent(node);
 441                 if (!node)
 442                         break;
 443         }
 444 out:
 445         return ret;
 446 }
 447
 448 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
 449                    int bits, int filled)
 450 {
 451         struct extent_state *state = NULL;
 452         struct cache_extent *node;
 453         int bitset = 0;
 454
 455         node = find_first_cache_extent(&tree->state, start);
 456         while (node && start <= end) {
 457                 state = container_of(node, struct extent_state, cache_node);
 458
 459                 if (filled && state->start > start) {
 460                         bitset = 0;
 461                         break;
 462                 }
 463                 if (state->start > end)
 464                         break;
 465                 if (state->state & bits) {
 466                         bitset = 1;
 467                         if (!filled)
 468                                 break;
 469                 } else if (filled) {
 470                         bitset = 0;
 471                         break;
 472                 }
 473                 start = state->end + 1;
 474                 if (start > end)
 475                         break;
 476                 node = next_cache_extent(node);
 477                 if (!node) {
 478                         if (filled)
 479                                 bitset = 0;
 480                         break;
 481                 }
 482         }
 483         return bitset;
 484 }
 485
 486 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
 487 {
 488         struct cache_extent *node;
 489         struct extent_state *state;
 490         int ret = 0;
 491
 492         node = find_first_cache_extent(&tree->state, start);
 493         if (!node) {
 494                 ret = -ENOENT;
 495                 goto out;
 496         }
 497         state = container_of(node, struct extent_state, cache_node);
 498         if (state->start != start) {
 499                 ret = -ENOENT;
 500                 goto out;
 501         }
 502         state->private = private;
 503 out:
 504         return ret;
 505 }
 506
 507 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
 508 {
 509         struct cache_extent *node;
 510         struct extent_state *state;
 511         int ret = 0;
 512
 513         node = find_first_cache_extent(&tree->state, start);
 514         if (!node) {
 515                 ret = -ENOENT;
 516                 goto out;
 517         }
 518         state = container_of(node, struct extent_state, cache_node);
 519         if (state->start != start) {
 520                 ret = -ENOENT;
 521                 goto out;
 522         }
 523         *private = state->private;
 524 out:
 525         return ret;
 526 }
 527
 528 static int free_some_buffers(struct extent_io_tree *tree)
 529 {
 530         u32 nrscan = 0;
 531         struct extent_buffer *eb;
 532         struct list_head *node, *next;
 533
 534         if (tree->cache_size < cache_max)
 535                 return 0;
 536         list_for_each_safe(node, next, &tree->lru) {
 537                 eb = list_entry(node, struct extent_buffer, lru);
 538                 if (eb->refs == 1) {
 539                         free_extent_buffer(eb);
 540                         if (tree->cache_size < cache_max)
 541                                 break;
 542                 }
 543                 if (nrscan++ > 64)
 544                         break;
 545         }
 546         return 0;
 547 }
 548
 549 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
 550                                                    u64 bytenr, u32 blocksize)
 551 {
 552         struct extent_buffer *eb;
 553         int ret;
 554
 555         eb = malloc(sizeof(struct extent_buffer) + blocksize);
 556         if (!eb) {
 557                 BUG();
 558                 return NULL;
 559         }
 560
 561         eb->start = bytenr;
 562         eb->len = blocksize;
 563         eb->refs = 2;
 564         eb->flags = 0;
 565         eb->tree = tree;
 566         eb->fd = -1;
 567         eb->dev_bytenr = (u64)-1;
 568         eb->cache_node.start = bytenr;
 569         eb->cache_node.size = blocksize;
 570
 571         free_some_buffers(tree);
 572         ret = insert_existing_cache_extent(&tree->cache, &eb->cache_node);
 573         if (ret) {
 574                 free(eb);
 575                 return NULL;
 576         }
 577         list_add_tail(&eb->lru, &tree->lru);
 578         tree->cache_size += blocksize;
 579         return eb;
 580 }
 581
 582 void free_extent_buffer(struct extent_buffer *eb)
 583 {
 584         if (!eb)
 585                 return;
 586
 587         eb->refs--;
 588         BUG_ON(eb->refs < 0);
 589         if (eb->refs == 0) {
 590                 struct extent_io_tree *tree = eb->tree;
 591                 BUG_ON(eb->flags & EXTENT_DIRTY);
 592                 list_del_init(&eb->lru);
 593                 remove_cache_extent(&tree->cache, &eb->cache_node);
 594                 BUG_ON(tree->cache_size < eb->len);
 595                 tree->cache_size -= eb->len;
 596                 free(eb);
 597         }
 598 }
 599
 600 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
 601                                          u64 bytenr, u32 blocksize)
 602 {
 603         struct extent_buffer *eb = NULL;
 604         struct cache_extent *cache;
 605
 606         cache = find_cache_extent(&tree->cache, bytenr, blocksize);
 607         if (cache && cache->start == bytenr && cache->size == blocksize) {
 608                 eb = container_of(cache, struct extent_buffer, cache_node);
 609                 list_move_tail(&eb->lru, &tree->lru);
 610                 eb->refs++;
 611         }
 612         return eb;
 613 }
 614
 615 struct extent_buffer *find_first_extent_buffer(struct extent_io_tree *tree,
 616                                                u64 start)
 617 {
 618         struct extent_buffer *eb = NULL;
 619         struct cache_extent *cache;
 620
 621         cache = find_first_cache_extent(&tree->cache, start);
 622         if (cache) {
 623                 eb = container_of(cache, struct extent_buffer, cache_node);
 624                 list_move_tail(&eb->lru, &tree->lru);
 625                 eb->refs++;
 626         }
 627         return eb;
 628 }
 629
 630 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
 631                                           u64 bytenr, u32 blocksize)
 632 {
 633         struct extent_buffer *eb;
 634         struct cache_extent *cache;
 635
 636         cache = find_cache_extent(&tree->cache, bytenr, blocksize);
 637         if (cache && cache->start == bytenr && cache->size == blocksize) {
 638                 eb = container_of(cache, struct extent_buffer, cache_node);
 639                 list_move_tail(&eb->lru, &tree->lru);
 640                 eb->refs++;
 641         } else {
 642                 if (cache) {
 643                         eb = container_of(cache, struct extent_buffer,
 644                                           cache_node);
 645                         BUG_ON(eb->refs != 1);
 646                         free_extent_buffer(eb);
 647                 }
 648                 eb = __alloc_extent_buffer(tree, bytenr, blocksize);
 649         }
 650         return eb;
 651 }
 652
 653 int read_extent_from_disk(struct extent_buffer *eb)
 654 {
 655         int ret;
 656         ret = pread(eb->fd, eb->data, eb->len, eb->dev_bytenr);
 657         if (ret < 0)
 658                 goto out;
 659         if (ret != eb->len) {
 660                 ret = -EIO;
 661                 goto out;
 662         }
 663         ret = 0;
 664 out:
 665         return ret;
 666 }
 667
 668 int write_extent_to_disk(struct extent_buffer *eb)
 669 {
 670         int ret;
 671         ret = pwrite(eb->fd, eb->data, eb->len, eb->dev_bytenr);
 672         if (ret < 0)
 673                 goto out;
 674         if (ret != eb->len) {
 675                 ret = -EIO;
 676                 goto out;
 677         }
 678         ret = 0;
 679 out:
 680         return ret;
 681 }
 682
 683 int set_extent_buffer_uptodate(struct extent_buffer *eb)
 684 {
 685         eb->flags |= EXTENT_UPTODATE;
 686         return 0;
 687 }
 688
 689 int extent_buffer_uptodate(struct extent_buffer *eb)
 690 {
 691         if (eb->flags & EXTENT_UPTODATE)
 692                 return 1;
 693         return 0;
 694 }
 695
 696 int set_extent_buffer_dirty(struct extent_buffer *eb)
 697 {
 698         struct extent_io_tree *tree = eb->tree;
 699         if (!(eb->flags & EXTENT_DIRTY)) {
 700                 eb->flags |= EXTENT_DIRTY;
 701                 set_extent_dirty(tree, eb->start, eb->start + eb->len - 1, 0);
 702                 extent_buffer_get(eb);
 703         }
 704         return 0;
 705 }
 706
 707 int clear_extent_buffer_dirty(struct extent_buffer *eb)
 708 {
 709         struct extent_io_tree *tree = eb->tree;
 710         if (eb->flags & EXTENT_DIRTY) {
 711                 eb->flags &= ~EXTENT_DIRTY;
 712                 clear_extent_dirty(tree, eb->start, eb->start + eb->len - 1, 0);
 713                 free_extent_buffer(eb);
 714         }
 715         return 0;
 716 }
 717
 718 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
 719                          unsigned long start, unsigned long len)
 720 {
 721         return memcmp(eb->data + start, ptrv, len);
 722 }
 723
 724 void read_extent_buffer(struct extent_buffer *eb, void *dst,
 725                         unsigned long start, unsigned long len)
 726 {
 727         memcpy(dst, eb->data + start, len);
 728 }
 729
 730 void write_extent_buffer(struct extent_buffer *eb, const void *src,
 731                          unsigned long start, unsigned long len)
 732 {
 733         memcpy(eb->data + start, src, len);
 734 }
 735
 736 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
 737                         unsigned long dst_offset, unsigned long src_offset,
 738                         unsigned long len)
 739 {
 740         memcpy(dst->data + dst_offset, src->data + src_offset, len);
 741 }
 742
 743 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
 744                           unsigned long src_offset, unsigned long len)
 745 {
 746         memcpy(dst->data + dst_offset, dst->data + src_offset, len);
 747 }
 748
 749 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
 750                            unsigned long src_offset, unsigned long len)
 751 {
 752         memmove(dst->data + dst_offset, dst->data + src_offset, len);
 753 }
 754
 755 void memset_extent_buffer(struct extent_buffer *eb, char c,
 756                           unsigned long start, unsigned long len)
 757 {
 758         memset(eb->data + start, c, len);
 759 }