btrfs-image.c

   1 /*
   2  * Copyright (C) 2008 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
  19 #define _XOPEN_SOURCE 500
  20 #define _GNU_SOURCE 1
  21 #include <pthread.h>
  22 #include <stdio.h>
  23 #include <stdlib.h>
  24 #include <sys/types.h>
  25 #include <sys/stat.h>
  26 #include <fcntl.h>
  27 #include <unistd.h>
  28 #include <dirent.h>
  29 #include <zlib.h>
  30 #include "kerncompat.h"
  31 #include "crc32c.h"
  32 #include "ctree.h"
  33 #include "disk-io.h"
  34 #include "transaction.h"
  35 #include "utils.h"
  36 #include "version.h"
  37
  38
  39 #define HEADER_MAGIC            0xbd5c25e27295668bULL
  40 #define MAX_PENDING_SIZE        (256 * 1024)
  41 #define BLOCK_SIZE              1024
  42 #define BLOCK_MASK              (BLOCK_SIZE - 1)
  43
  44 #define COMPRESS_NONE           0
  45 #define COMPRESS_ZLIB           1
  46
  47 struct meta_cluster_item {
  48         __le64 bytenr;
  49         __le32 size;
  50 } __attribute__ ((__packed__));
  51
  52 struct meta_cluster_header {
  53         __le64 magic;
  54         __le64 bytenr;
  55         __le32 nritems;
  56         u8 compress;
  57 } __attribute__ ((__packed__));
  58
  59 /* cluster header + index items + buffers */
  60 struct meta_cluster {
  61         struct meta_cluster_header header;
  62         struct meta_cluster_item items[];
  63 } __attribute__ ((__packed__));
  64
  65 #define ITEMS_PER_CLUSTER ((BLOCK_SIZE - sizeof(struct meta_cluster)) / \
  66                            sizeof(struct meta_cluster_item))
  67
  68 struct async_work {
  69         struct list_head list;
  70         struct list_head ordered;
  71         u64 start;
  72         u64 size;
  73         u8 *buffer;
  74         size_t bufsize;
  75         int error;
  76 };
  77
  78 struct metadump_struct {
  79         struct btrfs_root *root;
  80         FILE *out;
  81
  82         struct meta_cluster *cluster;
  83
  84         pthread_t *threads;
  85         size_t num_threads;
  86         pthread_mutex_t mutex;
  87         pthread_cond_t cond;
  88
  89         struct list_head list;
  90         struct list_head ordered;
  91         size_t num_items;
  92         size_t num_ready;
  93
  94         u64 pending_start;
  95         u64 pending_size;
  96
  97         int compress_level;
  98         int done;
  99 };
 100
 101 struct mdrestore_struct {
 102         FILE *in;
 103         FILE *out;
 104
 105         pthread_t *threads;
 106         size_t num_threads;
 107         pthread_mutex_t mutex;
 108         pthread_cond_t cond;
 109
 110         struct list_head list;
 111         size_t num_items;
 112
 113         int compress_method;
 114         int done;
 115         int error;
 116 };
 117
 118 static void csum_block(u8 *buf, size_t len)
 119 {
 120         char result[BTRFS_CRC32_SIZE];
 121         u32 crc = ~(u32)0;
 122         crc = crc32c(crc, buf + BTRFS_CSUM_SIZE, len - BTRFS_CSUM_SIZE);
 123         btrfs_csum_final(crc, result);
 124         memcpy(buf, result, BTRFS_CRC32_SIZE);
 125 }
 126
 127 /*
 128  * zero inline extents and csum items
 129  */
 130 static void zero_items(u8 *dst, struct extent_buffer *src)
 131 {
 132         struct btrfs_file_extent_item *fi;
 133         struct btrfs_item *item;
 134         struct btrfs_key key;
 135         u32 nritems = btrfs_header_nritems(src);
 136         size_t size;
 137         unsigned long ptr;
 138         int i, extent_type;
 139
 140         for (i = 0; i < nritems; i++) {
 141                 item = btrfs_item_nr(src, i);
 142                 btrfs_item_key_to_cpu(src, &key, i);
 143                 if (key.type == BTRFS_CSUM_ITEM_KEY) {
 144                         size = btrfs_item_size_nr(src, i);
 145                         memset(dst + btrfs_leaf_data(src) +
 146                                btrfs_item_offset_nr(src, i), 0, size);
 147                         continue;
 148                 }
 149                 if (key.type != BTRFS_EXTENT_DATA_KEY)
 150                         continue;
 151
 152                 fi = btrfs_item_ptr(src, i, struct btrfs_file_extent_item);
 153                 extent_type = btrfs_file_extent_type(src, fi);
 154                 if (extent_type != BTRFS_FILE_EXTENT_INLINE)
 155                         continue;
 156
 157                 ptr = btrfs_file_extent_inline_start(fi);
 158                 size = btrfs_file_extent_inline_item_len(src, item);
 159                 memset(dst + ptr, 0, size);
 160         }
 161 }
 162
 163 /*
 164  * copy buffer and zero useless data in the buffer
 165  */
 166 static void copy_buffer(u8 *dst, struct extent_buffer *src)
 167 {
 168         int level;
 169         size_t size;
 170         u32 nritems;
 171
 172         memcpy(dst, src->data, src->len);
 173         if (src->start == BTRFS_SUPER_INFO_OFFSET)
 174                 return;
 175
 176         level = btrfs_header_level(src);
 177         nritems = btrfs_header_nritems(src);
 178
 179         if (nritems == 0) {
 180                 size = sizeof(struct btrfs_header);
 181                 memset(dst + size, 0, src->len - size);
 182         } else if (level == 0) {
 183                 size = btrfs_leaf_data(src) +
 184                         btrfs_item_offset_nr(src, nritems - 1) -
 185                         btrfs_item_nr_offset(nritems);
 186                 memset(dst + btrfs_item_nr_offset(nritems), 0, size);
 187                 zero_items(dst, src);
 188         } else {
 189                 size = offsetof(struct btrfs_node, ptrs) +
 190                         sizeof(struct btrfs_key_ptr) * nritems;
 191                 memset(dst + size, 0, src->len - size);
 192         }
 193         csum_block(dst, src->len);
 194 }
 195
 196 static void *dump_worker(void *data)
 197 {
 198         struct metadump_struct *md = (struct metadump_struct *)data;
 199         struct async_work *async;
 200         int ret;
 201
 202         while (1) {
 203                 pthread_mutex_lock(&md->mutex);
 204                 while (list_empty(&md->list)) {
 205                         if (md->done) {
 206                                 pthread_mutex_unlock(&md->mutex);
 207                                 goto out;
 208                         }
 209                         pthread_cond_wait(&md->cond, &md->mutex);
 210                 }
 211                 async = list_entry(md->list.next, struct async_work, list);
 212                 list_del_init(&async->list);
 213                 pthread_mutex_unlock(&md->mutex);
 214
 215                 if (md->compress_level > 0) {
 216                         u8 *orig = async->buffer;
 217
 218                         async->bufsize = compressBound(async->size);
 219                         async->buffer = malloc(async->bufsize);
 220
 221                         ret = compress2(async->buffer,
 222                                          (unsigned long *)&async->bufsize,
 223                                          orig, async->size, md->compress_level);
 224
 225                         if (ret != Z_OK)
 226                                 async->error = 1;
 227
 228                         free(orig);
 229                 }
 230
 231                 pthread_mutex_lock(&md->mutex);
 232                 md->num_ready++;
 233                 pthread_mutex_unlock(&md->mutex);
 234         }
 235 out:
 236         pthread_exit(NULL);
 237 }
 238
 239 static void meta_cluster_init(struct metadump_struct *md, u64 start)
 240 {
 241         struct meta_cluster_header *header;
 242
 243         md->num_items = 0;
 244         md->num_ready = 0;
 245         header = &md->cluster->header;
 246         header->magic = cpu_to_le64(HEADER_MAGIC);
 247         header->bytenr = cpu_to_le64(start);
 248         header->nritems = cpu_to_le32(0);
 249         header->compress = md->compress_level > 0 ?
 250                            COMPRESS_ZLIB : COMPRESS_NONE;
 251 }
 252
 253 static int metadump_init(struct metadump_struct *md, struct btrfs_root *root,
 254                          FILE *out, int num_threads, int compress_level)
 255 {
 256         int i, ret;
 257
 258         memset(md, 0, sizeof(*md));
 259         pthread_cond_init(&md->cond, NULL);
 260         pthread_mutex_init(&md->mutex, NULL);
 261         INIT_LIST_HEAD(&md->list);
 262         INIT_LIST_HEAD(&md->ordered);
 263         md->root = root;
 264         md->out = out;
 265         md->pending_start = (u64)-1;
 266         md->compress_level = compress_level;
 267         md->cluster = calloc(1, BLOCK_SIZE);
 268         if (!md->cluster) {
 269                 pthread_cond_destroy(&md->cond);
 270                 pthread_mutex_destroy(&md->mutex);
 271                 return -ENOMEM;
 272         }
 273
 274         meta_cluster_init(md, 0);
 275         if (!num_threads)
 276                 return 0;
 277
 278         md->num_threads = num_threads;
 279         md->threads = calloc(num_threads, sizeof(pthread_t));
 280         if (!md->threads) {
 281                 free(md->cluster);
 282                 pthread_cond_destroy(&md->cond);
 283                 pthread_mutex_destroy(&md->mutex);
 284                 return -ENOMEM;
 285         }
 286
 287         for (i = 0; i < num_threads; i++) {
 288                 ret = pthread_create(md->threads + i, NULL, dump_worker, md);
 289                 if (ret)
 290                         break;
 291         }
 292
 293         if (ret) {
 294                 pthread_mutex_lock(&md->mutex);
 295                 md->done = 1;
 296                 pthread_cond_broadcast(&md->cond);
 297                 pthread_mutex_unlock(&md->mutex);
 298
 299                 for (i--; i >= 0; i--)
 300                         pthread_join(md->threads[i], NULL);
 301
 302                 pthread_cond_destroy(&md->cond);
 303                 pthread_mutex_destroy(&md->mutex);
 304                 free(md->cluster);
 305                 free(md->threads);
 306         }
 307
 308         return ret;
 309 }
 310
 311 static void metadump_destroy(struct metadump_struct *md)
 312 {
 313         int i;
 314         pthread_mutex_lock(&md->mutex);
 315         md->done = 1;
 316         pthread_cond_broadcast(&md->cond);
 317         pthread_mutex_unlock(&md->mutex);
 318
 319         for (i = 0; i < md->num_threads; i++)
 320                 pthread_join(md->threads[i], NULL);
 321
 322         pthread_cond_destroy(&md->cond);
 323         pthread_mutex_destroy(&md->mutex);
 324         free(md->threads);
 325         free(md->cluster);
 326 }
 327
 328 static int write_zero(FILE *out, size_t size)
 329 {
 330         static char zero[BLOCK_SIZE];
 331         return fwrite(zero, size, 1, out);
 332 }
 333
 334 static int write_buffers(struct metadump_struct *md, u64 *next)
 335 {
 336         struct meta_cluster_header *header = &md->cluster->header;
 337         struct meta_cluster_item *item;
 338         struct async_work *async;
 339         u64 bytenr = 0;
 340         u32 nritems = 0;
 341         int ret;
 342         int err = 0;
 343
 344         if (list_empty(&md->ordered))
 345                 goto out;
 346
 347         /* wait until all buffers are compressed */
 348         while (md->num_items > md->num_ready) {
 349                 struct timespec ts = {
 350                         .tv_sec = 0,
 351                         .tv_nsec = 10000000,
 352                 };
 353                 pthread_mutex_unlock(&md->mutex);
 354                 nanosleep(&ts, NULL);
 355                 pthread_mutex_lock(&md->mutex);
 356         }
 357
 358         /* setup and write index block */
 359         list_for_each_entry(async, &md->ordered, ordered) {
 360                 item = md->cluster->items + nritems;
 361                 item->bytenr = cpu_to_le64(async->start);
 362                 item->size = cpu_to_le32(async->bufsize);
 363                 nritems++;
 364         }
 365         header->nritems = cpu_to_le32(nritems);
 366
 367         ret = fwrite(md->cluster, BLOCK_SIZE, 1, md->out);
 368         if (ret != 1) {
 369                 fprintf(stderr, "Error writing out cluster: %d\n", errno);
 370                 return -EIO;
 371         }
 372
 373         /* write buffers */
 374         bytenr += le64_to_cpu(header->bytenr) + BLOCK_SIZE;
 375         while (!list_empty(&md->ordered)) {
 376                 async = list_entry(md->ordered.next, struct async_work,
 377                                    ordered);
 378                 list_del_init(&async->ordered);
 379
 380                 bytenr += async->bufsize;
 381                 if (!err)
 382                         ret = fwrite(async->buffer, async->bufsize, 1,
 383                                      md->out);
 384                 if (ret != 1) {
 385                         err = -EIO;
 386                         ret = 0;
 387                         fprintf(stderr, "Error writing out cluster: %d\n",
 388                                 errno);
 389                 }
 390
 391                 free(async->buffer);
 392                 free(async);
 393         }
 394
 395         /* zero unused space in the last block */
 396         if (!err && bytenr & BLOCK_MASK) {
 397                 size_t size = BLOCK_SIZE - (bytenr & BLOCK_MASK);
 398
 399                 bytenr += size;
 400                 ret = write_zero(md->out, size);
 401                 if (ret != 1) {
 402                         fprintf(stderr, "Error zeroing out buffer: %d\n",
 403                                 errno);
 404                         err = -EIO;
 405                 }
 406         }
 407 out:
 408         *next = bytenr;
 409         return err;
 410 }
 411
 412 static int flush_pending(struct metadump_struct *md, int done)
 413 {
 414         struct async_work *async = NULL;
 415         struct extent_buffer *eb;
 416         u64 blocksize = md->root->nodesize;
 417         u64 start;
 418         u64 size;
 419         size_t offset;
 420         int ret = 0;
 421
 422         if (md->pending_size) {
 423                 async = calloc(1, sizeof(*async));
 424                 if (!async)
 425                         return -ENOMEM;
 426
 427                 async->start = md->pending_start;
 428                 async->size = md->pending_size;
 429                 async->bufsize = async->size;
 430                 async->buffer = malloc(async->bufsize);
 431                 if (!async->buffer) {
 432                         free(async);
 433                         return -ENOMEM;
 434                 }
 435                 offset = 0;
 436                 start = async->start;
 437                 size = async->size;
 438                 while (size > 0) {
 439                         eb = read_tree_block(md->root, start, blocksize, 0);
 440                         if (!eb) {
 441                                 free(async->buffer);
 442                                 free(async);
 443                                 fprintf(stderr,
 444                                         "Error reading metadata block\n");
 445                                 return -EIO;
 446                         }
 447                         copy_buffer(async->buffer + offset, eb);
 448                         free_extent_buffer(eb);
 449                         start += blocksize;
 450                         offset += blocksize;
 451                         size -= blocksize;
 452                 }
 453
 454                 md->pending_start = (u64)-1;
 455                 md->pending_size = 0;
 456         } else if (!done) {
 457                 return 0;
 458         }
 459
 460         pthread_mutex_lock(&md->mutex);
 461         if (async) {
 462                 list_add_tail(&async->ordered, &md->ordered);
 463                 md->num_items++;
 464                 if (md->compress_level > 0) {
 465                         list_add_tail(&async->list, &md->list);
 466                         pthread_cond_signal(&md->cond);
 467                 } else {
 468                         md->num_ready++;
 469                 }
 470         }
 471         if (md->num_items >= ITEMS_PER_CLUSTER || done) {
 472                 ret = write_buffers(md, &start);
 473                 if (ret)
 474                         fprintf(stderr, "Error writing buffers %d\n",
 475                                 errno);
 476                 else
 477                         meta_cluster_init(md, start);
 478         }
 479         pthread_mutex_unlock(&md->mutex);
 480         return ret;
 481 }
 482
 483 static int add_metadata(u64 start, u64 size, struct metadump_struct *md)
 484 {
 485         int ret;
 486         if (md->pending_size + size > MAX_PENDING_SIZE ||
 487             md->pending_start + md->pending_size != start) {
 488                 ret = flush_pending(md, 0);
 489                 if (ret)
 490                         return ret;
 491                 md->pending_start = start;
 492         }
 493         readahead_tree_block(md->root, start, size, 0);
 494         md->pending_size += size;
 495         return 0;
 496 }
 497
 498 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 499 static int is_tree_block(struct btrfs_root *extent_root,
 500                          struct btrfs_path *path, u64 bytenr)
 501 {
 502         struct extent_buffer *leaf;
 503         struct btrfs_key key;
 504         u64 ref_objectid;
 505         int ret;
 506
 507         leaf = path->nodes[0];
 508         while (1) {
 509                 struct btrfs_extent_ref_v0 *ref_item;
 510                 path->slots[0]++;
 511                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 512                         ret = btrfs_next_leaf(extent_root, path);
 513                         if (ret < 0)
 514                                 return ret;
 515                         if (ret > 0)
 516                                 break;
 517                         leaf = path->nodes[0];
 518                 }
 519                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 520                 if (key.objectid != bytenr)
 521                         break;
 522                 if (key.type != BTRFS_EXTENT_REF_V0_KEY)
 523                         continue;
 524                 ref_item = btrfs_item_ptr(leaf, path->slots[0],
 525                                           struct btrfs_extent_ref_v0);
 526                 ref_objectid = btrfs_ref_objectid_v0(leaf, ref_item);
 527                 if (ref_objectid < BTRFS_FIRST_FREE_OBJECTID)
 528                         return 1;
 529                 break;
 530         }
 531         return 0;
 532 }
 533 #endif
 534
 535 static int copy_log_blocks(struct btrfs_root *root, struct extent_buffer *eb,
 536                            struct metadump_struct *metadump,
 537                            int log_root_tree)
 538 {
 539         struct extent_buffer *tmp;
 540         struct btrfs_root_item *ri;
 541         struct btrfs_key key;
 542         u64 bytenr;
 543         int level;
 544         int nritems = 0;
 545         int i = 0;
 546         int ret;
 547
 548         ret = add_metadata(btrfs_header_bytenr(eb), root->leafsize, metadump);
 549         if (ret) {
 550                 fprintf(stderr, "Error adding metadata block\n");
 551                 return ret;
 552         }
 553
 554         if (btrfs_header_level(eb) == 0 && !log_root_tree)
 555                 return 0;
 556
 557         level = btrfs_header_level(eb);
 558         nritems = btrfs_header_nritems(eb);
 559         for (i = 0; i < nritems; i++) {
 560                 if (level == 0) {
 561                         btrfs_item_key_to_cpu(eb, &key, i);
 562                         if (key.type != BTRFS_ROOT_ITEM_KEY)
 563                                 continue;
 564                         ri = btrfs_item_ptr(eb, i, struct btrfs_root_item);
 565                         bytenr = btrfs_disk_root_bytenr(eb, ri);
 566                         tmp = read_tree_block(root, bytenr, root->leafsize, 0);
 567                         if (!tmp) {
 568                                 fprintf(stderr,
 569                                         "Error reading log root block\n");
 570                                 return -EIO;
 571                         }
 572                         ret = copy_log_blocks(root, tmp, metadump, 0);
 573                         free_extent_buffer(tmp);
 574                         if (ret)
 575                                 return ret;
 576                 } else {
 577                         bytenr = btrfs_node_blockptr(eb, i);
 578                         tmp = read_tree_block(root, bytenr, root->leafsize, 0);
 579                         if (!tmp) {
 580                                 fprintf(stderr, "Error reading log block\n");
 581                                 return -EIO;
 582                         }
 583                         ret = copy_log_blocks(root, tmp, metadump,
 584                                               log_root_tree);
 585                         free_extent_buffer(tmp);
 586                         if (ret)
 587                                 return ret;
 588                 }
 589         }
 590
 591         return 0;
 592 }
 593
 594 static int copy_log_trees(struct btrfs_root *root,
 595                           struct metadump_struct *metadump,
 596                           struct btrfs_path *path)
 597 {
 598         u64 blocknr = btrfs_super_log_root(root->fs_info->super_copy);
 599
 600         if (blocknr == 0)
 601                 return 0;
 602
 603         if (!root->fs_info->log_root_tree ||
 604             !root->fs_info->log_root_tree->node) {
 605                 fprintf(stderr, "Error copying tree log, it wasn't setup\n");
 606                 return -EIO;
 607         }
 608
 609         return copy_log_blocks(root, root->fs_info->log_root_tree->node,
 610                                metadump, 1);
 611 }
 612
 613 static int create_metadump(const char *input, FILE *out, int num_threads,
 614                            int compress_level)
 615 {
 616         struct btrfs_root *root;
 617         struct btrfs_root *extent_root;
 618         struct btrfs_path *path = NULL;
 619         struct extent_buffer *leaf;
 620         struct btrfs_extent_item *ei;
 621         struct btrfs_key key;
 622         struct metadump_struct metadump;
 623         u64 bytenr;
 624         u64 num_bytes;
 625         int ret;
 626         int err = 0;
 627
 628         root = open_ctree(input, 0, 0);
 629         if (!root) {
 630                 fprintf(stderr, "Open ctree failed\n");
 631                 return -EIO;
 632         }
 633
 634         BUG_ON(root->nodesize != root->leafsize);
 635
 636         ret = metadump_init(&metadump, root, out, num_threads,
 637                             compress_level);
 638         if (ret) {
 639                 fprintf(stderr, "Error initing metadump %d\n", ret);
 640                 close_ctree(root);
 641                 return ret;
 642         }
 643
 644         ret = add_metadata(BTRFS_SUPER_INFO_OFFSET, 4096, &metadump);
 645         if (ret) {
 646                 fprintf(stderr, "Error adding metadata %d\n", ret);
 647                 err = ret;
 648                 goto out;
 649         }
 650
 651         extent_root = root->fs_info->extent_root;
 652         path = btrfs_alloc_path();
 653         if (!path) {
 654                 fprintf(stderr, "Out of memory allocing path\n");
 655                 err = -ENOMEM;
 656                 goto out;
 657         }
 658         bytenr = BTRFS_SUPER_INFO_OFFSET + 4096;
 659         key.objectid = bytenr;
 660         key.type = BTRFS_EXTENT_ITEM_KEY;
 661         key.offset = 0;
 662
 663         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 664         if (ret < 0) {
 665                 fprintf(stderr, "Error searching extent root %d\n", ret);
 666                 err = ret;
 667                 goto out;
 668         }
 669
 670         while (1) {
 671                 leaf = path->nodes[0];
 672                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 673                         ret = btrfs_next_leaf(extent_root, path);
 674                         if (ret < 0) {
 675                                 fprintf(stderr, "Error going to next leaf %d"
 676                                         "\n", ret);
 677                                 err = ret;
 678                                 goto out;
 679                         }
 680                         if (ret > 0)
 681                                 break;
 682                         leaf = path->nodes[0];
 683                 }
 684
 685                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 686                 if (key.objectid < bytenr ||
 687                     key.type != BTRFS_EXTENT_ITEM_KEY) {
 688                         path->slots[0]++;
 689                         continue;
 690                 }
 691
 692                 bytenr = key.objectid;
 693                 num_bytes = key.offset;
 694
 695                 if (btrfs_item_size_nr(leaf, path->slots[0]) > sizeof(*ei)) {
 696                         ei = btrfs_item_ptr(leaf, path->slots[0],
 697                                             struct btrfs_extent_item);
 698                         if (btrfs_extent_flags(leaf, ei) &
 699                             BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 700                                 ret = add_metadata(bytenr, num_bytes,
 701                                                    &metadump);
 702                                 if (ret) {
 703                                         fprintf(stderr, "Error adding block "
 704                                                 "%d\n", ret);
 705                                         err = ret;
 706                                         goto out;
 707                                 }
 708                         }
 709                 } else {
 710 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 711                         ret = is_tree_block(extent_root, path, bytenr);
 712                         if (ret < 0) {
 713                                 fprintf(stderr, "Error checking tree block "
 714                                         "%d\n", ret);
 715                                 err = ret;
 716                                 goto out;
 717                         }
 718
 719                         if (ret) {
 720                                 ret = add_metadata(bytenr, num_bytes,
 721                                                    &metadump);
 722                                 if (ret) {
 723                                         fprintf(stderr, "Error adding block "
 724                                                 "%d\n", ret);
 725                                         err = ret;
 726                                         goto out;
 727                                 }
 728                         }
 729 #else
 730                         fprintf(stderr, "Either extent tree corruption or "
 731                                 "you haven't built with V0 support\n");
 732                         err = -EIO;
 733                         goto out;
 734 #endif
 735                 }
 736                 bytenr += num_bytes;
 737         }
 738
 739         ret = copy_log_trees(root, &metadump, path);
 740         if (ret)
 741                 err = ret;
 742 out:
 743         ret = flush_pending(&metadump, 1);
 744         if (ret) {
 745                 if (!err)
 746                         ret = err;
 747                 fprintf(stderr, "Error flushing pending %d\n", ret);
 748         }
 749
 750         metadump_destroy(&metadump);
 751
 752         btrfs_free_path(path);
 753         ret = close_ctree(root);
 754         return err ? err : ret;
 755 }
 756
 757 static void update_super(u8 *buffer)
 758 {
 759         struct btrfs_super_block *super = (struct btrfs_super_block *)buffer;
 760         struct btrfs_chunk *chunk;
 761         struct btrfs_disk_key *key;
 762         u32 sectorsize = btrfs_super_sectorsize(super);
 763         u64 flags = btrfs_super_flags(super);
 764
 765         flags |= BTRFS_SUPER_FLAG_METADUMP;
 766         btrfs_set_super_flags(super, flags);
 767
 768         key = (struct btrfs_disk_key *)(super->sys_chunk_array);
 769         chunk = (struct btrfs_chunk *)(super->sys_chunk_array +
 770                                        sizeof(struct btrfs_disk_key));
 771
 772         btrfs_set_disk_key_objectid(key, BTRFS_FIRST_CHUNK_TREE_OBJECTID);
 773         btrfs_set_disk_key_type(key, BTRFS_CHUNK_ITEM_KEY);
 774         btrfs_set_disk_key_offset(key, 0);
 775
 776         btrfs_set_stack_chunk_length(chunk, (u64)-1);
 777         btrfs_set_stack_chunk_owner(chunk, BTRFS_EXTENT_TREE_OBJECTID);
 778         btrfs_set_stack_chunk_stripe_len(chunk, 64 * 1024);
 779         btrfs_set_stack_chunk_type(chunk, BTRFS_BLOCK_GROUP_SYSTEM);
 780         btrfs_set_stack_chunk_io_align(chunk, sectorsize);
 781         btrfs_set_stack_chunk_io_width(chunk, sectorsize);
 782         btrfs_set_stack_chunk_sector_size(chunk, sectorsize);
 783         btrfs_set_stack_chunk_num_stripes(chunk, 1);
 784         btrfs_set_stack_chunk_sub_stripes(chunk, 0);
 785         chunk->stripe.devid = super->dev_item.devid;
 786         chunk->stripe.offset = cpu_to_le64(0);
 787         memcpy(chunk->stripe.dev_uuid, super->dev_item.uuid, BTRFS_UUID_SIZE);
 788         btrfs_set_super_sys_array_size(super, sizeof(*key) + sizeof(*chunk));
 789         csum_block(buffer, 4096);
 790 }
 791
 792 static void *restore_worker(void *data)
 793 {
 794         struct mdrestore_struct *mdres = (struct mdrestore_struct *)data;
 795         struct async_work *async;
 796         size_t size;
 797         u8 *buffer;
 798         u8 *outbuf;
 799         int outfd;
 800         int ret;
 801
 802         outfd = fileno(mdres->out);
 803         buffer = malloc(MAX_PENDING_SIZE * 2);
 804         if (!buffer) {
 805                 fprintf(stderr, "Error allocing buffer\n");
 806                 pthread_mutex_lock(&mdres->mutex);
 807                 if (!mdres->error)
 808                         mdres->error = -ENOMEM;
 809                 pthread_mutex_unlock(&mdres->mutex);
 810                 goto out;
 811         }
 812
 813         while (1) {
 814                 int err = 0;
 815
 816                 pthread_mutex_lock(&mdres->mutex);
 817                 while (list_empty(&mdres->list)) {
 818                         if (mdres->done) {
 819                                 pthread_mutex_unlock(&mdres->mutex);
 820                                 goto out;
 821                         }
 822                         pthread_cond_wait(&mdres->cond, &mdres->mutex);
 823                 }
 824                 async = list_entry(mdres->list.next, struct async_work, list);
 825                 list_del_init(&async->list);
 826                 pthread_mutex_unlock(&mdres->mutex);
 827
 828                 if (mdres->compress_method == COMPRESS_ZLIB) {
 829                         size = MAX_PENDING_SIZE * 2;
 830                         ret = uncompress(buffer, (unsigned long *)&size,
 831                                          async->buffer, async->bufsize);
 832                         if (ret != Z_OK) {
 833                                 fprintf(stderr, "Error decompressing %d\n",
 834                                         ret);
 835                                 err = -EIO;
 836                         }
 837                         outbuf = buffer;
 838                 } else {
 839                         outbuf = async->buffer;
 840                         size = async->bufsize;
 841                 }
 842
 843                 if (async->start == BTRFS_SUPER_INFO_OFFSET)
 844                         update_super(outbuf);
 845
 846                 ret = pwrite64(outfd, outbuf, size, async->start);
 847                 if (ret < size) {
 848                         if (ret < 0) {
 849                                 fprintf(stderr, "Error writing to device %d\n",
 850                                         errno);
 851                                 err = errno;
 852                         } else {
 853                                 fprintf(stderr, "Short write\n");
 854                                 err = -EIO;
 855                         }
 856                 }
 857
 858                 pthread_mutex_lock(&mdres->mutex);
 859                 if (err && !mdres->error)
 860                         mdres->error = err;
 861                 mdres->num_items--;
 862                 pthread_mutex_unlock(&mdres->mutex);
 863
 864                 free(async->buffer);
 865                 free(async);
 866         }
 867 out:
 868         free(buffer);
 869         pthread_exit(NULL);
 870 }
 871
 872 static void mdrestore_destroy(struct mdrestore_struct *mdres)
 873 {
 874         int i;
 875         pthread_mutex_lock(&mdres->mutex);
 876         mdres->done = 1;
 877         pthread_cond_broadcast(&mdres->cond);
 878         pthread_mutex_unlock(&mdres->mutex);
 879
 880         for (i = 0; i < mdres->num_threads; i++)
 881                 pthread_join(mdres->threads[i], NULL);
 882
 883         pthread_cond_destroy(&mdres->cond);
 884         pthread_mutex_destroy(&mdres->mutex);
 885         free(mdres->threads);
 886 }
 887
 888 static int mdrestore_init(struct mdrestore_struct *mdres,
 889                           FILE *in, FILE *out, int num_threads)
 890 {
 891         int i, ret = 0;
 892
 893         memset(mdres, 0, sizeof(*mdres));
 894         pthread_cond_init(&mdres->cond, NULL);
 895         pthread_mutex_init(&mdres->mutex, NULL);
 896         INIT_LIST_HEAD(&mdres->list);
 897         mdres->in = in;
 898         mdres->out = out;
 899
 900         if (!num_threads)
 901                 return 0;
 902
 903         mdres->num_threads = num_threads;
 904         mdres->threads = calloc(num_threads, sizeof(pthread_t));
 905         if (!mdres->threads)
 906                 return -ENOMEM;
 907         for (i = 0; i < num_threads; i++) {
 908                 ret = pthread_create(mdres->threads + i, NULL, restore_worker,
 909                                      mdres);
 910                 if (ret)
 911                         break;
 912         }
 913         if (ret)
 914                 mdrestore_destroy(mdres);
 915         return ret;
 916 }
 917
 918 static int add_cluster(struct meta_cluster *cluster,
 919                        struct mdrestore_struct *mdres, u64 *next)
 920 {
 921         struct meta_cluster_item *item;
 922         struct meta_cluster_header *header = &cluster->header;
 923         struct async_work *async;
 924         u64 bytenr;
 925         u32 i, nritems;
 926         int ret;
 927
 928         BUG_ON(mdres->num_items);
 929         mdres->compress_method = header->compress;
 930
 931         bytenr = le64_to_cpu(header->bytenr) + BLOCK_SIZE;
 932         nritems = le32_to_cpu(header->nritems);
 933         for (i = 0; i < nritems; i++) {
 934                 item = &cluster->items[i];
 935                 async = calloc(1, sizeof(*async));
 936                 if (!async) {
 937                         fprintf(stderr, "Error allocating async\n");
 938                         return -ENOMEM;
 939                 }
 940                 async->start = le64_to_cpu(item->bytenr);
 941                 async->bufsize = le32_to_cpu(item->size);
 942                 async->buffer = malloc(async->bufsize);
 943                 if (!async->buffer) {
 944                         fprintf(stderr, "Error allocing async buffer\n");
 945                         free(async);
 946                         return -ENOMEM;
 947                 }
 948                 ret = fread(async->buffer, async->bufsize, 1, mdres->in);
 949                 if (ret != 1) {
 950                         fprintf(stderr, "Error reading buffer %d\n", errno);
 951                         free(async->buffer);
 952                         free(async);
 953                         return -EIO;
 954                 }
 955                 bytenr += async->bufsize;
 956
 957                 pthread_mutex_lock(&mdres->mutex);
 958                 list_add_tail(&async->list, &mdres->list);
 959                 mdres->num_items++;
 960                 pthread_cond_signal(&mdres->cond);
 961                 pthread_mutex_unlock(&mdres->mutex);
 962         }
 963         if (bytenr & BLOCK_MASK) {
 964                 char buffer[BLOCK_MASK];
 965                 size_t size = BLOCK_SIZE - (bytenr & BLOCK_MASK);
 966
 967                 bytenr += size;
 968                 ret = fread(buffer, size, 1, mdres->in);
 969                 if (ret != 1) {
 970                         fprintf(stderr, "Error reading in buffer %d\n", errno);
 971                         return -EIO;
 972                 }
 973         }
 974         *next = bytenr;
 975         return 0;
 976 }
 977
 978 static int wait_for_worker(struct mdrestore_struct *mdres)
 979 {
 980         int ret = 0;
 981
 982         pthread_mutex_lock(&mdres->mutex);
 983         ret = mdres->error;
 984         while (!ret && mdres->num_items > 0) {
 985                 struct timespec ts = {
 986                         .tv_sec = 0,
 987                         .tv_nsec = 10000000,
 988                 };
 989                 pthread_mutex_unlock(&mdres->mutex);
 990                 nanosleep(&ts, NULL);
 991                 pthread_mutex_lock(&mdres->mutex);
 992                 ret = mdres->error;
 993         }
 994         pthread_mutex_unlock(&mdres->mutex);
 995         return ret;
 996 }
 997
 998 static int restore_metadump(const char *input, FILE *out, int num_threads)
 999 {
1000         struct meta_cluster *cluster = NULL;
1001         struct meta_cluster_header *header;
1002         struct mdrestore_struct mdrestore;
1003         u64 bytenr = 0;
1004         FILE *in = NULL;
1005         int ret = 0;
1006
1007         if (!strcmp(input, "-")) {
1008                 in = stdin;
1009         } else {
1010                 in = fopen(input, "r");
1011                 if (!in) {
1012                         perror("unable to open metadump image");
1013                         return 1;
1014                 }
1015         }
1016
1017         cluster = malloc(BLOCK_SIZE);
1018         if (!cluster) {
1019                 fprintf(stderr, "Error allocating cluster\n");
1020                 if (in != stdin)
1021                         fclose(in);
1022                 return -ENOMEM;
1023         }
1024
1025         ret = mdrestore_init(&mdrestore, in, out, num_threads);
1026         if (ret) {
1027                 fprintf(stderr, "Error initing mdrestore %d\n", ret);
1028                 if (in != stdin)
1029                         fclose(in);
1030                 free(cluster);
1031                 return ret;
1032         }
1033
1034         while (1) {
1035                 ret = fread(cluster, BLOCK_SIZE, 1, in);
1036                 if (!ret)
1037                         break;
1038
1039                 header = &cluster->header;
1040                 if (le64_to_cpu(header->magic) != HEADER_MAGIC ||
1041                     le64_to_cpu(header->bytenr) != bytenr) {
1042                         fprintf(stderr, "bad header in metadump image\n");
1043                         ret = -EIO;
1044                         break;
1045                 }
1046                 ret = add_cluster(cluster, &mdrestore, &bytenr);
1047                 if (ret) {
1048                         fprintf(stderr, "Error adding cluster\n");
1049                         break;
1050                 }
1051
1052                 ret = wait_for_worker(&mdrestore);
1053                 if (ret) {
1054                         fprintf(stderr, "One of the threads errored out %d\n",
1055                                 ret);
1056                         break;
1057                 }
1058         }
1059
1060         mdrestore_destroy(&mdrestore);
1061         free(cluster);
1062         if (in != stdin)
1063                 fclose(in);
1064         return ret;
1065 }
1066
1067 static void print_usage(void)
1068 {
1069         fprintf(stderr, "usage: btrfs-image [options] source target\n");
1070         fprintf(stderr, "\t-r      \trestore metadump image\n");
1071         fprintf(stderr, "\t-c value\tcompression level (0 ~ 9)\n");
1072         fprintf(stderr, "\t-t value\tnumber of threads (1 ~ 32)\n");
1073         exit(1);
1074 }
1075
1076 int main(int argc, char *argv[])
1077 {
1078         char *source;
1079         char *target;
1080         int num_threads = 0;
1081         int compress_level = 0;
1082         int create = 1;
1083         int ret;
1084         FILE *out;
1085
1086         while (1) {
1087                 int c = getopt(argc, argv, "rc:t:");
1088                 if (c < 0)
1089                         break;
1090                 switch (c) {
1091                 case 'r':
1092                         create = 0;
1093                         break;
1094                 case 't':
1095                         num_threads = atoi(optarg);
1096                         if (num_threads <= 0 || num_threads > 32)
1097                                 print_usage();
1098                         break;
1099                 case 'c':
1100                         compress_level = atoi(optarg);
1101                         if (compress_level < 0 || compress_level > 9)
1102                                 print_usage();
1103                         break;
1104                 default:
1105                         print_usage();
1106                 }
1107         }
1108
1109         argc = argc - optind;
1110         if (argc != 2)
1111                 print_usage();
1112         source = argv[optind];
1113         target = argv[optind + 1];
1114
1115         if (create && !strcmp(target, "-")) {
1116                 out = stdout;
1117         } else {
1118                 out = fopen(target, "w+");
1119                 if (!out) {
1120                         perror("unable to create target file");
1121                         exit(1);
1122                 }
1123         }
1124
1125         if (num_threads == 0 && compress_level > 0) {
1126                 num_threads = sysconf(_SC_NPROCESSORS_ONLN);
1127                 if (num_threads <= 0)
1128                         num_threads = 1;
1129         }
1130
1131         if (create)
1132                 ret = create_metadump(source, out, num_threads,
1133                                       compress_level);
1134         else
1135                 ret = restore_metadump(source, out, 1);
1136
1137         if (out == stdout)
1138                 fflush(out);
1139         else
1140                 fclose(out);
1141
1142         return ret;
1143 }