sbin/hammer/ondisk.c

   1 /*
   2  * Copyright (c) 2007 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  */
  34
  35 #include <sys/diskslice.h>
  36 #include <sys/diskmbr.h>
  37
  38 #include "hammer_util.h"
  39
  40 static void check_volume(struct volume_info *vol);
  41 static void get_buffer_readahead(struct buffer_info *base);
  42 static __inline int readhammervol(struct volume_info *vol);
  43 static __inline int readhammerbuf(struct buffer_info *buf);
  44 static __inline int writehammervol(struct volume_info *vol);
  45 static __inline int writehammerbuf(struct buffer_info *buf);
  46
  47 uuid_t Hammer_FSType;
  48 uuid_t Hammer_FSId;
  49 int UseReadBehind = -4;
  50 int UseReadAhead = 4;
  51 int DebugOpt;
  52
  53 TAILQ_HEAD(volume_list, volume_info);
  54 static struct volume_list VolList = TAILQ_HEAD_INITIALIZER(VolList);
  55 static int valid_hammer_volumes;
  56
  57 static __inline
  58 int
  59 buffer_hash(hammer_off_t buf_offset)
  60 {
  61         int hi;
  62
  63         hi = (int)(buf_offset / HAMMER_BUFSIZE) & HAMMER_BUFLISTMASK;
  64         return(hi);
  65 }
  66
  67 static struct buffer_info*
  68 find_buffer(hammer_off_t buf_offset)
  69 {
  70         struct volume_info *volume;
  71         struct buffer_info *buf;
  72         int hi;
  73
  74         volume = get_volume(HAMMER_VOL_DECODE(buf_offset));
  75         assert(volume);
  76
  77         hi = buffer_hash(buf_offset);
  78         TAILQ_FOREACH(buf, &volume->buffer_lists[hi], entry)
  79                 if (buf->buf_offset == buf_offset)
  80                         return(buf);
  81         return(NULL);
  82 }
  83
  84 static
  85 struct volume_info *
  86 __alloc_volume(const char *volname, int oflags)
  87 {
  88         struct volume_info *vol;
  89         int i;
  90
  91         vol = malloc(sizeof(*vol));
  92         if (vol == NULL)
  93                 err(1, "alloc_volume");
  94         bzero(vol, sizeof(*vol));
  95
  96         vol->vol_no = -1;
  97         vol->rdonly = (oflags == O_RDONLY);
  98         vol->name = strdup(volname);
  99         vol->fd = open(vol->name, oflags);
 100         if (vol->fd < 0)
 101                 err(1, "alloc_volume: Failed to open %s", vol->name);
 102         check_volume(vol);
 103
 104         vol->ondisk = malloc(HAMMER_BUFSIZE);
 105         if (vol->ondisk == NULL)
 106                 err(1, "alloc_volume");
 107         bzero(vol->ondisk, HAMMER_BUFSIZE);
 108
 109         for (i = 0; i < HAMMER_BUFLISTS; ++i)
 110                 TAILQ_INIT(&vol->buffer_lists[i]);
 111
 112         return(vol);
 113 }
 114
 115 static void
 116 __add_volume(struct volume_info *vol)
 117 {
 118         struct volume_info *scan;
 119         struct stat st1, st2;
 120
 121         if (fstat(vol->fd, &st1) != 0)
 122                 errx(1, "add_volume: %s: Failed to stat", vol->name);
 123
 124         TAILQ_FOREACH(scan, &VolList, entry) {
 125                 if (scan->vol_no == vol->vol_no) {
 126                         errx(1, "add_volume: %s: Duplicate volume number %d "
 127                                 "against %s",
 128                                 vol->name, vol->vol_no, scan->name);
 129                 }
 130                 if (fstat(scan->fd, &st2) != 0) {
 131                         errx(1, "add_volume: %s: Failed to stat %s",
 132                                 vol->name, scan->name);
 133                 }
 134                 if ((st1.st_ino == st2.st_ino) && (st1.st_dev == st2.st_dev)) {
 135                         errx(1, "add_volume: %s: Specified more than once",
 136                                 vol->name);
 137                 }
 138         }
 139
 140         TAILQ_INSERT_TAIL(&VolList, vol, entry);
 141 }
 142
 143 static void
 144 __verify_volume(struct volume_info *vol)
 145 {
 146         hammer_volume_ondisk_t ondisk = vol->ondisk;
 147
 148         if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
 149                 errx(1, "verify_volume: Invalid volume signature %016jx",
 150                         ondisk->vol_signature);
 151         }
 152         if (ondisk->vol_rootvol != HAMMER_ROOT_VOLNO) {
 153                 errx(1, "verify_volume: Invalid root volume# %d",
 154                         ondisk->vol_rootvol);
 155         }
 156         if (bcmp(&Hammer_FSType, &ondisk->vol_fstype, sizeof(Hammer_FSType))) {
 157                 errx(1, "verify_volume: %s: Header does not indicate "
 158                         "that this is a HAMMER volume", vol->name);
 159         }
 160         if (bcmp(&Hammer_FSId, &ondisk->vol_fsid, sizeof(Hammer_FSId))) {
 161                 errx(1, "verify_volume: %s: FSId does not match other volumes!",
 162                         vol->name);
 163         }
 164 }
 165
 166 /*
 167  * Initialize a volume structure and ondisk vol_no field.
 168  */
 169 struct volume_info *
 170 init_volume(const char *filename, int oflags, int32_t vol_no)
 171 {
 172         struct volume_info *vol;
 173
 174         vol = __alloc_volume(filename, oflags);
 175         vol->vol_no = vol->ondisk->vol_no = vol_no;
 176
 177         __add_volume(vol);
 178
 179         return(vol);
 180 }
 181
 182 /*
 183  * Initialize a volume structure and read ondisk volume header.
 184  */
 185 struct volume_info*
 186 load_volume(const char *filename, int oflags, int verify)
 187 {
 188         struct volume_info *vol;
 189         int n;
 190
 191         vol = __alloc_volume(filename, oflags);
 192
 193         n = readhammervol(vol);
 194         if (n == -1) {
 195                 err(1, "load_volume: %s: Read failed at offset 0", vol->name);
 196         }
 197         vol->vol_no = vol->ondisk->vol_no;
 198
 199         if (valid_hammer_volumes++ == 0)
 200                 Hammer_FSId = vol->ondisk->vol_fsid;
 201         if (verify)
 202                 __verify_volume(vol);
 203
 204         __add_volume(vol);
 205
 206         return(vol);
 207 }
 208
 209 /*
 210  * Check basic volume characteristics.
 211  */
 212 static void
 213 check_volume(struct volume_info *vol)
 214 {
 215         struct partinfo pinfo;
 216         struct stat st;
 217
 218         /*
 219          * Get basic information about the volume
 220          */
 221         if (ioctl(vol->fd, DIOCGPART, &pinfo) < 0) {
 222                 /*
 223                  * Allow the formatting of regular files as HAMMER volumes
 224                  */
 225                 if (fstat(vol->fd, &st) < 0)
 226                         err(1, "Unable to stat %s", vol->name);
 227                 vol->size = st.st_size;
 228                 vol->type = "REGFILE";
 229         } else {
 230                 /*
 231                  * When formatting a block device as a HAMMER volume the
 232                  * sector size must be compatible.  HAMMER uses 16384 byte
 233                  * filesystem buffers.
 234                  */
 235                 if (pinfo.reserved_blocks) {
 236                         errx(1, "HAMMER cannot be placed in a partition "
 237                                 "which overlaps the disklabel or MBR");
 238                 }
 239                 if (pinfo.media_blksize > HAMMER_BUFSIZE ||
 240                     HAMMER_BUFSIZE % pinfo.media_blksize) {
 241                         errx(1, "A media sector size of %d is not supported",
 242                              pinfo.media_blksize);
 243                 }
 244
 245                 vol->size = pinfo.media_size;
 246                 vol->device_offset = pinfo.media_offset;
 247                 vol->type = "DEVICE";
 248         }
 249 }
 250
 251 void
 252 assert_volume_offset(struct volume_info *vol)
 253 {
 254         assert(hammer_is_zone_raw_buffer(vol->vol_free_off));
 255         assert(hammer_is_zone_raw_buffer(vol->vol_free_end));
 256         if (vol->vol_free_off >= vol->vol_free_end)
 257                 errx(1, "Ran out of room, filesystem too small");
 258 }
 259
 260 struct volume_info *
 261 get_volume(int32_t vol_no)
 262 {
 263         struct volume_info *vol;
 264
 265         TAILQ_FOREACH(vol, &VolList, entry) {
 266                 if (vol->vol_no == vol_no)
 267                         break;
 268         }
 269
 270         return(vol);
 271 }
 272
 273 struct volume_info *
 274 get_root_volume(void)
 275 {
 276         return(get_volume(HAMMER_ROOT_VOLNO));
 277 }
 278
 279 static struct buffer_info *
 280 __alloc_buffer(hammer_off_t buf_offset, int isnew)
 281 {
 282         struct volume_info *volume;
 283         struct buffer_info *buf;
 284         int hi;
 285
 286         volume = get_volume(HAMMER_VOL_DECODE(buf_offset));
 287         assert(volume != NULL);
 288
 289         buf = calloc(sizeof(*buf), 1);
 290         buf->buf_offset = buf_offset;
 291         buf->raw_offset = hammer_xlate_to_phys(volume->ondisk, buf_offset);
 292         buf->volume = volume;
 293         buf->ondisk = calloc(HAMMER_BUFSIZE, 1);
 294
 295         if (isnew <= 0) {
 296                 if (readhammerbuf(buf) == -1) {
 297                         err(1, "Failed to read %s:%016jx at %016jx",
 298                             volume->name,
 299                             (intmax_t)buf->buf_offset,
 300                             (intmax_t)buf->raw_offset);
 301                 }
 302         }
 303
 304         hi = buffer_hash(buf_offset);
 305         TAILQ_INSERT_TAIL(&volume->buffer_lists[hi], buf, entry);
 306         hammer_cache_add(&buf->cache);
 307
 308         return(buf);
 309 }
 310
 311 /*
 312  * Acquire the specified buffer.  isnew is -1 only when called
 313  * via get_buffer_readahead() to prevent another readahead.
 314  */
 315 static struct buffer_info *
 316 get_buffer(hammer_off_t buf_offset, int isnew)
 317 {
 318         struct buffer_info *buf;
 319         int zone;
 320         int dora = 0;
 321         int error = 0;
 322
 323         zone = HAMMER_ZONE_DECODE(buf_offset);
 324         if (zone > HAMMER_ZONE_RAW_BUFFER_INDEX)
 325                 buf_offset = blockmap_lookup(buf_offset, NULL, NULL, &error);
 326         if (error || buf_offset == HAMMER_OFF_BAD)
 327                 return(NULL);
 328         assert(hammer_is_zone_raw_buffer(buf_offset));
 329
 330         buf_offset &= ~HAMMER_BUFMASK64;
 331         buf = find_buffer(buf_offset);
 332
 333         if (buf == NULL) {
 334                 buf = __alloc_buffer(buf_offset, isnew);
 335                 dora = (isnew == 0);
 336         } else {
 337                 assert(isnew != -1);
 338                 hammer_cache_used(&buf->cache);
 339         }
 340         assert(buf->ondisk != NULL);
 341
 342         ++buf->cache.refs;
 343         hammer_cache_flush();
 344
 345         if (isnew > 0) {
 346                 assert(buf->cache.modified == 0);
 347                 bzero(buf->ondisk, HAMMER_BUFSIZE);
 348                 buf->cache.modified = 1;
 349         }
 350         if (dora)
 351                 get_buffer_readahead(buf);
 352         return(buf);
 353 }
 354
 355 static void
 356 get_buffer_readahead(struct buffer_info *base)
 357 {
 358         struct buffer_info *buf;
 359         struct volume_info *vol;
 360         hammer_off_t buf_offset;
 361         int64_t raw_offset;
 362         int ri = UseReadBehind;
 363         int re = UseReadAhead;
 364
 365         raw_offset = base->raw_offset + ri * HAMMER_BUFSIZE;
 366         vol = base->volume;
 367
 368         while (ri < re) {
 369                 if (raw_offset >= vol->ondisk->vol_buf_end)
 370                         break;
 371                 if (raw_offset < vol->ondisk->vol_buf_beg || ri == 0) {
 372                         ++ri;
 373                         raw_offset += HAMMER_BUFSIZE;
 374                         continue;
 375                 }
 376                 buf_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no,
 377                         raw_offset - vol->ondisk->vol_buf_beg);
 378                 buf = find_buffer(buf_offset);
 379                 if (buf == NULL) {
 380                         buf = get_buffer(buf_offset, -1);
 381                         rel_buffer(buf);
 382                 }
 383                 ++ri;
 384                 raw_offset += HAMMER_BUFSIZE;
 385         }
 386 }
 387
 388 void
 389 rel_buffer(struct buffer_info *buffer)
 390 {
 391         struct volume_info *volume;
 392         int hi;
 393
 394         if (buffer == NULL)
 395                 return;
 396         assert(buffer->cache.refs > 0);
 397         if (--buffer->cache.refs == 0) {
 398                 if (buffer->cache.delete) {
 399                         hi = buffer_hash(buffer->buf_offset);
 400                         volume = buffer->volume;
 401                         if (buffer->cache.modified)
 402                                 flush_buffer(buffer);
 403                         TAILQ_REMOVE(&volume->buffer_lists[hi], buffer, entry);
 404                         hammer_cache_del(&buffer->cache);
 405                         free(buffer->ondisk);
 406                         free(buffer);
 407                 }
 408         }
 409 }
 410
 411 /*
 412  * Retrieve a pointer to a buffer data given a buffer offset.  The underlying
 413  * bufferp is freed if isnew or the offset is out of range of the cached data.
 414  * If bufferp is freed a referenced buffer is loaded into it.
 415  */
 416 void *
 417 get_buffer_data(hammer_off_t buf_offset, struct buffer_info **bufferp,
 418                 int isnew)
 419 {
 420         if (*bufferp != NULL) {
 421                 if (isnew > 0 ||
 422                     (((*bufferp)->buf_offset ^ buf_offset) & ~HAMMER_BUFMASK64)) {
 423                         rel_buffer(*bufferp);
 424                         *bufferp = NULL;
 425                 }
 426         }
 427
 428         if (*bufferp == NULL) {
 429                 *bufferp = get_buffer(buf_offset, isnew);
 430                 if (*bufferp == NULL)
 431                         return(NULL);
 432         }
 433
 434         return(((char *)(*bufferp)->ondisk) +
 435                 ((int32_t)buf_offset & HAMMER_BUFMASK));
 436 }
 437
 438 /*
 439  * Allocate HAMMER elements - B-Tree nodes
 440  */
 441 hammer_node_ondisk_t
 442 alloc_btree_node(hammer_off_t *offp, struct buffer_info **data_bufferp)
 443 {
 444         hammer_node_ondisk_t node;
 445
 446         node = alloc_blockmap(HAMMER_ZONE_BTREE_INDEX, sizeof(*node),
 447                               offp, data_bufferp);
 448         bzero(node, sizeof(*node));
 449         return(node);
 450 }
 451
 452 /*
 453  * Allocate HAMMER elements - meta data (inode, direntry, PFS, etc)
 454  */
 455 void *
 456 alloc_meta_element(hammer_off_t *offp, int32_t data_len,
 457                    struct buffer_info **data_bufferp)
 458 {
 459         void *data;
 460
 461         data = alloc_blockmap(HAMMER_ZONE_META_INDEX, data_len,
 462                               offp, data_bufferp);
 463         bzero(data, data_len);
 464         return(data);
 465 }
 466
 467 /*
 468  * Format a new blockmap.  This is mostly a degenerate case because
 469  * all allocations are now actually done from the freemap.
 470  */
 471 void
 472 format_blockmap(struct volume_info *root_vol, int zone, hammer_off_t offset)
 473 {
 474         hammer_blockmap_t blockmap;
 475         hammer_off_t zone_base;
 476
 477         /* Only root volume needs formatting */
 478         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 479
 480         assert(hammer_is_zone2_mapped_index(zone));
 481
 482         blockmap = &root_vol->ondisk->vol0_blockmap[zone];
 483         zone_base = HAMMER_ZONE_ENCODE(zone, offset);
 484
 485         bzero(blockmap, sizeof(*blockmap));
 486         blockmap->phys_offset = 0;
 487         blockmap->first_offset = zone_base;
 488         blockmap->next_offset = zone_base;
 489         blockmap->alloc_offset = HAMMER_ENCODE(zone, 255, -1);
 490         hammer_crc_set_blockmap(blockmap);
 491 }
 492
 493 /*
 494  * Format a new freemap.  Set all layer1 entries to UNAVAIL.  The initialize
 495  * code will load each volume's freemap.
 496  */
 497 void
 498 format_freemap(struct volume_info *root_vol)
 499 {
 500         struct buffer_info *buffer = NULL;
 501         hammer_off_t layer1_offset;
 502         hammer_blockmap_t blockmap;
 503         hammer_blockmap_layer1_t layer1;
 504         int i, isnew;
 505
 506         /* Only root volume needs formatting */
 507         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 508
 509         layer1_offset = bootstrap_bigblock(root_vol);
 510         for (i = 0; i < HAMMER_BIGBLOCK_SIZE; i += sizeof(*layer1)) {
 511                 isnew = ((i % HAMMER_BUFSIZE) == 0);
 512                 layer1 = get_buffer_data(layer1_offset + i, &buffer, isnew);
 513                 bzero(layer1, sizeof(*layer1));
 514                 layer1->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 515                 layer1->blocks_free = 0;
 516                 hammer_crc_set_layer1(layer1);
 517         }
 518         assert(i == HAMMER_BIGBLOCK_SIZE);
 519         rel_buffer(buffer);
 520
 521         blockmap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 522         bzero(blockmap, sizeof(*blockmap));
 523         blockmap->phys_offset = layer1_offset;
 524         blockmap->first_offset = 0;
 525         blockmap->next_offset = HAMMER_ENCODE_RAW_BUFFER(0, 0);
 526         blockmap->alloc_offset = HAMMER_ENCODE_RAW_BUFFER(255, -1);
 527         hammer_crc_set_blockmap(blockmap);
 528 }
 529
 530 /*
 531  * Load the volume's remaining free space into the freemap.
 532  *
 533  * Returns the number of big-blocks available.
 534  */
 535 int64_t
 536 initialize_freemap(struct volume_info *vol)
 537 {
 538         struct volume_info *root_vol;
 539         struct buffer_info *buffer1 = NULL;
 540         struct buffer_info *buffer2 = NULL;
 541         hammer_blockmap_layer1_t layer1;
 542         hammer_blockmap_layer2_t layer2;
 543         hammer_off_t layer1_offset;
 544         hammer_off_t layer2_offset;
 545         hammer_off_t phys_offset;
 546         hammer_off_t block_offset;
 547         hammer_off_t aligned_vol_free_end;
 548         hammer_blockmap_t freemap;
 549         int64_t count = 0;
 550         int64_t layer1_count = 0;
 551
 552         root_vol = get_root_volume();
 553
 554         assert_volume_offset(vol);
 555         aligned_vol_free_end = HAMMER_BLOCKMAP_LAYER2_DOALIGN(vol->vol_free_end);
 556
 557         printf("initialize freemap volume %d\n", vol->vol_no);
 558
 559         /*
 560          * Initialize the freemap.  First preallocate the big-blocks required
 561          * to implement layer2.   This preallocation is a bootstrap allocation
 562          * using blocks from the target volume.
 563          */
 564         freemap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 565
 566         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 567              phys_offset < aligned_vol_free_end;
 568              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 569                 layer1_offset = freemap->phys_offset +
 570                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 571                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 572                 if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 573                         layer1->phys_offset = bootstrap_bigblock(vol);
 574                         layer1->blocks_free = 0;
 575                         buffer1->cache.modified = 1;
 576                         hammer_crc_set_layer1(layer1);
 577                 }
 578         }
 579
 580         /*
 581          * Now fill everything in.
 582          */
 583         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 584              phys_offset < aligned_vol_free_end;
 585              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 586                 layer1_count = 0;
 587                 layer1_offset = freemap->phys_offset +
 588                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 589                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 590                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 591
 592                 for (block_offset = 0;
 593                      block_offset < HAMMER_BLOCKMAP_LAYER2;
 594                      block_offset += HAMMER_BIGBLOCK_SIZE) {
 595                         layer2_offset = layer1->phys_offset +
 596                                         HAMMER_BLOCKMAP_LAYER2_OFFSET(block_offset);
 597                         layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 598                         bzero(layer2, sizeof(*layer2));
 599
 600                         if (phys_offset + block_offset < vol->vol_free_off) {
 601                                 /*
 602                                  * Big-blocks already allocated as part
 603                                  * of the freemap bootstrap.
 604                                  */
 605                                 layer2->zone = HAMMER_ZONE_FREEMAP_INDEX;
 606                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 607                                 layer2->bytes_free = 0;
 608                         } else if (phys_offset + block_offset < vol->vol_free_end) {
 609                                 layer2->zone = 0;
 610                                 layer2->append_off = 0;
 611                                 layer2->bytes_free = HAMMER_BIGBLOCK_SIZE;
 612                                 ++count;
 613                                 ++layer1_count;
 614                         } else {
 615                                 layer2->zone = HAMMER_ZONE_UNAVAIL_INDEX;
 616                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 617                                 layer2->bytes_free = 0;
 618                         }
 619                         hammer_crc_set_layer2(layer2);
 620                         buffer2->cache.modified = 1;
 621                 }
 622
 623                 layer1->blocks_free += layer1_count;
 624                 hammer_crc_set_layer1(layer1);
 625                 buffer1->cache.modified = 1;
 626         }
 627
 628         rel_buffer(buffer1);
 629         rel_buffer(buffer2);
 630         return(count);
 631 }
 632
 633 /*
 634  * Returns the number of big-blocks available for filesystem data and undos
 635  * without formatting.
 636  */
 637 int64_t
 638 count_freemap(struct volume_info *vol)
 639 {
 640         hammer_off_t phys_offset;
 641         hammer_off_t vol_free_off;
 642         hammer_off_t aligned_vol_free_end;
 643         int64_t count = 0;
 644
 645         vol_free_off = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 646
 647         assert_volume_offset(vol);
 648         aligned_vol_free_end = HAMMER_BLOCKMAP_LAYER2_DOALIGN(vol->vol_free_end);
 649
 650         if (vol->vol_no == HAMMER_ROOT_VOLNO)
 651                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 652
 653         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 654              phys_offset < aligned_vol_free_end;
 655              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 656                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 657         }
 658
 659         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 660              phys_offset < aligned_vol_free_end;
 661              phys_offset += HAMMER_BIGBLOCK_SIZE) {
 662                 if (phys_offset < vol_free_off) {
 663                         ;
 664                 } else if (phys_offset < vol->vol_free_end) {
 665                         ++count;
 666                 }
 667         }
 668
 669         return(count);
 670 }
 671
 672 /*
 673  * Format the undomap for the root volume.
 674  */
 675 void
 676 format_undomap(struct volume_info *root_vol, int64_t *undo_buffer_size)
 677 {
 678         hammer_off_t undo_limit;
 679         hammer_blockmap_t blockmap;
 680         hammer_volume_ondisk_t ondisk;
 681         struct buffer_info *buffer = NULL;
 682         hammer_off_t scan;
 683         int n;
 684         int limit_index;
 685         uint32_t seqno;
 686
 687         /* Only root volume needs formatting */
 688         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 689         ondisk = root_vol->ondisk;
 690
 691         /*
 692          * Size the undo buffer in multiples of HAMMER_BIGBLOCK_SIZE,
 693          * up to HAMMER_MAX_UNDO_BIGBLOCKS big-blocks.
 694          * Size to approximately 0.1% of the disk.
 695          *
 696          * The minimum UNDO fifo size is 512MB, or approximately 1% of
 697          * the recommended 50G disk.
 698          *
 699          * Changing this minimum is rather dangerous as complex filesystem
 700          * operations can cause the UNDO FIFO to fill up otherwise.
 701          */
 702         undo_limit = *undo_buffer_size;
 703         if (undo_limit == 0) {
 704                 undo_limit = HAMMER_VOL_BUF_SIZE(ondisk) / 1000;
 705                 if (undo_limit < HAMMER_BIGBLOCK_SIZE * HAMMER_MIN_UNDO_BIGBLOCKS)
 706                         undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_MIN_UNDO_BIGBLOCKS;
 707         }
 708         undo_limit = HAMMER_BIGBLOCK_DOALIGN(undo_limit);
 709         if (undo_limit < HAMMER_BIGBLOCK_SIZE)
 710                 undo_limit = HAMMER_BIGBLOCK_SIZE;
 711         if (undo_limit > HAMMER_BIGBLOCK_SIZE * HAMMER_MAX_UNDO_BIGBLOCKS)
 712                 undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_MAX_UNDO_BIGBLOCKS;
 713         *undo_buffer_size = undo_limit;
 714
 715         blockmap = &ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 716         bzero(blockmap, sizeof(*blockmap));
 717         blockmap->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 718         blockmap->first_offset = HAMMER_ENCODE_UNDO(0);
 719         blockmap->next_offset = blockmap->first_offset;
 720         blockmap->alloc_offset = HAMMER_ENCODE_UNDO(undo_limit);
 721         hammer_crc_set_blockmap(blockmap);
 722
 723         limit_index = undo_limit / HAMMER_BIGBLOCK_SIZE;
 724         assert(limit_index <= HAMMER_MAX_UNDO_BIGBLOCKS);
 725
 726         for (n = 0; n < limit_index; ++n) {
 727                 ondisk->vol0_undo_array[n] = alloc_undo_bigblock(root_vol);
 728         }
 729         while (n < HAMMER_MAX_UNDO_BIGBLOCKS) {
 730                 ondisk->vol0_undo_array[n++] = HAMMER_BLOCKMAP_UNAVAIL;
 731         }
 732
 733         /*
 734          * Pre-initialize the UNDO blocks (HAMMER version 4+)
 735          */
 736         printf("initializing the undo map (%jd MB)\n",
 737                 (intmax_t)HAMMER_OFF_LONG_ENCODE(blockmap->alloc_offset) /
 738                 (1024 * 1024));
 739
 740         scan = blockmap->first_offset;
 741         seqno = 0;
 742
 743         while (scan < blockmap->alloc_offset) {
 744                 hammer_fifo_head_t head;
 745                 hammer_fifo_tail_t tail;
 746                 int isnew;
 747                 int bytes = HAMMER_UNDO_ALIGN;
 748
 749                 isnew = ((scan & HAMMER_BUFMASK64) == 0);
 750                 head = get_buffer_data(scan, &buffer, isnew);
 751                 buffer->cache.modified = 1;
 752                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 753
 754                 bzero(head, bytes);
 755                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 756                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
 757                 head->hdr_size = bytes;
 758                 head->hdr_seq = seqno++;
 759
 760                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 761                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 762                 tail->tail_size = bytes;
 763
 764                 hammer_crc_set_fifo_head(head, bytes);
 765
 766                 scan += bytes;
 767         }
 768         rel_buffer(buffer);
 769 }
 770
 771 const char *zone_labels[] = {
 772         "",             /* 0 */
 773         "raw_volume",   /* 1 */
 774         "raw_buffer",   /* 2 */
 775         "undo",         /* 3 */
 776         "freemap",      /* 4 */
 777         "",             /* 5 */
 778         "",             /* 6 */
 779         "",             /* 7 */
 780         "btree",        /* 8 */
 781         "meta",         /* 9 */
 782         "large_data",   /* 10 */
 783         "small_data",   /* 11 */
 784         "",             /* 12 */
 785         "",             /* 13 */
 786         "",             /* 14 */
 787         "unavail",      /* 15 */
 788 };
 789
 790 void
 791 print_blockmap(const struct volume_info *vol)
 792 {
 793         hammer_blockmap_t blockmap;
 794         hammer_volume_ondisk_t ondisk;
 795         int64_t size, used;
 796         int i;
 797 #define INDENT ""
 798
 799         ondisk = vol->ondisk;
 800         printf(INDENT"vol_label\t%s\n", ondisk->vol_label);
 801         printf(INDENT"vol_count\t%d\n", ondisk->vol_count);
 802         printf(INDENT"vol_bot_beg\t%s\n", sizetostr(ondisk->vol_bot_beg));
 803         printf(INDENT"vol_mem_beg\t%s\n", sizetostr(ondisk->vol_mem_beg));
 804         printf(INDENT"vol_buf_beg\t%s\n", sizetostr(ondisk->vol_buf_beg));
 805         printf(INDENT"vol_buf_end\t%s\n", sizetostr(ondisk->vol_buf_end));
 806         printf(INDENT"vol0_next_tid\t%016jx\n",
 807                (uintmax_t)ondisk->vol0_next_tid);
 808
 809         blockmap = &ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 810         size = HAMMER_OFF_LONG_ENCODE(blockmap->alloc_offset);
 811         if (blockmap->first_offset <= blockmap->next_offset)
 812                 used = blockmap->next_offset - blockmap->first_offset;
 813         else
 814                 used = blockmap->alloc_offset - blockmap->first_offset +
 815                         HAMMER_OFF_LONG_ENCODE(blockmap->next_offset);
 816         printf(INDENT"undo_size\t%s\n", sizetostr(size));
 817         printf(INDENT"undo_used\t%s\n", sizetostr(used));
 818
 819         printf(INDENT"zone #             "
 820                "phys             first            next             alloc\n");
 821         for (i = 0; i < HAMMER_MAX_ZONES; i++) {
 822                 blockmap = &ondisk->vol0_blockmap[i];
 823                 printf(INDENT"zone %-2d %-10s %016jx %016jx %016jx %016jx\n",
 824                         i, zone_labels[i],
 825                         (uintmax_t)blockmap->phys_offset,
 826                         (uintmax_t)blockmap->first_offset,
 827                         (uintmax_t)blockmap->next_offset,
 828                         (uintmax_t)blockmap->alloc_offset);
 829         }
 830 }
 831
 832 /*
 833  * Flush various tracking structures to disk
 834  */
 835 void
 836 flush_all_volumes(void)
 837 {
 838         struct volume_info *vol;
 839
 840         TAILQ_FOREACH(vol, &VolList, entry)
 841                 flush_volume(vol);
 842 }
 843
 844 void
 845 flush_volume(struct volume_info *volume)
 846 {
 847         struct buffer_info *buffer;
 848         int i;
 849
 850         for (i = 0; i < HAMMER_BUFLISTS; ++i) {
 851                 TAILQ_FOREACH(buffer, &volume->buffer_lists[i], entry)
 852                         flush_buffer(buffer);
 853         }
 854         if (writehammervol(volume) == -1)
 855                 err(1, "Write volume %d (%s)", volume->vol_no, volume->name);
 856 }
 857
 858 void
 859 flush_buffer(struct buffer_info *buffer)
 860 {
 861         struct volume_info *vol;
 862
 863         vol = buffer->volume;
 864         if (writehammerbuf(buffer) == -1)
 865                 err(1, "Write volume %d (%s)", vol->vol_no, vol->name);
 866         buffer->cache.modified = 0;
 867 }
 868
 869 /*
 870  * Core I/O operations
 871  */
 872 static int
 873 __read(struct volume_info *vol, void *data, int64_t offset, int size)
 874 {
 875         ssize_t n;
 876
 877         n = pread(vol->fd, data, size, offset);
 878         if (n != size)
 879                 return(-1);
 880         return(0);
 881 }
 882
 883 static __inline int
 884 readhammervol(struct volume_info *vol)
 885 {
 886         return(__read(vol, vol->ondisk, 0, HAMMER_BUFSIZE));
 887 }
 888
 889 static __inline int
 890 readhammerbuf(struct buffer_info *buf)
 891 {
 892         return(__read(buf->volume, buf->ondisk, buf->raw_offset, HAMMER_BUFSIZE));
 893 }
 894
 895 static int
 896 __write(struct volume_info *vol, const void *data, int64_t offset, int size)
 897 {
 898         ssize_t n;
 899
 900         if (vol->rdonly)
 901                 return(0);
 902
 903         n = pwrite(vol->fd, data, size, offset);
 904         if (n != size)
 905                 return(-1);
 906         return(0);
 907 }
 908
 909 static __inline int
 910 writehammervol(struct volume_info *vol)
 911 {
 912         return(__write(vol, vol->ondisk, 0, HAMMER_BUFSIZE));
 913 }
 914
 915 static __inline int
 916 writehammerbuf(struct buffer_info *buf)
 917 {
 918         return(__write(buf->volume, buf->ondisk, buf->raw_offset, HAMMER_BUFSIZE));
 919 }
 920
 921 int64_t init_boot_area_size(int64_t value, off_t avg_vol_size)
 922 {
 923         if (value == 0) {
 924                 value = HAMMER_BOOT_NOMBYTES;
 925                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 926                         value >>= 1;
 927         }
 928
 929         if (value < HAMMER_BOOT_MINBYTES) {
 930                 value = HAMMER_BOOT_MINBYTES;
 931         } else if (value > HAMMER_BOOT_MAXBYTES) {
 932                 value = HAMMER_BOOT_MAXBYTES;
 933         }
 934
 935         return(value);
 936 }
 937
 938 int64_t init_memory_log_size(int64_t value, off_t avg_vol_size)
 939 {
 940         if (value == 0) {
 941                 value = HAMMER_MEM_NOMBYTES;
 942                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 943                         value >>= 1;
 944         }
 945
 946         if (value < HAMMER_MEM_MINBYTES) {
 947                 value = HAMMER_MEM_MINBYTES;
 948         } else if (value > HAMMER_MEM_MAXBYTES) {
 949                 value = HAMMER_MEM_MAXBYTES;
 950         }
 951
 952         return(value);
 953 }