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(volume_info_t volume);
  41 static void get_buffer_readahead(buffer_info_t base);
  42 static __inline int readhammervol(volume_info_t volume);
  43 static __inline int readhammerbuf(buffer_info_t buffer);
  44 static __inline int writehammervol(volume_info_t volume);
  45 static __inline int writehammerbuf(buffer_info_t buffer);
  46
  47 uuid_t Hammer_FSType;
  48 uuid_t Hammer_FSId;
  49 int UseReadBehind = -4;
  50 int UseReadAhead = 4;
  51 int DebugOpt;
  52 uint32_t HammerVersion = -1;
  53
  54 TAILQ_HEAD(volume_list, volume_info);
  55 static struct volume_list VolList = TAILQ_HEAD_INITIALIZER(VolList);
  56 static int valid_hammer_volumes;
  57
  58 static __inline
  59 int
  60 buffer_hash(hammer_off_t zone2_offset)
  61 {
  62         int hi;
  63
  64         hi = (int)(zone2_offset / HAMMER_BUFSIZE) & HAMMER_BUFLISTMASK;
  65         return(hi);
  66 }
  67
  68 static buffer_info_t
  69 find_buffer(hammer_off_t zone2_offset)
  70 {
  71         volume_info_t volume;
  72         buffer_info_t buffer;
  73         int hi;
  74
  75         volume = get_volume(HAMMER_VOL_DECODE(zone2_offset));
  76         assert(volume);
  77
  78         hi = buffer_hash(zone2_offset);
  79         TAILQ_FOREACH(buffer, &volume->buffer_lists[hi], entry)
  80                 if (buffer->zone2_offset == zone2_offset)
  81                         return(buffer);
  82         return(NULL);
  83 }
  84
  85 static
  86 volume_info_t
  87 __alloc_volume(const char *volname, int oflags)
  88 {
  89         volume_info_t volume;
  90         int i;
  91
  92         volume = calloc(1, sizeof(*volume));
  93         volume->vol_no = -1;
  94         volume->rdonly = (oflags == O_RDONLY);
  95         volume->name = strdup(volname);
  96         volume->fd = open(volume->name, oflags);
  97         if (volume->fd < 0) {
  98                 err(1, "alloc_volume: Failed to open %s", volume->name);
  99                 /* not reached */
 100         }
 101         check_volume(volume);
 102
 103         volume->ondisk = calloc(1, HAMMER_BUFSIZE);
 104
 105         for (i = 0; i < HAMMER_BUFLISTS; ++i)
 106                 TAILQ_INIT(&volume->buffer_lists[i]);
 107
 108         return(volume);
 109 }
 110
 111 static void
 112 __add_volume(const volume_info_t volume)
 113 {
 114         volume_info_t scan;
 115         struct stat st1, st2;
 116
 117         if (fstat(volume->fd, &st1) != 0) {
 118                 errx(1, "add_volume: %s: Failed to stat", volume->name);
 119                 /* not reached */
 120         }
 121
 122         TAILQ_FOREACH(scan, &VolList, entry) {
 123                 if (scan->vol_no == volume->vol_no) {
 124                         errx(1, "add_volume: %s: Duplicate volume number %d "
 125                                 "against %s",
 126                                 volume->name, volume->vol_no, scan->name);
 127                         /* not reached */
 128                 }
 129                 if (fstat(scan->fd, &st2) != 0) {
 130                         errx(1, "add_volume: %s: Failed to stat %s",
 131                                 volume->name, scan->name);
 132                         /* not reached */
 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                                 volume->name);
 137                         /* not reached */
 138                 }
 139         }
 140
 141         TAILQ_INSERT_TAIL(&VolList, volume, entry);
 142 }
 143
 144 static void
 145 __verify_volume(const volume_info_t volume)
 146 {
 147         hammer_volume_ondisk_t ondisk = volume->ondisk;
 148
 149         if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
 150                 errx(1, "verify_volume: Invalid volume signature %016jx",
 151                         ondisk->vol_signature);
 152                 /* not reached */
 153         }
 154         if (ondisk->vol_rootvol != HAMMER_ROOT_VOLNO) {
 155                 errx(1, "verify_volume: Invalid root volume# %d",
 156                         ondisk->vol_rootvol);
 157                 /* not reached */
 158         }
 159         if (bcmp(&Hammer_FSType, &ondisk->vol_fstype, sizeof(Hammer_FSType))) {
 160                 errx(1, "verify_volume: %s: Header does not indicate "
 161                         "that this is a HAMMER volume", volume->name);
 162                 /* not reached */
 163         }
 164         if (bcmp(&Hammer_FSId, &ondisk->vol_fsid, sizeof(Hammer_FSId))) {
 165                 errx(1, "verify_volume: %s: FSId does not match other volumes!",
 166                         volume->name);
 167                 /* not reached */
 168         }
 169         if (ondisk->vol_version < HAMMER_VOL_VERSION_MIN ||
 170             ondisk->vol_version >= HAMMER_VOL_VERSION_WIP) {
 171                 errx(1, "verify_volume: %s: Invalid volume version %u",
 172                         volume->name, ondisk->vol_version);
 173                 /* not reached */
 174         }
 175 }
 176
 177 /*
 178  * Initialize a volume structure and ondisk vol_no field.
 179  */
 180 volume_info_t
 181 init_volume(const char *filename, int oflags, int32_t vol_no)
 182 {
 183         volume_info_t volume;
 184
 185         volume = __alloc_volume(filename, oflags);
 186         volume->vol_no = volume->ondisk->vol_no = vol_no;
 187
 188         __add_volume(volume);
 189
 190         return(volume);
 191 }
 192
 193 /*
 194  * Initialize a volume structure and read ondisk volume header.
 195  */
 196 volume_info_t
 197 load_volume(const char *filename, int oflags, int verify_volume)
 198 {
 199         volume_info_t volume;
 200         int n;
 201
 202         volume = __alloc_volume(filename, oflags);
 203
 204         n = readhammervol(volume);
 205         if (n == -1) {
 206                 err(1, "load_volume: %s: Read failed at offset 0",
 207                     volume->name);
 208                 /* not reached */
 209         }
 210         volume->vol_no = volume->ondisk->vol_no;
 211         HammerVersion = volume->ondisk->vol_version;
 212
 213         if (valid_hammer_volumes++ == 0)
 214                 Hammer_FSId = volume->ondisk->vol_fsid;
 215         if (verify_volume)
 216                 __verify_volume(volume);
 217
 218         __add_volume(volume);
 219
 220         return(volume);
 221 }
 222
 223 /*
 224  * Check basic volume characteristics.
 225  */
 226 static void
 227 check_volume(volume_info_t volume)
 228 {
 229         struct partinfo pinfo;
 230         struct stat st;
 231
 232         /*
 233          * Allow the formatting of block devices or regular files
 234          */
 235         if (ioctl(volume->fd, DIOCGPART, &pinfo) < 0) {
 236                 if (fstat(volume->fd, &st) < 0) {
 237                         err(1, "Unable to stat %s", volume->name);
 238                         /* not reached */
 239                 }
 240                 if (S_ISREG(st.st_mode)) {
 241                         volume->size = st.st_size;
 242                         volume->type = "REGFILE";
 243                 } else {
 244                         errx(1, "Unsupported file type for %s", volume->name);
 245                         /* not reached */
 246                 }
 247         } else {
 248                 /*
 249                  * When formatting a block device as a HAMMER volume the
 250                  * sector size must be compatible.  HAMMER uses 16384 byte
 251                  * filesystem buffers.
 252                  */
 253                 if (pinfo.reserved_blocks) {
 254                         errx(1, "HAMMER cannot be placed in a partition "
 255                                 "which overlaps the disklabel or MBR");
 256                         /* not reached */
 257                 }
 258                 if (pinfo.media_blksize > HAMMER_BUFSIZE ||
 259                     HAMMER_BUFSIZE % pinfo.media_blksize) {
 260                         errx(1, "A media sector size of %d is not supported",
 261                              pinfo.media_blksize);
 262                         /* not reached */
 263                 }
 264
 265                 volume->size = pinfo.media_size;
 266                 volume->device_offset = pinfo.media_offset;
 267                 volume->type = "DEVICE";
 268         }
 269 }
 270
 271 int
 272 is_regfile(const volume_info_t volume)
 273 {
 274         return(strcmp(volume->type, "REGFILE") ? 0 : 1);
 275 }
 276
 277 void
 278 assert_volume_offset(const volume_info_t volume)
 279 {
 280         assert(hammer_is_zone_raw_buffer(volume->vol_free_off));
 281         assert(hammer_is_zone_raw_buffer(volume->vol_free_end));
 282         if (volume->vol_free_off >= volume->vol_free_end) {
 283                 errx(1, "Ran out of room, filesystem too small");
 284                 /* not reached */
 285         }
 286 }
 287
 288 volume_info_t
 289 get_volume(int32_t vol_no)
 290 {
 291         volume_info_t volume;
 292
 293         TAILQ_FOREACH(volume, &VolList, entry) {
 294                 if (volume->vol_no == vol_no)
 295                         break;
 296         }
 297
 298         return(volume);
 299 }
 300
 301 volume_info_t
 302 get_root_volume(void)
 303 {
 304         return(get_volume(HAMMER_ROOT_VOLNO));
 305 }
 306
 307 static hammer_off_t
 308 __blockmap_xlate_to_zone2(hammer_off_t buf_offset)
 309 {
 310         hammer_off_t zone2_offset;
 311         int error = 0;
 312
 313         if (hammer_is_zone_raw_buffer(buf_offset))
 314                 zone2_offset = buf_offset;
 315         else
 316                 zone2_offset = blockmap_lookup(buf_offset, &error);
 317
 318         if (error)
 319                 return(HAMMER_OFF_BAD);
 320         assert(hammer_is_zone_raw_buffer(zone2_offset));
 321
 322         return(zone2_offset);
 323 }
 324
 325 static buffer_info_t
 326 __alloc_buffer(hammer_off_t zone2_offset, int isnew)
 327 {
 328         volume_info_t volume;
 329         buffer_info_t buffer;
 330         int hi;
 331
 332         volume = get_volume(HAMMER_VOL_DECODE(zone2_offset));
 333         assert(volume != NULL);
 334
 335         buffer = calloc(1, sizeof(*buffer));
 336         buffer->zone2_offset = zone2_offset;
 337         buffer->raw_offset = hammer_xlate_to_phys(volume->ondisk, zone2_offset);
 338         buffer->volume = volume;
 339         buffer->ondisk = calloc(1, HAMMER_BUFSIZE);
 340
 341         if (isnew <= 0) {
 342                 if (readhammerbuf(buffer) == -1) {
 343                         err(1, "Failed to read %s:%016jx at %016jx",
 344                             volume->name,
 345                             (intmax_t)buffer->zone2_offset,
 346                             (intmax_t)buffer->raw_offset);
 347                         /* not reached */
 348                 }
 349         }
 350
 351         hi = buffer_hash(zone2_offset);
 352         TAILQ_INSERT_TAIL(&volume->buffer_lists[hi], buffer, entry);
 353         hammer_cache_add(&buffer->cache);
 354
 355         return(buffer);
 356 }
 357
 358 /*
 359  * Acquire the 16KB buffer for specified zone offset.
 360  */
 361 static buffer_info_t
 362 get_buffer(hammer_off_t buf_offset, int isnew)
 363 {
 364         buffer_info_t buffer;
 365         hammer_off_t zone2_offset;
 366         int dora = 0;
 367
 368         zone2_offset = __blockmap_xlate_to_zone2(buf_offset);
 369         if (zone2_offset == HAMMER_OFF_BAD)
 370                 return(NULL);
 371
 372         zone2_offset &= ~HAMMER_BUFMASK64;
 373         buffer = find_buffer(zone2_offset);
 374
 375         if (buffer == NULL) {
 376                 buffer = __alloc_buffer(zone2_offset, isnew);
 377                 dora = (isnew == 0);
 378         } else {
 379                 assert(isnew != -1);
 380                 hammer_cache_used(&buffer->cache);
 381         }
 382         assert(buffer->ondisk != NULL);
 383
 384         ++buffer->cache.refs;
 385         hammer_cache_flush();
 386
 387         if (isnew > 0) {
 388                 assert(buffer->cache.modified == 0);
 389                 bzero(buffer->ondisk, HAMMER_BUFSIZE);
 390                 buffer->cache.modified = 1;
 391         }
 392         if (dora)
 393                 get_buffer_readahead(buffer);
 394         return(buffer);
 395 }
 396
 397 static void
 398 get_buffer_readahead(const buffer_info_t base)
 399 {
 400         buffer_info_t buffer;
 401         volume_info_t volume;
 402         hammer_off_t zone2_offset;
 403         int64_t raw_offset;
 404         int ri = UseReadBehind;
 405         int re = UseReadAhead;
 406
 407         raw_offset = base->raw_offset + ri * HAMMER_BUFSIZE;
 408         volume = base->volume;
 409
 410         while (ri < re) {
 411                 if (raw_offset >= volume->ondisk->vol_buf_end)
 412                         break;
 413                 if (raw_offset < volume->ondisk->vol_buf_beg || ri == 0) {
 414                         ++ri;
 415                         raw_offset += HAMMER_BUFSIZE;
 416                         continue;
 417                 }
 418                 zone2_offset = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no,
 419                         raw_offset - volume->ondisk->vol_buf_beg);
 420                 buffer = find_buffer(zone2_offset);
 421                 if (buffer == NULL) {
 422                         /* call with -1 to prevent another readahead */
 423                         buffer = get_buffer(zone2_offset, -1);
 424                         rel_buffer(buffer);
 425                 }
 426                 ++ri;
 427                 raw_offset += HAMMER_BUFSIZE;
 428         }
 429 }
 430
 431 void
 432 rel_buffer(buffer_info_t buffer)
 433 {
 434         volume_info_t volume;
 435         int hi;
 436
 437         if (buffer == NULL)
 438                 return;
 439         assert(buffer->cache.refs > 0);
 440         if (--buffer->cache.refs == 0) {
 441                 if (buffer->cache.delete) {
 442                         hi = buffer_hash(buffer->zone2_offset);
 443                         volume = buffer->volume;
 444                         if (buffer->cache.modified)
 445                                 flush_buffer(buffer);
 446                         TAILQ_REMOVE(&volume->buffer_lists[hi], buffer, entry);
 447                         hammer_cache_del(&buffer->cache);
 448                         free(buffer->ondisk);
 449                         free(buffer);
 450                 }
 451         }
 452 }
 453
 454 /*
 455  * Retrieve a pointer to a buffer data given a buffer offset.  The underlying
 456  * bufferp is freed if isnew or the offset is out of range of the cached data.
 457  * If bufferp is freed a referenced buffer is loaded into it.
 458  */
 459 void *
 460 get_buffer_data(hammer_off_t buf_offset, buffer_info_t *bufferp, int isnew)
 461 {
 462         hammer_off_t xor;
 463
 464         if (*bufferp != NULL) {
 465                 /* XXX xor is always non zero for indirect zones */
 466                 xor = HAMMER_OFF_LONG_ENCODE(buf_offset) ^
 467                       HAMMER_OFF_LONG_ENCODE((*bufferp)->zone2_offset);
 468                 if (isnew > 0 || (xor & ~HAMMER_BUFMASK64)) {
 469                         rel_buffer(*bufferp);
 470                         *bufferp = NULL;
 471                 }
 472         }
 473
 474         if (*bufferp == NULL) {
 475                 *bufferp = get_buffer(buf_offset, isnew);
 476                 if (*bufferp == NULL)
 477                         return(NULL);
 478         }
 479
 480         return(((char *)(*bufferp)->ondisk) +
 481                 ((int32_t)buf_offset & HAMMER_BUFMASK));
 482 }
 483
 484 /*
 485  * Allocate HAMMER elements - B-Tree nodes
 486  */
 487 hammer_node_ondisk_t
 488 alloc_btree_node(hammer_off_t *offp, buffer_info_t *data_bufferp)
 489 {
 490         hammer_node_ondisk_t node;
 491
 492         node = alloc_blockmap(HAMMER_ZONE_BTREE_INDEX, sizeof(*node),
 493                               offp, data_bufferp);
 494         bzero(node, sizeof(*node));
 495         return(node);
 496 }
 497
 498 /*
 499  * Allocate HAMMER elements - meta data (inode, direntry, PFS, etc)
 500  */
 501 void *
 502 alloc_meta_element(hammer_off_t *offp, int32_t data_len,
 503                    buffer_info_t *data_bufferp)
 504 {
 505         void *data;
 506
 507         data = alloc_blockmap(HAMMER_ZONE_META_INDEX, data_len,
 508                               offp, data_bufferp);
 509         bzero(data, data_len);
 510         return(data);
 511 }
 512
 513 /*
 514  * Format a new blockmap.  This is mostly a degenerate case because
 515  * all allocations are now actually done from the freemap.
 516  */
 517 void
 518 format_blockmap(volume_info_t root_vol, int zone, hammer_off_t offset)
 519 {
 520         hammer_blockmap_t blockmap;
 521         hammer_off_t zone_base;
 522
 523         /* Only root volume needs formatting */
 524         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 525
 526         assert(hammer_is_index_record(zone));
 527
 528         blockmap = &root_vol->ondisk->vol0_blockmap[zone];
 529         zone_base = HAMMER_ZONE_ENCODE(zone, offset);
 530
 531         bzero(blockmap, sizeof(*blockmap));
 532         blockmap->phys_offset = 0;
 533         blockmap->first_offset = zone_base;
 534         blockmap->next_offset = zone_base;
 535         blockmap->alloc_offset = HAMMER_ENCODE(zone, 255, -1);
 536         hammer_crc_set_blockmap(HammerVersion, blockmap);
 537 }
 538
 539 /*
 540  * Format a new freemap.  Set all layer1 entries to UNAVAIL.  The initialize
 541  * code will load each volume's freemap.
 542  */
 543 void
 544 format_freemap(volume_info_t root_vol)
 545 {
 546         buffer_info_t buffer = NULL;
 547         hammer_off_t layer1_offset;
 548         hammer_blockmap_t blockmap;
 549         hammer_blockmap_layer1_t layer1;
 550         int i, isnew;
 551
 552         /* Only root volume needs formatting */
 553         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 554
 555         layer1_offset = bootstrap_bigblock(root_vol);
 556         for (i = 0; i < HAMMER_BIGBLOCK_SIZE; i += sizeof(*layer1)) {
 557                 isnew = ((i % HAMMER_BUFSIZE) == 0);
 558                 layer1 = get_buffer_data(layer1_offset + i, &buffer, isnew);
 559                 bzero(layer1, sizeof(*layer1));
 560                 layer1->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 561                 layer1->blocks_free = 0;
 562                 hammer_crc_set_layer1(HammerVersion, layer1);
 563         }
 564         assert(i == HAMMER_BIGBLOCK_SIZE);
 565         rel_buffer(buffer);
 566
 567         blockmap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 568         bzero(blockmap, sizeof(*blockmap));
 569         blockmap->phys_offset = layer1_offset;
 570         blockmap->first_offset = 0;
 571         blockmap->next_offset = HAMMER_ENCODE_RAW_BUFFER(0, 0);
 572         blockmap->alloc_offset = HAMMER_ENCODE_RAW_BUFFER(255, -1);
 573         hammer_crc_set_blockmap(HammerVersion, blockmap);
 574 }
 575
 576 /*
 577  * Load the volume's remaining free space into the freemap.
 578  *
 579  * Returns the number of big-blocks available.
 580  */
 581 int64_t
 582 initialize_freemap(volume_info_t volume)
 583 {
 584         volume_info_t root_vol;
 585         buffer_info_t buffer1 = NULL;
 586         buffer_info_t buffer2 = NULL;
 587         hammer_blockmap_layer1_t layer1;
 588         hammer_blockmap_layer2_t layer2;
 589         hammer_off_t layer1_offset;
 590         hammer_off_t layer2_offset;
 591         hammer_off_t phys_offset;
 592         hammer_off_t block_offset;
 593         hammer_off_t aligned_vol_free_end;
 594         hammer_blockmap_t freemap;
 595         int64_t count = 0;
 596         int64_t layer1_count = 0;
 597
 598         root_vol = get_root_volume();
 599
 600         assert_volume_offset(volume);
 601         aligned_vol_free_end = HAMMER_BLOCKMAP_LAYER2_DOALIGN(volume->vol_free_end);
 602
 603         printf("initialize freemap volume %d\n", volume->vol_no);
 604
 605         /*
 606          * Initialize the freemap.  First preallocate the big-blocks required
 607          * to implement layer2.   This preallocation is a bootstrap allocation
 608          * using blocks from the target volume.
 609          */
 610         freemap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 611
 612         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no, 0);
 613              phys_offset < aligned_vol_free_end;
 614              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 615                 layer1_offset = freemap->phys_offset +
 616                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 617                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 618                 if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 619                         layer1->phys_offset = bootstrap_bigblock(volume);
 620                         layer1->blocks_free = 0;
 621                         buffer1->cache.modified = 1;
 622                         hammer_crc_set_layer1(HammerVersion, layer1);
 623                 }
 624         }
 625
 626         /*
 627          * Now fill everything in.
 628          */
 629         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no, 0);
 630              phys_offset < aligned_vol_free_end;
 631              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 632                 layer1_count = 0;
 633                 layer1_offset = freemap->phys_offset +
 634                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 635                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 636                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 637
 638                 for (block_offset = 0;
 639                      block_offset < HAMMER_BLOCKMAP_LAYER2;
 640                      block_offset += HAMMER_BIGBLOCK_SIZE) {
 641                         layer2_offset = layer1->phys_offset +
 642                                         HAMMER_BLOCKMAP_LAYER2_OFFSET(block_offset);
 643                         layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 644                         bzero(layer2, sizeof(*layer2));
 645
 646                         if (phys_offset + block_offset < volume->vol_free_off) {
 647                                 /*
 648                                  * Big-blocks already allocated as part
 649                                  * of the freemap bootstrap.
 650                                  */
 651                                 layer2->zone = HAMMER_ZONE_FREEMAP_INDEX;
 652                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 653                                 layer2->bytes_free = 0;
 654                         } else if (phys_offset + block_offset < volume->vol_free_end) {
 655                                 layer2->zone = 0;
 656                                 layer2->append_off = 0;
 657                                 layer2->bytes_free = HAMMER_BIGBLOCK_SIZE;
 658                                 ++count;
 659                                 ++layer1_count;
 660                         } else {
 661                                 layer2->zone = HAMMER_ZONE_UNAVAIL_INDEX;
 662                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 663                                 layer2->bytes_free = 0;
 664                         }
 665                         hammer_crc_set_layer2(HammerVersion, layer2);
 666                         buffer2->cache.modified = 1;
 667                 }
 668
 669                 layer1->blocks_free += layer1_count;
 670                 hammer_crc_set_layer1(HammerVersion, layer1);
 671                 buffer1->cache.modified = 1;
 672         }
 673
 674         rel_buffer(buffer1);
 675         rel_buffer(buffer2);
 676         return(count);
 677 }
 678
 679 /*
 680  * Returns the number of big-blocks available for filesystem data and undos
 681  * without formatting.
 682  */
 683 int64_t
 684 count_freemap(const volume_info_t volume)
 685 {
 686         hammer_off_t phys_offset;
 687         hammer_off_t vol_free_off;
 688         hammer_off_t aligned_vol_free_end;
 689         int64_t count = 0;
 690
 691         vol_free_off = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no, 0);
 692
 693         assert_volume_offset(volume);
 694         aligned_vol_free_end = HAMMER_BLOCKMAP_LAYER2_DOALIGN(volume->vol_free_end);
 695
 696         if (volume->vol_no == HAMMER_ROOT_VOLNO)
 697                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 698
 699         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no, 0);
 700              phys_offset < aligned_vol_free_end;
 701              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 702                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 703         }
 704
 705         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no, 0);
 706              phys_offset < aligned_vol_free_end;
 707              phys_offset += HAMMER_BIGBLOCK_SIZE) {
 708                 if (phys_offset < vol_free_off)
 709                         ;
 710                 else if (phys_offset < volume->vol_free_end)
 711                         ++count;
 712         }
 713
 714         return(count);
 715 }
 716
 717 /*
 718  * Format the undomap for the root volume.
 719  */
 720 void
 721 format_undomap(volume_info_t root_vol, int64_t *undo_buffer_size)
 722 {
 723         hammer_off_t undo_limit;
 724         hammer_blockmap_t blockmap;
 725         hammer_volume_ondisk_t ondisk;
 726         buffer_info_t buffer = NULL;
 727         hammer_off_t scan;
 728         int n;
 729         int limit_index;
 730         uint32_t seqno;
 731
 732         /* Only root volume needs formatting */
 733         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 734         ondisk = root_vol->ondisk;
 735
 736         /*
 737          * Size the undo buffer in multiples of HAMMER_BIGBLOCK_SIZE,
 738          * up to HAMMER_MAX_UNDO_BIGBLOCKS big-blocks.
 739          * Size to approximately 0.1% of the disk.
 740          *
 741          * The minimum UNDO fifo size is 512MB, or approximately 1% of
 742          * the recommended 50G disk.
 743          *
 744          * Changing this minimum is rather dangerous as complex filesystem
 745          * operations can cause the UNDO FIFO to fill up otherwise.
 746          */
 747         undo_limit = *undo_buffer_size;
 748         if (undo_limit == 0) {
 749                 undo_limit = HAMMER_VOL_BUF_SIZE(ondisk) / 1000;
 750                 if (undo_limit < HAMMER_BIGBLOCK_SIZE * HAMMER_MIN_UNDO_BIGBLOCKS)
 751                         undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_MIN_UNDO_BIGBLOCKS;
 752         }
 753         undo_limit = HAMMER_BIGBLOCK_DOALIGN(undo_limit);
 754         if (undo_limit < HAMMER_BIGBLOCK_SIZE)
 755                 undo_limit = HAMMER_BIGBLOCK_SIZE;
 756         if (undo_limit > HAMMER_BIGBLOCK_SIZE * HAMMER_MAX_UNDO_BIGBLOCKS)
 757                 undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_MAX_UNDO_BIGBLOCKS;
 758         *undo_buffer_size = undo_limit;
 759
 760         blockmap = &ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 761         bzero(blockmap, sizeof(*blockmap));
 762         blockmap->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 763         blockmap->first_offset = HAMMER_ENCODE_UNDO(0);
 764         blockmap->next_offset = blockmap->first_offset;
 765         blockmap->alloc_offset = HAMMER_ENCODE_UNDO(undo_limit);
 766         hammer_crc_set_blockmap(HammerVersion, blockmap);
 767
 768         limit_index = undo_limit / HAMMER_BIGBLOCK_SIZE;
 769         assert(limit_index <= HAMMER_MAX_UNDO_BIGBLOCKS);
 770
 771         for (n = 0; n < limit_index; ++n)
 772                 ondisk->vol0_undo_array[n] = alloc_undo_bigblock(root_vol);
 773         while (n < HAMMER_MAX_UNDO_BIGBLOCKS)
 774                 ondisk->vol0_undo_array[n++] = HAMMER_BLOCKMAP_UNAVAIL;
 775
 776         /*
 777          * Pre-initialize the UNDO blocks (HAMMER version 4+)
 778          */
 779         printf("initializing the undo map (%jd MB)\n",
 780                 (intmax_t)HAMMER_OFF_LONG_ENCODE(blockmap->alloc_offset) /
 781                 (1024 * 1024));
 782
 783         scan = blockmap->first_offset;
 784         seqno = 0;
 785
 786         while (scan < blockmap->alloc_offset) {
 787                 hammer_fifo_head_t head;
 788                 hammer_fifo_tail_t tail;
 789                 int isnew;
 790                 int bytes = HAMMER_UNDO_ALIGN;
 791
 792                 isnew = ((scan & HAMMER_BUFMASK64) == 0);
 793                 head = get_buffer_data(scan, &buffer, isnew);
 794                 buffer->cache.modified = 1;
 795                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 796
 797                 bzero(head, bytes);
 798                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 799                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
 800                 head->hdr_size = bytes;
 801                 head->hdr_seq = seqno++;
 802
 803                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 804                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 805                 tail->tail_size = bytes;
 806
 807                 hammer_crc_set_fifo_head(HammerVersion, head, bytes);
 808
 809                 scan += bytes;
 810         }
 811         rel_buffer(buffer);
 812 }
 813
 814 const char *zone_labels[] = {
 815         "",             /* 0 */
 816         "raw_volume",   /* 1 */
 817         "raw_buffer",   /* 2 */
 818         "undo",         /* 3 */
 819         "freemap",      /* 4 */
 820         "",             /* 5 */
 821         "",             /* 6 */
 822         "",             /* 7 */
 823         "btree",        /* 8 */
 824         "meta",         /* 9 */
 825         "large_data",   /* 10 */
 826         "small_data",   /* 11 */
 827         "",             /* 12 */
 828         "",             /* 13 */
 829         "",             /* 14 */
 830         "unavail",      /* 15 */
 831 };
 832
 833 void
 834 print_blockmap(const volume_info_t volume)
 835 {
 836         hammer_blockmap_t blockmap;
 837         hammer_volume_ondisk_t ondisk;
 838         int64_t size, used;
 839         int i;
 840 #define INDENT ""
 841
 842         ondisk = volume->ondisk;
 843         printf(INDENT"vol_label\t%s\n", ondisk->vol_label);
 844         printf(INDENT"vol_count\t%d\n", ondisk->vol_count);
 845         printf(INDENT"vol_bot_beg\t%s\n", sizetostr(ondisk->vol_bot_beg));
 846         printf(INDENT"vol_mem_beg\t%s\n", sizetostr(ondisk->vol_mem_beg));
 847         printf(INDENT"vol_buf_beg\t%s\n", sizetostr(ondisk->vol_buf_beg));
 848         printf(INDENT"vol_buf_end\t%s\n", sizetostr(ondisk->vol_buf_end));
 849         printf(INDENT"vol0_next_tid\t%016jx\n",
 850                (uintmax_t)ondisk->vol0_next_tid);
 851
 852         blockmap = &ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 853         size = HAMMER_OFF_LONG_ENCODE(blockmap->alloc_offset);
 854         if (blockmap->first_offset <= blockmap->next_offset)
 855                 used = blockmap->next_offset - blockmap->first_offset;
 856         else
 857                 used = blockmap->alloc_offset - blockmap->first_offset +
 858                         HAMMER_OFF_LONG_ENCODE(blockmap->next_offset);
 859         printf(INDENT"undo_size\t%s\n", sizetostr(size));
 860         printf(INDENT"undo_used\t%s\n", sizetostr(used));
 861
 862         printf(INDENT"zone #             "
 863                "phys             first            next             alloc\n");
 864         for (i = 0; i < HAMMER_MAX_ZONES; i++) {
 865                 blockmap = &ondisk->vol0_blockmap[i];
 866                 printf(INDENT"zone %-2d %-10s %016jx %016jx %016jx %016jx\n",
 867                         i, zone_labels[i],
 868                         (uintmax_t)blockmap->phys_offset,
 869                         (uintmax_t)blockmap->first_offset,
 870                         (uintmax_t)blockmap->next_offset,
 871                         (uintmax_t)blockmap->alloc_offset);
 872         }
 873 }
 874
 875 /*
 876  * Flush various tracking structures to disk
 877  */
 878 void
 879 flush_all_volumes(void)
 880 {
 881         volume_info_t volume;
 882
 883         TAILQ_FOREACH(volume, &VolList, entry)
 884                 flush_volume(volume);
 885 }
 886
 887 void
 888 flush_volume(volume_info_t volume)
 889 {
 890         buffer_info_t buffer;
 891         int i;
 892
 893         for (i = 0; i < HAMMER_BUFLISTS; ++i) {
 894                 TAILQ_FOREACH(buffer, &volume->buffer_lists[i], entry)
 895                         flush_buffer(buffer);
 896         }
 897         if (writehammervol(volume) == -1) {
 898                 err(1, "Write volume %d (%s)", volume->vol_no, volume->name);
 899                 /* not reached */
 900         }
 901 }
 902
 903 void
 904 flush_buffer(buffer_info_t buffer)
 905 {
 906         volume_info_t volume;
 907
 908         volume = buffer->volume;
 909         if (writehammerbuf(buffer) == -1) {
 910                 err(1, "Write volume %d (%s)", volume->vol_no, volume->name);
 911                 /* not reached */
 912         }
 913         buffer->cache.modified = 0;
 914 }
 915
 916 /*
 917  * Core I/O operations
 918  */
 919 static int
 920 __read(volume_info_t volume, void *data, int64_t offset, int size)
 921 {
 922         ssize_t n;
 923
 924         n = pread(volume->fd, data, size, offset);
 925         if (n != size)
 926                 return(-1);
 927         return(0);
 928 }
 929
 930 static __inline int
 931 readhammervol(volume_info_t volume)
 932 {
 933         return(__read(volume, volume->ondisk, 0, HAMMER_BUFSIZE));
 934 }
 935
 936 static __inline int
 937 readhammerbuf(buffer_info_t buffer)
 938 {
 939         return(__read(buffer->volume, buffer->ondisk, buffer->raw_offset,
 940                 HAMMER_BUFSIZE));
 941 }
 942
 943 static int
 944 __write(volume_info_t volume, const void *data, int64_t offset, int size)
 945 {
 946         ssize_t n;
 947
 948         if (volume->rdonly)
 949                 return(0);
 950
 951         n = pwrite(volume->fd, data, size, offset);
 952         if (n != size)
 953                 return(-1);
 954         return(0);
 955 }
 956
 957 static __inline int
 958 writehammervol(volume_info_t volume)
 959 {
 960         return(__write(volume, volume->ondisk, 0, HAMMER_BUFSIZE));
 961 }
 962
 963 static __inline int
 964 writehammerbuf(buffer_info_t buffer)
 965 {
 966         return(__write(buffer->volume, buffer->ondisk, buffer->raw_offset,
 967                 HAMMER_BUFSIZE));
 968 }
 969
 970 int64_t init_boot_area_size(int64_t value, off_t avg_vol_size)
 971 {
 972         if (value == 0) {
 973                 value = HAMMER_BOOT_NOMBYTES;
 974                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 975                         value >>= 1;
 976         }
 977
 978         if (value < HAMMER_BOOT_MINBYTES)
 979                 value = HAMMER_BOOT_MINBYTES;
 980         else if (value > HAMMER_BOOT_MAXBYTES)
 981                 value = HAMMER_BOOT_MAXBYTES;
 982
 983         return(value);
 984 }
 985
 986 int64_t init_memory_log_size(int64_t value, off_t avg_vol_size)
 987 {
 988         if (value == 0) {
 989                 value = HAMMER_MEM_NOMBYTES;
 990                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 991                         value >>= 1;
 992         }
 993
 994         if (value < HAMMER_MEM_MINBYTES)
 995                 value = HAMMER_MEM_MINBYTES;
 996         else if (value > HAMMER_MEM_MAXBYTES)
 997                 value = HAMMER_MEM_MAXBYTES;
 998
 999         return(value);
1000 }