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 *volume);
  41 static void get_buffer_readahead(struct buffer_info *base);
  42 static __inline int readhammervol(struct volume_info *volume);
  43 static __inline int readhammerbuf(struct buffer_info *buffer);
  44 static __inline int writehammervol(struct volume_info *volume);
  45 static __inline int writehammerbuf(struct buffer_info *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 struct buffer_info*
  69 find_buffer(hammer_off_t zone2_offset)
  70 {
  71         struct volume_info *volume;
  72         struct buffer_info *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 struct volume_info *
  87 __alloc_volume(const char *volname, int oflags)
  88 {
  89         struct volume_info *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         check_volume(volume);
 100
 101         volume->ondisk = calloc(1, HAMMER_BUFSIZE);
 102
 103         for (i = 0; i < HAMMER_BUFLISTS; ++i)
 104                 TAILQ_INIT(&volume->buffer_lists[i]);
 105
 106         return(volume);
 107 }
 108
 109 static void
 110 __add_volume(struct volume_info *volume)
 111 {
 112         struct volume_info *scan;
 113         struct stat st1, st2;
 114
 115         if (fstat(volume->fd, &st1) != 0)
 116                 errx(1, "add_volume: %s: Failed to stat", volume->name);
 117
 118         TAILQ_FOREACH(scan, &VolList, entry) {
 119                 if (scan->vol_no == volume->vol_no) {
 120                         errx(1, "add_volume: %s: Duplicate volume number %d "
 121                                 "against %s",
 122                                 volume->name, volume->vol_no, scan->name);
 123                 }
 124                 if (fstat(scan->fd, &st2) != 0) {
 125                         errx(1, "add_volume: %s: Failed to stat %s",
 126                                 volume->name, scan->name);
 127                 }
 128                 if ((st1.st_ino == st2.st_ino) && (st1.st_dev == st2.st_dev)) {
 129                         errx(1, "add_volume: %s: Specified more than once",
 130                                 volume->name);
 131                 }
 132         }
 133
 134         TAILQ_INSERT_TAIL(&VolList, volume, entry);
 135 }
 136
 137 static void
 138 __verify_volume(struct volume_info *volume)
 139 {
 140         hammer_volume_ondisk_t ondisk = volume->ondisk;
 141
 142         if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
 143                 errx(1, "verify_volume: Invalid volume signature %016jx",
 144                         ondisk->vol_signature);
 145         }
 146         if (ondisk->vol_rootvol != HAMMER_ROOT_VOLNO) {
 147                 errx(1, "verify_volume: Invalid root volume# %d",
 148                         ondisk->vol_rootvol);
 149         }
 150         if (bcmp(&Hammer_FSType, &ondisk->vol_fstype, sizeof(Hammer_FSType))) {
 151                 errx(1, "verify_volume: %s: Header does not indicate "
 152                         "that this is a HAMMER volume", volume->name);
 153         }
 154         if (bcmp(&Hammer_FSId, &ondisk->vol_fsid, sizeof(Hammer_FSId))) {
 155                 errx(1, "verify_volume: %s: FSId does not match other volumes!",
 156                         volume->name);
 157         }
 158 }
 159
 160 /*
 161  * Initialize a volume structure and ondisk vol_no field.
 162  */
 163 struct volume_info *
 164 init_volume(const char *filename, int oflags, int32_t vol_no)
 165 {
 166         struct volume_info *volume;
 167
 168         volume = __alloc_volume(filename, oflags);
 169         volume->vol_no = volume->ondisk->vol_no = vol_no;
 170
 171         __add_volume(volume);
 172
 173         return(volume);
 174 }
 175
 176 /*
 177  * Initialize a volume structure and read ondisk volume header.
 178  */
 179 struct volume_info*
 180 load_volume(const char *filename, int oflags, int verify)
 181 {
 182         struct volume_info *volume;
 183         int n;
 184
 185         volume = __alloc_volume(filename, oflags);
 186
 187         n = readhammervol(volume);
 188         if (n == -1) {
 189                 err(1, "load_volume: %s: Read failed at offset 0",
 190                     volume->name);
 191         }
 192         volume->vol_no = volume->ondisk->vol_no;
 193         HammerVersion = volume->ondisk->vol_version;
 194
 195         if (valid_hammer_volumes++ == 0)
 196                 Hammer_FSId = volume->ondisk->vol_fsid;
 197         if (verify)
 198                 __verify_volume(volume);
 199
 200         __add_volume(volume);
 201
 202         return(volume);
 203 }
 204
 205 /*
 206  * Check basic volume characteristics.
 207  */
 208 static void
 209 check_volume(struct volume_info *volume)
 210 {
 211         struct partinfo pinfo;
 212         struct stat st;
 213
 214         /*
 215          * Allow the formatting of block devices or regular files
 216          */
 217         if (ioctl(volume->fd, DIOCGPART, &pinfo) < 0) {
 218                 if (fstat(volume->fd, &st) < 0)
 219                         err(1, "Unable to stat %s", volume->name);
 220                 if (S_ISREG(st.st_mode)) {
 221                         volume->size = st.st_size;
 222                         volume->type = "REGFILE";
 223                 } else {
 224                         errx(1, "Unsupported file type for %s", volume->name);
 225                         /* not reached */
 226                 }
 227         } else {
 228                 /*
 229                  * When formatting a block device as a HAMMER volume the
 230                  * sector size must be compatible.  HAMMER uses 16384 byte
 231                  * filesystem buffers.
 232                  */
 233                 if (pinfo.reserved_blocks) {
 234                         errx(1, "HAMMER cannot be placed in a partition "
 235                                 "which overlaps the disklabel or MBR");
 236                 }
 237                 if (pinfo.media_blksize > HAMMER_BUFSIZE ||
 238                     HAMMER_BUFSIZE % pinfo.media_blksize) {
 239                         errx(1, "A media sector size of %d is not supported",
 240                              pinfo.media_blksize);
 241                 }
 242
 243                 volume->size = pinfo.media_size;
 244                 volume->device_offset = pinfo.media_offset;
 245                 volume->type = "DEVICE";
 246         }
 247 }
 248
 249 int
 250 is_regfile(struct volume_info *volume)
 251 {
 252         return(strcmp(volume->type, "REGFILE") ? 0 : 1);
 253 }
 254
 255 void
 256 assert_volume_offset(struct volume_info *volume)
 257 {
 258         assert(hammer_is_zone_raw_buffer(volume->vol_free_off));
 259         assert(hammer_is_zone_raw_buffer(volume->vol_free_end));
 260         if (volume->vol_free_off >= volume->vol_free_end)
 261                 errx(1, "Ran out of room, filesystem too small");
 262 }
 263
 264 struct volume_info *
 265 get_volume(int32_t vol_no)
 266 {
 267         struct volume_info *volume;
 268
 269         TAILQ_FOREACH(volume, &VolList, entry) {
 270                 if (volume->vol_no == vol_no)
 271                         break;
 272         }
 273
 274         return(volume);
 275 }
 276
 277 struct volume_info *
 278 get_root_volume(void)
 279 {
 280         return(get_volume(HAMMER_ROOT_VOLNO));
 281 }
 282
 283 static hammer_off_t
 284 __blockmap_xlate_to_zone2(hammer_off_t buf_offset)
 285 {
 286         hammer_off_t zone2_offset;
 287         int error = 0;
 288
 289         if (hammer_is_zone_raw_buffer(buf_offset))
 290                 zone2_offset = buf_offset;
 291         else
 292                 zone2_offset = blockmap_lookup(buf_offset, &error);
 293
 294         if (error)
 295                 return(HAMMER_OFF_BAD);
 296         assert(hammer_is_zone_raw_buffer(zone2_offset));
 297
 298         return(zone2_offset);
 299 }
 300
 301 static struct buffer_info *
 302 __alloc_buffer(hammer_off_t zone2_offset, int isnew)
 303 {
 304         struct volume_info *volume;
 305         struct buffer_info *buffer;
 306         int hi;
 307
 308         volume = get_volume(HAMMER_VOL_DECODE(zone2_offset));
 309         assert(volume != NULL);
 310
 311         buffer = calloc(1, sizeof(*buffer));
 312         buffer->zone2_offset = zone2_offset;
 313         buffer->raw_offset = hammer_xlate_to_phys(volume->ondisk, zone2_offset);
 314         buffer->volume = volume;
 315         buffer->ondisk = calloc(1, HAMMER_BUFSIZE);
 316
 317         if (isnew <= 0) {
 318                 if (readhammerbuf(buffer) == -1) {
 319                         err(1, "Failed to read %s:%016jx at %016jx",
 320                             volume->name,
 321                             (intmax_t)buffer->zone2_offset,
 322                             (intmax_t)buffer->raw_offset);
 323                 }
 324         }
 325
 326         hi = buffer_hash(zone2_offset);
 327         TAILQ_INSERT_TAIL(&volume->buffer_lists[hi], buffer, entry);
 328         hammer_cache_add(&buffer->cache);
 329
 330         return(buffer);
 331 }
 332
 333 /*
 334  * Acquire the 16KB buffer for specified zone offset.
 335  */
 336 static struct buffer_info *
 337 get_buffer(hammer_off_t buf_offset, int isnew)
 338 {
 339         struct buffer_info *buffer;
 340         hammer_off_t zone2_offset;
 341         int dora = 0;
 342
 343         zone2_offset = __blockmap_xlate_to_zone2(buf_offset);
 344         if (zone2_offset == HAMMER_OFF_BAD)
 345                 return(NULL);
 346
 347         zone2_offset &= ~HAMMER_BUFMASK64;
 348         buffer = find_buffer(zone2_offset);
 349
 350         if (buffer == NULL) {
 351                 buffer = __alloc_buffer(zone2_offset, isnew);
 352                 dora = (isnew == 0);
 353         } else {
 354                 assert(isnew != -1);
 355                 hammer_cache_used(&buffer->cache);
 356         }
 357         assert(buffer->ondisk != NULL);
 358
 359         ++buffer->cache.refs;
 360         hammer_cache_flush();
 361
 362         if (isnew > 0) {
 363                 assert(buffer->cache.modified == 0);
 364                 bzero(buffer->ondisk, HAMMER_BUFSIZE);
 365                 buffer->cache.modified = 1;
 366         }
 367         if (dora)
 368                 get_buffer_readahead(buffer);
 369         return(buffer);
 370 }
 371
 372 static void
 373 get_buffer_readahead(struct buffer_info *base)
 374 {
 375         struct buffer_info *buffer;
 376         struct volume_info *volume;
 377         hammer_off_t zone2_offset;
 378         int64_t raw_offset;
 379         int ri = UseReadBehind;
 380         int re = UseReadAhead;
 381
 382         raw_offset = base->raw_offset + ri * HAMMER_BUFSIZE;
 383         volume = base->volume;
 384
 385         while (ri < re) {
 386                 if (raw_offset >= volume->ondisk->vol_buf_end)
 387                         break;
 388                 if (raw_offset < volume->ondisk->vol_buf_beg || ri == 0) {
 389                         ++ri;
 390                         raw_offset += HAMMER_BUFSIZE;
 391                         continue;
 392                 }
 393                 zone2_offset = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no,
 394                         raw_offset - volume->ondisk->vol_buf_beg);
 395                 buffer = find_buffer(zone2_offset);
 396                 if (buffer == NULL) {
 397                         /* call with -1 to prevent another readahead */
 398                         buffer = get_buffer(zone2_offset, -1);
 399                         rel_buffer(buffer);
 400                 }
 401                 ++ri;
 402                 raw_offset += HAMMER_BUFSIZE;
 403         }
 404 }
 405
 406 void
 407 rel_buffer(struct buffer_info *buffer)
 408 {
 409         struct volume_info *volume;
 410         int hi;
 411
 412         if (buffer == NULL)
 413                 return;
 414         assert(buffer->cache.refs > 0);
 415         if (--buffer->cache.refs == 0) {
 416                 if (buffer->cache.delete) {
 417                         hi = buffer_hash(buffer->zone2_offset);
 418                         volume = buffer->volume;
 419                         if (buffer->cache.modified)
 420                                 flush_buffer(buffer);
 421                         TAILQ_REMOVE(&volume->buffer_lists[hi], buffer, entry);
 422                         hammer_cache_del(&buffer->cache);
 423                         free(buffer->ondisk);
 424                         free(buffer);
 425                 }
 426         }
 427 }
 428
 429 /*
 430  * Retrieve a pointer to a buffer data given a buffer offset.  The underlying
 431  * bufferp is freed if isnew or the offset is out of range of the cached data.
 432  * If bufferp is freed a referenced buffer is loaded into it.
 433  */
 434 void *
 435 get_buffer_data(hammer_off_t buf_offset, struct buffer_info **bufferp,
 436                 int isnew)
 437 {
 438         hammer_off_t xor;
 439
 440         if (*bufferp != NULL) {
 441                 /* XXX xor is always non zero for indirect zones */
 442                 xor = HAMMER_OFF_LONG_ENCODE(buf_offset) ^
 443                       HAMMER_OFF_LONG_ENCODE((*bufferp)->zone2_offset);
 444                 if (isnew > 0 || (xor & ~HAMMER_BUFMASK64)) {
 445                         rel_buffer(*bufferp);
 446                         *bufferp = NULL;
 447                 }
 448         }
 449
 450         if (*bufferp == NULL) {
 451                 *bufferp = get_buffer(buf_offset, isnew);
 452                 if (*bufferp == NULL)
 453                         return(NULL);
 454         }
 455
 456         return(((char *)(*bufferp)->ondisk) +
 457                 ((int32_t)buf_offset & HAMMER_BUFMASK));
 458 }
 459
 460 /*
 461  * Allocate HAMMER elements - B-Tree nodes
 462  */
 463 hammer_node_ondisk_t
 464 alloc_btree_node(hammer_off_t *offp, struct buffer_info **data_bufferp)
 465 {
 466         hammer_node_ondisk_t node;
 467
 468         node = alloc_blockmap(HAMMER_ZONE_BTREE_INDEX, sizeof(*node),
 469                               offp, data_bufferp);
 470         bzero(node, sizeof(*node));
 471         return(node);
 472 }
 473
 474 /*
 475  * Allocate HAMMER elements - meta data (inode, direntry, PFS, etc)
 476  */
 477 void *
 478 alloc_meta_element(hammer_off_t *offp, int32_t data_len,
 479                    struct buffer_info **data_bufferp)
 480 {
 481         void *data;
 482
 483         data = alloc_blockmap(HAMMER_ZONE_META_INDEX, data_len,
 484                               offp, data_bufferp);
 485         bzero(data, data_len);
 486         return(data);
 487 }
 488
 489 /*
 490  * Format a new blockmap.  This is mostly a degenerate case because
 491  * all allocations are now actually done from the freemap.
 492  */
 493 void
 494 format_blockmap(struct volume_info *root_vol, int zone, hammer_off_t offset)
 495 {
 496         hammer_blockmap_t blockmap;
 497         hammer_off_t zone_base;
 498
 499         /* Only root volume needs formatting */
 500         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 501
 502         assert(hammer_is_index_record(zone));
 503
 504         blockmap = &root_vol->ondisk->vol0_blockmap[zone];
 505         zone_base = HAMMER_ZONE_ENCODE(zone, offset);
 506
 507         bzero(blockmap, sizeof(*blockmap));
 508         blockmap->phys_offset = 0;
 509         blockmap->first_offset = zone_base;
 510         blockmap->next_offset = zone_base;
 511         blockmap->alloc_offset = HAMMER_ENCODE(zone, 255, -1);
 512         hammer_crc_set_blockmap(HammerVersion, blockmap);
 513 }
 514
 515 /*
 516  * Format a new freemap.  Set all layer1 entries to UNAVAIL.  The initialize
 517  * code will load each volume's freemap.
 518  */
 519 void
 520 format_freemap(struct volume_info *root_vol)
 521 {
 522         struct buffer_info *buffer = NULL;
 523         hammer_off_t layer1_offset;
 524         hammer_blockmap_t blockmap;
 525         hammer_blockmap_layer1_t layer1;
 526         int i, isnew;
 527
 528         /* Only root volume needs formatting */
 529         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 530
 531         layer1_offset = bootstrap_bigblock(root_vol);
 532         for (i = 0; i < HAMMER_BIGBLOCK_SIZE; i += sizeof(*layer1)) {
 533                 isnew = ((i % HAMMER_BUFSIZE) == 0);
 534                 layer1 = get_buffer_data(layer1_offset + i, &buffer, isnew);
 535                 bzero(layer1, sizeof(*layer1));
 536                 layer1->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 537                 layer1->blocks_free = 0;
 538                 hammer_crc_set_layer1(HammerVersion, layer1);
 539         }
 540         assert(i == HAMMER_BIGBLOCK_SIZE);
 541         rel_buffer(buffer);
 542
 543         blockmap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 544         bzero(blockmap, sizeof(*blockmap));
 545         blockmap->phys_offset = layer1_offset;
 546         blockmap->first_offset = 0;
 547         blockmap->next_offset = HAMMER_ENCODE_RAW_BUFFER(0, 0);
 548         blockmap->alloc_offset = HAMMER_ENCODE_RAW_BUFFER(255, -1);
 549         hammer_crc_set_blockmap(HammerVersion, blockmap);
 550 }
 551
 552 /*
 553  * Load the volume's remaining free space into the freemap.
 554  *
 555  * Returns the number of big-blocks available.
 556  */
 557 int64_t
 558 initialize_freemap(struct volume_info *volume)
 559 {
 560         struct volume_info *root_vol;
 561         struct buffer_info *buffer1 = NULL;
 562         struct buffer_info *buffer2 = NULL;
 563         hammer_blockmap_layer1_t layer1;
 564         hammer_blockmap_layer2_t layer2;
 565         hammer_off_t layer1_offset;
 566         hammer_off_t layer2_offset;
 567         hammer_off_t phys_offset;
 568         hammer_off_t block_offset;
 569         hammer_off_t aligned_vol_free_end;
 570         hammer_blockmap_t freemap;
 571         int64_t count = 0;
 572         int64_t layer1_count = 0;
 573
 574         root_vol = get_root_volume();
 575
 576         assert_volume_offset(volume);
 577         aligned_vol_free_end = HAMMER_BLOCKMAP_LAYER2_DOALIGN(volume->vol_free_end);
 578
 579         printf("initialize freemap volume %d\n", volume->vol_no);
 580
 581         /*
 582          * Initialize the freemap.  First preallocate the big-blocks required
 583          * to implement layer2.   This preallocation is a bootstrap allocation
 584          * using blocks from the target volume.
 585          */
 586         freemap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 587
 588         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no, 0);
 589              phys_offset < aligned_vol_free_end;
 590              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 591                 layer1_offset = freemap->phys_offset +
 592                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 593                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 594                 if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 595                         layer1->phys_offset = bootstrap_bigblock(volume);
 596                         layer1->blocks_free = 0;
 597                         buffer1->cache.modified = 1;
 598                         hammer_crc_set_layer1(HammerVersion, layer1);
 599                 }
 600         }
 601
 602         /*
 603          * Now fill everything in.
 604          */
 605         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no, 0);
 606              phys_offset < aligned_vol_free_end;
 607              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 608                 layer1_count = 0;
 609                 layer1_offset = freemap->phys_offset +
 610                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 611                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 612                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 613
 614                 for (block_offset = 0;
 615                      block_offset < HAMMER_BLOCKMAP_LAYER2;
 616                      block_offset += HAMMER_BIGBLOCK_SIZE) {
 617                         layer2_offset = layer1->phys_offset +
 618                                         HAMMER_BLOCKMAP_LAYER2_OFFSET(block_offset);
 619                         layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 620                         bzero(layer2, sizeof(*layer2));
 621
 622                         if (phys_offset + block_offset < volume->vol_free_off) {
 623                                 /*
 624                                  * Big-blocks already allocated as part
 625                                  * of the freemap bootstrap.
 626                                  */
 627                                 layer2->zone = HAMMER_ZONE_FREEMAP_INDEX;
 628                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 629                                 layer2->bytes_free = 0;
 630                         } else if (phys_offset + block_offset < volume->vol_free_end) {
 631                                 layer2->zone = 0;
 632                                 layer2->append_off = 0;
 633                                 layer2->bytes_free = HAMMER_BIGBLOCK_SIZE;
 634                                 ++count;
 635                                 ++layer1_count;
 636                         } else {
 637                                 layer2->zone = HAMMER_ZONE_UNAVAIL_INDEX;
 638                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 639                                 layer2->bytes_free = 0;
 640                         }
 641                         hammer_crc_set_layer2(HammerVersion, layer2);
 642                         buffer2->cache.modified = 1;
 643                 }
 644
 645                 layer1->blocks_free += layer1_count;
 646                 hammer_crc_set_layer1(HammerVersion, layer1);
 647                 buffer1->cache.modified = 1;
 648         }
 649
 650         rel_buffer(buffer1);
 651         rel_buffer(buffer2);
 652         return(count);
 653 }
 654
 655 /*
 656  * Returns the number of big-blocks available for filesystem data and undos
 657  * without formatting.
 658  */
 659 int64_t
 660 count_freemap(struct volume_info *volume)
 661 {
 662         hammer_off_t phys_offset;
 663         hammer_off_t vol_free_off;
 664         hammer_off_t aligned_vol_free_end;
 665         int64_t count = 0;
 666
 667         vol_free_off = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no, 0);
 668
 669         assert_volume_offset(volume);
 670         aligned_vol_free_end = HAMMER_BLOCKMAP_LAYER2_DOALIGN(volume->vol_free_end);
 671
 672         if (volume->vol_no == HAMMER_ROOT_VOLNO)
 673                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 674
 675         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no, 0);
 676              phys_offset < aligned_vol_free_end;
 677              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 678                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 679         }
 680
 681         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no, 0);
 682              phys_offset < aligned_vol_free_end;
 683              phys_offset += HAMMER_BIGBLOCK_SIZE) {
 684                 if (phys_offset < vol_free_off)
 685                         ;
 686                 else if (phys_offset < volume->vol_free_end)
 687                         ++count;
 688         }
 689
 690         return(count);
 691 }
 692
 693 /*
 694  * Format the undomap for the root volume.
 695  */
 696 void
 697 format_undomap(struct volume_info *root_vol, int64_t *undo_buffer_size)
 698 {
 699         hammer_off_t undo_limit;
 700         hammer_blockmap_t blockmap;
 701         hammer_volume_ondisk_t ondisk;
 702         struct buffer_info *buffer = NULL;
 703         hammer_off_t scan;
 704         int n;
 705         int limit_index;
 706         uint32_t seqno;
 707
 708         /* Only root volume needs formatting */
 709         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 710         ondisk = root_vol->ondisk;
 711
 712         /*
 713          * Size the undo buffer in multiples of HAMMER_BIGBLOCK_SIZE,
 714          * up to HAMMER_MAX_UNDO_BIGBLOCKS big-blocks.
 715          * Size to approximately 0.1% of the disk.
 716          *
 717          * The minimum UNDO fifo size is 512MB, or approximately 1% of
 718          * the recommended 50G disk.
 719          *
 720          * Changing this minimum is rather dangerous as complex filesystem
 721          * operations can cause the UNDO FIFO to fill up otherwise.
 722          */
 723         undo_limit = *undo_buffer_size;
 724         if (undo_limit == 0) {
 725                 undo_limit = HAMMER_VOL_BUF_SIZE(ondisk) / 1000;
 726                 if (undo_limit < HAMMER_BIGBLOCK_SIZE * HAMMER_MIN_UNDO_BIGBLOCKS)
 727                         undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_MIN_UNDO_BIGBLOCKS;
 728         }
 729         undo_limit = HAMMER_BIGBLOCK_DOALIGN(undo_limit);
 730         if (undo_limit < HAMMER_BIGBLOCK_SIZE)
 731                 undo_limit = HAMMER_BIGBLOCK_SIZE;
 732         if (undo_limit > HAMMER_BIGBLOCK_SIZE * HAMMER_MAX_UNDO_BIGBLOCKS)
 733                 undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_MAX_UNDO_BIGBLOCKS;
 734         *undo_buffer_size = undo_limit;
 735
 736         blockmap = &ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 737         bzero(blockmap, sizeof(*blockmap));
 738         blockmap->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 739         blockmap->first_offset = HAMMER_ENCODE_UNDO(0);
 740         blockmap->next_offset = blockmap->first_offset;
 741         blockmap->alloc_offset = HAMMER_ENCODE_UNDO(undo_limit);
 742         hammer_crc_set_blockmap(HammerVersion, blockmap);
 743
 744         limit_index = undo_limit / HAMMER_BIGBLOCK_SIZE;
 745         assert(limit_index <= HAMMER_MAX_UNDO_BIGBLOCKS);
 746
 747         for (n = 0; n < limit_index; ++n)
 748                 ondisk->vol0_undo_array[n] = alloc_undo_bigblock(root_vol);
 749         while (n < HAMMER_MAX_UNDO_BIGBLOCKS)
 750                 ondisk->vol0_undo_array[n++] = HAMMER_BLOCKMAP_UNAVAIL;
 751
 752         /*
 753          * Pre-initialize the UNDO blocks (HAMMER version 4+)
 754          */
 755         printf("initializing the undo map (%jd MB)\n",
 756                 (intmax_t)HAMMER_OFF_LONG_ENCODE(blockmap->alloc_offset) /
 757                 (1024 * 1024));
 758
 759         scan = blockmap->first_offset;
 760         seqno = 0;
 761
 762         while (scan < blockmap->alloc_offset) {
 763                 hammer_fifo_head_t head;
 764                 hammer_fifo_tail_t tail;
 765                 int isnew;
 766                 int bytes = HAMMER_UNDO_ALIGN;
 767
 768                 isnew = ((scan & HAMMER_BUFMASK64) == 0);
 769                 head = get_buffer_data(scan, &buffer, isnew);
 770                 buffer->cache.modified = 1;
 771                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 772
 773                 bzero(head, bytes);
 774                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 775                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
 776                 head->hdr_size = bytes;
 777                 head->hdr_seq = seqno++;
 778
 779                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 780                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 781                 tail->tail_size = bytes;
 782
 783                 hammer_crc_set_fifo_head(HammerVersion, head, bytes);
 784
 785                 scan += bytes;
 786         }
 787         rel_buffer(buffer);
 788 }
 789
 790 const char *zone_labels[] = {
 791         "",             /* 0 */
 792         "raw_volume",   /* 1 */
 793         "raw_buffer",   /* 2 */
 794         "undo",         /* 3 */
 795         "freemap",      /* 4 */
 796         "",             /* 5 */
 797         "",             /* 6 */
 798         "",             /* 7 */
 799         "btree",        /* 8 */
 800         "meta",         /* 9 */
 801         "large_data",   /* 10 */
 802         "small_data",   /* 11 */
 803         "",             /* 12 */
 804         "",             /* 13 */
 805         "",             /* 14 */
 806         "unavail",      /* 15 */
 807 };
 808
 809 void
 810 print_blockmap(const struct volume_info *volume)
 811 {
 812         hammer_blockmap_t blockmap;
 813         hammer_volume_ondisk_t ondisk;
 814         int64_t size, used;
 815         int i;
 816 #define INDENT ""
 817
 818         ondisk = volume->ondisk;
 819         printf(INDENT"vol_label\t%s\n", ondisk->vol_label);
 820         printf(INDENT"vol_count\t%d\n", ondisk->vol_count);
 821         printf(INDENT"vol_bot_beg\t%s\n", sizetostr(ondisk->vol_bot_beg));
 822         printf(INDENT"vol_mem_beg\t%s\n", sizetostr(ondisk->vol_mem_beg));
 823         printf(INDENT"vol_buf_beg\t%s\n", sizetostr(ondisk->vol_buf_beg));
 824         printf(INDENT"vol_buf_end\t%s\n", sizetostr(ondisk->vol_buf_end));
 825         printf(INDENT"vol0_next_tid\t%016jx\n",
 826                (uintmax_t)ondisk->vol0_next_tid);
 827
 828         blockmap = &ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 829         size = HAMMER_OFF_LONG_ENCODE(blockmap->alloc_offset);
 830         if (blockmap->first_offset <= blockmap->next_offset)
 831                 used = blockmap->next_offset - blockmap->first_offset;
 832         else
 833                 used = blockmap->alloc_offset - blockmap->first_offset +
 834                         HAMMER_OFF_LONG_ENCODE(blockmap->next_offset);
 835         printf(INDENT"undo_size\t%s\n", sizetostr(size));
 836         printf(INDENT"undo_used\t%s\n", sizetostr(used));
 837
 838         printf(INDENT"zone #             "
 839                "phys             first            next             alloc\n");
 840         for (i = 0; i < HAMMER_MAX_ZONES; i++) {
 841                 blockmap = &ondisk->vol0_blockmap[i];
 842                 printf(INDENT"zone %-2d %-10s %016jx %016jx %016jx %016jx\n",
 843                         i, zone_labels[i],
 844                         (uintmax_t)blockmap->phys_offset,
 845                         (uintmax_t)blockmap->first_offset,
 846                         (uintmax_t)blockmap->next_offset,
 847                         (uintmax_t)blockmap->alloc_offset);
 848         }
 849 }
 850
 851 /*
 852  * Flush various tracking structures to disk
 853  */
 854 void
 855 flush_all_volumes(void)
 856 {
 857         struct volume_info *volume;
 858
 859         TAILQ_FOREACH(volume, &VolList, entry)
 860                 flush_volume(volume);
 861 }
 862
 863 void
 864 flush_volume(struct volume_info *volume)
 865 {
 866         struct buffer_info *buffer;
 867         int i;
 868
 869         for (i = 0; i < HAMMER_BUFLISTS; ++i) {
 870                 TAILQ_FOREACH(buffer, &volume->buffer_lists[i], entry)
 871                         flush_buffer(buffer);
 872         }
 873         if (writehammervol(volume) == -1)
 874                 err(1, "Write volume %d (%s)", volume->vol_no, volume->name);
 875 }
 876
 877 void
 878 flush_buffer(struct buffer_info *buffer)
 879 {
 880         struct volume_info *volume;
 881
 882         volume = buffer->volume;
 883         if (writehammerbuf(buffer) == -1)
 884                 err(1, "Write volume %d (%s)", volume->vol_no, volume->name);
 885         buffer->cache.modified = 0;
 886 }
 887
 888 /*
 889  * Core I/O operations
 890  */
 891 static int
 892 __read(struct volume_info *volume, void *data, int64_t offset, int size)
 893 {
 894         ssize_t n;
 895
 896         n = pread(volume->fd, data, size, offset);
 897         if (n != size)
 898                 return(-1);
 899         return(0);
 900 }
 901
 902 static __inline int
 903 readhammervol(struct volume_info *volume)
 904 {
 905         return(__read(volume, volume->ondisk, 0, HAMMER_BUFSIZE));
 906 }
 907
 908 static __inline int
 909 readhammerbuf(struct buffer_info *buffer)
 910 {
 911         return(__read(buffer->volume, buffer->ondisk, buffer->raw_offset,
 912                 HAMMER_BUFSIZE));
 913 }
 914
 915 static int
 916 __write(struct volume_info *volume, const void *data, int64_t offset, int size)
 917 {
 918         ssize_t n;
 919
 920         if (volume->rdonly)
 921                 return(0);
 922
 923         n = pwrite(volume->fd, data, size, offset);
 924         if (n != size)
 925                 return(-1);
 926         return(0);
 927 }
 928
 929 static __inline int
 930 writehammervol(struct volume_info *volume)
 931 {
 932         return(__write(volume, volume->ondisk, 0, HAMMER_BUFSIZE));
 933 }
 934
 935 static __inline int
 936 writehammerbuf(struct buffer_info *buffer)
 937 {
 938         return(__write(buffer->volume, buffer->ondisk, buffer->raw_offset,
 939                 HAMMER_BUFSIZE));
 940 }
 941
 942 int64_t init_boot_area_size(int64_t value, off_t avg_vol_size)
 943 {
 944         if (value == 0) {
 945                 value = HAMMER_BOOT_NOMBYTES;
 946                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 947                         value >>= 1;
 948         }
 949
 950         if (value < HAMMER_BOOT_MINBYTES)
 951                 value = HAMMER_BOOT_MINBYTES;
 952         else if (value > HAMMER_BOOT_MAXBYTES)
 953                 value = HAMMER_BOOT_MAXBYTES;
 954
 955         return(value);
 956 }
 957
 958 int64_t init_memory_log_size(int64_t value, off_t avg_vol_size)
 959 {
 960         if (value == 0) {
 961                 value = HAMMER_MEM_NOMBYTES;
 962                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 963                         value >>= 1;
 964         }
 965
 966         if (value < HAMMER_MEM_MINBYTES)
 967                 value = HAMMER_MEM_MINBYTES;
 968         else if (value > HAMMER_MEM_MAXBYTES)
 969                 value = HAMMER_MEM_MAXBYTES;
 970
 971         return(value);
 972 }