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 get_buffer_readahead(struct buffer_info *base);
  41 static void *get_ondisk(hammer_off_t buf_offset, struct buffer_info **bufferp,
  42                         int isnew);
  43 static __inline int readhammervol(struct volume_info *vol);
  44 static __inline int readhammerbuf(struct buffer_info *buf);
  45 static __inline int writehammervol(struct volume_info *vol);
  46 static __inline int writehammerbuf(struct buffer_info *buf);
  47
  48 uuid_t Hammer_FSType;
  49 uuid_t Hammer_FSId;
  50 int UseReadBehind = -4;
  51 int UseReadAhead = 4;
  52 int DebugOpt;
  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 buf_offset)
  61 {
  62         int hi;
  63
  64         hi = (int)(buf_offset / HAMMER_BUFSIZE) & HAMMER_BUFLISTMASK;
  65         return(hi);
  66 }
  67
  68 static struct buffer_info*
  69 find_buffer(struct volume_info *volume, hammer_off_t buf_offset)
  70 {
  71         int hi;
  72         struct buffer_info *buf;
  73
  74         hi = buffer_hash(buf_offset);
  75         TAILQ_FOREACH(buf, &volume->buffer_lists[hi], entry)
  76                 if (buf->buf_offset == buf_offset)
  77                         return(buf);
  78         return(NULL);
  79 }
  80
  81 static
  82 struct volume_info *
  83 __alloc_volume(const char *volname, int oflags)
  84 {
  85         struct volume_info *vol;
  86         int i;
  87
  88         vol = malloc(sizeof(*vol));
  89         if (vol == NULL)
  90                 err(1, "alloc_volume");
  91         bzero(vol, sizeof(*vol));
  92
  93         vol->vol_no = -1;
  94         vol->rdonly = (oflags == O_RDONLY);
  95         vol->name = strdup(volname);
  96         vol->fd = open(vol->name, oflags);
  97         if (vol->fd < 0)
  98                 err(1, "alloc_volume: Failed to open %s", vol->name);
  99
 100         vol->size = 0;
 101         vol->device_offset = 0;
 102         vol->type = NULL;
 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 /*
 144  * Initialize a volume structure and ondisk vol_no field.
 145  */
 146 struct volume_info *
 147 init_volume(int32_t vol_no, const char *filename, int oflags)
 148 {
 149         struct volume_info *vol;
 150
 151         vol = __alloc_volume(filename, oflags);
 152         vol->vol_no = vol->ondisk->vol_no = vol_no;
 153
 154         __add_volume(vol);
 155
 156         return(vol);
 157 }
 158
 159 /*
 160  * Initialize a volume structure and read ondisk volume header.
 161  */
 162 struct volume_info*
 163 load_volume(const char *filename, int oflags)
 164 {
 165         struct volume_info *vol;
 166         hammer_volume_ondisk_t ondisk;
 167         int n;
 168
 169         vol = __alloc_volume(filename, oflags);
 170
 171         n = readhammervol(vol);
 172         if (n == -1) {
 173                 err(1, "load_volume: %s: Read failed at offset 0", vol->name);
 174         }
 175         ondisk = vol->ondisk;
 176         vol->vol_no = ondisk->vol_no;
 177
 178         if (ondisk->vol_rootvol != HAMMER_ROOT_VOLNO) {
 179                 errx(1, "load_volume: Invalid root volume# %d",
 180                         ondisk->vol_rootvol);
 181         }
 182
 183         if (bcmp(&Hammer_FSType, &ondisk->vol_fstype, sizeof(Hammer_FSType))) {
 184                 errx(1, "load_volume: %s: Header does not indicate "
 185                         "that this is a hammer volume", vol->name);
 186         }
 187
 188         if (valid_hammer_volumes++ == 0) {
 189                 Hammer_FSId = ondisk->vol_fsid;
 190         } else if (bcmp(&Hammer_FSId, &ondisk->vol_fsid, sizeof(Hammer_FSId))) {
 191                 errx(1, "load_volume: %s: FSId does match other volumes!",
 192                         vol->name);
 193         }
 194
 195         __add_volume(vol);
 196
 197         return(vol);
 198 }
 199
 200 /*
 201  * Check basic volume characteristics.
 202  */
 203 void
 204 check_volume(struct volume_info *vol)
 205 {
 206         struct partinfo pinfo;
 207         struct stat st;
 208
 209         /*
 210          * Get basic information about the volume
 211          */
 212         if (ioctl(vol->fd, DIOCGPART, &pinfo) < 0) {
 213                 /*
 214                  * Allow the formatting of regular files as HAMMER volumes
 215                  */
 216                 if (fstat(vol->fd, &st) < 0)
 217                         err(1, "Unable to stat %s", vol->name);
 218                 vol->size = st.st_size;
 219                 vol->type = "REGFILE";
 220         } else {
 221                 /*
 222                  * When formatting a block device as a HAMMER volume the
 223                  * sector size must be compatible.  HAMMER uses 16384 byte
 224                  * filesystem buffers.
 225                  */
 226                 if (pinfo.reserved_blocks) {
 227                         errx(1, "HAMMER cannot be placed in a partition "
 228                                 "which overlaps the disklabel or MBR");
 229                 }
 230                 if (pinfo.media_blksize > HAMMER_BUFSIZE ||
 231                     HAMMER_BUFSIZE % pinfo.media_blksize) {
 232                         errx(1, "A media sector size of %d is not supported",
 233                              pinfo.media_blksize);
 234                 }
 235
 236                 vol->size = pinfo.media_size;
 237                 vol->device_offset = pinfo.media_offset;
 238                 vol->type = "DEVICE";
 239         }
 240 }
 241
 242 struct volume_info *
 243 get_volume(int32_t vol_no)
 244 {
 245         struct volume_info *vol;
 246
 247         TAILQ_FOREACH(vol, &VolList, entry) {
 248                 if (vol->vol_no == vol_no)
 249                         break;
 250         }
 251
 252         return(vol);
 253 }
 254
 255 struct volume_info *
 256 get_root_volume(void)
 257 {
 258         struct volume_info *root_vol;
 259
 260         root_vol = get_volume(HAMMER_ROOT_VOLNO);
 261         assert(root_vol != NULL);
 262
 263         return(root_vol);
 264 }
 265
 266 /*
 267  * Acquire the specified buffer.  isnew is -1 only when called
 268  * via get_buffer_readahead() to prevent another readahead.
 269  */
 270 static struct buffer_info *
 271 get_buffer(hammer_off_t buf_offset, int isnew)
 272 {
 273         struct buffer_info *buf;
 274         struct volume_info *volume;
 275         int vol_no;
 276         int zone;
 277         int hi;
 278         int dora = 0;
 279         int error = 0;
 280
 281         zone = HAMMER_ZONE_DECODE(buf_offset);
 282         if (zone > HAMMER_ZONE_RAW_BUFFER_INDEX)
 283                 buf_offset = blockmap_lookup(buf_offset, NULL, NULL, &error);
 284         if (error || buf_offset == HAMMER_OFF_BAD)
 285                 return(NULL);
 286         assert(hammer_is_zone_raw_buffer(buf_offset));
 287
 288         vol_no = HAMMER_VOL_DECODE(buf_offset);
 289         volume = get_volume(vol_no);
 290         assert(volume != NULL);
 291
 292         buf_offset &= ~HAMMER_BUFMASK64;
 293         buf = find_buffer(volume, buf_offset);
 294
 295         if (buf == NULL) {
 296                 buf = malloc(sizeof(*buf));
 297                 bzero(buf, sizeof(*buf));
 298                 buf->buf_offset = buf_offset;
 299                 buf->raw_offset = hammer_xlate_to_phys(volume->ondisk,
 300                                                         buf_offset);
 301                 buf->volume = volume;
 302                 buf->ondisk = malloc(HAMMER_BUFSIZE);
 303                 if (isnew <= 0) {
 304                         if (readhammerbuf(buf) == -1) {
 305                                 err(1, "get_buffer: %s:%016jx "
 306                                     "Read failed at offset %016jx",
 307                                     volume->name,
 308                                     (intmax_t)buf->buf_offset,
 309                                     (intmax_t)buf->raw_offset);
 310                         }
 311                 }
 312
 313                 hi = buffer_hash(buf_offset);
 314                 TAILQ_INSERT_TAIL(&volume->buffer_lists[hi], buf, entry);
 315                 hammer_cache_add(&buf->cache);
 316                 dora = (isnew == 0);
 317         } else {
 318                 assert(buf->ondisk != NULL);
 319                 assert(isnew != -1);
 320                 hammer_cache_used(&buf->cache);
 321         }
 322
 323         ++buf->cache.refs;
 324         hammer_cache_flush();
 325
 326         if (isnew > 0) {
 327                 assert(buf->cache.modified == 0);
 328                 bzero(buf->ondisk, HAMMER_BUFSIZE);
 329                 buf->cache.modified = 1;
 330         }
 331         if (dora)
 332                 get_buffer_readahead(buf);
 333         return(buf);
 334 }
 335
 336 static void
 337 get_buffer_readahead(struct buffer_info *base)
 338 {
 339         struct buffer_info *buf;
 340         struct volume_info *vol;
 341         hammer_off_t buf_offset;
 342         int64_t raw_offset;
 343         int ri = UseReadBehind;
 344         int re = UseReadAhead;
 345
 346         raw_offset = base->raw_offset + ri * HAMMER_BUFSIZE;
 347         vol = base->volume;
 348
 349         while (ri < re) {
 350                 if (raw_offset >= vol->ondisk->vol_buf_end)
 351                         break;
 352                 if (raw_offset < vol->ondisk->vol_buf_beg || ri == 0) {
 353                         ++ri;
 354                         raw_offset += HAMMER_BUFSIZE;
 355                         continue;
 356                 }
 357                 buf_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no,
 358                         raw_offset - vol->ondisk->vol_buf_beg);
 359                 buf = find_buffer(vol, buf_offset);
 360                 if (buf == NULL) {
 361                         buf = get_buffer(buf_offset, -1);
 362                         rel_buffer(buf);
 363                 }
 364                 ++ri;
 365                 raw_offset += HAMMER_BUFSIZE;
 366         }
 367 }
 368
 369 void
 370 rel_buffer(struct buffer_info *buffer)
 371 {
 372         struct volume_info *volume;
 373         int hi;
 374
 375         if (buffer == NULL)
 376                 return;
 377         assert(buffer->cache.refs > 0);
 378         if (--buffer->cache.refs == 0) {
 379                 if (buffer->cache.delete) {
 380                         hi = buffer_hash(buffer->buf_offset);
 381                         volume = buffer->volume;
 382                         if (buffer->cache.modified)
 383                                 flush_buffer(buffer);
 384                         TAILQ_REMOVE(&volume->buffer_lists[hi], buffer, entry);
 385                         hammer_cache_del(&buffer->cache);
 386                         free(buffer->ondisk);
 387                         free(buffer);
 388                 }
 389         }
 390 }
 391
 392 /*
 393  * Retrieve a pointer to a buffer data given a buffer offset.  The underlying
 394  * bufferp is freed if isnew or the offset is out of range of the cached data.
 395  * If bufferp is freed a referenced buffer is loaded into it.
 396  */
 397 void *
 398 get_buffer_data(hammer_off_t buf_offset, struct buffer_info **bufferp,
 399                 int isnew)
 400 {
 401         if (*bufferp != NULL) {
 402                 if (isnew > 0 ||
 403                     (((*bufferp)->buf_offset ^ buf_offset) & ~HAMMER_BUFMASK64)) {
 404                         rel_buffer(*bufferp);
 405                         *bufferp = NULL;
 406                 }
 407         }
 408         return(get_ondisk(buf_offset, bufferp, isnew));
 409 }
 410
 411 /*
 412  * Retrieve a pointer to a B-Tree node given a zone offset.  The underlying
 413  * bufferp is freed if non-NULL and a referenced buffer is loaded into it.
 414  */
 415 hammer_node_ondisk_t
 416 get_node(hammer_off_t node_offset, struct buffer_info **bufferp)
 417 {
 418         if (*bufferp != NULL) {
 419                 rel_buffer(*bufferp);
 420                 *bufferp = NULL;
 421         }
 422         return(get_ondisk(node_offset, bufferp, 0));
 423 }
 424
 425 /*
 426  * Return a pointer to a buffer data given a buffer offset.
 427  * If *bufferp is NULL acquire the buffer otherwise use that buffer.
 428  */
 429 static void *
 430 get_ondisk(hammer_off_t buf_offset, struct buffer_info **bufferp, int isnew)
 431 {
 432         if (*bufferp == NULL) {
 433                 *bufferp = get_buffer(buf_offset, isnew);
 434                 if (*bufferp == NULL)
 435                         return(NULL);
 436         }
 437
 438         return(((char *)(*bufferp)->ondisk) +
 439                 ((int32_t)buf_offset & HAMMER_BUFMASK));
 440 }
 441
 442 /*
 443  * Allocate HAMMER elements - B-Tree nodes
 444  */
 445 void *
 446 alloc_btree_element(hammer_off_t *offp, struct buffer_info **data_bufferp)
 447 {
 448         hammer_node_ondisk_t node;
 449
 450         node = alloc_blockmap(HAMMER_ZONE_BTREE_INDEX, sizeof(*node),
 451                               offp, data_bufferp);
 452         bzero(node, sizeof(*node));
 453         return (node);
 454 }
 455
 456 /*
 457  * Allocate HAMMER elements - meta data (inode, direntry, PFS, etc)
 458  */
 459 void *
 460 alloc_meta_element(hammer_off_t *offp, int32_t data_len,
 461                    struct buffer_info **data_bufferp)
 462 {
 463         void *data;
 464
 465         data = alloc_blockmap(HAMMER_ZONE_META_INDEX, data_len,
 466                               offp, data_bufferp);
 467         bzero(data, data_len);
 468         return (data);
 469 }
 470
 471 /*
 472  * Allocate HAMMER elements - data storage
 473  *
 474  * The only data_len supported by HAMMER userspace for large data zone
 475  * (zone 10) is HAMMER_BUFSIZE which is 16KB.  >16KB data does not fit
 476  * in a buffer allocated by get_buffer().  Also alloc_blockmap() does
 477  * not consider >16KB buffer size.
 478  */
 479 void *
 480 alloc_data_element(hammer_off_t *offp, int32_t data_len,
 481                    struct buffer_info **data_bufferp)
 482 {
 483         void *data;
 484         int zone;
 485
 486         if (data_len == 0)
 487                 return(NULL);
 488
 489         zone = hammer_data_zone_index(data_len);
 490         assert(data_len <= HAMMER_BUFSIZE); /* just one buffer */
 491         assert(zone == HAMMER_ZONE_LARGE_DATA_INDEX ||
 492                zone == HAMMER_ZONE_SMALL_DATA_INDEX);
 493
 494         data = alloc_blockmap(zone, data_len, offp, data_bufferp);
 495         bzero(data, data_len);
 496         return(data);
 497 }
 498
 499 /*
 500  * Format a new blockmap.  This is mostly a degenerate case because
 501  * all allocations are now actually done from the freemap.
 502  */
 503 void
 504 format_blockmap(struct volume_info *root_vol, int zone, hammer_off_t offset)
 505 {
 506         hammer_blockmap_t blockmap;
 507         hammer_off_t zone_base;
 508
 509         /* Only root volume needs formatting */
 510         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 511
 512         assert(hammer_is_zone2_mapped_index(zone));
 513
 514         blockmap = &root_vol->ondisk->vol0_blockmap[zone];
 515         zone_base = HAMMER_ZONE_ENCODE(zone, offset);
 516
 517         bzero(blockmap, sizeof(*blockmap));
 518         blockmap->phys_offset = 0;
 519         blockmap->first_offset = zone_base;
 520         blockmap->next_offset = zone_base;
 521         blockmap->alloc_offset = HAMMER_ENCODE(zone, 255, -1);
 522         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 523 }
 524
 525 /*
 526  * Format a new freemap.  Set all layer1 entries to UNAVAIL.  The initialize
 527  * code will load each volume's freemap.
 528  */
 529 void
 530 format_freemap(struct volume_info *root_vol)
 531 {
 532         struct buffer_info *buffer = NULL;
 533         hammer_off_t layer1_offset;
 534         hammer_blockmap_t blockmap;
 535         hammer_blockmap_layer1_t layer1;
 536         int i, isnew;
 537
 538         /* Only root volume needs formatting */
 539         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 540
 541         layer1_offset = alloc_bigblock(root_vol, HAMMER_ZONE_FREEMAP_INDEX);
 542         for (i = 0; i < HAMMER_BIGBLOCK_SIZE; i += sizeof(*layer1)) {
 543                 isnew = ((i % HAMMER_BUFSIZE) == 0);
 544                 layer1 = get_buffer_data(layer1_offset + i, &buffer, isnew);
 545                 bzero(layer1, sizeof(*layer1));
 546                 layer1->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 547                 layer1->blocks_free = 0;
 548                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 549         }
 550         assert(i == HAMMER_BIGBLOCK_SIZE);
 551         rel_buffer(buffer);
 552
 553         blockmap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 554         bzero(blockmap, sizeof(*blockmap));
 555         blockmap->phys_offset = layer1_offset;
 556         blockmap->first_offset = 0;
 557         blockmap->next_offset = HAMMER_ENCODE_RAW_BUFFER(0, 0);
 558         blockmap->alloc_offset = HAMMER_ENCODE_RAW_BUFFER(255, -1);
 559         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 560 }
 561
 562 /*
 563  * Load the volume's remaining free space into the freemap.
 564  *
 565  * Returns the number of big-blocks available.
 566  */
 567 int64_t
 568 initialize_freemap(struct volume_info *vol)
 569 {
 570         struct volume_info *root_vol;
 571         struct buffer_info *buffer1 = NULL;
 572         struct buffer_info *buffer2 = NULL;
 573         hammer_blockmap_layer1_t layer1;
 574         hammer_blockmap_layer2_t layer2;
 575         hammer_off_t layer1_offset;
 576         hammer_off_t layer2_offset;
 577         hammer_off_t phys_offset;
 578         hammer_off_t block_offset;
 579         hammer_off_t aligned_vol_free_end;
 580         hammer_blockmap_t freemap;
 581         int64_t count = 0;
 582         int64_t layer1_count = 0;
 583
 584         root_vol = get_root_volume();
 585         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 586                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 587
 588         printf("initialize freemap volume %d\n", vol->vol_no);
 589
 590         /*
 591          * Initialize the freemap.  First preallocate the big-blocks required
 592          * to implement layer2.   This preallocation is a bootstrap allocation
 593          * using blocks from the target volume.
 594          */
 595         freemap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 596
 597         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 598              phys_offset < aligned_vol_free_end;
 599              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 600                 layer1_offset = freemap->phys_offset +
 601                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 602                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 603                 if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 604                         layer1->phys_offset = alloc_bigblock(vol,
 605                                                 HAMMER_ZONE_FREEMAP_INDEX);
 606                         layer1->blocks_free = 0;
 607                         buffer1->cache.modified = 1;
 608                         layer1->layer1_crc = crc32(layer1,
 609                                                    HAMMER_LAYER1_CRCSIZE);
 610                 }
 611         }
 612
 613         /*
 614          * Now fill everything in.
 615          */
 616         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 617              phys_offset < aligned_vol_free_end;
 618              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 619                 layer1_count = 0;
 620                 layer1_offset = freemap->phys_offset +
 621                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 622                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 623                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 624
 625                 for (block_offset = 0;
 626                      block_offset < HAMMER_BLOCKMAP_LAYER2;
 627                      block_offset += HAMMER_BIGBLOCK_SIZE) {
 628                         layer2_offset = layer1->phys_offset +
 629                                         HAMMER_BLOCKMAP_LAYER2_OFFSET(block_offset);
 630                         layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 631                         bzero(layer2, sizeof(*layer2));
 632
 633                         if (phys_offset + block_offset < vol->vol_free_off) {
 634                                 /*
 635                                  * Fixups XXX - big-blocks already allocated as part
 636                                  * of the freemap bootstrap.
 637                                  */
 638                                 layer2->zone = HAMMER_ZONE_FREEMAP_INDEX;
 639                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 640                                 layer2->bytes_free = 0;
 641                         } else if (phys_offset + block_offset < vol->vol_free_end) {
 642                                 layer2->zone = 0;
 643                                 layer2->append_off = 0;
 644                                 layer2->bytes_free = HAMMER_BIGBLOCK_SIZE;
 645                                 ++count;
 646                                 ++layer1_count;
 647                         } else {
 648                                 layer2->zone = HAMMER_ZONE_UNAVAIL_INDEX;
 649                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 650                                 layer2->bytes_free = 0;
 651                         }
 652                         layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 653                         buffer2->cache.modified = 1;
 654                 }
 655
 656                 layer1->blocks_free += layer1_count;
 657                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 658                 buffer1->cache.modified = 1;
 659         }
 660
 661         rel_buffer(buffer1);
 662         rel_buffer(buffer2);
 663         return(count);
 664 }
 665
 666 /*
 667  * Returns the number of big-blocks available for filesystem data and undos
 668  * without formatting.
 669  */
 670 int64_t
 671 count_freemap(struct volume_info *vol)
 672 {
 673         hammer_off_t phys_offset;
 674         hammer_off_t vol_free_off;
 675         hammer_off_t aligned_vol_free_end;
 676         int64_t count = 0;
 677
 678         vol_free_off = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 679         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 680                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 681
 682         if (vol->vol_no == HAMMER_ROOT_VOLNO)
 683                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 684
 685         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 686              phys_offset < aligned_vol_free_end;
 687              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 688                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 689         }
 690
 691         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 692              phys_offset < aligned_vol_free_end;
 693              phys_offset += HAMMER_BIGBLOCK_SIZE) {
 694                 if (phys_offset < vol_free_off) {
 695                         ;
 696                 } else if (phys_offset < vol->vol_free_end) {
 697                         ++count;
 698                 }
 699         }
 700
 701         return(count);
 702 }
 703
 704 /*
 705  * Format the undomap for the root volume.
 706  */
 707 void
 708 format_undomap(struct volume_info *root_vol, int64_t *undo_buffer_size)
 709 {
 710         const int undo_zone = HAMMER_ZONE_UNDO_INDEX;
 711         hammer_off_t undo_limit;
 712         hammer_blockmap_t blockmap;
 713         hammer_volume_ondisk_t ondisk;
 714         struct buffer_info *buffer = NULL;
 715         hammer_off_t scan;
 716         int n;
 717         int limit_index;
 718         uint32_t seqno;
 719
 720         /* Only root volume needs formatting */
 721         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 722         ondisk = root_vol->ondisk;
 723
 724         /*
 725          * Size the undo buffer in multiples of HAMMER_BIGBLOCK_SIZE,
 726          * up to HAMMER_UNDO_LAYER2 big-blocks.  Size to approximately
 727          * 0.1% of the disk.
 728          *
 729          * The minimum UNDO fifo size is 500MB, or approximately 1% of
 730          * the recommended 50G disk.
 731          *
 732          * Changing this minimum is rather dangerous as complex filesystem
 733          * operations can cause the UNDO FIFO to fill up otherwise.
 734          */
 735         undo_limit = *undo_buffer_size;
 736         if (undo_limit == 0) {
 737                 undo_limit = HAMMER_VOL_BUF_SIZE(ondisk) / 1000;
 738                 if (undo_limit < 500*1024*1024)
 739                         undo_limit = 500*1024*1024;
 740         }
 741         undo_limit = (undo_limit + HAMMER_BIGBLOCK_MASK64) &
 742                      ~HAMMER_BIGBLOCK_MASK64;
 743         if (undo_limit < HAMMER_BIGBLOCK_SIZE)
 744                 undo_limit = HAMMER_BIGBLOCK_SIZE;
 745         if (undo_limit > HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2)
 746                 undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2;
 747         *undo_buffer_size = undo_limit;
 748
 749         blockmap = &ondisk->vol0_blockmap[undo_zone];
 750         bzero(blockmap, sizeof(*blockmap));
 751         blockmap->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 752         blockmap->first_offset = HAMMER_ZONE_ENCODE(undo_zone, 0);
 753         blockmap->next_offset = blockmap->first_offset;
 754         blockmap->alloc_offset = HAMMER_ZONE_ENCODE(undo_zone, undo_limit);
 755         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 756
 757         limit_index = undo_limit / HAMMER_BIGBLOCK_SIZE;
 758         assert(limit_index <= HAMMER_UNDO_LAYER2);
 759
 760         for (n = 0; n < limit_index; ++n) {
 761                 ondisk->vol0_undo_array[n] = alloc_bigblock(root_vol,
 762                                                         HAMMER_ZONE_UNDO_INDEX);
 763         }
 764         while (n < HAMMER_UNDO_LAYER2) {
 765                 ondisk->vol0_undo_array[n++] = HAMMER_BLOCKMAP_UNAVAIL;
 766         }
 767
 768         /*
 769          * Pre-initialize the UNDO blocks (HAMMER version 4+)
 770          */
 771         printf("initializing the undo map (%jd MB)\n",
 772                 (intmax_t)(blockmap->alloc_offset & HAMMER_OFF_LONG_MASK) /
 773                 (1024 * 1024));
 774
 775         scan = blockmap->first_offset;
 776         seqno = 0;
 777
 778         while (scan < blockmap->alloc_offset) {
 779                 hammer_fifo_head_t head;
 780                 hammer_fifo_tail_t tail;
 781                 int isnew;
 782                 int bytes = HAMMER_UNDO_ALIGN;
 783
 784                 isnew = ((scan & HAMMER_BUFMASK64) == 0);
 785                 head = get_buffer_data(scan, &buffer, isnew);
 786                 buffer->cache.modified = 1;
 787                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 788
 789                 bzero(head, bytes);
 790                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 791                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
 792                 head->hdr_size = bytes;
 793                 head->hdr_seq = seqno++;
 794
 795                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 796                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 797                 tail->tail_size = bytes;
 798
 799                 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
 800                                 crc32(head + 1, bytes - sizeof(*head));
 801
 802                 scan += bytes;
 803         }
 804         rel_buffer(buffer);
 805 }
 806
 807 const char *zone_labels[] = {
 808         "",             /* 0 */
 809         "raw_volume",   /* 1 */
 810         "raw_buffer",   /* 2 */
 811         "undo",         /* 3 */
 812         "freemap",      /* 4 */
 813         "",             /* 5 */
 814         "",             /* 6 */
 815         "",             /* 7 */
 816         "btree",        /* 8 */
 817         "meta",         /* 9 */
 818         "large_data",   /* 10 */
 819         "small_data",   /* 11 */
 820         "",             /* 12 */
 821         "",             /* 13 */
 822         "",             /* 14 */
 823         "unavail",      /* 15 */
 824 };
 825
 826 void
 827 print_blockmap(const struct volume_info *root_vol)
 828 {
 829         hammer_blockmap_t blockmap;
 830         hammer_volume_ondisk_t ondisk;
 831         int64_t size, used;
 832         int i;
 833 #define INDENT ""
 834
 835         ondisk = root_vol->ondisk;
 836         printf(INDENT"vol_label\t%s\n", ondisk->vol_label);
 837         printf(INDENT"vol_count\t%d\n", ondisk->vol_count);
 838         printf(INDENT"vol_bot_beg\t%s\n", sizetostr(ondisk->vol_bot_beg));
 839         printf(INDENT"vol_mem_beg\t%s\n", sizetostr(ondisk->vol_mem_beg));
 840         printf(INDENT"vol_buf_beg\t%s\n", sizetostr(ondisk->vol_buf_beg));
 841         printf(INDENT"vol_buf_end\t%s\n", sizetostr(ondisk->vol_buf_end));
 842         printf(INDENT"vol0_next_tid\t%016jx\n",
 843                (uintmax_t)ondisk->vol0_next_tid);
 844
 845         blockmap = &ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 846         size = blockmap->alloc_offset & HAMMER_OFF_LONG_MASK;
 847         if (blockmap->first_offset <= blockmap->next_offset)
 848                 used = blockmap->next_offset - blockmap->first_offset;
 849         else
 850                 used = blockmap->alloc_offset - blockmap->first_offset +
 851                         (blockmap->next_offset & HAMMER_OFF_LONG_MASK);
 852         printf(INDENT"undo_size\t%s\n", sizetostr(size));
 853         printf(INDENT"undo_used\t%s\n", sizetostr(used));
 854
 855         printf(INDENT"zone #             "
 856                "phys             first            next             alloc\n");
 857         for (i = 0; i < HAMMER_MAX_ZONES; i++) {
 858                 blockmap = &ondisk->vol0_blockmap[i];
 859                 printf(INDENT"zone %-2d %-10s %016jx %016jx %016jx %016jx\n",
 860                         i, zone_labels[i],
 861                         (uintmax_t)blockmap->phys_offset,
 862                         (uintmax_t)blockmap->first_offset,
 863                         (uintmax_t)blockmap->next_offset,
 864                         (uintmax_t)blockmap->alloc_offset);
 865         }
 866 }
 867
 868 /*
 869  * Flush various tracking structures to disk
 870  */
 871 void
 872 flush_all_volumes(void)
 873 {
 874         struct volume_info *vol;
 875
 876         TAILQ_FOREACH(vol, &VolList, entry)
 877                 flush_volume(vol);
 878 }
 879
 880 void
 881 flush_volume(struct volume_info *volume)
 882 {
 883         struct buffer_info *buffer;
 884         int i;
 885
 886         for (i = 0; i < HAMMER_BUFLISTS; ++i) {
 887                 TAILQ_FOREACH(buffer, &volume->buffer_lists[i], entry)
 888                         flush_buffer(buffer);
 889         }
 890         if (writehammervol(volume) == -1)
 891                 err(1, "Write volume %d (%s)", volume->vol_no, volume->name);
 892 }
 893
 894 void
 895 flush_buffer(struct buffer_info *buffer)
 896 {
 897         struct volume_info *vol;
 898
 899         vol = buffer->volume;
 900         if (writehammerbuf(buffer) == -1)
 901                 err(1, "Write volume %d (%s)", vol->vol_no, vol->name);
 902         buffer->cache.modified = 0;
 903 }
 904
 905 /*
 906  * Core I/O operations
 907  */
 908 static int
 909 __read(struct volume_info *vol, void *data, int64_t offset, int size)
 910 {
 911         ssize_t n;
 912
 913         n = pread(vol->fd, data, size, offset);
 914         if (n != size)
 915                 return(-1);
 916         return(0);
 917 }
 918
 919 static __inline int
 920 readhammervol(struct volume_info *vol)
 921 {
 922         return(__read(vol, vol->ondisk, 0, HAMMER_BUFSIZE));
 923 }
 924
 925 static __inline int
 926 readhammerbuf(struct buffer_info *buf)
 927 {
 928         return(__read(buf->volume, buf->ondisk, buf->raw_offset, HAMMER_BUFSIZE));
 929 }
 930
 931 static int
 932 __write(struct volume_info *vol, const void *data, int64_t offset, int size)
 933 {
 934         ssize_t n;
 935
 936         if (vol->rdonly)
 937                 return(0);
 938
 939         n = pwrite(vol->fd, data, size, offset);
 940         if (n != size)
 941                 return(-1);
 942         return(0);
 943 }
 944
 945 static __inline int
 946 writehammervol(struct volume_info *vol)
 947 {
 948         return(__write(vol, vol->ondisk, 0, HAMMER_BUFSIZE));
 949 }
 950
 951 static __inline int
 952 writehammerbuf(struct buffer_info *buf)
 953 {
 954         return(__write(buf->volume, buf->ondisk, buf->raw_offset, HAMMER_BUFSIZE));
 955 }
 956
 957 int64_t init_boot_area_size(int64_t value, off_t avg_vol_size)
 958 {
 959         if (value == 0) {
 960                 value = HAMMER_BOOT_NOMBYTES;
 961                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 962                         value >>= 1;
 963                 if (value < HAMMER_BOOT_MINBYTES)
 964                         value = 0;
 965         } else if (value < HAMMER_BOOT_MINBYTES) {
 966                 value = HAMMER_BOOT_MINBYTES;
 967         }
 968
 969         return(value);
 970 }
 971
 972 int64_t init_mem_area_size(int64_t value, off_t avg_vol_size)
 973 {
 974         if (value == 0) {
 975                 value = HAMMER_MEM_NOMBYTES;
 976                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 977                         value >>= 1;
 978                 if (value < HAMMER_MEM_MINBYTES)
 979                         value = 0;
 980         } else if (value < HAMMER_MEM_MINBYTES) {
 981                 value = HAMMER_MEM_MINBYTES;
 982         }
 983
 984         return(value);
 985 }