sbin/hammer/ondisk.c

   1 /*
   2  * Copyright (c) 2007 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  */
  34
  35 #include <sys/diskslice.h>
  36 #include <sys/diskmbr.h>
  37
  38 #include "hammer_util.h"
  39
  40 static void check_volume(struct volume_info *vol);
  41 static void get_buffer_readahead(struct buffer_info *base);
  42 static __inline int readhammervol(struct volume_info *vol);
  43 static __inline int readhammerbuf(struct buffer_info *buf);
  44 static __inline int writehammervol(struct volume_info *vol);
  45 static __inline int writehammerbuf(struct buffer_info *buf);
  46
  47 uuid_t Hammer_FSType;
  48 uuid_t Hammer_FSId;
  49 int UseReadBehind = -4;
  50 int UseReadAhead = 4;
  51 int DebugOpt;
  52
  53 TAILQ_HEAD(volume_list, volume_info);
  54 static struct volume_list VolList = TAILQ_HEAD_INITIALIZER(VolList);
  55 static int valid_hammer_volumes;
  56
  57 static __inline
  58 int
  59 buffer_hash(hammer_off_t buf_offset)
  60 {
  61         int hi;
  62
  63         hi = (int)(buf_offset / HAMMER_BUFSIZE) & HAMMER_BUFLISTMASK;
  64         return(hi);
  65 }
  66
  67 static struct buffer_info*
  68 find_buffer(hammer_off_t buf_offset)
  69 {
  70         struct volume_info *volume;
  71         struct buffer_info *buf;
  72         int hi;
  73
  74         volume = get_volume(HAMMER_VOL_DECODE(buf_offset));
  75         assert(volume);
  76
  77         hi = buffer_hash(buf_offset);
  78         TAILQ_FOREACH(buf, &volume->buffer_lists[hi], entry)
  79                 if (buf->buf_offset == buf_offset)
  80                         return(buf);
  81         return(NULL);
  82 }
  83
  84 static
  85 struct volume_info *
  86 __alloc_volume(const char *volname, int oflags)
  87 {
  88         struct volume_info *vol;
  89         int i;
  90
  91         vol = malloc(sizeof(*vol));
  92         if (vol == NULL)
  93                 err(1, "alloc_volume");
  94         bzero(vol, sizeof(*vol));
  95
  96         vol->vol_no = -1;
  97         vol->rdonly = (oflags == O_RDONLY);
  98         vol->name = strdup(volname);
  99         vol->fd = open(vol->name, oflags);
 100         if (vol->fd < 0)
 101                 err(1, "alloc_volume: Failed to open %s", vol->name);
 102         check_volume(vol);
 103
 104         vol->ondisk = malloc(HAMMER_BUFSIZE);
 105         if (vol->ondisk == NULL)
 106                 err(1, "alloc_volume");
 107         bzero(vol->ondisk, HAMMER_BUFSIZE);
 108
 109         for (i = 0; i < HAMMER_BUFLISTS; ++i)
 110                 TAILQ_INIT(&vol->buffer_lists[i]);
 111
 112         return(vol);
 113 }
 114
 115 static void
 116 __add_volume(struct volume_info *vol)
 117 {
 118         struct volume_info *scan;
 119         struct stat st1, st2;
 120
 121         if (fstat(vol->fd, &st1) != 0)
 122                 errx(1, "add_volume: %s: Failed to stat", vol->name);
 123
 124         TAILQ_FOREACH(scan, &VolList, entry) {
 125                 if (scan->vol_no == vol->vol_no) {
 126                         errx(1, "add_volume: %s: Duplicate volume number %d "
 127                                 "against %s",
 128                                 vol->name, vol->vol_no, scan->name);
 129                 }
 130                 if (fstat(scan->fd, &st2) != 0) {
 131                         errx(1, "add_volume: %s: Failed to stat %s",
 132                                 vol->name, scan->name);
 133                 }
 134                 if ((st1.st_ino == st2.st_ino) && (st1.st_dev == st2.st_dev)) {
 135                         errx(1, "add_volume: %s: Specified more than once",
 136                                 vol->name);
 137                 }
 138         }
 139
 140         TAILQ_INSERT_TAIL(&VolList, vol, entry);
 141 }
 142
 143 static void
 144 __verify_volume(struct volume_info *vol)
 145 {
 146         hammer_volume_ondisk_t ondisk = vol->ondisk;
 147
 148         if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
 149                 errx(1, "verify_volume: Invalid volume signature %016jx",
 150                         ondisk->vol_signature);
 151         }
 152         if (ondisk->vol_rootvol != HAMMER_ROOT_VOLNO) {
 153                 errx(1, "verify_volume: Invalid root volume# %d",
 154                         ondisk->vol_rootvol);
 155         }
 156         if (bcmp(&Hammer_FSType, &ondisk->vol_fstype, sizeof(Hammer_FSType))) {
 157                 errx(1, "verify_volume: %s: Header does not indicate "
 158                         "that this is a HAMMER volume", vol->name);
 159         }
 160         if (bcmp(&Hammer_FSId, &ondisk->vol_fsid, sizeof(Hammer_FSId))) {
 161                 errx(1, "verify_volume: %s: FSId does not match other volumes!",
 162                         vol->name);
 163         }
 164 }
 165
 166 /*
 167  * Initialize a volume structure and ondisk vol_no field.
 168  */
 169 struct volume_info *
 170 init_volume(const char *filename, int oflags, int32_t vol_no)
 171 {
 172         struct volume_info *vol;
 173
 174         vol = __alloc_volume(filename, oflags);
 175         vol->vol_no = vol->ondisk->vol_no = vol_no;
 176
 177         __add_volume(vol);
 178
 179         return(vol);
 180 }
 181
 182 /*
 183  * Initialize a volume structure and read ondisk volume header.
 184  */
 185 struct volume_info*
 186 load_volume(const char *filename, int oflags, int verify)
 187 {
 188         struct volume_info *vol;
 189         int n;
 190
 191         vol = __alloc_volume(filename, oflags);
 192
 193         n = readhammervol(vol);
 194         if (n == -1) {
 195                 err(1, "load_volume: %s: Read failed at offset 0", vol->name);
 196         }
 197         vol->vol_no = vol->ondisk->vol_no;
 198
 199         if (valid_hammer_volumes++ == 0)
 200                 Hammer_FSId = vol->ondisk->vol_fsid;
 201         if (verify)
 202                 __verify_volume(vol);
 203
 204         __add_volume(vol);
 205
 206         return(vol);
 207 }
 208
 209 /*
 210  * Check basic volume characteristics.
 211  */
 212 static void
 213 check_volume(struct volume_info *vol)
 214 {
 215         struct partinfo pinfo;
 216         struct stat st;
 217
 218         /*
 219          * Get basic information about the volume
 220          */
 221         if (ioctl(vol->fd, DIOCGPART, &pinfo) < 0) {
 222                 /*
 223                  * Allow the formatting of regular files as HAMMER volumes
 224                  */
 225                 if (fstat(vol->fd, &st) < 0)
 226                         err(1, "Unable to stat %s", vol->name);
 227                 vol->size = st.st_size;
 228                 vol->type = "REGFILE";
 229         } else {
 230                 /*
 231                  * When formatting a block device as a HAMMER volume the
 232                  * sector size must be compatible.  HAMMER uses 16384 byte
 233                  * filesystem buffers.
 234                  */
 235                 if (pinfo.reserved_blocks) {
 236                         errx(1, "HAMMER cannot be placed in a partition "
 237                                 "which overlaps the disklabel or MBR");
 238                 }
 239                 if (pinfo.media_blksize > HAMMER_BUFSIZE ||
 240                     HAMMER_BUFSIZE % pinfo.media_blksize) {
 241                         errx(1, "A media sector size of %d is not supported",
 242                              pinfo.media_blksize);
 243                 }
 244
 245                 vol->size = pinfo.media_size;
 246                 vol->device_offset = pinfo.media_offset;
 247                 vol->type = "DEVICE";
 248         }
 249 }
 250
 251 void
 252 assert_volume_offset(struct volume_info *vol)
 253 {
 254         assert(hammer_is_zone_raw_buffer(vol->vol_free_off));
 255         assert(hammer_is_zone_raw_buffer(vol->vol_free_end));
 256         if (vol->vol_free_off >= vol->vol_free_end)
 257                 errx(1, "Ran out of room, filesystem too small");
 258 }
 259
 260 struct volume_info *
 261 get_volume(int32_t vol_no)
 262 {
 263         struct volume_info *vol;
 264
 265         TAILQ_FOREACH(vol, &VolList, entry) {
 266                 if (vol->vol_no == vol_no)
 267                         break;
 268         }
 269
 270         return(vol);
 271 }
 272
 273 struct volume_info *
 274 get_root_volume(void)
 275 {
 276         return(get_volume(HAMMER_ROOT_VOLNO));
 277 }
 278
 279 static hammer_off_t
 280 __blockmap_xlate_to_zone2(hammer_off_t zone_offset)
 281 {
 282         hammer_off_t buf_offset;
 283         int error = 0;
 284
 285         if (hammer_is_zone_raw_buffer(zone_offset))
 286                 buf_offset = zone_offset;
 287         else
 288                 buf_offset = blockmap_lookup(zone_offset, &error);
 289
 290         if (error)
 291                 return(HAMMER_OFF_BAD);
 292         assert(hammer_is_zone_raw_buffer(buf_offset));
 293
 294         return(buf_offset);
 295 }
 296
 297 static struct buffer_info *
 298 __alloc_buffer(hammer_off_t buf_offset, int isnew)
 299 {
 300         struct volume_info *volume;
 301         struct buffer_info *buf;
 302         int hi;
 303
 304         volume = get_volume(HAMMER_VOL_DECODE(buf_offset));
 305         assert(volume != NULL);
 306
 307         buf = calloc(1, sizeof(*buf));
 308         buf->buf_offset = buf_offset;
 309         buf->raw_offset = hammer_xlate_to_phys(volume->ondisk, buf_offset);
 310         buf->volume = volume;
 311         buf->ondisk = calloc(1, HAMMER_BUFSIZE);
 312
 313         if (isnew <= 0) {
 314                 if (readhammerbuf(buf) == -1) {
 315                         err(1, "Failed to read %s:%016jx at %016jx",
 316                             volume->name,
 317                             (intmax_t)buf->buf_offset,
 318                             (intmax_t)buf->raw_offset);
 319                 }
 320         }
 321
 322         hi = buffer_hash(buf_offset);
 323         TAILQ_INSERT_TAIL(&volume->buffer_lists[hi], buf, entry);
 324         hammer_cache_add(&buf->cache);
 325
 326         return(buf);
 327 }
 328
 329 /*
 330  * Acquire the 16KB buffer for specified zone offset.
 331  */
 332 static struct buffer_info *
 333 get_buffer(hammer_off_t zone_offset, int isnew)
 334 {
 335         struct buffer_info *buf;
 336         hammer_off_t buf_offset;
 337         int dora = 0;
 338
 339         buf_offset = __blockmap_xlate_to_zone2(zone_offset);
 340         if (buf_offset == HAMMER_OFF_BAD)
 341                 return(NULL);
 342
 343         buf_offset &= ~HAMMER_BUFMASK64;
 344         buf = find_buffer(buf_offset);
 345
 346         if (buf == NULL) {
 347                 buf = __alloc_buffer(buf_offset, isnew);
 348                 dora = (isnew == 0);
 349         } else {
 350                 assert(isnew != -1);
 351                 hammer_cache_used(&buf->cache);
 352         }
 353         assert(buf->ondisk != NULL);
 354
 355         ++buf->cache.refs;
 356         hammer_cache_flush();
 357
 358         if (isnew > 0) {
 359                 assert(buf->cache.modified == 0);
 360                 bzero(buf->ondisk, HAMMER_BUFSIZE);
 361                 buf->cache.modified = 1;
 362         }
 363         if (dora)
 364                 get_buffer_readahead(buf);
 365         return(buf);
 366 }
 367
 368 static void
 369 get_buffer_readahead(struct buffer_info *base)
 370 {
 371         struct buffer_info *buf;
 372         struct volume_info *vol;
 373         hammer_off_t buf_offset;
 374         int64_t raw_offset;
 375         int ri = UseReadBehind;
 376         int re = UseReadAhead;
 377
 378         raw_offset = base->raw_offset + ri * HAMMER_BUFSIZE;
 379         vol = base->volume;
 380
 381         while (ri < re) {
 382                 if (raw_offset >= vol->ondisk->vol_buf_end)
 383                         break;
 384                 if (raw_offset < vol->ondisk->vol_buf_beg || ri == 0) {
 385                         ++ri;
 386                         raw_offset += HAMMER_BUFSIZE;
 387                         continue;
 388                 }
 389                 buf_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no,
 390                         raw_offset - vol->ondisk->vol_buf_beg);
 391                 buf = find_buffer(buf_offset);
 392                 if (buf == NULL) {
 393                         /* call with -1 to prevent another readahead */
 394                         buf = get_buffer(buf_offset, -1);
 395                         rel_buffer(buf);
 396                 }
 397                 ++ri;
 398                 raw_offset += HAMMER_BUFSIZE;
 399         }
 400 }
 401
 402 void
 403 rel_buffer(struct buffer_info *buffer)
 404 {
 405         struct volume_info *volume;
 406         int hi;
 407
 408         if (buffer == NULL)
 409                 return;
 410         assert(buffer->cache.refs > 0);
 411         if (--buffer->cache.refs == 0) {
 412                 if (buffer->cache.delete) {
 413                         hi = buffer_hash(buffer->buf_offset);
 414                         volume = buffer->volume;
 415                         if (buffer->cache.modified)
 416                                 flush_buffer(buffer);
 417                         TAILQ_REMOVE(&volume->buffer_lists[hi], buffer, entry);
 418                         hammer_cache_del(&buffer->cache);
 419                         free(buffer->ondisk);
 420                         free(buffer);
 421                 }
 422         }
 423 }
 424
 425 /*
 426  * Retrieve a pointer to a buffer data given a buffer offset.  The underlying
 427  * bufferp is freed if isnew or the offset is out of range of the cached data.
 428  * If bufferp is freed a referenced buffer is loaded into it.
 429  */
 430 void *
 431 get_buffer_data(hammer_off_t buf_offset, struct buffer_info **bufferp,
 432                 int isnew)
 433 {
 434         if (*bufferp != NULL) {
 435                 if (isnew > 0 ||
 436                     (((*bufferp)->buf_offset ^ buf_offset) & ~HAMMER_BUFMASK64)) {
 437                         rel_buffer(*bufferp);
 438                         *bufferp = NULL;
 439                 }
 440         }
 441
 442         if (*bufferp == NULL) {
 443                 *bufferp = get_buffer(buf_offset, isnew);
 444                 if (*bufferp == NULL)
 445                         return(NULL);
 446         }
 447
 448         return(((char *)(*bufferp)->ondisk) +
 449                 ((int32_t)buf_offset & HAMMER_BUFMASK));
 450 }
 451
 452 /*
 453  * Allocate HAMMER elements - B-Tree nodes
 454  */
 455 hammer_node_ondisk_t
 456 alloc_btree_node(hammer_off_t *offp, struct buffer_info **data_bufferp)
 457 {
 458         hammer_node_ondisk_t node;
 459
 460         node = alloc_blockmap(HAMMER_ZONE_BTREE_INDEX, sizeof(*node),
 461                               offp, data_bufferp);
 462         bzero(node, sizeof(*node));
 463         return(node);
 464 }
 465
 466 /*
 467  * Allocate HAMMER elements - meta data (inode, direntry, PFS, etc)
 468  */
 469 void *
 470 alloc_meta_element(hammer_off_t *offp, int32_t data_len,
 471                    struct buffer_info **data_bufferp)
 472 {
 473         void *data;
 474
 475         data = alloc_blockmap(HAMMER_ZONE_META_INDEX, data_len,
 476                               offp, data_bufferp);
 477         bzero(data, data_len);
 478         return(data);
 479 }
 480
 481 /*
 482  * Format a new blockmap.  This is mostly a degenerate case because
 483  * all allocations are now actually done from the freemap.
 484  */
 485 void
 486 format_blockmap(struct volume_info *root_vol, int zone, hammer_off_t offset)
 487 {
 488         hammer_blockmap_t blockmap;
 489         hammer_off_t zone_base;
 490
 491         /* Only root volume needs formatting */
 492         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 493
 494         assert(hammer_is_zone2_mapped_index(zone));
 495
 496         blockmap = &root_vol->ondisk->vol0_blockmap[zone];
 497         zone_base = HAMMER_ZONE_ENCODE(zone, offset);
 498
 499         bzero(blockmap, sizeof(*blockmap));
 500         blockmap->phys_offset = 0;
 501         blockmap->first_offset = zone_base;
 502         blockmap->next_offset = zone_base;
 503         blockmap->alloc_offset = HAMMER_ENCODE(zone, 255, -1);
 504         hammer_crc_set_blockmap(blockmap);
 505 }
 506
 507 /*
 508  * Format a new freemap.  Set all layer1 entries to UNAVAIL.  The initialize
 509  * code will load each volume's freemap.
 510  */
 511 void
 512 format_freemap(struct volume_info *root_vol)
 513 {
 514         struct buffer_info *buffer = NULL;
 515         hammer_off_t layer1_offset;
 516         hammer_blockmap_t blockmap;
 517         hammer_blockmap_layer1_t layer1;
 518         int i, isnew;
 519
 520         /* Only root volume needs formatting */
 521         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 522
 523         layer1_offset = bootstrap_bigblock(root_vol);
 524         for (i = 0; i < HAMMER_BIGBLOCK_SIZE; i += sizeof(*layer1)) {
 525                 isnew = ((i % HAMMER_BUFSIZE) == 0);
 526                 layer1 = get_buffer_data(layer1_offset + i, &buffer, isnew);
 527                 bzero(layer1, sizeof(*layer1));
 528                 layer1->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 529                 layer1->blocks_free = 0;
 530                 hammer_crc_set_layer1(layer1);
 531         }
 532         assert(i == HAMMER_BIGBLOCK_SIZE);
 533         rel_buffer(buffer);
 534
 535         blockmap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 536         bzero(blockmap, sizeof(*blockmap));
 537         blockmap->phys_offset = layer1_offset;
 538         blockmap->first_offset = 0;
 539         blockmap->next_offset = HAMMER_ENCODE_RAW_BUFFER(0, 0);
 540         blockmap->alloc_offset = HAMMER_ENCODE_RAW_BUFFER(255, -1);
 541         hammer_crc_set_blockmap(blockmap);
 542 }
 543
 544 /*
 545  * Load the volume's remaining free space into the freemap.
 546  *
 547  * Returns the number of big-blocks available.
 548  */
 549 int64_t
 550 initialize_freemap(struct volume_info *vol)
 551 {
 552         struct volume_info *root_vol;
 553         struct buffer_info *buffer1 = NULL;
 554         struct buffer_info *buffer2 = NULL;
 555         hammer_blockmap_layer1_t layer1;
 556         hammer_blockmap_layer2_t layer2;
 557         hammer_off_t layer1_offset;
 558         hammer_off_t layer2_offset;
 559         hammer_off_t phys_offset;
 560         hammer_off_t block_offset;
 561         hammer_off_t aligned_vol_free_end;
 562         hammer_blockmap_t freemap;
 563         int64_t count = 0;
 564         int64_t layer1_count = 0;
 565
 566         root_vol = get_root_volume();
 567
 568         assert_volume_offset(vol);
 569         aligned_vol_free_end = HAMMER_BLOCKMAP_LAYER2_DOALIGN(vol->vol_free_end);
 570
 571         printf("initialize freemap volume %d\n", vol->vol_no);
 572
 573         /*
 574          * Initialize the freemap.  First preallocate the big-blocks required
 575          * to implement layer2.   This preallocation is a bootstrap allocation
 576          * using blocks from the target volume.
 577          */
 578         freemap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 579
 580         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 581              phys_offset < aligned_vol_free_end;
 582              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 583                 layer1_offset = freemap->phys_offset +
 584                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 585                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 586                 if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 587                         layer1->phys_offset = bootstrap_bigblock(vol);
 588                         layer1->blocks_free = 0;
 589                         buffer1->cache.modified = 1;
 590                         hammer_crc_set_layer1(layer1);
 591                 }
 592         }
 593
 594         /*
 595          * Now fill everything in.
 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_count = 0;
 601                 layer1_offset = freemap->phys_offset +
 602                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 603                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 604                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 605
 606                 for (block_offset = 0;
 607                      block_offset < HAMMER_BLOCKMAP_LAYER2;
 608                      block_offset += HAMMER_BIGBLOCK_SIZE) {
 609                         layer2_offset = layer1->phys_offset +
 610                                         HAMMER_BLOCKMAP_LAYER2_OFFSET(block_offset);
 611                         layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 612                         bzero(layer2, sizeof(*layer2));
 613
 614                         if (phys_offset + block_offset < vol->vol_free_off) {
 615                                 /*
 616                                  * Big-blocks already allocated as part
 617                                  * of the freemap bootstrap.
 618                                  */
 619                                 layer2->zone = HAMMER_ZONE_FREEMAP_INDEX;
 620                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 621                                 layer2->bytes_free = 0;
 622                         } else if (phys_offset + block_offset < vol->vol_free_end) {
 623                                 layer2->zone = 0;
 624                                 layer2->append_off = 0;
 625                                 layer2->bytes_free = HAMMER_BIGBLOCK_SIZE;
 626                                 ++count;
 627                                 ++layer1_count;
 628                         } else {
 629                                 layer2->zone = HAMMER_ZONE_UNAVAIL_INDEX;
 630                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 631                                 layer2->bytes_free = 0;
 632                         }
 633                         hammer_crc_set_layer2(layer2);
 634                         buffer2->cache.modified = 1;
 635                 }
 636
 637                 layer1->blocks_free += layer1_count;
 638                 hammer_crc_set_layer1(layer1);
 639                 buffer1->cache.modified = 1;
 640         }
 641
 642         rel_buffer(buffer1);
 643         rel_buffer(buffer2);
 644         return(count);
 645 }
 646
 647 /*
 648  * Returns the number of big-blocks available for filesystem data and undos
 649  * without formatting.
 650  */
 651 int64_t
 652 count_freemap(struct volume_info *vol)
 653 {
 654         hammer_off_t phys_offset;
 655         hammer_off_t vol_free_off;
 656         hammer_off_t aligned_vol_free_end;
 657         int64_t count = 0;
 658
 659         vol_free_off = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 660
 661         assert_volume_offset(vol);
 662         aligned_vol_free_end = HAMMER_BLOCKMAP_LAYER2_DOALIGN(vol->vol_free_end);
 663
 664         if (vol->vol_no == HAMMER_ROOT_VOLNO)
 665                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 666
 667         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 668              phys_offset < aligned_vol_free_end;
 669              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 670                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 671         }
 672
 673         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 674              phys_offset < aligned_vol_free_end;
 675              phys_offset += HAMMER_BIGBLOCK_SIZE) {
 676                 if (phys_offset < vol_free_off) {
 677                         ;
 678                 } else if (phys_offset < vol->vol_free_end) {
 679                         ++count;
 680                 }
 681         }
 682
 683         return(count);
 684 }
 685
 686 /*
 687  * Format the undomap for the root volume.
 688  */
 689 void
 690 format_undomap(struct volume_info *root_vol, int64_t *undo_buffer_size)
 691 {
 692         hammer_off_t undo_limit;
 693         hammer_blockmap_t blockmap;
 694         hammer_volume_ondisk_t ondisk;
 695         struct buffer_info *buffer = NULL;
 696         hammer_off_t scan;
 697         int n;
 698         int limit_index;
 699         uint32_t seqno;
 700
 701         /* Only root volume needs formatting */
 702         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 703         ondisk = root_vol->ondisk;
 704
 705         /*
 706          * Size the undo buffer in multiples of HAMMER_BIGBLOCK_SIZE,
 707          * up to HAMMER_MAX_UNDO_BIGBLOCKS big-blocks.
 708          * Size to approximately 0.1% of the disk.
 709          *
 710          * The minimum UNDO fifo size is 512MB, or approximately 1% of
 711          * the recommended 50G disk.
 712          *
 713          * Changing this minimum is rather dangerous as complex filesystem
 714          * operations can cause the UNDO FIFO to fill up otherwise.
 715          */
 716         undo_limit = *undo_buffer_size;
 717         if (undo_limit == 0) {
 718                 undo_limit = HAMMER_VOL_BUF_SIZE(ondisk) / 1000;
 719                 if (undo_limit < HAMMER_BIGBLOCK_SIZE * HAMMER_MIN_UNDO_BIGBLOCKS)
 720                         undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_MIN_UNDO_BIGBLOCKS;
 721         }
 722         undo_limit = HAMMER_BIGBLOCK_DOALIGN(undo_limit);
 723         if (undo_limit < HAMMER_BIGBLOCK_SIZE)
 724                 undo_limit = HAMMER_BIGBLOCK_SIZE;
 725         if (undo_limit > HAMMER_BIGBLOCK_SIZE * HAMMER_MAX_UNDO_BIGBLOCKS)
 726                 undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_MAX_UNDO_BIGBLOCKS;
 727         *undo_buffer_size = undo_limit;
 728
 729         blockmap = &ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 730         bzero(blockmap, sizeof(*blockmap));
 731         blockmap->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 732         blockmap->first_offset = HAMMER_ENCODE_UNDO(0);
 733         blockmap->next_offset = blockmap->first_offset;
 734         blockmap->alloc_offset = HAMMER_ENCODE_UNDO(undo_limit);
 735         hammer_crc_set_blockmap(blockmap);
 736
 737         limit_index = undo_limit / HAMMER_BIGBLOCK_SIZE;
 738         assert(limit_index <= HAMMER_MAX_UNDO_BIGBLOCKS);
 739
 740         for (n = 0; n < limit_index; ++n) {
 741                 ondisk->vol0_undo_array[n] = alloc_undo_bigblock(root_vol);
 742         }
 743         while (n < HAMMER_MAX_UNDO_BIGBLOCKS) {
 744                 ondisk->vol0_undo_array[n++] = HAMMER_BLOCKMAP_UNAVAIL;
 745         }
 746
 747         /*
 748          * Pre-initialize the UNDO blocks (HAMMER version 4+)
 749          */
 750         printf("initializing the undo map (%jd MB)\n",
 751                 (intmax_t)HAMMER_OFF_LONG_ENCODE(blockmap->alloc_offset) /
 752                 (1024 * 1024));
 753
 754         scan = blockmap->first_offset;
 755         seqno = 0;
 756
 757         while (scan < blockmap->alloc_offset) {
 758                 hammer_fifo_head_t head;
 759                 hammer_fifo_tail_t tail;
 760                 int isnew;
 761                 int bytes = HAMMER_UNDO_ALIGN;
 762
 763                 isnew = ((scan & HAMMER_BUFMASK64) == 0);
 764                 head = get_buffer_data(scan, &buffer, isnew);
 765                 buffer->cache.modified = 1;
 766                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 767
 768                 bzero(head, bytes);
 769                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 770                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
 771                 head->hdr_size = bytes;
 772                 head->hdr_seq = seqno++;
 773
 774                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 775                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 776                 tail->tail_size = bytes;
 777
 778                 hammer_crc_set_fifo_head(head, bytes);
 779
 780                 scan += bytes;
 781         }
 782         rel_buffer(buffer);
 783 }
 784
 785 const char *zone_labels[] = {
 786         "",             /* 0 */
 787         "raw_volume",   /* 1 */
 788         "raw_buffer",   /* 2 */
 789         "undo",         /* 3 */
 790         "freemap",      /* 4 */
 791         "",             /* 5 */
 792         "",             /* 6 */
 793         "",             /* 7 */
 794         "btree",        /* 8 */
 795         "meta",         /* 9 */
 796         "large_data",   /* 10 */
 797         "small_data",   /* 11 */
 798         "",             /* 12 */
 799         "",             /* 13 */
 800         "",             /* 14 */
 801         "unavail",      /* 15 */
 802 };
 803
 804 void
 805 print_blockmap(const struct volume_info *vol)
 806 {
 807         hammer_blockmap_t blockmap;
 808         hammer_volume_ondisk_t ondisk;
 809         int64_t size, used;
 810         int i;
 811 #define INDENT ""
 812
 813         ondisk = vol->ondisk;
 814         printf(INDENT"vol_label\t%s\n", ondisk->vol_label);
 815         printf(INDENT"vol_count\t%d\n", ondisk->vol_count);
 816         printf(INDENT"vol_bot_beg\t%s\n", sizetostr(ondisk->vol_bot_beg));
 817         printf(INDENT"vol_mem_beg\t%s\n", sizetostr(ondisk->vol_mem_beg));
 818         printf(INDENT"vol_buf_beg\t%s\n", sizetostr(ondisk->vol_buf_beg));
 819         printf(INDENT"vol_buf_end\t%s\n", sizetostr(ondisk->vol_buf_end));
 820         printf(INDENT"vol0_next_tid\t%016jx\n",
 821                (uintmax_t)ondisk->vol0_next_tid);
 822
 823         blockmap = &ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 824         size = HAMMER_OFF_LONG_ENCODE(blockmap->alloc_offset);
 825         if (blockmap->first_offset <= blockmap->next_offset)
 826                 used = blockmap->next_offset - blockmap->first_offset;
 827         else
 828                 used = blockmap->alloc_offset - blockmap->first_offset +
 829                         HAMMER_OFF_LONG_ENCODE(blockmap->next_offset);
 830         printf(INDENT"undo_size\t%s\n", sizetostr(size));
 831         printf(INDENT"undo_used\t%s\n", sizetostr(used));
 832
 833         printf(INDENT"zone #             "
 834                "phys             first            next             alloc\n");
 835         for (i = 0; i < HAMMER_MAX_ZONES; i++) {
 836                 blockmap = &ondisk->vol0_blockmap[i];
 837                 printf(INDENT"zone %-2d %-10s %016jx %016jx %016jx %016jx\n",
 838                         i, zone_labels[i],
 839                         (uintmax_t)blockmap->phys_offset,
 840                         (uintmax_t)blockmap->first_offset,
 841                         (uintmax_t)blockmap->next_offset,
 842                         (uintmax_t)blockmap->alloc_offset);
 843         }
 844 }
 845
 846 /*
 847  * Flush various tracking structures to disk
 848  */
 849 void
 850 flush_all_volumes(void)
 851 {
 852         struct volume_info *vol;
 853
 854         TAILQ_FOREACH(vol, &VolList, entry)
 855                 flush_volume(vol);
 856 }
 857
 858 void
 859 flush_volume(struct volume_info *volume)
 860 {
 861         struct buffer_info *buffer;
 862         int i;
 863
 864         for (i = 0; i < HAMMER_BUFLISTS; ++i) {
 865                 TAILQ_FOREACH(buffer, &volume->buffer_lists[i], entry)
 866                         flush_buffer(buffer);
 867         }
 868         if (writehammervol(volume) == -1)
 869                 err(1, "Write volume %d (%s)", volume->vol_no, volume->name);
 870 }
 871
 872 void
 873 flush_buffer(struct buffer_info *buffer)
 874 {
 875         struct volume_info *vol;
 876
 877         vol = buffer->volume;
 878         if (writehammerbuf(buffer) == -1)
 879                 err(1, "Write volume %d (%s)", vol->vol_no, vol->name);
 880         buffer->cache.modified = 0;
 881 }
 882
 883 /*
 884  * Core I/O operations
 885  */
 886 static int
 887 __read(struct volume_info *vol, void *data, int64_t offset, int size)
 888 {
 889         ssize_t n;
 890
 891         n = pread(vol->fd, data, size, offset);
 892         if (n != size)
 893                 return(-1);
 894         return(0);
 895 }
 896
 897 static __inline int
 898 readhammervol(struct volume_info *vol)
 899 {
 900         return(__read(vol, vol->ondisk, 0, HAMMER_BUFSIZE));
 901 }
 902
 903 static __inline int
 904 readhammerbuf(struct buffer_info *buf)
 905 {
 906         return(__read(buf->volume, buf->ondisk, buf->raw_offset, HAMMER_BUFSIZE));
 907 }
 908
 909 static int
 910 __write(struct volume_info *vol, const void *data, int64_t offset, int size)
 911 {
 912         ssize_t n;
 913
 914         if (vol->rdonly)
 915                 return(0);
 916
 917         n = pwrite(vol->fd, data, size, offset);
 918         if (n != size)
 919                 return(-1);
 920         return(0);
 921 }
 922
 923 static __inline int
 924 writehammervol(struct volume_info *vol)
 925 {
 926         return(__write(vol, vol->ondisk, 0, HAMMER_BUFSIZE));
 927 }
 928
 929 static __inline int
 930 writehammerbuf(struct buffer_info *buf)
 931 {
 932         return(__write(buf->volume, buf->ondisk, buf->raw_offset, HAMMER_BUFSIZE));
 933 }
 934
 935 int64_t init_boot_area_size(int64_t value, off_t avg_vol_size)
 936 {
 937         if (value == 0) {
 938                 value = HAMMER_BOOT_NOMBYTES;
 939                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 940                         value >>= 1;
 941         }
 942
 943         if (value < HAMMER_BOOT_MINBYTES) {
 944                 value = HAMMER_BOOT_MINBYTES;
 945         } else if (value > HAMMER_BOOT_MAXBYTES) {
 946                 value = HAMMER_BOOT_MAXBYTES;
 947         }
 948
 949         return(value);
 950 }
 951
 952 int64_t init_memory_log_size(int64_t value, off_t avg_vol_size)
 953 {
 954         if (value == 0) {
 955                 value = HAMMER_MEM_NOMBYTES;
 956                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 957                         value >>= 1;
 958         }
 959
 960         if (value < HAMMER_MEM_MINBYTES) {
 961                 value = HAMMER_MEM_MINBYTES;
 962         } else if (value > HAMMER_MEM_MAXBYTES) {
 963                 value = HAMMER_MEM_MAXBYTES;
 964         }
 965
 966         return(value);
 967 }