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(struct volume_info *volume, hammer_off_t buf_offset)
  69 {
  70         int hi;
  71         struct buffer_info *buf;
  72
  73         hi = buffer_hash(buf_offset);
  74         TAILQ_FOREACH(buf, &volume->buffer_lists[hi], entry)
  75                 if (buf->buf_offset == buf_offset)
  76                         return(buf);
  77         return(NULL);
  78 }
  79
  80 static
  81 struct volume_info *
  82 __alloc_volume(const char *volname, int oflags)
  83 {
  84         struct volume_info *vol;
  85         int i;
  86
  87         vol = malloc(sizeof(*vol));
  88         if (vol == NULL)
  89                 err(1, "alloc_volume");
  90         bzero(vol, sizeof(*vol));
  91
  92         vol->vol_no = -1;
  93         vol->rdonly = (oflags == O_RDONLY);
  94         vol->name = strdup(volname);
  95         vol->fd = open(vol->name, oflags);
  96         if (vol->fd < 0)
  97                 err(1, "alloc_volume: Failed to open %s", vol->name);
  98         check_volume(vol);
  99
 100         vol->ondisk = malloc(HAMMER_BUFSIZE);
 101         if (vol->ondisk == NULL)
 102                 err(1, "alloc_volume");
 103         bzero(vol->ondisk, HAMMER_BUFSIZE);
 104
 105         for (i = 0; i < HAMMER_BUFLISTS; ++i)
 106                 TAILQ_INIT(&vol->buffer_lists[i]);
 107
 108         return(vol);
 109 }
 110
 111 static void
 112 __add_volume(struct volume_info *vol)
 113 {
 114         struct volume_info *scan;
 115         struct stat st1, st2;
 116
 117         if (fstat(vol->fd, &st1) != 0)
 118                 errx(1, "add_volume: %s: Failed to stat", vol->name);
 119
 120         TAILQ_FOREACH(scan, &VolList, entry) {
 121                 if (scan->vol_no == vol->vol_no) {
 122                         errx(1, "add_volume: %s: Duplicate volume number %d "
 123                                 "against %s",
 124                                 vol->name, vol->vol_no, scan->name);
 125                 }
 126                 if (fstat(scan->fd, &st2) != 0) {
 127                         errx(1, "add_volume: %s: Failed to stat %s",
 128                                 vol->name, scan->name);
 129                 }
 130                 if ((st1.st_ino == st2.st_ino) && (st1.st_dev == st2.st_dev)) {
 131                         errx(1, "add_volume: %s: Specified more than once",
 132                                 vol->name);
 133                 }
 134         }
 135
 136         TAILQ_INSERT_TAIL(&VolList, vol, entry);
 137 }
 138
 139 static void
 140 __verify_volume(struct volume_info *vol)
 141 {
 142         hammer_volume_ondisk_t ondisk = vol->ondisk;
 143
 144         if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
 145                 errx(1, "verify_volume: Invalid volume signature %016jx",
 146                         ondisk->vol_signature);
 147         }
 148         if (ondisk->vol_rootvol != HAMMER_ROOT_VOLNO) {
 149                 errx(1, "verify_volume: Invalid root volume# %d",
 150                         ondisk->vol_rootvol);
 151         }
 152         if (bcmp(&Hammer_FSType, &ondisk->vol_fstype, sizeof(Hammer_FSType))) {
 153                 errx(1, "verify_volume: %s: Header does not indicate "
 154                         "that this is a HAMMER volume", vol->name);
 155         }
 156         if (bcmp(&Hammer_FSId, &ondisk->vol_fsid, sizeof(Hammer_FSId))) {
 157                 errx(1, "verify_volume: %s: FSId does not match other volumes!",
 158                         vol->name);
 159         }
 160 }
 161
 162 /*
 163  * Initialize a volume structure and ondisk vol_no field.
 164  */
 165 struct volume_info *
 166 init_volume(const char *filename, int oflags, int32_t vol_no)
 167 {
 168         struct volume_info *vol;
 169
 170         vol = __alloc_volume(filename, oflags);
 171         vol->vol_no = vol->ondisk->vol_no = vol_no;
 172
 173         __add_volume(vol);
 174
 175         return(vol);
 176 }
 177
 178 /*
 179  * Initialize a volume structure and read ondisk volume header.
 180  */
 181 struct volume_info*
 182 load_volume(const char *filename, int oflags, int verify)
 183 {
 184         struct volume_info *vol;
 185         int n;
 186
 187         vol = __alloc_volume(filename, oflags);
 188
 189         n = readhammervol(vol);
 190         if (n == -1) {
 191                 err(1, "load_volume: %s: Read failed at offset 0", vol->name);
 192         }
 193         vol->vol_no = vol->ondisk->vol_no;
 194
 195         if (valid_hammer_volumes++ == 0)
 196                 Hammer_FSId = vol->ondisk->vol_fsid;
 197         if (verify)
 198                 __verify_volume(vol);
 199
 200         __add_volume(vol);
 201
 202         return(vol);
 203 }
 204
 205 /*
 206  * Check basic volume characteristics.
 207  */
 208 static void
 209 check_volume(struct volume_info *vol)
 210 {
 211         struct partinfo pinfo;
 212         struct stat st;
 213
 214         /*
 215          * Get basic information about the volume
 216          */
 217         if (ioctl(vol->fd, DIOCGPART, &pinfo) < 0) {
 218                 /*
 219                  * Allow the formatting of regular files as HAMMER volumes
 220                  */
 221                 if (fstat(vol->fd, &st) < 0)
 222                         err(1, "Unable to stat %s", vol->name);
 223                 vol->size = st.st_size;
 224                 vol->type = "REGFILE";
 225         } else {
 226                 /*
 227                  * When formatting a block device as a HAMMER volume the
 228                  * sector size must be compatible.  HAMMER uses 16384 byte
 229                  * filesystem buffers.
 230                  */
 231                 if (pinfo.reserved_blocks) {
 232                         errx(1, "HAMMER cannot be placed in a partition "
 233                                 "which overlaps the disklabel or MBR");
 234                 }
 235                 if (pinfo.media_blksize > HAMMER_BUFSIZE ||
 236                     HAMMER_BUFSIZE % pinfo.media_blksize) {
 237                         errx(1, "A media sector size of %d is not supported",
 238                              pinfo.media_blksize);
 239                 }
 240
 241                 vol->size = pinfo.media_size;
 242                 vol->device_offset = pinfo.media_offset;
 243                 vol->type = "DEVICE";
 244         }
 245 }
 246
 247 void
 248 assert_volume_offset(struct volume_info *vol)
 249 {
 250         assert(hammer_is_zone_raw_buffer(vol->vol_free_off));
 251         assert(hammer_is_zone_raw_buffer(vol->vol_free_end));
 252 }
 253
 254 struct volume_info *
 255 get_volume(int32_t vol_no)
 256 {
 257         struct volume_info *vol;
 258
 259         TAILQ_FOREACH(vol, &VolList, entry) {
 260                 if (vol->vol_no == vol_no)
 261                         break;
 262         }
 263
 264         return(vol);
 265 }
 266
 267 struct volume_info *
 268 get_root_volume(void)
 269 {
 270         return(get_volume(HAMMER_ROOT_VOLNO));
 271 }
 272
 273 /*
 274  * Acquire the specified buffer.  isnew is -1 only when called
 275  * via get_buffer_readahead() to prevent another readahead.
 276  */
 277 static struct buffer_info *
 278 get_buffer(hammer_off_t buf_offset, int isnew)
 279 {
 280         struct buffer_info *buf;
 281         struct volume_info *volume;
 282         int vol_no;
 283         int zone;
 284         int hi;
 285         int dora = 0;
 286         int error = 0;
 287
 288         zone = HAMMER_ZONE_DECODE(buf_offset);
 289         if (zone > HAMMER_ZONE_RAW_BUFFER_INDEX)
 290                 buf_offset = blockmap_lookup(buf_offset, NULL, NULL, &error);
 291         if (error || buf_offset == HAMMER_OFF_BAD)
 292                 return(NULL);
 293         assert(hammer_is_zone_raw_buffer(buf_offset));
 294
 295         vol_no = HAMMER_VOL_DECODE(buf_offset);
 296         volume = get_volume(vol_no);
 297         assert(volume != NULL);
 298
 299         buf_offset &= ~HAMMER_BUFMASK64;
 300         buf = find_buffer(volume, buf_offset);
 301
 302         if (buf == NULL) {
 303                 buf = malloc(sizeof(*buf));
 304                 bzero(buf, sizeof(*buf));
 305                 buf->buf_offset = buf_offset;
 306                 buf->raw_offset = hammer_xlate_to_phys(volume->ondisk,
 307                                                         buf_offset);
 308                 buf->volume = volume;
 309                 buf->ondisk = malloc(HAMMER_BUFSIZE);
 310                 if (isnew <= 0) {
 311                         if (readhammerbuf(buf) == -1) {
 312                                 err(1, "get_buffer: %s:%016jx "
 313                                     "Read failed at offset %016jx",
 314                                     volume->name,
 315                                     (intmax_t)buf->buf_offset,
 316                                     (intmax_t)buf->raw_offset);
 317                         }
 318                 }
 319
 320                 hi = buffer_hash(buf_offset);
 321                 TAILQ_INSERT_TAIL(&volume->buffer_lists[hi], buf, entry);
 322                 hammer_cache_add(&buf->cache);
 323                 dora = (isnew == 0);
 324         } else {
 325                 assert(buf->ondisk != NULL);
 326                 assert(isnew != -1);
 327                 hammer_cache_used(&buf->cache);
 328         }
 329
 330         ++buf->cache.refs;
 331         hammer_cache_flush();
 332
 333         if (isnew > 0) {
 334                 assert(buf->cache.modified == 0);
 335                 bzero(buf->ondisk, HAMMER_BUFSIZE);
 336                 buf->cache.modified = 1;
 337         }
 338         if (dora)
 339                 get_buffer_readahead(buf);
 340         return(buf);
 341 }
 342
 343 static void
 344 get_buffer_readahead(struct buffer_info *base)
 345 {
 346         struct buffer_info *buf;
 347         struct volume_info *vol;
 348         hammer_off_t buf_offset;
 349         int64_t raw_offset;
 350         int ri = UseReadBehind;
 351         int re = UseReadAhead;
 352
 353         raw_offset = base->raw_offset + ri * HAMMER_BUFSIZE;
 354         vol = base->volume;
 355
 356         while (ri < re) {
 357                 if (raw_offset >= vol->ondisk->vol_buf_end)
 358                         break;
 359                 if (raw_offset < vol->ondisk->vol_buf_beg || ri == 0) {
 360                         ++ri;
 361                         raw_offset += HAMMER_BUFSIZE;
 362                         continue;
 363                 }
 364                 buf_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no,
 365                         raw_offset - vol->ondisk->vol_buf_beg);
 366                 buf = find_buffer(vol, buf_offset);
 367                 if (buf == NULL) {
 368                         buf = get_buffer(buf_offset, -1);
 369                         rel_buffer(buf);
 370                 }
 371                 ++ri;
 372                 raw_offset += HAMMER_BUFSIZE;
 373         }
 374 }
 375
 376 void
 377 rel_buffer(struct buffer_info *buffer)
 378 {
 379         struct volume_info *volume;
 380         int hi;
 381
 382         if (buffer == NULL)
 383                 return;
 384         assert(buffer->cache.refs > 0);
 385         if (--buffer->cache.refs == 0) {
 386                 if (buffer->cache.delete) {
 387                         hi = buffer_hash(buffer->buf_offset);
 388                         volume = buffer->volume;
 389                         if (buffer->cache.modified)
 390                                 flush_buffer(buffer);
 391                         TAILQ_REMOVE(&volume->buffer_lists[hi], buffer, entry);
 392                         hammer_cache_del(&buffer->cache);
 393                         free(buffer->ondisk);
 394                         free(buffer);
 395                 }
 396         }
 397 }
 398
 399 /*
 400  * Retrieve a pointer to a buffer data given a buffer offset.  The underlying
 401  * bufferp is freed if isnew or the offset is out of range of the cached data.
 402  * If bufferp is freed a referenced buffer is loaded into it.
 403  */
 404 void *
 405 get_buffer_data(hammer_off_t buf_offset, struct buffer_info **bufferp,
 406                 int isnew)
 407 {
 408         if (*bufferp != NULL) {
 409                 if (isnew > 0 ||
 410                     (((*bufferp)->buf_offset ^ buf_offset) & ~HAMMER_BUFMASK64)) {
 411                         rel_buffer(*bufferp);
 412                         *bufferp = NULL;
 413                 }
 414         }
 415
 416         if (*bufferp == NULL) {
 417                 *bufferp = get_buffer(buf_offset, isnew);
 418                 if (*bufferp == NULL)
 419                         return(NULL);
 420         }
 421
 422         return(((char *)(*bufferp)->ondisk) +
 423                 ((int32_t)buf_offset & HAMMER_BUFMASK));
 424 }
 425
 426 /*
 427  * Allocate HAMMER elements - B-Tree nodes
 428  */
 429 hammer_node_ondisk_t
 430 alloc_btree_node(hammer_off_t *offp, struct buffer_info **data_bufferp)
 431 {
 432         hammer_node_ondisk_t node;
 433
 434         node = alloc_blockmap(HAMMER_ZONE_BTREE_INDEX, sizeof(*node),
 435                               offp, data_bufferp);
 436         bzero(node, sizeof(*node));
 437         return(node);
 438 }
 439
 440 /*
 441  * Allocate HAMMER elements - meta data (inode, direntry, PFS, etc)
 442  */
 443 void *
 444 alloc_meta_element(hammer_off_t *offp, int32_t data_len,
 445                    struct buffer_info **data_bufferp)
 446 {
 447         void *data;
 448
 449         data = alloc_blockmap(HAMMER_ZONE_META_INDEX, data_len,
 450                               offp, data_bufferp);
 451         bzero(data, data_len);
 452         return(data);
 453 }
 454
 455 /*
 456  * Format a new blockmap.  This is mostly a degenerate case because
 457  * all allocations are now actually done from the freemap.
 458  */
 459 void
 460 format_blockmap(struct volume_info *root_vol, int zone, hammer_off_t offset)
 461 {
 462         hammer_blockmap_t blockmap;
 463         hammer_off_t zone_base;
 464
 465         /* Only root volume needs formatting */
 466         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 467
 468         assert(hammer_is_zone2_mapped_index(zone));
 469
 470         blockmap = &root_vol->ondisk->vol0_blockmap[zone];
 471         zone_base = HAMMER_ZONE_ENCODE(zone, offset);
 472
 473         bzero(blockmap, sizeof(*blockmap));
 474         blockmap->phys_offset = 0;
 475         blockmap->first_offset = zone_base;
 476         blockmap->next_offset = zone_base;
 477         blockmap->alloc_offset = HAMMER_ENCODE(zone, 255, -1);
 478         hammer_crc_set_blockmap(blockmap);
 479 }
 480
 481 /*
 482  * Format a new freemap.  Set all layer1 entries to UNAVAIL.  The initialize
 483  * code will load each volume's freemap.
 484  */
 485 void
 486 format_freemap(struct volume_info *root_vol)
 487 {
 488         struct buffer_info *buffer = NULL;
 489         hammer_off_t layer1_offset;
 490         hammer_blockmap_t blockmap;
 491         hammer_blockmap_layer1_t layer1;
 492         int i, isnew;
 493
 494         /* Only root volume needs formatting */
 495         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 496
 497         layer1_offset = bootstrap_bigblock(root_vol);
 498         for (i = 0; i < HAMMER_BIGBLOCK_SIZE; i += sizeof(*layer1)) {
 499                 isnew = ((i % HAMMER_BUFSIZE) == 0);
 500                 layer1 = get_buffer_data(layer1_offset + i, &buffer, isnew);
 501                 bzero(layer1, sizeof(*layer1));
 502                 layer1->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 503                 layer1->blocks_free = 0;
 504                 hammer_crc_set_layer1(layer1);
 505         }
 506         assert(i == HAMMER_BIGBLOCK_SIZE);
 507         rel_buffer(buffer);
 508
 509         blockmap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 510         bzero(blockmap, sizeof(*blockmap));
 511         blockmap->phys_offset = layer1_offset;
 512         blockmap->first_offset = 0;
 513         blockmap->next_offset = HAMMER_ENCODE_RAW_BUFFER(0, 0);
 514         blockmap->alloc_offset = HAMMER_ENCODE_RAW_BUFFER(255, -1);
 515         hammer_crc_set_blockmap(blockmap);
 516 }
 517
 518 /*
 519  * Load the volume's remaining free space into the freemap.
 520  *
 521  * Returns the number of big-blocks available.
 522  */
 523 int64_t
 524 initialize_freemap(struct volume_info *vol)
 525 {
 526         struct volume_info *root_vol;
 527         struct buffer_info *buffer1 = NULL;
 528         struct buffer_info *buffer2 = NULL;
 529         hammer_blockmap_layer1_t layer1;
 530         hammer_blockmap_layer2_t layer2;
 531         hammer_off_t layer1_offset;
 532         hammer_off_t layer2_offset;
 533         hammer_off_t phys_offset;
 534         hammer_off_t block_offset;
 535         hammer_off_t aligned_vol_free_end;
 536         hammer_blockmap_t freemap;
 537         int64_t count = 0;
 538         int64_t layer1_count = 0;
 539
 540         root_vol = get_root_volume();
 541
 542         assert_volume_offset(vol);
 543         aligned_vol_free_end = HAMMER_BLOCKMAP_LAYER2_DOALIGN(vol->vol_free_end);
 544
 545         printf("initialize freemap volume %d\n", vol->vol_no);
 546
 547         /*
 548          * Initialize the freemap.  First preallocate the big-blocks required
 549          * to implement layer2.   This preallocation is a bootstrap allocation
 550          * using blocks from the target volume.
 551          */
 552         freemap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 553
 554         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 555              phys_offset < aligned_vol_free_end;
 556              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 557                 layer1_offset = freemap->phys_offset +
 558                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 559                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 560                 if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 561                         layer1->phys_offset = bootstrap_bigblock(vol);
 562                         layer1->blocks_free = 0;
 563                         buffer1->cache.modified = 1;
 564                         hammer_crc_set_layer1(layer1);
 565                 }
 566         }
 567
 568         /*
 569          * Now fill everything in.
 570          */
 571         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 572              phys_offset < aligned_vol_free_end;
 573              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 574                 layer1_count = 0;
 575                 layer1_offset = freemap->phys_offset +
 576                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 577                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 578                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 579
 580                 for (block_offset = 0;
 581                      block_offset < HAMMER_BLOCKMAP_LAYER2;
 582                      block_offset += HAMMER_BIGBLOCK_SIZE) {
 583                         layer2_offset = layer1->phys_offset +
 584                                         HAMMER_BLOCKMAP_LAYER2_OFFSET(block_offset);
 585                         layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 586                         bzero(layer2, sizeof(*layer2));
 587
 588                         if (phys_offset + block_offset < vol->vol_free_off) {
 589                                 /*
 590                                  * Big-blocks already allocated as part
 591                                  * of the freemap bootstrap.
 592                                  */
 593                                 layer2->zone = HAMMER_ZONE_FREEMAP_INDEX;
 594                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 595                                 layer2->bytes_free = 0;
 596                         } else if (phys_offset + block_offset < vol->vol_free_end) {
 597                                 layer2->zone = 0;
 598                                 layer2->append_off = 0;
 599                                 layer2->bytes_free = HAMMER_BIGBLOCK_SIZE;
 600                                 ++count;
 601                                 ++layer1_count;
 602                         } else {
 603                                 layer2->zone = HAMMER_ZONE_UNAVAIL_INDEX;
 604                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 605                                 layer2->bytes_free = 0;
 606                         }
 607                         hammer_crc_set_layer2(layer2);
 608                         buffer2->cache.modified = 1;
 609                 }
 610
 611                 layer1->blocks_free += layer1_count;
 612                 hammer_crc_set_layer1(layer1);
 613                 buffer1->cache.modified = 1;
 614         }
 615
 616         rel_buffer(buffer1);
 617         rel_buffer(buffer2);
 618         return(count);
 619 }
 620
 621 /*
 622  * Returns the number of big-blocks available for filesystem data and undos
 623  * without formatting.
 624  */
 625 int64_t
 626 count_freemap(struct volume_info *vol)
 627 {
 628         hammer_off_t phys_offset;
 629         hammer_off_t vol_free_off;
 630         hammer_off_t aligned_vol_free_end;
 631         int64_t count = 0;
 632
 633         vol_free_off = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 634
 635         assert_volume_offset(vol);
 636         aligned_vol_free_end = HAMMER_BLOCKMAP_LAYER2_DOALIGN(vol->vol_free_end);
 637
 638         if (vol->vol_no == HAMMER_ROOT_VOLNO)
 639                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 640
 641         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 642              phys_offset < aligned_vol_free_end;
 643              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 644                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 645         }
 646
 647         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 648              phys_offset < aligned_vol_free_end;
 649              phys_offset += HAMMER_BIGBLOCK_SIZE) {
 650                 if (phys_offset < vol_free_off) {
 651                         ;
 652                 } else if (phys_offset < vol->vol_free_end) {
 653                         ++count;
 654                 }
 655         }
 656
 657         return(count);
 658 }
 659
 660 /*
 661  * Format the undomap for the root volume.
 662  */
 663 void
 664 format_undomap(struct volume_info *root_vol, int64_t *undo_buffer_size)
 665 {
 666         hammer_off_t undo_limit;
 667         hammer_blockmap_t blockmap;
 668         hammer_volume_ondisk_t ondisk;
 669         struct buffer_info *buffer = NULL;
 670         hammer_off_t scan;
 671         int n;
 672         int limit_index;
 673         uint32_t seqno;
 674
 675         /* Only root volume needs formatting */
 676         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 677         ondisk = root_vol->ondisk;
 678
 679         /*
 680          * Size the undo buffer in multiples of HAMMER_BIGBLOCK_SIZE,
 681          * up to HAMMER_MAX_UNDO_BIGBLOCKS big-blocks.
 682          * Size to approximately 0.1% of the disk.
 683          *
 684          * The minimum UNDO fifo size is 512MB, or approximately 1% of
 685          * the recommended 50G disk.
 686          *
 687          * Changing this minimum is rather dangerous as complex filesystem
 688          * operations can cause the UNDO FIFO to fill up otherwise.
 689          */
 690         undo_limit = *undo_buffer_size;
 691         if (undo_limit == 0) {
 692                 undo_limit = HAMMER_VOL_BUF_SIZE(ondisk) / 1000;
 693                 if (undo_limit < HAMMER_BIGBLOCK_SIZE * HAMMER_MIN_UNDO_BIGBLOCKS)
 694                         undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_MIN_UNDO_BIGBLOCKS;
 695         }
 696         undo_limit = (undo_limit + HAMMER_BIGBLOCK_MASK64) &
 697                      ~HAMMER_BIGBLOCK_MASK64;
 698         if (undo_limit < HAMMER_BIGBLOCK_SIZE)
 699                 undo_limit = HAMMER_BIGBLOCK_SIZE;
 700         if (undo_limit > HAMMER_BIGBLOCK_SIZE * HAMMER_MAX_UNDO_BIGBLOCKS)
 701                 undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_MAX_UNDO_BIGBLOCKS;
 702         *undo_buffer_size = undo_limit;
 703
 704         blockmap = &ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 705         bzero(blockmap, sizeof(*blockmap));
 706         blockmap->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 707         blockmap->first_offset = HAMMER_ENCODE_UNDO(0);
 708         blockmap->next_offset = blockmap->first_offset;
 709         blockmap->alloc_offset = HAMMER_ENCODE_UNDO(undo_limit);
 710         hammer_crc_set_blockmap(blockmap);
 711
 712         limit_index = undo_limit / HAMMER_BIGBLOCK_SIZE;
 713         assert(limit_index <= HAMMER_MAX_UNDO_BIGBLOCKS);
 714
 715         for (n = 0; n < limit_index; ++n) {
 716                 ondisk->vol0_undo_array[n] = alloc_undo_bigblock(root_vol);
 717         }
 718         while (n < HAMMER_MAX_UNDO_BIGBLOCKS) {
 719                 ondisk->vol0_undo_array[n++] = HAMMER_BLOCKMAP_UNAVAIL;
 720         }
 721
 722         /*
 723          * Pre-initialize the UNDO blocks (HAMMER version 4+)
 724          */
 725         printf("initializing the undo map (%jd MB)\n",
 726                 (intmax_t)HAMMER_OFF_LONG_ENCODE(blockmap->alloc_offset) /
 727                 (1024 * 1024));
 728
 729         scan = blockmap->first_offset;
 730         seqno = 0;
 731
 732         while (scan < blockmap->alloc_offset) {
 733                 hammer_fifo_head_t head;
 734                 hammer_fifo_tail_t tail;
 735                 int isnew;
 736                 int bytes = HAMMER_UNDO_ALIGN;
 737
 738                 isnew = ((scan & HAMMER_BUFMASK64) == 0);
 739                 head = get_buffer_data(scan, &buffer, isnew);
 740                 buffer->cache.modified = 1;
 741                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 742
 743                 bzero(head, bytes);
 744                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 745                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
 746                 head->hdr_size = bytes;
 747                 head->hdr_seq = seqno++;
 748
 749                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 750                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 751                 tail->tail_size = bytes;
 752
 753                 hammer_crc_set_fifo_head(head, bytes);
 754
 755                 scan += bytes;
 756         }
 757         rel_buffer(buffer);
 758 }
 759
 760 const char *zone_labels[] = {
 761         "",             /* 0 */
 762         "raw_volume",   /* 1 */
 763         "raw_buffer",   /* 2 */
 764         "undo",         /* 3 */
 765         "freemap",      /* 4 */
 766         "",             /* 5 */
 767         "",             /* 6 */
 768         "",             /* 7 */
 769         "btree",        /* 8 */
 770         "meta",         /* 9 */
 771         "large_data",   /* 10 */
 772         "small_data",   /* 11 */
 773         "",             /* 12 */
 774         "",             /* 13 */
 775         "",             /* 14 */
 776         "unavail",      /* 15 */
 777 };
 778
 779 void
 780 print_blockmap(const struct volume_info *vol)
 781 {
 782         hammer_blockmap_t blockmap;
 783         hammer_volume_ondisk_t ondisk;
 784         int64_t size, used;
 785         int i;
 786 #define INDENT ""
 787
 788         ondisk = vol->ondisk;
 789         printf(INDENT"vol_label\t%s\n", ondisk->vol_label);
 790         printf(INDENT"vol_count\t%d\n", ondisk->vol_count);
 791         printf(INDENT"vol_bot_beg\t%s\n", sizetostr(ondisk->vol_bot_beg));
 792         printf(INDENT"vol_mem_beg\t%s\n", sizetostr(ondisk->vol_mem_beg));
 793         printf(INDENT"vol_buf_beg\t%s\n", sizetostr(ondisk->vol_buf_beg));
 794         printf(INDENT"vol_buf_end\t%s\n", sizetostr(ondisk->vol_buf_end));
 795         printf(INDENT"vol0_next_tid\t%016jx\n",
 796                (uintmax_t)ondisk->vol0_next_tid);
 797
 798         blockmap = &ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 799         size = HAMMER_OFF_LONG_ENCODE(blockmap->alloc_offset);
 800         if (blockmap->first_offset <= blockmap->next_offset)
 801                 used = blockmap->next_offset - blockmap->first_offset;
 802         else
 803                 used = blockmap->alloc_offset - blockmap->first_offset +
 804                         HAMMER_OFF_LONG_ENCODE(blockmap->next_offset);
 805         printf(INDENT"undo_size\t%s\n", sizetostr(size));
 806         printf(INDENT"undo_used\t%s\n", sizetostr(used));
 807
 808         printf(INDENT"zone #             "
 809                "phys             first            next             alloc\n");
 810         for (i = 0; i < HAMMER_MAX_ZONES; i++) {
 811                 blockmap = &ondisk->vol0_blockmap[i];
 812                 printf(INDENT"zone %-2d %-10s %016jx %016jx %016jx %016jx\n",
 813                         i, zone_labels[i],
 814                         (uintmax_t)blockmap->phys_offset,
 815                         (uintmax_t)blockmap->first_offset,
 816                         (uintmax_t)blockmap->next_offset,
 817                         (uintmax_t)blockmap->alloc_offset);
 818         }
 819 }
 820
 821 /*
 822  * Flush various tracking structures to disk
 823  */
 824 void
 825 flush_all_volumes(void)
 826 {
 827         struct volume_info *vol;
 828
 829         TAILQ_FOREACH(vol, &VolList, entry)
 830                 flush_volume(vol);
 831 }
 832
 833 void
 834 flush_volume(struct volume_info *volume)
 835 {
 836         struct buffer_info *buffer;
 837         int i;
 838
 839         for (i = 0; i < HAMMER_BUFLISTS; ++i) {
 840                 TAILQ_FOREACH(buffer, &volume->buffer_lists[i], entry)
 841                         flush_buffer(buffer);
 842         }
 843         if (writehammervol(volume) == -1)
 844                 err(1, "Write volume %d (%s)", volume->vol_no, volume->name);
 845 }
 846
 847 void
 848 flush_buffer(struct buffer_info *buffer)
 849 {
 850         struct volume_info *vol;
 851
 852         vol = buffer->volume;
 853         if (writehammerbuf(buffer) == -1)
 854                 err(1, "Write volume %d (%s)", vol->vol_no, vol->name);
 855         buffer->cache.modified = 0;
 856 }
 857
 858 /*
 859  * Core I/O operations
 860  */
 861 static int
 862 __read(struct volume_info *vol, void *data, int64_t offset, int size)
 863 {
 864         ssize_t n;
 865
 866         n = pread(vol->fd, data, size, offset);
 867         if (n != size)
 868                 return(-1);
 869         return(0);
 870 }
 871
 872 static __inline int
 873 readhammervol(struct volume_info *vol)
 874 {
 875         return(__read(vol, vol->ondisk, 0, HAMMER_BUFSIZE));
 876 }
 877
 878 static __inline int
 879 readhammerbuf(struct buffer_info *buf)
 880 {
 881         return(__read(buf->volume, buf->ondisk, buf->raw_offset, HAMMER_BUFSIZE));
 882 }
 883
 884 static int
 885 __write(struct volume_info *vol, const void *data, int64_t offset, int size)
 886 {
 887         ssize_t n;
 888
 889         if (vol->rdonly)
 890                 return(0);
 891
 892         n = pwrite(vol->fd, data, size, offset);
 893         if (n != size)
 894                 return(-1);
 895         return(0);
 896 }
 897
 898 static __inline int
 899 writehammervol(struct volume_info *vol)
 900 {
 901         return(__write(vol, vol->ondisk, 0, HAMMER_BUFSIZE));
 902 }
 903
 904 static __inline int
 905 writehammerbuf(struct buffer_info *buf)
 906 {
 907         return(__write(buf->volume, buf->ondisk, buf->raw_offset, HAMMER_BUFSIZE));
 908 }
 909
 910 int64_t init_boot_area_size(int64_t value, off_t avg_vol_size)
 911 {
 912         if (value == 0) {
 913                 value = HAMMER_BOOT_NOMBYTES;
 914                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 915                         value >>= 1;
 916         }
 917
 918         if (value < HAMMER_BOOT_MINBYTES) {
 919                 value = HAMMER_BOOT_MINBYTES;
 920         } else if (value > HAMMER_BOOT_MAXBYTES) {
 921                 value = HAMMER_BOOT_MAXBYTES;
 922         }
 923
 924         return(value);
 925 }
 926
 927 int64_t init_memory_log_size(int64_t value, off_t avg_vol_size)
 928 {
 929         if (value == 0) {
 930                 value = HAMMER_MEM_NOMBYTES;
 931                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 932                         value >>= 1;
 933         }
 934
 935         if (value < HAMMER_MEM_MINBYTES) {
 936                 value = HAMMER_MEM_MINBYTES;
 937         } else if (value > HAMMER_MEM_MAXBYTES) {
 938                 value = HAMMER_MEM_MAXBYTES;
 939         }
 940
 941         return(value);
 942 }