sbin/hammer/ondisk.c

   1 /*
   2  * Copyright (c) 2007 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  */
  34
  35 #include <sys/diskslice.h>
  36 #include <sys/diskmbr.h>
  37
  38 #include "hammer_util.h"
  39
  40 static void check_volume(struct volume_info *volume);
  41 static void get_buffer_readahead(struct buffer_info *base);
  42 static __inline int readhammervol(struct volume_info *volume);
  43 static __inline int readhammerbuf(struct buffer_info *buffer);
  44 static __inline int writehammervol(struct volume_info *volume);
  45 static __inline int writehammerbuf(struct buffer_info *buffer);
  46
  47 uuid_t Hammer_FSType;
  48 uuid_t Hammer_FSId;
  49 int UseReadBehind = -4;
  50 int UseReadAhead = 4;
  51 int DebugOpt;
  52 uint32_t HammerVersion = -1;
  53
  54 TAILQ_HEAD(volume_list, volume_info);
  55 static struct volume_list VolList = TAILQ_HEAD_INITIALIZER(VolList);
  56 static int valid_hammer_volumes;
  57
  58 static __inline
  59 int
  60 buffer_hash(hammer_off_t zone2_offset)
  61 {
  62         int hi;
  63
  64         hi = (int)(zone2_offset / HAMMER_BUFSIZE) & HAMMER_BUFLISTMASK;
  65         return(hi);
  66 }
  67
  68 static struct buffer_info*
  69 find_buffer(hammer_off_t zone2_offset)
  70 {
  71         struct volume_info *volume;
  72         struct buffer_info *buffer;
  73         int hi;
  74
  75         volume = get_volume(HAMMER_VOL_DECODE(zone2_offset));
  76         assert(volume);
  77
  78         hi = buffer_hash(zone2_offset);
  79         TAILQ_FOREACH(buffer, &volume->buffer_lists[hi], entry)
  80                 if (buffer->zone2_offset == zone2_offset)
  81                         return(buffer);
  82         return(NULL);
  83 }
  84
  85 static
  86 struct volume_info *
  87 __alloc_volume(const char *volname, int oflags)
  88 {
  89         struct volume_info *volume;
  90         int i;
  91
  92         volume = calloc(1, sizeof(*volume));
  93         volume->vol_no = -1;
  94         volume->rdonly = (oflags == O_RDONLY);
  95         volume->name = strdup(volname);
  96         volume->fd = open(volume->name, oflags);
  97         if (volume->fd < 0)
  98                 err(1, "alloc_volume: Failed to open %s", volume->name);
  99         check_volume(volume);
 100
 101         volume->ondisk = calloc(1, HAMMER_BUFSIZE);
 102
 103         for (i = 0; i < HAMMER_BUFLISTS; ++i)
 104                 TAILQ_INIT(&volume->buffer_lists[i]);
 105
 106         return(volume);
 107 }
 108
 109 static void
 110 __add_volume(struct volume_info *volume)
 111 {
 112         struct volume_info *scan;
 113         struct stat st1, st2;
 114
 115         if (fstat(volume->fd, &st1) != 0)
 116                 errx(1, "add_volume: %s: Failed to stat", volume->name);
 117
 118         TAILQ_FOREACH(scan, &VolList, entry) {
 119                 if (scan->vol_no == volume->vol_no) {
 120                         errx(1, "add_volume: %s: Duplicate volume number %d "
 121                                 "against %s",
 122                                 volume->name, volume->vol_no, scan->name);
 123                 }
 124                 if (fstat(scan->fd, &st2) != 0) {
 125                         errx(1, "add_volume: %s: Failed to stat %s",
 126                                 volume->name, scan->name);
 127                 }
 128                 if ((st1.st_ino == st2.st_ino) && (st1.st_dev == st2.st_dev)) {
 129                         errx(1, "add_volume: %s: Specified more than once",
 130                                 volume->name);
 131                 }
 132         }
 133
 134         TAILQ_INSERT_TAIL(&VolList, volume, entry);
 135 }
 136
 137 static void
 138 __verify_volume(struct volume_info *volume)
 139 {
 140         hammer_volume_ondisk_t ondisk = volume->ondisk;
 141
 142         if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
 143                 errx(1, "verify_volume: Invalid volume signature %016jx",
 144                         ondisk->vol_signature);
 145         }
 146         if (ondisk->vol_rootvol != HAMMER_ROOT_VOLNO) {
 147                 errx(1, "verify_volume: Invalid root volume# %d",
 148                         ondisk->vol_rootvol);
 149         }
 150         if (bcmp(&Hammer_FSType, &ondisk->vol_fstype, sizeof(Hammer_FSType))) {
 151                 errx(1, "verify_volume: %s: Header does not indicate "
 152                         "that this is a HAMMER volume", volume->name);
 153         }
 154         if (bcmp(&Hammer_FSId, &ondisk->vol_fsid, sizeof(Hammer_FSId))) {
 155                 errx(1, "verify_volume: %s: FSId does not match other volumes!",
 156                         volume->name);
 157         }
 158 }
 159
 160 /*
 161  * Initialize a volume structure and ondisk vol_no field.
 162  */
 163 struct volume_info *
 164 init_volume(const char *filename, int oflags, int32_t vol_no)
 165 {
 166         struct volume_info *volume;
 167
 168         volume = __alloc_volume(filename, oflags);
 169         volume->vol_no = volume->ondisk->vol_no = vol_no;
 170
 171         __add_volume(volume);
 172
 173         return(volume);
 174 }
 175
 176 /*
 177  * Initialize a volume structure and read ondisk volume header.
 178  */
 179 struct volume_info*
 180 load_volume(const char *filename, int oflags, int verify)
 181 {
 182         struct volume_info *volume;
 183         int n;
 184
 185         volume = __alloc_volume(filename, oflags);
 186
 187         n = readhammervol(volume);
 188         if (n == -1) {
 189                 err(1, "load_volume: %s: Read failed at offset 0",
 190                     volume->name);
 191         }
 192         volume->vol_no = volume->ondisk->vol_no;
 193         HammerVersion = volume->ondisk->vol_version;
 194
 195         if (valid_hammer_volumes++ == 0)
 196                 Hammer_FSId = volume->ondisk->vol_fsid;
 197         if (verify)
 198                 __verify_volume(volume);
 199
 200         __add_volume(volume);
 201
 202         return(volume);
 203 }
 204
 205 /*
 206  * Check basic volume characteristics.
 207  */
 208 static void
 209 check_volume(struct volume_info *volume)
 210 {
 211         struct partinfo pinfo;
 212         struct stat st;
 213
 214         /*
 215          * Get basic information about the volume
 216          */
 217         if (ioctl(volume->fd, DIOCGPART, &pinfo) < 0) {
 218                 /*
 219                  * Allow the formatting of regular files as HAMMER volumes
 220                  */
 221                 if (fstat(volume->fd, &st) < 0)
 222                         err(1, "Unable to stat %s", volume->name);
 223                 volume->size = st.st_size;
 224                 volume->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                 volume->size = pinfo.media_size;
 242                 volume->device_offset = pinfo.media_offset;
 243                 volume->type = "DEVICE";
 244         }
 245 }
 246
 247 int
 248 is_regfile(struct volume_info *volume)
 249 {
 250         return(strcmp(volume->type, "REGFILE") ? 0 : 1);
 251 }
 252
 253 void
 254 assert_volume_offset(struct volume_info *volume)
 255 {
 256         assert(hammer_is_zone_raw_buffer(volume->vol_free_off));
 257         assert(hammer_is_zone_raw_buffer(volume->vol_free_end));
 258         if (volume->vol_free_off >= volume->vol_free_end)
 259                 errx(1, "Ran out of room, filesystem too small");
 260 }
 261
 262 struct volume_info *
 263 get_volume(int32_t vol_no)
 264 {
 265         struct volume_info *volume;
 266
 267         TAILQ_FOREACH(volume, &VolList, entry) {
 268                 if (volume->vol_no == vol_no)
 269                         break;
 270         }
 271
 272         return(volume);
 273 }
 274
 275 struct volume_info *
 276 get_root_volume(void)
 277 {
 278         return(get_volume(HAMMER_ROOT_VOLNO));
 279 }
 280
 281 static hammer_off_t
 282 __blockmap_xlate_to_zone2(hammer_off_t buf_offset)
 283 {
 284         hammer_off_t zone2_offset;
 285         int error = 0;
 286
 287         if (hammer_is_zone_raw_buffer(buf_offset))
 288                 zone2_offset = buf_offset;
 289         else
 290                 zone2_offset = blockmap_lookup(buf_offset, &error);
 291
 292         if (error)
 293                 return(HAMMER_OFF_BAD);
 294         assert(hammer_is_zone_raw_buffer(zone2_offset));
 295
 296         return(zone2_offset);
 297 }
 298
 299 static struct buffer_info *
 300 __alloc_buffer(hammer_off_t zone2_offset, int isnew)
 301 {
 302         struct volume_info *volume;
 303         struct buffer_info *buffer;
 304         int hi;
 305
 306         volume = get_volume(HAMMER_VOL_DECODE(zone2_offset));
 307         assert(volume != NULL);
 308
 309         buffer = calloc(1, sizeof(*buffer));
 310         buffer->zone2_offset = zone2_offset;
 311         buffer->raw_offset = hammer_xlate_to_phys(volume->ondisk, zone2_offset);
 312         buffer->volume = volume;
 313         buffer->ondisk = calloc(1, HAMMER_BUFSIZE);
 314
 315         if (isnew <= 0) {
 316                 if (readhammerbuf(buffer) == -1) {
 317                         err(1, "Failed to read %s:%016jx at %016jx",
 318                             volume->name,
 319                             (intmax_t)buffer->zone2_offset,
 320                             (intmax_t)buffer->raw_offset);
 321                 }
 322         }
 323
 324         hi = buffer_hash(zone2_offset);
 325         TAILQ_INSERT_TAIL(&volume->buffer_lists[hi], buffer, entry);
 326         hammer_cache_add(&buffer->cache);
 327
 328         return(buffer);
 329 }
 330
 331 /*
 332  * Acquire the 16KB buffer for specified zone offset.
 333  */
 334 static struct buffer_info *
 335 get_buffer(hammer_off_t buf_offset, int isnew)
 336 {
 337         struct buffer_info *buffer;
 338         hammer_off_t zone2_offset;
 339         int dora = 0;
 340
 341         zone2_offset = __blockmap_xlate_to_zone2(buf_offset);
 342         if (zone2_offset == HAMMER_OFF_BAD)
 343                 return(NULL);
 344
 345         zone2_offset &= ~HAMMER_BUFMASK64;
 346         buffer = find_buffer(zone2_offset);
 347
 348         if (buffer == NULL) {
 349                 buffer = __alloc_buffer(zone2_offset, isnew);
 350                 dora = (isnew == 0);
 351         } else {
 352                 assert(isnew != -1);
 353                 hammer_cache_used(&buffer->cache);
 354         }
 355         assert(buffer->ondisk != NULL);
 356
 357         ++buffer->cache.refs;
 358         hammer_cache_flush();
 359
 360         if (isnew > 0) {
 361                 assert(buffer->cache.modified == 0);
 362                 bzero(buffer->ondisk, HAMMER_BUFSIZE);
 363                 buffer->cache.modified = 1;
 364         }
 365         if (dora)
 366                 get_buffer_readahead(buffer);
 367         return(buffer);
 368 }
 369
 370 static void
 371 get_buffer_readahead(struct buffer_info *base)
 372 {
 373         struct buffer_info *buffer;
 374         struct volume_info *volume;
 375         hammer_off_t zone2_offset;
 376         int64_t raw_offset;
 377         int ri = UseReadBehind;
 378         int re = UseReadAhead;
 379
 380         raw_offset = base->raw_offset + ri * HAMMER_BUFSIZE;
 381         volume = base->volume;
 382
 383         while (ri < re) {
 384                 if (raw_offset >= volume->ondisk->vol_buf_end)
 385                         break;
 386                 if (raw_offset < volume->ondisk->vol_buf_beg || ri == 0) {
 387                         ++ri;
 388                         raw_offset += HAMMER_BUFSIZE;
 389                         continue;
 390                 }
 391                 zone2_offset = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no,
 392                         raw_offset - volume->ondisk->vol_buf_beg);
 393                 buffer = find_buffer(zone2_offset);
 394                 if (buffer == NULL) {
 395                         /* call with -1 to prevent another readahead */
 396                         buffer = get_buffer(zone2_offset, -1);
 397                         rel_buffer(buffer);
 398                 }
 399                 ++ri;
 400                 raw_offset += HAMMER_BUFSIZE;
 401         }
 402 }
 403
 404 void
 405 rel_buffer(struct buffer_info *buffer)
 406 {
 407         struct volume_info *volume;
 408         int hi;
 409
 410         if (buffer == NULL)
 411                 return;
 412         assert(buffer->cache.refs > 0);
 413         if (--buffer->cache.refs == 0) {
 414                 if (buffer->cache.delete) {
 415                         hi = buffer_hash(buffer->zone2_offset);
 416                         volume = buffer->volume;
 417                         if (buffer->cache.modified)
 418                                 flush_buffer(buffer);
 419                         TAILQ_REMOVE(&volume->buffer_lists[hi], buffer, entry);
 420                         hammer_cache_del(&buffer->cache);
 421                         free(buffer->ondisk);
 422                         free(buffer);
 423                 }
 424         }
 425 }
 426
 427 /*
 428  * Retrieve a pointer to a buffer data given a buffer offset.  The underlying
 429  * bufferp is freed if isnew or the offset is out of range of the cached data.
 430  * If bufferp is freed a referenced buffer is loaded into it.
 431  */
 432 void *
 433 get_buffer_data(hammer_off_t buf_offset, struct buffer_info **bufferp,
 434                 int isnew)
 435 {
 436         hammer_off_t xor;
 437
 438         if (*bufferp != NULL) {
 439                 /* XXX xor is always non zero for indirect zones */
 440                 xor = HAMMER_OFF_LONG_ENCODE(buf_offset) ^
 441                       HAMMER_OFF_LONG_ENCODE((*bufferp)->zone2_offset);
 442                 if (isnew > 0 || (xor & ~HAMMER_BUFMASK64)) {
 443                         rel_buffer(*bufferp);
 444                         *bufferp = NULL;
 445                 }
 446         }
 447
 448         if (*bufferp == NULL) {
 449                 *bufferp = get_buffer(buf_offset, isnew);
 450                 if (*bufferp == NULL)
 451                         return(NULL);
 452         }
 453
 454         return(((char *)(*bufferp)->ondisk) +
 455                 ((int32_t)buf_offset & HAMMER_BUFMASK));
 456 }
 457
 458 /*
 459  * Allocate HAMMER elements - B-Tree nodes
 460  */
 461 hammer_node_ondisk_t
 462 alloc_btree_node(hammer_off_t *offp, struct buffer_info **data_bufferp)
 463 {
 464         hammer_node_ondisk_t node;
 465
 466         node = alloc_blockmap(HAMMER_ZONE_BTREE_INDEX, sizeof(*node),
 467                               offp, data_bufferp);
 468         bzero(node, sizeof(*node));
 469         return(node);
 470 }
 471
 472 /*
 473  * Allocate HAMMER elements - meta data (inode, direntry, PFS, etc)
 474  */
 475 void *
 476 alloc_meta_element(hammer_off_t *offp, int32_t data_len,
 477                    struct buffer_info **data_bufferp)
 478 {
 479         void *data;
 480
 481         data = alloc_blockmap(HAMMER_ZONE_META_INDEX, data_len,
 482                               offp, data_bufferp);
 483         bzero(data, data_len);
 484         return(data);
 485 }
 486
 487 /*
 488  * Format a new blockmap.  This is mostly a degenerate case because
 489  * all allocations are now actually done from the freemap.
 490  */
 491 void
 492 format_blockmap(struct volume_info *root_vol, int zone, hammer_off_t offset)
 493 {
 494         hammer_blockmap_t blockmap;
 495         hammer_off_t zone_base;
 496
 497         /* Only root volume needs formatting */
 498         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 499
 500         assert(hammer_is_index_record(zone));
 501
 502         blockmap = &root_vol->ondisk->vol0_blockmap[zone];
 503         zone_base = HAMMER_ZONE_ENCODE(zone, offset);
 504
 505         bzero(blockmap, sizeof(*blockmap));
 506         blockmap->phys_offset = 0;
 507         blockmap->first_offset = zone_base;
 508         blockmap->next_offset = zone_base;
 509         blockmap->alloc_offset = HAMMER_ENCODE(zone, 255, -1);
 510         hammer_crc_set_blockmap(HammerVersion, blockmap);
 511 }
 512
 513 /*
 514  * Format a new freemap.  Set all layer1 entries to UNAVAIL.  The initialize
 515  * code will load each volume's freemap.
 516  */
 517 void
 518 format_freemap(struct volume_info *root_vol)
 519 {
 520         struct buffer_info *buffer = NULL;
 521         hammer_off_t layer1_offset;
 522         hammer_blockmap_t blockmap;
 523         hammer_blockmap_layer1_t layer1;
 524         int i, isnew;
 525
 526         /* Only root volume needs formatting */
 527         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 528
 529         layer1_offset = bootstrap_bigblock(root_vol);
 530         for (i = 0; i < HAMMER_BIGBLOCK_SIZE; i += sizeof(*layer1)) {
 531                 isnew = ((i % HAMMER_BUFSIZE) == 0);
 532                 layer1 = get_buffer_data(layer1_offset + i, &buffer, isnew);
 533                 bzero(layer1, sizeof(*layer1));
 534                 layer1->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 535                 layer1->blocks_free = 0;
 536                 hammer_crc_set_layer1(HammerVersion, layer1);
 537         }
 538         assert(i == HAMMER_BIGBLOCK_SIZE);
 539         rel_buffer(buffer);
 540
 541         blockmap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 542         bzero(blockmap, sizeof(*blockmap));
 543         blockmap->phys_offset = layer1_offset;
 544         blockmap->first_offset = 0;
 545         blockmap->next_offset = HAMMER_ENCODE_RAW_BUFFER(0, 0);
 546         blockmap->alloc_offset = HAMMER_ENCODE_RAW_BUFFER(255, -1);
 547         hammer_crc_set_blockmap(HammerVersion, blockmap);
 548 }
 549
 550 /*
 551  * Load the volume's remaining free space into the freemap.
 552  *
 553  * Returns the number of big-blocks available.
 554  */
 555 int64_t
 556 initialize_freemap(struct volume_info *volume)
 557 {
 558         struct volume_info *root_vol;
 559         struct buffer_info *buffer1 = NULL;
 560         struct buffer_info *buffer2 = NULL;
 561         hammer_blockmap_layer1_t layer1;
 562         hammer_blockmap_layer2_t layer2;
 563         hammer_off_t layer1_offset;
 564         hammer_off_t layer2_offset;
 565         hammer_off_t phys_offset;
 566         hammer_off_t block_offset;
 567         hammer_off_t aligned_vol_free_end;
 568         hammer_blockmap_t freemap;
 569         int64_t count = 0;
 570         int64_t layer1_count = 0;
 571
 572         root_vol = get_root_volume();
 573
 574         assert_volume_offset(volume);
 575         aligned_vol_free_end = HAMMER_BLOCKMAP_LAYER2_DOALIGN(volume->vol_free_end);
 576
 577         printf("initialize freemap volume %d\n", volume->vol_no);
 578
 579         /*
 580          * Initialize the freemap.  First preallocate the big-blocks required
 581          * to implement layer2.   This preallocation is a bootstrap allocation
 582          * using blocks from the target volume.
 583          */
 584         freemap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 585
 586         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no, 0);
 587              phys_offset < aligned_vol_free_end;
 588              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 589                 layer1_offset = freemap->phys_offset +
 590                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 591                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 592                 if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 593                         layer1->phys_offset = bootstrap_bigblock(volume);
 594                         layer1->blocks_free = 0;
 595                         buffer1->cache.modified = 1;
 596                         hammer_crc_set_layer1(HammerVersion, layer1);
 597                 }
 598         }
 599
 600         /*
 601          * Now fill everything in.
 602          */
 603         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no, 0);
 604              phys_offset < aligned_vol_free_end;
 605              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 606                 layer1_count = 0;
 607                 layer1_offset = freemap->phys_offset +
 608                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 609                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 610                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 611
 612                 for (block_offset = 0;
 613                      block_offset < HAMMER_BLOCKMAP_LAYER2;
 614                      block_offset += HAMMER_BIGBLOCK_SIZE) {
 615                         layer2_offset = layer1->phys_offset +
 616                                         HAMMER_BLOCKMAP_LAYER2_OFFSET(block_offset);
 617                         layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 618                         bzero(layer2, sizeof(*layer2));
 619
 620                         if (phys_offset + block_offset < volume->vol_free_off) {
 621                                 /*
 622                                  * Big-blocks already allocated as part
 623                                  * of the freemap bootstrap.
 624                                  */
 625                                 layer2->zone = HAMMER_ZONE_FREEMAP_INDEX;
 626                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 627                                 layer2->bytes_free = 0;
 628                         } else if (phys_offset + block_offset < volume->vol_free_end) {
 629                                 layer2->zone = 0;
 630                                 layer2->append_off = 0;
 631                                 layer2->bytes_free = HAMMER_BIGBLOCK_SIZE;
 632                                 ++count;
 633                                 ++layer1_count;
 634                         } else {
 635                                 layer2->zone = HAMMER_ZONE_UNAVAIL_INDEX;
 636                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 637                                 layer2->bytes_free = 0;
 638                         }
 639                         hammer_crc_set_layer2(HammerVersion, layer2);
 640                         buffer2->cache.modified = 1;
 641                 }
 642
 643                 layer1->blocks_free += layer1_count;
 644                 hammer_crc_set_layer1(HammerVersion, layer1);
 645                 buffer1->cache.modified = 1;
 646         }
 647
 648         rel_buffer(buffer1);
 649         rel_buffer(buffer2);
 650         return(count);
 651 }
 652
 653 /*
 654  * Returns the number of big-blocks available for filesystem data and undos
 655  * without formatting.
 656  */
 657 int64_t
 658 count_freemap(struct volume_info *volume)
 659 {
 660         hammer_off_t phys_offset;
 661         hammer_off_t vol_free_off;
 662         hammer_off_t aligned_vol_free_end;
 663         int64_t count = 0;
 664
 665         vol_free_off = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no, 0);
 666
 667         assert_volume_offset(volume);
 668         aligned_vol_free_end = HAMMER_BLOCKMAP_LAYER2_DOALIGN(volume->vol_free_end);
 669
 670         if (volume->vol_no == HAMMER_ROOT_VOLNO)
 671                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 672
 673         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no, 0);
 674              phys_offset < aligned_vol_free_end;
 675              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 676                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 677         }
 678
 679         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no, 0);
 680              phys_offset < aligned_vol_free_end;
 681              phys_offset += HAMMER_BIGBLOCK_SIZE) {
 682                 if (phys_offset < vol_free_off)
 683                         ;
 684                 else if (phys_offset < volume->vol_free_end)
 685                         ++count;
 686         }
 687
 688         return(count);
 689 }
 690
 691 /*
 692  * Format the undomap for the root volume.
 693  */
 694 void
 695 format_undomap(struct volume_info *root_vol, int64_t *undo_buffer_size)
 696 {
 697         hammer_off_t undo_limit;
 698         hammer_blockmap_t blockmap;
 699         hammer_volume_ondisk_t ondisk;
 700         struct buffer_info *buffer = NULL;
 701         hammer_off_t scan;
 702         int n;
 703         int limit_index;
 704         uint32_t seqno;
 705
 706         /* Only root volume needs formatting */
 707         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 708         ondisk = root_vol->ondisk;
 709
 710         /*
 711          * Size the undo buffer in multiples of HAMMER_BIGBLOCK_SIZE,
 712          * up to HAMMER_MAX_UNDO_BIGBLOCKS big-blocks.
 713          * Size to approximately 0.1% of the disk.
 714          *
 715          * The minimum UNDO fifo size is 512MB, or approximately 1% of
 716          * the recommended 50G disk.
 717          *
 718          * Changing this minimum is rather dangerous as complex filesystem
 719          * operations can cause the UNDO FIFO to fill up otherwise.
 720          */
 721         undo_limit = *undo_buffer_size;
 722         if (undo_limit == 0) {
 723                 undo_limit = HAMMER_VOL_BUF_SIZE(ondisk) / 1000;
 724                 if (undo_limit < HAMMER_BIGBLOCK_SIZE * HAMMER_MIN_UNDO_BIGBLOCKS)
 725                         undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_MIN_UNDO_BIGBLOCKS;
 726         }
 727         undo_limit = HAMMER_BIGBLOCK_DOALIGN(undo_limit);
 728         if (undo_limit < HAMMER_BIGBLOCK_SIZE)
 729                 undo_limit = HAMMER_BIGBLOCK_SIZE;
 730         if (undo_limit > HAMMER_BIGBLOCK_SIZE * HAMMER_MAX_UNDO_BIGBLOCKS)
 731                 undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_MAX_UNDO_BIGBLOCKS;
 732         *undo_buffer_size = undo_limit;
 733
 734         blockmap = &ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 735         bzero(blockmap, sizeof(*blockmap));
 736         blockmap->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 737         blockmap->first_offset = HAMMER_ENCODE_UNDO(0);
 738         blockmap->next_offset = blockmap->first_offset;
 739         blockmap->alloc_offset = HAMMER_ENCODE_UNDO(undo_limit);
 740         hammer_crc_set_blockmap(HammerVersion, blockmap);
 741
 742         limit_index = undo_limit / HAMMER_BIGBLOCK_SIZE;
 743         assert(limit_index <= HAMMER_MAX_UNDO_BIGBLOCKS);
 744
 745         for (n = 0; n < limit_index; ++n)
 746                 ondisk->vol0_undo_array[n] = alloc_undo_bigblock(root_vol);
 747         while (n < HAMMER_MAX_UNDO_BIGBLOCKS)
 748                 ondisk->vol0_undo_array[n++] = HAMMER_BLOCKMAP_UNAVAIL;
 749
 750         /*
 751          * Pre-initialize the UNDO blocks (HAMMER version 4+)
 752          */
 753         printf("initializing the undo map (%jd MB)\n",
 754                 (intmax_t)HAMMER_OFF_LONG_ENCODE(blockmap->alloc_offset) /
 755                 (1024 * 1024));
 756
 757         scan = blockmap->first_offset;
 758         seqno = 0;
 759
 760         while (scan < blockmap->alloc_offset) {
 761                 hammer_fifo_head_t head;
 762                 hammer_fifo_tail_t tail;
 763                 int isnew;
 764                 int bytes = HAMMER_UNDO_ALIGN;
 765
 766                 isnew = ((scan & HAMMER_BUFMASK64) == 0);
 767                 head = get_buffer_data(scan, &buffer, isnew);
 768                 buffer->cache.modified = 1;
 769                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 770
 771                 bzero(head, bytes);
 772                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 773                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
 774                 head->hdr_size = bytes;
 775                 head->hdr_seq = seqno++;
 776
 777                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 778                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 779                 tail->tail_size = bytes;
 780
 781                 hammer_crc_set_fifo_head(HammerVersion, head, bytes);
 782
 783                 scan += bytes;
 784         }
 785         rel_buffer(buffer);
 786 }
 787
 788 const char *zone_labels[] = {
 789         "",             /* 0 */
 790         "raw_volume",   /* 1 */
 791         "raw_buffer",   /* 2 */
 792         "undo",         /* 3 */
 793         "freemap",      /* 4 */
 794         "",             /* 5 */
 795         "",             /* 6 */
 796         "",             /* 7 */
 797         "btree",        /* 8 */
 798         "meta",         /* 9 */
 799         "large_data",   /* 10 */
 800         "small_data",   /* 11 */
 801         "",             /* 12 */
 802         "",             /* 13 */
 803         "",             /* 14 */
 804         "unavail",      /* 15 */
 805 };
 806
 807 void
 808 print_blockmap(const struct volume_info *volume)
 809 {
 810         hammer_blockmap_t blockmap;
 811         hammer_volume_ondisk_t ondisk;
 812         int64_t size, used;
 813         int i;
 814 #define INDENT ""
 815
 816         ondisk = volume->ondisk;
 817         printf(INDENT"vol_label\t%s\n", ondisk->vol_label);
 818         printf(INDENT"vol_count\t%d\n", ondisk->vol_count);
 819         printf(INDENT"vol_bot_beg\t%s\n", sizetostr(ondisk->vol_bot_beg));
 820         printf(INDENT"vol_mem_beg\t%s\n", sizetostr(ondisk->vol_mem_beg));
 821         printf(INDENT"vol_buf_beg\t%s\n", sizetostr(ondisk->vol_buf_beg));
 822         printf(INDENT"vol_buf_end\t%s\n", sizetostr(ondisk->vol_buf_end));
 823         printf(INDENT"vol0_next_tid\t%016jx\n",
 824                (uintmax_t)ondisk->vol0_next_tid);
 825
 826         blockmap = &ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 827         size = HAMMER_OFF_LONG_ENCODE(blockmap->alloc_offset);
 828         if (blockmap->first_offset <= blockmap->next_offset)
 829                 used = blockmap->next_offset - blockmap->first_offset;
 830         else
 831                 used = blockmap->alloc_offset - blockmap->first_offset +
 832                         HAMMER_OFF_LONG_ENCODE(blockmap->next_offset);
 833         printf(INDENT"undo_size\t%s\n", sizetostr(size));
 834         printf(INDENT"undo_used\t%s\n", sizetostr(used));
 835
 836         printf(INDENT"zone #             "
 837                "phys             first            next             alloc\n");
 838         for (i = 0; i < HAMMER_MAX_ZONES; i++) {
 839                 blockmap = &ondisk->vol0_blockmap[i];
 840                 printf(INDENT"zone %-2d %-10s %016jx %016jx %016jx %016jx\n",
 841                         i, zone_labels[i],
 842                         (uintmax_t)blockmap->phys_offset,
 843                         (uintmax_t)blockmap->first_offset,
 844                         (uintmax_t)blockmap->next_offset,
 845                         (uintmax_t)blockmap->alloc_offset);
 846         }
 847 }
 848
 849 /*
 850  * Flush various tracking structures to disk
 851  */
 852 void
 853 flush_all_volumes(void)
 854 {
 855         struct volume_info *volume;
 856
 857         TAILQ_FOREACH(volume, &VolList, entry)
 858                 flush_volume(volume);
 859 }
 860
 861 void
 862 flush_volume(struct volume_info *volume)
 863 {
 864         struct buffer_info *buffer;
 865         int i;
 866
 867         for (i = 0; i < HAMMER_BUFLISTS; ++i) {
 868                 TAILQ_FOREACH(buffer, &volume->buffer_lists[i], entry)
 869                         flush_buffer(buffer);
 870         }
 871         if (writehammervol(volume) == -1)
 872                 err(1, "Write volume %d (%s)", volume->vol_no, volume->name);
 873 }
 874
 875 void
 876 flush_buffer(struct buffer_info *buffer)
 877 {
 878         struct volume_info *volume;
 879
 880         volume = buffer->volume;
 881         if (writehammerbuf(buffer) == -1)
 882                 err(1, "Write volume %d (%s)", volume->vol_no, volume->name);
 883         buffer->cache.modified = 0;
 884 }
 885
 886 /*
 887  * Core I/O operations
 888  */
 889 static int
 890 __read(struct volume_info *volume, void *data, int64_t offset, int size)
 891 {
 892         ssize_t n;
 893
 894         n = pread(volume->fd, data, size, offset);
 895         if (n != size)
 896                 return(-1);
 897         return(0);
 898 }
 899
 900 static __inline int
 901 readhammervol(struct volume_info *volume)
 902 {
 903         return(__read(volume, volume->ondisk, 0, HAMMER_BUFSIZE));
 904 }
 905
 906 static __inline int
 907 readhammerbuf(struct buffer_info *buffer)
 908 {
 909         return(__read(buffer->volume, buffer->ondisk, buffer->raw_offset,
 910                 HAMMER_BUFSIZE));
 911 }
 912
 913 static int
 914 __write(struct volume_info *volume, const void *data, int64_t offset, int size)
 915 {
 916         ssize_t n;
 917
 918         if (volume->rdonly)
 919                 return(0);
 920
 921         n = pwrite(volume->fd, data, size, offset);
 922         if (n != size)
 923                 return(-1);
 924         return(0);
 925 }
 926
 927 static __inline int
 928 writehammervol(struct volume_info *volume)
 929 {
 930         return(__write(volume, volume->ondisk, 0, HAMMER_BUFSIZE));
 931 }
 932
 933 static __inline int
 934 writehammerbuf(struct buffer_info *buffer)
 935 {
 936         return(__write(buffer->volume, buffer->ondisk, buffer->raw_offset,
 937                 HAMMER_BUFSIZE));
 938 }
 939
 940 int64_t init_boot_area_size(int64_t value, off_t avg_vol_size)
 941 {
 942         if (value == 0) {
 943                 value = HAMMER_BOOT_NOMBYTES;
 944                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 945                         value >>= 1;
 946         }
 947
 948         if (value < HAMMER_BOOT_MINBYTES)
 949                 value = HAMMER_BOOT_MINBYTES;
 950         else if (value > HAMMER_BOOT_MAXBYTES)
 951                 value = HAMMER_BOOT_MAXBYTES;
 952
 953         return(value);
 954 }
 955
 956 int64_t init_memory_log_size(int64_t value, off_t avg_vol_size)
 957 {
 958         if (value == 0) {
 959                 value = HAMMER_MEM_NOMBYTES;
 960                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 961                         value >>= 1;
 962         }
 963
 964         if (value < HAMMER_MEM_MINBYTES)
 965                 value = HAMMER_MEM_MINBYTES;
 966         else if (value > HAMMER_MEM_MAXBYTES)
 967                 value = HAMMER_MEM_MAXBYTES;
 968
 969         return(value);
 970 }