sbin/hammer/ondisk.c

   1 /*
   2  * Copyright (c) 2007 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  */
  34
  35 #include <sys/diskslice.h>
  36 #include <sys/diskmbr.h>
  37
  38 #include "hammer_util.h"
  39
  40 static void get_buffer_readahead(struct buffer_info *base);
  41 static __inline void *get_ondisk(hammer_off_t buf_offset,
  42                         struct buffer_info **bufferp, int isnew);
  43 static int readhammerbuf(struct volume_info *vol, void *data, int64_t offset);
  44 static int writehammerbuf(struct volume_info *vol, const void *data,
  45                         int64_t offset);
  46
  47 int DebugOpt;
  48
  49 uuid_t Hammer_FSType;
  50 uuid_t Hammer_FSId;
  51 int     UseReadBehind = -4;
  52 int     UseReadAhead = 4;
  53 int     AssertOnFailure = 1;
  54 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->name = strdup(volname);
  94         vol->fd = open(vol->name, oflags);
  95         if (vol->fd < 0)
  96                 err(1, "alloc_volume: Failed to open %s", vol->name);
  97
  98         vol->size = 0;
  99         vol->device_offset = 0;
 100         vol->type = NULL;
 101
 102         vol->ondisk = malloc(HAMMER_BUFSIZE);
 103         if (vol->ondisk == NULL)
 104                 err(1, "alloc_volume");
 105         bzero(vol->ondisk, HAMMER_BUFSIZE);
 106
 107         for (i = 0; i < HAMMER_BUFLISTS; ++i)
 108                 TAILQ_INIT(&vol->buffer_lists[i]);
 109
 110         return(vol);
 111 }
 112
 113 static void
 114 __add_volume(struct volume_info *vol)
 115 {
 116         struct volume_info *scan;
 117         struct stat st1, st2;
 118
 119         if (fstat(vol->fd, &st1) != 0)
 120                 errx(1, "add_volume: %s: Failed to stat", vol->name);
 121
 122         TAILQ_FOREACH(scan, &VolList, entry) {
 123                 if (scan->vol_no == vol->vol_no) {
 124                         errx(1, "add_volume: %s: Duplicate volume number %d "
 125                                 "against %s",
 126                                 vol->name, vol->vol_no, scan->name);
 127                 }
 128                 if (fstat(scan->fd, &st2) != 0) {
 129                         errx(1, "add_volume: %s: Failed to stat %s",
 130                                 vol->name, scan->name);
 131                 }
 132                 if ((st1.st_ino == st2.st_ino) && (st1.st_dev == st2.st_dev)) {
 133                         errx(1, "add_volume: %s: Specified more than once",
 134                                 vol->name);
 135                 }
 136         }
 137
 138         TAILQ_INSERT_TAIL(&VolList, vol, entry);
 139 }
 140
 141 /*
 142  * Initialize a volume structure and ondisk vol_no field.
 143  */
 144 struct volume_info *
 145 init_volume(int32_t vol_no, const char *filename, int oflags)
 146 {
 147         struct volume_info *vol;
 148
 149         vol = __alloc_volume(filename, oflags);
 150         vol->vol_no = vol->ondisk->vol_no = vol_no;
 151         vol->cache.modified = 1;
 152
 153         __add_volume(vol);
 154
 155         return(vol);
 156 }
 157
 158 /*
 159  * Initialize a volume structure and read ondisk volume header.
 160  */
 161 struct volume_info*
 162 load_volume(const char *filename, int oflags)
 163 {
 164         struct volume_info *vol;
 165         struct hammer_volume_ondisk *ondisk;
 166         int n;
 167
 168         vol = __alloc_volume(filename, oflags);
 169         ondisk = vol->ondisk;
 170
 171         n = readhammerbuf(vol, ondisk, 0);
 172         if (n == -1) {
 173                 err(1, "load_volume: %s: Read failed at offset 0", vol->name);
 174         }
 175         vol->vol_no = ondisk->vol_no;
 176
 177         if (ondisk->vol_rootvol != HAMMER_ROOT_VOLNO) {
 178                 errx(1, "load_volume: Invalid root volume# %d",
 179                         ondisk->vol_rootvol);
 180         }
 181
 182         if (bcmp(&Hammer_FSType, &ondisk->vol_fstype, sizeof(Hammer_FSType))) {
 183                 errx(1, "load_volume: %s: Header does not indicate "
 184                         "that this is a hammer volume", vol->name);
 185         }
 186
 187         if (valid_hammer_volumes++ == 0) {
 188                 Hammer_FSId = ondisk->vol_fsid;
 189         } else if (bcmp(&Hammer_FSId, &ondisk->vol_fsid, sizeof(Hammer_FSId))) {
 190                 errx(1, "load_volume: %s: FSId does match other volumes!",
 191                         vol->name);
 192         }
 193
 194         __add_volume(vol);
 195
 196         return(vol);
 197 }
 198
 199 /*
 200  * Check basic volume characteristics.
 201  */
 202 void
 203 check_volume(struct volume_info *vol)
 204 {
 205         struct partinfo pinfo;
 206         struct stat st;
 207
 208         /*
 209          * Get basic information about the volume
 210          */
 211         if (ioctl(vol->fd, DIOCGPART, &pinfo) < 0) {
 212                 /*
 213                  * Allow the formatting of regular files as HAMMER volumes
 214                  */
 215                 if (fstat(vol->fd, &st) < 0)
 216                         err(1, "Unable to stat %s", vol->name);
 217                 vol->size = st.st_size;
 218                 vol->type = "REGFILE";
 219         } else {
 220                 /*
 221                  * When formatting a block device as a HAMMER volume the
 222                  * sector size must be compatible.  HAMMER uses 16384 byte
 223                  * filesystem buffers.
 224                  */
 225                 if (pinfo.reserved_blocks) {
 226                         errx(1, "HAMMER cannot be placed in a partition "
 227                                 "which overlaps the disklabel or MBR");
 228                 }
 229                 if (pinfo.media_blksize > HAMMER_BUFSIZE ||
 230                     HAMMER_BUFSIZE % pinfo.media_blksize) {
 231                         errx(1, "A media sector size of %d is not supported",
 232                              pinfo.media_blksize);
 233                 }
 234
 235                 vol->size = pinfo.media_size;
 236                 vol->device_offset = pinfo.media_offset;
 237                 vol->type = "DEVICE";
 238         }
 239 }
 240
 241 struct volume_info *
 242 get_volume(int32_t vol_no)
 243 {
 244         struct volume_info *vol;
 245
 246         TAILQ_FOREACH(vol, &VolList, entry) {
 247                 if (vol->vol_no == vol_no)
 248                         break;
 249         }
 250         if (vol == NULL)
 251                 errx(1, "get_volume: Volume %d does not exist!", vol_no);
 252
 253         ++vol->cache.refs;
 254         /* not added to or removed from hammer cache */
 255         return(vol);
 256 }
 257
 258 struct volume_info *
 259 get_root_volume(void)
 260 {
 261         return(get_volume(HAMMER_ROOT_VOLNO));
 262 }
 263
 264 void
 265 rel_volume(struct volume_info *volume)
 266 {
 267         if (volume == NULL)
 268                 return;
 269         /* not added to or removed from hammer cache */
 270         --volume->cache.refs;
 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 struct buffer_info *
 278 get_buffer(hammer_off_t buf_offset, int isnew)
 279 {
 280         void *ondisk;
 281         struct buffer_info *buf;
 282         struct volume_info *volume;
 283         hammer_off_t orig_offset = buf_offset;
 284         int vol_no;
 285         int zone;
 286         int hi, n;
 287         int dora = 0;
 288
 289         zone = HAMMER_ZONE_DECODE(buf_offset);
 290         if (zone > HAMMER_ZONE_RAW_BUFFER_INDEX) {
 291                 buf_offset = blockmap_lookup(buf_offset, NULL, NULL, NULL);
 292         }
 293         if (buf_offset == HAMMER_OFF_BAD)
 294                 return(NULL);
 295         assert((buf_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_RAW_BUFFER);
 296
 297         vol_no = HAMMER_VOL_DECODE(buf_offset);
 298         volume = get_volume(vol_no);
 299         if (volume == NULL)
 300                 return(NULL);
 301
 302         buf_offset &= ~HAMMER_BUFMASK64;
 303         buf = find_buffer(volume, buf_offset);
 304
 305         if (buf == NULL) {
 306                 buf = malloc(sizeof(*buf));
 307                 bzero(buf, sizeof(*buf));
 308                 if (DebugOpt > 1) {
 309                         fprintf(stderr, "get_buffer: %016llx %016llx at %p\n",
 310                                 (long long)orig_offset, (long long)buf_offset,
 311                                 buf);
 312                 }
 313                 buf->buf_offset = buf_offset;
 314                 buf->raw_offset = hammer_xlate_to_phys(volume->ondisk,
 315                                                         buf_offset);
 316                 buf->volume = volume;
 317                 hi = buffer_hash(buf_offset);
 318                 TAILQ_INSERT_TAIL(&volume->buffer_lists[hi], buf, entry);
 319                 ++volume->cache.refs;
 320                 buf->cache.u.buffer = buf;
 321                 hammer_cache_add(&buf->cache, ISBUFFER);
 322                 dora = (isnew == 0);
 323         } else {
 324                 if (DebugOpt > 1) {
 325                         fprintf(stderr, "get_buffer: %016llx %016llx at %p *\n",
 326                                 (long long)orig_offset, (long long)buf_offset,
 327                                 buf);
 328                 }
 329                 hammer_cache_used(&buf->cache);
 330                 ++buf->use_count;
 331         }
 332         ++buf->cache.refs;
 333         hammer_cache_flush();
 334         if ((ondisk = buf->ondisk) == NULL) {
 335                 buf->ondisk = ondisk = malloc(HAMMER_BUFSIZE);
 336                 if (isnew <= 0) {
 337                         n = readhammerbuf(volume, ondisk, buf->raw_offset);
 338                         if (n == -1) {
 339                                 if (AssertOnFailure)
 340                                         err(1, "get_buffer: %s:%016llx "
 341                                             "Read failed at offset %016llx",
 342                                             volume->name,
 343                                             (long long)buf->buf_offset,
 344                                             (long long)buf->raw_offset);
 345                                 bzero(ondisk, HAMMER_BUFSIZE);
 346                         }
 347                 }
 348         }
 349         if (isnew > 0) {
 350                 bzero(ondisk, HAMMER_BUFSIZE);
 351                 buf->cache.modified = 1;
 352         }
 353         if (dora)
 354                 get_buffer_readahead(buf);
 355         return(buf);
 356 }
 357
 358 static void
 359 get_buffer_readahead(struct buffer_info *base)
 360 {
 361         struct buffer_info *buf;
 362         struct volume_info *vol;
 363         hammer_off_t buf_offset;
 364         int64_t raw_offset;
 365         int ri = UseReadBehind;
 366         int re = UseReadAhead;
 367
 368         raw_offset = base->raw_offset + ri * HAMMER_BUFSIZE;
 369         vol = base->volume;
 370
 371         while (ri < re) {
 372                 if (raw_offset >= vol->ondisk->vol_buf_end)
 373                         break;
 374                 if (raw_offset < vol->ondisk->vol_buf_beg || ri == 0) {
 375                         ++ri;
 376                         raw_offset += HAMMER_BUFSIZE;
 377                         continue;
 378                 }
 379                 buf_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no,
 380                         raw_offset - vol->ondisk->vol_buf_beg);
 381                 buf = find_buffer(vol, buf_offset);
 382                 if (buf == NULL) {
 383                         buf = get_buffer(buf_offset, -1);
 384                         rel_buffer(buf);
 385                 }
 386                 ++ri;
 387                 raw_offset += HAMMER_BUFSIZE;
 388         }
 389 }
 390
 391 void
 392 rel_buffer(struct buffer_info *buffer)
 393 {
 394         struct volume_info *volume;
 395         int hi;
 396
 397         if (buffer == NULL)
 398                 return;
 399         assert(buffer->cache.refs > 0);
 400         if (--buffer->cache.refs == 0) {
 401                 if (buffer->cache.delete) {
 402                         hi = buffer_hash(buffer->buf_offset);
 403                         volume = buffer->volume;
 404                         if (buffer->cache.modified)
 405                                 flush_buffer(buffer);
 406                         TAILQ_REMOVE(&volume->buffer_lists[hi], buffer, entry);
 407                         hammer_cache_del(&buffer->cache);
 408                         free(buffer->ondisk);
 409                         free(buffer);
 410                         rel_volume(volume);
 411                 }
 412         }
 413 }
 414
 415 /*
 416  * Retrieve a pointer to a buffer data given a buffer offset.  The underlying
 417  * bufferp is freed if isnew or the offset is out of range of the cached data.
 418  * If bufferp is freed a referenced buffer is loaded into it.
 419  */
 420 void *
 421 get_buffer_data(hammer_off_t buf_offset, struct buffer_info **bufferp,
 422                 int isnew)
 423 {
 424         if (*bufferp != NULL) {
 425                 if (isnew > 0 ||
 426                     (((*bufferp)->buf_offset ^ buf_offset) & ~HAMMER_BUFMASK64)) {
 427                         rel_buffer(*bufferp);
 428                         *bufferp = NULL;
 429                 }
 430         }
 431         return(get_ondisk(buf_offset, bufferp, isnew));
 432 }
 433
 434 /*
 435  * Retrieve a pointer to a B-Tree node given a zone offset.  The underlying
 436  * bufferp is freed if non-NULL and a referenced buffer is loaded into it.
 437  */
 438 hammer_node_ondisk_t
 439 get_node(hammer_off_t node_offset, struct buffer_info **bufferp)
 440 {
 441         if (*bufferp != NULL) {
 442                 rel_buffer(*bufferp);
 443                 *bufferp = NULL;
 444         }
 445         return(get_ondisk(node_offset, bufferp, 0));
 446 }
 447
 448 /*
 449  * Return a pointer to a buffer data given a buffer offset.
 450  * If *bufferp is NULL acquire the buffer otherwise use that buffer.
 451  */
 452 static __inline
 453 void *
 454 get_ondisk(hammer_off_t buf_offset, struct buffer_info **bufferp, int isnew)
 455 {
 456         struct buffer_info *buffer;
 457
 458         buffer = *bufferp;
 459         if (buffer == NULL) {
 460                 buffer = *bufferp = get_buffer(buf_offset, isnew);
 461                 if (buffer == NULL)
 462                         return(NULL);
 463         }
 464
 465         return((char *)buffer->ondisk +
 466                 ((int32_t)buf_offset & HAMMER_BUFMASK));
 467 }
 468
 469 /*
 470  * Allocate HAMMER elements - btree nodes, meta data, data storage
 471  */
 472 void *
 473 alloc_btree_element(hammer_off_t *offp, struct buffer_info **data_bufferp)
 474 {
 475         hammer_node_ondisk_t node;
 476
 477         node = alloc_blockmap(HAMMER_ZONE_BTREE_INDEX, sizeof(*node),
 478                               offp, data_bufferp);
 479         bzero(node, sizeof(*node));
 480         return (node);
 481 }
 482
 483 void *
 484 alloc_meta_element(hammer_off_t *offp, int32_t data_len,
 485                    struct buffer_info **data_bufferp)
 486 {
 487         void *data;
 488
 489         data = alloc_blockmap(HAMMER_ZONE_META_INDEX, data_len,
 490                               offp, data_bufferp);
 491         bzero(data, data_len);
 492         return (data);
 493 }
 494
 495 /*
 496  * The only data_len supported by HAMMER userspace for large data zone
 497  * (zone 10) is HAMMER_BUFSIZE which is 16KB.  >16KB data does not fit
 498  * in a buffer allocated by get_buffer().  Also alloc_blockmap() does
 499  * not consider >16KB buffer size.
 500  */
 501 void *
 502 alloc_data_element(hammer_off_t *offp, int32_t data_len,
 503                    struct buffer_info **data_bufferp)
 504 {
 505         void *data;
 506         int zone;
 507
 508         if (data_len == 0)
 509                 return(NULL);
 510
 511         zone = hammer_data_zone_index(data_len);
 512         assert(data_len <= HAMMER_BUFSIZE); /* just one buffer */
 513         assert(zone == HAMMER_ZONE_LARGE_DATA_INDEX ||
 514                zone == HAMMER_ZONE_SMALL_DATA_INDEX);
 515
 516         data = alloc_blockmap(zone, data_len, offp, data_bufferp);
 517         bzero(data, data_len);
 518         return(data);
 519 }
 520
 521 /*
 522  * Format a new blockmap.  This is mostly a degenerate case because
 523  * all allocations are now actually done from the freemap.
 524  */
 525 void
 526 format_blockmap(struct volume_info *root_vol, int zone, hammer_off_t offset)
 527 {
 528         hammer_blockmap_t blockmap;
 529         hammer_off_t zone_base;
 530
 531         /* Only root volume needs formatting */
 532         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 533
 534         assert(hammer_is_zone2_mapped_index(zone));
 535
 536         blockmap = &root_vol->ondisk->vol0_blockmap[zone];
 537         zone_base = HAMMER_ZONE_ENCODE(zone, offset);
 538
 539         bzero(blockmap, sizeof(*blockmap));
 540         blockmap->phys_offset = 0;
 541         blockmap->first_offset = zone_base;
 542         blockmap->next_offset = zone_base;
 543         blockmap->alloc_offset = HAMMER_ENCODE(zone, 255, -1);
 544         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 545         root_vol->cache.modified = 1;
 546 }
 547
 548 /*
 549  * Format a new freemap.  Set all layer1 entries to UNAVAIL.  The initialize
 550  * code will load each volume's freemap.
 551  */
 552 void
 553 format_freemap(struct volume_info *root_vol)
 554 {
 555         struct buffer_info *buffer = NULL;
 556         hammer_off_t layer1_offset;
 557         hammer_blockmap_t blockmap;
 558         struct hammer_blockmap_layer1 *layer1;
 559         int i, isnew;
 560
 561         /* Only root volume needs formatting */
 562         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 563
 564         layer1_offset = alloc_bigblock(root_vol, HAMMER_ZONE_FREEMAP_INDEX);
 565         for (i = 0; i < HAMMER_BIGBLOCK_SIZE; i += sizeof(*layer1)) {
 566                 isnew = ((i % HAMMER_BUFSIZE) == 0);
 567                 layer1 = get_buffer_data(layer1_offset + i, &buffer, isnew);
 568                 bzero(layer1, sizeof(*layer1));
 569                 layer1->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 570                 layer1->blocks_free = 0;
 571                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 572         }
 573         assert(i == HAMMER_BIGBLOCK_SIZE);
 574         rel_buffer(buffer);
 575
 576         blockmap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 577         bzero(blockmap, sizeof(*blockmap));
 578         blockmap->phys_offset = layer1_offset;
 579         blockmap->first_offset = 0;
 580         blockmap->next_offset = HAMMER_ENCODE_RAW_BUFFER(0, 0);
 581         blockmap->alloc_offset = HAMMER_ENCODE_RAW_BUFFER(255, -1);
 582         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 583         root_vol->cache.modified = 1;
 584 }
 585
 586 /*
 587  * Load the volume's remaining free space into the freemap.
 588  *
 589  * Returns the number of big-blocks available.
 590  */
 591 int64_t
 592 initialize_freemap(struct volume_info *vol)
 593 {
 594         struct volume_info *root_vol;
 595         struct buffer_info *buffer1 = NULL;
 596         struct buffer_info *buffer2 = NULL;
 597         struct hammer_blockmap_layer1 *layer1;
 598         struct hammer_blockmap_layer2 *layer2;
 599         hammer_off_t layer1_base;
 600         hammer_off_t layer1_offset;
 601         hammer_off_t layer2_offset;
 602         hammer_off_t phys_offset;
 603         hammer_off_t block_offset;
 604         hammer_off_t aligned_vol_free_end;
 605         hammer_blockmap_t freemap;
 606         int64_t count = 0;
 607         int64_t layer1_count = 0;
 608
 609         root_vol = get_root_volume();
 610         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 611                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 612
 613         printf("initialize freemap volume %d\n", vol->vol_no);
 614
 615         /*
 616          * Initialize the freemap.  First preallocate the big-blocks required
 617          * to implement layer2.   This preallocation is a bootstrap allocation
 618          * using blocks from the target volume.
 619          */
 620         freemap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 621         layer1_base = freemap->phys_offset;
 622
 623         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 624              phys_offset < aligned_vol_free_end;
 625              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 626                 layer1_offset = layer1_base +
 627                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 628                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 629                 if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 630                         layer1->phys_offset = alloc_bigblock(vol,
 631                                                 HAMMER_ZONE_FREEMAP_INDEX);
 632                         layer1->blocks_free = 0;
 633                         buffer1->cache.modified = 1;
 634                         layer1->layer1_crc = crc32(layer1,
 635                                                    HAMMER_LAYER1_CRCSIZE);
 636                 }
 637         }
 638
 639         /*
 640          * Now fill everything in.
 641          */
 642         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 643              phys_offset < aligned_vol_free_end;
 644              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 645                 layer1_count = 0;
 646                 layer1_offset = layer1_base +
 647                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 648                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 649                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 650
 651                 for (block_offset = 0;
 652                      block_offset < HAMMER_BLOCKMAP_LAYER2;
 653                      block_offset += HAMMER_BIGBLOCK_SIZE) {
 654                         layer2_offset = layer1->phys_offset +
 655                                         HAMMER_BLOCKMAP_LAYER2_OFFSET(block_offset);
 656                         layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 657                         bzero(layer2, sizeof(*layer2));
 658
 659                         if (phys_offset + block_offset < vol->vol_free_off) {
 660                                 /*
 661                                  * Fixups XXX - big-blocks already allocated as part
 662                                  * of the freemap bootstrap.
 663                                  */
 664                                 if (layer2->zone == 0) {
 665                                         layer2->zone = HAMMER_ZONE_FREEMAP_INDEX;
 666                                         layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 667                                         layer2->bytes_free = 0;
 668                                 }
 669                         } else if (phys_offset + block_offset < vol->vol_free_end) {
 670                                 layer2->zone = 0;
 671                                 layer2->append_off = 0;
 672                                 layer2->bytes_free = HAMMER_BIGBLOCK_SIZE;
 673                                 ++count;
 674                                 ++layer1_count;
 675                         } else {
 676                                 layer2->zone = HAMMER_ZONE_UNAVAIL_INDEX;
 677                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 678                                 layer2->bytes_free = 0;
 679                         }
 680                         layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 681                         buffer2->cache.modified = 1;
 682                 }
 683
 684                 layer1->blocks_free += layer1_count;
 685                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 686                 buffer1->cache.modified = 1;
 687         }
 688
 689         rel_buffer(buffer1);
 690         rel_buffer(buffer2);
 691         rel_volume(root_vol);
 692         return(count);
 693 }
 694
 695 /*
 696  * Returns the number of big-blocks available for filesystem data and undos
 697  * without formatting.
 698  */
 699 int64_t
 700 count_freemap(struct volume_info *vol)
 701 {
 702         hammer_off_t phys_offset;
 703         hammer_off_t vol_free_off;
 704         hammer_off_t aligned_vol_free_end;
 705         int64_t count = 0;
 706
 707         vol_free_off = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 708         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 709                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 710
 711         if (vol->vol_no == HAMMER_ROOT_VOLNO)
 712                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 713
 714         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 715              phys_offset < aligned_vol_free_end;
 716              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 717                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 718         }
 719
 720         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 721              phys_offset < aligned_vol_free_end;
 722              phys_offset += HAMMER_BIGBLOCK_SIZE) {
 723                 if (phys_offset < vol_free_off) {
 724                         ;
 725                 } else if (phys_offset < vol->vol_free_end) {
 726                         ++count;
 727                 }
 728         }
 729
 730         return(count);
 731 }
 732
 733 /*
 734  * Format the undomap for the root volume.
 735  */
 736 void
 737 format_undomap(struct volume_info *root_vol, int64_t *undo_buffer_size)
 738 {
 739         const int undo_zone = HAMMER_ZONE_UNDO_INDEX;
 740         hammer_off_t undo_limit;
 741         hammer_blockmap_t blockmap;
 742         struct hammer_volume_ondisk *ondisk;
 743         struct buffer_info *buffer = NULL;
 744         hammer_off_t scan;
 745         int n;
 746         int limit_index;
 747         uint32_t seqno;
 748
 749         /* Only root volume needs formatting */
 750         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 751         ondisk = root_vol->ondisk;
 752
 753         /*
 754          * Size the undo buffer in multiples of HAMMER_BIGBLOCK_SIZE,
 755          * up to HAMMER_UNDO_LAYER2 big-blocks.  Size to approximately
 756          * 0.1% of the disk.
 757          *
 758          * The minimum UNDO fifo size is 500MB, or approximately 1% of
 759          * the recommended 50G disk.
 760          *
 761          * Changing this minimum is rather dangerous as complex filesystem
 762          * operations can cause the UNDO FIFO to fill up otherwise.
 763          */
 764         undo_limit = *undo_buffer_size;
 765         if (undo_limit == 0) {
 766                 undo_limit = (ondisk->vol_buf_end - ondisk->vol_buf_beg) / 1000;
 767                 if (undo_limit < 500*1024*1024)
 768                         undo_limit = 500*1024*1024;
 769         }
 770         undo_limit = (undo_limit + HAMMER_BIGBLOCK_MASK64) &
 771                      ~HAMMER_BIGBLOCK_MASK64;
 772         if (undo_limit < HAMMER_BIGBLOCK_SIZE)
 773                 undo_limit = HAMMER_BIGBLOCK_SIZE;
 774         if (undo_limit > HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2)
 775                 undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2;
 776         *undo_buffer_size = undo_limit;
 777
 778         blockmap = &ondisk->vol0_blockmap[undo_zone];
 779         bzero(blockmap, sizeof(*blockmap));
 780         blockmap->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 781         blockmap->first_offset = HAMMER_ZONE_ENCODE(undo_zone, 0);
 782         blockmap->next_offset = blockmap->first_offset;
 783         blockmap->alloc_offset = HAMMER_ZONE_ENCODE(undo_zone, undo_limit);
 784         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 785
 786         limit_index = undo_limit / HAMMER_BIGBLOCK_SIZE;
 787         assert(limit_index <= HAMMER_UNDO_LAYER2);
 788
 789         for (n = 0; n < limit_index; ++n) {
 790                 ondisk->vol0_undo_array[n] = alloc_bigblock(NULL,
 791                                                         HAMMER_ZONE_UNDO_INDEX);
 792         }
 793         while (n < HAMMER_UNDO_LAYER2) {
 794                 ondisk->vol0_undo_array[n++] = HAMMER_BLOCKMAP_UNAVAIL;
 795         }
 796
 797         /*
 798          * Pre-initialize the UNDO blocks (HAMMER version 4+)
 799          */
 800         printf("initializing the undo map (%jd MB)\n",
 801                 (intmax_t)(blockmap->alloc_offset & HAMMER_OFF_LONG_MASK) /
 802                 (1024 * 1024));
 803
 804         scan = blockmap->first_offset;
 805         seqno = 0;
 806
 807         while (scan < blockmap->alloc_offset) {
 808                 hammer_fifo_head_t head;
 809                 hammer_fifo_tail_t tail;
 810                 int isnew;
 811                 int bytes = HAMMER_UNDO_ALIGN;
 812
 813                 isnew = ((scan & HAMMER_BUFMASK64) == 0);
 814                 head = get_buffer_data(scan, &buffer, isnew);
 815                 buffer->cache.modified = 1;
 816                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 817
 818                 bzero(head, bytes);
 819                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 820                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
 821                 head->hdr_size = bytes;
 822                 head->hdr_seq = seqno++;
 823
 824                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 825                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 826                 tail->tail_size = bytes;
 827
 828                 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
 829                                 crc32(head + 1, bytes - sizeof(*head));
 830
 831                 scan += bytes;
 832         }
 833         rel_buffer(buffer);
 834 }
 835
 836 /*
 837  * Flush various tracking structures to disk
 838  */
 839 void
 840 flush_all_volumes(void)
 841 {
 842         struct volume_info *vol;
 843
 844         TAILQ_FOREACH(vol, &VolList, entry)
 845                 flush_volume(vol);
 846 }
 847
 848 void
 849 flush_volume(struct volume_info *volume)
 850 {
 851         struct buffer_info *buffer;
 852         int i;
 853
 854         for (i = 0; i < HAMMER_BUFLISTS; ++i) {
 855                 TAILQ_FOREACH(buffer, &volume->buffer_lists[i], entry)
 856                         flush_buffer(buffer);
 857         }
 858         if (writehammerbuf(volume, volume->ondisk, 0) == -1)
 859                 err(1, "Write volume %d (%s)", volume->vol_no, volume->name);
 860         volume->cache.modified = 0;
 861 }
 862
 863 void
 864 flush_buffer(struct buffer_info *buffer)
 865 {
 866         struct volume_info *vol;
 867
 868         vol = buffer->volume;
 869         if (writehammerbuf(vol, buffer->ondisk, buffer->raw_offset) == -1)
 870                 err(1, "Write volume %d (%s)", vol->vol_no, vol->name);
 871         buffer->cache.modified = 0;
 872 }
 873
 874 /*
 875  * Core I/O operations
 876  */
 877 static int
 878 readhammerbuf(struct volume_info *vol, void *data, int64_t offset)
 879 {
 880         ssize_t n;
 881
 882         n = pread(vol->fd, data, HAMMER_BUFSIZE, offset);
 883         if (n != HAMMER_BUFSIZE)
 884                 return(-1);
 885         return(0);
 886 }
 887
 888 static int
 889 writehammerbuf(struct volume_info *vol, const void *data, int64_t offset)
 890 {
 891         ssize_t n;
 892
 893         n = pwrite(vol->fd, data, HAMMER_BUFSIZE, offset);
 894         if (n != HAMMER_BUFSIZE)
 895                 return(-1);
 896         return(0);
 897 }
 898
 899 int64_t init_boot_area_size(int64_t value, off_t avg_vol_size)
 900 {
 901         if (value == 0) {
 902                 value = HAMMER_BOOT_NOMBYTES;
 903                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 904                         value >>= 1;
 905                 if (value < HAMMER_BOOT_MINBYTES)
 906                         value = 0;
 907         } else if (value < HAMMER_BOOT_MINBYTES) {
 908                 value = HAMMER_BOOT_MINBYTES;
 909         }
 910
 911         return(value);
 912 }
 913
 914 int64_t init_mem_area_size(int64_t value, off_t avg_vol_size)
 915 {
 916         if (value == 0) {
 917                 value = HAMMER_MEM_NOMBYTES;
 918                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 919                         value >>= 1;
 920                 if (value < HAMMER_MEM_MINBYTES)
 921                         value = 0;
 922         } else if (value < HAMMER_MEM_MINBYTES) {
 923                 value = HAMMER_MEM_MINBYTES;
 924         }
 925
 926         return(value);
 927 }