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
  56 static __inline
  57 int
  58 buffer_hash(hammer_off_t buf_offset)
  59 {
  60         int hi;
  61
  62         hi = (int)(buf_offset / HAMMER_BUFSIZE) & HAMMER_BUFLISTMASK;
  63         return(hi);
  64 }
  65
  66 static struct buffer_info*
  67 find_buffer(struct volume_info *volume, hammer_off_t buf_offset)
  68 {
  69         int hi;
  70         struct buffer_info *buf;
  71
  72         hi = buffer_hash(buf_offset);
  73         TAILQ_FOREACH(buf, &volume->buffer_lists[hi], entry)
  74                 if (buf->buf_offset == buf_offset)
  75                         return(buf);
  76         return(NULL);
  77 }
  78
  79 static
  80 struct volume_info *
  81 __alloc_volume(const char *volname, int oflags)
  82 {
  83         struct volume_info *vol;
  84         int i;
  85
  86         vol = malloc(sizeof(*vol));
  87         if (vol == NULL)
  88                 err(1, "alloc_volume");
  89         bzero(vol, sizeof(*vol));
  90
  91         vol->vol_no = -1;
  92         vol->name = strdup(volname);
  93         vol->fd = open(vol->name, oflags);
  94         if (vol->fd < 0)
  95                 err(1, "alloc_volume: Failed to open %s", vol->name);
  96
  97         vol->size = 0;
  98         vol->device_offset = 0;
  99         vol->type = NULL;
 100
 101         vol->ondisk = malloc(HAMMER_BUFSIZE);
 102         if (vol->ondisk == NULL)
 103                 err(1, "alloc_volume");
 104         bzero(vol->ondisk, HAMMER_BUFSIZE);
 105
 106         for (i = 0; i < HAMMER_BUFLISTS; ++i)
 107                 TAILQ_INIT(&vol->buffer_lists[i]);
 108
 109         return(vol);
 110 }
 111
 112 static void
 113 __add_volume(struct volume_info *vol)
 114 {
 115         struct volume_info *scan;
 116         struct stat st1, st2;
 117
 118         if (fstat(vol->fd, &st1) != 0)
 119                 errx(1, "add_volume: %s: Failed to stat", vol->name);
 120
 121         TAILQ_FOREACH(scan, &VolList, entry) {
 122                 if (scan->vol_no == vol->vol_no) {
 123                         errx(1, "add_volume: %s: Duplicate volume number %d "
 124                                 "against %s",
 125                                 vol->name, vol->vol_no, scan->name);
 126                 }
 127                 if (fstat(scan->fd, &st2) != 0) {
 128                         errx(1, "add_volume: %s: Failed to stat %s",
 129                                 vol->name, scan->name);
 130                 }
 131                 if ((st1.st_ino == st2.st_ino) && (st1.st_dev == st2.st_dev)) {
 132                         errx(1, "add_volume: %s: Specified more than once",
 133                                 vol->name);
 134                 }
 135         }
 136
 137         TAILQ_INSERT_TAIL(&VolList, vol, entry);
 138 }
 139
 140 /*
 141  * Initialize a volume structure and ondisk vol_no field.
 142  */
 143 struct volume_info *
 144 init_volume(int32_t vol_no, const char *filename, int oflags)
 145 {
 146         struct volume_info *vol;
 147
 148         vol = __alloc_volume(filename, oflags);
 149         vol->vol_no = vol->ondisk->vol_no = vol_no;
 150         vol->cache.modified = 1;
 151
 152         __add_volume(vol);
 153
 154         return(vol);
 155 }
 156
 157 /*
 158  * Initialize a volume structure and read ondisk volume header.
 159  */
 160 struct volume_info*
 161 load_volume(const char *filename, int oflags)
 162 {
 163         struct volume_info *vol;
 164         struct hammer_volume_ondisk *ondisk;
 165         int n;
 166
 167         vol = __alloc_volume(filename, oflags);
 168         ondisk = vol->ondisk;
 169
 170         n = readhammerbuf(vol, ondisk, 0);
 171         if (n == -1) {
 172                 err(1, "load_volume: %s: Read failed at offset 0", vol->name);
 173         }
 174         vol->vol_no = ondisk->vol_no;
 175
 176         if (ondisk->vol_rootvol != HAMMER_ROOT_VOLNO) {
 177                 errx(1, "load_volume: Invalid root volume# %d",
 178                         ondisk->vol_rootvol);
 179         }
 180
 181         if (bcmp(&Hammer_FSType, &ondisk->vol_fstype, sizeof(Hammer_FSType))) {
 182                 errx(1, "load_volume: %s: Header does not indicate "
 183                         "that this is a hammer volume", vol->name);
 184         }
 185
 186         if (TAILQ_EMPTY(&VolList)) {
 187                 Hammer_FSId = ondisk->vol_fsid;
 188         } else if (bcmp(&Hammer_FSId, &ondisk->vol_fsid, sizeof(Hammer_FSId))) {
 189                 errx(1, "load_volume: %s: FSId does match other volumes!",
 190                         vol->name);
 191         }
 192
 193         __add_volume(vol);
 194
 195         return(vol);
 196 }
 197
 198 /*
 199  * Check basic volume characteristics.
 200  */
 201 void
 202 check_volume(struct volume_info *vol)
 203 {
 204         struct partinfo pinfo;
 205         struct stat st;
 206
 207         /*
 208          * Get basic information about the volume
 209          */
 210         if (ioctl(vol->fd, DIOCGPART, &pinfo) < 0) {
 211                 /*
 212                  * Allow the formatting of regular files as HAMMER volumes
 213                  */
 214                 if (fstat(vol->fd, &st) < 0)
 215                         err(1, "Unable to stat %s", vol->name);
 216                 vol->size = st.st_size;
 217                 vol->type = "REGFILE";
 218         } else {
 219                 /*
 220                  * When formatting a block device as a HAMMER volume the
 221                  * sector size must be compatible.  HAMMER uses 16384 byte
 222                  * filesystem buffers.
 223                  */
 224                 if (pinfo.reserved_blocks) {
 225                         errx(1, "HAMMER cannot be placed in a partition "
 226                                 "which overlaps the disklabel or MBR");
 227                 }
 228                 if (pinfo.media_blksize > HAMMER_BUFSIZE ||
 229                     HAMMER_BUFSIZE % pinfo.media_blksize) {
 230                         errx(1, "A media sector size of %d is not supported",
 231                              pinfo.media_blksize);
 232                 }
 233
 234                 vol->size = pinfo.media_size;
 235                 vol->device_offset = pinfo.media_offset;
 236                 vol->type = "DEVICE";
 237         }
 238 }
 239
 240 struct volume_info *
 241 get_volume(int32_t vol_no)
 242 {
 243         struct volume_info *vol;
 244
 245         TAILQ_FOREACH(vol, &VolList, entry) {
 246                 if (vol->vol_no == vol_no)
 247                         break;
 248         }
 249         if (vol == NULL)
 250                 errx(1, "get_volume: Volume %d does not exist!", vol_no);
 251
 252         ++vol->cache.refs;
 253         /* not added to or removed from hammer cache */
 254         return(vol);
 255 }
 256
 257 struct volume_info *
 258 get_root_volume(void)
 259 {
 260         return(get_volume(HAMMER_ROOT_VOLNO));
 261 }
 262
 263 void
 264 rel_volume(struct volume_info *volume)
 265 {
 266         if (volume == NULL)
 267                 return;
 268         /* not added to or removed from hammer cache */
 269         --volume->cache.refs;
 270 }
 271
 272 /*
 273  * Acquire the specified buffer.  isnew is -1 only when called
 274  * via get_buffer_readahead() to prevent another readahead.
 275  */
 276 struct buffer_info *
 277 get_buffer(hammer_off_t buf_offset, int isnew)
 278 {
 279         void *ondisk;
 280         struct buffer_info *buf;
 281         struct volume_info *volume;
 282         hammer_off_t orig_offset = buf_offset;
 283         int vol_no;
 284         int zone;
 285         int hi, n;
 286         int dora = 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, NULL);
 291         }
 292         if (buf_offset == HAMMER_OFF_BAD)
 293                 return(NULL);
 294         assert((buf_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_RAW_BUFFER);
 295
 296         vol_no = HAMMER_VOL_DECODE(buf_offset);
 297         volume = get_volume(vol_no);
 298         if (volume == NULL)
 299                 return(NULL);
 300
 301         buf_offset &= ~HAMMER_BUFMASK64;
 302         buf = find_buffer(volume, buf_offset);
 303
 304         if (buf == NULL) {
 305                 buf = malloc(sizeof(*buf));
 306                 bzero(buf, sizeof(*buf));
 307                 if (DebugOpt > 1) {
 308                         fprintf(stderr, "get_buffer: %016llx %016llx at %p\n",
 309                                 (long long)orig_offset, (long long)buf_offset,
 310                                 buf);
 311                 }
 312                 buf->buf_offset = buf_offset;
 313                 buf->raw_offset = hammer_xlate_to_phys(volume->ondisk,
 314                                                         buf_offset);
 315                 buf->volume = volume;
 316                 hi = buffer_hash(buf_offset);
 317                 TAILQ_INSERT_TAIL(&volume->buffer_lists[hi], buf, entry);
 318                 ++volume->cache.refs;
 319                 buf->cache.u.buffer = buf;
 320                 hammer_cache_add(&buf->cache, ISBUFFER);
 321                 dora = (isnew == 0);
 322         } else {
 323                 if (DebugOpt > 1) {
 324                         fprintf(stderr, "get_buffer: %016llx %016llx at %p *\n",
 325                                 (long long)orig_offset, (long long)buf_offset,
 326                                 buf);
 327                 }
 328                 hammer_cache_used(&buf->cache);
 329                 ++buf->use_count;
 330         }
 331         ++buf->cache.refs;
 332         hammer_cache_flush();
 333         if ((ondisk = buf->ondisk) == NULL) {
 334                 buf->ondisk = ondisk = malloc(HAMMER_BUFSIZE);
 335                 if (isnew <= 0) {
 336                         n = readhammerbuf(volume, ondisk, buf->raw_offset);
 337                         if (n == -1) {
 338                                 if (AssertOnFailure)
 339                                         err(1, "get_buffer: %s:%016llx "
 340                                             "Read failed at offset %016llx",
 341                                             volume->name,
 342                                             (long long)buf->buf_offset,
 343                                             (long long)buf->raw_offset);
 344                                 bzero(ondisk, HAMMER_BUFSIZE);
 345                         }
 346                 }
 347         }
 348         if (isnew > 0) {
 349                 bzero(ondisk, HAMMER_BUFSIZE);
 350                 buf->cache.modified = 1;
 351         }
 352         if (dora)
 353                 get_buffer_readahead(buf);
 354         return(buf);
 355 }
 356
 357 static void
 358 get_buffer_readahead(struct buffer_info *base)
 359 {
 360         struct buffer_info *buf;
 361         struct volume_info *vol;
 362         hammer_off_t buf_offset;
 363         int64_t raw_offset;
 364         int ri = UseReadBehind;
 365         int re = UseReadAhead;
 366
 367         raw_offset = base->raw_offset + ri * HAMMER_BUFSIZE;
 368         vol = base->volume;
 369
 370         while (ri < re) {
 371                 if (raw_offset >= vol->ondisk->vol_buf_end)
 372                         break;
 373                 if (raw_offset < vol->ondisk->vol_buf_beg || ri == 0) {
 374                         ++ri;
 375                         raw_offset += HAMMER_BUFSIZE;
 376                         continue;
 377                 }
 378                 buf_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no,
 379                         raw_offset - vol->ondisk->vol_buf_beg);
 380                 buf = find_buffer(vol, buf_offset);
 381                 if (buf == NULL) {
 382                         buf = get_buffer(buf_offset, -1);
 383                         rel_buffer(buf);
 384                 }
 385                 ++ri;
 386                 raw_offset += HAMMER_BUFSIZE;
 387         }
 388 }
 389
 390 void
 391 rel_buffer(struct buffer_info *buffer)
 392 {
 393         struct volume_info *volume;
 394         int hi;
 395
 396         if (buffer == NULL)
 397                 return;
 398         assert(buffer->cache.refs > 0);
 399         if (--buffer->cache.refs == 0) {
 400                 if (buffer->cache.delete) {
 401                         hi = buffer_hash(buffer->buf_offset);
 402                         volume = buffer->volume;
 403                         if (buffer->cache.modified)
 404                                 flush_buffer(buffer);
 405                         TAILQ_REMOVE(&volume->buffer_lists[hi], buffer, entry);
 406                         hammer_cache_del(&buffer->cache);
 407                         free(buffer->ondisk);
 408                         free(buffer);
 409                         rel_volume(volume);
 410                 }
 411         }
 412 }
 413
 414 /*
 415  * Retrieve a pointer to a buffer data given a buffer offset.  The underlying
 416  * bufferp is freed if isnew or the offset is out of range of the cached data.
 417  * If bufferp is freed a referenced buffer is loaded into it.
 418  */
 419 void *
 420 get_buffer_data(hammer_off_t buf_offset, struct buffer_info **bufferp,
 421                 int isnew)
 422 {
 423         if (*bufferp != NULL) {
 424                 if (isnew > 0 ||
 425                     (((*bufferp)->buf_offset ^ buf_offset) & ~HAMMER_BUFMASK64)) {
 426                         rel_buffer(*bufferp);
 427                         *bufferp = NULL;
 428                 }
 429         }
 430         return(get_ondisk(buf_offset, bufferp, isnew));
 431 }
 432
 433 /*
 434  * Retrieve a pointer to a B-Tree node given a zone offset.  The underlying
 435  * bufferp is freed if non-NULL and a referenced buffer is loaded into it.
 436  */
 437 hammer_node_ondisk_t
 438 get_node(hammer_off_t node_offset, struct buffer_info **bufferp)
 439 {
 440         if (*bufferp != NULL) {
 441                 rel_buffer(*bufferp);
 442                 *bufferp = NULL;
 443         }
 444         return(get_ondisk(node_offset, bufferp, 0));
 445 }
 446
 447 /*
 448  * Return a pointer to a buffer data given a buffer offset.
 449  * If *bufferp is NULL acquire the buffer otherwise use that buffer.
 450  */
 451 static __inline
 452 void *
 453 get_ondisk(hammer_off_t buf_offset, struct buffer_info **bufferp, int isnew)
 454 {
 455         struct buffer_info *buffer;
 456
 457         buffer = *bufferp;
 458         if (buffer == NULL) {
 459                 buffer = *bufferp = get_buffer(buf_offset, isnew);
 460                 if (buffer == NULL)
 461                         return(NULL);
 462         }
 463
 464         return((char *)buffer->ondisk +
 465                 ((int32_t)buf_offset & HAMMER_BUFMASK));
 466 }
 467
 468 /*
 469  * Allocate HAMMER elements - btree nodes, meta data, data storage
 470  */
 471 void *
 472 alloc_btree_element(hammer_off_t *offp, struct buffer_info **data_bufferp)
 473 {
 474         hammer_node_ondisk_t node;
 475
 476         node = alloc_blockmap(HAMMER_ZONE_BTREE_INDEX, sizeof(*node),
 477                               offp, data_bufferp);
 478         bzero(node, sizeof(*node));
 479         return (node);
 480 }
 481
 482 void *
 483 alloc_meta_element(hammer_off_t *offp, int32_t data_len,
 484                    struct buffer_info **data_bufferp)
 485 {
 486         void *data;
 487
 488         data = alloc_blockmap(HAMMER_ZONE_META_INDEX, data_len,
 489                               offp, data_bufferp);
 490         bzero(data, data_len);
 491         return (data);
 492 }
 493
 494 /*
 495  * The only data_len supported by HAMMER userspace for large data zone
 496  * (zone 10) is HAMMER_BUFSIZE which is 16KB.  >16KB data does not fit
 497  * in a buffer allocated by get_buffer().  Also alloc_blockmap() does
 498  * not consider >16KB buffer size.
 499  */
 500 void *
 501 alloc_data_element(hammer_off_t *offp, int32_t data_len,
 502                    struct buffer_info **data_bufferp)
 503 {
 504         void *data;
 505         int zone;
 506
 507         if (data_len == 0)
 508                 return(NULL);
 509
 510         zone = hammer_data_zone_index(data_len);
 511         assert(data_len <= HAMMER_BUFSIZE); /* just one buffer */
 512         assert(zone == HAMMER_ZONE_LARGE_DATA_INDEX ||
 513                zone == HAMMER_ZONE_SMALL_DATA_INDEX);
 514
 515         data = alloc_blockmap(zone, data_len, offp, data_bufferp);
 516         bzero(data, data_len);
 517         return(data);
 518 }
 519
 520 /*
 521  * Format a new blockmap.  This is mostly a degenerate case because
 522  * all allocations are now actually done from the freemap.
 523  */
 524 void
 525 format_blockmap(struct volume_info *root_vol, int zone, hammer_off_t offset)
 526 {
 527         hammer_blockmap_t blockmap;
 528         hammer_off_t zone_base;
 529
 530         /* Only root volume needs formatting */
 531         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 532
 533         assert(hammer_is_zone2_mapped_index(zone));
 534
 535         blockmap = &root_vol->ondisk->vol0_blockmap[zone];
 536         zone_base = HAMMER_ZONE_ENCODE(zone, offset);
 537
 538         bzero(blockmap, sizeof(*blockmap));
 539         blockmap->phys_offset = 0;
 540         blockmap->first_offset = zone_base;
 541         blockmap->next_offset = zone_base;
 542         blockmap->alloc_offset = HAMMER_ENCODE(zone, 255, -1);
 543         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 544         root_vol->cache.modified = 1;
 545 }
 546
 547 /*
 548  * Format a new freemap.  Set all layer1 entries to UNAVAIL.  The initialize
 549  * code will load each volume's freemap.
 550  */
 551 void
 552 format_freemap(struct volume_info *root_vol)
 553 {
 554         struct buffer_info *buffer = NULL;
 555         hammer_off_t layer1_offset;
 556         hammer_blockmap_t blockmap;
 557         struct hammer_blockmap_layer1 *layer1;
 558         int i, isnew;
 559
 560         /* Only root volume needs formatting */
 561         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 562
 563         layer1_offset = alloc_bigblock(root_vol, HAMMER_ZONE_FREEMAP_INDEX);
 564         for (i = 0; i < HAMMER_BIGBLOCK_SIZE; i += sizeof(*layer1)) {
 565                 isnew = ((i % HAMMER_BUFSIZE) == 0);
 566                 layer1 = get_buffer_data(layer1_offset + i, &buffer, isnew);
 567                 bzero(layer1, sizeof(*layer1));
 568                 layer1->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 569                 layer1->blocks_free = 0;
 570                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 571         }
 572         assert(i == HAMMER_BIGBLOCK_SIZE);
 573         rel_buffer(buffer);
 574
 575         blockmap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 576         bzero(blockmap, sizeof(*blockmap));
 577         blockmap->phys_offset = layer1_offset;
 578         blockmap->first_offset = 0;
 579         blockmap->next_offset = HAMMER_ENCODE_RAW_BUFFER(0, 0);
 580         blockmap->alloc_offset = HAMMER_ENCODE_RAW_BUFFER(255, -1);
 581         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 582         root_vol->cache.modified = 1;
 583 }
 584
 585 /*
 586  * Load the volume's remaining free space into the freemap.
 587  *
 588  * Returns the number of big-blocks available.
 589  */
 590 int64_t
 591 initialize_freemap(struct volume_info *vol)
 592 {
 593         struct volume_info *root_vol;
 594         struct buffer_info *buffer1 = NULL;
 595         struct buffer_info *buffer2 = NULL;
 596         struct hammer_blockmap_layer1 *layer1;
 597         struct hammer_blockmap_layer2 *layer2;
 598         hammer_off_t layer1_base;
 599         hammer_off_t layer1_offset;
 600         hammer_off_t layer2_offset;
 601         hammer_off_t phys_offset;
 602         hammer_off_t block_offset;
 603         hammer_off_t aligned_vol_free_end;
 604         hammer_blockmap_t freemap;
 605         int64_t count = 0;
 606         int64_t layer1_count = 0;
 607
 608         root_vol = get_root_volume();
 609         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 610                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 611
 612         printf("initialize freemap volume %d\n", vol->vol_no);
 613
 614         /*
 615          * Initialize the freemap.  First preallocate the big-blocks required
 616          * to implement layer2.   This preallocation is a bootstrap allocation
 617          * using blocks from the target volume.
 618          */
 619         freemap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 620         layer1_base = freemap->phys_offset;
 621
 622         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 623              phys_offset < aligned_vol_free_end;
 624              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 625                 layer1_offset = layer1_base +
 626                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 627                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 628                 if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 629                         layer1->phys_offset = alloc_bigblock(vol,
 630                                                 HAMMER_ZONE_FREEMAP_INDEX);
 631                         layer1->blocks_free = 0;
 632                         buffer1->cache.modified = 1;
 633                         layer1->layer1_crc = crc32(layer1,
 634                                                    HAMMER_LAYER1_CRCSIZE);
 635                 }
 636         }
 637
 638         /*
 639          * Now fill everything in.
 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                 layer1_count = 0;
 645                 layer1_offset = layer1_base +
 646                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 647                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 648                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 649
 650                 for (block_offset = 0;
 651                      block_offset < HAMMER_BLOCKMAP_LAYER2;
 652                      block_offset += HAMMER_BIGBLOCK_SIZE) {
 653                         layer2_offset = layer1->phys_offset +
 654                                         HAMMER_BLOCKMAP_LAYER2_OFFSET(block_offset);
 655                         layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 656                         bzero(layer2, sizeof(*layer2));
 657
 658                         if (phys_offset + block_offset < vol->vol_free_off) {
 659                                 /*
 660                                  * Fixups XXX - big-blocks already allocated as part
 661                                  * of the freemap bootstrap.
 662                                  */
 663                                 if (layer2->zone == 0) {
 664                                         layer2->zone = HAMMER_ZONE_FREEMAP_INDEX;
 665                                         layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 666                                         layer2->bytes_free = 0;
 667                                 }
 668                         } else if (phys_offset + block_offset < vol->vol_free_end) {
 669                                 layer2->zone = 0;
 670                                 layer2->append_off = 0;
 671                                 layer2->bytes_free = HAMMER_BIGBLOCK_SIZE;
 672                                 ++count;
 673                                 ++layer1_count;
 674                         } else {
 675                                 layer2->zone = HAMMER_ZONE_UNAVAIL_INDEX;
 676                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 677                                 layer2->bytes_free = 0;
 678                         }
 679                         layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 680                         buffer2->cache.modified = 1;
 681                 }
 682
 683                 layer1->blocks_free += layer1_count;
 684                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 685                 buffer1->cache.modified = 1;
 686         }
 687
 688         rel_buffer(buffer1);
 689         rel_buffer(buffer2);
 690         rel_volume(root_vol);
 691         return(count);
 692 }
 693
 694 /*
 695  * Returns the number of big-blocks available for filesystem data and undos
 696  * without formatting.
 697  */
 698 int64_t
 699 count_freemap(struct volume_info *vol)
 700 {
 701         hammer_off_t phys_offset;
 702         hammer_off_t vol_free_off;
 703         hammer_off_t aligned_vol_free_end;
 704         int64_t count = 0;
 705
 706         vol_free_off = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 707         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 708                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 709
 710         if (vol->vol_no == HAMMER_ROOT_VOLNO)
 711                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 712
 713         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 714              phys_offset < aligned_vol_free_end;
 715              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 716                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 717         }
 718
 719         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 720              phys_offset < aligned_vol_free_end;
 721              phys_offset += HAMMER_BIGBLOCK_SIZE) {
 722                 if (phys_offset < vol_free_off) {
 723                         ;
 724                 } else if (phys_offset < vol->vol_free_end) {
 725                         ++count;
 726                 }
 727         }
 728
 729         return(count);
 730 }
 731
 732 /*
 733  * Format the undomap for the root volume.
 734  */
 735 void
 736 format_undomap(struct volume_info *root_vol, int64_t *undo_buffer_size)
 737 {
 738         const int undo_zone = HAMMER_ZONE_UNDO_INDEX;
 739         hammer_off_t undo_limit;
 740         hammer_blockmap_t blockmap;
 741         struct hammer_volume_ondisk *ondisk;
 742         struct buffer_info *buffer = NULL;
 743         hammer_off_t scan;
 744         int n;
 745         int limit_index;
 746         uint32_t seqno;
 747
 748         /* Only root volume needs formatting */
 749         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 750         ondisk = root_vol->ondisk;
 751
 752         /*
 753          * Size the undo buffer in multiples of HAMMER_BIGBLOCK_SIZE,
 754          * up to HAMMER_UNDO_LAYER2 big-blocks.  Size to approximately
 755          * 0.1% of the disk.
 756          *
 757          * The minimum UNDO fifo size is 500MB, or approximately 1% of
 758          * the recommended 50G disk.
 759          *
 760          * Changing this minimum is rather dangerous as complex filesystem
 761          * operations can cause the UNDO FIFO to fill up otherwise.
 762          */
 763         undo_limit = *undo_buffer_size;
 764         if (undo_limit == 0) {
 765                 undo_limit = (ondisk->vol_buf_end - ondisk->vol_buf_beg) / 1000;
 766                 if (undo_limit < 500*1024*1024)
 767                         undo_limit = 500*1024*1024;
 768         }
 769         undo_limit = (undo_limit + HAMMER_BIGBLOCK_MASK64) &
 770                      ~HAMMER_BIGBLOCK_MASK64;
 771         if (undo_limit < HAMMER_BIGBLOCK_SIZE)
 772                 undo_limit = HAMMER_BIGBLOCK_SIZE;
 773         if (undo_limit > HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2)
 774                 undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2;
 775         *undo_buffer_size = undo_limit;
 776
 777         blockmap = &ondisk->vol0_blockmap[undo_zone];
 778         bzero(blockmap, sizeof(*blockmap));
 779         blockmap->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 780         blockmap->first_offset = HAMMER_ZONE_ENCODE(undo_zone, 0);
 781         blockmap->next_offset = blockmap->first_offset;
 782         blockmap->alloc_offset = HAMMER_ZONE_ENCODE(undo_zone, undo_limit);
 783         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 784
 785         limit_index = undo_limit / HAMMER_BIGBLOCK_SIZE;
 786         assert(limit_index <= HAMMER_UNDO_LAYER2);
 787
 788         for (n = 0; n < limit_index; ++n) {
 789                 ondisk->vol0_undo_array[n] = alloc_bigblock(NULL,
 790                                                         HAMMER_ZONE_UNDO_INDEX);
 791         }
 792         while (n < HAMMER_UNDO_LAYER2) {
 793                 ondisk->vol0_undo_array[n++] = HAMMER_BLOCKMAP_UNAVAIL;
 794         }
 795
 796         /*
 797          * Pre-initialize the UNDO blocks (HAMMER version 4+)
 798          */
 799         printf("initializing the undo map (%jd MB)\n",
 800                 (intmax_t)(blockmap->alloc_offset & HAMMER_OFF_LONG_MASK) /
 801                 (1024 * 1024));
 802
 803         scan = blockmap->first_offset;
 804         seqno = 0;
 805
 806         while (scan < blockmap->alloc_offset) {
 807                 hammer_fifo_head_t head;
 808                 hammer_fifo_tail_t tail;
 809                 int isnew;
 810                 int bytes = HAMMER_UNDO_ALIGN;
 811
 812                 isnew = ((scan & HAMMER_BUFMASK64) == 0);
 813                 head = get_buffer_data(scan, &buffer, isnew);
 814                 buffer->cache.modified = 1;
 815                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 816
 817                 bzero(head, bytes);
 818                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 819                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
 820                 head->hdr_size = bytes;
 821                 head->hdr_seq = seqno++;
 822
 823                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 824                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 825                 tail->tail_size = bytes;
 826
 827                 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
 828                                 crc32(head + 1, bytes - sizeof(*head));
 829
 830                 scan += bytes;
 831         }
 832         rel_buffer(buffer);
 833 }
 834
 835 /*
 836  * Flush various tracking structures to disk
 837  */
 838 void
 839 flush_all_volumes(void)
 840 {
 841         struct volume_info *vol;
 842
 843         TAILQ_FOREACH(vol, &VolList, entry)
 844                 flush_volume(vol);
 845 }
 846
 847 void
 848 flush_volume(struct volume_info *volume)
 849 {
 850         struct buffer_info *buffer;
 851         int i;
 852
 853         for (i = 0; i < HAMMER_BUFLISTS; ++i) {
 854                 TAILQ_FOREACH(buffer, &volume->buffer_lists[i], entry)
 855                         flush_buffer(buffer);
 856         }
 857         if (writehammerbuf(volume, volume->ondisk, 0) == -1)
 858                 err(1, "Write volume %d (%s)", volume->vol_no, volume->name);
 859         volume->cache.modified = 0;
 860 }
 861
 862 void
 863 flush_buffer(struct buffer_info *buffer)
 864 {
 865         struct volume_info *vol;
 866
 867         vol = buffer->volume;
 868         if (writehammerbuf(vol, buffer->ondisk, buffer->raw_offset) == -1)
 869                 err(1, "Write volume %d (%s)", vol->vol_no, vol->name);
 870         buffer->cache.modified = 0;
 871 }
 872
 873 /*
 874  * Core I/O operations
 875  */
 876 static int
 877 readhammerbuf(struct volume_info *vol, void *data, int64_t offset)
 878 {
 879         ssize_t n;
 880
 881         n = pread(vol->fd, data, HAMMER_BUFSIZE, offset);
 882         if (n != HAMMER_BUFSIZE)
 883                 return(-1);
 884         return(0);
 885 }
 886
 887 static int
 888 writehammerbuf(struct volume_info *vol, const void *data, int64_t offset)
 889 {
 890         ssize_t n;
 891
 892         n = pwrite(vol->fd, data, HAMMER_BUFSIZE, offset);
 893         if (n != HAMMER_BUFSIZE)
 894                 return(-1);
 895         return(0);
 896 }
 897
 898 int64_t init_boot_area_size(int64_t value, off_t avg_vol_size)
 899 {
 900         if (value == 0) {
 901                 value = HAMMER_BOOT_NOMBYTES;
 902                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 903                         value >>= 1;
 904                 if (value < HAMMER_BOOT_MINBYTES)
 905                         value = 0;
 906         } else if (value < HAMMER_BOOT_MINBYTES) {
 907                 value = HAMMER_BOOT_MINBYTES;
 908         }
 909
 910         return(value);
 911 }
 912
 913 int64_t init_mem_area_size(int64_t value, off_t avg_vol_size)
 914 {
 915         if (value == 0) {
 916                 value = HAMMER_MEM_NOMBYTES;
 917                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 918                         value >>= 1;
 919                 if (value < HAMMER_MEM_MINBYTES)
 920                         value = 0;
 921         } else if (value < HAMMER_MEM_MINBYTES) {
 922                 value = HAMMER_MEM_MINBYTES;
 923         }
 924
 925         return(value);
 926 }