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->rdonly = (oflags == O_RDONLY);
  94         vol->name = strdup(volname);
  95         vol->fd = open(vol->name, oflags);
  96         if (vol->fd < 0)
  97                 err(1, "alloc_volume: Failed to open %s", vol->name);
  98
  99         vol->size = 0;
 100         vol->device_offset = 0;
 101         vol->type = NULL;
 102
 103         vol->ondisk = malloc(HAMMER_BUFSIZE);
 104         if (vol->ondisk == NULL)
 105                 err(1, "alloc_volume");
 106         bzero(vol->ondisk, HAMMER_BUFSIZE);
 107
 108         for (i = 0; i < HAMMER_BUFLISTS; ++i)
 109                 TAILQ_INIT(&vol->buffer_lists[i]);
 110
 111         return(vol);
 112 }
 113
 114 static void
 115 __add_volume(struct volume_info *vol)
 116 {
 117         struct volume_info *scan;
 118         struct stat st1, st2;
 119
 120         if (fstat(vol->fd, &st1) != 0)
 121                 errx(1, "add_volume: %s: Failed to stat", vol->name);
 122
 123         TAILQ_FOREACH(scan, &VolList, entry) {
 124                 if (scan->vol_no == vol->vol_no) {
 125                         errx(1, "add_volume: %s: Duplicate volume number %d "
 126                                 "against %s",
 127                                 vol->name, vol->vol_no, scan->name);
 128                 }
 129                 if (fstat(scan->fd, &st2) != 0) {
 130                         errx(1, "add_volume: %s: Failed to stat %s",
 131                                 vol->name, scan->name);
 132                 }
 133                 if ((st1.st_ino == st2.st_ino) && (st1.st_dev == st2.st_dev)) {
 134                         errx(1, "add_volume: %s: Specified more than once",
 135                                 vol->name);
 136                 }
 137         }
 138
 139         TAILQ_INSERT_TAIL(&VolList, vol, entry);
 140 }
 141
 142 /*
 143  * Initialize a volume structure and ondisk vol_no field.
 144  */
 145 struct volume_info *
 146 init_volume(int32_t vol_no, const char *filename, int oflags)
 147 {
 148         struct volume_info *vol;
 149
 150         vol = __alloc_volume(filename, oflags);
 151         vol->vol_no = vol->ondisk->vol_no = vol_no;
 152         vol->cache.modified = 1;
 153
 154         __add_volume(vol);
 155
 156         return(vol);
 157 }
 158
 159 /*
 160  * Initialize a volume structure and read ondisk volume header.
 161  */
 162 struct volume_info*
 163 load_volume(const char *filename, int oflags)
 164 {
 165         struct volume_info *vol;
 166         struct hammer_volume_ondisk *ondisk;
 167         int n;
 168
 169         vol = __alloc_volume(filename, oflags);
 170         ondisk = vol->ondisk;
 171
 172         n = readhammerbuf(vol, ondisk, 0);
 173         if (n == -1) {
 174                 err(1, "load_volume: %s: Read failed at offset 0", vol->name);
 175         }
 176         vol->vol_no = ondisk->vol_no;
 177
 178         if (ondisk->vol_rootvol != HAMMER_ROOT_VOLNO) {
 179                 errx(1, "load_volume: Invalid root volume# %d",
 180                         ondisk->vol_rootvol);
 181         }
 182
 183         if (bcmp(&Hammer_FSType, &ondisk->vol_fstype, sizeof(Hammer_FSType))) {
 184                 errx(1, "load_volume: %s: Header does not indicate "
 185                         "that this is a hammer volume", vol->name);
 186         }
 187
 188         if (valid_hammer_volumes++ == 0) {
 189                 Hammer_FSId = ondisk->vol_fsid;
 190         } else if (bcmp(&Hammer_FSId, &ondisk->vol_fsid, sizeof(Hammer_FSId))) {
 191                 errx(1, "load_volume: %s: FSId does match other volumes!",
 192                         vol->name);
 193         }
 194
 195         __add_volume(vol);
 196
 197         return(vol);
 198 }
 199
 200 /*
 201  * Check basic volume characteristics.
 202  */
 203 void
 204 check_volume(struct volume_info *vol)
 205 {
 206         struct partinfo pinfo;
 207         struct stat st;
 208
 209         /*
 210          * Get basic information about the volume
 211          */
 212         if (ioctl(vol->fd, DIOCGPART, &pinfo) < 0) {
 213                 /*
 214                  * Allow the formatting of regular files as HAMMER volumes
 215                  */
 216                 if (fstat(vol->fd, &st) < 0)
 217                         err(1, "Unable to stat %s", vol->name);
 218                 vol->size = st.st_size;
 219                 vol->type = "REGFILE";
 220         } else {
 221                 /*
 222                  * When formatting a block device as a HAMMER volume the
 223                  * sector size must be compatible.  HAMMER uses 16384 byte
 224                  * filesystem buffers.
 225                  */
 226                 if (pinfo.reserved_blocks) {
 227                         errx(1, "HAMMER cannot be placed in a partition "
 228                                 "which overlaps the disklabel or MBR");
 229                 }
 230                 if (pinfo.media_blksize > HAMMER_BUFSIZE ||
 231                     HAMMER_BUFSIZE % pinfo.media_blksize) {
 232                         errx(1, "A media sector size of %d is not supported",
 233                              pinfo.media_blksize);
 234                 }
 235
 236                 vol->size = pinfo.media_size;
 237                 vol->device_offset = pinfo.media_offset;
 238                 vol->type = "DEVICE";
 239         }
 240 }
 241
 242 struct volume_info *
 243 get_volume(int32_t vol_no)
 244 {
 245         struct volume_info *vol;
 246
 247         TAILQ_FOREACH(vol, &VolList, entry) {
 248                 if (vol->vol_no == vol_no)
 249                         break;
 250         }
 251         if (vol == NULL)
 252                 errx(1, "get_volume: Volume %d does not exist!", vol_no);
 253
 254         ++vol->cache.refs;
 255         /* not added to or removed from hammer cache */
 256         return(vol);
 257 }
 258
 259 struct volume_info *
 260 get_root_volume(void)
 261 {
 262         return(get_volume(HAMMER_ROOT_VOLNO));
 263 }
 264
 265 void
 266 rel_volume(struct volume_info *volume)
 267 {
 268         if (volume == NULL)
 269                 return;
 270         /* not added to or removed from hammer cache */
 271         --volume->cache.refs;
 272 }
 273
 274 /*
 275  * Acquire the specified buffer.  isnew is -1 only when called
 276  * via get_buffer_readahead() to prevent another readahead.
 277  */
 278 struct buffer_info *
 279 get_buffer(hammer_off_t buf_offset, int isnew)
 280 {
 281         void *ondisk;
 282         struct buffer_info *buf;
 283         struct volume_info *volume;
 284         hammer_off_t orig_offset = buf_offset;
 285         int vol_no;
 286         int zone;
 287         int hi, n;
 288         int dora = 0;
 289
 290         zone = HAMMER_ZONE_DECODE(buf_offset);
 291         if (zone > HAMMER_ZONE_RAW_BUFFER_INDEX) {
 292                 buf_offset = blockmap_lookup(buf_offset, NULL, NULL, NULL);
 293         }
 294         if (buf_offset == HAMMER_OFF_BAD)
 295                 return(NULL);
 296         assert((buf_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_RAW_BUFFER);
 297
 298         vol_no = HAMMER_VOL_DECODE(buf_offset);
 299         volume = get_volume(vol_no);
 300         if (volume == NULL)
 301                 return(NULL);
 302
 303         buf_offset &= ~HAMMER_BUFMASK64;
 304         buf = find_buffer(volume, buf_offset);
 305
 306         if (buf == NULL) {
 307                 buf = malloc(sizeof(*buf));
 308                 bzero(buf, sizeof(*buf));
 309                 if (DebugOpt > 1) {
 310                         fprintf(stderr, "get_buffer: %016llx %016llx at %p\n",
 311                                 (long long)orig_offset, (long long)buf_offset,
 312                                 buf);
 313                 }
 314                 buf->buf_offset = buf_offset;
 315                 buf->raw_offset = hammer_xlate_to_phys(volume->ondisk,
 316                                                         buf_offset);
 317                 buf->volume = volume;
 318                 hi = buffer_hash(buf_offset);
 319                 TAILQ_INSERT_TAIL(&volume->buffer_lists[hi], buf, entry);
 320                 ++volume->cache.refs;
 321                 buf->cache.u.buffer = buf;
 322                 hammer_cache_add(&buf->cache, ISBUFFER);
 323                 dora = (isnew == 0);
 324         } else {
 325                 if (DebugOpt > 1) {
 326                         fprintf(stderr, "get_buffer: %016llx %016llx at %p *\n",
 327                                 (long long)orig_offset, (long long)buf_offset,
 328                                 buf);
 329                 }
 330                 hammer_cache_used(&buf->cache);
 331                 ++buf->use_count;
 332         }
 333         ++buf->cache.refs;
 334         hammer_cache_flush();
 335         if ((ondisk = buf->ondisk) == NULL) {
 336                 buf->ondisk = ondisk = malloc(HAMMER_BUFSIZE);
 337                 if (isnew <= 0) {
 338                         n = readhammerbuf(volume, ondisk, buf->raw_offset);
 339                         if (n == -1) {
 340                                 if (AssertOnFailure)
 341                                         err(1, "get_buffer: %s:%016llx "
 342                                             "Read failed at offset %016llx",
 343                                             volume->name,
 344                                             (long long)buf->buf_offset,
 345                                             (long long)buf->raw_offset);
 346                                 bzero(ondisk, HAMMER_BUFSIZE);
 347                         }
 348                 }
 349         }
 350         if (isnew > 0) {
 351                 bzero(ondisk, HAMMER_BUFSIZE);
 352                 buf->cache.modified = 1;
 353         }
 354         if (dora)
 355                 get_buffer_readahead(buf);
 356         return(buf);
 357 }
 358
 359 static void
 360 get_buffer_readahead(struct buffer_info *base)
 361 {
 362         struct buffer_info *buf;
 363         struct volume_info *vol;
 364         hammer_off_t buf_offset;
 365         int64_t raw_offset;
 366         int ri = UseReadBehind;
 367         int re = UseReadAhead;
 368
 369         raw_offset = base->raw_offset + ri * HAMMER_BUFSIZE;
 370         vol = base->volume;
 371
 372         while (ri < re) {
 373                 if (raw_offset >= vol->ondisk->vol_buf_end)
 374                         break;
 375                 if (raw_offset < vol->ondisk->vol_buf_beg || ri == 0) {
 376                         ++ri;
 377                         raw_offset += HAMMER_BUFSIZE;
 378                         continue;
 379                 }
 380                 buf_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no,
 381                         raw_offset - vol->ondisk->vol_buf_beg);
 382                 buf = find_buffer(vol, buf_offset);
 383                 if (buf == NULL) {
 384                         buf = get_buffer(buf_offset, -1);
 385                         rel_buffer(buf);
 386                 }
 387                 ++ri;
 388                 raw_offset += HAMMER_BUFSIZE;
 389         }
 390 }
 391
 392 void
 393 rel_buffer(struct buffer_info *buffer)
 394 {
 395         struct volume_info *volume;
 396         int hi;
 397
 398         if (buffer == NULL)
 399                 return;
 400         assert(buffer->cache.refs > 0);
 401         if (--buffer->cache.refs == 0) {
 402                 if (buffer->cache.delete) {
 403                         hi = buffer_hash(buffer->buf_offset);
 404                         volume = buffer->volume;
 405                         if (buffer->cache.modified)
 406                                 flush_buffer(buffer);
 407                         TAILQ_REMOVE(&volume->buffer_lists[hi], buffer, entry);
 408                         hammer_cache_del(&buffer->cache);
 409                         free(buffer->ondisk);
 410                         free(buffer);
 411                         rel_volume(volume);
 412                 }
 413         }
 414 }
 415
 416 /*
 417  * Retrieve a pointer to a buffer data given a buffer offset.  The underlying
 418  * bufferp is freed if isnew or the offset is out of range of the cached data.
 419  * If bufferp is freed a referenced buffer is loaded into it.
 420  */
 421 void *
 422 get_buffer_data(hammer_off_t buf_offset, struct buffer_info **bufferp,
 423                 int isnew)
 424 {
 425         if (*bufferp != NULL) {
 426                 if (isnew > 0 ||
 427                     (((*bufferp)->buf_offset ^ buf_offset) & ~HAMMER_BUFMASK64)) {
 428                         rel_buffer(*bufferp);
 429                         *bufferp = NULL;
 430                 }
 431         }
 432         return(get_ondisk(buf_offset, bufferp, isnew));
 433 }
 434
 435 /*
 436  * Retrieve a pointer to a B-Tree node given a zone offset.  The underlying
 437  * bufferp is freed if non-NULL and a referenced buffer is loaded into it.
 438  */
 439 hammer_node_ondisk_t
 440 get_node(hammer_off_t node_offset, struct buffer_info **bufferp)
 441 {
 442         if (*bufferp != NULL) {
 443                 rel_buffer(*bufferp);
 444                 *bufferp = NULL;
 445         }
 446         return(get_ondisk(node_offset, bufferp, 0));
 447 }
 448
 449 /*
 450  * Return a pointer to a buffer data given a buffer offset.
 451  * If *bufferp is NULL acquire the buffer otherwise use that buffer.
 452  */
 453 static __inline
 454 void *
 455 get_ondisk(hammer_off_t buf_offset, struct buffer_info **bufferp, int isnew)
 456 {
 457         struct buffer_info *buffer;
 458
 459         buffer = *bufferp;
 460         if (buffer == NULL) {
 461                 buffer = *bufferp = get_buffer(buf_offset, isnew);
 462                 if (buffer == NULL)
 463                         return(NULL);
 464         }
 465
 466         return((char *)buffer->ondisk +
 467                 ((int32_t)buf_offset & HAMMER_BUFMASK));
 468 }
 469
 470 /*
 471  * Allocate HAMMER elements - btree nodes, meta data, data storage
 472  */
 473 void *
 474 alloc_btree_element(hammer_off_t *offp, struct buffer_info **data_bufferp)
 475 {
 476         hammer_node_ondisk_t node;
 477
 478         node = alloc_blockmap(HAMMER_ZONE_BTREE_INDEX, sizeof(*node),
 479                               offp, data_bufferp);
 480         bzero(node, sizeof(*node));
 481         return (node);
 482 }
 483
 484 void *
 485 alloc_meta_element(hammer_off_t *offp, int32_t data_len,
 486                    struct buffer_info **data_bufferp)
 487 {
 488         void *data;
 489
 490         data = alloc_blockmap(HAMMER_ZONE_META_INDEX, data_len,
 491                               offp, data_bufferp);
 492         bzero(data, data_len);
 493         return (data);
 494 }
 495
 496 /*
 497  * The only data_len supported by HAMMER userspace for large data zone
 498  * (zone 10) is HAMMER_BUFSIZE which is 16KB.  >16KB data does not fit
 499  * in a buffer allocated by get_buffer().  Also alloc_blockmap() does
 500  * not consider >16KB buffer size.
 501  */
 502 void *
 503 alloc_data_element(hammer_off_t *offp, int32_t data_len,
 504                    struct buffer_info **data_bufferp)
 505 {
 506         void *data;
 507         int zone;
 508
 509         if (data_len == 0)
 510                 return(NULL);
 511
 512         zone = hammer_data_zone_index(data_len);
 513         assert(data_len <= HAMMER_BUFSIZE); /* just one buffer */
 514         assert(zone == HAMMER_ZONE_LARGE_DATA_INDEX ||
 515                zone == HAMMER_ZONE_SMALL_DATA_INDEX);
 516
 517         data = alloc_blockmap(zone, data_len, offp, data_bufferp);
 518         bzero(data, data_len);
 519         return(data);
 520 }
 521
 522 /*
 523  * Format a new blockmap.  This is mostly a degenerate case because
 524  * all allocations are now actually done from the freemap.
 525  */
 526 void
 527 format_blockmap(struct volume_info *root_vol, int zone, hammer_off_t offset)
 528 {
 529         hammer_blockmap_t blockmap;
 530         hammer_off_t zone_base;
 531
 532         /* Only root volume needs formatting */
 533         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 534
 535         assert(hammer_is_zone2_mapped_index(zone));
 536
 537         blockmap = &root_vol->ondisk->vol0_blockmap[zone];
 538         zone_base = HAMMER_ZONE_ENCODE(zone, offset);
 539
 540         bzero(blockmap, sizeof(*blockmap));
 541         blockmap->phys_offset = 0;
 542         blockmap->first_offset = zone_base;
 543         blockmap->next_offset = zone_base;
 544         blockmap->alloc_offset = HAMMER_ENCODE(zone, 255, -1);
 545         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 546         root_vol->cache.modified = 1;
 547 }
 548
 549 /*
 550  * Format a new freemap.  Set all layer1 entries to UNAVAIL.  The initialize
 551  * code will load each volume's freemap.
 552  */
 553 void
 554 format_freemap(struct volume_info *root_vol)
 555 {
 556         struct buffer_info *buffer = NULL;
 557         hammer_off_t layer1_offset;
 558         hammer_blockmap_t blockmap;
 559         struct hammer_blockmap_layer1 *layer1;
 560         int i, isnew;
 561
 562         /* Only root volume needs formatting */
 563         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 564
 565         layer1_offset = alloc_bigblock(root_vol, HAMMER_ZONE_FREEMAP_INDEX);
 566         for (i = 0; i < HAMMER_BIGBLOCK_SIZE; i += sizeof(*layer1)) {
 567                 isnew = ((i % HAMMER_BUFSIZE) == 0);
 568                 layer1 = get_buffer_data(layer1_offset + i, &buffer, isnew);
 569                 bzero(layer1, sizeof(*layer1));
 570                 layer1->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 571                 layer1->blocks_free = 0;
 572                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 573         }
 574         assert(i == HAMMER_BIGBLOCK_SIZE);
 575         rel_buffer(buffer);
 576
 577         blockmap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 578         bzero(blockmap, sizeof(*blockmap));
 579         blockmap->phys_offset = layer1_offset;
 580         blockmap->first_offset = 0;
 581         blockmap->next_offset = HAMMER_ENCODE_RAW_BUFFER(0, 0);
 582         blockmap->alloc_offset = HAMMER_ENCODE_RAW_BUFFER(255, -1);
 583         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 584         root_vol->cache.modified = 1;
 585 }
 586
 587 /*
 588  * Load the volume's remaining free space into the freemap.
 589  *
 590  * Returns the number of big-blocks available.
 591  */
 592 int64_t
 593 initialize_freemap(struct volume_info *vol)
 594 {
 595         struct volume_info *root_vol;
 596         struct buffer_info *buffer1 = NULL;
 597         struct buffer_info *buffer2 = NULL;
 598         struct hammer_blockmap_layer1 *layer1;
 599         struct hammer_blockmap_layer2 *layer2;
 600         hammer_off_t layer1_base;
 601         hammer_off_t layer1_offset;
 602         hammer_off_t layer2_offset;
 603         hammer_off_t phys_offset;
 604         hammer_off_t block_offset;
 605         hammer_off_t aligned_vol_free_end;
 606         hammer_blockmap_t freemap;
 607         int64_t count = 0;
 608         int64_t layer1_count = 0;
 609
 610         root_vol = get_root_volume();
 611         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 612                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 613
 614         printf("initialize freemap volume %d\n", vol->vol_no);
 615
 616         /*
 617          * Initialize the freemap.  First preallocate the big-blocks required
 618          * to implement layer2.   This preallocation is a bootstrap allocation
 619          * using blocks from the target volume.
 620          */
 621         freemap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 622         layer1_base = freemap->phys_offset;
 623
 624         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 625              phys_offset < aligned_vol_free_end;
 626              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 627                 layer1_offset = layer1_base +
 628                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 629                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 630                 if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 631                         layer1->phys_offset = alloc_bigblock(vol,
 632                                                 HAMMER_ZONE_FREEMAP_INDEX);
 633                         layer1->blocks_free = 0;
 634                         buffer1->cache.modified = 1;
 635                         layer1->layer1_crc = crc32(layer1,
 636                                                    HAMMER_LAYER1_CRCSIZE);
 637                 }
 638         }
 639
 640         /*
 641          * Now fill everything in.
 642          */
 643         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 644              phys_offset < aligned_vol_free_end;
 645              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 646                 layer1_count = 0;
 647                 layer1_offset = layer1_base +
 648                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 649                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 650                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 651
 652                 for (block_offset = 0;
 653                      block_offset < HAMMER_BLOCKMAP_LAYER2;
 654                      block_offset += HAMMER_BIGBLOCK_SIZE) {
 655                         layer2_offset = layer1->phys_offset +
 656                                         HAMMER_BLOCKMAP_LAYER2_OFFSET(block_offset);
 657                         layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 658                         bzero(layer2, sizeof(*layer2));
 659
 660                         if (phys_offset + block_offset < vol->vol_free_off) {
 661                                 /*
 662                                  * Fixups XXX - big-blocks already allocated as part
 663                                  * of the freemap bootstrap.
 664                                  */
 665                                 layer2->zone = HAMMER_ZONE_FREEMAP_INDEX;
 666                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 667                                 layer2->bytes_free = 0;
 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         if (vol->rdonly)
 893                 return(0);
 894
 895         n = pwrite(vol->fd, data, HAMMER_BUFSIZE, offset);
 896         if (n != HAMMER_BUFSIZE)
 897                 return(-1);
 898         return(0);
 899 }
 900
 901 int64_t init_boot_area_size(int64_t value, off_t avg_vol_size)
 902 {
 903         if (value == 0) {
 904                 value = HAMMER_BOOT_NOMBYTES;
 905                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 906                         value >>= 1;
 907                 if (value < HAMMER_BOOT_MINBYTES)
 908                         value = 0;
 909         } else if (value < HAMMER_BOOT_MINBYTES) {
 910                 value = HAMMER_BOOT_MINBYTES;
 911         }
 912
 913         return(value);
 914 }
 915
 916 int64_t init_mem_area_size(int64_t value, off_t avg_vol_size)
 917 {
 918         if (value == 0) {
 919                 value = HAMMER_MEM_NOMBYTES;
 920                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 921                         value >>= 1;
 922                 if (value < HAMMER_MEM_MINBYTES)
 923                         value = 0;
 924         } else if (value < HAMMER_MEM_MINBYTES) {
 925                 value = HAMMER_MEM_MINBYTES;
 926         }
 927
 928         return(value);
 929 }