sbin/hammer/ondisk.c

   1 /*
   2  * Copyright (c) 2007 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  */
  34
  35 #include <sys/diskslice.h>
  36 #include <sys/diskmbr.h>
  37
  38 #include "hammer_util.h"
  39
  40 static void check_volume(struct volume_info *vol);
  41 static void get_buffer_readahead(struct buffer_info *base);
  42 static void *get_ondisk(hammer_off_t buf_offset, struct buffer_info **bufferp,
  43                         int isnew);
  44 static __inline int readhammervol(struct volume_info *vol);
  45 static __inline int readhammerbuf(struct buffer_info *buf);
  46 static __inline int writehammervol(struct volume_info *vol);
  47 static __inline int writehammerbuf(struct buffer_info *buf);
  48
  49 uuid_t Hammer_FSType;
  50 uuid_t Hammer_FSId;
  51 int UseReadBehind = -4;
  52 int UseReadAhead = 4;
  53 int DebugOpt;
  54
  55 TAILQ_HEAD(volume_list, volume_info);
  56 static struct volume_list VolList = TAILQ_HEAD_INITIALIZER(VolList);
  57 static int valid_hammer_volumes;
  58
  59 static __inline
  60 int
  61 buffer_hash(hammer_off_t buf_offset)
  62 {
  63         int hi;
  64
  65         hi = (int)(buf_offset / HAMMER_BUFSIZE) & HAMMER_BUFLISTMASK;
  66         return(hi);
  67 }
  68
  69 static struct buffer_info*
  70 find_buffer(struct volume_info *volume, hammer_off_t buf_offset)
  71 {
  72         int hi;
  73         struct buffer_info *buf;
  74
  75         hi = buffer_hash(buf_offset);
  76         TAILQ_FOREACH(buf, &volume->buffer_lists[hi], entry)
  77                 if (buf->buf_offset == buf_offset)
  78                         return(buf);
  79         return(NULL);
  80 }
  81
  82 static
  83 struct volume_info *
  84 __alloc_volume(const char *volname, int oflags)
  85 {
  86         struct volume_info *vol;
  87         int i;
  88
  89         vol = malloc(sizeof(*vol));
  90         if (vol == NULL)
  91                 err(1, "alloc_volume");
  92         bzero(vol, sizeof(*vol));
  93
  94         vol->vol_no = -1;
  95         vol->rdonly = (oflags == O_RDONLY);
  96         vol->name = strdup(volname);
  97         vol->fd = open(vol->name, oflags);
  98         if (vol->fd < 0)
  99                 err(1, "alloc_volume: Failed to open %s", vol->name);
 100         check_volume(vol);
 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
 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         hammer_volume_ondisk_t ondisk;
 165         int n;
 166
 167         vol = __alloc_volume(filename, oflags);
 168
 169         n = readhammervol(vol);
 170         if (n == -1) {
 171                 err(1, "load_volume: %s: Read failed at offset 0", vol->name);
 172         }
 173         ondisk = vol->ondisk;
 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 (valid_hammer_volumes++ == 0) {
 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 static 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 void
 241 assert_volume_offset(struct volume_info *vol)
 242 {
 243         assert(hammer_is_zone_raw_buffer(vol->vol_free_off));
 244         assert(hammer_is_zone_raw_buffer(vol->vol_free_end));
 245 }
 246
 247 struct volume_info *
 248 get_volume(int32_t vol_no)
 249 {
 250         struct volume_info *vol;
 251
 252         TAILQ_FOREACH(vol, &VolList, entry) {
 253                 if (vol->vol_no == vol_no)
 254                         break;
 255         }
 256
 257         return(vol);
 258 }
 259
 260 struct volume_info *
 261 get_root_volume(void)
 262 {
 263         struct volume_info *root_vol;
 264
 265         root_vol = get_volume(HAMMER_ROOT_VOLNO);
 266         assert(root_vol != NULL);
 267
 268         return(root_vol);
 269 }
 270
 271 /*
 272  * Acquire the specified buffer.  isnew is -1 only when called
 273  * via get_buffer_readahead() to prevent another readahead.
 274  */
 275 static struct buffer_info *
 276 get_buffer(hammer_off_t buf_offset, int isnew)
 277 {
 278         struct buffer_info *buf;
 279         struct volume_info *volume;
 280         int vol_no;
 281         int zone;
 282         int hi;
 283         int dora = 0;
 284         int error = 0;
 285
 286         zone = HAMMER_ZONE_DECODE(buf_offset);
 287         if (zone > HAMMER_ZONE_RAW_BUFFER_INDEX)
 288                 buf_offset = blockmap_lookup(buf_offset, NULL, NULL, &error);
 289         if (error || buf_offset == HAMMER_OFF_BAD)
 290                 return(NULL);
 291         assert(hammer_is_zone_raw_buffer(buf_offset));
 292
 293         vol_no = HAMMER_VOL_DECODE(buf_offset);
 294         volume = get_volume(vol_no);
 295         assert(volume != NULL);
 296
 297         buf_offset &= ~HAMMER_BUFMASK64;
 298         buf = find_buffer(volume, buf_offset);
 299
 300         if (buf == NULL) {
 301                 buf = malloc(sizeof(*buf));
 302                 bzero(buf, sizeof(*buf));
 303                 buf->buf_offset = buf_offset;
 304                 buf->raw_offset = hammer_xlate_to_phys(volume->ondisk,
 305                                                         buf_offset);
 306                 buf->volume = volume;
 307                 buf->ondisk = malloc(HAMMER_BUFSIZE);
 308                 if (isnew <= 0) {
 309                         if (readhammerbuf(buf) == -1) {
 310                                 err(1, "get_buffer: %s:%016jx "
 311                                     "Read failed at offset %016jx",
 312                                     volume->name,
 313                                     (intmax_t)buf->buf_offset,
 314                                     (intmax_t)buf->raw_offset);
 315                         }
 316                 }
 317
 318                 hi = buffer_hash(buf_offset);
 319                 TAILQ_INSERT_TAIL(&volume->buffer_lists[hi], buf, entry);
 320                 hammer_cache_add(&buf->cache);
 321                 dora = (isnew == 0);
 322         } else {
 323                 assert(buf->ondisk != NULL);
 324                 assert(isnew != -1);
 325                 hammer_cache_used(&buf->cache);
 326         }
 327
 328         ++buf->cache.refs;
 329         hammer_cache_flush();
 330
 331         if (isnew > 0) {
 332                 assert(buf->cache.modified == 0);
 333                 bzero(buf->ondisk, HAMMER_BUFSIZE);
 334                 buf->cache.modified = 1;
 335         }
 336         if (dora)
 337                 get_buffer_readahead(buf);
 338         return(buf);
 339 }
 340
 341 static void
 342 get_buffer_readahead(struct buffer_info *base)
 343 {
 344         struct buffer_info *buf;
 345         struct volume_info *vol;
 346         hammer_off_t buf_offset;
 347         int64_t raw_offset;
 348         int ri = UseReadBehind;
 349         int re = UseReadAhead;
 350
 351         raw_offset = base->raw_offset + ri * HAMMER_BUFSIZE;
 352         vol = base->volume;
 353
 354         while (ri < re) {
 355                 if (raw_offset >= vol->ondisk->vol_buf_end)
 356                         break;
 357                 if (raw_offset < vol->ondisk->vol_buf_beg || ri == 0) {
 358                         ++ri;
 359                         raw_offset += HAMMER_BUFSIZE;
 360                         continue;
 361                 }
 362                 buf_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no,
 363                         raw_offset - vol->ondisk->vol_buf_beg);
 364                 buf = find_buffer(vol, buf_offset);
 365                 if (buf == NULL) {
 366                         buf = get_buffer(buf_offset, -1);
 367                         rel_buffer(buf);
 368                 }
 369                 ++ri;
 370                 raw_offset += HAMMER_BUFSIZE;
 371         }
 372 }
 373
 374 void
 375 rel_buffer(struct buffer_info *buffer)
 376 {
 377         struct volume_info *volume;
 378         int hi;
 379
 380         if (buffer == NULL)
 381                 return;
 382         assert(buffer->cache.refs > 0);
 383         if (--buffer->cache.refs == 0) {
 384                 if (buffer->cache.delete) {
 385                         hi = buffer_hash(buffer->buf_offset);
 386                         volume = buffer->volume;
 387                         if (buffer->cache.modified)
 388                                 flush_buffer(buffer);
 389                         TAILQ_REMOVE(&volume->buffer_lists[hi], buffer, entry);
 390                         hammer_cache_del(&buffer->cache);
 391                         free(buffer->ondisk);
 392                         free(buffer);
 393                 }
 394         }
 395 }
 396
 397 /*
 398  * Retrieve a pointer to a buffer data given a buffer offset.  The underlying
 399  * bufferp is freed if isnew or the offset is out of range of the cached data.
 400  * If bufferp is freed a referenced buffer is loaded into it.
 401  */
 402 void *
 403 get_buffer_data(hammer_off_t buf_offset, struct buffer_info **bufferp,
 404                 int isnew)
 405 {
 406         if (*bufferp != NULL) {
 407                 if (isnew > 0 ||
 408                     (((*bufferp)->buf_offset ^ buf_offset) & ~HAMMER_BUFMASK64)) {
 409                         rel_buffer(*bufferp);
 410                         *bufferp = NULL;
 411                 }
 412         }
 413         return(get_ondisk(buf_offset, bufferp, isnew));
 414 }
 415
 416 /*
 417  * Retrieve a pointer to a B-Tree node given a zone offset.  The underlying
 418  * bufferp is freed if non-NULL and a referenced buffer is loaded into it.
 419  */
 420 hammer_node_ondisk_t
 421 get_node(hammer_off_t node_offset, struct buffer_info **bufferp)
 422 {
 423         if (*bufferp != NULL) {
 424                 rel_buffer(*bufferp);
 425                 *bufferp = NULL;
 426         }
 427         return(get_ondisk(node_offset, bufferp, 0));
 428 }
 429
 430 /*
 431  * Return a pointer to a buffer data given a buffer offset.
 432  * If *bufferp is NULL acquire the buffer otherwise use that buffer.
 433  */
 434 static void *
 435 get_ondisk(hammer_off_t buf_offset, struct buffer_info **bufferp, int isnew)
 436 {
 437         if (*bufferp == NULL) {
 438                 *bufferp = get_buffer(buf_offset, isnew);
 439                 if (*bufferp == NULL)
 440                         return(NULL);
 441         }
 442
 443         return(((char *)(*bufferp)->ondisk) +
 444                 ((int32_t)buf_offset & HAMMER_BUFMASK));
 445 }
 446
 447 /*
 448  * Allocate HAMMER elements - B-Tree nodes
 449  */
 450 void *
 451 alloc_btree_element(hammer_off_t *offp, struct buffer_info **data_bufferp)
 452 {
 453         hammer_node_ondisk_t node;
 454
 455         node = alloc_blockmap(HAMMER_ZONE_BTREE_INDEX, sizeof(*node),
 456                               offp, data_bufferp);
 457         bzero(node, sizeof(*node));
 458         return (node);
 459 }
 460
 461 /*
 462  * Allocate HAMMER elements - meta data (inode, direntry, PFS, etc)
 463  */
 464 void *
 465 alloc_meta_element(hammer_off_t *offp, int32_t data_len,
 466                    struct buffer_info **data_bufferp)
 467 {
 468         void *data;
 469
 470         data = alloc_blockmap(HAMMER_ZONE_META_INDEX, data_len,
 471                               offp, data_bufferp);
 472         bzero(data, data_len);
 473         return (data);
 474 }
 475
 476 /*
 477  * Allocate HAMMER elements - data storage
 478  *
 479  * The only data_len supported by HAMMER userspace for large data zone
 480  * (zone 10) is HAMMER_BUFSIZE which is 16KB.  >16KB data does not fit
 481  * in a buffer allocated by get_buffer().  Also alloc_blockmap() does
 482  * not consider >16KB buffer size.
 483  */
 484 void *
 485 alloc_data_element(hammer_off_t *offp, int32_t data_len,
 486                    struct buffer_info **data_bufferp)
 487 {
 488         void *data;
 489         int zone;
 490
 491         if (data_len == 0)
 492                 return(NULL);
 493
 494         zone = hammer_data_zone_index(data_len);
 495         assert(data_len <= HAMMER_BUFSIZE); /* just one buffer */
 496         assert(zone == HAMMER_ZONE_LARGE_DATA_INDEX ||
 497                zone == HAMMER_ZONE_SMALL_DATA_INDEX);
 498
 499         data = alloc_blockmap(zone, data_len, offp, data_bufferp);
 500         bzero(data, data_len);
 501         return(data);
 502 }
 503
 504 /*
 505  * Format a new blockmap.  This is mostly a degenerate case because
 506  * all allocations are now actually done from the freemap.
 507  */
 508 void
 509 format_blockmap(struct volume_info *root_vol, int zone, hammer_off_t offset)
 510 {
 511         hammer_blockmap_t blockmap;
 512         hammer_off_t zone_base;
 513
 514         /* Only root volume needs formatting */
 515         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 516
 517         assert(hammer_is_zone2_mapped_index(zone));
 518
 519         blockmap = &root_vol->ondisk->vol0_blockmap[zone];
 520         zone_base = HAMMER_ZONE_ENCODE(zone, offset);
 521
 522         bzero(blockmap, sizeof(*blockmap));
 523         blockmap->phys_offset = 0;
 524         blockmap->first_offset = zone_base;
 525         blockmap->next_offset = zone_base;
 526         blockmap->alloc_offset = HAMMER_ENCODE(zone, 255, -1);
 527         hammer_crc_set_blockmap(blockmap);
 528 }
 529
 530 /*
 531  * Format a new freemap.  Set all layer1 entries to UNAVAIL.  The initialize
 532  * code will load each volume's freemap.
 533  */
 534 void
 535 format_freemap(struct volume_info *root_vol)
 536 {
 537         struct buffer_info *buffer = NULL;
 538         hammer_off_t layer1_offset;
 539         hammer_blockmap_t blockmap;
 540         hammer_blockmap_layer1_t layer1;
 541         int i, isnew;
 542
 543         /* Only root volume needs formatting */
 544         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 545
 546         layer1_offset = alloc_bigblock(root_vol, HAMMER_ZONE_FREEMAP_INDEX);
 547         for (i = 0; i < HAMMER_BIGBLOCK_SIZE; i += sizeof(*layer1)) {
 548                 isnew = ((i % HAMMER_BUFSIZE) == 0);
 549                 layer1 = get_buffer_data(layer1_offset + i, &buffer, isnew);
 550                 bzero(layer1, sizeof(*layer1));
 551                 layer1->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 552                 layer1->blocks_free = 0;
 553                 hammer_crc_set_layer1(layer1);
 554         }
 555         assert(i == HAMMER_BIGBLOCK_SIZE);
 556         rel_buffer(buffer);
 557
 558         blockmap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 559         bzero(blockmap, sizeof(*blockmap));
 560         blockmap->phys_offset = layer1_offset;
 561         blockmap->first_offset = 0;
 562         blockmap->next_offset = HAMMER_ENCODE_RAW_BUFFER(0, 0);
 563         blockmap->alloc_offset = HAMMER_ENCODE_RAW_BUFFER(255, -1);
 564         hammer_crc_set_blockmap(blockmap);
 565 }
 566
 567 /*
 568  * Load the volume's remaining free space into the freemap.
 569  *
 570  * Returns the number of big-blocks available.
 571  */
 572 int64_t
 573 initialize_freemap(struct volume_info *vol)
 574 {
 575         struct volume_info *root_vol;
 576         struct buffer_info *buffer1 = NULL;
 577         struct buffer_info *buffer2 = NULL;
 578         hammer_blockmap_layer1_t layer1;
 579         hammer_blockmap_layer2_t layer2;
 580         hammer_off_t layer1_offset;
 581         hammer_off_t layer2_offset;
 582         hammer_off_t phys_offset;
 583         hammer_off_t block_offset;
 584         hammer_off_t aligned_vol_free_end;
 585         hammer_blockmap_t freemap;
 586         int64_t count = 0;
 587         int64_t layer1_count = 0;
 588
 589         root_vol = get_root_volume();
 590
 591         assert_volume_offset(vol);
 592         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 593                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 594
 595         printf("initialize freemap volume %d\n", vol->vol_no);
 596
 597         /*
 598          * Initialize the freemap.  First preallocate the big-blocks required
 599          * to implement layer2.   This preallocation is a bootstrap allocation
 600          * using blocks from the target volume.
 601          */
 602         freemap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 603
 604         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 605              phys_offset < aligned_vol_free_end;
 606              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 607                 layer1_offset = freemap->phys_offset +
 608                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 609                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 610                 if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 611                         layer1->phys_offset = alloc_bigblock(vol,
 612                                                 HAMMER_ZONE_FREEMAP_INDEX);
 613                         layer1->blocks_free = 0;
 614                         buffer1->cache.modified = 1;
 615                         hammer_crc_set_layer1(layer1);
 616                 }
 617         }
 618
 619         /*
 620          * Now fill everything in.
 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_count = 0;
 626                 layer1_offset = freemap->phys_offset +
 627                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 628                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 629                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 630
 631                 for (block_offset = 0;
 632                      block_offset < HAMMER_BLOCKMAP_LAYER2;
 633                      block_offset += HAMMER_BIGBLOCK_SIZE) {
 634                         layer2_offset = layer1->phys_offset +
 635                                         HAMMER_BLOCKMAP_LAYER2_OFFSET(block_offset);
 636                         layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 637                         bzero(layer2, sizeof(*layer2));
 638
 639                         if (phys_offset + block_offset < vol->vol_free_off) {
 640                                 /*
 641                                  * Fixups XXX - big-blocks already allocated as part
 642                                  * of the freemap bootstrap.
 643                                  */
 644                                 layer2->zone = HAMMER_ZONE_FREEMAP_INDEX;
 645                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 646                                 layer2->bytes_free = 0;
 647                         } else if (phys_offset + block_offset < vol->vol_free_end) {
 648                                 layer2->zone = 0;
 649                                 layer2->append_off = 0;
 650                                 layer2->bytes_free = HAMMER_BIGBLOCK_SIZE;
 651                                 ++count;
 652                                 ++layer1_count;
 653                         } else {
 654                                 layer2->zone = HAMMER_ZONE_UNAVAIL_INDEX;
 655                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 656                                 layer2->bytes_free = 0;
 657                         }
 658                         hammer_crc_set_layer2(layer2);
 659                         buffer2->cache.modified = 1;
 660                 }
 661
 662                 layer1->blocks_free += layer1_count;
 663                 hammer_crc_set_layer1(layer1);
 664                 buffer1->cache.modified = 1;
 665         }
 666
 667         rel_buffer(buffer1);
 668         rel_buffer(buffer2);
 669         return(count);
 670 }
 671
 672 /*
 673  * Returns the number of big-blocks available for filesystem data and undos
 674  * without formatting.
 675  */
 676 int64_t
 677 count_freemap(struct volume_info *vol)
 678 {
 679         hammer_off_t phys_offset;
 680         hammer_off_t vol_free_off;
 681         hammer_off_t aligned_vol_free_end;
 682         int64_t count = 0;
 683
 684         vol_free_off = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 685
 686         assert_volume_offset(vol);
 687         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 688                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 689
 690         if (vol->vol_no == HAMMER_ROOT_VOLNO)
 691                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 692
 693         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 694              phys_offset < aligned_vol_free_end;
 695              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 696                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 697         }
 698
 699         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 700              phys_offset < aligned_vol_free_end;
 701              phys_offset += HAMMER_BIGBLOCK_SIZE) {
 702                 if (phys_offset < vol_free_off) {
 703                         ;
 704                 } else if (phys_offset < vol->vol_free_end) {
 705                         ++count;
 706                 }
 707         }
 708
 709         return(count);
 710 }
 711
 712 /*
 713  * Format the undomap for the root volume.
 714  */
 715 void
 716 format_undomap(struct volume_info *root_vol, int64_t *undo_buffer_size)
 717 {
 718         const int undo_zone = HAMMER_ZONE_UNDO_INDEX;
 719         hammer_off_t undo_limit;
 720         hammer_blockmap_t blockmap;
 721         hammer_volume_ondisk_t ondisk;
 722         struct buffer_info *buffer = NULL;
 723         hammer_off_t scan;
 724         int n;
 725         int limit_index;
 726         uint32_t seqno;
 727
 728         /* Only root volume needs formatting */
 729         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 730         ondisk = root_vol->ondisk;
 731
 732         /*
 733          * Size the undo buffer in multiples of HAMMER_BIGBLOCK_SIZE,
 734          * up to HAMMER_UNDO_LAYER2 big-blocks.  Size to approximately
 735          * 0.1% of the disk.
 736          *
 737          * The minimum UNDO fifo size is 500MB, or approximately 1% of
 738          * the recommended 50G disk.
 739          *
 740          * Changing this minimum is rather dangerous as complex filesystem
 741          * operations can cause the UNDO FIFO to fill up otherwise.
 742          */
 743         undo_limit = *undo_buffer_size;
 744         if (undo_limit == 0) {
 745                 undo_limit = HAMMER_VOL_BUF_SIZE(ondisk) / 1000;
 746                 if (undo_limit < 500*1024*1024)
 747                         undo_limit = 500*1024*1024;
 748         }
 749         undo_limit = (undo_limit + HAMMER_BIGBLOCK_MASK64) &
 750                      ~HAMMER_BIGBLOCK_MASK64;
 751         if (undo_limit < HAMMER_BIGBLOCK_SIZE)
 752                 undo_limit = HAMMER_BIGBLOCK_SIZE;
 753         if (undo_limit > HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2)
 754                 undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2;
 755         *undo_buffer_size = undo_limit;
 756
 757         blockmap = &ondisk->vol0_blockmap[undo_zone];
 758         bzero(blockmap, sizeof(*blockmap));
 759         blockmap->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 760         blockmap->first_offset = HAMMER_ZONE_ENCODE(undo_zone, 0);
 761         blockmap->next_offset = blockmap->first_offset;
 762         blockmap->alloc_offset = HAMMER_ZONE_ENCODE(undo_zone, undo_limit);
 763         hammer_crc_set_blockmap(blockmap);
 764
 765         limit_index = undo_limit / HAMMER_BIGBLOCK_SIZE;
 766         assert(limit_index <= HAMMER_UNDO_LAYER2);
 767
 768         for (n = 0; n < limit_index; ++n) {
 769                 ondisk->vol0_undo_array[n] = alloc_bigblock(root_vol,
 770                                                         HAMMER_ZONE_UNDO_INDEX);
 771         }
 772         while (n < HAMMER_UNDO_LAYER2) {
 773                 ondisk->vol0_undo_array[n++] = HAMMER_BLOCKMAP_UNAVAIL;
 774         }
 775
 776         /*
 777          * Pre-initialize the UNDO blocks (HAMMER version 4+)
 778          */
 779         printf("initializing the undo map (%jd MB)\n",
 780                 (intmax_t)(blockmap->alloc_offset & HAMMER_OFF_LONG_MASK) /
 781                 (1024 * 1024));
 782
 783         scan = blockmap->first_offset;
 784         seqno = 0;
 785
 786         while (scan < blockmap->alloc_offset) {
 787                 hammer_fifo_head_t head;
 788                 hammer_fifo_tail_t tail;
 789                 int isnew;
 790                 int bytes = HAMMER_UNDO_ALIGN;
 791
 792                 isnew = ((scan & HAMMER_BUFMASK64) == 0);
 793                 head = get_buffer_data(scan, &buffer, isnew);
 794                 buffer->cache.modified = 1;
 795                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 796
 797                 bzero(head, bytes);
 798                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 799                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
 800                 head->hdr_size = bytes;
 801                 head->hdr_seq = seqno++;
 802
 803                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 804                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 805                 tail->tail_size = bytes;
 806
 807                 hammer_crc_set_fifo_head(head, bytes);
 808
 809                 scan += bytes;
 810         }
 811         rel_buffer(buffer);
 812 }
 813
 814 const char *zone_labels[] = {
 815         "",             /* 0 */
 816         "raw_volume",   /* 1 */
 817         "raw_buffer",   /* 2 */
 818         "undo",         /* 3 */
 819         "freemap",      /* 4 */
 820         "",             /* 5 */
 821         "",             /* 6 */
 822         "",             /* 7 */
 823         "btree",        /* 8 */
 824         "meta",         /* 9 */
 825         "large_data",   /* 10 */
 826         "small_data",   /* 11 */
 827         "",             /* 12 */
 828         "",             /* 13 */
 829         "",             /* 14 */
 830         "unavail",      /* 15 */
 831 };
 832
 833 void
 834 print_blockmap(const struct volume_info *root_vol)
 835 {
 836         hammer_blockmap_t blockmap;
 837         hammer_volume_ondisk_t ondisk;
 838         int64_t size, used;
 839         int i;
 840 #define INDENT ""
 841
 842         ondisk = root_vol->ondisk;
 843         printf(INDENT"vol_label\t%s\n", ondisk->vol_label);
 844         printf(INDENT"vol_count\t%d\n", ondisk->vol_count);
 845         printf(INDENT"vol_bot_beg\t%s\n", sizetostr(ondisk->vol_bot_beg));
 846         printf(INDENT"vol_mem_beg\t%s\n", sizetostr(ondisk->vol_mem_beg));
 847         printf(INDENT"vol_buf_beg\t%s\n", sizetostr(ondisk->vol_buf_beg));
 848         printf(INDENT"vol_buf_end\t%s\n", sizetostr(ondisk->vol_buf_end));
 849         printf(INDENT"vol0_next_tid\t%016jx\n",
 850                (uintmax_t)ondisk->vol0_next_tid);
 851
 852         blockmap = &ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 853         size = blockmap->alloc_offset & HAMMER_OFF_LONG_MASK;
 854         if (blockmap->first_offset <= blockmap->next_offset)
 855                 used = blockmap->next_offset - blockmap->first_offset;
 856         else
 857                 used = blockmap->alloc_offset - blockmap->first_offset +
 858                         (blockmap->next_offset & HAMMER_OFF_LONG_MASK);
 859         printf(INDENT"undo_size\t%s\n", sizetostr(size));
 860         printf(INDENT"undo_used\t%s\n", sizetostr(used));
 861
 862         printf(INDENT"zone #             "
 863                "phys             first            next             alloc\n");
 864         for (i = 0; i < HAMMER_MAX_ZONES; i++) {
 865                 blockmap = &ondisk->vol0_blockmap[i];
 866                 printf(INDENT"zone %-2d %-10s %016jx %016jx %016jx %016jx\n",
 867                         i, zone_labels[i],
 868                         (uintmax_t)blockmap->phys_offset,
 869                         (uintmax_t)blockmap->first_offset,
 870                         (uintmax_t)blockmap->next_offset,
 871                         (uintmax_t)blockmap->alloc_offset);
 872         }
 873 }
 874
 875 /*
 876  * Flush various tracking structures to disk
 877  */
 878 void
 879 flush_all_volumes(void)
 880 {
 881         struct volume_info *vol;
 882
 883         TAILQ_FOREACH(vol, &VolList, entry)
 884                 flush_volume(vol);
 885 }
 886
 887 void
 888 flush_volume(struct volume_info *volume)
 889 {
 890         struct buffer_info *buffer;
 891         int i;
 892
 893         for (i = 0; i < HAMMER_BUFLISTS; ++i) {
 894                 TAILQ_FOREACH(buffer, &volume->buffer_lists[i], entry)
 895                         flush_buffer(buffer);
 896         }
 897         if (writehammervol(volume) == -1)
 898                 err(1, "Write volume %d (%s)", volume->vol_no, volume->name);
 899 }
 900
 901 void
 902 flush_buffer(struct buffer_info *buffer)
 903 {
 904         struct volume_info *vol;
 905
 906         vol = buffer->volume;
 907         if (writehammerbuf(buffer) == -1)
 908                 err(1, "Write volume %d (%s)", vol->vol_no, vol->name);
 909         buffer->cache.modified = 0;
 910 }
 911
 912 /*
 913  * Core I/O operations
 914  */
 915 static int
 916 __read(struct volume_info *vol, void *data, int64_t offset, int size)
 917 {
 918         ssize_t n;
 919
 920         n = pread(vol->fd, data, size, offset);
 921         if (n != size)
 922                 return(-1);
 923         return(0);
 924 }
 925
 926 static __inline int
 927 readhammervol(struct volume_info *vol)
 928 {
 929         return(__read(vol, vol->ondisk, 0, HAMMER_BUFSIZE));
 930 }
 931
 932 static __inline int
 933 readhammerbuf(struct buffer_info *buf)
 934 {
 935         return(__read(buf->volume, buf->ondisk, buf->raw_offset, HAMMER_BUFSIZE));
 936 }
 937
 938 static int
 939 __write(struct volume_info *vol, const void *data, int64_t offset, int size)
 940 {
 941         ssize_t n;
 942
 943         if (vol->rdonly)
 944                 return(0);
 945
 946         n = pwrite(vol->fd, data, size, offset);
 947         if (n != size)
 948                 return(-1);
 949         return(0);
 950 }
 951
 952 static __inline int
 953 writehammervol(struct volume_info *vol)
 954 {
 955         return(__write(vol, vol->ondisk, 0, HAMMER_BUFSIZE));
 956 }
 957
 958 static __inline int
 959 writehammerbuf(struct buffer_info *buf)
 960 {
 961         return(__write(buf->volume, buf->ondisk, buf->raw_offset, HAMMER_BUFSIZE));
 962 }
 963
 964 int64_t init_boot_area_size(int64_t value, off_t avg_vol_size)
 965 {
 966         if (value == 0) {
 967                 value = HAMMER_BOOT_NOMBYTES;
 968                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 969                         value >>= 1;
 970                 if (value < HAMMER_BOOT_MINBYTES)
 971                         value = 0;
 972         } else if (value < HAMMER_BOOT_MINBYTES) {
 973                 value = HAMMER_BOOT_MINBYTES;
 974         }
 975
 976         return(value);
 977 }
 978
 979 int64_t init_mem_area_size(int64_t value, off_t avg_vol_size)
 980 {
 981         if (value == 0) {
 982                 value = HAMMER_MEM_NOMBYTES;
 983                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 984                         value >>= 1;
 985                 if (value < HAMMER_MEM_MINBYTES)
 986                         value = 0;
 987         } else if (value < HAMMER_MEM_MINBYTES) {
 988                 value = HAMMER_MEM_MINBYTES;
 989         }
 990
 991         return(value);
 992 }