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 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
 250         return(vol);
 251 }
 252
 253 struct volume_info *
 254 get_root_volume(void)
 255 {
 256         struct volume_info *root_vol;
 257
 258         root_vol = get_volume(HAMMER_ROOT_VOLNO);
 259         assert(root_vol != NULL);
 260
 261         return(root_vol);
 262 }
 263
 264 /*
 265  * Acquire the specified buffer.  isnew is -1 only when called
 266  * via get_buffer_readahead() to prevent another readahead.
 267  */
 268 static struct buffer_info *
 269 get_buffer(hammer_off_t buf_offset, int isnew)
 270 {
 271         struct buffer_info *buf;
 272         struct volume_info *volume;
 273         int vol_no;
 274         int zone;
 275         int hi;
 276         int dora = 0;
 277         int error = 0;
 278
 279         zone = HAMMER_ZONE_DECODE(buf_offset);
 280         if (zone > HAMMER_ZONE_RAW_BUFFER_INDEX)
 281                 buf_offset = blockmap_lookup(buf_offset, NULL, NULL, &error);
 282         if (error || buf_offset == HAMMER_OFF_BAD)
 283                 return(NULL);
 284         assert(hammer_is_zone_raw_buffer(buf_offset));
 285
 286         vol_no = HAMMER_VOL_DECODE(buf_offset);
 287         volume = get_volume(vol_no);
 288         assert(volume != NULL);
 289
 290         buf_offset &= ~HAMMER_BUFMASK64;
 291         buf = find_buffer(volume, buf_offset);
 292
 293         if (buf == NULL) {
 294                 buf = malloc(sizeof(*buf));
 295                 bzero(buf, sizeof(*buf));
 296                 buf->buf_offset = buf_offset;
 297                 buf->raw_offset = hammer_xlate_to_phys(volume->ondisk,
 298                                                         buf_offset);
 299                 buf->volume = volume;
 300                 buf->ondisk = malloc(HAMMER_BUFSIZE);
 301                 if (isnew <= 0) {
 302                         if (readhammerbuf(buf) == -1) {
 303                                 err(1, "get_buffer: %s:%016jx "
 304                                     "Read failed at offset %016jx",
 305                                     volume->name,
 306                                     (intmax_t)buf->buf_offset,
 307                                     (intmax_t)buf->raw_offset);
 308                         }
 309                 }
 310
 311                 hi = buffer_hash(buf_offset);
 312                 TAILQ_INSERT_TAIL(&volume->buffer_lists[hi], buf, entry);
 313                 hammer_cache_add(&buf->cache);
 314                 dora = (isnew == 0);
 315         } else {
 316                 assert(buf->ondisk != NULL);
 317                 assert(isnew != -1);
 318                 hammer_cache_used(&buf->cache);
 319         }
 320
 321         ++buf->cache.refs;
 322         hammer_cache_flush();
 323
 324         if (isnew > 0) {
 325                 assert(buf->cache.modified == 0);
 326                 bzero(buf->ondisk, HAMMER_BUFSIZE);
 327                 buf->cache.modified = 1;
 328         }
 329         if (dora)
 330                 get_buffer_readahead(buf);
 331         return(buf);
 332 }
 333
 334 static void
 335 get_buffer_readahead(struct buffer_info *base)
 336 {
 337         struct buffer_info *buf;
 338         struct volume_info *vol;
 339         hammer_off_t buf_offset;
 340         int64_t raw_offset;
 341         int ri = UseReadBehind;
 342         int re = UseReadAhead;
 343
 344         raw_offset = base->raw_offset + ri * HAMMER_BUFSIZE;
 345         vol = base->volume;
 346
 347         while (ri < re) {
 348                 if (raw_offset >= vol->ondisk->vol_buf_end)
 349                         break;
 350                 if (raw_offset < vol->ondisk->vol_buf_beg || ri == 0) {
 351                         ++ri;
 352                         raw_offset += HAMMER_BUFSIZE;
 353                         continue;
 354                 }
 355                 buf_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no,
 356                         raw_offset - vol->ondisk->vol_buf_beg);
 357                 buf = find_buffer(vol, buf_offset);
 358                 if (buf == NULL) {
 359                         buf = get_buffer(buf_offset, -1);
 360                         rel_buffer(buf);
 361                 }
 362                 ++ri;
 363                 raw_offset += HAMMER_BUFSIZE;
 364         }
 365 }
 366
 367 void
 368 rel_buffer(struct buffer_info *buffer)
 369 {
 370         struct volume_info *volume;
 371         int hi;
 372
 373         if (buffer == NULL)
 374                 return;
 375         assert(buffer->cache.refs > 0);
 376         if (--buffer->cache.refs == 0) {
 377                 if (buffer->cache.delete) {
 378                         hi = buffer_hash(buffer->buf_offset);
 379                         volume = buffer->volume;
 380                         if (buffer->cache.modified)
 381                                 flush_buffer(buffer);
 382                         TAILQ_REMOVE(&volume->buffer_lists[hi], buffer, entry);
 383                         hammer_cache_del(&buffer->cache);
 384                         free(buffer->ondisk);
 385                         free(buffer);
 386                 }
 387         }
 388 }
 389
 390 /*
 391  * Retrieve a pointer to a buffer data given a buffer offset.  The underlying
 392  * bufferp is freed if isnew or the offset is out of range of the cached data.
 393  * If bufferp is freed a referenced buffer is loaded into it.
 394  */
 395 void *
 396 get_buffer_data(hammer_off_t buf_offset, struct buffer_info **bufferp,
 397                 int isnew)
 398 {
 399         if (*bufferp != NULL) {
 400                 if (isnew > 0 ||
 401                     (((*bufferp)->buf_offset ^ buf_offset) & ~HAMMER_BUFMASK64)) {
 402                         rel_buffer(*bufferp);
 403                         *bufferp = NULL;
 404                 }
 405         }
 406         return(get_ondisk(buf_offset, bufferp, isnew));
 407 }
 408
 409 /*
 410  * Retrieve a pointer to a B-Tree node given a zone offset.  The underlying
 411  * bufferp is freed if non-NULL and a referenced buffer is loaded into it.
 412  */
 413 hammer_node_ondisk_t
 414 get_node(hammer_off_t node_offset, struct buffer_info **bufferp)
 415 {
 416         if (*bufferp != NULL) {
 417                 rel_buffer(*bufferp);
 418                 *bufferp = NULL;
 419         }
 420         return(get_ondisk(node_offset, bufferp, 0));
 421 }
 422
 423 /*
 424  * Return a pointer to a buffer data given a buffer offset.
 425  * If *bufferp is NULL acquire the buffer otherwise use that buffer.
 426  */
 427 static void *
 428 get_ondisk(hammer_off_t buf_offset, struct buffer_info **bufferp, int isnew)
 429 {
 430         if (*bufferp == NULL) {
 431                 *bufferp = get_buffer(buf_offset, isnew);
 432                 if (*bufferp == NULL)
 433                         return(NULL);
 434         }
 435
 436         return(((char *)(*bufferp)->ondisk) +
 437                 ((int32_t)buf_offset & HAMMER_BUFMASK));
 438 }
 439
 440 /*
 441  * Allocate HAMMER elements - B-Tree nodes
 442  */
 443 void *
 444 alloc_btree_element(hammer_off_t *offp, struct buffer_info **data_bufferp)
 445 {
 446         hammer_node_ondisk_t node;
 447
 448         node = alloc_blockmap(HAMMER_ZONE_BTREE_INDEX, sizeof(*node),
 449                               offp, data_bufferp);
 450         bzero(node, sizeof(*node));
 451         return (node);
 452 }
 453
 454 /*
 455  * Allocate HAMMER elements - meta data (inode, direntry, PFS, etc)
 456  */
 457 void *
 458 alloc_meta_element(hammer_off_t *offp, int32_t data_len,
 459                    struct buffer_info **data_bufferp)
 460 {
 461         void *data;
 462
 463         data = alloc_blockmap(HAMMER_ZONE_META_INDEX, data_len,
 464                               offp, data_bufferp);
 465         bzero(data, data_len);
 466         return (data);
 467 }
 468
 469 /*
 470  * Allocate HAMMER elements - data storage
 471  *
 472  * The only data_len supported by HAMMER userspace for large data zone
 473  * (zone 10) is HAMMER_BUFSIZE which is 16KB.  >16KB data does not fit
 474  * in a buffer allocated by get_buffer().  Also alloc_blockmap() does
 475  * not consider >16KB buffer size.
 476  */
 477 void *
 478 alloc_data_element(hammer_off_t *offp, int32_t data_len,
 479                    struct buffer_info **data_bufferp)
 480 {
 481         void *data;
 482         int zone;
 483
 484         if (data_len == 0)
 485                 return(NULL);
 486
 487         zone = hammer_data_zone_index(data_len);
 488         assert(data_len <= HAMMER_BUFSIZE); /* just one buffer */
 489         assert(zone == HAMMER_ZONE_LARGE_DATA_INDEX ||
 490                zone == HAMMER_ZONE_SMALL_DATA_INDEX);
 491
 492         data = alloc_blockmap(zone, data_len, offp, data_bufferp);
 493         bzero(data, data_len);
 494         return(data);
 495 }
 496
 497 /*
 498  * Format a new blockmap.  This is mostly a degenerate case because
 499  * all allocations are now actually done from the freemap.
 500  */
 501 void
 502 format_blockmap(struct volume_info *root_vol, int zone, hammer_off_t offset)
 503 {
 504         hammer_blockmap_t blockmap;
 505         hammer_off_t zone_base;
 506
 507         /* Only root volume needs formatting */
 508         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 509
 510         assert(hammer_is_zone2_mapped_index(zone));
 511
 512         blockmap = &root_vol->ondisk->vol0_blockmap[zone];
 513         zone_base = HAMMER_ZONE_ENCODE(zone, offset);
 514
 515         bzero(blockmap, sizeof(*blockmap));
 516         blockmap->phys_offset = 0;
 517         blockmap->first_offset = zone_base;
 518         blockmap->next_offset = zone_base;
 519         blockmap->alloc_offset = HAMMER_ENCODE(zone, 255, -1);
 520         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 521 }
 522
 523 /*
 524  * Format a new freemap.  Set all layer1 entries to UNAVAIL.  The initialize
 525  * code will load each volume's freemap.
 526  */
 527 void
 528 format_freemap(struct volume_info *root_vol)
 529 {
 530         struct buffer_info *buffer = NULL;
 531         hammer_off_t layer1_offset;
 532         hammer_blockmap_t blockmap;
 533         hammer_blockmap_layer1_t layer1;
 534         int i, isnew;
 535
 536         /* Only root volume needs formatting */
 537         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 538
 539         layer1_offset = alloc_bigblock(root_vol, HAMMER_ZONE_FREEMAP_INDEX);
 540         for (i = 0; i < HAMMER_BIGBLOCK_SIZE; i += sizeof(*layer1)) {
 541                 isnew = ((i % HAMMER_BUFSIZE) == 0);
 542                 layer1 = get_buffer_data(layer1_offset + i, &buffer, isnew);
 543                 bzero(layer1, sizeof(*layer1));
 544                 layer1->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 545                 layer1->blocks_free = 0;
 546                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 547         }
 548         assert(i == HAMMER_BIGBLOCK_SIZE);
 549         rel_buffer(buffer);
 550
 551         blockmap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 552         bzero(blockmap, sizeof(*blockmap));
 553         blockmap->phys_offset = layer1_offset;
 554         blockmap->first_offset = 0;
 555         blockmap->next_offset = HAMMER_ENCODE_RAW_BUFFER(0, 0);
 556         blockmap->alloc_offset = HAMMER_ENCODE_RAW_BUFFER(255, -1);
 557         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 558 }
 559
 560 /*
 561  * Load the volume's remaining free space into the freemap.
 562  *
 563  * Returns the number of big-blocks available.
 564  */
 565 int64_t
 566 initialize_freemap(struct volume_info *vol)
 567 {
 568         struct volume_info *root_vol;
 569         struct buffer_info *buffer1 = NULL;
 570         struct buffer_info *buffer2 = NULL;
 571         hammer_blockmap_layer1_t layer1;
 572         hammer_blockmap_layer2_t layer2;
 573         hammer_off_t layer1_offset;
 574         hammer_off_t layer2_offset;
 575         hammer_off_t phys_offset;
 576         hammer_off_t block_offset;
 577         hammer_off_t aligned_vol_free_end;
 578         hammer_blockmap_t freemap;
 579         int64_t count = 0;
 580         int64_t layer1_count = 0;
 581
 582         root_vol = get_root_volume();
 583         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 584                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 585
 586         printf("initialize freemap volume %d\n", vol->vol_no);
 587
 588         /*
 589          * Initialize the freemap.  First preallocate the big-blocks required
 590          * to implement layer2.   This preallocation is a bootstrap allocation
 591          * using blocks from the target volume.
 592          */
 593         freemap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 594
 595         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 596              phys_offset < aligned_vol_free_end;
 597              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 598                 layer1_offset = freemap->phys_offset +
 599                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 600                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 601                 if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 602                         layer1->phys_offset = alloc_bigblock(vol,
 603                                                 HAMMER_ZONE_FREEMAP_INDEX);
 604                         layer1->blocks_free = 0;
 605                         buffer1->cache.modified = 1;
 606                         layer1->layer1_crc = crc32(layer1,
 607                                                    HAMMER_LAYER1_CRCSIZE);
 608                 }
 609         }
 610
 611         /*
 612          * Now fill everything in.
 613          */
 614         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 615              phys_offset < aligned_vol_free_end;
 616              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 617                 layer1_count = 0;
 618                 layer1_offset = freemap->phys_offset +
 619                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 620                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 621                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 622
 623                 for (block_offset = 0;
 624                      block_offset < HAMMER_BLOCKMAP_LAYER2;
 625                      block_offset += HAMMER_BIGBLOCK_SIZE) {
 626                         layer2_offset = layer1->phys_offset +
 627                                         HAMMER_BLOCKMAP_LAYER2_OFFSET(block_offset);
 628                         layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 629                         bzero(layer2, sizeof(*layer2));
 630
 631                         if (phys_offset + block_offset < vol->vol_free_off) {
 632                                 /*
 633                                  * Fixups XXX - big-blocks already allocated as part
 634                                  * of the freemap bootstrap.
 635                                  */
 636                                 layer2->zone = HAMMER_ZONE_FREEMAP_INDEX;
 637                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 638                                 layer2->bytes_free = 0;
 639                         } else if (phys_offset + block_offset < vol->vol_free_end) {
 640                                 layer2->zone = 0;
 641                                 layer2->append_off = 0;
 642                                 layer2->bytes_free = HAMMER_BIGBLOCK_SIZE;
 643                                 ++count;
 644                                 ++layer1_count;
 645                         } else {
 646                                 layer2->zone = HAMMER_ZONE_UNAVAIL_INDEX;
 647                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 648                                 layer2->bytes_free = 0;
 649                         }
 650                         layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 651                         buffer2->cache.modified = 1;
 652                 }
 653
 654                 layer1->blocks_free += layer1_count;
 655                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 656                 buffer1->cache.modified = 1;
 657         }
 658
 659         rel_buffer(buffer1);
 660         rel_buffer(buffer2);
 661         return(count);
 662 }
 663
 664 /*
 665  * Returns the number of big-blocks available for filesystem data and undos
 666  * without formatting.
 667  */
 668 int64_t
 669 count_freemap(struct volume_info *vol)
 670 {
 671         hammer_off_t phys_offset;
 672         hammer_off_t vol_free_off;
 673         hammer_off_t aligned_vol_free_end;
 674         int64_t count = 0;
 675
 676         vol_free_off = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 677         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 678                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 679
 680         if (vol->vol_no == HAMMER_ROOT_VOLNO)
 681                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 682
 683         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 684              phys_offset < aligned_vol_free_end;
 685              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 686                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 687         }
 688
 689         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 690              phys_offset < aligned_vol_free_end;
 691              phys_offset += HAMMER_BIGBLOCK_SIZE) {
 692                 if (phys_offset < vol_free_off) {
 693                         ;
 694                 } else if (phys_offset < vol->vol_free_end) {
 695                         ++count;
 696                 }
 697         }
 698
 699         return(count);
 700 }
 701
 702 /*
 703  * Format the undomap for the root volume.
 704  */
 705 void
 706 format_undomap(struct volume_info *root_vol, int64_t *undo_buffer_size)
 707 {
 708         const int undo_zone = HAMMER_ZONE_UNDO_INDEX;
 709         hammer_off_t undo_limit;
 710         hammer_blockmap_t blockmap;
 711         hammer_volume_ondisk_t ondisk;
 712         struct buffer_info *buffer = NULL;
 713         hammer_off_t scan;
 714         int n;
 715         int limit_index;
 716         uint32_t seqno;
 717
 718         /* Only root volume needs formatting */
 719         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 720         ondisk = root_vol->ondisk;
 721
 722         /*
 723          * Size the undo buffer in multiples of HAMMER_BIGBLOCK_SIZE,
 724          * up to HAMMER_UNDO_LAYER2 big-blocks.  Size to approximately
 725          * 0.1% of the disk.
 726          *
 727          * The minimum UNDO fifo size is 500MB, or approximately 1% of
 728          * the recommended 50G disk.
 729          *
 730          * Changing this minimum is rather dangerous as complex filesystem
 731          * operations can cause the UNDO FIFO to fill up otherwise.
 732          */
 733         undo_limit = *undo_buffer_size;
 734         if (undo_limit == 0) {
 735                 undo_limit = HAMMER_VOL_BUF_SIZE(ondisk) / 1000;
 736                 if (undo_limit < 500*1024*1024)
 737                         undo_limit = 500*1024*1024;
 738         }
 739         undo_limit = (undo_limit + HAMMER_BIGBLOCK_MASK64) &
 740                      ~HAMMER_BIGBLOCK_MASK64;
 741         if (undo_limit < HAMMER_BIGBLOCK_SIZE)
 742                 undo_limit = HAMMER_BIGBLOCK_SIZE;
 743         if (undo_limit > HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2)
 744                 undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2;
 745         *undo_buffer_size = undo_limit;
 746
 747         blockmap = &ondisk->vol0_blockmap[undo_zone];
 748         bzero(blockmap, sizeof(*blockmap));
 749         blockmap->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 750         blockmap->first_offset = HAMMER_ZONE_ENCODE(undo_zone, 0);
 751         blockmap->next_offset = blockmap->first_offset;
 752         blockmap->alloc_offset = HAMMER_ZONE_ENCODE(undo_zone, undo_limit);
 753         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 754
 755         limit_index = undo_limit / HAMMER_BIGBLOCK_SIZE;
 756         assert(limit_index <= HAMMER_UNDO_LAYER2);
 757
 758         for (n = 0; n < limit_index; ++n) {
 759                 ondisk->vol0_undo_array[n] = alloc_bigblock(root_vol,
 760                                                         HAMMER_ZONE_UNDO_INDEX);
 761         }
 762         while (n < HAMMER_UNDO_LAYER2) {
 763                 ondisk->vol0_undo_array[n++] = HAMMER_BLOCKMAP_UNAVAIL;
 764         }
 765
 766         /*
 767          * Pre-initialize the UNDO blocks (HAMMER version 4+)
 768          */
 769         printf("initializing the undo map (%jd MB)\n",
 770                 (intmax_t)(blockmap->alloc_offset & HAMMER_OFF_LONG_MASK) /
 771                 (1024 * 1024));
 772
 773         scan = blockmap->first_offset;
 774         seqno = 0;
 775
 776         while (scan < blockmap->alloc_offset) {
 777                 hammer_fifo_head_t head;
 778                 hammer_fifo_tail_t tail;
 779                 int isnew;
 780                 int bytes = HAMMER_UNDO_ALIGN;
 781
 782                 isnew = ((scan & HAMMER_BUFMASK64) == 0);
 783                 head = get_buffer_data(scan, &buffer, isnew);
 784                 buffer->cache.modified = 1;
 785                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 786
 787                 bzero(head, bytes);
 788                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 789                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
 790                 head->hdr_size = bytes;
 791                 head->hdr_seq = seqno++;
 792
 793                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 794                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 795                 tail->tail_size = bytes;
 796
 797                 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
 798                                 crc32(head + 1, bytes - sizeof(*head));
 799
 800                 scan += bytes;
 801         }
 802         rel_buffer(buffer);
 803 }
 804
 805 const char *zone_labels[] = {
 806         "",             /* 0 */
 807         "raw_volume",   /* 1 */
 808         "raw_buffer",   /* 2 */
 809         "undo",         /* 3 */
 810         "freemap",      /* 4 */
 811         "",             /* 5 */
 812         "",             /* 6 */
 813         "",             /* 7 */
 814         "btree",        /* 8 */
 815         "meta",         /* 9 */
 816         "large_data",   /* 10 */
 817         "small_data",   /* 11 */
 818         "",             /* 12 */
 819         "",             /* 13 */
 820         "",             /* 14 */
 821         "unavail",      /* 15 */
 822 };
 823
 824 void
 825 print_blockmap(const struct volume_info *root_vol)
 826 {
 827         hammer_blockmap_t blockmap;
 828         hammer_volume_ondisk_t ondisk;
 829         int64_t size, used;
 830         int i;
 831 #define INDENT ""
 832
 833         ondisk = root_vol->ondisk;
 834         printf(INDENT"vol_label\t%s\n", ondisk->vol_label);
 835         printf(INDENT"vol_count\t%d\n", ondisk->vol_count);
 836         printf(INDENT"vol_bot_beg\t%s\n", sizetostr(ondisk->vol_bot_beg));
 837         printf(INDENT"vol_mem_beg\t%s\n", sizetostr(ondisk->vol_mem_beg));
 838         printf(INDENT"vol_buf_beg\t%s\n", sizetostr(ondisk->vol_buf_beg));
 839         printf(INDENT"vol_buf_end\t%s\n", sizetostr(ondisk->vol_buf_end));
 840         printf(INDENT"vol0_next_tid\t%016jx\n",
 841                (uintmax_t)ondisk->vol0_next_tid);
 842
 843         blockmap = &ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 844         size = blockmap->alloc_offset & HAMMER_OFF_LONG_MASK;
 845         if (blockmap->first_offset <= blockmap->next_offset)
 846                 used = blockmap->next_offset - blockmap->first_offset;
 847         else
 848                 used = blockmap->alloc_offset - blockmap->first_offset +
 849                         (blockmap->next_offset & HAMMER_OFF_LONG_MASK);
 850         printf(INDENT"undo_size\t%s\n", sizetostr(size));
 851         printf(INDENT"undo_used\t%s\n", sizetostr(used));
 852
 853         printf(INDENT"zone #             "
 854                "phys             first            next             alloc\n");
 855         for (i = 0; i < HAMMER_MAX_ZONES; i++) {
 856                 blockmap = &ondisk->vol0_blockmap[i];
 857                 printf(INDENT"zone %-2d %-10s %016jx %016jx %016jx %016jx\n",
 858                         i, zone_labels[i],
 859                         (uintmax_t)blockmap->phys_offset,
 860                         (uintmax_t)blockmap->first_offset,
 861                         (uintmax_t)blockmap->next_offset,
 862                         (uintmax_t)blockmap->alloc_offset);
 863         }
 864 }
 865
 866 /*
 867  * Flush various tracking structures to disk
 868  */
 869 void
 870 flush_all_volumes(void)
 871 {
 872         struct volume_info *vol;
 873
 874         TAILQ_FOREACH(vol, &VolList, entry)
 875                 flush_volume(vol);
 876 }
 877
 878 void
 879 flush_volume(struct volume_info *volume)
 880 {
 881         struct buffer_info *buffer;
 882         int i;
 883
 884         for (i = 0; i < HAMMER_BUFLISTS; ++i) {
 885                 TAILQ_FOREACH(buffer, &volume->buffer_lists[i], entry)
 886                         flush_buffer(buffer);
 887         }
 888         if (writehammervol(volume) == -1)
 889                 err(1, "Write volume %d (%s)", volume->vol_no, volume->name);
 890 }
 891
 892 void
 893 flush_buffer(struct buffer_info *buffer)
 894 {
 895         struct volume_info *vol;
 896
 897         vol = buffer->volume;
 898         if (writehammerbuf(buffer) == -1)
 899                 err(1, "Write volume %d (%s)", vol->vol_no, vol->name);
 900         buffer->cache.modified = 0;
 901 }
 902
 903 /*
 904  * Core I/O operations
 905  */
 906 static int
 907 __read(struct volume_info *vol, void *data, int64_t offset, int size)
 908 {
 909         ssize_t n;
 910
 911         n = pread(vol->fd, data, size, offset);
 912         if (n != size)
 913                 return(-1);
 914         return(0);
 915 }
 916
 917 static __inline int
 918 readhammervol(struct volume_info *vol)
 919 {
 920         return(__read(vol, vol->ondisk, 0, HAMMER_BUFSIZE));
 921 }
 922
 923 static __inline int
 924 readhammerbuf(struct buffer_info *buf)
 925 {
 926         return(__read(buf->volume, buf->ondisk, buf->raw_offset, HAMMER_BUFSIZE));
 927 }
 928
 929 static int
 930 __write(struct volume_info *vol, const void *data, int64_t offset, int size)
 931 {
 932         ssize_t n;
 933
 934         if (vol->rdonly)
 935                 return(0);
 936
 937         n = pwrite(vol->fd, data, size, offset);
 938         if (n != size)
 939                 return(-1);
 940         return(0);
 941 }
 942
 943 static __inline int
 944 writehammervol(struct volume_info *vol)
 945 {
 946         return(__write(vol, vol->ondisk, 0, HAMMER_BUFSIZE));
 947 }
 948
 949 static __inline int
 950 writehammerbuf(struct buffer_info *buf)
 951 {
 952         return(__write(buf->volume, buf->ondisk, buf->raw_offset, HAMMER_BUFSIZE));
 953 }
 954
 955 int64_t init_boot_area_size(int64_t value, off_t avg_vol_size)
 956 {
 957         if (value == 0) {
 958                 value = HAMMER_BOOT_NOMBYTES;
 959                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 960                         value >>= 1;
 961                 if (value < HAMMER_BOOT_MINBYTES)
 962                         value = 0;
 963         } else if (value < HAMMER_BOOT_MINBYTES) {
 964                 value = HAMMER_BOOT_MINBYTES;
 965         }
 966
 967         return(value);
 968 }
 969
 970 int64_t init_mem_area_size(int64_t value, off_t avg_vol_size)
 971 {
 972         if (value == 0) {
 973                 value = HAMMER_MEM_NOMBYTES;
 974                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 975                         value >>= 1;
 976                 if (value < HAMMER_MEM_MINBYTES)
 977                         value = 0;
 978         } else if (value < HAMMER_MEM_MINBYTES) {
 979                 value = HAMMER_MEM_MINBYTES;
 980         }
 981
 982         return(value);
 983 }