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 int64_t UndoBufferSize;
  52 int     UseReadBehind = -4;
  53 int     UseReadAhead = 4;
  54 int     AssertOnFailure = 1;
  55 struct volume_list VolList = TAILQ_HEAD_INITIALIZER(VolList);
  56
  57 static __inline
  58 int
  59 buffer_hash(hammer_off_t buf_offset)
  60 {
  61         int hi;
  62
  63         hi = (int)(buf_offset / HAMMER_BUFSIZE) & HAMMER_BUFLISTMASK;
  64         return(hi);
  65 }
  66
  67 static struct buffer_info*
  68 find_buffer(struct volume_info *volume, hammer_off_t buf_offset)
  69 {
  70         int hi;
  71         struct buffer_info *buf;
  72
  73         hi = buffer_hash(buf_offset);
  74         TAILQ_FOREACH(buf, &volume->buffer_lists[hi], entry)
  75                 if (buf->buf_offset == buf_offset)
  76                         return(buf);
  77         return(NULL);
  78 }
  79
  80 static
  81 struct volume_info *
  82 __alloc_volume(const char *volname, int oflags)
  83 {
  84         struct volume_info *vol;
  85         int i;
  86
  87         vol = malloc(sizeof(*vol));
  88         if (vol == NULL)
  89                 err(1, "alloc_volume");
  90         bzero(vol, sizeof(*vol));
  91
  92         vol->vol_no = -1;
  93         vol->name = strdup(volname);
  94         vol->fd = open(vol->name, oflags);
  95         if (vol->fd < 0)
  96                 err(1, "alloc_volume: Failed to open %s", vol->name);
  97
  98         vol->size = 0;
  99         vol->device_offset = 0;
 100         vol->type = NULL;
 101
 102         vol->ondisk = malloc(HAMMER_BUFSIZE);
 103         if (vol->ondisk == NULL)
 104                 err(1, "alloc_volume");
 105         bzero(vol->ondisk, HAMMER_BUFSIZE);
 106
 107         for (i = 0; i < HAMMER_BUFLISTS; ++i)
 108                 TAILQ_INIT(&vol->buffer_lists[i]);
 109
 110         return(vol);
 111 }
 112
 113 static void
 114 __add_volume(struct volume_info *vol)
 115 {
 116         struct volume_info *scan;
 117         struct stat st1, st2;
 118
 119         if (fstat(vol->fd, &st1) != 0)
 120                 errx(1, "add_volume: %s: Failed to stat", vol->name);
 121
 122         TAILQ_FOREACH(scan, &VolList, entry) {
 123                 if (scan->vol_no == vol->vol_no) {
 124                         errx(1, "add_volume: %s: Duplicate volume number %d "
 125                                 "against %s",
 126                                 vol->name, vol->vol_no, scan->name);
 127                 }
 128                 if (fstat(scan->fd, &st2) != 0) {
 129                         errx(1, "add_volume: %s: Failed to stat %s",
 130                                 vol->name, scan->name);
 131                 }
 132                 if ((st1.st_ino == st2.st_ino) && (st1.st_dev == st2.st_dev)) {
 133                         errx(1, "add_volume: %s: Specified more than once",
 134                                 vol->name);
 135                 }
 136         }
 137
 138         TAILQ_INSERT_TAIL(&VolList, vol, entry);
 139 }
 140
 141 /*
 142  * Initialize a volume structure and ondisk vol_no field.
 143  */
 144 struct volume_info *
 145 init_volume(int32_t vol_no, const char *filename, int oflags)
 146 {
 147         struct volume_info *vol;
 148
 149         vol = __alloc_volume(filename, oflags);
 150         vol->vol_no = vol->ondisk->vol_no = vol_no;
 151         vol->cache.modified = 1;
 152
 153         __add_volume(vol);
 154
 155         return(vol);
 156 }
 157
 158 /*
 159  * Initialize a volume structure and read ondisk volume header.
 160  */
 161 struct volume_info*
 162 load_volume(const char *filename, int oflags)
 163 {
 164         struct volume_info *vol;
 165         struct hammer_volume_ondisk *ondisk;
 166         int n;
 167
 168         vol = __alloc_volume(filename, oflags);
 169         ondisk = vol->ondisk;
 170
 171         n = readhammerbuf(vol, ondisk, 0);
 172         if (n == -1) {
 173                 err(1, "load_volume: %s: Read failed at offset 0", vol->name);
 174         }
 175         vol->vol_no = ondisk->vol_no;
 176
 177         if (ondisk->vol_rootvol != HAMMER_ROOT_VOLNO) {
 178                 errx(1, "load_volume: Invalid root volume# %d",
 179                         ondisk->vol_rootvol);
 180         }
 181
 182         if (bcmp(&Hammer_FSType, &ondisk->vol_fstype, sizeof(Hammer_FSType))) {
 183                 errx(1, "load_volume: %s: Header does not indicate "
 184                         "that this is a hammer volume", vol->name);
 185         }
 186
 187         if (TAILQ_EMPTY(&VolList)) {
 188                 Hammer_FSId = ondisk->vol_fsid;
 189         } else if (bcmp(&Hammer_FSId, &ondisk->vol_fsid, sizeof(Hammer_FSId))) {
 190                 errx(1, "load_volume: %s: FSId does match other volumes!",
 191                         vol->name);
 192         }
 193
 194         __add_volume(vol);
 195
 196         return(vol);
 197 }
 198
 199 /*
 200  * Check basic volume characteristics.
 201  */
 202 void
 203 check_volume(struct volume_info *vol)
 204 {
 205         struct partinfo pinfo;
 206         struct stat st;
 207
 208         /*
 209          * Get basic information about the volume
 210          */
 211         if (ioctl(vol->fd, DIOCGPART, &pinfo) < 0) {
 212                 /*
 213                  * Allow the formatting of regular files as HAMMER volumes
 214                  */
 215                 if (fstat(vol->fd, &st) < 0)
 216                         err(1, "Unable to stat %s", vol->name);
 217                 vol->size = st.st_size;
 218                 vol->type = "REGFILE";
 219         } else {
 220                 /*
 221                  * When formatting a block device as a HAMMER volume the
 222                  * sector size must be compatible.  HAMMER uses 16384 byte
 223                  * filesystem buffers.
 224                  */
 225                 if (pinfo.reserved_blocks) {
 226                         errx(1, "HAMMER cannot be placed in a partition "
 227                                 "which overlaps the disklabel or MBR");
 228                 }
 229                 if (pinfo.media_blksize > HAMMER_BUFSIZE ||
 230                     HAMMER_BUFSIZE % pinfo.media_blksize) {
 231                         errx(1, "A media sector size of %d is not supported",
 232                              pinfo.media_blksize);
 233                 }
 234
 235                 vol->size = pinfo.media_size;
 236                 vol->device_offset = pinfo.media_offset;
 237                 vol->type = "DEVICE";
 238         }
 239 }
 240
 241 struct volume_info *
 242 get_volume(int32_t vol_no)
 243 {
 244         struct volume_info *vol;
 245
 246         TAILQ_FOREACH(vol, &VolList, entry) {
 247                 if (vol->vol_no == vol_no)
 248                         break;
 249         }
 250         if (vol == NULL) {
 251                 if (AssertOnFailure)
 252                         errx(1, "get_volume: Volume %d does not exist!",
 253                                 vol_no);
 254                 return(NULL);
 255         }
 256         ++vol->cache.refs;
 257         /* not added to or removed from hammer cache */
 258         return(vol);
 259 }
 260
 261 struct volume_info *
 262 get_root_volume(void)
 263 {
 264         return(get_volume(HAMMER_ROOT_VOLNO));
 265 }
 266
 267 void
 268 rel_volume(struct volume_info *volume)
 269 {
 270         if (volume == NULL)
 271                 return;
 272         /* not added to or removed from hammer cache */
 273         --volume->cache.refs;
 274 }
 275
 276 /*
 277  * Acquire the specified buffer.  isnew is -1 only when called
 278  * via get_buffer_readahead() to prevent another readahead.
 279  */
 280 struct buffer_info *
 281 get_buffer(hammer_off_t buf_offset, int isnew)
 282 {
 283         void *ondisk;
 284         struct buffer_info *buf;
 285         struct volume_info *volume;
 286         hammer_off_t orig_offset = buf_offset;
 287         int vol_no;
 288         int zone;
 289         int hi, n;
 290         int dora = 0;
 291
 292         zone = HAMMER_ZONE_DECODE(buf_offset);
 293         if (zone > HAMMER_ZONE_RAW_BUFFER_INDEX) {
 294                 buf_offset = blockmap_lookup(buf_offset, NULL, NULL, NULL);
 295         }
 296         if (buf_offset == HAMMER_OFF_BAD)
 297                 return(NULL);
 298
 299         if (AssertOnFailure) {
 300                 assert((buf_offset & HAMMER_OFF_ZONE_MASK) ==
 301                        HAMMER_ZONE_RAW_BUFFER);
 302         }
 303         vol_no = HAMMER_VOL_DECODE(buf_offset);
 304         volume = get_volume(vol_no);
 305         if (volume == NULL)
 306                 return(NULL);
 307
 308         buf_offset &= ~HAMMER_BUFMASK64;
 309         buf = find_buffer(volume, buf_offset);
 310
 311         if (buf == NULL) {
 312                 buf = malloc(sizeof(*buf));
 313                 bzero(buf, sizeof(*buf));
 314                 if (DebugOpt > 1) {
 315                         fprintf(stderr, "get_buffer: %016llx %016llx at %p\n",
 316                                 (long long)orig_offset, (long long)buf_offset,
 317                                 buf);
 318                 }
 319                 buf->buf_offset = buf_offset;
 320                 buf->raw_offset = hammer_xlate_to_phys(volume->ondisk,
 321                                                         buf_offset);
 322                 buf->volume = volume;
 323                 hi = buffer_hash(buf_offset);
 324                 TAILQ_INSERT_TAIL(&volume->buffer_lists[hi], buf, entry);
 325                 ++volume->cache.refs;
 326                 buf->cache.u.buffer = buf;
 327                 hammer_cache_add(&buf->cache, ISBUFFER);
 328                 dora = (isnew == 0);
 329         } else {
 330                 if (DebugOpt > 1) {
 331                         fprintf(stderr, "get_buffer: %016llx %016llx at %p *\n",
 332                                 (long long)orig_offset, (long long)buf_offset,
 333                                 buf);
 334                 }
 335                 hammer_cache_used(&buf->cache);
 336                 ++buf->use_count;
 337         }
 338         ++buf->cache.refs;
 339         hammer_cache_flush();
 340         if ((ondisk = buf->ondisk) == NULL) {
 341                 buf->ondisk = ondisk = malloc(HAMMER_BUFSIZE);
 342                 if (isnew <= 0) {
 343                         n = readhammerbuf(volume, ondisk, buf->raw_offset);
 344                         if (n == -1) {
 345                                 if (AssertOnFailure)
 346                                         err(1, "get_buffer: %s:%016llx "
 347                                             "Read failed at offset %016llx",
 348                                             volume->name,
 349                                             (long long)buf->buf_offset,
 350                                             (long long)buf->raw_offset);
 351                                 bzero(ondisk, HAMMER_BUFSIZE);
 352                         }
 353                 }
 354         }
 355         if (isnew > 0) {
 356                 bzero(ondisk, HAMMER_BUFSIZE);
 357                 buf->cache.modified = 1;
 358         }
 359         if (dora)
 360                 get_buffer_readahead(buf);
 361         return(buf);
 362 }
 363
 364 static void
 365 get_buffer_readahead(struct buffer_info *base)
 366 {
 367         struct buffer_info *buf;
 368         struct volume_info *vol;
 369         hammer_off_t buf_offset;
 370         int64_t raw_offset;
 371         int ri = UseReadBehind;
 372         int re = UseReadAhead;
 373
 374         raw_offset = base->raw_offset + ri * HAMMER_BUFSIZE;
 375         vol = base->volume;
 376
 377         while (ri < re) {
 378                 if (raw_offset >= vol->ondisk->vol_buf_end)
 379                         break;
 380                 if (raw_offset < vol->ondisk->vol_buf_beg || ri == 0) {
 381                         ++ri;
 382                         raw_offset += HAMMER_BUFSIZE;
 383                         continue;
 384                 }
 385                 buf_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no,
 386                         raw_offset - vol->ondisk->vol_buf_beg);
 387                 buf = find_buffer(vol, buf_offset);
 388                 if (buf == NULL) {
 389                         buf = get_buffer(buf_offset, -1);
 390                         rel_buffer(buf);
 391                 }
 392                 ++ri;
 393                 raw_offset += HAMMER_BUFSIZE;
 394         }
 395 }
 396
 397 void
 398 rel_buffer(struct buffer_info *buffer)
 399 {
 400         struct volume_info *volume;
 401         int hi;
 402
 403         if (buffer == NULL)
 404                 return;
 405         assert(buffer->cache.refs > 0);
 406         if (--buffer->cache.refs == 0) {
 407                 if (buffer->cache.delete) {
 408                         hi = buffer_hash(buffer->buf_offset);
 409                         volume = buffer->volume;
 410                         if (buffer->cache.modified)
 411                                 flush_buffer(buffer);
 412                         TAILQ_REMOVE(&volume->buffer_lists[hi], buffer, entry);
 413                         hammer_cache_del(&buffer->cache);
 414                         free(buffer->ondisk);
 415                         free(buffer);
 416                         rel_volume(volume);
 417                 }
 418         }
 419 }
 420
 421 /*
 422  * Retrieve a pointer to a buffer data given a buffer offset.  The underlying
 423  * bufferp is freed if isnew or the offset is out of range of the cached data.
 424  * If bufferp is freed a referenced buffer is loaded into it.
 425  */
 426 void *
 427 get_buffer_data(hammer_off_t buf_offset, struct buffer_info **bufferp,
 428                 int isnew)
 429 {
 430         if (*bufferp != NULL) {
 431                 if (isnew > 0 ||
 432                     (((*bufferp)->buf_offset ^ buf_offset) & ~HAMMER_BUFMASK64)) {
 433                         rel_buffer(*bufferp);
 434                         *bufferp = NULL;
 435                 }
 436         }
 437         return(get_ondisk(buf_offset, bufferp, isnew));
 438 }
 439
 440 /*
 441  * Retrieve a pointer to a B-Tree node given a zone offset.  The underlying
 442  * bufferp is freed if non-NULL and a referenced buffer is loaded into it.
 443  */
 444 hammer_node_ondisk_t
 445 get_node(hammer_off_t node_offset, struct buffer_info **bufferp)
 446 {
 447         if (*bufferp != NULL) {
 448                 rel_buffer(*bufferp);
 449                 *bufferp = NULL;
 450         }
 451         return(get_ondisk(node_offset, bufferp, 0));
 452 }
 453
 454 /*
 455  * Return a pointer to a buffer data given a buffer offset.
 456  * If *bufferp is NULL acquire the buffer otherwise use that buffer.
 457  */
 458 static __inline
 459 void *
 460 get_ondisk(hammer_off_t buf_offset, struct buffer_info **bufferp, int isnew)
 461 {
 462         struct buffer_info *buffer;
 463
 464         buffer = *bufferp;
 465         if (buffer == NULL) {
 466                 buffer = *bufferp = get_buffer(buf_offset, isnew);
 467                 if (buffer == NULL)
 468                         return(NULL);
 469         }
 470
 471         return((char *)buffer->ondisk +
 472                 ((int32_t)buf_offset & HAMMER_BUFMASK));
 473 }
 474
 475 /*
 476  * Allocate HAMMER elements - btree nodes, meta data, data storage
 477  */
 478 void *
 479 alloc_btree_element(hammer_off_t *offp, struct buffer_info **data_bufferp)
 480 {
 481         hammer_node_ondisk_t node;
 482
 483         node = alloc_blockmap(HAMMER_ZONE_BTREE_INDEX, sizeof(*node),
 484                               offp, data_bufferp);
 485         bzero(node, sizeof(*node));
 486         return (node);
 487 }
 488
 489 void *
 490 alloc_meta_element(hammer_off_t *offp, int32_t data_len,
 491                    struct buffer_info **data_bufferp)
 492 {
 493         void *data;
 494
 495         data = alloc_blockmap(HAMMER_ZONE_META_INDEX, data_len,
 496                               offp, data_bufferp);
 497         bzero(data, data_len);
 498         return (data);
 499 }
 500
 501 /*
 502  * The only data_len supported by HAMMER userspace for large data zone
 503  * (zone 10) is HAMMER_BUFSIZE which is 16KB.  >16KB data does not fit
 504  * in a buffer allocated by get_buffer().  Also alloc_blockmap() does
 505  * not consider >16KB buffer size.
 506  */
 507 void *
 508 alloc_data_element(hammer_off_t *offp, int32_t data_len,
 509                    struct buffer_info **data_bufferp)
 510 {
 511         void *data;
 512         int zone;
 513
 514         if (data_len == 0)
 515                 return(NULL);
 516
 517         zone = hammer_data_zone_index(data_len);
 518         assert(data_len <= HAMMER_BUFSIZE); /* just one buffer */
 519         assert(zone == HAMMER_ZONE_LARGE_DATA_INDEX ||
 520                zone == HAMMER_ZONE_SMALL_DATA_INDEX);
 521
 522         data = alloc_blockmap(zone, data_len, offp, data_bufferp);
 523         bzero(data, data_len);
 524         return(data);
 525 }
 526
 527 /*
 528  * Format a new blockmap.  This is mostly a degenerate case because
 529  * all allocations are now actually done from the freemap.
 530  */
 531 void
 532 format_blockmap(struct volume_info *root_vol, int zone, hammer_off_t offset)
 533 {
 534         hammer_blockmap_t blockmap;
 535         hammer_off_t zone_base;
 536
 537         /* Only root volume needs formatting */
 538         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 539
 540         assert(hammer_is_zone2_mapped_index(zone));
 541
 542         blockmap = &root_vol->ondisk->vol0_blockmap[zone];
 543         zone_base = HAMMER_ZONE_ENCODE(zone, offset);
 544
 545         bzero(blockmap, sizeof(*blockmap));
 546         blockmap->phys_offset = 0;
 547         blockmap->first_offset = zone_base;
 548         blockmap->next_offset = zone_base;
 549         blockmap->alloc_offset = HAMMER_ENCODE(zone, 255, -1);
 550         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 551         root_vol->cache.modified = 1;
 552 }
 553
 554 /*
 555  * Format a new freemap.  Set all layer1 entries to UNAVAIL.  The initialize
 556  * code will load each volume's freemap.
 557  */
 558 void
 559 format_freemap(struct volume_info *root_vol)
 560 {
 561         struct buffer_info *buffer = NULL;
 562         hammer_off_t layer1_offset;
 563         hammer_blockmap_t blockmap;
 564         struct hammer_blockmap_layer1 *layer1;
 565         int i, isnew;
 566
 567         /* Only root volume needs formatting */
 568         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 569
 570         layer1_offset = alloc_bigblock(root_vol, HAMMER_ZONE_FREEMAP_INDEX);
 571         for (i = 0; i < HAMMER_BIGBLOCK_SIZE; i += sizeof(*layer1)) {
 572                 isnew = ((i % HAMMER_BUFSIZE) == 0);
 573                 layer1 = get_buffer_data(layer1_offset + i, &buffer, isnew);
 574                 bzero(layer1, sizeof(*layer1));
 575                 layer1->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 576                 layer1->blocks_free = 0;
 577                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 578         }
 579         assert(i == HAMMER_BIGBLOCK_SIZE);
 580         rel_buffer(buffer);
 581
 582         blockmap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 583         bzero(blockmap, sizeof(*blockmap));
 584         blockmap->phys_offset = layer1_offset;
 585         blockmap->first_offset = 0;
 586         blockmap->next_offset = HAMMER_ENCODE_RAW_BUFFER(0, 0);
 587         blockmap->alloc_offset = HAMMER_ENCODE_RAW_BUFFER(255, -1);
 588         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 589         root_vol->cache.modified = 1;
 590 }
 591
 592 /*
 593  * Load the volume's remaining free space into the freemap.
 594  *
 595  * Returns the number of big-blocks available.
 596  */
 597 int64_t
 598 initialize_freemap(struct volume_info *vol)
 599 {
 600         struct volume_info *root_vol;
 601         struct buffer_info *buffer1 = NULL;
 602         struct buffer_info *buffer2 = NULL;
 603         struct hammer_blockmap_layer1 *layer1;
 604         struct hammer_blockmap_layer2 *layer2;
 605         hammer_off_t layer1_base;
 606         hammer_off_t layer1_offset;
 607         hammer_off_t layer2_offset;
 608         hammer_off_t phys_offset;
 609         hammer_off_t block_offset;
 610         hammer_off_t aligned_vol_free_end;
 611         hammer_blockmap_t freemap;
 612         int64_t count = 0;
 613         int64_t layer1_count = 0;
 614
 615         root_vol = get_root_volume();
 616         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 617                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 618
 619         printf("initialize freemap volume %d\n", vol->vol_no);
 620
 621         /*
 622          * Initialize the freemap.  First preallocate the big-blocks required
 623          * to implement layer2.   This preallocation is a bootstrap allocation
 624          * using blocks from the target volume.
 625          */
 626         freemap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 627         layer1_base = freemap->phys_offset;
 628
 629         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 630              phys_offset < aligned_vol_free_end;
 631              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 632                 layer1_offset = layer1_base +
 633                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 634                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 635                 if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 636                         layer1->phys_offset = alloc_bigblock(vol,
 637                                                 HAMMER_ZONE_FREEMAP_INDEX);
 638                         layer1->blocks_free = 0;
 639                         buffer1->cache.modified = 1;
 640                         layer1->layer1_crc = crc32(layer1,
 641                                                    HAMMER_LAYER1_CRCSIZE);
 642                 }
 643         }
 644
 645         /*
 646          * Now fill everything in.
 647          */
 648         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 649              phys_offset < aligned_vol_free_end;
 650              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 651                 layer1_count = 0;
 652                 layer1_offset = layer1_base +
 653                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 654                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 655                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 656
 657                 for (block_offset = 0;
 658                      block_offset < HAMMER_BLOCKMAP_LAYER2;
 659                      block_offset += HAMMER_BIGBLOCK_SIZE) {
 660                         layer2_offset = layer1->phys_offset +
 661                                         HAMMER_BLOCKMAP_LAYER2_OFFSET(block_offset);
 662                         layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 663                         bzero(layer2, sizeof(*layer2));
 664
 665                         if (phys_offset + block_offset < vol->vol_free_off) {
 666                                 /*
 667                                  * Fixups XXX - big-blocks already allocated as part
 668                                  * of the freemap bootstrap.
 669                                  */
 670                                 if (layer2->zone == 0) {
 671                                         layer2->zone = HAMMER_ZONE_FREEMAP_INDEX;
 672                                         layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 673                                         layer2->bytes_free = 0;
 674                                 }
 675                         } else if (phys_offset + block_offset < vol->vol_free_end) {
 676                                 layer2->zone = 0;
 677                                 layer2->append_off = 0;
 678                                 layer2->bytes_free = HAMMER_BIGBLOCK_SIZE;
 679                                 ++count;
 680                                 ++layer1_count;
 681                         } else {
 682                                 layer2->zone = HAMMER_ZONE_UNAVAIL_INDEX;
 683                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 684                                 layer2->bytes_free = 0;
 685                         }
 686                         layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 687                         buffer2->cache.modified = 1;
 688                 }
 689
 690                 layer1->blocks_free += layer1_count;
 691                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 692                 buffer1->cache.modified = 1;
 693         }
 694
 695         rel_buffer(buffer1);
 696         rel_buffer(buffer2);
 697         rel_volume(root_vol);
 698         return(count);
 699 }
 700
 701 /*
 702  * Returns the number of big-blocks available for filesystem data and undos
 703  * without formatting.
 704  */
 705 int64_t
 706 count_freemap(struct volume_info *vol)
 707 {
 708         hammer_off_t phys_offset;
 709         hammer_off_t vol_free_off;
 710         hammer_off_t aligned_vol_free_end;
 711         int64_t count = 0;
 712
 713         vol_free_off = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 714         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 715                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 716
 717         if (vol->vol_no == HAMMER_ROOT_VOLNO)
 718                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 719
 720         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 721              phys_offset < aligned_vol_free_end;
 722              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 723                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 724         }
 725
 726         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 727              phys_offset < aligned_vol_free_end;
 728              phys_offset += HAMMER_BIGBLOCK_SIZE) {
 729                 if (phys_offset < vol_free_off) {
 730                         ;
 731                 } else if (phys_offset < vol->vol_free_end) {
 732                         ++count;
 733                 }
 734         }
 735
 736         return(count);
 737 }
 738
 739 /*
 740  * Format the undomap for the root volume.
 741  */
 742 void
 743 format_undomap(struct volume_info *root_vol)
 744 {
 745         const int undo_zone = HAMMER_ZONE_UNDO_INDEX;
 746         hammer_off_t undo_limit;
 747         hammer_blockmap_t blockmap;
 748         struct hammer_volume_ondisk *ondisk;
 749         struct buffer_info *buffer = NULL;
 750         hammer_off_t scan;
 751         int n;
 752         int limit_index;
 753         uint32_t seqno;
 754
 755         /* Only root volume needs formatting */
 756         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 757         ondisk = root_vol->ondisk;
 758
 759         /*
 760          * Size the undo buffer in multiples of HAMMER_BIGBLOCK_SIZE,
 761          * up to HAMMER_UNDO_LAYER2 big-blocks.  Size to approximately
 762          * 0.1% of the disk.
 763          *
 764          * The minimum UNDO fifo size is 500MB, or approximately 1% of
 765          * the recommended 50G disk.
 766          *
 767          * Changing this minimum is rather dangerous as complex filesystem
 768          * operations can cause the UNDO FIFO to fill up otherwise.
 769          */
 770         undo_limit = UndoBufferSize;
 771         if (undo_limit == 0) {
 772                 undo_limit = (ondisk->vol_buf_end - ondisk->vol_buf_beg) / 1000;
 773                 if (undo_limit < 500*1024*1024)
 774                         undo_limit = 500*1024*1024;
 775         }
 776         undo_limit = (undo_limit + HAMMER_BIGBLOCK_MASK64) &
 777                      ~HAMMER_BIGBLOCK_MASK64;
 778         if (undo_limit < HAMMER_BIGBLOCK_SIZE)
 779                 undo_limit = HAMMER_BIGBLOCK_SIZE;
 780         if (undo_limit > HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2)
 781                 undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2;
 782         UndoBufferSize = undo_limit;
 783
 784         blockmap = &ondisk->vol0_blockmap[undo_zone];
 785         bzero(blockmap, sizeof(*blockmap));
 786         blockmap->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 787         blockmap->first_offset = HAMMER_ZONE_ENCODE(undo_zone, 0);
 788         blockmap->next_offset = blockmap->first_offset;
 789         blockmap->alloc_offset = HAMMER_ZONE_ENCODE(undo_zone, undo_limit);
 790         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 791
 792         limit_index = undo_limit / HAMMER_BIGBLOCK_SIZE;
 793         assert(limit_index <= HAMMER_UNDO_LAYER2);
 794
 795         for (n = 0; n < limit_index; ++n) {
 796                 ondisk->vol0_undo_array[n] = alloc_bigblock(NULL,
 797                                                         HAMMER_ZONE_UNDO_INDEX);
 798         }
 799         while (n < HAMMER_UNDO_LAYER2) {
 800                 ondisk->vol0_undo_array[n++] = HAMMER_BLOCKMAP_UNAVAIL;
 801         }
 802
 803         /*
 804          * Pre-initialize the UNDO blocks (HAMMER version 4+)
 805          */
 806         printf("initializing the undo map (%jd MB)\n",
 807                 (intmax_t)(blockmap->alloc_offset & HAMMER_OFF_LONG_MASK) /
 808                 (1024 * 1024));
 809
 810         scan = blockmap->first_offset;
 811         seqno = 0;
 812
 813         while (scan < blockmap->alloc_offset) {
 814                 hammer_fifo_head_t head;
 815                 hammer_fifo_tail_t tail;
 816                 int isnew;
 817                 int bytes = HAMMER_UNDO_ALIGN;
 818
 819                 isnew = ((scan & HAMMER_BUFMASK64) == 0);
 820                 head = get_buffer_data(scan, &buffer, isnew);
 821                 buffer->cache.modified = 1;
 822                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 823
 824                 bzero(head, bytes);
 825                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 826                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
 827                 head->hdr_size = bytes;
 828                 head->hdr_seq = seqno++;
 829
 830                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 831                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 832                 tail->tail_size = bytes;
 833
 834                 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
 835                                 crc32(head + 1, bytes - sizeof(*head));
 836
 837                 scan += bytes;
 838         }
 839         rel_buffer(buffer);
 840 }
 841
 842 /*
 843  * Flush various tracking structures to disk
 844  */
 845 void
 846 flush_all_volumes(void)
 847 {
 848         struct volume_info *vol;
 849
 850         TAILQ_FOREACH(vol, &VolList, entry)
 851                 flush_volume(vol);
 852 }
 853
 854 void
 855 flush_volume(struct volume_info *volume)
 856 {
 857         struct buffer_info *buffer;
 858         int i;
 859
 860         for (i = 0; i < HAMMER_BUFLISTS; ++i) {
 861                 TAILQ_FOREACH(buffer, &volume->buffer_lists[i], entry)
 862                         flush_buffer(buffer);
 863         }
 864         if (writehammerbuf(volume, volume->ondisk, 0) == -1)
 865                 err(1, "Write volume %d (%s)", volume->vol_no, volume->name);
 866         volume->cache.modified = 0;
 867 }
 868
 869 void
 870 flush_buffer(struct buffer_info *buffer)
 871 {
 872         struct volume_info *vol;
 873
 874         vol = buffer->volume;
 875         if (writehammerbuf(vol, buffer->ondisk, buffer->raw_offset) == -1)
 876                 err(1, "Write volume %d (%s)", vol->vol_no, vol->name);
 877         buffer->cache.modified = 0;
 878 }
 879
 880 /*
 881  * Core I/O operations
 882  */
 883 static int
 884 readhammerbuf(struct volume_info *vol, void *data, int64_t offset)
 885 {
 886         ssize_t n;
 887
 888         n = pread(vol->fd, data, HAMMER_BUFSIZE, offset);
 889         if (n != HAMMER_BUFSIZE)
 890                 return(-1);
 891         return(0);
 892 }
 893
 894 static int
 895 writehammerbuf(struct volume_info *vol, const void *data, int64_t offset)
 896 {
 897         ssize_t n;
 898
 899         n = pwrite(vol->fd, data, HAMMER_BUFSIZE, offset);
 900         if (n != HAMMER_BUFSIZE)
 901                 return(-1);
 902         return(0);
 903 }
 904
 905 int64_t init_boot_area_size(int64_t value, off_t avg_vol_size)
 906 {
 907         if (value == 0) {
 908                 value = HAMMER_BOOT_NOMBYTES;
 909                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 910                         value >>= 1;
 911                 if (value < HAMMER_BOOT_MINBYTES)
 912                         value = 0;
 913         } else if (value < HAMMER_BOOT_MINBYTES) {
 914                 value = HAMMER_BOOT_MINBYTES;
 915         }
 916
 917         return(value);
 918 }
 919
 920 int64_t init_mem_area_size(int64_t value, off_t avg_vol_size)
 921 {
 922         if (value == 0) {
 923                 value = HAMMER_MEM_NOMBYTES;
 924                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 925                         value >>= 1;
 926                 if (value < HAMMER_MEM_MINBYTES)
 927                         value = 0;
 928         } else if (value < HAMMER_MEM_MINBYTES) {
 929                 value = HAMMER_MEM_MINBYTES;
 930         }
 931
 932         return(value);
 933 }