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