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 void *get_ondisk(hammer_off_t buf_offset, struct buffer_info **bufferp,
  42                         int isnew);
  43 static __inline int readhammervol(struct volume_info *vol);
  44 static __inline int readhammerbuf(struct buffer_info *buf);
  45 static __inline int writehammervol(struct volume_info *vol);
  46 static __inline int writehammerbuf(struct buffer_info *buf);
  47
  48 uuid_t Hammer_FSType;
  49 uuid_t Hammer_FSId;
  50 int UseReadBehind = -4;
  51 int UseReadAhead = 4;
  52 int DebugOpt;
  53 struct volume_list VolList = TAILQ_HEAD_INITIALIZER(VolList);
  54 static int valid_hammer_volumes;
  55
  56 static __inline
  57 int
  58 buffer_hash(hammer_off_t buf_offset)
  59 {
  60         int hi;
  61
  62         hi = (int)(buf_offset / HAMMER_BUFSIZE) & HAMMER_BUFLISTMASK;
  63         return(hi);
  64 }
  65
  66 static struct buffer_info*
  67 find_buffer(struct volume_info *volume, hammer_off_t buf_offset)
  68 {
  69         int hi;
  70         struct buffer_info *buf;
  71
  72         hi = buffer_hash(buf_offset);
  73         TAILQ_FOREACH(buf, &volume->buffer_lists[hi], entry)
  74                 if (buf->buf_offset == buf_offset)
  75                         return(buf);
  76         return(NULL);
  77 }
  78
  79 static
  80 struct volume_info *
  81 __alloc_volume(const char *volname, int oflags)
  82 {
  83         struct volume_info *vol;
  84         int i;
  85
  86         vol = malloc(sizeof(*vol));
  87         if (vol == NULL)
  88                 err(1, "alloc_volume");
  89         bzero(vol, sizeof(*vol));
  90
  91         vol->vol_no = -1;
  92         vol->rdonly = (oflags == O_RDONLY);
  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
 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 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         if (vol == NULL)
 250                 errx(1, "get_volume: Volume %d does not exist!", vol_no);
 251
 252         /* not added to or removed from hammer cache */
 253         return(vol);
 254 }
 255
 256 struct volume_info *
 257 get_root_volume(void)
 258 {
 259         return(get_volume(HAMMER_ROOT_VOLNO));
 260 }
 261
 262 /*
 263  * Acquire the specified buffer.  isnew is -1 only when called
 264  * via get_buffer_readahead() to prevent another readahead.
 265  */
 266 static struct buffer_info *
 267 get_buffer(hammer_off_t buf_offset, int isnew)
 268 {
 269         struct buffer_info *buf;
 270         struct volume_info *volume;
 271         int vol_no;
 272         int zone;
 273         int hi;
 274         int dora = 0;
 275         int error = 0;
 276
 277         zone = HAMMER_ZONE_DECODE(buf_offset);
 278         if (zone > HAMMER_ZONE_RAW_BUFFER_INDEX)
 279                 buf_offset = blockmap_lookup(buf_offset, NULL, NULL, &error);
 280         if (error || buf_offset == HAMMER_OFF_BAD)
 281                 return(NULL);
 282         assert(hammer_is_zone_raw_buffer(buf_offset));
 283
 284         vol_no = HAMMER_VOL_DECODE(buf_offset);
 285         volume = get_volume(vol_no);
 286
 287         buf_offset &= ~HAMMER_BUFMASK64;
 288         buf = find_buffer(volume, buf_offset);
 289
 290         if (buf == NULL) {
 291                 buf = malloc(sizeof(*buf));
 292                 bzero(buf, sizeof(*buf));
 293                 buf->buf_offset = buf_offset;
 294                 buf->raw_offset = hammer_xlate_to_phys(volume->ondisk,
 295                                                         buf_offset);
 296                 buf->volume = volume;
 297                 buf->ondisk = malloc(HAMMER_BUFSIZE);
 298                 if (isnew <= 0) {
 299                         if (readhammerbuf(buf) == -1) {
 300                                 err(1, "get_buffer: %s:%016jx "
 301                                     "Read failed at offset %016jx",
 302                                     volume->name,
 303                                     (intmax_t)buf->buf_offset,
 304                                     (intmax_t)buf->raw_offset);
 305                         }
 306                 }
 307
 308                 hi = buffer_hash(buf_offset);
 309                 TAILQ_INSERT_TAIL(&volume->buffer_lists[hi], buf, entry);
 310                 buf->cache.buffer = buf;
 311                 hammer_cache_add(&buf->cache);
 312                 dora = (isnew == 0);
 313         } else {
 314                 assert(buf->ondisk != NULL);
 315                 assert(isnew != -1);
 316                 hammer_cache_used(&buf->cache);
 317         }
 318
 319         ++buf->cache.refs;
 320         hammer_cache_flush();
 321
 322         if (isnew > 0) {
 323                 assert(buf->cache.modified == 0);
 324                 bzero(buf->ondisk, HAMMER_BUFSIZE);
 325                 buf->cache.modified = 1;
 326         }
 327         if (dora)
 328                 get_buffer_readahead(buf);
 329         return(buf);
 330 }
 331
 332 static void
 333 get_buffer_readahead(struct buffer_info *base)
 334 {
 335         struct buffer_info *buf;
 336         struct volume_info *vol;
 337         hammer_off_t buf_offset;
 338         int64_t raw_offset;
 339         int ri = UseReadBehind;
 340         int re = UseReadAhead;
 341
 342         raw_offset = base->raw_offset + ri * HAMMER_BUFSIZE;
 343         vol = base->volume;
 344
 345         while (ri < re) {
 346                 if (raw_offset >= vol->ondisk->vol_buf_end)
 347                         break;
 348                 if (raw_offset < vol->ondisk->vol_buf_beg || ri == 0) {
 349                         ++ri;
 350                         raw_offset += HAMMER_BUFSIZE;
 351                         continue;
 352                 }
 353                 buf_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no,
 354                         raw_offset - vol->ondisk->vol_buf_beg);
 355                 buf = find_buffer(vol, buf_offset);
 356                 if (buf == NULL) {
 357                         buf = get_buffer(buf_offset, -1);
 358                         rel_buffer(buf);
 359                 }
 360                 ++ri;
 361                 raw_offset += HAMMER_BUFSIZE;
 362         }
 363 }
 364
 365 void
 366 rel_buffer(struct buffer_info *buffer)
 367 {
 368         struct volume_info *volume;
 369         int hi;
 370
 371         if (buffer == NULL)
 372                 return;
 373         assert(buffer->cache.refs > 0);
 374         if (--buffer->cache.refs == 0) {
 375                 if (buffer->cache.delete) {
 376                         hi = buffer_hash(buffer->buf_offset);
 377                         volume = buffer->volume;
 378                         if (buffer->cache.modified)
 379                                 flush_buffer(buffer);
 380                         TAILQ_REMOVE(&volume->buffer_lists[hi], buffer, entry);
 381                         hammer_cache_del(&buffer->cache);
 382                         free(buffer->ondisk);
 383                         free(buffer);
 384                 }
 385         }
 386 }
 387
 388 /*
 389  * Retrieve a pointer to a buffer data given a buffer offset.  The underlying
 390  * bufferp is freed if isnew or the offset is out of range of the cached data.
 391  * If bufferp is freed a referenced buffer is loaded into it.
 392  */
 393 void *
 394 get_buffer_data(hammer_off_t buf_offset, struct buffer_info **bufferp,
 395                 int isnew)
 396 {
 397         if (*bufferp != NULL) {
 398                 if (isnew > 0 ||
 399                     (((*bufferp)->buf_offset ^ buf_offset) & ~HAMMER_BUFMASK64)) {
 400                         rel_buffer(*bufferp);
 401                         *bufferp = NULL;
 402                 }
 403         }
 404         return(get_ondisk(buf_offset, bufferp, isnew));
 405 }
 406
 407 /*
 408  * Retrieve a pointer to a B-Tree node given a zone offset.  The underlying
 409  * bufferp is freed if non-NULL and a referenced buffer is loaded into it.
 410  */
 411 hammer_node_ondisk_t
 412 get_node(hammer_off_t node_offset, struct buffer_info **bufferp)
 413 {
 414         if (*bufferp != NULL) {
 415                 rel_buffer(*bufferp);
 416                 *bufferp = NULL;
 417         }
 418         return(get_ondisk(node_offset, bufferp, 0));
 419 }
 420
 421 /*
 422  * Return a pointer to a buffer data given a buffer offset.
 423  * If *bufferp is NULL acquire the buffer otherwise use that buffer.
 424  */
 425 static void *
 426 get_ondisk(hammer_off_t buf_offset, struct buffer_info **bufferp, int isnew)
 427 {
 428         if (*bufferp == NULL) {
 429                 *bufferp = get_buffer(buf_offset, isnew);
 430                 if (*bufferp == NULL)
 431                         return(NULL);
 432         }
 433
 434         return(((char *)(*bufferp)->ondisk) +
 435                 ((int32_t)buf_offset & HAMMER_BUFMASK));
 436 }
 437
 438 /*
 439  * Allocate HAMMER elements - B-Tree nodes
 440  */
 441 void *
 442 alloc_btree_element(hammer_off_t *offp, struct buffer_info **data_bufferp)
 443 {
 444         hammer_node_ondisk_t node;
 445
 446         node = alloc_blockmap(HAMMER_ZONE_BTREE_INDEX, sizeof(*node),
 447                               offp, data_bufferp);
 448         bzero(node, sizeof(*node));
 449         return (node);
 450 }
 451
 452 /*
 453  * Allocate HAMMER elements - meta data (inode, direntry, PFS, etc)
 454  */
 455 void *
 456 alloc_meta_element(hammer_off_t *offp, int32_t data_len,
 457                    struct buffer_info **data_bufferp)
 458 {
 459         void *data;
 460
 461         data = alloc_blockmap(HAMMER_ZONE_META_INDEX, data_len,
 462                               offp, data_bufferp);
 463         bzero(data, data_len);
 464         return (data);
 465 }
 466
 467 /*
 468  * Allocate HAMMER elements - data storage
 469  *
 470  * The only data_len supported by HAMMER userspace for large data zone
 471  * (zone 10) is HAMMER_BUFSIZE which is 16KB.  >16KB data does not fit
 472  * in a buffer allocated by get_buffer().  Also alloc_blockmap() does
 473  * not consider >16KB buffer size.
 474  */
 475 void *
 476 alloc_data_element(hammer_off_t *offp, int32_t data_len,
 477                    struct buffer_info **data_bufferp)
 478 {
 479         void *data;
 480         int zone;
 481
 482         if (data_len == 0)
 483                 return(NULL);
 484
 485         zone = hammer_data_zone_index(data_len);
 486         assert(data_len <= HAMMER_BUFSIZE); /* just one buffer */
 487         assert(zone == HAMMER_ZONE_LARGE_DATA_INDEX ||
 488                zone == HAMMER_ZONE_SMALL_DATA_INDEX);
 489
 490         data = alloc_blockmap(zone, data_len, offp, data_bufferp);
 491         bzero(data, data_len);
 492         return(data);
 493 }
 494
 495 /*
 496  * Format a new blockmap.  This is mostly a degenerate case because
 497  * all allocations are now actually done from the freemap.
 498  */
 499 void
 500 format_blockmap(struct volume_info *root_vol, int zone, hammer_off_t offset)
 501 {
 502         hammer_blockmap_t blockmap;
 503         hammer_off_t zone_base;
 504
 505         /* Only root volume needs formatting */
 506         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 507
 508         assert(hammer_is_zone2_mapped_index(zone));
 509
 510         blockmap = &root_vol->ondisk->vol0_blockmap[zone];
 511         zone_base = HAMMER_ZONE_ENCODE(zone, offset);
 512
 513         bzero(blockmap, sizeof(*blockmap));
 514         blockmap->phys_offset = 0;
 515         blockmap->first_offset = zone_base;
 516         blockmap->next_offset = zone_base;
 517         blockmap->alloc_offset = HAMMER_ENCODE(zone, 255, -1);
 518         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 519 }
 520
 521 /*
 522  * Format a new freemap.  Set all layer1 entries to UNAVAIL.  The initialize
 523  * code will load each volume's freemap.
 524  */
 525 void
 526 format_freemap(struct volume_info *root_vol)
 527 {
 528         struct buffer_info *buffer = NULL;
 529         hammer_off_t layer1_offset;
 530         hammer_blockmap_t blockmap;
 531         hammer_blockmap_layer1_t layer1;
 532         int i, isnew;
 533
 534         /* Only root volume needs formatting */
 535         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 536
 537         layer1_offset = alloc_bigblock(root_vol, HAMMER_ZONE_FREEMAP_INDEX);
 538         for (i = 0; i < HAMMER_BIGBLOCK_SIZE; i += sizeof(*layer1)) {
 539                 isnew = ((i % HAMMER_BUFSIZE) == 0);
 540                 layer1 = get_buffer_data(layer1_offset + i, &buffer, isnew);
 541                 bzero(layer1, sizeof(*layer1));
 542                 layer1->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 543                 layer1->blocks_free = 0;
 544                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 545         }
 546         assert(i == HAMMER_BIGBLOCK_SIZE);
 547         rel_buffer(buffer);
 548
 549         blockmap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 550         bzero(blockmap, sizeof(*blockmap));
 551         blockmap->phys_offset = layer1_offset;
 552         blockmap->first_offset = 0;
 553         blockmap->next_offset = HAMMER_ENCODE_RAW_BUFFER(0, 0);
 554         blockmap->alloc_offset = HAMMER_ENCODE_RAW_BUFFER(255, -1);
 555         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 556 }
 557
 558 /*
 559  * Load the volume's remaining free space into the freemap.
 560  *
 561  * Returns the number of big-blocks available.
 562  */
 563 int64_t
 564 initialize_freemap(struct volume_info *vol)
 565 {
 566         struct volume_info *root_vol;
 567         struct buffer_info *buffer1 = NULL;
 568         struct buffer_info *buffer2 = NULL;
 569         hammer_blockmap_layer1_t layer1;
 570         hammer_blockmap_layer2_t layer2;
 571         hammer_off_t layer1_offset;
 572         hammer_off_t layer2_offset;
 573         hammer_off_t phys_offset;
 574         hammer_off_t block_offset;
 575         hammer_off_t aligned_vol_free_end;
 576         hammer_blockmap_t freemap;
 577         int64_t count = 0;
 578         int64_t layer1_count = 0;
 579
 580         root_vol = get_root_volume();
 581         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 582                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 583
 584         printf("initialize freemap volume %d\n", vol->vol_no);
 585
 586         /*
 587          * Initialize the freemap.  First preallocate the big-blocks required
 588          * to implement layer2.   This preallocation is a bootstrap allocation
 589          * using blocks from the target volume.
 590          */
 591         freemap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 592
 593         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 594              phys_offset < aligned_vol_free_end;
 595              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 596                 layer1_offset = freemap->phys_offset +
 597                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 598                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 599                 if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 600                         layer1->phys_offset = alloc_bigblock(vol,
 601                                                 HAMMER_ZONE_FREEMAP_INDEX);
 602                         layer1->blocks_free = 0;
 603                         buffer1->cache.modified = 1;
 604                         layer1->layer1_crc = crc32(layer1,
 605                                                    HAMMER_LAYER1_CRCSIZE);
 606                 }
 607         }
 608
 609         /*
 610          * Now fill everything in.
 611          */
 612         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 613              phys_offset < aligned_vol_free_end;
 614              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 615                 layer1_count = 0;
 616                 layer1_offset = freemap->phys_offset +
 617                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 618                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 619                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 620
 621                 for (block_offset = 0;
 622                      block_offset < HAMMER_BLOCKMAP_LAYER2;
 623                      block_offset += HAMMER_BIGBLOCK_SIZE) {
 624                         layer2_offset = layer1->phys_offset +
 625                                         HAMMER_BLOCKMAP_LAYER2_OFFSET(block_offset);
 626                         layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 627                         bzero(layer2, sizeof(*layer2));
 628
 629                         if (phys_offset + block_offset < vol->vol_free_off) {
 630                                 /*
 631                                  * Fixups XXX - big-blocks already allocated as part
 632                                  * of the freemap bootstrap.
 633                                  */
 634                                 layer2->zone = HAMMER_ZONE_FREEMAP_INDEX;
 635                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 636                                 layer2->bytes_free = 0;
 637                         } else if (phys_offset + block_offset < vol->vol_free_end) {
 638                                 layer2->zone = 0;
 639                                 layer2->append_off = 0;
 640                                 layer2->bytes_free = HAMMER_BIGBLOCK_SIZE;
 641                                 ++count;
 642                                 ++layer1_count;
 643                         } else {
 644                                 layer2->zone = HAMMER_ZONE_UNAVAIL_INDEX;
 645                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 646                                 layer2->bytes_free = 0;
 647                         }
 648                         layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 649                         buffer2->cache.modified = 1;
 650                 }
 651
 652                 layer1->blocks_free += layer1_count;
 653                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 654                 buffer1->cache.modified = 1;
 655         }
 656
 657         rel_buffer(buffer1);
 658         rel_buffer(buffer2);
 659         return(count);
 660 }
 661
 662 /*
 663  * Returns the number of big-blocks available for filesystem data and undos
 664  * without formatting.
 665  */
 666 int64_t
 667 count_freemap(struct volume_info *vol)
 668 {
 669         hammer_off_t phys_offset;
 670         hammer_off_t vol_free_off;
 671         hammer_off_t aligned_vol_free_end;
 672         int64_t count = 0;
 673
 674         vol_free_off = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 675         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 676                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 677
 678         if (vol->vol_no == HAMMER_ROOT_VOLNO)
 679                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 680
 681         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 682              phys_offset < aligned_vol_free_end;
 683              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 684                 vol_free_off += HAMMER_BIGBLOCK_SIZE;
 685         }
 686
 687         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 688              phys_offset < aligned_vol_free_end;
 689              phys_offset += HAMMER_BIGBLOCK_SIZE) {
 690                 if (phys_offset < vol_free_off) {
 691                         ;
 692                 } else if (phys_offset < vol->vol_free_end) {
 693                         ++count;
 694                 }
 695         }
 696
 697         return(count);
 698 }
 699
 700 /*
 701  * Format the undomap for the root volume.
 702  */
 703 void
 704 format_undomap(struct volume_info *root_vol, int64_t *undo_buffer_size)
 705 {
 706         const int undo_zone = HAMMER_ZONE_UNDO_INDEX;
 707         hammer_off_t undo_limit;
 708         hammer_blockmap_t blockmap;
 709         hammer_volume_ondisk_t ondisk;
 710         struct buffer_info *buffer = NULL;
 711         hammer_off_t scan;
 712         int n;
 713         int limit_index;
 714         uint32_t seqno;
 715
 716         /* Only root volume needs formatting */
 717         assert(root_vol->vol_no == HAMMER_ROOT_VOLNO);
 718         ondisk = root_vol->ondisk;
 719
 720         /*
 721          * Size the undo buffer in multiples of HAMMER_BIGBLOCK_SIZE,
 722          * up to HAMMER_UNDO_LAYER2 big-blocks.  Size to approximately
 723          * 0.1% of the disk.
 724          *
 725          * The minimum UNDO fifo size is 500MB, or approximately 1% of
 726          * the recommended 50G disk.
 727          *
 728          * Changing this minimum is rather dangerous as complex filesystem
 729          * operations can cause the UNDO FIFO to fill up otherwise.
 730          */
 731         undo_limit = *undo_buffer_size;
 732         if (undo_limit == 0) {
 733                 undo_limit = HAMMER_VOL_BUF_SIZE(ondisk) / 1000;
 734                 if (undo_limit < 500*1024*1024)
 735                         undo_limit = 500*1024*1024;
 736         }
 737         undo_limit = (undo_limit + HAMMER_BIGBLOCK_MASK64) &
 738                      ~HAMMER_BIGBLOCK_MASK64;
 739         if (undo_limit < HAMMER_BIGBLOCK_SIZE)
 740                 undo_limit = HAMMER_BIGBLOCK_SIZE;
 741         if (undo_limit > HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2)
 742                 undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2;
 743         *undo_buffer_size = undo_limit;
 744
 745         blockmap = &ondisk->vol0_blockmap[undo_zone];
 746         bzero(blockmap, sizeof(*blockmap));
 747         blockmap->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 748         blockmap->first_offset = HAMMER_ZONE_ENCODE(undo_zone, 0);
 749         blockmap->next_offset = blockmap->first_offset;
 750         blockmap->alloc_offset = HAMMER_ZONE_ENCODE(undo_zone, undo_limit);
 751         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 752
 753         limit_index = undo_limit / HAMMER_BIGBLOCK_SIZE;
 754         assert(limit_index <= HAMMER_UNDO_LAYER2);
 755
 756         for (n = 0; n < limit_index; ++n) {
 757                 ondisk->vol0_undo_array[n] = alloc_bigblock(root_vol,
 758                                                         HAMMER_ZONE_UNDO_INDEX);
 759         }
 760         while (n < HAMMER_UNDO_LAYER2) {
 761                 ondisk->vol0_undo_array[n++] = HAMMER_BLOCKMAP_UNAVAIL;
 762         }
 763
 764         /*
 765          * Pre-initialize the UNDO blocks (HAMMER version 4+)
 766          */
 767         printf("initializing the undo map (%jd MB)\n",
 768                 (intmax_t)(blockmap->alloc_offset & HAMMER_OFF_LONG_MASK) /
 769                 (1024 * 1024));
 770
 771         scan = blockmap->first_offset;
 772         seqno = 0;
 773
 774         while (scan < blockmap->alloc_offset) {
 775                 hammer_fifo_head_t head;
 776                 hammer_fifo_tail_t tail;
 777                 int isnew;
 778                 int bytes = HAMMER_UNDO_ALIGN;
 779
 780                 isnew = ((scan & HAMMER_BUFMASK64) == 0);
 781                 head = get_buffer_data(scan, &buffer, isnew);
 782                 buffer->cache.modified = 1;
 783                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 784
 785                 bzero(head, bytes);
 786                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 787                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
 788                 head->hdr_size = bytes;
 789                 head->hdr_seq = seqno++;
 790
 791                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 792                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 793                 tail->tail_size = bytes;
 794
 795                 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
 796                                 crc32(head + 1, bytes - sizeof(*head));
 797
 798                 scan += bytes;
 799         }
 800         rel_buffer(buffer);
 801 }
 802
 803 const char *zone_labels[] = {
 804         "",             /* 0 */
 805         "raw_volume",   /* 1 */
 806         "raw_buffer",   /* 2 */
 807         "undo",         /* 3 */
 808         "freemap",      /* 4 */
 809         "",             /* 5 */
 810         "",             /* 6 */
 811         "",             /* 7 */
 812         "btree",        /* 8 */
 813         "meta",         /* 9 */
 814         "large_data",   /* 10 */
 815         "small_data",   /* 11 */
 816         "",             /* 12 */
 817         "",             /* 13 */
 818         "",             /* 14 */
 819         "unavail",      /* 15 */
 820 };
 821
 822 void
 823 print_blockmap(const struct volume_info *root_vol)
 824 {
 825         hammer_blockmap_t blockmap;
 826         hammer_volume_ondisk_t ondisk;
 827         int64_t size, used;
 828         int i;
 829 #define INDENT ""
 830
 831         ondisk = root_vol->ondisk;
 832         printf(INDENT"vol_label\t%s\n", ondisk->vol_label);
 833         printf(INDENT"vol_count\t%d\n", ondisk->vol_count);
 834         printf(INDENT"vol_bot_beg\t%s\n", sizetostr(ondisk->vol_bot_beg));
 835         printf(INDENT"vol_mem_beg\t%s\n", sizetostr(ondisk->vol_mem_beg));
 836         printf(INDENT"vol_buf_beg\t%s\n", sizetostr(ondisk->vol_buf_beg));
 837         printf(INDENT"vol_buf_end\t%s\n", sizetostr(ondisk->vol_buf_end));
 838         printf(INDENT"vol0_next_tid\t%016jx\n",
 839                (uintmax_t)ondisk->vol0_next_tid);
 840
 841         blockmap = &ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 842         size = blockmap->alloc_offset & HAMMER_OFF_LONG_MASK;
 843         if (blockmap->first_offset <= blockmap->next_offset)
 844                 used = blockmap->next_offset - blockmap->first_offset;
 845         else
 846                 used = blockmap->alloc_offset - blockmap->first_offset +
 847                         (blockmap->next_offset & HAMMER_OFF_LONG_MASK);
 848         printf(INDENT"undo_size\t%s\n", sizetostr(size));
 849         printf(INDENT"undo_used\t%s\n", sizetostr(used));
 850
 851         printf(INDENT"zone #             "
 852                "phys             first            next             alloc\n");
 853         for (i = 0; i < HAMMER_MAX_ZONES; i++) {
 854                 blockmap = &ondisk->vol0_blockmap[i];
 855                 printf(INDENT"zone %-2d %-10s %016jx %016jx %016jx %016jx\n",
 856                         i, zone_labels[i],
 857                         (uintmax_t)blockmap->phys_offset,
 858                         (uintmax_t)blockmap->first_offset,
 859                         (uintmax_t)blockmap->next_offset,
 860                         (uintmax_t)blockmap->alloc_offset);
 861         }
 862 }
 863
 864 /*
 865  * Flush various tracking structures to disk
 866  */
 867 void
 868 flush_all_volumes(void)
 869 {
 870         struct volume_info *vol;
 871
 872         TAILQ_FOREACH(vol, &VolList, entry)
 873                 flush_volume(vol);
 874 }
 875
 876 void
 877 flush_volume(struct volume_info *volume)
 878 {
 879         struct buffer_info *buffer;
 880         int i;
 881
 882         for (i = 0; i < HAMMER_BUFLISTS; ++i) {
 883                 TAILQ_FOREACH(buffer, &volume->buffer_lists[i], entry)
 884                         flush_buffer(buffer);
 885         }
 886         if (writehammervol(volume) == -1)
 887                 err(1, "Write volume %d (%s)", volume->vol_no, volume->name);
 888 }
 889
 890 void
 891 flush_buffer(struct buffer_info *buffer)
 892 {
 893         struct volume_info *vol;
 894
 895         vol = buffer->volume;
 896         if (writehammerbuf(buffer) == -1)
 897                 err(1, "Write volume %d (%s)", vol->vol_no, vol->name);
 898         buffer->cache.modified = 0;
 899 }
 900
 901 /*
 902  * Core I/O operations
 903  */
 904 static int
 905 __read(struct volume_info *vol, void *data, int64_t offset, int size)
 906 {
 907         ssize_t n;
 908
 909         n = pread(vol->fd, data, size, offset);
 910         if (n != size)
 911                 return(-1);
 912         return(0);
 913 }
 914
 915 static __inline int
 916 readhammervol(struct volume_info *vol)
 917 {
 918         return(__read(vol, vol->ondisk, 0, HAMMER_BUFSIZE));
 919 }
 920
 921 static __inline int
 922 readhammerbuf(struct buffer_info *buf)
 923 {
 924         return(__read(buf->volume, buf->ondisk, buf->raw_offset, HAMMER_BUFSIZE));
 925 }
 926
 927 static int
 928 __write(struct volume_info *vol, const void *data, int64_t offset, int size)
 929 {
 930         ssize_t n;
 931
 932         if (vol->rdonly)
 933                 return(0);
 934
 935         n = pwrite(vol->fd, data, size, offset);
 936         if (n != size)
 937                 return(-1);
 938         return(0);
 939 }
 940
 941 static __inline int
 942 writehammervol(struct volume_info *vol)
 943 {
 944         return(__write(vol, vol->ondisk, 0, HAMMER_BUFSIZE));
 945 }
 946
 947 static __inline int
 948 writehammerbuf(struct buffer_info *buf)
 949 {
 950         return(__write(buf->volume, buf->ondisk, buf->raw_offset, HAMMER_BUFSIZE));
 951 }
 952
 953 int64_t init_boot_area_size(int64_t value, off_t avg_vol_size)
 954 {
 955         if (value == 0) {
 956                 value = HAMMER_BOOT_NOMBYTES;
 957                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 958                         value >>= 1;
 959                 if (value < HAMMER_BOOT_MINBYTES)
 960                         value = 0;
 961         } else if (value < HAMMER_BOOT_MINBYTES) {
 962                 value = HAMMER_BOOT_MINBYTES;
 963         }
 964
 965         return(value);
 966 }
 967
 968 int64_t init_mem_area_size(int64_t value, off_t avg_vol_size)
 969 {
 970         if (value == 0) {
 971                 value = HAMMER_MEM_NOMBYTES;
 972                 while (value > avg_vol_size / HAMMER_MAX_VOLUMES)
 973                         value >>= 1;
 974                 if (value < HAMMER_MEM_MINBYTES)
 975                         value = 0;
 976         } else if (value < HAMMER_MEM_MINBYTES) {
 977                 value = HAMMER_MEM_MINBYTES;
 978         }
 979
 980         return(value);
 981 }