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  * $DragonFly: src/sbin/hammer/ondisk.c,v 1.21 2008/06/02 16:57:53 dillon Exp $
  35  */
  36
  37 #include <sys/types.h>
  38 #include <assert.h>
  39 #include <stdio.h>
  40 #include <stdlib.h>
  41 #include <stdarg.h>
  42 #include <string.h>
  43 #include <unistd.h>
  44 #include <stddef.h>
  45 #include <err.h>
  46 #include <fcntl.h>
  47 #include "hammer_util.h"
  48
  49 static void *alloc_blockmap(int zone, int bytes, hammer_off_t *result_offp,
  50                         struct buffer_info **bufferp);
  51 static hammer_off_t alloc_bigblock(struct volume_info *volume,
  52                         hammer_off_t owner);
  53 #if 0
  54 static void init_fifo_head(hammer_fifo_head_t head, u_int16_t hdr_type);
  55 static hammer_off_t hammer_alloc_fifo(int32_t base_bytes, int32_t ext_bytes,
  56                         struct buffer_info **bufp, u_int16_t hdr_type);
  57 static void readhammerbuf(struct volume_info *vol, void *data,
  58                         int64_t offset);
  59 #endif
  60 static void writehammerbuf(struct volume_info *vol, const void *data,
  61                         int64_t offset);
  62
  63 int DebugOpt;
  64
  65 uuid_t Hammer_FSType;
  66 uuid_t Hammer_FSId;
  67 int64_t BootAreaSize;
  68 int64_t MemAreaSize;
  69 int64_t UndoBufferSize;
  70 int     UsingSuperClusters;
  71 int     NumVolumes;
  72 int     RootVolNo = -1;
  73 struct volume_list VolList = TAILQ_HEAD_INITIALIZER(VolList);
  74
  75 static __inline
  76 int
  77 buffer_hash(hammer_off_t buf_offset)
  78 {
  79         int hi;
  80
  81         hi = (int)(buf_offset / HAMMER_BUFSIZE) & HAMMER_BUFLISTMASK;
  82         return(hi);
  83 }
  84
  85 /*
  86  * Lookup the requested information structure and related on-disk buffer.
  87  * Missing structures are created.
  88  */
  89 struct volume_info *
  90 setup_volume(int32_t vol_no, const char *filename, int isnew, int oflags)
  91 {
  92         struct volume_info *vol;
  93         struct volume_info *scan;
  94         struct hammer_volume_ondisk *ondisk;
  95         int i, n;
  96
  97         /*
  98          * Allocate the volume structure
  99          */
 100         vol = malloc(sizeof(*vol));
 101         bzero(vol, sizeof(*vol));
 102         for (i = 0; i < HAMMER_BUFLISTS; ++i)
 103                 TAILQ_INIT(&vol->buffer_lists[i]);
 104         vol->name = strdup(filename);
 105         vol->fd = open(filename, oflags);
 106         if (vol->fd < 0) {
 107                 free(vol->name);
 108                 free(vol);
 109                 err(1, "setup_volume: %s: Open failed", filename);
 110         }
 111
 112         /*
 113          * Read or initialize the volume header
 114          */
 115         vol->ondisk = ondisk = malloc(HAMMER_BUFSIZE);
 116         if (isnew) {
 117                 bzero(ondisk, HAMMER_BUFSIZE);
 118         } else {
 119                 n = pread(vol->fd, ondisk, HAMMER_BUFSIZE, 0);
 120                 if (n != HAMMER_BUFSIZE) {
 121                         err(1, "setup_volume: %s: Read failed at offset 0",
 122                             filename);
 123                 }
 124                 vol_no = ondisk->vol_no;
 125                 if (RootVolNo < 0) {
 126                         RootVolNo = ondisk->vol_rootvol;
 127                 } else if (RootVolNo != (int)ondisk->vol_rootvol) {
 128                         errx(1, "setup_volume: %s: root volume disagreement: "
 129                                 "%d vs %d",
 130                                 vol->name, RootVolNo, ondisk->vol_rootvol);
 131                 }
 132
 133                 if (bcmp(&Hammer_FSType, &ondisk->vol_fstype, sizeof(Hammer_FSType)) != 0) {
 134                         errx(1, "setup_volume: %s: Header does not indicate "
 135                                 "that this is a hammer volume", vol->name);
 136                 }
 137                 if (TAILQ_EMPTY(&VolList)) {
 138                         Hammer_FSId = vol->ondisk->vol_fsid;
 139                 } else if (bcmp(&Hammer_FSId, &ondisk->vol_fsid, sizeof(Hammer_FSId)) != 0) {
 140                         errx(1, "setup_volume: %s: FSId does match other "
 141                                 "volumes!", vol->name);
 142                 }
 143         }
 144         vol->vol_no = vol_no;
 145
 146         if (isnew) {
 147                 /*init_fifo_head(&ondisk->head, HAMMER_HEAD_TYPE_VOL);*/
 148                 vol->cache.modified = 1;
 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", filename, vol_no, scan->name);
 158                 }
 159         }
 160         TAILQ_INSERT_TAIL(&VolList, vol, entry);
 161         return(vol);
 162 }
 163
 164 struct volume_info *
 165 get_volume(int32_t vol_no)
 166 {
 167         struct volume_info *vol;
 168
 169         TAILQ_FOREACH(vol, &VolList, entry) {
 170                 if (vol->vol_no == vol_no)
 171                         break;
 172         }
 173         if (vol == NULL)
 174                 errx(1, "get_volume: Volume %d does not exist!", vol_no);
 175         ++vol->cache.refs;
 176         /* not added to or removed from hammer cache */
 177         return(vol);
 178 }
 179
 180 void
 181 rel_volume(struct volume_info *volume)
 182 {
 183         /* not added to or removed from hammer cache */
 184         --volume->cache.refs;
 185 }
 186
 187 /*
 188  * Acquire the specified buffer.
 189  */
 190 struct buffer_info *
 191 get_buffer(hammer_off_t buf_offset, int isnew)
 192 {
 193         void *ondisk;
 194         struct buffer_info *buf;
 195         struct volume_info *volume;
 196         hammer_off_t orig_offset = buf_offset;
 197         int vol_no;
 198         int zone;
 199         int hi, n;
 200
 201         zone = HAMMER_ZONE_DECODE(buf_offset);
 202         if (zone > HAMMER_ZONE_RAW_BUFFER_INDEX) {
 203                 buf_offset = blockmap_lookup(buf_offset, NULL, NULL);
 204         }
 205         assert((buf_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_RAW_BUFFER);
 206         vol_no = HAMMER_VOL_DECODE(buf_offset);
 207         volume = get_volume(vol_no);
 208         buf_offset &= ~HAMMER_BUFMASK64;
 209
 210         hi = buffer_hash(buf_offset);
 211
 212         TAILQ_FOREACH(buf, &volume->buffer_lists[hi], entry) {
 213                 if (buf->buf_offset == buf_offset)
 214                         break;
 215         }
 216         if (buf == NULL) {
 217                 buf = malloc(sizeof(*buf));
 218                 bzero(buf, sizeof(*buf));
 219                 if (DebugOpt) {
 220                         fprintf(stderr, "get_buffer %016llx %016llx\n",
 221                                 orig_offset, buf_offset);
 222                 }
 223                 buf->buf_offset = buf_offset;
 224                 buf->buf_disk_offset = volume->ondisk->vol_buf_beg +
 225                                         (buf_offset & HAMMER_OFF_SHORT_MASK);
 226                 buf->volume = volume;
 227                 TAILQ_INSERT_TAIL(&volume->buffer_lists[hi], buf, entry);
 228                 ++volume->cache.refs;
 229                 buf->cache.u.buffer = buf;
 230                 hammer_cache_add(&buf->cache, ISBUFFER);
 231         }
 232         ++buf->cache.refs;
 233         hammer_cache_flush();
 234         if ((ondisk = buf->ondisk) == NULL) {
 235                 buf->ondisk = ondisk = malloc(HAMMER_BUFSIZE);
 236                 if (isnew == 0) {
 237                         n = pread(volume->fd, ondisk, HAMMER_BUFSIZE,
 238                                   buf->buf_disk_offset);
 239                         if (n != HAMMER_BUFSIZE) {
 240                                 err(1, "get_buffer: %s:%016llx Read failed at "
 241                                        "offset %lld",
 242                                     volume->name, buf->buf_offset,
 243                                     buf->buf_disk_offset);
 244                         }
 245                 }
 246         }
 247         if (isnew) {
 248                 bzero(ondisk, HAMMER_BUFSIZE);
 249                 buf->cache.modified = 1;
 250         }
 251         return(buf);
 252 }
 253
 254 void
 255 rel_buffer(struct buffer_info *buffer)
 256 {
 257         struct volume_info *volume;
 258         int hi;
 259
 260         assert(buffer->cache.refs > 0);
 261         if (--buffer->cache.refs == 0) {
 262                 if (buffer->cache.delete) {
 263                         hi = buffer_hash(buffer->buf_offset);
 264                         volume = buffer->volume;
 265                         if (buffer->cache.modified)
 266                                 flush_buffer(buffer);
 267                         TAILQ_REMOVE(&volume->buffer_lists[hi], buffer, entry);
 268                         hammer_cache_del(&buffer->cache);
 269                         free(buffer->ondisk);
 270                         free(buffer);
 271                         rel_volume(volume);
 272                 }
 273         }
 274 }
 275
 276 void *
 277 get_buffer_data(hammer_off_t buf_offset, struct buffer_info **bufferp,
 278                 int isnew)
 279 {
 280         struct buffer_info *buffer;
 281
 282         if ((buffer = *bufferp) != NULL) {
 283                 if (isnew ||
 284                     ((buffer->buf_offset ^ buf_offset) & ~HAMMER_BUFMASK64)) {
 285                         rel_buffer(buffer);
 286                         buffer = *bufferp = NULL;
 287                 }
 288         }
 289         if (buffer == NULL)
 290                 buffer = *bufferp = get_buffer(buf_offset, isnew);
 291         return((char *)buffer->ondisk + ((int32_t)buf_offset & HAMMER_BUFMASK));
 292 }
 293
 294 /*
 295  * Retrieve a pointer to a B-Tree node given a cluster offset.  The underlying
 296  * bufp is freed if non-NULL and a referenced buffer is loaded into it.
 297  */
 298 hammer_node_ondisk_t
 299 get_node(hammer_off_t node_offset, struct buffer_info **bufp)
 300 {
 301         struct buffer_info *buf;
 302
 303         if (*bufp)
 304                 rel_buffer(*bufp);
 305         *bufp = buf = get_buffer(node_offset, 0);
 306         return((void *)((char *)buf->ondisk +
 307                         (int32_t)(node_offset & HAMMER_BUFMASK)));
 308 }
 309
 310 /*
 311  * Allocate HAMMER elements - btree nodes, data storage, and record elements
 312  *
 313  * NOTE: hammer_alloc_fifo() initializes the fifo header for the returned
 314  * item and zero's out the remainder, so don't bzero() it.
 315  */
 316 void *
 317 alloc_btree_element(hammer_off_t *offp)
 318 {
 319         struct buffer_info *buffer = NULL;
 320         hammer_node_ondisk_t node;
 321
 322         node = alloc_blockmap(HAMMER_ZONE_BTREE_INDEX, sizeof(*node),
 323                               offp, &buffer);
 324         bzero(node, sizeof(*node));
 325         /* XXX buffer not released, pointer remains valid */
 326         return(node);
 327 }
 328
 329 void *
 330 alloc_data_element(hammer_off_t *offp, int32_t data_len,
 331                    struct buffer_info **data_bufferp)
 332 {
 333         void *data;
 334
 335         if (data_len >= HAMMER_BUFSIZE) {
 336                 assert(data_len <= HAMMER_BUFSIZE); /* just one buffer */
 337                 data = alloc_blockmap(HAMMER_ZONE_LARGE_DATA_INDEX, data_len,
 338                                       offp, data_bufferp);
 339                 bzero(data, data_len);
 340         } else if (data_len) {
 341                 data = alloc_blockmap(HAMMER_ZONE_SMALL_DATA_INDEX, data_len,
 342                                       offp, data_bufferp);
 343                 bzero(data, data_len);
 344         } else {
 345                 data = NULL;
 346         }
 347         return (data);
 348 }
 349
 350 /*
 351  * Format a new freemap.  Set all layer1 entries to UNAVAIL.  The initialize
 352  * code will load each volume's freemap.
 353  */
 354 void
 355 format_freemap(struct volume_info *root_vol, hammer_blockmap_t blockmap)
 356 {
 357         struct buffer_info *buffer = NULL;
 358         hammer_off_t layer1_offset;
 359         struct hammer_blockmap_layer1 *layer1;
 360         int i, isnew;
 361
 362         layer1_offset = alloc_bigblock(root_vol, 0);
 363         for (i = 0; i < (int)HAMMER_BLOCKMAP_RADIX1; ++i) {
 364                 isnew = ((i % HAMMER_BLOCKMAP_RADIX1_PERBUFFER) == 0);
 365                 layer1 = get_buffer_data(layer1_offset + i * sizeof(*layer1),
 366                                          &buffer, isnew);
 367                 bzero(layer1, sizeof(*layer1));
 368                 layer1->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 369                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 370         }
 371         rel_buffer(buffer);
 372
 373         blockmap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 374         blockmap->phys_offset = layer1_offset;
 375         blockmap->alloc_offset = HAMMER_ENCODE_RAW_BUFFER(255, -1);
 376         blockmap->next_offset = HAMMER_ENCODE_RAW_BUFFER(0, 0);
 377         blockmap->reserved01 = 0;
 378         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 379         root_vol->cache.modified = 1;
 380 }
 381
 382 /*
 383  * Load the volume's remaining free space into the freemap.  If this is
 384  * the root volume, initialize the freemap owner for the layer1 bigblock.
 385  *
 386  * Returns the number of bigblocks available.
 387  */
 388 int64_t
 389 initialize_freemap(struct volume_info *vol)
 390 {
 391         struct volume_info *root_vol;
 392         struct buffer_info *buffer1 = NULL;
 393         struct buffer_info *buffer2 = NULL;
 394         struct hammer_blockmap_layer1 *layer1;
 395         struct hammer_blockmap_layer2 *layer2;
 396         hammer_off_t layer1_base;
 397         hammer_off_t layer1_offset;
 398         hammer_off_t layer2_offset;
 399         hammer_off_t phys_offset;
 400         hammer_off_t aligned_vol_free_end;
 401         int64_t count = 0;
 402         int modified1 = 0;
 403
 404         root_vol = get_volume(RootVolNo);
 405         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 406                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 407
 408         printf("initialize freemap volume %d\n", vol->vol_no);
 409
 410         /*
 411          * Initialize the freemap.  First preallocate the bigblocks required
 412          * to implement layer2.   This preallocation is a bootstrap allocation
 413          * using blocks from the target volume.
 414          */
 415         layer1_base = root_vol->ondisk->vol0_blockmap[
 416                                         HAMMER_ZONE_FREEMAP_INDEX].phys_offset;
 417         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 418              phys_offset < aligned_vol_free_end;
 419              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 420                 layer1_offset = layer1_base +
 421                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 422                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 423                 if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 424                         layer1->phys_offset = alloc_bigblock(vol, 0);
 425                         layer1->blocks_free = 0;
 426                         buffer1->cache.modified = 1;
 427                         layer1->layer1_crc = crc32(layer1,
 428                                                    HAMMER_LAYER1_CRCSIZE);
 429                 }
 430         }
 431
 432         /*
 433          * Now fill everything in.
 434          */
 435         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 436              phys_offset < aligned_vol_free_end;
 437              phys_offset += HAMMER_LARGEBLOCK_SIZE) {
 438                 modified1 = 0;
 439                 layer1_offset = layer1_base +
 440                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 441                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 442
 443                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 444                 layer2_offset = layer1->phys_offset +
 445                                 HAMMER_BLOCKMAP_LAYER2_OFFSET(phys_offset);
 446
 447                 layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 448                 if (phys_offset < vol->vol_free_off) {
 449                         /*
 450                          * Fixups XXX - bigblocks already allocated as part
 451                          * of the freemap bootstrap.
 452                          */
 453                         layer2->u.owner = HAMMER_ENCODE_FREEMAP(0, 0); /* XXX */
 454                 } else if (phys_offset < vol->vol_free_end) {
 455                         ++layer1->blocks_free;
 456                         buffer1->cache.modified = 1;
 457                         layer2->u.owner = HAMMER_BLOCKMAP_FREE;
 458                         ++count;
 459                         modified1 = 1;
 460                 } else {
 461                         layer2->u.owner = HAMMER_BLOCKMAP_UNAVAIL;
 462                 }
 463                 layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 464                 buffer2->cache.modified = 1;
 465
 466                 /*
 467                  * Finish-up layer 1
 468                  */
 469                 if (modified1) {
 470                         layer1->layer1_crc = crc32(layer1,
 471                                                    HAMMER_LAYER1_CRCSIZE);
 472                         buffer1->cache.modified = 1;
 473                 }
 474         }
 475         rel_buffer(buffer1);
 476         rel_buffer(buffer2);
 477         rel_volume(root_vol);
 478         return(count);
 479 }
 480
 481 /*
 482  * Allocate big-blocks using our poor-man's volume->vol_free_off and
 483  * update the freemap if owner != 0.
 484  *
 485  * An owner of 0 is used for bootstrapping the freemap.
 486  */
 487 hammer_off_t
 488 alloc_bigblock(struct volume_info *volume, hammer_off_t owner)
 489 {
 490         struct buffer_info *buffer = NULL;
 491         struct volume_info *root_vol;
 492         hammer_off_t result_offset;
 493         hammer_off_t layer_offset;
 494         struct hammer_blockmap_layer1 *layer1;
 495         struct hammer_blockmap_layer2 *layer2;
 496         int didget;
 497
 498         if (volume == NULL) {
 499                 volume = get_volume(RootVolNo);
 500                 didget = 1;
 501         } else {
 502                 didget = 0;
 503         }
 504         result_offset = volume->vol_free_off;
 505         if (result_offset >= volume->vol_free_end)
 506                 panic("alloc_bigblock: Ran out of room, filesystem too small");
 507         volume->vol_free_off += HAMMER_LARGEBLOCK_SIZE;
 508
 509         /*
 510          * Update the freemap
 511          */
 512         if (owner) {
 513                 root_vol = get_volume(RootVolNo);
 514                 layer_offset = root_vol->ondisk->vol0_blockmap[
 515                                         HAMMER_ZONE_FREEMAP_INDEX].phys_offset;
 516                 layer_offset += HAMMER_BLOCKMAP_LAYER1_OFFSET(result_offset);
 517                 layer1 = get_buffer_data(layer_offset, &buffer, 0);
 518                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 519                 --layer1->blocks_free;
 520                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 521                 buffer->cache.modified = 1;
 522                 layer_offset = layer1->phys_offset +
 523                                HAMMER_BLOCKMAP_LAYER2_OFFSET(result_offset);
 524                 layer2 = get_buffer_data(layer_offset, &buffer, 0);
 525                 assert(layer2->u.owner == HAMMER_BLOCKMAP_FREE);
 526                 layer2->u.owner = owner;
 527                 layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 528                 buffer->cache.modified = 1;
 529
 530                 --root_vol->ondisk->vol0_stat_freebigblocks;
 531                 root_vol->cache.modified = 1;
 532
 533                 rel_buffer(buffer);
 534                 rel_volume(root_vol);
 535         }
 536
 537         if (didget)
 538                 rel_volume(volume);
 539         return(result_offset);
 540 }
 541
 542 /*
 543  * Format the undo-map for the root volume.
 544  */
 545 void
 546 format_undomap(hammer_volume_ondisk_t ondisk)
 547 {
 548         const int undo_zone = HAMMER_ZONE_UNDO_INDEX;
 549         hammer_off_t undo_limit;
 550         hammer_blockmap_t blockmap;
 551         hammer_off_t scan;
 552         struct hammer_blockmap_layer2 *layer2;
 553         int n;
 554         int limit_index;
 555
 556         /*
 557          * Size the undo buffer in multiples of HAMMER_LARGEBLOCK_SIZE,
 558          * up to HAMMER_UNDO_LAYER2 large blocks.  Size to approximately
 559          * 0.1% of the disk.
 560          */
 561         undo_limit = UndoBufferSize;
 562         if (undo_limit == 0)
 563                 undo_limit = (ondisk->vol_buf_end - ondisk->vol_buf_beg) / 1000;
 564         undo_limit = (undo_limit + HAMMER_LARGEBLOCK_MASK64) &
 565                      ~HAMMER_LARGEBLOCK_MASK64;
 566         if (undo_limit < HAMMER_LARGEBLOCK_SIZE)
 567                 undo_limit = HAMMER_LARGEBLOCK_SIZE;
 568         if (undo_limit > HAMMER_LARGEBLOCK_SIZE * HAMMER_UNDO_LAYER2)
 569                 undo_limit = HAMMER_LARGEBLOCK_SIZE * HAMMER_UNDO_LAYER2;
 570         UndoBufferSize = undo_limit;
 571
 572         blockmap = &ondisk->vol0_blockmap[undo_zone];
 573         bzero(blockmap, sizeof(*blockmap));
 574         blockmap->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 575         blockmap->first_offset = HAMMER_ZONE_ENCODE(undo_zone, 0);
 576         blockmap->next_offset = blockmap->first_offset;
 577         blockmap->alloc_offset = HAMMER_ZONE_ENCODE(undo_zone, undo_limit);
 578         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 579
 580         layer2 = &ondisk->vol0_undo_array[0];
 581         n = 0;
 582         scan = blockmap->next_offset;
 583         limit_index = undo_limit / HAMMER_LARGEBLOCK_SIZE;
 584
 585         assert(limit_index <= HAMMER_UNDO_LAYER2);
 586
 587         for (n = 0; n < limit_index; ++n) {
 588                 layer2->u.phys_offset = alloc_bigblock(NULL, scan);
 589                 layer2->bytes_free = -1;        /* not used */
 590                 layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 591
 592                 scan += HAMMER_LARGEBLOCK_SIZE;
 593                 ++layer2;
 594         }
 595         while (n < HAMMER_UNDO_LAYER2) {
 596                 layer2->u.phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 597                 layer2->bytes_free = -1;
 598                 layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 599                 ++layer2;
 600                 ++n;
 601         }
 602 }
 603
 604 /*
 605  * Format a new blockmap.  Set the owner to the base of the blockmap
 606  * (meaning either the blockmap layer1 bigblock, layer2 bigblock, or
 607  * target bigblock).
 608  */
 609 void
 610 format_blockmap(hammer_blockmap_t blockmap, hammer_off_t zone_base)
 611 {
 612         blockmap->phys_offset = alloc_bigblock(NULL, zone_base);
 613         blockmap->alloc_offset = zone_base;
 614         blockmap->first_offset = zone_base;
 615         blockmap->next_offset = zone_base;
 616         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 617 }
 618
 619 /*
 620  * Allocate a chunk of data out of a blockmap.  This is a simplified
 621  * version which uses next_offset as a simple allocation iterator.
 622  */
 623 static
 624 void *
 625 alloc_blockmap(int zone, int bytes, hammer_off_t *result_offp,
 626                struct buffer_info **bufferp)
 627 {
 628         struct buffer_info *buffer1 = NULL;
 629         struct buffer_info *buffer2 = NULL;
 630         struct volume_info *volume;
 631         hammer_blockmap_t rootmap;
 632         struct hammer_blockmap_layer1 *layer1;
 633         struct hammer_blockmap_layer2 *layer2;
 634         hammer_off_t layer1_offset;
 635         hammer_off_t layer2_offset;
 636         hammer_off_t bigblock_offset;
 637         void *ptr;
 638
 639         volume = get_volume(RootVolNo);
 640
 641         rootmap = &volume->ondisk->vol0_blockmap[zone];
 642
 643         /*
 644          * Alignment and buffer-boundary issues.  If the allocation would
 645          * cross a buffer boundary we have to skip to the next buffer.
 646          */
 647         bytes = (bytes + 7) & ~7;
 648         if ((rootmap->next_offset ^ (rootmap->next_offset + bytes - 1)) &
 649             ~HAMMER_BUFMASK64) {
 650                 volume->cache.modified = 1;
 651                 rootmap->next_offset = (rootmap->next_offset + bytes) &
 652                                        ~HAMMER_BUFMASK64;
 653         }
 654
 655         /*
 656          * Dive layer 1
 657          */
 658         layer1_offset = rootmap->phys_offset +
 659                         HAMMER_BLOCKMAP_LAYER1_OFFSET(rootmap->next_offset);
 660
 661         layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 662         if ((rootmap->next_offset >= rootmap->alloc_offset &&
 663             (rootmap->next_offset & HAMMER_BLOCKMAP_LAYER2_MASK) == 0) ||
 664             layer1->phys_offset == HAMMER_BLOCKMAP_FREE
 665         ) {
 666                 assert(rootmap->next_offset >= rootmap->alloc_offset);
 667                 buffer1->cache.modified = 1;
 668                 bzero(layer1, sizeof(*layer1));
 669                 layer1->blocks_free = HAMMER_BLOCKMAP_RADIX2;
 670                 layer1->phys_offset = alloc_bigblock(NULL,
 671                                                      rootmap->next_offset);
 672         }
 673
 674         /*
 675          * Dive layer 2
 676          */
 677         layer2_offset = layer1->phys_offset +
 678                         HAMMER_BLOCKMAP_LAYER2_OFFSET(rootmap->next_offset);
 679
 680         layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 681
 682         if ((rootmap->next_offset & HAMMER_LARGEBLOCK_MASK64) == 0 &&
 683             (rootmap->next_offset >= rootmap->alloc_offset ||
 684              layer2->u.phys_offset == HAMMER_BLOCKMAP_FREE)
 685         ) {
 686                 assert(rootmap->next_offset >= rootmap->alloc_offset);
 687                 buffer2->cache.modified = 1;
 688                 bzero(layer2, sizeof(*layer2));
 689                 layer2->u.phys_offset = alloc_bigblock(NULL,
 690                                                        rootmap->next_offset);
 691                 layer2->bytes_free = HAMMER_LARGEBLOCK_SIZE;
 692                 --layer1->blocks_free;
 693         }
 694
 695         buffer1->cache.modified = 1;
 696         buffer2->cache.modified = 1;
 697         volume->cache.modified = 1;
 698         layer2->bytes_free -= bytes;
 699         *result_offp = rootmap->next_offset;
 700         rootmap->next_offset += bytes;
 701         rootmap->alloc_offset = rootmap->next_offset;
 702
 703         bigblock_offset = layer2->u.phys_offset +
 704                           (*result_offp & HAMMER_LARGEBLOCK_MASK);
 705
 706         layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 707         layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 708
 709         ptr = get_buffer_data(bigblock_offset, bufferp, 0);
 710         (*bufferp)->cache.modified = 1;
 711
 712         if (buffer1)
 713                 rel_buffer(buffer1);
 714         if (buffer2)
 715                 rel_buffer(buffer2);
 716
 717         rel_volume(volume);
 718         return(ptr);
 719 }
 720
 721 /*
 722  * Presize a blockmap.  Allocate all layer2 bigblocks required to map the
 723  * blockmap through the specified zone limit.
 724  *
 725  * Note: This code is typically called later, after some data may have
 726  *       already been allocated, but can be called or re-called at any time.
 727  *
 728  * Note: vol0_zone_limit is not zone-encoded.
 729  */
 730 void
 731 presize_blockmap(hammer_blockmap_t blockmap, hammer_off_t zone_base,
 732                  hammer_off_t vol0_zone_limit)
 733 {
 734         struct buffer_info *buffer1 = NULL;
 735         struct buffer_info *buffer2 = NULL;
 736         struct hammer_blockmap_layer1 *layer1;
 737         struct hammer_blockmap_layer2 *layer2;
 738         hammer_off_t zone_limit;
 739         hammer_off_t layer1_offset;
 740         hammer_off_t layer2_offset;
 741
 742         zone_limit = zone_base + vol0_zone_limit;
 743
 744         while (zone_base < zone_limit) {
 745                 layer1_offset = blockmap->phys_offset +
 746                         HAMMER_BLOCKMAP_LAYER1_OFFSET(zone_base);
 747                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 748
 749                 if ((zone_base >= blockmap->alloc_offset &&
 750                     (zone_base & HAMMER_BLOCKMAP_LAYER2_MASK) == 0) ||
 751                      layer1->phys_offset == HAMMER_BLOCKMAP_FREE
 752                 ) {
 753                         bzero(layer1, sizeof(*layer1));
 754                         layer1->blocks_free = HAMMER_BLOCKMAP_RADIX2;
 755                         layer1->phys_offset = alloc_bigblock(NULL, zone_base);
 756                         layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 757                         buffer1->cache.modified = 1;
 758                 }
 759                 layer2_offset = layer1->phys_offset +
 760                                 HAMMER_BLOCKMAP_LAYER2_OFFSET(zone_base);
 761                 layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 762                 if (zone_base >= blockmap->alloc_offset ||
 763                     layer2->u.phys_offset == HAMMER_BLOCKMAP_FREE) {
 764                         bzero(layer2, sizeof(*layer2));
 765                         layer2->u.phys_offset = HAMMER_BLOCKMAP_FREE;
 766                         layer2->bytes_free = HAMMER_LARGEBLOCK_SIZE;
 767                         layer2->entry_crc = crc32(layer2,
 768                                                   HAMMER_LAYER2_CRCSIZE);
 769                         buffer2->cache.modified = 1;
 770                 }
 771                 zone_base += HAMMER_LARGEBLOCK_SIZE64;
 772         }
 773         if (blockmap->alloc_offset < zone_limit)
 774                 blockmap->alloc_offset = zone_limit;
 775
 776         if (buffer1)
 777                 rel_buffer(buffer1);
 778         if (buffer2)
 779                 rel_buffer(buffer2);
 780 }
 781
 782 #if 0
 783 /*
 784  * Reserve space from the FIFO.  Make sure that bytes does not cross a
 785  * record boundary.
 786  *
 787  * Zero out base_bytes and initialize the fifo head and tail.  The
 788  * data area is not zerod.
 789  */
 790 static
 791 hammer_off_t
 792 hammer_alloc_fifo(int32_t base_bytes, int32_t ext_bytes,
 793                   struct buffer_info **bufp, u_int16_t hdr_type)
 794 {
 795         struct buffer_info *buf;
 796         struct volume_info *volume;
 797         hammer_fifo_head_t head;
 798         hammer_fifo_tail_t tail;
 799         hammer_off_t off;
 800         int32_t aligned_bytes;
 801
 802         aligned_bytes = (base_bytes + ext_bytes + HAMMER_TAIL_ONDISK_SIZE +
 803                          HAMMER_HEAD_ALIGN_MASK) & ~HAMMER_HEAD_ALIGN_MASK;
 804
 805         volume = get_volume(RootVolNo);
 806         off = volume->ondisk->vol0_fifo_end;
 807
 808         /*
 809          * For now don't deal with transitions across buffer boundaries,
 810          * only newfs_hammer uses this function.
 811          */
 812         assert((off & ~HAMMER_BUFMASK64) ==
 813                 ((off + aligned_bytes) & ~HAMMER_BUFMASK));
 814
 815         *bufp = buf = get_buffer(off, 0);
 816
 817         buf->cache.modified = 1;
 818         volume->cache.modified = 1;
 819
 820         head = (void *)((char *)buf->ondisk + ((int32_t)off & HAMMER_BUFMASK));
 821         bzero(head, base_bytes);
 822
 823         head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 824         head->hdr_type = hdr_type;
 825         head->hdr_size = aligned_bytes;
 826         head->hdr_seq = volume->ondisk->vol0_next_seq++;
 827
 828         tail = (void*)((char *)head + aligned_bytes - HAMMER_TAIL_ONDISK_SIZE);
 829         tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 830         tail->tail_type = hdr_type;
 831         tail->tail_size = aligned_bytes;
 832
 833         volume->ondisk->vol0_fifo_end += aligned_bytes;
 834         volume->cache.modified = 1;
 835
 836         rel_volume(volume);
 837
 838         return(off);
 839 }
 840
 841 #endif
 842
 843 /*
 844  * Flush various tracking structures to disk
 845  */
 846
 847 /*
 848  * Flush various tracking structures to disk
 849  */
 850 void
 851 flush_all_volumes(void)
 852 {
 853         struct volume_info *vol;
 854
 855         TAILQ_FOREACH(vol, &VolList, entry)
 856                 flush_volume(vol);
 857 }
 858
 859 void
 860 flush_volume(struct volume_info *volume)
 861 {
 862         struct buffer_info *buffer;
 863         int i;
 864
 865         for (i = 0; i < HAMMER_BUFLISTS; ++i) {
 866                 TAILQ_FOREACH(buffer, &volume->buffer_lists[i], entry)
 867                         flush_buffer(buffer);
 868         }
 869         writehammerbuf(volume, volume->ondisk, 0);
 870         volume->cache.modified = 0;
 871 }
 872
 873 void
 874 flush_buffer(struct buffer_info *buffer)
 875 {
 876         writehammerbuf(buffer->volume, buffer->ondisk, buffer->buf_disk_offset);
 877         buffer->cache.modified = 0;
 878 }
 879
 880 #if 0
 881 /*
 882  * Generic buffer initialization
 883  */
 884 static void
 885 init_fifo_head(hammer_fifo_head_t head, u_int16_t hdr_type)
 886 {
 887         head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 888         head->hdr_type = hdr_type;
 889         head->hdr_size = 0;
 890         head->hdr_crc = 0;
 891         head->hdr_seq = 0;
 892 }
 893
 894 #endif
 895
 896 #if 0
 897 /*
 898  * Core I/O operations
 899  */
 900 static void
 901 readhammerbuf(struct volume_info *vol, void *data, int64_t offset)
 902 {
 903         ssize_t n;
 904
 905         n = pread(vol->fd, data, HAMMER_BUFSIZE, offset);
 906         if (n != HAMMER_BUFSIZE)
 907                 err(1, "Read volume %d (%s)", vol->vol_no, vol->name);
 908 }
 909
 910 #endif
 911
 912 static void
 913 writehammerbuf(struct volume_info *vol, const void *data, int64_t offset)
 914 {
 915         ssize_t n;
 916
 917         n = pwrite(vol->fd, data, HAMMER_BUFSIZE, offset);
 918         if (n != HAMMER_BUFSIZE)
 919                 err(1, "Write volume %d (%s)", vol->vol_no, vol->name);
 920 }
 921
 922 void
 923 panic(const char *ctl, ...)
 924 {
 925         va_list va;
 926
 927         va_start(va, ctl);
 928         vfprintf(stderr, ctl, va);
 929         va_end(va);
 930         fprintf(stderr, "\n");
 931         exit(1);
 932 }
 933