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 1.25 2008/08/21 23:28:43 thomas 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, int zone);
  52 static void get_buffer_readahead(struct buffer_info *base);
  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 int     UseReadBehind = -4;
  74 int     UseReadAhead = 4;
  75 int     AssertOnFailure = 1;
  76 struct volume_list VolList = TAILQ_HEAD_INITIALIZER(VolList);
  77
  78 static __inline
  79 int
  80 buffer_hash(hammer_off_t buf_offset)
  81 {
  82         int hi;
  83
  84         hi = (int)(buf_offset / HAMMER_BUFSIZE) & HAMMER_BUFLISTMASK;
  85         return(hi);
  86 }
  87
  88 /*
  89  * Lookup the requested information structure and related on-disk buffer.
  90  * Missing structures are created.
  91  */
  92 struct volume_info *
  93 setup_volume(int32_t vol_no, const char *filename, int isnew, int oflags)
  94 {
  95         struct volume_info *vol;
  96         struct volume_info *scan;
  97         struct hammer_volume_ondisk *ondisk;
  98         int i, n;
  99
 100         /*
 101          * Allocate the volume structure
 102          */
 103         vol = malloc(sizeof(*vol));
 104         bzero(vol, sizeof(*vol));
 105         for (i = 0; i < HAMMER_BUFLISTS; ++i)
 106                 TAILQ_INIT(&vol->buffer_lists[i]);
 107         vol->name = strdup(filename);
 108         vol->fd = open(filename, oflags);
 109         if (vol->fd < 0) {
 110                 free(vol->name);
 111                 free(vol);
 112                 err(1, "setup_volume: %s: Open failed", filename);
 113         }
 114
 115         /*
 116          * Read or initialize the volume header
 117          */
 118         vol->ondisk = ondisk = malloc(HAMMER_BUFSIZE);
 119         if (isnew > 0) {
 120                 bzero(ondisk, HAMMER_BUFSIZE);
 121         } else {
 122                 n = pread(vol->fd, ondisk, HAMMER_BUFSIZE, 0);
 123                 if (n != HAMMER_BUFSIZE) {
 124                         err(1, "setup_volume: %s: Read failed at offset 0",
 125                             filename);
 126                 }
 127                 vol_no = ondisk->vol_no;
 128                 if (RootVolNo < 0) {
 129                         RootVolNo = ondisk->vol_rootvol;
 130                 } else if (RootVolNo != (int)ondisk->vol_rootvol) {
 131                         errx(1, "setup_volume: %s: root volume disagreement: "
 132                                 "%d vs %d",
 133                                 vol->name, RootVolNo, ondisk->vol_rootvol);
 134                 }
 135
 136                 if (bcmp(&Hammer_FSType, &ondisk->vol_fstype, sizeof(Hammer_FSType)) != 0) {
 137                         errx(1, "setup_volume: %s: Header does not indicate "
 138                                 "that this is a hammer volume", vol->name);
 139                 }
 140                 if (TAILQ_EMPTY(&VolList)) {
 141                         Hammer_FSId = vol->ondisk->vol_fsid;
 142                 } else if (bcmp(&Hammer_FSId, &ondisk->vol_fsid, sizeof(Hammer_FSId)) != 0) {
 143                         errx(1, "setup_volume: %s: FSId does match other "
 144                                 "volumes!", vol->name);
 145                 }
 146         }
 147         vol->vol_no = vol_no;
 148
 149         if (isnew > 0) {
 150                 /*init_fifo_head(&ondisk->head, HAMMER_HEAD_TYPE_VOL);*/
 151                 vol->cache.modified = 1;
 152         }
 153
 154         /*
 155          * Link the volume structure in
 156          */
 157         TAILQ_FOREACH(scan, &VolList, entry) {
 158                 if (scan->vol_no == vol_no) {
 159                         errx(1, "setup_volume %s: Duplicate volume number %d "
 160                                 "against %s", filename, vol_no, scan->name);
 161                 }
 162         }
 163         TAILQ_INSERT_TAIL(&VolList, vol, entry);
 164         return(vol);
 165 }
 166
 167 struct volume_info *
 168 get_volume(int32_t vol_no)
 169 {
 170         struct volume_info *vol;
 171
 172         TAILQ_FOREACH(vol, &VolList, entry) {
 173                 if (vol->vol_no == vol_no)
 174                         break;
 175         }
 176         if (vol == NULL)
 177                 errx(1, "get_volume: Volume %d does not exist!", vol_no);
 178         ++vol->cache.refs;
 179         /* not added to or removed from hammer cache */
 180         return(vol);
 181 }
 182
 183 void
 184 rel_volume(struct volume_info *volume)
 185 {
 186         /* not added to or removed from hammer cache */
 187         --volume->cache.refs;
 188 }
 189
 190 /*
 191  * Acquire the specified buffer.
 192  */
 193 struct buffer_info *
 194 get_buffer(hammer_off_t buf_offset, int isnew)
 195 {
 196         void *ondisk;
 197         struct buffer_info *buf;
 198         struct volume_info *volume;
 199         hammer_off_t orig_offset = buf_offset;
 200         int vol_no;
 201         int zone;
 202         int hi, n;
 203         int dora = 0;
 204
 205         zone = HAMMER_ZONE_DECODE(buf_offset);
 206         if (zone > HAMMER_ZONE_RAW_BUFFER_INDEX) {
 207                 buf_offset = blockmap_lookup(buf_offset, NULL, NULL, NULL);
 208         }
 209         if (buf_offset == HAMMER_OFF_BAD)
 210                 return(NULL);
 211
 212         if (AssertOnFailure) {
 213                 assert((buf_offset & HAMMER_OFF_ZONE_MASK) ==
 214                        HAMMER_ZONE_RAW_BUFFER);
 215         }
 216         vol_no = HAMMER_VOL_DECODE(buf_offset);
 217         volume = get_volume(vol_no);
 218         buf_offset &= ~HAMMER_BUFMASK64;
 219
 220         hi = buffer_hash(buf_offset);
 221
 222         TAILQ_FOREACH(buf, &volume->buffer_lists[hi], entry) {
 223                 if (buf->buf_offset == buf_offset)
 224                         break;
 225         }
 226         if (buf == NULL) {
 227                 buf = malloc(sizeof(*buf));
 228                 bzero(buf, sizeof(*buf));
 229                 if (DebugOpt) {
 230                         fprintf(stderr, "get_buffer %016llx %016llx\n",
 231                                 (long long)orig_offset, (long long)buf_offset);
 232                 }
 233                 buf->buf_offset = buf_offset;
 234                 buf->raw_offset = volume->ondisk->vol_buf_beg +
 235                                   (buf_offset & HAMMER_OFF_SHORT_MASK);
 236                 buf->volume = volume;
 237                 TAILQ_INSERT_TAIL(&volume->buffer_lists[hi], buf, entry);
 238                 ++volume->cache.refs;
 239                 buf->cache.u.buffer = buf;
 240                 hammer_cache_add(&buf->cache, ISBUFFER);
 241                 dora = (isnew == 0);
 242                 if (isnew < 0)
 243                         buf->flags |= HAMMER_BUFINFO_READAHEAD;
 244         } else {
 245                 if (isnew >= 0) {
 246                         buf->flags &= ~HAMMER_BUFINFO_READAHEAD;
 247                         hammer_cache_used(&buf->cache);
 248                 }
 249                 ++buf->use_count;
 250         }
 251         ++buf->cache.refs;
 252         hammer_cache_flush();
 253         if ((ondisk = buf->ondisk) == NULL) {
 254                 buf->ondisk = ondisk = malloc(HAMMER_BUFSIZE);
 255                 if (isnew <= 0) {
 256                         n = pread(volume->fd, ondisk, HAMMER_BUFSIZE,
 257                                   buf->raw_offset);
 258                         if (n != HAMMER_BUFSIZE) {
 259                                 if (AssertOnFailure)
 260                                 err(1, "get_buffer: %s:%016llx Read failed at "
 261                                        "offset %016llx",
 262                                     volume->name,
 263                                     (long long)buf->buf_offset,
 264                                     (long long)buf->raw_offset);
 265                                 bzero(ondisk, HAMMER_BUFSIZE);
 266                         }
 267                 }
 268         }
 269         if (isnew > 0) {
 270                 bzero(ondisk, HAMMER_BUFSIZE);
 271                 buf->cache.modified = 1;
 272         }
 273         if (dora)
 274                 get_buffer_readahead(buf);
 275         return(buf);
 276 }
 277
 278 static void
 279 get_buffer_readahead(struct buffer_info *base)
 280 {
 281         struct buffer_info *buf;
 282         struct volume_info *vol;
 283         hammer_off_t buf_offset;
 284         int64_t raw_offset;
 285         int ri = UseReadBehind;
 286         int re = UseReadAhead;
 287         int hi;
 288
 289         raw_offset = base->raw_offset + ri * HAMMER_BUFSIZE;
 290         vol = base->volume;
 291
 292         while (ri < re) {
 293                 if (raw_offset >= vol->ondisk->vol_buf_end)
 294                         break;
 295                 if (raw_offset < vol->ondisk->vol_buf_beg) {
 296                         ++ri;
 297                         raw_offset += HAMMER_BUFSIZE;
 298                         continue;
 299                 }
 300                 buf_offset = HAMMER_VOL_ENCODE(vol->vol_no) |
 301                              HAMMER_ZONE_RAW_BUFFER |
 302                              (raw_offset - vol->ondisk->vol_buf_beg);
 303                 hi = buffer_hash(raw_offset);
 304                 TAILQ_FOREACH(buf, &vol->buffer_lists[hi], entry) {
 305                         if (buf->raw_offset == raw_offset)
 306                                 break;
 307                 }
 308                 if (buf == NULL) {
 309                         buf = get_buffer(buf_offset, -1);
 310                         rel_buffer(buf);
 311                 }
 312                 ++ri;
 313                 raw_offset += HAMMER_BUFSIZE;
 314         }
 315 }
 316
 317 void
 318 rel_buffer(struct buffer_info *buffer)
 319 {
 320         struct volume_info *volume;
 321         int hi;
 322
 323         assert(buffer->cache.refs > 0);
 324         if (--buffer->cache.refs == 0) {
 325                 if (buffer->cache.delete) {
 326                         hi = buffer_hash(buffer->buf_offset);
 327                         volume = buffer->volume;
 328                         if (buffer->cache.modified)
 329                                 flush_buffer(buffer);
 330                         TAILQ_REMOVE(&volume->buffer_lists[hi], buffer, entry);
 331                         hammer_cache_del(&buffer->cache);
 332                         free(buffer->ondisk);
 333                         free(buffer);
 334                         rel_volume(volume);
 335                 }
 336         }
 337 }
 338
 339 void *
 340 get_buffer_data(hammer_off_t buf_offset, struct buffer_info **bufferp,
 341                 int isnew)
 342 {
 343         struct buffer_info *buffer;
 344
 345         if ((buffer = *bufferp) != NULL) {
 346                 if (isnew > 0 ||
 347                     ((buffer->buf_offset ^ buf_offset) & ~HAMMER_BUFMASK64)) {
 348                         rel_buffer(buffer);
 349                         buffer = *bufferp = NULL;
 350                 }
 351         }
 352         if (buffer == NULL)
 353                 buffer = *bufferp = get_buffer(buf_offset, isnew);
 354         return((char *)buffer->ondisk + ((int32_t)buf_offset & HAMMER_BUFMASK));
 355 }
 356
 357 /*
 358  * Retrieve a pointer to a B-Tree node given a cluster offset.  The underlying
 359  * bufp is freed if non-NULL and a referenced buffer is loaded into it.
 360  */
 361 hammer_node_ondisk_t
 362 get_node(hammer_off_t node_offset, struct buffer_info **bufp)
 363 {
 364         struct buffer_info *buf;
 365
 366         if (*bufp)
 367                 rel_buffer(*bufp);
 368         *bufp = buf = get_buffer(node_offset, 0);
 369         if (buf) {
 370                 return((void *)((char *)buf->ondisk +
 371                                 (int32_t)(node_offset & HAMMER_BUFMASK)));
 372         } else {
 373                 return(NULL);
 374         }
 375 }
 376
 377 /*
 378  * Allocate HAMMER elements - btree nodes, data storage, and record elements
 379  *
 380  * NOTE: hammer_alloc_fifo() initializes the fifo header for the returned
 381  * item and zero's out the remainder, so don't bzero() it.
 382  */
 383 void *
 384 alloc_btree_element(hammer_off_t *offp)
 385 {
 386         struct buffer_info *buffer = NULL;
 387         hammer_node_ondisk_t node;
 388
 389         node = alloc_blockmap(HAMMER_ZONE_BTREE_INDEX, sizeof(*node),
 390                               offp, &buffer);
 391         bzero(node, sizeof(*node));
 392         /* XXX buffer not released, pointer remains valid */
 393         return(node);
 394 }
 395
 396 void *
 397 alloc_data_element(hammer_off_t *offp, int32_t data_len,
 398                    struct buffer_info **data_bufferp)
 399 {
 400         void *data;
 401
 402         if (data_len >= HAMMER_BUFSIZE) {
 403                 assert(data_len <= HAMMER_BUFSIZE); /* just one buffer */
 404                 data = alloc_blockmap(HAMMER_ZONE_LARGE_DATA_INDEX, data_len,
 405                                       offp, data_bufferp);
 406                 bzero(data, data_len);
 407         } else if (data_len) {
 408                 data = alloc_blockmap(HAMMER_ZONE_SMALL_DATA_INDEX, data_len,
 409                                       offp, data_bufferp);
 410                 bzero(data, data_len);
 411         } else {
 412                 data = NULL;
 413         }
 414         return (data);
 415 }
 416
 417 /*
 418  * Format a new freemap.  Set all layer1 entries to UNAVAIL.  The initialize
 419  * code will load each volume's freemap.
 420  */
 421 void
 422 format_freemap(struct volume_info *root_vol, hammer_blockmap_t blockmap)
 423 {
 424         struct buffer_info *buffer = NULL;
 425         hammer_off_t layer1_offset;
 426         struct hammer_blockmap_layer1 *layer1;
 427         int i, isnew;
 428
 429         layer1_offset = alloc_bigblock(root_vol, HAMMER_ZONE_FREEMAP_INDEX);
 430         for (i = 0; i < (int)HAMMER_BLOCKMAP_RADIX1; ++i) {
 431                 isnew = ((i % HAMMER_BLOCKMAP_RADIX1_PERBUFFER) == 0);
 432                 layer1 = get_buffer_data(layer1_offset + i * sizeof(*layer1),
 433                                          &buffer, isnew);
 434                 bzero(layer1, sizeof(*layer1));
 435                 layer1->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 436                 layer1->blocks_free = 0;
 437                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 438         }
 439         rel_buffer(buffer);
 440
 441         blockmap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 442         blockmap->phys_offset = layer1_offset;
 443         blockmap->alloc_offset = HAMMER_ENCODE_RAW_BUFFER(255, -1);
 444         blockmap->next_offset = HAMMER_ENCODE_RAW_BUFFER(0, 0);
 445         blockmap->reserved01 = 0;
 446         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 447         root_vol->cache.modified = 1;
 448 }
 449
 450 /*
 451  * Load the volume's remaining free space into the freemap.
 452  *
 453  * Returns the number of bigblocks available.
 454  */
 455 int64_t
 456 initialize_freemap(struct volume_info *vol)
 457 {
 458         struct volume_info *root_vol;
 459         struct buffer_info *buffer1 = NULL;
 460         struct buffer_info *buffer2 = NULL;
 461         struct hammer_blockmap_layer1 *layer1;
 462         struct hammer_blockmap_layer2 *layer2;
 463         hammer_off_t layer1_base;
 464         hammer_off_t layer1_offset;
 465         hammer_off_t layer2_offset;
 466         hammer_off_t phys_offset;
 467         hammer_off_t aligned_vol_free_end;
 468         int64_t count = 0;
 469         int modified1 = 0;
 470
 471         root_vol = get_volume(RootVolNo);
 472         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 473                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 474
 475         printf("initialize freemap volume %d\n", vol->vol_no);
 476
 477         /*
 478          * Initialize the freemap.  First preallocate the bigblocks required
 479          * to implement layer2.   This preallocation is a bootstrap allocation
 480          * using blocks from the target volume.
 481          */
 482         layer1_base = root_vol->ondisk->vol0_blockmap[
 483                                         HAMMER_ZONE_FREEMAP_INDEX].phys_offset;
 484         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 485              phys_offset < aligned_vol_free_end;
 486              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 487                 layer1_offset = layer1_base +
 488                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 489                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 490                 if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 491                         layer1->phys_offset = alloc_bigblock(vol,
 492                                                 HAMMER_ZONE_FREEMAP_INDEX);
 493                         layer1->blocks_free = 0;
 494                         buffer1->cache.modified = 1;
 495                         layer1->layer1_crc = crc32(layer1,
 496                                                    HAMMER_LAYER1_CRCSIZE);
 497                 }
 498         }
 499
 500         /*
 501          * Now fill everything in.
 502          */
 503         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 504              phys_offset < aligned_vol_free_end;
 505              phys_offset += HAMMER_LARGEBLOCK_SIZE) {
 506                 modified1 = 0;
 507                 layer1_offset = layer1_base +
 508                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 509                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 510
 511                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 512                 layer2_offset = layer1->phys_offset +
 513                                 HAMMER_BLOCKMAP_LAYER2_OFFSET(phys_offset);
 514
 515                 layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 516                 bzero(layer2, sizeof(*layer2));
 517                 if (phys_offset < vol->vol_free_off) {
 518                         /*
 519                          * Fixups XXX - bigblocks already allocated as part
 520                          * of the freemap bootstrap.
 521                          */
 522                         if (layer2->zone == 0) {
 523                                 layer2->zone = HAMMER_ZONE_FREEMAP_INDEX;
 524                                 layer2->append_off = HAMMER_LARGEBLOCK_SIZE;
 525                                 layer2->bytes_free = 0;
 526                         }
 527                 } else if (phys_offset < vol->vol_free_end) {
 528                         ++layer1->blocks_free;
 529                         buffer1->cache.modified = 1;
 530                         layer2->zone = 0;
 531                         layer2->append_off = 0;
 532                         layer2->bytes_free = HAMMER_LARGEBLOCK_SIZE;
 533                         ++count;
 534                         modified1 = 1;
 535                 } else {
 536                         layer2->zone = HAMMER_ZONE_UNAVAIL_INDEX;
 537                         layer2->append_off = HAMMER_LARGEBLOCK_SIZE;
 538                         layer2->bytes_free = 0;
 539                 }
 540                 layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 541                 buffer2->cache.modified = 1;
 542
 543                 /*
 544                  * Finish-up layer 1
 545                  */
 546                 if (modified1) {
 547                         layer1->layer1_crc = crc32(layer1,
 548                                                    HAMMER_LAYER1_CRCSIZE);
 549                         buffer1->cache.modified = 1;
 550                 }
 551         }
 552         rel_buffer(buffer1);
 553         rel_buffer(buffer2);
 554         rel_volume(root_vol);
 555         return(count);
 556 }
 557
 558 /*
 559  * Allocate big-blocks using our poor-man's volume->vol_free_off.
 560  *
 561  * If the zone is HAMMER_ZONE_FREEMAP_INDEX we are bootstrapping the freemap
 562  * itself and cannot update it yet.
 563  */
 564 hammer_off_t
 565 alloc_bigblock(struct volume_info *volume, int zone)
 566 {
 567         struct buffer_info *buffer = NULL;
 568         struct volume_info *root_vol;
 569         hammer_off_t result_offset;
 570         hammer_off_t layer_offset;
 571         struct hammer_blockmap_layer1 *layer1;
 572         struct hammer_blockmap_layer2 *layer2;
 573         int didget;
 574
 575         if (volume == NULL) {
 576                 volume = get_volume(RootVolNo);
 577                 didget = 1;
 578         } else {
 579                 didget = 0;
 580         }
 581         result_offset = volume->vol_free_off;
 582         if (result_offset >= volume->vol_free_end)
 583                 panic("alloc_bigblock: Ran out of room, filesystem too small");
 584         volume->vol_free_off += HAMMER_LARGEBLOCK_SIZE;
 585
 586         /*
 587          * Update the freemap.
 588          */
 589         if (zone != HAMMER_ZONE_FREEMAP_INDEX) {
 590                 root_vol = get_volume(RootVolNo);
 591                 layer_offset = root_vol->ondisk->vol0_blockmap[
 592                                         HAMMER_ZONE_FREEMAP_INDEX].phys_offset;
 593                 layer_offset += HAMMER_BLOCKMAP_LAYER1_OFFSET(result_offset);
 594                 layer1 = get_buffer_data(layer_offset, &buffer, 0);
 595                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 596                 --layer1->blocks_free;
 597                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 598                 buffer->cache.modified = 1;
 599                 layer_offset = layer1->phys_offset +
 600                                HAMMER_BLOCKMAP_LAYER2_OFFSET(result_offset);
 601                 layer2 = get_buffer_data(layer_offset, &buffer, 0);
 602                 assert(layer2->zone == 0);
 603                 layer2->zone = zone;
 604                 layer2->append_off = HAMMER_LARGEBLOCK_SIZE;
 605                 layer2->bytes_free = 0;
 606                 layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 607                 buffer->cache.modified = 1;
 608
 609                 --root_vol->ondisk->vol0_stat_freebigblocks;
 610                 root_vol->cache.modified = 1;
 611
 612                 rel_buffer(buffer);
 613                 rel_volume(root_vol);
 614         }
 615
 616         if (didget)
 617                 rel_volume(volume);
 618         return(result_offset);
 619 }
 620
 621 /*
 622  * Format the undo-map for the root volume.
 623  */
 624 void
 625 format_undomap(hammer_volume_ondisk_t ondisk)
 626 {
 627         const int undo_zone = HAMMER_ZONE_UNDO_INDEX;
 628         hammer_off_t undo_limit;
 629         hammer_blockmap_t blockmap;
 630         struct buffer_info *buffer = NULL;
 631         hammer_off_t scan;
 632         int n;
 633         int limit_index;
 634         u_int32_t seqno;
 635
 636         /*
 637          * Size the undo buffer in multiples of HAMMER_LARGEBLOCK_SIZE,
 638          * up to HAMMER_UNDO_LAYER2 large blocks.  Size to approximately
 639          * 0.1% of the disk.
 640          *
 641          * The minimum UNDO fifo size is 100MB.
 642          */
 643         undo_limit = UndoBufferSize;
 644         if (undo_limit == 0) {
 645                 undo_limit = (ondisk->vol_buf_end - ondisk->vol_buf_beg) / 1000;
 646                 if (undo_limit < 100*1024*1024)
 647                         undo_limit = 100*1024*1024;
 648         }
 649         undo_limit = (undo_limit + HAMMER_LARGEBLOCK_MASK64) &
 650                      ~HAMMER_LARGEBLOCK_MASK64;
 651         if (undo_limit < HAMMER_LARGEBLOCK_SIZE)
 652                 undo_limit = HAMMER_LARGEBLOCK_SIZE;
 653         if (undo_limit > HAMMER_LARGEBLOCK_SIZE * HAMMER_UNDO_LAYER2)
 654                 undo_limit = HAMMER_LARGEBLOCK_SIZE * HAMMER_UNDO_LAYER2;
 655         UndoBufferSize = undo_limit;
 656
 657         blockmap = &ondisk->vol0_blockmap[undo_zone];
 658         bzero(blockmap, sizeof(*blockmap));
 659         blockmap->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 660         blockmap->first_offset = HAMMER_ZONE_ENCODE(undo_zone, 0);
 661         blockmap->next_offset = blockmap->first_offset;
 662         blockmap->alloc_offset = HAMMER_ZONE_ENCODE(undo_zone, undo_limit);
 663         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 664
 665         n = 0;
 666         scan = blockmap->next_offset;
 667         limit_index = undo_limit / HAMMER_LARGEBLOCK_SIZE;
 668
 669         assert(limit_index <= HAMMER_UNDO_LAYER2);
 670
 671         for (n = 0; n < limit_index; ++n) {
 672                 ondisk->vol0_undo_array[n] = alloc_bigblock(NULL,
 673                                                         HAMMER_ZONE_UNDO_INDEX);
 674                 scan += HAMMER_LARGEBLOCK_SIZE;
 675         }
 676         while (n < HAMMER_UNDO_LAYER2) {
 677                 ondisk->vol0_undo_array[n] = HAMMER_BLOCKMAP_UNAVAIL;
 678                 ++n;
 679         }
 680
 681         /*
 682          * Pre-initialize the UNDO blocks (HAMMER version 4+)
 683          */
 684         printf("initializing the undo map (%jd MB)\n",
 685                 (intmax_t)(blockmap->alloc_offset & HAMMER_OFF_LONG_MASK) /
 686                 (1024 * 1024));
 687
 688         scan = blockmap->first_offset;
 689         seqno = 0;
 690
 691         while (scan < blockmap->alloc_offset) {
 692                 hammer_fifo_head_t head;
 693                 hammer_fifo_tail_t tail;
 694                 int isnew;
 695                 int bytes = HAMMER_UNDO_ALIGN;
 696
 697                 isnew = ((scan & HAMMER_BUFMASK64) == 0);
 698                 head = get_buffer_data(scan, &buffer, isnew);
 699                 buffer->cache.modified = 1;
 700                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 701
 702                 bzero(head, bytes);
 703                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 704                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
 705                 head->hdr_size = bytes;
 706                 head->hdr_seq = seqno++;
 707
 708                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 709                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 710                 tail->tail_size = bytes;
 711
 712                 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
 713                                 crc32(head + 1, bytes - sizeof(*head));
 714
 715                 scan += bytes;
 716         }
 717         if (buffer)
 718                 rel_buffer(buffer);
 719 }
 720
 721 /*
 722  * Format a new blockmap.  This is mostly a degenerate case because
 723  * all allocations are now actually done from the freemap.
 724  */
 725 void
 726 format_blockmap(hammer_blockmap_t blockmap, hammer_off_t zone_base)
 727 {
 728         blockmap->phys_offset = 0;
 729         blockmap->alloc_offset = zone_base | HAMMER_VOL_ENCODE(255) |
 730                                  HAMMER_SHORT_OFF_ENCODE(-1);
 731         blockmap->first_offset = zone_base;
 732         blockmap->next_offset = zone_base;
 733         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 734 }
 735
 736 /*
 737  * Allocate a chunk of data out of a blockmap.  This is a simplified
 738  * version which uses next_offset as a simple allocation iterator.
 739  */
 740 static
 741 void *
 742 alloc_blockmap(int zone, int bytes, hammer_off_t *result_offp,
 743                struct buffer_info **bufferp)
 744 {
 745         struct buffer_info *buffer1 = NULL;
 746         struct buffer_info *buffer2 = NULL;
 747         struct volume_info *volume;
 748         hammer_blockmap_t blockmap;
 749         hammer_blockmap_t freemap;
 750         struct hammer_blockmap_layer1 *layer1;
 751         struct hammer_blockmap_layer2 *layer2;
 752         hammer_off_t layer1_offset;
 753         hammer_off_t layer2_offset;
 754         hammer_off_t zone2_offset;
 755         void *ptr;
 756
 757         volume = get_volume(RootVolNo);
 758
 759         blockmap = &volume->ondisk->vol0_blockmap[zone];
 760         freemap = &volume->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 761
 762         /*
 763          * Alignment and buffer-boundary issues.  If the allocation would
 764          * cross a buffer boundary we have to skip to the next buffer.
 765          */
 766         bytes = (bytes + 15) & ~15;
 767
 768 again:
 769         if ((blockmap->next_offset ^ (blockmap->next_offset + bytes - 1)) &
 770             ~HAMMER_BUFMASK64) {
 771                 volume->cache.modified = 1;
 772                 blockmap->next_offset = (blockmap->next_offset + bytes) &
 773                                         ~HAMMER_BUFMASK64;
 774         }
 775
 776         /*
 777          * Dive layer 1.  For now we can't allocate data outside of volume 0.
 778          */
 779         layer1_offset = freemap->phys_offset +
 780                         HAMMER_BLOCKMAP_LAYER1_OFFSET(blockmap->next_offset);
 781
 782         layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 783
 784         if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 785                 fprintf(stderr, "alloc_blockmap: ran out of space!\n");
 786                 exit(1);
 787         }
 788
 789         /*
 790          * Dive layer 2
 791          */
 792         layer2_offset = layer1->phys_offset +
 793                         HAMMER_BLOCKMAP_LAYER2_OFFSET(blockmap->next_offset);
 794
 795         layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 796
 797         if (layer2->zone == HAMMER_ZONE_UNAVAIL_INDEX) {
 798                 fprintf(stderr, "alloc_blockmap: ran out of space!\n");
 799                 exit(1);
 800         }
 801
 802         /*
 803          * If we are entering a new bigblock assign ownership to our
 804          * zone.  If the bigblock is owned by another zone skip it.
 805          */
 806         if (layer2->zone == 0) {
 807                 --layer1->blocks_free;
 808                 layer2->zone = zone;
 809                 assert(layer2->bytes_free == HAMMER_LARGEBLOCK_SIZE);
 810                 assert(layer2->append_off == 0);
 811         }
 812         if (layer2->zone != zone) {
 813                 blockmap->next_offset = (blockmap->next_offset + HAMMER_LARGEBLOCK_SIZE) &
 814                                         ~HAMMER_LARGEBLOCK_MASK64;
 815                 goto again;
 816         }
 817
 818         buffer1->cache.modified = 1;
 819         buffer2->cache.modified = 1;
 820         volume->cache.modified = 1;
 821         assert(layer2->append_off ==
 822                (blockmap->next_offset & HAMMER_LARGEBLOCK_MASK));
 823         layer2->bytes_free -= bytes;
 824         *result_offp = blockmap->next_offset;
 825         blockmap->next_offset += bytes;
 826         layer2->append_off = (int)blockmap->next_offset &
 827                               HAMMER_LARGEBLOCK_MASK;
 828
 829         layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 830         layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 831
 832         zone2_offset = (*result_offp & ~HAMMER_OFF_ZONE_MASK) |
 833                         HAMMER_ZONE_ENCODE(zone, 0);
 834
 835         ptr = get_buffer_data(zone2_offset, bufferp, 0);
 836         (*bufferp)->cache.modified = 1;
 837
 838         if (buffer1)
 839                 rel_buffer(buffer1);
 840         if (buffer2)
 841                 rel_buffer(buffer2);
 842
 843         rel_volume(volume);
 844         return(ptr);
 845 }
 846
 847 /*
 848  * Flush various tracking structures to disk
 849  */
 850
 851 /*
 852  * Flush various tracking structures to disk
 853  */
 854 void
 855 flush_all_volumes(void)
 856 {
 857         struct volume_info *vol;
 858
 859         TAILQ_FOREACH(vol, &VolList, entry)
 860                 flush_volume(vol);
 861 }
 862
 863 void
 864 flush_volume(struct volume_info *volume)
 865 {
 866         struct buffer_info *buffer;
 867         int i;
 868
 869         for (i = 0; i < HAMMER_BUFLISTS; ++i) {
 870                 TAILQ_FOREACH(buffer, &volume->buffer_lists[i], entry)
 871                         flush_buffer(buffer);
 872         }
 873         writehammerbuf(volume, volume->ondisk, 0);
 874         volume->cache.modified = 0;
 875 }
 876
 877 void
 878 flush_buffer(struct buffer_info *buffer)
 879 {
 880         writehammerbuf(buffer->volume, buffer->ondisk, buffer->raw_offset);
 881         buffer->cache.modified = 0;
 882 }
 883
 884 #if 0
 885 /*
 886  * Generic buffer initialization
 887  */
 888 static void
 889 init_fifo_head(hammer_fifo_head_t head, u_int16_t hdr_type)
 890 {
 891         head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 892         head->hdr_type = hdr_type;
 893         head->hdr_size = 0;
 894         head->hdr_crc = 0;
 895         head->hdr_seq = 0;
 896 }
 897
 898 #endif
 899
 900 #if 0
 901 /*
 902  * Core I/O operations
 903  */
 904 static void
 905 readhammerbuf(struct volume_info *vol, void *data, int64_t offset)
 906 {
 907         ssize_t n;
 908
 909         n = pread(vol->fd, data, HAMMER_BUFSIZE, offset);
 910         if (n != HAMMER_BUFSIZE)
 911                 err(1, "Read volume %d (%s)", vol->vol_no, vol->name);
 912 }
 913
 914 #endif
 915
 916 static void
 917 writehammerbuf(struct volume_info *vol, const void *data, int64_t offset)
 918 {
 919         ssize_t n;
 920
 921         n = pwrite(vol->fd, data, HAMMER_BUFSIZE, offset);
 922         if (n != HAMMER_BUFSIZE)
 923                 err(1, "Write volume %d (%s)", vol->vol_no, vol->name);
 924 }
 925
 926 void
 927 panic(const char *ctl, ...)
 928 {
 929         va_list va;
 930
 931         va_start(va, ctl);
 932         vfprintf(stderr, ctl, va);
 933         va_end(va);
 934         fprintf(stderr, "\n");
 935         exit(1);
 936 }
 937