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/types.h>
  36 #include <assert.h>
  37 #include <stdio.h>
  38 #include <stdlib.h>
  39 #include <stdarg.h>
  40 #include <string.h>
  41 #include <unistd.h>
  42 #include <stddef.h>
  43 #include <err.h>
  44 #include <fcntl.h>
  45 #include "hammer_util.h"
  46
  47 static void *alloc_blockmap(int zone, int bytes, hammer_off_t *result_offp,
  48                         struct buffer_info **bufferp);
  49 static hammer_off_t alloc_bigblock(struct volume_info *volume, int zone);
  50 static void get_buffer_readahead(struct buffer_info *base);
  51 static __inline void *get_ondisk(hammer_off_t buf_offset,
  52                         struct buffer_info **bufferp, int isnew);
  53 #if 0
  54 static void init_fifo_head(hammer_fifo_head_t head, u_int16_t hdr_type);
  55 static void readhammerbuf(struct volume_info *vol, void *data,
  56                         int64_t offset);
  57 #endif
  58 static void writehammerbuf(struct volume_info *vol, const void *data,
  59                         int64_t offset);
  60
  61 int DebugOpt;
  62
  63 uuid_t Hammer_FSType;
  64 uuid_t Hammer_FSId;
  65 int64_t BootAreaSize;
  66 int64_t MemAreaSize;
  67 int64_t UndoBufferSize;
  68 int     UsingSuperClusters;
  69 int     NumVolumes;
  70 int     RootVolNo = -1;
  71 int     UseReadBehind = -4;
  72 int     UseReadAhead = 4;
  73 int     AssertOnFailure = 1;
  74 struct volume_list VolList = TAILQ_HEAD_INITIALIZER(VolList);
  75
  76 static __inline
  77 int
  78 buffer_hash(hammer_off_t buf_offset)
  79 {
  80         int hi;
  81
  82         hi = (int)(buf_offset / HAMMER_BUFSIZE) & HAMMER_BUFLISTMASK;
  83         return(hi);
  84 }
  85
  86 /*
  87  * Lookup the requested information structure and related on-disk buffer.
  88  * Missing structures are created.
  89  */
  90 struct volume_info *
  91 setup_volume(int32_t vol_no, const char *filename, int isnew, int oflags)
  92 {
  93         struct volume_info *vol;
  94         struct volume_info *scan;
  95         struct hammer_volume_ondisk *ondisk;
  96         int i, n;
  97
  98         /*
  99          * Allocate the volume structure
 100          */
 101         vol = malloc(sizeof(*vol));
 102         bzero(vol, sizeof(*vol));
 103         for (i = 0; i < HAMMER_BUFLISTS; ++i)
 104                 TAILQ_INIT(&vol->buffer_lists[i]);
 105         vol->name = strdup(filename);
 106         vol->fd = open(filename, oflags);
 107         if (vol->fd < 0) {
 108                 free(vol->name);
 109                 free(vol);
 110                 err(1, "setup_volume: %s: Open failed", filename);
 111         }
 112
 113         /*
 114          * Read or initialize the volume header
 115          */
 116         vol->ondisk = ondisk = malloc(HAMMER_BUFSIZE);
 117         if (isnew > 0) {
 118                 bzero(ondisk, HAMMER_BUFSIZE);
 119         } else {
 120                 n = pread(vol->fd, ondisk, HAMMER_BUFSIZE, 0);
 121                 if (n != HAMMER_BUFSIZE) {
 122                         err(1, "setup_volume: %s: Read failed at offset 0",
 123                             filename);
 124                 }
 125                 vol_no = ondisk->vol_no;
 126                 if (RootVolNo < 0) {
 127                         RootVolNo = ondisk->vol_rootvol;
 128                 } else if (RootVolNo != (int)ondisk->vol_rootvol) {
 129                         errx(1, "setup_volume: %s: root volume disagreement: "
 130                                 "%d vs %d",
 131                                 vol->name, RootVolNo, ondisk->vol_rootvol);
 132                 }
 133
 134                 if (bcmp(&Hammer_FSType, &ondisk->vol_fstype, sizeof(Hammer_FSType)) != 0) {
 135                         errx(1, "setup_volume: %s: Header does not indicate "
 136                                 "that this is a hammer volume", vol->name);
 137                 }
 138                 if (TAILQ_EMPTY(&VolList)) {
 139                         Hammer_FSId = vol->ondisk->vol_fsid;
 140                 } else if (bcmp(&Hammer_FSId, &ondisk->vol_fsid, sizeof(Hammer_FSId)) != 0) {
 141                         errx(1, "setup_volume: %s: FSId does match other "
 142                                 "volumes!", vol->name);
 143                 }
 144         }
 145         vol->vol_no = vol_no;
 146
 147         if (isnew > 0) {
 148                 /*init_fifo_head(&ondisk->head, HAMMER_HEAD_TYPE_VOL);*/
 149                 vol->cache.modified = 1;
 150         }
 151
 152         /*
 153          * Link the volume structure in
 154          */
 155         TAILQ_FOREACH(scan, &VolList, entry) {
 156                 if (scan->vol_no == vol_no) {
 157                         errx(1, "setup_volume %s: Duplicate volume number %d "
 158                                 "against %s", filename, vol_no, scan->name);
 159                 }
 160         }
 161         TAILQ_INSERT_TAIL(&VolList, vol, entry);
 162         return(vol);
 163 }
 164
 165 struct volume_info *
 166 test_volume(int32_t vol_no)
 167 {
 168         struct volume_info *vol;
 169
 170         TAILQ_FOREACH(vol, &VolList, entry) {
 171                 if (vol->vol_no == vol_no)
 172                         break;
 173         }
 174         if (vol == NULL)
 175                 return(NULL);
 176         ++vol->cache.refs;
 177         /* not added to or removed from hammer cache */
 178         return(vol);
 179 }
 180
 181 struct volume_info *
 182 get_volume(int32_t vol_no)
 183 {
 184         struct volume_info *vol;
 185
 186         TAILQ_FOREACH(vol, &VolList, entry) {
 187                 if (vol->vol_no == vol_no)
 188                         break;
 189         }
 190         if (vol == NULL)
 191                 errx(1, "get_volume: Volume %d does not exist!", vol_no);
 192         ++vol->cache.refs;
 193         /* not added to or removed from hammer cache */
 194         return(vol);
 195 }
 196
 197 void
 198 rel_volume(struct volume_info *volume)
 199 {
 200         /* not added to or removed from hammer cache */
 201         --volume->cache.refs;
 202 }
 203
 204 /*
 205  * Acquire the specified buffer.
 206  */
 207 struct buffer_info *
 208 get_buffer(hammer_off_t buf_offset, int isnew)
 209 {
 210         void *ondisk;
 211         struct buffer_info *buf;
 212         struct volume_info *volume;
 213         hammer_off_t orig_offset = buf_offset;
 214         int vol_no;
 215         int zone;
 216         int hi, n;
 217         int dora = 0;
 218
 219         zone = HAMMER_ZONE_DECODE(buf_offset);
 220         if (zone > HAMMER_ZONE_RAW_BUFFER_INDEX) {
 221                 buf_offset = blockmap_lookup(buf_offset, NULL, NULL, NULL);
 222         }
 223         if (buf_offset == HAMMER_OFF_BAD)
 224                 return(NULL);
 225
 226         if (AssertOnFailure) {
 227                 assert((buf_offset & HAMMER_OFF_ZONE_MASK) ==
 228                        HAMMER_ZONE_RAW_BUFFER);
 229         }
 230         vol_no = HAMMER_VOL_DECODE(buf_offset);
 231         volume = test_volume(vol_no);
 232         if (volume == NULL) {
 233                 if (AssertOnFailure)
 234                         errx(1, "get_buffer: Volume %d not found!", vol_no);
 235                 return(NULL);
 236         }
 237
 238         buf_offset &= ~HAMMER_BUFMASK64;
 239
 240         hi = buffer_hash(buf_offset);
 241
 242         TAILQ_FOREACH(buf, &volume->buffer_lists[hi], entry) {
 243                 if (buf->buf_offset == buf_offset)
 244                         break;
 245         }
 246         if (buf == NULL) {
 247                 buf = malloc(sizeof(*buf));
 248                 bzero(buf, sizeof(*buf));
 249                 if (DebugOpt) {
 250                         fprintf(stderr, "get_buffer %016llx %016llx\n",
 251                                 (long long)orig_offset, (long long)buf_offset);
 252                 }
 253                 buf->buf_offset = buf_offset;
 254                 buf->raw_offset = volume->ondisk->vol_buf_beg +
 255                                   (buf_offset & HAMMER_OFF_SHORT_MASK);
 256                 buf->volume = volume;
 257                 TAILQ_INSERT_TAIL(&volume->buffer_lists[hi], buf, entry);
 258                 ++volume->cache.refs;
 259                 buf->cache.u.buffer = buf;
 260                 hammer_cache_add(&buf->cache, ISBUFFER);
 261                 dora = (isnew == 0);
 262                 if (isnew < 0)
 263                         buf->flags |= HAMMER_BUFINFO_READAHEAD;
 264         } else {
 265                 if (isnew >= 0) {
 266                         buf->flags &= ~HAMMER_BUFINFO_READAHEAD;
 267                         hammer_cache_used(&buf->cache);
 268                 }
 269                 ++buf->use_count;
 270         }
 271         ++buf->cache.refs;
 272         hammer_cache_flush();
 273         if ((ondisk = buf->ondisk) == NULL) {
 274                 buf->ondisk = ondisk = malloc(HAMMER_BUFSIZE);
 275                 if (isnew <= 0) {
 276                         n = pread(volume->fd, ondisk, HAMMER_BUFSIZE,
 277                                   buf->raw_offset);
 278                         if (n != HAMMER_BUFSIZE) {
 279                                 if (AssertOnFailure)
 280                                         err(1, "get_buffer: %s:%016llx "
 281                                             "Read failed at offset %016llx",
 282                                             volume->name,
 283                                             (long long)buf->buf_offset,
 284                                             (long long)buf->raw_offset);
 285                                 bzero(ondisk, HAMMER_BUFSIZE);
 286                         }
 287                 }
 288         }
 289         if (isnew > 0) {
 290                 bzero(ondisk, HAMMER_BUFSIZE);
 291                 buf->cache.modified = 1;
 292         }
 293         if (dora)
 294                 get_buffer_readahead(buf);
 295         return(buf);
 296 }
 297
 298 static void
 299 get_buffer_readahead(struct buffer_info *base)
 300 {
 301         struct buffer_info *buf;
 302         struct volume_info *vol;
 303         hammer_off_t buf_offset;
 304         int64_t raw_offset;
 305         int ri = UseReadBehind;
 306         int re = UseReadAhead;
 307         int hi;
 308
 309         raw_offset = base->raw_offset + ri * HAMMER_BUFSIZE;
 310         vol = base->volume;
 311
 312         while (ri < re) {
 313                 if (raw_offset >= vol->ondisk->vol_buf_end)
 314                         break;
 315                 if (raw_offset < vol->ondisk->vol_buf_beg) {
 316                         ++ri;
 317                         raw_offset += HAMMER_BUFSIZE;
 318                         continue;
 319                 }
 320                 buf_offset = HAMMER_VOL_ENCODE(vol->vol_no) |
 321                              HAMMER_ZONE_RAW_BUFFER |
 322                              (raw_offset - vol->ondisk->vol_buf_beg);
 323                 hi = buffer_hash(raw_offset);
 324                 TAILQ_FOREACH(buf, &vol->buffer_lists[hi], entry) {
 325                         if (buf->raw_offset == raw_offset)
 326                                 break;
 327                 }
 328                 if (buf == NULL) {
 329                         buf = get_buffer(buf_offset, -1);
 330                         rel_buffer(buf);
 331                 }
 332                 ++ri;
 333                 raw_offset += HAMMER_BUFSIZE;
 334         }
 335 }
 336
 337 void
 338 rel_buffer(struct buffer_info *buffer)
 339 {
 340         struct volume_info *volume;
 341         int hi;
 342
 343         assert(buffer->cache.refs > 0);
 344         if (--buffer->cache.refs == 0) {
 345                 if (buffer->cache.delete) {
 346                         hi = buffer_hash(buffer->buf_offset);
 347                         volume = buffer->volume;
 348                         if (buffer->cache.modified)
 349                                 flush_buffer(buffer);
 350                         TAILQ_REMOVE(&volume->buffer_lists[hi], buffer, entry);
 351                         hammer_cache_del(&buffer->cache);
 352                         free(buffer->ondisk);
 353                         free(buffer);
 354                         rel_volume(volume);
 355                 }
 356         }
 357 }
 358
 359 /*
 360  * Retrieve a pointer to a buffer data given a buffer offset.  The underlying
 361  * bufferp is freed if isnew or the offset is out of range of the cached data.
 362  * If bufferp is freed a referenced buffer is loaded into it.
 363  */
 364 void *
 365 get_buffer_data(hammer_off_t buf_offset, struct buffer_info **bufferp,
 366                 int isnew)
 367 {
 368         if (*bufferp != NULL) {
 369                 if (isnew > 0 ||
 370                     (((*bufferp)->buf_offset ^ buf_offset) & ~HAMMER_BUFMASK64)) {
 371                         rel_buffer(*bufferp);
 372                         *bufferp = NULL;
 373                 }
 374         }
 375         return(get_ondisk(buf_offset, bufferp, isnew));
 376 }
 377
 378 /*
 379  * Retrieve a pointer to a B-Tree node given a cluster offset.  The underlying
 380  * bufferp is freed if non-NULL and a referenced buffer is loaded into it.
 381  */
 382 hammer_node_ondisk_t
 383 get_node(hammer_off_t node_offset, struct buffer_info **bufferp)
 384 {
 385         if (*bufferp != NULL) {
 386                 rel_buffer(*bufferp);
 387                 *bufferp = NULL;
 388         }
 389         return(get_ondisk(node_offset, bufferp, 0));
 390 }
 391
 392 /*
 393  * Return a pointer to a buffer data given a buffer offset.
 394  * If *bufferp is NULL acquire the buffer otherwise use that buffer.
 395  */
 396 static __inline
 397 void *
 398 get_ondisk(hammer_off_t buf_offset, struct buffer_info **bufferp,
 399         int isnew)
 400 {
 401         struct buffer_info *buffer;
 402
 403         buffer = *bufferp;
 404         if (buffer == NULL) {
 405                 buffer = *bufferp = get_buffer(buf_offset, isnew);
 406                 if (buffer == NULL)
 407                         return(NULL);
 408         }
 409
 410         return((char *)buffer->ondisk +
 411                 ((int32_t)buf_offset & HAMMER_BUFMASK));
 412 }
 413
 414 /*
 415  * Allocate HAMMER elements - btree nodes, data storage
 416  */
 417 void *
 418 alloc_btree_element(hammer_off_t *offp)
 419 {
 420         struct buffer_info *buffer = NULL;
 421         hammer_node_ondisk_t node;
 422
 423         node = alloc_blockmap(HAMMER_ZONE_BTREE_INDEX, sizeof(*node),
 424                               offp, &buffer);
 425         bzero(node, sizeof(*node));
 426         /* XXX buffer not released, pointer remains valid */
 427         return(node);
 428 }
 429
 430 void *
 431 alloc_data_element(hammer_off_t *offp, int32_t data_len,
 432                    struct buffer_info **data_bufferp)
 433 {
 434         void *data;
 435
 436         if (data_len >= HAMMER_BUFSIZE) {
 437                 assert(data_len <= HAMMER_BUFSIZE); /* just one buffer */
 438                 data = alloc_blockmap(HAMMER_ZONE_LARGE_DATA_INDEX, data_len,
 439                                       offp, data_bufferp);
 440                 bzero(data, data_len);
 441         } else if (data_len) {
 442                 data = alloc_blockmap(HAMMER_ZONE_SMALL_DATA_INDEX, data_len,
 443                                       offp, data_bufferp);
 444                 bzero(data, data_len);
 445         } else {
 446                 data = NULL;
 447         }
 448         return (data);
 449 }
 450
 451 /*
 452  * Format a new freemap.  Set all layer1 entries to UNAVAIL.  The initialize
 453  * code will load each volume's freemap.
 454  */
 455 void
 456 format_freemap(struct volume_info *root_vol, hammer_blockmap_t blockmap)
 457 {
 458         struct buffer_info *buffer = NULL;
 459         hammer_off_t layer1_offset;
 460         struct hammer_blockmap_layer1 *layer1;
 461         int i, isnew;
 462
 463         layer1_offset = alloc_bigblock(root_vol, HAMMER_ZONE_FREEMAP_INDEX);
 464         for (i = 0; i < (int)HAMMER_BLOCKMAP_RADIX1; ++i) {
 465                 isnew = ((i % HAMMER_BLOCKMAP_RADIX1_PERBUFFER) == 0);
 466                 layer1 = get_buffer_data(layer1_offset + i * sizeof(*layer1),
 467                                          &buffer, isnew);
 468                 bzero(layer1, sizeof(*layer1));
 469                 layer1->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 470                 layer1->blocks_free = 0;
 471                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 472         }
 473         rel_buffer(buffer);
 474
 475         blockmap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 476         blockmap->phys_offset = layer1_offset;
 477         blockmap->alloc_offset = HAMMER_ENCODE_RAW_BUFFER(255, -1);
 478         blockmap->next_offset = HAMMER_ENCODE_RAW_BUFFER(0, 0);
 479         blockmap->reserved01 = 0;
 480         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 481         root_vol->cache.modified = 1;
 482 }
 483
 484 /*
 485  * Load the volume's remaining free space into the freemap.
 486  *
 487  * Returns the number of bigblocks available.
 488  */
 489 int64_t
 490 initialize_freemap(struct volume_info *vol)
 491 {
 492         struct volume_info *root_vol;
 493         struct buffer_info *buffer1 = NULL;
 494         struct buffer_info *buffer2 = NULL;
 495         struct hammer_blockmap_layer1 *layer1;
 496         struct hammer_blockmap_layer2 *layer2;
 497         hammer_off_t layer1_base;
 498         hammer_off_t layer1_offset;
 499         hammer_off_t layer2_offset;
 500         hammer_off_t phys_offset;
 501         hammer_off_t aligned_vol_free_end;
 502         int64_t count = 0;
 503         int modified1 = 0;
 504
 505         root_vol = get_volume(RootVolNo);
 506         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 507                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 508
 509         printf("initialize freemap volume %d\n", vol->vol_no);
 510
 511         /*
 512          * Initialize the freemap.  First preallocate the bigblocks required
 513          * to implement layer2.   This preallocation is a bootstrap allocation
 514          * using blocks from the target volume.
 515          */
 516         layer1_base = root_vol->ondisk->vol0_blockmap[
 517                                         HAMMER_ZONE_FREEMAP_INDEX].phys_offset;
 518         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 519              phys_offset < aligned_vol_free_end;
 520              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 521                 layer1_offset = layer1_base +
 522                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 523                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 524                 if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 525                         layer1->phys_offset = alloc_bigblock(vol,
 526                                                 HAMMER_ZONE_FREEMAP_INDEX);
 527                         layer1->blocks_free = 0;
 528                         buffer1->cache.modified = 1;
 529                         layer1->layer1_crc = crc32(layer1,
 530                                                    HAMMER_LAYER1_CRCSIZE);
 531                 }
 532         }
 533
 534         /*
 535          * Now fill everything in.
 536          */
 537         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 538              phys_offset < aligned_vol_free_end;
 539              phys_offset += HAMMER_BIGBLOCK_SIZE) {
 540                 modified1 = 0;
 541                 layer1_offset = layer1_base +
 542                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 543                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 544
 545                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 546                 layer2_offset = layer1->phys_offset +
 547                                 HAMMER_BLOCKMAP_LAYER2_OFFSET(phys_offset);
 548
 549                 layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 550                 bzero(layer2, sizeof(*layer2));
 551                 if (phys_offset < vol->vol_free_off) {
 552                         /*
 553                          * Fixups XXX - bigblocks already allocated as part
 554                          * of the freemap bootstrap.
 555                          */
 556                         if (layer2->zone == 0) {
 557                                 layer2->zone = HAMMER_ZONE_FREEMAP_INDEX;
 558                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 559                                 layer2->bytes_free = 0;
 560                         }
 561                 } else if (phys_offset < vol->vol_free_end) {
 562                         ++layer1->blocks_free;
 563                         buffer1->cache.modified = 1;
 564                         layer2->zone = 0;
 565                         layer2->append_off = 0;
 566                         layer2->bytes_free = HAMMER_BIGBLOCK_SIZE;
 567                         ++count;
 568                         modified1 = 1;
 569                 } else {
 570                         layer2->zone = HAMMER_ZONE_UNAVAIL_INDEX;
 571                         layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 572                         layer2->bytes_free = 0;
 573                 }
 574                 layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 575                 buffer2->cache.modified = 1;
 576
 577                 /*
 578                  * Finish-up layer 1
 579                  */
 580                 if (modified1) {
 581                         layer1->layer1_crc = crc32(layer1,
 582                                                    HAMMER_LAYER1_CRCSIZE);
 583                         buffer1->cache.modified = 1;
 584                 }
 585         }
 586         rel_buffer(buffer1);
 587         rel_buffer(buffer2);
 588         rel_volume(root_vol);
 589         return(count);
 590 }
 591
 592 /*
 593  * Allocate big-blocks using our poor-man's volume->vol_free_off.
 594  *
 595  * If the zone is HAMMER_ZONE_FREEMAP_INDEX we are bootstrapping the freemap
 596  * itself and cannot update it yet.
 597  */
 598 hammer_off_t
 599 alloc_bigblock(struct volume_info *volume, int zone)
 600 {
 601         struct buffer_info *buffer = NULL;
 602         struct volume_info *root_vol;
 603         hammer_off_t result_offset;
 604         hammer_off_t layer_offset;
 605         struct hammer_blockmap_layer1 *layer1;
 606         struct hammer_blockmap_layer2 *layer2;
 607         int didget;
 608
 609         if (volume == NULL) {
 610                 volume = get_volume(RootVolNo);
 611                 didget = 1;
 612         } else {
 613                 didget = 0;
 614         }
 615         result_offset = volume->vol_free_off;
 616         if (result_offset >= volume->vol_free_end)
 617                 panic("alloc_bigblock: Ran out of room, filesystem too small");
 618         volume->vol_free_off += HAMMER_BIGBLOCK_SIZE;
 619
 620         /*
 621          * Update the freemap.
 622          */
 623         if (zone != HAMMER_ZONE_FREEMAP_INDEX) {
 624                 root_vol = get_volume(RootVolNo);
 625                 layer_offset = root_vol->ondisk->vol0_blockmap[
 626                                         HAMMER_ZONE_FREEMAP_INDEX].phys_offset;
 627                 layer_offset += HAMMER_BLOCKMAP_LAYER1_OFFSET(result_offset);
 628                 layer1 = get_buffer_data(layer_offset, &buffer, 0);
 629                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 630                 --layer1->blocks_free;
 631                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 632                 buffer->cache.modified = 1;
 633                 layer_offset = layer1->phys_offset +
 634                                HAMMER_BLOCKMAP_LAYER2_OFFSET(result_offset);
 635                 layer2 = get_buffer_data(layer_offset, &buffer, 0);
 636                 assert(layer2->zone == 0);
 637                 layer2->zone = zone;
 638                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 639                 layer2->bytes_free = 0;
 640                 layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 641                 buffer->cache.modified = 1;
 642
 643                 --root_vol->ondisk->vol0_stat_freebigblocks;
 644                 root_vol->cache.modified = 1;
 645
 646                 rel_buffer(buffer);
 647                 rel_volume(root_vol);
 648         }
 649
 650         if (didget)
 651                 rel_volume(volume);
 652         return(result_offset);
 653 }
 654
 655 /*
 656  * Format the undo-map for the root volume.
 657  */
 658 void
 659 format_undomap(hammer_volume_ondisk_t ondisk)
 660 {
 661         const int undo_zone = HAMMER_ZONE_UNDO_INDEX;
 662         hammer_off_t undo_limit;
 663         hammer_blockmap_t blockmap;
 664         struct buffer_info *buffer = NULL;
 665         hammer_off_t scan;
 666         int n;
 667         int limit_index;
 668         u_int32_t seqno;
 669
 670         /*
 671          * Size the undo buffer in multiples of HAMMER_BIGBLOCK_SIZE,
 672          * up to HAMMER_UNDO_LAYER2 big blocks.  Size to approximately
 673          * 0.1% of the disk.
 674          *
 675          * The minimum UNDO fifo size is 500MB, or approximately 1% of
 676          * the recommended 50G disk.
 677          *
 678          * Changing this minimum is rather dangerous as complex filesystem
 679          * operations can cause the UNDO FIFO to fill up otherwise.
 680          */
 681         undo_limit = UndoBufferSize;
 682         if (undo_limit == 0) {
 683                 undo_limit = (ondisk->vol_buf_end - ondisk->vol_buf_beg) / 1000;
 684                 if (undo_limit < 500*1024*1024)
 685                         undo_limit = 500*1024*1024;
 686         }
 687         undo_limit = (undo_limit + HAMMER_BIGBLOCK_MASK64) &
 688                      ~HAMMER_BIGBLOCK_MASK64;
 689         if (undo_limit < HAMMER_BIGBLOCK_SIZE)
 690                 undo_limit = HAMMER_BIGBLOCK_SIZE;
 691         if (undo_limit > HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2)
 692                 undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2;
 693         UndoBufferSize = undo_limit;
 694
 695         blockmap = &ondisk->vol0_blockmap[undo_zone];
 696         bzero(blockmap, sizeof(*blockmap));
 697         blockmap->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 698         blockmap->first_offset = HAMMER_ZONE_ENCODE(undo_zone, 0);
 699         blockmap->next_offset = blockmap->first_offset;
 700         blockmap->alloc_offset = HAMMER_ZONE_ENCODE(undo_zone, undo_limit);
 701         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 702
 703         n = 0;
 704         scan = blockmap->next_offset;
 705         limit_index = undo_limit / HAMMER_BIGBLOCK_SIZE;
 706
 707         assert(limit_index <= HAMMER_UNDO_LAYER2);
 708
 709         for (n = 0; n < limit_index; ++n) {
 710                 ondisk->vol0_undo_array[n] = alloc_bigblock(NULL,
 711                                                         HAMMER_ZONE_UNDO_INDEX);
 712                 scan += HAMMER_BIGBLOCK_SIZE;
 713         }
 714         while (n < HAMMER_UNDO_LAYER2) {
 715                 ondisk->vol0_undo_array[n] = HAMMER_BLOCKMAP_UNAVAIL;
 716                 ++n;
 717         }
 718
 719         /*
 720          * Pre-initialize the UNDO blocks (HAMMER version 4+)
 721          */
 722         printf("initializing the undo map (%jd MB)\n",
 723                 (intmax_t)(blockmap->alloc_offset & HAMMER_OFF_LONG_MASK) /
 724                 (1024 * 1024));
 725
 726         scan = blockmap->first_offset;
 727         seqno = 0;
 728
 729         while (scan < blockmap->alloc_offset) {
 730                 hammer_fifo_head_t head;
 731                 hammer_fifo_tail_t tail;
 732                 int isnew;
 733                 int bytes = HAMMER_UNDO_ALIGN;
 734
 735                 isnew = ((scan & HAMMER_BUFMASK64) == 0);
 736                 head = get_buffer_data(scan, &buffer, isnew);
 737                 buffer->cache.modified = 1;
 738                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 739
 740                 bzero(head, bytes);
 741                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 742                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
 743                 head->hdr_size = bytes;
 744                 head->hdr_seq = seqno++;
 745
 746                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 747                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 748                 tail->tail_size = bytes;
 749
 750                 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
 751                                 crc32(head + 1, bytes - sizeof(*head));
 752
 753                 scan += bytes;
 754         }
 755         if (buffer)
 756                 rel_buffer(buffer);
 757 }
 758
 759 /*
 760  * Format a new blockmap.  This is mostly a degenerate case because
 761  * all allocations are now actually done from the freemap.
 762  */
 763 void
 764 format_blockmap(hammer_blockmap_t blockmap, hammer_off_t zone_base)
 765 {
 766         blockmap->phys_offset = 0;
 767         blockmap->alloc_offset = zone_base | HAMMER_VOL_ENCODE(255) |
 768                                  HAMMER_SHORT_OFF_ENCODE(-1);
 769         blockmap->first_offset = zone_base;
 770         blockmap->next_offset = zone_base;
 771         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 772 }
 773
 774 /*
 775  * Allocate a chunk of data out of a blockmap.  This is a simplified
 776  * version which uses next_offset as a simple allocation iterator.
 777  */
 778 static
 779 void *
 780 alloc_blockmap(int zone, int bytes, hammer_off_t *result_offp,
 781                struct buffer_info **bufferp)
 782 {
 783         struct buffer_info *buffer1 = NULL;
 784         struct buffer_info *buffer2 = NULL;
 785         struct volume_info *volume;
 786         hammer_blockmap_t blockmap;
 787         hammer_blockmap_t freemap;
 788         struct hammer_blockmap_layer1 *layer1;
 789         struct hammer_blockmap_layer2 *layer2;
 790         hammer_off_t layer1_offset;
 791         hammer_off_t layer2_offset;
 792         hammer_off_t zone2_offset;
 793         void *ptr;
 794
 795         volume = get_volume(RootVolNo);
 796
 797         blockmap = &volume->ondisk->vol0_blockmap[zone];
 798         freemap = &volume->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 799
 800         /*
 801          * Alignment and buffer-boundary issues.  If the allocation would
 802          * cross a buffer boundary we have to skip to the next buffer.
 803          */
 804         bytes = (bytes + 15) & ~15;
 805
 806 again:
 807         if ((blockmap->next_offset ^ (blockmap->next_offset + bytes - 1)) &
 808             ~HAMMER_BUFMASK64) {
 809                 volume->cache.modified = 1;
 810                 blockmap->next_offset = (blockmap->next_offset + bytes) &
 811                                         ~HAMMER_BUFMASK64;
 812         }
 813
 814         /*
 815          * Dive layer 1.  For now we can't allocate data outside of volume 0.
 816          */
 817         layer1_offset = freemap->phys_offset +
 818                         HAMMER_BLOCKMAP_LAYER1_OFFSET(blockmap->next_offset);
 819
 820         layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 821
 822         if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 823                 fprintf(stderr, "alloc_blockmap: ran out of space!\n");
 824                 exit(1);
 825         }
 826
 827         /*
 828          * Dive layer 2
 829          */
 830         layer2_offset = layer1->phys_offset +
 831                         HAMMER_BLOCKMAP_LAYER2_OFFSET(blockmap->next_offset);
 832
 833         layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 834
 835         if (layer2->zone == HAMMER_ZONE_UNAVAIL_INDEX) {
 836                 fprintf(stderr, "alloc_blockmap: ran out of space!\n");
 837                 exit(1);
 838         }
 839
 840         /*
 841          * If we are entering a new bigblock assign ownership to our
 842          * zone.  If the bigblock is owned by another zone skip it.
 843          */
 844         if (layer2->zone == 0) {
 845                 --layer1->blocks_free;
 846                 layer2->zone = zone;
 847                 assert(layer2->bytes_free == HAMMER_BIGBLOCK_SIZE);
 848                 assert(layer2->append_off == 0);
 849         }
 850         if (layer2->zone != zone) {
 851                 blockmap->next_offset = (blockmap->next_offset + HAMMER_BIGBLOCK_SIZE) &
 852                                         ~HAMMER_BIGBLOCK_MASK64;
 853                 goto again;
 854         }
 855
 856         buffer1->cache.modified = 1;
 857         buffer2->cache.modified = 1;
 858         volume->cache.modified = 1;
 859         assert(layer2->append_off ==
 860                (blockmap->next_offset & HAMMER_BIGBLOCK_MASK));
 861         layer2->bytes_free -= bytes;
 862         *result_offp = blockmap->next_offset;
 863         blockmap->next_offset += bytes;
 864         layer2->append_off = (int)blockmap->next_offset &
 865                               HAMMER_BIGBLOCK_MASK;
 866
 867         layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 868         layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 869
 870         zone2_offset = (*result_offp & ~HAMMER_OFF_ZONE_MASK) |
 871                         HAMMER_ZONE_ENCODE(zone, 0);
 872
 873         ptr = get_buffer_data(zone2_offset, bufferp, 0);
 874         (*bufferp)->cache.modified = 1;
 875
 876         if (buffer1)
 877                 rel_buffer(buffer1);
 878         if (buffer2)
 879                 rel_buffer(buffer2);
 880
 881         rel_volume(volume);
 882         return(ptr);
 883 }
 884
 885 /*
 886  * Flush various tracking structures to disk
 887  */
 888
 889 /*
 890  * Flush various tracking structures to disk
 891  */
 892 void
 893 flush_all_volumes(void)
 894 {
 895         struct volume_info *vol;
 896
 897         TAILQ_FOREACH(vol, &VolList, entry)
 898                 flush_volume(vol);
 899 }
 900
 901 void
 902 flush_volume(struct volume_info *volume)
 903 {
 904         struct buffer_info *buffer;
 905         int i;
 906
 907         for (i = 0; i < HAMMER_BUFLISTS; ++i) {
 908                 TAILQ_FOREACH(buffer, &volume->buffer_lists[i], entry)
 909                         flush_buffer(buffer);
 910         }
 911         writehammerbuf(volume, volume->ondisk, 0);
 912         volume->cache.modified = 0;
 913 }
 914
 915 void
 916 flush_buffer(struct buffer_info *buffer)
 917 {
 918         writehammerbuf(buffer->volume, buffer->ondisk, buffer->raw_offset);
 919         buffer->cache.modified = 0;
 920 }
 921
 922 #if 0
 923 /*
 924  * Generic buffer initialization
 925  */
 926 static void
 927 init_fifo_head(hammer_fifo_head_t head, u_int16_t hdr_type)
 928 {
 929         head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 930         head->hdr_type = hdr_type;
 931         head->hdr_size = 0;
 932         head->hdr_crc = 0;
 933         head->hdr_seq = 0;
 934 }
 935
 936 #endif
 937
 938 #if 0
 939 /*
 940  * Core I/O operations
 941  */
 942 static void
 943 readhammerbuf(struct volume_info *vol, void *data, int64_t offset)
 944 {
 945         ssize_t n;
 946
 947         n = pread(vol->fd, data, HAMMER_BUFSIZE, offset);
 948         if (n != HAMMER_BUFSIZE)
 949                 err(1, "Read volume %d (%s)", vol->vol_no, vol->name);
 950 }
 951
 952 #endif
 953
 954 static void
 955 writehammerbuf(struct volume_info *vol, const void *data, int64_t offset)
 956 {
 957         ssize_t n;
 958
 959         n = pwrite(vol->fd, data, HAMMER_BUFSIZE, offset);
 960         if (n != HAMMER_BUFSIZE)
 961                 err(1, "Write volume %d (%s)", vol->vol_no, vol->name);
 962 }
 963
 964 void
 965 panic(const char *ctl, ...)
 966 {
 967         va_list va;
 968
 969         va_start(va, ctl);
 970         vfprintf(stderr, ctl, va);
 971         va_end(va);
 972         fprintf(stderr, "\n");
 973         exit(1);
 974 }
 975