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         if (volume == NULL)
 201                 return;
 202         /* not added to or removed from hammer cache */
 203         --volume->cache.refs;
 204 }
 205
 206 /*
 207  * Acquire the specified buffer.
 208  */
 209 struct buffer_info *
 210 get_buffer(hammer_off_t buf_offset, int isnew)
 211 {
 212         void *ondisk;
 213         struct buffer_info *buf;
 214         struct volume_info *volume;
 215         hammer_off_t orig_offset = buf_offset;
 216         int vol_no;
 217         int zone;
 218         int hi, n;
 219         int dora = 0;
 220
 221         zone = HAMMER_ZONE_DECODE(buf_offset);
 222         if (zone > HAMMER_ZONE_RAW_BUFFER_INDEX) {
 223                 buf_offset = blockmap_lookup(buf_offset, NULL, NULL, NULL);
 224         }
 225         if (buf_offset == HAMMER_OFF_BAD)
 226                 return(NULL);
 227
 228         if (AssertOnFailure) {
 229                 assert((buf_offset & HAMMER_OFF_ZONE_MASK) ==
 230                        HAMMER_ZONE_RAW_BUFFER);
 231         }
 232         vol_no = HAMMER_VOL_DECODE(buf_offset);
 233         volume = test_volume(vol_no);
 234         if (volume == NULL) {
 235                 if (AssertOnFailure)
 236                         errx(1, "get_buffer: Volume %d not found!", vol_no);
 237                 return(NULL);
 238         }
 239
 240         buf_offset &= ~HAMMER_BUFMASK64;
 241
 242         hi = buffer_hash(buf_offset);
 243
 244         TAILQ_FOREACH(buf, &volume->buffer_lists[hi], entry) {
 245                 if (buf->buf_offset == buf_offset)
 246                         break;
 247         }
 248         if (buf == NULL) {
 249                 buf = malloc(sizeof(*buf));
 250                 bzero(buf, sizeof(*buf));
 251                 if (DebugOpt) {
 252                         fprintf(stderr, "get_buffer: %016llx %016llx at %p\n",
 253                                 (long long)orig_offset, (long long)buf_offset,
 254                                 buf);
 255                 }
 256                 buf->buf_offset = buf_offset;
 257                 buf->raw_offset = volume->ondisk->vol_buf_beg +
 258                                   (buf_offset & HAMMER_OFF_SHORT_MASK);
 259                 buf->volume = volume;
 260                 TAILQ_INSERT_TAIL(&volume->buffer_lists[hi], buf, entry);
 261                 ++volume->cache.refs;
 262                 buf->cache.u.buffer = buf;
 263                 hammer_cache_add(&buf->cache, ISBUFFER);
 264                 dora = (isnew == 0);
 265                 if (isnew < 0)
 266                         buf->flags |= HAMMER_BUFINFO_READAHEAD;
 267         } else {
 268                 if (DebugOpt) {
 269                         fprintf(stderr, "get_buffer: %016llx %016llx at %p *\n",
 270                                 (long long)orig_offset, (long long)buf_offset,
 271                                 buf);
 272                 }
 273                 if (isnew >= 0) {
 274                         buf->flags &= ~HAMMER_BUFINFO_READAHEAD;
 275                         hammer_cache_used(&buf->cache);
 276                 }
 277                 ++buf->use_count;
 278         }
 279         ++buf->cache.refs;
 280         hammer_cache_flush();
 281         if ((ondisk = buf->ondisk) == NULL) {
 282                 buf->ondisk = ondisk = malloc(HAMMER_BUFSIZE);
 283                 if (isnew <= 0) {
 284                         n = pread(volume->fd, ondisk, HAMMER_BUFSIZE,
 285                                   buf->raw_offset);
 286                         if (n != HAMMER_BUFSIZE) {
 287                                 if (AssertOnFailure)
 288                                         err(1, "get_buffer: %s:%016llx "
 289                                             "Read failed at offset %016llx",
 290                                             volume->name,
 291                                             (long long)buf->buf_offset,
 292                                             (long long)buf->raw_offset);
 293                                 bzero(ondisk, HAMMER_BUFSIZE);
 294                         }
 295                 }
 296         }
 297         if (isnew > 0) {
 298                 bzero(ondisk, HAMMER_BUFSIZE);
 299                 buf->cache.modified = 1;
 300         }
 301         if (dora)
 302                 get_buffer_readahead(buf);
 303         return(buf);
 304 }
 305
 306 static void
 307 get_buffer_readahead(struct buffer_info *base)
 308 {
 309         struct buffer_info *buf;
 310         struct volume_info *vol;
 311         hammer_off_t buf_offset;
 312         int64_t raw_offset;
 313         int ri = UseReadBehind;
 314         int re = UseReadAhead;
 315         int hi;
 316
 317         raw_offset = base->raw_offset + ri * HAMMER_BUFSIZE;
 318         vol = base->volume;
 319
 320         while (ri < re) {
 321                 if (raw_offset >= vol->ondisk->vol_buf_end)
 322                         break;
 323                 if (raw_offset < vol->ondisk->vol_buf_beg) {
 324                         ++ri;
 325                         raw_offset += HAMMER_BUFSIZE;
 326                         continue;
 327                 }
 328                 buf_offset = HAMMER_VOL_ENCODE(vol->vol_no) |
 329                              HAMMER_ZONE_RAW_BUFFER |
 330                              (raw_offset - vol->ondisk->vol_buf_beg);
 331                 hi = buffer_hash(raw_offset);
 332                 TAILQ_FOREACH(buf, &vol->buffer_lists[hi], entry) {
 333                         if (buf->raw_offset == raw_offset)
 334                                 break;
 335                 }
 336                 if (buf == NULL) {
 337                         buf = get_buffer(buf_offset, -1);
 338                         rel_buffer(buf);
 339                 }
 340                 ++ri;
 341                 raw_offset += HAMMER_BUFSIZE;
 342         }
 343 }
 344
 345 void
 346 rel_buffer(struct buffer_info *buffer)
 347 {
 348         struct volume_info *volume;
 349         int hi;
 350
 351         if (buffer == NULL)
 352                 return;
 353         assert(buffer->cache.refs > 0);
 354         if (--buffer->cache.refs == 0) {
 355                 if (buffer->cache.delete) {
 356                         hi = buffer_hash(buffer->buf_offset);
 357                         volume = buffer->volume;
 358                         if (buffer->cache.modified)
 359                                 flush_buffer(buffer);
 360                         TAILQ_REMOVE(&volume->buffer_lists[hi], buffer, entry);
 361                         hammer_cache_del(&buffer->cache);
 362                         free(buffer->ondisk);
 363                         free(buffer);
 364                         rel_volume(volume);
 365                 }
 366         }
 367 }
 368
 369 /*
 370  * Retrieve a pointer to a buffer data given a buffer offset.  The underlying
 371  * bufferp is freed if isnew or the offset is out of range of the cached data.
 372  * If bufferp is freed a referenced buffer is loaded into it.
 373  */
 374 void *
 375 get_buffer_data(hammer_off_t buf_offset, struct buffer_info **bufferp,
 376                 int isnew)
 377 {
 378         if (*bufferp != NULL) {
 379                 if (isnew > 0 ||
 380                     (((*bufferp)->buf_offset ^ buf_offset) & ~HAMMER_BUFMASK64)) {
 381                         rel_buffer(*bufferp);
 382                         *bufferp = NULL;
 383                 }
 384         }
 385         return(get_ondisk(buf_offset, bufferp, isnew));
 386 }
 387
 388 /*
 389  * Retrieve a pointer to a B-Tree node given a cluster offset.  The underlying
 390  * bufferp is freed if non-NULL and a referenced buffer is loaded into it.
 391  */
 392 hammer_node_ondisk_t
 393 get_node(hammer_off_t node_offset, struct buffer_info **bufferp)
 394 {
 395         if (*bufferp != NULL) {
 396                 rel_buffer(*bufferp);
 397                 *bufferp = NULL;
 398         }
 399         return(get_ondisk(node_offset, bufferp, 0));
 400 }
 401
 402 /*
 403  * Return a pointer to a buffer data given a buffer offset.
 404  * If *bufferp is NULL acquire the buffer otherwise use that buffer.
 405  */
 406 static __inline
 407 void *
 408 get_ondisk(hammer_off_t buf_offset, struct buffer_info **bufferp,
 409         int isnew)
 410 {
 411         struct buffer_info *buffer;
 412
 413         buffer = *bufferp;
 414         if (buffer == NULL) {
 415                 buffer = *bufferp = get_buffer(buf_offset, isnew);
 416                 if (buffer == NULL)
 417                         return(NULL);
 418         }
 419
 420         return((char *)buffer->ondisk +
 421                 ((int32_t)buf_offset & HAMMER_BUFMASK));
 422 }
 423
 424 /*
 425  * Allocate HAMMER elements - btree nodes, meta data, data storage
 426  */
 427 void *
 428 alloc_btree_element(hammer_off_t *offp)
 429 {
 430         struct buffer_info *buffer = NULL;
 431         hammer_node_ondisk_t node;
 432
 433         node = alloc_blockmap(HAMMER_ZONE_BTREE_INDEX, sizeof(*node),
 434                               offp, &buffer);
 435         bzero(node, sizeof(*node));
 436         /* XXX buffer not released, pointer remains valid */
 437         return(node);
 438 }
 439
 440 void *
 441 alloc_meta_element(hammer_off_t *offp, int32_t data_len,
 442                    struct buffer_info **data_bufferp)
 443 {
 444         void *data;
 445
 446         data = alloc_blockmap(HAMMER_ZONE_META_INDEX, data_len,
 447                               offp, data_bufferp);
 448         bzero(data, data_len);
 449         return (data);
 450 }
 451
 452 void *
 453 alloc_data_element(hammer_off_t *offp, int32_t data_len,
 454                    struct buffer_info **data_bufferp)
 455 {
 456         void *data;
 457
 458         if (data_len >= HAMMER_BUFSIZE) {
 459                 assert(data_len <= HAMMER_BUFSIZE); /* just one buffer */
 460                 data = alloc_blockmap(HAMMER_ZONE_LARGE_DATA_INDEX, data_len,
 461                                       offp, data_bufferp);
 462                 bzero(data, data_len);
 463         } else if (data_len) {
 464                 data = alloc_blockmap(HAMMER_ZONE_SMALL_DATA_INDEX, data_len,
 465                                       offp, data_bufferp);
 466                 bzero(data, data_len);
 467         } else {
 468                 data = NULL;
 469         }
 470         return (data);
 471 }
 472
 473 /*
 474  * Format a new freemap.  Set all layer1 entries to UNAVAIL.  The initialize
 475  * code will load each volume's freemap.
 476  */
 477 void
 478 format_freemap(struct volume_info *root_vol, hammer_blockmap_t blockmap)
 479 {
 480         struct buffer_info *buffer = NULL;
 481         hammer_off_t layer1_offset;
 482         struct hammer_blockmap_layer1 *layer1;
 483         int i, isnew;
 484
 485         layer1_offset = alloc_bigblock(root_vol, HAMMER_ZONE_FREEMAP_INDEX);
 486         for (i = 0; i < (int)HAMMER_BLOCKMAP_RADIX1; ++i) {
 487                 isnew = ((i % HAMMER_BLOCKMAP_RADIX1_PERBUFFER) == 0);
 488                 layer1 = get_buffer_data(layer1_offset + i * sizeof(*layer1),
 489                                          &buffer, isnew);
 490                 bzero(layer1, sizeof(*layer1));
 491                 layer1->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 492                 layer1->blocks_free = 0;
 493                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 494         }
 495         rel_buffer(buffer);
 496
 497         blockmap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 498         blockmap->phys_offset = layer1_offset;
 499         blockmap->alloc_offset = HAMMER_ENCODE_RAW_BUFFER(255, -1);
 500         blockmap->next_offset = HAMMER_ENCODE_RAW_BUFFER(0, 0);
 501         blockmap->reserved01 = 0;
 502         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 503         root_vol->cache.modified = 1;
 504 }
 505
 506 /*
 507  * Load the volume's remaining free space into the freemap.
 508  *
 509  * Returns the number of bigblocks available.
 510  */
 511 int64_t
 512 initialize_freemap(struct volume_info *vol)
 513 {
 514         struct volume_info *root_vol;
 515         struct buffer_info *buffer1 = NULL;
 516         struct buffer_info *buffer2 = NULL;
 517         struct hammer_blockmap_layer1 *layer1;
 518         struct hammer_blockmap_layer2 *layer2;
 519         hammer_off_t layer1_base;
 520         hammer_off_t layer1_offset;
 521         hammer_off_t layer2_offset;
 522         hammer_off_t phys_offset;
 523         hammer_off_t aligned_vol_free_end;
 524         int64_t count = 0;
 525         int modified1 = 0;
 526
 527         root_vol = get_volume(RootVolNo);
 528         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 529                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 530
 531         printf("initialize freemap volume %d\n", vol->vol_no);
 532
 533         /*
 534          * Initialize the freemap.  First preallocate the bigblocks required
 535          * to implement layer2.   This preallocation is a bootstrap allocation
 536          * using blocks from the target volume.
 537          */
 538         layer1_base = root_vol->ondisk->vol0_blockmap[
 539                                         HAMMER_ZONE_FREEMAP_INDEX].phys_offset;
 540         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 541              phys_offset < aligned_vol_free_end;
 542              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 543                 layer1_offset = layer1_base +
 544                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 545                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 546                 if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 547                         layer1->phys_offset = alloc_bigblock(vol,
 548                                                 HAMMER_ZONE_FREEMAP_INDEX);
 549                         layer1->blocks_free = 0;
 550                         buffer1->cache.modified = 1;
 551                         layer1->layer1_crc = crc32(layer1,
 552                                                    HAMMER_LAYER1_CRCSIZE);
 553                 }
 554         }
 555
 556         /*
 557          * Now fill everything in.
 558          */
 559         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 560              phys_offset < aligned_vol_free_end;
 561              phys_offset += HAMMER_BIGBLOCK_SIZE) {
 562                 modified1 = 0;
 563                 layer1_offset = layer1_base +
 564                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 565                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 566
 567                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 568                 layer2_offset = layer1->phys_offset +
 569                                 HAMMER_BLOCKMAP_LAYER2_OFFSET(phys_offset);
 570
 571                 layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 572                 bzero(layer2, sizeof(*layer2));
 573                 if (phys_offset < vol->vol_free_off) {
 574                         /*
 575                          * Fixups XXX - bigblocks already allocated as part
 576                          * of the freemap bootstrap.
 577                          */
 578                         if (layer2->zone == 0) {
 579                                 layer2->zone = HAMMER_ZONE_FREEMAP_INDEX;
 580                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 581                                 layer2->bytes_free = 0;
 582                         }
 583                 } else if (phys_offset < vol->vol_free_end) {
 584                         ++layer1->blocks_free;
 585                         buffer1->cache.modified = 1;
 586                         layer2->zone = 0;
 587                         layer2->append_off = 0;
 588                         layer2->bytes_free = HAMMER_BIGBLOCK_SIZE;
 589                         ++count;
 590                         modified1 = 1;
 591                 } else {
 592                         layer2->zone = HAMMER_ZONE_UNAVAIL_INDEX;
 593                         layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 594                         layer2->bytes_free = 0;
 595                 }
 596                 layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 597                 buffer2->cache.modified = 1;
 598
 599                 /*
 600                  * Finish-up layer 1
 601                  */
 602                 if (modified1) {
 603                         layer1->layer1_crc = crc32(layer1,
 604                                                    HAMMER_LAYER1_CRCSIZE);
 605                         buffer1->cache.modified = 1;
 606                 }
 607         }
 608         rel_buffer(buffer1);
 609         rel_buffer(buffer2);
 610         rel_volume(root_vol);
 611         return(count);
 612 }
 613
 614 /*
 615  * Allocate big-blocks using our poor-man's volume->vol_free_off.
 616  *
 617  * If the zone is HAMMER_ZONE_FREEMAP_INDEX we are bootstrapping the freemap
 618  * itself and cannot update it yet.
 619  */
 620 hammer_off_t
 621 alloc_bigblock(struct volume_info *volume, int zone)
 622 {
 623         struct buffer_info *buffer1 = NULL;
 624         struct buffer_info *buffer2 = NULL;
 625         struct volume_info *root_vol;
 626         hammer_off_t result_offset;
 627         hammer_off_t layer_offset;
 628         struct hammer_blockmap_layer1 *layer1;
 629         struct hammer_blockmap_layer2 *layer2;
 630
 631         if (volume == NULL)
 632                 volume = get_volume(RootVolNo);
 633
 634         result_offset = volume->vol_free_off;
 635         if (result_offset >= volume->vol_free_end)
 636                 panic("alloc_bigblock: Ran out of room, filesystem too small");
 637         volume->vol_free_off += HAMMER_BIGBLOCK_SIZE;
 638
 639         /*
 640          * Update the freemap.
 641          */
 642         if (zone != HAMMER_ZONE_FREEMAP_INDEX) {
 643                 root_vol = get_volume(RootVolNo);
 644                 layer_offset = root_vol->ondisk->vol0_blockmap[
 645                                         HAMMER_ZONE_FREEMAP_INDEX].phys_offset;
 646                 layer_offset += HAMMER_BLOCKMAP_LAYER1_OFFSET(result_offset);
 647                 layer1 = get_buffer_data(layer_offset, &buffer1, 0);
 648                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 649                 --layer1->blocks_free;
 650                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 651                 buffer1->cache.modified = 1;
 652                 layer_offset = layer1->phys_offset +
 653                                HAMMER_BLOCKMAP_LAYER2_OFFSET(result_offset);
 654                 layer2 = get_buffer_data(layer_offset, &buffer2, 0);
 655                 assert(layer2->zone == 0);
 656                 layer2->zone = zone;
 657                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 658                 layer2->bytes_free = 0;
 659                 layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 660                 buffer2->cache.modified = 1;
 661
 662                 --root_vol->ondisk->vol0_stat_freebigblocks;
 663                 root_vol->cache.modified = 1;
 664
 665                 rel_buffer(buffer1);
 666                 rel_buffer(buffer2);
 667                 rel_volume(root_vol);
 668         }
 669
 670         rel_volume(volume);
 671         return(result_offset);
 672 }
 673
 674 /*
 675  * Format the undo-map for the root volume.
 676  */
 677 void
 678 format_undomap(hammer_volume_ondisk_t ondisk)
 679 {
 680         const int undo_zone = HAMMER_ZONE_UNDO_INDEX;
 681         hammer_off_t undo_limit;
 682         hammer_blockmap_t blockmap;
 683         struct buffer_info *buffer = NULL;
 684         hammer_off_t scan;
 685         int n;
 686         int limit_index;
 687         u_int32_t seqno;
 688
 689         /*
 690          * Size the undo buffer in multiples of HAMMER_BIGBLOCK_SIZE,
 691          * up to HAMMER_UNDO_LAYER2 big blocks.  Size to approximately
 692          * 0.1% of the disk.
 693          *
 694          * The minimum UNDO fifo size is 500MB, or approximately 1% of
 695          * the recommended 50G disk.
 696          *
 697          * Changing this minimum is rather dangerous as complex filesystem
 698          * operations can cause the UNDO FIFO to fill up otherwise.
 699          */
 700         undo_limit = UndoBufferSize;
 701         if (undo_limit == 0) {
 702                 undo_limit = (ondisk->vol_buf_end - ondisk->vol_buf_beg) / 1000;
 703                 if (undo_limit < 500*1024*1024)
 704                         undo_limit = 500*1024*1024;
 705         }
 706         undo_limit = (undo_limit + HAMMER_BIGBLOCK_MASK64) &
 707                      ~HAMMER_BIGBLOCK_MASK64;
 708         if (undo_limit < HAMMER_BIGBLOCK_SIZE)
 709                 undo_limit = HAMMER_BIGBLOCK_SIZE;
 710         if (undo_limit > HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2)
 711                 undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2;
 712         UndoBufferSize = undo_limit;
 713
 714         blockmap = &ondisk->vol0_blockmap[undo_zone];
 715         bzero(blockmap, sizeof(*blockmap));
 716         blockmap->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 717         blockmap->first_offset = HAMMER_ZONE_ENCODE(undo_zone, 0);
 718         blockmap->next_offset = blockmap->first_offset;
 719         blockmap->alloc_offset = HAMMER_ZONE_ENCODE(undo_zone, undo_limit);
 720         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 721
 722         n = 0;
 723         scan = blockmap->next_offset;
 724         limit_index = undo_limit / HAMMER_BIGBLOCK_SIZE;
 725
 726         assert(limit_index <= HAMMER_UNDO_LAYER2);
 727
 728         for (n = 0; n < limit_index; ++n) {
 729                 ondisk->vol0_undo_array[n] = alloc_bigblock(NULL,
 730                                                         HAMMER_ZONE_UNDO_INDEX);
 731                 scan += HAMMER_BIGBLOCK_SIZE;
 732         }
 733         while (n < HAMMER_UNDO_LAYER2) {
 734                 ondisk->vol0_undo_array[n] = HAMMER_BLOCKMAP_UNAVAIL;
 735                 ++n;
 736         }
 737
 738         /*
 739          * Pre-initialize the UNDO blocks (HAMMER version 4+)
 740          */
 741         printf("initializing the undo map (%jd MB)\n",
 742                 (intmax_t)(blockmap->alloc_offset & HAMMER_OFF_LONG_MASK) /
 743                 (1024 * 1024));
 744
 745         scan = blockmap->first_offset;
 746         seqno = 0;
 747
 748         while (scan < blockmap->alloc_offset) {
 749                 hammer_fifo_head_t head;
 750                 hammer_fifo_tail_t tail;
 751                 int isnew;
 752                 int bytes = HAMMER_UNDO_ALIGN;
 753
 754                 isnew = ((scan & HAMMER_BUFMASK64) == 0);
 755                 head = get_buffer_data(scan, &buffer, isnew);
 756                 buffer->cache.modified = 1;
 757                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 758
 759                 bzero(head, bytes);
 760                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 761                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
 762                 head->hdr_size = bytes;
 763                 head->hdr_seq = seqno++;
 764
 765                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 766                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 767                 tail->tail_size = bytes;
 768
 769                 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
 770                                 crc32(head + 1, bytes - sizeof(*head));
 771
 772                 scan += bytes;
 773         }
 774         rel_buffer(buffer);
 775 }
 776
 777 /*
 778  * Format a new blockmap.  This is mostly a degenerate case because
 779  * all allocations are now actually done from the freemap.
 780  */
 781 void
 782 format_blockmap(hammer_blockmap_t blockmap, hammer_off_t zone_base)
 783 {
 784         blockmap->phys_offset = 0;
 785         blockmap->alloc_offset = zone_base | HAMMER_VOL_ENCODE(255) |
 786                                  HAMMER_SHORT_OFF_ENCODE(-1);
 787         blockmap->first_offset = zone_base;
 788         blockmap->next_offset = zone_base;
 789         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 790 }
 791
 792 /*
 793  * Allocate a chunk of data out of a blockmap.  This is a simplified
 794  * version which uses next_offset as a simple allocation iterator.
 795  */
 796 static
 797 void *
 798 alloc_blockmap(int zone, int bytes, hammer_off_t *result_offp,
 799                struct buffer_info **bufferp)
 800 {
 801         struct buffer_info *buffer1 = NULL;
 802         struct buffer_info *buffer2 = NULL;
 803         struct volume_info *volume;
 804         hammer_blockmap_t blockmap;
 805         hammer_blockmap_t freemap;
 806         struct hammer_blockmap_layer1 *layer1;
 807         struct hammer_blockmap_layer2 *layer2;
 808         hammer_off_t layer1_offset;
 809         hammer_off_t layer2_offset;
 810         hammer_off_t zone2_offset;
 811         void *ptr;
 812
 813         volume = get_volume(RootVolNo);
 814
 815         blockmap = &volume->ondisk->vol0_blockmap[zone];
 816         freemap = &volume->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 817
 818         /*
 819          * Alignment and buffer-boundary issues.  If the allocation would
 820          * cross a buffer boundary we have to skip to the next buffer.
 821          */
 822         bytes = (bytes + 15) & ~15;
 823
 824 again:
 825         if ((blockmap->next_offset ^ (blockmap->next_offset + bytes - 1)) &
 826             ~HAMMER_BUFMASK64) {
 827                 volume->cache.modified = 1;
 828                 blockmap->next_offset = (blockmap->next_offset + bytes) &
 829                                         ~HAMMER_BUFMASK64;
 830         }
 831
 832         /*
 833          * Dive layer 1.  For now we can't allocate data outside of volume 0.
 834          */
 835         layer1_offset = freemap->phys_offset +
 836                         HAMMER_BLOCKMAP_LAYER1_OFFSET(blockmap->next_offset);
 837
 838         layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 839
 840         if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 841                 fprintf(stderr, "alloc_blockmap: ran out of space!\n");
 842                 exit(1);
 843         }
 844
 845         /*
 846          * Dive layer 2
 847          */
 848         layer2_offset = layer1->phys_offset +
 849                         HAMMER_BLOCKMAP_LAYER2_OFFSET(blockmap->next_offset);
 850
 851         layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 852
 853         if (layer2->zone == HAMMER_ZONE_UNAVAIL_INDEX) {
 854                 fprintf(stderr, "alloc_blockmap: ran out of space!\n");
 855                 exit(1);
 856         }
 857
 858         /*
 859          * If we are entering a new bigblock assign ownership to our
 860          * zone.  If the bigblock is owned by another zone skip it.
 861          */
 862         if (layer2->zone == 0) {
 863                 --layer1->blocks_free;
 864                 layer2->zone = zone;
 865                 assert(layer2->bytes_free == HAMMER_BIGBLOCK_SIZE);
 866                 assert(layer2->append_off == 0);
 867         }
 868         if (layer2->zone != zone) {
 869                 blockmap->next_offset = (blockmap->next_offset + HAMMER_BIGBLOCK_SIZE) &
 870                                         ~HAMMER_BIGBLOCK_MASK64;
 871                 goto again;
 872         }
 873
 874         buffer1->cache.modified = 1;
 875         buffer2->cache.modified = 1;
 876         volume->cache.modified = 1;
 877         assert(layer2->append_off ==
 878                (blockmap->next_offset & HAMMER_BIGBLOCK_MASK));
 879         layer2->bytes_free -= bytes;
 880         *result_offp = blockmap->next_offset;
 881         blockmap->next_offset += bytes;
 882         layer2->append_off = (int)blockmap->next_offset &
 883                               HAMMER_BIGBLOCK_MASK;
 884
 885         layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 886         layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 887
 888         zone2_offset = (*result_offp & ~HAMMER_OFF_ZONE_MASK) |
 889                         HAMMER_ZONE_ENCODE(zone, 0);
 890
 891         ptr = get_buffer_data(zone2_offset, bufferp, 0);
 892         (*bufferp)->cache.modified = 1;
 893
 894         rel_buffer(buffer1);
 895         rel_buffer(buffer2);
 896         rel_volume(volume);
 897         return(ptr);
 898 }
 899
 900 /*
 901  * Flush various tracking structures to disk
 902  */
 903
 904 /*
 905  * Flush various tracking structures to disk
 906  */
 907 void
 908 flush_all_volumes(void)
 909 {
 910         struct volume_info *vol;
 911
 912         TAILQ_FOREACH(vol, &VolList, entry)
 913                 flush_volume(vol);
 914 }
 915
 916 void
 917 flush_volume(struct volume_info *volume)
 918 {
 919         struct buffer_info *buffer;
 920         int i;
 921
 922         for (i = 0; i < HAMMER_BUFLISTS; ++i) {
 923                 TAILQ_FOREACH(buffer, &volume->buffer_lists[i], entry)
 924                         flush_buffer(buffer);
 925         }
 926         writehammerbuf(volume, volume->ondisk, 0);
 927         volume->cache.modified = 0;
 928 }
 929
 930 void
 931 flush_buffer(struct buffer_info *buffer)
 932 {
 933         writehammerbuf(buffer->volume, buffer->ondisk, buffer->raw_offset);
 934         buffer->cache.modified = 0;
 935 }
 936
 937 #if 0
 938 /*
 939  * Generic buffer initialization
 940  */
 941 static void
 942 init_fifo_head(hammer_fifo_head_t head, u_int16_t hdr_type)
 943 {
 944         head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 945         head->hdr_type = hdr_type;
 946         head->hdr_size = 0;
 947         head->hdr_crc = 0;
 948         head->hdr_seq = 0;
 949 }
 950
 951 #endif
 952
 953 #if 0
 954 /*
 955  * Core I/O operations
 956  */
 957 static void
 958 readhammerbuf(struct volume_info *vol, void *data, int64_t offset)
 959 {
 960         ssize_t n;
 961
 962         n = pread(vol->fd, data, HAMMER_BUFSIZE, offset);
 963         if (n != HAMMER_BUFSIZE)
 964                 err(1, "Read volume %d (%s)", vol->vol_no, vol->name);
 965 }
 966
 967 #endif
 968
 969 static void
 970 writehammerbuf(struct volume_info *vol, const void *data, int64_t offset)
 971 {
 972         ssize_t n;
 973
 974         n = pwrite(vol->fd, data, HAMMER_BUFSIZE, offset);
 975         if (n != HAMMER_BUFSIZE)
 976                 err(1, "Write volume %d (%s)", vol->vol_no, vol->name);
 977 }
 978
 979 void
 980 panic(const char *ctl, ...)
 981 {
 982         va_list va;
 983
 984         va_start(va, ctl);
 985         vfprintf(stderr, ctl, va);
 986         va_end(va);
 987         fprintf(stderr, "\n");
 988         exit(1);
 989 }
 990