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