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 static struct buffer_info*
  87 find_buffer(struct volume_info *volume, hammer_off_t buf_offset)
  88 {
  89         int hi;
  90         struct buffer_info *buf;
  91
  92         hi = buffer_hash(buf_offset);
  93         TAILQ_FOREACH(buf, &volume->buffer_lists[hi], entry)
  94                 if (buf->buf_offset == buf_offset)
  95                         return(buf);
  96         return(NULL);
  97 }
  98
  99 /*
 100  * Lookup the requested information structure and related on-disk buffer.
 101  * Missing structures are created.
 102  */
 103 struct volume_info *
 104 setup_volume(int32_t vol_no, const char *filename, int isnew, int oflags)
 105 {
 106         struct volume_info *vol;
 107         struct volume_info *scan;
 108         struct hammer_volume_ondisk *ondisk;
 109         int i, n;
 110
 111         /*
 112          * Allocate the volume structure
 113          */
 114         vol = malloc(sizeof(*vol));
 115         bzero(vol, sizeof(*vol));
 116         for (i = 0; i < HAMMER_BUFLISTS; ++i)
 117                 TAILQ_INIT(&vol->buffer_lists[i]);
 118         vol->name = strdup(filename);
 119         vol->fd = open(filename, oflags);
 120         if (vol->fd < 0) {
 121                 free(vol->name);
 122                 free(vol);
 123                 err(1, "setup_volume: %s: Open failed", filename);
 124         }
 125
 126         /*
 127          * Read or initialize the volume header
 128          */
 129         vol->ondisk = ondisk = malloc(HAMMER_BUFSIZE);
 130         if (isnew > 0) {
 131                 bzero(ondisk, HAMMER_BUFSIZE);
 132         } else {
 133                 n = pread(vol->fd, ondisk, HAMMER_BUFSIZE, 0);
 134                 if (n != HAMMER_BUFSIZE) {
 135                         err(1, "setup_volume: %s: Read failed at offset 0",
 136                             filename);
 137                 }
 138                 vol_no = ondisk->vol_no;
 139                 if (RootVolNo < 0) {
 140                         RootVolNo = ondisk->vol_rootvol;
 141                 } else if (RootVolNo != (int)ondisk->vol_rootvol) {
 142                         errx(1, "setup_volume: %s: root volume disagreement: "
 143                                 "%d vs %d",
 144                                 vol->name, RootVolNo, ondisk->vol_rootvol);
 145                 }
 146
 147                 if (bcmp(&Hammer_FSType, &ondisk->vol_fstype, sizeof(Hammer_FSType)) != 0) {
 148                         errx(1, "setup_volume: %s: Header does not indicate "
 149                                 "that this is a hammer volume", vol->name);
 150                 }
 151                 if (TAILQ_EMPTY(&VolList)) {
 152                         Hammer_FSId = vol->ondisk->vol_fsid;
 153                 } else if (bcmp(&Hammer_FSId, &ondisk->vol_fsid, sizeof(Hammer_FSId)) != 0) {
 154                         errx(1, "setup_volume: %s: FSId does match other "
 155                                 "volumes!", vol->name);
 156                 }
 157         }
 158         vol->vol_no = vol_no;
 159
 160         if (isnew > 0) {
 161                 /*init_fifo_head(&ondisk->head, HAMMER_HEAD_TYPE_VOL);*/
 162                 vol->cache.modified = 1;
 163         }
 164
 165         /*
 166          * Link the volume structure in
 167          */
 168         TAILQ_FOREACH(scan, &VolList, entry) {
 169                 if (scan->vol_no == vol_no) {
 170                         errx(1, "setup_volume %s: Duplicate volume number %d "
 171                                 "against %s", filename, vol_no, scan->name);
 172                 }
 173         }
 174         TAILQ_INSERT_TAIL(&VolList, vol, entry);
 175         return(vol);
 176 }
 177
 178 struct volume_info *
 179 test_volume(int32_t vol_no)
 180 {
 181         struct volume_info *vol;
 182
 183         TAILQ_FOREACH(vol, &VolList, entry) {
 184                 if (vol->vol_no == vol_no)
 185                         break;
 186         }
 187         if (vol == NULL)
 188                 return(NULL);
 189         ++vol->cache.refs;
 190         /* not added to or removed from hammer cache */
 191         return(vol);
 192 }
 193
 194 struct volume_info *
 195 get_volume(int32_t vol_no)
 196 {
 197         struct volume_info *vol;
 198
 199         TAILQ_FOREACH(vol, &VolList, entry) {
 200                 if (vol->vol_no == vol_no)
 201                         break;
 202         }
 203         if (vol == NULL)
 204                 errx(1, "get_volume: Volume %d does not exist!", vol_no);
 205         ++vol->cache.refs;
 206         /* not added to or removed from hammer cache */
 207         return(vol);
 208 }
 209
 210 void
 211 rel_volume(struct volume_info *volume)
 212 {
 213         if (volume == NULL)
 214                 return;
 215         /* not added to or removed from hammer cache */
 216         --volume->cache.refs;
 217 }
 218
 219 /*
 220  * Acquire the specified buffer.
 221  */
 222 struct buffer_info *
 223 get_buffer(hammer_off_t buf_offset, int isnew)
 224 {
 225         void *ondisk;
 226         struct buffer_info *buf;
 227         struct volume_info *volume;
 228         hammer_off_t orig_offset = buf_offset;
 229         int vol_no;
 230         int zone;
 231         int hi, n;
 232         int dora = 0;
 233
 234         zone = HAMMER_ZONE_DECODE(buf_offset);
 235         if (zone > HAMMER_ZONE_RAW_BUFFER_INDEX) {
 236                 buf_offset = blockmap_lookup(buf_offset, NULL, NULL, NULL);
 237         }
 238         if (buf_offset == HAMMER_OFF_BAD)
 239                 return(NULL);
 240
 241         if (AssertOnFailure) {
 242                 assert((buf_offset & HAMMER_OFF_ZONE_MASK) ==
 243                        HAMMER_ZONE_RAW_BUFFER);
 244         }
 245         vol_no = HAMMER_VOL_DECODE(buf_offset);
 246         volume = test_volume(vol_no);
 247         if (volume == NULL) {
 248                 if (AssertOnFailure)
 249                         errx(1, "get_buffer: Volume %d not found!", vol_no);
 250                 return(NULL);
 251         }
 252
 253         buf_offset &= ~HAMMER_BUFMASK64;
 254         buf = find_buffer(volume, buf_offset);
 255
 256         if (buf == NULL) {
 257                 buf = malloc(sizeof(*buf));
 258                 bzero(buf, sizeof(*buf));
 259                 if (DebugOpt) {
 260                         fprintf(stderr, "get_buffer: %016llx %016llx at %p\n",
 261                                 (long long)orig_offset, (long long)buf_offset,
 262                                 buf);
 263                 }
 264                 buf->buf_offset = buf_offset;
 265                 buf->raw_offset = volume->ondisk->vol_buf_beg +
 266                                   (buf_offset & HAMMER_OFF_SHORT_MASK);
 267                 buf->volume = volume;
 268                 hi = buffer_hash(buf_offset);
 269                 TAILQ_INSERT_TAIL(&volume->buffer_lists[hi], buf, entry);
 270                 ++volume->cache.refs;
 271                 buf->cache.u.buffer = buf;
 272                 hammer_cache_add(&buf->cache, ISBUFFER);
 273                 dora = (isnew == 0);
 274                 if (isnew < 0)
 275                         buf->flags |= HAMMER_BUFINFO_READAHEAD;
 276         } else {
 277                 if (DebugOpt) {
 278                         fprintf(stderr, "get_buffer: %016llx %016llx at %p *\n",
 279                                 (long long)orig_offset, (long long)buf_offset,
 280                                 buf);
 281                 }
 282                 if (isnew >= 0) {
 283                         buf->flags &= ~HAMMER_BUFINFO_READAHEAD;
 284                         hammer_cache_used(&buf->cache);
 285                 }
 286                 ++buf->use_count;
 287         }
 288         ++buf->cache.refs;
 289         hammer_cache_flush();
 290         if ((ondisk = buf->ondisk) == NULL) {
 291                 buf->ondisk = ondisk = malloc(HAMMER_BUFSIZE);
 292                 if (isnew <= 0) {
 293                         n = pread(volume->fd, ondisk, HAMMER_BUFSIZE,
 294                                   buf->raw_offset);
 295                         if (n != HAMMER_BUFSIZE) {
 296                                 if (AssertOnFailure)
 297                                         err(1, "get_buffer: %s:%016llx "
 298                                             "Read failed at offset %016llx",
 299                                             volume->name,
 300                                             (long long)buf->buf_offset,
 301                                             (long long)buf->raw_offset);
 302                                 bzero(ondisk, HAMMER_BUFSIZE);
 303                         }
 304                 }
 305         }
 306         if (isnew > 0) {
 307                 bzero(ondisk, HAMMER_BUFSIZE);
 308                 buf->cache.modified = 1;
 309         }
 310         if (dora)
 311                 get_buffer_readahead(buf);
 312         return(buf);
 313 }
 314
 315 static void
 316 get_buffer_readahead(struct buffer_info *base)
 317 {
 318         struct buffer_info *buf;
 319         struct volume_info *vol;
 320         hammer_off_t buf_offset;
 321         int64_t raw_offset;
 322         int ri = UseReadBehind;
 323         int re = UseReadAhead;
 324
 325         raw_offset = base->raw_offset + ri * HAMMER_BUFSIZE;
 326         vol = base->volume;
 327
 328         while (ri < re) {
 329                 if (raw_offset >= vol->ondisk->vol_buf_end)
 330                         break;
 331                 if (raw_offset < vol->ondisk->vol_buf_beg || ri == 0) {
 332                         ++ri;
 333                         raw_offset += HAMMER_BUFSIZE;
 334                         continue;
 335                 }
 336                 buf_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no,
 337                         raw_offset - vol->ondisk->vol_buf_beg);
 338                 buf = find_buffer(vol, buf_offset);
 339                 if (buf == NULL) {
 340                         buf = get_buffer(buf_offset, -1);
 341                         rel_buffer(buf);
 342                 }
 343                 ++ri;
 344                 raw_offset += HAMMER_BUFSIZE;
 345         }
 346 }
 347
 348 void
 349 rel_buffer(struct buffer_info *buffer)
 350 {
 351         struct volume_info *volume;
 352         int hi;
 353
 354         if (buffer == NULL)
 355                 return;
 356         assert(buffer->cache.refs > 0);
 357         if (--buffer->cache.refs == 0) {
 358                 if (buffer->cache.delete) {
 359                         hi = buffer_hash(buffer->buf_offset);
 360                         volume = buffer->volume;
 361                         if (buffer->cache.modified)
 362                                 flush_buffer(buffer);
 363                         TAILQ_REMOVE(&volume->buffer_lists[hi], buffer, entry);
 364                         hammer_cache_del(&buffer->cache);
 365                         free(buffer->ondisk);
 366                         free(buffer);
 367                         rel_volume(volume);
 368                 }
 369         }
 370 }
 371
 372 /*
 373  * Retrieve a pointer to a buffer data given a buffer offset.  The underlying
 374  * bufferp is freed if isnew or the offset is out of range of the cached data.
 375  * If bufferp is freed a referenced buffer is loaded into it.
 376  */
 377 void *
 378 get_buffer_data(hammer_off_t buf_offset, struct buffer_info **bufferp,
 379                 int isnew)
 380 {
 381         if (*bufferp != NULL) {
 382                 if (isnew > 0 ||
 383                     (((*bufferp)->buf_offset ^ buf_offset) & ~HAMMER_BUFMASK64)) {
 384                         rel_buffer(*bufferp);
 385                         *bufferp = NULL;
 386                 }
 387         }
 388         return(get_ondisk(buf_offset, bufferp, isnew));
 389 }
 390
 391 /*
 392  * Retrieve a pointer to a B-Tree node given a cluster offset.  The underlying
 393  * bufferp is freed if non-NULL and a referenced buffer is loaded into it.
 394  */
 395 hammer_node_ondisk_t
 396 get_node(hammer_off_t node_offset, struct buffer_info **bufferp)
 397 {
 398         if (*bufferp != NULL) {
 399                 rel_buffer(*bufferp);
 400                 *bufferp = NULL;
 401         }
 402         return(get_ondisk(node_offset, bufferp, 0));
 403 }
 404
 405 /*
 406  * Return a pointer to a buffer data given a buffer offset.
 407  * If *bufferp is NULL acquire the buffer otherwise use that buffer.
 408  */
 409 static __inline
 410 void *
 411 get_ondisk(hammer_off_t buf_offset, struct buffer_info **bufferp,
 412         int isnew)
 413 {
 414         struct buffer_info *buffer;
 415
 416         buffer = *bufferp;
 417         if (buffer == NULL) {
 418                 buffer = *bufferp = get_buffer(buf_offset, isnew);
 419                 if (buffer == NULL)
 420                         return(NULL);
 421         }
 422
 423         return((char *)buffer->ondisk +
 424                 ((int32_t)buf_offset & HAMMER_BUFMASK));
 425 }
 426
 427 /*
 428  * Allocate HAMMER elements - btree nodes, meta data, data storage
 429  */
 430 void *
 431 alloc_btree_element(hammer_off_t *offp,
 432                     struct buffer_info **data_bufferp)
 433 {
 434         hammer_node_ondisk_t node;
 435
 436         node = alloc_blockmap(HAMMER_ZONE_BTREE_INDEX, sizeof(*node),
 437                               offp, data_bufferp);
 438         bzero(node, sizeof(*node));
 439         return (node);
 440 }
 441
 442 void *
 443 alloc_meta_element(hammer_off_t *offp, int32_t data_len,
 444                    struct buffer_info **data_bufferp)
 445 {
 446         void *data;
 447
 448         data = alloc_blockmap(HAMMER_ZONE_META_INDEX, data_len,
 449                               offp, data_bufferp);
 450         bzero(data, data_len);
 451         return (data);
 452 }
 453
 454 void *
 455 alloc_data_element(hammer_off_t *offp, int32_t data_len,
 456                    struct buffer_info **data_bufferp)
 457 {
 458         void *data;
 459
 460         if (data_len >= HAMMER_BUFSIZE) {
 461                 assert(data_len <= HAMMER_BUFSIZE); /* just one buffer */
 462                 data = alloc_blockmap(HAMMER_ZONE_LARGE_DATA_INDEX, data_len,
 463                                       offp, data_bufferp);
 464                 bzero(data, data_len);
 465         } else if (data_len) {
 466                 data = alloc_blockmap(HAMMER_ZONE_SMALL_DATA_INDEX, data_len,
 467                                       offp, data_bufferp);
 468                 bzero(data, data_len);
 469         } else {
 470                 data = NULL;
 471         }
 472         return (data);
 473 }
 474
 475 /*
 476  * Format a new freemap.  Set all layer1 entries to UNAVAIL.  The initialize
 477  * code will load each volume's freemap.
 478  */
 479 void
 480 format_freemap(struct volume_info *root_vol, hammer_blockmap_t blockmap)
 481 {
 482         struct buffer_info *buffer = NULL;
 483         hammer_off_t layer1_offset;
 484         struct hammer_blockmap_layer1 *layer1;
 485         int i, isnew;
 486
 487         layer1_offset = alloc_bigblock(root_vol, HAMMER_ZONE_FREEMAP_INDEX);
 488         for (i = 0; i < (int)HAMMER_BLOCKMAP_RADIX1; ++i) {
 489                 isnew = ((i % HAMMER_BLOCKMAP_RADIX1_PERBUFFER) == 0);
 490                 layer1 = get_buffer_data(layer1_offset + i * sizeof(*layer1),
 491                                          &buffer, isnew);
 492                 bzero(layer1, sizeof(*layer1));
 493                 layer1->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 494                 layer1->blocks_free = 0;
 495                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 496         }
 497         rel_buffer(buffer);
 498
 499         blockmap = &root_vol->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 500         blockmap->phys_offset = layer1_offset;
 501         blockmap->alloc_offset = HAMMER_ENCODE_RAW_BUFFER(255, -1);
 502         blockmap->next_offset = HAMMER_ENCODE_RAW_BUFFER(0, 0);
 503         blockmap->reserved01 = 0;
 504         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 505         root_vol->cache.modified = 1;
 506 }
 507
 508 /*
 509  * Load the volume's remaining free space into the freemap.
 510  *
 511  * Returns the number of bigblocks available.
 512  */
 513 int64_t
 514 initialize_freemap(struct volume_info *vol)
 515 {
 516         struct volume_info *root_vol;
 517         struct buffer_info *buffer1 = NULL;
 518         struct buffer_info *buffer2 = NULL;
 519         struct hammer_blockmap_layer1 *layer1;
 520         struct hammer_blockmap_layer2 *layer2;
 521         hammer_off_t layer1_base;
 522         hammer_off_t layer1_offset;
 523         hammer_off_t layer2_offset;
 524         hammer_off_t phys_offset;
 525         hammer_off_t aligned_vol_free_end;
 526         int64_t count = 0;
 527         int modified1 = 0;
 528
 529         root_vol = get_volume(RootVolNo);
 530         aligned_vol_free_end = (vol->vol_free_end + HAMMER_BLOCKMAP_LAYER2_MASK)
 531                                 & ~HAMMER_BLOCKMAP_LAYER2_MASK;
 532
 533         printf("initialize freemap volume %d\n", vol->vol_no);
 534
 535         /*
 536          * Initialize the freemap.  First preallocate the bigblocks required
 537          * to implement layer2.   This preallocation is a bootstrap allocation
 538          * using blocks from the target volume.
 539          */
 540         layer1_base = root_vol->ondisk->vol0_blockmap[
 541                                         HAMMER_ZONE_FREEMAP_INDEX].phys_offset;
 542         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 543              phys_offset < aligned_vol_free_end;
 544              phys_offset += HAMMER_BLOCKMAP_LAYER2) {
 545                 layer1_offset = layer1_base +
 546                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 547                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 548                 if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 549                         layer1->phys_offset = alloc_bigblock(vol,
 550                                                 HAMMER_ZONE_FREEMAP_INDEX);
 551                         layer1->blocks_free = 0;
 552                         buffer1->cache.modified = 1;
 553                         layer1->layer1_crc = crc32(layer1,
 554                                                    HAMMER_LAYER1_CRCSIZE);
 555                 }
 556         }
 557
 558         /*
 559          * Now fill everything in.
 560          */
 561         for (phys_offset = HAMMER_ENCODE_RAW_BUFFER(vol->vol_no, 0);
 562              phys_offset < aligned_vol_free_end;
 563              phys_offset += HAMMER_BIGBLOCK_SIZE) {
 564                 modified1 = 0;
 565                 layer1_offset = layer1_base +
 566                                 HAMMER_BLOCKMAP_LAYER1_OFFSET(phys_offset);
 567                 layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 568
 569                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 570                 layer2_offset = layer1->phys_offset +
 571                                 HAMMER_BLOCKMAP_LAYER2_OFFSET(phys_offset);
 572
 573                 layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 574                 bzero(layer2, sizeof(*layer2));
 575                 if (phys_offset < vol->vol_free_off) {
 576                         /*
 577                          * Fixups XXX - bigblocks already allocated as part
 578                          * of the freemap bootstrap.
 579                          */
 580                         if (layer2->zone == 0) {
 581                                 layer2->zone = HAMMER_ZONE_FREEMAP_INDEX;
 582                                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 583                                 layer2->bytes_free = 0;
 584                         }
 585                 } else if (phys_offset < vol->vol_free_end) {
 586                         ++layer1->blocks_free;
 587                         buffer1->cache.modified = 1;
 588                         layer2->zone = 0;
 589                         layer2->append_off = 0;
 590                         layer2->bytes_free = HAMMER_BIGBLOCK_SIZE;
 591                         ++count;
 592                         modified1 = 1;
 593                 } else {
 594                         layer2->zone = HAMMER_ZONE_UNAVAIL_INDEX;
 595                         layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 596                         layer2->bytes_free = 0;
 597                 }
 598                 layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 599                 buffer2->cache.modified = 1;
 600
 601                 /*
 602                  * Finish-up layer 1
 603                  */
 604                 if (modified1) {
 605                         layer1->layer1_crc = crc32(layer1,
 606                                                    HAMMER_LAYER1_CRCSIZE);
 607                         buffer1->cache.modified = 1;
 608                 }
 609         }
 610         rel_buffer(buffer1);
 611         rel_buffer(buffer2);
 612         rel_volume(root_vol);
 613         return(count);
 614 }
 615
 616 /*
 617  * Allocate big-blocks using our poor-man's volume->vol_free_off.
 618  *
 619  * If the zone is HAMMER_ZONE_FREEMAP_INDEX we are bootstrapping the freemap
 620  * itself and cannot update it yet.
 621  */
 622 hammer_off_t
 623 alloc_bigblock(struct volume_info *volume, int zone)
 624 {
 625         struct buffer_info *buffer1 = NULL;
 626         struct buffer_info *buffer2 = NULL;
 627         struct volume_info *root_vol;
 628         hammer_off_t result_offset;
 629         hammer_off_t layer_offset;
 630         struct hammer_blockmap_layer1 *layer1;
 631         struct hammer_blockmap_layer2 *layer2;
 632
 633         if (volume == NULL)
 634                 volume = get_volume(RootVolNo);
 635
 636         result_offset = volume->vol_free_off;
 637         if (result_offset >= volume->vol_free_end)
 638                 panic("alloc_bigblock: Ran out of room, filesystem too small");
 639         volume->vol_free_off += HAMMER_BIGBLOCK_SIZE;
 640
 641         /*
 642          * Update the freemap.
 643          */
 644         if (zone != HAMMER_ZONE_FREEMAP_INDEX) {
 645                 root_vol = get_volume(RootVolNo);
 646                 layer_offset = root_vol->ondisk->vol0_blockmap[
 647                                         HAMMER_ZONE_FREEMAP_INDEX].phys_offset;
 648                 layer_offset += HAMMER_BLOCKMAP_LAYER1_OFFSET(result_offset);
 649                 layer1 = get_buffer_data(layer_offset, &buffer1, 0);
 650                 assert(layer1->phys_offset != HAMMER_BLOCKMAP_UNAVAIL);
 651                 --layer1->blocks_free;
 652                 layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 653                 buffer1->cache.modified = 1;
 654                 layer_offset = layer1->phys_offset +
 655                                HAMMER_BLOCKMAP_LAYER2_OFFSET(result_offset);
 656                 layer2 = get_buffer_data(layer_offset, &buffer2, 0);
 657                 assert(layer2->zone == 0);
 658                 layer2->zone = zone;
 659                 layer2->append_off = HAMMER_BIGBLOCK_SIZE;
 660                 layer2->bytes_free = 0;
 661                 layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 662                 buffer2->cache.modified = 1;
 663
 664                 --root_vol->ondisk->vol0_stat_freebigblocks;
 665                 root_vol->cache.modified = 1;
 666
 667                 rel_buffer(buffer1);
 668                 rel_buffer(buffer2);
 669                 rel_volume(root_vol);
 670         }
 671
 672         rel_volume(volume);
 673         return(result_offset);
 674 }
 675
 676 /*
 677  * Format the undo-map for the root volume.
 678  */
 679 void
 680 format_undomap(hammer_volume_ondisk_t ondisk)
 681 {
 682         const int undo_zone = HAMMER_ZONE_UNDO_INDEX;
 683         hammer_off_t undo_limit;
 684         hammer_blockmap_t blockmap;
 685         struct buffer_info *buffer = NULL;
 686         hammer_off_t scan;
 687         int n;
 688         int limit_index;
 689         u_int32_t seqno;
 690
 691         /*
 692          * Size the undo buffer in multiples of HAMMER_BIGBLOCK_SIZE,
 693          * up to HAMMER_UNDO_LAYER2 big blocks.  Size to approximately
 694          * 0.1% of the disk.
 695          *
 696          * The minimum UNDO fifo size is 500MB, or approximately 1% of
 697          * the recommended 50G disk.
 698          *
 699          * Changing this minimum is rather dangerous as complex filesystem
 700          * operations can cause the UNDO FIFO to fill up otherwise.
 701          */
 702         undo_limit = UndoBufferSize;
 703         if (undo_limit == 0) {
 704                 undo_limit = (ondisk->vol_buf_end - ondisk->vol_buf_beg) / 1000;
 705                 if (undo_limit < 500*1024*1024)
 706                         undo_limit = 500*1024*1024;
 707         }
 708         undo_limit = (undo_limit + HAMMER_BIGBLOCK_MASK64) &
 709                      ~HAMMER_BIGBLOCK_MASK64;
 710         if (undo_limit < HAMMER_BIGBLOCK_SIZE)
 711                 undo_limit = HAMMER_BIGBLOCK_SIZE;
 712         if (undo_limit > HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2)
 713                 undo_limit = HAMMER_BIGBLOCK_SIZE * HAMMER_UNDO_LAYER2;
 714         UndoBufferSize = undo_limit;
 715
 716         blockmap = &ondisk->vol0_blockmap[undo_zone];
 717         bzero(blockmap, sizeof(*blockmap));
 718         blockmap->phys_offset = HAMMER_BLOCKMAP_UNAVAIL;
 719         blockmap->first_offset = HAMMER_ZONE_ENCODE(undo_zone, 0);
 720         blockmap->next_offset = blockmap->first_offset;
 721         blockmap->alloc_offset = HAMMER_ZONE_ENCODE(undo_zone, undo_limit);
 722         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 723
 724         n = 0;
 725         scan = blockmap->next_offset;
 726         limit_index = undo_limit / HAMMER_BIGBLOCK_SIZE;
 727
 728         assert(limit_index <= HAMMER_UNDO_LAYER2);
 729
 730         for (n = 0; n < limit_index; ++n) {
 731                 ondisk->vol0_undo_array[n] = alloc_bigblock(NULL,
 732                                                         HAMMER_ZONE_UNDO_INDEX);
 733                 scan += HAMMER_BIGBLOCK_SIZE;
 734         }
 735         while (n < HAMMER_UNDO_LAYER2) {
 736                 ondisk->vol0_undo_array[n] = HAMMER_BLOCKMAP_UNAVAIL;
 737                 ++n;
 738         }
 739
 740         /*
 741          * Pre-initialize the UNDO blocks (HAMMER version 4+)
 742          */
 743         printf("initializing the undo map (%jd MB)\n",
 744                 (intmax_t)(blockmap->alloc_offset & HAMMER_OFF_LONG_MASK) /
 745                 (1024 * 1024));
 746
 747         scan = blockmap->first_offset;
 748         seqno = 0;
 749
 750         while (scan < blockmap->alloc_offset) {
 751                 hammer_fifo_head_t head;
 752                 hammer_fifo_tail_t tail;
 753                 int isnew;
 754                 int bytes = HAMMER_UNDO_ALIGN;
 755
 756                 isnew = ((scan & HAMMER_BUFMASK64) == 0);
 757                 head = get_buffer_data(scan, &buffer, isnew);
 758                 buffer->cache.modified = 1;
 759                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 760
 761                 bzero(head, bytes);
 762                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 763                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
 764                 head->hdr_size = bytes;
 765                 head->hdr_seq = seqno++;
 766
 767                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 768                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 769                 tail->tail_size = bytes;
 770
 771                 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
 772                                 crc32(head + 1, bytes - sizeof(*head));
 773
 774                 scan += bytes;
 775         }
 776         rel_buffer(buffer);
 777 }
 778
 779 /*
 780  * Format a new blockmap.  This is mostly a degenerate case because
 781  * all allocations are now actually done from the freemap.
 782  */
 783 void
 784 format_blockmap(hammer_blockmap_t blockmap, hammer_off_t zone_base)
 785 {
 786         blockmap->phys_offset = 0;
 787         blockmap->alloc_offset = zone_base | HAMMER_VOL_ENCODE(255) |
 788                                  HAMMER_SHORT_OFF_ENCODE(-1);
 789         blockmap->first_offset = zone_base;
 790         blockmap->next_offset = zone_base;
 791         blockmap->entry_crc = crc32(blockmap, HAMMER_BLOCKMAP_CRCSIZE);
 792 }
 793
 794 /*
 795  * Allocate a chunk of data out of a blockmap.  This is a simplified
 796  * version which uses next_offset as a simple allocation iterator.
 797  */
 798 static
 799 void *
 800 alloc_blockmap(int zone, int bytes, hammer_off_t *result_offp,
 801                struct buffer_info **bufferp)
 802 {
 803         struct buffer_info *buffer1 = NULL;
 804         struct buffer_info *buffer2 = NULL;
 805         struct volume_info *volume;
 806         hammer_blockmap_t blockmap;
 807         hammer_blockmap_t freemap;
 808         struct hammer_blockmap_layer1 *layer1;
 809         struct hammer_blockmap_layer2 *layer2;
 810         hammer_off_t layer1_offset;
 811         hammer_off_t layer2_offset;
 812         hammer_off_t zone2_offset;
 813         void *ptr;
 814
 815         volume = get_volume(RootVolNo);
 816
 817         blockmap = &volume->ondisk->vol0_blockmap[zone];
 818         freemap = &volume->ondisk->vol0_blockmap[HAMMER_ZONE_FREEMAP_INDEX];
 819
 820         /*
 821          * Alignment and buffer-boundary issues.  If the allocation would
 822          * cross a buffer boundary we have to skip to the next buffer.
 823          */
 824         bytes = (bytes + 15) & ~15;
 825
 826 again:
 827         if ((blockmap->next_offset ^ (blockmap->next_offset + bytes - 1)) &
 828             ~HAMMER_BUFMASK64) {
 829                 volume->cache.modified = 1;
 830                 blockmap->next_offset = (blockmap->next_offset + bytes) &
 831                                         ~HAMMER_BUFMASK64;
 832         }
 833
 834         /*
 835          * Dive layer 1.  For now we can't allocate data outside of volume 0.
 836          */
 837         layer1_offset = freemap->phys_offset +
 838                         HAMMER_BLOCKMAP_LAYER1_OFFSET(blockmap->next_offset);
 839
 840         layer1 = get_buffer_data(layer1_offset, &buffer1, 0);
 841
 842         if (layer1->phys_offset == HAMMER_BLOCKMAP_UNAVAIL) {
 843                 fprintf(stderr, "alloc_blockmap: ran out of space!\n");
 844                 exit(1);
 845         }
 846
 847         /*
 848          * Dive layer 2
 849          */
 850         layer2_offset = layer1->phys_offset +
 851                         HAMMER_BLOCKMAP_LAYER2_OFFSET(blockmap->next_offset);
 852
 853         layer2 = get_buffer_data(layer2_offset, &buffer2, 0);
 854
 855         if (layer2->zone == HAMMER_ZONE_UNAVAIL_INDEX) {
 856                 fprintf(stderr, "alloc_blockmap: ran out of space!\n");
 857                 exit(1);
 858         }
 859
 860         /*
 861          * If we are entering a new bigblock assign ownership to our
 862          * zone.  If the bigblock is owned by another zone skip it.
 863          */
 864         if (layer2->zone == 0) {
 865                 --layer1->blocks_free;
 866                 layer2->zone = zone;
 867                 assert(layer2->bytes_free == HAMMER_BIGBLOCK_SIZE);
 868                 assert(layer2->append_off == 0);
 869         }
 870         if (layer2->zone != zone) {
 871                 blockmap->next_offset = (blockmap->next_offset + HAMMER_BIGBLOCK_SIZE) &
 872                                         ~HAMMER_BIGBLOCK_MASK64;
 873                 goto again;
 874         }
 875
 876         buffer1->cache.modified = 1;
 877         buffer2->cache.modified = 1;
 878         volume->cache.modified = 1;
 879         assert(layer2->append_off ==
 880                (blockmap->next_offset & HAMMER_BIGBLOCK_MASK));
 881         layer2->bytes_free -= bytes;
 882         *result_offp = blockmap->next_offset;
 883         blockmap->next_offset += bytes;
 884         layer2->append_off = (int)blockmap->next_offset &
 885                               HAMMER_BIGBLOCK_MASK;
 886
 887         layer1->layer1_crc = crc32(layer1, HAMMER_LAYER1_CRCSIZE);
 888         layer2->entry_crc = crc32(layer2, HAMMER_LAYER2_CRCSIZE);
 889
 890         zone2_offset = HAMMER_ZONE_ENCODE(zone,
 891                         *result_offp & ~HAMMER_OFF_ZONE_MASK);
 892
 893         ptr = get_buffer_data(zone2_offset, bufferp, 0);
 894         (*bufferp)->cache.modified = 1;
 895
 896         rel_buffer(buffer1);
 897         rel_buffer(buffer2);
 898         rel_volume(volume);
 899         return(ptr);
 900 }
 901
 902 /*
 903  * Flush various tracking structures to disk
 904  */
 905 void
 906 flush_all_volumes(void)
 907 {
 908         struct volume_info *vol;
 909
 910         TAILQ_FOREACH(vol, &VolList, entry)
 911                 flush_volume(vol);
 912 }
 913
 914 void
 915 flush_volume(struct volume_info *volume)
 916 {
 917         struct buffer_info *buffer;
 918         int i;
 919
 920         for (i = 0; i < HAMMER_BUFLISTS; ++i) {
 921                 TAILQ_FOREACH(buffer, &volume->buffer_lists[i], entry)
 922                         flush_buffer(buffer);
 923         }
 924         writehammerbuf(volume, volume->ondisk, 0);
 925         volume->cache.modified = 0;
 926 }
 927
 928 void
 929 flush_buffer(struct buffer_info *buffer)
 930 {
 931         writehammerbuf(buffer->volume, buffer->ondisk, buffer->raw_offset);
 932         buffer->cache.modified = 0;
 933 }
 934
 935 #if 0
 936 /*
 937  * Generic buffer initialization
 938  */
 939 static void
 940 init_fifo_head(hammer_fifo_head_t head, u_int16_t hdr_type)
 941 {
 942         head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 943         head->hdr_type = hdr_type;
 944         head->hdr_size = 0;
 945         head->hdr_crc = 0;
 946         head->hdr_seq = 0;
 947 }
 948
 949 #endif
 950
 951 #if 0
 952 /*
 953  * Core I/O operations
 954  */
 955 static void
 956 readhammerbuf(struct volume_info *vol, void *data, int64_t offset)
 957 {
 958         ssize_t n;
 959
 960         n = pread(vol->fd, data, HAMMER_BUFSIZE, offset);
 961         if (n != HAMMER_BUFSIZE)
 962                 err(1, "Read volume %d (%s)", vol->vol_no, vol->name);
 963 }
 964
 965 #endif
 966
 967 static void
 968 writehammerbuf(struct volume_info *vol, const void *data, int64_t offset)
 969 {
 970         ssize_t n;
 971
 972         n = pwrite(vol->fd, data, HAMMER_BUFSIZE, offset);
 973         if (n != HAMMER_BUFSIZE)
 974                 err(1, "Write volume %d (%s)", vol->vol_no, vol->name);
 975 }
 976
 977 void
 978 panic(const char *ctl, ...)
 979 {
 980         va_list va;
 981
 982         va_start(va, ctl);
 983         vfprintf(stderr, ctl, va);
 984         va_end(va);
 985         fprintf(stderr, "\n");
 986         exit(1);
 987 }
 988