git.sesse.net Git - plocate/blob - database-builder.cpp

   1 #include "database-builder.h"
   2
   3 #include "dprintf.h"
   4 #include "turbopfor-encode.h"
   5
   6 #include <algorithm>
   7 #include <assert.h>
   8 #ifdef HAS_ENDIAN_H
   9 #include <endian.h>
  10 #endif
  11 #include <fcntl.h>
  12 #include <string.h>
  13 #include <string_view>
  14 #include <sys/stat.h>
  15 #include <sys/time.h>
  16 #include <sys/types.h>
  17 #include <unistd.h>
  18 #include <zdict.h>
  19 #include <zstd.h>
  20
  21 #define P4NENC_BOUND(n) ((n + 127) / 128 + (n + 32) * sizeof(uint32_t))
  22
  23 #define NUM_TRIGRAMS 16777216
  24
  25 using namespace std;
  26 using namespace std::chrono;
  27
  28 constexpr unsigned num_overflow_slots = 16;
  29
  30 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf);
  31
  32 class PostingListBuilder {
  33 public:
  34         inline void add_docid(uint32_t docid);
  35         inline void add_first_docid(uint32_t docid);
  36         void finish();
  37
  38         vector<unsigned char> encoded;
  39         size_t get_num_docids() const
  40         {
  41                 // Updated only when we flush, so check that we're finished.
  42                 assert(pending_deltas.empty());
  43                 return num_docids;
  44         }
  45
  46 private:
  47         void write_header(uint32_t docid);
  48         void append_block();
  49
  50         vector<uint32_t> pending_deltas;
  51
  52         uint32_t num_docids = 0;  // Should be size_t, except the format only supports 2^32 docids per posting list anyway.
  53         uint32_t last_docid = -1;
  54 };
  55
  56 void PostingListBuilder::add_docid(uint32_t docid)
  57 {
  58         // Deduplicate against the last inserted value, if any.
  59         if (docid == last_docid) {
  60                 return;
  61         }
  62
  63         pending_deltas.push_back(docid - last_docid - 1);
  64         last_docid = docid;
  65         if (pending_deltas.size() == 128) {
  66                 append_block();
  67                 pending_deltas.clear();
  68                 num_docids += 128;
  69         }
  70 }
  71
  72 void PostingListBuilder::add_first_docid(uint32_t docid)
  73 {
  74         write_header(docid);
  75         ++num_docids;
  76         last_docid = docid;
  77 }
  78
  79 void PostingListBuilder::finish()
  80 {
  81         if (pending_deltas.empty()) {
  82                 return;
  83         }
  84
  85         assert(!encoded.empty());  // write_header() should already have run.
  86
  87         // No interleaving for partial blocks.
  88         unsigned char buf[P4NENC_BOUND(128)];
  89         unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), pending_deltas.size(), /*interleaved=*/false, buf);
  90         encoded.insert(encoded.end(), buf, end);
  91
  92         num_docids += pending_deltas.size();
  93         pending_deltas.clear();
  94 }
  95
  96 void PostingListBuilder::append_block()
  97 {
  98         unsigned char buf[P4NENC_BOUND(128)];
  99         assert(pending_deltas.size() == 128);
 100         unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), 128, /*interleaved=*/true, buf);
 101         encoded.insert(encoded.end(), buf, end);
 102 }
 103
 104 void PostingListBuilder::write_header(uint32_t docid)
 105 {
 106         unsigned char buf[P4NENC_BOUND(1)];
 107         unsigned char *end = write_baseval(docid, buf);
 108         encoded.insert(encoded.end(), buf, end);
 109 }
 110
 111 void DictionaryBuilder::add_file(string filename, dir_time)
 112 {
 113         if (keep_current_block) {  // Only bother saving the filenames if we're actually keeping the block.
 114                 if (!current_block.empty()) {
 115                         current_block.push_back('\0');
 116                 }
 117                 current_block += filename;
 118         }
 119         if (++num_files_in_block == block_size) {
 120                 flush_block();
 121         }
 122 }
 123
 124 void DictionaryBuilder::flush_block()
 125 {
 126         if (keep_current_block) {
 127                 if (slot_for_current_block == -1) {
 128                         lengths.push_back(current_block.size());
 129                         sampled_blocks.push_back(move(current_block));
 130                 } else {
 131                         lengths[slot_for_current_block] = current_block.size();
 132                         sampled_blocks[slot_for_current_block] = move(current_block);
 133                 }
 134         }
 135         current_block.clear();
 136         num_files_in_block = 0;
 137         ++block_num;
 138
 139         if (block_num < blocks_to_keep) {
 140                 keep_current_block = true;
 141                 slot_for_current_block = -1;
 142         } else {
 143                 // Keep every block with equal probability (reservoir sampling).
 144                 uint64_t idx = uniform_int_distribution<uint64_t>(0, block_num)(reservoir_rand);
 145                 keep_current_block = (idx < blocks_to_keep);
 146                 slot_for_current_block = idx;
 147         }
 148 }
 149
 150 string DictionaryBuilder::train(size_t buf_size)
 151 {
 152         string dictionary_buf;
 153         sort(sampled_blocks.begin(), sampled_blocks.end());  // Seemingly important for decompression speed.
 154         for (const string &block : sampled_blocks) {
 155                 dictionary_buf += block;
 156         }
 157
 158         string buf;
 159         buf.resize(buf_size);
 160         size_t ret = ZDICT_trainFromBuffer(&buf[0], buf_size, dictionary_buf.data(), lengths.data(), lengths.size());
 161         if (ZDICT_isError(ret)) {
 162                 return "";
 163         }
 164         dprintf("Sampled %zu bytes in %zu blocks, built a dictionary of size %zu\n", dictionary_buf.size(), lengths.size(), ret);
 165         buf.resize(ret);
 166
 167         sampled_blocks.clear();
 168         lengths.clear();
 169
 170         return buf;
 171 }
 172
 173 class EncodingCorpus : public DatabaseReceiver {
 174 public:
 175         EncodingCorpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict, bool store_dir_times);
 176         ~EncodingCorpus();
 177
 178         void add_file(std::string filename, dir_time dt) override;
 179         void flush_block() override;
 180         void finish() override;
 181
 182         std::vector<uint64_t> filename_blocks;
 183         size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
 184         bool seen_trigram(uint32_t trgm)
 185         {
 186                 return invindex[trgm] != nullptr;
 187         }
 188         size_t num_files_seen() const override { return num_files; }
 189         PostingListBuilder &get_pl_builder(uint32_t trgm)
 190         {
 191                 return *invindex[trgm];
 192         }
 193
 194         void add_docid(uint32_t trgm, uint32_t docid)
 195         {
 196                 if (invindex[trgm] == nullptr) {
 197                         invindex[trgm] = new PostingListBuilder;
 198                         invindex[trgm]->add_first_docid(docid);
 199                 } else {
 200                         invindex[trgm]->add_docid(docid);
 201                 }
 202         }
 203
 204         size_t num_trigrams() const;
 205         std::string get_compressed_dir_times();
 206
 207 private:
 208         void compress_dir_times(size_t allowed_slop);
 209
 210         std::unique_ptr<PostingListBuilder *[]> invindex;
 211         FILE *outfp;
 212         off_t outfp_pos;  // Cheaper than calling ftell(outfp) all the time.
 213         std::string current_block;
 214         std::string tempbuf;
 215         const size_t block_size;
 216         const bool store_dir_times;
 217         ZSTD_CDict *cdict;
 218
 219         ZSTD_CStream *dir_time_ctx = nullptr;
 220         std::string dir_times;  // Buffer of still-uncompressed data.
 221         std::string dir_times_compressed;
 222 };
 223
 224 EncodingCorpus::EncodingCorpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict, bool store_dir_times)
 225         : invindex(new PostingListBuilder *[NUM_TRIGRAMS]), outfp(outfp), outfp_pos(ftell(outfp)), block_size(block_size), store_dir_times(store_dir_times), cdict(cdict)
 226 {
 227         fill(invindex.get(), invindex.get() + NUM_TRIGRAMS, nullptr);
 228         if (store_dir_times) {
 229                 dir_time_ctx = ZSTD_createCStream();
 230                 ZSTD_initCStream(dir_time_ctx, /*level=*/6);
 231         }
 232 }
 233
 234 EncodingCorpus::~EncodingCorpus()
 235 {
 236         for (unsigned i = 0; i < NUM_TRIGRAMS; ++i) {
 237                 delete invindex[i];
 238         }
 239 }
 240
 241 void EncodingCorpus::add_file(string filename, dir_time dt)
 242 {
 243         ++num_files;
 244         if (!current_block.empty()) {
 245                 current_block.push_back('\0');
 246         }
 247         current_block += filename;
 248         if (++num_files_in_block == block_size) {
 249                 flush_block();
 250         }
 251
 252         if (store_dir_times) {
 253                 if (dt.sec == -1) {
 254                         // Not a directory.
 255                         dir_times.push_back('\0');
 256                 } else {
 257                         dir_times.push_back('\1');
 258                         dir_times.append(reinterpret_cast<char *>(&dt.sec), sizeof(dt.sec));
 259                         dir_times.append(reinterpret_cast<char *>(&dt.nsec), sizeof(dt.nsec));
 260                 }
 261                 compress_dir_times(/*allowed_slop=*/4096);
 262         }
 263 }
 264
 265 void EncodingCorpus::compress_dir_times(size_t allowed_slop)
 266 {
 267         while (dir_times.size() >= allowed_slop) {
 268                 size_t old_size = dir_times_compressed.size();
 269                 dir_times_compressed.resize(old_size + 4096);
 270
 271                 ZSTD_outBuffer outbuf;
 272                 outbuf.dst = dir_times_compressed.data() + old_size;
 273                 outbuf.size = 4096;
 274                 outbuf.pos = 0;
 275
 276                 ZSTD_inBuffer inbuf;
 277                 inbuf.src = dir_times.data();
 278                 inbuf.size = dir_times.size();
 279                 inbuf.pos = 0;
 280
 281                 int ret = ZSTD_compressStream(dir_time_ctx, &outbuf, &inbuf);
 282                 if (ret < 0) {
 283                         fprintf(stderr, "ZSTD_compressStream() failed\n");
 284                         exit(1);
 285                 }
 286
 287                 dir_times_compressed.resize(old_size + outbuf.pos);
 288                 dir_times.erase(dir_times.begin(), dir_times.begin() + inbuf.pos);
 289
 290                 if (outbuf.pos == 0 && inbuf.pos == 0) {
 291                         // Nothing happened (not enough data?), try again later.
 292                         return;
 293                 }
 294         }
 295 }
 296
 297 void EncodingCorpus::flush_block()
 298 {
 299         if (current_block.empty()) {
 300                 return;
 301         }
 302
 303         uint32_t docid = num_blocks;
 304
 305         // Create trigrams.
 306         const char *ptr = current_block.c_str();
 307         const char *end = ptr + current_block.size();
 308         while (ptr < end - 3) {  // Must be at least one filename left, that's at least three bytes.
 309                 if (ptr[0] == '\0') {
 310                         // This filename is zero bytes, so skip it (and the zero terminator).
 311                         ++ptr;
 312                         continue;
 313                 } else if (ptr[1] == '\0') {
 314                         // This filename is one byte, so skip it (and the zero terminator).
 315                         ptr += 2;
 316                         continue;
 317                 } else if (ptr[2] == '\0') {
 318                         // This filename is two bytes, so skip it (and the zero terminator).
 319                         ptr += 3;
 320                         continue;
 321                 }
 322                 for (;;) {
 323                         // NOTE: Will read one byte past the end of the trigram, but it's OK,
 324                         // since we always call it from contexts where there's a terminating zero byte.
 325                         uint32_t trgm;
 326                         memcpy(&trgm, ptr, sizeof(trgm));
 327                         ++ptr;
 328                         trgm = le32toh(trgm);
 329                         add_docid(trgm & 0xffffff, docid);
 330                         if (trgm <= 0xffffff) {
 331                                 // Terminating zero byte, so we're done with this filename.
 332                                 // Skip the remaining two bytes, and the zero terminator.
 333                                 ptr += 3;
 334                                 break;
 335                         }
 336                 }
 337         }
 338
 339         // Compress and add the filename block.
 340         filename_blocks.push_back(outfp_pos);
 341         string compressed = zstd_compress(current_block, cdict, &tempbuf);
 342         if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
 343                 perror("fwrite()");
 344                 exit(1);
 345         }
 346         outfp_pos += compressed.size();
 347
 348         current_block.clear();
 349         num_files_in_block = 0;
 350         ++num_blocks;
 351 }
 352
 353 void EncodingCorpus::finish()
 354 {
 355         flush_block();
 356 }
 357
 358 size_t EncodingCorpus::num_trigrams() const
 359 {
 360         size_t num = 0;
 361         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 362                 if (invindex[trgm] != nullptr) {
 363                         ++num;
 364                 }
 365         }
 366         return num;
 367 }
 368
 369 string EncodingCorpus::get_compressed_dir_times()
 370 {
 371         if (!store_dir_times) {
 372                 return "";
 373         }
 374         compress_dir_times(/*allowed_slop=*/0);
 375         assert(dir_times.empty());
 376
 377         for (;;) {
 378                 size_t old_size = dir_times_compressed.size();
 379                 dir_times_compressed.resize(old_size + 4096);
 380
 381                 ZSTD_outBuffer outbuf;
 382                 outbuf.dst = dir_times_compressed.data() + old_size;
 383                 outbuf.size = 4096;
 384                 outbuf.pos = 0;
 385
 386                 int ret = ZSTD_endStream(dir_time_ctx, &outbuf);
 387                 if (ret < 0) {
 388                         fprintf(stderr, "ZSTD_compressStream() failed\n");
 389                         exit(1);
 390                 }
 391
 392                 dir_times_compressed.resize(old_size + outbuf.pos);
 393
 394                 if (ret == 0) {
 395                         // All done.
 396                         break;
 397                 }
 398         }
 399
 400         return dir_times_compressed;
 401 }
 402
 403 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf)
 404 {
 405         static ZSTD_CCtx *ctx = nullptr;
 406         if (ctx == nullptr) {
 407                 ctx = ZSTD_createCCtx();
 408         }
 409
 410         size_t max_size = ZSTD_compressBound(src.size());
 411         if (tempbuf->size() < max_size) {
 412                 tempbuf->resize(max_size);
 413         }
 414         size_t size;
 415         if (cdict == nullptr) {
 416                 size = ZSTD_compressCCtx(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
 417         } else {
 418                 size = ZSTD_compress_usingCDict(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), cdict);
 419         }
 420         return string(tempbuf->data(), size);
 421 }
 422
 423 bool is_prime(uint32_t x)
 424 {
 425         if ((x % 2) == 0 || (x % 3) == 0) {
 426                 return false;
 427         }
 428         uint32_t limit = ceil(sqrt(x));
 429         for (uint32_t factor = 5; factor <= limit; ++factor) {
 430                 if ((x % factor) == 0) {
 431                         return false;
 432                 }
 433         }
 434         return true;
 435 }
 436
 437 uint32_t next_prime(uint32_t x)
 438 {
 439         if ((x % 2) == 0) {
 440                 ++x;
 441         }
 442         while (!is_prime(x)) {
 443                 x += 2;
 444         }
 445         return x;
 446 }
 447
 448 unique_ptr<Trigram[]> create_hashtable(EncodingCorpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
 449 {
 450         unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]);  // 1 for the sentinel element at the end.
 451         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 452                 ht[i].trgm = uint32_t(-1);
 453                 ht[i].num_docids = 0;
 454                 ht[i].offset = 0;
 455         }
 456         for (uint32_t trgm : all_trigrams) {
 457                 // We don't know offset yet, so set it to zero.
 458                 Trigram to_insert{ trgm, uint32_t(corpus.get_pl_builder(trgm).get_num_docids()), 0 };
 459
 460                 uint32_t bucket = hash_trigram(trgm, ht_size);
 461                 unsigned distance = 0;
 462                 while (ht[bucket].num_docids != 0) {
 463                         // Robin Hood hashing; reduces the longest distance by a lot.
 464                         unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
 465                         if (distance > other_distance) {
 466                                 swap(to_insert, ht[bucket]);
 467                                 distance = other_distance;
 468                         }
 469
 470                         ++bucket, ++distance;
 471                         if (distance > num_overflow_slots) {
 472                                 return nullptr;
 473                         }
 474                 }
 475                 ht[bucket] = to_insert;
 476         }
 477         return ht;
 478 }
 479
 480 DatabaseBuilder::DatabaseBuilder(const char *outfile, gid_t owner, int block_size, string dictionary, bool check_visibility)
 481         : outfile(outfile), block_size(block_size)
 482 {
 483         umask(0027);
 484
 485         string path = outfile;
 486         path.resize(path.find_last_of('/') + 1);
 487         if (path.empty()) {
 488                 path = ".";
 489         }
 490         int fd = -1;
 491 #ifdef O_TMPFILE
 492         fd = open(path.c_str(), O_WRONLY | O_TMPFILE, 0640);
 493         if (fd == -1 && errno != EOPNOTSUPP && errno != EISDIR) {
 494                 perror(path.c_str());
 495                 exit(1);
 496         }
 497 #endif
 498         if (fd == -1) {
 499                 temp_filename = string(outfile) + ".XXXXXX";
 500                 fd = mkstemp(&temp_filename[0]);
 501                 if (fd == -1) {
 502                         perror(temp_filename.c_str());
 503                         exit(1);
 504                 }
 505                 if (fchmod(fd, 0640) == -1) {
 506                         perror("fchmod");
 507                         exit(1);
 508                 }
 509         }
 510
 511         if (owner != (gid_t)-1) {
 512                 if (fchown(fd, (uid_t)-1, owner) == -1) {
 513                         perror("fchown");
 514                         exit(1);
 515                 }
 516         }
 517
 518         outfp = fdopen(fd, "wb");
 519         if (outfp == nullptr) {
 520                 perror(outfile);
 521                 exit(1);
 522         }
 523
 524         // Write the header.
 525         memcpy(hdr.magic, "\0plocate", 8);
 526         hdr.version = -1;  // Mark as broken.
 527         hdr.hashtable_size = 0;  // Not known yet.
 528         hdr.extra_ht_slots = num_overflow_slots;
 529         hdr.num_docids = 0;
 530         hdr.hash_table_offset_bytes = -1;  // We don't know these offsets yet.
 531         hdr.max_version = 2;
 532         hdr.filename_index_offset_bytes = -1;
 533         hdr.zstd_dictionary_length_bytes = -1;
 534         hdr.check_visibility = check_visibility;
 535         fwrite(&hdr, sizeof(hdr), 1, outfp);
 536
 537         if (dictionary.empty()) {
 538                 hdr.zstd_dictionary_offset_bytes = 0;
 539                 hdr.zstd_dictionary_length_bytes = 0;
 540         } else {
 541                 hdr.zstd_dictionary_offset_bytes = ftell(outfp);
 542                 fwrite(dictionary.data(), dictionary.size(), 1, outfp);
 543                 hdr.zstd_dictionary_length_bytes = dictionary.size();
 544                 cdict = ZSTD_createCDict(dictionary.data(), dictionary.size(), /*level=*/6);
 545         }
 546
 547         hdr.directory_data_length_bytes = 0;
 548         hdr.directory_data_offset_bytes = 0;
 549         hdr.next_zstd_dictionary_length_bytes = 0;
 550         hdr.next_zstd_dictionary_offset_bytes = 0;
 551         hdr.conf_block_length_bytes = 0;
 552         hdr.conf_block_offset_bytes = 0;
 553 }
 554
 555 DatabaseReceiver *DatabaseBuilder::start_corpus(bool store_dir_times)
 556 {
 557         corpus_start = steady_clock::now();
 558         corpus = new EncodingCorpus(outfp, block_size, cdict, store_dir_times);
 559         return corpus;
 560 }
 561
 562 void DatabaseBuilder::set_next_dictionary(std::string next_dictionary)
 563 {
 564         this->next_dictionary = move(next_dictionary);
 565 }
 566
 567 void DatabaseBuilder::set_conf_block(std::string conf_block)
 568 {
 569         this->conf_block = move(conf_block);
 570 }
 571
 572 void DatabaseBuilder::finish_corpus()
 573 {
 574         corpus->finish();
 575         hdr.num_docids = corpus->filename_blocks.size();
 576
 577         // Stick an empty block at the end as sentinel.
 578         corpus->filename_blocks.push_back(ftell(outfp));
 579         const size_t bytes_for_filenames = corpus->filename_blocks.back() - corpus->filename_blocks.front();
 580
 581         // Write the offsets to the filenames.
 582         hdr.filename_index_offset_bytes = ftell(outfp);
 583         const size_t bytes_for_filename_index = corpus->filename_blocks.size() * sizeof(uint64_t);
 584         fwrite(corpus->filename_blocks.data(), corpus->filename_blocks.size(), sizeof(uint64_t), outfp);
 585         corpus->filename_blocks.clear();
 586         corpus->filename_blocks.shrink_to_fit();
 587
 588         // Finish up encoding the posting lists.
 589         size_t trigrams = 0, longest_posting_list = 0;
 590         size_t bytes_for_posting_lists = 0;
 591         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 592                 if (!corpus->seen_trigram(trgm))
 593                         continue;
 594                 PostingListBuilder &pl_builder = corpus->get_pl_builder(trgm);
 595                 pl_builder.finish();
 596                 longest_posting_list = max(longest_posting_list, pl_builder.get_num_docids());
 597                 trigrams += pl_builder.get_num_docids();
 598                 bytes_for_posting_lists += pl_builder.encoded.size();
 599         }
 600         size_t num_trigrams = corpus->num_trigrams();
 601         dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
 602                 corpus->num_files, num_trigrams, trigrams, double(trigrams) / num_trigrams, longest_posting_list);
 603         dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));
 604
 605         dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - corpus_start).count());
 606
 607         // Find the used trigrams.
 608         vector<uint32_t> all_trigrams;
 609         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 610                 if (corpus->seen_trigram(trgm)) {
 611                         all_trigrams.push_back(trgm);
 612                 }
 613         }
 614
 615         // Create the hash table.
 616         unique_ptr<Trigram[]> hashtable;
 617         uint32_t ht_size = next_prime(all_trigrams.size());
 618         for (;;) {
 619                 hashtable = create_hashtable(*corpus, all_trigrams, ht_size, num_overflow_slots);
 620                 if (hashtable == nullptr) {
 621                         dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
 622                         ht_size = next_prime(ht_size * 1.05);
 623                 } else {
 624                         dprintf("Created hash table of size %u.\n\n", ht_size);
 625                         break;
 626                 }
 627         }
 628
 629         // Find the offsets for each posting list.
 630         size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
 631         uint64_t offset = ftell(outfp) + bytes_for_hashtable;
 632         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 633                 hashtable[i].offset = offset;  // Needs to be there even for empty slots.
 634                 if (hashtable[i].num_docids == 0) {
 635                         continue;
 636                 }
 637
 638                 const vector<unsigned char> &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
 639                 offset += encoded.size();
 640         }
 641
 642         // Write the hash table.
 643         hdr.hash_table_offset_bytes = ftell(outfp);
 644         hdr.hashtable_size = ht_size;
 645         fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
 646
 647         // Write the actual posting lists.
 648         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 649                 if (hashtable[i].num_docids == 0) {
 650                         continue;
 651                 }
 652                 const vector<unsigned char> &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
 653                 fwrite(encoded.data(), encoded.size(), 1, outfp);
 654         }
 655
 656         // Finally, write the directory times (for updatedb).
 657         string compressed_dir_times = corpus->get_compressed_dir_times();
 658         size_t bytes_for_compressed_dir_times = 0;
 659         if (!compressed_dir_times.empty()) {
 660                 hdr.directory_data_offset_bytes = ftell(outfp);
 661                 hdr.directory_data_length_bytes = compressed_dir_times.size();
 662                 fwrite(compressed_dir_times.data(), compressed_dir_times.size(), 1, outfp);
 663                 bytes_for_compressed_dir_times = compressed_dir_times.size();
 664                 compressed_dir_times.clear();
 665         }
 666
 667         // Write the recommended dictionary for next update.
 668         if (!next_dictionary.empty()) {
 669                 hdr.next_zstd_dictionary_offset_bytes = ftell(outfp);
 670                 hdr.next_zstd_dictionary_length_bytes = next_dictionary.size();
 671                 fwrite(next_dictionary.data(), next_dictionary.size(), 1, outfp);
 672         }
 673
 674         // And the configuration block.
 675         if (!conf_block.empty()) {
 676                 hdr.conf_block_offset_bytes = ftell(outfp);
 677                 hdr.conf_block_length_bytes = conf_block.size();
 678                 fwrite(conf_block.data(), conf_block.size(), 1, outfp);
 679         }
 680
 681         // Rewind, and write the updated header.
 682         hdr.version = 1;
 683         fseek(outfp, 0, SEEK_SET);
 684         fwrite(&hdr, sizeof(hdr), 1, outfp);
 685
 686         if (!temp_filename.empty()) {
 687                 if (rename(temp_filename.c_str(), outfile.c_str()) == -1) {
 688                         perror("rename");
 689                         exit(1);
 690                 }
 691         } else {
 692 #ifdef O_TMPFILE
 693                 // Give the file a proper name, making it visible in the file system.
 694                 // TODO: It would be nice to be able to do this atomically, like with rename.
 695                 unlink(outfile.c_str());
 696                 char procpath[256];
 697                 snprintf(procpath, sizeof(procpath), "/proc/self/fd/%d", fileno(outfp));
 698                 if (linkat(AT_FDCWD, procpath, AT_FDCWD, outfile.c_str(), AT_SYMLINK_FOLLOW) == -1) {
 699                         perror("linkat");
 700                         exit(1);
 701                 }
 702 #endif
 703         }
 704
 705         fclose(outfp);
 706
 707         size_t total_bytes = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames + bytes_for_compressed_dir_times);
 708
 709         dprintf("Block size:     %7d files\n", block_size);
 710         dprintf("Dictionary:     %'7.1f MB\n", hdr.zstd_dictionary_length_bytes / 1048576.0);
 711         dprintf("Hash table:     %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
 712         dprintf("Posting lists:  %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
 713         dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
 714         dprintf("Filenames:      %'7.1f MB\n", bytes_for_filenames / 1048576.0);
 715         if (bytes_for_compressed_dir_times != 0) {
 716                 dprintf("Modify times:   %'7.1f MB\n", bytes_for_compressed_dir_times / 1048576.0);
 717         }
 718         dprintf("Total:          %'7.1f MB\n", total_bytes / 1048576.0);
 719         dprintf("\n");
 720 }