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