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