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