git.sesse.net Git - plocate/blob - plocate-build.cpp

   1 #include "db.h"
   2 #include "vp4.h"
   3
   4 #include <algorithm>
   5 #include <arpa/inet.h>
   6 #include <assert.h>
   7 #include <chrono>
   8 #include <endian.h>
   9 #include <fcntl.h>
  10 #include <math.h>
  11 #include <memory>
  12 #include <stdio.h>
  13 #include <string.h>
  14 #include <string>
  15 #include <sys/mman.h>
  16 #include <sys/stat.h>
  17 #include <sys/types.h>
  18 #include <unistd.h>
  19 #include <unordered_map>
  20 #include <vector>
  21 #include <zstd.h>
  22
  23 #define P4NENC_BOUND(n) ((n + 127) / 128 + (n + 32) * sizeof(uint32_t))
  24 #define dprintf(...)
  25 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
  26
  27 using namespace std;
  28 using namespace std::chrono;
  29
  30 string zstd_compress(const string &src, string *tempbuf);
  31
  32 constexpr unsigned num_overflow_slots = 16;
  33
  34 static inline uint32_t read_unigram(const string_view s, size_t idx)
  35 {
  36         if (idx < s.size()) {
  37                 return (unsigned char)s[idx];
  38         } else {
  39                 return 0;
  40         }
  41 }
  42
  43 static inline uint32_t read_trigram(const string_view s, size_t start)
  44 {
  45         return read_unigram(s, start) |
  46                 (read_unigram(s, start + 1) << 8) |
  47                 (read_unigram(s, start + 2) << 16);
  48 }
  49
  50 enum {
  51         DBE_NORMAL = 0, /* A non-directory file */
  52         DBE_DIRECTORY = 1, /* A directory */
  53         DBE_END = 2 /* End of directory contents; contains no name */
  54 };
  55
  56 // From mlocate.
  57 struct db_header {
  58         uint8_t magic[8];
  59         uint32_t conf_size;
  60         uint8_t version;
  61         uint8_t check_visibility;
  62         uint8_t pad[2];
  63 };
  64
  65 // From mlocate.
  66 struct db_directory {
  67         uint64_t time_sec;
  68         uint32_t time_nsec;
  69         uint8_t pad[4];
  70 };
  71
  72 class PostingListBuilder {
  73 public:
  74         void add_docid(uint32_t docid);
  75         void finish();
  76
  77         string encoded;
  78         size_t num_docids = 0;
  79
  80 private:
  81         void write_header(uint32_t docid);
  82         void append_block();
  83
  84         vector<uint32_t> pending_docids;
  85
  86         uint32_t last_block_end;
  87 };
  88
  89 void PostingListBuilder::add_docid(uint32_t docid)
  90 {
  91         // Deduplicate against the last inserted value, if any.
  92         if (pending_docids.empty()) {
  93                 if (encoded.empty()) {
  94                         // Very first docid.
  95                         write_header(docid);
  96                         ++num_docids;
  97                         last_block_end = docid;
  98                         return;
  99                 } else if (docid == last_block_end) {
 100                         return;
 101                 }
 102         } else {
 103                 if (docid == pending_docids.back()) {
 104                         return;
 105                 }
 106         }
 107
 108         pending_docids.push_back(docid);
 109         if (pending_docids.size() == 128) {
 110                 append_block();
 111                 pending_docids.clear();
 112                 last_block_end = docid;
 113         }
 114         ++num_docids;
 115 }
 116
 117 void PostingListBuilder::finish()
 118 {
 119         if (pending_docids.empty()) {
 120                 return;
 121         }
 122
 123         assert(!encoded.empty());  // write_header() should already have run.
 124
 125         // No interleaving for partial blocks.
 126         unsigned char buf[P4NENC_BOUND(128)];
 127         unsigned char *end = p4d1enc32(pending_docids.data(), pending_docids.size(), buf, last_block_end);
 128         encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
 129 }
 130
 131 void PostingListBuilder::append_block()
 132 {
 133         unsigned char buf[P4NENC_BOUND(128)];
 134         assert(pending_docids.size() == 128);
 135         unsigned char *end = p4d1enc128v32(pending_docids.data(), 128, buf, last_block_end);
 136         encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
 137 }
 138
 139 void PostingListBuilder::write_header(uint32_t docid)
 140 {
 141         unsigned char buf[P4NENC_BOUND(1)];
 142         size_t bytes = p4nd1enc128v32(&docid, 1, buf);
 143         encoded.append(reinterpret_cast<char *>(buf), bytes);
 144 }
 145
 146 class Corpus {
 147 public:
 148         Corpus(FILE *outfp, size_t block_size)
 149                 : outfp(outfp), block_size(block_size) {}
 150         void add_file(string filename);
 151         void flush_block();
 152
 153         vector<uint64_t> filename_blocks;
 154         unordered_map<uint32_t, PostingListBuilder> invindex;
 155         size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
 156
 157 private:
 158         FILE *outfp;
 159         string current_block;
 160         string tempbuf;
 161         const size_t block_size;
 162 };
 163
 164 void Corpus::add_file(string filename)
 165 {
 166         ++num_files;
 167         if (!current_block.empty()) {
 168                 current_block.push_back('\0');
 169         }
 170         current_block += filename;
 171         if (++num_files_in_block == block_size) {
 172                 flush_block();
 173         }
 174 }
 175
 176 void Corpus::flush_block()
 177 {
 178         if (current_block.empty()) {
 179                 return;
 180         }
 181
 182         uint32_t docid = num_blocks;
 183
 184         // Create trigrams.
 185         const char *ptr = current_block.c_str();
 186         while (ptr < current_block.c_str() + current_block.size()) {
 187                 string_view s(ptr);
 188                 if (s.size() >= 3) {
 189                         for (size_t j = 0; j < s.size() - 2; ++j) {
 190                                 uint32_t trgm = read_trigram(s, j);
 191                                 invindex[trgm].add_docid(docid);
 192                         }
 193                 }
 194                 ptr += s.size() + 1;
 195         }
 196
 197         // Compress and add the filename block.
 198         filename_blocks.push_back(ftell(outfp));
 199         string compressed = zstd_compress(current_block, &tempbuf);
 200         if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
 201                 perror("fwrite()");
 202                 exit(1);
 203         }
 204
 205         current_block.clear();
 206         num_files_in_block = 0;
 207         ++num_blocks;
 208 }
 209
 210 const char *handle_directory(const char *ptr, Corpus *corpus)
 211 {
 212         ptr += sizeof(db_directory);
 213
 214         string dir_path = ptr;
 215         ptr += dir_path.size() + 1;
 216         if (dir_path == "/") {
 217                 dir_path = "";
 218         }
 219
 220         for (;;) {
 221                 uint8_t type = *ptr++;
 222                 if (type == DBE_NORMAL) {
 223                         string filename = ptr;
 224                         corpus->add_file(dir_path + "/" + filename);
 225                         ptr += filename.size() + 1;
 226                 } else if (type == DBE_DIRECTORY) {
 227                         string dirname = ptr;
 228                         corpus->add_file(dir_path + "/" + dirname);
 229                         ptr += dirname.size() + 1;
 230                 } else {
 231                         return ptr;
 232                 }
 233         }
 234 }
 235
 236 void read_mlocate(const char *filename, Corpus *corpus)
 237 {
 238         int fd = open(filename, O_RDONLY);
 239         if (fd == -1) {
 240                 perror(filename);
 241                 exit(1);
 242         }
 243         off_t len = lseek(fd, 0, SEEK_END);
 244         if (len == -1) {
 245                 perror("lseek");
 246                 exit(1);
 247         }
 248         const char *data = (char *)mmap(nullptr, len, PROT_READ, MAP_SHARED, fd, /*offset=*/0);
 249         if (data == MAP_FAILED) {
 250                 perror("mmap");
 251                 exit(1);
 252         }
 253
 254         const db_header *hdr = (const db_header *)data;
 255
 256         // TODO: Care about the base path.
 257         string path = data + sizeof(db_header);
 258         uint64_t offset = sizeof(db_header) + path.size() + 1 + ntohl(hdr->conf_size);
 259
 260         const char *ptr = data + offset;
 261         while (ptr < data + len) {
 262                 ptr = handle_directory(ptr, corpus);
 263         }
 264
 265         munmap((void *)data, len);
 266         close(fd);
 267 }
 268
 269 string zstd_compress(const string &src, string *tempbuf)
 270 {
 271         size_t max_size = ZSTD_compressBound(src.size());
 272         if (tempbuf->size() < max_size) {
 273                 tempbuf->resize(max_size);
 274         }
 275         size_t size = ZSTD_compress(&(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
 276         return string(tempbuf->data(), size);
 277 }
 278
 279 bool is_prime(uint32_t x)
 280 {
 281         if ((x % 2) == 0 || (x % 3) == 0) {
 282                 return false;
 283         }
 284         uint32_t limit = ceil(sqrt(x));
 285         for (uint32_t factor = 5; factor <= limit; ++factor) {
 286                 if ((x % factor) == 0) {
 287                         return false;
 288                 }
 289         }
 290         return true;
 291 }
 292
 293 uint32_t next_prime(uint32_t x)
 294 {
 295         if ((x % 2) == 0) {
 296                 ++x;
 297         }
 298         while (!is_prime(x)) {
 299                 x += 2;
 300         }
 301         return x;
 302 }
 303
 304 unique_ptr<Trigram[]> create_hashtable(const Corpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
 305 {
 306         unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]);  // 1 for the sentinel element at the end.
 307         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 308                 ht[i].trgm = uint32_t(-1);
 309                 ht[i].num_docids = 0;
 310                 ht[i].offset = 0;
 311         }
 312         for (uint32_t trgm : all_trigrams) {
 313                 // We don't know offset yet, so set it to zero.
 314                 Trigram to_insert{ trgm, uint32_t(corpus.invindex.find(trgm)->second.num_docids), 0 };
 315
 316                 uint32_t bucket = hash_trigram(trgm, ht_size);
 317                 unsigned distance = 0;
 318                 while (ht[bucket].num_docids != 0) {
 319                         // Robin Hood hashing; reduces the longest distance by a lot.
 320                         unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
 321                         if (distance > other_distance) {
 322                                 swap(to_insert, ht[bucket]);
 323                                 distance = other_distance;
 324                         }
 325
 326                         ++bucket, ++distance;
 327                         if (distance > num_overflow_slots) {
 328                                 return nullptr;
 329                         }
 330                 }
 331                 ht[bucket] = to_insert;
 332         }
 333         return ht;
 334 }
 335
 336 void do_build(const char *infile, const char *outfile, int block_size)
 337 {
 338         steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
 339
 340         umask(0027);
 341         FILE *outfp = fopen(outfile, "wb");
 342
 343         // Write the header.
 344         Header hdr;
 345         memcpy(hdr.magic, "\0plocate", 8);
 346         hdr.version = -1;  // Mark as broken.
 347         hdr.hashtable_size = 0;  // Not known yet.
 348         hdr.extra_ht_slots = num_overflow_slots;
 349         hdr.hash_table_offset_bytes = -1;  // We don't know these offsets yet.
 350         hdr.filename_index_offset_bytes = -1;
 351         fwrite(&hdr, sizeof(hdr), 1, outfp);
 352
 353         Corpus corpus(outfp, block_size);
 354
 355         read_mlocate(infile, &corpus);
 356         if (false) {  // To read a plain text file.
 357                 FILE *fp = fopen(infile, "r");
 358                 while (!feof(fp)) {
 359                         char buf[1024];
 360                         if (fgets(buf, 1024, fp) == nullptr || feof(fp)) {
 361                                 break;
 362                         }
 363                         string s(buf);
 364                         if (s.back() == '\n')
 365                                 s.pop_back();
 366                         corpus.add_file(move(s));
 367                 }
 368                 fclose(fp);
 369         }
 370         corpus.flush_block();
 371         dprintf("Read %zu files from %s\n", corpus.num_files, infile);
 372
 373         // Stick an empty block at the end as sentinel.
 374         corpus.filename_blocks.push_back(ftell(outfp));
 375         const size_t bytes_for_filenames = corpus.filename_blocks.back() - corpus.filename_blocks.front();
 376
 377         // Write the offsets to the filenames.
 378         hdr.filename_index_offset_bytes = ftell(outfp);
 379         const size_t bytes_for_filename_index = corpus.filename_blocks.size() * sizeof(uint64_t);
 380         fwrite(corpus.filename_blocks.data(), corpus.filename_blocks.size(), sizeof(uint64_t), outfp);
 381         corpus.filename_blocks.clear();
 382         corpus.filename_blocks.shrink_to_fit();
 383
 384         // Finish up encoding the posting lists.
 385         size_t trigrams = 0, longest_posting_list = 0;
 386         size_t bytes_for_posting_lists = 0;
 387         for (auto &[trigram, pl_builder] : corpus.invindex) {
 388                 pl_builder.finish();
 389                 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
 390                 trigrams += pl_builder.num_docids;
 391                 bytes_for_posting_lists += pl_builder.encoded.size();
 392         }
 393         dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
 394                 corpus.num_files, corpus.invindex.size(), trigrams, double(trigrams) / corpus.invindex.size(), longest_posting_list);
 395         dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));
 396
 397         dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 398
 399         // Sort the trigrams, mostly to get a consistent result every time
 400         // (the hash table will put things in random order anyway).
 401         vector<uint32_t> all_trigrams;
 402         for (auto &[trigram, pl_builder] : corpus.invindex) {
 403                 all_trigrams.push_back(trigram);
 404         }
 405         sort(all_trigrams.begin(), all_trigrams.end());
 406
 407         // Create the hash table.
 408         unique_ptr<Trigram[]> hashtable;
 409         uint32_t ht_size = next_prime(all_trigrams.size());
 410         for (;;) {
 411                 hashtable = create_hashtable(corpus, all_trigrams, ht_size, num_overflow_slots);
 412                 if (hashtable == nullptr) {
 413                         dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
 414                         ht_size = next_prime(ht_size * 1.05);
 415                 } else {
 416                         dprintf("Created hash table of size %u.\n\n", ht_size);
 417                         break;
 418                 }
 419         }
 420
 421         // Find the offsets for each posting list.
 422         size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
 423         uint64_t offset = ftell(outfp) + bytes_for_hashtable;
 424         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 425                 hashtable[i].offset = offset;  // Needs to be there even for empty slots.
 426                 if (hashtable[i].num_docids == 0) {
 427                         continue;
 428                 }
 429
 430                 const string &encoded = corpus.invindex[hashtable[i].trgm].encoded;
 431                 offset += encoded.size();
 432         }
 433
 434         // Write the hash table.
 435         hdr.hash_table_offset_bytes = ftell(outfp);
 436         hdr.hashtable_size = ht_size;
 437         fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
 438
 439         // Write the actual posting lists.
 440         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 441                 if (hashtable[i].num_docids == 0) {
 442                         continue;
 443                 }
 444                 const string &encoded = corpus.invindex[hashtable[i].trgm].encoded;
 445                 fwrite(encoded.data(), encoded.size(), 1, outfp);
 446         }
 447
 448         // Rewind, and write the updated header.
 449         hdr.version = 0;
 450         fseek(outfp, 0, SEEK_SET);
 451         fwrite(&hdr, sizeof(hdr), 1, outfp);
 452         fclose(outfp);
 453
 454         size_t total_bytes __attribute__((unused)) = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
 455
 456         dprintf("Block size:     %7d files\n", block_size);
 457         dprintf("Hash table:     %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
 458         dprintf("Posting lists:  %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
 459         dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
 460         dprintf("Filenames:      %'7.1f MB\n", bytes_for_filenames / 1048576.0);
 461         dprintf("Total:          %'7.1f MB\n", total_bytes / 1048576.0);
 462         dprintf("\n");
 463 }
 464
 465 int main(int argc, char **argv)
 466 {
 467         do_build(argv[1], argv[2], 32);
 468         exit(EXIT_SUCCESS);
 469 }