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/stat.h>
  16 #include <sys/types.h>
  17 #include <unistd.h>
  18 #include <unordered_map>
  19 #include <vector>
  20 #include <zstd.h>
  21
  22 #define P4NENC_BOUND(n) ((n + 127) / 128 + (n + 32) * sizeof(uint32_t))
  23 #define dprintf(...)
  24 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
  25
  26 using namespace std;
  27 using namespace std::chrono;
  28
  29 string zstd_compress(const string &src, string *tempbuf);
  30
  31 constexpr unsigned num_overflow_slots = 16;
  32
  33 static inline uint32_t read_unigram(const string_view s, size_t idx)
  34 {
  35         if (idx < s.size()) {
  36                 return (unsigned char)s[idx];
  37         } else {
  38                 return 0;
  39         }
  40 }
  41
  42 static inline uint32_t read_trigram(const string_view s, size_t start)
  43 {
  44         return read_unigram(s, start) |
  45                 (read_unigram(s, start + 1) << 8) |
  46                 (read_unigram(s, start + 2) << 16);
  47 }
  48
  49 enum {
  50         DBE_NORMAL = 0, /* A non-directory file */
  51         DBE_DIRECTORY = 1, /* A directory */
  52         DBE_END = 2 /* End of directory contents; contains no name */
  53 };
  54
  55 // From mlocate.
  56 struct db_header {
  57         uint8_t magic[8];
  58         uint32_t conf_size;
  59         uint8_t version;
  60         uint8_t check_visibility;
  61         uint8_t pad[2];
  62 };
  63
  64 // From mlocate.
  65 struct db_directory {
  66         uint64_t time_sec;
  67         uint32_t time_nsec;
  68         uint8_t pad[4];
  69 };
  70
  71 class PostingListBuilder {
  72 public:
  73         void add_docid(uint32_t docid);
  74         void finish();
  75
  76         string encoded;
  77         size_t num_docids = 0;
  78
  79 private:
  80         void write_header(uint32_t docid);
  81         void append_block();
  82
  83         vector<uint32_t> pending_docids;
  84
  85         uint32_t last_block_end;
  86 };
  87
  88 void PostingListBuilder::add_docid(uint32_t docid)
  89 {
  90         // Deduplicate against the last inserted value, if any.
  91         if (pending_docids.empty()) {
  92                 if (encoded.empty()) {
  93                         // Very first docid.
  94                         write_header(docid);
  95                         ++num_docids;
  96                         last_block_end = docid;
  97                         return;
  98                 } else if (docid == last_block_end) {
  99                         return;
 100                 }
 101         } else {
 102                 if (docid == pending_docids.back()) {
 103                         return;
 104                 }
 105         }
 106
 107         pending_docids.push_back(docid);
 108         if (pending_docids.size() == 128) {
 109                 append_block();
 110                 pending_docids.clear();
 111                 last_block_end = docid;
 112         }
 113         ++num_docids;
 114 }
 115
 116 void PostingListBuilder::finish()
 117 {
 118         if (pending_docids.empty()) {
 119                 return;
 120         }
 121
 122         assert(!encoded.empty());  // write_header() should already have run.
 123
 124         // No interleaving for partial blocks.
 125         unsigned char buf[P4NENC_BOUND(128)];
 126         unsigned char *end = p4d1enc32(pending_docids.data(), pending_docids.size(), buf, last_block_end);
 127         encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
 128 }
 129
 130 void PostingListBuilder::append_block()
 131 {
 132         unsigned char buf[P4NENC_BOUND(128)];
 133         assert(pending_docids.size() == 128);
 134         unsigned char *end = p4d1enc128v32(pending_docids.data(), 128, buf, last_block_end);
 135         encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
 136 }
 137
 138 void PostingListBuilder::write_header(uint32_t docid)
 139 {
 140         unsigned char buf[P4NENC_BOUND(1)];
 141         size_t bytes = p4nd1enc128v32(&docid, 1, buf);
 142         encoded.append(reinterpret_cast<char *>(buf), bytes);
 143 }
 144
 145 class Corpus {
 146 public:
 147         Corpus(FILE *outfp, size_t block_size)
 148                 : outfp(outfp), block_size(block_size) {}
 149         void add_file(string filename);
 150         void flush_block();
 151
 152         vector<uint64_t> filename_blocks;
 153         unordered_map<uint32_t, PostingListBuilder> invindex;
 154         size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
 155
 156 private:
 157         FILE *outfp;
 158         string current_block;
 159         string tempbuf;
 160         const size_t block_size;
 161 };
 162
 163 void Corpus::add_file(string filename)
 164 {
 165         ++num_files;
 166         if (!current_block.empty()) {
 167                 current_block.push_back('\0');
 168         }
 169         current_block += filename;
 170         if (++num_files_in_block == block_size) {
 171                 flush_block();
 172         }
 173 }
 174
 175 void Corpus::flush_block()
 176 {
 177         if (current_block.empty()) {
 178                 return;
 179         }
 180
 181         uint32_t docid = num_blocks;
 182
 183         // Create trigrams.
 184         const char *ptr = current_block.c_str();
 185         while (ptr < current_block.c_str() + current_block.size()) {
 186                 string_view s(ptr);
 187                 if (s.size() >= 3) {
 188                         for (size_t j = 0; j < s.size() - 2; ++j) {
 189                                 uint32_t trgm = read_trigram(s, j);
 190                                 invindex[trgm].add_docid(docid);
 191                         }
 192                 }
 193                 ptr += s.size() + 1;
 194         }
 195
 196         // Compress and add the filename block.
 197         filename_blocks.push_back(ftell(outfp));
 198         string compressed = zstd_compress(current_block, &tempbuf);
 199         if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
 200                 perror("fwrite()");
 201                 exit(1);
 202         }
 203
 204         current_block.clear();
 205         num_files_in_block = 0;
 206         ++num_blocks;
 207 }
 208
 209 string read_cstr(FILE *fp)
 210 {
 211         string ret;
 212         for ( ;; ) {
 213                 int ch = getc(fp);
 214                 if (ch == -1) {
 215                         perror("getc");
 216                         exit(1);
 217                 }
 218                 if (ch == 0) {
 219                         return ret;
 220                 }
 221                 ret.push_back(ch);
 222         }
 223 }
 224
 225 void handle_directory(FILE *fp, Corpus *corpus)
 226 {
 227         db_directory dummy;
 228         if (fread(&dummy, sizeof(dummy), 1, fp) != 1) {
 229                 if (feof(fp)) {
 230                         return;
 231                 } else {
 232                         perror("fread");
 233                 }
 234         }
 235
 236         string dir_path = read_cstr(fp);
 237         if (dir_path == "/") {
 238                 dir_path = "";
 239         }
 240
 241         for (;;) {
 242                 int type = getc(fp);
 243                 if (type == DBE_NORMAL) {
 244                         string filename = read_cstr(fp);
 245                         corpus->add_file(dir_path + "/" + filename);
 246                 } else if (type == DBE_DIRECTORY) {
 247                         string dirname = read_cstr(fp);
 248                         corpus->add_file(dir_path + "/" + dirname);
 249                 } else {
 250                         return;  // Probably end.
 251                 }
 252         }
 253 }
 254
 255 void read_mlocate(const char *filename, Corpus *corpus)
 256 {
 257         FILE *fp = fopen(filename, "rb");
 258         if (fp == nullptr) {
 259                 perror(filename);
 260                 exit(1);
 261         }
 262
 263         db_header hdr;
 264         if (fread(&hdr, sizeof(hdr), 1, fp) != 1) {
 265                 perror("short read");
 266                 exit(1);
 267         }
 268
 269         // TODO: Care about the base path.
 270         string path = read_cstr(fp);
 271         while (!feof(fp)) {
 272                 handle_directory(fp, corpus);
 273         }
 274         fclose(fp);
 275 }
 276
 277 string zstd_compress(const string &src, string *tempbuf)
 278 {
 279         size_t max_size = ZSTD_compressBound(src.size());
 280         if (tempbuf->size() < max_size) {
 281                 tempbuf->resize(max_size);
 282         }
 283         size_t size = ZSTD_compress(&(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
 284         return string(tempbuf->data(), size);
 285 }
 286
 287 bool is_prime(uint32_t x)
 288 {
 289         if ((x % 2) == 0 || (x % 3) == 0) {
 290                 return false;
 291         }
 292         uint32_t limit = ceil(sqrt(x));
 293         for (uint32_t factor = 5; factor <= limit; ++factor) {
 294                 if ((x % factor) == 0) {
 295                         return false;
 296                 }
 297         }
 298         return true;
 299 }
 300
 301 uint32_t next_prime(uint32_t x)
 302 {
 303         if ((x % 2) == 0) {
 304                 ++x;
 305         }
 306         while (!is_prime(x)) {
 307                 x += 2;
 308         }
 309         return x;
 310 }
 311
 312 unique_ptr<Trigram[]> create_hashtable(const Corpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
 313 {
 314         unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]);  // 1 for the sentinel element at the end.
 315         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 316                 ht[i].trgm = uint32_t(-1);
 317                 ht[i].num_docids = 0;
 318                 ht[i].offset = 0;
 319         }
 320         for (uint32_t trgm : all_trigrams) {
 321                 // We don't know offset yet, so set it to zero.
 322                 Trigram to_insert{ trgm, uint32_t(corpus.invindex.find(trgm)->second.num_docids), 0 };
 323
 324                 uint32_t bucket = hash_trigram(trgm, ht_size);
 325                 unsigned distance = 0;
 326                 while (ht[bucket].num_docids != 0) {
 327                         // Robin Hood hashing; reduces the longest distance by a lot.
 328                         unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
 329                         if (distance > other_distance) {
 330                                 swap(to_insert, ht[bucket]);
 331                                 distance = other_distance;
 332                         }
 333
 334                         ++bucket, ++distance;
 335                         if (distance > num_overflow_slots) {
 336                                 return nullptr;
 337                         }
 338                 }
 339                 ht[bucket] = to_insert;
 340         }
 341         return ht;
 342 }
 343
 344 void do_build(const char *infile, const char *outfile, int block_size)
 345 {
 346         steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
 347
 348         umask(0027);
 349         FILE *outfp = fopen(outfile, "wb");
 350
 351         // Write the header.
 352         Header hdr;
 353         memcpy(hdr.magic, "\0plocate", 8);
 354         hdr.version = -1;  // Mark as broken.
 355         hdr.hashtable_size = 0;  // Not known yet.
 356         hdr.extra_ht_slots = num_overflow_slots;
 357         hdr.hash_table_offset_bytes = -1;  // We don't know these offsets yet.
 358         hdr.filename_index_offset_bytes = -1;
 359         fwrite(&hdr, sizeof(hdr), 1, outfp);
 360
 361         Corpus corpus(outfp, block_size);
 362
 363         read_mlocate(infile, &corpus);
 364         if (false) {  // To read a plain text file.
 365                 FILE *fp = fopen(infile, "r");
 366                 while (!feof(fp)) {
 367                         char buf[1024];
 368                         if (fgets(buf, 1024, fp) == nullptr || feof(fp)) {
 369                                 break;
 370                         }
 371                         string s(buf);
 372                         if (s.back() == '\n')
 373                                 s.pop_back();
 374                         corpus.add_file(move(s));
 375                 }
 376                 fclose(fp);
 377         }
 378         corpus.flush_block();
 379         dprintf("Read %zu files from %s\n", corpus.num_files, infile);
 380         hdr.num_docids = corpus.filename_blocks.size();
 381
 382         // Stick an empty block at the end as sentinel.
 383         corpus.filename_blocks.push_back(ftell(outfp));
 384         const size_t bytes_for_filenames = corpus.filename_blocks.back() - corpus.filename_blocks.front();
 385
 386         // Write the offsets to the filenames.
 387         hdr.filename_index_offset_bytes = ftell(outfp);
 388         const size_t bytes_for_filename_index = corpus.filename_blocks.size() * sizeof(uint64_t);
 389         fwrite(corpus.filename_blocks.data(), corpus.filename_blocks.size(), sizeof(uint64_t), outfp);
 390         corpus.filename_blocks.clear();
 391         corpus.filename_blocks.shrink_to_fit();
 392
 393         // Finish up encoding the posting lists.
 394         size_t trigrams = 0, longest_posting_list = 0;
 395         size_t bytes_for_posting_lists = 0;
 396         for (auto &[trigram, pl_builder] : corpus.invindex) {
 397                 pl_builder.finish();
 398                 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
 399                 trigrams += pl_builder.num_docids;
 400                 bytes_for_posting_lists += pl_builder.encoded.size();
 401         }
 402         dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
 403                 corpus.num_files, corpus.invindex.size(), trigrams, double(trigrams) / corpus.invindex.size(), longest_posting_list);
 404         dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));
 405
 406         dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 407
 408         // Sort the trigrams, mostly to get a consistent result every time
 409         // (the hash table will put things in random order anyway).
 410         vector<uint32_t> all_trigrams;
 411         for (auto &[trigram, pl_builder] : corpus.invindex) {
 412                 all_trigrams.push_back(trigram);
 413         }
 414         sort(all_trigrams.begin(), all_trigrams.end());
 415
 416         // Create the hash table.
 417         unique_ptr<Trigram[]> hashtable;
 418         uint32_t ht_size = next_prime(all_trigrams.size());
 419         for (;;) {
 420                 hashtable = create_hashtable(corpus, all_trigrams, ht_size, num_overflow_slots);
 421                 if (hashtable == nullptr) {
 422                         dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
 423                         ht_size = next_prime(ht_size * 1.05);
 424                 } else {
 425                         dprintf("Created hash table of size %u.\n\n", ht_size);
 426                         break;
 427                 }
 428         }
 429
 430         // Find the offsets for each posting list.
 431         size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
 432         uint64_t offset = ftell(outfp) + bytes_for_hashtable;
 433         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 434                 hashtable[i].offset = offset;  // Needs to be there even for empty slots.
 435                 if (hashtable[i].num_docids == 0) {
 436                         continue;
 437                 }
 438
 439                 const string &encoded = corpus.invindex[hashtable[i].trgm].encoded;
 440                 offset += encoded.size();
 441         }
 442
 443         // Write the hash table.
 444         hdr.hash_table_offset_bytes = ftell(outfp);
 445         hdr.hashtable_size = ht_size;
 446         fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
 447
 448         // Write the actual posting lists.
 449         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 450                 if (hashtable[i].num_docids == 0) {
 451                         continue;
 452                 }
 453                 const string &encoded = corpus.invindex[hashtable[i].trgm].encoded;
 454                 fwrite(encoded.data(), encoded.size(), 1, outfp);
 455         }
 456
 457         // Rewind, and write the updated header.
 458         hdr.version = 0;
 459         fseek(outfp, 0, SEEK_SET);
 460         fwrite(&hdr, sizeof(hdr), 1, outfp);
 461         fclose(outfp);
 462
 463         size_t total_bytes __attribute__((unused)) = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
 464
 465         dprintf("Block size:     %7d files\n", block_size);
 466         dprintf("Hash table:     %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
 467         dprintf("Posting lists:  %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
 468         dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
 469         dprintf("Filenames:      %'7.1f MB\n", bytes_for_filenames / 1048576.0);
 470         dprintf("Total:          %'7.1f MB\n", total_bytes / 1048576.0);
 471         dprintf("\n");
 472 }
 473
 474 int main(int argc, char **argv)
 475 {
 476         do_build(argv[1], argv[2], 32);
 477         exit(EXIT_SUCCESS);
 478 }