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

   1 #include "db.h"
   2 #include "turbopfor-encode.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 #define NUM_TRIGRAMS 16777216
  27
  28 using namespace std;
  29 using namespace std::chrono;
  30
  31 string zstd_compress(const string &src, string *tempbuf);
  32
  33 constexpr unsigned num_overflow_slots = 16;
  34
  35 static inline uint32_t read_unigram(const string_view s, size_t idx)
  36 {
  37         if (idx < s.size()) {
  38                 return (unsigned char)s[idx];
  39         } else {
  40                 return 0;
  41         }
  42 }
  43
  44 static inline uint32_t read_trigram(const string_view s, size_t start)
  45 {
  46         return read_unigram(s, start) |
  47                 (read_unigram(s, start + 1) << 8) |
  48                 (read_unigram(s, start + 2) << 16);
  49 }
  50
  51 enum {
  52         DBE_NORMAL = 0, /* A non-directory file */
  53         DBE_DIRECTORY = 1, /* A directory */
  54         DBE_END = 2 /* End of directory contents; contains no name */
  55 };
  56
  57 // From mlocate.
  58 struct db_header {
  59         uint8_t magic[8];
  60         uint32_t conf_size;
  61         uint8_t version;
  62         uint8_t check_visibility;
  63         uint8_t pad[2];
  64 };
  65
  66 // From mlocate.
  67 struct db_directory {
  68         uint64_t time_sec;
  69         uint32_t time_nsec;
  70         uint8_t pad[4];
  71 };
  72
  73 class PostingListBuilder {
  74 public:
  75         inline void add_docid(uint32_t docid);
  76         void finish();
  77
  78         string encoded;
  79         size_t num_docids = 0;
  80
  81 private:
  82         void write_header(uint32_t docid);
  83         void append_block();
  84
  85         vector<uint32_t> pending_deltas;
  86
  87         uint32_t last_block_end, last_docid = -1;
  88 };
  89
  90 void PostingListBuilder::add_docid(uint32_t docid)
  91 {
  92         // Deduplicate against the last inserted value, if any.
  93         if (docid == last_docid) {
  94                 return;
  95         }
  96
  97         if (num_docids == 0) {
  98                 // Very first docid.
  99                 write_header(docid);
 100                 ++num_docids;
 101                 last_block_end = last_docid = docid;
 102                 return;
 103         }
 104
 105         pending_deltas.push_back(docid - last_docid - 1);
 106         last_docid = docid;
 107         if (pending_deltas.size() == 128) {
 108                 append_block();
 109                 pending_deltas.clear();
 110                 last_block_end = docid;
 111         }
 112         ++num_docids;
 113 }
 114
 115 void PostingListBuilder::finish()
 116 {
 117         if (pending_deltas.empty()) {
 118                 return;
 119         }
 120
 121         assert(!encoded.empty());  // write_header() should already have run.
 122
 123         // No interleaving for partial blocks.
 124         unsigned char buf[P4NENC_BOUND(128)];
 125         unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), pending_deltas.size(), /*interleaved=*/false, buf);
 126         encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
 127 }
 128
 129 void PostingListBuilder::append_block()
 130 {
 131         unsigned char buf[P4NENC_BOUND(128)];
 132         assert(pending_deltas.size() == 128);
 133         unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), 128, /*interleaved=*/true, buf);
 134         encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
 135 }
 136
 137 void PostingListBuilder::write_header(uint32_t docid)
 138 {
 139         unsigned char buf[P4NENC_BOUND(1)];
 140         unsigned char *end = write_baseval(docid, buf);
 141         encoded.append(reinterpret_cast<char *>(buf), end - buf);
 142 }
 143
 144 class Corpus {
 145 public:
 146         Corpus(FILE *outfp, size_t block_size)
 147                 : invindex(new PostingListBuilder *[NUM_TRIGRAMS]), outfp(outfp), block_size(block_size) {}
 148         void add_file(string filename);
 149         void flush_block();
 150
 151         vector<uint64_t> filename_blocks;
 152         size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
 153         bool seen_trigram(uint32_t trgm)
 154         {
 155                 return invindex[trgm] != nullptr;
 156         }
 157         PostingListBuilder &get_pl_builder(uint32_t trgm)
 158         {
 159                 if (invindex[trgm] == nullptr) {
 160                         invindex[trgm] = new PostingListBuilder;
 161                 }
 162                 return *invindex[trgm];
 163         }
 164
 165 private:
 166         unique_ptr<PostingListBuilder *[]> invindex;
 167         FILE *outfp;
 168         string current_block;
 169         string tempbuf;
 170         const size_t block_size;
 171 };
 172
 173 void Corpus::add_file(string filename)
 174 {
 175         ++num_files;
 176         if (!current_block.empty()) {
 177                 current_block.push_back('\0');
 178         }
 179         current_block += filename;
 180         if (++num_files_in_block == block_size) {
 181                 flush_block();
 182         }
 183 }
 184
 185 void Corpus::flush_block()
 186 {
 187         if (current_block.empty()) {
 188                 return;
 189         }
 190
 191         uint32_t docid = num_blocks;
 192
 193         // Create trigrams.
 194         const char *ptr = current_block.c_str();
 195         while (ptr < current_block.c_str() + current_block.size()) {
 196                 string_view s(ptr);
 197                 if (s.size() >= 3) {
 198                         for (size_t j = 0; j < s.size() - 2; ++j) {
 199                                 uint32_t trgm = read_trigram(s, j);
 200                                 get_pl_builder(trgm).add_docid(docid);
 201                         }
 202                 }
 203                 ptr += s.size() + 1;
 204         }
 205
 206         // Compress and add the filename block.
 207         filename_blocks.push_back(ftell(outfp));
 208         string compressed = zstd_compress(current_block, &tempbuf);
 209         if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
 210                 perror("fwrite()");
 211                 exit(1);
 212         }
 213
 214         current_block.clear();
 215         num_files_in_block = 0;
 216         ++num_blocks;
 217 }
 218
 219 string read_cstr(FILE *fp)
 220 {
 221         string ret;
 222         for (;;) {
 223                 int ch = getc(fp);
 224                 if (ch == -1) {
 225                         perror("getc");
 226                         exit(1);
 227                 }
 228                 if (ch == 0) {
 229                         return ret;
 230                 }
 231                 ret.push_back(ch);
 232         }
 233 }
 234
 235 void handle_directory(FILE *fp, Corpus *corpus)
 236 {
 237         db_directory dummy;
 238         if (fread(&dummy, sizeof(dummy), 1, fp) != 1) {
 239                 if (feof(fp)) {
 240                         return;
 241                 } else {
 242                         perror("fread");
 243                 }
 244         }
 245
 246         string dir_path = read_cstr(fp);
 247         if (dir_path == "/") {
 248                 dir_path = "";
 249         }
 250
 251         for (;;) {
 252                 int type = getc(fp);
 253                 if (type == DBE_NORMAL) {
 254                         string filename = read_cstr(fp);
 255                         corpus->add_file(dir_path + "/" + filename);
 256                 } else if (type == DBE_DIRECTORY) {
 257                         string dirname = read_cstr(fp);
 258                         corpus->add_file(dir_path + "/" + dirname);
 259                 } else {
 260                         return;  // Probably end.
 261                 }
 262         }
 263 }
 264
 265 void read_mlocate(const char *filename, Corpus *corpus)
 266 {
 267         FILE *fp = fopen(filename, "rb");
 268         if (fp == nullptr) {
 269                 perror(filename);
 270                 exit(1);
 271         }
 272
 273         db_header hdr;
 274         if (fread(&hdr, sizeof(hdr), 1, fp) != 1) {
 275                 perror("short read");
 276                 exit(1);
 277         }
 278
 279         // TODO: Care about the base path.
 280         string path = read_cstr(fp);
 281         while (!feof(fp)) {
 282                 handle_directory(fp, corpus);
 283         }
 284         fclose(fp);
 285 }
 286
 287 string zstd_compress(const string &src, string *tempbuf)
 288 {
 289         size_t max_size = ZSTD_compressBound(src.size());
 290         if (tempbuf->size() < max_size) {
 291                 tempbuf->resize(max_size);
 292         }
 293         size_t size = ZSTD_compress(&(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
 294         return string(tempbuf->data(), size);
 295 }
 296
 297 bool is_prime(uint32_t x)
 298 {
 299         if ((x % 2) == 0 || (x % 3) == 0) {
 300                 return false;
 301         }
 302         uint32_t limit = ceil(sqrt(x));
 303         for (uint32_t factor = 5; factor <= limit; ++factor) {
 304                 if ((x % factor) == 0) {
 305                         return false;
 306                 }
 307         }
 308         return true;
 309 }
 310
 311 uint32_t next_prime(uint32_t x)
 312 {
 313         if ((x % 2) == 0) {
 314                 ++x;
 315         }
 316         while (!is_prime(x)) {
 317                 x += 2;
 318         }
 319         return x;
 320 }
 321
 322 unique_ptr<Trigram[]> create_hashtable(Corpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
 323 {
 324         unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]);  // 1 for the sentinel element at the end.
 325         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 326                 ht[i].trgm = uint32_t(-1);
 327                 ht[i].num_docids = 0;
 328                 ht[i].offset = 0;
 329         }
 330         for (uint32_t trgm : all_trigrams) {
 331                 // We don't know offset yet, so set it to zero.
 332                 Trigram to_insert{ trgm, uint32_t(corpus.get_pl_builder(trgm).num_docids), 0 };
 333
 334                 uint32_t bucket = hash_trigram(trgm, ht_size);
 335                 unsigned distance = 0;
 336                 while (ht[bucket].num_docids != 0) {
 337                         // Robin Hood hashing; reduces the longest distance by a lot.
 338                         unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
 339                         if (distance > other_distance) {
 340                                 swap(to_insert, ht[bucket]);
 341                                 distance = other_distance;
 342                         }
 343
 344                         ++bucket, ++distance;
 345                         if (distance > num_overflow_slots) {
 346                                 return nullptr;
 347                         }
 348                 }
 349                 ht[bucket] = to_insert;
 350         }
 351         return ht;
 352 }
 353
 354 void do_build(const char *infile, const char *outfile, int block_size)
 355 {
 356         steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
 357
 358         umask(0027);
 359         FILE *outfp = fopen(outfile, "wb");
 360
 361         // Write the header.
 362         Header hdr;
 363         memcpy(hdr.magic, "\0plocate", 8);
 364         hdr.version = -1;  // Mark as broken.
 365         hdr.hashtable_size = 0;  // Not known yet.
 366         hdr.extra_ht_slots = num_overflow_slots;
 367         hdr.hash_table_offset_bytes = -1;  // We don't know these offsets yet.
 368         hdr.filename_index_offset_bytes = -1;
 369         fwrite(&hdr, sizeof(hdr), 1, outfp);
 370
 371         Corpus corpus(outfp, block_size);
 372
 373         read_mlocate(infile, &corpus);
 374         if (false) {  // To read a plain text file.
 375                 FILE *fp = fopen(infile, "r");
 376                 while (!feof(fp)) {
 377                         char buf[1024];
 378                         if (fgets(buf, 1024, fp) == nullptr || feof(fp)) {
 379                                 break;
 380                         }
 381                         string s(buf);
 382                         if (s.back() == '\n')
 383                                 s.pop_back();
 384                         corpus.add_file(move(s));
 385                 }
 386                 fclose(fp);
 387         }
 388         corpus.flush_block();
 389         dprintf("Read %zu files from %s\n", corpus.num_files, infile);
 390         hdr.num_docids = corpus.filename_blocks.size();
 391
 392         // Stick an empty block at the end as sentinel.
 393         corpus.filename_blocks.push_back(ftell(outfp));
 394         const size_t bytes_for_filenames = corpus.filename_blocks.back() - corpus.filename_blocks.front();
 395
 396         // Write the offsets to the filenames.
 397         hdr.filename_index_offset_bytes = ftell(outfp);
 398         const size_t bytes_for_filename_index = corpus.filename_blocks.size() * sizeof(uint64_t);
 399         fwrite(corpus.filename_blocks.data(), corpus.filename_blocks.size(), sizeof(uint64_t), outfp);
 400         corpus.filename_blocks.clear();
 401         corpus.filename_blocks.shrink_to_fit();
 402
 403         // Finish up encoding the posting lists.
 404         size_t trigrams = 0, longest_posting_list = 0;
 405         size_t bytes_for_posting_lists = 0;
 406         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 407                 if (!corpus.seen_trigram(trgm))
 408                         continue;
 409                 PostingListBuilder &pl_builder = corpus.get_pl_builder(trgm);
 410                 pl_builder.finish();
 411                 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
 412                 trigrams += pl_builder.num_docids;
 413                 bytes_for_posting_lists += pl_builder.encoded.size();
 414         }
 415         dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
 416                 corpus.num_files, corpus.invindex.size(), trigrams, double(trigrams) / corpus.invindex.size(), longest_posting_list);
 417         dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));
 418
 419         dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 420
 421         // Find the used trigrams.
 422         vector<uint32_t> all_trigrams;
 423         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 424                 if (corpus.seen_trigram(trgm)) {
 425                         all_trigrams.push_back(trgm);
 426                 }
 427         }
 428
 429         // Create the hash table.
 430         unique_ptr<Trigram[]> hashtable;
 431         uint32_t ht_size = next_prime(all_trigrams.size());
 432         for (;;) {
 433                 hashtable = create_hashtable(corpus, all_trigrams, ht_size, num_overflow_slots);
 434                 if (hashtable == nullptr) {
 435                         dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
 436                         ht_size = next_prime(ht_size * 1.05);
 437                 } else {
 438                         dprintf("Created hash table of size %u.\n\n", ht_size);
 439                         break;
 440                 }
 441         }
 442
 443         // Find the offsets for each posting list.
 444         size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
 445         uint64_t offset = ftell(outfp) + bytes_for_hashtable;
 446         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 447                 hashtable[i].offset = offset;  // Needs to be there even for empty slots.
 448                 if (hashtable[i].num_docids == 0) {
 449                         continue;
 450                 }
 451
 452                 const string &encoded = corpus.get_pl_builder(hashtable[i].trgm).encoded;
 453                 offset += encoded.size();
 454         }
 455
 456         // Write the hash table.
 457         hdr.hash_table_offset_bytes = ftell(outfp);
 458         hdr.hashtable_size = ht_size;
 459         fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
 460
 461         // Write the actual posting lists.
 462         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 463                 if (hashtable[i].num_docids == 0) {
 464                         continue;
 465                 }
 466                 const string &encoded = corpus.get_pl_builder(hashtable[i].trgm).encoded;
 467                 fwrite(encoded.data(), encoded.size(), 1, outfp);
 468         }
 469
 470         // Rewind, and write the updated header.
 471         hdr.version = 0;
 472         fseek(outfp, 0, SEEK_SET);
 473         fwrite(&hdr, sizeof(hdr), 1, outfp);
 474         fclose(outfp);
 475
 476         size_t total_bytes __attribute__((unused)) = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
 477
 478         dprintf("Block size:     %7d files\n", block_size);
 479         dprintf("Hash table:     %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
 480         dprintf("Posting lists:  %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
 481         dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
 482         dprintf("Filenames:      %'7.1f MB\n", bytes_for_filenames / 1048576.0);
 483         dprintf("Total:          %'7.1f MB\n", total_bytes / 1048576.0);
 484         dprintf("\n");
 485 }
 486
 487 int main(int argc, char **argv)
 488 {
 489         do_build(argv[1], argv[2], 32);
 490         exit(EXIT_SUCCESS);
 491 }