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                 fill(invindex.get(), invindex.get() + NUM_TRIGRAMS, nullptr);
 149         }
 150         ~Corpus() {
 151                 for (unsigned i = 0; i < NUM_TRIGRAMS; ++i) {
 152                         delete invindex[i];
 153                 }
 154         }
 155
 156         void add_file(string filename);
 157         void flush_block();
 158
 159         vector<uint64_t> filename_blocks;
 160         size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
 161         bool seen_trigram(uint32_t trgm)
 162         {
 163                 return invindex[trgm] != nullptr;
 164         }
 165         PostingListBuilder &get_pl_builder(uint32_t trgm)
 166         {
 167                 if (invindex[trgm] == nullptr) {
 168                         invindex[trgm] = new PostingListBuilder;
 169                 }
 170                 return *invindex[trgm];
 171         }
 172
 173 private:
 174         unique_ptr<PostingListBuilder *[]> invindex;
 175         FILE *outfp;
 176         string current_block;
 177         string tempbuf;
 178         const size_t block_size;
 179 };
 180
 181 void Corpus::add_file(string filename)
 182 {
 183         ++num_files;
 184         if (!current_block.empty()) {
 185                 current_block.push_back('\0');
 186         }
 187         current_block += filename;
 188         if (++num_files_in_block == block_size) {
 189                 flush_block();
 190         }
 191 }
 192
 193 void Corpus::flush_block()
 194 {
 195         if (current_block.empty()) {
 196                 return;
 197         }
 198
 199         uint32_t docid = num_blocks;
 200
 201         // Create trigrams.
 202         const char *ptr = current_block.c_str();
 203         while (ptr < current_block.c_str() + current_block.size()) {
 204                 string_view s(ptr);
 205                 if (s.size() >= 3) {
 206                         for (size_t j = 0; j < s.size() - 2; ++j) {
 207                                 uint32_t trgm = read_trigram(s, j);
 208                                 get_pl_builder(trgm).add_docid(docid);
 209                         }
 210                 }
 211                 ptr += s.size() + 1;
 212         }
 213
 214         // Compress and add the filename block.
 215         filename_blocks.push_back(ftell(outfp));
 216         string compressed = zstd_compress(current_block, &tempbuf);
 217         if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
 218                 perror("fwrite()");
 219                 exit(1);
 220         }
 221
 222         current_block.clear();
 223         num_files_in_block = 0;
 224         ++num_blocks;
 225 }
 226
 227 string read_cstr(FILE *fp)
 228 {
 229         string ret;
 230         for (;;) {
 231                 int ch = getc(fp);
 232                 if (ch == -1) {
 233                         perror("getc");
 234                         exit(1);
 235                 }
 236                 if (ch == 0) {
 237                         return ret;
 238                 }
 239                 ret.push_back(ch);
 240         }
 241 }
 242
 243 void handle_directory(FILE *fp, Corpus *corpus)
 244 {
 245         db_directory dummy;
 246         if (fread(&dummy, sizeof(dummy), 1, fp) != 1) {
 247                 if (feof(fp)) {
 248                         return;
 249                 } else {
 250                         perror("fread");
 251                 }
 252         }
 253
 254         string dir_path = read_cstr(fp);
 255         if (dir_path == "/") {
 256                 dir_path = "";
 257         }
 258
 259         for (;;) {
 260                 int type = getc(fp);
 261                 if (type == DBE_NORMAL) {
 262                         string filename = read_cstr(fp);
 263                         corpus->add_file(dir_path + "/" + filename);
 264                 } else if (type == DBE_DIRECTORY) {
 265                         string dirname = read_cstr(fp);
 266                         corpus->add_file(dir_path + "/" + dirname);
 267                 } else {
 268                         return;  // Probably end.
 269                 }
 270         }
 271 }
 272
 273 void read_mlocate(const char *filename, Corpus *corpus)
 274 {
 275         FILE *fp = fopen(filename, "rb");
 276         if (fp == nullptr) {
 277                 perror(filename);
 278                 exit(1);
 279         }
 280
 281         db_header hdr;
 282         if (fread(&hdr, sizeof(hdr), 1, fp) != 1) {
 283                 perror("short read");
 284                 exit(1);
 285         }
 286
 287         // TODO: Care about the base path.
 288         string path = read_cstr(fp);
 289         while (!feof(fp)) {
 290                 handle_directory(fp, corpus);
 291         }
 292         fclose(fp);
 293 }
 294
 295 string zstd_compress(const string &src, string *tempbuf)
 296 {
 297         size_t max_size = ZSTD_compressBound(src.size());
 298         if (tempbuf->size() < max_size) {
 299                 tempbuf->resize(max_size);
 300         }
 301         size_t size = ZSTD_compress(&(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
 302         return string(tempbuf->data(), size);
 303 }
 304
 305 bool is_prime(uint32_t x)
 306 {
 307         if ((x % 2) == 0 || (x % 3) == 0) {
 308                 return false;
 309         }
 310         uint32_t limit = ceil(sqrt(x));
 311         for (uint32_t factor = 5; factor <= limit; ++factor) {
 312                 if ((x % factor) == 0) {
 313                         return false;
 314                 }
 315         }
 316         return true;
 317 }
 318
 319 uint32_t next_prime(uint32_t x)
 320 {
 321         if ((x % 2) == 0) {
 322                 ++x;
 323         }
 324         while (!is_prime(x)) {
 325                 x += 2;
 326         }
 327         return x;
 328 }
 329
 330 unique_ptr<Trigram[]> create_hashtable(Corpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
 331 {
 332         unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]);  // 1 for the sentinel element at the end.
 333         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 334                 ht[i].trgm = uint32_t(-1);
 335                 ht[i].num_docids = 0;
 336                 ht[i].offset = 0;
 337         }
 338         for (uint32_t trgm : all_trigrams) {
 339                 // We don't know offset yet, so set it to zero.
 340                 Trigram to_insert{ trgm, uint32_t(corpus.get_pl_builder(trgm).num_docids), 0 };
 341
 342                 uint32_t bucket = hash_trigram(trgm, ht_size);
 343                 unsigned distance = 0;
 344                 while (ht[bucket].num_docids != 0) {
 345                         // Robin Hood hashing; reduces the longest distance by a lot.
 346                         unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
 347                         if (distance > other_distance) {
 348                                 swap(to_insert, ht[bucket]);
 349                                 distance = other_distance;
 350                         }
 351
 352                         ++bucket, ++distance;
 353                         if (distance > num_overflow_slots) {
 354                                 return nullptr;
 355                         }
 356                 }
 357                 ht[bucket] = to_insert;
 358         }
 359         return ht;
 360 }
 361
 362 void do_build(const char *infile, const char *outfile, int block_size)
 363 {
 364         steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
 365
 366         umask(0027);
 367         FILE *outfp = fopen(outfile, "wb");
 368
 369         // Write the header.
 370         Header hdr;
 371         memcpy(hdr.magic, "\0plocate", 8);
 372         hdr.version = -1;  // Mark as broken.
 373         hdr.hashtable_size = 0;  // Not known yet.
 374         hdr.extra_ht_slots = num_overflow_slots;
 375         hdr.num_docids = 0;
 376         hdr.hash_table_offset_bytes = -1;  // We don't know these offsets yet.
 377         hdr.filename_index_offset_bytes = -1;
 378         fwrite(&hdr, sizeof(hdr), 1, outfp);
 379
 380         Corpus corpus(outfp, block_size);
 381
 382         read_mlocate(infile, &corpus);
 383         if (false) {  // To read a plain text file.
 384                 FILE *fp = fopen(infile, "r");
 385                 while (!feof(fp)) {
 386                         char buf[1024];
 387                         if (fgets(buf, 1024, fp) == nullptr || feof(fp)) {
 388                                 break;
 389                         }
 390                         string s(buf);
 391                         if (s.back() == '\n')
 392                                 s.pop_back();
 393                         corpus.add_file(move(s));
 394                 }
 395                 fclose(fp);
 396         }
 397         corpus.flush_block();
 398         dprintf("Read %zu files from %s\n", corpus.num_files, infile);
 399         hdr.num_docids = corpus.filename_blocks.size();
 400
 401         // Stick an empty block at the end as sentinel.
 402         corpus.filename_blocks.push_back(ftell(outfp));
 403         const size_t bytes_for_filenames = corpus.filename_blocks.back() - corpus.filename_blocks.front();
 404
 405         // Write the offsets to the filenames.
 406         hdr.filename_index_offset_bytes = ftell(outfp);
 407         const size_t bytes_for_filename_index = corpus.filename_blocks.size() * sizeof(uint64_t);
 408         fwrite(corpus.filename_blocks.data(), corpus.filename_blocks.size(), sizeof(uint64_t), outfp);
 409         corpus.filename_blocks.clear();
 410         corpus.filename_blocks.shrink_to_fit();
 411
 412         // Finish up encoding the posting lists.
 413         size_t trigrams = 0, longest_posting_list = 0;
 414         size_t bytes_for_posting_lists = 0;
 415         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 416                 if (!corpus.seen_trigram(trgm))
 417                         continue;
 418                 PostingListBuilder &pl_builder = corpus.get_pl_builder(trgm);
 419                 pl_builder.finish();
 420                 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
 421                 trigrams += pl_builder.num_docids;
 422                 bytes_for_posting_lists += pl_builder.encoded.size();
 423         }
 424         dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
 425                 corpus.num_files, corpus.invindex.size(), trigrams, double(trigrams) / corpus.invindex.size(), longest_posting_list);
 426         dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));
 427
 428         dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 429
 430         // Find the used trigrams.
 431         vector<uint32_t> all_trigrams;
 432         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 433                 if (corpus.seen_trigram(trgm)) {
 434                         all_trigrams.push_back(trgm);
 435                 }
 436         }
 437
 438         // Create the hash table.
 439         unique_ptr<Trigram[]> hashtable;
 440         uint32_t ht_size = next_prime(all_trigrams.size());
 441         for (;;) {
 442                 hashtable = create_hashtable(corpus, all_trigrams, ht_size, num_overflow_slots);
 443                 if (hashtable == nullptr) {
 444                         dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
 445                         ht_size = next_prime(ht_size * 1.05);
 446                 } else {
 447                         dprintf("Created hash table of size %u.\n\n", ht_size);
 448                         break;
 449                 }
 450         }
 451
 452         // Find the offsets for each posting list.
 453         size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
 454         uint64_t offset = ftell(outfp) + bytes_for_hashtable;
 455         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 456                 hashtable[i].offset = offset;  // Needs to be there even for empty slots.
 457                 if (hashtable[i].num_docids == 0) {
 458                         continue;
 459                 }
 460
 461                 const string &encoded = corpus.get_pl_builder(hashtable[i].trgm).encoded;
 462                 offset += encoded.size();
 463         }
 464
 465         // Write the hash table.
 466         hdr.hash_table_offset_bytes = ftell(outfp);
 467         hdr.hashtable_size = ht_size;
 468         fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
 469
 470         // Write the actual posting lists.
 471         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 472                 if (hashtable[i].num_docids == 0) {
 473                         continue;
 474                 }
 475                 const string &encoded = corpus.get_pl_builder(hashtable[i].trgm).encoded;
 476                 fwrite(encoded.data(), encoded.size(), 1, outfp);
 477         }
 478
 479         // Rewind, and write the updated header.
 480         hdr.version = 0;
 481         fseek(outfp, 0, SEEK_SET);
 482         fwrite(&hdr, sizeof(hdr), 1, outfp);
 483         fclose(outfp);
 484
 485         size_t total_bytes __attribute__((unused)) = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
 486
 487         dprintf("Block size:     %7d files\n", block_size);
 488         dprintf("Hash table:     %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
 489         dprintf("Posting lists:  %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
 490         dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
 491         dprintf("Filenames:      %'7.1f MB\n", bytes_for_filenames / 1048576.0);
 492         dprintf("Total:          %'7.1f MB\n", total_bytes / 1048576.0);
 493         dprintf("\n");
 494 }
 495
 496 int main(int argc, char **argv)
 497 {
 498         do_build(argv[1], argv[2], 32);
 499         exit(EXIT_SUCCESS);
 500 }