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

   1 #include "vp4.h"
   2
   3 #include <algorithm>
   4 #include <arpa/inet.h>
   5 #include <assert.h>
   6 #include <chrono>
   7 #include <endian.h>
   8 #include <fcntl.h>
   9 #include <math.h>
  10 #include <memory>
  11 #include <stdio.h>
  12 #include <string.h>
  13 #include <string>
  14 #include <sys/mman.h>
  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 #include "db.h"
  23
  24 #define P4NENC_BOUND(n) ((n + 127) / 128 + (n + 32) * sizeof(uint32_t))
  25 #define dprintf(...)
  26 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
  27
  28 using namespace std;
  29 using namespace std::chrono;
  30
  31 string zstd_compress(const string &src, string *tempbuf);
  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(size_t block_size)
 148                 : block_size(block_size) {}
 149         void add_file(string filename);
 150         void flush_block();
 151
 152         vector<string> 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         string current_block;
 158         string tempbuf;
 159         const size_t block_size;
 160 };
 161
 162 void Corpus::add_file(string filename)
 163 {
 164         ++num_files;
 165         if (!current_block.empty()) {
 166                 current_block.push_back('\0');
 167         }
 168         current_block += filename;
 169         if (++num_files_in_block == block_size) {
 170                 flush_block();
 171         }
 172 }
 173
 174 void Corpus::flush_block()
 175 {
 176         if (current_block.empty()) {
 177                 return;
 178         }
 179
 180         uint32_t docid = num_blocks;
 181
 182         // Create trigrams.
 183         const char *ptr = current_block.c_str();
 184         while (ptr < current_block.c_str() + current_block.size()) {
 185                 string_view s(ptr);
 186                 if (s.size() >= 3) {
 187                         for (size_t j = 0; j < s.size() - 2; ++j) {
 188                                 uint32_t trgm = read_trigram(s, j);
 189                                 invindex[trgm].add_docid(docid);
 190                         }
 191                 }
 192                 ptr += s.size() + 1;
 193         }
 194
 195         // Compress and add the filename block.
 196         filename_blocks.push_back(zstd_compress(current_block, &tempbuf));
 197
 198         current_block.clear();
 199         num_files_in_block = 0;
 200         ++num_blocks;
 201 }
 202
 203 const char *handle_directory(const char *ptr, Corpus *corpus)
 204 {
 205         ptr += sizeof(db_directory);
 206
 207         string dir_path = ptr;
 208         ptr += dir_path.size() + 1;
 209         if (dir_path == "/") {
 210                 dir_path = "";
 211         }
 212
 213         for (;;) {
 214                 uint8_t type = *ptr++;
 215                 if (type == DBE_NORMAL) {
 216                         string filename = ptr;
 217                         corpus->add_file(dir_path + "/" + filename);
 218                         ptr += filename.size() + 1;
 219                 } else if (type == DBE_DIRECTORY) {
 220                         string dirname = ptr;
 221                         corpus->add_file(dir_path + "/" + dirname);
 222                         ptr += dirname.size() + 1;
 223                 } else {
 224                         return ptr;
 225                 }
 226         }
 227 }
 228
 229 void read_mlocate(const char *filename, Corpus *corpus)
 230 {
 231         int fd = open(filename, O_RDONLY);
 232         if (fd == -1) {
 233                 perror(filename);
 234                 exit(1);
 235         }
 236         off_t len = lseek(fd, 0, SEEK_END);
 237         if (len == -1) {
 238                 perror("lseek");
 239                 exit(1);
 240         }
 241         const char *data = (char *)mmap(nullptr, len, PROT_READ, MAP_SHARED, fd, /*offset=*/0);
 242         if (data == MAP_FAILED) {
 243                 perror("mmap");
 244                 exit(1);
 245         }
 246
 247         const db_header *hdr = (const db_header *)data;
 248
 249         // TODO: Care about the base path.
 250         string path = data + sizeof(db_header);
 251         uint64_t offset = sizeof(db_header) + path.size() + 1 + ntohl(hdr->conf_size);
 252
 253         const char *ptr = data + offset;
 254         while (ptr < data + len) {
 255                 ptr = handle_directory(ptr, corpus);
 256         }
 257
 258         munmap((void *)data, len);
 259         close(fd);
 260 }
 261
 262 string zstd_compress(const string &src, string *tempbuf)
 263 {
 264         size_t max_size = ZSTD_compressBound(src.size());
 265         if (tempbuf->size() < max_size) {
 266                 tempbuf->resize(max_size);
 267         }
 268         size_t size = ZSTD_compress(&(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
 269         return string(tempbuf->data(), size);
 270 }
 271
 272 bool is_prime(uint32_t x)
 273 {
 274         if ((x % 2) == 0 || (x % 3) == 0) {
 275                 return false;
 276         }
 277         uint32_t limit = ceil(sqrt(x));
 278         for (uint32_t factor = 5; factor <= limit; ++factor) {
 279                 if ((x % factor) == 0) {
 280                         return false;
 281                 }
 282         }
 283         return true;
 284 }
 285
 286 uint32_t next_prime(uint32_t x)
 287 {
 288         if ((x % 2) == 0) {
 289                 ++x;
 290         }
 291         while (!is_prime(x)) {
 292                 x += 2;
 293         }
 294         return x;
 295 }
 296
 297 unique_ptr<Trigram[]> create_hashtable(const Corpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
 298 {
 299         unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]);  // 1 for the sentinel element at the end.
 300         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 301                 ht[i].trgm = uint32_t(-1);
 302                 ht[i].num_docids = 0;
 303                 ht[i].offset = 0;
 304         }
 305         for (uint32_t trgm : all_trigrams) {
 306                 // We don't know offset yet, so set it to zero.
 307                 Trigram to_insert{trgm, uint32_t(corpus.invindex.find(trgm)->second.num_docids), 0};
 308
 309                 uint32_t bucket = hash_trigram(trgm, ht_size);
 310                 unsigned distance = 0;
 311                 while (ht[bucket].num_docids != 0) {
 312                         // Robin Hood hashing; reduces the longest distance by a lot.
 313                         unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
 314                         if (distance > other_distance) {
 315                                 swap(to_insert, ht[bucket]);
 316                                 distance = other_distance;
 317                         }
 318
 319                         ++bucket, ++distance;
 320                         if (distance > num_overflow_slots) {
 321                                 return nullptr;
 322                         }
 323                 }
 324                 ht[bucket] = to_insert;
 325         }
 326         return ht;
 327 }
 328
 329 void do_build(const char *infile, const char *outfile, int block_size)
 330 {
 331         steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
 332
 333         Corpus corpus(block_size);
 334
 335         read_mlocate(infile, &corpus);
 336         if (false) {  // To read a plain text file.
 337                 FILE *fp = fopen(infile, "r");
 338                 while (!feof(fp)) {
 339                         char buf[1024];
 340                         if (fgets(buf, 1024, fp) == nullptr || feof(fp)) {
 341                                 break;
 342                         }
 343                         string s(buf);
 344                         if (s.back() == '\n')
 345                                 s.pop_back();
 346                         corpus.add_file(move(s));
 347                 }
 348                 fclose(fp);
 349         }
 350         corpus.flush_block();
 351         dprintf("Read %zu files from %s\n", corpus.num_files, infile);
 352
 353         size_t trigrams = 0, longest_posting_list = 0;
 354         size_t bytes_used = 0;
 355         for (auto &[trigram, pl_builder] : corpus.invindex) {
 356                 pl_builder.finish();
 357                 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
 358                 trigrams += pl_builder.num_docids;
 359                 bytes_used += pl_builder.encoded.size();
 360         }
 361         dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
 362                 corpus.num_files, corpus.invindex.size(), trigrams, double(trigrams) / corpus.invindex.size(), longest_posting_list);
 363         dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_used, 8 * bytes_used / double(trigrams));
 364
 365         dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 366
 367         // Sort the trigrams, mostly to get a consistent result every time
 368         // (the hash table will put things in random order anyway).
 369         vector<uint32_t> all_trigrams;
 370         for (auto &[trigram, pl_builder] : corpus.invindex) {
 371                 all_trigrams.push_back(trigram);
 372         }
 373         sort(all_trigrams.begin(), all_trigrams.end());
 374
 375         unique_ptr<Trigram[]> hashtable;
 376         uint32_t ht_size = next_prime(all_trigrams.size());
 377         constexpr unsigned num_overflow_slots = 16;
 378         for ( ;; ) {
 379                 hashtable = create_hashtable(corpus, all_trigrams, ht_size, num_overflow_slots);
 380                 if (hashtable == nullptr) {
 381                         dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
 382                         ht_size = next_prime(ht_size * 1.05);
 383                 } else {
 384                         dprintf("Created hash table of size %u.\n\n", ht_size);
 385                         break;
 386                 }
 387         }
 388
 389         umask(0027);
 390         FILE *outfp = fopen(outfile, "wb");
 391
 392         // Find the offsets for each posting list.
 393         size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
 394         uint64_t offset = sizeof(Header) + bytes_for_hashtable;
 395         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 396                 hashtable[i].offset = offset;  // Needs to be there even for empty slots.
 397                 if (hashtable[i].num_docids == 0) {
 398                         continue;
 399                 }
 400
 401                 const PostingListBuilder &pl_builder = corpus.invindex[hashtable[i].trgm];
 402                 offset += pl_builder.encoded.size();
 403         }
 404
 405         // Write the header.
 406         Header hdr;
 407         memcpy(hdr.magic, "\0plocate", 8);
 408         hdr.version = 0;
 409         hdr.hashtable_size = ht_size;
 410         hdr.extra_ht_slots = num_overflow_slots;
 411         hdr.hash_table_offset_bytes = sizeof(hdr);  // This member is just there for flexibility.
 412         hdr.filename_index_offset_bytes = offset;
 413         fwrite(&hdr, sizeof(hdr), 1, outfp);
 414
 415         // Write the hash table.
 416         fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
 417
 418         // Write the actual posting lists.
 419         size_t bytes_for_posting_lists = 0;
 420         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 421                 if (hashtable[i].num_docids == 0) {
 422                         continue;
 423                 }
 424                 const string &encoded = corpus.invindex[hashtable[i].trgm].encoded;
 425                 fwrite(encoded.data(), encoded.size(), 1, outfp);
 426                 bytes_for_posting_lists += encoded.size();
 427         }
 428
 429         // Stick an empty block at the end as sentinel.
 430         corpus.filename_blocks.push_back("");
 431
 432         // Write the offsets to the filenames.
 433         size_t bytes_for_filename_index = 0, bytes_for_filenames = 0;
 434         offset = hdr.filename_index_offset_bytes + corpus.filename_blocks.size() * sizeof(offset);
 435         for (const string &filename : corpus.filename_blocks) {
 436                 fwrite(&offset, sizeof(offset), 1, outfp);
 437                 offset += filename.size();
 438                 bytes_for_filename_index += sizeof(offset);
 439                 bytes_for_filenames += filename.size();
 440         }
 441
 442         // Write the actual filenames.
 443         for (const string &filename : corpus.filename_blocks) {
 444                 fwrite(filename.data(), filename.size(), 1, outfp);
 445         }
 446
 447         fclose(outfp);
 448
 449         size_t total_bytes __attribute__((unused)) = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
 450
 451         dprintf("Block size:     %7d files\n", block_size);
 452         dprintf("Hash table:     %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
 453         dprintf("Posting lists:  %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
 454         dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
 455         dprintf("Filenames:      %'7.1f MB\n", bytes_for_filenames / 1048576.0);
 456         dprintf("Total:          %'7.1f MB\n", total_bytes / 1048576.0);
 457         dprintf("\n");
 458 }
 459
 460 int main(int argc, char **argv)
 461 {
 462         do_build(argv[1], argv[2], 32);
 463         exit(EXIT_SUCCESS);
 464 }