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