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 <stdio.h>
  10 #include <string.h>
  11 #include <string>
  12 #include <sys/mman.h>
  13 #include <sys/stat.h>
  14 #include <sys/types.h>
  15 #include <unistd.h>
  16 #include <unordered_map>
  17 #include <vector>
  18 #include <zstd.h>
  19
  20 #define P4NENC_BOUND(n) ((n + 127) / 128 + (n + 32) * sizeof(uint32_t))
  21 #define dprintf(...)
  22 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
  23
  24 using namespace std;
  25 using namespace std::chrono;
  26
  27 string zstd_compress(const string &src, string *tempbuf);
  28
  29 static inline uint32_t read_unigram(const string_view s, size_t idx)
  30 {
  31         if (idx < s.size()) {
  32                 return (unsigned char)s[idx];
  33         } else {
  34                 return 0;
  35         }
  36 }
  37
  38 static inline uint32_t read_trigram(const string_view s, size_t start)
  39 {
  40         return read_unigram(s, start) |
  41                 (read_unigram(s, start + 1) << 8) |
  42                 (read_unigram(s, start + 2) << 16);
  43 }
  44
  45 enum {
  46         DBE_NORMAL = 0, /* A non-directory file */
  47         DBE_DIRECTORY = 1, /* A directory */
  48         DBE_END = 2 /* End of directory contents; contains no name */
  49 };
  50
  51 // From mlocate.
  52 struct db_header {
  53         uint8_t magic[8];
  54         uint32_t conf_size;
  55         uint8_t version;
  56         uint8_t check_visibility;
  57         uint8_t pad[2];
  58 };
  59
  60 // From mlocate.
  61 struct db_directory {
  62         uint64_t time_sec;
  63         uint32_t time_nsec;
  64         uint8_t pad[4];
  65 };
  66
  67 class PostingListBuilder {
  68 public:
  69         void add_docid(uint32_t docid);
  70         void finish();
  71
  72         string encoded;
  73         size_t num_docids = 0;
  74
  75 private:
  76         void write_header(uint32_t docid);
  77         void append_block();
  78
  79         vector<uint32_t> pending_docids;
  80
  81         uint32_t last_block_end;
  82 };
  83
  84 void PostingListBuilder::add_docid(uint32_t docid)
  85 {
  86         // Deduplicate against the last inserted value, if any.
  87         if (pending_docids.empty()) {
  88                 if (encoded.empty()) {
  89                         // Very first docid.
  90                         write_header(docid);
  91                         ++num_docids;
  92                         last_block_end = docid;
  93                         return;
  94                 } else if (docid == last_block_end) {
  95                         return;
  96                 }
  97         } else {
  98                 if (docid == pending_docids.back()) {
  99                         return;
 100                 }
 101         }
 102
 103         pending_docids.push_back(docid);
 104         if (pending_docids.size() == 128) {
 105                 append_block();
 106                 pending_docids.clear();
 107                 last_block_end = docid;
 108         }
 109         ++num_docids;
 110 }
 111
 112 void PostingListBuilder::finish()
 113 {
 114         if (pending_docids.empty()) {
 115                 return;
 116         }
 117
 118         assert(!encoded.empty());  // write_header() should already have run.
 119
 120         // No interleaving for partial blocks.
 121         unsigned char buf[P4NENC_BOUND(128)];
 122         unsigned char *end = p4d1enc32(pending_docids.data(), pending_docids.size(), buf, last_block_end);
 123         encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
 124 }
 125
 126 void PostingListBuilder::append_block()
 127 {
 128         unsigned char buf[P4NENC_BOUND(128)];
 129         assert(pending_docids.size() == 128);
 130         unsigned char *end = p4d1enc128v32(pending_docids.data(), 128, buf, last_block_end);
 131         encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
 132 }
 133
 134 void PostingListBuilder::write_header(uint32_t docid)
 135 {
 136         unsigned char buf[P4NENC_BOUND(1)];
 137         size_t bytes = p4nd1enc128v32(&docid, 1, buf);
 138         encoded.append(reinterpret_cast<char *>(buf), bytes);
 139 }
 140
 141 class Corpus {
 142 public:
 143         Corpus(size_t block_size)
 144                 : block_size(block_size) {}
 145         void add_file(string filename);
 146         void flush_block();
 147
 148         vector<string> filename_blocks;
 149         unordered_map<uint32_t, PostingListBuilder> invindex;
 150         size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
 151
 152 private:
 153         string current_block;
 154         string tempbuf;
 155         const size_t block_size;
 156 };
 157
 158 void Corpus::add_file(string filename)
 159 {
 160         ++num_files;
 161         if (!current_block.empty()) {
 162                 current_block.push_back('\0');
 163         }
 164         current_block += filename;
 165         if (++num_files_in_block == block_size) {
 166                 flush_block();
 167         }
 168 }
 169
 170 void Corpus::flush_block()
 171 {
 172         if (current_block.empty()) {
 173                 return;
 174         }
 175
 176         uint32_t docid = num_blocks;
 177
 178         // Create trigrams.
 179         const char *ptr = current_block.c_str();
 180         while (ptr < current_block.c_str() + current_block.size()) {
 181                 string_view s(ptr);
 182                 if (s.size() >= 3) {
 183                         for (size_t j = 0; j < s.size() - 2; ++j) {
 184                                 uint32_t trgm = read_trigram(s, j);
 185                                 invindex[trgm].add_docid(docid);
 186                         }
 187                 }
 188                 ptr += s.size() + 1;
 189         }
 190
 191         // Compress and add the filename block.
 192         filename_blocks.push_back(zstd_compress(current_block, &tempbuf));
 193
 194         current_block.clear();
 195         num_files_in_block = 0;
 196         ++num_blocks;
 197 }
 198
 199 const char *handle_directory(const char *ptr, Corpus *corpus)
 200 {
 201         ptr += sizeof(db_directory);
 202
 203         string dir_path = ptr;
 204         ptr += dir_path.size() + 1;
 205         if (dir_path == "/") {
 206                 dir_path = "";
 207         }
 208
 209         for (;;) {
 210                 uint8_t type = *ptr++;
 211                 if (type == DBE_NORMAL) {
 212                         string filename = ptr;
 213                         corpus->add_file(dir_path + "/" + filename);
 214                         ptr += filename.size() + 1;
 215                 } else if (type == DBE_DIRECTORY) {
 216                         string dirname = ptr;
 217                         corpus->add_file(dir_path + "/" + dirname);
 218                         ptr += dirname.size() + 1;
 219                 } else {
 220                         return ptr;
 221                 }
 222         }
 223 }
 224
 225 void read_mlocate(const char *filename, Corpus *corpus)
 226 {
 227         int fd = open(filename, O_RDONLY);
 228         if (fd == -1) {
 229                 perror(filename);
 230                 exit(1);
 231         }
 232         off_t len = lseek(fd, 0, SEEK_END);
 233         if (len == -1) {
 234                 perror("lseek");
 235                 exit(1);
 236         }
 237         const char *data = (char *)mmap(nullptr, len, PROT_READ, MAP_SHARED, fd, /*offset=*/0);
 238         if (data == MAP_FAILED) {
 239                 perror("mmap");
 240                 exit(1);
 241         }
 242
 243         const db_header *hdr = (const db_header *)data;
 244
 245         // TODO: Care about the base path.
 246         string path = data + sizeof(db_header);
 247         uint64_t offset = sizeof(db_header) + path.size() + 1 + ntohl(hdr->conf_size);
 248
 249         const char *ptr = data + offset;
 250         while (ptr < data + len) {
 251                 ptr = handle_directory(ptr, corpus);
 252         }
 253
 254         munmap((void *)data, len);
 255         close(fd);
 256 }
 257
 258 string zstd_compress(const string &src, string *tempbuf)
 259 {
 260         size_t max_size = ZSTD_compressBound(src.size());
 261         if (tempbuf->size() < max_size) {
 262                 tempbuf->resize(max_size);
 263         }
 264         size_t size = ZSTD_compress(&(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
 265         return string(tempbuf->data(), size);
 266 }
 267
 268 void do_build(const char *infile, const char *outfile, int block_size)
 269 {
 270         //steady_clock::time_point start = steady_clock::now();
 271
 272         Corpus corpus(block_size);
 273
 274         read_mlocate(infile, &corpus);
 275         if (false) {  // To read a plain text file.
 276                 FILE *fp = fopen(infile, "r");
 277                 while (!feof(fp)) {
 278                         char buf[1024];
 279                         if (fgets(buf, 1024, fp) == nullptr || feof(fp)) {
 280                                 break;
 281                         }
 282                         string s(buf);
 283                         if (s.back() == '\n')
 284                                 s.pop_back();
 285                         corpus.add_file(move(s));
 286                 }
 287                 fclose(fp);
 288         }
 289         corpus.flush_block();
 290         dprintf("Read %zu files from %s\n", corpus.num_files, infile);
 291
 292         size_t trigrams = 0, longest_posting_list = 0;
 293         size_t bytes_used = 0;
 294         for (auto &[trigram, pl_builder] : corpus.invindex) {
 295                 pl_builder.finish();
 296                 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
 297                 trigrams += pl_builder.num_docids;
 298                 bytes_used += pl_builder.encoded.size();
 299         }
 300         dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
 301                 corpus.num_files, corpus.invindex.size(), trigrams, double(trigrams) / corpus.invindex.size(), longest_posting_list);
 302         dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_used, 8 * bytes_used / double(trigrams));
 303
 304         //steady_clock::time_point end = steady_clock::now();
 305         dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(end - start).count());
 306
 307         vector<uint32_t> all_trigrams;
 308         for (auto &[trigram, pl_builder] : corpus.invindex) {
 309                 all_trigrams.push_back(trigram);
 310         }
 311         sort(all_trigrams.begin(), all_trigrams.end());
 312
 313         umask(0027);
 314         FILE *outfp = fopen(outfile, "wb");
 315
 316         // Write the header.
 317         uint64_t num_trigrams = corpus.invindex.size();  // 64 to get alignment.
 318         fwrite(&num_trigrams, sizeof(num_trigrams), 1, outfp);
 319
 320         // Find out where the posting lists will end.
 321         uint64_t offset = sizeof(uint64_t) * 2 + 16 * all_trigrams.size();
 322         uint64_t filename_index_offset = offset;
 323         for (auto &[trigram, pl_builder] : corpus.invindex) {
 324                 filename_index_offset += pl_builder.encoded.size();
 325         }
 326         fwrite(&filename_index_offset, sizeof(filename_index_offset), 1, outfp);
 327
 328         size_t bytes_for_trigrams = 0, bytes_for_posting_lists = 0, bytes_for_filename_index = 0, bytes_for_filenames = 0;
 329         for (uint32_t trgm : all_trigrams) {
 330                 // 16 bytes: trgm (4), number of docids (4), byte offset (8)
 331                 fwrite(&trgm, sizeof(trgm), 1, outfp);  // One byte wasted, but OK.
 332                 const PostingListBuilder &pl_builder = corpus.invindex[trgm];
 333                 uint32_t num_docids = pl_builder.num_docids;
 334                 fwrite(&num_docids, sizeof(num_docids), 1, outfp);
 335                 fwrite(&offset, sizeof(offset), 1, outfp);
 336                 offset += pl_builder.encoded.size();
 337                 bytes_for_trigrams += 16;
 338         }
 339
 340         // Write the actual posting lists.
 341         for (uint32_t trgm : all_trigrams) {
 342                 const string &encoded = corpus.invindex[trgm].encoded;
 343                 fwrite(encoded.data(), encoded.size(), 1, outfp);
 344                 bytes_for_posting_lists += encoded.size();
 345         }
 346
 347         // Stick an empty block at the end as sentinel.
 348         corpus.filename_blocks.push_back("");
 349
 350         // Write the offsets to the filenames.
 351         offset = filename_index_offset + corpus.filename_blocks.size() * sizeof(offset);
 352         for (const string &filename : corpus.filename_blocks) {
 353                 fwrite(&offset, sizeof(offset), 1, outfp);
 354                 offset += filename.size();
 355                 bytes_for_filename_index += sizeof(offset);
 356                 bytes_for_filenames += filename.size();
 357         }
 358
 359         // Write the actual filenames.
 360         for (const string &filename : corpus.filename_blocks) {
 361                 fwrite(filename.data(), filename.size(), 1, outfp);
 362         }
 363
 364         fclose(outfp);
 365
 366         //size_t total_bytes = (bytes_for_trigrams + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
 367
 368         dprintf("Block size:     %7d files\n", block_size);
 369         dprintf("Trigrams:       %'7.1f MB\n", bytes_for_trigrams / 1048576.0);
 370         dprintf("Posting lists:  %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
 371         dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
 372         dprintf("Filenames:      %'7.1f MB\n", bytes_for_filenames / 1048576.0);
 373         dprintf("Total:          %'7.1f MB\n", total_bytes / 1048576.0);
 374         dprintf("\n");
 375 }
 376
 377 int main(int argc, char **argv)
 378 {
 379         do_build(argv[1], argv[2], 32);
 380         exit(EXIT_SUCCESS);
 381 }