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

   1 #include <stdio.h>
   2 #include <string.h>
   3 #include <algorithm>
   4 #include <unordered_map>
   5 #include <string>
   6 #include <vector>
   7 #include <chrono>
   8 #include <unistd.h>
   9 #include <fcntl.h>
  10 #include <sys/mman.h>
  11 #include <arpa/inet.h>
  12 #include <endian.h>
  13 #include <sys/types.h>
  14 #include <sys/stat.h>
  15 #include <zstd.h>
  16
  17 #include "vp4.h"
  18
  19 #define P4NENC_BOUND(n) ((n+127)/128+(n+32)*sizeof(uint32_t))
  20 #define dprintf(...)
  21 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
  22
  23 using namespace std;
  24 using namespace std::chrono;
  25
  26 string zstd_compress(const string &src, string *tempbuf);
  27
  28 static inline uint32_t read_unigram(const string_view s, size_t idx)
  29 {
  30         if (idx < s.size()) {
  31                 return (unsigned char)s[idx];
  32         } else {
  33                 return 0;
  34         }
  35 }
  36
  37 static inline uint32_t read_trigram(const string_view s, size_t start)
  38 {
  39         return read_unigram(s, start) |
  40                 (read_unigram(s, start + 1) << 8) |
  41                 (read_unigram(s, start + 2) << 16);
  42 }
  43
  44 enum
  45 {
  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 {
  54         uint8_t magic[8];
  55         uint32_t conf_size;
  56         uint8_t version;
  57         uint8_t check_visibility;
  58         uint8_t pad[2];
  59 };
  60
  61 // From mlocate.
  62 struct db_directory
  63 {
  64         uint64_t time_sec;
  65         uint32_t time_nsec;
  66         uint8_t pad[4];
  67 };
  68
  69 class Corpus {
  70 public:
  71         Corpus(size_t block_size) : block_size(block_size) {}
  72         void add_file(string filename);
  73         void flush_block();
  74
  75         vector<string> filename_blocks;
  76         unordered_map<uint32_t, vector<uint32_t>> invindex;
  77         size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
  78
  79 private:
  80         string current_block;
  81         string tempbuf;
  82         const size_t block_size;
  83 };
  84
  85 void Corpus::add_file(string filename)
  86 {
  87         ++num_files;
  88         if (!current_block.empty()) {
  89                 current_block.push_back('\0');
  90         }
  91         current_block += filename;
  92         if (++num_files_in_block == block_size) {
  93                 flush_block();
  94         }
  95 }
  96
  97 void Corpus::flush_block()
  98 {
  99         if (current_block.empty()) {
 100                 return;
 101         }
 102
 103         uint32_t docid = num_blocks;
 104
 105         // Create trigrams.
 106         const char *ptr = current_block.c_str();
 107         while (ptr < current_block.c_str() + current_block.size()) {
 108                 string_view s(ptr);
 109                 if (s.size() >= 3) {
 110                         for (size_t j = 0; j < s.size() - 2; ++j) {
 111                                 uint32_t trgm = read_trigram(s, j);
 112                                 vector<uint32_t> &docids = invindex[trgm];
 113                                 if (docids.empty() || docids.back() != docid) {
 114                                         docids.push_back(docid);
 115                                 }
 116                         }
 117                 }
 118                 ptr += s.size() + 1;
 119         }
 120
 121         // Compress and add the filename block.
 122         filename_blocks.push_back(zstd_compress(current_block, &tempbuf));
 123
 124         current_block.clear();
 125         num_files_in_block = 0;
 126         ++num_blocks;
 127 }
 128
 129 const char *handle_directory(const char *ptr, Corpus *corpus)
 130 {
 131         ptr += sizeof(db_directory);
 132
 133         string dir_path = ptr;
 134         ptr += dir_path.size() + 1;
 135         if (dir_path == "/") {
 136                 dir_path = "";
 137         }
 138
 139         for ( ;; ) {
 140                 uint8_t type = *ptr++;
 141                 if (type == DBE_NORMAL) {
 142                         string filename = ptr;
 143                         corpus->add_file(dir_path + "/" + filename);
 144                         ptr += filename.size() + 1;
 145                 } else if (type == DBE_DIRECTORY) {
 146                         string dirname = ptr;
 147                         corpus->add_file(dir_path + "/" + dirname);
 148                         ptr += dirname.size() + 1;
 149                 } else {
 150                         return ptr;
 151                 }
 152         }
 153 }
 154
 155 void read_mlocate(const char *filename, Corpus *corpus)
 156 {
 157         int fd = open(filename, O_RDONLY);
 158         if (fd == -1) {
 159                 perror(filename);
 160                 exit(1);
 161         }
 162         off_t len = lseek(fd, 0, SEEK_END);
 163         if (len == -1) {
 164                 perror("lseek");
 165                 exit(1);
 166         }
 167         const char *data = (char *)mmap(nullptr, len, PROT_READ, MAP_SHARED, fd, /*offset=*/0);
 168         if (data == MAP_FAILED) {
 169                 perror("mmap");
 170                 exit(1);
 171         }
 172
 173         const db_header *hdr = (const db_header *)data;
 174
 175         // TODO: Care about the base path.
 176         string path = data + sizeof(db_header);
 177         uint64_t offset = sizeof(db_header) + path.size() + 1 + ntohl(hdr->conf_size);
 178
 179         const char *ptr = data + offset;
 180         while (ptr < data + len) {
 181                 ptr = handle_directory(ptr, corpus);
 182         }
 183
 184         munmap((void *)data, len);
 185         close(fd);
 186 }
 187
 188 string zstd_compress(const string &src, string *tempbuf)
 189 {
 190         size_t max_size = ZSTD_compressBound(src.size());
 191         if (tempbuf->size() < max_size) {
 192                 tempbuf->resize(max_size);
 193         }
 194         size_t size = ZSTD_compress(&(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
 195         return string(tempbuf->data(), size);
 196 }
 197
 198 void do_build(const char *infile, const char *outfile, int block_size)
 199 {
 200         //steady_clock::time_point start = steady_clock::now();
 201
 202         Corpus corpus(block_size);
 203
 204         read_mlocate(infile, &corpus);
 205         if (false) {  // To read a plain text file.
 206                 FILE *fp = fopen(infile, "r");
 207                 while (!feof(fp)) {
 208                         char buf[1024];
 209                         if (fgets(buf, 1024, fp) == nullptr || feof(fp)) {
 210                                 break;
 211                         }
 212                         string s(buf);
 213                         if (s.back() == '\n') s.pop_back();
 214                         corpus.add_file(move(s));
 215                 }
 216                 fclose(fp);
 217         }
 218         corpus.flush_block();
 219         dprintf("Read %zu files from %s\n", corpus.num_files, infile);
 220
 221         // Compress posting lists.
 222         string buf;
 223         size_t trigrams = 0, longest_posting_list = 0;
 224         size_t bytes_used = 0;
 225         unordered_map<uint32_t, string> pl;
 226         for (auto &[trigram, docids] : corpus.invindex) {
 227                 longest_posting_list = max(longest_posting_list, docids.size());
 228                 trigrams += docids.size();
 229
 230                 size_t bytes_needed = P4NENC_BOUND(docids.size());
 231                 if (buf.size() < bytes_needed) buf.resize(bytes_needed);
 232                 size_t bytes = p4nd1enc128v32(&docids[0], docids.size(), reinterpret_cast<unsigned char *>(&buf[0]));
 233                 pl[trigram] = string(buf.data(), bytes);
 234                 bytes_used += bytes;
 235         }
 236         dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
 237                 corpus.num_files, corpus.invindex.size(), trigrams, double(trigrams) / corpus.invindex.size(), longest_posting_list);
 238
 239         dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_used, 8 * bytes_used / double(trigrams));
 240         //steady_clock::time_point end = steady_clock::now();
 241         dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(end - start).count());
 242
 243         vector<uint32_t> all_trigrams;
 244         for (auto &[trigram, docids] : corpus.invindex) {
 245                 all_trigrams.push_back(trigram);
 246         }
 247         sort(all_trigrams.begin(), all_trigrams.end());
 248
 249         umask(0027);
 250         FILE *outfp = fopen(outfile, "wb");
 251
 252         // Write the header.
 253         uint64_t num_trigrams = corpus.invindex.size();  // 64 to get alignment.
 254         fwrite(&num_trigrams, sizeof(num_trigrams), 1, outfp);
 255
 256         // Find out where the posting lists will end.
 257         uint64_t offset = sizeof(uint64_t) * 2 + 16 * all_trigrams.size();
 258         uint64_t filename_index_offset = offset;
 259         for (uint32_t trgm : all_trigrams) {
 260                 filename_index_offset += pl[trgm].size();
 261         }
 262         fwrite(&filename_index_offset, sizeof(filename_index_offset), 1, outfp);
 263
 264         size_t bytes_for_trigrams = 0, bytes_for_posting_lists = 0, bytes_for_filename_index = 0, bytes_for_filenames = 0;
 265
 266         for (uint32_t trgm : all_trigrams) {
 267                 // 16 bytes: trgm (4), number of docids (4), byte offset (8)
 268                 fwrite(&trgm, sizeof(trgm), 1, outfp);  // One byte wasted, but OK.
 269                 uint32_t num_docids = corpus.invindex[trgm].size();
 270                 fwrite(&num_docids, sizeof(num_docids), 1, outfp);
 271                 fwrite(&offset, sizeof(offset), 1, outfp);
 272                 offset += pl[trgm].size();
 273                 bytes_for_trigrams += 16;
 274         }
 275
 276         // Write the actual posting lists.
 277         for (uint32_t trgm : all_trigrams) {
 278                 fwrite(pl[trgm].data(), pl[trgm].size(), 1, outfp);
 279                 bytes_for_posting_lists += pl[trgm].size();
 280         }
 281
 282         // Stick an empty block at the end as sentinel.
 283         corpus.filename_blocks.push_back("");
 284
 285         // Write the offsets to the filenames.
 286         offset = filename_index_offset + corpus.filename_blocks.size() * sizeof(offset);
 287         for (const string &filename : corpus.filename_blocks) {
 288                 fwrite(&offset, sizeof(offset), 1, outfp);
 289                 offset += filename.size();
 290                 bytes_for_filename_index += sizeof(offset);
 291                 bytes_for_filenames += filename.size();
 292         }
 293
 294         // Write the actual filenames.
 295         for (const string &filename : corpus.filename_blocks) {
 296                 fwrite(filename.data(), filename.size(), 1, outfp);
 297         }
 298
 299         fclose(outfp);
 300
 301         //size_t total_bytes = (bytes_for_trigrams + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
 302
 303         dprintf("Block size:     %7d files\n", block_size);
 304         dprintf("Trigrams:       %'7.1f MB\n", bytes_for_trigrams / 1048576.0);
 305         dprintf("Posting lists:  %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
 306         dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
 307         dprintf("Filenames:      %'7.1f MB\n", bytes_for_filenames / 1048576.0);
 308         dprintf("Total:          %'7.1f MB\n", total_bytes / 1048576.0);
 309         dprintf("\n");
 310 }
 311
 312 int main(int argc, char **argv)
 313 {
 314         do_build(argv[1], argv[2], 32);
 315         exit(EXIT_SUCCESS);
 316 }