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

   1 #include "db.h"
   2 #include "io_uring_engine.h"
   3 #include "vp4.h"
   4
   5 #include <algorithm>
   6 #include <arpa/inet.h>
   7 #include <chrono>
   8 #include <endian.h>
   9 #include <fcntl.h>
  10 #include <functional>
  11 #include <getopt.h>
  12 #include <memory>
  13 #include <stdio.h>
  14 #include <string.h>
  15 #include <string>
  16 #include <unistd.h>
  17 #include <unordered_map>
  18 #include <vector>
  19 #include <zstd.h>
  20
  21 using namespace std;
  22 using namespace std::chrono;
  23
  24 #define dprintf(...)
  25 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
  26
  27 class Serializer {
  28 public:
  29         void do_or_wait(int seq, function<void()> cb);
  30
  31 private:
  32         int next_seq = 0;
  33         struct Element {
  34                 int seq;
  35                 function<void()> cb;
  36
  37                 bool operator<(const Element &other) const
  38                 {
  39                         return seq > other.seq;
  40                 }
  41         };
  42         priority_queue<Element> pending;
  43 };
  44
  45 void Serializer::do_or_wait(int seq, function<void()> cb)
  46 {
  47         if (seq != next_seq) {
  48                 pending.emplace(Element{ seq, move(cb) });
  49                 return;
  50         }
  51
  52         cb();
  53         ++next_seq;
  54
  55         while (!pending.empty() && pending.top().seq == next_seq) {
  56                 pending.top().cb();
  57                 pending.pop();
  58                 ++next_seq;
  59         }
  60 }
  61
  62 static inline uint32_t read_unigram(const string &s, size_t idx)
  63 {
  64         if (idx < s.size()) {
  65                 return (unsigned char)s[idx];
  66         } else {
  67                 return 0;
  68         }
  69 }
  70
  71 static inline uint32_t read_trigram(const string &s, size_t start)
  72 {
  73         return read_unigram(s, start) | (read_unigram(s, start + 1) << 8) |
  74                 (read_unigram(s, start + 2) << 16);
  75 }
  76
  77 bool has_access(const char *filename,
  78                 unordered_map<string, bool> *access_rx_cache)
  79 {
  80         const char *end = strchr(filename + 1, '/');
  81         while (end != nullptr) {
  82                 string parent_path(filename, end);
  83                 auto it = access_rx_cache->find(parent_path);
  84                 bool ok;
  85                 if (it == access_rx_cache->end()) {
  86                         ok = access(parent_path.c_str(), R_OK | X_OK) == 0;
  87                         access_rx_cache->emplace(move(parent_path), ok);
  88                 } else {
  89                         ok = it->second;
  90                 }
  91                 if (!ok) {
  92                         return false;
  93                 }
  94                 end = strchr(end + 1, '/');
  95         }
  96
  97         return true;
  98 }
  99
 100 class Corpus {
 101 public:
 102         Corpus(int fd, IOUringEngine *engine);
 103         ~Corpus();
 104         void find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb);
 105         void get_compressed_filename_block(uint32_t docid, function<void(string)> cb) const;
 106         size_t get_num_filename_blocks() const;
 107         off_t offset_for_block(uint32_t docid) const
 108         {
 109                 return hdr.filename_index_offset_bytes + docid * sizeof(uint64_t);
 110         }
 111
 112 public:
 113         const int fd;
 114         IOUringEngine *const engine;
 115
 116         Header hdr;
 117 };
 118
 119 Corpus::Corpus(int fd, IOUringEngine *engine)
 120         : fd(fd), engine(engine)
 121 {
 122         // Enable to test cold-cache behavior (except for access()).
 123         if (false) {
 124                 off_t len = lseek(fd, 0, SEEK_END);
 125                 if (len == -1) {
 126                         perror("lseek");
 127                         exit(1);
 128                 }
 129                 posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED);
 130         }
 131
 132         complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
 133         if (memcmp(hdr.magic, "\0plocate", 8) != 0) {
 134                 fprintf(stderr, "plocate.db is corrupt or an old version; please rebuild it.\n");
 135                 exit(1);
 136         }
 137         if (hdr.version != 0) {
 138                 fprintf(stderr, "plocate.db has version %u, expected 0; please rebuild it.\n", hdr.version);
 139                 exit(1);
 140         }
 141 }
 142
 143 Corpus::~Corpus()
 144 {
 145         close(fd);
 146 }
 147
 148 void Corpus::find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb)
 149 {
 150         uint32_t bucket = hash_trigram(trgm, hdr.hashtable_size);
 151         engine->submit_read(fd, sizeof(Trigram) * (hdr.extra_ht_slots + 2), hdr.hash_table_offset_bytes + sizeof(Trigram) * bucket, [this, trgm, bucket, cb{ move(cb) }](string s) {
 152                 const Trigram *trgmptr = reinterpret_cast<const Trigram *>(s.data());
 153                 for (unsigned i = 0; i < hdr.extra_ht_slots + 1; ++i) {
 154                         if (trgmptr[i].trgm == trgm) {
 155                                 cb(trgmptr + i, trgmptr[i + 1].offset - trgmptr[i].offset);
 156                                 return;
 157                         }
 158                 }
 159
 160                 // Not found.
 161                 cb(nullptr, 0);
 162         });
 163 }
 164
 165 void Corpus::get_compressed_filename_block(uint32_t docid, function<void(string)> cb) const
 166 {
 167         // Read the file offset from this docid and the next one.
 168         // This is always allowed, since we have a sentinel block at the end.
 169         engine->submit_read(fd, sizeof(uint64_t) * 2, offset_for_block(docid), [this, cb{ move(cb) }](string s) {
 170                 const uint64_t *ptr = reinterpret_cast<const uint64_t *>(s.data());
 171                 off_t offset = ptr[0];
 172                 size_t len = ptr[1] - ptr[0];
 173                 engine->submit_read(fd, len, offset, cb);
 174         });
 175 }
 176
 177 size_t Corpus::get_num_filename_blocks() const
 178 {
 179         // The beginning of the filename blocks is the end of the filename index blocks.
 180         uint64_t end;
 181         complete_pread(fd, &end, sizeof(end), hdr.filename_index_offset_bytes);
 182
 183         // Subtract the sentinel block.
 184         return (end - hdr.filename_index_offset_bytes) / sizeof(uint64_t) - 1;
 185 }
 186
 187 size_t scan_file_block(const string &needle, string_view compressed,
 188                        unordered_map<string, bool> *access_rx_cache)
 189 {
 190         size_t matched = 0;
 191
 192         unsigned long long uncompressed_len = ZSTD_getFrameContentSize(compressed.data(), compressed.size());
 193         if (uncompressed_len == ZSTD_CONTENTSIZE_UNKNOWN || uncompressed_len == ZSTD_CONTENTSIZE_ERROR) {
 194                 fprintf(stderr, "ZSTD_getFrameContentSize() failed\n");
 195                 exit(1);
 196         }
 197
 198         string block;
 199         block.resize(uncompressed_len + 1);
 200
 201         size_t err = ZSTD_decompress(&block[0], block.size(), compressed.data(),
 202                                      compressed.size());
 203         if (ZSTD_isError(err)) {
 204                 fprintf(stderr, "ZSTD_decompress(): %s\n", ZSTD_getErrorName(err));
 205                 exit(1);
 206         }
 207         block[block.size() - 1] = '\0';
 208
 209         for (const char *filename = block.data();
 210              filename != block.data() + block.size();
 211              filename += strlen(filename) + 1) {
 212                 if (strstr(filename, needle.c_str()) == nullptr) {
 213                         continue;
 214                 }
 215                 if (has_access(filename, access_rx_cache)) {
 216                         ++matched;
 217                         printf("%s\n", filename);
 218                 }
 219         }
 220         return matched;
 221 }
 222
 223 size_t scan_docids(const string &needle, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
 224 {
 225         Serializer docids_in_order;
 226         unordered_map<string, bool> access_rx_cache;
 227         size_t matched = 0;
 228         for (size_t i = 0; i < docids.size(); ++i) {
 229                 uint32_t docid = docids[i];
 230                 corpus.get_compressed_filename_block(docid, [i, &matched, &needle, &access_rx_cache, &docids_in_order](string compressed) {
 231                         docids_in_order.do_or_wait(i, [&matched, &needle, compressed{ move(compressed) }, &access_rx_cache] {
 232                                 matched += scan_file_block(needle, compressed, &access_rx_cache);
 233                         });
 234                 });
 235         }
 236         engine->finish();
 237         return matched;
 238 }
 239
 240 // We do this sequentially, as it's faster than scattering
 241 // a lot of I/O through io_uring and hoping the kernel will
 242 // coalesce it plus readahead for us.
 243 void scan_all_docids(const string &needle, int fd, const Corpus &corpus, IOUringEngine *engine)
 244 {
 245         unordered_map<string, bool> access_rx_cache;
 246         uint32_t num_blocks = corpus.get_num_filename_blocks();
 247         unique_ptr<uint64_t[]> offsets(new uint64_t[num_blocks + 1]);
 248         complete_pread(fd, offsets.get(), (num_blocks + 1) * sizeof(uint64_t), corpus.offset_for_block(0));
 249         string compressed;
 250         for (uint32_t io_docid = 0; io_docid < num_blocks; io_docid += 32) {
 251                 uint32_t last_docid = std::min(io_docid + 32, num_blocks);
 252                 size_t io_len = offsets[last_docid] - offsets[io_docid];
 253                 if (compressed.size() < io_len) {
 254                         compressed.resize(io_len);
 255                 }
 256                 complete_pread(fd, &compressed[0], io_len, offsets[io_docid]);
 257
 258                 for (uint32_t docid = io_docid; docid < last_docid; ++docid) {
 259                         size_t relative_offset = offsets[docid] - offsets[io_docid];
 260                         size_t len = offsets[docid + 1] - offsets[docid];
 261                         scan_file_block(needle, { &compressed[relative_offset], len }, &access_rx_cache);
 262                 }
 263         }
 264 }
 265
 266 void do_search_file(const string &needle, const char *filename)
 267 {
 268         int fd = open(filename, O_RDONLY);
 269         if (fd == -1) {
 270                 perror(filename);
 271                 exit(1);
 272         }
 273
 274         // Drop privileges.
 275         if (setgid(getgid()) != 0) {
 276                 perror("setgid");
 277                 exit(EXIT_FAILURE);
 278         }
 279
 280         steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
 281         if (access("/", R_OK | X_OK)) {
 282                 // We can't find anything, no need to bother...
 283                 return;
 284         }
 285
 286         IOUringEngine engine;
 287         Corpus corpus(fd, &engine);
 288         dprintf("Corpus init done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 289
 290         if (needle.size() < 3) {
 291                 // Too short for trigram matching. Apply brute force.
 292                 // (We could have searched through all trigrams that matched
 293                 // the pattern and done a union of them, but that's a lot of
 294                 // work for fairly unclear gain.)
 295                 scan_all_docids(needle, fd, corpus, &engine);
 296                 return;
 297         }
 298
 299         vector<pair<Trigram, size_t>> trigrams;
 300         for (size_t i = 0; i < needle.size() - 2; ++i) {
 301                 uint32_t trgm = read_trigram(needle, i);
 302                 corpus.find_trigram(trgm, [trgm, &trigrams](const Trigram *trgmptr, size_t len) {
 303                         if (trgmptr == nullptr) {
 304                                 dprintf("trigram %06x isn't found, we abort the search\n", trgm);
 305                                 return;
 306                         }
 307                         trigrams.emplace_back(*trgmptr, len);
 308                 });
 309         }
 310         engine.finish();
 311         dprintf("Hashtable lookups done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 312
 313         sort(trigrams.begin(), trigrams.end());
 314         {
 315                 auto last = unique(trigrams.begin(), trigrams.end());
 316                 trigrams.erase(last, trigrams.end());
 317         }
 318         sort(trigrams.begin(), trigrams.end(),
 319              [&](const pair<Trigram, size_t> &a, const pair<Trigram, size_t> &b) {
 320                      return a.first.num_docids < b.first.num_docids;
 321              });
 322
 323         vector<uint32_t> in1, in2, out;
 324         bool done = false;
 325         for (auto [trgmptr, len] : trigrams) {
 326                 if (!in1.empty() && trgmptr.num_docids > in1.size() * 100) {
 327                         uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
 328                         dprintf("trigram '%c%c%c' (%zu bytes) has %u entries, ignoring the rest (will "
 329                                 "weed out false positives later)\n",
 330                                 trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff,
 331                                 len, trgmptr.num_docids);
 332                         break;
 333                 }
 334
 335                 // Only stay a certain amount ahead, so that we don't spend I/O
 336                 // on reading the latter, large posting lists. We are unlikely
 337                 // to need them anyway, even if they should come in first.
 338                 if (engine.get_waiting_reads() >= 5) {
 339                         engine.finish();
 340                         if (done)
 341                                 break;
 342                 }
 343                 engine.submit_read(fd, len, trgmptr.offset, [trgmptr, len, &done, &in1, &in2, &out](string s) {
 344                         if (done)
 345                                 return;
 346                         uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
 347                         size_t num = trgmptr.num_docids;
 348                         unsigned char *pldata = reinterpret_cast<unsigned char *>(s.data());
 349                         if (in1.empty()) {
 350                                 in1.resize(num + 128);
 351                                 p4nd1dec128v32(pldata, num, &in1[0]);
 352                                 in1.resize(num);
 353                                 dprintf("trigram '%c%c%c' (%zu bytes) decoded to %zu entries\n", trgm & 0xff,
 354                                         (trgm >> 8) & 0xff, (trgm >> 16) & 0xff, len, num);
 355                         } else {
 356                                 if (in2.size() < num + 128) {
 357                                         in2.resize(num + 128);
 358                                 }
 359                                 p4nd1dec128v32(pldata, num, &in2[0]);
 360
 361                                 out.clear();
 362                                 set_intersection(in1.begin(), in1.end(), in2.begin(), in2.begin() + num,
 363                                                  back_inserter(out));
 364                                 swap(in1, out);
 365                                 dprintf("trigram '%c%c%c' (%zu bytes) decoded to %zu entries, %zu left\n",
 366                                         trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff,
 367                                         len, num, in1.size());
 368                                 if (in1.empty()) {
 369                                         dprintf("no matches (intersection list is empty)\n");
 370                                         done = true;
 371                                 }
 372                         }
 373                 });
 374         }
 375         engine.finish();
 376         if (done) {
 377                 return;
 378         }
 379         dprintf("Intersection done after %.1f ms. Doing final verification and printing:\n",
 380                 1e3 * duration<float>(steady_clock::now() - start).count());
 381
 382         size_t matched __attribute__((unused)) = scan_docids(needle, in1, corpus, &engine);
 383         dprintf("Done in %.1f ms, found %zu matches.\n",
 384                 1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 385 }
 386
 387 const char *dbpath = "/var/lib/mlocate/plocate.db";
 388
 389 void usage()
 390 {
 391         printf("Usage: slocate [OPTION]... PATTERN\n");
 392         printf("  -d, --database DBPATH  use DBPATH instead of default database (which is\n");
 393         printf("                         %s)\n", dbpath);
 394         printf("  -h, --help             print this help\n");
 395 }
 396
 397 int main(int argc, char **argv)
 398 {
 399         static const struct option long_options[] = {
 400                 { "help", no_argument, 0, 'h' },
 401                 { "database", required_argument, 0, 'd' },
 402                 { 0, 0, 0, 0 }
 403         };
 404
 405         for (;;) {
 406                 int option_index = 0;
 407                 int c = getopt_long(argc, argv, "d:h:", long_options, &option_index);
 408                 if (c == -1) {
 409                         break;
 410                 }
 411                 switch (c) {
 412                 case 'd':
 413                         dbpath = strdup(optarg);
 414                         break;
 415                 case 'h':
 416                         usage();
 417                         exit(0);
 418                 }
 419         }
 420
 421         if (optind == argc) {
 422                 fprintf(stderr, "slocate: no pattern to search for specified\n");
 423                 exit(0);
 424         }
 425         do_search_file(argv[optind], dbpath);
 426 }