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

   1 #include "db.h"
   2 #include "io_uring_engine.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 <functional>
  11 #include <getopt.h>
  12 #include <limits.h>
  13 #include <memory>
  14 #include <stdio.h>
  15 #include <string.h>
  16 #include <string>
  17 #include <unistd.h>
  18 #include <unordered_map>
  19 #include <vector>
  20 #include <zstd.h>
  21
  22 using namespace std;
  23 using namespace std::chrono;
  24
  25 #define dprintf(...)
  26 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
  27
  28 #include "turbopfor.h"
  29
  30 const char *dbpath = "/var/lib/mlocate/plocate.db";
  31 bool only_count = false;
  32 bool print_nul = false;
  33 int64_t limit_matches = numeric_limits<int64_t>::max();
  34
  35 class Serializer {
  36 public:
  37         bool ready_to_print(int seq) { return next_seq == seq; }
  38         void print_delayed(int seq, const vector<string> msg);
  39         void release_current();
  40
  41 private:
  42         int next_seq = 0;
  43         struct Element {
  44                 int seq;
  45                 vector<string> msg;
  46
  47                 bool operator<(const Element &other) const
  48                 {
  49                         return seq > other.seq;
  50                 }
  51         };
  52         priority_queue<Element> pending;
  53 };
  54
  55 void Serializer::print_delayed(int seq, const vector<string> msg)
  56 {
  57         pending.push(Element{ seq, move(msg) });
  58 }
  59
  60 void Serializer::release_current()
  61 {
  62         ++next_seq;
  63
  64         // See if any delayed prints can now be dealt with.
  65         while (!pending.empty() && pending.top().seq == next_seq) {
  66                 if (limit_matches-- <= 0)
  67                         return;
  68                 for (const string &msg : pending.top().msg) {
  69                         if (print_nul) {
  70                                 printf("%s%c", msg.c_str(), 0);
  71                         } else {
  72                                 printf("%s\n", msg.c_str());
  73                         }
  74                 }
  75                 pending.pop();
  76                 ++next_seq;
  77         }
  78 }
  79
  80 static inline uint32_t read_unigram(const string &s, size_t idx)
  81 {
  82         if (idx < s.size()) {
  83                 return (unsigned char)s[idx];
  84         } else {
  85                 return 0;
  86         }
  87 }
  88
  89 static inline uint32_t read_trigram(const string &s, size_t start)
  90 {
  91         return read_unigram(s, start) | (read_unigram(s, start + 1) << 8) |
  92                 (read_unigram(s, start + 2) << 16);
  93 }
  94
  95 bool has_access(const char *filename,
  96                 unordered_map<string, bool> *access_rx_cache)
  97 {
  98         const char *end = strchr(filename + 1, '/');
  99         while (end != nullptr) {
 100                 string parent_path(filename, end);
 101                 auto it = access_rx_cache->find(parent_path);
 102                 bool ok;
 103                 if (it == access_rx_cache->end()) {
 104                         ok = access(parent_path.c_str(), R_OK | X_OK) == 0;
 105                         access_rx_cache->emplace(move(parent_path), ok);
 106                 } else {
 107                         ok = it->second;
 108                 }
 109                 if (!ok) {
 110                         return false;
 111                 }
 112                 end = strchr(end + 1, '/');
 113         }
 114
 115         return true;
 116 }
 117
 118 class Corpus {
 119 public:
 120         Corpus(int fd, IOUringEngine *engine);
 121         ~Corpus();
 122         void find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb);
 123         void get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const;
 124         size_t get_num_filename_blocks() const;
 125         off_t offset_for_block(uint32_t docid) const
 126         {
 127                 return hdr.filename_index_offset_bytes + docid * sizeof(uint64_t);
 128         }
 129
 130 public:
 131         const int fd;
 132         IOUringEngine *const engine;
 133
 134         Header hdr;
 135 };
 136
 137 Corpus::Corpus(int fd, IOUringEngine *engine)
 138         : fd(fd), engine(engine)
 139 {
 140         // Enable to test cold-cache behavior (except for access()).
 141         if (false) {
 142                 off_t len = lseek(fd, 0, SEEK_END);
 143                 if (len == -1) {
 144                         perror("lseek");
 145                         exit(1);
 146                 }
 147                 posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED);
 148         }
 149
 150         complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
 151         if (memcmp(hdr.magic, "\0plocate", 8) != 0) {
 152                 fprintf(stderr, "plocate.db is corrupt or an old version; please rebuild it.\n");
 153                 exit(1);
 154         }
 155         if (hdr.version != 0) {
 156                 fprintf(stderr, "plocate.db has version %u, expected 0; please rebuild it.\n", hdr.version);
 157                 exit(1);
 158         }
 159 }
 160
 161 Corpus::~Corpus()
 162 {
 163         close(fd);
 164 }
 165
 166 void Corpus::find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb)
 167 {
 168         uint32_t bucket = hash_trigram(trgm, hdr.hashtable_size);
 169         engine->submit_read(fd, sizeof(Trigram) * (hdr.extra_ht_slots + 2), hdr.hash_table_offset_bytes + sizeof(Trigram) * bucket, [this, trgm, cb{ move(cb) }](string_view s) {
 170                 const Trigram *trgmptr = reinterpret_cast<const Trigram *>(s.data());
 171                 for (unsigned i = 0; i < hdr.extra_ht_slots + 1; ++i) {
 172                         if (trgmptr[i].trgm == trgm) {
 173                                 cb(trgmptr + i, trgmptr[i + 1].offset - trgmptr[i].offset);
 174                                 return;
 175                         }
 176                 }
 177
 178                 // Not found.
 179                 cb(nullptr, 0);
 180         });
 181 }
 182
 183 void Corpus::get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const
 184 {
 185         // Read the file offset from this docid and the next one.
 186         // This is always allowed, since we have a sentinel block at the end.
 187         engine->submit_read(fd, sizeof(uint64_t) * 2, offset_for_block(docid), [this, cb{ move(cb) }](string_view s) {
 188                 const uint64_t *ptr = reinterpret_cast<const uint64_t *>(s.data());
 189                 off_t offset = ptr[0];
 190                 size_t len = ptr[1] - ptr[0];
 191                 engine->submit_read(fd, len, offset, cb);
 192         });
 193 }
 194
 195 size_t Corpus::get_num_filename_blocks() const
 196 {
 197         return hdr.num_docids;
 198 }
 199
 200 uint64_t scan_file_block(const vector<string> &needles, string_view compressed,
 201                          unordered_map<string, bool> *access_rx_cache, int seq,
 202                          Serializer *serializer)
 203 {
 204         uint64_t matched = 0;
 205
 206         unsigned long long uncompressed_len = ZSTD_getFrameContentSize(compressed.data(), compressed.size());
 207         if (uncompressed_len == ZSTD_CONTENTSIZE_UNKNOWN || uncompressed_len == ZSTD_CONTENTSIZE_ERROR) {
 208                 fprintf(stderr, "ZSTD_getFrameContentSize() failed\n");
 209                 exit(1);
 210         }
 211
 212         string block;
 213         block.resize(uncompressed_len + 1);
 214
 215         size_t err = ZSTD_decompress(&block[0], block.size(), compressed.data(),
 216                                      compressed.size());
 217         if (ZSTD_isError(err)) {
 218                 fprintf(stderr, "ZSTD_decompress(): %s\n", ZSTD_getErrorName(err));
 219                 exit(1);
 220         }
 221         block[block.size() - 1] = '\0';
 222
 223         bool immediate_print = (serializer == nullptr || serializer->ready_to_print(seq));
 224         vector<string> delayed;
 225
 226         for (const char *filename = block.data();
 227              filename != block.data() + block.size();
 228              filename += strlen(filename) + 1) {
 229                 bool found = true;
 230                 for (const string &needle : needles) {
 231                         if (strstr(filename, needle.c_str()) == nullptr) {
 232                                 found = false;
 233                                 break;
 234                         }
 235                 }
 236                 if (found && has_access(filename, access_rx_cache)) {
 237                         if (limit_matches-- <= 0)
 238                                 break;
 239                         ++matched;
 240                         if (only_count)
 241                                 continue;
 242                         if (immediate_print) {
 243                                 if (print_nul) {
 244                                         printf("%s%c", filename, 0);
 245                                 } else {
 246                                         printf("%s\n", filename);
 247                                 }
 248                         } else {
 249                                 delayed.push_back(filename);
 250                         }
 251                 }
 252         }
 253         if (serializer != nullptr && !only_count) {
 254                 if (immediate_print) {
 255                         serializer->release_current();
 256                 } else {
 257                         serializer->print_delayed(seq, move(delayed));
 258                 }
 259         }
 260         return matched;
 261 }
 262
 263 size_t scan_docids(const vector<string> &needles, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
 264 {
 265         Serializer docids_in_order;
 266         unordered_map<string, bool> access_rx_cache;
 267         uint64_t matched = 0;
 268         for (size_t i = 0; i < docids.size(); ++i) {
 269                 uint32_t docid = docids[i];
 270                 corpus.get_compressed_filename_block(docid, [i, &matched, &needles, &access_rx_cache, &docids_in_order](string_view compressed) {
 271                         matched += scan_file_block(needles, compressed, &access_rx_cache, i, &docids_in_order);
 272                 });
 273         }
 274         engine->finish();
 275         return matched;
 276 }
 277
 278 // We do this sequentially, as it's faster than scattering
 279 // a lot of I/O through io_uring and hoping the kernel will
 280 // coalesce it plus readahead for us.
 281 uint64_t scan_all_docids(const vector<string> &needles, int fd, const Corpus &corpus, IOUringEngine *engine)
 282 {
 283         unordered_map<string, bool> access_rx_cache;
 284         uint32_t num_blocks = corpus.get_num_filename_blocks();
 285         unique_ptr<uint64_t[]> offsets(new uint64_t[num_blocks + 1]);
 286         complete_pread(fd, offsets.get(), (num_blocks + 1) * sizeof(uint64_t), corpus.offset_for_block(0));
 287         string compressed;
 288         uint64_t matched = 0;
 289         for (uint32_t io_docid = 0; io_docid < num_blocks; io_docid += 32) {
 290                 uint32_t last_docid = std::min(io_docid + 32, num_blocks);
 291                 size_t io_len = offsets[last_docid] - offsets[io_docid];
 292                 if (compressed.size() < io_len) {
 293                         compressed.resize(io_len);
 294                 }
 295                 complete_pread(fd, &compressed[0], io_len, offsets[io_docid]);
 296
 297                 for (uint32_t docid = io_docid; docid < last_docid; ++docid) {
 298                         size_t relative_offset = offsets[docid] - offsets[io_docid];
 299                         size_t len = offsets[docid + 1] - offsets[docid];
 300                         matched += scan_file_block(needles, { &compressed[relative_offset], len }, &access_rx_cache, 0, nullptr);
 301                         if (limit_matches <= 0)
 302                                 return matched;
 303                 }
 304         }
 305         return matched;
 306 }
 307
 308 void do_search_file(const vector<string> &needles, const char *filename)
 309 {
 310         int fd = open(filename, O_RDONLY);
 311         if (fd == -1) {
 312                 perror(filename);
 313                 exit(1);
 314         }
 315
 316         // Drop privileges.
 317         if (setgid(getgid()) != 0) {
 318                 perror("setgid");
 319                 exit(EXIT_FAILURE);
 320         }
 321
 322         steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
 323         if (access("/", R_OK | X_OK)) {
 324                 // We can't find anything, no need to bother...
 325                 return;
 326         }
 327
 328         IOUringEngine engine(/*slop_bytes=*/16);  // 16 slop bytes as described in turbopfor.h.
 329         Corpus corpus(fd, &engine);
 330         dprintf("Corpus init done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 331
 332         vector<pair<Trigram, size_t>> trigrams;
 333         uint64_t shortest_so_far = numeric_limits<uint32_t>::max();
 334         for (const string &needle : needles) {
 335                 if (needle.size() < 3)
 336                         continue;
 337                 for (size_t i = 0; i < needle.size() - 2; ++i) {
 338                         uint32_t trgm = read_trigram(needle, i);
 339                         corpus.find_trigram(trgm, [trgm, &trigrams, &shortest_so_far](const Trigram *trgmptr, size_t len) {
 340                                 if (trgmptr == nullptr) {
 341                                         dprintf("trigram '%c%c%c' isn't found, we abort the search\n",
 342                                                 trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff);
 343                                         if (only_count) {
 344                                                 printf("0\n");
 345                                         }
 346                                         exit(0);
 347                                 }
 348                                 if (trgmptr->num_docids > shortest_so_far * 100) {
 349                                         dprintf("not loading trigram '%c%c%c' with %u docids, it would be ignored later anyway\n",
 350                                                 trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff,
 351                                                 trgmptr->num_docids);
 352                                 } else {
 353                                         trigrams.emplace_back(*trgmptr, len);
 354                                         shortest_so_far = std::min<uint64_t>(shortest_so_far, trgmptr->num_docids);
 355                                 }
 356                         });
 357                 }
 358         }
 359         engine.finish();
 360         dprintf("Hashtable lookups done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 361
 362         if (trigrams.empty()) {
 363                 // Too short for trigram matching. Apply brute force.
 364                 // (We could have searched through all trigrams that matched
 365                 // the pattern and done a union of them, but that's a lot of
 366                 // work for fairly unclear gain.)
 367                 uint64_t matched = scan_all_docids(needles, fd, corpus, &engine);
 368                 printf("%zu\n", matched);
 369                 return;
 370         }
 371         sort(trigrams.begin(), trigrams.end());
 372         {
 373                 auto last = unique(trigrams.begin(), trigrams.end());
 374                 trigrams.erase(last, trigrams.end());
 375         }
 376         sort(trigrams.begin(), trigrams.end(),
 377              [&](const pair<Trigram, size_t> &a, const pair<Trigram, size_t> &b) {
 378                      return a.first.num_docids < b.first.num_docids;
 379              });
 380
 381         vector<uint32_t> in1, in2, out;
 382         bool done = false;
 383         for (auto [trgmptr, len] : trigrams) {
 384                 if (!in1.empty() && trgmptr.num_docids > in1.size() * 100) {
 385                         uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
 386                         dprintf("trigram '%c%c%c' (%zu bytes) has %u entries, ignoring the rest (will "
 387                                 "weed out false positives later)\n",
 388                                 trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff,
 389                                 len, trgmptr.num_docids);
 390                         break;
 391                 }
 392
 393                 // Only stay a certain amount ahead, so that we don't spend I/O
 394                 // on reading the latter, large posting lists. We are unlikely
 395                 // to need them anyway, even if they should come in first.
 396                 if (engine.get_waiting_reads() >= 5) {
 397                         engine.finish();
 398                         if (done)
 399                                 break;
 400                 }
 401                 engine.submit_read(fd, len, trgmptr.offset, [trgmptr{ trgmptr }, len{ len }, &done, &in1, &in2, &out](string_view s) {
 402                         if (done)
 403                                 return;
 404                         uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
 405                         size_t num = trgmptr.num_docids;
 406                         const unsigned char *pldata = reinterpret_cast<const unsigned char *>(s.data());
 407                         if (in1.empty()) {
 408                                 in1.resize(num + 128);
 409                                 decode_pfor_delta1_128(pldata, num, /*interleaved=*/true, &in1[0]);
 410                                 in1.resize(num);
 411                                 dprintf("trigram '%c%c%c' (%zu bytes) decoded to %zu entries\n", trgm & 0xff,
 412                                         (trgm >> 8) & 0xff, (trgm >> 16) & 0xff, len, num);
 413                         } else {
 414                                 if (in2.size() < num + 128) {
 415                                         in2.resize(num + 128);
 416                                 }
 417                                 decode_pfor_delta1_128(pldata, num, /*interleaved=*/true, &in2[0]);
 418
 419                                 out.clear();
 420                                 set_intersection(in1.begin(), in1.end(), in2.begin(), in2.begin() + num,
 421                                                  back_inserter(out));
 422                                 swap(in1, out);
 423                                 dprintf("trigram '%c%c%c' (%zu bytes) decoded to %zu entries, %zu left\n",
 424                                         trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff,
 425                                         len, num, in1.size());
 426                                 if (in1.empty()) {
 427                                         dprintf("no matches (intersection list is empty)\n");
 428                                         done = true;
 429                                 }
 430                         }
 431                 });
 432         }
 433         engine.finish();
 434         if (done) {
 435                 return;
 436         }
 437         dprintf("Intersection done after %.1f ms. Doing final verification and printing:\n",
 438                 1e3 * duration<float>(steady_clock::now() - start).count());
 439
 440         uint64_t matched = scan_docids(needles, in1, corpus, &engine);
 441         dprintf("Done in %.1f ms, found %zu matches.\n",
 442                 1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 443
 444         if (only_count) {
 445                 printf("%zu\n", matched);
 446         }
 447 }
 448
 449 void usage()
 450 {
 451         // The help text comes from mlocate.
 452         printf("Usage: plocate [OPTION]... PATTERN...\n");
 453         printf("\n");
 454         printf("  -c, --count            only print number of found entries\n");
 455         printf("  -d, --database DBPATH  use DBPATH instead of default database (which is\n");
 456         printf("                         %s)\n", dbpath);
 457         printf("  -h, --help             print this help\n");
 458         printf("  -l, --limit, -n LIMIT  limit output (or counting) to LIMIT entries\n");
 459         printf("  -0, --null             separate entries with NUL on output\n");
 460 }
 461
 462 int main(int argc, char **argv)
 463 {
 464         static const struct option long_options[] = {
 465                 { "help", no_argument, 0, 'h' },
 466                 { "count", no_argument, 0, 'c' },
 467                 { "database", required_argument, 0, 'd' },
 468                 { "limit", required_argument, 0, 'l' },
 469                 { nullptr, required_argument, 0, 'n' },
 470                 { "null", no_argument, 0, '0' },
 471                 { 0, 0, 0, 0 }
 472         };
 473
 474         for (;;) {
 475                 int option_index = 0;
 476                 int c = getopt_long(argc, argv, "cd:hl:n:0", long_options, &option_index);
 477                 if (c == -1) {
 478                         break;
 479                 }
 480                 switch (c) {
 481                 case 'c':
 482                         only_count = true;
 483                         break;
 484                 case 'd':
 485                         dbpath = strdup(optarg);
 486                         break;
 487                 case 'h':
 488                         usage();
 489                         exit(0);
 490                 case 'l':
 491                 case 'n':
 492                         limit_matches = atoll(optarg);
 493                         break;
 494                 case '0':
 495                         print_nul = true;
 496                         break;
 497                 default:
 498                         exit(1);
 499                 }
 500         }
 501
 502         vector<string> needles;
 503         for (int i = optind; i < argc; ++i) {
 504                 needles.push_back(argv[i]);
 505         }
 506         if (needles.empty()) {
 507                 fprintf(stderr, "plocate: no pattern to search for specified\n");
 508                 exit(0);
 509         }
 510         do_search_file(needles, dbpath);
 511 }