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

   1 #include "db.h"
   2 #include "io_uring_engine.h"
   3
   4 #include <algorithm>
   5 #include <chrono>
   6 #include <fcntl.h>
   7 #include <functional>
   8 #include <getopt.h>
   9 #include <iosfwd>
  10 #include <iterator>
  11 #include <limits>
  12 #include <memory>
  13 #include <queue>
  14 #include <stdint.h>
  15 #include <stdio.h>
  16 #include <stdlib.h>
  17 #include <string.h>
  18 #include <string>
  19 #include <string_view>
  20 #include <unistd.h>
  21 #include <unordered_map>
  22 #include <utility>
  23 #include <vector>
  24 #include <zstd.h>
  25
  26 using namespace std;
  27 using namespace std::chrono;
  28
  29 #define dprintf(...)
  30 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
  31
  32 #include "turbopfor.h"
  33
  34 const char *dbpath = "/var/lib/mlocate/plocate.db";
  35 bool only_count = false;
  36 bool print_nul = false;
  37 int64_t limit_matches = numeric_limits<int64_t>::max();
  38
  39 class Serializer {
  40 public:
  41         bool ready_to_print(int seq) { return next_seq == seq; }
  42         void print_delayed(int seq, const vector<string> msg);
  43         void release_current();
  44
  45 private:
  46         int next_seq = 0;
  47         struct Element {
  48                 int seq;
  49                 vector<string> msg;
  50
  51                 bool operator<(const Element &other) const
  52                 {
  53                         return seq > other.seq;
  54                 }
  55         };
  56         priority_queue<Element> pending;
  57 };
  58
  59 void Serializer::print_delayed(int seq, const vector<string> msg)
  60 {
  61         pending.push(Element{ seq, move(msg) });
  62 }
  63
  64 void Serializer::release_current()
  65 {
  66         ++next_seq;
  67
  68         // See if any delayed prints can now be dealt with.
  69         while (!pending.empty() && pending.top().seq == next_seq) {
  70                 if (limit_matches-- <= 0)
  71                         return;
  72                 for (const string &msg : pending.top().msg) {
  73                         if (print_nul) {
  74                                 printf("%s%c", msg.c_str(), 0);
  75                         } else {
  76                                 printf("%s\n", msg.c_str());
  77                         }
  78                 }
  79                 pending.pop();
  80                 ++next_seq;
  81         }
  82 }
  83
  84 static inline uint32_t read_unigram(const string &s, size_t idx)
  85 {
  86         if (idx < s.size()) {
  87                 return (unsigned char)s[idx];
  88         } else {
  89                 return 0;
  90         }
  91 }
  92
  93 static inline uint32_t read_trigram(const string &s, size_t start)
  94 {
  95         return read_unigram(s, start) | (read_unigram(s, start + 1) << 8) |
  96                 (read_unigram(s, start + 2) << 16);
  97 }
  98
  99 bool has_access(const char *filename,
 100                 unordered_map<string, bool> *access_rx_cache)
 101 {
 102         const char *end = strchr(filename + 1, '/');
 103         while (end != nullptr) {
 104                 string parent_path(filename, end);
 105                 auto it = access_rx_cache->find(parent_path);
 106                 bool ok;
 107                 if (it == access_rx_cache->end()) {
 108                         ok = access(parent_path.c_str(), R_OK | X_OK) == 0;
 109                         access_rx_cache->emplace(move(parent_path), ok);
 110                 } else {
 111                         ok = it->second;
 112                 }
 113                 if (!ok) {
 114                         return false;
 115                 }
 116                 end = strchr(end + 1, '/');
 117         }
 118
 119         return true;
 120 }
 121
 122 class Corpus {
 123 public:
 124         Corpus(int fd, IOUringEngine *engine);
 125         ~Corpus();
 126         void find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb);
 127         void get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const;
 128         size_t get_num_filename_blocks() const;
 129         off_t offset_for_block(uint32_t docid) const
 130         {
 131                 return hdr.filename_index_offset_bytes + docid * sizeof(uint64_t);
 132         }
 133
 134 public:
 135         const int fd;
 136         IOUringEngine *const engine;
 137
 138         Header hdr;
 139 };
 140
 141 Corpus::Corpus(int fd, IOUringEngine *engine)
 142         : fd(fd), engine(engine)
 143 {
 144         // Enable to test cold-cache behavior (except for access()).
 145         if (false) {
 146                 off_t len = lseek(fd, 0, SEEK_END);
 147                 if (len == -1) {
 148                         perror("lseek");
 149                         exit(1);
 150                 }
 151                 posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED);
 152         }
 153
 154         complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
 155         if (memcmp(hdr.magic, "\0plocate", 8) != 0) {
 156                 fprintf(stderr, "plocate.db is corrupt or an old version; please rebuild it.\n");
 157                 exit(1);
 158         }
 159         if (hdr.version != 0) {
 160                 fprintf(stderr, "plocate.db has version %u, expected 0; please rebuild it.\n", hdr.version);
 161                 exit(1);
 162         }
 163 }
 164
 165 Corpus::~Corpus()
 166 {
 167         close(fd);
 168 }
 169
 170 void Corpus::find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb)
 171 {
 172         uint32_t bucket = hash_trigram(trgm, hdr.hashtable_size);
 173         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) {
 174                 const Trigram *trgmptr = reinterpret_cast<const Trigram *>(s.data());
 175                 for (unsigned i = 0; i < hdr.extra_ht_slots + 1; ++i) {
 176                         if (trgmptr[i].trgm == trgm) {
 177                                 cb(trgmptr + i, trgmptr[i + 1].offset - trgmptr[i].offset);
 178                                 return;
 179                         }
 180                 }
 181
 182                 // Not found.
 183                 cb(nullptr, 0);
 184         });
 185 }
 186
 187 void Corpus::get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const
 188 {
 189         // Read the file offset from this docid and the next one.
 190         // This is always allowed, since we have a sentinel block at the end.
 191         engine->submit_read(fd, sizeof(uint64_t) * 2, offset_for_block(docid), [this, cb{ move(cb) }](string_view s) {
 192                 const uint64_t *ptr = reinterpret_cast<const uint64_t *>(s.data());
 193                 off_t offset = ptr[0];
 194                 size_t len = ptr[1] - ptr[0];
 195                 engine->submit_read(fd, len, offset, cb);
 196         });
 197 }
 198
 199 size_t Corpus::get_num_filename_blocks() const
 200 {
 201         return hdr.num_docids;
 202 }
 203
 204 uint64_t scan_file_block(const vector<string> &needles, string_view compressed,
 205                          unordered_map<string, bool> *access_rx_cache, int seq,
 206                          Serializer *serializer)
 207 {
 208         uint64_t matched = 0;
 209
 210         unsigned long long uncompressed_len = ZSTD_getFrameContentSize(compressed.data(), compressed.size());
 211         if (uncompressed_len == ZSTD_CONTENTSIZE_UNKNOWN || uncompressed_len == ZSTD_CONTENTSIZE_ERROR) {
 212                 fprintf(stderr, "ZSTD_getFrameContentSize() failed\n");
 213                 exit(1);
 214         }
 215
 216         string block;
 217         block.resize(uncompressed_len + 1);
 218
 219         size_t err = ZSTD_decompress(&block[0], block.size(), compressed.data(),
 220                                      compressed.size());
 221         if (ZSTD_isError(err)) {
 222                 fprintf(stderr, "ZSTD_decompress(): %s\n", ZSTD_getErrorName(err));
 223                 exit(1);
 224         }
 225         block[block.size() - 1] = '\0';
 226
 227         bool immediate_print = (serializer == nullptr || serializer->ready_to_print(seq));
 228         vector<string> delayed;
 229
 230         for (const char *filename = block.data();
 231              filename != block.data() + block.size();
 232              filename += strlen(filename) + 1) {
 233                 bool found = true;
 234                 for (const string &needle : needles) {
 235                         if (strstr(filename, needle.c_str()) == nullptr) {
 236                                 found = false;
 237                                 break;
 238                         }
 239                 }
 240                 if (found && has_access(filename, access_rx_cache)) {
 241                         if (limit_matches-- <= 0)
 242                                 break;
 243                         ++matched;
 244                         if (only_count)
 245                                 continue;
 246                         if (immediate_print) {
 247                                 if (print_nul) {
 248                                         printf("%s%c", filename, 0);
 249                                 } else {
 250                                         printf("%s\n", filename);
 251                                 }
 252                         } else {
 253                                 delayed.push_back(filename);
 254                         }
 255                 }
 256         }
 257         if (serializer != nullptr && !only_count) {
 258                 if (immediate_print) {
 259                         serializer->release_current();
 260                 } else {
 261                         serializer->print_delayed(seq, move(delayed));
 262                 }
 263         }
 264         return matched;
 265 }
 266
 267 size_t scan_docids(const vector<string> &needles, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
 268 {
 269         Serializer docids_in_order;
 270         unordered_map<string, bool> access_rx_cache;
 271         uint64_t matched = 0;
 272         for (size_t i = 0; i < docids.size(); ++i) {
 273                 uint32_t docid = docids[i];
 274                 corpus.get_compressed_filename_block(docid, [i, &matched, &needles, &access_rx_cache, &docids_in_order](string_view compressed) {
 275                         matched += scan_file_block(needles, compressed, &access_rx_cache, i, &docids_in_order);
 276                 });
 277         }
 278         engine->finish();
 279         return matched;
 280 }
 281
 282 // We do this sequentially, as it's faster than scattering
 283 // a lot of I/O through io_uring and hoping the kernel will
 284 // coalesce it plus readahead for us.
 285 uint64_t scan_all_docids(const vector<string> &needles, int fd, const Corpus &corpus, IOUringEngine *engine)
 286 {
 287         unordered_map<string, bool> access_rx_cache;
 288         uint32_t num_blocks = corpus.get_num_filename_blocks();
 289         unique_ptr<uint64_t[]> offsets(new uint64_t[num_blocks + 1]);
 290         complete_pread(fd, offsets.get(), (num_blocks + 1) * sizeof(uint64_t), corpus.offset_for_block(0));
 291         string compressed;
 292         uint64_t matched = 0;
 293         for (uint32_t io_docid = 0; io_docid < num_blocks; io_docid += 32) {
 294                 uint32_t last_docid = std::min(io_docid + 32, num_blocks);
 295                 size_t io_len = offsets[last_docid] - offsets[io_docid];
 296                 if (compressed.size() < io_len) {
 297                         compressed.resize(io_len);
 298                 }
 299                 complete_pread(fd, &compressed[0], io_len, offsets[io_docid]);
 300
 301                 for (uint32_t docid = io_docid; docid < last_docid; ++docid) {
 302                         size_t relative_offset = offsets[docid] - offsets[io_docid];
 303                         size_t len = offsets[docid + 1] - offsets[docid];
 304                         matched += scan_file_block(needles, { &compressed[relative_offset], len }, &access_rx_cache, 0, nullptr);
 305                         if (limit_matches <= 0)
 306                                 return matched;
 307                 }
 308         }
 309         return matched;
 310 }
 311
 312 void do_search_file(const vector<string> &needles, const char *filename)
 313 {
 314         int fd = open(filename, O_RDONLY);
 315         if (fd == -1) {
 316                 perror(filename);
 317                 exit(1);
 318         }
 319
 320         // Drop privileges.
 321         if (setgid(getgid()) != 0) {
 322                 perror("setgid");
 323                 exit(EXIT_FAILURE);
 324         }
 325
 326         steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
 327         if (access("/", R_OK | X_OK)) {
 328                 // We can't find anything, no need to bother...
 329                 return;
 330         }
 331
 332         IOUringEngine engine(/*slop_bytes=*/16);  // 16 slop bytes as described in turbopfor.h.
 333         Corpus corpus(fd, &engine);
 334         dprintf("Corpus init done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 335
 336         vector<pair<Trigram, size_t>> trigrams;
 337         uint64_t shortest_so_far = numeric_limits<uint32_t>::max();
 338         for (const string &needle : needles) {
 339                 if (needle.size() < 3)
 340                         continue;
 341                 for (size_t i = 0; i < needle.size() - 2; ++i) {
 342                         uint32_t trgm = read_trigram(needle, i);
 343                         corpus.find_trigram(trgm, [trgm, &trigrams, &shortest_so_far](const Trigram *trgmptr, size_t len) {
 344                                 if (trgmptr == nullptr) {
 345                                         dprintf("trigram '%c%c%c' isn't found, we abort the search\n",
 346                                                 trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff);
 347                                         if (only_count) {
 348                                                 printf("0\n");
 349                                         }
 350                                         exit(0);
 351                                 }
 352                                 if (trgmptr->num_docids > shortest_so_far * 100) {
 353                                         dprintf("not loading trigram '%c%c%c' with %u docids, it would be ignored later anyway\n",
 354                                                 trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff,
 355                                                 trgmptr->num_docids);
 356                                 } else {
 357                                         trigrams.emplace_back(*trgmptr, len);
 358                                         shortest_so_far = std::min<uint64_t>(shortest_so_far, trgmptr->num_docids);
 359                                 }
 360                         });
 361                 }
 362         }
 363         engine.finish();
 364         dprintf("Hashtable lookups done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 365
 366         if (trigrams.empty()) {
 367                 // Too short for trigram matching. Apply brute force.
 368                 // (We could have searched through all trigrams that matched
 369                 // the pattern and done a union of them, but that's a lot of
 370                 // work for fairly unclear gain.)
 371                 uint64_t matched = scan_all_docids(needles, fd, corpus, &engine);
 372                 if (only_count) {
 373                         printf("%zu\n", matched);
 374                 }
 375                 return;
 376         }
 377         sort(trigrams.begin(), trigrams.end());
 378         {
 379                 auto last = unique(trigrams.begin(), trigrams.end());
 380                 trigrams.erase(last, trigrams.end());
 381         }
 382         sort(trigrams.begin(), trigrams.end(),
 383              [&](const pair<Trigram, size_t> &a, const pair<Trigram, size_t> &b) {
 384                      return a.first.num_docids < b.first.num_docids;
 385              });
 386
 387         vector<uint32_t> in1, in2, out;
 388         bool done = false;
 389         for (auto [trgmptr, len] : trigrams) {
 390                 if (!in1.empty() && trgmptr.num_docids > in1.size() * 100) {
 391                         uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
 392                         dprintf("trigram '%c%c%c' (%zu bytes) has %u entries, ignoring the rest (will "
 393                                 "weed out false positives later)\n",
 394                                 trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff,
 395                                 len, trgmptr.num_docids);
 396                         break;
 397                 }
 398
 399                 // Only stay a certain amount ahead, so that we don't spend I/O
 400                 // on reading the latter, large posting lists. We are unlikely
 401                 // to need them anyway, even if they should come in first.
 402                 if (engine.get_waiting_reads() >= 5) {
 403                         engine.finish();
 404                         if (done)
 405                                 break;
 406                 }
 407                 engine.submit_read(fd, len, trgmptr.offset, [trgmptr{ trgmptr }, len{ len }, &done, &in1, &in2, &out](string_view s) {
 408                         if (done)
 409                                 return;
 410                         uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
 411                         size_t num = trgmptr.num_docids;
 412                         const unsigned char *pldata = reinterpret_cast<const unsigned char *>(s.data());
 413                         if (in1.empty()) {
 414                                 in1.resize(num + 128);
 415                                 decode_pfor_delta1_128(pldata, num, /*interleaved=*/true, &in1[0]);
 416                                 in1.resize(num);
 417                                 dprintf("trigram '%c%c%c' (%zu bytes) decoded to %zu entries\n", trgm & 0xff,
 418                                         (trgm >> 8) & 0xff, (trgm >> 16) & 0xff, len, num);
 419                         } else {
 420                                 if (in2.size() < num + 128) {
 421                                         in2.resize(num + 128);
 422                                 }
 423                                 decode_pfor_delta1_128(pldata, num, /*interleaved=*/true, &in2[0]);
 424
 425                                 out.clear();
 426                                 set_intersection(in1.begin(), in1.end(), in2.begin(), in2.begin() + num,
 427                                                  back_inserter(out));
 428                                 swap(in1, out);
 429                                 dprintf("trigram '%c%c%c' (%zu bytes) decoded to %zu entries, %zu left\n",
 430                                         trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff,
 431                                         len, num, in1.size());
 432                                 if (in1.empty()) {
 433                                         dprintf("no matches (intersection list is empty)\n");
 434                                         done = true;
 435                                 }
 436                         }
 437                 });
 438         }
 439         engine.finish();
 440         if (done) {
 441                 return;
 442         }
 443         dprintf("Intersection done after %.1f ms. Doing final verification and printing:\n",
 444                 1e3 * duration<float>(steady_clock::now() - start).count());
 445
 446         uint64_t matched = scan_docids(needles, in1, corpus, &engine);
 447         dprintf("Done in %.1f ms, found %zu matches.\n",
 448                 1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 449
 450         if (only_count) {
 451                 printf("%zu\n", matched);
 452         }
 453 }
 454
 455 void usage()
 456 {
 457         // The help text comes from mlocate.
 458         printf("Usage: plocate [OPTION]... PATTERN...\n");
 459         printf("\n");
 460         printf("  -c, --count            only print number of found entries\n");
 461         printf("  -d, --database DBPATH  use DBPATH instead of default database (which is\n");
 462         printf("                         %s)\n", dbpath);
 463         printf("  -h, --help             print this help\n");
 464         printf("  -l, --limit, -n LIMIT  limit output (or counting) to LIMIT entries\n");
 465         printf("  -0, --null             separate entries with NUL on output\n");
 466 }
 467
 468 int main(int argc, char **argv)
 469 {
 470         static const struct option long_options[] = {
 471                 { "help", no_argument, 0, 'h' },
 472                 { "count", no_argument, 0, 'c' },
 473                 { "database", required_argument, 0, 'd' },
 474                 { "limit", required_argument, 0, 'l' },
 475                 { nullptr, required_argument, 0, 'n' },
 476                 { "null", no_argument, 0, '0' },
 477                 { 0, 0, 0, 0 }
 478         };
 479
 480         for (;;) {
 481                 int option_index = 0;
 482                 int c = getopt_long(argc, argv, "cd:hl:n:0", long_options, &option_index);
 483                 if (c == -1) {
 484                         break;
 485                 }
 486                 switch (c) {
 487                 case 'c':
 488                         only_count = true;
 489                         break;
 490                 case 'd':
 491                         dbpath = strdup(optarg);
 492                         break;
 493                 case 'h':
 494                         usage();
 495                         exit(0);
 496                 case 'l':
 497                 case 'n':
 498                         limit_matches = atoll(optarg);
 499                         break;
 500                 case '0':
 501                         print_nul = true;
 502                         break;
 503                 default:
 504                         exit(1);
 505                 }
 506         }
 507
 508         vector<string> needles;
 509         for (int i = optind; i < argc; ++i) {
 510                 needles.push_back(argv[i]);
 511         }
 512         if (needles.empty()) {
 513                 fprintf(stderr, "plocate: no pattern to search for specified\n");
 514                 exit(0);
 515         }
 516         do_search_file(needles, dbpath);
 517 }