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

   1 #include "db.h"
   2 #include "dprintf.h"
   3 #include "io_uring_engine.h"
   4 #include "parse_trigrams.h"
   5 #include "turbopfor.h"
   6 #include "unique_sort.h"
   7
   8 #include <algorithm>
   9 #include <assert.h>
  10 #include <chrono>
  11 #include <fcntl.h>
  12 #include <fnmatch.h>
  13 #include <functional>
  14 #include <getopt.h>
  15 #include <inttypes.h>
  16 #include <iosfwd>
  17 #include <iterator>
  18 #include <limits>
  19 #include <map>
  20 #include <memory>
  21 #include <queue>
  22 #include <regex.h>
  23 #include <stdint.h>
  24 #include <stdio.h>
  25 #include <stdlib.h>
  26 #include <string.h>
  27 #include <string>
  28 #include <string_view>
  29 #include <unistd.h>
  30 #include <unordered_map>
  31 #include <unordered_set>
  32 #include <utility>
  33 #include <vector>
  34 #include <zstd.h>
  35
  36 using namespace std;
  37 using namespace std::chrono;
  38
  39 #define DEFAULT_DBPATH "/var/lib/mlocate/plocate.db"
  40
  41 const char *dbpath = DEFAULT_DBPATH;
  42 bool ignore_case = false;
  43 bool only_count = false;
  44 bool print_nul = false;
  45 bool use_debug = false;
  46 bool patterns_are_regex = false;
  47 bool use_extended_regex = false;
  48 int64_t limit_matches = numeric_limits<int64_t>::max();
  49 int64_t limit_left = numeric_limits<int64_t>::max();
  50
  51 steady_clock::time_point start;
  52
  53 void apply_limit()
  54 {
  55         if (--limit_left > 0) {
  56                 return;
  57         }
  58         dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
  59                 1e3 * duration<float>(steady_clock::now() - start).count(), limit_matches);
  60         if (only_count) {
  61                 printf("%" PRId64 "\n", limit_matches);
  62         }
  63         exit(0);
  64 }
  65
  66 class Serializer {
  67 public:
  68         ~Serializer() { assert(limit_left <= 0 || pending.empty()); }
  69         void print(uint64_t seq, uint64_t skip, const string msg);
  70
  71 private:
  72         uint64_t next_seq = 0;
  73         struct Element {
  74                 uint64_t seq, skip;
  75                 string msg;
  76
  77                 bool operator<(const Element &other) const
  78                 {
  79                         return seq > other.seq;
  80                 }
  81         };
  82         priority_queue<Element> pending;
  83 };
  84
  85 void Serializer::print(uint64_t seq, uint64_t skip, const string msg)
  86 {
  87         if (only_count) {
  88                 if (!msg.empty()) {
  89                         apply_limit();
  90                 }
  91                 return;
  92         }
  93
  94         if (next_seq != seq) {
  95                 pending.push(Element{ seq, skip, move(msg) });
  96                 return;
  97         }
  98
  99         if (!msg.empty()) {
 100                 if (print_nul) {
 101                         printf("%s%c", msg.c_str(), 0);
 102                 } else {
 103                         printf("%s\n", msg.c_str());
 104                 }
 105                 apply_limit();
 106         }
 107         next_seq += skip;
 108
 109         // See if any delayed prints can now be dealt with.
 110         while (!pending.empty() && pending.top().seq == next_seq) {
 111                 if (!pending.top().msg.empty()) {
 112                         if (print_nul) {
 113                                 printf("%s%c", pending.top().msg.c_str(), 0);
 114                         } else {
 115                                 printf("%s\n", pending.top().msg.c_str());
 116                         }
 117                         apply_limit();
 118                 }
 119                 next_seq += pending.top().skip;
 120                 pending.pop();
 121         }
 122 }
 123
 124 struct Needle {
 125         enum { STRSTR,
 126                REGEX,
 127                GLOB } type;
 128         string str;  // Filled in no matter what.
 129         regex_t re;  // For REGEX.
 130 };
 131
 132 bool matches(const Needle &needle, const char *haystack)
 133 {
 134         if (needle.type == Needle::STRSTR) {
 135                 return strstr(haystack, needle.str.c_str()) != nullptr;
 136         } else if (needle.type == Needle::GLOB) {
 137                 int flags = ignore_case ? FNM_CASEFOLD : 0;
 138                 return fnmatch(needle.str.c_str(), haystack, flags) == 0;
 139         } else {
 140                 assert(needle.type == Needle::REGEX);
 141                 return regexec(&needle.re, haystack, /*nmatch=*/0, /*pmatch=*/nullptr, /*flags=*/0) == 0;
 142         }
 143 }
 144
 145 class AccessRXCache {
 146 public:
 147         AccessRXCache(IOUringEngine *engine)
 148                 : engine(engine) {}
 149         void check_access(const char *filename, bool allow_async, function<void(bool)> cb);
 150
 151 private:
 152         unordered_map<string, bool> cache;
 153         struct PendingStat {
 154                 string filename;
 155                 function<void(bool)> cb;
 156         };
 157         map<string, vector<PendingStat>> pending_stats;
 158         IOUringEngine *engine;
 159 };
 160
 161 void AccessRXCache::check_access(const char *filename, bool allow_async, function<void(bool)> cb)
 162 {
 163         if (engine == nullptr || !engine->get_supports_stat()) {
 164                 allow_async = false;
 165         }
 166
 167         for (const char *end = strchr(filename + 1, '/'); end != nullptr; end = strchr(end + 1, '/')) {
 168                 string parent_path(filename, end - filename);  // string_view from C++20.
 169                 auto cache_it = cache.find(parent_path);
 170                 if (cache_it != cache.end()) {
 171                         // Found in the cache.
 172                         if (!cache_it->second) {
 173                                 cb(false);
 174                                 return;
 175                         }
 176                         continue;
 177                 }
 178
 179                 if (!allow_async) {
 180                         bool ok = access(parent_path.c_str(), R_OK | X_OK) == 0;
 181                         cache.emplace(parent_path, ok);
 182                         if (!ok) {
 183                                 cb(false);
 184                                 return;
 185                         }
 186                         continue;
 187                 }
 188
 189                 // We want to call access(), but it could block on I/O. io_uring doesn't support
 190                 // access(), but we can do a dummy asynchonous statx() to populate the kernel's cache,
 191                 // which nearly always makes the next access() instantaneous.
 192
 193                 // See if there's already a pending stat that matches this,
 194                 // or is a subdirectory.
 195                 auto it = pending_stats.lower_bound(parent_path);
 196                 if (it != pending_stats.end() && it->first.size() >= parent_path.size() &&
 197                     it->first.compare(0, parent_path.size(), parent_path) == 0) {
 198                         it->second.emplace_back(PendingStat{ filename, move(cb) });
 199                 } else {
 200                         it = pending_stats.emplace(filename, vector<PendingStat>{}).first;
 201                         engine->submit_stat(filename, [this, it, filename{ strdup(filename) }, cb{ move(cb) }] {
 202                                 // The stat returned, so now do the actual access() calls.
 203                                 // All of them should be in cache, so don't fire off new statx()
 204                                 // calls during that check.
 205                                 check_access(filename, /*allow_async=*/false, move(cb));
 206                                 free(filename);
 207
 208                                 // Call all others that waited for the same stat() to finish.
 209                                 // They may fire off new stat() calls if needed.
 210                                 vector<PendingStat> pending = move(it->second);
 211                                 pending_stats.erase(it);
 212                                 for (PendingStat &ps : pending) {
 213                                         check_access(ps.filename.c_str(), /*allow_async=*/true, move(ps.cb));
 214                                 }
 215                         });
 216                 }
 217                 return;  // The rest will happen in async context.
 218         }
 219
 220         // Passed all checks.
 221         cb(true);
 222 }
 223
 224 class Corpus {
 225 public:
 226         Corpus(int fd, IOUringEngine *engine);
 227         ~Corpus();
 228         void find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb);
 229         void get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const;
 230         size_t get_num_filename_blocks() const;
 231         off_t offset_for_block(uint32_t docid) const
 232         {
 233                 return hdr.filename_index_offset_bytes + docid * sizeof(uint64_t);
 234         }
 235
 236 public:
 237         const int fd;
 238         IOUringEngine *const engine;
 239
 240         Header hdr;
 241 };
 242
 243 Corpus::Corpus(int fd, IOUringEngine *engine)
 244         : fd(fd), engine(engine)
 245 {
 246         // Enable to test cold-cache behavior (except for access()).
 247         if (false) {
 248                 off_t len = lseek(fd, 0, SEEK_END);
 249                 if (len == -1) {
 250                         perror("lseek");
 251                         exit(1);
 252                 }
 253                 posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED);
 254         }
 255
 256         complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
 257         if (memcmp(hdr.magic, "\0plocate", 8) != 0) {
 258                 fprintf(stderr, "plocate.db is corrupt or an old version; please rebuild it.\n");
 259                 exit(1);
 260         }
 261         if (hdr.version != 0) {
 262                 fprintf(stderr, "plocate.db has version %u, expected 0; please rebuild it.\n", hdr.version);
 263                 exit(1);
 264         }
 265 }
 266
 267 Corpus::~Corpus()
 268 {
 269         close(fd);
 270 }
 271
 272 void Corpus::find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb)
 273 {
 274         uint32_t bucket = hash_trigram(trgm, hdr.hashtable_size);
 275         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) {
 276                 const Trigram *trgmptr = reinterpret_cast<const Trigram *>(s.data());
 277                 for (unsigned i = 0; i < hdr.extra_ht_slots + 1; ++i) {
 278                         if (trgmptr[i].trgm == trgm) {
 279                                 cb(trgmptr + i, trgmptr[i + 1].offset - trgmptr[i].offset);
 280                                 return;
 281                         }
 282                 }
 283
 284                 // Not found.
 285                 cb(nullptr, 0);
 286         });
 287 }
 288
 289 void Corpus::get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const
 290 {
 291         // Read the file offset from this docid and the next one.
 292         // This is always allowed, since we have a sentinel block at the end.
 293         engine->submit_read(fd, sizeof(uint64_t) * 2, offset_for_block(docid), [this, cb{ move(cb) }](string_view s) {
 294                 const uint64_t *ptr = reinterpret_cast<const uint64_t *>(s.data());
 295                 off_t offset = ptr[0];
 296                 size_t len = ptr[1] - ptr[0];
 297                 engine->submit_read(fd, len, offset, cb);
 298         });
 299 }
 300
 301 size_t Corpus::get_num_filename_blocks() const
 302 {
 303         return hdr.num_docids;
 304 }
 305
 306 void scan_file_block(const vector<Needle> &needles, string_view compressed,
 307                      AccessRXCache *access_rx_cache, uint64_t seq, Serializer *serializer,
 308                      uint64_t *matched)
 309 {
 310         unsigned long long uncompressed_len = ZSTD_getFrameContentSize(compressed.data(), compressed.size());
 311         if (uncompressed_len == ZSTD_CONTENTSIZE_UNKNOWN || uncompressed_len == ZSTD_CONTENTSIZE_ERROR) {
 312                 fprintf(stderr, "ZSTD_getFrameContentSize() failed\n");
 313                 exit(1);
 314         }
 315
 316         string block;
 317         block.resize(uncompressed_len + 1);
 318
 319         static ZSTD_DCtx *ctx = ZSTD_createDCtx();  // Reused across calls.
 320         size_t err = ZSTD_decompressDCtx(ctx, &block[0], block.size(), compressed.data(),
 321                                          compressed.size());
 322         if (ZSTD_isError(err)) {
 323                 fprintf(stderr, "ZSTD_decompress(): %s\n", ZSTD_getErrorName(err));
 324                 exit(1);
 325         }
 326         block[block.size() - 1] = '\0';
 327
 328         auto test_candidate = [&](const char *filename, uint64_t local_seq, uint64_t next_seq) {
 329                 access_rx_cache->check_access(filename, /*allow_async=*/true, [matched, serializer, local_seq, next_seq, filename{ strdup(filename) }](bool ok) {
 330                         if (ok) {
 331                                 ++*matched;
 332                                 serializer->print(local_seq, next_seq - local_seq, filename);
 333                         } else {
 334                                 serializer->print(local_seq, next_seq - local_seq, "");
 335                         }
 336                         free(filename);
 337                 });
 338         };
 339
 340         // We need to know the next sequence number before inserting into Serializer,
 341         // so always buffer one candidate.
 342         const char *pending_candidate = nullptr;
 343
 344         uint64_t local_seq = seq << 32;
 345         for (const char *filename = block.data();
 346              filename != block.data() + block.size();
 347              filename += strlen(filename) + 1) {
 348                 bool found = true;
 349                 for (const Needle &needle : needles) {
 350                         if (!matches(needle, filename)) {
 351                                 found = false;
 352                                 break;
 353                         }
 354                 }
 355                 if (found) {
 356                         if (pending_candidate != nullptr) {
 357                                 test_candidate(pending_candidate, local_seq, local_seq + 1);
 358                                 ++local_seq;
 359                         }
 360                         pending_candidate = filename;
 361                 }
 362         }
 363         if (pending_candidate == nullptr) {
 364                 serializer->print(seq << 32, 1ULL << 32, "");
 365         } else {
 366                 test_candidate(pending_candidate, local_seq, (seq + 1) << 32);
 367         }
 368 }
 369
 370 size_t scan_docids(const vector<Needle> &needles, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
 371 {
 372         Serializer docids_in_order;
 373         AccessRXCache access_rx_cache(engine);
 374         uint64_t matched = 0;
 375         for (size_t i = 0; i < docids.size(); ++i) {
 376                 uint32_t docid = docids[i];
 377                 corpus.get_compressed_filename_block(docid, [i, &matched, &needles, &access_rx_cache, &docids_in_order](string_view compressed) {
 378                         scan_file_block(needles, compressed, &access_rx_cache, i, &docids_in_order, &matched);
 379                 });
 380         }
 381         engine->finish();
 382         return matched;
 383 }
 384
 385 // We do this sequentially, as it's faster than scattering
 386 // a lot of I/O through io_uring and hoping the kernel will
 387 // coalesce it plus readahead for us.
 388 uint64_t scan_all_docids(const vector<Needle> &needles, int fd, const Corpus &corpus)
 389 {
 390         AccessRXCache access_rx_cache(nullptr);
 391         Serializer serializer;  // Mostly dummy; handles only the limit.
 392         uint32_t num_blocks = corpus.get_num_filename_blocks();
 393         unique_ptr<uint64_t[]> offsets(new uint64_t[num_blocks + 1]);
 394         complete_pread(fd, offsets.get(), (num_blocks + 1) * sizeof(uint64_t), corpus.offset_for_block(0));
 395         string compressed;
 396         uint64_t matched = 0;
 397         for (uint32_t io_docid = 0; io_docid < num_blocks; io_docid += 32) {
 398                 uint32_t last_docid = std::min(io_docid + 32, num_blocks);
 399                 size_t io_len = offsets[last_docid] - offsets[io_docid];
 400                 if (compressed.size() < io_len) {
 401                         compressed.resize(io_len);
 402                 }
 403                 complete_pread(fd, &compressed[0], io_len, offsets[io_docid]);
 404
 405                 for (uint32_t docid = io_docid; docid < last_docid; ++docid) {
 406                         size_t relative_offset = offsets[docid] - offsets[io_docid];
 407                         size_t len = offsets[docid + 1] - offsets[docid];
 408                         scan_file_block(needles, { &compressed[relative_offset], len }, &access_rx_cache, docid, &serializer, &matched);
 409                 }
 410         }
 411         return matched;
 412 }
 413
 414 // Takes the given posting list, unions it into the parts of the trigram disjunction
 415 // already read; if the list is complete, intersects with “cur_candidates”.
 416 //
 417 // Returns true if the search should be aborted (we are done).
 418 bool new_posting_list_read(TrigramDisjunction *td, vector<uint32_t> decoded, vector<uint32_t> *cur_candidates, vector<uint32_t> *tmp)
 419 {
 420         if (td->docids.empty()) {
 421                 td->docids = move(decoded);
 422         } else {
 423                 tmp->clear();
 424                 set_union(decoded.begin(), decoded.end(), td->docids.begin(), td->docids.end(), back_inserter(*tmp));
 425                 swap(*tmp, td->docids);
 426         }
 427         if (--td->remaining_trigrams_to_read > 0) {
 428                 // Need to wait for more.
 429                 if (ignore_case) {
 430                         dprintf("  ... %u reads left in OR group %u (%zu docids in list)\n",
 431                                 td->remaining_trigrams_to_read, td->index, td->docids.size());
 432                 }
 433                 return false;
 434         }
 435         if (cur_candidates->empty()) {
 436                 if (ignore_case) {
 437                         dprintf("  ... all reads done for OR group %u (%zu docids)\n",
 438                                 td->index, td->docids.size());
 439                 }
 440                 *cur_candidates = move(td->docids);
 441         } else {
 442                 tmp->clear();
 443                 set_intersection(cur_candidates->begin(), cur_candidates->end(),
 444                                  td->docids.begin(), td->docids.end(),
 445                                  back_inserter(*tmp));
 446                 swap(*cur_candidates, *tmp);
 447                 if (ignore_case) {
 448                         if (cur_candidates->empty()) {
 449                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (none left, search is done)\n",
 450                                         td->index, td->docids.size());
 451                                 return true;
 452                         } else {
 453                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (%zu left)\n",
 454                                         td->index, td->docids.size(), cur_candidates->size());
 455                         }
 456                 }
 457         }
 458         return false;
 459 }
 460
 461 void do_search_file(const vector<Needle> &needles, const char *filename)
 462 {
 463         int fd = open(filename, O_RDONLY);
 464         if (fd == -1) {
 465                 perror(filename);
 466                 exit(1);
 467         }
 468
 469         // Drop privileges.
 470         if (setgid(getgid()) != 0) {
 471                 perror("setgid");
 472                 exit(EXIT_FAILURE);
 473         }
 474
 475         start = steady_clock::now();
 476         if (access("/", R_OK | X_OK)) {
 477                 // We can't find anything, no need to bother...
 478                 return;
 479         }
 480
 481         IOUringEngine engine(/*slop_bytes=*/16);  // 16 slop bytes as described in turbopfor.h.
 482         Corpus corpus(fd, &engine);
 483         dprintf("Corpus init done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 484
 485         vector<TrigramDisjunction> trigram_groups;
 486         if (patterns_are_regex) {
 487                 // We could parse the regex to find trigrams that have to be there
 488                 // (there are actually known algorithms to deal with disjunctions
 489                 // and such, too), but for now, we just go brute force.
 490                 // Using locate with regexes is pretty niche.
 491         } else {
 492                 for (const Needle &needle : needles) {
 493                         parse_trigrams(needle.str, ignore_case, &trigram_groups);
 494                 }
 495         }
 496
 497         unique_sort(
 498                 &trigram_groups,
 499                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives < b.trigram_alternatives; },
 500                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives == b.trigram_alternatives; });
 501
 502         // Give them names for debugging.
 503         unsigned td_index = 0;
 504         for (TrigramDisjunction &td : trigram_groups) {
 505                 td.index = td_index++;
 506         }
 507
 508         // Collect which trigrams we need to look up in the hash table.
 509         unordered_map<uint32_t, vector<TrigramDisjunction *>> trigrams_to_lookup;
 510         for (TrigramDisjunction &td : trigram_groups) {
 511                 for (uint32_t trgm : td.trigram_alternatives) {
 512                         trigrams_to_lookup[trgm].push_back(&td);
 513                 }
 514         }
 515         if (trigrams_to_lookup.empty()) {
 516                 // Too short for trigram matching. Apply brute force.
 517                 // (We could have searched through all trigrams that matched
 518                 // the pattern and done a union of them, but that's a lot of
 519                 // work for fairly unclear gain.)
 520                 uint64_t matched = scan_all_docids(needles, fd, corpus);
 521                 if (only_count) {
 522                         printf("%" PRId64 "\n", matched);
 523                 }
 524                 return;
 525         }
 526
 527         // Look them all up on disk.
 528         for (auto &[trgm, trigram_groups] : trigrams_to_lookup) {
 529                 corpus.find_trigram(trgm, [trgm{ trgm }, trigram_groups{ &trigram_groups }](const Trigram *trgmptr, size_t len) {
 530                         if (trgmptr == nullptr) {
 531                                 dprintf("trigram %s isn't found\n", print_trigram(trgm).c_str());
 532                                 for (TrigramDisjunction *td : *trigram_groups) {
 533                                         --td->remaining_trigrams_to_read;
 534                                         if (td->remaining_trigrams_to_read == 0 && td->read_trigrams.empty()) {
 535                                                 dprintf("zero matches in %s, so we are done\n", print_td(*td).c_str());
 536                                                 if (only_count) {
 537                                                         printf("0\n");
 538                                                 }
 539                                                 exit(0);
 540                                         }
 541                                 }
 542                                 return;
 543                         }
 544                         for (TrigramDisjunction *td : *trigram_groups) {
 545                                 --td->remaining_trigrams_to_read;
 546                                 td->max_num_docids += trgmptr->num_docids;
 547                                 td->read_trigrams.emplace_back(*trgmptr, len);
 548                         }
 549                 });
 550         }
 551         engine.finish();
 552         dprintf("Hashtable lookups done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 553
 554         for (TrigramDisjunction &td : trigram_groups) {
 555                 // Reset for reads.
 556                 td.remaining_trigrams_to_read = td.read_trigrams.size();
 557
 558                 if (ignore_case) {  // If case-sensitive, they'll all be pretty obvious single-entry groups.
 559                         dprintf("OR group %u (max_num_docids=%u): %s\n", td.index, td.max_num_docids, print_td(td).c_str());
 560                 }
 561         }
 562
 563         // TODO: For case-insensitive (ie. more than one alternative in each),
 564         // prioritize the ones with fewer seeks?
 565         sort(trigram_groups.begin(), trigram_groups.end(),
 566              [&](const TrigramDisjunction &a, const TrigramDisjunction &b) {
 567                      return a.max_num_docids < b.max_num_docids;
 568              });
 569
 570         unordered_map<uint32_t, vector<TrigramDisjunction *>> uses_trigram;
 571         for (TrigramDisjunction &td : trigram_groups) {
 572                 for (uint32_t trgm : td.trigram_alternatives) {
 573                         uses_trigram[trgm].push_back(&td);
 574                 }
 575         }
 576
 577         unordered_set<uint32_t> trigrams_submitted_read;
 578         vector<uint32_t> cur_candidates, tmp, decoded;
 579         bool done = false;
 580         for (TrigramDisjunction &td : trigram_groups) {
 581                 if (!cur_candidates.empty() && td.max_num_docids > cur_candidates.size() * 100) {
 582                         dprintf("%s has up to %u entries, ignoring the rest (will "
 583                                 "weed out false positives later)\n",
 584                                 print_td(td).c_str(), td.max_num_docids);
 585                         break;
 586                 }
 587
 588                 for (auto &[trgmptr, len] : td.read_trigrams) {
 589                         if (trigrams_submitted_read.count(trgmptr.trgm) != 0) {
 590                                 continue;
 591                         }
 592                         trigrams_submitted_read.insert(trgmptr.trgm);
 593                         // Only stay a certain amount ahead, so that we don't spend I/O
 594                         // on reading the latter, large posting lists. We are unlikely
 595                         // to need them anyway, even if they should come in first.
 596                         if (engine.get_waiting_reads() >= 5) {
 597                                 engine.finish();
 598                                 if (done)
 599                                         break;
 600                         }
 601                         engine.submit_read(fd, len, trgmptr.offset, [trgmptr{ trgmptr }, len{ len }, &done, &cur_candidates, &tmp, &decoded, &uses_trigram](string_view s) {
 602                                 if (done)
 603                                         return;
 604
 605                                 uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
 606                                 const unsigned char *pldata = reinterpret_cast<const unsigned char *>(s.data());
 607                                 size_t num = trgmptr.num_docids;
 608                                 decoded.resize(num);
 609                                 decode_pfor_delta1_128(pldata, num, /*interleaved=*/true, &decoded[0]);
 610
 611                                 assert(uses_trigram.count(trgm) != 0);
 612                                 bool was_empty = cur_candidates.empty();
 613                                 if (ignore_case) {
 614                                         dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 615                                 }
 616
 617                                 for (TrigramDisjunction *td : uses_trigram[trgm]) {
 618                                         done |= new_posting_list_read(td, decoded, &cur_candidates, &tmp);
 619                                         if (done)
 620                                                 break;
 621                                 }
 622                                 if (!ignore_case) {
 623                                         if (was_empty) {
 624                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 625                                         } else if (cur_candidates.empty()) {
 626                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (none left, search is done)\n", print_trigram(trgm).c_str(), len, num);
 627                                         } else {
 628                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (%zu left)\n", print_trigram(trgm).c_str(), len, num, cur_candidates.size());
 629                                         }
 630                                 }
 631                         });
 632                 }
 633         }
 634         engine.finish();
 635         if (done) {
 636                 return;
 637         }
 638         dprintf("Intersection done after %.1f ms. Doing final verification and printing:\n",
 639                 1e3 * duration<float>(steady_clock::now() - start).count());
 640
 641         uint64_t matched = scan_docids(needles, cur_candidates, corpus, &engine);
 642         dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
 643                 1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 644
 645         if (only_count) {
 646                 printf("%" PRId64 "\n", matched);
 647         }
 648 }
 649
 650 string unescape_glob_to_plain_string(const string &needle)
 651 {
 652         string unescaped;
 653         for (size_t i = 0; i < needle.size(); i += read_unigram(needle, i).second) {
 654                 uint32_t ch = read_unigram(needle, i).first;
 655                 assert(ch != WILDCARD_UNIGRAM);
 656                 if (ch == PREMATURE_END_UNIGRAM) {
 657                         fprintf(stderr, "Pattern '%s' ended prematurely\n", needle.c_str());
 658                         exit(1);
 659                 }
 660                 unescaped.push_back(ch);
 661         }
 662         return unescaped;
 663 }
 664
 665 regex_t compile_regex(const string &needle)
 666 {
 667         regex_t re;
 668         int flags = REG_NOSUB;
 669         if (ignore_case) {
 670                 flags |= REG_ICASE;
 671         }
 672         if (use_extended_regex) {
 673                 flags |= REG_EXTENDED;
 674         }
 675         int err = regcomp(&re, needle.c_str(), flags);
 676         if (err != 0) {
 677                 char errbuf[256];
 678                 regerror(err, &re, errbuf, sizeof(errbuf));
 679                 fprintf(stderr, "Error when compiling regex '%s': %s\n", needle.c_str(), errbuf);
 680                 exit(1);
 681         }
 682         return re;
 683 }
 684
 685 void usage()
 686 {
 687         printf(
 688                 "Usage: plocate [OPTION]... PATTERN...\n"
 689                 "\n"
 690                 "  -c, --count            print number of matches instead of the matches\n"
 691                 "  -d, --database DBPATH  search for files in DBPATH\n"
 692                 "                         (default is " DEFAULT_DBPATH ")\n"
 693                 "  -i, --ignore-case      search case-insensitively\n"
 694                 "  -l, --limit LIMIT      stop after LIMIT matches\n"
 695                 "  -0, --null             delimit matches by NUL instead of newline\n"
 696                 "  -r, --regexp           interpret patterns as basic regexps (slow)\n"
 697                 "      --regex            interpret patterns as extended regexps (slow)\n"
 698                 "      --help             print this help\n"
 699                 "      --version          print version information\n");
 700 }
 701
 702 void version()
 703 {
 704         printf("plocate %s\n", PLOCATE_VERSION);
 705         printf("Copyright 2020 Steinar H. Gunderson\n");
 706         printf("License GPLv2+: GNU GPL version 2 or later <https://gnu.org/licenses/gpl.html>.\n");
 707         printf("This is free software: you are free to change and redistribute it.\n");
 708         printf("There is NO WARRANTY, to the extent permitted by law.\n");
 709         exit(0);
 710 }
 711
 712 int main(int argc, char **argv)
 713 {
 714         constexpr int EXTENDED_REGEX = 1000;
 715         static const struct option long_options[] = {
 716                 { "help", no_argument, 0, 'h' },
 717                 { "count", no_argument, 0, 'c' },
 718                 { "database", required_argument, 0, 'd' },
 719                 { "ignore-case", no_argument, 0, 'i' },
 720                 { "limit", required_argument, 0, 'l' },
 721                 { "null", no_argument, 0, '0' },
 722                 { "version", no_argument, 0, 'V' },
 723                 { "regexp", no_argument, 0, 'r' },
 724                 { "regex", no_argument, 0, EXTENDED_REGEX },
 725                 { "debug", no_argument, 0, 'D' },  // Not documented.
 726                 { 0, 0, 0, 0 }
 727         };
 728
 729         setlocale(LC_ALL, "");
 730         for (;;) {
 731                 int option_index = 0;
 732                 int c = getopt_long(argc, argv, "cd:hil:n:0VD", long_options, &option_index);
 733                 if (c == -1) {
 734                         break;
 735                 }
 736                 switch (c) {
 737                 case 'c':
 738                         only_count = true;
 739                         break;
 740                 case 'd':
 741                         dbpath = strdup(optarg);
 742                         break;
 743                 case 'h':
 744                         usage();
 745                         exit(0);
 746                 case 'i':
 747                         ignore_case = true;
 748                         break;
 749                 case 'l':
 750                 case 'n':
 751                         limit_matches = limit_left = atoll(optarg);
 752                         if (limit_matches <= 0) {
 753                                 fprintf(stderr, "Error: limit must be a strictly positive number.\n");
 754                                 exit(1);
 755                         }
 756                         break;
 757                 case '0':
 758                         print_nul = true;
 759                         break;
 760                 case 'r':
 761                         patterns_are_regex = true;
 762                         break;
 763                 case EXTENDED_REGEX:
 764                         patterns_are_regex = true;
 765                         use_extended_regex = true;
 766                         break;
 767                 case 'D':
 768                         use_debug = true;
 769                         break;
 770                 case 'V':
 771                         version();
 772                         break;
 773                 default:
 774                         exit(1);
 775                 }
 776         }
 777
 778         if (use_debug) {
 779                 // Debug information would leak information about which files exist,
 780                 // so drop setgid before we open the file; one would either need to run
 781                 // as root, or use a locally-built file.
 782                 if (setgid(getgid()) != 0) {
 783                         perror("setgid");
 784                         exit(EXIT_FAILURE);
 785                 }
 786         }
 787
 788         vector<Needle> needles;
 789         for (int i = optind; i < argc; ++i) {
 790                 Needle needle;
 791                 needle.str = argv[i];
 792
 793                 // See if there are any wildcard characters, which indicates we should treat it
 794                 // as an (anchored) glob.
 795                 bool any_wildcard = false;
 796                 for (size_t i = 0; i < needle.str.size(); i += read_unigram(needle.str, i).second) {
 797                         if (read_unigram(needle.str, i).first == WILDCARD_UNIGRAM) {
 798                                 any_wildcard = true;
 799                                 break;
 800                         }
 801                 }
 802
 803                 if (patterns_are_regex) {
 804                         needle.type = Needle::REGEX;
 805                         needle.re = compile_regex(needle.str);
 806                 } else if (any_wildcard) {
 807                         needle.type = Needle::GLOB;
 808                 } else if (ignore_case) {
 809                         // strcasestr() doesn't handle locales correctly (even though LSB
 810                         // claims it should), but somehow, fnmatch() does, and it's about
 811                         // the same speed as using a regex.
 812                         needle.type = Needle::GLOB;
 813                         needle.str = "*" + needle.str + "*";
 814                 } else {
 815                         needle.type = Needle::STRSTR;
 816                         needle.str = unescape_glob_to_plain_string(needle.str);
 817                 }
 818                 needles.push_back(move(needle));
 819         }
 820         if (needles.empty()) {
 821                 fprintf(stderr, "plocate: no pattern to search for specified\n");
 822                 exit(0);
 823         }
 824         do_search_file(needles, dbpath);
 825 }