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 ZSTD_DDict *ddict = nullptr;
  53
  54 void apply_limit()
  55 {
  56         if (--limit_left > 0) {
  57                 return;
  58         }
  59         dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
  60                 1e3 * duration<float>(steady_clock::now() - start).count(), limit_matches);
  61         if (only_count) {
  62                 printf("%" PRId64 "\n", limit_matches);
  63         }
  64         exit(0);
  65 }
  66
  67 class Serializer {
  68 public:
  69         ~Serializer() { assert(limit_left <= 0 || pending.empty()); }
  70         void print(uint64_t seq, uint64_t skip, const string msg);
  71
  72 private:
  73         uint64_t next_seq = 0;
  74         struct Element {
  75                 uint64_t seq, skip;
  76                 string msg;
  77
  78                 bool operator<(const Element &other) const
  79                 {
  80                         return seq > other.seq;
  81                 }
  82         };
  83         priority_queue<Element> pending;
  84 };
  85
  86 void Serializer::print(uint64_t seq, uint64_t skip, const string msg)
  87 {
  88         if (only_count) {
  89                 if (!msg.empty()) {
  90                         apply_limit();
  91                 }
  92                 return;
  93         }
  94
  95         if (next_seq != seq) {
  96                 pending.push(Element{ seq, skip, move(msg) });
  97                 return;
  98         }
  99
 100         if (!msg.empty()) {
 101                 if (print_nul) {
 102                         printf("%s%c", msg.c_str(), 0);
 103                 } else {
 104                         printf("%s\n", msg.c_str());
 105                 }
 106                 apply_limit();
 107         }
 108         next_seq += skip;
 109
 110         // See if any delayed prints can now be dealt with.
 111         while (!pending.empty() && pending.top().seq == next_seq) {
 112                 if (!pending.top().msg.empty()) {
 113                         if (print_nul) {
 114                                 printf("%s%c", pending.top().msg.c_str(), 0);
 115                         } else {
 116                                 printf("%s\n", pending.top().msg.c_str());
 117                         }
 118                         apply_limit();
 119                 }
 120                 next_seq += pending.top().skip;
 121                 pending.pop();
 122         }
 123 }
 124
 125 struct Needle {
 126         enum { STRSTR,
 127                REGEX,
 128                GLOB } type;
 129         string str;  // Filled in no matter what.
 130         regex_t re;  // For REGEX.
 131 };
 132
 133 bool matches(const Needle &needle, const char *haystack)
 134 {
 135         if (needle.type == Needle::STRSTR) {
 136                 return strstr(haystack, needle.str.c_str()) != nullptr;
 137         } else if (needle.type == Needle::GLOB) {
 138                 int flags = ignore_case ? FNM_CASEFOLD : 0;
 139                 return fnmatch(needle.str.c_str(), haystack, flags) == 0;
 140         } else {
 141                 assert(needle.type == Needle::REGEX);
 142                 return regexec(&needle.re, haystack, /*nmatch=*/0, /*pmatch=*/nullptr, /*flags=*/0) == 0;
 143         }
 144 }
 145
 146 class AccessRXCache {
 147 public:
 148         AccessRXCache(IOUringEngine *engine)
 149                 : engine(engine) {}
 150         void check_access(const char *filename, bool allow_async, function<void(bool)> cb);
 151
 152 private:
 153         unordered_map<string, bool> cache;
 154         struct PendingStat {
 155                 string filename;
 156                 function<void(bool)> cb;
 157         };
 158         map<string, vector<PendingStat>> pending_stats;
 159         IOUringEngine *engine;
 160 };
 161
 162 void AccessRXCache::check_access(const char *filename, bool allow_async, function<void(bool)> cb)
 163 {
 164         if (engine == nullptr || !engine->get_supports_stat()) {
 165                 allow_async = false;
 166         }
 167
 168         for (const char *end = strchr(filename + 1, '/'); end != nullptr; end = strchr(end + 1, '/')) {
 169                 string parent_path(filename, end - filename);  // string_view from C++20.
 170                 auto cache_it = cache.find(parent_path);
 171                 if (cache_it != cache.end()) {
 172                         // Found in the cache.
 173                         if (!cache_it->second) {
 174                                 cb(false);
 175                                 return;
 176                         }
 177                         continue;
 178                 }
 179
 180                 if (!allow_async) {
 181                         bool ok = access(parent_path.c_str(), R_OK | X_OK) == 0;
 182                         cache.emplace(parent_path, ok);
 183                         if (!ok) {
 184                                 cb(false);
 185                                 return;
 186                         }
 187                         continue;
 188                 }
 189
 190                 // We want to call access(), but it could block on I/O. io_uring doesn't support
 191                 // access(), but we can do a dummy asynchonous statx() to populate the kernel's cache,
 192                 // which nearly always makes the next access() instantaneous.
 193
 194                 // See if there's already a pending stat that matches this,
 195                 // or is a subdirectory.
 196                 auto it = pending_stats.lower_bound(parent_path);
 197                 if (it != pending_stats.end() && it->first.size() >= parent_path.size() &&
 198                     it->first.compare(0, parent_path.size(), parent_path) == 0) {
 199                         it->second.emplace_back(PendingStat{ filename, move(cb) });
 200                 } else {
 201                         it = pending_stats.emplace(filename, vector<PendingStat>{}).first;
 202                         engine->submit_stat(filename, [this, it, filename{ strdup(filename) }, cb{ move(cb) }] {
 203                                 // The stat returned, so now do the actual access() calls.
 204                                 // All of them should be in cache, so don't fire off new statx()
 205                                 // calls during that check.
 206                                 check_access(filename, /*allow_async=*/false, move(cb));
 207                                 free(filename);
 208
 209                                 // Call all others that waited for the same stat() to finish.
 210                                 // They may fire off new stat() calls if needed.
 211                                 vector<PendingStat> pending = move(it->second);
 212                                 pending_stats.erase(it);
 213                                 for (PendingStat &ps : pending) {
 214                                         check_access(ps.filename.c_str(), /*allow_async=*/true, move(ps.cb));
 215                                 }
 216                         });
 217                 }
 218                 return;  // The rest will happen in async context.
 219         }
 220
 221         // Passed all checks.
 222         cb(true);
 223 }
 224
 225 class Corpus {
 226 public:
 227         Corpus(int fd, IOUringEngine *engine);
 228         ~Corpus();
 229         void find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb);
 230         void get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const;
 231         size_t get_num_filename_blocks() const;
 232         off_t offset_for_block(uint32_t docid) const
 233         {
 234                 return hdr.filename_index_offset_bytes + docid * sizeof(uint64_t);
 235         }
 236         const Header &get_hdr() const { return hdr; }
 237
 238 public:
 239         const int fd;
 240         IOUringEngine *const engine;
 241
 242         Header hdr;
 243 };
 244
 245 Corpus::Corpus(int fd, IOUringEngine *engine)
 246         : fd(fd), engine(engine)
 247 {
 248         // Enable to test cold-cache behavior (except for access()).
 249         if (false) {
 250                 off_t len = lseek(fd, 0, SEEK_END);
 251                 if (len == -1) {
 252                         perror("lseek");
 253                         exit(1);
 254                 }
 255                 posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED);
 256         }
 257
 258         complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
 259         if (memcmp(hdr.magic, "\0plocate", 8) != 0) {
 260                 fprintf(stderr, "plocate.db is corrupt or an old version; please rebuild it.\n");
 261                 exit(1);
 262         }
 263         if (hdr.version != 0 && hdr.version != 1) {
 264                 fprintf(stderr, "plocate.db has version %u, expected 0 or 1; please rebuild it.\n", hdr.version);
 265                 exit(1);
 266         }
 267         if (hdr.version == 0) {
 268                 // These will be junk data.
 269                 hdr.zstd_dictionary_offset_bytes = 0;
 270                 hdr.zstd_dictionary_length_bytes = 0;
 271         }
 272 }
 273
 274 Corpus::~Corpus()
 275 {
 276         close(fd);
 277 }
 278
 279 void Corpus::find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb)
 280 {
 281         uint32_t bucket = hash_trigram(trgm, hdr.hashtable_size);
 282         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) {
 283                 const Trigram *trgmptr = reinterpret_cast<const Trigram *>(s.data());
 284                 for (unsigned i = 0; i < hdr.extra_ht_slots + 1; ++i) {
 285                         if (trgmptr[i].trgm == trgm) {
 286                                 cb(trgmptr + i, trgmptr[i + 1].offset - trgmptr[i].offset);
 287                                 return;
 288                         }
 289                 }
 290
 291                 // Not found.
 292                 cb(nullptr, 0);
 293         });
 294 }
 295
 296 void Corpus::get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const
 297 {
 298         // Read the file offset from this docid and the next one.
 299         // This is always allowed, since we have a sentinel block at the end.
 300         engine->submit_read(fd, sizeof(uint64_t) * 2, offset_for_block(docid), [this, cb{ move(cb) }](string_view s) {
 301                 const uint64_t *ptr = reinterpret_cast<const uint64_t *>(s.data());
 302                 off_t offset = ptr[0];
 303                 size_t len = ptr[1] - ptr[0];
 304                 engine->submit_read(fd, len, offset, cb);
 305         });
 306 }
 307
 308 size_t Corpus::get_num_filename_blocks() const
 309 {
 310         return hdr.num_docids;
 311 }
 312
 313 void scan_file_block(const vector<Needle> &needles, string_view compressed,
 314                      AccessRXCache *access_rx_cache, uint64_t seq, Serializer *serializer,
 315                      uint64_t *matched)
 316 {
 317         unsigned long long uncompressed_len = ZSTD_getFrameContentSize(compressed.data(), compressed.size());
 318         if (uncompressed_len == ZSTD_CONTENTSIZE_UNKNOWN || uncompressed_len == ZSTD_CONTENTSIZE_ERROR) {
 319                 fprintf(stderr, "ZSTD_getFrameContentSize() failed\n");
 320                 exit(1);
 321         }
 322
 323         string block;
 324         block.resize(uncompressed_len + 1);
 325
 326         static ZSTD_DCtx *ctx = ZSTD_createDCtx();  // Reused across calls.
 327         size_t err;
 328
 329         if (ddict != nullptr) {
 330                 err = ZSTD_decompress_usingDDict(ctx, &block[0], block.size(), compressed.data(),
 331                                                  compressed.size(), ddict);
 332         } else {
 333                 err = ZSTD_decompressDCtx(ctx, &block[0], block.size(), compressed.data(),
 334                                           compressed.size());
 335         }
 336         if (ZSTD_isError(err)) {
 337                 fprintf(stderr, "ZSTD_decompress(): %s\n", ZSTD_getErrorName(err));
 338                 exit(1);
 339         }
 340         block[block.size() - 1] = '\0';
 341
 342         auto test_candidate = [&](const char *filename, uint64_t local_seq, uint64_t next_seq) {
 343                 access_rx_cache->check_access(filename, /*allow_async=*/true, [matched, serializer, local_seq, next_seq, filename{ strdup(filename) }](bool ok) {
 344                         if (ok) {
 345                                 ++*matched;
 346                                 serializer->print(local_seq, next_seq - local_seq, filename);
 347                         } else {
 348                                 serializer->print(local_seq, next_seq - local_seq, "");
 349                         }
 350                         free(filename);
 351                 });
 352         };
 353
 354         // We need to know the next sequence number before inserting into Serializer,
 355         // so always buffer one candidate.
 356         const char *pending_candidate = nullptr;
 357
 358         uint64_t local_seq = seq << 32;
 359         for (const char *filename = block.data();
 360              filename != block.data() + block.size();
 361              filename += strlen(filename) + 1) {
 362                 bool found = true;
 363                 for (const Needle &needle : needles) {
 364                         if (!matches(needle, filename)) {
 365                                 found = false;
 366                                 break;
 367                         }
 368                 }
 369                 if (found) {
 370                         if (pending_candidate != nullptr) {
 371                                 test_candidate(pending_candidate, local_seq, local_seq + 1);
 372                                 ++local_seq;
 373                         }
 374                         pending_candidate = filename;
 375                 }
 376         }
 377         if (pending_candidate == nullptr) {
 378                 serializer->print(seq << 32, 1ULL << 32, "");
 379         } else {
 380                 test_candidate(pending_candidate, local_seq, (seq + 1) << 32);
 381         }
 382 }
 383
 384 size_t scan_docids(const vector<Needle> &needles, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
 385 {
 386         Serializer docids_in_order;
 387         AccessRXCache access_rx_cache(engine);
 388         uint64_t matched = 0;
 389         for (size_t i = 0; i < docids.size(); ++i) {
 390                 uint32_t docid = docids[i];
 391                 corpus.get_compressed_filename_block(docid, [i, &matched, &needles, &access_rx_cache, &docids_in_order](string_view compressed) {
 392                         scan_file_block(needles, compressed, &access_rx_cache, i, &docids_in_order, &matched);
 393                 });
 394         }
 395         engine->finish();
 396         return matched;
 397 }
 398
 399 // We do this sequentially, as it's faster than scattering
 400 // a lot of I/O through io_uring and hoping the kernel will
 401 // coalesce it plus readahead for us.
 402 uint64_t scan_all_docids(const vector<Needle> &needles, int fd, const Corpus &corpus)
 403 {
 404         {
 405                 const Header &hdr = corpus.get_hdr();
 406                 if (hdr.zstd_dictionary_length_bytes > 0) {
 407                         string dictionary;
 408                         dictionary.resize(hdr.zstd_dictionary_length_bytes);
 409                         complete_pread(fd, &dictionary[0], hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes);
 410                         ddict = ZSTD_createDDict(dictionary.data(), dictionary.size());
 411                 }
 412         }
 413
 414         AccessRXCache access_rx_cache(nullptr);
 415         Serializer serializer;  // Mostly dummy; handles only the limit.
 416         uint32_t num_blocks = corpus.get_num_filename_blocks();
 417         unique_ptr<uint64_t[]> offsets(new uint64_t[num_blocks + 1]);
 418         complete_pread(fd, offsets.get(), (num_blocks + 1) * sizeof(uint64_t), corpus.offset_for_block(0));
 419         string compressed;
 420         uint64_t matched = 0;
 421         for (uint32_t io_docid = 0; io_docid < num_blocks; io_docid += 32) {
 422                 uint32_t last_docid = std::min(io_docid + 32, num_blocks);
 423                 size_t io_len = offsets[last_docid] - offsets[io_docid];
 424                 if (compressed.size() < io_len) {
 425                         compressed.resize(io_len);
 426                 }
 427                 complete_pread(fd, &compressed[0], io_len, offsets[io_docid]);
 428
 429                 for (uint32_t docid = io_docid; docid < last_docid; ++docid) {
 430                         size_t relative_offset = offsets[docid] - offsets[io_docid];
 431                         size_t len = offsets[docid + 1] - offsets[docid];
 432                         scan_file_block(needles, { &compressed[relative_offset], len }, &access_rx_cache, docid, &serializer, &matched);
 433                 }
 434         }
 435         return matched;
 436 }
 437
 438 // Takes the given posting list, unions it into the parts of the trigram disjunction
 439 // already read; if the list is complete, intersects with “cur_candidates”.
 440 //
 441 // Returns true if the search should be aborted (we are done).
 442 bool new_posting_list_read(TrigramDisjunction *td, vector<uint32_t> decoded, vector<uint32_t> *cur_candidates, vector<uint32_t> *tmp)
 443 {
 444         if (td->docids.empty()) {
 445                 td->docids = move(decoded);
 446         } else {
 447                 tmp->clear();
 448                 set_union(decoded.begin(), decoded.end(), td->docids.begin(), td->docids.end(), back_inserter(*tmp));
 449                 swap(*tmp, td->docids);
 450         }
 451         if (--td->remaining_trigrams_to_read > 0) {
 452                 // Need to wait for more.
 453                 if (ignore_case) {
 454                         dprintf("  ... %u reads left in OR group %u (%zu docids in list)\n",
 455                                 td->remaining_trigrams_to_read, td->index, td->docids.size());
 456                 }
 457                 return false;
 458         }
 459         if (cur_candidates->empty()) {
 460                 if (ignore_case) {
 461                         dprintf("  ... all reads done for OR group %u (%zu docids)\n",
 462                                 td->index, td->docids.size());
 463                 }
 464                 *cur_candidates = move(td->docids);
 465         } else {
 466                 tmp->clear();
 467                 set_intersection(cur_candidates->begin(), cur_candidates->end(),
 468                                  td->docids.begin(), td->docids.end(),
 469                                  back_inserter(*tmp));
 470                 swap(*cur_candidates, *tmp);
 471                 if (ignore_case) {
 472                         if (cur_candidates->empty()) {
 473                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (none left, search is done)\n",
 474                                         td->index, td->docids.size());
 475                                 return true;
 476                         } else {
 477                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (%zu left)\n",
 478                                         td->index, td->docids.size(), cur_candidates->size());
 479                         }
 480                 }
 481         }
 482         return false;
 483 }
 484
 485 void do_search_file(const vector<Needle> &needles, const char *filename)
 486 {
 487         int fd = open(filename, O_RDONLY);
 488         if (fd == -1) {
 489                 perror(filename);
 490                 exit(1);
 491         }
 492
 493         // Drop privileges.
 494         if (setgid(getgid()) != 0) {
 495                 perror("setgid");
 496                 exit(EXIT_FAILURE);
 497         }
 498
 499         start = steady_clock::now();
 500         if (access("/", R_OK | X_OK)) {
 501                 // We can't find anything, no need to bother...
 502                 return;
 503         }
 504
 505         IOUringEngine engine(/*slop_bytes=*/16);  // 16 slop bytes as described in turbopfor.h.
 506         Corpus corpus(fd, &engine);
 507         dprintf("Corpus init done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 508
 509         vector<TrigramDisjunction> trigram_groups;
 510         if (patterns_are_regex) {
 511                 // We could parse the regex to find trigrams that have to be there
 512                 // (there are actually known algorithms to deal with disjunctions
 513                 // and such, too), but for now, we just go brute force.
 514                 // Using locate with regexes is pretty niche.
 515         } else {
 516                 for (const Needle &needle : needles) {
 517                         parse_trigrams(needle.str, ignore_case, &trigram_groups);
 518                 }
 519         }
 520
 521         unique_sort(
 522                 &trigram_groups,
 523                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives < b.trigram_alternatives; },
 524                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives == b.trigram_alternatives; });
 525
 526         // Give them names for debugging.
 527         unsigned td_index = 0;
 528         for (TrigramDisjunction &td : trigram_groups) {
 529                 td.index = td_index++;
 530         }
 531
 532         // Collect which trigrams we need to look up in the hash table.
 533         unordered_map<uint32_t, vector<TrigramDisjunction *>> trigrams_to_lookup;
 534         for (TrigramDisjunction &td : trigram_groups) {
 535                 for (uint32_t trgm : td.trigram_alternatives) {
 536                         trigrams_to_lookup[trgm].push_back(&td);
 537                 }
 538         }
 539         if (trigrams_to_lookup.empty()) {
 540                 // Too short for trigram matching. Apply brute force.
 541                 // (We could have searched through all trigrams that matched
 542                 // the pattern and done a union of them, but that's a lot of
 543                 // work for fairly unclear gain.)
 544                 uint64_t matched = scan_all_docids(needles, fd, corpus);
 545                 if (only_count) {
 546                         printf("%" PRId64 "\n", matched);
 547                 }
 548                 return;
 549         }
 550
 551         // Sneak in fetching the dictionary, if present. It's not necessarily clear
 552         // exactly where it would be cheapest to get it, but it needs to be present
 553         // before we can decode any of the posting lists. Most likely, it's
 554         // in the same filesystem block as the header anyway, so it should be
 555         // present in the cache.
 556         {
 557                 const Header &hdr = corpus.get_hdr();
 558                 if (hdr.zstd_dictionary_length_bytes > 0) {
 559                         engine.submit_read(fd, hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes, [](string_view s) {
 560                                 ddict = ZSTD_createDDict(s.data(), s.size());
 561                                 dprintf("Dictionary initialized after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 562                         });
 563                 }
 564         }
 565
 566         // Look them all up on disk.
 567         for (auto &[trgm, trigram_groups] : trigrams_to_lookup) {
 568                 corpus.find_trigram(trgm, [trgm{ trgm }, trigram_groups{ &trigram_groups }](const Trigram *trgmptr, size_t len) {
 569                         if (trgmptr == nullptr) {
 570                                 dprintf("trigram %s isn't found\n", print_trigram(trgm).c_str());
 571                                 for (TrigramDisjunction *td : *trigram_groups) {
 572                                         --td->remaining_trigrams_to_read;
 573                                         if (td->remaining_trigrams_to_read == 0 && td->read_trigrams.empty()) {
 574                                                 dprintf("zero matches in %s, so we are done\n", print_td(*td).c_str());
 575                                                 if (only_count) {
 576                                                         printf("0\n");
 577                                                 }
 578                                                 exit(0);
 579                                         }
 580                                 }
 581                                 return;
 582                         }
 583                         for (TrigramDisjunction *td : *trigram_groups) {
 584                                 --td->remaining_trigrams_to_read;
 585                                 td->max_num_docids += trgmptr->num_docids;
 586                                 td->read_trigrams.emplace_back(*trgmptr, len);
 587                         }
 588                 });
 589         }
 590         engine.finish();
 591         dprintf("Hashtable lookups done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 592
 593         for (TrigramDisjunction &td : trigram_groups) {
 594                 // Reset for reads.
 595                 td.remaining_trigrams_to_read = td.read_trigrams.size();
 596
 597                 if (ignore_case) {  // If case-sensitive, they'll all be pretty obvious single-entry groups.
 598                         dprintf("OR group %u (max_num_docids=%u): %s\n", td.index, td.max_num_docids, print_td(td).c_str());
 599                 }
 600         }
 601
 602         // TODO: For case-insensitive (ie. more than one alternative in each),
 603         // prioritize the ones with fewer seeks?
 604         sort(trigram_groups.begin(), trigram_groups.end(),
 605              [&](const TrigramDisjunction &a, const TrigramDisjunction &b) {
 606                      return a.max_num_docids < b.max_num_docids;
 607              });
 608
 609         unordered_map<uint32_t, vector<TrigramDisjunction *>> uses_trigram;
 610         for (TrigramDisjunction &td : trigram_groups) {
 611                 for (uint32_t trgm : td.trigram_alternatives) {
 612                         uses_trigram[trgm].push_back(&td);
 613                 }
 614         }
 615
 616         unordered_set<uint32_t> trigrams_submitted_read;
 617         vector<uint32_t> cur_candidates, tmp, decoded;
 618         bool done = false;
 619         for (TrigramDisjunction &td : trigram_groups) {
 620                 if (!cur_candidates.empty() && td.max_num_docids > cur_candidates.size() * 100) {
 621                         dprintf("%s has up to %u entries, ignoring the rest (will "
 622                                 "weed out false positives later)\n",
 623                                 print_td(td).c_str(), td.max_num_docids);
 624                         break;
 625                 }
 626
 627                 for (auto &[trgmptr, len] : td.read_trigrams) {
 628                         if (trigrams_submitted_read.count(trgmptr.trgm) != 0) {
 629                                 continue;
 630                         }
 631                         trigrams_submitted_read.insert(trgmptr.trgm);
 632                         // Only stay a certain amount ahead, so that we don't spend I/O
 633                         // on reading the latter, large posting lists. We are unlikely
 634                         // to need them anyway, even if they should come in first.
 635                         if (engine.get_waiting_reads() >= 5) {
 636                                 engine.finish();
 637                                 if (done)
 638                                         break;
 639                         }
 640                         engine.submit_read(fd, len, trgmptr.offset, [trgmptr{ trgmptr }, len{ len }, &done, &cur_candidates, &tmp, &decoded, &uses_trigram](string_view s) {
 641                                 if (done)
 642                                         return;
 643
 644                                 uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
 645                                 const unsigned char *pldata = reinterpret_cast<const unsigned char *>(s.data());
 646                                 size_t num = trgmptr.num_docids;
 647                                 decoded.resize(num);
 648                                 decode_pfor_delta1_128(pldata, num, /*interleaved=*/true, &decoded[0]);
 649
 650                                 assert(uses_trigram.count(trgm) != 0);
 651                                 bool was_empty = cur_candidates.empty();
 652                                 if (ignore_case) {
 653                                         dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 654                                 }
 655
 656                                 for (TrigramDisjunction *td : uses_trigram[trgm]) {
 657                                         done |= new_posting_list_read(td, decoded, &cur_candidates, &tmp);
 658                                         if (done)
 659                                                 break;
 660                                 }
 661                                 if (!ignore_case) {
 662                                         if (was_empty) {
 663                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 664                                         } else if (cur_candidates.empty()) {
 665                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (none left, search is done)\n", print_trigram(trgm).c_str(), len, num);
 666                                         } else {
 667                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (%zu left)\n", print_trigram(trgm).c_str(), len, num, cur_candidates.size());
 668                                         }
 669                                 }
 670                         });
 671                 }
 672         }
 673         engine.finish();
 674         if (done) {
 675                 return;
 676         }
 677         dprintf("Intersection done after %.1f ms. Doing final verification and printing:\n",
 678                 1e3 * duration<float>(steady_clock::now() - start).count());
 679
 680         uint64_t matched = scan_docids(needles, cur_candidates, corpus, &engine);
 681         dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
 682                 1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 683
 684         if (only_count) {
 685                 printf("%" PRId64 "\n", matched);
 686         }
 687 }
 688
 689 string unescape_glob_to_plain_string(const string &needle)
 690 {
 691         string unescaped;
 692         for (size_t i = 0; i < needle.size(); i += read_unigram(needle, i).second) {
 693                 uint32_t ch = read_unigram(needle, i).first;
 694                 assert(ch != WILDCARD_UNIGRAM);
 695                 if (ch == PREMATURE_END_UNIGRAM) {
 696                         fprintf(stderr, "Pattern '%s' ended prematurely\n", needle.c_str());
 697                         exit(1);
 698                 }
 699                 unescaped.push_back(ch);
 700         }
 701         return unescaped;
 702 }
 703
 704 regex_t compile_regex(const string &needle)
 705 {
 706         regex_t re;
 707         int flags = REG_NOSUB;
 708         if (ignore_case) {
 709                 flags |= REG_ICASE;
 710         }
 711         if (use_extended_regex) {
 712                 flags |= REG_EXTENDED;
 713         }
 714         int err = regcomp(&re, needle.c_str(), flags);
 715         if (err != 0) {
 716                 char errbuf[256];
 717                 regerror(err, &re, errbuf, sizeof(errbuf));
 718                 fprintf(stderr, "Error when compiling regex '%s': %s\n", needle.c_str(), errbuf);
 719                 exit(1);
 720         }
 721         return re;
 722 }
 723
 724 void usage()
 725 {
 726         printf(
 727                 "Usage: plocate [OPTION]... PATTERN...\n"
 728                 "\n"
 729                 "  -c, --count            print number of matches instead of the matches\n"
 730                 "  -d, --database DBPATH  search for files in DBPATH\n"
 731                 "                         (default is " DEFAULT_DBPATH ")\n"
 732                 "  -i, --ignore-case      search case-insensitively\n"
 733                 "  -l, --limit LIMIT      stop after LIMIT matches\n"
 734                 "  -0, --null             delimit matches by NUL instead of newline\n"
 735                 "  -r, --regexp           interpret patterns as basic regexps (slow)\n"
 736                 "      --regex            interpret patterns as extended regexps (slow)\n"
 737                 "      --help             print this help\n"
 738                 "      --version          print version information\n");
 739 }
 740
 741 void version()
 742 {
 743         printf("plocate %s\n", PLOCATE_VERSION);
 744         printf("Copyright 2020 Steinar H. Gunderson\n");
 745         printf("License GPLv2+: GNU GPL version 2 or later <https://gnu.org/licenses/gpl.html>.\n");
 746         printf("This is free software: you are free to change and redistribute it.\n");
 747         printf("There is NO WARRANTY, to the extent permitted by law.\n");
 748         exit(0);
 749 }
 750
 751 int main(int argc, char **argv)
 752 {
 753         constexpr int EXTENDED_REGEX = 1000;
 754         static const struct option long_options[] = {
 755                 { "help", no_argument, 0, 'h' },
 756                 { "count", no_argument, 0, 'c' },
 757                 { "database", required_argument, 0, 'd' },
 758                 { "ignore-case", no_argument, 0, 'i' },
 759                 { "limit", required_argument, 0, 'l' },
 760                 { "null", no_argument, 0, '0' },
 761                 { "version", no_argument, 0, 'V' },
 762                 { "regexp", no_argument, 0, 'r' },
 763                 { "regex", no_argument, 0, EXTENDED_REGEX },
 764                 { "debug", no_argument, 0, 'D' },  // Not documented.
 765                 { 0, 0, 0, 0 }
 766         };
 767
 768         setlocale(LC_ALL, "");
 769         for (;;) {
 770                 int option_index = 0;
 771                 int c = getopt_long(argc, argv, "cd:hil:n:0VD", long_options, &option_index);
 772                 if (c == -1) {
 773                         break;
 774                 }
 775                 switch (c) {
 776                 case 'c':
 777                         only_count = true;
 778                         break;
 779                 case 'd':
 780                         dbpath = strdup(optarg);
 781                         break;
 782                 case 'h':
 783                         usage();
 784                         exit(0);
 785                 case 'i':
 786                         ignore_case = true;
 787                         break;
 788                 case 'l':
 789                 case 'n':
 790                         limit_matches = limit_left = atoll(optarg);
 791                         if (limit_matches <= 0) {
 792                                 fprintf(stderr, "Error: limit must be a strictly positive number.\n");
 793                                 exit(1);
 794                         }
 795                         break;
 796                 case '0':
 797                         print_nul = true;
 798                         break;
 799                 case 'r':
 800                         patterns_are_regex = true;
 801                         break;
 802                 case EXTENDED_REGEX:
 803                         patterns_are_regex = true;
 804                         use_extended_regex = true;
 805                         break;
 806                 case 'D':
 807                         use_debug = true;
 808                         break;
 809                 case 'V':
 810                         version();
 811                         break;
 812                 default:
 813                         exit(1);
 814                 }
 815         }
 816
 817         if (use_debug) {
 818                 // Debug information would leak information about which files exist,
 819                 // so drop setgid before we open the file; one would either need to run
 820                 // as root, or use a locally-built file.
 821                 if (setgid(getgid()) != 0) {
 822                         perror("setgid");
 823                         exit(EXIT_FAILURE);
 824                 }
 825         }
 826
 827         vector<Needle> needles;
 828         for (int i = optind; i < argc; ++i) {
 829                 Needle needle;
 830                 needle.str = argv[i];
 831
 832                 // See if there are any wildcard characters, which indicates we should treat it
 833                 // as an (anchored) glob.
 834                 bool any_wildcard = false;
 835                 for (size_t i = 0; i < needle.str.size(); i += read_unigram(needle.str, i).second) {
 836                         if (read_unigram(needle.str, i).first == WILDCARD_UNIGRAM) {
 837                                 any_wildcard = true;
 838                                 break;
 839                         }
 840                 }
 841
 842                 if (patterns_are_regex) {
 843                         needle.type = Needle::REGEX;
 844                         needle.re = compile_regex(needle.str);
 845                 } else if (any_wildcard) {
 846                         needle.type = Needle::GLOB;
 847                 } else if (ignore_case) {
 848                         // strcasestr() doesn't handle locales correctly (even though LSB
 849                         // claims it should), but somehow, fnmatch() does, and it's about
 850                         // the same speed as using a regex.
 851                         needle.type = Needle::GLOB;
 852                         needle.str = "*" + needle.str + "*";
 853                 } else {
 854                         needle.type = Needle::STRSTR;
 855                         needle.str = unescape_glob_to_plain_string(needle.str);
 856                 }
 857                 needles.push_back(move(needle));
 858         }
 859         if (needles.empty()) {
 860                 fprintf(stderr, "plocate: no pattern to search for specified\n");
 861                 exit(0);
 862         }
 863         do_search_file(needles, dbpath);
 864 }