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

   1 #include "access_rx_cache.h"
   2 #include "db.h"
   3 #include "dprintf.h"
   4 #include "io_uring_engine.h"
   5 #include "needle.h"
   6 #include "parse_trigrams.h"
   7 #include "serializer.h"
   8 #include "turbopfor.h"
   9 #include "unique_sort.h"
  10
  11 #include <algorithm>
  12 #include <assert.h>
  13 #include <atomic>
  14 #include <chrono>
  15 #include <condition_variable>
  16 #include <deque>
  17 #include <fcntl.h>
  18 #include <functional>
  19 #include <getopt.h>
  20 #include <inttypes.h>
  21 #include <iterator>
  22 #include <limits>
  23 #include <locale.h>
  24 #include <memory>
  25 #include <mutex>
  26 #include <regex.h>
  27 #include <stdint.h>
  28 #include <stdio.h>
  29 #include <stdlib.h>
  30 #include <string.h>
  31 #include <string>
  32 #include <string_view>
  33 #include <thread>
  34 #include <tuple>
  35 #include <unistd.h>
  36 #include <unordered_map>
  37 #include <unordered_set>
  38 #include <utility>
  39 #include <vector>
  40 #include <zstd.h>
  41
  42 using namespace std;
  43 using namespace std::chrono;
  44
  45 #define DEFAULT_DBPATH "/var/lib/mlocate/plocate.db"
  46
  47 const char *dbpath = DEFAULT_DBPATH;
  48 bool ignore_case = false;
  49 bool only_count = false;
  50 bool print_nul = false;
  51 bool use_debug = false;
  52 bool flush_cache = false;
  53 bool patterns_are_regex = false;
  54 bool use_extended_regex = false;
  55 bool match_basename = false;
  56 int64_t limit_matches = numeric_limits<int64_t>::max();
  57 int64_t limit_left = numeric_limits<int64_t>::max();
  58
  59 steady_clock::time_point start;
  60 ZSTD_DDict *ddict = nullptr;
  61
  62 class Corpus {
  63 public:
  64         Corpus(int fd, IOUringEngine *engine);
  65         ~Corpus();
  66         void find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb);
  67         void get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const;
  68         size_t get_num_filename_blocks() const;
  69         off_t offset_for_block(uint32_t docid) const
  70         {
  71                 return hdr.filename_index_offset_bytes + docid * sizeof(uint64_t);
  72         }
  73         const Header &get_hdr() const { return hdr; }
  74
  75 public:
  76         const int fd;
  77         IOUringEngine *const engine;
  78
  79         Header hdr;
  80 };
  81
  82 Corpus::Corpus(int fd, IOUringEngine *engine)
  83         : fd(fd), engine(engine)
  84 {
  85         if (flush_cache) {
  86                 off_t len = lseek(fd, 0, SEEK_END);
  87                 if (len == -1) {
  88                         perror("lseek");
  89                         exit(1);
  90                 }
  91                 posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED);
  92         }
  93
  94         complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
  95         if (memcmp(hdr.magic, "\0plocate", 8) != 0) {
  96                 fprintf(stderr, "plocate.db is corrupt or an old version; please rebuild it.\n");
  97                 exit(1);
  98         }
  99         if (hdr.version != 0 && hdr.version != 1) {
 100                 fprintf(stderr, "plocate.db has version %u, expected 0 or 1; please rebuild it.\n", hdr.version);
 101                 exit(1);
 102         }
 103         if (hdr.version == 0) {
 104                 // These will be junk data.
 105                 hdr.zstd_dictionary_offset_bytes = 0;
 106                 hdr.zstd_dictionary_length_bytes = 0;
 107         }
 108 }
 109
 110 Corpus::~Corpus()
 111 {
 112         close(fd);
 113 }
 114
 115 void Corpus::find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb)
 116 {
 117         uint32_t bucket = hash_trigram(trgm, hdr.hashtable_size);
 118         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) {
 119                 const Trigram *trgmptr = reinterpret_cast<const Trigram *>(s.data());
 120                 for (unsigned i = 0; i < hdr.extra_ht_slots + 1; ++i) {
 121                         if (trgmptr[i].trgm == trgm) {
 122                                 cb(trgmptr + i, trgmptr[i + 1].offset - trgmptr[i].offset);
 123                                 return;
 124                         }
 125                 }
 126
 127                 // Not found.
 128                 cb(nullptr, 0);
 129         });
 130 }
 131
 132 void Corpus::get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const
 133 {
 134         // Read the file offset from this docid and the next one.
 135         // This is always allowed, since we have a sentinel block at the end.
 136         engine->submit_read(fd, sizeof(uint64_t) * 2, offset_for_block(docid), [this, cb{ move(cb) }](string_view s) {
 137                 const uint64_t *ptr = reinterpret_cast<const uint64_t *>(s.data());
 138                 off_t offset = ptr[0];
 139                 size_t len = ptr[1] - ptr[0];
 140                 engine->submit_read(fd, len, offset, cb);
 141         });
 142 }
 143
 144 size_t Corpus::get_num_filename_blocks() const
 145 {
 146         return hdr.num_docids;
 147 }
 148
 149 void scan_file_block(const vector<Needle> &needles, string_view compressed,
 150                      AccessRXCache *access_rx_cache, uint64_t seq, ResultReceiver *serializer,
 151                      atomic<uint64_t> *matched)
 152 {
 153         unsigned long long uncompressed_len = ZSTD_getFrameContentSize(compressed.data(), compressed.size());
 154         if (uncompressed_len == ZSTD_CONTENTSIZE_UNKNOWN || uncompressed_len == ZSTD_CONTENTSIZE_ERROR) {
 155                 fprintf(stderr, "ZSTD_getFrameContentSize() failed\n");
 156                 exit(1);
 157         }
 158
 159         string block;
 160         block.resize(uncompressed_len + 1);
 161
 162         static thread_local ZSTD_DCtx *ctx = ZSTD_createDCtx();  // Reused across calls.
 163         size_t err;
 164
 165         if (ddict != nullptr) {
 166                 err = ZSTD_decompress_usingDDict(ctx, &block[0], block.size(), compressed.data(),
 167                                                  compressed.size(), ddict);
 168         } else {
 169                 err = ZSTD_decompressDCtx(ctx, &block[0], block.size(), compressed.data(),
 170                                           compressed.size());
 171         }
 172         if (ZSTD_isError(err)) {
 173                 fprintf(stderr, "ZSTD_decompress(): %s\n", ZSTD_getErrorName(err));
 174                 exit(1);
 175         }
 176         block[block.size() - 1] = '\0';
 177
 178         auto test_candidate = [&](const char *filename, uint64_t local_seq, uint64_t next_seq) {
 179                 access_rx_cache->check_access(filename, /*allow_async=*/true, [matched, serializer, local_seq, next_seq, filename{ strdup(filename) }](bool ok) {
 180                         if (ok) {
 181                                 ++*matched;
 182                                 serializer->print(local_seq, next_seq - local_seq, filename);
 183                         } else {
 184                                 serializer->print(local_seq, next_seq - local_seq, "");
 185                         }
 186                         free(filename);
 187                 });
 188         };
 189
 190         // We need to know the next sequence number before inserting into Serializer,
 191         // so always buffer one candidate.
 192         const char *pending_candidate = nullptr;
 193
 194         uint64_t local_seq = seq << 32;
 195         for (const char *filename = block.data();
 196              filename != block.data() + block.size();
 197              filename += strlen(filename) + 1) {
 198                 const char *haystack = filename;
 199                 if (match_basename) {
 200                         haystack = strrchr(filename, '/');
 201                         if (haystack == nullptr) {
 202                                 haystack = filename;
 203                         } else {
 204                                 ++haystack;
 205                         }
 206                 }
 207
 208                 bool found = true;
 209                 for (const Needle &needle : needles) {
 210                         if (!matches(needle, haystack)) {
 211                                 found = false;
 212                                 break;
 213                         }
 214                 }
 215                 if (found) {
 216                         if (pending_candidate != nullptr) {
 217                                 test_candidate(pending_candidate, local_seq, local_seq + 1);
 218                                 ++local_seq;
 219                         }
 220                         pending_candidate = filename;
 221                 }
 222         }
 223         if (pending_candidate == nullptr) {
 224                 serializer->print(seq << 32, 1ULL << 32, "");
 225         } else {
 226                 test_candidate(pending_candidate, local_seq, (seq + 1) << 32);
 227         }
 228 }
 229
 230 size_t scan_docids(const vector<Needle> &needles, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
 231 {
 232         Serializer docids_in_order;
 233         AccessRXCache access_rx_cache(engine);
 234         atomic<uint64_t> matched{ 0 };
 235         for (size_t i = 0; i < docids.size(); ++i) {
 236                 uint32_t docid = docids[i];
 237                 corpus.get_compressed_filename_block(docid, [i, &matched, &needles, &access_rx_cache, &docids_in_order](string_view compressed) {
 238                         scan_file_block(needles, compressed, &access_rx_cache, i, &docids_in_order, &matched);
 239                 });
 240         }
 241         engine->finish();
 242         return matched;
 243 }
 244
 245 struct WorkerThread {
 246         thread t;
 247
 248         // We use a result queue instead of synchronizing Serializer,
 249         // since a lock on it becomes a huge choke point if there are
 250         // lots of threads.
 251         mutex result_mu;
 252         struct Result {
 253                 uint64_t seq;
 254                 uint64_t skip;
 255                 string msg;
 256         };
 257         vector<Result> results;
 258 };
 259
 260 class WorkerThreadReceiver : public ResultReceiver {
 261 public:
 262         WorkerThreadReceiver(WorkerThread *wt)
 263                 : wt(wt) {}
 264
 265         void print(uint64_t seq, uint64_t skip, const string msg) override
 266         {
 267                 lock_guard<mutex> lock(wt->result_mu);
 268                 if (msg.empty() && !wt->results.empty() && wt->results.back().seq + wt->results.back().skip == seq) {
 269                         wt->results.back().skip += skip;
 270                 } else {
 271                         wt->results.emplace_back(WorkerThread::Result{ seq, skip, move(msg) });
 272                 }
 273         }
 274
 275 private:
 276         WorkerThread *wt;
 277 };
 278
 279 void deliver_results(WorkerThread *wt, Serializer *serializer)
 280 {
 281         vector<WorkerThread::Result> results;
 282         {
 283                 lock_guard<mutex> lock(wt->result_mu);
 284                 results = move(wt->results);
 285         }
 286         for (const WorkerThread::Result &result : results) {
 287                 serializer->print(result.seq, result.skip, move(result.msg));
 288         }
 289 }
 290
 291 // We do this sequentially, as it's faster than scattering
 292 // a lot of I/O through io_uring and hoping the kernel will
 293 // coalesce it plus readahead for us. Since we assume that
 294 // we will primarily be CPU-bound, we'll be firing up one
 295 // worker thread for each spare core (the last one will
 296 // only be doing I/O). access() is still synchronous.
 297 uint64_t scan_all_docids(const vector<Needle> &needles, int fd, const Corpus &corpus)
 298 {
 299         {
 300                 const Header &hdr = corpus.get_hdr();
 301                 if (hdr.zstd_dictionary_length_bytes > 0) {
 302                         string dictionary;
 303                         dictionary.resize(hdr.zstd_dictionary_length_bytes);
 304                         complete_pread(fd, &dictionary[0], hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes);
 305                         ddict = ZSTD_createDDict(dictionary.data(), dictionary.size());
 306                 }
 307         }
 308
 309         AccessRXCache access_rx_cache(nullptr);
 310         Serializer serializer;
 311         uint32_t num_blocks = corpus.get_num_filename_blocks();
 312         unique_ptr<uint64_t[]> offsets(new uint64_t[num_blocks + 1]);
 313         complete_pread(fd, offsets.get(), (num_blocks + 1) * sizeof(uint64_t), corpus.offset_for_block(0));
 314         atomic<uint64_t> matched{ 0 };
 315
 316         mutex mu;
 317         condition_variable queue_added, queue_removed;
 318         deque<tuple<int, int, string>> work_queue;  // Under mu.
 319         bool done = false;  // Under mu.
 320
 321         unsigned num_threads = max<int>(sysconf(_SC_NPROCESSORS_ONLN) - 1, 1);
 322         dprintf("Using %u worker threads for linear scan.\n", num_threads);
 323         unique_ptr<WorkerThread[]> threads(new WorkerThread[num_threads]);
 324         for (unsigned i = 0; i < num_threads; ++i) {
 325                 threads[i].t = thread([&threads, &mu, &queue_added, &queue_removed, &work_queue, &done, &offsets, &needles, &access_rx_cache, &matched, i] {
 326                         // regcomp() takes a lock on the regex, so each thread will need its own.
 327                         const vector<Needle> *use_needles = &needles;
 328                         vector<Needle> recompiled_needles;
 329                         if (i != 0 && patterns_are_regex) {
 330                                 recompiled_needles = needles;
 331                                 for (Needle &needle : recompiled_needles) {
 332                                         needle.re = compile_regex(needle.str);
 333                                 }
 334                                 use_needles = &recompiled_needles;
 335                         }
 336
 337                         WorkerThreadReceiver receiver(&threads[i]);
 338                         for (;;) {
 339                                 uint32_t io_docid, last_docid;
 340                                 string compressed;
 341
 342                                 {
 343                                         unique_lock<mutex> lock(mu);
 344                                         queue_added.wait(lock, [&work_queue, &done] { return !work_queue.empty() || done; });
 345                                         if (done && work_queue.empty()) {
 346                                                 return;
 347                                         }
 348                                         tie(io_docid, last_docid, compressed) = move(work_queue.front());
 349                                         work_queue.pop_front();
 350                                         queue_removed.notify_all();
 351                                 }
 352
 353                                 for (uint32_t docid = io_docid; docid < last_docid; ++docid) {
 354                                         size_t relative_offset = offsets[docid] - offsets[io_docid];
 355                                         size_t len = offsets[docid + 1] - offsets[docid];
 356                                         scan_file_block(*use_needles, { &compressed[relative_offset], len }, &access_rx_cache, docid, &receiver, &matched);
 357                                 }
 358                         }
 359                 });
 360         }
 361
 362         string compressed;
 363         for (uint32_t io_docid = 0; io_docid < num_blocks; io_docid += 32) {
 364                 uint32_t last_docid = std::min(io_docid + 32, num_blocks);
 365                 size_t io_len = offsets[last_docid] - offsets[io_docid];
 366                 if (compressed.size() < io_len) {
 367                         compressed.resize(io_len);
 368                 }
 369                 complete_pread(fd, &compressed[0], io_len, offsets[io_docid]);
 370
 371                 {
 372                         unique_lock<mutex> lock(mu);
 373                         queue_removed.wait(lock, [&work_queue] { return work_queue.size() < 256; });  // Allow ~2MB of data queued up.
 374                         work_queue.emplace_back(io_docid, last_docid, move(compressed));
 375                         queue_added.notify_one();  // Avoid the thundering herd.
 376                 }
 377
 378                 // Pick up some results, so that we are sure that we won't just overload.
 379                 // (Seemingly, going through all of these causes slowness with many threads,
 380                 // but taking only one is OK.)
 381                 unsigned i = io_docid / 32;
 382                 deliver_results(&threads[i % num_threads], &serializer);
 383         }
 384         {
 385                 lock_guard<mutex> lock(mu);
 386                 done = true;
 387                 queue_added.notify_all();
 388         }
 389         for (unsigned i = 0; i < num_threads; ++i) {
 390                 threads[i].t.join();
 391                 deliver_results(&threads[i], &serializer);
 392         }
 393         return matched;
 394 }
 395
 396 // Takes the given posting list, unions it into the parts of the trigram disjunction
 397 // already read; if the list is complete, intersects with “cur_candidates”.
 398 //
 399 // Returns true if the search should be aborted (we are done).
 400 bool new_posting_list_read(TrigramDisjunction *td, vector<uint32_t> decoded, vector<uint32_t> *cur_candidates, vector<uint32_t> *tmp)
 401 {
 402         if (td->docids.empty()) {
 403                 td->docids = move(decoded);
 404         } else {
 405                 tmp->clear();
 406                 set_union(decoded.begin(), decoded.end(), td->docids.begin(), td->docids.end(), back_inserter(*tmp));
 407                 swap(*tmp, td->docids);
 408         }
 409         if (--td->remaining_trigrams_to_read > 0) {
 410                 // Need to wait for more.
 411                 if (ignore_case) {
 412                         dprintf("  ... %u reads left in OR group %u (%zu docids in list)\n",
 413                                 td->remaining_trigrams_to_read, td->index, td->docids.size());
 414                 }
 415                 return false;
 416         }
 417         if (cur_candidates->empty()) {
 418                 if (ignore_case) {
 419                         dprintf("  ... all reads done for OR group %u (%zu docids)\n",
 420                                 td->index, td->docids.size());
 421                 }
 422                 *cur_candidates = move(td->docids);
 423         } else {
 424                 tmp->clear();
 425                 set_intersection(cur_candidates->begin(), cur_candidates->end(),
 426                                  td->docids.begin(), td->docids.end(),
 427                                  back_inserter(*tmp));
 428                 swap(*cur_candidates, *tmp);
 429                 if (ignore_case) {
 430                         if (cur_candidates->empty()) {
 431                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (none left, search is done)\n",
 432                                         td->index, td->docids.size());
 433                                 return true;
 434                         } else {
 435                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (%zu left)\n",
 436                                         td->index, td->docids.size(), cur_candidates->size());
 437                         }
 438                 }
 439         }
 440         return false;
 441 }
 442
 443 void do_search_file(const vector<Needle> &needles, const char *filename)
 444 {
 445         int fd = open(filename, O_RDONLY);
 446         if (fd == -1) {
 447                 perror(filename);
 448                 exit(1);
 449         }
 450
 451         // Drop privileges.
 452         if (setgid(getgid()) != 0) {
 453                 perror("setgid");
 454                 exit(EXIT_FAILURE);
 455         }
 456
 457         start = steady_clock::now();
 458         if (access("/", R_OK | X_OK)) {
 459                 // We can't find anything, no need to bother...
 460                 return;
 461         }
 462
 463         IOUringEngine engine(/*slop_bytes=*/16);  // 16 slop bytes as described in turbopfor.h.
 464         Corpus corpus(fd, &engine);
 465         dprintf("Corpus init done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 466
 467         vector<TrigramDisjunction> trigram_groups;
 468         if (patterns_are_regex) {
 469                 // We could parse the regex to find trigrams that have to be there
 470                 // (there are actually known algorithms to deal with disjunctions
 471                 // and such, too), but for now, we just go brute force.
 472                 // Using locate with regexes is pretty niche.
 473         } else {
 474                 for (const Needle &needle : needles) {
 475                         parse_trigrams(needle.str, ignore_case, &trigram_groups);
 476                 }
 477         }
 478
 479         unique_sort(
 480                 &trigram_groups,
 481                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives < b.trigram_alternatives; },
 482                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives == b.trigram_alternatives; });
 483
 484         // Give them names for debugging.
 485         unsigned td_index = 0;
 486         for (TrigramDisjunction &td : trigram_groups) {
 487                 td.index = td_index++;
 488         }
 489
 490         // Collect which trigrams we need to look up in the hash table.
 491         unordered_map<uint32_t, vector<TrigramDisjunction *>> trigrams_to_lookup;
 492         for (TrigramDisjunction &td : trigram_groups) {
 493                 for (uint32_t trgm : td.trigram_alternatives) {
 494                         trigrams_to_lookup[trgm].push_back(&td);
 495                 }
 496         }
 497         if (trigrams_to_lookup.empty()) {
 498                 // Too short for trigram matching. Apply brute force.
 499                 // (We could have searched through all trigrams that matched
 500                 // the pattern and done a union of them, but that's a lot of
 501                 // work for fairly unclear gain.)
 502                 uint64_t matched = scan_all_docids(needles, fd, corpus);
 503                 dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
 504                         1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 505                 if (only_count) {
 506                         printf("%" PRId64 "\n", matched);
 507                 }
 508                 return;
 509         }
 510
 511         // Sneak in fetching the dictionary, if present. It's not necessarily clear
 512         // exactly where it would be cheapest to get it, but it needs to be present
 513         // before we can decode any of the posting lists. Most likely, it's
 514         // in the same filesystem block as the header anyway, so it should be
 515         // present in the cache.
 516         {
 517                 const Header &hdr = corpus.get_hdr();
 518                 if (hdr.zstd_dictionary_length_bytes > 0) {
 519                         engine.submit_read(fd, hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes, [](string_view s) {
 520                                 ddict = ZSTD_createDDict(s.data(), s.size());
 521                                 dprintf("Dictionary initialized after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 522                         });
 523                 }
 524         }
 525
 526         // Look them all up on disk.
 527         for (auto &[trgm, trigram_groups] : trigrams_to_lookup) {
 528                 corpus.find_trigram(trgm, [trgm{ trgm }, trigram_groups{ &trigram_groups }](const Trigram *trgmptr, size_t len) {
 529                         if (trgmptr == nullptr) {
 530                                 dprintf("trigram %s isn't found\n", print_trigram(trgm).c_str());
 531                                 for (TrigramDisjunction *td : *trigram_groups) {
 532                                         --td->remaining_trigrams_to_read;
 533                                         if (td->remaining_trigrams_to_read == 0 && td->read_trigrams.empty()) {
 534                                                 dprintf("zero matches in %s, so we are done\n", print_td(*td).c_str());
 535                                                 if (only_count) {
 536                                                         printf("0\n");
 537                                                 }
 538                                                 exit(0);
 539                                         }
 540                                 }
 541                                 return;
 542                         }
 543                         for (TrigramDisjunction *td : *trigram_groups) {
 544                                 --td->remaining_trigrams_to_read;
 545                                 td->max_num_docids += trgmptr->num_docids;
 546                                 td->read_trigrams.emplace_back(*trgmptr, len);
 547                         }
 548                 });
 549         }
 550         engine.finish();
 551         dprintf("Hashtable lookups done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 552
 553         for (TrigramDisjunction &td : trigram_groups) {
 554                 // Reset for reads.
 555                 td.remaining_trigrams_to_read = td.read_trigrams.size();
 556
 557                 if (ignore_case) {  // If case-sensitive, they'll all be pretty obvious single-entry groups.
 558                         dprintf("OR group %u (max_num_docids=%u): %s\n", td.index, td.max_num_docids, print_td(td).c_str());
 559                 }
 560         }
 561
 562         // TODO: For case-insensitive (ie. more than one alternative in each),
 563         // prioritize the ones with fewer seeks?
 564         sort(trigram_groups.begin(), trigram_groups.end(),
 565              [&](const TrigramDisjunction &a, const TrigramDisjunction &b) {
 566                      return a.max_num_docids < b.max_num_docids;
 567              });
 568
 569         unordered_map<uint32_t, vector<TrigramDisjunction *>> uses_trigram;
 570         for (TrigramDisjunction &td : trigram_groups) {
 571                 for (uint32_t trgm : td.trigram_alternatives) {
 572                         uses_trigram[trgm].push_back(&td);
 573                 }
 574         }
 575
 576         unordered_set<uint32_t> trigrams_submitted_read;
 577         vector<uint32_t> cur_candidates, tmp, decoded;
 578         bool done = false;
 579         for (TrigramDisjunction &td : trigram_groups) {
 580                 if (!cur_candidates.empty() && td.max_num_docids > cur_candidates.size() * 100) {
 581                         dprintf("%s has up to %u entries, ignoring the rest (will "
 582                                 "weed out false positives later)\n",
 583                                 print_td(td).c_str(), td.max_num_docids);
 584                         break;
 585                 }
 586
 587                 for (auto &[trgmptr, len] : td.read_trigrams) {
 588                         if (trigrams_submitted_read.count(trgmptr.trgm) != 0) {
 589                                 continue;
 590                         }
 591                         trigrams_submitted_read.insert(trgmptr.trgm);
 592                         // Only stay a certain amount ahead, so that we don't spend I/O
 593                         // on reading the latter, large posting lists. We are unlikely
 594                         // to need them anyway, even if they should come in first.
 595                         if (engine.get_waiting_reads() >= 5) {
 596                                 engine.finish();
 597                                 if (done)
 598                                         break;
 599                         }
 600                         engine.submit_read(fd, len, trgmptr.offset, [trgmptr{ trgmptr }, len{ len }, &done, &cur_candidates, &tmp, &decoded, &uses_trigram](string_view s) {
 601                                 if (done)
 602                                         return;
 603
 604                                 uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
 605                                 const unsigned char *pldata = reinterpret_cast<const unsigned char *>(s.data());
 606                                 size_t num = trgmptr.num_docids;
 607                                 decoded.resize(num);
 608                                 decode_pfor_delta1_128(pldata, num, /*interleaved=*/true, &decoded[0]);
 609
 610                                 assert(uses_trigram.count(trgm) != 0);
 611                                 bool was_empty = cur_candidates.empty();
 612                                 if (ignore_case) {
 613                                         dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 614                                 }
 615
 616                                 for (TrigramDisjunction *td : uses_trigram[trgm]) {
 617                                         done |= new_posting_list_read(td, decoded, &cur_candidates, &tmp);
 618                                         if (done)
 619                                                 break;
 620                                 }
 621                                 if (!ignore_case) {
 622                                         if (was_empty) {
 623                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 624                                         } else if (cur_candidates.empty()) {
 625                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (none left, search is done)\n", print_trigram(trgm).c_str(), len, num);
 626                                         } else {
 627                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (%zu left)\n", print_trigram(trgm).c_str(), len, num, cur_candidates.size());
 628                                         }
 629                                 }
 630                         });
 631                 }
 632         }
 633         engine.finish();
 634         if (done) {
 635                 return;
 636         }
 637         dprintf("Intersection done after %.1f ms. Doing final verification and printing:\n",
 638                 1e3 * duration<float>(steady_clock::now() - start).count());
 639
 640         uint64_t matched = scan_docids(needles, cur_candidates, corpus, &engine);
 641         dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
 642                 1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 643
 644         if (only_count) {
 645                 printf("%" PRId64 "\n", matched);
 646         }
 647 }
 648
 649 void usage()
 650 {
 651         printf(
 652                 "Usage: plocate [OPTION]... PATTERN...\n"
 653                 "\n"
 654                 "  -b, --basename         search only the file name portion of path names\n"
 655                 "  -c, --count            print number of matches instead of the matches\n"
 656                 "  -d, --database DBPATH  search for files in DBPATH\n"
 657                 "                         (default is " DEFAULT_DBPATH ")\n"
 658                 "  -i, --ignore-case      search case-insensitively\n"
 659                 "  -l, --limit LIMIT      stop after LIMIT matches\n"
 660                 "  -0, --null             delimit matches by NUL instead of newline\n"
 661                 "  -r, --regexp           interpret patterns as basic regexps (slow)\n"
 662                 "      --regex            interpret patterns as extended regexps (slow)\n"
 663                 "  -w, --wholename        search the entire path name (default; see -b)\n"
 664                 "      --help             print this help\n"
 665                 "      --version          print version information\n");
 666 }
 667
 668 void version()
 669 {
 670         printf("plocate %s\n", PLOCATE_VERSION);
 671         printf("Copyright 2020 Steinar H. Gunderson\n");
 672         printf("License GPLv2+: GNU GPL version 2 or later <https://gnu.org/licenses/gpl.html>.\n");
 673         printf("This is free software: you are free to change and redistribute it.\n");
 674         printf("There is NO WARRANTY, to the extent permitted by law.\n");
 675         exit(0);
 676 }
 677
 678 int main(int argc, char **argv)
 679 {
 680         constexpr int EXTENDED_REGEX = 1000;
 681         constexpr int FLUSH_CACHE = 1001;
 682         static const struct option long_options[] = {
 683                 { "help", no_argument, 0, 'h' },
 684                 { "count", no_argument, 0, 'c' },
 685                 { "basename", no_argument, 0, 'b' },
 686                 { "database", required_argument, 0, 'd' },
 687                 { "ignore-case", no_argument, 0, 'i' },
 688                 { "limit", required_argument, 0, 'l' },
 689                 { "null", no_argument, 0, '0' },
 690                 { "version", no_argument, 0, 'V' },
 691                 { "regexp", no_argument, 0, 'r' },
 692                 { "regex", no_argument, 0, EXTENDED_REGEX },
 693                 { "wholename", no_argument, 0, 'w' },
 694                 { "debug", no_argument, 0, 'D' },  // Not documented.
 695                 // Enable to test cold-cache behavior (except for access()). Not documented.
 696                 { "flush-cache", no_argument, 0, FLUSH_CACHE },
 697                 { 0, 0, 0, 0 }
 698         };
 699
 700         setlocale(LC_ALL, "");
 701         for (;;) {
 702                 int option_index = 0;
 703                 int c = getopt_long(argc, argv, "bcd:hil:n:0wVD", long_options, &option_index);
 704                 if (c == -1) {
 705                         break;
 706                 }
 707                 switch (c) {
 708                 case 'b':
 709                         match_basename = true;
 710                         break;
 711                 case 'c':
 712                         only_count = true;
 713                         break;
 714                 case 'd':
 715                         dbpath = strdup(optarg);
 716                         break;
 717                 case 'h':
 718                         usage();
 719                         exit(0);
 720                 case 'i':
 721                         ignore_case = true;
 722                         break;
 723                 case 'l':
 724                 case 'n':
 725                         limit_matches = limit_left = atoll(optarg);
 726                         if (limit_matches <= 0) {
 727                                 fprintf(stderr, "Error: limit must be a strictly positive number.\n");
 728                                 exit(1);
 729                         }
 730                         break;
 731                 case '0':
 732                         print_nul = true;
 733                         break;
 734                 case 'r':
 735                         patterns_are_regex = true;
 736                         break;
 737                 case EXTENDED_REGEX:
 738                         patterns_are_regex = true;
 739                         use_extended_regex = true;
 740                         break;
 741                 case 'w':
 742                         match_basename = false;  // No-op unless -b is given first.
 743                         break;
 744                 case 'D':
 745                         use_debug = true;
 746                         break;
 747                 case FLUSH_CACHE:
 748                         flush_cache = true;
 749                         break;
 750                 case 'V':
 751                         version();
 752                         break;
 753                 default:
 754                         exit(1);
 755                 }
 756         }
 757
 758         if (use_debug || flush_cache) {
 759                 // Debug information would leak information about which files exist,
 760                 // so drop setgid before we open the file; one would either need to run
 761                 // as root, or use a locally-built file. Doing the same thing for
 762                 // flush_cache is mostly paranoia, in an attempt to prevent random users
 763                 // from making plocate slow for everyone else.
 764                 if (setgid(getgid()) != 0) {
 765                         perror("setgid");
 766                         exit(EXIT_FAILURE);
 767                 }
 768         }
 769
 770         vector<Needle> needles;
 771         for (int i = optind; i < argc; ++i) {
 772                 Needle needle;
 773                 needle.str = argv[i];
 774
 775                 // See if there are any wildcard characters, which indicates we should treat it
 776                 // as an (anchored) glob.
 777                 bool any_wildcard = false;
 778                 for (size_t i = 0; i < needle.str.size(); i += read_unigram(needle.str, i).second) {
 779                         if (read_unigram(needle.str, i).first == WILDCARD_UNIGRAM) {
 780                                 any_wildcard = true;
 781                                 break;
 782                         }
 783                 }
 784
 785                 if (patterns_are_regex) {
 786                         needle.type = Needle::REGEX;
 787                         needle.re = compile_regex(needle.str);
 788                 } else if (any_wildcard) {
 789                         needle.type = Needle::GLOB;
 790                 } else if (ignore_case) {
 791                         // strcasestr() doesn't handle locales correctly (even though LSB
 792                         // claims it should), but somehow, fnmatch() does, and it's about
 793                         // the same speed as using a regex.
 794                         needle.type = Needle::GLOB;
 795                         needle.str = "*" + needle.str + "*";
 796                 } else {
 797                         needle.type = Needle::STRSTR;
 798                         needle.str = unescape_glob_to_plain_string(needle.str);
 799                 }
 800                 needles.push_back(move(needle));
 801         }
 802         if (needles.empty()) {
 803                 fprintf(stderr, "plocate: no pattern to search for specified\n");
 804                 exit(0);
 805         }
 806         do_search_file(needles, dbpath);
 807 }