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

   1 #include "access_rx_cache.h"
   2 #include "complete_pread.h"
   3 #include "db.h"
   4 #include "dprintf.h"
   5 #include "io_uring_engine.h"
   6 #include "needle.h"
   7 #include "parse_trigrams.h"
   8 #include "serializer.h"
   9 #include "turbopfor.h"
  10 #include "unique_sort.h"
  11
  12 #include <algorithm>
  13 #include <assert.h>
  14 #include <atomic>
  15 #include <chrono>
  16 #include <condition_variable>
  17 #include <deque>
  18 #include <fcntl.h>
  19 #include <functional>
  20 #include <getopt.h>
  21 #include <inttypes.h>
  22 #include <iterator>
  23 #include <limits>
  24 #include <locale.h>
  25 #include <memory>
  26 #include <mutex>
  27 #include <regex.h>
  28 #include <stdint.h>
  29 #include <stdio.h>
  30 #include <stdlib.h>
  31 #include <string.h>
  32 #include <string>
  33 #include <string_view>
  34 #include <thread>
  35 #include <tuple>
  36 #include <unistd.h>
  37 #include <unordered_map>
  38 #include <unordered_set>
  39 #include <utility>
  40 #include <vector>
  41 #include <zstd.h>
  42
  43 using namespace std;
  44 using namespace std::chrono;
  45
  46 const char *dbpath = DBFILE;
  47 bool ignore_case = false;
  48 bool only_count = false;
  49 bool print_nul = false;
  50 bool use_debug = false;
  51 bool flush_cache = false;
  52 bool patterns_are_regex = false;
  53 bool use_extended_regex = false;
  54 bool match_basename = false;
  55 int64_t limit_matches = numeric_limits<int64_t>::max();
  56 int64_t limit_left = numeric_limits<int64_t>::max();
  57 bool stdout_is_tty = false;
  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         if (hdr.max_version < 2) {
 109                 // This too. (We ignore the other max_version 2 fields.)
 110                 hdr.check_visibility = true;
 111         }
 112 }
 113
 114 Corpus::~Corpus()
 115 {
 116         close(fd);
 117 }
 118
 119 void Corpus::find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb)
 120 {
 121         uint32_t bucket = hash_trigram(trgm, hdr.hashtable_size);
 122         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) {
 123                 const Trigram *trgmptr = reinterpret_cast<const Trigram *>(s.data());
 124                 for (unsigned i = 0; i < hdr.extra_ht_slots + 1; ++i) {
 125                         if (trgmptr[i].trgm == trgm) {
 126                                 cb(trgmptr + i, trgmptr[i + 1].offset - trgmptr[i].offset);
 127                                 return;
 128                         }
 129                 }
 130
 131                 // Not found.
 132                 cb(nullptr, 0);
 133         });
 134 }
 135
 136 void Corpus::get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const
 137 {
 138         // Read the file offset from this docid and the next one.
 139         // This is always allowed, since we have a sentinel block at the end.
 140         engine->submit_read(fd, sizeof(uint64_t) * 2, offset_for_block(docid), [this, cb{ move(cb) }](string_view s) {
 141                 const uint64_t *ptr = reinterpret_cast<const uint64_t *>(s.data());
 142                 off_t offset = ptr[0];
 143                 size_t len = ptr[1] - ptr[0];
 144                 engine->submit_read(fd, len, offset, cb);
 145         });
 146 }
 147
 148 size_t Corpus::get_num_filename_blocks() const
 149 {
 150         return hdr.num_docids;
 151 }
 152
 153 void scan_file_block(const vector<Needle> &needles, string_view compressed,
 154                      AccessRXCache *access_rx_cache, uint64_t seq, ResultReceiver *serializer,
 155                      atomic<uint64_t> *matched)
 156 {
 157         unsigned long long uncompressed_len = ZSTD_getFrameContentSize(compressed.data(), compressed.size());
 158         if (uncompressed_len == ZSTD_CONTENTSIZE_UNKNOWN || uncompressed_len == ZSTD_CONTENTSIZE_ERROR) {
 159                 fprintf(stderr, "ZSTD_getFrameContentSize() failed\n");
 160                 exit(1);
 161         }
 162
 163         string block;
 164         block.resize(uncompressed_len + 1);
 165
 166         static thread_local ZSTD_DCtx *ctx = ZSTD_createDCtx();  // Reused across calls.
 167         size_t err;
 168
 169         if (ddict != nullptr) {
 170                 err = ZSTD_decompress_usingDDict(ctx, &block[0], block.size(), compressed.data(),
 171                                                  compressed.size(), ddict);
 172         } else {
 173                 err = ZSTD_decompressDCtx(ctx, &block[0], block.size(), compressed.data(),
 174                                           compressed.size());
 175         }
 176         if (ZSTD_isError(err)) {
 177                 fprintf(stderr, "ZSTD_decompress(): %s\n", ZSTD_getErrorName(err));
 178                 exit(1);
 179         }
 180         block[block.size() - 1] = '\0';
 181
 182         auto test_candidate = [&](const char *filename, uint64_t local_seq, uint64_t next_seq) {
 183                 access_rx_cache->check_access(filename, /*allow_async=*/true, [matched, serializer, local_seq, next_seq, filename{ strdup(filename) }](bool ok) {
 184                         if (ok) {
 185                                 ++*matched;
 186                                 serializer->print(local_seq, next_seq - local_seq, filename);
 187                         } else {
 188                                 serializer->print(local_seq, next_seq - local_seq, "");
 189                         }
 190                         free(filename);
 191                 });
 192         };
 193
 194         // We need to know the next sequence number before inserting into Serializer,
 195         // so always buffer one candidate.
 196         const char *pending_candidate = nullptr;
 197
 198         uint64_t local_seq = seq << 32;
 199         for (const char *filename = block.data();
 200              filename != block.data() + block.size();
 201              filename += strlen(filename) + 1) {
 202                 const char *haystack = filename;
 203                 if (match_basename) {
 204                         haystack = strrchr(filename, '/');
 205                         if (haystack == nullptr) {
 206                                 haystack = filename;
 207                         } else {
 208                                 ++haystack;
 209                         }
 210                 }
 211
 212                 bool found = true;
 213                 for (const Needle &needle : needles) {
 214                         if (!matches(needle, haystack)) {
 215                                 found = false;
 216                                 break;
 217                         }
 218                 }
 219                 if (found) {
 220                         if (pending_candidate != nullptr) {
 221                                 test_candidate(pending_candidate, local_seq, local_seq + 1);
 222                                 ++local_seq;
 223                         }
 224                         pending_candidate = filename;
 225                 }
 226         }
 227         if (pending_candidate == nullptr) {
 228                 serializer->print(seq << 32, 1ULL << 32, "");
 229         } else {
 230                 test_candidate(pending_candidate, local_seq, (seq + 1) << 32);
 231         }
 232 }
 233
 234 size_t scan_docids(const vector<Needle> &needles, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
 235 {
 236         Serializer docids_in_order;
 237         AccessRXCache access_rx_cache(engine, corpus.get_hdr().check_visibility);
 238         atomic<uint64_t> matched{ 0 };
 239         for (size_t i = 0; i < docids.size(); ++i) {
 240                 uint32_t docid = docids[i];
 241                 corpus.get_compressed_filename_block(docid, [i, &matched, &needles, &access_rx_cache, &docids_in_order](string_view compressed) {
 242                         scan_file_block(needles, compressed, &access_rx_cache, i, &docids_in_order, &matched);
 243                 });
 244         }
 245         engine->finish();
 246         return matched;
 247 }
 248
 249 struct WorkerThread {
 250         thread t;
 251
 252         // We use a result queue instead of synchronizing Serializer,
 253         // since a lock on it becomes a huge choke point if there are
 254         // lots of threads.
 255         mutex result_mu;
 256         struct Result {
 257                 uint64_t seq;
 258                 uint64_t skip;
 259                 string msg;
 260         };
 261         vector<Result> results;
 262 };
 263
 264 class WorkerThreadReceiver : public ResultReceiver {
 265 public:
 266         WorkerThreadReceiver(WorkerThread *wt)
 267                 : wt(wt) {}
 268
 269         void print(uint64_t seq, uint64_t skip, const string msg) override
 270         {
 271                 lock_guard<mutex> lock(wt->result_mu);
 272                 if (msg.empty() && !wt->results.empty() && wt->results.back().seq + wt->results.back().skip == seq) {
 273                         wt->results.back().skip += skip;
 274                 } else {
 275                         wt->results.emplace_back(WorkerThread::Result{ seq, skip, move(msg) });
 276                 }
 277         }
 278
 279 private:
 280         WorkerThread *wt;
 281 };
 282
 283 void deliver_results(WorkerThread *wt, Serializer *serializer)
 284 {
 285         vector<WorkerThread::Result> results;
 286         {
 287                 lock_guard<mutex> lock(wt->result_mu);
 288                 results = move(wt->results);
 289         }
 290         for (const WorkerThread::Result &result : results) {
 291                 serializer->print(result.seq, result.skip, move(result.msg));
 292         }
 293 }
 294
 295 // We do this sequentially, as it's faster than scattering
 296 // a lot of I/O through io_uring and hoping the kernel will
 297 // coalesce it plus readahead for us. Since we assume that
 298 // we will primarily be CPU-bound, we'll be firing up one
 299 // worker thread for each spare core (the last one will
 300 // only be doing I/O). access() is still synchronous.
 301 uint64_t scan_all_docids(const vector<Needle> &needles, int fd, const Corpus &corpus)
 302 {
 303         {
 304                 const Header &hdr = corpus.get_hdr();
 305                 if (hdr.zstd_dictionary_length_bytes > 0) {
 306                         string dictionary;
 307                         dictionary.resize(hdr.zstd_dictionary_length_bytes);
 308                         complete_pread(fd, &dictionary[0], hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes);
 309                         ddict = ZSTD_createDDict(dictionary.data(), dictionary.size());
 310                 }
 311         }
 312
 313         AccessRXCache access_rx_cache(nullptr, corpus.get_hdr().check_visibility);
 314         Serializer serializer;
 315         uint32_t num_blocks = corpus.get_num_filename_blocks();
 316         unique_ptr<uint64_t[]> offsets(new uint64_t[num_blocks + 1]);
 317         complete_pread(fd, offsets.get(), (num_blocks + 1) * sizeof(uint64_t), corpus.offset_for_block(0));
 318         atomic<uint64_t> matched{ 0 };
 319
 320         mutex mu;
 321         condition_variable queue_added, queue_removed;
 322         deque<tuple<int, int, string>> work_queue;  // Under mu.
 323         bool done = false;  // Under mu.
 324
 325         unsigned num_threads = max<int>(sysconf(_SC_NPROCESSORS_ONLN) - 1, 1);
 326         dprintf("Using %u worker threads for linear scan.\n", num_threads);
 327         unique_ptr<WorkerThread[]> threads(new WorkerThread[num_threads]);
 328         for (unsigned i = 0; i < num_threads; ++i) {
 329                 threads[i].t = thread([&threads, &mu, &queue_added, &queue_removed, &work_queue, &done, &offsets, &needles, &access_rx_cache, &matched, i] {
 330                         // regcomp() takes a lock on the regex, so each thread will need its own.
 331                         const vector<Needle> *use_needles = &needles;
 332                         vector<Needle> recompiled_needles;
 333                         if (i != 0 && patterns_are_regex) {
 334                                 recompiled_needles = needles;
 335                                 for (Needle &needle : recompiled_needles) {
 336                                         needle.re = compile_regex(needle.str);
 337                                 }
 338                                 use_needles = &recompiled_needles;
 339                         }
 340
 341                         WorkerThreadReceiver receiver(&threads[i]);
 342                         for (;;) {
 343                                 uint32_t io_docid, last_docid;
 344                                 string compressed;
 345
 346                                 {
 347                                         unique_lock<mutex> lock(mu);
 348                                         queue_added.wait(lock, [&work_queue, &done] { return !work_queue.empty() || done; });
 349                                         if (done && work_queue.empty()) {
 350                                                 return;
 351                                         }
 352                                         tie(io_docid, last_docid, compressed) = move(work_queue.front());
 353                                         work_queue.pop_front();
 354                                         queue_removed.notify_all();
 355                                 }
 356
 357                                 for (uint32_t docid = io_docid; docid < last_docid; ++docid) {
 358                                         size_t relative_offset = offsets[docid] - offsets[io_docid];
 359                                         size_t len = offsets[docid + 1] - offsets[docid];
 360                                         scan_file_block(*use_needles, { &compressed[relative_offset], len }, &access_rx_cache, docid, &receiver, &matched);
 361                                 }
 362                         }
 363                 });
 364         }
 365
 366         string compressed;
 367         for (uint32_t io_docid = 0; io_docid < num_blocks; io_docid += 32) {
 368                 uint32_t last_docid = std::min(io_docid + 32, num_blocks);
 369                 size_t io_len = offsets[last_docid] - offsets[io_docid];
 370                 if (compressed.size() < io_len) {
 371                         compressed.resize(io_len);
 372                 }
 373                 complete_pread(fd, &compressed[0], io_len, offsets[io_docid]);
 374
 375                 {
 376                         unique_lock<mutex> lock(mu);
 377                         queue_removed.wait(lock, [&work_queue] { return work_queue.size() < 256; });  // Allow ~2MB of data queued up.
 378                         work_queue.emplace_back(io_docid, last_docid, move(compressed));
 379                         queue_added.notify_one();  // Avoid the thundering herd.
 380                 }
 381
 382                 // Pick up some results, so that we are sure that we won't just overload.
 383                 // (Seemingly, going through all of these causes slowness with many threads,
 384                 // but taking only one is OK.)
 385                 unsigned i = io_docid / 32;
 386                 deliver_results(&threads[i % num_threads], &serializer);
 387         }
 388         {
 389                 lock_guard<mutex> lock(mu);
 390                 done = true;
 391                 queue_added.notify_all();
 392         }
 393         for (unsigned i = 0; i < num_threads; ++i) {
 394                 threads[i].t.join();
 395                 deliver_results(&threads[i], &serializer);
 396         }
 397         return matched;
 398 }
 399
 400 // Takes the given posting list, unions it into the parts of the trigram disjunction
 401 // already read; if the list is complete, intersects with “cur_candidates”.
 402 //
 403 // Returns true if the search should be aborted (we are done).
 404 bool new_posting_list_read(TrigramDisjunction *td, vector<uint32_t> decoded, vector<uint32_t> *cur_candidates, vector<uint32_t> *tmp)
 405 {
 406         if (td->docids.empty()) {
 407                 td->docids = move(decoded);
 408         } else {
 409                 tmp->clear();
 410                 set_union(decoded.begin(), decoded.end(), td->docids.begin(), td->docids.end(), back_inserter(*tmp));
 411                 swap(*tmp, td->docids);
 412         }
 413         if (--td->remaining_trigrams_to_read > 0) {
 414                 // Need to wait for more.
 415                 if (ignore_case) {
 416                         dprintf("  ... %u reads left in OR group %u (%zu docids in list)\n",
 417                                 td->remaining_trigrams_to_read, td->index, td->docids.size());
 418                 }
 419                 return false;
 420         }
 421         if (cur_candidates->empty()) {
 422                 if (ignore_case) {
 423                         dprintf("  ... all reads done for OR group %u (%zu docids)\n",
 424                                 td->index, td->docids.size());
 425                 }
 426                 *cur_candidates = move(td->docids);
 427         } else {
 428                 tmp->clear();
 429                 set_intersection(cur_candidates->begin(), cur_candidates->end(),
 430                                  td->docids.begin(), td->docids.end(),
 431                                  back_inserter(*tmp));
 432                 swap(*cur_candidates, *tmp);
 433                 if (ignore_case) {
 434                         if (cur_candidates->empty()) {
 435                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (none left, search is done)\n",
 436                                         td->index, td->docids.size());
 437                                 return true;
 438                         } else {
 439                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (%zu left)\n",
 440                                         td->index, td->docids.size(), cur_candidates->size());
 441                         }
 442                 }
 443         }
 444         return false;
 445 }
 446
 447 void do_search_file(const vector<Needle> &needles, const char *filename)
 448 {
 449         int fd = open(filename, O_RDONLY);
 450         if (fd == -1) {
 451                 perror(filename);
 452                 exit(1);
 453         }
 454
 455         // Drop privileges.
 456         if (setgid(getgid()) != 0) {
 457                 perror("setgid");
 458                 exit(EXIT_FAILURE);
 459         }
 460
 461         start = steady_clock::now();
 462         if (access("/", R_OK | X_OK)) {
 463                 // We can't find anything, no need to bother...
 464                 return;
 465         }
 466
 467         IOUringEngine engine(/*slop_bytes=*/16);  // 16 slop bytes as described in turbopfor.h.
 468         Corpus corpus(fd, &engine);
 469         dprintf("Corpus init done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 470
 471         vector<TrigramDisjunction> trigram_groups;
 472         if (patterns_are_regex) {
 473                 // We could parse the regex to find trigrams that have to be there
 474                 // (there are actually known algorithms to deal with disjunctions
 475                 // and such, too), but for now, we just go brute force.
 476                 // Using locate with regexes is pretty niche.
 477         } else {
 478                 for (const Needle &needle : needles) {
 479                         parse_trigrams(needle.str, ignore_case, &trigram_groups);
 480                 }
 481         }
 482
 483         unique_sort(
 484                 &trigram_groups,
 485                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives < b.trigram_alternatives; },
 486                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives == b.trigram_alternatives; });
 487
 488         // Give them names for debugging.
 489         unsigned td_index = 0;
 490         for (TrigramDisjunction &td : trigram_groups) {
 491                 td.index = td_index++;
 492         }
 493
 494         // Collect which trigrams we need to look up in the hash table.
 495         unordered_map<uint32_t, vector<TrigramDisjunction *>> trigrams_to_lookup;
 496         for (TrigramDisjunction &td : trigram_groups) {
 497                 for (uint32_t trgm : td.trigram_alternatives) {
 498                         trigrams_to_lookup[trgm].push_back(&td);
 499                 }
 500         }
 501         if (trigrams_to_lookup.empty()) {
 502                 // Too short for trigram matching. Apply brute force.
 503                 // (We could have searched through all trigrams that matched
 504                 // the pattern and done a union of them, but that's a lot of
 505                 // work for fairly unclear gain.)
 506                 uint64_t matched = scan_all_docids(needles, fd, corpus);
 507                 dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
 508                         1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 509                 if (only_count) {
 510                         printf("%" PRId64 "\n", matched);
 511                 }
 512                 return;
 513         }
 514
 515         // Sneak in fetching the dictionary, if present. It's not necessarily clear
 516         // exactly where it would be cheapest to get it, but it needs to be present
 517         // before we can decode any of the posting lists. Most likely, it's
 518         // in the same filesystem block as the header anyway, so it should be
 519         // present in the cache.
 520         {
 521                 const Header &hdr = corpus.get_hdr();
 522                 if (hdr.zstd_dictionary_length_bytes > 0) {
 523                         engine.submit_read(fd, hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes, [](string_view s) {
 524                                 ddict = ZSTD_createDDict(s.data(), s.size());
 525                                 dprintf("Dictionary initialized after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 526                         });
 527                 }
 528         }
 529
 530         // Look them all up on disk.
 531         for (auto &[trgm, trigram_groups] : trigrams_to_lookup) {
 532                 corpus.find_trigram(trgm, [trgm{ trgm }, trigram_groups{ &trigram_groups }](const Trigram *trgmptr, size_t len) {
 533                         if (trgmptr == nullptr) {
 534                                 dprintf("trigram %s isn't found\n", print_trigram(trgm).c_str());
 535                                 for (TrigramDisjunction *td : *trigram_groups) {
 536                                         --td->remaining_trigrams_to_read;
 537                                         if (td->remaining_trigrams_to_read == 0 && td->read_trigrams.empty()) {
 538                                                 dprintf("zero matches in %s, so we are done\n", print_td(*td).c_str());
 539                                                 if (only_count) {
 540                                                         printf("0\n");
 541                                                 }
 542                                                 exit(0);
 543                                         }
 544                                 }
 545                                 return;
 546                         }
 547                         for (TrigramDisjunction *td : *trigram_groups) {
 548                                 --td->remaining_trigrams_to_read;
 549                                 td->max_num_docids += trgmptr->num_docids;
 550                                 td->read_trigrams.emplace_back(*trgmptr, len);
 551                         }
 552                 });
 553         }
 554         engine.finish();
 555         dprintf("Hashtable lookups done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 556
 557         for (TrigramDisjunction &td : trigram_groups) {
 558                 // Reset for reads.
 559                 td.remaining_trigrams_to_read = td.read_trigrams.size();
 560
 561                 if (ignore_case) {  // If case-sensitive, they'll all be pretty obvious single-entry groups.
 562                         dprintf("OR group %u (max_num_docids=%u): %s\n", td.index, td.max_num_docids, print_td(td).c_str());
 563                 }
 564         }
 565
 566         // TODO: For case-insensitive (ie. more than one alternative in each),
 567         // prioritize the ones with fewer seeks?
 568         sort(trigram_groups.begin(), trigram_groups.end(),
 569              [&](const TrigramDisjunction &a, const TrigramDisjunction &b) {
 570                      return a.max_num_docids < b.max_num_docids;
 571              });
 572
 573         unordered_map<uint32_t, vector<TrigramDisjunction *>> uses_trigram;
 574         for (TrigramDisjunction &td : trigram_groups) {
 575                 for (uint32_t trgm : td.trigram_alternatives) {
 576                         uses_trigram[trgm].push_back(&td);
 577                 }
 578         }
 579
 580         unordered_set<uint32_t> trigrams_submitted_read;
 581         vector<uint32_t> cur_candidates, tmp, decoded;
 582         bool done = false;
 583         for (TrigramDisjunction &td : trigram_groups) {
 584                 if (!cur_candidates.empty() && td.max_num_docids > cur_candidates.size() * 100) {
 585                         dprintf("%s has up to %u entries, ignoring the rest (will "
 586                                 "weed out false positives later)\n",
 587                                 print_td(td).c_str(), td.max_num_docids);
 588                         break;
 589                 }
 590
 591                 for (auto &[trgmptr, len] : td.read_trigrams) {
 592                         if (trigrams_submitted_read.count(trgmptr.trgm) != 0) {
 593                                 continue;
 594                         }
 595                         trigrams_submitted_read.insert(trgmptr.trgm);
 596                         // Only stay a certain amount ahead, so that we don't spend I/O
 597                         // on reading the latter, large posting lists. We are unlikely
 598                         // to need them anyway, even if they should come in first.
 599                         if (engine.get_waiting_reads() >= 5) {
 600                                 engine.finish();
 601                                 if (done)
 602                                         break;
 603                         }
 604                         engine.submit_read(fd, len, trgmptr.offset, [trgmptr{ trgmptr }, len{ len }, &done, &cur_candidates, &tmp, &decoded, &uses_trigram](string_view s) {
 605                                 if (done)
 606                                         return;
 607
 608                                 uint32_t trgm = trgmptr.trgm;
 609                                 const unsigned char *pldata = reinterpret_cast<const unsigned char *>(s.data());
 610                                 size_t num = trgmptr.num_docids;
 611                                 decoded.resize(num);
 612                                 decode_pfor_delta1_128(pldata, num, /*interleaved=*/true, &decoded[0]);
 613
 614                                 assert(uses_trigram.count(trgm) != 0);
 615                                 bool was_empty = cur_candidates.empty();
 616                                 if (ignore_case) {
 617                                         dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 618                                 }
 619
 620                                 for (TrigramDisjunction *td : uses_trigram[trgm]) {
 621                                         done |= new_posting_list_read(td, decoded, &cur_candidates, &tmp);
 622                                         if (done)
 623                                                 break;
 624                                 }
 625                                 if (!ignore_case) {
 626                                         if (was_empty) {
 627                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 628                                         } else if (cur_candidates.empty()) {
 629                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (none left, search is done)\n", print_trigram(trgm).c_str(), len, num);
 630                                         } else {
 631                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (%zu left)\n", print_trigram(trgm).c_str(), len, num, cur_candidates.size());
 632                                         }
 633                                 }
 634                         });
 635                 }
 636         }
 637         engine.finish();
 638         if (done) {
 639                 return;
 640         }
 641         dprintf("Intersection done after %.1f ms. Doing final verification and printing:\n",
 642                 1e3 * duration<float>(steady_clock::now() - start).count());
 643
 644         uint64_t matched = scan_docids(needles, cur_candidates, corpus, &engine);
 645         dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
 646                 1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 647
 648         if (only_count) {
 649                 printf("%" PRId64 "\n", matched);
 650         }
 651 }
 652
 653 void usage()
 654 {
 655         printf(
 656                 "Usage: plocate [OPTION]... PATTERN...\n"
 657                 "\n"
 658                 "  -b, --basename         search only the file name portion of path names\n"
 659                 "  -c, --count            print number of matches instead of the matches\n"
 660                 "  -d, --database DBPATH  search for files in DBPATH\n"
 661                 "                         (default is " DBFILE ")\n"
 662                 "  -i, --ignore-case      search case-insensitively\n"
 663                 "  -l, --limit LIMIT      stop after LIMIT matches\n"
 664                 "  -0, --null             delimit matches by NUL instead of newline\n"
 665                 "  -r, --regexp           interpret patterns as basic regexps (slow)\n"
 666                 "      --regex            interpret patterns as extended regexps (slow)\n"
 667                 "  -w, --wholename        search the entire path name (default; see -b)\n"
 668                 "      --help             print this help\n"
 669                 "      --version          print version information\n");
 670 }
 671
 672 void version()
 673 {
 674         printf("%s %s\n", PACKAGE_NAME, PACKAGE_VERSION);
 675         printf("Copyright 2020 Steinar H. Gunderson\n");
 676         printf("License GPLv2+: GNU GPL version 2 or later <https://gnu.org/licenses/gpl.html>.\n");
 677         printf("This is free software: you are free to change and redistribute it.\n");
 678         printf("There is NO WARRANTY, to the extent permitted by law.\n");
 679         exit(0);
 680 }
 681
 682 int main(int argc, char **argv)
 683 {
 684         constexpr int EXTENDED_REGEX = 1000;
 685         constexpr int FLUSH_CACHE = 1001;
 686         static const struct option long_options[] = {
 687                 { "help", no_argument, 0, 'h' },
 688                 { "count", no_argument, 0, 'c' },
 689                 { "basename", no_argument, 0, 'b' },
 690                 { "database", required_argument, 0, 'd' },
 691                 { "ignore-case", no_argument, 0, 'i' },
 692                 { "limit", required_argument, 0, 'l' },
 693                 { "null", no_argument, 0, '0' },
 694                 { "version", no_argument, 0, 'V' },
 695                 { "regexp", no_argument, 0, 'r' },
 696                 { "regex", no_argument, 0, EXTENDED_REGEX },
 697                 { "wholename", no_argument, 0, 'w' },
 698                 { "debug", no_argument, 0, 'D' },  // Not documented.
 699                 // Enable to test cold-cache behavior (except for access()). Not documented.
 700                 { "flush-cache", no_argument, 0, FLUSH_CACHE },
 701                 { 0, 0, 0, 0 }
 702         };
 703
 704         setlocale(LC_ALL, "");
 705         for (;;) {
 706                 int option_index = 0;
 707                 int c = getopt_long(argc, argv, "bcd:hil:n:0rwVD", long_options, &option_index);
 708                 if (c == -1) {
 709                         break;
 710                 }
 711                 switch (c) {
 712                 case 'b':
 713                         match_basename = true;
 714                         break;
 715                 case 'c':
 716                         only_count = true;
 717                         break;
 718                 case 'd':
 719                         dbpath = strdup(optarg);
 720                         break;
 721                 case 'h':
 722                         usage();
 723                         exit(0);
 724                 case 'i':
 725                         ignore_case = true;
 726                         break;
 727                 case 'l':
 728                 case 'n':
 729                         limit_matches = limit_left = atoll(optarg);
 730                         if (limit_matches <= 0) {
 731                                 fprintf(stderr, "Error: limit must be a strictly positive number.\n");
 732                                 exit(1);
 733                         }
 734                         break;
 735                 case '0':
 736                         print_nul = true;
 737                         break;
 738                 case 'r':
 739                         patterns_are_regex = true;
 740                         break;
 741                 case EXTENDED_REGEX:
 742                         patterns_are_regex = true;
 743                         use_extended_regex = true;
 744                         break;
 745                 case 'w':
 746                         match_basename = false;  // No-op unless -b is given first.
 747                         break;
 748                 case 'D':
 749                         use_debug = true;
 750                         break;
 751                 case FLUSH_CACHE:
 752                         flush_cache = true;
 753                         break;
 754                 case 'V':
 755                         version();
 756                         break;
 757                 default:
 758                         exit(1);
 759                 }
 760         }
 761
 762         if (use_debug || flush_cache) {
 763                 // Debug information would leak information about which files exist,
 764                 // so drop setgid before we open the file; one would either need to run
 765                 // as root, or use a locally-built file. Doing the same thing for
 766                 // flush_cache is mostly paranoia, in an attempt to prevent random users
 767                 // from making plocate slow for everyone else.
 768                 if (setgid(getgid()) != 0) {
 769                         perror("setgid");
 770                         exit(EXIT_FAILURE);
 771                 }
 772         }
 773
 774         if (!print_nul) {
 775                 stdout_is_tty = isatty(1);
 776         }
 777
 778         vector<Needle> needles;
 779         for (int i = optind; i < argc; ++i) {
 780                 Needle needle;
 781                 needle.str = argv[i];
 782
 783                 // See if there are any wildcard characters, which indicates we should treat it
 784                 // as an (anchored) glob.
 785                 bool any_wildcard = false;
 786                 for (size_t i = 0; i < needle.str.size(); i += read_unigram(needle.str, i).second) {
 787                         if (read_unigram(needle.str, i).first == WILDCARD_UNIGRAM) {
 788                                 any_wildcard = true;
 789                                 break;
 790                         }
 791                 }
 792
 793                 if (patterns_are_regex) {
 794                         needle.type = Needle::REGEX;
 795                         needle.re = compile_regex(needle.str);
 796                 } else if (any_wildcard) {
 797                         needle.type = Needle::GLOB;
 798                 } else if (ignore_case) {
 799                         // strcasestr() doesn't handle locales correctly (even though LSB
 800                         // claims it should), but somehow, fnmatch() does, and it's about
 801                         // the same speed as using a regex.
 802                         needle.type = Needle::GLOB;
 803                         needle.str = "*" + needle.str + "*";
 804                 } else {
 805                         needle.type = Needle::STRSTR;
 806                         needle.str = unescape_glob_to_plain_string(needle.str);
 807                 }
 808                 needles.push_back(move(needle));
 809         }
 810         if (needles.empty()) {
 811                 fprintf(stderr, "plocate: no pattern to search for specified\n");
 812                 exit(0);
 813         }
 814         do_search_file(needles, dbpath);
 815 }