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 #define DEFAULT_DBPATH "/var/lib/mlocate/plocate.db"
  47
  48 const char *dbpath = DEFAULT_DBPATH;
  49 bool ignore_case = false;
  50 bool only_count = false;
  51 bool print_nul = false;
  52 bool use_debug = false;
  53 bool flush_cache = false;
  54 bool patterns_are_regex = false;
  55 bool use_extended_regex = false;
  56 bool match_basename = false;
  57 int64_t limit_matches = numeric_limits<int64_t>::max();
  58 int64_t limit_left = numeric_limits<int64_t>::max();
  59 bool stdout_is_tty = false;
  60
  61 steady_clock::time_point start;
  62 ZSTD_DDict *ddict = nullptr;
  63
  64 class Corpus {
  65 public:
  66         Corpus(int fd, IOUringEngine *engine);
  67         ~Corpus();
  68         void find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb);
  69         void get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const;
  70         size_t get_num_filename_blocks() const;
  71         off_t offset_for_block(uint32_t docid) const
  72         {
  73                 return hdr.filename_index_offset_bytes + docid * sizeof(uint64_t);
  74         }
  75         const Header &get_hdr() const { return hdr; }
  76
  77 public:
  78         const int fd;
  79         IOUringEngine *const engine;
  80
  81         Header hdr;
  82 };
  83
  84 Corpus::Corpus(int fd, IOUringEngine *engine)
  85         : fd(fd), engine(engine)
  86 {
  87         if (flush_cache) {
  88                 off_t len = lseek(fd, 0, SEEK_END);
  89                 if (len == -1) {
  90                         perror("lseek");
  91                         exit(1);
  92                 }
  93                 posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED);
  94         }
  95
  96         complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
  97         if (memcmp(hdr.magic, "\0plocate", 8) != 0) {
  98                 fprintf(stderr, "plocate.db is corrupt or an old version; please rebuild it.\n");
  99                 exit(1);
 100         }
 101         if (hdr.version != 0 && hdr.version != 1) {
 102                 fprintf(stderr, "plocate.db has version %u, expected 0 or 1; please rebuild it.\n", hdr.version);
 103                 exit(1);
 104         }
 105         if (hdr.version == 0) {
 106                 // These will be junk data.
 107                 hdr.zstd_dictionary_offset_bytes = 0;
 108                 hdr.zstd_dictionary_length_bytes = 0;
 109         }
 110 }
 111
 112 Corpus::~Corpus()
 113 {
 114         close(fd);
 115 }
 116
 117 void Corpus::find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb)
 118 {
 119         uint32_t bucket = hash_trigram(trgm, hdr.hashtable_size);
 120         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) {
 121                 const Trigram *trgmptr = reinterpret_cast<const Trigram *>(s.data());
 122                 for (unsigned i = 0; i < hdr.extra_ht_slots + 1; ++i) {
 123                         if (trgmptr[i].trgm == trgm) {
 124                                 cb(trgmptr + i, trgmptr[i + 1].offset - trgmptr[i].offset);
 125                                 return;
 126                         }
 127                 }
 128
 129                 // Not found.
 130                 cb(nullptr, 0);
 131         });
 132 }
 133
 134 void Corpus::get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const
 135 {
 136         // Read the file offset from this docid and the next one.
 137         // This is always allowed, since we have a sentinel block at the end.
 138         engine->submit_read(fd, sizeof(uint64_t) * 2, offset_for_block(docid), [this, cb{ move(cb) }](string_view s) {
 139                 const uint64_t *ptr = reinterpret_cast<const uint64_t *>(s.data());
 140                 off_t offset = ptr[0];
 141                 size_t len = ptr[1] - ptr[0];
 142                 engine->submit_read(fd, len, offset, cb);
 143         });
 144 }
 145
 146 size_t Corpus::get_num_filename_blocks() const
 147 {
 148         return hdr.num_docids;
 149 }
 150
 151 void scan_file_block(const vector<Needle> &needles, string_view compressed,
 152                      AccessRXCache *access_rx_cache, uint64_t seq, ResultReceiver *serializer,
 153                      atomic<uint64_t> *matched)
 154 {
 155         unsigned long long uncompressed_len = ZSTD_getFrameContentSize(compressed.data(), compressed.size());
 156         if (uncompressed_len == ZSTD_CONTENTSIZE_UNKNOWN || uncompressed_len == ZSTD_CONTENTSIZE_ERROR) {
 157                 fprintf(stderr, "ZSTD_getFrameContentSize() failed\n");
 158                 exit(1);
 159         }
 160
 161         string block;
 162         block.resize(uncompressed_len + 1);
 163
 164         static thread_local ZSTD_DCtx *ctx = ZSTD_createDCtx();  // Reused across calls.
 165         size_t err;
 166
 167         if (ddict != nullptr) {
 168                 err = ZSTD_decompress_usingDDict(ctx, &block[0], block.size(), compressed.data(),
 169                                                  compressed.size(), ddict);
 170         } else {
 171                 err = ZSTD_decompressDCtx(ctx, &block[0], block.size(), compressed.data(),
 172                                           compressed.size());
 173         }
 174         if (ZSTD_isError(err)) {
 175                 fprintf(stderr, "ZSTD_decompress(): %s\n", ZSTD_getErrorName(err));
 176                 exit(1);
 177         }
 178         block[block.size() - 1] = '\0';
 179
 180         auto test_candidate = [&](const char *filename, uint64_t local_seq, uint64_t next_seq) {
 181                 access_rx_cache->check_access(filename, /*allow_async=*/true, [matched, serializer, local_seq, next_seq, filename{ strdup(filename) }](bool ok) {
 182                         if (ok) {
 183                                 ++*matched;
 184                                 serializer->print(local_seq, next_seq - local_seq, filename);
 185                         } else {
 186                                 serializer->print(local_seq, next_seq - local_seq, "");
 187                         }
 188                         free(filename);
 189                 });
 190         };
 191
 192         // We need to know the next sequence number before inserting into Serializer,
 193         // so always buffer one candidate.
 194         const char *pending_candidate = nullptr;
 195
 196         uint64_t local_seq = seq << 32;
 197         for (const char *filename = block.data();
 198              filename != block.data() + block.size();
 199              filename += strlen(filename) + 1) {
 200                 const char *haystack = filename;
 201                 if (match_basename) {
 202                         haystack = strrchr(filename, '/');
 203                         if (haystack == nullptr) {
 204                                 haystack = filename;
 205                         } else {
 206                                 ++haystack;
 207                         }
 208                 }
 209
 210                 bool found = true;
 211                 for (const Needle &needle : needles) {
 212                         if (!matches(needle, haystack)) {
 213                                 found = false;
 214                                 break;
 215                         }
 216                 }
 217                 if (found) {
 218                         if (pending_candidate != nullptr) {
 219                                 test_candidate(pending_candidate, local_seq, local_seq + 1);
 220                                 ++local_seq;
 221                         }
 222                         pending_candidate = filename;
 223                 }
 224         }
 225         if (pending_candidate == nullptr) {
 226                 serializer->print(seq << 32, 1ULL << 32, "");
 227         } else {
 228                 test_candidate(pending_candidate, local_seq, (seq + 1) << 32);
 229         }
 230 }
 231
 232 size_t scan_docids(const vector<Needle> &needles, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
 233 {
 234         Serializer docids_in_order;
 235         AccessRXCache access_rx_cache(engine);
 236         atomic<uint64_t> matched{ 0 };
 237         for (size_t i = 0; i < docids.size(); ++i) {
 238                 uint32_t docid = docids[i];
 239                 corpus.get_compressed_filename_block(docid, [i, &matched, &needles, &access_rx_cache, &docids_in_order](string_view compressed) {
 240                         scan_file_block(needles, compressed, &access_rx_cache, i, &docids_in_order, &matched);
 241                 });
 242         }
 243         engine->finish();
 244         return matched;
 245 }
 246
 247 struct WorkerThread {
 248         thread t;
 249
 250         // We use a result queue instead of synchronizing Serializer,
 251         // since a lock on it becomes a huge choke point if there are
 252         // lots of threads.
 253         mutex result_mu;
 254         struct Result {
 255                 uint64_t seq;
 256                 uint64_t skip;
 257                 string msg;
 258         };
 259         vector<Result> results;
 260 };
 261
 262 class WorkerThreadReceiver : public ResultReceiver {
 263 public:
 264         WorkerThreadReceiver(WorkerThread *wt)
 265                 : wt(wt) {}
 266
 267         void print(uint64_t seq, uint64_t skip, const string msg) override
 268         {
 269                 lock_guard<mutex> lock(wt->result_mu);
 270                 if (msg.empty() && !wt->results.empty() && wt->results.back().seq + wt->results.back().skip == seq) {
 271                         wt->results.back().skip += skip;
 272                 } else {
 273                         wt->results.emplace_back(WorkerThread::Result{ seq, skip, move(msg) });
 274                 }
 275         }
 276
 277 private:
 278         WorkerThread *wt;
 279 };
 280
 281 void deliver_results(WorkerThread *wt, Serializer *serializer)
 282 {
 283         vector<WorkerThread::Result> results;
 284         {
 285                 lock_guard<mutex> lock(wt->result_mu);
 286                 results = move(wt->results);
 287         }
 288         for (const WorkerThread::Result &result : results) {
 289                 serializer->print(result.seq, result.skip, move(result.msg));
 290         }
 291 }
 292
 293 // We do this sequentially, as it's faster than scattering
 294 // a lot of I/O through io_uring and hoping the kernel will
 295 // coalesce it plus readahead for us. Since we assume that
 296 // we will primarily be CPU-bound, we'll be firing up one
 297 // worker thread for each spare core (the last one will
 298 // only be doing I/O). access() is still synchronous.
 299 uint64_t scan_all_docids(const vector<Needle> &needles, int fd, const Corpus &corpus)
 300 {
 301         {
 302                 const Header &hdr = corpus.get_hdr();
 303                 if (hdr.zstd_dictionary_length_bytes > 0) {
 304                         string dictionary;
 305                         dictionary.resize(hdr.zstd_dictionary_length_bytes);
 306                         complete_pread(fd, &dictionary[0], hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes);
 307                         ddict = ZSTD_createDDict(dictionary.data(), dictionary.size());
 308                 }
 309         }
 310
 311         AccessRXCache access_rx_cache(nullptr);
 312         Serializer serializer;
 313         uint32_t num_blocks = corpus.get_num_filename_blocks();
 314         unique_ptr<uint64_t[]> offsets(new uint64_t[num_blocks + 1]);
 315         complete_pread(fd, offsets.get(), (num_blocks + 1) * sizeof(uint64_t), corpus.offset_for_block(0));
 316         atomic<uint64_t> matched{ 0 };
 317
 318         mutex mu;
 319         condition_variable queue_added, queue_removed;
 320         deque<tuple<int, int, string>> work_queue;  // Under mu.
 321         bool done = false;  // Under mu.
 322
 323         unsigned num_threads = max<int>(sysconf(_SC_NPROCESSORS_ONLN) - 1, 1);
 324         dprintf("Using %u worker threads for linear scan.\n", num_threads);
 325         unique_ptr<WorkerThread[]> threads(new WorkerThread[num_threads]);
 326         for (unsigned i = 0; i < num_threads; ++i) {
 327                 threads[i].t = thread([&threads, &mu, &queue_added, &queue_removed, &work_queue, &done, &offsets, &needles, &access_rx_cache, &matched, i] {
 328                         // regcomp() takes a lock on the regex, so each thread will need its own.
 329                         const vector<Needle> *use_needles = &needles;
 330                         vector<Needle> recompiled_needles;
 331                         if (i != 0 && patterns_are_regex) {
 332                                 recompiled_needles = needles;
 333                                 for (Needle &needle : recompiled_needles) {
 334                                         needle.re = compile_regex(needle.str);
 335                                 }
 336                                 use_needles = &recompiled_needles;
 337                         }
 338
 339                         WorkerThreadReceiver receiver(&threads[i]);
 340                         for (;;) {
 341                                 uint32_t io_docid, last_docid;
 342                                 string compressed;
 343
 344                                 {
 345                                         unique_lock<mutex> lock(mu);
 346                                         queue_added.wait(lock, [&work_queue, &done] { return !work_queue.empty() || done; });
 347                                         if (done && work_queue.empty()) {
 348                                                 return;
 349                                         }
 350                                         tie(io_docid, last_docid, compressed) = move(work_queue.front());
 351                                         work_queue.pop_front();
 352                                         queue_removed.notify_all();
 353                                 }
 354
 355                                 for (uint32_t docid = io_docid; docid < last_docid; ++docid) {
 356                                         size_t relative_offset = offsets[docid] - offsets[io_docid];
 357                                         size_t len = offsets[docid + 1] - offsets[docid];
 358                                         scan_file_block(*use_needles, { &compressed[relative_offset], len }, &access_rx_cache, docid, &receiver, &matched);
 359                                 }
 360                         }
 361                 });
 362         }
 363
 364         string compressed;
 365         for (uint32_t io_docid = 0; io_docid < num_blocks; io_docid += 32) {
 366                 uint32_t last_docid = std::min(io_docid + 32, num_blocks);
 367                 size_t io_len = offsets[last_docid] - offsets[io_docid];
 368                 if (compressed.size() < io_len) {
 369                         compressed.resize(io_len);
 370                 }
 371                 complete_pread(fd, &compressed[0], io_len, offsets[io_docid]);
 372
 373                 {
 374                         unique_lock<mutex> lock(mu);
 375                         queue_removed.wait(lock, [&work_queue] { return work_queue.size() < 256; });  // Allow ~2MB of data queued up.
 376                         work_queue.emplace_back(io_docid, last_docid, move(compressed));
 377                         queue_added.notify_one();  // Avoid the thundering herd.
 378                 }
 379
 380                 // Pick up some results, so that we are sure that we won't just overload.
 381                 // (Seemingly, going through all of these causes slowness with many threads,
 382                 // but taking only one is OK.)
 383                 unsigned i = io_docid / 32;
 384                 deliver_results(&threads[i % num_threads], &serializer);
 385         }
 386         {
 387                 lock_guard<mutex> lock(mu);
 388                 done = true;
 389                 queue_added.notify_all();
 390         }
 391         for (unsigned i = 0; i < num_threads; ++i) {
 392                 threads[i].t.join();
 393                 deliver_results(&threads[i], &serializer);
 394         }
 395         return matched;
 396 }
 397
 398 // Takes the given posting list, unions it into the parts of the trigram disjunction
 399 // already read; if the list is complete, intersects with “cur_candidates”.
 400 //
 401 // Returns true if the search should be aborted (we are done).
 402 bool new_posting_list_read(TrigramDisjunction *td, vector<uint32_t> decoded, vector<uint32_t> *cur_candidates, vector<uint32_t> *tmp)
 403 {
 404         if (td->docids.empty()) {
 405                 td->docids = move(decoded);
 406         } else {
 407                 tmp->clear();
 408                 set_union(decoded.begin(), decoded.end(), td->docids.begin(), td->docids.end(), back_inserter(*tmp));
 409                 swap(*tmp, td->docids);
 410         }
 411         if (--td->remaining_trigrams_to_read > 0) {
 412                 // Need to wait for more.
 413                 if (ignore_case) {
 414                         dprintf("  ... %u reads left in OR group %u (%zu docids in list)\n",
 415                                 td->remaining_trigrams_to_read, td->index, td->docids.size());
 416                 }
 417                 return false;
 418         }
 419         if (cur_candidates->empty()) {
 420                 if (ignore_case) {
 421                         dprintf("  ... all reads done for OR group %u (%zu docids)\n",
 422                                 td->index, td->docids.size());
 423                 }
 424                 *cur_candidates = move(td->docids);
 425         } else {
 426                 tmp->clear();
 427                 set_intersection(cur_candidates->begin(), cur_candidates->end(),
 428                                  td->docids.begin(), td->docids.end(),
 429                                  back_inserter(*tmp));
 430                 swap(*cur_candidates, *tmp);
 431                 if (ignore_case) {
 432                         if (cur_candidates->empty()) {
 433                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (none left, search is done)\n",
 434                                         td->index, td->docids.size());
 435                                 return true;
 436                         } else {
 437                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (%zu left)\n",
 438                                         td->index, td->docids.size(), cur_candidates->size());
 439                         }
 440                 }
 441         }
 442         return false;
 443 }
 444
 445 void do_search_file(const vector<Needle> &needles, const char *filename)
 446 {
 447         int fd = open(filename, O_RDONLY);
 448         if (fd == -1) {
 449                 perror(filename);
 450                 exit(1);
 451         }
 452
 453         // Drop privileges.
 454         if (setgid(getgid()) != 0) {
 455                 perror("setgid");
 456                 exit(EXIT_FAILURE);
 457         }
 458
 459         start = steady_clock::now();
 460         if (access("/", R_OK | X_OK)) {
 461                 // We can't find anything, no need to bother...
 462                 return;
 463         }
 464
 465         IOUringEngine engine(/*slop_bytes=*/16);  // 16 slop bytes as described in turbopfor.h.
 466         Corpus corpus(fd, &engine);
 467         dprintf("Corpus init done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 468
 469         vector<TrigramDisjunction> trigram_groups;
 470         if (patterns_are_regex) {
 471                 // We could parse the regex to find trigrams that have to be there
 472                 // (there are actually known algorithms to deal with disjunctions
 473                 // and such, too), but for now, we just go brute force.
 474                 // Using locate with regexes is pretty niche.
 475         } else {
 476                 for (const Needle &needle : needles) {
 477                         parse_trigrams(needle.str, ignore_case, &trigram_groups);
 478                 }
 479         }
 480
 481         unique_sort(
 482                 &trigram_groups,
 483                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives < b.trigram_alternatives; },
 484                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives == b.trigram_alternatives; });
 485
 486         // Give them names for debugging.
 487         unsigned td_index = 0;
 488         for (TrigramDisjunction &td : trigram_groups) {
 489                 td.index = td_index++;
 490         }
 491
 492         // Collect which trigrams we need to look up in the hash table.
 493         unordered_map<uint32_t, vector<TrigramDisjunction *>> trigrams_to_lookup;
 494         for (TrigramDisjunction &td : trigram_groups) {
 495                 for (uint32_t trgm : td.trigram_alternatives) {
 496                         trigrams_to_lookup[trgm].push_back(&td);
 497                 }
 498         }
 499         if (trigrams_to_lookup.empty()) {
 500                 // Too short for trigram matching. Apply brute force.
 501                 // (We could have searched through all trigrams that matched
 502                 // the pattern and done a union of them, but that's a lot of
 503                 // work for fairly unclear gain.)
 504                 uint64_t matched = scan_all_docids(needles, fd, corpus);
 505                 dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
 506                         1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 507                 if (only_count) {
 508                         printf("%" PRId64 "\n", matched);
 509                 }
 510                 return;
 511         }
 512
 513         // Sneak in fetching the dictionary, if present. It's not necessarily clear
 514         // exactly where it would be cheapest to get it, but it needs to be present
 515         // before we can decode any of the posting lists. Most likely, it's
 516         // in the same filesystem block as the header anyway, so it should be
 517         // present in the cache.
 518         {
 519                 const Header &hdr = corpus.get_hdr();
 520                 if (hdr.zstd_dictionary_length_bytes > 0) {
 521                         engine.submit_read(fd, hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes, [](string_view s) {
 522                                 ddict = ZSTD_createDDict(s.data(), s.size());
 523                                 dprintf("Dictionary initialized after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 524                         });
 525                 }
 526         }
 527
 528         // Look them all up on disk.
 529         for (auto &[trgm, trigram_groups] : trigrams_to_lookup) {
 530                 corpus.find_trigram(trgm, [trgm{ trgm }, trigram_groups{ &trigram_groups }](const Trigram *trgmptr, size_t len) {
 531                         if (trgmptr == nullptr) {
 532                                 dprintf("trigram %s isn't found\n", print_trigram(trgm).c_str());
 533                                 for (TrigramDisjunction *td : *trigram_groups) {
 534                                         --td->remaining_trigrams_to_read;
 535                                         if (td->remaining_trigrams_to_read == 0 && td->read_trigrams.empty()) {
 536                                                 dprintf("zero matches in %s, so we are done\n", print_td(*td).c_str());
 537                                                 if (only_count) {
 538                                                         printf("0\n");
 539                                                 }
 540                                                 exit(0);
 541                                         }
 542                                 }
 543                                 return;
 544                         }
 545                         for (TrigramDisjunction *td : *trigram_groups) {
 546                                 --td->remaining_trigrams_to_read;
 547                                 td->max_num_docids += trgmptr->num_docids;
 548                                 td->read_trigrams.emplace_back(*trgmptr, len);
 549                         }
 550                 });
 551         }
 552         engine.finish();
 553         dprintf("Hashtable lookups done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 554
 555         for (TrigramDisjunction &td : trigram_groups) {
 556                 // Reset for reads.
 557                 td.remaining_trigrams_to_read = td.read_trigrams.size();
 558
 559                 if (ignore_case) {  // If case-sensitive, they'll all be pretty obvious single-entry groups.
 560                         dprintf("OR group %u (max_num_docids=%u): %s\n", td.index, td.max_num_docids, print_td(td).c_str());
 561                 }
 562         }
 563
 564         // TODO: For case-insensitive (ie. more than one alternative in each),
 565         // prioritize the ones with fewer seeks?
 566         sort(trigram_groups.begin(), trigram_groups.end(),
 567              [&](const TrigramDisjunction &a, const TrigramDisjunction &b) {
 568                      return a.max_num_docids < b.max_num_docids;
 569              });
 570
 571         unordered_map<uint32_t, vector<TrigramDisjunction *>> uses_trigram;
 572         for (TrigramDisjunction &td : trigram_groups) {
 573                 for (uint32_t trgm : td.trigram_alternatives) {
 574                         uses_trigram[trgm].push_back(&td);
 575                 }
 576         }
 577
 578         unordered_set<uint32_t> trigrams_submitted_read;
 579         vector<uint32_t> cur_candidates, tmp, decoded;
 580         bool done = false;
 581         for (TrigramDisjunction &td : trigram_groups) {
 582                 if (!cur_candidates.empty() && td.max_num_docids > cur_candidates.size() * 100) {
 583                         dprintf("%s has up to %u entries, ignoring the rest (will "
 584                                 "weed out false positives later)\n",
 585                                 print_td(td).c_str(), td.max_num_docids);
 586                         break;
 587                 }
 588
 589                 for (auto &[trgmptr, len] : td.read_trigrams) {
 590                         if (trigrams_submitted_read.count(trgmptr.trgm) != 0) {
 591                                 continue;
 592                         }
 593                         trigrams_submitted_read.insert(trgmptr.trgm);
 594                         // Only stay a certain amount ahead, so that we don't spend I/O
 595                         // on reading the latter, large posting lists. We are unlikely
 596                         // to need them anyway, even if they should come in first.
 597                         if (engine.get_waiting_reads() >= 5) {
 598                                 engine.finish();
 599                                 if (done)
 600                                         break;
 601                         }
 602                         engine.submit_read(fd, len, trgmptr.offset, [trgmptr{ trgmptr }, len{ len }, &done, &cur_candidates, &tmp, &decoded, &uses_trigram](string_view s) {
 603                                 if (done)
 604                                         return;
 605
 606                                 uint32_t trgm = trgmptr.trgm;
 607                                 const unsigned char *pldata = reinterpret_cast<const unsigned char *>(s.data());
 608                                 size_t num = trgmptr.num_docids;
 609                                 decoded.resize(num);
 610                                 decode_pfor_delta1_128(pldata, num, /*interleaved=*/true, &decoded[0]);
 611
 612                                 assert(uses_trigram.count(trgm) != 0);
 613                                 bool was_empty = cur_candidates.empty();
 614                                 if (ignore_case) {
 615                                         dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 616                                 }
 617
 618                                 for (TrigramDisjunction *td : uses_trigram[trgm]) {
 619                                         done |= new_posting_list_read(td, decoded, &cur_candidates, &tmp);
 620                                         if (done)
 621                                                 break;
 622                                 }
 623                                 if (!ignore_case) {
 624                                         if (was_empty) {
 625                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 626                                         } else if (cur_candidates.empty()) {
 627                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (none left, search is done)\n", print_trigram(trgm).c_str(), len, num);
 628                                         } else {
 629                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (%zu left)\n", print_trigram(trgm).c_str(), len, num, cur_candidates.size());
 630                                         }
 631                                 }
 632                         });
 633                 }
 634         }
 635         engine.finish();
 636         if (done) {
 637                 return;
 638         }
 639         dprintf("Intersection done after %.1f ms. Doing final verification and printing:\n",
 640                 1e3 * duration<float>(steady_clock::now() - start).count());
 641
 642         uint64_t matched = scan_docids(needles, cur_candidates, corpus, &engine);
 643         dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
 644                 1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 645
 646         if (only_count) {
 647                 printf("%" PRId64 "\n", matched);
 648         }
 649 }
 650
 651 void usage()
 652 {
 653         printf(
 654                 "Usage: plocate [OPTION]... PATTERN...\n"
 655                 "\n"
 656                 "  -b, --basename         search only the file name portion of path names\n"
 657                 "  -c, --count            print number of matches instead of the matches\n"
 658                 "  -d, --database DBPATH  search for files in DBPATH\n"
 659                 "                         (default is " DEFAULT_DBPATH ")\n"
 660                 "  -i, --ignore-case      search case-insensitively\n"
 661                 "  -l, --limit LIMIT      stop after LIMIT matches\n"
 662                 "  -0, --null             delimit matches by NUL instead of newline\n"
 663                 "  -r, --regexp           interpret patterns as basic regexps (slow)\n"
 664                 "      --regex            interpret patterns as extended regexps (slow)\n"
 665                 "  -w, --wholename        search the entire path name (default; see -b)\n"
 666                 "      --help             print this help\n"
 667                 "      --version          print version information\n");
 668 }
 669
 670 void version()
 671 {
 672         printf("plocate %s\n", PLOCATE_VERSION);
 673         printf("Copyright 2020 Steinar H. Gunderson\n");
 674         printf("License GPLv2+: GNU GPL version 2 or later <https://gnu.org/licenses/gpl.html>.\n");
 675         printf("This is free software: you are free to change and redistribute it.\n");
 676         printf("There is NO WARRANTY, to the extent permitted by law.\n");
 677         exit(0);
 678 }
 679
 680 int main(int argc, char **argv)
 681 {
 682         constexpr int EXTENDED_REGEX = 1000;
 683         constexpr int FLUSH_CACHE = 1001;
 684         static const struct option long_options[] = {
 685                 { "help", no_argument, 0, 'h' },
 686                 { "count", no_argument, 0, 'c' },
 687                 { "basename", no_argument, 0, 'b' },
 688                 { "database", required_argument, 0, 'd' },
 689                 { "ignore-case", no_argument, 0, 'i' },
 690                 { "limit", required_argument, 0, 'l' },
 691                 { "null", no_argument, 0, '0' },
 692                 { "version", no_argument, 0, 'V' },
 693                 { "regexp", no_argument, 0, 'r' },
 694                 { "regex", no_argument, 0, EXTENDED_REGEX },
 695                 { "wholename", no_argument, 0, 'w' },
 696                 { "debug", no_argument, 0, 'D' },  // Not documented.
 697                 // Enable to test cold-cache behavior (except for access()). Not documented.
 698                 { "flush-cache", no_argument, 0, FLUSH_CACHE },
 699                 { 0, 0, 0, 0 }
 700         };
 701
 702         setlocale(LC_ALL, "");
 703         for (;;) {
 704                 int option_index = 0;
 705                 int c = getopt_long(argc, argv, "bcd:hil:n:0rwVD", long_options, &option_index);
 706                 if (c == -1) {
 707                         break;
 708                 }
 709                 switch (c) {
 710                 case 'b':
 711                         match_basename = true;
 712                         break;
 713                 case 'c':
 714                         only_count = true;
 715                         break;
 716                 case 'd':
 717                         dbpath = strdup(optarg);
 718                         break;
 719                 case 'h':
 720                         usage();
 721                         exit(0);
 722                 case 'i':
 723                         ignore_case = true;
 724                         break;
 725                 case 'l':
 726                 case 'n':
 727                         limit_matches = limit_left = atoll(optarg);
 728                         if (limit_matches <= 0) {
 729                                 fprintf(stderr, "Error: limit must be a strictly positive number.\n");
 730                                 exit(1);
 731                         }
 732                         break;
 733                 case '0':
 734                         print_nul = true;
 735                         break;
 736                 case 'r':
 737                         patterns_are_regex = true;
 738                         break;
 739                 case EXTENDED_REGEX:
 740                         patterns_are_regex = true;
 741                         use_extended_regex = true;
 742                         break;
 743                 case 'w':
 744                         match_basename = false;  // No-op unless -b is given first.
 745                         break;
 746                 case 'D':
 747                         use_debug = true;
 748                         break;
 749                 case FLUSH_CACHE:
 750                         flush_cache = true;
 751                         break;
 752                 case 'V':
 753                         version();
 754                         break;
 755                 default:
 756                         exit(1);
 757                 }
 758         }
 759
 760         if (use_debug || flush_cache) {
 761                 // Debug information would leak information about which files exist,
 762                 // so drop setgid before we open the file; one would either need to run
 763                 // as root, or use a locally-built file. Doing the same thing for
 764                 // flush_cache is mostly paranoia, in an attempt to prevent random users
 765                 // from making plocate slow for everyone else.
 766                 if (setgid(getgid()) != 0) {
 767                         perror("setgid");
 768                         exit(EXIT_FAILURE);
 769                 }
 770         }
 771
 772         if (!print_nul) {
 773                 stdout_is_tty = isatty(1);
 774         }
 775
 776         vector<Needle> needles;
 777         for (int i = optind; i < argc; ++i) {
 778                 Needle needle;
 779                 needle.str = argv[i];
 780
 781                 // See if there are any wildcard characters, which indicates we should treat it
 782                 // as an (anchored) glob.
 783                 bool any_wildcard = false;
 784                 for (size_t i = 0; i < needle.str.size(); i += read_unigram(needle.str, i).second) {
 785                         if (read_unigram(needle.str, i).first == WILDCARD_UNIGRAM) {
 786                                 any_wildcard = true;
 787                                 break;
 788                         }
 789                 }
 790
 791                 if (patterns_are_regex) {
 792                         needle.type = Needle::REGEX;
 793                         needle.re = compile_regex(needle.str);
 794                 } else if (any_wildcard) {
 795                         needle.type = Needle::GLOB;
 796                 } else if (ignore_case) {
 797                         // strcasestr() doesn't handle locales correctly (even though LSB
 798                         // claims it should), but somehow, fnmatch() does, and it's about
 799                         // the same speed as using a regex.
 800                         needle.type = Needle::GLOB;
 801                         needle.str = "*" + needle.str + "*";
 802                 } else {
 803                         needle.type = Needle::STRSTR;
 804                         needle.str = unescape_glob_to_plain_string(needle.str);
 805                 }
 806                 needles.push_back(move(needle));
 807         }
 808         if (needles.empty()) {
 809                 fprintf(stderr, "plocate: no pattern to search for specified\n");
 810                 exit(0);
 811         }
 812         do_search_file(needles, dbpath);
 813 }