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