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 "parse_trigrams.h"
   6 #include "serializer.h"
   7 #include "turbopfor.h"
   8 #include "unique_sort.h"
   9
  10 #include <algorithm>
  11 #include <assert.h>
  12 #include <atomic>
  13 #include <chrono>
  14 #include <condition_variable>
  15 #include <deque>
  16 #include <fcntl.h>
  17 #include <fnmatch.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 patterns_are_regex = false;
  53 bool use_extended_regex = false;
  54 int64_t limit_matches = numeric_limits<int64_t>::max();
  55 int64_t limit_left = numeric_limits<int64_t>::max();
  56
  57 steady_clock::time_point start;
  58 ZSTD_DDict *ddict = nullptr;
  59
  60 regex_t compile_regex(const string &needle);
  61
  62 struct Needle {
  63         enum { STRSTR,
  64                REGEX,
  65                GLOB } type;
  66         string str;  // Filled in no matter what.
  67         regex_t re;  // For REGEX.
  68 };
  69
  70 bool matches(const Needle &needle, const char *haystack)
  71 {
  72         if (needle.type == Needle::STRSTR) {
  73                 return strstr(haystack, needle.str.c_str()) != nullptr;
  74         } else if (needle.type == Needle::GLOB) {
  75                 int flags = ignore_case ? FNM_CASEFOLD : 0;
  76                 return fnmatch(needle.str.c_str(), haystack, flags) == 0;
  77         } else {
  78                 assert(needle.type == Needle::REGEX);
  79                 return regexec(&needle.re, haystack, /*nmatch=*/0, /*pmatch=*/nullptr, /*flags=*/0) == 0;
  80         }
  81 }
  82
  83 class Corpus {
  84 public:
  85         Corpus(int fd, IOUringEngine *engine);
  86         ~Corpus();
  87         void find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb);
  88         void get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const;
  89         size_t get_num_filename_blocks() const;
  90         off_t offset_for_block(uint32_t docid) const
  91         {
  92                 return hdr.filename_index_offset_bytes + docid * sizeof(uint64_t);
  93         }
  94         const Header &get_hdr() const { return hdr; }
  95
  96 public:
  97         const int fd;
  98         IOUringEngine *const engine;
  99
 100         Header hdr;
 101 };
 102
 103 Corpus::Corpus(int fd, IOUringEngine *engine)
 104         : fd(fd), engine(engine)
 105 {
 106         // Enable to test cold-cache behavior (except for access()).
 107         if (false) {
 108                 off_t len = lseek(fd, 0, SEEK_END);
 109                 if (len == -1) {
 110                         perror("lseek");
 111                         exit(1);
 112                 }
 113                 posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED);
 114         }
 115
 116         complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
 117         if (memcmp(hdr.magic, "\0plocate", 8) != 0) {
 118                 fprintf(stderr, "plocate.db is corrupt or an old version; please rebuild it.\n");
 119                 exit(1);
 120         }
 121         if (hdr.version != 0 && hdr.version != 1) {
 122                 fprintf(stderr, "plocate.db has version %u, expected 0 or 1; please rebuild it.\n", hdr.version);
 123                 exit(1);
 124         }
 125         if (hdr.version == 0) {
 126                 // These will be junk data.
 127                 hdr.zstd_dictionary_offset_bytes = 0;
 128                 hdr.zstd_dictionary_length_bytes = 0;
 129         }
 130 }
 131
 132 Corpus::~Corpus()
 133 {
 134         close(fd);
 135 }
 136
 137 void Corpus::find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb)
 138 {
 139         uint32_t bucket = hash_trigram(trgm, hdr.hashtable_size);
 140         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) {
 141                 const Trigram *trgmptr = reinterpret_cast<const Trigram *>(s.data());
 142                 for (unsigned i = 0; i < hdr.extra_ht_slots + 1; ++i) {
 143                         if (trgmptr[i].trgm == trgm) {
 144                                 cb(trgmptr + i, trgmptr[i + 1].offset - trgmptr[i].offset);
 145                                 return;
 146                         }
 147                 }
 148
 149                 // Not found.
 150                 cb(nullptr, 0);
 151         });
 152 }
 153
 154 void Corpus::get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const
 155 {
 156         // Read the file offset from this docid and the next one.
 157         // This is always allowed, since we have a sentinel block at the end.
 158         engine->submit_read(fd, sizeof(uint64_t) * 2, offset_for_block(docid), [this, cb{ move(cb) }](string_view s) {
 159                 const uint64_t *ptr = reinterpret_cast<const uint64_t *>(s.data());
 160                 off_t offset = ptr[0];
 161                 size_t len = ptr[1] - ptr[0];
 162                 engine->submit_read(fd, len, offset, cb);
 163         });
 164 }
 165
 166 size_t Corpus::get_num_filename_blocks() const
 167 {
 168         return hdr.num_docids;
 169 }
 170
 171 void scan_file_block(const vector<Needle> &needles, string_view compressed,
 172                      AccessRXCache *access_rx_cache, uint64_t seq, ResultReceiver *serializer,
 173                      atomic<uint64_t> *matched)
 174 {
 175         unsigned long long uncompressed_len = ZSTD_getFrameContentSize(compressed.data(), compressed.size());
 176         if (uncompressed_len == ZSTD_CONTENTSIZE_UNKNOWN || uncompressed_len == ZSTD_CONTENTSIZE_ERROR) {
 177                 fprintf(stderr, "ZSTD_getFrameContentSize() failed\n");
 178                 exit(1);
 179         }
 180
 181         string block;
 182         block.resize(uncompressed_len + 1);
 183
 184         static thread_local ZSTD_DCtx *ctx = ZSTD_createDCtx();  // Reused across calls.
 185         size_t err;
 186
 187         if (ddict != nullptr) {
 188                 err = ZSTD_decompress_usingDDict(ctx, &block[0], block.size(), compressed.data(),
 189                                                  compressed.size(), ddict);
 190         } else {
 191                 err = ZSTD_decompressDCtx(ctx, &block[0], block.size(), compressed.data(),
 192                                           compressed.size());
 193         }
 194         if (ZSTD_isError(err)) {
 195                 fprintf(stderr, "ZSTD_decompress(): %s\n", ZSTD_getErrorName(err));
 196                 exit(1);
 197         }
 198         block[block.size() - 1] = '\0';
 199
 200         auto test_candidate = [&](const char *filename, uint64_t local_seq, uint64_t next_seq) {
 201                 access_rx_cache->check_access(filename, /*allow_async=*/true, [matched, serializer, local_seq, next_seq, filename{ strdup(filename) }](bool ok) {
 202                         if (ok) {
 203                                 ++*matched;
 204                                 serializer->print(local_seq, next_seq - local_seq, filename);
 205                         } else {
 206                                 serializer->print(local_seq, next_seq - local_seq, "");
 207                         }
 208                         free(filename);
 209                 });
 210         };
 211
 212         // We need to know the next sequence number before inserting into Serializer,
 213         // so always buffer one candidate.
 214         const char *pending_candidate = nullptr;
 215
 216         uint64_t local_seq = seq << 32;
 217         for (const char *filename = block.data();
 218              filename != block.data() + block.size();
 219              filename += strlen(filename) + 1) {
 220                 bool found = true;
 221                 for (const Needle &needle : needles) {
 222                         if (!matches(needle, filename)) {
 223                                 found = false;
 224                                 break;
 225                         }
 226                 }
 227                 if (found) {
 228                         if (pending_candidate != nullptr) {
 229                                 test_candidate(pending_candidate, local_seq, local_seq + 1);
 230                                 ++local_seq;
 231                         }
 232                         pending_candidate = filename;
 233                 }
 234         }
 235         if (pending_candidate == nullptr) {
 236                 serializer->print(seq << 32, 1ULL << 32, "");
 237         } else {
 238                 test_candidate(pending_candidate, local_seq, (seq + 1) << 32);
 239         }
 240 }
 241
 242 size_t scan_docids(const vector<Needle> &needles, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
 243 {
 244         Serializer docids_in_order;
 245         AccessRXCache access_rx_cache(engine);
 246         atomic<uint64_t> matched{ 0 };
 247         for (size_t i = 0; i < docids.size(); ++i) {
 248                 uint32_t docid = docids[i];
 249                 corpus.get_compressed_filename_block(docid, [i, &matched, &needles, &access_rx_cache, &docids_in_order](string_view compressed) {
 250                         scan_file_block(needles, compressed, &access_rx_cache, i, &docids_in_order, &matched);
 251                 });
 252         }
 253         engine->finish();
 254         return matched;
 255 }
 256
 257 struct WorkerThread {
 258         thread t;
 259
 260         // We use a result queue instead of synchronizing Serializer,
 261         // since a lock on it becomes a huge choke point if there are
 262         // lots of threads.
 263         mutex result_mu;
 264         struct Result {
 265                 uint64_t seq;
 266                 uint64_t skip;
 267                 string msg;
 268         };
 269         vector<Result> results;
 270 };
 271
 272 class WorkerThreadReceiver : public ResultReceiver {
 273 public:
 274         WorkerThreadReceiver(WorkerThread *wt)
 275                 : wt(wt) {}
 276
 277         void print(uint64_t seq, uint64_t skip, const string msg) override
 278         {
 279                 lock_guard<mutex> lock(wt->result_mu);
 280                 if (msg.empty() && !wt->results.empty() && wt->results.back().seq + wt->results.back().skip == seq) {
 281                         wt->results.back().skip += skip;
 282                 } else {
 283                         wt->results.emplace_back(WorkerThread::Result{ seq, skip, move(msg) });
 284                 }
 285         }
 286
 287 private:
 288         WorkerThread *wt;
 289 };
 290
 291 void deliver_results(WorkerThread *wt, Serializer *serializer)
 292 {
 293         vector<WorkerThread::Result> results;
 294         {
 295                 lock_guard<mutex> lock(wt->result_mu);
 296                 results = move(wt->results);
 297         }
 298         for (const WorkerThread::Result &result : results) {
 299                 serializer->print(result.seq, result.skip, move(result.msg));
 300         }
 301 }
 302
 303 // We do this sequentially, as it's faster than scattering
 304 // a lot of I/O through io_uring and hoping the kernel will
 305 // coalesce it plus readahead for us. Since we assume that
 306 // we will primarily be CPU-bound, we'll be firing up one
 307 // worker thread for each spare core (the last one will
 308 // only be doing I/O). access() is still synchronous.
 309 uint64_t scan_all_docids(const vector<Needle> &needles, int fd, const Corpus &corpus)
 310 {
 311         {
 312                 const Header &hdr = corpus.get_hdr();
 313                 if (hdr.zstd_dictionary_length_bytes > 0) {
 314                         string dictionary;
 315                         dictionary.resize(hdr.zstd_dictionary_length_bytes);
 316                         complete_pread(fd, &dictionary[0], hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes);
 317                         ddict = ZSTD_createDDict(dictionary.data(), dictionary.size());
 318                 }
 319         }
 320
 321         AccessRXCache access_rx_cache(nullptr);
 322         Serializer serializer;
 323         uint32_t num_blocks = corpus.get_num_filename_blocks();
 324         unique_ptr<uint64_t[]> offsets(new uint64_t[num_blocks + 1]);
 325         complete_pread(fd, offsets.get(), (num_blocks + 1) * sizeof(uint64_t), corpus.offset_for_block(0));
 326         atomic<uint64_t> matched{ 0 };
 327
 328         mutex mu;
 329         condition_variable queue_added, queue_removed;
 330         deque<tuple<int, int, string>> work_queue;  // Under mu.
 331         bool done = false;  // Under mu.
 332
 333         unsigned num_threads = max<int>(sysconf(_SC_NPROCESSORS_ONLN) - 1, 1);
 334         dprintf("Using %u worker threads for linear scan.\n", num_threads);
 335         unique_ptr<WorkerThread[]> threads(new WorkerThread[num_threads]);
 336         for (unsigned i = 0; i < num_threads; ++i) {
 337                 threads[i].t = thread([&threads, &mu, &queue_added, &queue_removed, &work_queue, &done, &offsets, &needles, &access_rx_cache, &matched, i] {
 338                         // regcomp() takes a lock on the regex, so each thread will need its own.
 339                         const vector<Needle> *use_needles = &needles;
 340                         vector<Needle> recompiled_needles;
 341                         if (i != 0 && patterns_are_regex) {
 342                                 recompiled_needles = needles;
 343                                 for (Needle &needle : recompiled_needles) {
 344                                         needle.re = compile_regex(needle.str);
 345                                 }
 346                                 use_needles = &recompiled_needles;
 347                         }
 348
 349                         WorkerThreadReceiver receiver(&threads[i]);
 350                         for (;;) {
 351                                 uint32_t io_docid, last_docid;
 352                                 string compressed;
 353
 354                                 {
 355                                         unique_lock<mutex> lock(mu);
 356                                         queue_added.wait(lock, [&work_queue, &done] { return !work_queue.empty() || done; });
 357                                         if (done && work_queue.empty()) {
 358                                                 return;
 359                                         }
 360                                         tie(io_docid, last_docid, compressed) = move(work_queue.front());
 361                                         work_queue.pop_front();
 362                                         queue_removed.notify_all();
 363                                 }
 364
 365                                 for (uint32_t docid = io_docid; docid < last_docid; ++docid) {
 366                                         size_t relative_offset = offsets[docid] - offsets[io_docid];
 367                                         size_t len = offsets[docid + 1] - offsets[docid];
 368                                         scan_file_block(*use_needles, { &compressed[relative_offset], len }, &access_rx_cache, docid, &receiver, &matched);
 369                                 }
 370                         }
 371                 });
 372         }
 373
 374         string compressed;
 375         for (uint32_t io_docid = 0; io_docid < num_blocks; io_docid += 32) {
 376                 uint32_t last_docid = std::min(io_docid + 32, num_blocks);
 377                 size_t io_len = offsets[last_docid] - offsets[io_docid];
 378                 if (compressed.size() < io_len) {
 379                         compressed.resize(io_len);
 380                 }
 381                 complete_pread(fd, &compressed[0], io_len, offsets[io_docid]);
 382
 383                 {
 384                         unique_lock<mutex> lock(mu);
 385                         queue_removed.wait(lock, [&work_queue] { return work_queue.size() < 256; });  // Allow ~2MB of data queued up.
 386                         work_queue.emplace_back(io_docid, last_docid, move(compressed));
 387                         queue_added.notify_one();  // Avoid the thundering herd.
 388                 }
 389
 390                 // Pick up some results, so that we are sure that we won't just overload.
 391                 // (Seemingly, going through all of these causes slowness with many threads,
 392                 // but taking only one is OK.)
 393                 unsigned i = io_docid / 32;
 394                 deliver_results(&threads[i % num_threads], &serializer);
 395         }
 396         {
 397                 lock_guard<mutex> lock(mu);
 398                 done = true;
 399                 queue_added.notify_all();
 400         }
 401         for (unsigned i = 0; i < num_threads; ++i) {
 402                 threads[i].t.join();
 403                 deliver_results(&threads[i], &serializer);
 404         }
 405         return matched;
 406 }
 407
 408 // Takes the given posting list, unions it into the parts of the trigram disjunction
 409 // already read; if the list is complete, intersects with “cur_candidates”.
 410 //
 411 // Returns true if the search should be aborted (we are done).
 412 bool new_posting_list_read(TrigramDisjunction *td, vector<uint32_t> decoded, vector<uint32_t> *cur_candidates, vector<uint32_t> *tmp)
 413 {
 414         if (td->docids.empty()) {
 415                 td->docids = move(decoded);
 416         } else {
 417                 tmp->clear();
 418                 set_union(decoded.begin(), decoded.end(), td->docids.begin(), td->docids.end(), back_inserter(*tmp));
 419                 swap(*tmp, td->docids);
 420         }
 421         if (--td->remaining_trigrams_to_read > 0) {
 422                 // Need to wait for more.
 423                 if (ignore_case) {
 424                         dprintf("  ... %u reads left in OR group %u (%zu docids in list)\n",
 425                                 td->remaining_trigrams_to_read, td->index, td->docids.size());
 426                 }
 427                 return false;
 428         }
 429         if (cur_candidates->empty()) {
 430                 if (ignore_case) {
 431                         dprintf("  ... all reads done for OR group %u (%zu docids)\n",
 432                                 td->index, td->docids.size());
 433                 }
 434                 *cur_candidates = move(td->docids);
 435         } else {
 436                 tmp->clear();
 437                 set_intersection(cur_candidates->begin(), cur_candidates->end(),
 438                                  td->docids.begin(), td->docids.end(),
 439                                  back_inserter(*tmp));
 440                 swap(*cur_candidates, *tmp);
 441                 if (ignore_case) {
 442                         if (cur_candidates->empty()) {
 443                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (none left, search is done)\n",
 444                                         td->index, td->docids.size());
 445                                 return true;
 446                         } else {
 447                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (%zu left)\n",
 448                                         td->index, td->docids.size(), cur_candidates->size());
 449                         }
 450                 }
 451         }
 452         return false;
 453 }
 454
 455 void do_search_file(const vector<Needle> &needles, const char *filename)
 456 {
 457         int fd = open(filename, O_RDONLY);
 458         if (fd == -1) {
 459                 perror(filename);
 460                 exit(1);
 461         }
 462
 463         // Drop privileges.
 464         if (setgid(getgid()) != 0) {
 465                 perror("setgid");
 466                 exit(EXIT_FAILURE);
 467         }
 468
 469         start = steady_clock::now();
 470         if (access("/", R_OK | X_OK)) {
 471                 // We can't find anything, no need to bother...
 472                 return;
 473         }
 474
 475         IOUringEngine engine(/*slop_bytes=*/16);  // 16 slop bytes as described in turbopfor.h.
 476         Corpus corpus(fd, &engine);
 477         dprintf("Corpus init done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 478
 479         vector<TrigramDisjunction> trigram_groups;
 480         if (patterns_are_regex) {
 481                 // We could parse the regex to find trigrams that have to be there
 482                 // (there are actually known algorithms to deal with disjunctions
 483                 // and such, too), but for now, we just go brute force.
 484                 // Using locate with regexes is pretty niche.
 485         } else {
 486                 for (const Needle &needle : needles) {
 487                         parse_trigrams(needle.str, ignore_case, &trigram_groups);
 488                 }
 489         }
 490
 491         unique_sort(
 492                 &trigram_groups,
 493                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives < b.trigram_alternatives; },
 494                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives == b.trigram_alternatives; });
 495
 496         // Give them names for debugging.
 497         unsigned td_index = 0;
 498         for (TrigramDisjunction &td : trigram_groups) {
 499                 td.index = td_index++;
 500         }
 501
 502         // Collect which trigrams we need to look up in the hash table.
 503         unordered_map<uint32_t, vector<TrigramDisjunction *>> trigrams_to_lookup;
 504         for (TrigramDisjunction &td : trigram_groups) {
 505                 for (uint32_t trgm : td.trigram_alternatives) {
 506                         trigrams_to_lookup[trgm].push_back(&td);
 507                 }
 508         }
 509         if (trigrams_to_lookup.empty()) {
 510                 // Too short for trigram matching. Apply brute force.
 511                 // (We could have searched through all trigrams that matched
 512                 // the pattern and done a union of them, but that's a lot of
 513                 // work for fairly unclear gain.)
 514                 uint64_t matched = scan_all_docids(needles, fd, corpus);
 515                 if (only_count) {
 516                         printf("%" PRId64 "\n", matched);
 517                 }
 518                 return;
 519         }
 520
 521         // Sneak in fetching the dictionary, if present. It's not necessarily clear
 522         // exactly where it would be cheapest to get it, but it needs to be present
 523         // before we can decode any of the posting lists. Most likely, it's
 524         // in the same filesystem block as the header anyway, so it should be
 525         // present in the cache.
 526         {
 527                 const Header &hdr = corpus.get_hdr();
 528                 if (hdr.zstd_dictionary_length_bytes > 0) {
 529                         engine.submit_read(fd, hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes, [](string_view s) {
 530                                 ddict = ZSTD_createDDict(s.data(), s.size());
 531                                 dprintf("Dictionary initialized after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 532                         });
 533                 }
 534         }
 535
 536         // Look them all up on disk.
 537         for (auto &[trgm, trigram_groups] : trigrams_to_lookup) {
 538                 corpus.find_trigram(trgm, [trgm{ trgm }, trigram_groups{ &trigram_groups }](const Trigram *trgmptr, size_t len) {
 539                         if (trgmptr == nullptr) {
 540                                 dprintf("trigram %s isn't found\n", print_trigram(trgm).c_str());
 541                                 for (TrigramDisjunction *td : *trigram_groups) {
 542                                         --td->remaining_trigrams_to_read;
 543                                         if (td->remaining_trigrams_to_read == 0 && td->read_trigrams.empty()) {
 544                                                 dprintf("zero matches in %s, so we are done\n", print_td(*td).c_str());
 545                                                 if (only_count) {
 546                                                         printf("0\n");
 547                                                 }
 548                                                 exit(0);
 549                                         }
 550                                 }
 551                                 return;
 552                         }
 553                         for (TrigramDisjunction *td : *trigram_groups) {
 554                                 --td->remaining_trigrams_to_read;
 555                                 td->max_num_docids += trgmptr->num_docids;
 556                                 td->read_trigrams.emplace_back(*trgmptr, len);
 557                         }
 558                 });
 559         }
 560         engine.finish();
 561         dprintf("Hashtable lookups done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 562
 563         for (TrigramDisjunction &td : trigram_groups) {
 564                 // Reset for reads.
 565                 td.remaining_trigrams_to_read = td.read_trigrams.size();
 566
 567                 if (ignore_case) {  // If case-sensitive, they'll all be pretty obvious single-entry groups.
 568                         dprintf("OR group %u (max_num_docids=%u): %s\n", td.index, td.max_num_docids, print_td(td).c_str());
 569                 }
 570         }
 571
 572         // TODO: For case-insensitive (ie. more than one alternative in each),
 573         // prioritize the ones with fewer seeks?
 574         sort(trigram_groups.begin(), trigram_groups.end(),
 575              [&](const TrigramDisjunction &a, const TrigramDisjunction &b) {
 576                      return a.max_num_docids < b.max_num_docids;
 577              });
 578
 579         unordered_map<uint32_t, vector<TrigramDisjunction *>> uses_trigram;
 580         for (TrigramDisjunction &td : trigram_groups) {
 581                 for (uint32_t trgm : td.trigram_alternatives) {
 582                         uses_trigram[trgm].push_back(&td);
 583                 }
 584         }
 585
 586         unordered_set<uint32_t> trigrams_submitted_read;
 587         vector<uint32_t> cur_candidates, tmp, decoded;
 588         bool done = false;
 589         for (TrigramDisjunction &td : trigram_groups) {
 590                 if (!cur_candidates.empty() && td.max_num_docids > cur_candidates.size() * 100) {
 591                         dprintf("%s has up to %u entries, ignoring the rest (will "
 592                                 "weed out false positives later)\n",
 593                                 print_td(td).c_str(), td.max_num_docids);
 594                         break;
 595                 }
 596
 597                 for (auto &[trgmptr, len] : td.read_trigrams) {
 598                         if (trigrams_submitted_read.count(trgmptr.trgm) != 0) {
 599                                 continue;
 600                         }
 601                         trigrams_submitted_read.insert(trgmptr.trgm);
 602                         // Only stay a certain amount ahead, so that we don't spend I/O
 603                         // on reading the latter, large posting lists. We are unlikely
 604                         // to need them anyway, even if they should come in first.
 605                         if (engine.get_waiting_reads() >= 5) {
 606                                 engine.finish();
 607                                 if (done)
 608                                         break;
 609                         }
 610                         engine.submit_read(fd, len, trgmptr.offset, [trgmptr{ trgmptr }, len{ len }, &done, &cur_candidates, &tmp, &decoded, &uses_trigram](string_view s) {
 611                                 if (done)
 612                                         return;
 613
 614                                 uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
 615                                 const unsigned char *pldata = reinterpret_cast<const unsigned char *>(s.data());
 616                                 size_t num = trgmptr.num_docids;
 617                                 decoded.resize(num);
 618                                 decode_pfor_delta1_128(pldata, num, /*interleaved=*/true, &decoded[0]);
 619
 620                                 assert(uses_trigram.count(trgm) != 0);
 621                                 bool was_empty = cur_candidates.empty();
 622                                 if (ignore_case) {
 623                                         dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 624                                 }
 625
 626                                 for (TrigramDisjunction *td : uses_trigram[trgm]) {
 627                                         done |= new_posting_list_read(td, decoded, &cur_candidates, &tmp);
 628                                         if (done)
 629                                                 break;
 630                                 }
 631                                 if (!ignore_case) {
 632                                         if (was_empty) {
 633                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 634                                         } else if (cur_candidates.empty()) {
 635                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (none left, search is done)\n", print_trigram(trgm).c_str(), len, num);
 636                                         } else {
 637                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (%zu left)\n", print_trigram(trgm).c_str(), len, num, cur_candidates.size());
 638                                         }
 639                                 }
 640                         });
 641                 }
 642         }
 643         engine.finish();
 644         if (done) {
 645                 return;
 646         }
 647         dprintf("Intersection done after %.1f ms. Doing final verification and printing:\n",
 648                 1e3 * duration<float>(steady_clock::now() - start).count());
 649
 650         uint64_t matched = scan_docids(needles, cur_candidates, corpus, &engine);
 651         dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
 652                 1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 653
 654         if (only_count) {
 655                 printf("%" PRId64 "\n", matched);
 656         }
 657 }
 658
 659 string unescape_glob_to_plain_string(const string &needle)
 660 {
 661         string unescaped;
 662         for (size_t i = 0; i < needle.size(); i += read_unigram(needle, i).second) {
 663                 uint32_t ch = read_unigram(needle, i).first;
 664                 assert(ch != WILDCARD_UNIGRAM);
 665                 if (ch == PREMATURE_END_UNIGRAM) {
 666                         fprintf(stderr, "Pattern '%s' ended prematurely\n", needle.c_str());
 667                         exit(1);
 668                 }
 669                 unescaped.push_back(ch);
 670         }
 671         return unescaped;
 672 }
 673
 674 regex_t compile_regex(const string &needle)
 675 {
 676         regex_t re;
 677         int flags = REG_NOSUB;
 678         if (ignore_case) {
 679                 flags |= REG_ICASE;
 680         }
 681         if (use_extended_regex) {
 682                 flags |= REG_EXTENDED;
 683         }
 684         int err = regcomp(&re, needle.c_str(), flags);
 685         if (err != 0) {
 686                 char errbuf[256];
 687                 regerror(err, &re, errbuf, sizeof(errbuf));
 688                 fprintf(stderr, "Error when compiling regex '%s': %s\n", needle.c_str(), errbuf);
 689                 exit(1);
 690         }
 691         return re;
 692 }
 693
 694 void usage()
 695 {
 696         printf(
 697                 "Usage: plocate [OPTION]... PATTERN...\n"
 698                 "\n"
 699                 "  -c, --count            print number of matches instead of the matches\n"
 700                 "  -d, --database DBPATH  search for files in DBPATH\n"
 701                 "                         (default is " DEFAULT_DBPATH ")\n"
 702                 "  -i, --ignore-case      search case-insensitively\n"
 703                 "  -l, --limit LIMIT      stop after LIMIT matches\n"
 704                 "  -0, --null             delimit matches by NUL instead of newline\n"
 705                 "  -r, --regexp           interpret patterns as basic regexps (slow)\n"
 706                 "      --regex            interpret patterns as extended regexps (slow)\n"
 707                 "      --help             print this help\n"
 708                 "      --version          print version information\n");
 709 }
 710
 711 void version()
 712 {
 713         printf("plocate %s\n", PLOCATE_VERSION);
 714         printf("Copyright 2020 Steinar H. Gunderson\n");
 715         printf("License GPLv2+: GNU GPL version 2 or later <https://gnu.org/licenses/gpl.html>.\n");
 716         printf("This is free software: you are free to change and redistribute it.\n");
 717         printf("There is NO WARRANTY, to the extent permitted by law.\n");
 718         exit(0);
 719 }
 720
 721 int main(int argc, char **argv)
 722 {
 723         constexpr int EXTENDED_REGEX = 1000;
 724         static const struct option long_options[] = {
 725                 { "help", no_argument, 0, 'h' },
 726                 { "count", no_argument, 0, 'c' },
 727                 { "database", required_argument, 0, 'd' },
 728                 { "ignore-case", no_argument, 0, 'i' },
 729                 { "limit", required_argument, 0, 'l' },
 730                 { "null", no_argument, 0, '0' },
 731                 { "version", no_argument, 0, 'V' },
 732                 { "regexp", no_argument, 0, 'r' },
 733                 { "regex", no_argument, 0, EXTENDED_REGEX },
 734                 { "debug", no_argument, 0, 'D' },  // Not documented.
 735                 { 0, 0, 0, 0 }
 736         };
 737
 738         setlocale(LC_ALL, "");
 739         for (;;) {
 740                 int option_index = 0;
 741                 int c = getopt_long(argc, argv, "cd:hil:n:0VD", long_options, &option_index);
 742                 if (c == -1) {
 743                         break;
 744                 }
 745                 switch (c) {
 746                 case 'c':
 747                         only_count = true;
 748                         break;
 749                 case 'd':
 750                         dbpath = strdup(optarg);
 751                         break;
 752                 case 'h':
 753                         usage();
 754                         exit(0);
 755                 case 'i':
 756                         ignore_case = true;
 757                         break;
 758                 case 'l':
 759                 case 'n':
 760                         limit_matches = limit_left = atoll(optarg);
 761                         if (limit_matches <= 0) {
 762                                 fprintf(stderr, "Error: limit must be a strictly positive number.\n");
 763                                 exit(1);
 764                         }
 765                         break;
 766                 case '0':
 767                         print_nul = true;
 768                         break;
 769                 case 'r':
 770                         patterns_are_regex = true;
 771                         break;
 772                 case EXTENDED_REGEX:
 773                         patterns_are_regex = true;
 774                         use_extended_regex = true;
 775                         break;
 776                 case 'D':
 777                         use_debug = true;
 778                         break;
 779                 case 'V':
 780                         version();
 781                         break;
 782                 default:
 783                         exit(1);
 784                 }
 785         }
 786
 787         if (use_debug) {
 788                 // Debug information would leak information about which files exist,
 789                 // so drop setgid before we open the file; one would either need to run
 790                 // as root, or use a locally-built file.
 791                 if (setgid(getgid()) != 0) {
 792                         perror("setgid");
 793                         exit(EXIT_FAILURE);
 794                 }
 795         }
 796
 797         vector<Needle> needles;
 798         for (int i = optind; i < argc; ++i) {
 799                 Needle needle;
 800                 needle.str = argv[i];
 801
 802                 // See if there are any wildcard characters, which indicates we should treat it
 803                 // as an (anchored) glob.
 804                 bool any_wildcard = false;
 805                 for (size_t i = 0; i < needle.str.size(); i += read_unigram(needle.str, i).second) {
 806                         if (read_unigram(needle.str, i).first == WILDCARD_UNIGRAM) {
 807                                 any_wildcard = true;
 808                                 break;
 809                         }
 810                 }
 811
 812                 if (patterns_are_regex) {
 813                         needle.type = Needle::REGEX;
 814                         needle.re = compile_regex(needle.str);
 815                 } else if (any_wildcard) {
 816                         needle.type = Needle::GLOB;
 817                 } else if (ignore_case) {
 818                         // strcasestr() doesn't handle locales correctly (even though LSB
 819                         // claims it should), but somehow, fnmatch() does, and it's about
 820                         // the same speed as using a regex.
 821                         needle.type = Needle::GLOB;
 822                         needle.str = "*" + needle.str + "*";
 823                 } else {
 824                         needle.type = Needle::STRSTR;
 825                         needle.str = unescape_glob_to_plain_string(needle.str);
 826                 }
 827                 needles.push_back(move(needle));
 828         }
 829         if (needles.empty()) {
 830                 fprintf(stderr, "plocate: no pattern to search for specified\n");
 831                 exit(0);
 832         }
 833         do_search_file(needles, dbpath);
 834 }