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 <sys/stat.h>
  35 #include <sys/types.h>
  36 #include <sys/wait.h>
  37 #include <thread>
  38 #include <tuple>
  39 #include <unistd.h>
  40 #include <unordered_map>
  41 #include <unordered_set>
  42 #include <utility>
  43 #include <vector>
  44 #include <zstd.h>
  45
  46 using namespace std;
  47 using namespace std::chrono;
  48
  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 bool check_existence = false;
  58 int64_t limit_matches = numeric_limits<int64_t>::max();
  59 int64_t limit_left = numeric_limits<int64_t>::max();
  60 bool stdout_is_tty = false;
  61 static bool in_forked_child = false;
  62
  63 steady_clock::time_point start;
  64 ZSTD_DDict *ddict = nullptr;
  65
  66 class Corpus {
  67 public:
  68         Corpus(int fd, IOUringEngine *engine);
  69         ~Corpus();
  70         void find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb);
  71         void get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const;
  72         size_t get_num_filename_blocks() const;
  73         off_t offset_for_block(uint32_t docid) const
  74         {
  75                 return hdr.filename_index_offset_bytes + docid * sizeof(uint64_t);
  76         }
  77         const Header &get_hdr() const { return hdr; }
  78
  79 public:
  80         const int fd;
  81         IOUringEngine *const engine;
  82
  83         Header hdr;
  84 };
  85
  86 Corpus::Corpus(int fd, IOUringEngine *engine)
  87         : fd(fd), engine(engine)
  88 {
  89         if (flush_cache) {
  90                 off_t len = lseek(fd, 0, SEEK_END);
  91                 if (len == -1) {
  92                         perror("lseek");
  93                         exit(1);
  94                 }
  95                 posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED);
  96         }
  97
  98         complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
  99         if (memcmp(hdr.magic, "\0plocate", 8) != 0) {
 100                 fprintf(stderr, "plocate.db is corrupt or an old version; please rebuild it.\n");
 101                 exit(1);
 102         }
 103         if (hdr.version != 0 && hdr.version != 1) {
 104                 fprintf(stderr, "plocate.db has version %u, expected 0 or 1; please rebuild it.\n", hdr.version);
 105                 exit(1);
 106         }
 107         if (hdr.version == 0) {
 108                 // These will be junk data.
 109                 hdr.zstd_dictionary_offset_bytes = 0;
 110                 hdr.zstd_dictionary_length_bytes = 0;
 111         }
 112         if (hdr.max_version < 2) {
 113                 // This too. (We ignore the other max_version 2 fields.)
 114                 hdr.check_visibility = true;
 115         }
 116 }
 117
 118 Corpus::~Corpus()
 119 {
 120         close(fd);
 121 }
 122
 123 void Corpus::find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb)
 124 {
 125         uint32_t bucket = hash_trigram(trgm, hdr.hashtable_size);
 126         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) {
 127                 const Trigram *trgmptr = reinterpret_cast<const Trigram *>(s.data());
 128                 for (unsigned i = 0; i < hdr.extra_ht_slots + 1; ++i) {
 129                         if (trgmptr[i].trgm == trgm) {
 130                                 cb(trgmptr + i, trgmptr[i + 1].offset - trgmptr[i].offset);
 131                                 return;
 132                         }
 133                 }
 134
 135                 // Not found.
 136                 cb(nullptr, 0);
 137         });
 138 }
 139
 140 void Corpus::get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const
 141 {
 142         // Read the file offset from this docid and the next one.
 143         // This is always allowed, since we have a sentinel block at the end.
 144         engine->submit_read(fd, sizeof(uint64_t) * 2, offset_for_block(docid), [this, cb{ move(cb) }](string_view s) {
 145                 const uint64_t *ptr = reinterpret_cast<const uint64_t *>(s.data());
 146                 off_t offset = ptr[0];
 147                 size_t len = ptr[1] - ptr[0];
 148                 engine->submit_read(fd, len, offset, cb);
 149         });
 150 }
 151
 152 size_t Corpus::get_num_filename_blocks() const
 153 {
 154         return hdr.num_docids;
 155 }
 156
 157 template<class T>
 158 void stat_if_needed(const char *filename, bool access_ok, IOUringEngine *engine, T cb)
 159 {
 160         if (!access_ok || !check_existence) {
 161                 // Doesn't have access or doesn't care about existence, so no need to stat.
 162                 cb(access_ok);
 163         } else if (engine == nullptr || !engine->get_supports_stat()) {
 164                 // Do a synchronous stat.
 165                 struct stat buf;
 166                 bool ok = lstat(filename, &buf) == 0;
 167                 cb(ok);
 168         } else {
 169                 engine->submit_stat(filename, cb);
 170         }
 171 }
 172
 173 void scan_file_block(const vector<Needle> &needles, string_view compressed,
 174                      IOUringEngine *engine, AccessRXCache *access_rx_cache, uint64_t seq, ResultReceiver *serializer,
 175                      atomic<uint64_t> *matched)
 176 {
 177         unsigned long long uncompressed_len = ZSTD_getFrameContentSize(compressed.data(), compressed.size());
 178         if (uncompressed_len == ZSTD_CONTENTSIZE_UNKNOWN || uncompressed_len == ZSTD_CONTENTSIZE_ERROR) {
 179                 fprintf(stderr, "ZSTD_getFrameContentSize() failed\n");
 180                 exit(1);
 181         }
 182
 183         string block;
 184         block.resize(uncompressed_len + 1);
 185
 186         static thread_local ZSTD_DCtx *ctx = ZSTD_createDCtx();  // Reused across calls.
 187         size_t err;
 188
 189         if (ddict != nullptr) {
 190                 err = ZSTD_decompress_usingDDict(ctx, &block[0], block.size(), compressed.data(),
 191                                                  compressed.size(), ddict);
 192         } else {
 193                 err = ZSTD_decompressDCtx(ctx, &block[0], block.size(), compressed.data(),
 194                                           compressed.size());
 195         }
 196         if (ZSTD_isError(err)) {
 197                 fprintf(stderr, "ZSTD_decompress(): %s\n", ZSTD_getErrorName(err));
 198                 exit(1);
 199         }
 200         block[block.size() - 1] = '\0';
 201
 202         auto test_candidate = [&](const char *filename, uint64_t local_seq, uint64_t next_seq) {
 203                 access_rx_cache->check_access(filename, /*allow_async=*/true, [matched, engine, serializer, local_seq, next_seq, filename{ strdup(filename) }](bool ok) {
 204                         stat_if_needed(filename, ok, engine, [matched, serializer, local_seq, next_seq, filename](bool ok) {
 205                                 if (ok) {
 206                                         ++*matched;
 207                                         serializer->print(local_seq, next_seq - local_seq, filename);
 208                                 } else {
 209                                         serializer->print(local_seq, next_seq - local_seq, "");
 210                                 }
 211                                 free(filename);
 212                         });
 213                 });
 214         };
 215
 216         // We need to know the next sequence number before inserting into Serializer,
 217         // so always buffer one candidate.
 218         const char *pending_candidate = nullptr;
 219
 220         uint64_t local_seq = seq << 32;
 221         for (const char *filename = block.data();
 222              filename != block.data() + block.size();
 223              filename += strlen(filename) + 1) {
 224                 const char *haystack = filename;
 225                 if (match_basename) {
 226                         haystack = strrchr(filename, '/');
 227                         if (haystack == nullptr) {
 228                                 haystack = filename;
 229                         } else {
 230                                 ++haystack;
 231                         }
 232                 }
 233
 234                 bool found = true;
 235                 for (const Needle &needle : needles) {
 236                         if (!matches(needle, haystack)) {
 237                                 found = false;
 238                                 break;
 239                         }
 240                 }
 241                 if (found) {
 242                         if (pending_candidate != nullptr) {
 243                                 test_candidate(pending_candidate, local_seq, local_seq + 1);
 244                                 ++local_seq;
 245                         }
 246                         pending_candidate = filename;
 247                 }
 248         }
 249         if (pending_candidate == nullptr) {
 250                 serializer->print(seq << 32, 1ULL << 32, "");
 251         } else {
 252                 test_candidate(pending_candidate, local_seq, (seq + 1) << 32);
 253         }
 254 }
 255
 256 size_t scan_docids(const vector<Needle> &needles, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
 257 {
 258         Serializer docids_in_order;
 259         AccessRXCache access_rx_cache(engine, corpus.get_hdr().check_visibility);
 260         atomic<uint64_t> matched{ 0 };
 261         for (size_t i = 0; i < docids.size(); ++i) {
 262                 uint32_t docid = docids[i];
 263                 corpus.get_compressed_filename_block(docid, [i, &matched, &needles, &access_rx_cache, engine, &docids_in_order](string_view compressed) {
 264                         scan_file_block(needles, compressed, engine, &access_rx_cache, i, &docids_in_order, &matched);
 265                 });
 266         }
 267         engine->finish();
 268         return matched;
 269 }
 270
 271 struct WorkerThread {
 272         thread t;
 273
 274         // We use a result queue instead of synchronizing Serializer,
 275         // since a lock on it becomes a huge choke point if there are
 276         // lots of threads.
 277         mutex result_mu;
 278         struct Result {
 279                 uint64_t seq;
 280                 uint64_t skip;
 281                 string msg;
 282         };
 283         vector<Result> results;
 284 };
 285
 286 class WorkerThreadReceiver : public ResultReceiver {
 287 public:
 288         WorkerThreadReceiver(WorkerThread *wt)
 289                 : wt(wt) {}
 290
 291         void print(uint64_t seq, uint64_t skip, const string msg) override
 292         {
 293                 lock_guard<mutex> lock(wt->result_mu);
 294                 if (msg.empty() && !wt->results.empty() && wt->results.back().seq + wt->results.back().skip == seq) {
 295                         wt->results.back().skip += skip;
 296                 } else {
 297                         wt->results.emplace_back(WorkerThread::Result{ seq, skip, move(msg) });
 298                 }
 299         }
 300
 301 private:
 302         WorkerThread *wt;
 303 };
 304
 305 void deliver_results(WorkerThread *wt, Serializer *serializer)
 306 {
 307         vector<WorkerThread::Result> results;
 308         {
 309                 lock_guard<mutex> lock(wt->result_mu);
 310                 results = move(wt->results);
 311         }
 312         for (const WorkerThread::Result &result : results) {
 313                 serializer->print(result.seq, result.skip, move(result.msg));
 314         }
 315 }
 316
 317 // We do this sequentially, as it's faster than scattering
 318 // a lot of I/O through io_uring and hoping the kernel will
 319 // coalesce it plus readahead for us. Since we assume that
 320 // we will primarily be CPU-bound, we'll be firing up one
 321 // worker thread for each spare core (the last one will
 322 // only be doing I/O). access() is still synchronous.
 323 uint64_t scan_all_docids(const vector<Needle> &needles, int fd, const Corpus &corpus)
 324 {
 325         {
 326                 const Header &hdr = corpus.get_hdr();
 327                 if (hdr.zstd_dictionary_length_bytes > 0) {
 328                         string dictionary;
 329                         dictionary.resize(hdr.zstd_dictionary_length_bytes);
 330                         complete_pread(fd, &dictionary[0], hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes);
 331                         ddict = ZSTD_createDDict(dictionary.data(), dictionary.size());
 332                 }
 333         }
 334
 335         AccessRXCache access_rx_cache(nullptr, corpus.get_hdr().check_visibility);
 336         Serializer serializer;
 337         uint32_t num_blocks = corpus.get_num_filename_blocks();
 338         unique_ptr<uint64_t[]> offsets(new uint64_t[num_blocks + 1]);
 339         complete_pread(fd, offsets.get(), (num_blocks + 1) * sizeof(uint64_t), corpus.offset_for_block(0));
 340         atomic<uint64_t> matched{ 0 };
 341
 342         mutex mu;
 343         condition_variable queue_added, queue_removed;
 344         deque<tuple<int, int, string>> work_queue;  // Under mu.
 345         bool done = false;  // Under mu.
 346
 347         unsigned num_threads = max<int>(sysconf(_SC_NPROCESSORS_ONLN) - 1, 1);
 348         dprintf("Using %u worker threads for linear scan.\n", num_threads);
 349         unique_ptr<WorkerThread[]> threads(new WorkerThread[num_threads]);
 350         for (unsigned i = 0; i < num_threads; ++i) {
 351                 threads[i].t = thread([&threads, &mu, &queue_added, &queue_removed, &work_queue, &done, &offsets, &needles, &access_rx_cache, engine{ corpus.engine }, &matched, i] {
 352                         // regcomp() takes a lock on the regex, so each thread will need its own.
 353                         const vector<Needle> *use_needles = &needles;
 354                         vector<Needle> recompiled_needles;
 355                         if (i != 0 && patterns_are_regex) {
 356                                 recompiled_needles = needles;
 357                                 for (Needle &needle : recompiled_needles) {
 358                                         needle.re = compile_regex(needle.str);
 359                                 }
 360                                 use_needles = &recompiled_needles;
 361                         }
 362
 363                         WorkerThreadReceiver receiver(&threads[i]);
 364                         for (;;) {
 365                                 uint32_t io_docid, last_docid;
 366                                 string compressed;
 367
 368                                 {
 369                                         unique_lock<mutex> lock(mu);
 370                                         queue_added.wait(lock, [&work_queue, &done] { return !work_queue.empty() || done; });
 371                                         if (done && work_queue.empty()) {
 372                                                 return;
 373                                         }
 374                                         tie(io_docid, last_docid, compressed) = move(work_queue.front());
 375                                         work_queue.pop_front();
 376                                         queue_removed.notify_all();
 377                                 }
 378
 379                                 for (uint32_t docid = io_docid; docid < last_docid; ++docid) {
 380                                         size_t relative_offset = offsets[docid] - offsets[io_docid];
 381                                         size_t len = offsets[docid + 1] - offsets[docid];
 382                                         scan_file_block(*use_needles, { &compressed[relative_offset], len }, engine, &access_rx_cache, docid, &receiver, &matched);
 383                                 }
 384                         }
 385                 });
 386         }
 387
 388         string compressed;
 389         for (uint32_t io_docid = 0; io_docid < num_blocks; io_docid += 32) {
 390                 uint32_t last_docid = std::min(io_docid + 32, num_blocks);
 391                 size_t io_len = offsets[last_docid] - offsets[io_docid];
 392                 if (compressed.size() < io_len) {
 393                         compressed.resize(io_len);
 394                 }
 395                 complete_pread(fd, &compressed[0], io_len, offsets[io_docid]);
 396
 397                 {
 398                         unique_lock<mutex> lock(mu);
 399                         queue_removed.wait(lock, [&work_queue] { return work_queue.size() < 256; });  // Allow ~2MB of data queued up.
 400                         work_queue.emplace_back(io_docid, last_docid, move(compressed));
 401                         queue_added.notify_one();  // Avoid the thundering herd.
 402                 }
 403
 404                 // Pick up some results, so that we are sure that we won't just overload.
 405                 // (Seemingly, going through all of these causes slowness with many threads,
 406                 // but taking only one is OK.)
 407                 unsigned i = io_docid / 32;
 408                 deliver_results(&threads[i % num_threads], &serializer);
 409         }
 410         {
 411                 lock_guard<mutex> lock(mu);
 412                 done = true;
 413                 queue_added.notify_all();
 414         }
 415         for (unsigned i = 0; i < num_threads; ++i) {
 416                 threads[i].t.join();
 417                 deliver_results(&threads[i], &serializer);
 418         }
 419         return matched;
 420 }
 421
 422 // Takes the given posting list, unions it into the parts of the trigram disjunction
 423 // already read; if the list is complete, intersects with “cur_candidates”.
 424 //
 425 // Returns true if the search should be aborted (we are done).
 426 bool new_posting_list_read(TrigramDisjunction *td, vector<uint32_t> decoded, vector<uint32_t> *cur_candidates, vector<uint32_t> *tmp)
 427 {
 428         if (td->docids.empty()) {
 429                 td->docids = move(decoded);
 430         } else {
 431                 tmp->clear();
 432                 set_union(decoded.begin(), decoded.end(), td->docids.begin(), td->docids.end(), back_inserter(*tmp));
 433                 swap(*tmp, td->docids);
 434         }
 435         if (--td->remaining_trigrams_to_read > 0) {
 436                 // Need to wait for more.
 437                 if (ignore_case) {
 438                         dprintf("  ... %u reads left in OR group %u (%zu docids in list)\n",
 439                                 td->remaining_trigrams_to_read, td->index, td->docids.size());
 440                 }
 441                 return false;
 442         }
 443         if (cur_candidates->empty()) {
 444                 if (ignore_case) {
 445                         dprintf("  ... all reads done for OR group %u (%zu docids)\n",
 446                                 td->index, td->docids.size());
 447                 }
 448                 *cur_candidates = move(td->docids);
 449         } else {
 450                 tmp->clear();
 451                 set_intersection(cur_candidates->begin(), cur_candidates->end(),
 452                                  td->docids.begin(), td->docids.end(),
 453                                  back_inserter(*tmp));
 454                 swap(*cur_candidates, *tmp);
 455                 if (ignore_case) {
 456                         if (cur_candidates->empty()) {
 457                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (none left, search is done)\n",
 458                                         td->index, td->docids.size());
 459                                 return true;
 460                         } else {
 461                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (%zu left)\n",
 462                                         td->index, td->docids.size(), cur_candidates->size());
 463                         }
 464                 }
 465         }
 466         return false;
 467 }
 468
 469 uint64_t do_search_file(const vector<Needle> &needles, const std::string &filename)
 470 {
 471         int fd = open(filename.c_str(), O_RDONLY);
 472         if (fd == -1) {
 473                 perror(filename.c_str());
 474                 exit(1);
 475         }
 476
 477         // Drop privileges.
 478         if (setgid(getgid()) != 0) {
 479                 perror("setgid");
 480                 exit(EXIT_FAILURE);
 481         }
 482
 483         start = steady_clock::now();
 484         if (access("/", R_OK | X_OK)) {
 485                 // We can't find anything, no need to bother...
 486                 return 0;
 487         }
 488
 489         IOUringEngine engine(/*slop_bytes=*/16);  // 16 slop bytes as described in turbopfor.h.
 490         Corpus corpus(fd, &engine);
 491         dprintf("Corpus init done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 492
 493         vector<TrigramDisjunction> trigram_groups;
 494         if (patterns_are_regex) {
 495                 // We could parse the regex to find trigrams that have to be there
 496                 // (there are actually known algorithms to deal with disjunctions
 497                 // and such, too), but for now, we just go brute force.
 498                 // Using locate with regexes is pretty niche.
 499         } else {
 500                 for (const Needle &needle : needles) {
 501                         parse_trigrams(needle.str, ignore_case, &trigram_groups);
 502                 }
 503         }
 504
 505         unique_sort(
 506                 &trigram_groups,
 507                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives < b.trigram_alternatives; },
 508                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives == b.trigram_alternatives; });
 509
 510         // Give them names for debugging.
 511         unsigned td_index = 0;
 512         for (TrigramDisjunction &td : trigram_groups) {
 513                 td.index = td_index++;
 514         }
 515
 516         // Collect which trigrams we need to look up in the hash table.
 517         unordered_map<uint32_t, vector<TrigramDisjunction *>> trigrams_to_lookup;
 518         for (TrigramDisjunction &td : trigram_groups) {
 519                 for (uint32_t trgm : td.trigram_alternatives) {
 520                         trigrams_to_lookup[trgm].push_back(&td);
 521                 }
 522         }
 523         if (trigrams_to_lookup.empty()) {
 524                 // Too short for trigram matching. Apply brute force.
 525                 // (We could have searched through all trigrams that matched
 526                 // the pattern and done a union of them, but that's a lot of
 527                 // work for fairly unclear gain.)
 528                 uint64_t matched = scan_all_docids(needles, fd, corpus);
 529                 dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
 530                         1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 531                 return matched;
 532         }
 533
 534         // Sneak in fetching the dictionary, if present. It's not necessarily clear
 535         // exactly where it would be cheapest to get it, but it needs to be present
 536         // before we can decode any of the posting lists. Most likely, it's
 537         // in the same filesystem block as the header anyway, so it should be
 538         // present in the cache.
 539         {
 540                 const Header &hdr = corpus.get_hdr();
 541                 if (hdr.zstd_dictionary_length_bytes > 0) {
 542                         engine.submit_read(fd, hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes, [](string_view s) {
 543                                 ddict = ZSTD_createDDict(s.data(), s.size());
 544                                 dprintf("Dictionary initialized after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 545                         });
 546                 }
 547         }
 548
 549         // Look them all up on disk.
 550         bool should_early_exit = false;
 551         for (auto &[trgm, trigram_groups] : trigrams_to_lookup) {
 552                 corpus.find_trigram(trgm, [trgm{ trgm }, trigram_groups{ &trigram_groups }, &should_early_exit](const Trigram *trgmptr, size_t len) {
 553                         if (trgmptr == nullptr) {
 554                                 dprintf("trigram %s isn't found\n", print_trigram(trgm).c_str());
 555                                 for (TrigramDisjunction *td : *trigram_groups) {
 556                                         --td->remaining_trigrams_to_read;
 557
 558                                         // If we now know this trigram group doesn't match anything at all,
 559                                         // we can do early exit; however, if we're in a forked child,
 560                                         // that would confuse the parent process (since we don't write
 561                                         // our count to the pipe), so we wait until we're back in to the
 562                                         // regular (non-async) context. This is a fairly rare case anyway,
 563                                         // and the gains from dropping the remaining trigram reads are limited.
 564                                         if (td->remaining_trigrams_to_read == 0 && td->read_trigrams.empty()) {
 565                                                 if (in_forked_child) {
 566                                                         should_early_exit = true;
 567                                                 } else {
 568                                                         dprintf("zero matches in %s, so we are done\n", print_td(*td).c_str());
 569                                                         if (only_count) {
 570                                                                 printf("0\n");
 571                                                         }
 572                                                         exit(0);
 573                                                 }
 574                                         }
 575                                 }
 576                                 return;
 577                         }
 578                         for (TrigramDisjunction *td : *trigram_groups) {
 579                                 --td->remaining_trigrams_to_read;
 580                                 td->max_num_docids += trgmptr->num_docids;
 581                                 td->read_trigrams.emplace_back(*trgmptr, len);
 582                         }
 583                 });
 584         }
 585         engine.finish();
 586         dprintf("Hashtable lookups done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 587
 588         if (should_early_exit) {
 589                 return 0;
 590         }
 591
 592         for (TrigramDisjunction &td : trigram_groups) {
 593                 // Reset for reads.
 594                 td.remaining_trigrams_to_read = td.read_trigrams.size();
 595
 596                 if (ignore_case) {  // If case-sensitive, they'll all be pretty obvious single-entry groups.
 597                         dprintf("OR group %u (max_num_docids=%u): %s\n", td.index, td.max_num_docids, print_td(td).c_str());
 598                 }
 599         }
 600
 601         // TODO: For case-insensitive (ie. more than one alternative in each),
 602         // prioritize the ones with fewer seeks?
 603         sort(trigram_groups.begin(), trigram_groups.end(),
 604              [&](const TrigramDisjunction &a, const TrigramDisjunction &b) {
 605                      return a.max_num_docids < b.max_num_docids;
 606              });
 607
 608         unordered_map<uint32_t, vector<TrigramDisjunction *>> uses_trigram;
 609         for (TrigramDisjunction &td : trigram_groups) {
 610                 for (uint32_t trgm : td.trigram_alternatives) {
 611                         uses_trigram[trgm].push_back(&td);
 612                 }
 613         }
 614
 615         unordered_set<uint32_t> trigrams_submitted_read;
 616         vector<uint32_t> cur_candidates, tmp, decoded;
 617         bool done = false;
 618         for (TrigramDisjunction &td : trigram_groups) {
 619                 if (!cur_candidates.empty() && td.max_num_docids > cur_candidates.size() * 100) {
 620                         dprintf("%s has up to %u entries, ignoring the rest (will "
 621                                 "weed out false positives later)\n",
 622                                 print_td(td).c_str(), td.max_num_docids);
 623                         break;
 624                 }
 625
 626                 for (auto &[trgmptr, len] : td.read_trigrams) {
 627                         if (trigrams_submitted_read.count(trgmptr.trgm) != 0) {
 628                                 continue;
 629                         }
 630                         trigrams_submitted_read.insert(trgmptr.trgm);
 631                         // Only stay a certain amount ahead, so that we don't spend I/O
 632                         // on reading the latter, large posting lists. We are unlikely
 633                         // to need them anyway, even if they should come in first.
 634                         if (engine.get_waiting_reads() >= 5) {
 635                                 engine.finish();
 636                                 if (done)
 637                                         break;
 638                         }
 639                         engine.submit_read(fd, len, trgmptr.offset, [trgmptr{ trgmptr }, len{ len }, &done, &cur_candidates, &tmp, &decoded, &uses_trigram](string_view s) {
 640                                 if (done)
 641                                         return;
 642
 643                                 uint32_t trgm = trgmptr.trgm;
 644                                 const unsigned char *pldata = reinterpret_cast<const unsigned char *>(s.data());
 645                                 size_t num = trgmptr.num_docids;
 646                                 decoded.resize(num);
 647                                 decode_pfor_delta1_128(pldata, num, /*interleaved=*/true, &decoded[0]);
 648
 649                                 assert(uses_trigram.count(trgm) != 0);
 650                                 bool was_empty = cur_candidates.empty();
 651                                 if (ignore_case) {
 652                                         dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 653                                 }
 654
 655                                 for (TrigramDisjunction *td : uses_trigram[trgm]) {
 656                                         done |= new_posting_list_read(td, decoded, &cur_candidates, &tmp);
 657                                         if (done)
 658                                                 break;
 659                                 }
 660                                 if (!ignore_case) {
 661                                         if (was_empty) {
 662                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 663                                         } else if (cur_candidates.empty()) {
 664                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (none left, search is done)\n", print_trigram(trgm).c_str(), len, num);
 665                                         } else {
 666                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (%zu left)\n", print_trigram(trgm).c_str(), len, num, cur_candidates.size());
 667                                         }
 668                                 }
 669                         });
 670                 }
 671         }
 672         engine.finish();
 673         if (done) {
 674                 return 0;
 675         }
 676         dprintf("Intersection done after %.1f ms. Doing final verification and printing:\n",
 677                 1e3 * duration<float>(steady_clock::now() - start).count());
 678
 679         uint64_t matched = scan_docids(needles, cur_candidates, corpus, &engine);
 680         dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
 681                 1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 682         return matched;
 683 }
 684
 685 // Run do_search_file() in a child process.
 686 //
 687 // The reason for this is that we're not robust against malicious input, so we need
 688 // to drop privileges after opening the file. (Otherwise, we could fall prey to an attack
 689 // where a user does locate -d badfile.db:/var/lib/plocate/plocate.db, badfile.db contains
 690 // a buffer overflow that takes over the process, and then uses the elevated privileges
 691 // to print out otherwise inaccessible paths.) We solve this by forking and treating the
 692 // child process as untrusted after it has dropped its privileges (which it does before
 693 // reading any data from the file); it returns a single 64-bit number over a pipe,
 694 // and that's it. The parent keeps its privileges, and can then fork out new children
 695 // without fear of being taken over. (The child keeps stdout for outputting results.)
 696 //
 697 // The count is returned over the pipe, because it's needed both for --limit and --count.
 698 uint64_t do_search_file_in_child(const vector<Needle> &needles, const std::string &filename)
 699 {
 700         int pipefd[2];
 701         if (pipe(pipefd) == -1) {
 702                 perror("pipe");
 703                 exit(EXIT_FAILURE);
 704         }
 705
 706         pid_t child_pid = fork();
 707         switch (child_pid) {
 708         case 0: {
 709                 // Child.
 710                 close(pipefd[0]);
 711                 in_forked_child = true;
 712                 uint64_t matched = do_search_file(needles, filename);
 713                 int ret;
 714                 do {
 715                         ret = write(pipefd[1], &matched, sizeof(matched));
 716                 } while (ret == -1 && errno == EINTR);
 717                 if (ret != sizeof(matched)) {
 718                         perror("write");
 719                         _exit(EXIT_FAILURE);
 720                 }
 721                 fflush(stdout);
 722                 _exit(EXIT_SUCCESS);
 723         }
 724         case -1:
 725                 // Error.
 726                 perror("fork");
 727                 exit(EXIT_FAILURE);
 728         default:
 729                 // Parent.
 730                 close(pipefd[1]);
 731                 break;
 732         }
 733
 734         // Wait for the child to finish.
 735         int wstatus;
 736         pid_t err;
 737         do {
 738                 err = waitpid(child_pid, &wstatus, 0);
 739         } while (err == -1 && errno == EINTR);
 740         if (err == -1) {
 741                 perror("waitpid");
 742                 exit(EXIT_FAILURE);
 743         }
 744         if (WIFEXITED(wstatus)) {
 745                 if (WEXITSTATUS(wstatus) != 0) {
 746                         // The child has probably already printed out its error, so just propagate the exit status.
 747                         exit(WEXITSTATUS(wstatus));
 748                 }
 749                 // Success!
 750         } else if (!WIFEXITED(wstatus)) {
 751                 fprintf(stderr, "FATAL: Child died unexpectedly while processing %s\n", filename.c_str());
 752                 exit(1);
 753         }
 754
 755         // Now get the number of matches from the child.
 756         uint64_t matched;
 757         int ret;
 758         do {
 759                 ret = read(pipefd[0], &matched, sizeof(matched));
 760         } while (ret == -1 && errno == EINTR);
 761         if (ret == -1) {
 762                 perror("read");
 763                 exit(EXIT_FAILURE);
 764         } else if (ret != sizeof(matched)) {
 765                 fprintf(stderr, "FATAL: Short read through pipe (got %d bytes)\n", ret);
 766                 exit(EXIT_FAILURE);
 767         }
 768         close(pipefd[0]);
 769         return matched;
 770 }
 771
 772 // Parses a colon-separated list of strings and appends them onto the given vector.
 773 // Backslash escapes whatever comes after it.
 774 void parse_dbpaths(const char *ptr, vector<string> *output)
 775 {
 776         string str;
 777         while (*ptr != '\0') {
 778                 if (*ptr == '\\') {
 779                         if (ptr[1] == '\0') {
 780                                 fprintf(stderr, "ERROR: Escape character at the end of string\n");
 781                                 exit(EXIT_FAILURE);
 782                         }
 783                         // Escape.
 784                         str.push_back(ptr[1]);
 785                         ptr += 2;
 786                         continue;
 787                 }
 788                 if (*ptr == ':') {
 789                         // Separator.
 790                         output->push_back(move(str));
 791                         ++ptr;
 792                         continue;
 793                 }
 794                 str.push_back(*ptr++);
 795         }
 796         output->push_back(move(str));
 797 }
 798
 799 void usage()
 800 {
 801         printf(
 802                 "Usage: plocate [OPTION]... PATTERN...\n"
 803                 "\n"
 804                 "  -b, --basename         search only the file name portion of path names\n"
 805                 "  -c, --count            print number of matches instead of the matches\n"
 806                 "  -d, --database DBPATH  search for files in DBPATH\n"
 807                 "                         (default is " DBFILE ")\n"
 808                 "  -i, --ignore-case      search case-insensitively\n"
 809                 "  -l, --limit LIMIT      stop after LIMIT matches\n"
 810                 "  -0, --null             delimit matches by NUL instead of newline\n"
 811                 "  -r, --regexp           interpret patterns as basic regexps (slow)\n"
 812                 "      --regex            interpret patterns as extended regexps (slow)\n"
 813                 "  -w, --wholename        search the entire path name (default; see -b)\n"
 814                 "      --help             print this help\n"
 815                 "      --version          print version information\n");
 816 }
 817
 818 void version()
 819 {
 820         printf("%s %s\n", PACKAGE_NAME, PACKAGE_VERSION);
 821         printf("Copyright 2020 Steinar H. Gunderson\n");
 822         printf("License GPLv2+: GNU GPL version 2 or later <https://gnu.org/licenses/gpl.html>.\n");
 823         printf("This is free software: you are free to change and redistribute it.\n");
 824         printf("There is NO WARRANTY, to the extent permitted by law.\n");
 825         exit(0);
 826 }
 827
 828 int main(int argc, char **argv)
 829 {
 830         vector<string> dbpaths;
 831
 832         constexpr int EXTENDED_REGEX = 1000;
 833         constexpr int FLUSH_CACHE = 1001;
 834         static const struct option long_options[] = {
 835                 { "help", no_argument, 0, 'h' },
 836                 { "count", no_argument, 0, 'c' },
 837                 { "basename", no_argument, 0, 'b' },
 838                 { "database", required_argument, 0, 'd' },
 839                 { "existing", no_argument, 0, 'e' },
 840                 { "ignore-case", no_argument, 0, 'i' },
 841                 { "limit", required_argument, 0, 'l' },
 842                 { "null", no_argument, 0, '0' },
 843                 { "version", no_argument, 0, 'V' },
 844                 { "regexp", no_argument, 0, 'r' },
 845                 { "regex", no_argument, 0, EXTENDED_REGEX },
 846                 { "wholename", no_argument, 0, 'w' },
 847                 { "debug", no_argument, 0, 'D' },  // Not documented.
 848                 // Enable to test cold-cache behavior (except for access()). Not documented.
 849                 { "flush-cache", no_argument, 0, FLUSH_CACHE },
 850                 { 0, 0, 0, 0 }
 851         };
 852
 853         setlocale(LC_ALL, "");
 854         for (;;) {
 855                 int option_index = 0;
 856                 int c = getopt_long(argc, argv, "bcd:ehil:n:0rwVD", long_options, &option_index);
 857                 if (c == -1) {
 858                         break;
 859                 }
 860                 switch (c) {
 861                 case 'b':
 862                         match_basename = true;
 863                         break;
 864                 case 'c':
 865                         only_count = true;
 866                         break;
 867                 case 'd':
 868                         parse_dbpaths(optarg, &dbpaths);
 869                         break;
 870                 case 'e':
 871                         check_existence = true;
 872                         break;
 873                 case 'h':
 874                         usage();
 875                         exit(0);
 876                 case 'i':
 877                         ignore_case = true;
 878                         break;
 879                 case 'l':
 880                 case 'n':
 881                         limit_matches = limit_left = atoll(optarg);
 882                         if (limit_matches <= 0) {
 883                                 fprintf(stderr, "Error: limit must be a strictly positive number.\n");
 884                                 exit(1);
 885                         }
 886                         break;
 887                 case '0':
 888                         print_nul = true;
 889                         break;
 890                 case 'r':
 891                         patterns_are_regex = true;
 892                         break;
 893                 case EXTENDED_REGEX:
 894                         patterns_are_regex = true;
 895                         use_extended_regex = true;
 896                         break;
 897                 case 'w':
 898                         match_basename = false;  // No-op unless -b is given first.
 899                         break;
 900                 case 'D':
 901                         use_debug = true;
 902                         break;
 903                 case FLUSH_CACHE:
 904                         flush_cache = true;
 905                         break;
 906                 case 'V':
 907                         version();
 908                         break;
 909                 default:
 910                         exit(1);
 911                 }
 912         }
 913
 914         if (use_debug || flush_cache) {
 915                 // Debug information would leak information about which files exist,
 916                 // so drop setgid before we open the file; one would either need to run
 917                 // as root, or use a locally-built file. Doing the same thing for
 918                 // flush_cache is mostly paranoia, in an attempt to prevent random users
 919                 // from making plocate slow for everyone else.
 920                 if (setgid(getgid()) != 0) {
 921                         perror("setgid");
 922                         exit(EXIT_FAILURE);
 923                 }
 924         }
 925
 926         if (!print_nul) {
 927                 stdout_is_tty = isatty(1);
 928         }
 929
 930         vector<Needle> needles;
 931         for (int i = optind; i < argc; ++i) {
 932                 Needle needle;
 933                 needle.str = argv[i];
 934
 935                 // See if there are any wildcard characters, which indicates we should treat it
 936                 // as an (anchored) glob.
 937                 bool any_wildcard = false;
 938                 for (size_t i = 0; i < needle.str.size(); i += read_unigram(needle.str, i).second) {
 939                         if (read_unigram(needle.str, i).first == WILDCARD_UNIGRAM) {
 940                                 any_wildcard = true;
 941                                 break;
 942                         }
 943                 }
 944
 945                 if (patterns_are_regex) {
 946                         needle.type = Needle::REGEX;
 947                         needle.re = compile_regex(needle.str);
 948                 } else if (any_wildcard) {
 949                         needle.type = Needle::GLOB;
 950                 } else if (ignore_case) {
 951                         // strcasestr() doesn't handle locales correctly (even though LSB
 952                         // claims it should), but somehow, fnmatch() does, and it's about
 953                         // the same speed as using a regex.
 954                         needle.type = Needle::GLOB;
 955                         needle.str = "*" + needle.str + "*";
 956                 } else {
 957                         needle.type = Needle::STRSTR;
 958                         needle.str = unescape_glob_to_plain_string(needle.str);
 959                 }
 960                 needles.push_back(move(needle));
 961         }
 962         if (needles.empty()) {
 963                 fprintf(stderr, "plocate: no pattern to search for specified\n");
 964                 exit(0);
 965         }
 966
 967         if (dbpaths.empty()) {
 968                 // No -d given, so use our default. Note that this happens
 969                 // even if LOCATE_PATH exists, to match mlocate behavior.
 970                 dbpaths.push_back(DBFILE);
 971         }
 972
 973         const char *locate_path = getenv("LOCATE_PATH");
 974         if (locate_path != nullptr) {
 975                 parse_dbpaths(locate_path, &dbpaths);
 976         }
 977
 978         uint64_t matched = 0;
 979         for (size_t i = 0; i < dbpaths.size(); ++i) {
 980                 uint64_t this_matched;
 981                 if (i != dbpaths.size() - 1) {
 982                         this_matched = do_search_file_in_child(needles, dbpaths[i]);
 983                 } else {
 984                         this_matched = do_search_file(needles, dbpaths[i]);
 985                 }
 986                 matched += this_matched;
 987                 limit_left -= this_matched;
 988         }
 989         if (only_count) {
 990                 printf("%" PRId64 "\n", matched);
 991         }
 992 }