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 bool literal_printing = false;
  62 static bool in_forked_child = false;
  63
  64 steady_clock::time_point start;
  65 ZSTD_DDict *ddict = nullptr;
  66
  67 class Corpus {
  68 public:
  69         Corpus(int fd, const char *filename_for_errors, IOUringEngine *engine);
  70         ~Corpus();
  71         void find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb);
  72         void get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const;
  73         size_t get_num_filename_blocks() const;
  74         off_t offset_for_block(uint32_t docid) const
  75         {
  76                 return hdr.filename_index_offset_bytes + docid * sizeof(uint64_t);
  77         }
  78         const Header &get_hdr() const { return hdr; }
  79
  80 public:
  81         const int fd;
  82         IOUringEngine *const engine;
  83
  84         Header hdr;
  85 };
  86
  87 Corpus::Corpus(int fd, const char *filename_for_errors, IOUringEngine *engine)
  88         : fd(fd), engine(engine)
  89 {
  90         if (flush_cache) {
  91                 off_t len = lseek(fd, 0, SEEK_END);
  92                 if (len == -1) {
  93                         perror("lseek");
  94                         exit(1);
  95                 }
  96                 posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED);
  97         }
  98
  99         complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
 100         if (memcmp(hdr.magic, "\0plocate", 8) != 0) {
 101                 fprintf(stderr, "%s: database is corrupt or not a plocate database; please rebuild it.\n", filename_for_errors);
 102                 exit(1);
 103         }
 104         if (hdr.version != 0 && hdr.version != 1) {
 105                 fprintf(stderr, "%s: has version %u, expected 0 or 1; please rebuild it.\n", filename_for_errors, hdr.version);
 106                 exit(1);
 107         }
 108         if (hdr.version == 0) {
 109                 // These will be junk data.
 110                 hdr.zstd_dictionary_offset_bytes = 0;
 111                 hdr.zstd_dictionary_length_bytes = 0;
 112         }
 113         if (hdr.max_version < 2) {
 114                 // This too. (We ignore the other max_version 2 fields.)
 115                 hdr.check_visibility = true;
 116         }
 117 }
 118
 119 Corpus::~Corpus()
 120 {
 121         close(fd);
 122 }
 123
 124 void Corpus::find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb)
 125 {
 126         uint32_t bucket = hash_trigram(trgm, hdr.hashtable_size);
 127         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) {
 128                 const Trigram *trgmptr = reinterpret_cast<const Trigram *>(s.data());
 129                 for (unsigned i = 0; i < hdr.extra_ht_slots + 1; ++i) {
 130                         if (trgmptr[i].trgm == trgm) {
 131                                 cb(trgmptr + i, trgmptr[i + 1].offset - trgmptr[i].offset);
 132                                 return;
 133                         }
 134                 }
 135
 136                 // Not found.
 137                 cb(nullptr, 0);
 138         });
 139 }
 140
 141 void Corpus::get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const
 142 {
 143         // Read the file offset from this docid and the next one.
 144         // This is always allowed, since we have a sentinel block at the end.
 145         engine->submit_read(fd, sizeof(uint64_t) * 2, offset_for_block(docid), [this, cb{ move(cb) }](string_view s) {
 146                 const uint64_t *ptr = reinterpret_cast<const uint64_t *>(s.data());
 147                 off_t offset = ptr[0];
 148                 size_t len = ptr[1] - ptr[0];
 149                 engine->submit_read(fd, len, offset, cb);
 150         });
 151 }
 152
 153 size_t Corpus::get_num_filename_blocks() const
 154 {
 155         return hdr.num_docids;
 156 }
 157
 158 template<class T>
 159 void stat_if_needed(const char *filename, bool access_ok, IOUringEngine *engine, T cb)
 160 {
 161         if (!access_ok || !check_existence) {
 162                 // Doesn't have access or doesn't care about existence, so no need to stat.
 163                 cb(access_ok);
 164         } else if (engine == nullptr || !engine->get_supports_stat()) {
 165                 // Do a synchronous stat.
 166                 struct stat buf;
 167                 bool ok = lstat(filename, &buf) == 0;
 168                 cb(ok);
 169         } else {
 170                 engine->submit_stat(filename, cb);
 171         }
 172 }
 173
 174 void scan_file_block(const vector<Needle> &needles, string_view compressed,
 175                      IOUringEngine *engine, AccessRXCache *access_rx_cache, uint64_t seq, ResultReceiver *serializer,
 176                      atomic<uint64_t> *matched)
 177 {
 178         unsigned long long uncompressed_len = ZSTD_getFrameContentSize(compressed.data(), compressed.size());
 179         if (uncompressed_len == ZSTD_CONTENTSIZE_UNKNOWN || uncompressed_len == ZSTD_CONTENTSIZE_ERROR) {
 180                 fprintf(stderr, "ZSTD_getFrameContentSize() failed\n");
 181                 exit(1);
 182         }
 183
 184         string block;
 185         block.resize(uncompressed_len + 1);
 186
 187         static thread_local ZSTD_DCtx *ctx = ZSTD_createDCtx();  // Reused across calls.
 188         size_t err;
 189
 190         if (ddict != nullptr) {
 191                 err = ZSTD_decompress_usingDDict(ctx, &block[0], block.size(), compressed.data(),
 192                                                  compressed.size(), ddict);
 193         } else {
 194                 err = ZSTD_decompressDCtx(ctx, &block[0], block.size(), compressed.data(),
 195                                           compressed.size());
 196         }
 197         if (ZSTD_isError(err)) {
 198                 fprintf(stderr, "ZSTD_decompress(): %s\n", ZSTD_getErrorName(err));
 199                 exit(1);
 200         }
 201         block[block.size() - 1] = '\0';
 202
 203         auto test_candidate = [&](const char *filename, uint64_t local_seq, uint64_t next_seq) {
 204                 access_rx_cache->check_access(filename, /*allow_async=*/true, [matched, engine, serializer, local_seq, next_seq, filename{ strdup(filename) }](bool ok) {
 205                         stat_if_needed(filename, ok, engine, [matched, serializer, local_seq, next_seq, filename](bool ok) {
 206                                 if (ok) {
 207                                         ++*matched;
 208                                         serializer->print(local_seq, next_seq - local_seq, filename);
 209                                 } else {
 210                                         serializer->print(local_seq, next_seq - local_seq, "");
 211                                 }
 212                                 free(filename);
 213                         });
 214                 });
 215         };
 216
 217         // We need to know the next sequence number before inserting into Serializer,
 218         // so always buffer one candidate.
 219         const char *pending_candidate = nullptr;
 220
 221         uint64_t local_seq = seq << 32;
 222         for (const char *filename = block.data();
 223              filename != block.data() + block.size();
 224              filename += strlen(filename) + 1) {
 225                 const char *haystack = filename;
 226                 if (match_basename) {
 227                         haystack = strrchr(filename, '/');
 228                         if (haystack == nullptr) {
 229                                 haystack = filename;
 230                         } else {
 231                                 ++haystack;
 232                         }
 233                 }
 234
 235                 bool found = true;
 236                 for (const Needle &needle : needles) {
 237                         if (!matches(needle, haystack)) {
 238                                 found = false;
 239                                 break;
 240                         }
 241                 }
 242                 if (found) {
 243                         if (pending_candidate != nullptr) {
 244                                 test_candidate(pending_candidate, local_seq, local_seq + 1);
 245                                 ++local_seq;
 246                         }
 247                         pending_candidate = filename;
 248                 }
 249         }
 250         if (pending_candidate == nullptr) {
 251                 serializer->print(seq << 32, 1ULL << 32, "");
 252         } else {
 253                 test_candidate(pending_candidate, local_seq, (seq + 1) << 32);
 254         }
 255 }
 256
 257 size_t scan_docids(const vector<Needle> &needles, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
 258 {
 259         Serializer docids_in_order;
 260         AccessRXCache access_rx_cache(engine, corpus.get_hdr().check_visibility);
 261         atomic<uint64_t> matched{ 0 };
 262         for (size_t i = 0; i < docids.size(); ++i) {
 263                 uint32_t docid = docids[i];
 264                 corpus.get_compressed_filename_block(docid, [i, &matched, &needles, &access_rx_cache, engine, &docids_in_order](string_view compressed) {
 265                         scan_file_block(needles, compressed, engine, &access_rx_cache, i, &docids_in_order, &matched);
 266                 });
 267         }
 268         engine->finish();
 269         return matched;
 270 }
 271
 272 struct WorkerThread {
 273         thread t;
 274
 275         // We use a result queue instead of synchronizing Serializer,
 276         // since a lock on it becomes a huge choke point if there are
 277         // lots of threads.
 278         mutex result_mu;
 279         struct Result {
 280                 uint64_t seq;
 281                 uint64_t skip;
 282                 string msg;
 283         };
 284         vector<Result> results;
 285 };
 286
 287 class WorkerThreadReceiver : public ResultReceiver {
 288 public:
 289         WorkerThreadReceiver(WorkerThread *wt)
 290                 : wt(wt) {}
 291
 292         void print(uint64_t seq, uint64_t skip, const string msg) override
 293         {
 294                 lock_guard<mutex> lock(wt->result_mu);
 295                 if (msg.empty() && !wt->results.empty() && wt->results.back().seq + wt->results.back().skip == seq) {
 296                         wt->results.back().skip += skip;
 297                 } else {
 298                         wt->results.emplace_back(WorkerThread::Result{ seq, skip, move(msg) });
 299                 }
 300         }
 301
 302 private:
 303         WorkerThread *wt;
 304 };
 305
 306 void deliver_results(WorkerThread *wt, Serializer *serializer)
 307 {
 308         vector<WorkerThread::Result> results;
 309         {
 310                 lock_guard<mutex> lock(wt->result_mu);
 311                 results = move(wt->results);
 312         }
 313         for (const WorkerThread::Result &result : results) {
 314                 serializer->print(result.seq, result.skip, move(result.msg));
 315         }
 316 }
 317
 318 // We do this sequentially, as it's faster than scattering
 319 // a lot of I/O through io_uring and hoping the kernel will
 320 // coalesce it plus readahead for us. Since we assume that
 321 // we will primarily be CPU-bound, we'll be firing up one
 322 // worker thread for each spare core (the last one will
 323 // only be doing I/O). access() is still synchronous.
 324 uint64_t scan_all_docids(const vector<Needle> &needles, int fd, const Corpus &corpus)
 325 {
 326         {
 327                 const Header &hdr = corpus.get_hdr();
 328                 if (hdr.zstd_dictionary_length_bytes > 0) {
 329                         string dictionary;
 330                         dictionary.resize(hdr.zstd_dictionary_length_bytes);
 331                         complete_pread(fd, &dictionary[0], hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes);
 332                         ddict = ZSTD_createDDict(dictionary.data(), dictionary.size());
 333                 }
 334         }
 335
 336         AccessRXCache access_rx_cache(nullptr, corpus.get_hdr().check_visibility);
 337         Serializer serializer;
 338         uint32_t num_blocks = corpus.get_num_filename_blocks();
 339         unique_ptr<uint64_t[]> offsets(new uint64_t[num_blocks + 1]);
 340         complete_pread(fd, offsets.get(), (num_blocks + 1) * sizeof(uint64_t), corpus.offset_for_block(0));
 341         atomic<uint64_t> matched{ 0 };
 342
 343         mutex mu;
 344         condition_variable queue_added, queue_removed;
 345         deque<tuple<int, int, string>> work_queue;  // Under mu.
 346         bool done = false;  // Under mu.
 347
 348         unsigned num_threads = max<int>(sysconf(_SC_NPROCESSORS_ONLN) - 1, 1);
 349         dprintf("Using %u worker threads for linear scan.\n", num_threads);
 350         unique_ptr<WorkerThread[]> threads(new WorkerThread[num_threads]);
 351         for (unsigned i = 0; i < num_threads; ++i) {
 352                 threads[i].t = thread([&threads, &mu, &queue_added, &queue_removed, &work_queue, &done, &offsets, &needles, &access_rx_cache, &matched, i] {
 353                         // regcomp() takes a lock on the regex, so each thread will need its own.
 354                         const vector<Needle> *use_needles = &needles;
 355                         vector<Needle> recompiled_needles;
 356                         if (i != 0 && patterns_are_regex) {
 357                                 recompiled_needles = needles;
 358                                 for (Needle &needle : recompiled_needles) {
 359                                         needle.re = compile_regex(needle.str);
 360                                 }
 361                                 use_needles = &recompiled_needles;
 362                         }
 363
 364                         WorkerThreadReceiver receiver(&threads[i]);
 365                         for (;;) {
 366                                 uint32_t io_docid, last_docid;
 367                                 string compressed;
 368
 369                                 {
 370                                         unique_lock<mutex> lock(mu);
 371                                         queue_added.wait(lock, [&work_queue, &done] { return !work_queue.empty() || done; });
 372                                         if (done && work_queue.empty()) {
 373                                                 return;
 374                                         }
 375                                         tie(io_docid, last_docid, compressed) = move(work_queue.front());
 376                                         work_queue.pop_front();
 377                                         queue_removed.notify_all();
 378                                 }
 379
 380                                 for (uint32_t docid = io_docid; docid < last_docid; ++docid) {
 381                                         size_t relative_offset = offsets[docid] - offsets[io_docid];
 382                                         size_t len = offsets[docid + 1] - offsets[docid];
 383                                         // IOUringEngine isn't thread-safe, so we do any needed stat()s synchronously (nullptr engine).
 384                                         scan_file_block(*use_needles, { &compressed[relative_offset], len }, /*engine=*/nullptr, &access_rx_cache, docid, &receiver, &matched);
 385                                 }
 386                         }
 387                 });
 388         }
 389
 390         string compressed;
 391         for (uint32_t io_docid = 0; io_docid < num_blocks; io_docid += 32) {
 392                 uint32_t last_docid = std::min(io_docid + 32, num_blocks);
 393                 size_t io_len = offsets[last_docid] - offsets[io_docid];
 394                 if (compressed.size() < io_len) {
 395                         compressed.resize(io_len);
 396                 }
 397                 complete_pread(fd, &compressed[0], io_len, offsets[io_docid]);
 398
 399                 {
 400                         unique_lock<mutex> lock(mu);
 401                         queue_removed.wait(lock, [&work_queue] { return work_queue.size() < 256; });  // Allow ~2MB of data queued up.
 402                         work_queue.emplace_back(io_docid, last_docid, move(compressed));
 403                         queue_added.notify_one();  // Avoid the thundering herd.
 404                 }
 405
 406                 // Pick up some results, so that we are sure that we won't just overload.
 407                 // (Seemingly, going through all of these causes slowness with many threads,
 408                 // but taking only one is OK.)
 409                 unsigned i = io_docid / 32;
 410                 deliver_results(&threads[i % num_threads], &serializer);
 411         }
 412         {
 413                 lock_guard<mutex> lock(mu);
 414                 done = true;
 415                 queue_added.notify_all();
 416         }
 417         for (unsigned i = 0; i < num_threads; ++i) {
 418                 threads[i].t.join();
 419                 deliver_results(&threads[i], &serializer);
 420         }
 421         return matched;
 422 }
 423
 424 // Takes the given posting list, unions it into the parts of the trigram disjunction
 425 // already read; if the list is complete, intersects with “cur_candidates”.
 426 //
 427 // Returns true if the search should be aborted (we are done).
 428 bool new_posting_list_read(TrigramDisjunction *td, vector<uint32_t> decoded, vector<uint32_t> *cur_candidates, vector<uint32_t> *tmp)
 429 {
 430         if (td->docids.empty()) {
 431                 td->docids = move(decoded);
 432         } else {
 433                 tmp->clear();
 434                 set_union(decoded.begin(), decoded.end(), td->docids.begin(), td->docids.end(), back_inserter(*tmp));
 435                 swap(*tmp, td->docids);
 436         }
 437         if (--td->remaining_trigrams_to_read > 0) {
 438                 // Need to wait for more.
 439                 if (ignore_case) {
 440                         dprintf("  ... %u reads left in OR group %u (%zu docids in list)\n",
 441                                 td->remaining_trigrams_to_read, td->index, td->docids.size());
 442                 }
 443                 return false;
 444         }
 445         if (cur_candidates->empty()) {
 446                 if (ignore_case) {
 447                         dprintf("  ... all reads done for OR group %u (%zu docids)\n",
 448                                 td->index, td->docids.size());
 449                 }
 450                 *cur_candidates = move(td->docids);
 451         } else {
 452                 tmp->clear();
 453                 set_intersection(cur_candidates->begin(), cur_candidates->end(),
 454                                  td->docids.begin(), td->docids.end(),
 455                                  back_inserter(*tmp));
 456                 swap(*cur_candidates, *tmp);
 457                 if (ignore_case) {
 458                         if (cur_candidates->empty()) {
 459                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (none left, search is done)\n",
 460                                         td->index, td->docids.size());
 461                                 return true;
 462                         } else {
 463                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (%zu left)\n",
 464                                         td->index, td->docids.size(), cur_candidates->size());
 465                         }
 466                 }
 467         }
 468         return false;
 469 }
 470
 471 uint64_t do_search_file(const vector<Needle> &needles, const std::string &filename)
 472 {
 473         int fd = open(filename.c_str(), O_RDONLY);
 474         if (fd == -1) {
 475                 perror(filename.c_str());
 476                 exit(1);
 477         }
 478
 479         // Drop privileges.
 480         if (setgid(getgid()) != 0) {
 481                 perror("setgid");
 482                 exit(EXIT_FAILURE);
 483         }
 484
 485         start = steady_clock::now();
 486         if (access("/", R_OK | X_OK)) {
 487                 // We can't find anything, no need to bother...
 488                 return 0;
 489         }
 490
 491         IOUringEngine engine(/*slop_bytes=*/16);  // 16 slop bytes as described in turbopfor.h.
 492         Corpus corpus(fd, filename.c_str(), &engine);
 493         dprintf("Corpus init done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 494
 495         vector<TrigramDisjunction> trigram_groups;
 496         if (patterns_are_regex) {
 497                 // We could parse the regex to find trigrams that have to be there
 498                 // (there are actually known algorithms to deal with disjunctions
 499                 // and such, too), but for now, we just go brute force.
 500                 // Using locate with regexes is pretty niche.
 501         } else {
 502                 for (const Needle &needle : needles) {
 503                         parse_trigrams(needle.str, ignore_case, &trigram_groups);
 504                 }
 505         }
 506
 507         unique_sort(
 508                 &trigram_groups,
 509                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives < b.trigram_alternatives; },
 510                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives == b.trigram_alternatives; });
 511
 512         // Give them names for debugging.
 513         unsigned td_index = 0;
 514         for (TrigramDisjunction &td : trigram_groups) {
 515                 td.index = td_index++;
 516         }
 517
 518         // Collect which trigrams we need to look up in the hash table.
 519         unordered_map<uint32_t, vector<TrigramDisjunction *>> trigrams_to_lookup;
 520         for (TrigramDisjunction &td : trigram_groups) {
 521                 for (uint32_t trgm : td.trigram_alternatives) {
 522                         trigrams_to_lookup[trgm].push_back(&td);
 523                 }
 524         }
 525         if (trigrams_to_lookup.empty()) {
 526                 // Too short for trigram matching. Apply brute force.
 527                 // (We could have searched through all trigrams that matched
 528                 // the pattern and done a union of them, but that's a lot of
 529                 // work for fairly unclear gain.)
 530                 uint64_t matched = scan_all_docids(needles, fd, corpus);
 531                 dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
 532                         1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 533                 return matched;
 534         }
 535
 536         // Sneak in fetching the dictionary, if present. It's not necessarily clear
 537         // exactly where it would be cheapest to get it, but it needs to be present
 538         // before we can decode any of the posting lists. Most likely, it's
 539         // in the same filesystem block as the header anyway, so it should be
 540         // present in the cache.
 541         {
 542                 const Header &hdr = corpus.get_hdr();
 543                 if (hdr.zstd_dictionary_length_bytes > 0) {
 544                         engine.submit_read(fd, hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes, [](string_view s) {
 545                                 ddict = ZSTD_createDDict(s.data(), s.size());
 546                                 dprintf("Dictionary initialized after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 547                         });
 548                 }
 549         }
 550
 551         // Look them all up on disk.
 552         bool should_early_exit = false;
 553         for (auto &[trgm, trigram_groups] : trigrams_to_lookup) {
 554                 corpus.find_trigram(trgm, [trgm{ trgm }, trigram_groups{ &trigram_groups }, &should_early_exit](const Trigram *trgmptr, size_t len) {
 555                         if (trgmptr == nullptr) {
 556                                 dprintf("trigram %s isn't found\n", print_trigram(trgm).c_str());
 557                                 for (TrigramDisjunction *td : *trigram_groups) {
 558                                         --td->remaining_trigrams_to_read;
 559
 560                                         // If we now know this trigram group doesn't match anything at all,
 561                                         // we can do early exit; however, if we're in a forked child,
 562                                         // that would confuse the parent process (since we don't write
 563                                         // our count to the pipe), so we wait until we're back in to the
 564                                         // regular (non-async) context. This is a fairly rare case anyway,
 565                                         // and the gains from dropping the remaining trigram reads are limited.
 566                                         if (td->remaining_trigrams_to_read == 0 && td->read_trigrams.empty()) {
 567                                                 if (in_forked_child) {
 568                                                         should_early_exit = true;
 569                                                 } else {
 570                                                         dprintf("zero matches in %s, so we are done\n", print_td(*td).c_str());
 571                                                         if (only_count) {
 572                                                                 printf("0\n");
 573                                                         }
 574                                                         exit(0);
 575                                                 }
 576                                         }
 577                                 }
 578                                 return;
 579                         }
 580                         for (TrigramDisjunction *td : *trigram_groups) {
 581                                 --td->remaining_trigrams_to_read;
 582                                 td->max_num_docids += trgmptr->num_docids;
 583                                 td->read_trigrams.emplace_back(*trgmptr, len);
 584                         }
 585                 });
 586         }
 587         engine.finish();
 588         dprintf("Hashtable lookups done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 589
 590         if (should_early_exit) {
 591                 return 0;
 592         }
 593
 594         for (TrigramDisjunction &td : trigram_groups) {
 595                 // Reset for reads.
 596                 td.remaining_trigrams_to_read = td.read_trigrams.size();
 597
 598                 if (ignore_case) {  // If case-sensitive, they'll all be pretty obvious single-entry groups.
 599                         dprintf("OR group %u (max_num_docids=%u): %s\n", td.index, td.max_num_docids, print_td(td).c_str());
 600                 }
 601         }
 602
 603         // TODO: For case-insensitive (ie. more than one alternative in each),
 604         // prioritize the ones with fewer seeks?
 605         sort(trigram_groups.begin(), trigram_groups.end(),
 606              [&](const TrigramDisjunction &a, const TrigramDisjunction &b) {
 607                      return a.max_num_docids < b.max_num_docids;
 608              });
 609
 610         unordered_map<uint32_t, vector<TrigramDisjunction *>> uses_trigram;
 611         for (TrigramDisjunction &td : trigram_groups) {
 612                 for (uint32_t trgm : td.trigram_alternatives) {
 613                         uses_trigram[trgm].push_back(&td);
 614                 }
 615         }
 616
 617         unordered_set<uint32_t> trigrams_submitted_read;
 618         vector<uint32_t> cur_candidates, tmp, decoded;
 619         bool done = false;
 620         for (TrigramDisjunction &td : trigram_groups) {
 621                 if (!cur_candidates.empty() && td.max_num_docids > cur_candidates.size() * 100) {
 622                         dprintf("%s has up to %u entries, ignoring the rest (will "
 623                                 "weed out false positives later)\n",
 624                                 print_td(td).c_str(), td.max_num_docids);
 625                         break;
 626                 }
 627
 628                 for (auto &[trgmptr, len] : td.read_trigrams) {
 629                         if (trigrams_submitted_read.count(trgmptr.trgm) != 0) {
 630                                 continue;
 631                         }
 632                         trigrams_submitted_read.insert(trgmptr.trgm);
 633                         // Only stay a certain amount ahead, so that we don't spend I/O
 634                         // on reading the latter, large posting lists. We are unlikely
 635                         // to need them anyway, even if they should come in first.
 636                         if (engine.get_waiting_reads() >= 5) {
 637                                 engine.finish();
 638                                 if (done)
 639                                         break;
 640                         }
 641                         engine.submit_read(fd, len, trgmptr.offset, [trgmptr{ trgmptr }, len{ len }, &done, &cur_candidates, &tmp, &decoded, &uses_trigram](string_view s) {
 642                                 if (done)
 643                                         return;
 644
 645                                 uint32_t trgm = trgmptr.trgm;
 646                                 const unsigned char *pldata = reinterpret_cast<const unsigned char *>(s.data());
 647                                 size_t num = trgmptr.num_docids;
 648                                 decoded.resize(num);
 649                                 decode_pfor_delta1_128(pldata, num, /*interleaved=*/true, &decoded[0]);
 650
 651                                 assert(uses_trigram.count(trgm) != 0);
 652                                 bool was_empty = cur_candidates.empty();
 653                                 if (ignore_case) {
 654                                         dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 655                                 }
 656
 657                                 for (TrigramDisjunction *td : uses_trigram[trgm]) {
 658                                         done |= new_posting_list_read(td, decoded, &cur_candidates, &tmp);
 659                                         if (done)
 660                                                 break;
 661                                 }
 662                                 if (!ignore_case) {
 663                                         if (was_empty) {
 664                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 665                                         } else if (cur_candidates.empty()) {
 666                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (none left, search is done)\n", print_trigram(trgm).c_str(), len, num);
 667                                         } else {
 668                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (%zu left)\n", print_trigram(trgm).c_str(), len, num, cur_candidates.size());
 669                                         }
 670                                 }
 671                         });
 672                 }
 673         }
 674         engine.finish();
 675         if (done) {
 676                 return 0;
 677         }
 678         dprintf("Intersection done after %.1f ms. Doing final verification and printing:\n",
 679                 1e3 * duration<float>(steady_clock::now() - start).count());
 680
 681         uint64_t matched = scan_docids(needles, cur_candidates, corpus, &engine);
 682         dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
 683                 1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 684         return matched;
 685 }
 686
 687 // Run do_search_file() in a child process.
 688 //
 689 // The reason for this is that we're not robust against malicious input, so we need
 690 // to drop privileges after opening the file. (Otherwise, we could fall prey to an attack
 691 // where a user does locate -d badfile.db:/var/lib/plocate/plocate.db, badfile.db contains
 692 // a buffer overflow that takes over the process, and then uses the elevated privileges
 693 // to print out otherwise inaccessible paths.) We solve this by forking and treating the
 694 // child process as untrusted after it has dropped its privileges (which it does before
 695 // reading any data from the file); it returns a single 64-bit number over a pipe,
 696 // and that's it. The parent keeps its privileges, and can then fork out new children
 697 // without fear of being taken over. (The child keeps stdout for outputting results.)
 698 //
 699 // The count is returned over the pipe, because it's needed both for --limit and --count.
 700 uint64_t do_search_file_in_child(const vector<Needle> &needles, const std::string &filename)
 701 {
 702         int pipefd[2];
 703         if (pipe(pipefd) == -1) {
 704                 perror("pipe");
 705                 exit(EXIT_FAILURE);
 706         }
 707
 708         pid_t child_pid = fork();
 709         switch (child_pid) {
 710         case 0: {
 711                 // Child.
 712                 close(pipefd[0]);
 713                 in_forked_child = true;
 714                 uint64_t matched = do_search_file(needles, filename);
 715                 int ret;
 716                 do {
 717                         ret = write(pipefd[1], &matched, sizeof(matched));
 718                 } while (ret == -1 && errno == EINTR);
 719                 if (ret != sizeof(matched)) {
 720                         perror("write");
 721                         _exit(EXIT_FAILURE);
 722                 }
 723                 fflush(stdout);
 724                 _exit(EXIT_SUCCESS);
 725         }
 726         case -1:
 727                 // Error.
 728                 perror("fork");
 729                 exit(EXIT_FAILURE);
 730         default:
 731                 // Parent.
 732                 close(pipefd[1]);
 733                 break;
 734         }
 735
 736         // Wait for the child to finish.
 737         int wstatus;
 738         pid_t err;
 739         do {
 740                 err = waitpid(child_pid, &wstatus, 0);
 741         } while (err == -1 && errno == EINTR);
 742         if (err == -1) {
 743                 perror("waitpid");
 744                 exit(EXIT_FAILURE);
 745         }
 746         if (WIFEXITED(wstatus)) {
 747                 if (WEXITSTATUS(wstatus) != 0) {
 748                         // The child has probably already printed out its error, so just propagate the exit status.
 749                         exit(WEXITSTATUS(wstatus));
 750                 }
 751                 // Success!
 752         } else if (!WIFEXITED(wstatus)) {
 753                 fprintf(stderr, "FATAL: Child died unexpectedly while processing %s\n", filename.c_str());
 754                 exit(1);
 755         }
 756
 757         // Now get the number of matches from the child.
 758         uint64_t matched;
 759         int ret;
 760         do {
 761                 ret = read(pipefd[0], &matched, sizeof(matched));
 762         } while (ret == -1 && errno == EINTR);
 763         if (ret == -1) {
 764                 perror("read");
 765                 exit(EXIT_FAILURE);
 766         } else if (ret != sizeof(matched)) {
 767                 fprintf(stderr, "FATAL: Short read through pipe (got %d bytes)\n", ret);
 768                 exit(EXIT_FAILURE);
 769         }
 770         close(pipefd[0]);
 771         return matched;
 772 }
 773
 774 // Parses a colon-separated list of strings and appends them onto the given vector.
 775 // Backslash escapes whatever comes after it.
 776 void parse_dbpaths(const char *ptr, vector<string> *output)
 777 {
 778         string str;
 779         while (*ptr != '\0') {
 780                 if (*ptr == '\\') {
 781                         if (ptr[1] == '\0') {
 782                                 fprintf(stderr, "ERROR: Escape character at the end of string\n");
 783                                 exit(EXIT_FAILURE);
 784                         }
 785                         // Escape.
 786                         str.push_back(ptr[1]);
 787                         ptr += 2;
 788                         continue;
 789                 }
 790                 if (*ptr == ':') {
 791                         // Separator.
 792                         output->push_back(move(str));
 793                         ++ptr;
 794                         continue;
 795                 }
 796                 str.push_back(*ptr++);
 797         }
 798         output->push_back(move(str));
 799 }
 800
 801 void usage()
 802 {
 803         printf(
 804                 "Usage: plocate [OPTION]... PATTERN...\n"
 805                 "\n"
 806                 "  -b, --basename         search only the file name portion of path names\n"
 807                 "  -c, --count            print number of matches instead of the matches\n"
 808                 "  -d, --database DBPATH  search for files in DBPATH\n"
 809                 "                         (default is " DBFILE ")\n"
 810                 "  -i, --ignore-case      search case-insensitively\n"
 811                 "  -l, --limit LIMIT      stop after LIMIT matches\n"
 812                 "  -0, --null             delimit matches by NUL instead of newline\n"
 813                 "  -N, --literal          do not quote filenames, even if printing to a tty\n"
 814                 "  -r, --regexp           interpret patterns as basic regexps (slow)\n"
 815                 "      --regex            interpret patterns as extended regexps (slow)\n"
 816                 "  -w, --wholename        search the entire path name (default; see -b)\n"
 817                 "      --help             print this help\n"
 818                 "      --version          print version information\n");
 819 }
 820
 821 void version()
 822 {
 823         printf("%s %s\n", PACKAGE_NAME, PACKAGE_VERSION);
 824         printf("Copyright 2020 Steinar H. Gunderson\n");
 825         printf("License GPLv2+: GNU GPL version 2 or later <https://gnu.org/licenses/gpl.html>.\n");
 826         printf("This is free software: you are free to change and redistribute it.\n");
 827         printf("There is NO WARRANTY, to the extent permitted by law.\n");
 828         exit(0);
 829 }
 830
 831 int main(int argc, char **argv)
 832 {
 833         vector<string> dbpaths;
 834
 835         constexpr int EXTENDED_REGEX = 1000;
 836         constexpr int FLUSH_CACHE = 1001;
 837         static const struct option long_options[] = {
 838                 { "help", no_argument, 0, 'h' },
 839                 { "count", no_argument, 0, 'c' },
 840                 { "all", no_argument, 0, 'A' },
 841                 { "basename", no_argument, 0, 'b' },
 842                 { "database", required_argument, 0, 'd' },
 843                 { "existing", no_argument, 0, 'e' },
 844                 { "ignore-case", no_argument, 0, 'i' },
 845                 { "limit", required_argument, 0, 'l' },
 846                 { "literal", no_argument, 0, 'N' },
 847                 { "null", no_argument, 0, '0' },
 848                 { "version", no_argument, 0, 'V' },
 849                 { "regexp", no_argument, 0, 'r' },
 850                 { "regex", no_argument, 0, EXTENDED_REGEX },
 851                 { "wholename", no_argument, 0, 'w' },
 852                 { "debug", no_argument, 0, 'D' },  // Not documented.
 853                 // Enable to test cold-cache behavior (except for access()). Not documented.
 854                 { "flush-cache", no_argument, 0, FLUSH_CACHE },
 855                 { 0, 0, 0, 0 }
 856         };
 857
 858         setlocale(LC_ALL, "");
 859         for (;;) {
 860                 int option_index = 0;
 861                 int c = getopt_long(argc, argv, "Abcd:ehil:n:N0rwVD", long_options, &option_index);
 862                 if (c == -1) {
 863                         break;
 864                 }
 865                 switch (c) {
 866                 case 'A':
 867                         // Ignored.
 868                         break;
 869                 case 'b':
 870                         match_basename = true;
 871                         break;
 872                 case 'c':
 873                         only_count = true;
 874                         break;
 875                 case 'd':
 876                         parse_dbpaths(optarg, &dbpaths);
 877                         break;
 878                 case 'e':
 879                         check_existence = true;
 880                         break;
 881                 case 'h':
 882                         usage();
 883                         exit(0);
 884                 case 'i':
 885                         ignore_case = true;
 886                         break;
 887                 case 'l':
 888                 case 'n':
 889                         limit_matches = limit_left = atoll(optarg);
 890                         if (limit_matches <= 0) {
 891                                 fprintf(stderr, "Error: limit must be a strictly positive number.\n");
 892                                 exit(1);
 893                         }
 894                         break;
 895                 case 'N':
 896                         literal_printing = true;
 897                         break;
 898                 case '0':
 899                         print_nul = true;
 900                         break;
 901                 case 'r':
 902                         patterns_are_regex = true;
 903                         break;
 904                 case EXTENDED_REGEX:
 905                         patterns_are_regex = true;
 906                         use_extended_regex = true;
 907                         break;
 908                 case 'w':
 909                         match_basename = false;  // No-op unless -b is given first.
 910                         break;
 911                 case 'D':
 912                         use_debug = true;
 913                         break;
 914                 case FLUSH_CACHE:
 915                         flush_cache = true;
 916                         break;
 917                 case 'V':
 918                         version();
 919                         break;
 920                 default:
 921                         exit(1);
 922                 }
 923         }
 924
 925         if (use_debug || flush_cache) {
 926                 // Debug information would leak information about which files exist,
 927                 // so drop setgid before we open the file; one would either need to run
 928                 // as root, or use a locally-built file. Doing the same thing for
 929                 // flush_cache is mostly paranoia, in an attempt to prevent random users
 930                 // from making plocate slow for everyone else.
 931                 if (setgid(getgid()) != 0) {
 932                         perror("setgid");
 933                         exit(EXIT_FAILURE);
 934                 }
 935         }
 936
 937         if (!print_nul) {
 938                 stdout_is_tty = isatty(1);
 939         }
 940
 941         vector<Needle> needles;
 942         for (int i = optind; i < argc; ++i) {
 943                 Needle needle;
 944                 needle.str = argv[i];
 945
 946                 // See if there are any wildcard characters, which indicates we should treat it
 947                 // as an (anchored) glob.
 948                 bool any_wildcard = false;
 949                 for (size_t i = 0; i < needle.str.size(); i += read_unigram(needle.str, i).second) {
 950                         if (read_unigram(needle.str, i).first == WILDCARD_UNIGRAM) {
 951                                 any_wildcard = true;
 952                                 break;
 953                         }
 954                 }
 955
 956                 if (patterns_are_regex) {
 957                         needle.type = Needle::REGEX;
 958                         needle.re = compile_regex(needle.str);
 959                 } else if (any_wildcard) {
 960                         needle.type = Needle::GLOB;
 961                 } else if (ignore_case) {
 962                         // strcasestr() doesn't handle locales correctly (even though LSB
 963                         // claims it should), but somehow, fnmatch() does, and it's about
 964                         // the same speed as using a regex.
 965                         needle.type = Needle::GLOB;
 966                         needle.str = "*" + needle.str + "*";
 967                 } else {
 968                         needle.type = Needle::STRSTR;
 969                         needle.str = unescape_glob_to_plain_string(needle.str);
 970                 }
 971                 needles.push_back(move(needle));
 972         }
 973         if (needles.empty()) {
 974                 fprintf(stderr, "plocate: no pattern to search for specified\n");
 975                 exit(0);
 976         }
 977
 978         if (dbpaths.empty()) {
 979                 // No -d given, so use our default. Note that this happens
 980                 // even if LOCATE_PATH exists, to match mlocate behavior.
 981                 dbpaths.push_back(DBFILE);
 982         }
 983
 984         const char *locate_path = getenv("LOCATE_PATH");
 985         if (locate_path != nullptr) {
 986                 parse_dbpaths(locate_path, &dbpaths);
 987         }
 988
 989         uint64_t matched = 0;
 990         for (size_t i = 0; i < dbpaths.size(); ++i) {
 991                 uint64_t this_matched;
 992                 if (i != dbpaths.size() - 1) {
 993                         this_matched = do_search_file_in_child(needles, dbpaths[i]);
 994                 } else {
 995                         this_matched = do_search_file(needles, dbpaths[i]);
 996                 }
 997                 matched += this_matched;
 998                 limit_left -= this_matched;
 999         }
1000         if (only_count) {
1001                 printf("%" PRId64 "\n", matched);
1002         }
1003 }