git.sesse.net Git - plocate/blob - plocate.cpp

   1 #include "db.h"
   2 #include "dprintf.h"
   3 #include "io_uring_engine.h"
   4 #include "parse_trigrams.h"
   5 #include "serializer.h"
   6 #include "turbopfor.h"
   7 #include "unique_sort.h"
   8
   9 #include <algorithm>
  10 #include <assert.h>
  11 #include <atomic>
  12 #include <chrono>
  13 #include <condition_variable>
  14 #include <deque>
  15 #include <fcntl.h>
  16 #include <fnmatch.h>
  17 #include <functional>
  18 #include <getopt.h>
  19 #include <inttypes.h>
  20 #include <iterator>
  21 #include <limits>
  22 #include <locale.h>
  23 #include <map>
  24 #include <memory>
  25 #include <mutex>
  26 #include <regex.h>
  27 #include <stdio.h>
  28 #include <stdlib.h>
  29 #include <stdint.h>
  30 #include <string.h>
  31 #include <string>
  32 #include <string_view>
  33 #include <thread>
  34 #include <tuple>
  35 #include <unistd.h>
  36 #include <unordered_map>
  37 #include <unordered_set>
  38 #include <utility>
  39 #include <vector>
  40 #include <zstd.h>
  41
  42 using namespace std;
  43 using namespace std::chrono;
  44
  45 #define DEFAULT_DBPATH "/var/lib/mlocate/plocate.db"
  46
  47 const char *dbpath = DEFAULT_DBPATH;
  48 bool ignore_case = false;
  49 bool only_count = false;
  50 bool print_nul = false;
  51 bool use_debug = false;
  52 bool patterns_are_regex = false;
  53 bool use_extended_regex = false;
  54 int64_t limit_matches = numeric_limits<int64_t>::max();
  55 int64_t limit_left = numeric_limits<int64_t>::max();
  56
  57 steady_clock::time_point start;
  58 ZSTD_DDict *ddict = nullptr;
  59
  60 regex_t compile_regex(const string &needle);
  61
  62 struct Needle {
  63         enum { STRSTR,
  64                REGEX,
  65                GLOB } type;
  66         string str;  // Filled in no matter what.
  67         regex_t re;  // For REGEX.
  68 };
  69
  70 bool matches(const Needle &needle, const char *haystack)
  71 {
  72         if (needle.type == Needle::STRSTR) {
  73                 return strstr(haystack, needle.str.c_str()) != nullptr;
  74         } else if (needle.type == Needle::GLOB) {
  75                 int flags = ignore_case ? FNM_CASEFOLD : 0;
  76                 return fnmatch(needle.str.c_str(), haystack, flags) == 0;
  77         } else {
  78                 assert(needle.type == Needle::REGEX);
  79                 return regexec(&needle.re, haystack, /*nmatch=*/0, /*pmatch=*/nullptr, /*flags=*/0) == 0;
  80         }
  81 }
  82
  83 class AccessRXCache {
  84 public:
  85         AccessRXCache(IOUringEngine *engine)
  86                 : engine(engine) {}
  87         void check_access(const char *filename, bool allow_async, function<void(bool)> cb);
  88
  89 private:
  90         unordered_map<string, bool> cache;
  91         struct PendingStat {
  92                 string filename;
  93                 function<void(bool)> cb;
  94         };
  95         map<string, vector<PendingStat>> pending_stats;
  96         IOUringEngine *engine;
  97         mutex mu;
  98 };
  99
 100 void AccessRXCache::check_access(const char *filename, bool allow_async, function<void(bool)> cb)
 101 {
 102         lock_guard<mutex> lock(mu);
 103         if (engine == nullptr || !engine->get_supports_stat()) {
 104                 allow_async = false;
 105         }
 106
 107         for (const char *end = strchr(filename + 1, '/'); end != nullptr; end = strchr(end + 1, '/')) {
 108                 string parent_path(filename, end - filename);  // string_view from C++20.
 109                 auto cache_it = cache.find(parent_path);
 110                 if (cache_it != cache.end()) {
 111                         // Found in the cache.
 112                         if (!cache_it->second) {
 113                                 cb(false);
 114                                 return;
 115                         }
 116                         continue;
 117                 }
 118
 119                 if (!allow_async) {
 120                         bool ok = access(parent_path.c_str(), R_OK | X_OK) == 0;
 121                         cache.emplace(parent_path, ok);
 122                         if (!ok) {
 123                                 cb(false);
 124                                 return;
 125                         }
 126                         continue;
 127                 }
 128
 129                 // We want to call access(), but it could block on I/O. io_uring doesn't support
 130                 // access(), but we can do a dummy asynchonous statx() to populate the kernel's cache,
 131                 // which nearly always makes the next access() instantaneous.
 132
 133                 // See if there's already a pending stat that matches this,
 134                 // or is a subdirectory.
 135                 auto it = pending_stats.lower_bound(parent_path);
 136                 if (it != pending_stats.end() && it->first.size() >= parent_path.size() &&
 137                     it->first.compare(0, parent_path.size(), parent_path) == 0) {
 138                         it->second.emplace_back(PendingStat{ filename, move(cb) });
 139                 } else {
 140                         it = pending_stats.emplace(filename, vector<PendingStat>{}).first;
 141                         engine->submit_stat(filename, [this, it, filename{ strdup(filename) }, cb{ move(cb) }] {
 142                                 // The stat returned, so now do the actual access() calls.
 143                                 // All of them should be in cache, so don't fire off new statx()
 144                                 // calls during that check.
 145                                 check_access(filename, /*allow_async=*/false, move(cb));
 146                                 free(filename);
 147
 148                                 // Call all others that waited for the same stat() to finish.
 149                                 // They may fire off new stat() calls if needed.
 150                                 vector<PendingStat> pending = move(it->second);
 151                                 pending_stats.erase(it);
 152                                 for (PendingStat &ps : pending) {
 153                                         check_access(ps.filename.c_str(), /*allow_async=*/true, move(ps.cb));
 154                                 }
 155                         });
 156                 }
 157                 return;  // The rest will happen in async context.
 158         }
 159
 160         // Passed all checks.
 161         cb(true);
 162 }
 163
 164 class Corpus {
 165 public:
 166         Corpus(int fd, IOUringEngine *engine);
 167         ~Corpus();
 168         void find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb);
 169         void get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const;
 170         size_t get_num_filename_blocks() const;
 171         off_t offset_for_block(uint32_t docid) const
 172         {
 173                 return hdr.filename_index_offset_bytes + docid * sizeof(uint64_t);
 174         }
 175         const Header &get_hdr() const { return hdr; }
 176
 177 public:
 178         const int fd;
 179         IOUringEngine *const engine;
 180
 181         Header hdr;
 182 };
 183
 184 Corpus::Corpus(int fd, IOUringEngine *engine)
 185         : fd(fd), engine(engine)
 186 {
 187         // Enable to test cold-cache behavior (except for access()).
 188         if (false) {
 189                 off_t len = lseek(fd, 0, SEEK_END);
 190                 if (len == -1) {
 191                         perror("lseek");
 192                         exit(1);
 193                 }
 194                 posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED);
 195         }
 196
 197         complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
 198         if (memcmp(hdr.magic, "\0plocate", 8) != 0) {
 199                 fprintf(stderr, "plocate.db is corrupt or an old version; please rebuild it.\n");
 200                 exit(1);
 201         }
 202         if (hdr.version != 0 && hdr.version != 1) {
 203                 fprintf(stderr, "plocate.db has version %u, expected 0 or 1; please rebuild it.\n", hdr.version);
 204                 exit(1);
 205         }
 206         if (hdr.version == 0) {
 207                 // These will be junk data.
 208                 hdr.zstd_dictionary_offset_bytes = 0;
 209                 hdr.zstd_dictionary_length_bytes = 0;
 210         }
 211 }
 212
 213 Corpus::~Corpus()
 214 {
 215         close(fd);
 216 }
 217
 218 void Corpus::find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb)
 219 {
 220         uint32_t bucket = hash_trigram(trgm, hdr.hashtable_size);
 221         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) {
 222                 const Trigram *trgmptr = reinterpret_cast<const Trigram *>(s.data());
 223                 for (unsigned i = 0; i < hdr.extra_ht_slots + 1; ++i) {
 224                         if (trgmptr[i].trgm == trgm) {
 225                                 cb(trgmptr + i, trgmptr[i + 1].offset - trgmptr[i].offset);
 226                                 return;
 227                         }
 228                 }
 229
 230                 // Not found.
 231                 cb(nullptr, 0);
 232         });
 233 }
 234
 235 void Corpus::get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const
 236 {
 237         // Read the file offset from this docid and the next one.
 238         // This is always allowed, since we have a sentinel block at the end.
 239         engine->submit_read(fd, sizeof(uint64_t) * 2, offset_for_block(docid), [this, cb{ move(cb) }](string_view s) {
 240                 const uint64_t *ptr = reinterpret_cast<const uint64_t *>(s.data());
 241                 off_t offset = ptr[0];
 242                 size_t len = ptr[1] - ptr[0];
 243                 engine->submit_read(fd, len, offset, cb);
 244         });
 245 }
 246
 247 size_t Corpus::get_num_filename_blocks() const
 248 {
 249         return hdr.num_docids;
 250 }
 251
 252 void scan_file_block(const vector<Needle> &needles, string_view compressed,
 253                      AccessRXCache *access_rx_cache, uint64_t seq, ResultReceiver *serializer,
 254                      atomic<uint64_t> *matched)
 255 {
 256         unsigned long long uncompressed_len = ZSTD_getFrameContentSize(compressed.data(), compressed.size());
 257         if (uncompressed_len == ZSTD_CONTENTSIZE_UNKNOWN || uncompressed_len == ZSTD_CONTENTSIZE_ERROR) {
 258                 fprintf(stderr, "ZSTD_getFrameContentSize() failed\n");
 259                 exit(1);
 260         }
 261
 262         string block;
 263         block.resize(uncompressed_len + 1);
 264
 265         static thread_local ZSTD_DCtx *ctx = ZSTD_createDCtx();  // Reused across calls.
 266         size_t err;
 267
 268         if (ddict != nullptr) {
 269                 err = ZSTD_decompress_usingDDict(ctx, &block[0], block.size(), compressed.data(),
 270                                                  compressed.size(), ddict);
 271         } else {
 272                 err = ZSTD_decompressDCtx(ctx, &block[0], block.size(), compressed.data(),
 273                                           compressed.size());
 274         }
 275         if (ZSTD_isError(err)) {
 276                 fprintf(stderr, "ZSTD_decompress(): %s\n", ZSTD_getErrorName(err));
 277                 exit(1);
 278         }
 279         block[block.size() - 1] = '\0';
 280
 281         auto test_candidate = [&](const char *filename, uint64_t local_seq, uint64_t next_seq) {
 282                 access_rx_cache->check_access(filename, /*allow_async=*/true, [matched, serializer, local_seq, next_seq, filename{ strdup(filename) }](bool ok) {
 283                         if (ok) {
 284                                 ++*matched;
 285                                 serializer->print(local_seq, next_seq - local_seq, filename);
 286                         } else {
 287                                 serializer->print(local_seq, next_seq - local_seq, "");
 288                         }
 289                         free(filename);
 290                 });
 291         };
 292
 293         // We need to know the next sequence number before inserting into Serializer,
 294         // so always buffer one candidate.
 295         const char *pending_candidate = nullptr;
 296
 297         uint64_t local_seq = seq << 32;
 298         for (const char *filename = block.data();
 299              filename != block.data() + block.size();
 300              filename += strlen(filename) + 1) {
 301                 bool found = true;
 302                 for (const Needle &needle : needles) {
 303                         if (!matches(needle, filename)) {
 304                                 found = false;
 305                                 break;
 306                         }
 307                 }
 308                 if (found) {
 309                         if (pending_candidate != nullptr) {
 310                                 test_candidate(pending_candidate, local_seq, local_seq + 1);
 311                                 ++local_seq;
 312                         }
 313                         pending_candidate = filename;
 314                 }
 315         }
 316         if (pending_candidate == nullptr) {
 317                 serializer->print(seq << 32, 1ULL << 32, "");
 318         } else {
 319                 test_candidate(pending_candidate, local_seq, (seq + 1) << 32);
 320         }
 321 }
 322
 323 size_t scan_docids(const vector<Needle> &needles, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
 324 {
 325         Serializer docids_in_order;
 326         AccessRXCache access_rx_cache(engine);
 327         atomic<uint64_t> matched{ 0 };
 328         for (size_t i = 0; i < docids.size(); ++i) {
 329                 uint32_t docid = docids[i];
 330                 corpus.get_compressed_filename_block(docid, [i, &matched, &needles, &access_rx_cache, &docids_in_order](string_view compressed) {
 331                         scan_file_block(needles, compressed, &access_rx_cache, i, &docids_in_order, &matched);
 332                 });
 333         }
 334         engine->finish();
 335         return matched;
 336 }
 337
 338 struct WorkerThread {
 339         thread t;
 340
 341         // We use a result queue instead of synchronizing Serializer,
 342         // since a lock on it becomes a huge choke point if there are
 343         // lots of threads.
 344         mutex result_mu;
 345         struct Result {
 346                 uint64_t seq;
 347                 uint64_t skip;
 348                 string msg;
 349         };
 350         vector<Result> results;
 351 };
 352
 353 class WorkerThreadReceiver : public ResultReceiver {
 354 public:
 355         WorkerThreadReceiver(WorkerThread *wt)
 356                 : wt(wt) {}
 357
 358         void print(uint64_t seq, uint64_t skip, const string msg) override
 359         {
 360                 lock_guard<mutex> lock(wt->result_mu);
 361                 if (msg.empty() && !wt->results.empty() && wt->results.back().seq + wt->results.back().skip == seq) {
 362                         wt->results.back().skip += skip;
 363                 } else {
 364                         wt->results.emplace_back(WorkerThread::Result{ seq, skip, move(msg) });
 365                 }
 366         }
 367
 368 private:
 369         WorkerThread *wt;
 370 };
 371
 372 void deliver_results(WorkerThread *wt, Serializer *serializer)
 373 {
 374         vector<WorkerThread::Result> results;
 375         {
 376                 lock_guard<mutex> lock(wt->result_mu);
 377                 results = move(wt->results);
 378         }
 379         for (const WorkerThread::Result &result : results) {
 380                 serializer->print(result.seq, result.skip, move(result.msg));
 381         }
 382 }
 383
 384 // We do this sequentially, as it's faster than scattering
 385 // a lot of I/O through io_uring and hoping the kernel will
 386 // coalesce it plus readahead for us. Since we assume that
 387 // we will primarily be CPU-bound, we'll be firing up one
 388 // worker thread for each spare core (the last one will
 389 // only be doing I/O). access() is still synchronous.
 390 uint64_t scan_all_docids(const vector<Needle> &needles, int fd, const Corpus &corpus)
 391 {
 392         {
 393                 const Header &hdr = corpus.get_hdr();
 394                 if (hdr.zstd_dictionary_length_bytes > 0) {
 395                         string dictionary;
 396                         dictionary.resize(hdr.zstd_dictionary_length_bytes);
 397                         complete_pread(fd, &dictionary[0], hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes);
 398                         ddict = ZSTD_createDDict(dictionary.data(), dictionary.size());
 399                 }
 400         }
 401
 402         AccessRXCache access_rx_cache(nullptr);
 403         Serializer serializer;
 404         uint32_t num_blocks = corpus.get_num_filename_blocks();
 405         unique_ptr<uint64_t[]> offsets(new uint64_t[num_blocks + 1]);
 406         complete_pread(fd, offsets.get(), (num_blocks + 1) * sizeof(uint64_t), corpus.offset_for_block(0));
 407         atomic<uint64_t> matched{0};
 408
 409         mutex mu;
 410         condition_variable queue_added, queue_removed;
 411         deque<tuple<int, int, string>> work_queue;  // Under mu.
 412         bool done = false;  // Under mu.
 413
 414         unsigned num_threads = max<int>(sysconf(_SC_NPROCESSORS_ONLN) - 1, 1);
 415         dprintf("Using %u worker threads for linear scan.\n", num_threads);
 416         unique_ptr<WorkerThread[]> threads(new WorkerThread[num_threads]);
 417         for (unsigned i = 0; i < num_threads; ++i) {
 418                 threads[i].t = thread([&threads, &mu, &queue_added, &queue_removed, &work_queue, &done, &offsets, &needles, &access_rx_cache, &matched, i] {
 419                         // regcomp() takes a lock on the regex, so each thread will need its own.
 420                         const vector<Needle> *use_needles = &needles;
 421                         vector<Needle> recompiled_needles;
 422                         if (i != 0 && patterns_are_regex) {
 423                                 recompiled_needles = needles;
 424                                 for (Needle &needle : recompiled_needles) {
 425                                         needle.re = compile_regex(needle.str);
 426                                 }
 427                                 use_needles = &recompiled_needles;
 428                         }
 429
 430                         WorkerThreadReceiver receiver(&threads[i]);
 431                         for (;;) {
 432                                 uint32_t io_docid, last_docid;
 433                                 string compressed;
 434
 435                                 {
 436                                         unique_lock<mutex> lock(mu);
 437                                         queue_added.wait(lock, [&work_queue, &done] { return !work_queue.empty() || done; });
 438                                         if (done && work_queue.empty()) {
 439                                                 return;
 440                                         }
 441                                         tie(io_docid, last_docid, compressed) = move(work_queue.front());
 442                                         work_queue.pop_front();
 443                                         queue_removed.notify_all();
 444                                 }
 445
 446                                 for (uint32_t docid = io_docid; docid < last_docid; ++docid) {
 447                                         size_t relative_offset = offsets[docid] - offsets[io_docid];
 448                                         size_t len = offsets[docid + 1] - offsets[docid];
 449                                         scan_file_block(*use_needles, { &compressed[relative_offset], len }, &access_rx_cache, docid, &receiver, &matched);
 450                                 }
 451                         }
 452                 });
 453         }
 454
 455         string compressed;
 456         for (uint32_t io_docid = 0; io_docid < num_blocks; io_docid += 32) {
 457                 uint32_t last_docid = std::min(io_docid + 32, num_blocks);
 458                 size_t io_len = offsets[last_docid] - offsets[io_docid];
 459                 if (compressed.size() < io_len) {
 460                         compressed.resize(io_len);
 461                 }
 462                 complete_pread(fd, &compressed[0], io_len, offsets[io_docid]);
 463
 464                 {
 465                         unique_lock<mutex> lock(mu);
 466                         queue_removed.wait(lock, [&work_queue] { return work_queue.size() < 256; });  // Allow ~2MB of data queued up.
 467                         work_queue.emplace_back(io_docid, last_docid, move(compressed));
 468                         queue_added.notify_one();  // Avoid the thundering herd.
 469                 }
 470
 471                 // Pick up some results, so that we are sure that we won't just overload.
 472                 // (Seemingly, going through all of these causes slowness with many threads,
 473                 // but taking only one is OK.)
 474                 unsigned i = io_docid / 32;
 475                 deliver_results(&threads[i % num_threads], &serializer);
 476         }
 477         {
 478                 lock_guard<mutex> lock(mu);
 479                 done = true;
 480                 queue_added.notify_all();
 481         }
 482         for (unsigned i = 0; i < num_threads; ++i) {
 483                 threads[i].t.join();
 484                 deliver_results(&threads[i], &serializer);
 485         }
 486         return matched;
 487 }
 488
 489 // Takes the given posting list, unions it into the parts of the trigram disjunction
 490 // already read; if the list is complete, intersects with “cur_candidates”.
 491 //
 492 // Returns true if the search should be aborted (we are done).
 493 bool new_posting_list_read(TrigramDisjunction *td, vector<uint32_t> decoded, vector<uint32_t> *cur_candidates, vector<uint32_t> *tmp)
 494 {
 495         if (td->docids.empty()) {
 496                 td->docids = move(decoded);
 497         } else {
 498                 tmp->clear();
 499                 set_union(decoded.begin(), decoded.end(), td->docids.begin(), td->docids.end(), back_inserter(*tmp));
 500                 swap(*tmp, td->docids);
 501         }
 502         if (--td->remaining_trigrams_to_read > 0) {
 503                 // Need to wait for more.
 504                 if (ignore_case) {
 505                         dprintf("  ... %u reads left in OR group %u (%zu docids in list)\n",
 506                                 td->remaining_trigrams_to_read, td->index, td->docids.size());
 507                 }
 508                 return false;
 509         }
 510         if (cur_candidates->empty()) {
 511                 if (ignore_case) {
 512                         dprintf("  ... all reads done for OR group %u (%zu docids)\n",
 513                                 td->index, td->docids.size());
 514                 }
 515                 *cur_candidates = move(td->docids);
 516         } else {
 517                 tmp->clear();
 518                 set_intersection(cur_candidates->begin(), cur_candidates->end(),
 519                                  td->docids.begin(), td->docids.end(),
 520                                  back_inserter(*tmp));
 521                 swap(*cur_candidates, *tmp);
 522                 if (ignore_case) {
 523                         if (cur_candidates->empty()) {
 524                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (none left, search is done)\n",
 525                                         td->index, td->docids.size());
 526                                 return true;
 527                         } else {
 528                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (%zu left)\n",
 529                                         td->index, td->docids.size(), cur_candidates->size());
 530                         }
 531                 }
 532         }
 533         return false;
 534 }
 535
 536 void do_search_file(const vector<Needle> &needles, const char *filename)
 537 {
 538         int fd = open(filename, O_RDONLY);
 539         if (fd == -1) {
 540                 perror(filename);
 541                 exit(1);
 542         }
 543
 544         // Drop privileges.
 545         if (setgid(getgid()) != 0) {
 546                 perror("setgid");
 547                 exit(EXIT_FAILURE);
 548         }
 549
 550         start = steady_clock::now();
 551         if (access("/", R_OK | X_OK)) {
 552                 // We can't find anything, no need to bother...
 553                 return;
 554         }
 555
 556         IOUringEngine engine(/*slop_bytes=*/16);  // 16 slop bytes as described in turbopfor.h.
 557         Corpus corpus(fd, &engine);
 558         dprintf("Corpus init done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 559
 560         vector<TrigramDisjunction> trigram_groups;
 561         if (patterns_are_regex) {
 562                 // We could parse the regex to find trigrams that have to be there
 563                 // (there are actually known algorithms to deal with disjunctions
 564                 // and such, too), but for now, we just go brute force.
 565                 // Using locate with regexes is pretty niche.
 566         } else {
 567                 for (const Needle &needle : needles) {
 568                         parse_trigrams(needle.str, ignore_case, &trigram_groups);
 569                 }
 570         }
 571
 572         unique_sort(
 573                 &trigram_groups,
 574                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives < b.trigram_alternatives; },
 575                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives == b.trigram_alternatives; });
 576
 577         // Give them names for debugging.
 578         unsigned td_index = 0;
 579         for (TrigramDisjunction &td : trigram_groups) {
 580                 td.index = td_index++;
 581         }
 582
 583         // Collect which trigrams we need to look up in the hash table.
 584         unordered_map<uint32_t, vector<TrigramDisjunction *>> trigrams_to_lookup;
 585         for (TrigramDisjunction &td : trigram_groups) {
 586                 for (uint32_t trgm : td.trigram_alternatives) {
 587                         trigrams_to_lookup[trgm].push_back(&td);
 588                 }
 589         }
 590         if (trigrams_to_lookup.empty()) {
 591                 // Too short for trigram matching. Apply brute force.
 592                 // (We could have searched through all trigrams that matched
 593                 // the pattern and done a union of them, but that's a lot of
 594                 // work for fairly unclear gain.)
 595                 uint64_t matched = scan_all_docids(needles, fd, corpus);
 596                 if (only_count) {
 597                         printf("%" PRId64 "\n", matched);
 598                 }
 599                 return;
 600         }
 601
 602         // Sneak in fetching the dictionary, if present. It's not necessarily clear
 603         // exactly where it would be cheapest to get it, but it needs to be present
 604         // before we can decode any of the posting lists. Most likely, it's
 605         // in the same filesystem block as the header anyway, so it should be
 606         // present in the cache.
 607         {
 608                 const Header &hdr = corpus.get_hdr();
 609                 if (hdr.zstd_dictionary_length_bytes > 0) {
 610                         engine.submit_read(fd, hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes, [](string_view s) {
 611                                 ddict = ZSTD_createDDict(s.data(), s.size());
 612                                 dprintf("Dictionary initialized after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 613                         });
 614                 }
 615         }
 616
 617         // Look them all up on disk.
 618         for (auto &[trgm, trigram_groups] : trigrams_to_lookup) {
 619                 corpus.find_trigram(trgm, [trgm{ trgm }, trigram_groups{ &trigram_groups }](const Trigram *trgmptr, size_t len) {
 620                         if (trgmptr == nullptr) {
 621                                 dprintf("trigram %s isn't found\n", print_trigram(trgm).c_str());
 622                                 for (TrigramDisjunction *td : *trigram_groups) {
 623                                         --td->remaining_trigrams_to_read;
 624                                         if (td->remaining_trigrams_to_read == 0 && td->read_trigrams.empty()) {
 625                                                 dprintf("zero matches in %s, so we are done\n", print_td(*td).c_str());
 626                                                 if (only_count) {
 627                                                         printf("0\n");
 628                                                 }
 629                                                 exit(0);
 630                                         }
 631                                 }
 632                                 return;
 633                         }
 634                         for (TrigramDisjunction *td : *trigram_groups) {
 635                                 --td->remaining_trigrams_to_read;
 636                                 td->max_num_docids += trgmptr->num_docids;
 637                                 td->read_trigrams.emplace_back(*trgmptr, len);
 638                         }
 639                 });
 640         }
 641         engine.finish();
 642         dprintf("Hashtable lookups done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 643
 644         for (TrigramDisjunction &td : trigram_groups) {
 645                 // Reset for reads.
 646                 td.remaining_trigrams_to_read = td.read_trigrams.size();
 647
 648                 if (ignore_case) {  // If case-sensitive, they'll all be pretty obvious single-entry groups.
 649                         dprintf("OR group %u (max_num_docids=%u): %s\n", td.index, td.max_num_docids, print_td(td).c_str());
 650                 }
 651         }
 652
 653         // TODO: For case-insensitive (ie. more than one alternative in each),
 654         // prioritize the ones with fewer seeks?
 655         sort(trigram_groups.begin(), trigram_groups.end(),
 656              [&](const TrigramDisjunction &a, const TrigramDisjunction &b) {
 657                      return a.max_num_docids < b.max_num_docids;
 658              });
 659
 660         unordered_map<uint32_t, vector<TrigramDisjunction *>> uses_trigram;
 661         for (TrigramDisjunction &td : trigram_groups) {
 662                 for (uint32_t trgm : td.trigram_alternatives) {
 663                         uses_trigram[trgm].push_back(&td);
 664                 }
 665         }
 666
 667         unordered_set<uint32_t> trigrams_submitted_read;
 668         vector<uint32_t> cur_candidates, tmp, decoded;
 669         bool done = false;
 670         for (TrigramDisjunction &td : trigram_groups) {
 671                 if (!cur_candidates.empty() && td.max_num_docids > cur_candidates.size() * 100) {
 672                         dprintf("%s has up to %u entries, ignoring the rest (will "
 673                                 "weed out false positives later)\n",
 674                                 print_td(td).c_str(), td.max_num_docids);
 675                         break;
 676                 }
 677
 678                 for (auto &[trgmptr, len] : td.read_trigrams) {
 679                         if (trigrams_submitted_read.count(trgmptr.trgm) != 0) {
 680                                 continue;
 681                         }
 682                         trigrams_submitted_read.insert(trgmptr.trgm);
 683                         // Only stay a certain amount ahead, so that we don't spend I/O
 684                         // on reading the latter, large posting lists. We are unlikely
 685                         // to need them anyway, even if they should come in first.
 686                         if (engine.get_waiting_reads() >= 5) {
 687                                 engine.finish();
 688                                 if (done)
 689                                         break;
 690                         }
 691                         engine.submit_read(fd, len, trgmptr.offset, [trgmptr{ trgmptr }, len{ len }, &done, &cur_candidates, &tmp, &decoded, &uses_trigram](string_view s) {
 692                                 if (done)
 693                                         return;
 694
 695                                 uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
 696                                 const unsigned char *pldata = reinterpret_cast<const unsigned char *>(s.data());
 697                                 size_t num = trgmptr.num_docids;
 698                                 decoded.resize(num);
 699                                 decode_pfor_delta1_128(pldata, num, /*interleaved=*/true, &decoded[0]);
 700
 701                                 assert(uses_trigram.count(trgm) != 0);
 702                                 bool was_empty = cur_candidates.empty();
 703                                 if (ignore_case) {
 704                                         dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 705                                 }
 706
 707                                 for (TrigramDisjunction *td : uses_trigram[trgm]) {
 708                                         done |= new_posting_list_read(td, decoded, &cur_candidates, &tmp);
 709                                         if (done)
 710                                                 break;
 711                                 }
 712                                 if (!ignore_case) {
 713                                         if (was_empty) {
 714                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 715                                         } else if (cur_candidates.empty()) {
 716                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (none left, search is done)\n", print_trigram(trgm).c_str(), len, num);
 717                                         } else {
 718                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (%zu left)\n", print_trigram(trgm).c_str(), len, num, cur_candidates.size());
 719                                         }
 720                                 }
 721                         });
 722                 }
 723         }
 724         engine.finish();
 725         if (done) {
 726                 return;
 727         }
 728         dprintf("Intersection done after %.1f ms. Doing final verification and printing:\n",
 729                 1e3 * duration<float>(steady_clock::now() - start).count());
 730
 731         uint64_t matched = scan_docids(needles, cur_candidates, corpus, &engine);
 732         dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
 733                 1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 734
 735         if (only_count) {
 736                 printf("%" PRId64 "\n", matched);
 737         }
 738 }
 739
 740 string unescape_glob_to_plain_string(const string &needle)
 741 {
 742         string unescaped;
 743         for (size_t i = 0; i < needle.size(); i += read_unigram(needle, i).second) {
 744                 uint32_t ch = read_unigram(needle, i).first;
 745                 assert(ch != WILDCARD_UNIGRAM);
 746                 if (ch == PREMATURE_END_UNIGRAM) {
 747                         fprintf(stderr, "Pattern '%s' ended prematurely\n", needle.c_str());
 748                         exit(1);
 749                 }
 750                 unescaped.push_back(ch);
 751         }
 752         return unescaped;
 753 }
 754
 755 regex_t compile_regex(const string &needle)
 756 {
 757         regex_t re;
 758         int flags = REG_NOSUB;
 759         if (ignore_case) {
 760                 flags |= REG_ICASE;
 761         }
 762         if (use_extended_regex) {
 763                 flags |= REG_EXTENDED;
 764         }
 765         int err = regcomp(&re, needle.c_str(), flags);
 766         if (err != 0) {
 767                 char errbuf[256];
 768                 regerror(err, &re, errbuf, sizeof(errbuf));
 769                 fprintf(stderr, "Error when compiling regex '%s': %s\n", needle.c_str(), errbuf);
 770                 exit(1);
 771         }
 772         return re;
 773 }
 774
 775 void usage()
 776 {
 777         printf(
 778                 "Usage: plocate [OPTION]... PATTERN...\n"
 779                 "\n"
 780                 "  -c, --count            print number of matches instead of the matches\n"
 781                 "  -d, --database DBPATH  search for files in DBPATH\n"
 782                 "                         (default is " DEFAULT_DBPATH ")\n"
 783                 "  -i, --ignore-case      search case-insensitively\n"
 784                 "  -l, --limit LIMIT      stop after LIMIT matches\n"
 785                 "  -0, --null             delimit matches by NUL instead of newline\n"
 786                 "  -r, --regexp           interpret patterns as basic regexps (slow)\n"
 787                 "      --regex            interpret patterns as extended regexps (slow)\n"
 788                 "      --help             print this help\n"
 789                 "      --version          print version information\n");
 790 }
 791
 792 void version()
 793 {
 794         printf("plocate %s\n", PLOCATE_VERSION);
 795         printf("Copyright 2020 Steinar H. Gunderson\n");
 796         printf("License GPLv2+: GNU GPL version 2 or later <https://gnu.org/licenses/gpl.html>.\n");
 797         printf("This is free software: you are free to change and redistribute it.\n");
 798         printf("There is NO WARRANTY, to the extent permitted by law.\n");
 799         exit(0);
 800 }
 801
 802 int main(int argc, char **argv)
 803 {
 804         constexpr int EXTENDED_REGEX = 1000;
 805         static const struct option long_options[] = {
 806                 { "help", no_argument, 0, 'h' },
 807                 { "count", no_argument, 0, 'c' },
 808                 { "database", required_argument, 0, 'd' },
 809                 { "ignore-case", no_argument, 0, 'i' },
 810                 { "limit", required_argument, 0, 'l' },
 811                 { "null", no_argument, 0, '0' },
 812                 { "version", no_argument, 0, 'V' },
 813                 { "regexp", no_argument, 0, 'r' },
 814                 { "regex", no_argument, 0, EXTENDED_REGEX },
 815                 { "debug", no_argument, 0, 'D' },  // Not documented.
 816                 { 0, 0, 0, 0 }
 817         };
 818
 819         setlocale(LC_ALL, "");
 820         for (;;) {
 821                 int option_index = 0;
 822                 int c = getopt_long(argc, argv, "cd:hil:n:0VD", long_options, &option_index);
 823                 if (c == -1) {
 824                         break;
 825                 }
 826                 switch (c) {
 827                 case 'c':
 828                         only_count = true;
 829                         break;
 830                 case 'd':
 831                         dbpath = strdup(optarg);
 832                         break;
 833                 case 'h':
 834                         usage();
 835                         exit(0);
 836                 case 'i':
 837                         ignore_case = true;
 838                         break;
 839                 case 'l':
 840                 case 'n':
 841                         limit_matches = limit_left = atoll(optarg);
 842                         if (limit_matches <= 0) {
 843                                 fprintf(stderr, "Error: limit must be a strictly positive number.\n");
 844                                 exit(1);
 845                         }
 846                         break;
 847                 case '0':
 848                         print_nul = true;
 849                         break;
 850                 case 'r':
 851                         patterns_are_regex = true;
 852                         break;
 853                 case EXTENDED_REGEX:
 854                         patterns_are_regex = true;
 855                         use_extended_regex = true;
 856                         break;
 857                 case 'D':
 858                         use_debug = true;
 859                         break;
 860                 case 'V':
 861                         version();
 862                         break;
 863                 default:
 864                         exit(1);
 865                 }
 866         }
 867
 868         if (use_debug) {
 869                 // Debug information would leak information about which files exist,
 870                 // so drop setgid before we open the file; one would either need to run
 871                 // as root, or use a locally-built file.
 872                 if (setgid(getgid()) != 0) {
 873                         perror("setgid");
 874                         exit(EXIT_FAILURE);
 875                 }
 876         }
 877
 878         vector<Needle> needles;
 879         for (int i = optind; i < argc; ++i) {
 880                 Needle needle;
 881                 needle.str = argv[i];
 882
 883                 // See if there are any wildcard characters, which indicates we should treat it
 884                 // as an (anchored) glob.
 885                 bool any_wildcard = false;
 886                 for (size_t i = 0; i < needle.str.size(); i += read_unigram(needle.str, i).second) {
 887                         if (read_unigram(needle.str, i).first == WILDCARD_UNIGRAM) {
 888                                 any_wildcard = true;
 889                                 break;
 890                         }
 891                 }
 892
 893                 if (patterns_are_regex) {
 894                         needle.type = Needle::REGEX;
 895                         needle.re = compile_regex(needle.str);
 896                 } else if (any_wildcard) {
 897                         needle.type = Needle::GLOB;
 898                 } else if (ignore_case) {
 899                         // strcasestr() doesn't handle locales correctly (even though LSB
 900                         // claims it should), but somehow, fnmatch() does, and it's about
 901                         // the same speed as using a regex.
 902                         needle.type = Needle::GLOB;
 903                         needle.str = "*" + needle.str + "*";
 904                 } else {
 905                         needle.type = Needle::STRSTR;
 906                         needle.str = unescape_glob_to_plain_string(needle.str);
 907                 }
 908                 needles.push_back(move(needle));
 909         }
 910         if (needles.empty()) {
 911                 fprintf(stderr, "plocate: no pattern to search for specified\n");
 912                 exit(0);
 913         }
 914         do_search_file(needles, dbpath);
 915 }