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 "turbopfor.h"
   6 #include "unique_sort.h"
   7
   8 #include <algorithm>
   9 #include <atomic>
  10 #include <assert.h>
  11 #include <chrono>
  12 #include <condition_variable>
  13 #include <fcntl.h>
  14 #include <fnmatch.h>
  15 #include <functional>
  16 #include <getopt.h>
  17 #include <inttypes.h>
  18 #include <iosfwd>
  19 #include <iterator>
  20 #include <limits>
  21 #include <map>
  22 #include <memory>
  23 #include <mutex>
  24 #include <queue>
  25 #include <regex.h>
  26 #include <stdint.h>
  27 #include <stdio.h>
  28 #include <stdlib.h>
  29 #include <string.h>
  30 #include <string>
  31 #include <string_view>
  32 #include <thread>
  33 #include <unistd.h>
  34 #include <unordered_map>
  35 #include <unordered_set>
  36 #include <utility>
  37 #include <vector>
  38 #include <zstd.h>
  39
  40 using namespace std;
  41 using namespace std::chrono;
  42
  43 #define DEFAULT_DBPATH "/var/lib/mlocate/plocate.db"
  44
  45 const char *dbpath = DEFAULT_DBPATH;
  46 bool ignore_case = false;
  47 bool only_count = false;
  48 bool print_nul = false;
  49 bool use_debug = false;
  50 bool patterns_are_regex = false;
  51 bool use_extended_regex = false;
  52 int64_t limit_matches = numeric_limits<int64_t>::max();
  53 int64_t limit_left = numeric_limits<int64_t>::max();
  54
  55 steady_clock::time_point start;
  56 ZSTD_DDict *ddict = nullptr;
  57
  58 regex_t compile_regex(const string &needle);
  59
  60 void apply_limit()
  61 {
  62         if (--limit_left > 0) {
  63                 return;
  64         }
  65         dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
  66                 1e3 * duration<float>(steady_clock::now() - start).count(), limit_matches);
  67         if (only_count) {
  68                 printf("%" PRId64 "\n", limit_matches);
  69         }
  70         exit(0);
  71 }
  72
  73 class ResultReceiver {
  74 public:
  75         virtual ~ResultReceiver() = default;
  76         virtual void print(uint64_t seq, uint64_t skip, const string msg) = 0;
  77 };
  78
  79 class Serializer : public ResultReceiver {
  80 public:
  81         ~Serializer() { assert(limit_left <= 0 || pending.empty()); }
  82         void print(uint64_t seq, uint64_t skip, const string msg) override;
  83
  84 private:
  85         uint64_t next_seq = 0;
  86         struct Element {
  87                 uint64_t seq, skip;
  88                 string msg;
  89
  90                 bool operator<(const Element &other) const
  91                 {
  92                         return seq > other.seq;
  93                 }
  94         };
  95         priority_queue<Element> pending;
  96 };
  97
  98 void Serializer::print(uint64_t seq, uint64_t skip, const string msg)
  99 {
 100         if (only_count) {
 101                 if (!msg.empty()) {
 102                         apply_limit();
 103                 }
 104                 return;
 105         }
 106
 107         if (next_seq != seq) {
 108                 pending.push(Element{ seq, skip, move(msg) });
 109                 return;
 110         }
 111
 112         if (!msg.empty()) {
 113                 if (print_nul) {
 114                         printf("%s%c", msg.c_str(), 0);
 115                 } else {
 116                         printf("%s\n", msg.c_str());
 117                 }
 118                 apply_limit();
 119         }
 120         next_seq += skip;
 121
 122         // See if any delayed prints can now be dealt with.
 123         while (!pending.empty() && pending.top().seq == next_seq) {
 124                 if (!pending.top().msg.empty()) {
 125                         if (print_nul) {
 126                                 printf("%s%c", pending.top().msg.c_str(), 0);
 127                         } else {
 128                                 printf("%s\n", pending.top().msg.c_str());
 129                         }
 130                         apply_limit();
 131                 }
 132                 next_seq += pending.top().skip;
 133                 pending.pop();
 134         }
 135 }
 136
 137 struct Needle {
 138         enum { STRSTR,
 139                REGEX,
 140                GLOB } type;
 141         string str;  // Filled in no matter what.
 142         regex_t re;  // For REGEX.
 143 };
 144
 145 bool matches(const Needle &needle, const char *haystack)
 146 {
 147         if (needle.type == Needle::STRSTR) {
 148                 return strstr(haystack, needle.str.c_str()) != nullptr;
 149         } else if (needle.type == Needle::GLOB) {
 150                 int flags = ignore_case ? FNM_CASEFOLD : 0;
 151                 return fnmatch(needle.str.c_str(), haystack, flags) == 0;
 152         } else {
 153                 assert(needle.type == Needle::REGEX);
 154                 return regexec(&needle.re, haystack, /*nmatch=*/0, /*pmatch=*/nullptr, /*flags=*/0) == 0;
 155         }
 156 }
 157
 158 class AccessRXCache {
 159 public:
 160         AccessRXCache(IOUringEngine *engine)
 161                 : engine(engine) {}
 162         void check_access(const char *filename, bool allow_async, function<void(bool)> cb);
 163
 164 private:
 165         unordered_map<string, bool> cache;
 166         struct PendingStat {
 167                 string filename;
 168                 function<void(bool)> cb;
 169         };
 170         map<string, vector<PendingStat>> pending_stats;
 171         IOUringEngine *engine;
 172         mutex mu;
 173 };
 174
 175 void AccessRXCache::check_access(const char *filename, bool allow_async, function<void(bool)> cb)
 176 {
 177         lock_guard<mutex> lock(mu);
 178         if (engine == nullptr || !engine->get_supports_stat()) {
 179                 allow_async = false;
 180         }
 181
 182         for (const char *end = strchr(filename + 1, '/'); end != nullptr; end = strchr(end + 1, '/')) {
 183                 string parent_path(filename, end - filename);  // string_view from C++20.
 184                 auto cache_it = cache.find(parent_path);
 185                 if (cache_it != cache.end()) {
 186                         // Found in the cache.
 187                         if (!cache_it->second) {
 188                                 cb(false);
 189                                 return;
 190                         }
 191                         continue;
 192                 }
 193
 194                 if (!allow_async) {
 195                         bool ok = access(parent_path.c_str(), R_OK | X_OK) == 0;
 196                         cache.emplace(parent_path, ok);
 197                         if (!ok) {
 198                                 cb(false);
 199                                 return;
 200                         }
 201                         continue;
 202                 }
 203
 204                 // We want to call access(), but it could block on I/O. io_uring doesn't support
 205                 // access(), but we can do a dummy asynchonous statx() to populate the kernel's cache,
 206                 // which nearly always makes the next access() instantaneous.
 207
 208                 // See if there's already a pending stat that matches this,
 209                 // or is a subdirectory.
 210                 auto it = pending_stats.lower_bound(parent_path);
 211                 if (it != pending_stats.end() && it->first.size() >= parent_path.size() &&
 212                     it->first.compare(0, parent_path.size(), parent_path) == 0) {
 213                         it->second.emplace_back(PendingStat{ filename, move(cb) });
 214                 } else {
 215                         it = pending_stats.emplace(filename, vector<PendingStat>{}).first;
 216                         engine->submit_stat(filename, [this, it, filename{ strdup(filename) }, cb{ move(cb) }] {
 217                                 // The stat returned, so now do the actual access() calls.
 218                                 // All of them should be in cache, so don't fire off new statx()
 219                                 // calls during that check.
 220                                 check_access(filename, /*allow_async=*/false, move(cb));
 221                                 free(filename);
 222
 223                                 // Call all others that waited for the same stat() to finish.
 224                                 // They may fire off new stat() calls if needed.
 225                                 vector<PendingStat> pending = move(it->second);
 226                                 pending_stats.erase(it);
 227                                 for (PendingStat &ps : pending) {
 228                                         check_access(ps.filename.c_str(), /*allow_async=*/true, move(ps.cb));
 229                                 }
 230                         });
 231                 }
 232                 return;  // The rest will happen in async context.
 233         }
 234
 235         // Passed all checks.
 236         cb(true);
 237 }
 238
 239 class Corpus {
 240 public:
 241         Corpus(int fd, IOUringEngine *engine);
 242         ~Corpus();
 243         void find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb);
 244         void get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const;
 245         size_t get_num_filename_blocks() const;
 246         off_t offset_for_block(uint32_t docid) const
 247         {
 248                 return hdr.filename_index_offset_bytes + docid * sizeof(uint64_t);
 249         }
 250         const Header &get_hdr() const { return hdr; }
 251
 252 public:
 253         const int fd;
 254         IOUringEngine *const engine;
 255
 256         Header hdr;
 257 };
 258
 259 Corpus::Corpus(int fd, IOUringEngine *engine)
 260         : fd(fd), engine(engine)
 261 {
 262         // Enable to test cold-cache behavior (except for access()).
 263         if (false) {
 264                 off_t len = lseek(fd, 0, SEEK_END);
 265                 if (len == -1) {
 266                         perror("lseek");
 267                         exit(1);
 268                 }
 269                 posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED);
 270         }
 271
 272         complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
 273         if (memcmp(hdr.magic, "\0plocate", 8) != 0) {
 274                 fprintf(stderr, "plocate.db is corrupt or an old version; please rebuild it.\n");
 275                 exit(1);
 276         }
 277         if (hdr.version != 0 && hdr.version != 1) {
 278                 fprintf(stderr, "plocate.db has version %u, expected 0 or 1; please rebuild it.\n", hdr.version);
 279                 exit(1);
 280         }
 281         if (hdr.version == 0) {
 282                 // These will be junk data.
 283                 hdr.zstd_dictionary_offset_bytes = 0;
 284                 hdr.zstd_dictionary_length_bytes = 0;
 285         }
 286 }
 287
 288 Corpus::~Corpus()
 289 {
 290         close(fd);
 291 }
 292
 293 void Corpus::find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb)
 294 {
 295         uint32_t bucket = hash_trigram(trgm, hdr.hashtable_size);
 296         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) {
 297                 const Trigram *trgmptr = reinterpret_cast<const Trigram *>(s.data());
 298                 for (unsigned i = 0; i < hdr.extra_ht_slots + 1; ++i) {
 299                         if (trgmptr[i].trgm == trgm) {
 300                                 cb(trgmptr + i, trgmptr[i + 1].offset - trgmptr[i].offset);
 301                                 return;
 302                         }
 303                 }
 304
 305                 // Not found.
 306                 cb(nullptr, 0);
 307         });
 308 }
 309
 310 void Corpus::get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const
 311 {
 312         // Read the file offset from this docid and the next one.
 313         // This is always allowed, since we have a sentinel block at the end.
 314         engine->submit_read(fd, sizeof(uint64_t) * 2, offset_for_block(docid), [this, cb{ move(cb) }](string_view s) {
 315                 const uint64_t *ptr = reinterpret_cast<const uint64_t *>(s.data());
 316                 off_t offset = ptr[0];
 317                 size_t len = ptr[1] - ptr[0];
 318                 engine->submit_read(fd, len, offset, cb);
 319         });
 320 }
 321
 322 size_t Corpus::get_num_filename_blocks() const
 323 {
 324         return hdr.num_docids;
 325 }
 326
 327 void scan_file_block(const vector<Needle> &needles, string_view compressed,
 328                      AccessRXCache *access_rx_cache, uint64_t seq, ResultReceiver *serializer,
 329                      atomic<uint64_t> *matched)
 330 {
 331         unsigned long long uncompressed_len = ZSTD_getFrameContentSize(compressed.data(), compressed.size());
 332         if (uncompressed_len == ZSTD_CONTENTSIZE_UNKNOWN || uncompressed_len == ZSTD_CONTENTSIZE_ERROR) {
 333                 fprintf(stderr, "ZSTD_getFrameContentSize() failed\n");
 334                 exit(1);
 335         }
 336
 337         string block;
 338         block.resize(uncompressed_len + 1);
 339
 340         static thread_local ZSTD_DCtx *ctx = ZSTD_createDCtx();  // Reused across calls.
 341         size_t err;
 342
 343         if (ddict != nullptr) {
 344                 err = ZSTD_decompress_usingDDict(ctx, &block[0], block.size(), compressed.data(),
 345                                                  compressed.size(), ddict);
 346         } else {
 347                 err = ZSTD_decompressDCtx(ctx, &block[0], block.size(), compressed.data(),
 348                                           compressed.size());
 349         }
 350         if (ZSTD_isError(err)) {
 351                 fprintf(stderr, "ZSTD_decompress(): %s\n", ZSTD_getErrorName(err));
 352                 exit(1);
 353         }
 354         block[block.size() - 1] = '\0';
 355
 356         auto test_candidate = [&](const char *filename, uint64_t local_seq, uint64_t next_seq) {
 357                 access_rx_cache->check_access(filename, /*allow_async=*/true, [matched, serializer, local_seq, next_seq, filename{ strdup(filename) }](bool ok) {
 358                         if (ok) {
 359                                 ++*matched;
 360                                 serializer->print(local_seq, next_seq - local_seq, filename);
 361                         } else {
 362                                 serializer->print(local_seq, next_seq - local_seq, "");
 363                         }
 364                         free(filename);
 365                 });
 366         };
 367
 368         // We need to know the next sequence number before inserting into Serializer,
 369         // so always buffer one candidate.
 370         const char *pending_candidate = nullptr;
 371
 372         uint64_t local_seq = seq << 32;
 373         for (const char *filename = block.data();
 374              filename != block.data() + block.size();
 375              filename += strlen(filename) + 1) {
 376                 bool found = true;
 377                 for (const Needle &needle : needles) {
 378                         if (!matches(needle, filename)) {
 379                                 found = false;
 380                                 break;
 381                         }
 382                 }
 383                 if (found) {
 384                         if (pending_candidate != nullptr) {
 385                                 test_candidate(pending_candidate, local_seq, local_seq + 1);
 386                                 ++local_seq;
 387                         }
 388                         pending_candidate = filename;
 389                 }
 390         }
 391         if (pending_candidate == nullptr) {
 392                 serializer->print(seq << 32, 1ULL << 32, "");
 393         } else {
 394                 test_candidate(pending_candidate, local_seq, (seq + 1) << 32);
 395         }
 396 }
 397
 398 size_t scan_docids(const vector<Needle> &needles, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
 399 {
 400         Serializer docids_in_order;
 401         AccessRXCache access_rx_cache(engine);
 402         atomic<uint64_t> matched{0};
 403         for (size_t i = 0; i < docids.size(); ++i) {
 404                 uint32_t docid = docids[i];
 405                 corpus.get_compressed_filename_block(docid, [i, &matched, &needles, &access_rx_cache, &docids_in_order](string_view compressed) {
 406                         scan_file_block(needles, compressed, &access_rx_cache, i, &docids_in_order, &matched);
 407                 });
 408         }
 409         engine->finish();
 410         return matched;
 411 }
 412
 413 struct WorkerThread {
 414         thread t;
 415
 416         // We use a result queue instead of synchronizing Serializer,
 417         // since a lock on it becomes a huge choke point if there are
 418         // lots of threads.
 419         mutex result_mu;
 420         struct Result {
 421                 uint64_t seq;
 422                 uint64_t skip;
 423                 string msg;
 424         };
 425         vector<Result> results;
 426 };
 427
 428 class WorkerThreadReceiver : public ResultReceiver {
 429 public:
 430         WorkerThreadReceiver(WorkerThread *wt) : wt(wt) {}
 431
 432         void print(uint64_t seq, uint64_t skip, const string msg) override
 433         {
 434                 lock_guard<mutex> lock(wt->result_mu);
 435                 wt->results.emplace_back(WorkerThread::Result{ seq, skip, move(msg) });
 436         }
 437
 438 private:
 439         WorkerThread *wt;
 440 };
 441
 442 void deliver_results(WorkerThread *wt, Serializer *serializer)
 443 {
 444         vector<WorkerThread::Result> results;
 445         {
 446                 lock_guard<mutex> lock(wt->result_mu);
 447                 results = move(wt->results);
 448         }
 449         for (const WorkerThread::Result &result : results) {
 450                 serializer->print(result.seq, result.skip, move(result.msg));
 451         }
 452 }
 453
 454 // We do this sequentially, as it's faster than scattering
 455 // a lot of I/O through io_uring and hoping the kernel will
 456 // coalesce it plus readahead for us. Since we assume that
 457 // we will primarily be CPU-bound, we'll be firing up one
 458 // worker thread for each spare core (the last one will
 459 // only be doing I/O). access() is still synchronous.
 460 uint64_t scan_all_docids(const vector<Needle> &needles, int fd, const Corpus &corpus)
 461 {
 462         {
 463                 const Header &hdr = corpus.get_hdr();
 464                 if (hdr.zstd_dictionary_length_bytes > 0) {
 465                         string dictionary;
 466                         dictionary.resize(hdr.zstd_dictionary_length_bytes);
 467                         complete_pread(fd, &dictionary[0], hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes);
 468                         ddict = ZSTD_createDDict(dictionary.data(), dictionary.size());
 469                 }
 470         }
 471
 472         AccessRXCache access_rx_cache(nullptr);
 473         Serializer serializer;
 474         uint32_t num_blocks = corpus.get_num_filename_blocks();
 475         unique_ptr<uint64_t[]> offsets(new uint64_t[num_blocks + 1]);
 476         complete_pread(fd, offsets.get(), (num_blocks + 1) * sizeof(uint64_t), corpus.offset_for_block(0));
 477         atomic<uint64_t> matched{0};
 478
 479         mutex mu;
 480         condition_variable queue_added, queue_removed;
 481         deque<tuple<int, int, string>> work_queue;  // Under mu.
 482         bool done = false;  // Under mu.
 483
 484         unsigned num_threads = max<int>(sysconf(_SC_NPROCESSORS_ONLN) - 1, 1);
 485         dprintf("Using %u worker threads for linear scan.\n", num_threads);
 486         unique_ptr<WorkerThread[]> threads(new WorkerThread[num_threads]);
 487         for (unsigned i = 0; i < num_threads; ++i) {
 488                 threads[i].t = thread([&threads, &mu, &queue_added, &queue_removed, &work_queue, &done, &offsets, &needles, &access_rx_cache, &matched, i] {
 489                         // regcomp() takes a lock on the regex, so each thread will need its own.
 490                         const vector<Needle> *use_needles = &needles;
 491                         vector<Needle> recompiled_needles;
 492                         if (i != 0 && patterns_are_regex) {
 493                                 recompiled_needles = needles;
 494                                 for (Needle &needle : recompiled_needles) {
 495                                         needle.re = compile_regex(needle.str);
 496                                 }
 497                                 use_needles = &recompiled_needles;
 498                         }
 499
 500                         WorkerThreadReceiver receiver(&threads[i]);
 501                         for (;;) {
 502                                 uint32_t io_docid, last_docid;
 503                                 string compressed;
 504
 505                                 {
 506                                         unique_lock<mutex> lock(mu);
 507                                         queue_added.wait(lock, [&work_queue, &done] { return !work_queue.empty() || done; });
 508                                         if (done && work_queue.empty()) {
 509                                                 return;
 510                                         }
 511                                         tie(io_docid, last_docid, compressed) = move(work_queue.front());
 512                                         work_queue.pop_front();
 513                                         queue_removed.notify_all();
 514                                 }
 515
 516                                 for (uint32_t docid = io_docid; docid < last_docid; ++docid) {
 517                                         size_t relative_offset = offsets[docid] - offsets[io_docid];
 518                                         size_t len = offsets[docid + 1] - offsets[docid];
 519                                         scan_file_block(*use_needles, { &compressed[relative_offset], len }, &access_rx_cache, docid, &receiver, &matched);
 520                                 }
 521                         }
 522                 });
 523         }
 524
 525         string compressed;
 526         for (uint32_t io_docid = 0; io_docid < num_blocks; io_docid += 32) {
 527                 uint32_t last_docid = std::min(io_docid + 32, num_blocks);
 528                 size_t io_len = offsets[last_docid] - offsets[io_docid];
 529                 if (compressed.size() < io_len) {
 530                         compressed.resize(io_len);
 531                 }
 532                 complete_pread(fd, &compressed[0], io_len, offsets[io_docid]);
 533
 534                 {
 535                         unique_lock<mutex> lock(mu);
 536                         queue_removed.wait(lock, [&work_queue] { return work_queue.size() < 256; });  // Allow ~2MB of data queued up.
 537                         work_queue.emplace_back(io_docid, last_docid, move(compressed));
 538                         queue_added.notify_one();  // Avoid the thundering herd.
 539                 }
 540
 541                 // Pick up some results, so that we are sure that we won't just overload.
 542                 // (Seemingly, going through all of these causes slowness with many threads,
 543                 // but taking only one is OK.)
 544                 unsigned i = io_docid / 32;
 545                 deliver_results(&threads[i % num_threads], &serializer);
 546         }
 547         {
 548                 lock_guard<mutex> lock(mu);
 549                 done = true;
 550                 queue_added.notify_all();
 551         }
 552         for (unsigned i = 0; i < num_threads; ++i) {
 553                 threads[i].t.join();
 554                 deliver_results(&threads[i], &serializer);
 555         }
 556         return matched;
 557 }
 558
 559 // Takes the given posting list, unions it into the parts of the trigram disjunction
 560 // already read; if the list is complete, intersects with “cur_candidates”.
 561 //
 562 // Returns true if the search should be aborted (we are done).
 563 bool new_posting_list_read(TrigramDisjunction *td, vector<uint32_t> decoded, vector<uint32_t> *cur_candidates, vector<uint32_t> *tmp)
 564 {
 565         if (td->docids.empty()) {
 566                 td->docids = move(decoded);
 567         } else {
 568                 tmp->clear();
 569                 set_union(decoded.begin(), decoded.end(), td->docids.begin(), td->docids.end(), back_inserter(*tmp));
 570                 swap(*tmp, td->docids);
 571         }
 572         if (--td->remaining_trigrams_to_read > 0) {
 573                 // Need to wait for more.
 574                 if (ignore_case) {
 575                         dprintf("  ... %u reads left in OR group %u (%zu docids in list)\n",
 576                                 td->remaining_trigrams_to_read, td->index, td->docids.size());
 577                 }
 578                 return false;
 579         }
 580         if (cur_candidates->empty()) {
 581                 if (ignore_case) {
 582                         dprintf("  ... all reads done for OR group %u (%zu docids)\n",
 583                                 td->index, td->docids.size());
 584                 }
 585                 *cur_candidates = move(td->docids);
 586         } else {
 587                 tmp->clear();
 588                 set_intersection(cur_candidates->begin(), cur_candidates->end(),
 589                                  td->docids.begin(), td->docids.end(),
 590                                  back_inserter(*tmp));
 591                 swap(*cur_candidates, *tmp);
 592                 if (ignore_case) {
 593                         if (cur_candidates->empty()) {
 594                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (none left, search is done)\n",
 595                                         td->index, td->docids.size());
 596                                 return true;
 597                         } else {
 598                                 dprintf("  ... all reads done for OR group %u (%zu docids), intersected (%zu left)\n",
 599                                         td->index, td->docids.size(), cur_candidates->size());
 600                         }
 601                 }
 602         }
 603         return false;
 604 }
 605
 606 void do_search_file(const vector<Needle> &needles, const char *filename)
 607 {
 608         int fd = open(filename, O_RDONLY);
 609         if (fd == -1) {
 610                 perror(filename);
 611                 exit(1);
 612         }
 613
 614         // Drop privileges.
 615         if (setgid(getgid()) != 0) {
 616                 perror("setgid");
 617                 exit(EXIT_FAILURE);
 618         }
 619
 620         start = steady_clock::now();
 621         if (access("/", R_OK | X_OK)) {
 622                 // We can't find anything, no need to bother...
 623                 return;
 624         }
 625
 626         IOUringEngine engine(/*slop_bytes=*/16);  // 16 slop bytes as described in turbopfor.h.
 627         Corpus corpus(fd, &engine);
 628         dprintf("Corpus init done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 629
 630         vector<TrigramDisjunction> trigram_groups;
 631         if (patterns_are_regex) {
 632                 // We could parse the regex to find trigrams that have to be there
 633                 // (there are actually known algorithms to deal with disjunctions
 634                 // and such, too), but for now, we just go brute force.
 635                 // Using locate with regexes is pretty niche.
 636         } else {
 637                 for (const Needle &needle : needles) {
 638                         parse_trigrams(needle.str, ignore_case, &trigram_groups);
 639                 }
 640         }
 641
 642         unique_sort(
 643                 &trigram_groups,
 644                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives < b.trigram_alternatives; },
 645                 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives == b.trigram_alternatives; });
 646
 647         // Give them names for debugging.
 648         unsigned td_index = 0;
 649         for (TrigramDisjunction &td : trigram_groups) {
 650                 td.index = td_index++;
 651         }
 652
 653         // Collect which trigrams we need to look up in the hash table.
 654         unordered_map<uint32_t, vector<TrigramDisjunction *>> trigrams_to_lookup;
 655         for (TrigramDisjunction &td : trigram_groups) {
 656                 for (uint32_t trgm : td.trigram_alternatives) {
 657                         trigrams_to_lookup[trgm].push_back(&td);
 658                 }
 659         }
 660         if (trigrams_to_lookup.empty()) {
 661                 // Too short for trigram matching. Apply brute force.
 662                 // (We could have searched through all trigrams that matched
 663                 // the pattern and done a union of them, but that's a lot of
 664                 // work for fairly unclear gain.)
 665                 uint64_t matched = scan_all_docids(needles, fd, corpus);
 666                 if (only_count) {
 667                         printf("%" PRId64 "\n", matched);
 668                 }
 669                 return;
 670         }
 671
 672         // Sneak in fetching the dictionary, if present. It's not necessarily clear
 673         // exactly where it would be cheapest to get it, but it needs to be present
 674         // before we can decode any of the posting lists. Most likely, it's
 675         // in the same filesystem block as the header anyway, so it should be
 676         // present in the cache.
 677         {
 678                 const Header &hdr = corpus.get_hdr();
 679                 if (hdr.zstd_dictionary_length_bytes > 0) {
 680                         engine.submit_read(fd, hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes, [](string_view s) {
 681                                 ddict = ZSTD_createDDict(s.data(), s.size());
 682                                 dprintf("Dictionary initialized after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 683                         });
 684                 }
 685         }
 686
 687         // Look them all up on disk.
 688         for (auto &[trgm, trigram_groups] : trigrams_to_lookup) {
 689                 corpus.find_trigram(trgm, [trgm{ trgm }, trigram_groups{ &trigram_groups }](const Trigram *trgmptr, size_t len) {
 690                         if (trgmptr == nullptr) {
 691                                 dprintf("trigram %s isn't found\n", print_trigram(trgm).c_str());
 692                                 for (TrigramDisjunction *td : *trigram_groups) {
 693                                         --td->remaining_trigrams_to_read;
 694                                         if (td->remaining_trigrams_to_read == 0 && td->read_trigrams.empty()) {
 695                                                 dprintf("zero matches in %s, so we are done\n", print_td(*td).c_str());
 696                                                 if (only_count) {
 697                                                         printf("0\n");
 698                                                 }
 699                                                 exit(0);
 700                                         }
 701                                 }
 702                                 return;
 703                         }
 704                         for (TrigramDisjunction *td : *trigram_groups) {
 705                                 --td->remaining_trigrams_to_read;
 706                                 td->max_num_docids += trgmptr->num_docids;
 707                                 td->read_trigrams.emplace_back(*trgmptr, len);
 708                         }
 709                 });
 710         }
 711         engine.finish();
 712         dprintf("Hashtable lookups done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 713
 714         for (TrigramDisjunction &td : trigram_groups) {
 715                 // Reset for reads.
 716                 td.remaining_trigrams_to_read = td.read_trigrams.size();
 717
 718                 if (ignore_case) {  // If case-sensitive, they'll all be pretty obvious single-entry groups.
 719                         dprintf("OR group %u (max_num_docids=%u): %s\n", td.index, td.max_num_docids, print_td(td).c_str());
 720                 }
 721         }
 722
 723         // TODO: For case-insensitive (ie. more than one alternative in each),
 724         // prioritize the ones with fewer seeks?
 725         sort(trigram_groups.begin(), trigram_groups.end(),
 726              [&](const TrigramDisjunction &a, const TrigramDisjunction &b) {
 727                      return a.max_num_docids < b.max_num_docids;
 728              });
 729
 730         unordered_map<uint32_t, vector<TrigramDisjunction *>> uses_trigram;
 731         for (TrigramDisjunction &td : trigram_groups) {
 732                 for (uint32_t trgm : td.trigram_alternatives) {
 733                         uses_trigram[trgm].push_back(&td);
 734                 }
 735         }
 736
 737         unordered_set<uint32_t> trigrams_submitted_read;
 738         vector<uint32_t> cur_candidates, tmp, decoded;
 739         bool done = false;
 740         for (TrigramDisjunction &td : trigram_groups) {
 741                 if (!cur_candidates.empty() && td.max_num_docids > cur_candidates.size() * 100) {
 742                         dprintf("%s has up to %u entries, ignoring the rest (will "
 743                                 "weed out false positives later)\n",
 744                                 print_td(td).c_str(), td.max_num_docids);
 745                         break;
 746                 }
 747
 748                 for (auto &[trgmptr, len] : td.read_trigrams) {
 749                         if (trigrams_submitted_read.count(trgmptr.trgm) != 0) {
 750                                 continue;
 751                         }
 752                         trigrams_submitted_read.insert(trgmptr.trgm);
 753                         // Only stay a certain amount ahead, so that we don't spend I/O
 754                         // on reading the latter, large posting lists. We are unlikely
 755                         // to need them anyway, even if they should come in first.
 756                         if (engine.get_waiting_reads() >= 5) {
 757                                 engine.finish();
 758                                 if (done)
 759                                         break;
 760                         }
 761                         engine.submit_read(fd, len, trgmptr.offset, [trgmptr{ trgmptr }, len{ len }, &done, &cur_candidates, &tmp, &decoded, &uses_trigram](string_view s) {
 762                                 if (done)
 763                                         return;
 764
 765                                 uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
 766                                 const unsigned char *pldata = reinterpret_cast<const unsigned char *>(s.data());
 767                                 size_t num = trgmptr.num_docids;
 768                                 decoded.resize(num);
 769                                 decode_pfor_delta1_128(pldata, num, /*interleaved=*/true, &decoded[0]);
 770
 771                                 assert(uses_trigram.count(trgm) != 0);
 772                                 bool was_empty = cur_candidates.empty();
 773                                 if (ignore_case) {
 774                                         dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 775                                 }
 776
 777                                 for (TrigramDisjunction *td : uses_trigram[trgm]) {
 778                                         done |= new_posting_list_read(td, decoded, &cur_candidates, &tmp);
 779                                         if (done)
 780                                                 break;
 781                                 }
 782                                 if (!ignore_case) {
 783                                         if (was_empty) {
 784                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
 785                                         } else if (cur_candidates.empty()) {
 786                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (none left, search is done)\n", print_trigram(trgm).c_str(), len, num);
 787                                         } else {
 788                                                 dprintf("trigram %s (%zu bytes) decoded to %zu entries (%zu left)\n", print_trigram(trgm).c_str(), len, num, cur_candidates.size());
 789                                         }
 790                                 }
 791                         });
 792                 }
 793         }
 794         engine.finish();
 795         if (done) {
 796                 return;
 797         }
 798         dprintf("Intersection done after %.1f ms. Doing final verification and printing:\n",
 799                 1e3 * duration<float>(steady_clock::now() - start).count());
 800
 801         uint64_t matched = scan_docids(needles, cur_candidates, corpus, &engine);
 802         dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
 803                 1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 804
 805         if (only_count) {
 806                 printf("%" PRId64 "\n", matched);
 807         }
 808 }
 809
 810 string unescape_glob_to_plain_string(const string &needle)
 811 {
 812         string unescaped;
 813         for (size_t i = 0; i < needle.size(); i += read_unigram(needle, i).second) {
 814                 uint32_t ch = read_unigram(needle, i).first;
 815                 assert(ch != WILDCARD_UNIGRAM);
 816                 if (ch == PREMATURE_END_UNIGRAM) {
 817                         fprintf(stderr, "Pattern '%s' ended prematurely\n", needle.c_str());
 818                         exit(1);
 819                 }
 820                 unescaped.push_back(ch);
 821         }
 822         return unescaped;
 823 }
 824
 825 regex_t compile_regex(const string &needle)
 826 {
 827         regex_t re;
 828         int flags = REG_NOSUB;
 829         if (ignore_case) {
 830                 flags |= REG_ICASE;
 831         }
 832         if (use_extended_regex) {
 833                 flags |= REG_EXTENDED;
 834         }
 835         int err = regcomp(&re, needle.c_str(), flags);
 836         if (err != 0) {
 837                 char errbuf[256];
 838                 regerror(err, &re, errbuf, sizeof(errbuf));
 839                 fprintf(stderr, "Error when compiling regex '%s': %s\n", needle.c_str(), errbuf);
 840                 exit(1);
 841         }
 842         return re;
 843 }
 844
 845 void usage()
 846 {
 847         printf(
 848                 "Usage: plocate [OPTION]... PATTERN...\n"
 849                 "\n"
 850                 "  -c, --count            print number of matches instead of the matches\n"
 851                 "  -d, --database DBPATH  search for files in DBPATH\n"
 852                 "                         (default is " DEFAULT_DBPATH ")\n"
 853                 "  -i, --ignore-case      search case-insensitively\n"
 854                 "  -l, --limit LIMIT      stop after LIMIT matches\n"
 855                 "  -0, --null             delimit matches by NUL instead of newline\n"
 856                 "  -r, --regexp           interpret patterns as basic regexps (slow)\n"
 857                 "      --regex            interpret patterns as extended regexps (slow)\n"
 858                 "      --help             print this help\n"
 859                 "      --version          print version information\n");
 860 }
 861
 862 void version()
 863 {
 864         printf("plocate %s\n", PLOCATE_VERSION);
 865         printf("Copyright 2020 Steinar H. Gunderson\n");
 866         printf("License GPLv2+: GNU GPL version 2 or later <https://gnu.org/licenses/gpl.html>.\n");
 867         printf("This is free software: you are free to change and redistribute it.\n");
 868         printf("There is NO WARRANTY, to the extent permitted by law.\n");
 869         exit(0);
 870 }
 871
 872 int main(int argc, char **argv)
 873 {
 874         constexpr int EXTENDED_REGEX = 1000;
 875         static const struct option long_options[] = {
 876                 { "help", no_argument, 0, 'h' },
 877                 { "count", no_argument, 0, 'c' },
 878                 { "database", required_argument, 0, 'd' },
 879                 { "ignore-case", no_argument, 0, 'i' },
 880                 { "limit", required_argument, 0, 'l' },
 881                 { "null", no_argument, 0, '0' },
 882                 { "version", no_argument, 0, 'V' },
 883                 { "regexp", no_argument, 0, 'r' },
 884                 { "regex", no_argument, 0, EXTENDED_REGEX },
 885                 { "debug", no_argument, 0, 'D' },  // Not documented.
 886                 { 0, 0, 0, 0 }
 887         };
 888
 889         setlocale(LC_ALL, "");
 890         for (;;) {
 891                 int option_index = 0;
 892                 int c = getopt_long(argc, argv, "cd:hil:n:0VD", long_options, &option_index);
 893                 if (c == -1) {
 894                         break;
 895                 }
 896                 switch (c) {
 897                 case 'c':
 898                         only_count = true;
 899                         break;
 900                 case 'd':
 901                         dbpath = strdup(optarg);
 902                         break;
 903                 case 'h':
 904                         usage();
 905                         exit(0);
 906                 case 'i':
 907                         ignore_case = true;
 908                         break;
 909                 case 'l':
 910                 case 'n':
 911                         limit_matches = limit_left = atoll(optarg);
 912                         if (limit_matches <= 0) {
 913                                 fprintf(stderr, "Error: limit must be a strictly positive number.\n");
 914                                 exit(1);
 915                         }
 916                         break;
 917                 case '0':
 918                         print_nul = true;
 919                         break;
 920                 case 'r':
 921                         patterns_are_regex = true;
 922                         break;
 923                 case EXTENDED_REGEX:
 924                         patterns_are_regex = true;
 925                         use_extended_regex = true;
 926                         break;
 927                 case 'D':
 928                         use_debug = true;
 929                         break;
 930                 case 'V':
 931                         version();
 932                         break;
 933                 default:
 934                         exit(1);
 935                 }
 936         }
 937
 938         if (use_debug) {
 939                 // Debug information would leak information about which files exist,
 940                 // so drop setgid before we open the file; one would either need to run
 941                 // as root, or use a locally-built file.
 942                 if (setgid(getgid()) != 0) {
 943                         perror("setgid");
 944                         exit(EXIT_FAILURE);
 945                 }
 946         }
 947
 948         vector<Needle> needles;
 949         for (int i = optind; i < argc; ++i) {
 950                 Needle needle;
 951                 needle.str = argv[i];
 952
 953                 // See if there are any wildcard characters, which indicates we should treat it
 954                 // as an (anchored) glob.
 955                 bool any_wildcard = false;
 956                 for (size_t i = 0; i < needle.str.size(); i += read_unigram(needle.str, i).second) {
 957                         if (read_unigram(needle.str, i).first == WILDCARD_UNIGRAM) {
 958                                 any_wildcard = true;
 959                                 break;
 960                         }
 961                 }
 962
 963                 if (patterns_are_regex) {
 964                         needle.type = Needle::REGEX;
 965                         needle.re = compile_regex(needle.str);
 966                 } else if (any_wildcard) {
 967                         needle.type = Needle::GLOB;
 968                 } else if (ignore_case) {
 969                         // strcasestr() doesn't handle locales correctly (even though LSB
 970                         // claims it should), but somehow, fnmatch() does, and it's about
 971                         // the same speed as using a regex.
 972                         needle.type = Needle::GLOB;
 973                         needle.str = "*" + needle.str + "*";
 974                 } else {
 975                         needle.type = Needle::STRSTR;
 976                         needle.str = unescape_glob_to_plain_string(needle.str);
 977                 }
 978                 needles.push_back(move(needle));
 979         }
 980         if (needles.empty()) {
 981                 fprintf(stderr, "plocate: no pattern to search for specified\n");
 982                 exit(0);
 983         }
 984         do_search_file(needles, dbpath);
 985 }