3 #include "io_uring_engine.h"
4 #include "parse_trigrams.h"
6 #include "unique_sort.h"
27 #include <string_view>
29 #include <unordered_map>
30 #include <unordered_set>
36 using namespace std::chrono;
38 #define DEFAULT_DBPATH "/var/lib/mlocate/plocate.db"
40 const char *dbpath = DEFAULT_DBPATH;
41 bool ignore_case = false;
42 bool only_count = false;
43 bool print_nul = false;
44 bool use_debug = false;
45 bool patterns_are_regex = false;
46 bool use_extended_regex = false;
47 int64_t limit_matches = numeric_limits<int64_t>::max();
51 bool ready_to_print(int seq) { return next_seq == seq; }
52 void print_delayed(int seq, const vector<string> msg);
53 void release_current();
61 bool operator<(const Element &other) const
63 return seq > other.seq;
66 priority_queue<Element> pending;
69 void Serializer::print_delayed(int seq, const vector<string> msg)
71 pending.push(Element{ seq, move(msg) });
74 void Serializer::release_current()
78 // See if any delayed prints can now be dealt with.
79 while (!pending.empty() && pending.top().seq == next_seq) {
80 if (limit_matches-- <= 0)
82 for (const string &msg : pending.top().msg) {
84 printf("%s%c", msg.c_str(), 0);
86 printf("%s\n", msg.c_str());
98 string str; // Filled in no matter what.
99 regex_t re; // For REGEX.
102 bool matches(const Needle &needle, const char *haystack)
104 if (needle.type == Needle::STRSTR) {
105 return strstr(haystack, needle.str.c_str()) != nullptr;
106 } else if (needle.type == Needle::GLOB) {
107 int flags = ignore_case ? FNM_CASEFOLD : 0;
108 return fnmatch(needle.str.c_str(), haystack, flags) == 0;
110 assert(needle.type == Needle::REGEX);
111 return regexec(&needle.re, haystack, /*nmatch=*/0, /*pmatch=*/nullptr, /*flags=*/0) == 0;
115 bool has_access(const char *filename,
116 unordered_map<string, bool> *access_rx_cache)
118 const char *end = strchr(filename + 1, '/');
119 while (end != nullptr) {
120 string parent_path(filename, end);
121 auto it = access_rx_cache->find(parent_path);
123 if (it == access_rx_cache->end()) {
124 ok = access(parent_path.c_str(), R_OK | X_OK) == 0;
125 access_rx_cache->emplace(move(parent_path), ok);
132 end = strchr(end + 1, '/');
140 Corpus(int fd, IOUringEngine *engine);
142 void find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb);
143 void get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const;
144 size_t get_num_filename_blocks() const;
145 off_t offset_for_block(uint32_t docid) const
147 return hdr.filename_index_offset_bytes + docid * sizeof(uint64_t);
152 IOUringEngine *const engine;
157 Corpus::Corpus(int fd, IOUringEngine *engine)
158 : fd(fd), engine(engine)
160 // Enable to test cold-cache behavior (except for access()).
162 off_t len = lseek(fd, 0, SEEK_END);
167 posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED);
170 complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
171 if (memcmp(hdr.magic, "\0plocate", 8) != 0) {
172 fprintf(stderr, "plocate.db is corrupt or an old version; please rebuild it.\n");
175 if (hdr.version != 0) {
176 fprintf(stderr, "plocate.db has version %u, expected 0; please rebuild it.\n", hdr.version);
186 void Corpus::find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb)
188 uint32_t bucket = hash_trigram(trgm, hdr.hashtable_size);
189 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) {
190 const Trigram *trgmptr = reinterpret_cast<const Trigram *>(s.data());
191 for (unsigned i = 0; i < hdr.extra_ht_slots + 1; ++i) {
192 if (trgmptr[i].trgm == trgm) {
193 cb(trgmptr + i, trgmptr[i + 1].offset - trgmptr[i].offset);
203 void Corpus::get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const
205 // Read the file offset from this docid and the next one.
206 // This is always allowed, since we have a sentinel block at the end.
207 engine->submit_read(fd, sizeof(uint64_t) * 2, offset_for_block(docid), [this, cb{ move(cb) }](string_view s) {
208 const uint64_t *ptr = reinterpret_cast<const uint64_t *>(s.data());
209 off_t offset = ptr[0];
210 size_t len = ptr[1] - ptr[0];
211 engine->submit_read(fd, len, offset, cb);
215 size_t Corpus::get_num_filename_blocks() const
217 return hdr.num_docids;
220 uint64_t scan_file_block(const vector<Needle> &needles, string_view compressed,
221 unordered_map<string, bool> *access_rx_cache, int seq,
222 Serializer *serializer)
224 uint64_t matched = 0;
226 unsigned long long uncompressed_len = ZSTD_getFrameContentSize(compressed.data(), compressed.size());
227 if (uncompressed_len == ZSTD_CONTENTSIZE_UNKNOWN || uncompressed_len == ZSTD_CONTENTSIZE_ERROR) {
228 fprintf(stderr, "ZSTD_getFrameContentSize() failed\n");
233 block.resize(uncompressed_len + 1);
235 size_t err = ZSTD_decompress(&block[0], block.size(), compressed.data(),
237 if (ZSTD_isError(err)) {
238 fprintf(stderr, "ZSTD_decompress(): %s\n", ZSTD_getErrorName(err));
241 block[block.size() - 1] = '\0';
243 bool immediate_print = (serializer == nullptr || serializer->ready_to_print(seq));
244 vector<string> delayed;
246 for (const char *filename = block.data();
247 filename != block.data() + block.size();
248 filename += strlen(filename) + 1) {
250 for (const Needle &needle : needles) {
251 if (!matches(needle, filename)) {
256 if (found && has_access(filename, access_rx_cache)) {
257 if (limit_matches-- <= 0)
262 if (immediate_print) {
264 printf("%s%c", filename, 0);
266 printf("%s\n", filename);
269 delayed.push_back(filename);
273 if (serializer != nullptr && !only_count) {
274 if (immediate_print) {
275 serializer->release_current();
277 serializer->print_delayed(seq, move(delayed));
283 size_t scan_docids(const vector<Needle> &needles, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
285 Serializer docids_in_order;
286 unordered_map<string, bool> access_rx_cache;
287 uint64_t matched = 0;
288 for (size_t i = 0; i < docids.size(); ++i) {
289 uint32_t docid = docids[i];
290 corpus.get_compressed_filename_block(docid, [i, &matched, &needles, &access_rx_cache, &docids_in_order](string_view compressed) {
291 matched += scan_file_block(needles, compressed, &access_rx_cache, i, &docids_in_order);
298 // We do this sequentially, as it's faster than scattering
299 // a lot of I/O through io_uring and hoping the kernel will
300 // coalesce it plus readahead for us.
301 uint64_t scan_all_docids(const vector<Needle> &needles, int fd, const Corpus &corpus, IOUringEngine *engine)
303 unordered_map<string, bool> access_rx_cache;
304 uint32_t num_blocks = corpus.get_num_filename_blocks();
305 unique_ptr<uint64_t[]> offsets(new uint64_t[num_blocks + 1]);
306 complete_pread(fd, offsets.get(), (num_blocks + 1) * sizeof(uint64_t), corpus.offset_for_block(0));
308 uint64_t matched = 0;
309 for (uint32_t io_docid = 0; io_docid < num_blocks; io_docid += 32) {
310 uint32_t last_docid = std::min(io_docid + 32, num_blocks);
311 size_t io_len = offsets[last_docid] - offsets[io_docid];
312 if (compressed.size() < io_len) {
313 compressed.resize(io_len);
315 complete_pread(fd, &compressed[0], io_len, offsets[io_docid]);
317 for (uint32_t docid = io_docid; docid < last_docid; ++docid) {
318 size_t relative_offset = offsets[docid] - offsets[io_docid];
319 size_t len = offsets[docid + 1] - offsets[docid];
320 matched += scan_file_block(needles, { &compressed[relative_offset], len }, &access_rx_cache, 0, nullptr);
321 if (limit_matches <= 0)
328 // Takes the given posting list, unions it into the parts of the trigram disjunction
329 // already read; if the list is complete, intersects with “cur_candidates”.
331 // Returns true if the search should be aborted (we are done).
332 bool new_posting_list_read(TrigramDisjunction *td, vector<uint32_t> decoded, vector<uint32_t> *cur_candidates, vector<uint32_t> *tmp)
334 if (td->docids.empty()) {
335 td->docids = move(decoded);
338 set_union(decoded.begin(), decoded.end(), td->docids.begin(), td->docids.end(), back_inserter(*tmp));
339 swap(*tmp, td->docids);
341 if (--td->remaining_trigrams_to_read > 0) {
342 // Need to wait for more.
344 dprintf(" ... %u reads left in OR group %u (%zu docids in list)\n",
345 td->remaining_trigrams_to_read, td->index, td->docids.size());
349 if (cur_candidates->empty()) {
351 dprintf(" ... all reads done for OR group %u (%zu docids)\n",
352 td->index, td->docids.size());
354 *cur_candidates = move(td->docids);
357 set_intersection(cur_candidates->begin(), cur_candidates->end(),
358 td->docids.begin(), td->docids.end(),
359 back_inserter(*tmp));
360 swap(*cur_candidates, *tmp);
362 if (cur_candidates->empty()) {
363 dprintf(" ... all reads done for OR group %u (%zu docids), intersected (none left, search is done)\n",
364 td->index, td->docids.size());
367 dprintf(" ... all reads done for OR group %u (%zu docids), intersected (%zu left)\n",
368 td->index, td->docids.size(), cur_candidates->size());
375 void do_search_file(const vector<Needle> &needles, const char *filename)
377 int fd = open(filename, O_RDONLY);
384 if (setgid(getgid()) != 0) {
389 steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
390 if (access("/", R_OK | X_OK)) {
391 // We can't find anything, no need to bother...
395 IOUringEngine engine(/*slop_bytes=*/16); // 16 slop bytes as described in turbopfor.h.
396 Corpus corpus(fd, &engine);
397 dprintf("Corpus init done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
399 vector<TrigramDisjunction> trigram_groups;
400 if (patterns_are_regex) {
401 // We could parse the regex to find trigrams that have to be there
402 // (there are actually known algorithms to deal with disjunctions
403 // and such, too), but for now, we just go brute force.
404 // Using locate with regexes is pretty niche.
406 for (const Needle &needle : needles) {
407 parse_trigrams(needle.str, ignore_case, &trigram_groups);
413 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives < b.trigram_alternatives; },
414 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives == b.trigram_alternatives; });
416 // Give them names for debugging.
417 unsigned td_index = 0;
418 for (TrigramDisjunction &td : trigram_groups) {
419 td.index = td_index++;
422 // Collect which trigrams we need to look up in the hash table.
423 unordered_map<uint32_t, vector<TrigramDisjunction *>> trigrams_to_lookup;
424 for (TrigramDisjunction &td : trigram_groups) {
425 for (uint32_t trgm : td.trigram_alternatives) {
426 trigrams_to_lookup[trgm].push_back(&td);
429 if (trigrams_to_lookup.empty()) {
430 // Too short for trigram matching. Apply brute force.
431 // (We could have searched through all trigrams that matched
432 // the pattern and done a union of them, but that's a lot of
433 // work for fairly unclear gain.)
434 uint64_t matched = scan_all_docids(needles, fd, corpus, &engine);
436 printf("%" PRId64 "\n", matched);
441 // Look them all up on disk.
442 for (auto &[trgm, trigram_groups] : trigrams_to_lookup) {
443 corpus.find_trigram(trgm, [trgm{ trgm }, trigram_groups{ &trigram_groups }](const Trigram *trgmptr, size_t len) {
444 if (trgmptr == nullptr) {
445 dprintf("trigram %s isn't found\n", print_trigram(trgm).c_str());
446 for (TrigramDisjunction *td : *trigram_groups) {
447 --td->remaining_trigrams_to_read;
448 if (td->remaining_trigrams_to_read == 0 && td->read_trigrams.empty()) {
449 dprintf("zero matches in %s, so we are done\n", print_td(*td).c_str());
458 for (TrigramDisjunction *td : *trigram_groups) {
459 --td->remaining_trigrams_to_read;
460 td->max_num_docids += trgmptr->num_docids;
461 td->read_trigrams.emplace_back(*trgmptr, len);
466 dprintf("Hashtable lookups done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
468 for (TrigramDisjunction &td : trigram_groups) {
470 td.remaining_trigrams_to_read = td.read_trigrams.size();
472 if (ignore_case) { // If case-sensitive, they'll all be pretty obvious single-entry groups.
473 dprintf("OR group %u (max_num_docids=%u): %s\n", td.index, td.max_num_docids, print_td(td).c_str());
477 // TODO: For case-insensitive (ie. more than one alternative in each),
478 // prioritize the ones with fewer seeks?
479 sort(trigram_groups.begin(), trigram_groups.end(),
480 [&](const TrigramDisjunction &a, const TrigramDisjunction &b) {
481 return a.max_num_docids < b.max_num_docids;
484 unordered_map<uint32_t, vector<TrigramDisjunction *>> uses_trigram;
485 for (TrigramDisjunction &td : trigram_groups) {
486 for (uint32_t trgm : td.trigram_alternatives) {
487 uses_trigram[trgm].push_back(&td);
491 unordered_set<uint32_t> trigrams_submitted_read;
492 vector<uint32_t> cur_candidates, tmp, decoded;
494 for (TrigramDisjunction &td : trigram_groups) {
495 if (!cur_candidates.empty() && td.max_num_docids > cur_candidates.size() * 100) {
496 dprintf("%s has up to %u entries, ignoring the rest (will "
497 "weed out false positives later)\n",
498 print_td(td).c_str(), td.max_num_docids);
502 for (auto &[trgmptr, len] : td.read_trigrams) {
503 if (trigrams_submitted_read.count(trgmptr.trgm) != 0) {
506 trigrams_submitted_read.insert(trgmptr.trgm);
507 // Only stay a certain amount ahead, so that we don't spend I/O
508 // on reading the latter, large posting lists. We are unlikely
509 // to need them anyway, even if they should come in first.
510 if (engine.get_waiting_reads() >= 5) {
515 engine.submit_read(fd, len, trgmptr.offset, [trgmptr{ trgmptr }, len{ len }, &done, &cur_candidates, &tmp, &decoded, &uses_trigram](string_view s) {
519 uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
520 const unsigned char *pldata = reinterpret_cast<const unsigned char *>(s.data());
521 size_t num = trgmptr.num_docids;
523 decode_pfor_delta1_128(pldata, num, /*interleaved=*/true, &decoded[0]);
525 assert(uses_trigram.count(trgm) != 0);
526 bool was_empty = cur_candidates.empty();
528 dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
531 for (TrigramDisjunction *td : uses_trigram[trgm]) {
532 done |= new_posting_list_read(td, decoded, &cur_candidates, &tmp);
538 dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
539 } else if (cur_candidates.empty()) {
540 dprintf("trigram %s (%zu bytes) decoded to %zu entries (none left, search is done)\n", print_trigram(trgm).c_str(), len, num);
542 dprintf("trigram %s (%zu bytes) decoded to %zu entries (%zu left)\n", print_trigram(trgm).c_str(), len, num, cur_candidates.size());
552 dprintf("Intersection done after %.1f ms. Doing final verification and printing:\n",
553 1e3 * duration<float>(steady_clock::now() - start).count());
555 uint64_t matched = scan_docids(needles, cur_candidates, corpus, &engine);
556 dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
557 1e3 * duration<float>(steady_clock::now() - start).count(), matched);
560 printf("%" PRId64 "\n", matched);
564 string unescape_glob_to_plain_string(const string &needle)
567 for (size_t i = 0; i < needle.size(); i += read_unigram(needle, i).second) {
568 uint32_t ch = read_unigram(needle, i).first;
569 assert(ch != WILDCARD_UNIGRAM);
570 if (ch == PREMATURE_END_UNIGRAM) {
571 fprintf(stderr, "Pattern '%s' ended prematurely\n", needle.c_str());
574 unescaped.push_back(ch);
579 regex_t compile_regex(const string &needle)
582 int flags = REG_NOSUB;
586 if (use_extended_regex) {
587 flags |= REG_EXTENDED;
589 int err = regcomp(&re, needle.c_str(), flags);
592 regerror(err, &re, errbuf, sizeof(errbuf));
593 fprintf(stderr, "Error when compiling regex '%s': %s\n", needle.c_str(), errbuf);
602 "Usage: plocate [OPTION]... PATTERN...\n"
604 " -c, --count print number of matches instead of the matches\n"
605 " -d, --database DBPATH search for files in DBPATH\n"
606 " (default is " DEFAULT_DBPATH ")\n"
607 " -i, --ignore-case search case-insensitively\n"
608 " -l, --limit LIMIT stop after LIMIT matches\n"
609 " -0, --null delimit matches by NUL instead of newline\n"
610 " -r, --regexp interpret patterns as basic regexps (slow)\n"
611 " --regex interpret patterns as extended regexps (slow)\n"
612 " --help print this help\n"
613 " --version print version information\n");
618 printf("plocate %s\n", PLOCATE_VERSION);
619 printf("Copyright 2020 Steinar H. Gunderson\n");
620 printf("License GPLv2+: GNU GPL version 2 or later <https://gnu.org/licenses/gpl.html>.\n");
621 printf("This is free software: you are free to change and redistribute it.\n");
622 printf("There is NO WARRANTY, to the extent permitted by law.\n");
626 int main(int argc, char **argv)
628 constexpr int EXTENDED_REGEX = 1000;
629 static const struct option long_options[] = {
630 { "help", no_argument, 0, 'h' },
631 { "count", no_argument, 0, 'c' },
632 { "database", required_argument, 0, 'd' },
633 { "ignore-case", no_argument, 0, 'i' },
634 { "limit", required_argument, 0, 'l' },
635 { "null", no_argument, 0, '0' },
636 { "version", no_argument, 0, 'V' },
637 { "regexp", no_argument, 0, 'r' },
638 { "regex", no_argument, 0, EXTENDED_REGEX },
639 { "debug", no_argument, 0, 'D' }, // Not documented.
643 setlocale(LC_ALL, "");
645 int option_index = 0;
646 int c = getopt_long(argc, argv, "cd:hil:n:0VD", long_options, &option_index);
655 dbpath = strdup(optarg);
665 limit_matches = atoll(optarg);
666 if (limit_matches <= 0) {
667 fprintf(stderr, "Error: limit must be a strictly positive number.\n");
675 patterns_are_regex = true;
678 patterns_are_regex = true;
679 use_extended_regex = true;
693 // Debug information would leak information about which files exist,
694 // so drop setgid before we open the file; one would either need to run
695 // as root, or use a locally-built file.
696 if (setgid(getgid()) != 0) {
702 vector<Needle> needles;
703 for (int i = optind; i < argc; ++i) {
705 needle.str = argv[i];
707 // See if there are any wildcard characters, which indicates we should treat it
708 // as an (anchored) glob.
709 bool any_wildcard = false;
710 for (size_t i = 0; i < needle.str.size(); i += read_unigram(needle.str, i).second) {
711 if (read_unigram(needle.str, i).first == WILDCARD_UNIGRAM) {
717 if (patterns_are_regex) {
718 needle.type = Needle::REGEX;
719 needle.re = compile_regex(needle.str);
720 } else if (any_wildcard) {
721 needle.type = Needle::GLOB;
722 } else if (ignore_case) {
723 // strcasestr() doesn't handle locales correctly (even though LSB
724 // claims it should), but somehow, fnmatch() does, and it's about
725 // the same speed as using a regex.
726 needle.type = Needle::GLOB;
727 needle.str = "*" + needle.str + "*";
729 needle.type = Needle::STRSTR;
730 needle.str = unescape_glob_to_plain_string(needle.str);
732 needles.push_back(move(needle));
734 if (needles.empty()) {
735 fprintf(stderr, "plocate: no pattern to search for specified\n");
738 do_search_file(needles, dbpath);