3 #include "io_uring_engine.h"
4 #include "parse_trigrams.h"
6 #include "unique_sort.h"
26 #include <string_view>
28 #include <unordered_map>
29 #include <unordered_set>
35 using namespace std::chrono;
37 #define DEFAULT_DBPATH "/var/lib/mlocate/plocate.db"
39 const char *dbpath = DEFAULT_DBPATH;
40 bool ignore_case = false;
41 bool only_count = false;
42 bool print_nul = false;
43 bool use_debug = false;
44 bool patterns_are_regex = false;
45 bool use_extended_regex = false;
46 int64_t limit_matches = numeric_limits<int64_t>::max();
50 bool ready_to_print(int seq) { return next_seq == seq; }
51 void print_delayed(int seq, const vector<string> msg);
52 void release_current();
60 bool operator<(const Element &other) const
62 return seq > other.seq;
65 priority_queue<Element> pending;
68 void Serializer::print_delayed(int seq, const vector<string> msg)
70 pending.push(Element{ seq, move(msg) });
73 void Serializer::release_current()
77 // See if any delayed prints can now be dealt with.
78 while (!pending.empty() && pending.top().seq == next_seq) {
79 if (limit_matches-- <= 0)
81 for (const string &msg : pending.top().msg) {
83 printf("%s%c", msg.c_str(), 0);
85 printf("%s\n", msg.c_str());
97 string str; // Filled in no matter what.
98 regex_t re; // For REGEX.
101 bool matches(const Needle &needle, const char *haystack)
103 if (needle.type == Needle::STRSTR) {
104 return strstr(haystack, needle.str.c_str()) != nullptr;
105 } else if (needle.type == Needle::GLOB) {
106 int flags = ignore_case ? FNM_CASEFOLD : 0;
107 return fnmatch(needle.str.c_str(), haystack, flags) == 0;
109 assert(needle.type == Needle::REGEX);
110 return regexec(&needle.re, haystack, /*nmatch=*/0, /*pmatch=*/nullptr, /*flags=*/0) == 0;
114 bool has_access(const char *filename,
115 unordered_map<string, bool> *access_rx_cache)
117 const char *end = strchr(filename + 1, '/');
118 while (end != nullptr) {
119 string parent_path(filename, end);
120 auto it = access_rx_cache->find(parent_path);
122 if (it == access_rx_cache->end()) {
123 ok = access(parent_path.c_str(), R_OK | X_OK) == 0;
124 access_rx_cache->emplace(move(parent_path), ok);
131 end = strchr(end + 1, '/');
139 Corpus(int fd, IOUringEngine *engine);
141 void find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb);
142 void get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const;
143 size_t get_num_filename_blocks() const;
144 off_t offset_for_block(uint32_t docid) const
146 return hdr.filename_index_offset_bytes + docid * sizeof(uint64_t);
151 IOUringEngine *const engine;
156 Corpus::Corpus(int fd, IOUringEngine *engine)
157 : fd(fd), engine(engine)
159 // Enable to test cold-cache behavior (except for access()).
161 off_t len = lseek(fd, 0, SEEK_END);
166 posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED);
169 complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
170 if (memcmp(hdr.magic, "\0plocate", 8) != 0) {
171 fprintf(stderr, "plocate.db is corrupt or an old version; please rebuild it.\n");
174 if (hdr.version != 0) {
175 fprintf(stderr, "plocate.db has version %u, expected 0; please rebuild it.\n", hdr.version);
185 void Corpus::find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb)
187 uint32_t bucket = hash_trigram(trgm, hdr.hashtable_size);
188 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) {
189 const Trigram *trgmptr = reinterpret_cast<const Trigram *>(s.data());
190 for (unsigned i = 0; i < hdr.extra_ht_slots + 1; ++i) {
191 if (trgmptr[i].trgm == trgm) {
192 cb(trgmptr + i, trgmptr[i + 1].offset - trgmptr[i].offset);
202 void Corpus::get_compressed_filename_block(uint32_t docid, function<void(string_view)> cb) const
204 // Read the file offset from this docid and the next one.
205 // This is always allowed, since we have a sentinel block at the end.
206 engine->submit_read(fd, sizeof(uint64_t) * 2, offset_for_block(docid), [this, cb{ move(cb) }](string_view s) {
207 const uint64_t *ptr = reinterpret_cast<const uint64_t *>(s.data());
208 off_t offset = ptr[0];
209 size_t len = ptr[1] - ptr[0];
210 engine->submit_read(fd, len, offset, cb);
214 size_t Corpus::get_num_filename_blocks() const
216 return hdr.num_docids;
219 uint64_t scan_file_block(const vector<Needle> &needles, string_view compressed,
220 unordered_map<string, bool> *access_rx_cache, int seq,
221 Serializer *serializer)
223 uint64_t matched = 0;
225 unsigned long long uncompressed_len = ZSTD_getFrameContentSize(compressed.data(), compressed.size());
226 if (uncompressed_len == ZSTD_CONTENTSIZE_UNKNOWN || uncompressed_len == ZSTD_CONTENTSIZE_ERROR) {
227 fprintf(stderr, "ZSTD_getFrameContentSize() failed\n");
232 block.resize(uncompressed_len + 1);
234 size_t err = ZSTD_decompress(&block[0], block.size(), compressed.data(),
236 if (ZSTD_isError(err)) {
237 fprintf(stderr, "ZSTD_decompress(): %s\n", ZSTD_getErrorName(err));
240 block[block.size() - 1] = '\0';
242 bool immediate_print = (serializer == nullptr || serializer->ready_to_print(seq));
243 vector<string> delayed;
245 for (const char *filename = block.data();
246 filename != block.data() + block.size();
247 filename += strlen(filename) + 1) {
249 for (const Needle &needle : needles) {
250 if (!matches(needle, filename)) {
255 if (found && has_access(filename, access_rx_cache)) {
256 if (limit_matches-- <= 0)
261 if (immediate_print) {
263 printf("%s%c", filename, 0);
265 printf("%s\n", filename);
268 delayed.push_back(filename);
272 if (serializer != nullptr && !only_count) {
273 if (immediate_print) {
274 serializer->release_current();
276 serializer->print_delayed(seq, move(delayed));
282 size_t scan_docids(const vector<Needle> &needles, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
284 Serializer docids_in_order;
285 unordered_map<string, bool> access_rx_cache;
286 uint64_t matched = 0;
287 for (size_t i = 0; i < docids.size(); ++i) {
288 uint32_t docid = docids[i];
289 corpus.get_compressed_filename_block(docid, [i, &matched, &needles, &access_rx_cache, &docids_in_order](string_view compressed) {
290 matched += scan_file_block(needles, compressed, &access_rx_cache, i, &docids_in_order);
297 // We do this sequentially, as it's faster than scattering
298 // a lot of I/O through io_uring and hoping the kernel will
299 // coalesce it plus readahead for us.
300 uint64_t scan_all_docids(const vector<Needle> &needles, int fd, const Corpus &corpus, IOUringEngine *engine)
302 unordered_map<string, bool> access_rx_cache;
303 uint32_t num_blocks = corpus.get_num_filename_blocks();
304 unique_ptr<uint64_t[]> offsets(new uint64_t[num_blocks + 1]);
305 complete_pread(fd, offsets.get(), (num_blocks + 1) * sizeof(uint64_t), corpus.offset_for_block(0));
307 uint64_t matched = 0;
308 for (uint32_t io_docid = 0; io_docid < num_blocks; io_docid += 32) {
309 uint32_t last_docid = std::min(io_docid + 32, num_blocks);
310 size_t io_len = offsets[last_docid] - offsets[io_docid];
311 if (compressed.size() < io_len) {
312 compressed.resize(io_len);
314 complete_pread(fd, &compressed[0], io_len, offsets[io_docid]);
316 for (uint32_t docid = io_docid; docid < last_docid; ++docid) {
317 size_t relative_offset = offsets[docid] - offsets[io_docid];
318 size_t len = offsets[docid + 1] - offsets[docid];
319 matched += scan_file_block(needles, { &compressed[relative_offset], len }, &access_rx_cache, 0, nullptr);
320 if (limit_matches <= 0)
327 // Takes the given posting list, unions it into the parts of the trigram disjunction
328 // already read; if the list is complete, intersects with “cur_candidates”.
330 // Returns true if the search should be aborted (we are done).
331 bool new_posting_list_read(TrigramDisjunction *td, vector<uint32_t> decoded, vector<uint32_t> *cur_candidates, vector<uint32_t> *tmp)
333 if (td->docids.empty()) {
334 td->docids = move(decoded);
337 set_union(decoded.begin(), decoded.end(), td->docids.begin(), td->docids.end(), back_inserter(*tmp));
338 swap(*tmp, td->docids);
340 if (--td->remaining_trigrams_to_read > 0) {
341 // Need to wait for more.
343 dprintf(" ... %u reads left in OR group %u (%zu docids in list)\n",
344 td->remaining_trigrams_to_read, td->index, td->docids.size());
348 if (cur_candidates->empty()) {
350 dprintf(" ... all reads done for OR group %u (%zu docids)\n",
351 td->index, td->docids.size());
353 *cur_candidates = move(td->docids);
356 set_intersection(cur_candidates->begin(), cur_candidates->end(),
357 td->docids.begin(), td->docids.end(),
358 back_inserter(*tmp));
359 swap(*cur_candidates, *tmp);
361 if (cur_candidates->empty()) {
362 dprintf(" ... all reads done for OR group %u (%zu docids), intersected (none left, search is done)\n",
363 td->index, td->docids.size());
366 dprintf(" ... all reads done for OR group %u (%zu docids), intersected (%zu left)\n",
367 td->index, td->docids.size(), cur_candidates->size());
374 void do_search_file(const vector<Needle> &needles, const char *filename)
376 int fd = open(filename, O_RDONLY);
383 if (setgid(getgid()) != 0) {
388 steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
389 if (access("/", R_OK | X_OK)) {
390 // We can't find anything, no need to bother...
394 IOUringEngine engine(/*slop_bytes=*/16); // 16 slop bytes as described in turbopfor.h.
395 Corpus corpus(fd, &engine);
396 dprintf("Corpus init done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
398 vector<TrigramDisjunction> trigram_groups;
399 if (patterns_are_regex) {
400 // We could parse the regex to find trigrams that have to be there
401 // (there are actually known algorithms to deal with disjunctions
402 // and such, too), but for now, we just go brute force.
403 // Using locate with regexes is pretty niche.
405 for (const Needle &needle : needles) {
406 parse_trigrams(needle.str, ignore_case, &trigram_groups);
412 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives < b.trigram_alternatives; },
413 [](const TrigramDisjunction &a, const TrigramDisjunction &b) { return a.trigram_alternatives == b.trigram_alternatives; });
415 // Give them names for debugging.
416 unsigned td_index = 0;
417 for (TrigramDisjunction &td : trigram_groups) {
418 td.index = td_index++;
421 // Collect which trigrams we need to look up in the hash table.
422 unordered_map<uint32_t, vector<TrigramDisjunction *>> trigrams_to_lookup;
423 for (TrigramDisjunction &td : trigram_groups) {
424 for (uint32_t trgm : td.trigram_alternatives) {
425 trigrams_to_lookup[trgm].push_back(&td);
428 if (trigrams_to_lookup.empty()) {
429 // Too short for trigram matching. Apply brute force.
430 // (We could have searched through all trigrams that matched
431 // the pattern and done a union of them, but that's a lot of
432 // work for fairly unclear gain.)
433 uint64_t matched = scan_all_docids(needles, fd, corpus, &engine);
435 printf("%zu\n", matched);
440 // Look them all up on disk.
441 for (auto &[trgm, trigram_groups] : trigrams_to_lookup) {
442 corpus.find_trigram(trgm, [trgm{ trgm }, trigram_groups{ &trigram_groups }](const Trigram *trgmptr, size_t len) {
443 if (trgmptr == nullptr) {
444 dprintf("trigram %s isn't found\n", print_trigram(trgm).c_str());
445 for (TrigramDisjunction *td : *trigram_groups) {
446 --td->remaining_trigrams_to_read;
447 if (td->remaining_trigrams_to_read == 0 && td->read_trigrams.empty()) {
448 dprintf("zero matches in %s, so we are done\n", print_td(*td).c_str());
457 for (TrigramDisjunction *td : *trigram_groups) {
458 --td->remaining_trigrams_to_read;
459 td->max_num_docids += trgmptr->num_docids;
460 td->read_trigrams.emplace_back(*trgmptr, len);
465 dprintf("Hashtable lookups done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
467 for (TrigramDisjunction &td : trigram_groups) {
469 td.remaining_trigrams_to_read = td.read_trigrams.size();
471 if (ignore_case) { // If case-sensitive, they'll all be pretty obvious single-entry groups.
472 dprintf("OR group %u (max_num_docids=%u): %s\n", td.index, td.max_num_docids, print_td(td).c_str());
476 // TODO: For case-insensitive (ie. more than one alternative in each),
477 // prioritize the ones with fewer seeks?
478 sort(trigram_groups.begin(), trigram_groups.end(),
479 [&](const TrigramDisjunction &a, const TrigramDisjunction &b) {
480 return a.max_num_docids < b.max_num_docids;
483 unordered_map<uint32_t, vector<TrigramDisjunction *>> uses_trigram;
484 for (TrigramDisjunction &td : trigram_groups) {
485 for (uint32_t trgm : td.trigram_alternatives) {
486 uses_trigram[trgm].push_back(&td);
490 unordered_set<uint32_t> trigrams_submitted_read;
491 vector<uint32_t> cur_candidates, tmp, decoded;
493 for (TrigramDisjunction &td : trigram_groups) {
494 if (!cur_candidates.empty() && td.max_num_docids > cur_candidates.size() * 100) {
495 dprintf("%s has up to %u entries, ignoring the rest (will "
496 "weed out false positives later)\n",
497 print_td(td).c_str(), td.max_num_docids);
501 for (auto &[trgmptr, len] : td.read_trigrams) {
502 if (trigrams_submitted_read.count(trgmptr.trgm) != 0) {
505 trigrams_submitted_read.insert(trgmptr.trgm);
506 // Only stay a certain amount ahead, so that we don't spend I/O
507 // on reading the latter, large posting lists. We are unlikely
508 // to need them anyway, even if they should come in first.
509 if (engine.get_waiting_reads() >= 5) {
514 engine.submit_read(fd, len, trgmptr.offset, [trgmptr{ trgmptr }, len{ len }, &done, &cur_candidates, &tmp, &decoded, &uses_trigram](string_view s) {
518 uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
519 const unsigned char *pldata = reinterpret_cast<const unsigned char *>(s.data());
520 size_t num = trgmptr.num_docids;
522 decode_pfor_delta1_128(pldata, num, /*interleaved=*/true, &decoded[0]);
524 assert(uses_trigram.count(trgm) != 0);
525 bool was_empty = cur_candidates.empty();
527 dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
530 for (TrigramDisjunction *td : uses_trigram[trgm]) {
531 done |= new_posting_list_read(td, decoded, &cur_candidates, &tmp);
537 dprintf("trigram %s (%zu bytes) decoded to %zu entries\n", print_trigram(trgm).c_str(), len, num);
538 } else if (cur_candidates.empty()) {
539 dprintf("trigram %s (%zu bytes) decoded to %zu entries (none left, search is done)\n", print_trigram(trgm).c_str(), len, num);
541 dprintf("trigram %s (%zu bytes) decoded to %zu entries (%zu left)\n", print_trigram(trgm).c_str(), len, num, cur_candidates.size());
551 dprintf("Intersection done after %.1f ms. Doing final verification and printing:\n",
552 1e3 * duration<float>(steady_clock::now() - start).count());
554 uint64_t matched = scan_docids(needles, cur_candidates, corpus, &engine);
555 dprintf("Done in %.1f ms, found %zu matches.\n",
556 1e3 * duration<float>(steady_clock::now() - start).count(), matched);
559 printf("%zu\n", matched);
563 string unescape_glob_to_plain_string(const string &needle)
566 for (size_t i = 0; i < needle.size(); i += read_unigram(needle, i).second) {
567 uint32_t ch = read_unigram(needle, i).first;
568 assert(ch != WILDCARD_UNIGRAM);
569 if (ch == PREMATURE_END_UNIGRAM) {
570 fprintf(stderr, "Pattern '%s' ended prematurely\n", needle.c_str());
573 unescaped.push_back(ch);
578 regex_t compile_regex(const string &needle)
581 int flags = REG_NOSUB;
585 if (use_extended_regex) {
586 flags |= REG_EXTENDED;
588 int err = regcomp(&re, needle.c_str(), flags);
591 regerror(err, &re, errbuf, sizeof(errbuf));
592 fprintf(stderr, "Error when compiling regex '%s': %s\n", needle.c_str(), errbuf);
601 "Usage: plocate [OPTION]... PATTERN...\n"
603 " -c, --count print number of matches instead of the matches\n"
604 " -d, --database DBPATH search for files in DBPATH\n"
605 " (default is " DEFAULT_DBPATH ")\n"
606 " -i, --ignore-case search case-insensitively\n"
607 " -l, --limit LIMIT stop after LIMIT matches\n"
608 " -0, --null delimit matches by NUL instead of newline\n"
609 " -r, --regexp interpret patterns as basic regexps (slow)\n"
610 " --regex interpret patterns as extended regexps (slow)\n"
611 " --help print this help\n"
612 " --version print version information\n");
617 printf("plocate %s\n", PLOCATE_VERSION);
618 printf("Copyright 2020 Steinar H. Gunderson\n");
619 printf("License GPLv2+: GNU GPL version 2 or later <https://gnu.org/licenses/gpl.html>.\n");
620 printf("This is free software: you are free to change and redistribute it.\n");
621 printf("There is NO WARRANTY, to the extent permitted by law.\n");
625 int main(int argc, char **argv)
627 constexpr int EXTENDED_REGEX = 1000;
628 static const struct option long_options[] = {
629 { "help", no_argument, 0, 'h' },
630 { "count", no_argument, 0, 'c' },
631 { "database", required_argument, 0, 'd' },
632 { "ignore-case", no_argument, 0, 'i' },
633 { "limit", required_argument, 0, 'l' },
634 { "null", no_argument, 0, '0' },
635 { "version", no_argument, 0, 'V' },
636 { "regexp", no_argument, 0, 'r' },
637 { "regex", no_argument, 0, EXTENDED_REGEX },
638 { "debug", no_argument, 0, 'D' }, // Not documented.
642 setlocale(LC_ALL, "");
644 int option_index = 0;
645 int c = getopt_long(argc, argv, "cd:hil:n:0VD", long_options, &option_index);
654 dbpath = strdup(optarg);
664 limit_matches = atoll(optarg);
665 if (limit_matches <= 0) {
666 fprintf(stderr, "Error: limit must be a strictly positive number.\n");
674 patterns_are_regex = true;
677 patterns_are_regex = true;
678 use_extended_regex = true;
692 // Debug information would leak information about which files exist,
693 // so drop setgid before we open the file; one would either need to run
694 // as root, or use a locally-built file.
695 if (setgid(getgid()) != 0) {
701 vector<Needle> needles;
702 for (int i = optind; i < argc; ++i) {
704 needle.str = argv[i];
706 // See if there are any wildcard characters, which indicates we should treat it
707 // as an (anchored) glob.
708 bool any_wildcard = false;
709 for (size_t i = 0; i < needle.str.size(); i += read_unigram(needle.str, i).second) {
710 if (read_unigram(needle.str, i).first == WILDCARD_UNIGRAM) {
716 if (patterns_are_regex) {
717 needle.type = Needle::REGEX;
718 needle.re = compile_regex(needle.str);
719 } else if (any_wildcard) {
720 needle.type = Needle::GLOB;
721 } else if (ignore_case) {
722 // strcasestr() doesn't handle locales correctly (even though LSB
723 // claims it should), but somehow, fnmatch() does, and it's about
724 // the same speed as using a regex.
725 needle.type = Needle::GLOB;
726 needle.str = "*" + needle.str + "*";
728 needle.type = Needle::STRSTR;
729 needle.str = unescape_glob_to_plain_string(needle.str);
731 needles.push_back(move(needle));
733 if (needles.empty()) {
734 fprintf(stderr, "plocate: no pattern to search for specified\n");
737 do_search_file(needles, dbpath);