2 #include "io_uring_engine.h"
16 #include <unordered_map>
21 using namespace std::chrono;
24 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
28 void do_or_wait(int seq, function<void()> cb);
36 bool operator<(const Element &other) const
38 return seq > other.seq;
41 priority_queue<Element> pending;
44 void Serializer::do_or_wait(int seq, function<void()> cb)
46 if (seq != next_seq) {
47 pending.emplace(Element{ seq, move(cb) });
54 while (!pending.empty() && pending.top().seq == next_seq) {
61 static inline uint32_t read_unigram(const string &s, size_t idx)
64 return (unsigned char)s[idx];
70 static inline uint32_t read_trigram(const string &s, size_t start)
72 return read_unigram(s, start) | (read_unigram(s, start + 1) << 8) |
73 (read_unigram(s, start + 2) << 16);
76 bool has_access(const char *filename,
77 unordered_map<string, bool> *access_rx_cache)
79 const char *end = strchr(filename + 1, '/');
80 while (end != nullptr) {
81 string parent_path(filename, end);
82 auto it = access_rx_cache->find(parent_path);
84 if (it == access_rx_cache->end()) {
85 ok = access(parent_path.c_str(), R_OK | X_OK) == 0;
86 access_rx_cache->emplace(move(parent_path), ok);
93 end = strchr(end + 1, '/');
101 Corpus(int fd, IOUringEngine *engine);
103 void find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb);
104 void get_compressed_filename_block(uint32_t docid, function<void(string)> cb) const;
105 size_t get_num_filename_blocks() const;
106 off_t offset_for_block(uint32_t docid) const
108 return hdr.filename_index_offset_bytes + docid * sizeof(uint64_t);
113 IOUringEngine *const engine;
118 Corpus::Corpus(int fd, IOUringEngine *engine)
119 : fd(fd), engine(engine)
121 // Enable to test cold-cache behavior (except for access()).
123 off_t len = lseek(fd, 0, SEEK_END);
128 posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED);
131 complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
132 if (memcmp(hdr.magic, "\0plocate", 8) != 0) {
133 fprintf(stderr, "plocate.db is corrupt or an old version; please rebuild it.\n");
136 if (hdr.version != 0) {
137 fprintf(stderr, "plocate.db has version %u, expected 0; please rebuild it.\n", hdr.version);
147 void Corpus::find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb)
149 uint32_t bucket = hash_trigram(trgm, hdr.hashtable_size);
150 engine->submit_read(fd, sizeof(Trigram) * (hdr.extra_ht_slots + 2), hdr.hash_table_offset_bytes + sizeof(Trigram) * bucket, [this, trgm, bucket, cb{ move(cb) }](string s) {
151 const Trigram *trgmptr = reinterpret_cast<const Trigram *>(s.data());
152 for (unsigned i = 0; i < hdr.extra_ht_slots + 1; ++i) {
153 if (trgmptr[i].trgm == trgm) {
154 cb(trgmptr + i, trgmptr[i + 1].offset - trgmptr[i].offset);
164 void Corpus::get_compressed_filename_block(uint32_t docid, function<void(string)> cb) const
166 // Read the file offset from this docid and the next one.
167 // This is always allowed, since we have a sentinel block at the end.
168 engine->submit_read(fd, sizeof(uint64_t) * 2, offset_for_block(docid), [this, cb{ move(cb) }](string s) {
169 const uint64_t *ptr = reinterpret_cast<const uint64_t *>(s.data());
170 off_t offset = ptr[0];
171 size_t len = ptr[1] - ptr[0];
172 engine->submit_read(fd, len, offset, cb);
176 size_t Corpus::get_num_filename_blocks() const
178 // The beginning of the filename blocks is the end of the filename index blocks.
180 complete_pread(fd, &end, sizeof(end), hdr.filename_index_offset_bytes);
182 // Subtract the sentinel block.
183 return (end - hdr.filename_index_offset_bytes) / sizeof(uint64_t) - 1;
186 size_t scan_file_block(const string &needle, string_view compressed,
187 unordered_map<string, bool> *access_rx_cache)
191 unsigned long long uncompressed_len = ZSTD_getFrameContentSize(compressed.data(), compressed.size());
192 if (uncompressed_len == ZSTD_CONTENTSIZE_UNKNOWN || uncompressed_len == ZSTD_CONTENTSIZE_ERROR) {
193 fprintf(stderr, "ZSTD_getFrameContentSize() failed\n");
198 block.resize(uncompressed_len + 1);
200 size_t err = ZSTD_decompress(&block[0], block.size(), compressed.data(),
202 if (ZSTD_isError(err)) {
203 fprintf(stderr, "ZSTD_decompress(): %s\n", ZSTD_getErrorName(err));
206 block[block.size() - 1] = '\0';
208 for (const char *filename = block.data();
209 filename != block.data() + block.size();
210 filename += strlen(filename) + 1) {
211 if (strstr(filename, needle.c_str()) == nullptr) {
214 if (has_access(filename, access_rx_cache)) {
216 printf("%s\n", filename);
222 size_t scan_docids(const string &needle, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
224 Serializer docids_in_order;
225 unordered_map<string, bool> access_rx_cache;
227 for (size_t i = 0; i < docids.size(); ++i) {
228 uint32_t docid = docids[i];
229 corpus.get_compressed_filename_block(docid, [i, &matched, &needle, &access_rx_cache, &docids_in_order](string compressed) {
230 docids_in_order.do_or_wait(i, [&matched, &needle, compressed{ move(compressed) }, &access_rx_cache] {
231 matched += scan_file_block(needle, compressed, &access_rx_cache);
239 // We do this sequentially, as it's faster than scattering
240 // a lot of I/O through io_uring and hoping the kernel will
241 // coalesce it plus readahead for us.
242 void scan_all_docids(const string &needle, int fd, const Corpus &corpus, IOUringEngine *engine)
244 unordered_map<string, bool> access_rx_cache;
245 uint32_t num_blocks = corpus.get_num_filename_blocks();
246 unique_ptr<uint64_t[]> offsets(new uint64_t[num_blocks + 1]);
247 complete_pread(fd, offsets.get(), (num_blocks + 1) * sizeof(uint64_t), corpus.offset_for_block(0));
249 for (uint32_t io_docid = 0; io_docid < num_blocks; io_docid += 32) {
250 uint32_t last_docid = std::min(io_docid + 32, num_blocks);
251 size_t io_len = offsets[last_docid] - offsets[io_docid];
252 if (compressed.size() < io_len) {
253 compressed.resize(io_len);
255 complete_pread(fd, &compressed[0], io_len, offsets[io_docid]);
257 for (uint32_t docid = io_docid; docid < last_docid; ++docid) {
258 size_t relative_offset = offsets[docid] - offsets[io_docid];
259 size_t len = offsets[docid + 1] - offsets[docid];
260 scan_file_block(needle, { &compressed[relative_offset], len }, &access_rx_cache);
265 void do_search_file(const string &needle, const char *filename)
267 int fd = open(filename, O_RDONLY);
274 if (setgid(getgid()) != 0) {
279 steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
280 if (access("/", R_OK | X_OK)) {
281 // We can't find anything, no need to bother...
285 IOUringEngine engine;
286 Corpus corpus(fd, &engine);
287 dprintf("Corpus init done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
289 if (needle.size() < 3) {
290 // Too short for trigram matching. Apply brute force.
291 // (We could have searched through all trigrams that matched
292 // the pattern and done a union of them, but that's a lot of
293 // work for fairly unclear gain.)
294 scan_all_docids(needle, fd, corpus, &engine);
298 vector<pair<Trigram, size_t>> trigrams;
299 for (size_t i = 0; i < needle.size() - 2; ++i) {
300 uint32_t trgm = read_trigram(needle, i);
301 corpus.find_trigram(trgm, [trgm, &trigrams](const Trigram *trgmptr, size_t len) {
302 if (trgmptr == nullptr) {
303 dprintf("trigram %06x isn't found, we abort the search\n", trgm);
306 trigrams.emplace_back(*trgmptr, len);
310 dprintf("Hashtable lookups done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
312 sort(trigrams.begin(), trigrams.end());
314 auto last = unique(trigrams.begin(), trigrams.end());
315 trigrams.erase(last, trigrams.end());
317 sort(trigrams.begin(), trigrams.end(),
318 [&](const pair<Trigram, size_t> &a, const pair<Trigram, size_t> &b) {
319 return a.first.num_docids < b.first.num_docids;
322 vector<uint32_t> in1, in2, out;
324 for (auto [trgmptr, len] : trigrams) {
325 if (!in1.empty() && trgmptr.num_docids > in1.size() * 100) {
326 uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
327 dprintf("trigram '%c%c%c' (%zu bytes) has %u entries, ignoring the rest (will "
328 "weed out false positives later)\n",
329 trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff,
330 len, trgmptr.num_docids);
334 // Only stay a certain amount ahead, so that we don't spend I/O
335 // on reading the latter, large posting lists. We are unlikely
336 // to need them anyway, even if they should come in first.
337 if (engine.get_waiting_reads() >= 5) {
342 engine.submit_read(fd, len, trgmptr.offset, [trgmptr, len, &done, &in1, &in2, &out](string s) {
345 uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
346 size_t num = trgmptr.num_docids;
347 unsigned char *pldata = reinterpret_cast<unsigned char *>(s.data());
349 in1.resize(num + 128);
350 p4nd1dec128v32(pldata, num, &in1[0]);
352 dprintf("trigram '%c%c%c' (%zu bytes) decoded to %zu entries\n", trgm & 0xff,
353 (trgm >> 8) & 0xff, (trgm >> 16) & 0xff, len, num);
355 if (in2.size() < num + 128) {
356 in2.resize(num + 128);
358 p4nd1dec128v32(pldata, num, &in2[0]);
361 set_intersection(in1.begin(), in1.end(), in2.begin(), in2.begin() + num,
364 dprintf("trigram '%c%c%c' (%zu bytes) decoded to %zu entries, %zu left\n",
365 trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff,
366 len, num, in1.size());
368 dprintf("no matches (intersection list is empty)\n");
378 dprintf("Intersection done after %.1f ms. Doing final verification and printing:\n",
379 1e3 * duration<float>(steady_clock::now() - start).count());
381 size_t matched __attribute__((unused)) = scan_docids(needle, in1, corpus, &engine);
382 dprintf("Done in %.1f ms, found %zu matches.\n",
383 1e3 * duration<float>(steady_clock::now() - start).count(), matched);
386 int main(int argc, char **argv)
388 do_search_file(argv[1], "/var/lib/mlocate/plocate.db");