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

   1 #include "db.h"
   2 #include "io_uring_engine.h"
   3 #include "vp4.h"
   4
   5 #include <algorithm>
   6 #include <arpa/inet.h>
   7 #include <chrono>
   8 #include <endian.h>
   9 #include <fcntl.h>
  10 #include <functional>
  11 #include <getopt.h>
  12 #include <memory>
  13 #include <stdio.h>
  14 #include <string.h>
  15 #include <string>
  16 #include <unistd.h>
  17 #include <unordered_map>
  18 #include <vector>
  19 #include <zstd.h>
  20
  21 using namespace std;
  22 using namespace std::chrono;
  23
  24 #define dprintf(...)
  25 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
  26
  27 const char *dbpath = "/var/lib/mlocate/plocate.db";
  28 bool print_nul = false;
  29
  30 class Serializer {
  31 public:
  32         void do_or_wait(int seq, function<void()> cb);
  33
  34 private:
  35         int next_seq = 0;
  36         struct Element {
  37                 int seq;
  38                 function<void()> cb;
  39
  40                 bool operator<(const Element &other) const
  41                 {
  42                         return seq > other.seq;
  43                 }
  44         };
  45         priority_queue<Element> pending;
  46 };
  47
  48 void Serializer::do_or_wait(int seq, function<void()> cb)
  49 {
  50         if (seq != next_seq) {
  51                 pending.emplace(Element{ seq, move(cb) });
  52                 return;
  53         }
  54
  55         cb();
  56         ++next_seq;
  57
  58         while (!pending.empty() && pending.top().seq == next_seq) {
  59                 pending.top().cb();
  60                 pending.pop();
  61                 ++next_seq;
  62         }
  63 }
  64
  65 static inline uint32_t read_unigram(const string &s, size_t idx)
  66 {
  67         if (idx < s.size()) {
  68                 return (unsigned char)s[idx];
  69         } else {
  70                 return 0;
  71         }
  72 }
  73
  74 static inline uint32_t read_trigram(const string &s, size_t start)
  75 {
  76         return read_unigram(s, start) | (read_unigram(s, start + 1) << 8) |
  77                 (read_unigram(s, start + 2) << 16);
  78 }
  79
  80 bool has_access(const char *filename,
  81                 unordered_map<string, bool> *access_rx_cache)
  82 {
  83         const char *end = strchr(filename + 1, '/');
  84         while (end != nullptr) {
  85                 string parent_path(filename, end);
  86                 auto it = access_rx_cache->find(parent_path);
  87                 bool ok;
  88                 if (it == access_rx_cache->end()) {
  89                         ok = access(parent_path.c_str(), R_OK | X_OK) == 0;
  90                         access_rx_cache->emplace(move(parent_path), ok);
  91                 } else {
  92                         ok = it->second;
  93                 }
  94                 if (!ok) {
  95                         return false;
  96                 }
  97                 end = strchr(end + 1, '/');
  98         }
  99
 100         return true;
 101 }
 102
 103 class Corpus {
 104 public:
 105         Corpus(int fd, IOUringEngine *engine);
 106         ~Corpus();
 107         void find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb);
 108         void get_compressed_filename_block(uint32_t docid, function<void(string)> cb) const;
 109         size_t get_num_filename_blocks() const;
 110         off_t offset_for_block(uint32_t docid) const
 111         {
 112                 return hdr.filename_index_offset_bytes + docid * sizeof(uint64_t);
 113         }
 114
 115 public:
 116         const int fd;
 117         IOUringEngine *const engine;
 118
 119         Header hdr;
 120 };
 121
 122 Corpus::Corpus(int fd, IOUringEngine *engine)
 123         : fd(fd), engine(engine)
 124 {
 125         // Enable to test cold-cache behavior (except for access()).
 126         if (false) {
 127                 off_t len = lseek(fd, 0, SEEK_END);
 128                 if (len == -1) {
 129                         perror("lseek");
 130                         exit(1);
 131                 }
 132                 posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED);
 133         }
 134
 135         complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
 136         if (memcmp(hdr.magic, "\0plocate", 8) != 0) {
 137                 fprintf(stderr, "plocate.db is corrupt or an old version; please rebuild it.\n");
 138                 exit(1);
 139         }
 140         if (hdr.version != 0) {
 141                 fprintf(stderr, "plocate.db has version %u, expected 0; please rebuild it.\n", hdr.version);
 142                 exit(1);
 143         }
 144 }
 145
 146 Corpus::~Corpus()
 147 {
 148         close(fd);
 149 }
 150
 151 void Corpus::find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb)
 152 {
 153         uint32_t bucket = hash_trigram(trgm, hdr.hashtable_size);
 154         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) {
 155                 const Trigram *trgmptr = reinterpret_cast<const Trigram *>(s.data());
 156                 for (unsigned i = 0; i < hdr.extra_ht_slots + 1; ++i) {
 157                         if (trgmptr[i].trgm == trgm) {
 158                                 cb(trgmptr + i, trgmptr[i + 1].offset - trgmptr[i].offset);
 159                                 return;
 160                         }
 161                 }
 162
 163                 // Not found.
 164                 cb(nullptr, 0);
 165         });
 166 }
 167
 168 void Corpus::get_compressed_filename_block(uint32_t docid, function<void(string)> cb) const
 169 {
 170         // Read the file offset from this docid and the next one.
 171         // This is always allowed, since we have a sentinel block at the end.
 172         engine->submit_read(fd, sizeof(uint64_t) * 2, offset_for_block(docid), [this, cb{ move(cb) }](string s) {
 173                 const uint64_t *ptr = reinterpret_cast<const uint64_t *>(s.data());
 174                 off_t offset = ptr[0];
 175                 size_t len = ptr[1] - ptr[0];
 176                 engine->submit_read(fd, len, offset, cb);
 177         });
 178 }
 179
 180 size_t Corpus::get_num_filename_blocks() const
 181 {
 182         // The beginning of the filename blocks is the end of the filename index blocks.
 183         uint64_t end;
 184         complete_pread(fd, &end, sizeof(end), hdr.filename_index_offset_bytes);
 185
 186         // Subtract the sentinel block.
 187         return (end - hdr.filename_index_offset_bytes) / sizeof(uint64_t) - 1;
 188 }
 189
 190 size_t scan_file_block(const vector<string> &needles, string_view compressed,
 191                        unordered_map<string, bool> *access_rx_cache)
 192 {
 193         size_t matched = 0;
 194
 195         unsigned long long uncompressed_len = ZSTD_getFrameContentSize(compressed.data(), compressed.size());
 196         if (uncompressed_len == ZSTD_CONTENTSIZE_UNKNOWN || uncompressed_len == ZSTD_CONTENTSIZE_ERROR) {
 197                 fprintf(stderr, "ZSTD_getFrameContentSize() failed\n");
 198                 exit(1);
 199         }
 200
 201         string block;
 202         block.resize(uncompressed_len + 1);
 203
 204         size_t err = ZSTD_decompress(&block[0], block.size(), compressed.data(),
 205                                      compressed.size());
 206         if (ZSTD_isError(err)) {
 207                 fprintf(stderr, "ZSTD_decompress(): %s\n", ZSTD_getErrorName(err));
 208                 exit(1);
 209         }
 210         block[block.size() - 1] = '\0';
 211
 212         for (const char *filename = block.data();
 213              filename != block.data() + block.size();
 214              filename += strlen(filename) + 1) {
 215                 bool found = true;
 216                 for (const string &needle : needles) {
 217                         if (strstr(filename, needle.c_str()) == nullptr) {
 218                                 found = false;
 219                                 break;
 220                         }
 221                 }
 222                 if (found && has_access(filename, access_rx_cache)) {
 223                         ++matched;
 224                         if (print_nul) {
 225                                 printf("%s%c", filename, 0);
 226                         } else {
 227                                 printf("%s\n", filename);
 228                         }
 229                 }
 230         }
 231         return matched;
 232 }
 233
 234 size_t scan_docids(const vector<string> &needles, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
 235 {
 236         Serializer docids_in_order;
 237         unordered_map<string, bool> access_rx_cache;
 238         size_t matched = 0;
 239         for (size_t i = 0; i < docids.size(); ++i) {
 240                 uint32_t docid = docids[i];
 241                 corpus.get_compressed_filename_block(docid, [i, &matched, &needles, &access_rx_cache, &docids_in_order](string compressed) {
 242                         docids_in_order.do_or_wait(i, [&matched, &needles, compressed{ move(compressed) }, &access_rx_cache] {
 243                                 matched += scan_file_block(needles, compressed, &access_rx_cache);
 244                         });
 245                 });
 246         }
 247         engine->finish();
 248         return matched;
 249 }
 250
 251 // We do this sequentially, as it's faster than scattering
 252 // a lot of I/O through io_uring and hoping the kernel will
 253 // coalesce it plus readahead for us.
 254 void scan_all_docids(const vector<string> &needles, int fd, const Corpus &corpus, IOUringEngine *engine)
 255 {
 256         unordered_map<string, bool> access_rx_cache;
 257         uint32_t num_blocks = corpus.get_num_filename_blocks();
 258         unique_ptr<uint64_t[]> offsets(new uint64_t[num_blocks + 1]);
 259         complete_pread(fd, offsets.get(), (num_blocks + 1) * sizeof(uint64_t), corpus.offset_for_block(0));
 260         string compressed;
 261         for (uint32_t io_docid = 0; io_docid < num_blocks; io_docid += 32) {
 262                 uint32_t last_docid = std::min(io_docid + 32, num_blocks);
 263                 size_t io_len = offsets[last_docid] - offsets[io_docid];
 264                 if (compressed.size() < io_len) {
 265                         compressed.resize(io_len);
 266                 }
 267                 complete_pread(fd, &compressed[0], io_len, offsets[io_docid]);
 268
 269                 for (uint32_t docid = io_docid; docid < last_docid; ++docid) {
 270                         size_t relative_offset = offsets[docid] - offsets[io_docid];
 271                         size_t len = offsets[docid + 1] - offsets[docid];
 272                         scan_file_block(needles, { &compressed[relative_offset], len }, &access_rx_cache);
 273                 }
 274         }
 275 }
 276
 277 void do_search_file(const vector<string> &needles, const char *filename)
 278 {
 279         int fd = open(filename, O_RDONLY);
 280         if (fd == -1) {
 281                 perror(filename);
 282                 exit(1);
 283         }
 284
 285         // Drop privileges.
 286         if (setgid(getgid()) != 0) {
 287                 perror("setgid");
 288                 exit(EXIT_FAILURE);
 289         }
 290
 291         steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
 292         if (access("/", R_OK | X_OK)) {
 293                 // We can't find anything, no need to bother...
 294                 return;
 295         }
 296
 297         IOUringEngine engine;
 298         Corpus corpus(fd, &engine);
 299         dprintf("Corpus init done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 300
 301         vector<pair<Trigram, size_t>> trigrams;
 302         for (const string &needle : needles) {
 303                 if (needle.size() < 3) continue;
 304                 for (size_t i = 0; i < needle.size() - 2; ++i) {
 305                         uint32_t trgm = read_trigram(needle, i);
 306                         corpus.find_trigram(trgm, [trgm, &trigrams](const Trigram *trgmptr, size_t len) {
 307                                 if (trgmptr == nullptr) {
 308                                         dprintf("trigram %06x isn't found, we abort the search\n", trgm);
 309                                         return;
 310                                 }
 311                                 trigrams.emplace_back(*trgmptr, len);
 312                         });
 313                 }
 314         }
 315         engine.finish();
 316         dprintf("Hashtable lookups done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 317
 318         if (trigrams.empty()) {
 319                 // Too short for trigram matching. Apply brute force.
 320                 // (We could have searched through all trigrams that matched
 321                 // the pattern and done a union of them, but that's a lot of
 322                 // work for fairly unclear gain.)
 323                 scan_all_docids(needles, fd, corpus, &engine);
 324                 return;
 325         }
 326         sort(trigrams.begin(), trigrams.end());
 327         {
 328                 auto last = unique(trigrams.begin(), trigrams.end());
 329                 trigrams.erase(last, trigrams.end());
 330         }
 331         sort(trigrams.begin(), trigrams.end(),
 332              [&](const pair<Trigram, size_t> &a, const pair<Trigram, size_t> &b) {
 333                      return a.first.num_docids < b.first.num_docids;
 334              });
 335
 336         vector<uint32_t> in1, in2, out;
 337         bool done = false;
 338         for (auto [trgmptr, len] : trigrams) {
 339                 if (!in1.empty() && trgmptr.num_docids > in1.size() * 100) {
 340                         uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
 341                         dprintf("trigram '%c%c%c' (%zu bytes) has %u entries, ignoring the rest (will "
 342                                 "weed out false positives later)\n",
 343                                 trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff,
 344                                 len, trgmptr.num_docids);
 345                         break;
 346                 }
 347
 348                 // Only stay a certain amount ahead, so that we don't spend I/O
 349                 // on reading the latter, large posting lists. We are unlikely
 350                 // to need them anyway, even if they should come in first.
 351                 if (engine.get_waiting_reads() >= 5) {
 352                         engine.finish();
 353                         if (done)
 354                                 break;
 355                 }
 356                 engine.submit_read(fd, len, trgmptr.offset, [trgmptr, len, &done, &in1, &in2, &out](string s) {
 357                         if (done)
 358                                 return;
 359                         uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
 360                         size_t num = trgmptr.num_docids;
 361                         unsigned char *pldata = reinterpret_cast<unsigned char *>(s.data());
 362                         if (in1.empty()) {
 363                                 in1.resize(num + 128);
 364                                 p4nd1dec128v32(pldata, num, &in1[0]);
 365                                 in1.resize(num);
 366                                 dprintf("trigram '%c%c%c' (%zu bytes) decoded to %zu entries\n", trgm & 0xff,
 367                                         (trgm >> 8) & 0xff, (trgm >> 16) & 0xff, len, num);
 368                         } else {
 369                                 if (in2.size() < num + 128) {
 370                                         in2.resize(num + 128);
 371                                 }
 372                                 p4nd1dec128v32(pldata, num, &in2[0]);
 373
 374                                 out.clear();
 375                                 set_intersection(in1.begin(), in1.end(), in2.begin(), in2.begin() + num,
 376                                                  back_inserter(out));
 377                                 swap(in1, out);
 378                                 dprintf("trigram '%c%c%c' (%zu bytes) decoded to %zu entries, %zu left\n",
 379                                         trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff,
 380                                         len, num, in1.size());
 381                                 if (in1.empty()) {
 382                                         dprintf("no matches (intersection list is empty)\n");
 383                                         done = true;
 384                                 }
 385                         }
 386                 });
 387         }
 388         engine.finish();
 389         if (done) {
 390                 return;
 391         }
 392         dprintf("Intersection done after %.1f ms. Doing final verification and printing:\n",
 393                 1e3 * duration<float>(steady_clock::now() - start).count());
 394
 395         size_t matched __attribute__((unused)) = scan_docids(needles, in1, corpus, &engine);
 396         dprintf("Done in %.1f ms, found %zu matches.\n",
 397                 1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 398 }
 399
 400 void usage()
 401 {
 402         // The help text comes from mlocate.
 403         printf("Usage: plocate [OPTION]... PATTERN...\n");
 404         printf("  -d, --database DBPATH  use DBPATH instead of default database (which is\n");
 405         printf("                         %s)\n", dbpath);
 406         printf("  -h, --help             print this help\n");
 407         printf("  -0, --null             separate entries with NUL on output\n");
 408 }
 409
 410 int main(int argc, char **argv)
 411 {
 412         static const struct option long_options[] = {
 413                 { "help", no_argument, 0, 'h' },
 414                 { "database", required_argument, 0, 'd' },
 415                 { "null", no_argument, 0, '0' },
 416                 { 0, 0, 0, 0 }
 417         };
 418
 419         for (;;) {
 420                 int option_index = 0;
 421                 int c = getopt_long(argc, argv, "d:h0", long_options, &option_index);
 422                 if (c == -1) {
 423                         break;
 424                 }
 425                 switch (c) {
 426                 case 'd':
 427                         dbpath = strdup(optarg);
 428                         break;
 429                 case 'h':
 430                         usage();
 431                         exit(0);
 432                 case '0':
 433                         print_nul = true;
 434                         break;
 435                 default:
 436                         exit(1);
 437                 }
 438         }
 439
 440         vector<string> needles;
 441         for (int i = optind; i < argc; ++i) {
 442                 needles.push_back(argv[i]);
 443         }
 444         if (needles.empty()) {
 445                 fprintf(stderr, "plocate: no pattern to search for specified\n");
 446                 exit(0);
 447         }
 448         do_search_file(needles, dbpath);
 449 }