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

   1 #include "db.h"
   2 #include "io_uring_engine.h"
   3
   4 #include <algorithm>
   5 #include <arpa/inet.h>
   6 #include <assert.h>
   7 #include <chrono>
   8 #include <endian.h>
   9 #include <fcntl.h>
  10 #include <functional>
  11 #include <getopt.h>
  12 #include <limits.h>
  13 #include <memory>
  14 #include <stdio.h>
  15 #include <string.h>
  16 #include <string>
  17 #include <unistd.h>
  18 #include <unordered_map>
  19 #include <vector>
  20 #include <zstd.h>
  21
  22 using namespace std;
  23 using namespace std::chrono;
  24
  25 #define dprintf(...)
  26 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
  27
  28 #include "turbopfor.h"
  29
  30 const char *dbpath = "/var/lib/mlocate/plocate.db";
  31 bool only_count = false;
  32 bool print_nul = false;
  33
  34 class Serializer {
  35 public:
  36         bool ready_to_print(int seq) { return next_seq == seq; }
  37         void print_delayed(int seq, const vector<string> msg);
  38         void release_current();
  39
  40 private:
  41         int next_seq = 0;
  42         struct Element {
  43                 int seq;
  44                 vector<string> msg;
  45
  46                 bool operator<(const Element &other) const
  47                 {
  48                         return seq > other.seq;
  49                 }
  50         };
  51         priority_queue<Element> pending;
  52 };
  53
  54 void Serializer::print_delayed(int seq, const vector<string> msg)
  55 {
  56         pending.push(Element{ seq, move(msg) });
  57 }
  58
  59 void Serializer::release_current()
  60 {
  61         ++next_seq;
  62
  63         // See if any delayed prints can now be dealt with.
  64         while (!pending.empty() && pending.top().seq == next_seq) {
  65                 for (const string &msg : pending.top().msg) {
  66                         if (print_nul) {
  67                                 printf("%s%c", msg.c_str(), 0);
  68                         } else {
  69                                 printf("%s\n", msg.c_str());
  70                         }
  71                 }
  72                 pending.pop();
  73                 ++next_seq;
  74         }
  75 }
  76
  77 static inline uint32_t read_unigram(const string &s, size_t idx)
  78 {
  79         if (idx < s.size()) {
  80                 return (unsigned char)s[idx];
  81         } else {
  82                 return 0;
  83         }
  84 }
  85
  86 static inline uint32_t read_trigram(const string &s, size_t start)
  87 {
  88         return read_unigram(s, start) | (read_unigram(s, start + 1) << 8) |
  89                 (read_unigram(s, start + 2) << 16);
  90 }
  91
  92 bool has_access(const char *filename,
  93                 unordered_map<string, bool> *access_rx_cache)
  94 {
  95         const char *end = strchr(filename + 1, '/');
  96         while (end != nullptr) {
  97                 string parent_path(filename, end);
  98                 auto it = access_rx_cache->find(parent_path);
  99                 bool ok;
 100                 if (it == access_rx_cache->end()) {
 101                         ok = access(parent_path.c_str(), R_OK | X_OK) == 0;
 102                         access_rx_cache->emplace(move(parent_path), ok);
 103                 } else {
 104                         ok = it->second;
 105                 }
 106                 if (!ok) {
 107                         return false;
 108                 }
 109                 end = strchr(end + 1, '/');
 110         }
 111
 112         return true;
 113 }
 114
 115 class Corpus {
 116 public:
 117         Corpus(int fd, IOUringEngine *engine);
 118         ~Corpus();
 119         void find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb);
 120         void get_compressed_filename_block(uint32_t docid, function<void(string)> cb) const;
 121         size_t get_num_filename_blocks() const;
 122         off_t offset_for_block(uint32_t docid) const
 123         {
 124                 return hdr.filename_index_offset_bytes + docid * sizeof(uint64_t);
 125         }
 126
 127 public:
 128         const int fd;
 129         IOUringEngine *const engine;
 130
 131         Header hdr;
 132 };
 133
 134 Corpus::Corpus(int fd, IOUringEngine *engine)
 135         : fd(fd), engine(engine)
 136 {
 137         // Enable to test cold-cache behavior (except for access()).
 138         if (false) {
 139                 off_t len = lseek(fd, 0, SEEK_END);
 140                 if (len == -1) {
 141                         perror("lseek");
 142                         exit(1);
 143                 }
 144                 posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED);
 145         }
 146
 147         complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
 148         if (memcmp(hdr.magic, "\0plocate", 8) != 0) {
 149                 fprintf(stderr, "plocate.db is corrupt or an old version; please rebuild it.\n");
 150                 exit(1);
 151         }
 152         if (hdr.version != 0) {
 153                 fprintf(stderr, "plocate.db has version %u, expected 0; please rebuild it.\n", hdr.version);
 154                 exit(1);
 155         }
 156 }
 157
 158 Corpus::~Corpus()
 159 {
 160         close(fd);
 161 }
 162
 163 void Corpus::find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb)
 164 {
 165         uint32_t bucket = hash_trigram(trgm, hdr.hashtable_size);
 166         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 s) {
 167                 const Trigram *trgmptr = reinterpret_cast<const Trigram *>(s.data());
 168                 for (unsigned i = 0; i < hdr.extra_ht_slots + 1; ++i) {
 169                         if (trgmptr[i].trgm == trgm) {
 170                                 cb(trgmptr + i, trgmptr[i + 1].offset - trgmptr[i].offset);
 171                                 return;
 172                         }
 173                 }
 174
 175                 // Not found.
 176                 cb(nullptr, 0);
 177         });
 178 }
 179
 180 void Corpus::get_compressed_filename_block(uint32_t docid, function<void(string)> cb) const
 181 {
 182         // Read the file offset from this docid and the next one.
 183         // This is always allowed, since we have a sentinel block at the end.
 184         engine->submit_read(fd, sizeof(uint64_t) * 2, offset_for_block(docid), [this, cb{ move(cb) }](string s) {
 185                 const uint64_t *ptr = reinterpret_cast<const uint64_t *>(s.data());
 186                 off_t offset = ptr[0];
 187                 size_t len = ptr[1] - ptr[0];
 188                 engine->submit_read(fd, len, offset, cb);
 189         });
 190 }
 191
 192 size_t Corpus::get_num_filename_blocks() const
 193 {
 194         return hdr.num_docids;
 195 }
 196
 197 uint64_t scan_file_block(const vector<string> &needles, string_view compressed,
 198                        unordered_map<string, bool> *access_rx_cache, int seq,
 199                        Serializer *serializer)
 200 {
 201         uint64_t matched = 0;
 202
 203         unsigned long long uncompressed_len = ZSTD_getFrameContentSize(compressed.data(), compressed.size());
 204         if (uncompressed_len == ZSTD_CONTENTSIZE_UNKNOWN || uncompressed_len == ZSTD_CONTENTSIZE_ERROR) {
 205                 fprintf(stderr, "ZSTD_getFrameContentSize() failed\n");
 206                 exit(1);
 207         }
 208
 209         string block;
 210         block.resize(uncompressed_len + 1);
 211
 212         size_t err = ZSTD_decompress(&block[0], block.size(), compressed.data(),
 213                                      compressed.size());
 214         if (ZSTD_isError(err)) {
 215                 fprintf(stderr, "ZSTD_decompress(): %s\n", ZSTD_getErrorName(err));
 216                 exit(1);
 217         }
 218         block[block.size() - 1] = '\0';
 219
 220         bool immediate_print = (serializer == nullptr || serializer->ready_to_print(seq));
 221         vector<string> delayed;
 222
 223         for (const char *filename = block.data();
 224              filename != block.data() + block.size();
 225              filename += strlen(filename) + 1) {
 226                 bool found = true;
 227                 for (const string &needle : needles) {
 228                         if (strstr(filename, needle.c_str()) == nullptr) {
 229                                 found = false;
 230                                 break;
 231                         }
 232                 }
 233                 if (found && has_access(filename, access_rx_cache)) {
 234                         ++matched;
 235                         if (only_count) continue;
 236                         if (immediate_print) {
 237                                 if (print_nul) {
 238                                         printf("%s%c", filename, 0);
 239                                 } else {
 240                                         printf("%s\n", filename);
 241                                 }
 242                         } else {
 243                                 delayed.push_back(filename);
 244                         }
 245                 }
 246         }
 247         if (serializer != nullptr && !only_count) {
 248                 if (immediate_print) {
 249                         serializer->release_current();
 250                 } else {
 251                         serializer->print_delayed(seq, move(delayed));
 252                 }
 253         }
 254         return matched;
 255 }
 256
 257 size_t scan_docids(const vector<string> &needles, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
 258 {
 259         Serializer docids_in_order;
 260         unordered_map<string, bool> access_rx_cache;
 261         uint64_t matched = 0;
 262         for (size_t i = 0; i < docids.size(); ++i) {
 263                 uint32_t docid = docids[i];
 264                 corpus.get_compressed_filename_block(docid, [i, &matched, &needles, &access_rx_cache, &docids_in_order](string compressed) {
 265                         matched += scan_file_block(needles, compressed, &access_rx_cache, i, &docids_in_order);
 266                 });
 267         }
 268         engine->finish();
 269         return matched;
 270 }
 271
 272 // We do this sequentially, as it's faster than scattering
 273 // a lot of I/O through io_uring and hoping the kernel will
 274 // coalesce it plus readahead for us.
 275 uint64_t scan_all_docids(const vector<string> &needles, int fd, const Corpus &corpus, IOUringEngine *engine)
 276 {
 277         unordered_map<string, bool> access_rx_cache;
 278         uint32_t num_blocks = corpus.get_num_filename_blocks();
 279         unique_ptr<uint64_t[]> offsets(new uint64_t[num_blocks + 1]);
 280         complete_pread(fd, offsets.get(), (num_blocks + 1) * sizeof(uint64_t), corpus.offset_for_block(0));
 281         string compressed;
 282         uint64_t matched = 0;
 283         for (uint32_t io_docid = 0; io_docid < num_blocks; io_docid += 32) {
 284                 uint32_t last_docid = std::min(io_docid + 32, num_blocks);
 285                 size_t io_len = offsets[last_docid] - offsets[io_docid];
 286                 if (compressed.size() < io_len) {
 287                         compressed.resize(io_len);
 288                 }
 289                 complete_pread(fd, &compressed[0], io_len, offsets[io_docid]);
 290
 291                 for (uint32_t docid = io_docid; docid < last_docid; ++docid) {
 292                         size_t relative_offset = offsets[docid] - offsets[io_docid];
 293                         size_t len = offsets[docid + 1] - offsets[docid];
 294                         matched += scan_file_block(needles, { &compressed[relative_offset], len }, &access_rx_cache, 0, nullptr);
 295                 }
 296         }
 297         return matched;
 298 }
 299
 300 void do_search_file(const vector<string> &needles, const char *filename)
 301 {
 302         int fd = open(filename, O_RDONLY);
 303         if (fd == -1) {
 304                 perror(filename);
 305                 exit(1);
 306         }
 307
 308         // Drop privileges.
 309         if (setgid(getgid()) != 0) {
 310                 perror("setgid");
 311                 exit(EXIT_FAILURE);
 312         }
 313
 314         steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
 315         if (access("/", R_OK | X_OK)) {
 316                 // We can't find anything, no need to bother...
 317                 return;
 318         }
 319
 320         IOUringEngine engine;
 321         Corpus corpus(fd, &engine);
 322         dprintf("Corpus init done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 323
 324         vector<pair<Trigram, size_t>> trigrams;
 325         uint64_t shortest_so_far = numeric_limits<uint32_t>::max();
 326         for (const string &needle : needles) {
 327                 if (needle.size() < 3)
 328                         continue;
 329                 for (size_t i = 0; i < needle.size() - 2; ++i) {
 330                         uint32_t trgm = read_trigram(needle, i);
 331                         corpus.find_trigram(trgm, [trgm, &trigrams, &shortest_so_far](const Trigram *trgmptr, size_t len) {
 332                                 if (trgmptr == nullptr) {
 333                                         dprintf("trigram '%c%c%c' isn't found, we abort the search\n",
 334                                                 trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff);
 335                                         if (only_count) {
 336                                                 printf("0\n");
 337                                         }
 338                                         exit(0);
 339                                 }
 340                                 if (trgmptr->num_docids > shortest_so_far * 100) {
 341                                         dprintf("not loading trigram '%c%c%c' with %u docids, it would be ignored later anyway\n",
 342                                                 trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff,
 343                                                 trgmptr->num_docids);
 344                                 } else {
 345                                         trigrams.emplace_back(*trgmptr, len);
 346                                         shortest_so_far = std::min<uint64_t>(shortest_so_far, trgmptr->num_docids);
 347                                 }
 348                         });
 349                 }
 350         }
 351         engine.finish();
 352         dprintf("Hashtable lookups done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
 353
 354         if (trigrams.empty()) {
 355                 // Too short for trigram matching. Apply brute force.
 356                 // (We could have searched through all trigrams that matched
 357                 // the pattern and done a union of them, but that's a lot of
 358                 // work for fairly unclear gain.)
 359                 uint64_t matched = scan_all_docids(needles, fd, corpus, &engine);
 360                 printf("%zu\n", matched);
 361                 return;
 362         }
 363         sort(trigrams.begin(), trigrams.end());
 364         {
 365                 auto last = unique(trigrams.begin(), trigrams.end());
 366                 trigrams.erase(last, trigrams.end());
 367         }
 368         sort(trigrams.begin(), trigrams.end(),
 369              [&](const pair<Trigram, size_t> &a, const pair<Trigram, size_t> &b) {
 370                      return a.first.num_docids < b.first.num_docids;
 371              });
 372
 373         vector<uint32_t> in1, in2, out;
 374         bool done = false;
 375         for (auto [trgmptr, len] : trigrams) {
 376                 if (!in1.empty() && trgmptr.num_docids > in1.size() * 100) {
 377                         uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
 378                         dprintf("trigram '%c%c%c' (%zu bytes) has %u entries, ignoring the rest (will "
 379                                 "weed out false positives later)\n",
 380                                 trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff,
 381                                 len, trgmptr.num_docids);
 382                         break;
 383                 }
 384
 385                 // Only stay a certain amount ahead, so that we don't spend I/O
 386                 // on reading the latter, large posting lists. We are unlikely
 387                 // to need them anyway, even if they should come in first.
 388                 if (engine.get_waiting_reads() >= 5) {
 389                         engine.finish();
 390                         if (done)
 391                                 break;
 392                 }
 393                 engine.submit_read(fd, len, trgmptr.offset, [trgmptr{trgmptr}, len{len}, &done, &in1, &in2, &out](string s) {
 394                         if (done)
 395                                 return;
 396                         uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
 397                         size_t num = trgmptr.num_docids;
 398                         unsigned char *pldata = reinterpret_cast<unsigned char *>(s.data());
 399                         if (in1.empty()) {
 400                                 in1.resize(num + 128);
 401                                 decode_pfor_delta1<128>(pldata, num, /*interleaved=*/true, &in1[0]);
 402                                 in1.resize(num);
 403                                 dprintf("trigram '%c%c%c' (%zu bytes) decoded to %zu entries\n", trgm & 0xff,
 404                                         (trgm >> 8) & 0xff, (trgm >> 16) & 0xff, len, num);
 405                         } else {
 406                                 if (in2.size() < num + 128) {
 407                                         in2.resize(num + 128);
 408                                 }
 409                                 decode_pfor_delta1<128>(pldata, num, /*interleaved=*/true, &in2[0]);
 410
 411                                 out.clear();
 412                                 set_intersection(in1.begin(), in1.end(), in2.begin(), in2.begin() + num,
 413                                                  back_inserter(out));
 414                                 swap(in1, out);
 415                                 dprintf("trigram '%c%c%c' (%zu bytes) decoded to %zu entries, %zu left\n",
 416                                         trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff,
 417                                         len, num, in1.size());
 418                                 if (in1.empty()) {
 419                                         dprintf("no matches (intersection list is empty)\n");
 420                                         done = true;
 421                                 }
 422                         }
 423                 });
 424         }
 425         engine.finish();
 426         if (done) {
 427                 return;
 428         }
 429         dprintf("Intersection done after %.1f ms. Doing final verification and printing:\n",
 430                 1e3 * duration<float>(steady_clock::now() - start).count());
 431
 432         uint64_t matched = scan_docids(needles, in1, corpus, &engine);
 433         dprintf("Done in %.1f ms, found %zu matches.\n",
 434                 1e3 * duration<float>(steady_clock::now() - start).count(), matched);
 435
 436         if (only_count) {
 437                 printf("%zu\n", matched);
 438         }
 439 }
 440
 441 void usage()
 442 {
 443         // The help text comes from mlocate.
 444         printf("Usage: plocate [OPTION]... PATTERN...\n");
 445         printf("\n");
 446         printf("  -c, --count            only print number of found entries\n");
 447         printf("  -d, --database DBPATH  use DBPATH instead of default database (which is\n");
 448         printf("                         %s)\n", dbpath);
 449         printf("  -h, --help             print this help\n");
 450         printf("  -0, --null             separate entries with NUL on output\n");
 451 }
 452
 453 int main(int argc, char **argv)
 454 {
 455         static const struct option long_options[] = {
 456                 { "help", no_argument, 0, 'h' },
 457                 { "count", no_argument, 0, 'c' },
 458                 { "database", required_argument, 0, 'd' },
 459                 { "null", no_argument, 0, '0' },
 460                 { 0, 0, 0, 0 }
 461         };
 462
 463         for (;;) {
 464                 int option_index = 0;
 465                 int c = getopt_long(argc, argv, "cd:h0", long_options, &option_index);
 466                 if (c == -1) {
 467                         break;
 468                 }
 469                 switch (c) {
 470                 case 'c':
 471                         only_count = true;
 472                         break;
 473                 case 'd':
 474                         dbpath = strdup(optarg);
 475                         break;
 476                 case 'h':
 477                         usage();
 478                         exit(0);
 479                 case '0':
 480                         print_nul = true;
 481                         break;
 482                 default:
 483                         exit(1);
 484                 }
 485         }
 486
 487         vector<string> needles;
 488         for (int i = optind; i < argc; ++i) {
 489                 needles.push_back(argv[i]);
 490         }
 491         if (needles.empty()) {
 492                 fprintf(stderr, "plocate: no pattern to search for specified\n");
 493                 exit(0);
 494         }
 495         do_search_file(needles, dbpath);
 496 }