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

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