src/leveldb/db/repair.cc

   1 // Copyright (c) 2011 The LevelDB Authors. All rights reserved.
   2 // Use of this source code is governed by a BSD-style license that can be
   3 // found in the LICENSE file. See the AUTHORS file for names of contributors.
   4 //
   5 // We recover the contents of the descriptor from the other files we find.
   6 // (1) Any log files are first converted to tables
   7 // (2) We scan every table to compute
   8 //     (a) smallest/largest for the table
   9 //     (b) largest sequence number in the table
  10 // (3) We generate descriptor contents:
  11 //      - log number is set to zero
  12 //      - next-file-number is set to 1 + largest file number we found
  13 //      - last-sequence-number is set to largest sequence# found across
  14 //        all tables (see 2c)
  15 //      - compaction pointers are cleared
  16 //      - every table file is added at level 0
  17 //
  18 // Possible optimization 1:
  19 //   (a) Compute total size and use to pick appropriate max-level M
  20 //   (b) Sort tables by largest sequence# in the table
  21 //   (c) For each table: if it overlaps earlier table, place in level-0,
  22 //       else place in level-M.
  23 // Possible optimization 2:
  24 //   Store per-table metadata (smallest, largest, largest-seq#, ...)
  25 //   in the table's meta section to speed up ScanTable.
  26
  27 #include "db/builder.h"
  28 #include "db/db_impl.h"
  29 #include "db/dbformat.h"
  30 #include "db/filename.h"
  31 #include "db/log_reader.h"
  32 #include "db/log_writer.h"
  33 #include "db/memtable.h"
  34 #include "db/table_cache.h"
  35 #include "db/version_edit.h"
  36 #include "db/write_batch_internal.h"
  37 #include "leveldb/comparator.h"
  38 #include "leveldb/db.h"
  39 #include "leveldb/env.h"
  40
  41 namespace leveldb {
  42
  43 namespace {
  44
  45 class Repairer {
  46  public:
  47   Repairer(const std::string& dbname, const Options& options)
  48       : dbname_(dbname),
  49         env_(options.env),
  50         icmp_(options.comparator),
  51         ipolicy_(options.filter_policy),
  52         options_(SanitizeOptions(dbname, &icmp_, &ipolicy_, options)),
  53         owns_info_log_(options_.info_log != options.info_log),
  54         owns_cache_(options_.block_cache != options.block_cache),
  55         next_file_number_(1) {
  56     // TableCache can be small since we expect each table to be opened once.
  57     table_cache_ = new TableCache(dbname_, &options_, 10);
  58   }
  59
  60   ~Repairer() {
  61     delete table_cache_;
  62     if (owns_info_log_) {
  63       delete options_.info_log;
  64     }
  65     if (owns_cache_) {
  66       delete options_.block_cache;
  67     }
  68   }
  69
  70   Status Run() {
  71     Status status = FindFiles();
  72     if (status.ok()) {
  73       ConvertLogFilesToTables();
  74       ExtractMetaData();
  75       status = WriteDescriptor();
  76     }
  77     if (status.ok()) {
  78       unsigned long long bytes = 0;
  79       for (size_t i = 0; i < tables_.size(); i++) {
  80         bytes += tables_[i].meta.file_size;
  81       }
  82       Log(options_.info_log,
  83           "**** Repaired leveldb %s; "
  84           "recovered %d files; %llu bytes. "
  85           "Some data may have been lost. "
  86           "****",
  87           dbname_.c_str(),
  88           static_cast<int>(tables_.size()),
  89           bytes);
  90     }
  91     return status;
  92   }
  93
  94  private:
  95   struct TableInfo {
  96     FileMetaData meta;
  97     SequenceNumber max_sequence;
  98   };
  99
 100   std::string const dbname_;
 101   Env* const env_;
 102   InternalKeyComparator const icmp_;
 103   InternalFilterPolicy const ipolicy_;
 104   Options const options_;
 105   bool owns_info_log_;
 106   bool owns_cache_;
 107   TableCache* table_cache_;
 108   VersionEdit edit_;
 109
 110   std::vector<std::string> manifests_;
 111   std::vector<uint64_t> table_numbers_;
 112   std::vector<uint64_t> logs_;
 113   std::vector<TableInfo> tables_;
 114   uint64_t next_file_number_;
 115
 116   Status FindFiles() {
 117     std::vector<std::string> filenames;
 118     Status status = env_->GetChildren(dbname_, &filenames);
 119     if (!status.ok()) {
 120       return status;
 121     }
 122     if (filenames.empty()) {
 123       return Status::IOError(dbname_, "repair found no files");
 124     }
 125
 126     uint64_t number;
 127     FileType type;
 128     for (size_t i = 0; i < filenames.size(); i++) {
 129       if (ParseFileName(filenames[i], &number, &type)) {
 130         if (type == kDescriptorFile) {
 131           manifests_.push_back(filenames[i]);
 132         } else {
 133           if (number + 1 > next_file_number_) {
 134             next_file_number_ = number + 1;
 135           }
 136           if (type == kLogFile) {
 137             logs_.push_back(number);
 138           } else if (type == kTableFile) {
 139             table_numbers_.push_back(number);
 140           } else {
 141             // Ignore other files
 142           }
 143         }
 144       }
 145     }
 146     return status;
 147   }
 148
 149   void ConvertLogFilesToTables() {
 150     for (size_t i = 0; i < logs_.size(); i++) {
 151       std::string logname = LogFileName(dbname_, logs_[i]);
 152       Status status = ConvertLogToTable(logs_[i]);
 153       if (!status.ok()) {
 154         Log(options_.info_log, "Log #%llu: ignoring conversion error: %s",
 155             (unsigned long long) logs_[i],
 156             status.ToString().c_str());
 157       }
 158       ArchiveFile(logname);
 159     }
 160   }
 161
 162   Status ConvertLogToTable(uint64_t log) {
 163     struct LogReporter : public log::Reader::Reporter {
 164       Env* env;
 165       Logger* info_log;
 166       uint64_t lognum;
 167       virtual void Corruption(size_t bytes, const Status& s) {
 168         // We print error messages for corruption, but continue repairing.
 169         Log(info_log, "Log #%llu: dropping %d bytes; %s",
 170             (unsigned long long) lognum,
 171             static_cast<int>(bytes),
 172             s.ToString().c_str());
 173       }
 174     };
 175
 176     // Open the log file
 177     std::string logname = LogFileName(dbname_, log);
 178     SequentialFile* lfile;
 179     Status status = env_->NewSequentialFile(logname, &lfile);
 180     if (!status.ok()) {
 181       return status;
 182     }
 183
 184     // Create the log reader.
 185     LogReporter reporter;
 186     reporter.env = env_;
 187     reporter.info_log = options_.info_log;
 188     reporter.lognum = log;
 189     // We intentionally make log::Reader do checksumming so that
 190     // corruptions cause entire commits to be skipped instead of
 191     // propagating bad information (like overly large sequence
 192     // numbers).
 193     log::Reader reader(lfile, &reporter, false/*do not checksum*/,
 194                        0/*initial_offset*/);
 195
 196     // Read all the records and add to a memtable
 197     std::string scratch;
 198     Slice record;
 199     WriteBatch batch;
 200     MemTable* mem = new MemTable(icmp_);
 201     mem->Ref();
 202     int counter = 0;
 203     while (reader.ReadRecord(&record, &scratch)) {
 204       if (record.size() < 12) {
 205         reporter.Corruption(
 206             record.size(), Status::Corruption("log record too small"));
 207         continue;
 208       }
 209       WriteBatchInternal::SetContents(&batch, record);
 210       status = WriteBatchInternal::InsertInto(&batch, mem);
 211       if (status.ok()) {
 212         counter += WriteBatchInternal::Count(&batch);
 213       } else {
 214         Log(options_.info_log, "Log #%llu: ignoring %s",
 215             (unsigned long long) log,
 216             status.ToString().c_str());
 217         status = Status::OK();  // Keep going with rest of file
 218       }
 219     }
 220     delete lfile;
 221
 222     // Do not record a version edit for this conversion to a Table
 223     // since ExtractMetaData() will also generate edits.
 224     FileMetaData meta;
 225     meta.number = next_file_number_++;
 226     Iterator* iter = mem->NewIterator();
 227     status = BuildTable(dbname_, env_, options_, table_cache_, iter, &meta);
 228     delete iter;
 229     mem->Unref();
 230     mem = NULL;
 231     if (status.ok()) {
 232       if (meta.file_size > 0) {
 233         table_numbers_.push_back(meta.number);
 234       }
 235     }
 236     Log(options_.info_log, "Log #%llu: %d ops saved to Table #%llu %s",
 237         (unsigned long long) log,
 238         counter,
 239         (unsigned long long) meta.number,
 240         status.ToString().c_str());
 241     return status;
 242   }
 243
 244   void ExtractMetaData() {
 245     for (size_t i = 0; i < table_numbers_.size(); i++) {
 246       ScanTable(table_numbers_[i]);
 247     }
 248   }
 249
 250   Iterator* NewTableIterator(const FileMetaData& meta) {
 251     // Same as compaction iterators: if paranoid_checks are on, turn
 252     // on checksum verification.
 253     ReadOptions r;
 254     r.verify_checksums = options_.paranoid_checks;
 255     return table_cache_->NewIterator(r, meta.number, meta.file_size);
 256   }
 257
 258   void ScanTable(uint64_t number) {
 259     TableInfo t;
 260     t.meta.number = number;
 261     std::string fname = TableFileName(dbname_, number);
 262     Status status = env_->GetFileSize(fname, &t.meta.file_size);
 263     if (!status.ok()) {
 264       // Try alternate file name.
 265       fname = SSTTableFileName(dbname_, number);
 266       Status s2 = env_->GetFileSize(fname, &t.meta.file_size);
 267       if (s2.ok()) {
 268         status = Status::OK();
 269       }
 270     }
 271     if (!status.ok()) {
 272       ArchiveFile(TableFileName(dbname_, number));
 273       ArchiveFile(SSTTableFileName(dbname_, number));
 274       Log(options_.info_log, "Table #%llu: dropped: %s",
 275           (unsigned long long) t.meta.number,
 276           status.ToString().c_str());
 277       return;
 278     }
 279
 280     // Extract metadata by scanning through table.
 281     int counter = 0;
 282     Iterator* iter = NewTableIterator(t.meta);
 283     bool empty = true;
 284     ParsedInternalKey parsed;
 285     t.max_sequence = 0;
 286     for (iter->SeekToFirst(); iter->Valid(); iter->Next()) {
 287       Slice key = iter->key();
 288       if (!ParseInternalKey(key, &parsed)) {
 289         Log(options_.info_log, "Table #%llu: unparsable key %s",
 290             (unsigned long long) t.meta.number,
 291             EscapeString(key).c_str());
 292         continue;
 293       }
 294
 295       counter++;
 296       if (empty) {
 297         empty = false;
 298         t.meta.smallest.DecodeFrom(key);
 299       }
 300       t.meta.largest.DecodeFrom(key);
 301       if (parsed.sequence > t.max_sequence) {
 302         t.max_sequence = parsed.sequence;
 303       }
 304     }
 305     if (!iter->status().ok()) {
 306       status = iter->status();
 307     }
 308     delete iter;
 309     Log(options_.info_log, "Table #%llu: %d entries %s",
 310         (unsigned long long) t.meta.number,
 311         counter,
 312         status.ToString().c_str());
 313
 314     if (status.ok()) {
 315       tables_.push_back(t);
 316     } else {
 317       RepairTable(fname, t);  // RepairTable archives input file.
 318     }
 319   }
 320
 321   void RepairTable(const std::string& src, TableInfo t) {
 322     // We will copy src contents to a new table and then rename the
 323     // new table over the source.
 324
 325     // Create builder.
 326     std::string copy = TableFileName(dbname_, next_file_number_++);
 327     WritableFile* file;
 328     Status s = env_->NewWritableFile(copy, &file);
 329     if (!s.ok()) {
 330       return;
 331     }
 332     TableBuilder* builder = new TableBuilder(options_, file);
 333
 334     // Copy data.
 335     Iterator* iter = NewTableIterator(t.meta);
 336     int counter = 0;
 337     for (iter->SeekToFirst(); iter->Valid(); iter->Next()) {
 338       builder->Add(iter->key(), iter->value());
 339       counter++;
 340     }
 341     delete iter;
 342
 343     ArchiveFile(src);
 344     if (counter == 0) {
 345       builder->Abandon();  // Nothing to save
 346     } else {
 347       s = builder->Finish();
 348       if (s.ok()) {
 349         t.meta.file_size = builder->FileSize();
 350       }
 351     }
 352     delete builder;
 353     builder = NULL;
 354
 355     if (s.ok()) {
 356       s = file->Close();
 357     }
 358     delete file;
 359     file = NULL;
 360
 361     if (counter > 0 && s.ok()) {
 362       std::string orig = TableFileName(dbname_, t.meta.number);
 363       s = env_->RenameFile(copy, orig);
 364       if (s.ok()) {
 365         Log(options_.info_log, "Table #%llu: %d entries repaired",
 366             (unsigned long long) t.meta.number, counter);
 367         tables_.push_back(t);
 368       }
 369     }
 370     if (!s.ok()) {
 371       env_->DeleteFile(copy);
 372     }
 373   }
 374
 375   Status WriteDescriptor() {
 376     std::string tmp = TempFileName(dbname_, 1);
 377     WritableFile* file;
 378     Status status = env_->NewWritableFile(tmp, &file);
 379     if (!status.ok()) {
 380       return status;
 381     }
 382
 383     SequenceNumber max_sequence = 0;
 384     for (size_t i = 0; i < tables_.size(); i++) {
 385       if (max_sequence < tables_[i].max_sequence) {
 386         max_sequence = tables_[i].max_sequence;
 387       }
 388     }
 389
 390     edit_.SetComparatorName(icmp_.user_comparator()->Name());
 391     edit_.SetLogNumber(0);
 392     edit_.SetNextFile(next_file_number_);
 393     edit_.SetLastSequence(max_sequence);
 394
 395     for (size_t i = 0; i < tables_.size(); i++) {
 396       // TODO(opt): separate out into multiple levels
 397       const TableInfo& t = tables_[i];
 398       edit_.AddFile(0, t.meta.number, t.meta.file_size,
 399                     t.meta.smallest, t.meta.largest);
 400     }
 401
 402     //fprintf(stderr, "NewDescriptor:\n%s\n", edit_.DebugString().c_str());
 403     {
 404       log::Writer log(file);
 405       std::string record;
 406       edit_.EncodeTo(&record);
 407       status = log.AddRecord(record);
 408     }
 409     if (status.ok()) {
 410       status = file->Close();
 411     }
 412     delete file;
 413     file = NULL;
 414
 415     if (!status.ok()) {
 416       env_->DeleteFile(tmp);
 417     } else {
 418       // Discard older manifests
 419       for (size_t i = 0; i < manifests_.size(); i++) {
 420         ArchiveFile(dbname_ + "/" + manifests_[i]);
 421       }
 422
 423       // Install new manifest
 424       status = env_->RenameFile(tmp, DescriptorFileName(dbname_, 1));
 425       if (status.ok()) {
 426         status = SetCurrentFile(env_, dbname_, 1);
 427       } else {
 428         env_->DeleteFile(tmp);
 429       }
 430     }
 431     return status;
 432   }
 433
 434   void ArchiveFile(const std::string& fname) {
 435     // Move into another directory.  E.g., for
 436     //    dir/foo
 437     // rename to
 438     //    dir/lost/foo
 439     const char* slash = strrchr(fname.c_str(), '/');
 440     std::string new_dir;
 441     if (slash != NULL) {
 442       new_dir.assign(fname.data(), slash - fname.data());
 443     }
 444     new_dir.append("/lost");
 445     env_->CreateDir(new_dir);  // Ignore error
 446     std::string new_file = new_dir;
 447     new_file.append("/");
 448     new_file.append((slash == NULL) ? fname.c_str() : slash + 1);
 449     Status s = env_->RenameFile(fname, new_file);
 450     Log(options_.info_log, "Archiving %s: %s\n",
 451         fname.c_str(), s.ToString().c_str());
 452   }
 453 };
 454 }  // namespace
 455
 456 Status RepairDB(const std::string& dbname, const Options& options) {
 457   Repairer repairer(dbname, options);
 458   return repairer.Run();
 459 }
 460
 461 }  // namespace leveldb