@nxtedition/rocksdb 5.2.21 → 5.2.28

package/deps/rocksdb/rocksdb/trace_replay/trace_replay.h

@@ -5,13 +5,17 @@
 
 #pragma once
 
+#include <atomic>
 #include <memory>
+#include <mutex>
 #include <unordered_map>
 #include <utility>
 
-#include "rocksdb/env.h"
 #include "rocksdb/options.h"
-#include "rocksdb/trace_reader_writer.h"
+#include "rocksdb/rocksdb_namespace.h"
+#include "rocksdb/status.h"
+#include "rocksdb/trace_record.h"
+#include "rocksdb/utilities/replayer.h"
 
 namespace ROCKSDB_NAMESPACE {
 
@@ -22,9 +26,17 @@ class ColumnFamilyHandle;
 class ColumnFamilyData;
 class DB;
 class DBImpl;
+class Env;
 class Slice;
+class SystemClock;
+class TraceReader;
+class TraceWriter;
 class WriteBatch;
 
+struct ReadOptions;
+struct TraceOptions;
+struct WriteOptions;
+
 extern const std::string kTraceMagic;
 const unsigned int kTraceTimestampSize = 8;
 const unsigned int kTraceTypeSize = 1;
@@ -32,53 +44,74 @@ const unsigned int kTracePayloadLengthSize = 4;
 const unsigned int kTraceMetadataSize =
     kTraceTimestampSize + kTraceTypeSize + kTracePayloadLengthSize;
 
-// Supported Trace types.
-enum TraceType : char {
-  kTraceBegin = 1,
-  kTraceEnd = 2,
-  kTraceWrite = 3,
-  kTraceGet = 4,
-  kTraceIteratorSeek = 5,
-  kTraceIteratorSeekForPrev = 6,
-  // Block cache related types.
-  kBlockTraceIndexBlock = 7,
-  kBlockTraceFilterBlock = 8,
-  kBlockTraceDataBlock = 9,
-  kBlockTraceUncompressionDictBlock = 10,
-  kBlockTraceRangeDeletionBlock = 11,
-  // IO Trace related types based on options that will be added in trace file.
-  kIOGeneral = 12,
-  kIOFileName = 13,
-  kIOFileNameAndFileSize = 14,
-  kIOLen = 15,
-  kIOLenAndOffset = 16,
-  // All trace types should be added before kTraceMax
-  kTraceMax,
-};
+static const int kTraceFileMajorVersion = 0;
+static const int kTraceFileMinorVersion = 2;
 
-// TODO: This should also be made part of public interface to help users build
-// custom TracerReaders and TraceWriters.
-//
 // The data structure that defines a single trace.
 struct Trace {
   uint64_t ts;  // timestamp
   TraceType type;
+  // Each bit in payload_map stores which corresponding struct member added in
+  // the payload. Each TraceType has its corresponding payload struct. For
+  // example, if bit at position 0 is set in write payload, then the write batch
+  // will be addedd.
+  uint64_t payload_map = 0;
+  // Each trace type has its own payload_struct, which will be serilized in the
+  // payload.
   std::string payload;
 
   void reset() {
     ts = 0;
     type = kTraceMax;
+    payload_map = 0;
     payload.clear();
   }
 };
 
+enum TracePayloadType : char {
+  // Each member of all query payload structs should have a corresponding flag
+  // here. Make sure to add them sequentially in the order of it is added.
+  kEmptyPayload = 0,
+  kWriteBatchData = 1,
+  kGetCFID = 2,
+  kGetKey = 3,
+  kIterCFID = 4,
+  kIterKey = 5,
+  kIterLowerBound = 6,
+  kIterUpperBound = 7,
+  kMultiGetSize = 8,
+  kMultiGetCFIDs = 9,
+  kMultiGetKeys = 10,
+};
+
 class TracerHelper {
  public:
-  // Encode a trace object into the given string.
+  // Parse the string with major and minor version only
+  static Status ParseVersionStr(std::string& v_string, int* v_num);
+
+  // Parse the trace file version and db version in trace header
+  static Status ParseTraceHeader(const Trace& header, int* trace_version,
+                                 int* db_version);
+
+  // Encode a version 0.1 trace object into the given string.
   static void EncodeTrace(const Trace& trace, std::string* encoded_trace);
 
   // Decode a string into the given trace object.
   static Status DecodeTrace(const std::string& encoded_trace, Trace* trace);
+
+  // Decode a string into the given trace header.
+  static Status DecodeHeader(const std::string& encoded_trace, Trace* header);
+
+  // Set the payload map based on the payload type
+  static bool SetPayloadMap(uint64_t& payload_map,
+                            const TracePayloadType payload_type);
+
+  // Decode a Trace object into the corresponding TraceRecord.
+  // Return Status::OK() if nothing is wrong, record will be set accordingly.
+  // Return Status::NotSupported() if the trace type is not support, or the
+  // corresponding error status, record will be set to nullptr.
+  static Status DecodeTraceRecord(Trace* trace, int trace_file_version,
+                                  std::unique_ptr<TraceRecord>* record);
 };
 
 // Tracer captures all RocksDB operations using a user-provided TraceWriter.
@@ -86,7 +119,7 @@ class TracerHelper {
 // timestamp and type, followed by the trace payload.
 class Tracer {
  public:
-  Tracer(Env* env, const TraceOptions& trace_options,
+  Tracer(SystemClock* clock, const TraceOptions& trace_options,
          std::unique_ptr<TraceWriter>&& trace_writer);
   ~Tracer();
 
@@ -97,13 +130,30 @@ class Tracer {
   Status Get(ColumnFamilyHandle* cfname, const Slice& key);
 
   // Trace Iterators.
-  Status IteratorSeek(const uint32_t& cf_id, const Slice& key);
-  Status IteratorSeekForPrev(const uint32_t& cf_id, const Slice& key);
+  Status IteratorSeek(const uint32_t& cf_id, const Slice& key,
+                      const Slice& lower_bound, const Slice upper_bound);
+  Status IteratorSeekForPrev(const uint32_t& cf_id, const Slice& key,
+                             const Slice& lower_bound, const Slice upper_bound);
+
+  // Trace MultiGet
+
+  Status MultiGet(const size_t num_keys, ColumnFamilyHandle** column_families,
+                  const Slice* keys);
+
+  Status MultiGet(const size_t num_keys, ColumnFamilyHandle* column_family,
+                  const Slice* keys);
+
+  Status MultiGet(const std::vector<ColumnFamilyHandle*>& column_family,
+                  const std::vector<Slice>& keys);
 
   // Returns true if the trace is over the configured max trace file limit.
   // False otherwise.
   bool IsTraceFileOverMax();
 
+  // Returns true if the order of write trace records must match the order of
+  // the corresponding records logged to WAL and applied to the DB.
+  bool IsWriteOrderPreserved() { return trace_options_.preserve_write_order; }
+
   // Writes a trace footer at the end of the tracing
   Status Close();
 
@@ -124,72 +174,10 @@ class Tracer {
   // Returns true if a trace should be skipped, false otherwise.
   bool ShouldSkipTrace(const TraceType& type);
 
-  Env* env_;
+  SystemClock* clock_;
   TraceOptions trace_options_;
   std::unique_ptr<TraceWriter> trace_writer_;
   uint64_t trace_request_count_;
 };
 
-// Replayer helps to replay the captured RocksDB operations, using a user
-// provided TraceReader.
-// The Replayer is instantiated via db_bench today, on using "replay" benchmark.
-class Replayer {
- public:
-  Replayer(DB* db, const std::vector<ColumnFamilyHandle*>& handles,
-           std::unique_ptr<TraceReader>&& reader);
-  ~Replayer();
-
-  // Replay all the traces from the provided trace stream, taking the delay
-  // between the traces into consideration.
-  Status Replay();
-
-  // Replay the provide trace stream, which is the same as Replay(), with
-  // multi-threads. Queries are scheduled in the thread pool job queue.
-  // User can set the number of threads in the thread pool.
-  Status MultiThreadReplay(uint32_t threads_num);
-
-  // Enables fast forwarding a replay by reducing the delay between the ingested
-  // traces.
-  // fast_forward : Rate of replay speedup.
-  //   If 1, replay the operations at the same rate as in the trace stream.
-  //   If > 1, speed up the replay by this amount.
-  Status SetFastForward(uint32_t fast_forward);
-
- private:
-  Status ReadHeader(Trace* header);
-  Status ReadFooter(Trace* footer);
-  Status ReadTrace(Trace* trace);
-
-  // The background function for MultiThreadReplay to execute Get query
-  // based on the trace records.
-  static void BGWorkGet(void* arg);
-
-  // The background function for MultiThreadReplay to execute WriteBatch
-  // (Put, Delete, SingleDelete, DeleteRange) based on the trace records.
-  static void BGWorkWriteBatch(void* arg);
-
-  // The background function for MultiThreadReplay to execute Iterator (Seek)
-  // based on the trace records.
-  static void BGWorkIterSeek(void* arg);
-
-  // The background function for MultiThreadReplay to execute Iterator
-  // (SeekForPrev) based on the trace records.
-  static void BGWorkIterSeekForPrev(void* arg);
-
-  DBImpl* db_;
-  Env* env_;
-  std::unique_ptr<TraceReader> trace_reader_;
-  std::unordered_map<uint32_t, ColumnFamilyHandle*> cf_map_;
-  uint32_t fast_forward_;
-};
-
-// The passin arg of MultiThreadRepkay for each trace record.
-struct ReplayerWorkerArg {
-  DB* db;
-  Trace trace_entry;
-  std::unordered_map<uint32_t, ColumnFamilyHandle*>* cf_map;
-  WriteOptions woptions;
-  ReadOptions roptions;
-};
-
 }  // namespace ROCKSDB_NAMESPACE

package/deps/rocksdb/rocksdb/util/autovector.h

@@ -11,6 +11,7 @@
 #include <stdexcept>
 #include <vector>
 
+#include "port/lang.h"
 #include "rocksdb/rocksdb_namespace.h"
 
 namespace ROCKSDB_NAMESPACE {
@@ -51,25 +52,25 @@ template <class T, size_t kSize = 8>
 class autovector {
  public:
   // General STL-style container member types.
-  typedef T value_type;
-  typedef typename std::vector<T>::difference_type difference_type;
-  typedef typename std::vector<T>::size_type size_type;
-  typedef value_type& reference;
-  typedef const value_type& const_reference;
-  typedef value_type* pointer;
-  typedef const value_type* const_pointer;
+  using value_type = T;
+  using difference_type = typename std::vector<T>::difference_type;
+  using size_type = typename std::vector<T>::size_type;
+  using reference = value_type&;
+  using const_reference = const value_type&;
+  using pointer = value_type*;
+  using const_pointer = const value_type*;
 
   // This class is the base for regular/const iterator
   template <class TAutoVector, class TValueType>
   class iterator_impl {
    public:
     // -- iterator traits
-    typedef iterator_impl<TAutoVector, TValueType> self_type;
-    typedef TValueType value_type;
-    typedef TValueType& reference;
-    typedef TValueType* pointer;
-    typedef typename TAutoVector::difference_type difference_type;
-    typedef std::random_access_iterator_tag iterator_category;
+    using self_type = iterator_impl<TAutoVector, TValueType>;
+    using value_type = TValueType;
+    using reference = TValueType&;
+    using pointer = TValueType*;
+    using difference_type = typename TAutoVector::difference_type;
+    using iterator_category = std::random_access_iterator_tag;
 
     iterator_impl(TAutoVector* vect, size_t index)
         : vect_(vect), index_(index) {};
@@ -175,10 +176,10 @@ class autovector {
     size_t index_ = 0;
   };
 
-  typedef iterator_impl<autovector, value_type> iterator;
-  typedef iterator_impl<const autovector, const value_type> const_iterator;
-  typedef std::reverse_iterator<iterator> reverse_iterator;
-  typedef std::reverse_iterator<const_iterator> const_reverse_iterator;
+  using iterator = iterator_impl<autovector, value_type>;
+  using const_iterator = iterator_impl<const autovector, const value_type>;
+  using reverse_iterator = std::reverse_iterator<iterator>;
+  using const_reverse_iterator = std::reverse_iterator<const_iterator>;
 
   autovector() : values_(reinterpret_cast<pointer>(buf_)) {}
 
@@ -320,6 +321,9 @@ class autovector {
 
   autovector& operator=(const autovector& other) { return assign(other); }
 
+  autovector(autovector&& other) noexcept { *this = std::move(other); }
+  autovector& operator=(autovector&& other);
+
   // -- Iterator Operations
   iterator begin() { return iterator(this, 0); }
 
@@ -352,7 +356,8 @@ class autovector {
 };
 
 template <class T, size_t kSize>
-autovector<T, kSize>& autovector<T, kSize>::assign(const autovector& other) {
+autovector<T, kSize>& autovector<T, kSize>::assign(
+    const autovector<T, kSize>& other) {
   values_ = reinterpret_cast<pointer>(buf_);
   // copy the internal vector
   vect_.assign(other.vect_.begin(), other.vect_.end());
@@ -363,5 +368,20 @@ autovector<T, kSize>& autovector<T, kSize>::assign(const autovector& other) {
 
   return *this;
 }
+
+template <class T, size_t kSize>
+autovector<T, kSize>& autovector<T, kSize>::operator=(
+    autovector<T, kSize>&& other) {
+  values_ = reinterpret_cast<pointer>(buf_);
+  vect_ = std::move(other.vect_);
+  size_t n = other.num_stack_items_;
+  num_stack_items_ = n;
+  other.num_stack_items_ = 0;
+  for (size_t i = 0; i < n; ++i) {
+    values_[i] = std::move(other.values_[i]);
+  }
+  return *this;
+}
+
 #endif  // ROCKSDB_LITE
 }  // namespace ROCKSDB_NAMESPACE

package/deps/rocksdb/rocksdb/util/autovector_test.cc

@@ -64,7 +64,7 @@ TEST_F(AutoVectorTest, PushBackAndPopBack) {
 }
 
 TEST_F(AutoVectorTest, EmplaceBack) {
-  typedef std::pair<size_t, std::string> ValType;
+  using ValType = std::pair<size_t, std::string>;
   autovector<ValType, kSize> vec;
 
   for (size_t i = 0; i < 1000 * kSize; ++i) {

package/deps/rocksdb/rocksdb/util/bloom_impl.h

@@ -41,6 +41,10 @@ class BloomMath {
   // cache line size.
   static double CacheLocalFpRate(double bits_per_key, int num_probes,
                                  int cache_line_bits) {
+    if (bits_per_key <= 0.0) {
+      // Fix a discontinuity
+      return 1.0;
+    }
     double keys_per_cache_line = cache_line_bits / bits_per_key;
     // A reasonable estimate is the average of the FP rates for one standard
     // deviation above and below the mean bucket occupancy. See