rocksdb/trace_replay/io_tracer.h

//  Copyright (c) 2011-present, Facebook, Inc.  All rights reserved.
//  This source code is licensed under both the GPLv2 (found in the
//  COPYING file in the root directory) and Apache 2.0 License
//  (found in the LICENSE.Apache file in the root directory).

#pragma once

#include <atomic>
#include <fstream>

#include "monitoring/instrumented_mutex.h"
#include "rocksdb/env.h"
#include "rocksdb/options.h"
#include "rocksdb/trace_reader_writer.h"
#include "trace_replay/trace_replay.h"

namespace ROCKSDB_NAMESPACE {

struct IOTraceRecord {
  // Required fields for all accesses.
  uint64_t access_timestamp = 0;
  TraceType trace_type = TraceType::kTraceMax;
  std::string file_operation;
  uint64_t latency = 0;
  std::string io_status;
  // Required fields for read.
  std::string file_name;
  uint64_t len = 0;
  uint64_t offset = 0;
  uint64_t file_size = 0;

  IOTraceRecord() {}

  IOTraceRecord(const uint64_t& _access_timestamp, const TraceType& _trace_type,
                const std::string& _file_operation, const uint64_t& _latency,
                const std::string& _io_status, const std::string& _file_name)
      : access_timestamp(_access_timestamp),
        trace_type(_trace_type),
        file_operation(_file_operation),
        latency(_latency),
        io_status(_io_status),
        file_name(_file_name) {}

  IOTraceRecord(const uint64_t& _access_timestamp, const TraceType& _trace_type,
                const std::string& _file_operation, const uint64_t& _latency,
                const std::string& _io_status, const std::string& _file_name,
                const uint64_t& _file_size)
      : access_timestamp(_access_timestamp),
        trace_type(_trace_type),
        file_operation(_file_operation),
        latency(_latency),
        io_status(_io_status),
        file_name(_file_name),
        file_size(_file_size) {}

  IOTraceRecord(const uint64_t& _access_timestamp, const TraceType& _trace_type,
                const std::string& _file_operation, const uint64_t& _latency,
                const std::string& _io_status, const uint64_t& _len = 0,
                const uint64_t& _offset = 0)
      : access_timestamp(_access_timestamp),
        trace_type(_trace_type),
        file_operation(_file_operation),
        latency(_latency),
        io_status(_io_status),
        len(_len),
        offset(_offset) {}
};

struct IOTraceHeader {
  uint64_t start_time;
  uint32_t rocksdb_major_version;
  uint32_t rocksdb_minor_version;
};

// IOTraceWriter writes IO operation as a single trace. Each trace will have a
// timestamp and type, followed by the trace payload.
class IOTraceWriter {
 public:
  IOTraceWriter(Env* env, const TraceOptions& trace_options,
                std::unique_ptr<TraceWriter>&& trace_writer);
  ~IOTraceWriter() = default;
  // No copy and move.
  IOTraceWriter(const IOTraceWriter&) = delete;
  IOTraceWriter& operator=(const IOTraceWriter&) = delete;
  IOTraceWriter(IOTraceWriter&&) = delete;
  IOTraceWriter& operator=(IOTraceWriter&&) = delete;

  Status WriteIOOp(const IOTraceRecord& record);

  // Write a trace header at the beginning, typically on initiating a trace,
  // with some metadata like a magic number and RocksDB version.
  Status WriteHeader();

 private:
  Env* env_;
  TraceOptions trace_options_;
  std::unique_ptr<TraceWriter> trace_writer_;
};

// IOTraceReader helps read the trace file generated by IOTraceWriter.
class IOTraceReader {
 public:
  explicit IOTraceReader(std::unique_ptr<TraceReader>&& reader);
  ~IOTraceReader() = default;
  // No copy and move.
  IOTraceReader(const IOTraceReader&) = delete;
  IOTraceReader& operator=(const IOTraceReader&) = delete;
  IOTraceReader(IOTraceReader&&) = delete;
  IOTraceReader& operator=(IOTraceReader&&) = delete;

  Status ReadHeader(IOTraceHeader* header);

  Status ReadIOOp(IOTraceRecord* record);

 private:
  std::unique_ptr<TraceReader> trace_reader_;
};

// An IO tracer. It uses IOTraceWriter to write the access record to the
// trace file.
class IOTracer {
 public:
  IOTracer();
  ~IOTracer();
  // No copy and move.
  IOTracer(const IOTracer&) = delete;
  IOTracer& operator=(const IOTracer&) = delete;
  IOTracer(IOTracer&&) = delete;
  IOTracer& operator=(IOTracer&&) = delete;

  // no_sanitize is added for tracing_enabled. writer_ is protected under mutex
  // so even if user call Start/EndIOTrace and tracing_enabled is not updated in
  // the meanwhile, WriteIOOp will anyways check the writer_ protected under
  // mutex and ignore the operation if writer_is null. So its ok if
  // tracing_enabled shows non updated value.

#if defined(__clang__)
#if defined(__has_feature) && __has_feature(thread_sanitizer)
#define TSAN_SUPPRESSION __attribute__((no_sanitize("thread")))
#endif  // __has_feature(thread_sanitizer)
#else   // __clang__
#ifdef __SANITIZE_THREAD__
#define TSAN_SUPPRESSION __attribute__((no_sanitize("thread")))
#endif  // __SANITIZE_THREAD__
#endif  // __clang__

#ifndef TSAN_SUPPRESSION
#define TSAN_SUPPRESSION
#endif  // TSAN_SUPPRESSION

  // Start writing IO operations to the trace_writer.
  TSAN_SUPPRESSION Status
  StartIOTrace(Env* env, const TraceOptions& trace_options,
               std::unique_ptr<TraceWriter>&& trace_writer);

  // Stop writing IO operations to the trace_writer.
  TSAN_SUPPRESSION void EndIOTrace();

  TSAN_SUPPRESSION bool is_tracing_enabled() const { return tracing_enabled; }

  Status WriteIOOp(const IOTraceRecord& record);

 private:
  TraceOptions trace_options_;
  // A mutex protects the writer_.
  InstrumentedMutex trace_writer_mutex_;
  std::atomic<IOTraceWriter*> writer_;
  // bool tracing_enabled is added to avoid costly operation of checking atomic
  // variable 'writer_' is nullptr or not in is_tracing_enabled().
  // is_tracing_enabled() is invoked multiple times by FileSystem classes.
  bool tracing_enabled;
};

}  // namespace ROCKSDB_NAMESPACE