integrate rate limiter into rocksdb

Summary:
Add option and plugin rate limiter for PosixWritableFile. The rate
limiter only applies to flush and compaction. WAL and MANIFEST are
excluded from this enforcement.

Test Plan: db_test

Reviewers: igor, yhchiang, sdong

Reviewed By: sdong

Subscribers: leveldb

Differential Revision: https://reviews.facebook.net/D19425

include/rocksdb/env.h

@@ -35,6 +35,7 @@ class WritableFile;
 class RandomRWFile;
 class Directory;
 struct DBOptions;
+class RateLimiter;
 
 using std::unique_ptr;
 using std::shared_ptr;
@@ -74,6 +75,9 @@ struct EnvOptions {
   // write. By default, we set it to true for MANIFEST writes and false for
   // WAL writes
   bool fallocate_with_keep_size = true;
+
+  // If not nullptr, write rate limiting is enabled for flush and compaction
+  RateLimiter* rate_limiter = nullptr;
 };
 
 class Env {
@@ -379,7 +383,10 @@ class RandomAccessFile {
 // at a time to the file.
 class WritableFile {
  public:
-  WritableFile() : last_preallocated_block_(0), preallocation_block_size_ (0) {
+  WritableFile()
+    : last_preallocated_block_(0),
+      preallocation_block_size_(0),
+      io_priority_(Env::IO_TOTAL) {
   }
   virtual ~WritableFile();
 
@@ -398,6 +405,14 @@ class WritableFile {
     return Sync();
   }
 
+  /*
+   * Change the priority in rate limiter if rate limiting is enabled.
+   * If rate limiting is not enabled, this call has no effect.
+   */
+  virtual void SetIOPriority(Env::IOPriority pri) {
+    io_priority_ = pri;
+  }
+
   /*
    * Get the size of valid data in the file.
    */
@@ -482,6 +497,9 @@ class WritableFile {
   // No copying allowed
   WritableFile(const WritableFile&);
   void operator=(const WritableFile&);
+
+ protected:
+  Env::IOPriority io_priority_;
 };
 
 // A file abstraction for random reading and writing.

include/rocksdb/options.h

@@ -39,8 +39,7 @@ class Slice;
 class SliceTransform;
 class Statistics;
 class InternalKeyComparator;
-
-using std::shared_ptr;
+class RateLimiter;
 
 // DB contents are stored in a set of blocks, each of which holds a
 // sequence of key,value pairs.  Each block may be compressed before
@@ -133,7 +132,7 @@ struct ColumnFamilyOptions {
   // for the first time. It's necessary to specify a merge operator when
   // openning the DB in this case.
   // Default: nullptr
-  shared_ptr<MergeOperator> merge_operator;
+  std::shared_ptr<MergeOperator> merge_operator;
 
   // A single CompactionFilter instance to call into during compaction.
   // Allows an application to modify/delete a key-value during background
@@ -206,12 +205,12 @@ struct ColumnFamilyOptions {
   // If non-NULL use the specified cache for blocks.
   // If NULL, rocksdb will automatically create and use an 8MB internal cache.
   // Default: nullptr
-  shared_ptr<Cache> block_cache;
+  std::shared_ptr<Cache> block_cache;
 
   // If non-NULL use the specified cache for compressed blocks.
   // If NULL, rocksdb will not use a compressed block cache.
   // Default: nullptr
-  shared_ptr<Cache> block_cache_compressed;
+  std::shared_ptr<Cache> block_cache_compressed;
 
   // Approximate size of user data packed per block.  Note that the
   // block size specified here corresponds to uncompressed data.  The
@@ -626,11 +625,16 @@ struct DBOptions {
   // Default: Env::Default()
   Env* env;
 
+  // Use to control write rate of flush and compaction. Flush has higher
+  // priority than compaction. Rate limiting is disabled if nullptr.
+  // Default: nullptr
+  std::shared_ptr<RateLimiter> rate_limiter;
+
   // Any internal progress/error information generated by the db will
   // be written to info_log if it is non-nullptr, or to a file stored
   // in the same directory as the DB contents if info_log is nullptr.
   // Default: nullptr
-  shared_ptr<Logger> info_log;
+  std::shared_ptr<Logger> info_log;
 
   InfoLogLevel info_log_level;
 
@@ -653,7 +657,7 @@ struct DBOptions {
   // If non-null, then we should collect metrics about database operations
   // Statistics objects should not be shared between DB instances as
   // it does not use any locks to prevent concurrent updates.
-  shared_ptr<Statistics> statistics;
+  std::shared_ptr<Statistics> statistics;
 
   // If true, then the contents of data files are not synced
   // to stable storage. Their contents remain in the OS buffers till the