kv/KeyValueDB: add column family

src/common/options.cc

@@ -3447,6 +3447,10 @@ std::vector<Option> get_global_options() {
     .set_default("compression=kNoCompression,max_write_buffer_number=4,min_write_buffer_number_to_merge=1,recycle_log_file_num=4,write_buffer_size=268435456,writable_file_max_buffer_size=0,compaction_readahead_size=2097152")
     .set_description("Rocksdb options"),
 
+    Option("bluestore_rocksdb_cf", Option::TYPE_BOOL, Option::LEVEL_ADVANCED)
+    .set_default(false)
+    .set_description("Enable use of rocksdb column families for bluestore metadata"),
+
     Option("bluestore_fsck_on_mount", Option::TYPE_BOOL, Option::LEVEL_DEV)
     .set_default(false)
     .set_description("Run fsck at mount"),

src/kv/KeyValueDB.h

@@ -15,18 +15,31 @@
 #include "common/perf_counters.h"
 
 using std::string;
+using std::vector;
 /**
  * Defines virtual interface to be implemented by key value store
  *
  * Kyoto Cabinet or LevelDB should implement this
  */
 class KeyValueDB {
 public:
+  /*
+   *  See RocksDB's definition of a column family(CF) and how to use it.
+   *  The interfaces of KeyValueDB is extended, when a column family is created.
+   *  Prefix will be the name of column family to use.
+   */
+  struct ColumnFamily {
+    string name;      //< name of this individual column family
+    string option;    //< configure option string for this CF
+    ColumnFamily(const string &name, const string &option)
+      : name(name), option(option) {}
+  };
+
   class TransactionImpl {
   public:
     /// Set Keys
     void set(
-      const std::string &prefix,                 ///< [in] Prefix for keys
+      const std::string &prefix,                      ///< [in] Prefix for keys, or CF name
       const std::map<std::string, bufferlist> &to_set ///< [in] keys/values to set
     ) {
       std::map<std::string, bufferlist>::const_iterator it;
@@ -36,8 +49,8 @@ class KeyValueDB {
 
     /// Set Keys (via encoded bufferlist)
     void set(
-      const std::string &prefix,      ///< [in] prefix
-      bufferlist& to_set_bl     ///< [in] encoded key/values to set
+      const std::string &prefix,      ///< [in] prefix, or CF name
+      bufferlist& to_set_bl           ///< [in] encoded key/values to set
       ) {
       bufferlist::iterator p = to_set_bl.begin();
       uint32_t num;
@@ -53,9 +66,9 @@ class KeyValueDB {
 
     /// Set Key
     virtual void set(
-      const std::string &prefix,   ///< [in] Prefix for the key
+      const std::string &prefix,      ///< [in] Prefix or CF for the key
       const std::string &k,	      ///< [in] Key to set
-      const bufferlist &bl    ///< [in] Value to set
+      const bufferlist &bl            ///< [in] Value to set
       ) = 0;
     virtual void set(
       const std::string &prefix,
@@ -67,8 +80,8 @@ class KeyValueDB {
 
     /// Removes Keys (via encoded bufferlist)
     void rmkeys(
-      const std::string &prefix,   ///< [in] Prefix to search for
-      bufferlist &keys_bl ///< [in] Keys to remove
+      const std::string &prefix,     ///< [in] Prefix or CF to search for
+      bufferlist &keys_bl            ///< [in] Keys to remove
     ) {
       bufferlist::iterator p = keys_bl.begin();
       uint32_t num;
@@ -82,7 +95,7 @@ class KeyValueDB {
 
     /// Removes Keys
     void rmkeys(
-      const std::string &prefix,   ///< [in] Prefix to search for
+      const std::string &prefix,        ///< [in] Prefix/CF to search for
       const std::set<std::string> &keys ///< [in] Keys to remove
     ) {
       std::set<std::string>::const_iterator it;
@@ -92,8 +105,8 @@ class KeyValueDB {
 
     /// Remove Key
     virtual void rmkey(
-      const std::string &prefix,   ///< [in] Prefix to search for
-      const std::string &k	      ///< [in] Key to remove
+      const std::string &prefix,       ///< [in] Prefix/CF to search for
+      const std::string &k	       ///< [in] Key to remove
       ) = 0;
     virtual void rmkey(
       const std::string &prefix,   ///< [in] Prefix to search for
@@ -109,13 +122,13 @@ class KeyValueDB {
     /// If a key is overwritten (by calling set multiple times), then the result
     /// of calling rm_single_key on this key is undefined.
     virtual void rm_single_key(
-      const std::string &prefix,   ///< [in] Prefix to search for
+      const std::string &prefix,      ///< [in] Prefix/CF to search for
       const std::string &k	      ///< [in] Key to remove
       ) { return rmkey(prefix, k);}
 
     /// Removes keys beginning with prefix
     virtual void rmkeys_by_prefix(
-      const std::string &prefix ///< [in] Prefix by which to remove keys
+      const std::string &prefix       ///< [in] Prefix/CF by which to remove keys
       ) = 0;
 
     virtual void rm_range_keys(
@@ -126,7 +139,7 @@ class KeyValueDB {
 
     /// Merge value into key
     virtual void merge(
-      const std::string &prefix,   ///< [in] Prefix ==> MUST match some established merge operator
+      const std::string &prefix,   ///< [in] Prefix/CF ==> MUST match some established merge operator
       const std::string &key,      ///< [in] Key to be merged
       const bufferlist  &value     ///< [in] value to be merged into key
     ) { assert(0 == "Not implemented"); }
@@ -144,8 +157,16 @@ class KeyValueDB {
   static int test_init(const std::string& type, const std::string& dir);
   virtual int init(string option_str="") = 0;
   virtual int open(std::ostream &out) = 0;
+  virtual int open(std::ostream &out, const vector<ColumnFamily>& cfs) {
+    assert(0 == "Not implemented");
+  }
   virtual int create_and_open(std::ostream &out) = 0;
   virtual void close() { }
+  // vector cfs contains column families to be created when db is created.
+  virtual int create_and_open(std::ostream &out,
+			      const vector<ColumnFamily>& cfs) {
+    assert(0 == "Not implemented");
+  }
 
   virtual Transaction get_transaction() = 0;
   virtual int submit_transaction(Transaction) = 0;
@@ -155,13 +176,13 @@ class KeyValueDB {
 
   /// Retrieve Keys
   virtual int get(
-    const std::string &prefix,        ///< [in] Prefix for key
-    const std::set<std::string> &key,      ///< [in] Key to retrieve
-    std::map<std::string, bufferlist> *out ///< [out] Key value retrieved
+    const std::string &prefix,               ///< [in] Prefix/CF for key
+    const std::set<std::string> &key,        ///< [in] Key to retrieve
+    std::map<std::string, bufferlist> *out   ///< [out] Key value retrieved
     ) = 0;
-  virtual int get(const std::string &prefix, ///< [in] prefix
+  virtual int get(const std::string &prefix, ///< [in] prefix or CF name
 		  const std::string &key,    ///< [in] key
-		  bufferlist *value) {  ///< [out] value
+		  bufferlist *value) {       ///< [out] value
     std::set<std::string> ks;
     ks.insert(key);
     std::map<std::string,bufferlist> om;
@@ -180,7 +201,10 @@ class KeyValueDB {
     return get(prefix, string(key, keylen), value);
   }
 
-  class GenericIteratorImpl {
+  // This superclass is used both by kv iterators *and* by the ObjectMap
+  // omap iterator.  The class hiearchies are unfortunatley tied together
+  // by the legacy DBOjectMap implementation :(.
+  class SimplestIteratorImpl {
   public:
     virtual int seek_to_first() = 0;
     virtual int upper_bound(const std::string &after) = 0;
@@ -190,9 +214,29 @@ class KeyValueDB {
     virtual std::string key() = 0;
     virtual bufferlist value() = 0;
     virtual int status() = 0;
-    virtual ~GenericIteratorImpl() {}
+    virtual ~SimplestIteratorImpl() {}
   };
 
+  class IteratorImpl : public SimplestIteratorImpl {
+  public:
+    virtual ~IteratorImpl() {}
+    virtual int seek_to_last() = 0;
+    virtual int prev(bool validate=true) = 0;
+    virtual std::pair<std::string, std::string> raw_key() = 0;
+    virtual bufferptr value_as_ptr() {
+      bufferlist bl = value();
+      if (bl.length() == 1) {
+        return *bl.buffers().begin();
+      } else if (bl.length() == 0) {
+        return bufferptr();
+      } else {
+	ceph_abort();
+      }
+    }
+  };
+  typedef ceph::shared_ptr< IteratorImpl > Iterator;
+
+  // This is the low-level iterator implemented by the underlying KV store.
   class WholeSpaceIteratorImpl {
   public:
     virtual int seek_to_first() = 0;
@@ -227,18 +271,20 @@ class KeyValueDB {
   };
   typedef ceph::shared_ptr< WholeSpaceIteratorImpl > WholeSpaceIterator;
 
-  class IteratorImpl : public GenericIteratorImpl {
+private:
+  // This class filters a WholeSpaceIterator by a prefix.
+  class PrefixIteratorImpl : public IteratorImpl {
     const std::string prefix;
     WholeSpaceIterator generic_iter;
   public:
-    IteratorImpl(const std::string &prefix, WholeSpaceIterator iter) :
+    PrefixIteratorImpl(const std::string &prefix, WholeSpaceIterator iter) :
       prefix(prefix), generic_iter(iter) { }
-    ~IteratorImpl() override { }
+    ~PrefixIteratorImpl() override { }
 
     int seek_to_first() override {
       return generic_iter->seek_to_first(prefix);
     }
-    int seek_to_last() {
+    int seek_to_last() override {
       return generic_iter->seek_to_last(prefix);
     }
     int upper_bound(const std::string &after) override {
@@ -264,7 +310,7 @@ class KeyValueDB {
       }      
     }
 
-    int prev(bool validate=true) {
+    int prev(bool validate=true) override {
       if (validate) {
         if (valid())
           return generic_iter->prev();
@@ -276,28 +322,34 @@ class KeyValueDB {
     std::string key() override {
       return generic_iter->key();
     }
-    std::pair<std::string, std::string> raw_key() {
+    std::pair<std::string, std::string> raw_key() override {
       return generic_iter->raw_key();
     }
     bufferlist value() override {
       return generic_iter->value();
     }
-    bufferptr value_as_ptr() {
+    bufferptr value_as_ptr() override {
       return generic_iter->value_as_ptr();
     }
     int status() override {
       return generic_iter->status();
     }
   };
+public:
 
-  typedef ceph::shared_ptr< IteratorImpl > Iterator;
+  virtual WholeSpaceIterator get_wholespace_iterator() = 0;
+  virtual Iterator get_iterator(const std::string &prefix) {
+    return std::make_shared<PrefixIteratorImpl>(
+      prefix,
+      get_wholespace_iterator());
+  }
 
-  WholeSpaceIterator get_iterator() {
-    return _get_iterator();
+  void add_column_family(const std::string& cf_name, void *handle) {
+    cf_handles.insert(std::make_pair(cf_name, handle));
   }
 
-  Iterator get_iterator(const std::string &prefix) {
-    return std::make_shared<IteratorImpl>(prefix, get_iterator());
+  bool is_column_family(const std::string& prefix) {
+    return cf_handles.count(prefix);
   }
 
   virtual uint64_t get_estimated_size(std::map<std::string,uint64_t> &extra) = 0;
@@ -361,11 +413,12 @@ class KeyValueDB {
     return nullptr;
   }
 protected:
-  /// List of matching prefixes and merge operators
+  /// List of matching prefixes/ColumnFamilies and merge operators
   std::vector<std::pair<std::string,
 			std::shared_ptr<MergeOperator> > > merge_ops;
 
-  virtual WholeSpaceIterator _get_iterator() = 0;
+  /// column families in use, name->handle
+  std::unordered_map<std::string, void *> cf_handles;
 };
 
 #endif

src/kv/KineticStore.h

@@ -56,7 +56,7 @@ class KineticStore : public KeyValueDB {
     return do_open(out, false);
   }
   /// Creates underlying db if missing and opens it
-  int create_and_open(ostream &out) {
+  int create_and_open(ostream &out) override {
     return do_open(out, true);
   }
 
@@ -149,8 +149,7 @@ class KineticStore : public KeyValueDB {
   }
 
 
-protected:
-  WholeSpaceIterator _get_iterator() {
+  WholeSpaceIterator get_wholespace_iterator() {
     return std::make_shared<KineticWholeSpaceIteratorImpl>(kinetic_conn.get());
   }
 };

src/kv/LevelDBStore.h

@@ -402,8 +402,7 @@ class LevelDBStore : public KeyValueDB {
   }
 
 
-protected:
-  WholeSpaceIterator _get_iterator() override {
+  WholeSpaceIterator get_wholespace_iterator() override {
     return std::make_shared<LevelDBWholeSpaceIteratorImpl>(
 	db->NewIterator(leveldb::ReadOptions()));
   }

src/kv/MemDB.h

@@ -201,9 +201,7 @@ class MemDB : public KeyValueDB
     return 0;
   }
 
-protected:
-
-  WholeSpaceIterator _get_iterator() override {
+  WholeSpaceIterator get_wholespace_iterator() override {
     return std::shared_ptr<KeyValueDB::WholeSpaceIteratorImpl>(
       new MDBWholeSpaceIteratorImpl(&m_map, &m_lock, &iterator_seq_no, m_using_btree));
   }