From fe834c21002ba7d0661d57f67c3a3a6eed08f2b1 Mon Sep 17 00:00:00 2001
From: Adam Kupczyk <akupczyk@redhat.com>
Date: Thu, 24 Sep 2020 12:54:46 +0200
Subject: [PATCH] BlueStore: Add block_cache logic to column family definition

Modified original onode column family cache into generic feature.
Now 2 options are possible for each column family:
1) use generic block cache but apply different BlockBasedTableOptions
2) create separate block cache. it will be applied for all shards of column family

It is done by specifying special option "block_cache" to CF definition:
Example: O(3,0-13)=block_cache={high_ratio=1.000}

"block_cache" accepts all BlockBasedTableOptions options with additions:
-"type" - binned_lru/lru/clock (default: ceph_options.rocksdb_cache_type)
-"size" - e.g.: 100M           (default: ceph_options.rocksdb_cache_size)
-"high_ratio" - e.g.: 0.75     (default: 0.0)
If any of above is set, new block cache is created, otherwise default is used.

Signed-off-by: Adam Kupczyk <akupczyk@redhat.com>
---
 src/common/options.cc  |   4 +-
 src/kv/RocksDBStore.cc | 177 ++++++++++++++++++++++++++---------------
 src/kv/RocksDBStore.h  |  42 +++++-----
 3 files changed, 134 insertions(+), 89 deletions(-)

diff --git a/src/common/options.cc b/src/common/options.cc
index 2033f72da3e..2465e529e0a 100644
--- a/src/common/options.cc
+++ b/src/common/options.cc
@@ -3790,7 +3790,7 @@ std::vector<Option> get_global_options() {
     .set_long_description("By default RocksDB will load an SST file's index and bloom filters into memory when it is opened and remove them from memory when an SST file is closed.  Thus, memory consumption by indices and bloom filters is directly tied to the number of concurrent SST files allowed to be kept open.  This option instead stores cached indicies and filters in the block cache where they directly compete with other cached data.  By default we set this option to true to better account for and bound rocksdb memory usage and keep filters in memory even when an SST file is closed."),
 
     Option("rocksdb_cache_index_and_filter_blocks_with_high_priority", Option::TYPE_BOOL, Option::LEVEL_DEV)
-    .set_default(true)
+    .set_default(false)
     .set_description("Whether to cache indices and filters in the block cache with high priority")
     .set_long_description("A downside of setting rocksdb_cache_index_and_filter_blocks to true is that regular data can push indices and filters out of memory.  Setting this option to true means they are cached with higher priority than other data and should typically stay in the block cache."),
 
@@ -4531,7 +4531,7 @@ std::vector<Option> get_global_options() {
     .set_description("Enable use of rocksdb column families for bluestore metadata"),
 
     Option("bluestore_rocksdb_cfs", Option::TYPE_STR, Option::LEVEL_DEV)
-    .set_default("m(3) p(3,0-12) O(3,0-13) L")
+    .set_default("m(3) p(3,0-12) O(3,0-13)=block_cache={type=binned_lru} L")
     .set_description("Definition of column families and their sharding")
     .set_long_description("Space separated list of elements: column_def [ '=' rocksdb_options ]. "
 			  "column_def := column_name [ '(' shard_count [ ',' hash_begin '-' [ hash_end ] ] ')' ]. "
diff --git a/src/kv/RocksDBStore.cc b/src/kv/RocksDBStore.cc
index beba8b8ed1e..ac96368f251 100644
--- a/src/kv/RocksDBStore.cc
+++ b/src/kv/RocksDBStore.cc
@@ -456,26 +456,25 @@ int RocksDBStore::create_and_open(ostream &out,
   return do_open(out, true, false, cfs);
 }
 
-int RocksDBStore::init_block_cache(uint64_t size, rocksdb::BlockBasedTableOptions& bbto) {
-  auto shard_bits = cct->_conf()->rocksdb_cache_shard_bits;
-  if (cct->_conf()->rocksdb_cache_type == "binned_lru") {
-    bbto.block_cache = rocksdb_cache::NewBinnedLRUCache(cct, size, shard_bits);
-  } else if (cct->_conf()->rocksdb_cache_type == "lru") {
-    bbto.block_cache = rocksdb::NewLRUCache(size, shard_bits);
-  } else if (cct->_conf()->rocksdb_cache_type == "clock") {
-    bbto.block_cache = rocksdb::NewClockCache(size, shard_bits);
-    if (!bbto.block_cache) {
-      derr << "rocksdb_cache_type '" << cct->_conf()->rocksdb_cache_type
+std::shared_ptr<rocksdb::Cache> RocksDBStore::create_block_cache(
+    const std::string& cache_type, size_t cache_size, double cache_prio_high) {
+  std::shared_ptr<rocksdb::Cache> cache;
+  auto shard_bits = cct->_conf->rocksdb_cache_shard_bits;
+  if (cache_type == "binned_lru") {
+    cache = rocksdb_cache::NewBinnedLRUCache(cct, cache_size, shard_bits, false, cache_prio_high);
+  } else if (cache_type == "lru") {
+    cache = rocksdb::NewLRUCache(cache_size, shard_bits);
+  } else if (cache_type == "clock") {
+    cache = rocksdb::NewClockCache(cache_size, shard_bits);
+    if (!cache) {
+      derr << "rocksdb_cache_type '" << cache
            << "' chosen, but RocksDB not compiled with LibTBB. "
            << dendl;
-      return -EINVAL;
     }
   } else {
-    derr << "unrecognized rocksdb_cache_type '" << g_conf()->rocksdb_cache_type
-      << "'" << dendl;
-    return -EINVAL;
+    derr << "unrecognized rocksdb_cache_type '" << cache_type << "'" << dendl;
   }
-  return 0;
+  return cache;
 }
 
 int RocksDBStore::load_rocksdb_options(bool create_if_missing, rocksdb::Options& opt)
@@ -547,8 +546,11 @@ int RocksDBStore::load_rocksdb_options(bool create_if_missing, rocksdb::Options&
   }
   uint64_t row_cache_size = cache_size * cct->_conf->rocksdb_cache_row_ratio;
   uint64_t block_cache_size = cache_size - row_cache_size;
-  init_block_cache(block_cache_size, bbt_opts);
 
+  bbt_opts.block_cache = create_block_cache(cct->_conf->rocksdb_cache_type, block_cache_size);
+  if (!bbt_opts.block_cache) {
+    return -EINVAL;
+  }
   bbt_opts.block_size = cct->_conf->rocksdb_block_size;
 
   if (row_cache_size > 0)
@@ -772,12 +774,19 @@ int RocksDBStore::create_shards(const rocksdb::Options& opt,
     // the base for new CF
     rocksdb::ColumnFamilyOptions cf_opt(opt);
     // user input options will override the base options
+    std::unordered_map<std::string, std::string> column_opts_map;
+    std::string block_cache_opts;
+    int r = extract_block_cache_options(p.options, &column_opts_map, &block_cache_opts);
+    if (r != 0) {
+      derr << __func__ << " failed to parse options; column family=" << p.name <<
+	" options=" << p.options << dendl;
+      return -EINVAL;
+    }
     rocksdb::Status status;
-    status = rocksdb::GetColumnFamilyOptionsFromString(
-						       cf_opt, p.options, &cf_opt);
+    status = rocksdb::GetColumnFamilyOptionsFromMap(cf_opt, column_opts_map, &cf_opt);
     if (!status.ok()) {
-      derr << __func__ << " invalid db column family option string for CF: "
-	   << p.name << dendl;
+      derr << __func__ << " invalid db options; column family="
+	   << p.name << " options=" << p.options << dendl;
       return -EINVAL;
     }
     install_cf_mergeop(p.name, &cf_opt);
@@ -821,6 +830,7 @@ int RocksDBStore::apply_sharding(const rocksdb::Options& opt,
     }
     r = create_shards(opt, sharding_def);
     if (r != 0 ) {
+      derr << __func__ << " create_shards failed error=" << r << dendl;
       return r;
     }
     opt.env->CreateDir(sharding_def_dir);
@@ -837,9 +847,33 @@ int RocksDBStore::apply_sharding(const rocksdb::Options& opt,
 }
 // linking to rocksdb function defined in options_helper.cc
 // it can parse nested params like "nested_opt={opt1=1;opt2=2}"
-extern rocksdb::Status StringToMap(const std::string& opts_str,
+
+extern rocksdb::Status rocksdb::StringToMap(const std::string& opts_str,
 				   std::unordered_map<std::string, std::string>* opts_map);
 
+int RocksDBStore::extract_block_cache_options(const std::string& opts_str,
+					      std::unordered_map<std::string, std::string>* column_opts_map,
+					      std::string* block_cache_opt)
+{
+  dout(5) << __func__ << " opts_str=" << opts_str << dendl;
+  rocksdb::Status status = rocksdb::StringToMap(opts_str, column_opts_map);
+  if (!status.ok()) {
+    dout(5) << __func__ << " error '" << status.getState() <<
+      "' while parsing options '" << opts_str << "'" << dendl;
+    return -EINVAL;
+  }
+  //extract "block_cache" option
+  if (auto it = column_opts_map->find("block_cache"); it != column_opts_map->end()) {
+    *block_cache_opt = it->second;
+    column_opts_map->erase(it);
+  } else {
+    block_cache_opt->clear();
+  }
+  return 0;
+}
+
+
+
 int RocksDBStore::verify_sharding(const rocksdb::Options& opt,
 				  std::vector<rocksdb::ColumnFamilyDescriptor>& existing_cfs,
 				  std::vector<std::pair<size_t, RocksDBStore::ColumnFamily> >& existing_cfs_shard,
@@ -857,7 +891,9 @@ int RocksDBStore::verify_sharding(const rocksdb::Options& opt,
       derr << __func__ << " cannot read from " << sharding_def_file << dendl;
       return -EIO;
     }
+    dout(20) << __func__ << " sharding=" << stored_sharding_text << dendl;
   } else {
+    dout(30) << __func__ << " no sharding" << dendl;
     //no "sharding_def" present
   }
   //check if sharding_def matches stored_sharding_def
@@ -890,72 +926,87 @@ int RocksDBStore::verify_sharding(const rocksdb::Options& opt,
 
   for (auto& column : stored_sharding_def) {
     rocksdb::ColumnFamilyOptions cf_opt(opt);
-    //sift column.options into 2 categories:
-    // - column family options
-    // - block cache options
     std::unordered_map<std::string, std::string> options_map;
-    status = StringToMap(column.options, &options_map);
-    if (!status.ok()) {
-      dout(5) << __func__ << " error '" << status.getState() << "' while parsing options '" <<
-	column.options << dendl;
-      return -EIO;
-    }
-    //extract "block_cache" options
-    std::unordered_map<std::string, std::string> cache_options_map;
-    for (auto it = options_map.begin; it!= options_map.end() ; /*nop*/ ) {
-      if (it->first.find("block_cache.") == 0) {
-	cache_options_map.insert(it->first.substr(strlen("block_cache.")), it->second);
-	it = erase(it);
-      } else {
-	++it;
-      }
+    std::string block_cache_opt;
+
+    int r = extract_block_cache_options(column.options, &options_map, &block_cache_opt);
+    if (r != 0) {
+      derr << __func__ << " failed to parse options; column family=" << column.name <<
+	" options=" << column.options << dendl;
+      return -EINVAL;
     }
-    status = rocksdb::GetColumnFamilyOptionsFromMap(
-						    cf_opt, options_map, &cf_opt);
+    status = rocksdb::GetColumnFamilyOptionsFromMap(cf_opt, options_map, &cf_opt);
     if (!status.ok()) {
       derr << __func__ << " invalid db column family options for CF '"
 	   << column.name << "': " << column.options << dendl;
-      derr << __func__ << "error = '" << status.getState() << "'" << dendl;
+      derr << __func__ << " error = '" << status.getState() << "'" << dendl;
       return -EINVAL;
     }
     install_cf_mergeop(column.name, &cf_opt);
-    if (!cache_options_map.empty()) {
+
+    if (!block_cache_opt.empty()) {
+      std::unordered_map<std::string, std::string> cache_options_map;
+      status = rocksdb::StringToMap(block_cache_opt, &cache_options_map);
+      if (!status.ok()) {
+	derr << __func__ << " invalid block cache options; column=" << column.name <<
+	  " options=" << block_cache_opt << dendl;
+	derr << __func__ << " error = '" << status.getState() << "'" << dendl;
+	return -EINVAL;
+      }
       bool require_new_block_cache = false;
-      std::string cache_type = cct->_conf()->rocksdb_cache_type;      
-      if (auto it = cache_options_map.find("cache_type"); it !=cache_options_map.end()) {
-	cache_type = it->second();
+      std::string cache_type = cct->_conf->rocksdb_cache_type;
+      if (const auto it = cache_options_map.find("type"); it !=cache_options_map.end()) {
+	cache_type = it->second;
 	cache_options_map.erase(it);
 	require_new_block_cache = true;
       }
-      size_t cache_size = cct->_conf()->rocksdb_cache_size;
-      if (auto it = cache_options_map.find("cache_size"); it !=cache_options_map.end()) {
+      size_t cache_size = cct->_conf->rocksdb_cache_size;
+      if (auto it = cache_options_map.find("size"); it !=cache_options_map.end()) {
 	std::string error;
-	cache_size = strict_iecstrtoll(it->second(), &error);
+	cache_size = strict_iecstrtoll(it->second.c_str(), &error);
 	if (!error.empty()) {
-	  derr << __func__ << " invalid size: '" << it->second() << "'" << dendl;
+	  derr << __func__ << " invalid size: '" << it->second << "'" << dendl;
 	}
 	cache_options_map.erase(it);
 	require_new_block_cache = true;
       }
-      std::shared_ptr<rocksdb::Cache> block_cache;
-      if (require_new_block_cache) {
-	block_cache = create_block_cache(cache_type, case_size);
-	if (!block_cache) {
-	  return -EINVAL;
+      double high_pri_pool_ratio = 0.0;
+      if (auto it = cache_options_map.find("high_ratio"); it !=cache_options_map.end()) {
+	std::string error;
+	high_pri_pool_ratio = strict_strtod(it->second.c_str(), &error);
+	if (!error.empty()) {
+	  derr << __func__ << " invalid high_pri (float): '" << it->second << "'" << dendl;
 	}
-	
-      } else {
-	block_cache = bbt_opts.block_cache;
+	cache_options_map.erase(it);
+	require_new_block_cache = true;
       }
+
       rocksdb::BlockBasedTableOptions column_bbt_opts;
-      status = GetBlockBasedTableOptionsFromMap(bbt_opts, cache_options_map, column_bbt_opts);
+      status = GetBlockBasedTableOptionsFromMap(bbt_opts, cache_options_map, &column_bbt_opts);
       if (!status.ok()) {
-	derr << __func__ << " invalid cache options for CF '"
-	     << column.name << "': " << cache_options_map << dendl;
-	derr << __func__ << "error = '" << status.getState() << "'" << dendl;
+	derr << __func__ << " invalid block cache options; column=" << column.name <<
+	  " options=" << block_cache_opt << dendl;
+	derr << __func__ << " error = '" << status.getState() << "'" << dendl;
 	return -EINVAL;
       }
-         
+      std::shared_ptr<rocksdb::Cache> block_cache;
+      if (column_bbt_opts.no_block_cache) {
+	// clear all settings except no_block_cache
+	// rocksdb does not like then
+	column_bbt_opts = rocksdb::BlockBasedTableOptions();
+	column_bbt_opts.no_block_cache = true;
+      } else {
+	if (require_new_block_cache) {
+	  block_cache = create_block_cache(cache_type, cache_size, high_pri_pool_ratio);
+	  if (!block_cache) {
+	    dout(5) << __func__ << " failed to create block cache for params: " << block_cache_opt << dendl;
+	    return -EINVAL;
+	  }
+	} else {
+	  block_cache = bbt_opts.block_cache;
+	}
+      }
+      column_bbt_opts.block_cache = block_cache;
       cf_bbt_opts[column.name] = column_bbt_opts;
       cf_opt.table_factory.reset(NewBlockBasedTableFactory(cf_bbt_opts[column.name]));
     }
diff --git a/src/kv/RocksDBStore.h b/src/kv/RocksDBStore.h
index d4c0be4149c..969fe5bfe46 100644
--- a/src/kv/RocksDBStore.h
+++ b/src/kv/RocksDBStore.h
@@ -150,8 +150,10 @@ private:
 		      std::vector<std::pair<size_t, RocksDBStore::ColumnFamily> >& existing_cfs_shard,
 		      std::vector<rocksdb::ColumnFamilyDescriptor>& missing_cfs,
 		      std::vector<std::pair<size_t, RocksDBStore::ColumnFamily> >& missing_cfs_shard);
-  int init_block_cache(uint64_t size, rocksdb::BlockBasedTableOptions& bbto);
-
+  std::shared_ptr<rocksdb::Cache> create_block_cache(const std::string& cache_type, size_t cache_size, double cache_prio_high = 0.0);
+  int extract_block_cache_options(const std::string& opts_str,
+				  std::unordered_map<std::string, std::string>* column_opts_map,
+				  std::string* block_cache_opt);
   // manage async compactions
   ceph::mutex compact_queue_lock =
     ceph::make_mutex("RocksDBStore::compact_thread_lock");
@@ -464,7 +466,7 @@ err:
 
   virtual int64_t get_cache_usage(string prefix) const override {
     auto it = cf_bbt_opts.find(prefix);
-    if (it != cf_bbt_opts.end()) {
+    if (it != cf_bbt_opts.end() && it->second.block_cache) {
       return static_cast<int64_t>(it->second.block_cache->GetUsage());
     }
     return -EINVAL;
@@ -476,12 +478,22 @@ err:
     return 0;
   }
 
-  virtual std::shared_ptr<PriorityCache::PriCache> get_priority_cache() 
-      const override {
-    return std::dynamic_pointer_cast<PriorityCache::PriCache>(
+  virtual std::shared_ptr<PriorityCache::PriCache>
+      get_priority_cache() const override {
+    return dynamic_pointer_cast<PriorityCache::PriCache>(
         bbt_opts.block_cache);
   }
 
+  virtual std::shared_ptr<PriorityCache::PriCache>
+      get_priority_cache(string prefix) const override {
+    auto it = cf_bbt_opts.find(prefix);
+    if (it != cf_bbt_opts.end()) {
+      return dynamic_pointer_cast<PriorityCache::PriCache>(
+          it->second.block_cache);
+    }
+    return nullptr;
+  }
+
   WholeSpaceIterator get_wholespace_iterator(IteratorOpts opts = 0) override;
 private:
   WholeSpaceIterator get_default_cf_iterator();
@@ -505,24 +517,6 @@ public:
   };
   int reshard(const std::string& new_sharding, const resharding_ctrl* ctrl = nullptr);
 
-  int set_cache_capacity(int64_t capacity);
-  int64_t get_cache_capacity();
-
-  virtual std::shared_ptr<PriorityCache::PriCache>
-      get_priority_cache() const override {
-    return dynamic_pointer_cast<PriorityCache::PriCache>(
-        bbt_opts.block_cache);
-  }
-
-  virtual std::shared_ptr<PriorityCache::PriCache>
-      get_priority_cache(string prefix) const override {
-    auto it = cf_bbt_opts.find(prefix);
-    if (it != cf_bbt_opts.end()) {
-      return dynamic_pointer_cast<PriorityCache::PriCache>(
-          it->second.block_cache);
-    }
-    return nullptr;
-  }
 
 };
 
-- 
2.39.5