path: root/zenserver/cache/structuredcachestore.h

// Copyright Epic Games, Inc. All Rights Reserved.

#pragma once

#include <zencore/compactbinary.h>
#include <zencore/iobuffer.h>
#include <zencore/iohash.h>
#include <zencore/thread.h>
#include <zencore/uid.h>
#include <zenstore/blockstore.h>
#include <zenstore/caslog.h>
#include <zenstore/gc.h>

ZEN_THIRD_PARTY_INCLUDES_START
#include <tsl/robin_map.h>
ZEN_THIRD_PARTY_INCLUDES_END

#include <atomic>
#include <compare>
#include <filesystem>
#include <unordered_map>

#define ZEN_USE_CACHE_TRACKER 0

namespace zen {

class PathBuilderBase;
class GcManager;
class ZenCacheTracker;
class ScrubContext;

/******************************************************************************

  /$$$$$$$$                        /$$$$$$                   /$$
 |_____ $$                        /$$__  $$                 | $$
	  /$$/  /$$$$$$ /$$$$$$$     | $$  \__/ /$$$$$$  /$$$$$$| $$$$$$$  /$$$$$$
	 /$$/  /$$__  $| $$__  $$    | $$      |____  $$/$$_____| $$__  $$/$$__  $$
	/$$/  | $$$$$$$| $$  \ $$    | $$       /$$$$$$| $$     | $$  \ $| $$$$$$$$
   /$$/   | $$_____| $$  | $$    | $$    $$/$$__  $| $$     | $$  | $| $$_____/
  /$$$$$$$|  $$$$$$| $$  | $$    |  $$$$$$|  $$$$$$|  $$$$$$| $$  | $|  $$$$$$$
 |________/\_______|__/  |__/     \______/ \_______/\_______|__/  |__/\_______/

  Cache store for UE5. Restricts keys to "{bucket}/{hash}" pairs where the hash
  is 40 (hex) chars in size. Values may be opaque blobs or structured objects
  which can in turn contain references to other objects (or blobs).

******************************************************************************/

namespace access_tracking {

	struct KeyAccessTime
	{
		IoHash		  Key;
		GcClock::Tick LastAccess{};
	};

	struct AccessTimes
	{
		std::unordered_map<std::string, std::vector<KeyAccessTime>> Buckets;
	};
};	// namespace access_tracking

struct ZenCacheValue
{
	IoBuffer Value;
	CbObject IndexData;
};

//////////////////////////////////////////////////////////////////////////

#pragma pack(push)
#pragma pack(1)

struct DiskLocation
{
	inline DiskLocation() = default;

	inline DiskLocation(uint64_t ValueSize, uint8_t Flags) : Flags(Flags | kStandaloneFile) { Location.StandaloneSize = ValueSize; }

	inline DiskLocation(const BlockStoreLocation& Location, uint64_t PayloadAlignment, uint8_t Flags) : Flags(Flags & ~kStandaloneFile)
	{
		this->Location.BlockLocation = BlockStoreDiskLocation(Location, PayloadAlignment);
	}

	inline BlockStoreLocation GetBlockLocation(uint64_t PayloadAlignment) const
	{
		ZEN_ASSERT(!(Flags & kStandaloneFile));
		return Location.BlockLocation.Get(PayloadAlignment);
	}

	inline uint64_t		  Size() const { return (Flags & kStandaloneFile) ? Location.StandaloneSize : Location.BlockLocation.GetSize(); }
	inline uint8_t		  IsFlagSet(uint64_t Flag) const { return Flags & Flag; }
	inline uint8_t		  GetFlags() const { return Flags; }
	inline ZenContentType GetContentType() const
	{
		ZenContentType ContentType = ZenContentType::kBinary;

		if (IsFlagSet(kStructured))
		{
			ContentType = ZenContentType::kCbObject;
		}

		if (IsFlagSet(kCompressed))
		{
			ContentType = ZenContentType::kCompressedBinary;
		}

		return ContentType;
	}

	union
	{
		BlockStoreDiskLocation BlockLocation;		// 10 bytes
		uint64_t			   StandaloneSize = 0;	// 8 bytes
	} Location;

	static const uint8_t kStandaloneFile = 0x80u;  // Stored as a separate file
	static const uint8_t kStructured	 = 0x40u;  // Serialized as compact binary
	static const uint8_t kTombStone		 = 0x20u;  // Represents a deleted key/value
	static const uint8_t kCompressed	 = 0x10u;  // Stored in compressed buffer format

	uint8_t Flags	 = 0;
	uint8_t Reserved = 0;
};

struct DiskIndexEntry
{
	IoHash		 Key;		// 20 bytes
	DiskLocation Location;	// 12 bytes
};
#pragma pack(pop)

static_assert(sizeof(DiskIndexEntry) == 32);

/** In-memory cache storage

	Intended for small values which are frequently accessed

	This should have a better memory management policy to maintain reasonable
	footprint.

 */
class ZenCacheMemoryLayer
{
public:
	struct Configuration
	{
		uint64_t TargetFootprintBytes = 16 * 1024 * 1024;
		uint64_t ScavengeThreshold	  = 4 * 1024 * 1024;
	};

	struct BucketInfo
	{
		uint64_t EntryCount = 0;
		uint64_t TotalSize	= 0;
	};

	struct Info
	{
		Configuration			 Config;
		std::vector<std::string> BucketNames;
		uint64_t				 EntryCount = 0;
		uint64_t				 TotalSize	= 0;
	};

	ZenCacheMemoryLayer();
	~ZenCacheMemoryLayer();

	bool	 Get(std::string_view Bucket, const IoHash& HashKey, ZenCacheValue& OutValue);
	void	 Put(std::string_view Bucket, const IoHash& HashKey, const ZenCacheValue& Value);
	void	 Drop();
	bool	 DropBucket(std::string_view Bucket);
	void	 Scrub(ScrubContext& Ctx);
	void	 GatherAccessTimes(zen::access_tracking::AccessTimes& AccessTimes);
	void	 Reset();
	uint64_t TotalSize() const;

	Info					  GetInfo() const;
	std::optional<BucketInfo> GetBucketInfo(std::string_view Bucket) const;

	const Configuration& GetConfiguration() const { return m_Configuration; }
	void				 SetConfiguration(const Configuration& NewConfig) { m_Configuration = NewConfig; }

private:
	struct CacheBucket
	{
		struct BucketValue
		{
			IoBuffer			Payload;
			std::atomic_int64_t LastAccess;

			BucketValue() : Payload(), LastAccess() {}
			BucketValue(IoBuffer Value, const int64_t Timestamp) : Payload(Value), LastAccess(Timestamp) {}
			BucketValue(const BucketValue& V) : Payload(V.Payload), LastAccess(V.LastAccess.load(std::memory_order_relaxed)) {}
			BucketValue(BucketValue&& V) : Payload(std::move(V.Payload)), LastAccess(V.LastAccess.load(std::memory_order_relaxed)) {}

			BucketValue& operator=(const BucketValue& V) { return *this = BucketValue(V); }
			BucketValue& operator=(BucketValue&& V)
			{
				Payload = std::move(V.Payload);
				LastAccess.store(V.LastAccess.load(), std::memory_order_relaxed);
				return *this;
			}
		};

		mutable RwLock						m_BucketLock;
		tsl::robin_map<IoHash, BucketValue> m_CacheMap;
		std::atomic_uint64_t				m_TotalSize{};

		bool			Get(const IoHash& HashKey, ZenCacheValue& OutValue);
		void			Put(const IoHash& HashKey, const ZenCacheValue& Value);
		void			Drop();
		void			Scrub(ScrubContext& Ctx);
		void			GatherAccessTimes(std::vector<zen::access_tracking::KeyAccessTime>& AccessTimes);
		inline uint64_t TotalSize() const { return m_TotalSize; }
		uint64_t		EntryCount() const;
	};

	mutable RwLock												  m_Lock;
	std::unordered_map<std::string, std::unique_ptr<CacheBucket>> m_Buckets;
	std::vector<std::unique_ptr<CacheBucket>>					  m_DroppedBuckets;
	Configuration												  m_Configuration;

	ZenCacheMemoryLayer(const ZenCacheMemoryLayer&) = delete;
	ZenCacheMemoryLayer& operator=(const ZenCacheMemoryLayer&) = delete;
};

class ZenCacheDiskLayer
{
public:
	struct Configuration
	{
		std::filesystem::path RootDir;
	};

	struct BucketInfo
	{
		uint64_t EntryCount = 0;
		uint64_t TotalSize	= 0;
	};

	struct Info
	{
		Configuration			 Config;
		std::vector<std::string> BucketNames;
		uint64_t				 EntryCount = 0;
		uint64_t				 TotalSize	= 0;
	};

	explicit ZenCacheDiskLayer(const std::filesystem::path& RootDir);
	~ZenCacheDiskLayer();

	bool Get(std::string_view Bucket, const IoHash& HashKey, ZenCacheValue& OutValue);
	void Put(std::string_view Bucket, const IoHash& HashKey, const ZenCacheValue& Value);
	bool Drop();
	bool DropBucket(std::string_view Bucket);
	void Flush();
	void Scrub(ScrubContext& Ctx);
	void GatherReferences(GcContext& GcCtx);
	void CollectGarbage(GcContext& GcCtx);
	void UpdateAccessTimes(const zen::access_tracking::AccessTimes& AccessTimes);
	// void IterateBuckets(const std::function<void(std::string_view Bucket)>& Callback) const;

	void	 DiscoverBuckets();
	uint64_t TotalSize() const;

	Info					  GetInfo() const;
	std::optional<BucketInfo> GetBucketInfo(std::string_view Bucket) const;

private:
	/** A cache bucket manages a single directory containing
		metadata and data for that bucket
	  */
	struct CacheBucket
	{
		CacheBucket(std::string BucketName);
		~CacheBucket();

		bool OpenOrCreate(std::filesystem::path BucketDir, bool AllowCreate = true);
		bool Get(const IoHash& HashKey, ZenCacheValue& OutValue);
		void Put(const IoHash& HashKey, const ZenCacheValue& Value);
		bool Drop();
		void Flush();
		void Scrub(ScrubContext& Ctx);
		void GatherReferences(GcContext& GcCtx);
		void CollectGarbage(GcContext& GcCtx);
		void UpdateAccessTimes(const std::vector<zen::access_tracking::KeyAccessTime>& AccessTimes);

		inline uint64_t TotalSize() const { return m_TotalStandaloneSize.load(std::memory_order::relaxed) + m_BlockStore.TotalSize(); }
		uint64_t		EntryCount() const;

	private:
		const uint64_t MaxBlockSize		  = 1ull << 30;
		uint64_t	   m_PayloadAlignment = 1ull << 4;

		std::string			  m_BucketName;
		std::filesystem::path m_BucketDir;
		std::filesystem::path m_BlocksBasePath;
		BlockStore			  m_BlockStore;
		Oid					  m_BucketId;
		uint64_t			  m_LargeObjectThreshold = 128 * 1024;

		// These files are used to manage storage of small objects for this bucket

		TCasLogFile<DiskIndexEntry> m_SlogFile;
		uint64_t					m_LogFlushPosition = 0;

		struct IndexEntry
		{
			DiskLocation		Location;
			std::atomic_int64_t LastAccess;

			IndexEntry() : Location(), LastAccess() {}
			IndexEntry(const DiskLocation& Loc, const int64_t Timestamp) : Location(Loc), LastAccess(Timestamp) {}
			IndexEntry(const IndexEntry& E) : Location(E.Location), LastAccess(E.LastAccess.load(std::memory_order_relaxed)) {}
			IndexEntry(IndexEntry&& E) noexcept : Location(std::move(E.Location)), LastAccess(E.LastAccess.load(std::memory_order_relaxed))
			{
			}

			IndexEntry& operator=(const IndexEntry& E) { return *this = IndexEntry(E); }
			IndexEntry& operator=(IndexEntry&& E) noexcept
			{
				Location = std::move(E.Location);
				LastAccess.store(E.LastAccess.load(), std::memory_order_relaxed);
				return *this;
			}
		};

		using IndexMap = tsl::robin_map<IoHash, IndexEntry, IoHash::Hasher>;

		mutable RwLock m_IndexLock;
		IndexMap	   m_Index;

		std::atomic_uint64_t m_TotalStandaloneSize{};

		void	 BuildPath(PathBuilderBase& Path, const IoHash& HashKey);
		void	 PutStandaloneCacheValue(const IoHash& HashKey, const ZenCacheValue& Value);
		bool	 GetStandaloneCacheValue(const DiskLocation& Loc, const IoHash& HashKey, ZenCacheValue& OutValue);
		void	 PutInlineCacheValue(const IoHash& HashKey, const ZenCacheValue& Value);
		bool	 GetInlineCacheValue(const DiskLocation& Loc, ZenCacheValue& OutValue);
		void	 MakeIndexSnapshot();
		uint64_t ReadIndexFile();
		uint64_t ReadLog(uint64_t LogPosition);
		void	 OpenLog(const std::filesystem::path& BucketDir, const bool IsNew);
		void	 SaveManifest();

		// These locks are here to avoid contention on file creation, therefore it's sufficient
		// that we take the same lock for the same hash
		//
		// These locks are small and should really be spaced out so they don't share cache lines,
		// but we don't currently access them at particularly high frequency so it should not be
		// an issue in practice

		RwLock		   m_ShardedLocks[256];
		inline RwLock& LockForHash(const IoHash& Hash) { return m_ShardedLocks[Hash.Hash[19]]; }
	};

	std::filesystem::path										  m_RootDir;
	mutable RwLock												  m_Lock;
	std::unordered_map<std::string, std::unique_ptr<CacheBucket>> m_Buckets;  // TODO: make this case insensitive
	std::vector<std::unique_ptr<CacheBucket>>					  m_DroppedBuckets;

	ZenCacheDiskLayer(const ZenCacheDiskLayer&) = delete;
	ZenCacheDiskLayer& operator=(const ZenCacheDiskLayer&) = delete;
};

class ZenCacheNamespace final : public RefCounted, public GcStorage, public GcContributor
{
public:
	struct Configuration
	{
		std::filesystem::path RootDir;
		uint64_t			  DiskLayerThreshold = 0;
	};
	struct BucketInfo
	{
		ZenCacheDiskLayer::BucketInfo	DiskLayerInfo;
		ZenCacheMemoryLayer::BucketInfo MemoryLayerInfo;
	};
	struct Info
	{
		Configuration			  Config;
		std::vector<std::string>  BucketNames;
		ZenCacheDiskLayer::Info	  DiskLayerInfo;
		ZenCacheMemoryLayer::Info MemoryLayerInfo;
	};

	ZenCacheNamespace(GcManager& Gc, const std::filesystem::path& RootDir);
	~ZenCacheNamespace();

	bool					  Get(std::string_view Bucket, const IoHash& HashKey, ZenCacheValue& OutValue);
	void					  Put(std::string_view Bucket, const IoHash& HashKey, const ZenCacheValue& Value);
	bool					  Drop();
	bool					  DropBucket(std::string_view Bucket);
	void					  Flush();
	void					  Scrub(ScrubContext& Ctx);
	uint64_t				  DiskLayerThreshold() const { return m_DiskLayerSizeThreshold; }
	virtual void			  GatherReferences(GcContext& GcCtx) override;
	virtual void			  CollectGarbage(GcContext& GcCtx) override;
	virtual GcStorageSize	  StorageSize() const override;
	Info					  GetInfo() const;
	std::optional<BucketInfo> GetBucketInfo(std::string_view Bucket) const;

private:
	std::filesystem::path m_RootDir;
	ZenCacheMemoryLayer	  m_MemLayer;
	ZenCacheDiskLayer	  m_DiskLayer;
	uint64_t			  m_DiskLayerSizeThreshold = 1 * 1024;
	uint64_t			  m_LastScrubTime		   = 0;

#if ZEN_USE_CACHE_TRACKER
	std::unique_ptr<ZenCacheTracker> m_AccessTracker;
#endif

	ZenCacheNamespace(const ZenCacheNamespace&) = delete;
	ZenCacheNamespace& operator=(const ZenCacheNamespace&) = delete;
};

class ZenCacheStore final
{
public:
	static constexpr std::string_view DefaultNamespace =
		"!default!";  // This is intentionally not a valid namespace name and will only be used for mapping when no namespace is given
	static constexpr std::string_view NamespaceDiskPrefix = "ns_";

	struct Configuration
	{
		std::filesystem::path BasePath;
		bool				  AllowAutomaticCreationOfNamespaces = false;
	};

	struct Info
	{
		Configuration			 Config;
		std::vector<std::string> NamespaceNames;
		uint64_t				 DiskEntryCount	  = 0;
		uint64_t				 MemoryEntryCount = 0;
		GcStorageSize			 StorageSize;
	};

	ZenCacheStore(GcManager& Gc, const Configuration& Configuration);
	~ZenCacheStore();

	bool Get(std::string_view Namespace, std::string_view Bucket, const IoHash& HashKey, ZenCacheValue& OutValue);
	void Put(std::string_view Namespace, std::string_view Bucket, const IoHash& HashKey, const ZenCacheValue& Value);
	bool DropBucket(std::string_view Namespace, std::string_view Bucket);
	bool DropNamespace(std::string_view Namespace);
	void Flush();
	void Scrub(ScrubContext& Ctx);

	GcStorageSize StorageSize() const;
	//	const Configuration& GetConfiguration() const { return m_Configuration; }

	Info										 GetInfo() const;
	std::optional<ZenCacheNamespace::Info>		 GetNamespaceInfo(std::string_view Namespace);
	std::optional<ZenCacheNamespace::BucketInfo> GetBucketInfo(std::string_view Namespace, std::string_view Bucket);

private:
	const ZenCacheNamespace* FindNamespace(std::string_view Namespace) const;
	ZenCacheNamespace*		 GetNamespace(std::string_view Namespace);
	void IterateNamespaces(const std::function<void(std::string_view Namespace, ZenCacheNamespace& Store)>& Callback) const;

	typedef std::unordered_map<std::string, std::unique_ptr<ZenCacheNamespace>> NamespaceMap;

	mutable RwLock									m_NamespacesLock;
	NamespaceMap									m_Namespaces;
	std::vector<std::unique_ptr<ZenCacheNamespace>> m_DroppedNamespaces;

	GcManager&	  m_Gc;
	Configuration m_Configuration;
};

void z$_forcelink();

}  // namespace zen