Files
UnrealEngineUWP/Engine/Source/Developer/DerivedDataCache/Private/HttpCacheStore.cpp

5159 lines
164 KiB
C++
Raw Normal View History

// Copyright Epic Games, Inc. All Rights Reserved.
#include "DerivedDataBackendInterface.h"
#if WITH_HTTP_DDC_BACKEND
#if PLATFORM_WINDOWS || PLATFORM_HOLOLENS
#include "Windows/WindowsHWrapper.h"
#include "Windows/AllowWindowsPlatformTypes.h"
#endif
#include "curl/curl.h"
#if PLATFORM_WINDOWS || PLATFORM_HOLOLENS
#include "Windows/HideWindowsPlatformTypes.h"
#endif
#include "Algo/Accumulate.h"
#include "Algo/Find.h"
#include "Algo/Transform.h"
#include "Compression/CompressedBuffer.h"
#include "Containers/DepletableMpscQueue.h"
#include "Containers/StaticArray.h"
#include "Containers/StringView.h"
#include "Containers/Ticker.h"
#include "DerivedDataCacheKey.h"
#include "DerivedDataCachePrivate.h"
#include "DerivedDataCacheRecord.h"
#include "DerivedDataCacheUsageStats.h"
#include "DerivedDataChunk.h"
#include "DerivedDataRequest.h"
#include "DerivedDataRequestOwner.h"
#include "DerivedDataValue.h"
#include "Dom/JsonObject.h"
#include "Experimental/Async/LazyEvent.h"
#include "Experimental/Containers/FAAArrayQueue.h"
#include "GenericPlatform/GenericPlatformFile.h"
#include "HAL/IConsoleManager.h"
#include "HAL/PlatformFileManager.h"
#include "HAL/Runnable.h"
#include "HAL/RunnableThread.h"
#include "IO/IoHash.h"
#include "Memory/SharedBuffer.h"
#include "Misc/FileHelper.h"
#include "Misc/Optional.h"
#include "Misc/ScopeLock.h"
#include "Misc/SecureHash.h"
#include "Misc/StringBuilder.h"
#include "Policies/CondensedJsonPrintPolicy.h"
#include "ProfilingDebugging/CountersTrace.h"
#include "ProfilingDebugging/CpuProfilerTrace.h"
#include "Serialization/BufferArchive.h"
#include "Serialization/CompactBinary.h"
#include "Serialization/CompactBinaryPackage.h"
#include "Serialization/CompactBinaryValidation.h"
#include "Serialization/CompactBinaryWriter.h"
#include "Serialization/JsonReader.h"
#include "Serialization/JsonSerializer.h"
#include "Serialization/JsonWriter.h"
#include "Tasks/Task.h"
#if WITH_SSL
#include "Ssl.h"
#include <openssl/ssl.h>
#endif
// Enables data request helpers that internally
// batch requests to reduce the number of concurrent
// connections.
#ifndef WITH_DATAREQUEST_HELPER
#define WITH_DATAREQUEST_HELPER 1
#endif
#define UE_HTTPDDC_BACKEND_WAIT_INTERVAL 0.01f
#define UE_HTTPDDC_BACKEND_WAIT_INTERVAL_MS ((uint32)(UE_HTTPDDC_BACKEND_WAIT_INTERVAL*1000))
#define UE_HTTPDDC_HTTP_REQUEST_TIMEOUT_SECONDS 30L
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
#define UE_HTTPDDC_HTTP_REQUEST_TIMEOUT_ENABLED 1
#define UE_HTTPDDC_HTTP_DEBUG 0
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
#define UE_HTTPDDC_GET_REQUEST_POOL_SIZE 48
#define UE_HTTPDDC_PUT_REQUEST_POOL_SIZE 16
#define UE_HTTPDDC_NONBLOCKING_REQUEST_POOL_SIZE 128
#define UE_HTTPDDC_MAX_FAILED_LOGIN_ATTEMPTS 16
#define UE_HTTPDDC_MAX_ATTEMPTS 4
#define UE_HTTPDDC_MAX_BUFFER_RESERVE 104857600u
#define UE_HTTPDDC_BATCH_SIZE 12
#define UE_HTTPDDC_BATCH_NUM 64
#define UE_HTTPDDC_BATCH_GET_WEIGHT 4
#define UE_HTTPDDC_BATCH_HEAD_WEIGHT 1
#define UE_HTTPDDC_BATCH_WEIGHT_HINT 12
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
namespace UE::DerivedData
{
TRACE_DECLARE_INT_COUNTER(HttpDDC_Exist, TEXT("HttpDDC Exist"));
TRACE_DECLARE_INT_COUNTER(HttpDDC_ExistHit, TEXT("HttpDDC Exist Hit"));
TRACE_DECLARE_INT_COUNTER(HttpDDC_Get, TEXT("HttpDDC Get"));
TRACE_DECLARE_INT_COUNTER(HttpDDC_GetHit, TEXT("HttpDDC Get Hit"));
TRACE_DECLARE_INT_COUNTER(HttpDDC_Put, TEXT("HttpDDC Put"));
TRACE_DECLARE_INT_COUNTER(HttpDDC_PutHit, TEXT("HttpDDC Put Hit"));
TRACE_DECLARE_INT_COUNTER(HttpDDC_BytesReceived, TEXT("HttpDDC Bytes Received"));
TRACE_DECLARE_INT_COUNTER(HttpDDC_BytesSent, TEXT("HttpDDC Bytes Sent"));
static bool bHttpEnableAsync = true;
static FAutoConsoleVariableRef CVarHttpEnableAsync(
TEXT("DDC.Http.EnableAsync"),
bHttpEnableAsync,
TEXT("If true, async operations are permitted, otherwise all operations are forced to be synchronous."),
ECVF_Default);
static CURLcode sslctx_function(CURL * curl, void * sslctx, void * parm);
typedef TSharedPtr<class IHttpRequest> FHttpRequestPtr;
typedef TSharedPtr<class IHttpResponse, ESPMode::ThreadSafe> FHttpResponsePtr;
/**
* Encapsulation for access token shared by all requests.
*/
struct FHttpAccessToken
{
public:
FHttpAccessToken() = default;
FString GetHeader();
void SetHeader(const TCHAR*);
uint32 GetSerial() const;
private:
FRWLock Lock;
FString Token;
uint32 Serial;
};
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
struct FHttpSharedData
{
FHttpSharedData()
: PendingRequestEvent(EEventMode::AutoReset)
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
{
FMemory::Memset(WriteLocked, 0, sizeof(WriteLocked));
CurlShare = curl_share_init();
curl_share_setopt(CurlShare, CURLSHOPT_USERDATA, this);
curl_share_setopt(CurlShare, CURLSHOPT_LOCKFUNC, LockFn);
curl_share_setopt(CurlShare, CURLSHOPT_UNLOCKFUNC, UnlockFn);
curl_share_setopt(CurlShare, CURLSHOPT_SHARE, CURL_LOCK_DATA_DNS);
curl_share_setopt(CurlShare, CURLSHOPT_SHARE, CURL_LOCK_DATA_SSL_SESSION);
CurlMulti = curl_multi_init();
curl_multi_setopt(CurlMulti, CURLMOPT_PIPELINING, CURLPIPE_MULTIPLEX);
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
}
~FHttpSharedData()
{
curl_multi_cleanup(CurlMulti);
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
curl_share_cleanup(CurlShare);
}
CURLSH* CurlShare;
CURLM* CurlMulti;
TDepletableMpscQueue<CURL*> PendingRequestAdditions;
FLazyEvent PendingRequestEvent;
FRunnableThread* AsyncServiceThread = nullptr;
TUniquePtr<class FHttpCacheStoreRunnable> AsyncRunnable;
std::atomic<bool> bAsyncThreadStarting = false;
static inline std::atomic<bool> bAsyncThreadShutdownRequested = false;
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
private:
FRWLock Locks[CURL_LOCK_DATA_LAST];
bool WriteLocked[CURL_LOCK_DATA_LAST];
static void LockFn(CURL* Handle, curl_lock_data Data, curl_lock_access Access, void* User)
{
FHttpSharedData* SharedData = (FHttpSharedData*)User;
if (Access == CURL_LOCK_ACCESS_SHARED)
{
SharedData->Locks[Data].ReadLock();
}
else
{
SharedData->Locks[Data].WriteLock();
SharedData->WriteLocked[Data] = true;
}
}
static void UnlockFn(CURL* Handle, curl_lock_data Data, void* User)
{
FHttpSharedData* SharedData = (FHttpSharedData*)User;
if (!SharedData->WriteLocked[Data])
{
SharedData->Locks[Data].ReadUnlock();
}
else
{
SharedData->WriteLocked[Data] = false;
SharedData->Locks[Data].WriteUnlock();
}
}
};
template <typename T>
class TRefCountedUniqueFunction final : public FThreadSafeRefCountedObject
{
public:
explicit TRefCountedUniqueFunction(T&& InFunction) : Function(MoveTemp(InFunction))
{
}
const T& GetFunction() const { return Function; }
private:
T Function;
};
class FHttpCacheStoreRunnable final : public FRunnable
{
public:
explicit FHttpCacheStoreRunnable(FHttpSharedData& InSharedData) : SharedData(InSharedData)
{
}
virtual uint32 Run() override final;
protected:
FHttpSharedData& SharedData;
};
/**
* Minimal HTTP request type wrapping CURL without the need for managers. This request
* is written to allow reuse of request objects, in order to allow connections to be reused.
*
* CURL has a global library initialization (curl_global_init). We rely on this happening in
* the Online/HTTP library which is a dependency on this module.
*/
class FHttpRequest
{
public:
/**
* Supported request verbs
*/
enum RequestVerb
{
Get,
Put,
PutCompactBinary,
PutCompressedBlob,
Post,
PostCompactBinary,
PostJson,
Delete,
Head
};
/**
* Convenience result type interpreted from HTTP response code.
*/
enum Result
{
Success,
Failed,
FailedTimeout
};
enum class ECompletionBehavior : uint8
{
Done,
Retry
};
using FOnHttpRequestComplete = TUniqueFunction<ECompletionBehavior(Result HttpResult, FHttpRequest* Request)>;
FHttpRequest(const TCHAR* InDomain, const TCHAR* InEffectiveDomain, FHttpAccessToken* InAuthorizationToken, bool bInLogErrors)
: AsyncData(nullptr)
, bLogErrors(bInLogErrors)
, Domain(InDomain)
, EffectiveDomain(InEffectiveDomain)
, AuthorizationToken(InAuthorizationToken)
{
Curl = curl_easy_init();
Reset();
}
~FHttpRequest()
{
curl_easy_cleanup(Curl);
check(!AsyncData);
}
/**
* Resets all options on the request except those that should always be set.
*/
void Reset()
{
Headers.Reset();
ResponseHeader.Reset();
ResponseBuffer.Reset();
ResponseCode = 0;
ReadDataView = FMemoryView();
WriteDataBufferPtr = nullptr;
WriteHeaderBufferPtr = nullptr;
BytesSent = 0;
BytesReceived = 0;
Attempts = 0;
CurlResult = CURL_LAST;
static FHttpSharedData StaticSharedData;
SharedData = &StaticSharedData;
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
curl_easy_reset(Curl);
check(!AsyncData);
// Options that are always set for all connections.
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
#if UE_HTTPDDC_HTTP_REQUEST_TIMEOUT_ENABLED
curl_easy_setopt(Curl, CURLOPT_CONNECTTIMEOUT, UE_HTTPDDC_HTTP_REQUEST_TIMEOUT_SECONDS);
#endif
curl_easy_setopt(Curl, CURLOPT_FOLLOWLOCATION, 1L);
curl_easy_setopt(Curl, CURLOPT_NOSIGNAL, 1L);
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
curl_easy_setopt(Curl, CURLOPT_DNS_CACHE_TIMEOUT, 300L); // Don't re-resolve every minute
curl_easy_setopt(Curl, CURLOPT_SHARE, SharedData->CurlShare);
// SSL options
curl_easy_setopt(Curl, CURLOPT_USE_SSL, CURLUSESSL_ALL);
curl_easy_setopt(Curl, CURLOPT_SSL_VERIFYPEER, 1);
curl_easy_setopt(Curl, CURLOPT_SSL_VERIFYHOST, 1);
curl_easy_setopt(Curl, CURLOPT_SSLCERTTYPE, "PEM");
// Response functions
curl_easy_setopt(Curl, CURLOPT_HEADERDATA, this);
curl_easy_setopt(Curl, CURLOPT_HEADERFUNCTION, &FHttpRequest::StaticWriteHeaderFn);
curl_easy_setopt(Curl, CURLOPT_WRITEDATA, this);
curl_easy_setopt(Curl, CURLOPT_WRITEFUNCTION, StaticWriteBodyFn);
// SSL certification verification
curl_easy_setopt(Curl, CURLOPT_CAINFO, nullptr);
curl_easy_setopt(Curl, CURLOPT_SSL_CTX_FUNCTION, *sslctx_function);
curl_easy_setopt(Curl, CURLOPT_SSL_CTX_DATA, this);
// Allow compressed data
curl_easy_setopt(Curl, CURLOPT_ACCEPT_ENCODING, "gzip");
// Rewind method, handle special error case where request need to rewind data stream
curl_easy_setopt(Curl, CURLOPT_SEEKFUNCTION, StaticSeekFn);
curl_easy_setopt(Curl, CURLOPT_SEEKDATA, this);
// Set minimum speed behavior to allow operations to abort if the transfer speed is poor for the given duration (1kbps over a 30 second span)
curl_easy_setopt(Curl, CURLOPT_LOW_SPEED_TIME, 30L);
curl_easy_setopt(Curl, CURLOPT_LOW_SPEED_LIMIT, 1024L);
// Debug hooks
#if UE_HTTPDDC_HTTP_DEBUG
curl_easy_setopt(Curl, CURLOPT_DEBUGDATA, this);
curl_easy_setopt(Curl, CURLOPT_DEBUGFUNCTION, StaticDebugCallback);
curl_easy_setopt(Curl, CURLOPT_VERBOSE, 1L);
#endif
curl_easy_setopt(Curl, CURLOPT_PRIVATE, this);
curl_easy_setopt(Curl, CURLOPT_HTTP_VERSION, CURL_HTTP_VERSION_2_0);
}
void PrepareToRetry()
{
ResponseHeader.Reset();
ResponseBuffer.Reset();
ResponseCode = 0;
BytesSent = 0;
BytesReceived = 0;
CurlResult = CURL_LAST;
++Attempts;
}
/** Gets the domain name for this request */
const FString& GetName() const
{
return Domain;
}
/** Gets the domain name for this request */
const FString& GetDomain() const
{
return Domain;
}
/** Gets the effective domain name for this request */
const FString& GetEffectiveDomain() const
{
return EffectiveDomain;
}
/** Returns the HTTP response code.*/
const int64 GetResponseCode() const
{
return ResponseCode;
}
/** Returns the number of bytes received this request (headers withstanding). */
const size_t GetBytesReceived() const
{
return BytesReceived;
}
/** Returns the number of attempts we've made issuing this request (currently tracked for async requests only). */
const size_t GetAttempts() const
{
return Attempts;
}
/** Returns the number of bytes sent during this request (headers withstanding). */
const size_t GetBytesSent() const
{
return BytesSent;
}
/**
* Upload buffer using the request, using either "Put" or "Post" verbs.
* @param Uri Url to use.
* @param Buffer Data to upload
* @return Result of the request
*/
template<RequestVerb V>
Result PerformBlockingUpload(const TCHAR* Uri, TArrayView<const uint8> Buffer, TConstArrayView<long> ExpectedErrorCodes = {})
{
static_assert(V == Put || V == PutCompactBinary || V == PutCompressedBlob || V == Post || V == PostCompactBinary || V == PostJson, "Upload should use either Put or Post verbs.");
uint64 ContentLength = 0u;
if constexpr (V == Put || V == PutCompactBinary || V == PutCompressedBlob)
{
curl_easy_setopt(Curl, CURLOPT_UPLOAD, 1L);
curl_easy_setopt(Curl, CURLOPT_INFILESIZE, Buffer.Num());
curl_easy_setopt(Curl, CURLOPT_READDATA, this);
curl_easy_setopt(Curl, CURLOPT_READFUNCTION, StaticReadFn);
if constexpr (V == PutCompactBinary)
{
Headers.Add(FString(TEXT("Content-Type: application/x-ue-cb")));
}
else if constexpr (V == PutCompressedBlob)
{
Headers.Add(FString(TEXT("Content-Type: application/x-ue-comp")));
}
else
{
Headers.Add(FString(TEXT("Content-Type: application/octet-stream")));
}
ContentLength = Buffer.Num();
ReadDataView = MakeMemoryView(Buffer);
}
else if constexpr (V == Post || V == PostCompactBinary || V == PostJson)
{
curl_easy_setopt(Curl, CURLOPT_POST, 1L);
curl_easy_setopt(Curl, CURLOPT_INFILESIZE, Buffer.Num());
curl_easy_setopt(Curl, CURLOPT_READDATA, this);
curl_easy_setopt(Curl, CURLOPT_READFUNCTION, StaticReadFn);
if constexpr (V == PostCompactBinary)
{
Headers.Add(FString(TEXT("Content-Type: application/x-ue-cb")));
}
else if constexpr (V == PostJson)
{
Headers.Add(FString(TEXT("Content-Type: application/json")));
}
else
{
Headers.Add(FString(TEXT("Content-Type: application/x-www-form-urlencoded")));
}
ContentLength = Buffer.Num();
ReadDataView = MakeMemoryView(Buffer);
}
return PerformBlocking(Uri, V, ContentLength, ExpectedErrorCodes);
}
template<RequestVerb V>
void EnqueueAsyncUpload(IRequestOwner& Owner, struct FRequestPool* Pool, const TCHAR* Uri, FSharedBuffer Buffer, FOnHttpRequestComplete&& OnComplete, TConstArrayView<long> ExpectedErrorCodes = {})
{
static_assert(V == Put || V == PutCompactBinary || V == PutCompressedBlob || V == Post || V == PostCompactBinary || V == PostJson, "Upload should use either Put or Post verbs.");
uint64 ContentLength = 0u;
if constexpr (V == Put || V == PutCompactBinary || V == PutCompressedBlob)
{
curl_easy_setopt(Curl, CURLOPT_UPLOAD, 1L);
curl_easy_setopt(Curl, CURLOPT_INFILESIZE, Buffer.GetSize());
curl_easy_setopt(Curl, CURLOPT_READDATA, this);
curl_easy_setopt(Curl, CURLOPT_READFUNCTION, StaticReadFn);
if constexpr (V == PutCompactBinary)
{
Headers.Add(FString(TEXT("Content-Type: application/x-ue-cb")));
}
else if constexpr (V == PutCompressedBlob)
{
Headers.Add(FString(TEXT("Content-Type: application/x-ue-comp")));
}
else
{
Headers.Add(FString(TEXT("Content-Type: application/octet-stream")));
}
ReadSharedBuffer = Buffer;
ContentLength = Buffer.GetSize();
ReadDataView = Buffer.GetView();
}
else if constexpr (V == Post || V == PostCompactBinary || V == PostJson)
{
curl_easy_setopt(Curl, CURLOPT_POST, 1L);
curl_easy_setopt(Curl, CURLOPT_INFILESIZE, Buffer.GetSize());
curl_easy_setopt(Curl, CURLOPT_READDATA, this);
curl_easy_setopt(Curl, CURLOPT_READFUNCTION, StaticReadFn);
if constexpr (V == PostCompactBinary)
{
Headers.Add(FString(TEXT("Content-Type: application/x-ue-cb")));
}
else if constexpr (V == PostJson)
{
Headers.Add(FString(TEXT("Content-Type: application/json")));
}
else
{
Headers.Add(FString(TEXT("Content-Type: application/x-www-form-urlencoded")));
}
ReadSharedBuffer = Buffer;
ContentLength = Buffer.GetSize();
ReadDataView = Buffer.GetView();
}
return EnqueueAsync(Owner, Pool, Uri, V, ContentLength, MoveTemp(OnComplete), ExpectedErrorCodes);
}
/**
* Download an url into a buffer using the request.
* @param Uri Url to use.
* @param Buffer Optional buffer where data should be downloaded to. If empty downloaded data will
* be stored in an internal buffer and accessed GetResponse* methods.
* @return Result of the request
*/
Result PerformBlockingDownload(const TCHAR* Uri, TArray<uint8>* Buffer, TConstArrayView<long> ExpectedErrorCodes = {400})
{
curl_easy_setopt(Curl, CURLOPT_HTTPGET, 1L);
WriteDataBufferPtr = Buffer;
return PerformBlocking(Uri, Get, 0u, ExpectedErrorCodes);
}
void EnqueueAsyncDownload(IRequestOwner& Owner, struct FRequestPool* Pool, const TCHAR* Uri, FOnHttpRequestComplete&& OnComplete, TConstArrayView<long> ExpectedErrorCodes = {400})
{
curl_easy_setopt(Curl, CURLOPT_HTTPGET, 1L);
return EnqueueAsync(Owner, Pool, Uri, Get, 0u, MoveTemp(OnComplete), ExpectedErrorCodes);
}
/**
* Query an url using the request. Queries can use either "Head" or "Delete" verbs.
* @param Uri Url to use.
* @return Result of the request
*/
template<RequestVerb V>
Result PerformBlockingQuery(const TCHAR* Uri, TConstArrayView<long> ExpectedErrorCodes = {400})
{
static_assert(V == Head || V == Delete, "Queries should use either Head or Delete verbs.");
if (V == Delete)
{
curl_easy_setopt(Curl, CURLOPT_CUSTOMREQUEST, "DELETE");
}
else if (V == Head)
{
curl_easy_setopt(Curl, CURLOPT_NOBODY, 1L);
}
return PerformBlocking(Uri, V, 0u, ExpectedErrorCodes);
}
template<RequestVerb V>
void EnqueueAsyncQuery(const TCHAR* Uri, FOnHttpRequestComplete&& OnComplete, TConstArrayView<long> ExpectedErrorCodes = {400})
{
static_assert(V == Head || V == Delete, "Queries should use either Head or Delete verbs.");
if (V == Delete)
{
curl_easy_setopt(Curl, CURLOPT_CUSTOMREQUEST, "DELETE");
}
else if (V == Head)
{
curl_easy_setopt(Curl, CURLOPT_NOBODY, 1L);
}
return EnqueueAsync(Uri, V, 0u, MoveTemp(OnComplete), ExpectedErrorCodes);
}
/**
* Set a header to send with the request.
*/
void SetHeader(const TCHAR* Header, const TCHAR* Value)
{
check(CurlResult == CURL_LAST); // Cannot set header after request is sent
Headers.Add(FString::Printf(TEXT("%s: %s"), Header, Value));
}
/**
* Attempts to find the header from the response. Returns false if header is not present.
*/
bool GetHeader(const ANSICHAR* Header, FString& OutValue) const
{
check(CurlResult != CURL_LAST); // Cannot query headers before request is sent
const ANSICHAR* HeadersBuffer = (const ANSICHAR*) ResponseHeader.GetData();
size_t HeaderLen = strlen(Header);
// Find the header key in the (ANSI) response buffer. If not found we can exist immediately
if (const ANSICHAR* Found = strstr(HeadersBuffer, Header))
{
const ANSICHAR* Linebreak = strchr(Found, '\r');
const ANSICHAR* ValueStart = Found + HeaderLen + 2; //colon and space
const size_t ValueSize = Linebreak - ValueStart;
FUTF8ToTCHAR TCHARData(ValueStart, ValueSize);
OutValue = FString(TCHARData.Length(), TCHARData.Get());
return true;
}
return false;
}
/**
* Returns the response buffer. Note that is the request is performed
* with an external buffer as target buffer this string will be empty.
*/
const TArray<uint8>& GetResponseBuffer() const
{
return ResponseBuffer;
}
FSharedBuffer MoveResponseBufferToShared()
{
return MakeSharedBufferFromArray(MoveTemp(ResponseBuffer));
}
/**
* Returns the response buffer as a string. Note that is the request is performed
* with an external buffer as target buffer this string will be empty.
*/
FString GetResponseAsString() const
{
return GetAnsiBufferAsString(ResponseBuffer);
}
/**
* Returns the response header as a string.
*/
FString GetResponseHeaderAsString()
{
return GetAnsiBufferAsString(ResponseHeader);
}
/**
* Tries to parse the response buffer as a JsonObject. Return empty pointer if
* parse error occurs.
*/
TSharedPtr<FJsonObject> GetResponseAsJsonObject() const
{
FString Response = GetAnsiBufferAsString(ResponseBuffer);
TSharedPtr<FJsonObject> JsonObject;
TSharedRef<TJsonReader<> > JsonReader = TJsonReaderFactory<>::Create(Response);
if (!FJsonSerializer::Deserialize(JsonReader, JsonObject) || !JsonObject.IsValid())
{
return TSharedPtr<FJsonObject>(nullptr);
}
return JsonObject;
}
/**
* Tries to parse the response buffer as a JsonArray. Return empty array if
* parse error occurs.
*/
TArray<TSharedPtr<FJsonValue>> GetResponseAsJsonArray() const
{
FString Response = GetAnsiBufferAsString(ResponseBuffer);
TArray<TSharedPtr<FJsonValue>> JsonArray;
TSharedRef<TJsonReader<>> JsonReader = TJsonReaderFactory<>::Create(Response);
FJsonSerializer::Deserialize(JsonReader, JsonArray);
return JsonArray;
}
void CompleteAsync(CURLcode Result);
/** Will return true if the response code is considered a success */
static bool IsSuccessResponse(long ResponseCode)
{
// We consider anything in the 1XX or 2XX range a success
return ResponseCode >= 100 && ResponseCode < 300;
}
static bool AllowAsync();
private:
struct FAsyncRequestData final : public FRequestBase
{
IRequestOwner* Owner = nullptr;
FRequestPool* Pool = nullptr;
curl_slist* CurlHeaders = nullptr;
FString Uri;
RequestVerb Verb;
TArray<long, TInlineAllocator<4>> ExpectedErrorCodes;
FOnHttpRequestComplete OnComplete;
FLazyEvent Event {EEventMode::ManualReset};
void Reset()
{
if (CurlHeaders)
{
curl_slist_free_all(CurlHeaders);
CurlHeaders = nullptr;
}
Uri.Empty();
ExpectedErrorCodes.Empty();
}
void SetPriority(EPriority Priority) final {}
void Cancel() final
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_Cancel);
Event.Wait();
}
void Wait() final
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_Wait);
Event.Wait();
}
};
CURL* Curl;
CURLcode CurlResult;
FHttpSharedData* SharedData;
FAsyncRequestData* AsyncData;
long ResponseCode;
size_t BytesSent;
size_t BytesReceived;
size_t Attempts;
bool bLogErrors;
FSharedBuffer ReadSharedBuffer;
FMemoryView ReadDataView;
TArray<uint8>* WriteDataBufferPtr;
TArray<uint8>* WriteHeaderBufferPtr;
TArray<uint8> ResponseHeader;
TArray<uint8> ResponseBuffer;
TArray<FString> Headers;
FString Domain;
FString EffectiveDomain;
FHttpAccessToken* AuthorizationToken;
curl_slist* PrepareToIssueRequest(const TCHAR* Uri, uint64 ContentLength)
{
static const char* CommonHeaders[] = {
"User-Agent: Unreal Engine",
nullptr
};
// Setup request options
FString Url = FString::Printf(TEXT("%s/%s"), *EffectiveDomain, Uri);
curl_easy_setopt(Curl, CURLOPT_URL, TCHAR_TO_ANSI(*Url));
// Setup response header buffer. If caller has not setup a response data buffer, use internal.
WriteHeaderBufferPtr = &ResponseHeader;
if (WriteDataBufferPtr == nullptr)
{
WriteDataBufferPtr = &ResponseBuffer;
}
// Content-Length should always be set
Headers.Add(FString::Printf(TEXT("Content-Length: %d"), ContentLength));
// And auth token if it's set
if (AuthorizationToken)
{
Headers.Add(AuthorizationToken->GetHeader());
}
// Build headers list
curl_slist* CurlHeaders = nullptr;
// Add common headers
for (uint8 i = 0; CommonHeaders[i] != nullptr; ++i)
{
CurlHeaders = curl_slist_append(CurlHeaders, CommonHeaders[i]);
}
// Setup added headers
for (const FString& Header : Headers)
{
CurlHeaders = curl_slist_append(CurlHeaders, TCHAR_TO_ANSI(*Header));
}
curl_easy_setopt(Curl, CURLOPT_HTTPHEADER, CurlHeaders);
return CurlHeaders;
}
/**
* Performs the request, blocking until finished.
* @param Uri Address on the domain to query
* @param Verb HTTP verb to use
* @param ContentLength The number of bytes being uploaded for the body of this request.
* @param ExpectedErrorCodes An array of expected return codes outside of the success range that should NOT be logged as an abnormal/exceptional outcome.
* If unset the response body will be stored in the request.
*/
Result PerformBlocking(const TCHAR* Uri, RequestVerb Verb, uint64 ContentLength, TConstArrayView<long> ExpectedErrorCodes)
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_CurlPerform);
// Build headers list
curl_slist* CurlHeaders = PrepareToIssueRequest(Uri, ContentLength);
// Shots fired!
CurlResult = curl_easy_perform(Curl);
// Get response code
curl_easy_getinfo(Curl, CURLINFO_RESPONSE_CODE, &ResponseCode);
LogResult(CurlResult, Uri, Verb, ExpectedErrorCodes);
// Clean up
curl_slist_free_all(CurlHeaders);
return CurlResult == CURLE_OK ? Success : Failed;
}
void EnqueueAsync(IRequestOwner& Owner, FRequestPool* Pool, const TCHAR* Uri, RequestVerb Verb, uint64 ContentLength, FOnHttpRequestComplete&& OnComplete, TConstArrayView<long> ExpectedErrorCodes);
void LogResult(CURLcode Result, const TCHAR* Uri, RequestVerb Verb, TConstArrayView<long> ExpectedErrorCodes) const
{
if (Result == CURLE_OK)
{
bool bSuccess = false;
const TCHAR* VerbStr = nullptr;
FString AdditionalInfo;
switch (Verb)
{
case Head:
bSuccess = (ExpectedErrorCodes.Contains(ResponseCode) || IsSuccessResponse(ResponseCode));
VerbStr = TEXT("querying");
break;
case Get:
bSuccess = (ExpectedErrorCodes.Contains(ResponseCode) || IsSuccessResponse(ResponseCode));
VerbStr = TEXT("fetching");
AdditionalInfo = FString::Printf(TEXT("Received: %d bytes."), BytesReceived);
break;
case Put:
case PutCompactBinary:
case PutCompressedBlob:
bSuccess = (ExpectedErrorCodes.Contains(ResponseCode) || IsSuccessResponse(ResponseCode));
VerbStr = TEXT("updating");
AdditionalInfo = FString::Printf(TEXT("Sent: %d bytes."), BytesSent);
break;
case Post:
case PostCompactBinary:
case PostJson:
bSuccess = (ExpectedErrorCodes.Contains(ResponseCode) || IsSuccessResponse(ResponseCode));
VerbStr = TEXT("posting");
break;
case Delete:
bSuccess = (ExpectedErrorCodes.Contains(ResponseCode) || IsSuccessResponse(ResponseCode));
VerbStr = TEXT("deleting");
break;
}
if (bSuccess)
{
UE_LOG(
LogDerivedDataCache,
Verbose,
TEXT("%s: Finished %s HTTP cache entry (response %d) from %s. %s"),
*GetName(),
VerbStr,
ResponseCode,
Uri,
*AdditionalInfo
);
}
else if(bLogErrors)
{
// Print the response body if we got one, otherwise print header.
FString Response = GetAnsiBufferAsString(ResponseBuffer.Num() > 0 ? ResponseBuffer : ResponseHeader);
Response.ReplaceCharInline(TEXT('\n'), TEXT(' '));
Response.ReplaceCharInline(TEXT('\r'), TEXT(' '));
// Dont log access denied as error, since tokens can expire mid session
if (ResponseCode == 401)
{
UE_LOG(
LogDerivedDataCache,
Verbose,
TEXT("%s: Failed %s HTTP cache entry (response %d) from %s. Response: %s"),
*GetName(),
VerbStr,
ResponseCode,
Uri,
*Response
);
}
else
{
UE_LOG(
LogDerivedDataCache,
Display,
TEXT("%s: Failed %s HTTP cache entry (response %d) from %s. Response: %s"),
*GetName(),
VerbStr,
ResponseCode,
Uri,
*Response
);
}
}
}
else if(bLogErrors)
{
UE_LOG(
LogDerivedDataCache,
Display,
TEXT("%s: Error while connecting to %s: %s"),
*GetName(),
*EffectiveDomain,
ANSI_TO_TCHAR(curl_easy_strerror(Result))
);
}
}
FString GetAnsiBufferAsString(const TArray<uint8>& Buffer) const
{
// Content is NOT null-terminated; we need to specify lengths here
FUTF8ToTCHAR TCHARData(reinterpret_cast<const ANSICHAR*>(Buffer.GetData()), Buffer.Num());
return FString(TCHARData.Length(), TCHARData.Get());
}
static size_t StaticDebugCallback(CURL * Handle, curl_infotype DebugInfoType, char * DebugInfo, size_t DebugInfoSize, void* UserData)
{
FHttpRequest* Request = static_cast<FHttpRequest*>(UserData);
switch (DebugInfoType)
{
case CURLINFO_TEXT:
{
// Truncate at 1023 characters. This is just an arbitrary number based on a buffer size seen in
// the libcurl code.
DebugInfoSize = FMath::Min(DebugInfoSize, (size_t)1023);
// Calculate the actual length of the string due to incorrect use of snprintf() in lib/vtls/openssl.c.
char* FoundNulPtr = (char*)memchr(DebugInfo, 0, DebugInfoSize);
int CalculatedSize = FoundNulPtr != nullptr ? FoundNulPtr - DebugInfo : DebugInfoSize;
auto ConvertedString = StringCast<TCHAR>(static_cast<const ANSICHAR*>(DebugInfo), CalculatedSize);
FString DebugText(ConvertedString.Length(), ConvertedString.Get());
DebugText.ReplaceInline(TEXT("\n"), TEXT(""), ESearchCase::CaseSensitive);
DebugText.ReplaceInline(TEXT("\r"), TEXT(""), ESearchCase::CaseSensitive);
UE_LOG(LogDerivedDataCache, VeryVerbose, TEXT("%s: %p: '%s'"), *Request->GetName(), Request, *DebugText);
}
break;
case CURLINFO_HEADER_IN:
UE_LOG(LogDerivedDataCache, VeryVerbose, TEXT("%s: %p: Received header (%d bytes)"), *Request->GetName(), Request, DebugInfoSize);
break;
case CURLINFO_DATA_IN:
UE_LOG(LogDerivedDataCache, VeryVerbose, TEXT("%s: %p: Received data (%d bytes)"), *Request->GetName(), Request, DebugInfoSize);
break;
case CURLINFO_DATA_OUT:
UE_LOG(LogDerivedDataCache, VeryVerbose, TEXT("%s: %p: Sent data (%d bytes)"), *Request->GetName(), Request, DebugInfoSize);
break;
case CURLINFO_SSL_DATA_IN:
UE_LOG(LogDerivedDataCache, VeryVerbose, TEXT("%s: %p: Received SSL data (%d bytes)"), *Request->GetName(), Request, DebugInfoSize);
break;
case CURLINFO_SSL_DATA_OUT:
UE_LOG(LogDerivedDataCache, VeryVerbose, TEXT("%s: %p: Sent SSL data (%d bytes)"), *Request->GetName(), Request, DebugInfoSize);
break;
}
return 0;
}
static size_t StaticReadFn(void* Ptr, size_t SizeInBlocks, size_t BlockSizeInBytes, void* UserData)
{
FHttpRequest* Request = static_cast<FHttpRequest*>(UserData);
const size_t MaxReadSize = SizeInBlocks * BlockSizeInBytes;
const FMemoryView SourceView = Request->ReadDataView.Mid(Request->BytesSent, MaxReadSize);
MakeMemoryView(Ptr, MaxReadSize).CopyFrom(SourceView);
Request->BytesSent += SourceView.GetSize();
return SourceView.GetSize();
}
static size_t StaticWriteHeaderFn(void* Ptr, size_t SizeInBlocks, size_t BlockSizeInBytes, void* UserData)
{
FHttpRequest* Request = static_cast<FHttpRequest*>(UserData);
const size_t WriteSize = SizeInBlocks * BlockSizeInBytes;
TArray<uint8>* WriteHeaderBufferPtr = Request->WriteHeaderBufferPtr;
if (WriteHeaderBufferPtr && WriteSize > 0)
{
const size_t CurrentBufferLength = WriteHeaderBufferPtr->Num();
if (CurrentBufferLength > 0)
{
// Remove the previous zero termination
(*WriteHeaderBufferPtr)[CurrentBufferLength-1] = ' ';
}
// Write the header
WriteHeaderBufferPtr->Append((const uint8*)Ptr, WriteSize + 1);
(*WriteHeaderBufferPtr)[WriteHeaderBufferPtr->Num()-1] = 0; // Zero terminate string
return WriteSize;
}
return 0;
}
static size_t StaticWriteBodyFn(void* Ptr, size_t SizeInBlocks, size_t BlockSizeInBytes, void* UserData)
{
FHttpRequest* Request = static_cast<FHttpRequest*>(UserData);
const size_t WriteSize = SizeInBlocks * BlockSizeInBytes;
TArray<uint8>* WriteDataBufferPtr = Request->WriteDataBufferPtr;
if (WriteDataBufferPtr && WriteSize > 0)
{
// If this is the first part of the body being received, try to reserve
// memory if content length is defined in the header.
if (Request->BytesReceived == 0 && Request->WriteHeaderBufferPtr)
{
static const ANSICHAR* ContentLengthHeaderStr = "Content-Length: ";
const ANSICHAR* Header = (const ANSICHAR*)Request->WriteHeaderBufferPtr->GetData();
if (const ANSICHAR* ContentLengthHeader = FCStringAnsi::Strstr(Header, ContentLengthHeaderStr))
{
size_t ContentLength = (size_t)FCStringAnsi::Atoi64(ContentLengthHeader + strlen(ContentLengthHeaderStr));
if (ContentLength > 0u && ContentLength < UE_HTTPDDC_MAX_BUFFER_RESERVE)
{
WriteDataBufferPtr->Reserve(ContentLength);
}
}
}
// Write to the target buffer
WriteDataBufferPtr->Append((const uint8*)Ptr, WriteSize);
Request->BytesReceived += WriteSize;
return WriteSize;
}
return 0;
}
static size_t StaticSeekFn(void* UserData, curl_off_t Offset, int Origin)
{
FHttpRequest* Request = static_cast<FHttpRequest*>(UserData);
size_t NewPosition = 0;
switch (Origin)
{
case SEEK_SET: NewPosition = Offset; break;
case SEEK_CUR: NewPosition = Request->BytesSent + Offset; break;
case SEEK_END: NewPosition = Request->ReadDataView.GetSize() + Offset; break;
}
// Make sure we don't seek outside of the buffer
if (NewPosition < 0 || NewPosition >= Request->ReadDataView.GetSize())
{
return CURL_SEEKFUNC_FAIL;
}
// Update the used offset
Request->BytesSent = NewPosition;
return CURL_SEEKFUNC_OK;
}
};
//----------------------------------------------------------------------------------------------------------
// Forward declarations
//----------------------------------------------------------------------------------------------------------
bool VerifyPayload(const FSHAHash& Hash, const TCHAR* Namespace, const TCHAR* Bucket, const TCHAR* CacheKey, const TArray<uint8>& Payload);
bool VerifyPayload(const FIoHash& Hash, const TCHAR* Namespace, const TCHAR* Bucket, const TCHAR* CacheKey, const TArray<uint8>& Payload);
bool VerifyRequest(const class FHttpRequest* Request, const TCHAR* Namespace, const TCHAR* Bucket, const TCHAR* CacheKey, const TArray<uint8>& Payload);
bool HashPayload(class FHttpRequest* Request, const TArrayView<const uint8> Payload);
bool ShouldAbortForShutdown();
//----------------------------------------------------------------------------------------------------------
// Request pool
//----------------------------------------------------------------------------------------------------------
/**
* Pool that manages a fixed set of requests. Users are required to release requests that have been
* acquired. Usable with \ref FScopedRequestPtr which handles this automatically.
*/
struct FRequestPool
{
FRequestPool(const TCHAR* InServiceUrl, const TCHAR* InEffectiveServiceUrl, FHttpAccessToken* InAuthorizationToken, uint32 PoolSize, uint32 InOverflowLimit = 0)
: ActiveOverflowRequests(0)
, OverflowLimit(InOverflowLimit)
{
Pool.AddUninitialized(PoolSize);
Requests.AddUninitialized(PoolSize);
for (uint8 i = 0; i < Pool.Num(); ++i)
{
Pool[i].Usage = 0u;
Pool[i].Request = new(&Requests[i]) FHttpRequest(InServiceUrl, InEffectiveServiceUrl, InAuthorizationToken, true);
}
InitData = MakeUnique<FInitData>(InServiceUrl, InEffectiveServiceUrl, InAuthorizationToken);
}
~FRequestPool()
{
check(ActiveOverflowRequests.load() == 0);
for (uint8 i = 0; i < Pool.Num(); ++i)
{
// No requests should be in use by now.
check(Pool[i].Usage.load(std::memory_order_acquire) == 0u);
}
}
/**
* Attempts to get a request is free. Once a request has been returned it is
* "owned by the caller and need to release it to the pool when work has been completed.
* @return Usable request instance if one is available, otherwise null.
*/
FHttpRequest* GetFreeRequest(bool bUnboundedOverflow = false)
{
for (uint8 i = 0; i < Pool.Num(); ++i)
{
if (!Pool[i].Usage.load(std::memory_order_relaxed))
{
uint8 Expected = 0u;
if (Pool[i].Usage.compare_exchange_strong(Expected, 1u))
{
Pool[i].Request->Reset();
return Pool[i].Request;
}
}
}
if (bUnboundedOverflow || (OverflowLimit > 0))
{
// The use of two operations here (load, then fetch_add) implies that we can exceed the overflow limit because the combined operation
// is not atomic. This is acceptable for our use case. If we wanted to enforce the hard limit, we could use a loop instead.
if (bUnboundedOverflow || (ActiveOverflowRequests.load(std::memory_order_relaxed) < OverflowLimit))
{
// Create an overflow request (outside of the pre-allocated range of requests)
ActiveOverflowRequests.fetch_add(1, std::memory_order_relaxed);
return new FHttpRequest(*InitData->ServiceUrl, *InitData->EffectiveServiceUrl, InitData->AccessToken, true);
}
}
return nullptr;
}
class FWaiter : public FThreadSafeRefCountedObject
{
public:
std::atomic<FHttpRequest*> Request{ nullptr };
FWaiter(FRequestPool* InPool)
: Event(FPlatformProcess::GetSynchEventFromPool(true))
, Pool(InPool)
{
}
bool Wait(uint32 TimeMS)
{
return Event->Wait(TimeMS);
}
void Trigger()
{
Event->Trigger();
}
private:
~FWaiter()
{
FPlatformProcess::ReturnSynchEventToPool(Event);
if (Request)
{
Pool->ReleaseRequestToPool(Request.exchange(nullptr));
}
}
FEvent* Event;
FRequestPool* Pool;
};
/**
* Block until a request is free. Once a request has been returned it is
* "owned by the caller and need to release it to the pool when work has been completed.
* @return Usable request instance.
*/
FHttpRequest* WaitForFreeRequest(bool bUnboundedOverflow = false)
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_WaitForConnection);
FHttpRequest* Request = GetFreeRequest(bUnboundedOverflow);
if (Request == nullptr)
{
// Make it a fair by allowing each thread to register itself in a FIFO
// so that the first thread to start waiting is the first one to get a request.
FWaiter* Waiter = new FWaiter(this);
Waiter->AddRef(); // One ref for the thread that will dequeue
Waiter->AddRef(); // One ref for us
Waiters.enqueue(Waiter);
while (!Waiter->Wait(UE_HTTPDDC_BACKEND_WAIT_INTERVAL_MS))
{
// While waiting, allow us to check if a race occurred and a request has been freed
// between the time we checked for free requests and the time we queued ourself as a Waiter.
if ((Request = GetFreeRequest(bUnboundedOverflow)) != nullptr)
{
// We abandon the FWaiter, it will be freed by the next dequeue
// and if it has a request, it will be queued back to the pool.
Waiter->Release();
return Request;
}
}
Request = Waiter->Request.exchange(nullptr);
Request->Reset();
Waiter->Release();
}
check(Request);
return Request;
}
/**
* Release request to the pool.
* @param Request Request that should be freed. Note that any buffer owened by the request can now be reset.
*/
void ReleaseRequestToPool(FHttpRequest* Request)
{
if ((Request < Requests.GetData()) || (Request >= (Requests.GetData() + Requests.Num())))
{
// For overflow requests (outside of the pre-allocated range of requests), just delete it immediately
delete Request;
ActiveOverflowRequests.fetch_sub(1, std::memory_order_relaxed);
return;
}
for (uint8 i = 0; i < Pool.Num(); ++i)
{
if (Pool[i].Request == Request)
{
// If only 1 user is remaining, we can give it to a waiter
// instead of releasing it back to the pool.
if (Pool[i].Usage == 1u)
{
if (FWaiter* Waiter = Waiters.dequeue())
{
Waiter->Request = Request;
Waiter->Trigger();
Waiter->Release();
return;
}
}
Pool[i].Usage--;
return;
}
}
check(false);
}
/**
* While holding a request, make it shared across many users.
*/
void MakeRequestShared(FHttpRequest* Request, uint8 Users)
{
if ((Request < Requests.GetData()) || (Request >= (Requests.GetData() + Requests.Num())))
{
// Overflow requests (outside of the pre-allocated range of requests), cannot be made shared
check(false);
}
check(Users != 0);
for (uint8 i = 0; i < Pool.Num(); ++i)
{
if (Pool[i].Request == Request)
{
Pool[i].Usage = Users;
return;
}
}
check(false);
}
private:
struct FEntry
{
std::atomic<uint8> Usage;
FHttpRequest* Request;
};
TArray<FEntry> Pool;
TArray<FHttpRequest> Requests;
FAAArrayQueue<FWaiter> Waiters;
std::atomic<uint32> ActiveOverflowRequests;
struct FInitData
{
FString ServiceUrl;
FString EffectiveServiceUrl;
FHttpAccessToken* AccessToken;
FInitData(const TCHAR* InServiceUrl, const TCHAR* InEffectiveServiceUrl, FHttpAccessToken* InAccessToken)
: ServiceUrl(InServiceUrl)
, EffectiveServiceUrl(InEffectiveServiceUrl)
, AccessToken(InAccessToken)
{
}
};
TUniquePtr<const FInitData> InitData;
const uint32 OverflowLimit;
FRequestPool() = delete;
};
//----------------------------------------------------------------------------------------------------------
// FScopedRequestPtr
//----------------------------------------------------------------------------------------------------------
/**
* Utility class to manage requesting and releasing requests from the \ref FRequestPool.
*/
struct FScopedRequestPtr
{
public:
FScopedRequestPtr(FRequestPool* InPool)
: Request(InPool->WaitForFreeRequest())
, Pool(InPool)
{
}
~FScopedRequestPtr()
{
Pool->ReleaseRequestToPool(Request);
}
bool IsValid() const
{
return Request != nullptr;
}
FHttpRequest* Get() const
{
check(IsValid());
return Request;
}
FHttpRequest* operator->()
{
check(IsValid());
return Request;
}
private:
FHttpRequest* Request;
FRequestPool* Pool;
};
#if WITH_DATAREQUEST_HELPER
//----------------------------------------------------------------------------------------------------------
// FDataRequestHelper
//----------------------------------------------------------------------------------------------------------
/**
* Helper class for requesting data. Will batch requests once the number of concurrent requests reach a threshold.
*/
struct FDataRequestHelper
{
FDataRequestHelper(FRequestPool* InPool, const TCHAR* InNamespace, const TCHAR* InBucket, const TCHAR* InCacheKey, TArray<uint8>* OutData)
: Request(nullptr)
, Pool(InPool)
, bVerified(false, 1)
{
Request = Pool->GetFreeRequest();
if (Request && OutData != nullptr)
{
// We are below the threshold, make the connection immediately. OutData is set so this is a get.
FString Uri = FString::Printf(TEXT("api/v1/c/ddc/%s/%s/%s.raw"), InNamespace, InBucket, InCacheKey);
const FHttpRequest::Result Result = Request->PerformBlockingDownload(*Uri, OutData);
if (FHttpRequest::IsSuccessResponse(Request->GetResponseCode()))
{
if (VerifyRequest(Request, InNamespace, InBucket, InCacheKey, *OutData))
{
TRACE_COUNTER_ADD(HttpDDC_GetHit, int64(1));
TRACE_COUNTER_ADD(HttpDDC_BytesReceived, int64(Request->GetBytesReceived()));
bVerified[0] = true;
}
}
}
else if (Request)
{
// We are below the threshold, make the connection immediately. OutData is missing so this is a head.
FString Uri = FString::Printf(TEXT("api/v1/c/ddc/%s/%s/%s"), InNamespace, InBucket, InCacheKey);
const FHttpRequest::Result Result = Request->PerformBlockingQuery<FHttpRequest::Head>(*Uri);
if (FHttpRequest::IsSuccessResponse(Request->GetResponseCode()))
{
TRACE_COUNTER_ADD(HttpDDC_ExistHit, int64(1));
bVerified[0] = true;
}
}
else
{
// We have exceeded the threshold for concurrent connections, start or add this request
// to a batched request.
if (IsQueueCandidate(1, OutData && !OutData->IsEmpty()))
{
Request = QueueBatchRequest(
InPool,
InNamespace,
InBucket,
TConstArrayView<const TCHAR*>({InCacheKey}),
OutData ? TConstArrayView<TArray<uint8>*>({OutData}) : TConstArrayView<TArray<uint8>*>(),
bVerified
);
}
if (!Request)
{
Request = Pool->WaitForFreeRequest();
FQueuedBatchEntry Entry{
InNamespace,
InBucket,
TConstArrayView<const TCHAR*>({InCacheKey}),
OutData ? TConstArrayView<TArray<uint8>*>({OutData}) : TConstArrayView<TArray<uint8>*>(),
OutData && !OutData->IsEmpty() ? FHttpRequest::RequestVerb::Get : FHttpRequest::RequestVerb::Head,
&bVerified
};
PerformBatchQuery(Request, TArrayView<FQueuedBatchEntry>(&Entry, 1));
}
}
}
// Constructor specifically for batched head queries
FDataRequestHelper(FRequestPool* InPool, const TCHAR* InNamespace, const TCHAR* InBucket, TConstArrayView<FString> InCacheKeys)
: Request(nullptr)
, Pool(InPool)
, bVerified(false, InCacheKeys.Num())
{
// Transform the FString array to char pointers
TArray<const TCHAR*> CacheKeys;
Algo::Transform(InCacheKeys, CacheKeys, [](const FString& Key) { return *Key; });
Request = Pool->GetFreeRequest();
if (Request || !IsQueueCandidate(InCacheKeys.Num(), false))
{
// If the request is too big for existing batches, wait for a free connection and create our own.
if (!Request)
{
Request = Pool->WaitForFreeRequest();
}
FQueuedBatchEntry Entry{
InNamespace,
InBucket,
CacheKeys,
TConstArrayView<TArray<uint8>*>(),
FHttpRequest::RequestVerb::Head,
&bVerified
};
PerformBatchQuery(Request, TArrayView<FQueuedBatchEntry>(&Entry, 1));
}
else
{
Request = QueueBatchRequest(
InPool,
InNamespace,
InBucket,
CacheKeys,
TConstArrayView<TArray<uint8>*>(),
bVerified
);
if (!Request)
{
Request = Pool->WaitForFreeRequest();
FQueuedBatchEntry Entry{
InNamespace,
InBucket,
CacheKeys,
TConstArrayView<TArray<uint8>*>(),
FHttpRequest::RequestVerb::Head,
&bVerified
};
PerformBatchQuery(Request, TArrayView<FQueuedBatchEntry>(&Entry, 1));
}
}
}
~FDataRequestHelper()
{
if (Request)
{
Pool->ReleaseRequestToPool(Request);
}
}
static void StaticInitialize()
{
static bool bInitialized = false;
check(!bInitialized);
for (FBatch& Batch : Batches)
{
Batch.Reserved = 0;
Batch.Ready = 0;
Batch.Complete = TUniquePtr<FEvent, FBatch::FEventDeleter>(FPlatformProcess::GetSynchEventFromPool(true));
}
bInitialized = true;
}
static void StaticShutdown()
{
for (FBatch& Batch : Batches)
{
Batch.Complete.Reset();
}
}
bool IsSuccess() const
{
return bVerified[0];
}
const TBitArray<>& IsBatchSuccess() const
{
return bVerified;
}
int64 GetResponseCode() const
{
return Request ? Request->GetResponseCode() : 0;
}
private:
struct FQueuedBatchEntry
{
const TCHAR* Namespace;
const TCHAR* Bucket;
TConstArrayView<const TCHAR*> CacheKeys;
TConstArrayView<TArray<uint8>*> OutDatas;
FHttpRequest::RequestVerb Verb;
TBitArray<>* bSuccess;
};
struct FBatch
{
struct FEventDeleter
{
void operator()(FEvent* Event)
{
FPlatformProcess::ReturnSynchEventToPool(Event);
}
};
FQueuedBatchEntry Entries[UE_HTTPDDC_BATCH_SIZE];
std::atomic<uint32> Reserved;
std::atomic<uint32> Ready;
std::atomic<uint32> WeightHint;
FHttpRequest* Request;
TUniquePtr<FEvent, FEventDeleter> Complete;
};
FHttpRequest* Request;
FRequestPool* Pool;
TBitArray<> bVerified;
static std::atomic<uint32> FirstAvailableBatch;
static TStaticArray<FBatch, UE_HTTPDDC_BATCH_NUM> Batches;
static uint32 ComputeWeight(int32 NumKeys, bool bHasDatas)
{
return NumKeys * (bHasDatas ? UE_HTTPDDC_BATCH_GET_WEIGHT : UE_HTTPDDC_BATCH_HEAD_WEIGHT);
}
static bool IsQueueCandidate(int32 NumKeys, bool bHasDatas)
{
if (NumKeys > UE_HTTPDDC_BATCH_SIZE)
{
return false;
}
const uint32 Weight = ComputeWeight(NumKeys, bHasDatas);
if (Weight > UE_HTTPDDC_BATCH_WEIGHT_HINT)
{
return false;
}
return true;
}
/**
* Queues up a request to be batched. Blocks until the query is made.
*/
static FHttpRequest* QueueBatchRequest(FRequestPool* InPool,
const TCHAR* InNamespace,
const TCHAR* InBucket,
TConstArrayView<const TCHAR*> InCacheKeys,
TConstArrayView<TArray<uint8>*> OutDatas,
TBitArray<>& bOutVerified)
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_BatchQuery);
check(InCacheKeys.Num() == OutDatas.Num() || OutDatas.Num() == 0);
const uint32 RequestNum = InCacheKeys.Num();
const uint32 RequestWeight = ComputeWeight(InCacheKeys.Num(), !OutDatas.IsEmpty());
for (int32 i = 0; i < Batches.Num(); i++)
{
uint32 Index = (FirstAvailableBatch.load(std::memory_order_relaxed) + i) % Batches.Num();
FBatch& Batch = Batches[Index];
//Assign different weights to head vs. get queries
if (Batch.WeightHint.load(std::memory_order_acquire) + RequestWeight > UE_HTTPDDC_BATCH_WEIGHT_HINT)
{
continue;
}
// Attempt to reserve a spot in the batch
const uint32 Reserve = Batch.Reserved.fetch_add(1, std::memory_order_acquire);
if (Reserve >= UE_HTTPDDC_BATCH_SIZE)
{
// We didn't manage to snag a valid reserve index try next batch
continue;
}
// Add our weight to the batch. Note we are treating it as a hint, so don't syncronize.
const uint32 ActualWeight = Batch.WeightHint.fetch_add(RequestWeight, std::memory_order_release);
TAnsiStringBuilder<64> BatchString;
BatchString << "HttpDDC_Batch" << Index;
TRACE_CPUPROFILER_EVENT_SCOPE_TEXT(*BatchString);
if (Reserve == (UE_HTTPDDC_BATCH_SIZE - 1))
{
FirstAvailableBatch++;
}
Batch.Entries[Reserve] = FQueuedBatchEntry{
InNamespace,
InBucket,
InCacheKeys,
OutDatas,
OutDatas.Num() ? FHttpRequest::RequestVerb::Get : FHttpRequest::RequestVerb::Head,
&bOutVerified
};
// Signal we are ready for batch to be submitted
Batch.Ready.fetch_add(1u, std::memory_order_release);
FHttpRequest* Request = nullptr;
// The first to reserve a slot is the "driver" of the batch
if (Reserve == 0)
{
Batch.Request = InPool->WaitForFreeRequest();
// Make sure no new requests are added
const uint32 Reserved = FMath::Min((uint32)UE_HTTPDDC_BATCH_SIZE, Batch.Reserved.fetch_add(UE_HTTPDDC_BATCH_SIZE, std::memory_order_acquire));
// Give other threads time to copy their data to batch
while (Batch.Ready.load(std::memory_order_acquire) < Reserved)
{
}
// Increment request ref count to reflect all waiting threads
InPool->MakeRequestShared(Batch.Request, Reserved);
// Do the actual query and write response to respective target arrays
PerformBatchQuery(Batch.Request, TArrayView<FQueuedBatchEntry>(Batch.Entries, Batch.Ready));
// Signal to waiting threads the batch is complete
Batch.Complete->Trigger();
// Store away the request and wait until other threads have too
Request = Batch.Request;
while (Batch.Ready.load(std::memory_order_acquire) > 1)
{
}
//Reset batch for next use
Batch.Complete->Reset();
Batch.WeightHint.store(0, std::memory_order_release);
Batch.Ready.store(0, std::memory_order_release);
Batch.Reserved.store(0, std::memory_order_release);
}
else
{
// Wait until "driver" has done query
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_WaitForMasterOfBatch);
Batch.Complete->Wait(~0);
}
// Store away request and signal we are done
Request = Batch.Request;
Batch.Ready.fetch_sub(1u, std::memory_order_release);
}
return Request;
}
return nullptr;
}
/**
* Creates request uri and headers and submits the request
*/
static void PerformBatchQuery(FHttpRequest* Request, TArrayView<FQueuedBatchEntry> Entries)
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_BatchGet);
const TCHAR* Uri(TEXT("api/v1/c/ddc-rpc/batchget"));
int64 ResponseCode = 0; uint32 Attempts = 0;
//Prepare request object
TArray<TSharedPtr<FJsonValue>> Operations;
for (const FQueuedBatchEntry& Entry : Entries)
{
for (int32 KeyIdx = 0; KeyIdx < Entry.CacheKeys.Num(); KeyIdx++)
{
TSharedPtr<FJsonObject> Object = MakeShared<FJsonObject>();
Object->SetField(TEXT("bucket"), MakeShared<FJsonValueString>(Entry.Bucket));
Object->SetField(TEXT("key"), MakeShared<FJsonValueString>(Entry.CacheKeys[KeyIdx]));
if (Entry.Verb == FHttpRequest::RequestVerb::Head)
{
Object->SetField(TEXT("verb"), MakeShared<FJsonValueString>(TEXT("HEAD")));
}
Operations.Add(MakeShared<FJsonValueObject>(Object));
}
}
TSharedPtr<FJsonObject> RequestObject = MakeShared<FJsonObject>();
RequestObject->SetField(TEXT("namespace"), MakeShared<FJsonValueString>(Entries[0].Namespace));
RequestObject->SetField(TEXT("operations"), MakeShared<FJsonValueArray>(Operations));
//Serialize to a buffer
FBufferArchive RequestData;
if (FJsonSerializer::Serialize(RequestObject.ToSharedRef(), TJsonWriterFactory<ANSICHAR, TCondensedJsonPrintPolicy<ANSICHAR>>::Create(&RequestData)))
{
Request->PerformBlockingUpload<FHttpRequest::PostJson>(Uri, MakeArrayView(RequestData));
ResponseCode = Request->GetResponseCode();
if (ResponseCode == 200)
{
const TArray<uint8>& ResponseBuffer = Request->GetResponseBuffer();
const uint8* Response = ResponseBuffer.GetData();
const int32 ResponseSize = ResponseBuffer.Num();
// Parse the response and move the data to the target requests.
if (ParseBatchedResponse(Response, ResponseSize, Entries))
{
UE_LOG(LogDerivedDataCache, VeryVerbose, TEXT("%s: Batch query with %d operations completed."), *Request->GetName(), Entries.Num());
return;
}
}
}
// If we get here the request failed.
UE_LOG(LogDerivedDataCache, Display, TEXT("%s: Batch query failed. Query: %s"), *Request->GetName(), ANSI_TO_TCHAR((ANSICHAR*)RequestData.GetData()));
// Set all batch operations to failures
for (FQueuedBatchEntry Entry : Entries)
{
Entry.bSuccess->SetRange(0, Entry.CacheKeys.Num(), false);
}
}
// Above result value
enum class OpResult : uint8
{
Ok = 0, // Op finished succesfully
Error = 1, // Error during op
NotFound = 2, // Key was not found
Exists = 3 // Used to indicate head op success
};
// Searches for potentially multiple key requests that are satisfied the given cache key result
// Search strategy is exhaustive forward search from the last found entry. If the results come in ordered the same as the requests,
// and there are no duplicates, the search will be somewhat efficient (still has to do exhaustive searching looking for duplicates).
// If the results are unordered or there are duplicates, search will become more inefficient.
struct FRequestSearchHelper
{
FRequestSearchHelper(TArrayView<FQueuedBatchEntry> InRequests, const FUTF8ToTCHAR& InCacheKey, int32 InEntryIdx, int32 InKeyIdx, OpResult InRequestResult)
: Requests(InRequests)
, CacheKey(InCacheKey)
, StartEntryIdx(InEntryIdx)
, StartKeyIdx(InKeyIdx)
, RequestResult(InRequestResult)
{
}
bool FindNext(int32& EntryIdx, int32& KeyIdx)
{
int32 CurrentEntryIdx = EntryIdx;
int32 CurrentKeyIdx = KeyIdx;
do
{
// Do not match a get request with a head response code (i.e. Exists)
// or a head request with a get response code (i.e. Ok)
// if the response code is an error or not found they can be matched to both head or get request it doesn't matter
const FQueuedBatchEntry& CurrentRequest = Requests[CurrentEntryIdx];
bool bRequestTypeMatch = !((CurrentRequest.Verb == FHttpRequest::Get) && (RequestResult == OpResult::Exists))
&& !((CurrentRequest.Verb == FHttpRequest::Head) && (RequestResult == OpResult::Ok));
if (bRequestTypeMatch && FCString::Stricmp(CurrentRequest.CacheKeys[CurrentKeyIdx], CacheKey.Get()) == 0)
{
EntryIdx = CurrentEntryIdx;
KeyIdx = CurrentKeyIdx;
return true;
}
}
while (AdvanceIndices(CurrentEntryIdx, CurrentKeyIdx));
return false;
}
bool AdvanceIndices(int32& EntryIdx, int32& KeyIdx)
{
if (++KeyIdx >= Requests[EntryIdx].CacheKeys.Num())
{
EntryIdx = (EntryIdx + 1) % Requests.Num();
KeyIdx = 0;
}
return !((EntryIdx == StartEntryIdx) && (KeyIdx == StartKeyIdx));
}
TArrayView<FQueuedBatchEntry> Requests;
const FUTF8ToTCHAR& CacheKey;
int32 StartEntryIdx;
int32 StartKeyIdx;
OpResult RequestResult;
};
/**
* Parses a batched response stream, moves the data to target requests and marks them with result.
* @param Response Pointer to Response buffer
* @param ResponseSize Size of response buffer
* @param Requests Requests that will be filled with data.
* @return True if response was successfully parsed, false otherwise.
*/
static bool ParseBatchedResponse(const uint8* ResponseStart, const int32 ResponseSize, TArrayView<FQueuedBatchEntry> Requests)
{
// The expected data stream is structured accordingly
// {"JPTR"} {PayloadCount:uint32} {{"JPEE"} {Name:cstr} {Result:uint8} {Hash:IoHash} {Size:uint64} {Payload...}} ...
const auto& ResponseErrorMessage = TEXT("Malformed response from server.");
const ANSICHAR* ProtocolMagic = "JPTR";
const ANSICHAR* PayloadMagic = "JPEE";
const uint32 MagicSize = 4;
const uint8* Response = ResponseStart;
const uint8* ResponseEnd = Response + ResponseSize;
// Check that the stream starts with the protocol magic
if (FMemory::Memcmp(ProtocolMagic, Response, MagicSize) != 0)
{
UE_LOG(LogDerivedDataCache, Display, ResponseErrorMessage);
return false;
}
Response += MagicSize;
// Number of payloads recieved
uint32 PayloadCount = *(uint32*)Response;
Response += sizeof(uint32);
uint32 PayloadIdx = 0; // Current processed result
int32 EntryIdx = 0; // Current Entry index
int32 KeyIdx = 0; // Current Key index for current Entry
while (Response < ResponseEnd && FMemory::Memcmp(PayloadMagic, Response, MagicSize) == 0)
{
PayloadIdx++;
Response += MagicSize;
const ANSICHAR* PayloadNameA = (const ANSICHAR*)Response;
Response += FCStringAnsi::Strlen(PayloadNameA) + 1; //String and zero termination
const ANSICHAR* CacheKeyA = FCStringAnsi::Strrchr(PayloadNameA, '.') + 1; // "namespace.bucket.cachekey"
// Result of the operation is used to match to the appropriate request (i.e. get or head)
OpResult PayloadResult = static_cast<OpResult>(*Response);
Response += sizeof(uint8);
const uint8* ResponseRewindMark = Response;
// Find the payload among the requests. Payloads may be returned in any order and if the same cache key was part of two requests,
// a single payload may satisfy multiple cache keys in multiple requests.
FUTF8ToTCHAR CacheKey(CacheKeyA);
FRequestSearchHelper RequestSearch(Requests, CacheKey, EntryIdx, KeyIdx, PayloadResult);
bool bFoundAny = false;
while (RequestSearch.FindNext(EntryIdx, KeyIdx))
{
Response = ResponseRewindMark;
bFoundAny = true;
FQueuedBatchEntry& RequestOp = Requests[EntryIdx];
TBitArray<>& bSuccess = *RequestOp.bSuccess;
switch (PayloadResult)
{
case OpResult::Ok:
{
// Payload hash of the following payload data
FIoHash PayloadHash = *(FIoHash*)Response;
Response += sizeof(FIoHash);
// Size of the following payload data
const uint64 PayloadSize = *(uint64*)Response;
Response += sizeof(uint64);
if (PayloadSize > 0)
{
if (Response + PayloadSize > ResponseEnd)
{
UE_LOG(LogDerivedDataCache, Display, ResponseErrorMessage);
return false;
}
if (bSuccess[KeyIdx])
{
Response += PayloadSize;
}
else
{
TArray<uint8>* OutData = RequestOp.OutDatas[KeyIdx];
OutData->Append(Response, PayloadSize);
Response += PayloadSize;
// Verify the received and parsed payload
if (VerifyPayload(PayloadHash, RequestOp.Namespace, RequestOp.Bucket, RequestOp.CacheKeys[KeyIdx], *OutData))
{
TRACE_COUNTER_ADD(HttpDDC_GetHit, int64(1));
TRACE_COUNTER_ADD(HttpDDC_BytesReceived, int64(PayloadSize));
bSuccess[KeyIdx] = true;
}
else
{
OutData->Empty();
bSuccess[KeyIdx] = false;
}
}
}
else
{
bSuccess[KeyIdx] = false;
}
}
break;
case OpResult::Exists:
{
TRACE_COUNTER_ADD(HttpDDC_ExistHit, int64(1));
bSuccess[KeyIdx] = true;
}
break;
default:
case OpResult::Error:
UE_LOG(LogDerivedDataCache, Display, TEXT("Server error while getting %s"), CacheKey.Get());
// intentional falltrough
case OpResult::NotFound:
bSuccess[KeyIdx] = false;
break;
}
if (!RequestSearch.AdvanceIndices(EntryIdx, KeyIdx))
{
break;
}
}
if (!bFoundAny)
{
UE_LOG(LogDerivedDataCache, Error, ResponseErrorMessage);
return false;
}
}
// Have we parsed all the payloads from the message?
if (PayloadIdx != PayloadCount)
{
UE_LOG(LogDerivedDataCache, Display, TEXT("%s: Found %d payloads but %d was reported."), ResponseErrorMessage, PayloadIdx, PayloadCount);
}
return true;
}
};
TStaticArray<FDataRequestHelper::FBatch, UE_HTTPDDC_BATCH_NUM> FDataRequestHelper::Batches;
std::atomic<uint32> FDataRequestHelper::FirstAvailableBatch;
//----------------------------------------------------------------------------------------------------------
// FDataUploadHelper
//----------------------------------------------------------------------------------------------------------
struct FDataUploadHelper
{
FDataUploadHelper(FRequestPool* InPool,
const TCHAR* InNamespace,
const TCHAR* InBucket,
const TCHAR* InCacheKey,
const TArrayView<const uint8>& InData,
FDerivedDataCacheUsageStats& InUsageStats)
: ResponseCode(0)
, bSuccess(false)
, bQueued(false)
{
FHttpRequest* Request = InPool->GetFreeRequest();
if (Request)
{
ResponseCode = PerformPut(Request, InNamespace, InBucket, InCacheKey, InData, InUsageStats);
bSuccess = FHttpRequest::IsSuccessResponse(Request->GetResponseCode());
ProcessQueuedPutsAndReleaseRequest(InPool, Request, InUsageStats);
}
else
{
FQueuedEntry* Entry = new FQueuedEntry(InNamespace, InBucket, InCacheKey, InData);
QueuedPuts.Push(Entry);
bSuccess = true;
bQueued = true;
// A request may have been released while the entry was being queued.
Request = InPool->GetFreeRequest();
if (Request)
{
ProcessQueuedPutsAndReleaseRequest(InPool, Request, InUsageStats);
}
}
}
bool IsSuccess() const
{
return bSuccess;
}
int64 GetResponseCode() const
{
return ResponseCode;
}
bool IsQueued() const
{
return bQueued;
}
private:
struct FQueuedEntry
{
FString Namespace;
FString Bucket;
FString CacheKey;
TArray<uint8> Data;
FQueuedEntry(const TCHAR* InNamespace, const TCHAR* InBucket, const TCHAR* InCacheKey, const TArrayView<const uint8> InData)
: Namespace(InNamespace)
, Bucket(InBucket)
, CacheKey(InCacheKey)
, Data(InData) // Copies the data!
{
}
};
static TLockFreePointerListUnordered<FQueuedEntry, PLATFORM_CACHE_LINE_SIZE> QueuedPuts;
int64 ResponseCode;
bool bSuccess;
bool bQueued;
static void ProcessQueuedPutsAndReleaseRequest(FRequestPool* Pool, FHttpRequest* Request, FDerivedDataCacheUsageStats& UsageStats)
{
while (Request)
{
// Make sure that whether we early exit or execute past the end of this scope that
// the request is released back to the pool.
{
ON_SCOPE_EXIT
{
Pool->ReleaseRequestToPool(Request);
};
if (ShouldAbortForShutdown())
{
return;
}
while (FQueuedEntry* Entry = QueuedPuts.Pop())
{
Request->Reset();
PerformPut(Request, *Entry->Namespace, *Entry->Bucket, *Entry->CacheKey, Entry->Data, UsageStats);
delete Entry;
if (ShouldAbortForShutdown())
{
return;
}
}
}
// An entry may have been queued while the request was being released.
if (QueuedPuts.IsEmpty())
{
break;
}
// Process the queue again if a request is free, otherwise the thread that got the request will process it.
Request = Pool->GetFreeRequest();
}
}
static int64 PerformPut(FHttpRequest* Request, const TCHAR* Namespace, const TCHAR* Bucket, const TCHAR* CacheKey, const TArrayView<const uint8> Data, FDerivedDataCacheUsageStats& UsageStats)
{
COOK_STAT(auto Timer = UsageStats.TimePut());
HashPayload(Request, Data);
TStringBuilder<256> Uri;
Uri.Appendf(TEXT("api/v1/c/ddc/%s/%s/%s"), Namespace, Bucket, CacheKey);
Request->PerformBlockingUpload<FHttpRequest::Put>(*Uri, Data);
const int64 ResponseCode = Request->GetResponseCode();
if (FHttpRequest::IsSuccessResponse(ResponseCode))
{
TRACE_COUNTER_ADD(HttpDDC_BytesSent, int64(Request->GetBytesSent()));
COOK_STAT(Timer.AddHit(Request->GetBytesSent()));
}
return Request->GetResponseCode();
}
};
TLockFreePointerListUnordered<FDataUploadHelper::FQueuedEntry, PLATFORM_CACHE_LINE_SIZE> FDataUploadHelper::QueuedPuts;
#endif // WITH_DATAREQUEST_HELPER
//----------------------------------------------------------------------------------------------------------
// Certificate checking
//----------------------------------------------------------------------------------------------------------
#if WITH_SSL
static int SslCertVerify(int PreverifyOk, X509_STORE_CTX* Context)
{
if (PreverifyOk == 1)
{
SSL* Handle = static_cast<SSL*>(X509_STORE_CTX_get_ex_data(Context, SSL_get_ex_data_X509_STORE_CTX_idx()));
check(Handle);
SSL_CTX* SslContext = SSL_get_SSL_CTX(Handle);
check(SslContext);
FHttpRequest* Request = static_cast<FHttpRequest*>(SSL_CTX_get_app_data(SslContext));
check(Request);
const FString& Domain = Request->GetDomain();
if (!FSslModule::Get().GetCertificateManager().VerifySslCertificates(Context, Domain))
{
PreverifyOk = 0;
}
}
return PreverifyOk;
}
static CURLcode sslctx_function(CURL * curl, void * sslctx, void * parm)
{
SSL_CTX* Context = static_cast<SSL_CTX*>(sslctx);
const ISslCertificateManager& CertificateManager = FSslModule::Get().GetCertificateManager();
CertificateManager.AddCertificatesToSslContext(Context);
SSL_CTX_set_verify(Context, SSL_CTX_get_verify_mode(Context), SslCertVerify);
SSL_CTX_set_app_data(Context, parm);
/* all set to go */
return CURLE_OK;
}
#endif //#if WITH_SSL
//----------------------------------------------------------------------------------------------------------
// Content parsing and checking
//----------------------------------------------------------------------------------------------------------
/**
* Verifies the integrity of the received data using supplied checksum.
* @param Hash received hash value.
* @param Namespace The namespace string used when originally fetching the request.
* @param Bucket The bucket string used when originally fetching the request.
* @param CacheKey The cache key string used when originally fetching the request.
* @param Payload Payload received.
* @return True if the data is correct, false if checksums doesn't match.
*/
bool VerifyPayload(const FSHAHash& Hash, const TCHAR* Namespace, const TCHAR* Bucket, const TCHAR* CacheKey, const TArray<uint8>& Payload)
{
FSHAHash PayloadHash;
FSHA1::HashBuffer(Payload.GetData(), Payload.Num(), PayloadHash.Hash);
if (Hash != PayloadHash)
{
UE_LOG(LogDerivedDataCache,
Display,
TEXT("Checksum from server did not match received data (%s vs %s). Discarding cached result. Namespace: %s, Bucket: %s, Key: %s."),
*WriteToString<48>(Hash),
*WriteToString<48>(PayloadHash),
Namespace,
Bucket,
CacheKey
);
return false;
}
return true;
}
/**
* Verifies the integrity of the received data using supplied checksum.
* @param Hash received hash value.
* @param Namespace The namespace string used when originally fetching the request.
* @param Bucket The bucket string used when originally fetching the request.
* @param CacheKey The cache key string used when originally fetching the request.
* @param Payload Payload received.
* @return True if the data is correct, false if checksums doesn't match.
*/
bool VerifyPayload(const FIoHash& Hash, const TCHAR* Namespace, const TCHAR* Bucket, const TCHAR* CacheKey, const TArray<uint8>& Payload)
{
FIoHash PayloadHash = FIoHash::HashBuffer(Payload.GetData(), Payload.Num());
if (Hash != PayloadHash)
{
UE_LOG(LogDerivedDataCache,
Display,
TEXT("Checksum from server did not match received data (%s vs %s). Discarding cached result. Namespace: %s, Bucket: %s, Key: %s."),
*WriteToString<48>(Hash),
*WriteToString<48>(PayloadHash),
Namespace,
Bucket,
CacheKey
);
return false;
}
return true;
}
/**
* Verifies the integrity of the received data using supplied checksum.
* @param Request Request that the data was be received with.
* @param Namespace The namespace string used when originally fetching the request.
* @param Bucket The bucket string used when originally fetching the request.
* @param CacheKey The cache key string used when originally fetching the request.
* @param Payload Payload received.
* @return True if the data is correct, false if checksums doesn't match.
*/
bool VerifyRequest(const FHttpRequest* Request, const TCHAR* Namespace, const TCHAR* Bucket, const TCHAR* CacheKey, const TArray<uint8>& Payload)
{
FString ReceivedHashStr;
if (Request->GetHeader("X-Jupiter-Sha1", ReceivedHashStr))
{
FSHAHash ReceivedHash;
ReceivedHash.FromString(ReceivedHashStr);
return VerifyPayload(ReceivedHash, Namespace, Bucket, CacheKey, Payload);
}
if (Request->GetHeader("X-Jupiter-IoHash", ReceivedHashStr))
{
FIoHash ReceivedHash(ReceivedHashStr);
return VerifyPayload(ReceivedHash, Namespace, Bucket, CacheKey, Payload);
}
UE_LOG(LogDerivedDataCache, Warning, TEXT("%s: HTTP server did not send a content hash. Wrong server version?"), *Request->GetName());
return true;
}
/**
* Adds a checksum (as request header) for a given payload. Jupiter will use this to verify the integrity
* of the received data.
* @param Request Request that the data will be sent with.
* @param Payload Payload that will be sent.
* @return True on success, false on failure.
*/
bool HashPayload(FHttpRequest* Request, const TArrayView<const uint8> Payload)
{
FIoHash PayloadHash = FIoHash::HashBuffer(Payload.GetData(), Payload.Num());
Request->SetHeader(TEXT("X-Jupiter-IoHash"), *WriteToString<48>(PayloadHash));
return true;
}
bool ShouldAbortForShutdown()
{
return !GIsBuildMachine && FDerivedDataBackend::Get().IsShuttingDown();
}
TConstArrayView<uint8> MakeConstArrayView(FSharedBuffer Buffer)
{
return TConstArrayView<uint8>(reinterpret_cast<const uint8*>(Buffer.GetData()), Buffer.GetSize());
}
static bool IsValueDataReady(FValue& Value, const ECachePolicy Policy)
{
if (!EnumHasAnyFlags(Policy, ECachePolicy::Query))
{
Value = Value.RemoveData();
return true;
}
if (Value.HasData())
{
if (EnumHasAnyFlags(Policy, ECachePolicy::SkipData))
{
Value = Value.RemoveData();
}
return true;
}
return false;
};
//----------------------------------------------------------------------------------------------------------
// FHttpAccessToken
//----------------------------------------------------------------------------------------------------------
FString FHttpAccessToken::GetHeader()
{
Lock.ReadLock();
FString Header = FString::Printf(TEXT("Authorization: Bearer %s"), *Token);
Lock.ReadUnlock();
return Header;
}
void FHttpAccessToken::SetHeader(const TCHAR* InToken)
{
Lock.WriteLock();
Token = InToken;
Serial++;
Lock.WriteUnlock();
}
uint32 FHttpAccessToken::GetSerial() const
{
return Serial;
}
//----------------------------------------------------------------------------------------------------------
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
// FHttpCacheStore
//----------------------------------------------------------------------------------------------------------
/**
* Backend for a HTTP based caching service (Jupiter).
**/
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
class FHttpCacheStore final : public FDerivedDataBackendInterface
{
public:
/**
* Creates the backend, checks health status and attempts to acquire an access token.
*
* @param ServiceUrl Base url to the service including schema.
* @param Namespace Namespace to use.
* @param StructuredNamespace Namespace to use for structured cache operations.
* @param OAuthProvider Url to OAuth provider, for example "https://myprovider.com/oauth2/v1/token".
* @param OAuthClientId OAuth client identifier.
* @param OAuthData OAuth form data to send to login service. Can either be the raw form data or a Windows network file address (starting with "\\").
* @param OAuthScope OAuth scope identifier
*/
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
FHttpCacheStore(
const TCHAR* ServiceUrl,
bool bResolveHostCanonicalName,
const TCHAR* Namespace,
const TCHAR* StructuredNamespace,
const TCHAR* OAuthProvider,
const TCHAR* OAuthClientId,
const TCHAR* OAuthData,
const TCHAR* OAuthScope,
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
EBackendLegacyMode LegacyMode,
bool bReadOnly);
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
~FHttpCacheStore();
/**
* Checks is backend is usable (reachable and accessible).
* @return true if usable
*/
bool IsUsable() const { return bIsUsable; }
/** return true if this cache is writable **/
virtual bool IsWritable() const override
{
return !bReadOnly && bIsUsable;
}
virtual bool CachedDataProbablyExists(const TCHAR* CacheKey) override;
virtual TBitArray<> CachedDataProbablyExistsBatch(TConstArrayView<FString> CacheKeys) override;
virtual bool GetCachedData(const TCHAR* CacheKey, TArray<uint8>& OutData) override;
virtual EPutStatus PutCachedData(const TCHAR* CacheKey, TArrayView<const uint8> InData, bool bPutEvenIfExists) override;
virtual void RemoveCachedData(const TCHAR* CacheKey, bool bTransient) override;
virtual TSharedRef<FDerivedDataCacheStatsNode> GatherUsageStats() const override;
virtual FString GetName() const override;
virtual TBitArray<> TryToPrefetch(TConstArrayView<FString> CacheKeys) override;
virtual bool WouldCache(const TCHAR* CacheKey, TArrayView<const uint8> InData) override;
virtual ESpeedClass GetSpeedClass() const override;
virtual bool ApplyDebugOptions(FBackendDebugOptions& InOptions) override;
void SetSpeedClass(ESpeedClass InSpeedClass) { SpeedClass = InSpeedClass; }
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
EBackendLegacyMode GetLegacyMode() const final { return LegacyMode; }
virtual void Put(
TConstArrayView<FCachePutRequest> Requests,
IRequestOwner& Owner,
FOnCachePutComplete&& OnComplete) override;
virtual void Get(
TConstArrayView<FCacheGetRequest> Requests,
IRequestOwner& Owner,
FOnCacheGetComplete&& OnComplete) override;
virtual void PutValue(
TConstArrayView<FCachePutValueRequest> Requests,
IRequestOwner& Owner,
FOnCachePutValueComplete&& OnComplete) override;
virtual void GetValue(
TConstArrayView<FCacheGetValueRequest> Requests,
IRequestOwner& Owner,
FOnCacheGetValueComplete&& OnComplete) override;
virtual void GetChunks(
TConstArrayView<FCacheGetChunkRequest> Requests,
IRequestOwner& Owner,
FOnCacheGetChunkComplete&& OnComplete) override;
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
static FHttpCacheStore* GetAny()
{
return AnyInstance;
}
const FString& GetDomain() const { return Domain; }
const FString& GetNamespace() const { return Namespace; }
const FString& GetStructuredNamespace() const { return StructuredNamespace; }
const FString& GetOAuthProvider() const { return OAuthProvider; }
const FString& GetOAuthClientId() const { return OAuthClientId; }
const FString& GetOAuthSecret() const { return OAuthSecret; }
const FString& GetOAuthScope() const { return OAuthScope; }
private:
FString Domain;
FString EffectiveDomain;
FString Namespace;
FString StructuredNamespace;
FString DefaultBucket;
FString OAuthProvider;
FString OAuthClientId;
FString OAuthSecret;
FString OAuthScope;
FCriticalSection AccessCs;
FDerivedDataCacheUsageStats UsageStats;
FBackendDebugOptions DebugOptions;
TUniquePtr<struct FRequestPool> GetRequestPools[2];
TUniquePtr<struct FRequestPool> PutRequestPools[2];
TUniquePtr<struct FRequestPool> NonBlockingRequestPools;
TUniquePtr<struct FHttpAccessToken> Access;
bool bIsUsable;
bool bReadOnly;
uint32 FailedLoginAttempts;
ESpeedClass SpeedClass;
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
EBackendLegacyMode LegacyMode;
static inline FHttpCacheStore* AnyInstance = nullptr;
bool IsServiceReady();
bool AcquireAccessToken();
bool ShouldRetryOnError(FHttpRequest::Result Result, int64 ResponseCode);
bool ShouldRetryOnError(int64 ResponseCode) { return ShouldRetryOnError(FHttpRequest::Success, ResponseCode); }
enum class OperationCategory
{
Get,
Put
};
template<OperationCategory Category>
FHttpRequest* WaitForHttpRequestForOwner(IRequestOwner& Owner, bool bUnboundedOverflow, FRequestPool*& OutPool)
{
if (!FHttpRequest::AllowAsync())
{
if (Category == OperationCategory::Get)
{
OutPool = GetRequestPools[IsInGameThread()].Get();
}
else
{
OutPool = PutRequestPools[IsInGameThread()].Get();
}
return OutPool->WaitForFreeRequest();
}
else
{
OutPool = NonBlockingRequestPools.Get();
return OutPool->WaitForFreeRequest(bUnboundedOverflow);
}
}
struct FGetCacheRecordOnlyResponse
{
FSharedString Name;
FCacheKey Key;
uint64 UserData = 0;
uint64 BytesReceived = 0;
FOptionalCacheRecord Record;
EStatus Status = EStatus::Error;
};
using FOnGetCacheRecordOnlyComplete = TUniqueFunction<void(FGetCacheRecordOnlyResponse&& Response)>;
void GetCacheRecordOnlyAsync(
IRequestOwner& Owner,
const FSharedString& Name,
const FCacheKey& Key,
const FCacheRecordPolicy& Policy,
uint64 UserData,
FOnGetCacheRecordOnlyComplete&& OnComplete);
void GetCacheRecordAsync(
IRequestOwner& Owner,
const FSharedString& Name,
const FCacheKey& Key,
const FCacheRecordPolicy& Policy,
uint64 UserData,
TUniqueFunction<void(FCacheGetResponse&& Response, uint64 BytesReceived)>&& OnComplete);
void PutCacheRecordAsync(
IRequestOwner& Owner,
const FSharedString& Name,
const FCacheRecord& Record,
const FCacheRecordPolicy& Policy,
uint64 UserData,
TUniqueFunction<void(FCachePutResponse&& Response, uint64 BytesSent)>&& OnComplete);
void PutCacheValueAsync(
IRequestOwner& Owner,
const FSharedString& Name,
const FCacheKey& Key,
const FValue& Value,
ECachePolicy Policy,
uint64 UserData,
TUniqueFunction<void(FCachePutValueResponse&& Response, uint64 BytesSent)>&& OnComplete);
void GetCacheValueAsync(
IRequestOwner& Owner,
FSharedString Name,
const FCacheKey& Key,
ECachePolicy Policy,
uint64 UserData,
FOnCacheGetValueComplete&& OnComplete);
void RefCachedDataProbablyExistsBatchAsync(
IRequestOwner& Owner,
TConstArrayView<FCacheGetValueRequest> ValueRefs,
FOnCacheGetValueComplete&& OnComplete);
class FPutPackageOp;
class FGetRecordOp;
};
//----------------------------------------------------------------------------------------------------------
// FHttpCacheStore::FPutPackageOp
//----------------------------------------------------------------------------------------------------------
class FHttpCacheStore::FPutPackageOp final : public FThreadSafeRefCountedObject
{
public:
struct FCachePutPackageResponse
{
FSharedString Name;
FCacheKey Key;
uint64 UserData = 0;
uint64 BytesSent = 0;
EStatus Status = EStatus::Error;
};
using FOnCachePutPackageComplete = TUniqueFunction<void(FCachePutPackageResponse&& Response)>;
/** Performs a multi-request operation for uploading a package of content. */
static void PutPackage(
FHttpCacheStore& CacheStore,
IRequestOwner& Owner,
const FSharedString& Name,
FCacheKey Key,
FCbPackage&& Package,
FCacheRecordPolicy Policy,
uint64 UserData,
FOnCachePutPackageComplete&& OnComplete);
private:
FHttpCacheStore& CacheStore;
IRequestOwner& Owner;
const FSharedString Name;
const FCacheKey Key;
const uint64 UserData;
std::atomic<uint64> BytesSent;
const FCbObject PackageObject;
const FIoHash PackageObjectHash;
const uint32 TotalBlobUploads;
std::atomic<uint32> SuccessfulBlobUploads;
std::atomic<uint32> PendingBlobUploads;
FOnCachePutPackageComplete OnComplete;
struct FCachePutRefResponse
{
FSharedString Name;
FCacheKey Key;
uint64 UserData = 0;
uint64 BytesSent = 0;
TConstArrayView<FIoHash> NeededBlobHashes;
EStatus Status = EStatus::Error;
};
using FOnCachePutRefComplete = TUniqueFunction<void(FCachePutRefResponse&& Response)>;
FPutPackageOp(
FHttpCacheStore& InCacheStore,
IRequestOwner& InOwner,
const FSharedString& InName,
const FCacheKey& InKey,
uint64 InUserData,
uint64 InBytesSent,
const FCbObject& InPackageObject,
const FIoHash& InPackageObjectHash,
uint32 InTotalBlobUploads,
FOnCachePutPackageComplete&& InOnComplete);
static void PutRefAsync(
FHttpCacheStore& CacheStore,
IRequestOwner& Owner,
const FSharedString& Name,
FCacheKey Key,
FCbObject Object,
FIoHash ObjectHash,
uint64 UserData,
bool bFinalize,
FOnCachePutRefComplete&& OnComplete);
static void OnPackagePutRefComplete(
FHttpCacheStore& CacheStore,
IRequestOwner& Owner,
const FSharedString& Name,
const FCacheKey& Key,
FCbPackage&& Package,
FCacheRecordPolicy Policy,
uint64 UserData,
FOnCachePutPackageComplete&& OnComplete,
FCachePutRefResponse&& Response);
FHttpRequest::ECompletionBehavior OnCompressedBlobUploadComplete(
FHttpRequest::Result HttpResult,
FHttpRequest* Request);
void OnPutRefFinalizationComplete(
FCachePutRefResponse&& Response);
FCachePutPackageResponse MakeResponse(uint64 InBytesSent, EStatus Status)
{
return FCachePutPackageResponse{ Name, Key, UserData, InBytesSent, Status };
};
};
//----------------------------------------------------------------------------------------------------------
// FHttpCacheStore::FGetRecordOp
//----------------------------------------------------------------------------------------------------------
class FHttpCacheStore::FGetRecordOp final : public FThreadSafeRefCountedObject
{
public:
/** Performs a multi-request operation for downloading a record. */
static void GetRecord(
FHttpCacheStore& CacheStore,
IRequestOwner& Owner,
const FSharedString& Name,
const FCacheKey& Key,
const FCacheRecordPolicy& Policy,
uint64 UserData,
TUniqueFunction<void(FCacheGetResponse&& Response, uint64 BytesReceived)>&& OnComplete);
struct FGetCachedDataBatchResponse
{
FSharedString Name;
FCacheKey Key;
int32 ValueIndex;
uint64 BytesReceived = 0;
FCompressedBuffer DataBuffer;
EStatus Status = EStatus::Error;
};
using FOnGetCachedDataBatchComplete = TUniqueFunction<void(FGetCachedDataBatchResponse&& Response)>;
/** Utility method for fetching a batch of value data. */
template<typename ValueType, typename ValueIdGetterType>
static void GetDataBatch(
FHttpCacheStore& CacheStore,
IRequestOwner& Owner,
FSharedString Name,
const FCacheKey& Key,
TConstArrayView<ValueType> Values,
ValueIdGetterType ValueIdGetter,
FOnGetCachedDataBatchComplete&& OnComplete);
private:
FHttpCacheStore& CacheStore;
IRequestOwner& Owner;
const FSharedString Name;
const FCacheKey Key;
const uint64 UserData;
std::atomic<uint64> BytesReceived;
TArray<FCompressedBuffer> FetchedBuffers;
const TArray<FValueWithId> RequiredGets;
const TArray<FValueWithId> RequiredHeads;
FCacheRecordBuilder RecordBuilder;
const uint32 TotalOperations;
std::atomic<uint32> SuccessfulOperations;
std::atomic<uint32> PendingOperations;
TUniqueFunction<void(FCacheGetResponse&& Response, uint64 BytesReceived)> OnComplete;
FGetRecordOp(
FHttpCacheStore& InCacheStore,
IRequestOwner& InOwner,
const FSharedString& InName,
const FCacheKey& InKey,
uint64 InUserData,
uint64 InBytesReceived,
TArray<FValueWithId>&& InRequiredGets,
TArray<FValueWithId>&& InRequiredHeads,
FCacheRecordBuilder&& InRecordBuilder,
TUniqueFunction<void(FCacheGetResponse&& Response, uint64 BytesReceived)>&& InOnComplete);
static void OnOnlyRecordComplete(
FHttpCacheStore& CacheStore,
IRequestOwner& Owner,
const FCacheRecordPolicy& Policy,
TUniqueFunction<void(FCacheGetResponse&& Response, uint64 BytesReceived)>&& OnComplete,
FGetCacheRecordOnlyResponse&& Response);
struct FCachedDataProbablyExistsBatchResponse
{
FSharedString Name;
FCacheKey Key;
int32 ValueIndex;
EStatus Status = EStatus::Error;
};
using FOnCachedDataProbablyExistsBatchComplete = TUniqueFunction<void(FCachedDataProbablyExistsBatchResponse&& Response)>;
void DataProbablyExistsBatch(
TConstArrayView<FValueWithId> Values,
FOnCachedDataProbablyExistsBatchComplete&& OnComplete);
void FinishDataStep(bool bSuccess, uint64 InBytesReceived);
};
void FHttpCacheStore::FPutPackageOp::PutPackage(
FHttpCacheStore& CacheStore,
IRequestOwner& Owner,
const FSharedString& Name,
FCacheKey Key,
FCbPackage&& Package,
FCacheRecordPolicy Policy,
uint64 UserData,
FOnCachePutPackageComplete&& OnComplete)
{
// TODO: Jupiter currently always overwrites. It doesn't have a "write if not present" feature (for records or attachments),
// but would require one to implement all policy correctly.
// Initial record upload
PutRefAsync(CacheStore, Owner, Name, Key, Package.GetObject(), Package.GetObjectHash(), UserData, false,
[&CacheStore, &Owner, Name = FSharedString(Name), Key, Package = MoveTemp(Package), Policy, UserData, OnComplete = MoveTemp(OnComplete)](FCachePutRefResponse&& Response) mutable
{
return OnPackagePutRefComplete(CacheStore, Owner, Name, Key, MoveTemp(Package), Policy, UserData, MoveTemp(OnComplete), MoveTemp(Response));
});
}
FHttpCacheStore::FPutPackageOp::FPutPackageOp(
FHttpCacheStore& InCacheStore,
IRequestOwner& InOwner,
const FSharedString& InName,
const FCacheKey& InKey,
uint64 InUserData,
uint64 InBytesSent,
const FCbObject& InPackageObject,
const FIoHash& InPackageObjectHash,
uint32 InTotalBlobUploads,
FOnCachePutPackageComplete&& InOnComplete)
: CacheStore(InCacheStore)
, Owner(InOwner)
, Name(InName)
, Key(InKey)
, UserData(InUserData)
, BytesSent(InBytesSent)
, PackageObject(InPackageObject)
, PackageObjectHash(InPackageObjectHash)
, TotalBlobUploads(InTotalBlobUploads)
, SuccessfulBlobUploads(0)
, PendingBlobUploads(InTotalBlobUploads)
, OnComplete(MoveTemp(InOnComplete))
{
}
void FHttpCacheStore::FPutPackageOp::PutRefAsync(
FHttpCacheStore& CacheStore,
IRequestOwner& Owner,
const FSharedString& Name,
FCacheKey Key,
FCbObject Object,
FIoHash ObjectHash,
uint64 UserData,
bool bFinalize,
FOnCachePutRefComplete&& OnComplete)
{
FString Bucket(Key.Bucket.ToString());
Bucket.ToLowerInline();
TStringBuilder<256> RefsUri;
RefsUri << "api/v1/refs/" << CacheStore.StructuredNamespace << "/" << Bucket << "/" << Key.Hash;
if (bFinalize)
{
RefsUri << "/finalize/" << ObjectHash;
}
FRequestPool* Pool = nullptr;
FHttpRequest* Request = CacheStore.WaitForHttpRequestForOwner<OperationCategory::Put>(Owner, bFinalize /* bUnboundedOverflow */, Pool);
auto OnHttpRequestComplete = [&CacheStore, &Owner, Name = FSharedString(Name), Key, Object, UserData, bFinalize, OnComplete = MoveTemp(OnComplete)]
(FHttpRequest::Result HttpResult, FHttpRequest* Request)
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_PutRefAsync_OnHttpRequestComplete);
if (Owner.IsCanceled())
{
OnComplete({ Name, Key, UserData, Request->GetBytesSent(), {}, EStatus::Canceled });
return FHttpRequest::ECompletionBehavior::Done;
}
int64 ResponseCode = Request->GetResponseCode();
if (FHttpRequest::IsSuccessResponse(ResponseCode))
{
TArray<FIoHash> NeededBlobHashes;
// Useful when debugging issues related to compressed/uncompressed blobs being returned from Jupiter
const bool bPutRefBlobsAlways = false;
if (bPutRefBlobsAlways && !bFinalize)
{
Object.IterateAttachments([&NeededBlobHashes](FCbFieldView AttachmentFieldView)
{
FIoHash AttachmentHash = AttachmentFieldView.AsHash();
if (!AttachmentHash.IsZero())
{
NeededBlobHashes.Add(AttachmentHash);
}
});
}
else if (TSharedPtr<FJsonObject> ResponseObject = Request->GetResponseAsJsonObject())
{
TArray<FString> NeedsArrayStrings;
ResponseObject->TryGetStringArrayField(TEXT("needs"), NeedsArrayStrings);
NeededBlobHashes.Reserve(NeedsArrayStrings.Num());
for (const FString& NeededString : NeedsArrayStrings)
{
FIoHash BlobHash;
LexFromString(BlobHash, *NeededString);
if (!BlobHash.IsZero())
{
NeededBlobHashes.Add(BlobHash);
}
}
}
OnComplete({ Name, Key, UserData, Request->GetBytesSent(), NeededBlobHashes, EStatus::Ok });
return FHttpRequest::ECompletionBehavior::Done;
}
if (!ShouldAbortForShutdown() && CacheStore.ShouldRetryOnError(HttpResult, ResponseCode) && ((Request->GetAttempts()+1) < UE_HTTPDDC_MAX_ATTEMPTS))
{
return FHttpRequest::ECompletionBehavior::Retry;
}
OnComplete({ Name, Key, UserData, Request->GetBytesSent(), {}, EStatus::Error });
return FHttpRequest::ECompletionBehavior::Done;
};
if (bFinalize)
{
Request->EnqueueAsyncUpload<FHttpRequest::Post>(Owner, Pool, *RefsUri, FSharedBuffer(), MoveTemp(OnHttpRequestComplete));
}
else
{
Request->SetHeader(TEXT("X-Jupiter-IoHash"), *WriteToString<48>(ObjectHash));
Request->EnqueueAsyncUpload<FHttpRequest::PutCompactBinary>(Owner, Pool, *RefsUri, Object.GetBuffer().ToShared(), MoveTemp(OnHttpRequestComplete));
}
}
void FHttpCacheStore::FPutPackageOp::OnPackagePutRefComplete(
FHttpCacheStore& CacheStore,
IRequestOwner& Owner,
const FSharedString& Name,
const FCacheKey& Key,
FCbPackage&& Package,
FCacheRecordPolicy Policy,
uint64 UserData,
FOnCachePutPackageComplete&& OnComplete,
FCachePutRefResponse&& Response)
{
if (Response.Status != EStatus::Ok)
{
if (Response.Status == EStatus::Error)
{
UE_LOG(LogDerivedDataCache, Log, TEXT("%s: Failed to put reference object for put of %s from '%s'"),
*CacheStore.GetName(), *WriteToString<96>(Response.Key), *Response.Name);
}
return OnComplete(FCachePutPackageResponse{ Name, Key, UserData, Response.BytesSent, Response.Status });
}
struct FCompressedBlobUpload
{
FIoHash Hash;
FSharedBuffer BlobBuffer;
FCompressedBlobUpload(const FIoHash& InHash, FSharedBuffer&& InBlobBuffer) : Hash(InHash), BlobBuffer(InBlobBuffer)
{
}
};
TArray<FCompressedBlobUpload> CompressedBlobUploads;
// TODO: blob uploading and finalization should be replaced with a single batch compressed blob upload endpoint in the future.
TStringBuilder<128> ExpectedHashes;
bool bExpectedHashesSerialized = false;
// Needed blob upload (if any missing)
for (const FIoHash& NeededBlobHash : Response.NeededBlobHashes)
{
if (const FCbAttachment* Attachment = Package.FindAttachment(NeededBlobHash))
{
FSharedBuffer TempBuffer;
if (Attachment->IsCompressedBinary())
{
TempBuffer = Attachment->AsCompressedBinary().GetCompressed().ToShared();
}
else if (Attachment->IsBinary())
{
TempBuffer = FValue::Compress(Attachment->AsCompositeBinary()).GetData().GetCompressed().ToShared();
}
else
{
TempBuffer = FValue::Compress(Attachment->AsObject().GetBuffer()).GetData().GetCompressed().ToShared();
}
CompressedBlobUploads.Emplace(NeededBlobHash, MoveTemp(TempBuffer));
}
else
{
if (!bExpectedHashesSerialized)
{
bool bFirstHash = true;
for (const FCbAttachment& PackageAttachment : Package.GetAttachments())
{
if (!bFirstHash)
{
ExpectedHashes << TEXT(", ");
}
ExpectedHashes << PackageAttachment.GetHash();
bFirstHash = false;
}
bExpectedHashesSerialized = true;
}
UE_LOG(LogDerivedDataCache, Log, TEXT("%s: Server reported needed hash '%s' that is outside the set of expected hashes (%s) for put of %s from '%s'"),
*CacheStore.GetName(), *WriteToString<96>(NeededBlobHash), ExpectedHashes.ToString(), *WriteToString<96>(Response.Key), *Response.Name);
}
}
if (CompressedBlobUploads.IsEmpty())
{
// No blobs need to be uploaded. No finalization necessary.
return OnComplete(FCachePutPackageResponse{ Name, Key, UserData, Response.BytesSent, EStatus::Ok });
}
// Having this be a ref ensures we don't have the op reach 0 ref count as we queue up multiple operations which MAY execute synchronously
TRefCountPtr<FPutPackageOp> PutPackageOp = new FPutPackageOp(
CacheStore,
Owner,
Response.Name,
Response.Key,
Response.UserData,
Response.BytesSent,
Package.GetObject(),
Package.GetObjectHash(),
(uint32)CompressedBlobUploads.Num(),
MoveTemp(OnComplete)
);
FRequestBarrier Barrier(Owner);
for (const FCompressedBlobUpload& CompressedBlobUpload : CompressedBlobUploads)
{
TStringBuilder<256> CompressedBlobsUri;
CompressedBlobsUri << "api/v1/compressed-blobs/" << CacheStore.StructuredNamespace << "/" << CompressedBlobUpload.Hash;
FRequestPool* Pool = nullptr;
FHttpRequest* Request = CacheStore.WaitForHttpRequestForOwner<OperationCategory::Put>(Owner, true /* bUnboundedOverflow */, Pool);
Request->EnqueueAsyncUpload<FHttpRequest::PutCompressedBlob>(Owner, Pool, *CompressedBlobsUri, CompressedBlobUpload.BlobBuffer,
[PutPackageOp](FHttpRequest::Result HttpResult, FHttpRequest* Request)
{
return PutPackageOp->OnCompressedBlobUploadComplete(HttpResult, Request);
});
}
}
FHttpRequest::ECompletionBehavior FHttpCacheStore::FPutPackageOp::OnCompressedBlobUploadComplete(
FHttpRequest::Result HttpResult,
FHttpRequest* Request)
{
int64 ResponseCode = Request->GetResponseCode();
bool bIsSuccessResponse = FHttpRequest::IsSuccessResponse(ResponseCode);
if (!bIsSuccessResponse && !ShouldAbortForShutdown() && !Owner.IsCanceled() && CacheStore.ShouldRetryOnError(HttpResult, ResponseCode) && ((Request->GetAttempts()+1) < UE_HTTPDDC_MAX_ATTEMPTS))
{
return FHttpRequest::ECompletionBehavior::Retry;
}
BytesSent.fetch_add(Request->GetBytesSent(), std::memory_order_relaxed);
if (bIsSuccessResponse)
{
SuccessfulBlobUploads.fetch_add(1, std::memory_order_relaxed);
}
if (PendingBlobUploads.fetch_sub(1, std::memory_order_relaxed) == 1)
{
if (Owner.IsCanceled())
{
OnComplete(MakeResponse(BytesSent.load(std::memory_order_relaxed), EStatus::Canceled));
return FHttpRequest::ECompletionBehavior::Done;
}
uint32 LocalSuccessfulBlobUploads = SuccessfulBlobUploads.load(std::memory_order_relaxed);
if (LocalSuccessfulBlobUploads == TotalBlobUploads)
{
// Perform finalization
PutRefAsync(CacheStore, Owner, Name, Key, PackageObject, PackageObjectHash, UserData, true,
[PutPackageOp = TRefCountPtr<FPutPackageOp>(this)](FCachePutRefResponse&& Response)
{
return PutPackageOp->OnPutRefFinalizationComplete(MoveTemp(Response));
});
}
else
{
uint32 FailedBlobUploads = (uint32)(TotalBlobUploads - LocalSuccessfulBlobUploads);
UE_LOG(LogDerivedDataCache, Log, TEXT("%s: Failed to put %d/%d blobs for put of %s from '%s'"),
*CacheStore.GetName(), FailedBlobUploads, TotalBlobUploads, *WriteToString<96>(Key), *Name);
OnComplete(MakeResponse(BytesSent.load(std::memory_order_relaxed), EStatus::Error));
}
}
return FHttpRequest::ECompletionBehavior::Done;
}
void FHttpCacheStore::FPutPackageOp::OnPutRefFinalizationComplete(
FCachePutRefResponse&& Response)
{
BytesSent.fetch_add(Response.BytesSent, std::memory_order_relaxed);
if (Response.Status == EStatus::Error)
{
UE_LOG(LogDerivedDataCache, Log, TEXT("%s: Failed to finalize reference object for put of %s from '%s'"),
*CacheStore.GetName(), *WriteToString<96>(Key), *Name);
}
return OnComplete(MakeResponse(BytesSent.load(std::memory_order_relaxed), Response.Status));
}
void FHttpCacheStore::FGetRecordOp::GetRecord(
FHttpCacheStore& CacheStore,
IRequestOwner& Owner,
const FSharedString& Name,
const FCacheKey& Key,
const FCacheRecordPolicy& Policy,
uint64 UserData,
TUniqueFunction<void(FCacheGetResponse&& Response, uint64 BytesReceived)>&& OnComplete)
{
CacheStore.GetCacheRecordOnlyAsync(Owner, Name, Key, Policy, UserData, [&CacheStore, &Owner, Policy = FCacheRecordPolicy(Policy), OnComplete = MoveTemp(OnComplete)](FGetCacheRecordOnlyResponse&& Response) mutable
{
OnOnlyRecordComplete(CacheStore, Owner, Policy, MoveTemp(OnComplete), MoveTemp(Response));
});
}
template<typename ValueType, typename ValueIdGetterType>
void FHttpCacheStore::FGetRecordOp::GetDataBatch(
FHttpCacheStore& CacheStore,
IRequestOwner& Owner,
FSharedString Name,
const FCacheKey& Key,
TConstArrayView<ValueType> Values,
ValueIdGetterType ValueIdGetter,
FOnGetCachedDataBatchComplete&& OnComplete)
{
if (Values.IsEmpty())
{
return;
}
FRequestBarrier Barrier(Owner);
TRefCountedUniqueFunction<FOnGetCachedDataBatchComplete>* CompletionFunction = new TRefCountedUniqueFunction<FOnGetCachedDataBatchComplete>(MoveTemp(OnComplete));
TRefCountPtr<TRefCountedUniqueFunction<FOnGetCachedDataBatchComplete>> BatchOnCompleteRef(CompletionFunction);
for (int32 ValueIndex = 0; ValueIndex < Values.Num(); ++ValueIndex)
{
const ValueType& Value = Values[ValueIndex];
const FIoHash& RawHash = Value.GetRawHash();
FRequestPool* Pool = nullptr;
FHttpRequest* Request = CacheStore.WaitForHttpRequestForOwner<OperationCategory::Get>(Owner, true /* bUnboundedOverflow */, Pool);
auto OnHttpRequestComplete = [&CacheStore, &Owner, Name = FSharedString(Name), Key = FCacheKey(Key), ValueIndex, Value = Value.RemoveData(), ValueIdGetter, OnCompletePtr = TRefCountPtr<TRefCountedUniqueFunction<FOnGetCachedDataBatchComplete>>(CompletionFunction)]
(FHttpRequest::Result HttpResult, FHttpRequest* Request)
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_GetDataBatch_OnHttpRequestComplete);
int64 ResponseCode = Request->GetResponseCode();
bool bHit = false;
FCompressedBuffer CompressedBuffer;
if (FHttpRequest::IsSuccessResponse(ResponseCode))
{
FString ReceivedContentType;
if (Request->GetHeader("Content-Type", ReceivedContentType))
{
if (ReceivedContentType == TEXT("application/x-ue-comp"))
{
CompressedBuffer = FCompressedBuffer::FromCompressed(Request->MoveResponseBufferToShared());
bHit = true;
}
else if (ReceivedContentType == TEXT("application/octet-stream"))
{
CompressedBuffer = FValue::Compress(Request->MoveResponseBufferToShared()).GetData();
bHit = true;
}
else
{
bHit = false;
}
}
else
{
CompressedBuffer = FCompressedBuffer::FromCompressed(Request->MoveResponseBufferToShared());
bHit = true;
}
}
if (!ShouldAbortForShutdown() && !Owner.IsCanceled() && CacheStore.ShouldRetryOnError(HttpResult, ResponseCode) && ((Request->GetAttempts()+1) < UE_HTTPDDC_MAX_ATTEMPTS))
{
return FHttpRequest::ECompletionBehavior::Retry;
}
if (!bHit)
{
UE_LOG(LogDerivedDataCache, Verbose,
TEXT("%s: Cache miss with missing value %s with hash %s for %s from '%s'"),
*CacheStore.GetName(), *ValueIdGetter(Value), *WriteToString<48>(Value.GetRawHash()), *WriteToString<96>(Key),
*Name);
OnCompletePtr->GetFunction()({ Name, Key, ValueIndex, Request->GetBytesReceived(), {}, EStatus::Error });
}
else if (CompressedBuffer.GetRawHash() != Value.GetRawHash())
{
UE_LOG(LogDerivedDataCache, Display,
TEXT("%s: Cache miss with corrupted value %s with hash %s for %s from '%s'"),
*CacheStore.GetName(), *ValueIdGetter(Value), *WriteToString<48>(Value.GetRawHash()),
*WriteToString<96>(Key), *Name);
OnCompletePtr->GetFunction()({ Name, Key, ValueIndex, Request->GetBytesReceived(), {}, EStatus::Error });
}
else
{
OnCompletePtr->GetFunction()({ Name, Key, ValueIndex, Request->GetBytesReceived(), MoveTemp(CompressedBuffer), EStatus::Ok });
}
return FHttpRequest::ECompletionBehavior::Done;
};
TStringBuilder<256> CompressedBlobsUri;
CompressedBlobsUri << "api/v1/compressed-blobs/" << CacheStore.StructuredNamespace << "/" << RawHash;
Request->SetHeader(TEXT("Accept"), TEXT("*/*"));
Request->EnqueueAsyncDownload(Owner, Pool, *CompressedBlobsUri, MoveTemp(OnHttpRequestComplete), { 404 });
}
}
FHttpCacheStore::FGetRecordOp::FGetRecordOp(
FHttpCacheStore& InCacheStore,
IRequestOwner& InOwner,
const FSharedString& InName,
const FCacheKey& InKey,
uint64 InUserData,
uint64 InBytesReceived,
TArray<FValueWithId>&& InRequiredGets,
TArray<FValueWithId>&& InRequiredHeads,
FCacheRecordBuilder&& InRecordBuilder,
TUniqueFunction<void(FCacheGetResponse&& Response, uint64 BytesReceived)>&& InOnComplete)
: CacheStore(InCacheStore)
, Owner(InOwner)
, Name(InName)
, Key(InKey)
, UserData(InUserData)
, BytesReceived(InBytesReceived)
, RequiredGets(MoveTemp(InRequiredGets))
, RequiredHeads(MoveTemp(InRequiredHeads))
, RecordBuilder(MoveTemp(InRecordBuilder))
, TotalOperations(RequiredGets.Num() + RequiredHeads.Num())
, SuccessfulOperations(0)
, PendingOperations(TotalOperations)
, OnComplete(MoveTemp(InOnComplete))
{
FetchedBuffers.AddDefaulted(RequiredGets.Num());
}
void FHttpCacheStore::FGetRecordOp::OnOnlyRecordComplete(
FHttpCacheStore& CacheStore,
IRequestOwner& Owner,
const FCacheRecordPolicy& Policy,
TUniqueFunction<void(FCacheGetResponse&& Response, uint64 BytesReceived)>&& OnComplete,
FGetCacheRecordOnlyResponse&& Response)
{
FCacheRecordBuilder RecordBuilder(Response.Key);
if (Response.Status != EStatus::Ok)
{
return OnComplete({ Response.Name, RecordBuilder.Build(), Response.UserData, Response.Status }, Response.BytesReceived);
}
if (!EnumHasAnyFlags(Policy.GetRecordPolicy(), ECachePolicy::SkipMeta))
{
RecordBuilder.SetMeta(FCbObject(Response.Record.Get().GetMeta()));
}
// TODO: There is not currently a batched GET endpoint for Jupiter. Once there is, all payload data should be fetched in one call.
// In the meantime, we try to keep the code structured in a way that is friendly to future batching of GETs.
TArray<FValueWithId> RequiredGets;
TArray<FValueWithId> RequiredHeads;
for (FValueWithId Value : Response.Record.Get().GetValues())
{
const ECachePolicy ValuePolicy = Policy.GetValuePolicy(Value.GetId());
if (IsValueDataReady(Value, ValuePolicy))
{
RecordBuilder.AddValue(MoveTemp(Value));
}
else
{
if (EnumHasAnyFlags(ValuePolicy, ECachePolicy::SkipData))
{
RequiredHeads.Emplace(Value);
}
else
{
RequiredGets.Emplace(Value);
}
}
}
if (RequiredGets.IsEmpty() && RequiredHeads.IsEmpty())
{
return OnComplete({ Response.Name, RecordBuilder.Build(), Response.UserData, Response.Status }, Response.BytesReceived);
}
// Having this be a ref ensures we don't have the op reach 0 ref count in between the start of the exist batch operation and the get batch operation
TRefCountPtr<FGetRecordOp> GetRecordOp = new FGetRecordOp(
CacheStore,
Owner,
Response.Name,
Response.Key,
Response.UserData,
Response.BytesReceived,
MoveTemp(RequiredGets),
MoveTemp(RequiredHeads),
MoveTemp(RecordBuilder),
MoveTemp(OnComplete)
);
auto IdGetter = [](const FValueWithId& Value)
{
return FString(WriteToString<16>(Value.GetId()));
};
{
FRequestBarrier Barrier(Owner);
GetRecordOp->DataProbablyExistsBatch(GetRecordOp->RequiredHeads, [GetRecordOp](FCachedDataProbablyExistsBatchResponse&& Response)
{
GetRecordOp->FinishDataStep(Response.Status == EStatus::Ok, 0);
});
GetDataBatch<FValueWithId>(CacheStore, Owner, Response.Name, Response.Key, GetRecordOp->RequiredGets, IdGetter, [GetRecordOp](FGetCachedDataBatchResponse&& Response)
{
GetRecordOp->FetchedBuffers[Response.ValueIndex] = MoveTemp(Response.DataBuffer);
GetRecordOp->FinishDataStep(Response.Status == EStatus::Ok, Response.BytesReceived);
});
}
}
void FHttpCacheStore::FGetRecordOp::DataProbablyExistsBatch(
TConstArrayView<FValueWithId> Values,
FOnCachedDataProbablyExistsBatchComplete&& InOnComplete)
{
if (Values.IsEmpty())
{
return;
}
FRequestPool* Pool = nullptr;
FHttpRequest* Request = CacheStore.WaitForHttpRequestForOwner<OperationCategory::Get>(Owner, true /* bUnboundedOverflow */, Pool);
TStringBuilder<256> CompressedBlobsUri;
CompressedBlobsUri << "api/v1/compressed-blobs/" << CacheStore.StructuredNamespace << "/exists?";
bool bFirstItem = true;
for (const FValueWithId& Value : Values)
{
if (!bFirstItem)
{
CompressedBlobsUri << "&";
}
CompressedBlobsUri << "id=" << Value.GetRawHash();
bFirstItem = false;
}
auto OnHttpRequestComplete = [this, Values = TArray<FValueWithId>(Values), InOnComplete = MoveTemp(InOnComplete)](FHttpRequest::Result HttpResult, FHttpRequest* Request)
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_DataProbablyExistsBatch_OnHttpRequestComplete);
int64 ResponseCode = Request->GetResponseCode();
if (FHttpRequest::IsSuccessResponse(ResponseCode))
{
if (TSharedPtr<FJsonObject> ResponseObject = Request->GetResponseAsJsonObject())
{
TArray<FString> NeedsArrayStrings;
if (ResponseObject->TryGetStringArrayField(TEXT("needs"), NeedsArrayStrings))
{
if (NeedsArrayStrings.IsEmpty())
{
for (int32 ValueIndex = 0; ValueIndex < Values.Num(); ++ValueIndex)
{
const FValueWithId& Value = Values[ValueIndex];
UE_LOG(LogDerivedDataCache, Verbose,
TEXT("%s: Cache exists miss with missing value %s with hash %s for %s from '%s'"),
*CacheStore.GetName(), *WriteToString<16>(Value.GetId()), *WriteToString<48>(Value.GetRawHash()), *WriteToString<96>(Key),
*Name);
InOnComplete({ Name, Key, ValueIndex, EStatus::Error });
}
return FHttpRequest::ECompletionBehavior::Done;
}
}
TBitArray<> ResultStatus(true, Values.Num());
for (const FString& NeedsString : NeedsArrayStrings)
{
FIoHash NeedHash;
LexFromString(NeedHash, *NeedsString);
for (int32 ValueIndex = 0; ValueIndex < Values.Num(); ++ValueIndex)
{
const FValueWithId& Value = Values[ValueIndex];
if (ResultStatus[ValueIndex] && NeedHash == Value.GetRawHash())
{
ResultStatus[ValueIndex] = false;
break;
}
}
}
for (int32 ValueIndex = 0; ValueIndex < Values.Num(); ++ValueIndex)
{
const FValueWithId& Value = Values[ValueIndex];
if (ResultStatus[ValueIndex])
{
UE_LOG(LogDerivedDataCache, VeryVerbose,
TEXT("%s: Cache exists hit for %s with hash %s for %s from '%s'"),
*CacheStore.GetName(), *WriteToString<16>(Value.GetId()), *WriteToString<48>(Value.GetRawHash()), *WriteToString<96>(Key),
*Name);
InOnComplete({ Name, Key, ValueIndex, EStatus::Ok });
}
else
{
UE_LOG(LogDerivedDataCache, Verbose,
TEXT("%s: Cache exists miss with missing value %s with hash %s for %s from '%s'"),
*CacheStore.GetName(), *WriteToString<16>(Value.GetId()), *WriteToString<48>(Value.GetRawHash()), *WriteToString<96>(Key),
*Name);
InOnComplete({ Name, Key, ValueIndex, EStatus::Error });
}
}
}
else
{
for (int32 ValueIndex = 0; ValueIndex < Values.Num(); ++ValueIndex)
{
UE_LOG(LogDerivedDataCache, Log,
TEXT("%s: Cache exists returned invalid results."),
*CacheStore.GetName());
InOnComplete({ Name, Key, ValueIndex, EStatus::Error });
}
}
return FHttpRequest::ECompletionBehavior::Done;
}
if (!ShouldAbortForShutdown() && !Owner.IsCanceled() && CacheStore.ShouldRetryOnError(HttpResult, ResponseCode) && ((Request->GetAttempts()+1) < UE_HTTPDDC_MAX_ATTEMPTS))
{
return FHttpRequest::ECompletionBehavior::Retry;
}
for (int32 ValueIndex = 0; ValueIndex < Values.Num(); ++ValueIndex)
{
const FValueWithId& Value = Values[ValueIndex];
UE_LOG(LogDerivedDataCache, Verbose, TEXT("%s: Cache miss with failed HTTP request for %s from '%s'"),
*CacheStore.GetName(), *WriteToString<96>(Key), *Name);
InOnComplete({Name, Key, ValueIndex, EStatus::Error});
}
return FHttpRequest::ECompletionBehavior::Done;
};
FSharedBuffer DummyBuffer;
Request->EnqueueAsyncUpload<FHttpRequest::Post>(Owner, Pool, *CompressedBlobsUri, DummyBuffer, MoveTemp(OnHttpRequestComplete));
}
void FHttpCacheStore::FGetRecordOp::FinishDataStep(bool bSuccess, uint64 InBytesReceived)
{
BytesReceived.fetch_add(InBytesReceived, std::memory_order_relaxed);
if (bSuccess)
{
SuccessfulOperations.fetch_add(1, std::memory_order_relaxed);
}
if (PendingOperations.fetch_sub(1, std::memory_order_acq_rel) == 1)
{
EStatus Status = EStatus::Error;
uint32 LocalSuccessfulOperations = SuccessfulOperations.load(std::memory_order_relaxed);
if (LocalSuccessfulOperations == TotalOperations)
{
for (int32 Index = 0; Index < RequiredHeads.Num(); ++Index)
{
RecordBuilder.AddValue(RequiredHeads[Index].RemoveData());
}
for (int32 Index = 0; Index < RequiredGets.Num(); ++Index)
{
RecordBuilder.AddValue(FValueWithId(RequiredGets[Index].GetId(), FetchedBuffers[Index]));
}
Status = EStatus::Ok;
}
OnComplete({Name, RecordBuilder.Build(), UserData, Status}, BytesReceived.load(std::memory_order_relaxed));
}
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
FHttpCacheStore::FHttpCacheStore(
const TCHAR* InServiceUrl,
bool bResolveHostCanonicalName,
const TCHAR* InNamespace,
const TCHAR* InStructuredNamespace,
const TCHAR* InOAuthProvider,
const TCHAR* InOAuthClientId,
const TCHAR* InOAuthSecret,
const TCHAR* InOAuthScope,
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
const EBackendLegacyMode InLegacyMode,
const bool bInReadOnly)
: Domain(InServiceUrl)
, EffectiveDomain(InServiceUrl)
, Namespace(InNamespace)
, StructuredNamespace(InStructuredNamespace)
, DefaultBucket(TEXT("default"))
, OAuthProvider(InOAuthProvider)
, OAuthClientId(InOAuthClientId)
, OAuthSecret(InOAuthSecret)
, OAuthScope(InOAuthScope)
, Access(nullptr)
, bIsUsable(false)
, bReadOnly(bInReadOnly)
, FailedLoginAttempts(0)
, SpeedClass(ESpeedClass::Slow)
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
, LegacyMode(InLegacyMode)
{
#if WITH_DATAREQUEST_HELPER
FDataRequestHelper::StaticInitialize();
#endif
if (IsServiceReady() && AcquireAccessToken())
{
FString OriginalDomainPrefix;
TAnsiStringBuilder<64> DomainResolveName;
if (Domain.StartsWith(TEXT("http://")))
{
DomainResolveName << Domain.RightChop(7);
OriginalDomainPrefix = TEXT("http://");
}
else if (Domain.StartsWith(TEXT("https://")))
{
DomainResolveName << Domain.RightChop(8);
OriginalDomainPrefix = TEXT("https://");
}
else
{
DomainResolveName << Domain;
}
addrinfo* AddrResult = nullptr;
addrinfo AddrHints;
FMemory::Memset(&AddrHints, 0, sizeof(AddrHints));
AddrHints.ai_flags = AI_CANONNAME;
AddrHints.ai_family = AF_UNSPEC;
if (bResolveHostCanonicalName && !::getaddrinfo(*DomainResolveName, nullptr, &AddrHints, &AddrResult))
{
if (AddrResult->ai_canonname)
{
// Swap the domain with a canonical name from DNS so that if we are using regional redirection, we pin to a region.
EffectiveDomain = OriginalDomainPrefix + ANSI_TO_TCHAR(AddrResult->ai_canonname);
UE_LOG(LogDerivedDataCache, Display,
TEXT("%s: Pinned to %s based on DNS canonical name."),
*Domain, *EffectiveDomain);
}
else
{
EffectiveDomain = Domain;
}
::freeaddrinfo(AddrResult);
}
else
{
EffectiveDomain = Domain;
}
GetRequestPools[0] = MakeUnique<FRequestPool>(*Domain, *EffectiveDomain, Access.Get(), UE_HTTPDDC_GET_REQUEST_POOL_SIZE);
GetRequestPools[1] = MakeUnique<FRequestPool>(*Domain, *EffectiveDomain, Access.Get(), UE_HTTPDDC_GET_REQUEST_POOL_SIZE);
PutRequestPools[0] = MakeUnique<FRequestPool>(*Domain, *EffectiveDomain, Access.Get(), UE_HTTPDDC_PUT_REQUEST_POOL_SIZE);
PutRequestPools[1] = MakeUnique<FRequestPool>(*Domain, *EffectiveDomain, Access.Get(), UE_HTTPDDC_PUT_REQUEST_POOL_SIZE);
// Allowing the non-blocking requests to overflow to double their pre-allocated size before we start waiting for one to free up.
NonBlockingRequestPools = MakeUnique<FRequestPool>(*Domain, *EffectiveDomain, Access.Get(), UE_HTTPDDC_NONBLOCKING_REQUEST_POOL_SIZE, UE_HTTPDDC_NONBLOCKING_REQUEST_POOL_SIZE);
bIsUsable = true;
}
AnyInstance = this;
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
FHttpCacheStore::~FHttpCacheStore()
{
if (AnyInstance == this)
{
AnyInstance = nullptr;
}
#if WITH_DATAREQUEST_HELPER
FDataRequestHelper::StaticShutdown();
#endif
FHttpSharedData::bAsyncThreadShutdownRequested.store(true, std::memory_order_relaxed);
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
FString FHttpCacheStore::GetName() const
{
return Domain;
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
TBitArray<> FHttpCacheStore::TryToPrefetch(TConstArrayView<FString> CacheKeys)
{
return CachedDataProbablyExistsBatch(CacheKeys);
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
bool FHttpCacheStore::WouldCache(const TCHAR* CacheKey, TArrayView<const uint8> InData)
{
return IsWritable();
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
FHttpCacheStore::ESpeedClass FHttpCacheStore::GetSpeedClass() const
{
return SpeedClass;
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
bool FHttpCacheStore::ApplyDebugOptions(FBackendDebugOptions& InOptions)
{
DebugOptions = InOptions;
return true;
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
bool FHttpCacheStore::IsServiceReady()
{
FHttpRequest Request(*Domain, *Domain, nullptr, false);
FHttpRequest::Result Result = Request.PerformBlockingDownload(TEXT("health/ready"), nullptr);
if (Result == FHttpRequest::Success && Request.GetResponseCode() == 200)
{
UE_LOG(LogDerivedDataCache, Display, TEXT("%s: HTTP DDC service status: %s."), *Request.GetName(), *Request.GetResponseAsString());
return true;
}
else
{
UE_LOG(LogDerivedDataCache, Warning, TEXT("%s: Unable to reach HTTP DDC service at %s. Status: %d . Response: %s"), *Request.GetName(), *Domain, Request.GetResponseCode(), *Request.GetResponseAsString());
}
return false;
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
bool FHttpCacheStore::AcquireAccessToken()
{
if (Domain.StartsWith(TEXT("http://localhost")))
{
UE_LOG(LogDerivedDataCache, Log, TEXT("Connecting to a local host '%s', so skipping authorization"), *Domain);
return true;
}
// Avoid spamming the this if the service is down
if (FailedLoginAttempts > UE_HTTPDDC_MAX_FAILED_LOGIN_ATTEMPTS)
{
return false;
}
ensureMsgf(OAuthProvider.StartsWith(TEXT("http://")) || OAuthProvider.StartsWith(TEXT("https://")),
TEXT("The OAuth provider %s is not valid. Needs to be a fully qualified url."),
*OAuthProvider
);
// In case many requests wants to update the token at the same time
// get the current serial while we wait to take the CS.
const uint32 WantsToUpdateTokenSerial = Access.IsValid() ? Access->GetSerial() : 0u;
{
FScopeLock Lock(&AccessCs);
// Check if someone has beaten us to update the token, then it
// should now be valid.
if (Access.IsValid() && Access->GetSerial() > WantsToUpdateTokenSerial)
{
return true;
}
const uint32 SchemeEnd = OAuthProvider.Find(TEXT("://")) + 3;
const uint32 DomainEnd = OAuthProvider.Find(TEXT("/"), ESearchCase::CaseSensitive, ESearchDir::FromStart, SchemeEnd);
FString AuthDomain(DomainEnd, *OAuthProvider);
FString Uri(*OAuthProvider + DomainEnd + 1);
FHttpRequest Request(*AuthDomain, *AuthDomain, nullptr, false);
FHttpRequest::Result Result = FHttpRequest::Success;
if (OAuthProvider.StartsWith(TEXT("http://localhost")))
{
// Simple unauthenticated call to a local endpoint that mimics
// the result from an OIDC provider.
Result = Request.PerformBlockingDownload(*Uri, nullptr);
}
else
{
// Needs client id and secret to authenticate with an actual OIDC provider.
// If contents of the secret string is a file path, resolve and read form data.
if (OAuthSecret.StartsWith(TEXT("file://")))
{
FString FilePath = OAuthSecret.Mid(7, OAuthSecret.Len() - 7);
FString SecretFileContents;
if (FFileHelper::LoadFileToString(SecretFileContents, *FilePath))
{
// Overwrite the filepath with the actual content.
OAuthSecret = SecretFileContents;
}
else
{
UE_LOG(LogDerivedDataCache, Warning, TEXT("%s: Failed to read OAuth form data file (%s)."), *Request.GetName(), *OAuthSecret);
return false;
}
}
FString OAuthFormData = FString::Printf(
TEXT("client_id=%s&scope=%s&grant_type=client_credentials&client_secret=%s"),
*OAuthClientId,
*OAuthScope,
*OAuthSecret
);
TArray<uint8> FormData;
auto OAuthFormDataUTF8 = FTCHARToUTF8(*OAuthFormData);
FormData.Append((uint8*)OAuthFormDataUTF8.Get(), OAuthFormDataUTF8.Length());
Result = Request.PerformBlockingUpload<FHttpRequest::Post>(*Uri, MakeArrayView(FormData));
}
if (Result == FHttpRequest::Success && Request.GetResponseCode() == 200)
{
TSharedPtr<FJsonObject> ResponseObject = Request.GetResponseAsJsonObject();
if (ResponseObject)
{
FString AccessTokenString;
int32 ExpiryTimeSeconds = 0;
int32 CurrentTimeSeconds = int32(FPlatformTime::ToSeconds(FPlatformTime::Cycles()));
if (ResponseObject->TryGetStringField(TEXT("access_token"), AccessTokenString) &&
ResponseObject->TryGetNumberField(TEXT("expires_in"), ExpiryTimeSeconds))
{
if (!Access)
{
Access = MakeUnique<FHttpAccessToken>();
}
Access->SetHeader(*AccessTokenString);
UE_LOG(LogDerivedDataCache, Display, TEXT("%s: Logged in to HTTP DDC services. Expires in %d seconds."), *Request.GetName(), ExpiryTimeSeconds);
//Schedule a refresh of the token ahead of expiry time (this will not work in commandlets)
if (!IsRunningCommandlet())
{
FTSTicker::GetCoreTicker().AddTicker(FTickerDelegate::CreateLambda(
[this](float DeltaTime)
{
this->AcquireAccessToken();
return false;
}
), ExpiryTimeSeconds - 20.0f);
}
// Reset failed login attempts, the service is indeed alive.
FailedLoginAttempts = 0;
return true;
}
}
}
else
{
UE_LOG(LogDerivedDataCache, Warning, TEXT("%s: Failed to log in to HTTP services. Server responed with code %d."), *Request.GetName(), Request.GetResponseCode());
FailedLoginAttempts++;
}
}
return false;
}
bool FHttpCacheStore::ShouldRetryOnError(FHttpRequest::Result Result, int64 ResponseCode)
{
if (Result == FHttpRequest::FailedTimeout)
{
return true;
}
// Access token might have expired, request a new token and try again.
if (ResponseCode == 401 && AcquireAccessToken())
{
return true;
}
// Too many requests, make a new attempt
if (ResponseCode == 429)
{
return true;
}
return false;
}
void FHttpCacheStore::GetCacheRecordOnlyAsync(
IRequestOwner& Owner,
const FSharedString& Name,
const FCacheKey& Key,
const FCacheRecordPolicy& Policy,
uint64 UserData,
FOnGetCacheRecordOnlyComplete&& OnComplete)
{
auto MakeResponse = [Name = FSharedString(Name), Key, UserData](uint64 BytesReceived, EStatus Status)
{
return FGetCacheRecordOnlyResponse{ Name, Key, UserData, BytesReceived, {}, Status };
};
if (!IsUsable())
{
UE_LOG(LogDerivedDataCache, VeryVerbose,
TEXT("%s: Skipped get of %s from '%s' because this cache store is not available"),
*GetName(), *WriteToString<96>(Key), *Name);
return OnComplete(MakeResponse(0, EStatus::Error));
}
// Skip the request if querying the cache is disabled.
const ECachePolicy QueryPolicy = SpeedClass == ESpeedClass::Local
? ECachePolicy::QueryLocal : ECachePolicy::QueryRemote;
if (!EnumHasAnyFlags(Policy.GetRecordPolicy(), QueryPolicy))
{
UE_LOG(LogDerivedDataCache, VeryVerbose, TEXT("%s: Skipped get of %s from '%s' due to cache policy"),
*GetName(), *WriteToString<96>(Key), *Name);
return OnComplete(MakeResponse(0, EStatus::Error));
}
if (DebugOptions.ShouldSimulateGetMiss(Key))
{
UE_LOG(LogDerivedDataCache, Verbose, TEXT("%s: Simulated miss for get of %s from '%s'"),
*GetName(), *WriteToString<96>(Key), *Name);
return OnComplete(MakeResponse(0, EStatus::Error));
}
FString Bucket(Key.Bucket.ToString());
Bucket.ToLowerInline();
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
TStringBuilder<256> RefsUri;
RefsUri << "api/v1/refs/" << StructuredNamespace << "/" << Bucket << "/" << Key.Hash;
FRequestPool* Pool = nullptr;
FHttpRequest* Request = WaitForHttpRequestForOwner<OperationCategory::Get>(Owner, false /* bUnboundedOverflow */, Pool);
auto OnHttpRequestComplete = [this, &Owner, Name = FSharedString(Name), Key, UserData, OnComplete = MoveTemp(OnComplete)]
(FHttpRequest::Result HttpResult, FHttpRequest* Request)
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_GetCacheRecordOnlyAsync_OnHttpRequestComplete);
int64 ResponseCode = Request->GetResponseCode();
if (FHttpRequest::IsSuccessResponse(ResponseCode))
{
FSharedBuffer ResponseBuffer = Request->MoveResponseBufferToShared();
if (ValidateCompactBinary(ResponseBuffer, ECbValidateMode::Default) != ECbValidateError::None)
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
{
UE_LOG(LogDerivedDataCache, Log, TEXT("%s: Cache miss with invalid package for %s from '%s'"),
*GetName(), *WriteToString<96>(Key), *Name);
OnComplete({ Name, Key, UserData, Request->GetBytesReceived(), {}, EStatus::Error });
return FHttpRequest::ECompletionBehavior::Done;
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
}
FOptionalCacheRecord Record = FCacheRecord::Load(FCbPackage(FCbObject(ResponseBuffer)));
if (Record.IsNull())
{
UE_LOG(LogDerivedDataCache, Log, TEXT("%s: Cache miss with record load failure for %s from '%s'"),
*GetName(), *WriteToString<96>(Key), *Name);
OnComplete({ Name, Key, UserData, Request->GetBytesReceived(), {}, EStatus::Error });
return FHttpRequest::ECompletionBehavior::Done;
}
OnComplete({ Name, Key, UserData, Request->GetBytesReceived(), MoveTemp(Record), EStatus::Ok });
return FHttpRequest::ECompletionBehavior::Done;
}
if (!ShouldAbortForShutdown() && !Owner.IsCanceled() && ShouldRetryOnError(HttpResult, ResponseCode) && ((Request->GetAttempts()+1) < UE_HTTPDDC_MAX_ATTEMPTS))
{
return FHttpRequest::ECompletionBehavior::Retry;
}
UE_LOG(LogDerivedDataCache, Verbose, TEXT("%s: Cache miss with missing package for %s from '%s'"),
*GetName(), *WriteToString<96>(Key), *Name);
OnComplete({ Name, Key, UserData, Request->GetBytesReceived(), {}, EStatus::Error });
return FHttpRequest::ECompletionBehavior::Done;
};
Request->SetHeader(TEXT("Accept"), TEXT("application/x-ue-cb"));
Request->EnqueueAsyncDownload(Owner, Pool, *RefsUri, MoveTemp(OnHttpRequestComplete), { 401, 404 });
}
void FHttpCacheStore::PutCacheRecordAsync(
IRequestOwner& Owner,
const FSharedString& Name,
const FCacheRecord& Record,
const FCacheRecordPolicy& Policy,
uint64 UserData,
TUniqueFunction<void(FCachePutResponse&& Response, uint64 BytesSent)>&& OnComplete)
{
const FCacheKey& Key = Record.GetKey();
auto MakeResponse = [Name = FSharedString(Name), Key = FCacheKey(Key), UserData](EStatus Status)
{
return FCachePutResponse{ Name, Key, UserData, Status };
};
if (!IsWritable())
{
UE_LOG(LogDerivedDataCache, VeryVerbose,
TEXT("%s: Skipped put of %s from '%s' because this cache store is read-only"),
*GetName(), *WriteToString<96>(Key), *Name);
return OnComplete(MakeResponse(EStatus::Error), 0);
}
// Skip the request if storing to the cache is disabled.
const ECachePolicy RecordPolicy = Policy.GetRecordPolicy();
const ECachePolicy StoreFlag = SpeedClass == ESpeedClass::Local ? ECachePolicy::StoreLocal : ECachePolicy::StoreRemote;
if (!EnumHasAnyFlags(RecordPolicy, StoreFlag))
{
UE_LOG(LogDerivedDataCache, VeryVerbose, TEXT("%s: Skipped put of %s from '%s' due to cache policy"),
*GetName(), *WriteToString<96>(Key), *Name);
return OnComplete(MakeResponse(EStatus::Error), 0);
}
if (DebugOptions.ShouldSimulatePutMiss(Key))
{
UE_LOG(LogDerivedDataCache, Verbose, TEXT("%s: Simulated miss for put of %s from '%s'"),
*GetName(), *WriteToString<96>(Key), *Name);
return OnComplete(MakeResponse(EStatus::Error), 0);
}
// TODO: Jupiter currently always overwrites. It doesn't have a "write if not present" feature (for records or attachments),
// but would require one to implement all policy correctly.
FString Bucket(Key.Bucket.ToString());
Bucket.ToLowerInline();
FCbPackage Package = Record.Save();
FPutPackageOp::PutPackage(*this, Owner, Name, Key, MoveTemp(Package), Policy, UserData, [MakeResponse = MoveTemp(MakeResponse), OnComplete = MoveTemp(OnComplete)](FPutPackageOp::FCachePutPackageResponse&& Response)
{
OnComplete(MakeResponse(Response.Status), Response.BytesSent);
});
}
void FHttpCacheStore::PutCacheValueAsync(
IRequestOwner& Owner,
const FSharedString& Name,
const FCacheKey& Key,
const FValue& Value,
const ECachePolicy Policy,
uint64 UserData,
TUniqueFunction<void(FCachePutValueResponse&& Response, uint64 BytesSent)>&& OnComplete)
{
auto MakeResponse = [Name = FSharedString(Name), Key = FCacheKey(Key), UserData](EStatus Status)
{
return FCachePutValueResponse{ Name, Key, UserData, Status };
};
if (!IsWritable())
{
UE_LOG(LogDerivedDataCache, VeryVerbose,
TEXT("%s: Skipped put of %s from '%s' because this cache store is read-only"),
*GetName(), *WriteToString<96>(Key), *Name);
return OnComplete(MakeResponse(EStatus::Error), 0);
}
// Skip the request if storing to the cache is disabled.
const ECachePolicy StoreFlag = SpeedClass == ESpeedClass::Local ? ECachePolicy::StoreLocal : ECachePolicy::StoreRemote;
if (!EnumHasAnyFlags(Policy, StoreFlag))
{
UE_LOG(LogDerivedDataCache, VeryVerbose, TEXT("%s: Skipped put of %s from '%s' due to cache policy"),
*GetName(), *WriteToString<96>(Key), *Name);
return OnComplete(MakeResponse(EStatus::Error), 0);
}
if (DebugOptions.ShouldSimulatePutMiss(Key))
{
UE_LOG(LogDerivedDataCache, Verbose, TEXT("%s: Simulated miss for put of %s from '%s'"),
*GetName(), *WriteToString<96>(Key), *Name);
return OnComplete(MakeResponse(EStatus::Error), 0);
}
// TODO: Jupiter currently always overwrites. It doesn't have a "write if not present" feature (for records or attachments),
// but would require one to implement all policy correctly.
FString Bucket(Key.Bucket.ToString());
Bucket.ToLowerInline();
FCbWriter Writer;
Writer.BeginObject();
Writer.AddBinaryAttachment("RawHash", Value.GetRawHash());
Writer.AddInteger("RawSize", Value.GetRawSize());
Writer.EndObject();
FCbPackage Package(Writer.Save().AsObject());
Package.AddAttachment(FCbAttachment(Value.GetData()));
FPutPackageOp::PutPackage(*this, Owner, Name, Key, MoveTemp(Package), Policy, UserData, [MakeResponse = MoveTemp(MakeResponse), OnComplete = MoveTemp(OnComplete)](FPutPackageOp::FCachePutPackageResponse&& Response)
{
OnComplete(MakeResponse(Response.Status), Response.BytesSent);
});
}
void FHttpCacheStore::GetCacheValueAsync(
IRequestOwner& Owner,
FSharedString Name,
const FCacheKey& Key,
ECachePolicy Policy,
uint64 UserData,
FOnCacheGetValueComplete&& OnComplete)
{
if (!IsUsable())
{
UE_LOG(LogDerivedDataCache, VeryVerbose,
TEXT("%s: Skipped get of %s from '%s' because this cache store is not available"),
*GetName(), *WriteToString<96>(Key), *Name);
OnComplete({Name, Key, {}, UserData, EStatus::Error});
return;
}
// Skip the request if querying the cache is disabled.
const ECachePolicy QueryFlag = SpeedClass == ESpeedClass::Local ? ECachePolicy::QueryLocal : ECachePolicy::QueryRemote;
if (!EnumHasAnyFlags(Policy, QueryFlag))
{
UE_LOG(LogDerivedDataCache, VeryVerbose, TEXT("%s: Skipped get of %s from '%s' due to cache policy"),
*GetName(), *WriteToString<96>(Key), *Name);
OnComplete({Name, Key, {}, UserData, EStatus::Error});
return;
}
if (DebugOptions.ShouldSimulateGetMiss(Key))
{
UE_LOG(LogDerivedDataCache, Verbose, TEXT("%s: Simulated miss for get of %s from '%s'"),
*GetName(), *WriteToString<96>(Key), *Name);
OnComplete({Name, Key, {}, UserData, EStatus::Error});
return;
}
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
const bool bSkipData = EnumHasAnyFlags(Policy, ECachePolicy::SkipData);
FString Bucket(Key.Bucket.ToString());
Bucket.ToLowerInline();
TStringBuilder<256> RefsUri;
RefsUri << "api/v1/refs/" << StructuredNamespace << "/" << Bucket << "/" << Key.Hash;
FRequestPool* Pool = nullptr;
FHttpRequest* Request = WaitForHttpRequestForOwner<OperationCategory::Get>(Owner, false /* bUnboundedOverflow */, Pool);
if (bSkipData)
{
Request->SetHeader(TEXT("Accept"), TEXT("application/x-ue-cb"));
}
else
{
Request->SetHeader(TEXT("Accept"), TEXT("application/x-jupiter-inline"));
}
auto OnHttpRequestComplete = [this, &Owner, Name, Key, UserData, bSkipData, OnComplete = MoveTemp(OnComplete)] (FHttpRequest::Result HttpResult, FHttpRequest* Request)
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_GetCacheValueAsync_OnHttpRequestComplete);
int64 ResponseCode = Request->GetResponseCode();
if (FHttpRequest::IsSuccessResponse(ResponseCode))
{
FValue ResultValue;
FSharedBuffer ResponseBuffer = Request->MoveResponseBufferToShared();
if (bSkipData)
{
if (ValidateCompactBinary(ResponseBuffer, ECbValidateMode::Default) != ECbValidateError::None)
{
UE_LOG(LogDerivedDataCache, Display, TEXT("%s: Cache miss with invalid package for %s from '%s'"),
*GetName(), *WriteToString<96>(Key), *Name);
OnComplete({Name, Key, {}, UserData, EStatus::Error});
return FHttpRequest::ECompletionBehavior::Done;
}
const FCbObjectView Object = FCbObject(ResponseBuffer);
const FIoHash RawHash = Object["RawHash"].AsHash();
const uint64 RawSize = Object["RawSize"].AsUInt64(MAX_uint64);
if (RawHash.IsZero() || RawSize == MAX_uint64)
{
UE_LOG(LogDerivedDataCache, Display, TEXT("%s: Cache miss with invalid value for %s from '%'"),
*GetName(), *WriteToString<96>(Key), *Name);
OnComplete({Name, Key, {}, UserData, EStatus::Error});
return FHttpRequest::ECompletionBehavior::Done;
}
ResultValue = FValue(RawHash, RawSize);
}
else
{
FCompressedBuffer CompressedBuffer = FCompressedBuffer::FromCompressed(ResponseBuffer);
if (!CompressedBuffer)
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
{
FString ReceivedHashStr;
if (Request->GetHeader("X-Jupiter-InlinePayloadHash", ReceivedHashStr))
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
{
FIoHash ReceivedHash(ReceivedHashStr);
FIoHash ComputedHash = FIoHash::HashBuffer(ResponseBuffer.GetView());
if (ReceivedHash == ComputedHash)
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
{
CompressedBuffer = FCompressedBuffer::Compress(ResponseBuffer);
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
}
}
}
if (!CompressedBuffer)
{
UE_LOG(LogDerivedDataCache, Display, TEXT("%s: Cache miss with invalid package for %s from '%s'"),
*GetName(), *WriteToString<96>(Key), *Name);
OnComplete({Name, Key, {}, UserData, EStatus::Error});
return FHttpRequest::ECompletionBehavior::Done;
}
ResultValue = FValue(CompressedBuffer);
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
}
OnComplete({Name, Key, ResultValue, UserData, EStatus::Ok});
return FHttpRequest::ECompletionBehavior::Done;
}
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
if (!ShouldAbortForShutdown() && !Owner.IsCanceled() && ShouldRetryOnError(HttpResult, ResponseCode) && ((Request->GetAttempts()+1) < UE_HTTPDDC_MAX_ATTEMPTS))
{
return FHttpRequest::ECompletionBehavior::Retry;
}
UE_LOG(LogDerivedDataCache, Verbose, TEXT("%s: Cache miss with failed HTTP request for %s from '%s'"),
*GetName(), *WriteToString<96>(Key), *Name);
OnComplete({Name, Key, {}, UserData, EStatus::Error});
return FHttpRequest::ECompletionBehavior::Done;
};
Request->EnqueueAsyncDownload(Owner, Pool, *RefsUri, MoveTemp(OnHttpRequestComplete), { 401, 404 });
}
void FHttpCacheStore::GetCacheRecordAsync(
IRequestOwner& Owner,
const FSharedString& Name,
const FCacheKey& Key,
const FCacheRecordPolicy& Policy,
uint64 UserData,
TUniqueFunction<void(FCacheGetResponse&& Response, uint64 BytesReceived)>&& OnComplete)
{
FGetRecordOp::GetRecord(*this, Owner, Name, Key, Policy, UserData, MoveTemp(OnComplete));
}
void FHttpCacheStore::RefCachedDataProbablyExistsBatchAsync(
IRequestOwner& Owner,
TConstArrayView<FCacheGetValueRequest> ValueRefs,
FOnCacheGetValueComplete&& OnComplete)
{
if (ValueRefs.IsEmpty())
{
return;
}
if (!IsUsable())
{
for (const FCacheGetValueRequest& ValueRef : ValueRefs)
{
UE_LOG(LogDerivedDataCache, VeryVerbose,
TEXT("%s: Skipped exists check of %s from '%s' because this cache store is not available"),
*GetName(), *WriteToString<96>(ValueRef.Key), *ValueRef.Name);
OnComplete(ValueRef.MakeResponse(EStatus::Error));
}
return;
}
TStringBuilder<256> RefsUri;
RefsUri << "api/v1/refs/" << StructuredNamespace;
FCbWriter RequestWriter;
RequestWriter.BeginObject();
RequestWriter.BeginArray(ANSITEXTVIEW("ops"));
uint32 OpIndex = 0;
for (const FCacheGetValueRequest& ValueRef : ValueRefs)
{
RequestWriter.BeginObject();
RequestWriter.AddInteger(ANSITEXTVIEW("opId"), OpIndex);
RequestWriter.AddString(ANSITEXTVIEW("op"), ANSITEXTVIEW("GET"));
FCacheKey Key = ValueRef.Key;
FString Bucket(Key.Bucket.ToString());
Bucket.ToLowerInline();
RequestWriter.AddString(ANSITEXTVIEW("bucket"), Bucket);
RequestWriter.AddString(ANSITEXTVIEW("key"), LexToString(Key.Hash));
RequestWriter.AddBool(ANSITEXTVIEW("resolveAttachments"), true);
RequestWriter.EndObject();
++OpIndex;
}
RequestWriter.EndArray();
RequestWriter.EndObject();
FCbFieldIterator RequestFields = RequestWriter.Save();
FRequestPool* Pool = nullptr;
FHttpRequest* Request = WaitForHttpRequestForOwner<OperationCategory::Get>(Owner, false /* bUnboundedOverflow */, Pool);
Request->SetHeader(TEXT("Accept"), TEXT("application/x-ue-cb"));
auto OnHttpRequestComplete = [this, &Owner, ValueRefs = TArray<FCacheGetValueRequest>(ValueRefs), OnComplete = MoveTemp(OnComplete)](FHttpRequest::Result HttpResult, FHttpRequest* Request)
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_RefCachedDataProbablyExistsBatchAsync_OnHttpRequestComplete);
int64 ResponseCode = Request->GetResponseCode();
if (FHttpRequest::IsSuccessResponse(ResponseCode))
{
FMemoryView ResponseView = MakeMemoryView(Request->GetResponseBuffer().GetData(), Request->GetResponseBuffer().Num());
if (ValidateCompactBinary(ResponseView, ECbValidateMode::Default) != ECbValidateError::None)
{
for (const FCacheGetValueRequest& ValueRef : ValueRefs)
{
UE_LOG(LogDerivedDataCache, Log,
TEXT("%s: Cache exists returned invalid results."),
*GetName());
OnComplete(ValueRef.MakeResponse(EStatus::Error));
}
return FHttpRequest::ECompletionBehavior::Done;
}
const FCbObjectView ResponseObject = FCbObjectView(Request->GetResponseBuffer().GetData());
FCbArrayView ResultsArrayView = ResponseObject[ANSITEXTVIEW("results")].AsArrayView();
if (ResultsArrayView.Num() != ValueRefs.Num())
{
for (const FCacheGetValueRequest& ValueRef : ValueRefs)
{
UE_LOG(LogDerivedDataCache, Log,
TEXT("%s: Cache exists returned unexpected quantity of results (expected %d, got %d)."),
*GetName(), ValueRefs.Num(), ResultsArrayView.Num());
OnComplete(ValueRef.MakeResponse(EStatus::Error));
}
return FHttpRequest::ECompletionBehavior::Done;
}
for (FCbFieldView ResultFieldView : ResultsArrayView)
{
FCbObjectView ResultObjectView = ResultFieldView.AsObjectView();
uint32 OpId = ResultObjectView[ANSITEXTVIEW("opId")].AsUInt32();
FCbObjectView ResponseObjectView = ResultObjectView[ANSITEXTVIEW("response")].AsObjectView();
int32 StatusCode = ResultObjectView[ANSITEXTVIEW("statusCode")].AsInt32();
if (OpId >= (uint32)ValueRefs.Num())
{
UE_LOG(LogDerivedDataCache, Display, TEXT("%s: Encountered invalid opId %d while querying %d values"),
*GetName(), OpId, ValueRefs.Num());
continue;
}
const FCacheGetValueRequest& ValueRef = ValueRefs[OpId];
if (!FHttpRequest::IsSuccessResponse(StatusCode))
{
UE_LOG(LogDerivedDataCache, Verbose, TEXT("%s: Cache miss with unsuccessful response code %d for %s from '%s'"),
*GetName(), StatusCode, *WriteToString<96>(ValueRef.Key), *ValueRef.Name);
OnComplete(ValueRef.MakeResponse(EStatus::Error));
continue;
}
const ECachePolicy QueryFlag = SpeedClass == ESpeedClass::Local ? ECachePolicy::QueryLocal : ECachePolicy::QueryRemote;
if (!EnumHasAnyFlags(ValueRef.Policy, QueryFlag))
{
UE_LOG(LogDerivedDataCache, VeryVerbose, TEXT("%s: Skipped exists check of %s from '%s' due to cache policy"),
*GetName(), *WriteToString<96>(ValueRef.Key), *ValueRef.Name);
OnComplete(ValueRef.MakeResponse(EStatus::Error));
continue;
}
const FIoHash RawHash = ResponseObjectView[ANSITEXTVIEW("RawHash")].AsHash();
const uint64 RawSize = ResponseObjectView[ANSITEXTVIEW("RawSize")].AsUInt64(MAX_uint64);
if (RawHash.IsZero() || RawSize == MAX_uint64)
{
UE_LOG(LogDerivedDataCache, Display, TEXT("%s: Cache miss with invalid value for %s from '%s'"),
*GetName(), *WriteToString<96>(ValueRef.Key), *ValueRef.Name);
OnComplete(ValueRef.MakeResponse(EStatus::Error));
continue;
}
OnComplete({ValueRef.Name, ValueRef.Key, FValue(RawHash, RawSize), ValueRef.UserData, EStatus::Ok});
}
return FHttpRequest::ECompletionBehavior::Done;
}
if (!ShouldAbortForShutdown() && !Owner.IsCanceled() && ShouldRetryOnError(HttpResult, ResponseCode) && ((Request->GetAttempts()+1) < UE_HTTPDDC_MAX_ATTEMPTS))
{
return FHttpRequest::ECompletionBehavior::Retry;
}
for (const FCacheGetValueRequest& ValueRef : ValueRefs)
{
UE_LOG(LogDerivedDataCache, Verbose, TEXT("%s: Cache miss with failed HTTP request for %s from '%s'"),
*GetName(), *WriteToString<96>(ValueRef.Key), *ValueRef.Name);
OnComplete(ValueRef.MakeResponse(EStatus::Error));
}
return FHttpRequest::ECompletionBehavior::Done;
};
Request->EnqueueAsyncUpload<FHttpRequest::PostCompactBinary>(Owner, Pool, *RefsUri, RequestFields.GetOuterBuffer(), MoveTemp(OnHttpRequestComplete));
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
bool FHttpCacheStore::CachedDataProbablyExists(const TCHAR* CacheKey)
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_Exist);
TRACE_COUNTER_ADD(HttpDDC_Exist, int64(1));
COOK_STAT(auto Timer = UsageStats.TimeProbablyExists());
if (DebugOptions.ShouldSimulateGetMiss(CacheKey))
{
return false;
}
#if WITH_DATAREQUEST_HELPER
// Retry request until we get an accepted response or exhaust allowed number of attempts.
for (int32 Attempts = 0; Attempts < UE_HTTPDDC_MAX_ATTEMPTS; ++Attempts)
{
FDataRequestHelper RequestHelper(GetRequestPools[IsInGameThread()].Get(), *Namespace, *DefaultBucket, CacheKey, nullptr);
const int64 ResponseCode = RequestHelper.GetResponseCode();
if (FHttpRequest::IsSuccessResponse(ResponseCode) && RequestHelper.IsSuccess())
{
COOK_STAT(Timer.AddHit(0));
return true;
}
if (!ShouldRetryOnError(ResponseCode))
{
return false;
}
}
#else
FString Uri = FString::Printf(TEXT("api/v1/c/ddc/%s/%s/%s"), *Namespace, *DefaultBucket, CacheKey);
// Retry request until we get an accepted response or exhaust allowed number of attempts.
for (int32 Attempts = 0; Attempts < UE_HTTPDDC_MAX_ATTEMPTS; ++Attempts)
{
FScopedRequestPtr Request(GetRequestPools[IsInGameThread()].Get());
const FHttpRequest::Result Result = Request->PerformBlockingQuery<FHttpRequest::Head>(*Uri);
const int64 ResponseCode = Request->GetResponseCode();
if (FHttpRequest::IsSuccessResponse(ResponseCode) || ResponseCode == 400)
{
const bool bIsHit = (Result == FHttpRequest::Success && FHttpRequest::IsSuccessResponse(ResponseCode));
if (bIsHit)
{
TRACE_COUNTER_ADD(HttpDDC_ExistHit, int64(1));
COOK_STAT(Timer.AddHit(0));
}
return bIsHit;
}
if (!ShouldRetryOnError(ResponseCode))
{
break;
}
}
#endif
return false;
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
TBitArray<> FHttpCacheStore::CachedDataProbablyExistsBatch(TConstArrayView<FString> CacheKeys)
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_Exist);
TRACE_COUNTER_ADD(HttpDDC_Exist, int64(1));
COOK_STAT(auto Timer = UsageStats.TimeProbablyExists());
#if WITH_DATAREQUEST_HELPER
for (int32 Attempts = 0; Attempts < UE_HTTPDDC_MAX_ATTEMPTS; ++Attempts)
{
FDataRequestHelper RequestHelper(GetRequestPools[IsInGameThread()].Get(), *Namespace, *DefaultBucket, CacheKeys);
const int64 ResponseCode = RequestHelper.GetResponseCode();
if (FHttpRequest::IsSuccessResponse(ResponseCode) && RequestHelper.IsSuccess())
{
COOK_STAT(Timer.AddHit(0));
TBitArray<> Results = RequestHelper.IsBatchSuccess();
int32 ResultIndex = 0;
for (const FString& CacheKey : CacheKeys)
{
if (DebugOptions.ShouldSimulateGetMiss(*CacheKey))
{
Results[ResultIndex] = false;
}
ResultIndex++;
}
return Results;
}
if (!ShouldRetryOnError(ResponseCode))
{
TBitArray<> Results = RequestHelper.IsBatchSuccess();
int32 ResultIndex = 0;
for (const FString& CacheKey : CacheKeys)
{
if (DebugOptions.ShouldSimulateGetMiss(*CacheKey))
{
Results[ResultIndex] = false;
}
ResultIndex++;
}
return Results;
}
}
#else
const TCHAR* const Uri = TEXT("api/v1/c/ddc-rpc");
TAnsiStringBuilder<512> Body;
const FTCHARToUTF8 AnsiNamespace(*Namespace);
const FTCHARToUTF8 AnsiBucket(*DefaultBucket);
Body << "{\"Operations\":[";
for (const FString& CacheKey : CacheKeys)
{
Body << "{\"Namespace\":\"" << AnsiNamespace.Get() << "\",\"Bucket\":\"" << AnsiBucket.Get() << "\",";
Body << "\"Id\":\"" << FTCHARToUTF8(*CacheKey).Get() << "\",\"Op\":\"HEAD\"},";
}
Body.RemoveSuffix(1);
Body << "]}";
TConstArrayView<uint8> BodyView(reinterpret_cast<const uint8*>(Body.ToString()), Body.Len());
// Retry request until we get an accepted response or exhaust allowed number of attempts.
for (int32 Attempts = 0; Attempts < UE_HTTPDDC_MAX_ATTEMPTS; ++Attempts)
{
FScopedRequestPtr Request(GetRequestPools[IsInGameThread()].Get());
const FHttpRequest::Result Result = Request->PerformBlockingUpload<FHttpRequest::PostJson>(Uri, BodyView);
const int64 ResponseCode = Request->GetResponseCode();
if (Result == FHttpRequest::Success && ResponseCode == 200)
{
TArray<TSharedPtr<FJsonValue>> ResponseArray = Request->GetResponseAsJsonArray();
TBitArray<> Exists;
Exists.Reserve(CacheKeys.Num());
for (const FString& CacheKey : CacheKeys)
{
if (DebugOptions.ShouldSimulateGetMiss(*CacheKey))
{
Exists.Add(false);
}
else
{
const TSharedPtr<FJsonValue>* FoundResponse = Algo::FindByPredicate(ResponseArray, [&CacheKey](const TSharedPtr<FJsonValue>& Response) {
FString Key;
Response->TryGetString(Key);
return Key == CacheKey;
});
Exists.Add(FoundResponse != nullptr);
}
}
if (Exists.CountSetBits() == CacheKeys.Num())
{
TRACE_COUNTER_ADD(HttpDDC_ExistHit, int64(1));
COOK_STAT(Timer.AddHit(0));
}
return Exists;
}
if (!ShouldRetryOnError(ResponseCode))
{
break;
}
}
#endif
return TBitArray<>(false, CacheKeys.Num());
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
bool FHttpCacheStore::GetCachedData(const TCHAR* CacheKey, TArray<uint8>& OutData)
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_GetCachedData);
TRACE_COUNTER_ADD(HttpDDC_Get, int64(1));
COOK_STAT(auto Timer = UsageStats.TimeGet());
if (DebugOptions.ShouldSimulateGetMiss(CacheKey))
{
return false;
}
#if WITH_DATAREQUEST_HELPER
// Retry request until we get an accepted response or exhaust allowed number of attempts.
for (int32 Attempts = 0; Attempts < UE_HTTPDDC_MAX_ATTEMPTS; ++Attempts)
{
FDataRequestHelper RequestHelper(GetRequestPools[IsInGameThread()].Get(), *Namespace, *DefaultBucket, CacheKey, &OutData);
const int64 ResponseCode = RequestHelper.GetResponseCode();
if (FHttpRequest::IsSuccessResponse(ResponseCode) && RequestHelper.IsSuccess())
{
COOK_STAT(Timer.AddHit(OutData.Num()));
check(OutData.Num() > 0);
return true;
}
if (!ShouldRetryOnError(ResponseCode))
{
return false;
}
}
#else
FString Uri = FString::Printf(TEXT("api/v1/c/ddc/%s/%s/%s.raw"), *Namespace, *DefaultBucket, CacheKey);
// Retry request until we get an accepted response or exhaust allowed number of attempts.
for (int32 Attempts = 0; Attempts < UE_HTTPDDC_MAX_ATTEMPTS; ++Attempts)
{
FScopedRequestPtr Request(GetRequestPools[IsInGameThread()].Get());
if (Request.IsValid())
{
FHttpRequest::Result Result = Request->PerformBlockingDownload(*Uri, &OutData);
const uint64 ResponseCode = Request->GetResponseCode();
// Request was successful, make sure we got all the expected data.
if (FHttpRequest::IsSuccessResponse(ResponseCode) && VerifyRequest(Request.Get(), *Namespace, *DefaultBucket, CacheKey, OutData))
{
TRACE_COUNTER_ADD(HttpDDC_GetHit, int64(1));
TRACE_COUNTER_ADD(HttpDDC_BytesReceived, int64(Request->GetBytesReceived()));
COOK_STAT(Timer.AddHit(Request->GetBytesReceived()));
return true;
}
if (!ShouldRetryOnError(ResponseCode))
{
return false;
}
}
}
#endif
return false;
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
FDerivedDataBackendInterface::EPutStatus FHttpCacheStore::PutCachedData(const TCHAR* CacheKey, TArrayView<const uint8> InData, bool bPutEvenIfExists)
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_PutCachedData);
if (!IsWritable())
{
UE_LOG(LogDerivedDataCache, Verbose, TEXT("%s is read only. Skipping put of %s"), *GetName(), CacheKey);
return EPutStatus::NotCached;
}
// don't put anything we pretended didn't exist
if (DebugOptions.ShouldSimulatePutMiss(CacheKey))
{
return EPutStatus::Skipped;
}
#if 0 // No longer WITH_DATAREQUEST_HELPER as async puts are unsupported except through the AsyncPutWrapper which expects the inner backend to perform the put synchronously
for (int32 Attempts = 0; Attempts < UE_HTTPDDC_MAX_ATTEMPTS; ++Attempts)
{
FDataUploadHelper Request(PutRequestPools[IsInGameThread()].Get(), *Namespace, *DefaultBucket, CacheKey, InData, UsageStats);
if (ShouldAbortForShutdown())
{
return EPutStatus::NotCached;
}
const int64 ResponseCode = Request.GetResponseCode();
if (Request.IsSuccess() && (Request.IsQueued() || FHttpRequest::IsSuccessResponse(ResponseCode)))
{
return Request.IsQueued() ? EPutStatus::Executing : EPutStatus::Cached;
}
if (!ShouldRetryOnError(ResponseCode))
{
return EPutStatus::NotCached;
}
}
#else
COOK_STAT(auto Timer = UsageStats.TimePut());
FString Uri = FString::Printf(TEXT("api/v1/c/ddc/%s/%s/%s"), *Namespace, *DefaultBucket, CacheKey);
int64 ResponseCode = 0; uint32 Attempts = 0;
// Retry request until we get an accepted response or exhaust allowed number of attempts.
while (ResponseCode == 0 && ++Attempts < UE_HTTPDDC_MAX_ATTEMPTS)
{
if (ShouldAbortForShutdown())
{
return EPutStatus::NotCached;
}
FScopedRequestPtr Request(PutRequestPools[IsInGameThread()].Get());
if (Request.IsValid())
{
// Append the content hash to the header
HashPayload(Request.Get(), InData);
Request->PerformBlockingUpload<FHttpRequest::Put>(*Uri, InData);
ResponseCode = Request->GetResponseCode();
if (FHttpRequest::IsSuccessResponse(ResponseCode))
{
TRACE_COUNTER_ADD(HttpDDC_BytesSent, int64(Request->GetBytesSent()));
COOK_STAT(Timer.AddHit(Request->GetBytesSent()));
return EPutStatus::Cached;
}
if (!ShouldRetryOnError(ResponseCode))
{
return EPutStatus::NotCached;
}
ResponseCode = 0;
}
}
#endif // WITH_DATAREQUEST_HELPER
return EPutStatus::NotCached;
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
void FHttpCacheStore::RemoveCachedData(const TCHAR* CacheKey, bool bTransient)
{
// do not remove transient data as Jupiter does its own verification of the content and cleans itself up
if (!IsWritable() || bTransient)
return;
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_Remove);
FString Uri = FString::Printf(TEXT("api/v1/c/ddc/%s/%s/%s"), *Namespace, *DefaultBucket, CacheKey);
int64 ResponseCode = 0; uint32 Attempts = 0;
// Retry request until we get an accepted response or exhaust allowed number of attempts.
while (ResponseCode == 0 && ++Attempts < UE_HTTPDDC_MAX_ATTEMPTS)
{
FScopedRequestPtr Request(PutRequestPools[IsInGameThread()].Get());
if (Request.IsValid())
{
FHttpRequest::Result Result = Request->PerformBlockingQuery<FHttpRequest::Delete>(*Uri, {});
ResponseCode = Request->GetResponseCode();
if (ResponseCode == 200)
{
return;
}
if (!ShouldRetryOnError(ResponseCode))
{
return;
}
ResponseCode = 0;
}
}
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
TSharedRef<FDerivedDataCacheStatsNode> FHttpCacheStore::GatherUsageStats() const
{
TSharedRef<FDerivedDataCacheStatsNode> Usage =
MakeShared<FDerivedDataCacheStatsNode>(TEXT("Horde Storage"), FString::Printf(TEXT("%s (%s)"), *Domain, *Namespace), /*bIsLocal*/ false);
Usage->Stats.Add(TEXT(""), UsageStats);
return Usage;
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
void FHttpCacheStore::Put(
const TConstArrayView<FCachePutRequest> Requests,
IRequestOwner& Owner,
FOnCachePutComplete&& OnComplete)
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_Put);
FRequestBarrier Barrier(Owner);
TRefCountedUniqueFunction<FOnCachePutComplete>* CompletionFunction = new TRefCountedUniqueFunction<FOnCachePutComplete>(MoveTemp(OnComplete));
TRefCountPtr<TRefCountedUniqueFunction<FOnCachePutComplete>> BatchOnCompleteRef(CompletionFunction);
for (const FCachePutRequest& Request : Requests)
{
PutCacheRecordAsync(Owner, Request.Name, Request.Record, Request.Policy, Request.UserData, [COOK_STAT(Timer = UsageStats.TimePut(), ) OnCompletePtr = TRefCountPtr<TRefCountedUniqueFunction<FOnCachePutComplete>>(CompletionFunction)](FCachePutResponse&& Response, uint64 BytesSent) mutable
{
TRACE_COUNTER_ADD(HttpDDC_BytesSent, BytesSent);
if (Response.Status == EStatus::Ok)
{
COOK_STAT(if (BytesSent) { Timer.AddHit(BytesSent); });
}
OnCompletePtr->GetFunction()(MoveTemp(Response));
});
}
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
void FHttpCacheStore::Get(
const TConstArrayView<FCacheGetRequest> Requests,
IRequestOwner& Owner,
FOnCacheGetComplete&& OnComplete)
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_Get);
FRequestBarrier Barrier(Owner);
TRefCountedUniqueFunction<FOnCacheGetComplete>* CompletionFunction = new TRefCountedUniqueFunction<FOnCacheGetComplete>(MoveTemp(OnComplete));
TRefCountPtr<TRefCountedUniqueFunction<FOnCacheGetComplete>> BatchOnCompleteRef(CompletionFunction);
for (const FCacheGetRequest& Request : Requests)
{
GetCacheRecordAsync(Owner, Request.Name, Request.Key, Request.Policy, Request.UserData, [COOK_STAT(Timer = UsageStats.TimePut(), ) OnCompletePtr = TRefCountPtr<TRefCountedUniqueFunction<FOnCacheGetComplete>>(CompletionFunction)](FCacheGetResponse&& Response, uint64 BytesReceived) mutable
{
TRACE_COUNTER_ADD(HttpDDC_BytesReceived, BytesReceived);
if (Response.Status == EStatus::Ok)
{
COOK_STAT(Timer.AddHit(BytesReceived););
}
OnCompletePtr->GetFunction()(MoveTemp(Response));
});
}
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
void FHttpCacheStore::PutValue(
const TConstArrayView<FCachePutValueRequest> Requests,
IRequestOwner& Owner,
FOnCachePutValueComplete&& OnComplete)
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_PutValue);
FRequestBarrier Barrier(Owner);
TRefCountedUniqueFunction<FOnCachePutValueComplete>* CompletionFunction = new TRefCountedUniqueFunction<FOnCachePutValueComplete>(MoveTemp(OnComplete));
TRefCountPtr<TRefCountedUniqueFunction<FOnCachePutValueComplete>> BatchOnCompleteRef(CompletionFunction);
for (const FCachePutValueRequest& Request : Requests)
{
PutCacheValueAsync(Owner, Request.Name, Request.Key, Request.Value, Request.Policy, Request.UserData, [COOK_STAT(Timer = UsageStats.TimePut(),) OnCompletePtr = TRefCountPtr<TRefCountedUniqueFunction<FOnCachePutValueComplete>>(CompletionFunction)](FCachePutValueResponse&& Response, uint64 BytesSent) mutable
{
TRACE_COUNTER_ADD(HttpDDC_BytesSent, BytesSent);
if (Response.Status == EStatus::Ok)
{
COOK_STAT(if (BytesSent) { Timer.AddHit(BytesSent); });
}
OnCompletePtr->GetFunction()(MoveTemp(Response));
});
}
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
void FHttpCacheStore::GetValue(
const TConstArrayView<FCacheGetValueRequest> Requests,
IRequestOwner& Owner,
FOnCacheGetValueComplete&& OnComplete)
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_GetValue);
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
COOK_STAT(double StartTime = FPlatformTime::Seconds());
COOK_STAT(bool bIsInGameThread = IsInGameThread());
bool bBatchExistsCandidate = true;
for (const FCacheGetValueRequest& Request : Requests)
{
if (!EnumHasAnyFlags(Request.Policy, ECachePolicy::SkipData))
{
bBatchExistsCandidate = false;
break;
}
}
if (bBatchExistsCandidate)
{
RefCachedDataProbablyExistsBatchAsync(Owner, Requests,
[this, COOK_STAT(StartTime, bIsInGameThread, ) OnComplete = MoveTemp(OnComplete)](FCacheGetValueResponse&& Response)
{
if (Response.Status != EStatus::Ok)
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
{
COOK_STAT(UsageStats.GetStats.Accumulate(FCookStats::CallStats::EHitOrMiss::Miss, FCookStats::CallStats::EStatType::Counter, 1l, bIsInGameThread));
OnComplete(MoveTemp(Response));
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
}
else
{
UE_LOG(LogDerivedDataCache, Verbose, TEXT("%s: Cache hit for %s from '%s'"),
*GetName(), *WriteToString<96>(Response.Key), *Response.Name);
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
COOK_STAT(UsageStats.GetStats.Accumulate(FCookStats::CallStats::EHitOrMiss::Hit, FCookStats::CallStats::EStatType::Counter, 1l, bIsInGameThread));
OnComplete(MoveTemp(Response));
}
COOK_STAT(const int64 CyclesUsed = int64((FPlatformTime::Seconds() - StartTime) / FPlatformTime::GetSecondsPerCycle()));
COOK_STAT(UsageStats.GetStats.Accumulate(FCookStats::CallStats::EHitOrMiss::Hit, FCookStats::CallStats::EStatType::Cycles, CyclesUsed, bIsInGameThread));
});
}
else
{
FRequestBarrier Barrier(Owner);
TRefCountedUniqueFunction<FOnCacheGetValueComplete>* CompletionFunction = new TRefCountedUniqueFunction<FOnCacheGetValueComplete>(MoveTemp(OnComplete));
TRefCountPtr<TRefCountedUniqueFunction<FOnCacheGetValueComplete>> BatchOnCompleteRef(CompletionFunction);
int64 HitBytes = 0;
for (const FCacheGetValueRequest& Request : Requests)
{
GetCacheValueAsync(Owner, Request.Name, Request.Key, Request.Policy, Request.UserData,
[this, COOK_STAT(StartTime, bIsInGameThread,) Policy = Request.Policy, OnCompletePtr = TRefCountPtr<TRefCountedUniqueFunction<FOnCacheGetValueComplete>>(CompletionFunction)] (FCacheGetValueResponse&& Response)
{
const FOnCacheGetValueComplete& OnComplete = OnCompletePtr->GetFunction();
check(OnComplete);
if (Response.Status != EStatus::Ok)
{
COOK_STAT(UsageStats.GetStats.Accumulate(FCookStats::CallStats::EHitOrMiss::Miss, FCookStats::CallStats::EStatType::Counter, 1l, bIsInGameThread));
OnComplete(MoveTemp(Response));
}
else
{
if (!IsValueDataReady(Response.Value, Policy) && !EnumHasAnyFlags(Policy, ECachePolicy::SkipData))
{
// With inline fetching, expect we will always have a value we can use. Even SkipData/Exists can rely on the blob existing if the ref is reported to exist.
UE_LOG(LogDerivedDataCache, Log, TEXT("%s: Cache miss due to inlining failure for %s from '%s'"),
*GetName(), *WriteToString<96>(Response.Key), *Response.Name);
COOK_STAT(UsageStats.GetStats.Accumulate(FCookStats::CallStats::EHitOrMiss::Miss, FCookStats::CallStats::EStatType::Counter, 1l, bIsInGameThread));
OnComplete(MoveTemp(Response));
}
else
{
UE_LOG(LogDerivedDataCache, Verbose, TEXT("%s: Cache hit for %s from '%s'"),
*GetName(), *WriteToString<96>(Response.Key), *Response.Name);
uint64 ValueSize = Response.Value.GetData().GetCompressedSize();
TRACE_COUNTER_ADD(HttpDDC_BytesReceived, ValueSize);
COOK_STAT(UsageStats.GetStats.Accumulate(FCookStats::CallStats::EHitOrMiss::Hit, FCookStats::CallStats::EStatType::Counter, 1l, bIsInGameThread));
OnComplete({ Response.Name, Response.Key, Response.Value, Response.UserData, EStatus::Ok });
COOK_STAT(UsageStats.GetStats.Accumulate(FCookStats::CallStats::EHitOrMiss::Hit, FCookStats::CallStats::EStatType::Bytes, ValueSize, bIsInGameThread));
}
}
COOK_STAT(const int64 CyclesUsed = int64((FPlatformTime::Seconds() - StartTime) / FPlatformTime::GetSecondsPerCycle()));
COOK_STAT(UsageStats.GetStats.Accumulate(FCookStats::CallStats::EHitOrMiss::Hit, FCookStats::CallStats::EStatType::Cycles, CyclesUsed, bIsInGameThread));
});
}
}
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
void FHttpCacheStore::GetChunks(
const TConstArrayView<FCacheGetChunkRequest> Requests,
IRequestOwner& Owner,
FOnCacheGetChunkComplete&& OnComplete)
{
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_GetChunks);
// TODO: This is inefficient because Jupiter doesn't allow us to get only part of a compressed blob, so we have to
// get the whole thing and then decompress only the portion we need. Furthermore, because there is no propagation
// between cache stores during chunk requests, the fetched result won't end up in the local store.
// These efficiency issues will be addressed by changes to the Hierarchy that translate chunk requests that
// are missing in local/fast stores and have to be retrieved from slow stores into record requests instead. That
// will make this code path unused/uncommon as Jupiter will most always be a slow store with a local/fast store in front of it.
// Regardless, to adhere to the functional contract, this implementation must exist.
TArray<FCacheGetChunkRequest, TInlineAllocator<16>> SortedRequests(Requests);
SortedRequests.StableSort(TChunkLess());
bool bHasValue = false;
FValue Value;
FValueId ValueId;
FCacheKey ValueKey;
FCompressedBuffer ValueBuffer;
FCompressedBufferReader ValueReader;
EStatus ValueStatus = EStatus::Error;
FOptionalCacheRecord Record;
for (const FCacheGetChunkRequest& Request : SortedRequests)
{
const bool bExistsOnly = EnumHasAnyFlags(Request.Policy, ECachePolicy::SkipData);
COOK_STAT(auto Timer = bExistsOnly ? UsageStats.TimeProbablyExists() : UsageStats.TimeGet());
if (!(bHasValue && ValueKey == Request.Key && ValueId == Request.Id) || ValueReader.HasSource() < !bExistsOnly)
{
ValueStatus = EStatus::Error;
ValueReader.ResetSource();
ValueKey = {};
ValueId.Reset();
Value.Reset();
bHasValue = false;
if (Request.Id.IsValid())
{
if (!(Record && Record.Get().GetKey() == Request.Key))
{
FCacheRecordPolicyBuilder PolicyBuilder(ECachePolicy::None);
PolicyBuilder.AddValuePolicy(Request.Id, Request.Policy);
Record.Reset();
FRequestOwner BlockingOwner(EPriority::Blocking);
GetCacheRecordOnlyAsync(BlockingOwner, Request.Name, Request.Key, PolicyBuilder.Build(), 0, [&Record](FGetCacheRecordOnlyResponse&& Response)
{
Record = MoveTemp(Response.Record);
});
BlockingOwner.Wait();
}
if (Record)
{
const FValueWithId& ValueWithId = Record.Get().GetValue(Request.Id);
bHasValue = ValueWithId.IsValid();
Value = ValueWithId;
ValueId = Request.Id;
ValueKey = Request.Key;
if (IsValueDataReady(Value, Request.Policy))
{
ValueReader.SetSource(Value.GetData());
}
else
{
auto IdGetter = [](const FValueWithId& Value)
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
{
return FString(WriteToString<16>(Value.GetId()));
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
};
FRequestOwner BlockingOwner(EPriority::Blocking);
bool bSucceeded = false;
FCompressedBuffer NewBuffer;
FGetRecordOp::GetDataBatch(*this, BlockingOwner, Request.Name, Request.Key, ::MakeArrayView({ ValueWithId }), IdGetter, [&bSucceeded, &NewBuffer](FGetRecordOp::FGetCachedDataBatchResponse&& Response)
{
if (Response.Status == EStatus::Ok)
{
bSucceeded = true;
NewBuffer = MoveTemp(Response.DataBuffer);
}
});
BlockingOwner.Wait();
if (bSucceeded)
{
ValueBuffer = MoveTemp(NewBuffer);
ValueReader.SetSource(ValueBuffer);
}
else
{
ValueBuffer.Reset();
ValueReader.ResetSource();
}
}
}
}
else
{
ValueKey = Request.Key;
{
FRequestOwner BlockingOwner(EPriority::Blocking);
bool bSucceeded = false;
GetCacheValueAsync(BlockingOwner, Request.Name, Request.Key, Request.Policy, 0, [&bSucceeded, &Value](FCacheGetValueResponse&& Response)
{
Value = MoveTemp(Response.Value);
bSucceeded = Response.Status == EStatus::Ok;
});
BlockingOwner.Wait();
bHasValue = bSucceeded;
}
if (bHasValue)
{
if (IsValueDataReady(Value, Request.Policy))
Avoid wasted time when pulling data from Jupiter by: 1) Increasing connection pool size 2) Allowing DNS and SSL/TLS data to be shared by connections 3) Switching to ValueOnly mode for Jupiter backends by default 4) Using "inline get" feature when querying Value data Tested: - .run editor -- -execCmds="Automation RunTests System.DerivedDataCache.HttpDerivedDataBackend.CacheStore;Quit" -NullRHI -stdout -unattended -ddc=DerivedDataBackendGraph_Jupiter -Launch CitySample editor with empty local cache and open Small_City_LVL -Launch CitySample editor with hot local cache and open Small_City_LVL Can be improved further with a batched existence API for refs (instead of blobs). The current one does not work due to a Jupiter server bug, so this is an improvement that will have to come afterwards. Note: There is a known issue involving Jupiter where the server will sometimes return uncompressed blobs. This manifests as cache misses due to invalid/rejected results returned from Jupiter. The client handles the miss and builds the data and stores it in its local cache but does not upload it to Jupiter (because Jupiter says it already has the blob). Subsequent runs will not suffer from the issue as the data will be present in the local cache. There are discussions in progress for how to address this issue on server or client side. No immediate mitigation is needed for it. #rb devin.doucette #jira UE-143616 #lockdown aurel.cordonnier #preflight #ROBOMERGE-OWNER: zousar.shaker #ROBOMERGE-AUTHOR: zousar.shaker #ROBOMERGE-SOURCE: CL 19104205 via CL 19104462 via CL 19104579 via CL 19104585 via CL 19110006 #ROBOMERGE-BOT: UE5 (Release-Engine-Staging -> Main) (v921-19075845) [CL 19113840 by zousar shaker in ue5-main branch]
2022-02-24 10:24:10 -05:00
{
ValueReader.SetSource(Value.GetData());
}
else
{
auto IdGetter = [](const FValue& Value)
{
return FString(TEXT("Default"));
};
FRequestOwner BlockingOwner(EPriority::Blocking);
bool bSucceeded = false;
FCompressedBuffer NewBuffer;
FGetRecordOp::GetDataBatch(*this, BlockingOwner, Request.Name, Request.Key, ::MakeArrayView({ Value }), IdGetter, [&bSucceeded, &NewBuffer](FGetRecordOp::FGetCachedDataBatchResponse&& Response)
{
if (Response.Status == EStatus::Ok)
{
bSucceeded = true;
NewBuffer = MoveTemp(Response.DataBuffer);
}
});
BlockingOwner.Wait();
if (bSucceeded)
{
ValueBuffer = MoveTemp(NewBuffer);
ValueReader.SetSource(ValueBuffer);
}
else
{
ValueBuffer.Reset();
ValueReader.ResetSource();
}
}
}
else
{
ValueBuffer.Reset();
ValueReader.ResetSource();
}
}
}
if (bHasValue)
{
const uint64 RawOffset = FMath::Min(Value.GetRawSize(), Request.RawOffset);
const uint64 RawSize = FMath::Min(Value.GetRawSize() - RawOffset, Request.RawSize);
UE_LOG(LogDerivedDataCache, Verbose, TEXT("%s: Cache hit for %s from '%s'"),
*GetName(), *WriteToString<96>(Request.Key, '/', Request.Id), *Request.Name);
COOK_STAT(Timer.AddHit(!bExistsOnly ? RawSize : 0));
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
FSharedBuffer Buffer;
if (!bExistsOnly)
{
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
Buffer = ValueReader.Decompress(RawOffset, RawSize);
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
const EStatus ChunkStatus = bExistsOnly || Buffer.GetSize() == RawSize ? EStatus::Ok : EStatus::Error;
OnComplete({Request.Name, Request.Key, Request.Id, Request.RawOffset,
RawSize, Value.GetRawHash(), MoveTemp(Buffer), Request.UserData, ChunkStatus});
continue;
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
OnComplete(Request.MakeResponse(EStatus::Error));
}
}
bool FHttpRequest::AllowAsync()
{
if (!FGenericPlatformProcess::SupportsMultithreading() || !bHttpEnableAsync)
{
return false;
}
return true;
}
void FHttpRequest::EnqueueAsync(IRequestOwner& Owner, FRequestPool* Pool, const TCHAR* Uri, RequestVerb Verb, uint64 ContentLength, FOnHttpRequestComplete&& OnComplete, TConstArrayView<long> ExpectedErrorCodes)
{
if (!AllowAsync())
{
while (OnComplete(PerformBlocking(Uri, Verb, ContentLength, ExpectedErrorCodes), this) == ECompletionBehavior::Retry)
{
PrepareToRetry();
}
if (Pool)
{
Pool->ReleaseRequestToPool(this);
}
return;
}
TRACE_CPUPROFILER_EVENT_SCOPE(HttpDDC_CurlEnqueueAsync);
AsyncData = new FAsyncRequestData;
AsyncData->Owner = &Owner;
AsyncData->Pool = Pool;
AsyncData->CurlHeaders = PrepareToIssueRequest(Uri, ContentLength);
AsyncData->Uri = Uri;
AsyncData->Verb = Verb;
AsyncData->ExpectedErrorCodes = ExpectedErrorCodes;
AsyncData->OnComplete = MoveTemp(OnComplete);
AsyncData->Owner->Begin(AsyncData);
if (SharedData->PendingRequestAdditions.EnqueueAndReturnWasEmpty(Curl))
{
SharedData->PendingRequestEvent.Trigger();
}
bool ExistingValue = false;
if (SharedData->bAsyncThreadStarting.compare_exchange_weak(ExistingValue, true, std::memory_order_relaxed))
{
SharedData->AsyncRunnable = MakeUnique<FHttpCacheStoreRunnable>(*SharedData);
SharedData->AsyncServiceThread = FRunnableThread::Create(SharedData->AsyncRunnable.Get(), TEXT("HttpCacheStoreThread"), 64 * 1024, TPri_Normal);
}
}
void FHttpRequest::CompleteAsync(CURLcode Result)
{
CurlResult = Result;
// Get response code
curl_easy_getinfo(Curl, CURLINFO_RESPONSE_CODE, &ResponseCode);
LogResult(Result, *AsyncData->Uri, AsyncData->Verb, AsyncData->ExpectedErrorCodes);
ECompletionBehavior Behavior;
{
FRequestBarrier Barrier(*AsyncData->Owner, ERequestBarrierFlags::Priority);
FHttpRequest::Result HttpResult;
switch (CurlResult)
{
case CURLE_OK:
HttpResult = Success;
break;
case CURLE_OPERATION_TIMEDOUT:
HttpResult = FailedTimeout;
break;
default:
HttpResult = Failed;
break;
}
Behavior = AsyncData->OnComplete(HttpResult, this);
}
if (Behavior == ECompletionBehavior::Retry)
{
PrepareToRetry();
if (SharedData->PendingRequestAdditions.EnqueueAndReturnWasEmpty(Curl))
{
SharedData->PendingRequestEvent.Trigger();
}
}
else
{
// Clean up
curl_slist_free_all(AsyncData->CurlHeaders);
AsyncData->Owner->End(AsyncData, [this, Behavior]
{
AsyncData->Event.Trigger();
FRequestPool* Pool = AsyncData->Pool;
AsyncData = nullptr;
if (Pool)
{
Pool->ReleaseRequestToPool(this);
}
});
}
}
uint32 FHttpCacheStoreRunnable::Run()
{
int ActiveTransfers = 0;
auto ProcessPendingRequests = [this, &ActiveTransfers]()
{
int CurrentActiveTransfers = -1;
do
{
SharedData.PendingRequestAdditions.Deplete([this, &ActiveTransfers](CURL* Curl)
{
curl_multi_add_handle(SharedData.CurlMulti, Curl);
++ActiveTransfers;
});
curl_multi_perform(SharedData.CurlMulti, &CurrentActiveTransfers);
if (CurrentActiveTransfers == 0 || ActiveTransfers != CurrentActiveTransfers)
{
for (;;)
{
int MsgsStillInQueue = 0; // may use that to impose some upper limit we may spend in that loop
CURLMsg* Message = curl_multi_info_read(SharedData.CurlMulti, &MsgsStillInQueue);
if (!Message)
{
break;
}
// find out which requests have completed
if (Message->msg == CURLMSG_DONE)
{
CURL* CompletedHandle = Message->easy_handle;
curl_multi_remove_handle(SharedData.CurlMulti, CompletedHandle);
void* PrivateData = nullptr;
curl_easy_getinfo(CompletedHandle, CURLINFO_PRIVATE, &PrivateData);
FHttpRequest* CompletedRequest = (FHttpRequest*)PrivateData;
if (CompletedRequest)
{
// It is important that the CompleteAsync call doesn't happen on this thread as it is possible it will block waiting
// for a free HTTP request, and if that happens on this thread, we can deadlock as no HTTP requests will become
// available while this thread is blocked.
UE::Tasks::Launch(TEXT("FHttpRequest::CompleteAsync"), [CompletedRequest, Result = Message->data.result]() mutable
{
CompletedRequest->CompleteAsync(Result);
});
}
}
}
ActiveTransfers = CurrentActiveTransfers;
}
if (CurrentActiveTransfers > 0)
{
curl_multi_wait(SharedData.CurlMulti, nullptr, 0, 1, nullptr);
}
}
while (CurrentActiveTransfers > 0);
};
do
{
ProcessPendingRequests();
SharedData.PendingRequestEvent.Wait(100);
}
while (!FHttpSharedData::bAsyncThreadShutdownRequested.load(std::memory_order_relaxed));
// Process last requests before shutdown. May want these to be aborted instead.
ProcessPendingRequests();
return 0;
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
} // UE::DerivedData
#endif // WITH_HTTP_DDC_BACKEND
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
namespace UE::DerivedData
{
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
ILegacyCacheStore* CreateHttpCacheStore(
const TCHAR* NodeName,
const TCHAR* ServiceUrl,
bool bResolveHostCanonicalName,
const TCHAR* Namespace,
const TCHAR* StructuredNamespace,
const TCHAR* OAuthProvider,
const TCHAR* OAuthClientId,
const TCHAR* OAuthData,
const TCHAR* OAuthScope,
const FDerivedDataBackendInterface::ESpeedClass* ForceSpeedClass,
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
EBackendLegacyMode LegacyMode,
bool bReadOnly)
{
#if WITH_HTTP_DDC_BACKEND
FHttpCacheStore* Backend = new FHttpCacheStore(ServiceUrl, bResolveHostCanonicalName, Namespace, StructuredNamespace, OAuthProvider, OAuthClientId, OAuthData, OAuthScope, LegacyMode, bReadOnly);
if (Backend->IsUsable())
{
return Backend;
}
UE_LOG(LogDerivedDataCache, Warning, TEXT("Node %s could not contact the service (%s), will not use it"), NodeName, ServiceUrl);
delete Backend;
return nullptr;
#else
UE_LOG(LogDerivedDataCache, Warning, TEXT("HTTP backend is not yet supported in the current build configuration."));
return nullptr;
#endif
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
FDerivedDataBackendInterface* GetAnyHttpCacheStore(
FString& OutDomain,
FString& OutOAuthProvider,
FString& OutOAuthClientId,
FString& OutOAuthSecret,
FString& OutOAuthScope,
FString& OutNamespace,
FString& OutStructuredNamespace)
{
#if WITH_HTTP_DDC_BACKEND
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
if (FHttpCacheStore* HttpBackend = FHttpCacheStore::GetAny())
{
OutDomain = HttpBackend->GetDomain();
OutOAuthProvider = HttpBackend->GetOAuthProvider();
OutOAuthClientId = HttpBackend->GetOAuthClientId();
OutOAuthSecret = HttpBackend->GetOAuthSecret();
OutOAuthScope = HttpBackend->GetOAuthScope();
OutNamespace = HttpBackend->GetNamespace();
OutStructuredNamespace = HttpBackend->GetStructuredNamespace();
return HttpBackend;
}
return nullptr;
#else
return nullptr;
#endif
}
DDC: Enabled compression of legacy cache data - FileSystem, Http, Pak, S3 use the ValueWithLegacyFallback mode by default, which cause them to fall back to searching for uncompressed data if compressed data is not found. - FileSystem has been fixed to store up to 1 MiB of compressed data inline with the value package rather than separately in content-addressable storage. - Pak has been optimized to have GetChunks only load the required region of the requested value, rather than the whole value. - Pak has been changed to stop storing data inline in the record package, since it will end up in the same file anyway when stored separately. - Pak will upgrade the compressor and compression level when a compressed pak file is requested. Default cache compression uses Oodle Mermaid VeryFast and will upgrade to Oodle Kraken Optimal2. - Zen does not have compression enabled by default, pending deployment of a new version that stores compressed values to Horde Storage. - Added a missing request barrier when saving uncompressed data as compressed. Example reduction in file system cache size when cooking for Windows: - CitySample dropped from 66.5 GiB to 21.8 GiB. - Lyra dropped from 2.54 GiB to 672 MiB. - ShooterGame dropped from 1.21 GiB to 380 MiB. Example reduction in compressed pak file cache size when cooking for Windows: - CitySample dropped from 22.3 GiB to 18.5 GiB. - Lyra dropped from 691 MiB to 543 MiB. - ShooterGame dropped from 387 MiB to 313 MiB. #jira UE-134381 #preflight 620a703f583261b0a658e043, 620a6fb2803d9066e6805310, 620a733117632e948459b6af #lockdown Aurel.Cordonnier #rb Zousar.Shaker #ROBOMERGE-AUTHOR: devin.doucette #ROBOMERGE-SOURCE: CL 18983671 in //UE5/Release-5.0/... via CL 18983890 via CL 18984096 #ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v917-18934589) [CL 18984126 by devin doucette in ue5-main branch]
2022-02-14 14:43:39 -05:00
} // UE::DerivedData