Add shaping throttler delay to CallContext

cloud/blockstore/libs/client_rdma/rdma_client.cpp

@@ -66,19 +66,25 @@ struct IRequestHandler: public NRdma::TNullContext
 ////////////////////////////////////////////////////////////////////////////////
 
 template <typename TResponse>
-void ProcessPostponeTime(
+void ProcessThrottleTime(
     const TCallContextPtr& callContext,
     TResponse& localResponse)
 {
+    NProto::TThrottlerInfo& throttler =
+        *localResponse.MutableHeaders()->MutableThrottler();
     const ui64 throttlerDelay =
-        Max(localResponse.GetDeprecatedThrottlerDelay(),
-            localResponse.GetHeaders().GetThrottler().GetDelay());
+        Max(localResponse.GetDeprecatedThrottlerDelay(), throttler.GetDelay());
     callContext->AddTime(
         EProcessingStage::Postponed,
         TDuration::MicroSeconds(throttlerDelay));
     localResponse.SetDeprecatedThrottlerDelay(0);
-    localResponse.MutableHeaders()->MutableThrottler()->SetDelay(0);
+    throttler.SetDelay(0);
     callContext->SetPossiblePostponeDuration(TDuration::Zero());
+
+    callContext->AddTime(
+        EProcessingStage::Shaping,
+        TDuration::MicroSeconds(throttler.GetShapingDelay()));
+    throttler.SetShapingDelay(0);
 }
 
 ////////////////////////////////////////////////////////////////////////////////
@@ -181,7 +187,7 @@ class TReadBlocksHandler final
             responseMsg.MutableHeaders()->ClearTrace();
         }
 
-        ProcessPostponeTime(CallContext, localResponse);
+        ProcessThrottleTime(CallContext, localResponse);
 
         Response.SetValue(std::move(localResponse));
     }
@@ -316,7 +322,7 @@ class TWriteBlocksHandler final
             responseMsg.MutableHeaders()->ClearTrace();
         }
 
-        ProcessPostponeTime(CallContext, responseMsg);
+        ProcessThrottleTime(CallContext, responseMsg);
 
         Response.SetValue(std::move(responseMsg));
     }
@@ -418,7 +424,7 @@ class TZeroBlocksHandler final
         responseMsg.MutableHeaders()->ClearTrace();
         responseMsg.ClearDeprecatedTrace();
 
-        ProcessPostponeTime(CallContext, responseMsg);
+        ProcessThrottleTime(CallContext, responseMsg);
 
         Response.SetValue(std::move(responseMsg));
     }

cloud/blockstore/libs/service_rdma/rdma_target.cpp

@@ -67,6 +67,10 @@ void FillResponse(const TCallContextPtr& callContext, TResponse& response)
         callContext->Time(EProcessingStage::Postponed).MicroSeconds();
     response.SetDeprecatedThrottlerDelay(postponeTime);
     response.MutableHeaders()->MutableThrottler()->SetDelay(postponeTime);
+
+    const ui64 shapingTime =
+        callContext->Time(EProcessingStage::Shaping).MicroSeconds();
+    response.MutableHeaders()->MutableThrottler()->SetShapingDelay(shapingTime);
 }
 
 ////////////////////////////////////////////////////////////////////////////////

cloud/blockstore/libs/storage/service/volume_client_actor.cpp

@@ -378,15 +378,22 @@ void TVolumeClientActor::HandleResponse(
         THasGetDeprecatedThrottlerDelay<TProtoType>::value);
 
     if constexpr (RequiresThrottling<TMethod>) {
+        NProto::TThrottlerInfo& throttler =
+            *msg->Record.MutableHeaders()->MutableThrottler();
         const ui64 throttlerDelay =
             Max(msg->Record.GetDeprecatedThrottlerDelay(),
-                msg->Record.GetHeaders().GetThrottler().GetDelay());
+                throttler.GetDelay());
         it->second.CallContext->AddTime(
             EProcessingStage::Postponed,
             TDuration::MicroSeconds(throttlerDelay));
         msg->Record.SetDeprecatedThrottlerDelay(0);
-        msg->Record.MutableHeaders()->MutableThrottler()->SetDelay(0);
+        throttler.SetDelay(0);
         it->second.CallContext->SetPossiblePostponeDuration(TDuration::Zero());
+
+        it->second.CallContext->AddTime(
+            EProcessingStage::Shaping,
+            TDuration::MicroSeconds(throttler.GetShapingDelay()));
+        throttler.SetShapingDelay(0);
     }
 
     LWTRACK(

cloud/blockstore/libs/storage/volume/volume_actor_forward.cpp

@@ -31,7 +31,10 @@ namespace {
 Y_HAS_MEMBER(SetDeprecatedThrottlerDelay);
 
 template <typename TMethod>
-void StoreThrottlerDelay(typename TMethod::TResponse& response, TDuration delay)
+void StoreThrottlerDelay(
+    typename TMethod::TResponse& response,
+    TDuration delay,
+    TDuration shapingDelay)
 {
     using TProtoType = decltype(TMethod::TResponse::Record);
     static_assert(
@@ -41,6 +44,8 @@ void StoreThrottlerDelay(typename TMethod::TResponse& response, TDuration delay)
         response.Record.SetDeprecatedThrottlerDelay(delay.MicroSeconds());
         response.Record.MutableHeaders()->MutableThrottler()->SetDelay(
             delay.MicroSeconds());
+        response.Record.MutableHeaders()->MutableThrottler()->SetShapingDelay(
+            shapingDelay.MicroSeconds());
     }
 }
 
@@ -59,7 +64,8 @@ void RejectVolumeRequest(
 
     StoreThrottlerDelay<TMethod>(
         *response,
-        callContext.Time(EProcessingStage::Postponed));
+        callContext.Time(EProcessingStage::Postponed),
+        callContext.Time(EProcessingStage::Shaping));
 
     NCloud::Send(ctx, caller, std::move(response), callerCookie);
 }
@@ -321,7 +327,8 @@ void TVolumeActor::FillResponse(
 
     StoreThrottlerDelay<TMethod>(
         response,
-        callContext.Time(EProcessingStage::Postponed));
+        callContext.Time(EProcessingStage::Postponed),
+        callContext.Time(EProcessingStage::Shaping));
 }
 
 ////////////////////////////////////////////////////////////////////////////////

cloud/blockstore/libs/storage/volume_proxy/volume_proxy.cpp

@@ -826,15 +826,22 @@ void TVolumeProxyActor::HandleResponse(
         THasGetDeprecatedThrottlerDelay<TProtoType>::value ==
         RequiresThrottling<TMethod>);
     if constexpr (RequiresThrottling<TMethod>) {
+        NProto::TThrottlerInfo& throttler =
+            *msg->Record.MutableHeaders()->MutableThrottler();
         const ui64 throttlerDelay =
             Max(msg->Record.GetDeprecatedThrottlerDelay(),
-                msg->Record.GetHeaders().GetThrottler().GetDelay());
+                throttler.GetDelay());
         it->second.CallContext->AddTime(
             EProcessingStage::Postponed,
             TDuration::MicroSeconds(throttlerDelay));
         msg->Record.SetDeprecatedThrottlerDelay(0);
-        msg->Record.MutableHeaders()->MutableThrottler()->SetDelay(0);
+        throttler.SetDelay(0);
         it->second.CallContext->SetPossiblePostponeDuration(TDuration::Zero());
+
+        it->second.CallContext->AddTime(
+            EProcessingStage::Shaping,
+            TDuration::MicroSeconds(throttler.GetShapingDelay()));
+        throttler.SetShapingDelay(0);
     }
 
     LWTRACK(

cloud/storage/core/libs/common/context.cpp

@@ -30,13 +30,14 @@ TRequestTime TCallContextBase::CalcRequestTime(ui64 nowCycles) const
 
     const auto postponeDuration = Time(EProcessingStage::Postponed);
     const auto backoffTime = Time(EProcessingStage::Backoff);
+    const auto shapingTime = Time(EProcessingStage::Shaping);
 
     auto responseSentCycles = GetResponseSentCycles();
     auto responseDuration = CyclesToDurationSafe(
         (responseSentCycles ? responseSentCycles : nowCycles) - startCycles);
 
     requestTime.ExecutionTime = responseDuration - postponeDuration -
-        backoffTime - GetPossiblePostponeDuration();
+        backoffTime - shapingTime - GetPossiblePostponeDuration();
 
     return requestTime;
 }

cloud/storage/core/libs/common/context.h

@@ -13,7 +13,9 @@ enum class EProcessingStage
 {
     Postponed,
     Backoff,
-    Last
+    Shaping,
+
+    Last = Shaping,
 };
 
 struct TRequestTime
@@ -33,7 +35,7 @@ struct TCallContextBase
     : public TThrRefBase
 {
 private:
-    TAtomic Stage2Time[static_cast<int>(EProcessingStage::Last)] = {};
+    TAtomic Stage2Time[static_cast<int>(EProcessingStage::Last) + 1] = {};
     TAtomic RequestStartedCycles = 0;
     TAtomic ResponseSentCycles = 0;
     TAtomic PossiblePostponeMicroSeconds = 0;

cloud/storage/core/protos/throttler.proto

@@ -11,4 +11,7 @@ message TThrottlerInfo
 {
     // Time spent by a request in the throttler queue in us.
     uint64 Delay = 1;
+
+    // Time spent by a request in the shaping queue in us.
+    uint64 ShapingDelay = 2;
 }