Add PDisk error reason to VDisk's PDisk error state (#9302)

4 files changed, 71 insertions, 42 deletions

diff --git a/ydb/core/blobstorage/vdisk/common/vdisk_context.h b/ydb/core/blobstorage/vdisk/common/vdisk_context.h
index 24faf02ce4..d0b463c38f 100644
--- a/ydb/core/blobstorage/vdisk/common/vdisk_context.h
+++ b/ydb/core/blobstorage/vdisk/common/vdisk_context.h
@@ -80,8 +80,6 @@ namespace NKikimr {
         TOutOfSpaceState OutOfSpaceState;
         // Global stat about huge heap fragmentation
         THugeHeapFragmentation HugeHeapFragmentation;
-        // Tracks PDisk errors
-        TPDiskErrorState PDiskErrorState;
         friend class TDskSpaceTrackerActor;
 
         NMonGroup::TCostGroup CostMonGroup;
@@ -89,10 +87,6 @@ namespace NKikimr {
     public:
         TLogger Logger;
 
-        TPDiskErrorState::EState GetPDiskErrorState() const {
-            return PDiskErrorState.GetState();
-        }
-
     public:
         TVDiskContext(
                 const TActorId &vdiskActorId,
@@ -137,14 +131,11 @@ namespace NKikimr {
                 case NKikimrProto::CORRUPTED:
                 case NKikimrProto::OUT_OF_SPACE: {
                     // Device is out of order
-                    PDiskErrorState.Set(ev.Status, ev.StatusFlags);
-                    auto newState = PDiskErrorState.GetState();
                     LOG_ERROR(actorSystemOrCtx, NKikimrServices::BS_VDISK_OTHER,
                             VDISKP(VDiskLogPrefix,
-                                "CheckPDiskResponse: Recoverable error from PDisk: %s newState# %s",
-                                FormatMessage(ev.Status, ev.ErrorReason, ev.StatusFlags, message).data(),
-                                TPDiskErrorState::StateToString(newState)));
-                    actorSystemOrCtx.Send(VDiskActorId, new TEvPDiskErrorStateChange(newState));
+                                "CheckPDiskResponse: Recoverable error from PDisk: %s",
+                                FormatMessage(ev.Status, ev.ErrorReason, ev.StatusFlags, message).data()));
+                    actorSystemOrCtx.Send(VDiskActorId, new TEvPDiskErrorStateChange(ev.Status, ev.StatusFlags, ev.ErrorReason));
                     return false;
                 }
                 default:
diff --git a/ydb/core/blobstorage/vdisk/common/vdisk_pdisk_error.h b/ydb/core/blobstorage/vdisk/common/vdisk_pdisk_error.h
index d4474b21f7..86c83cc03e 100644
--- a/ydb/core/blobstorage/vdisk/common/vdisk_pdisk_error.h
+++ b/ydb/core/blobstorage/vdisk/common/vdisk_pdisk_error.h
@@ -1,8 +1,10 @@
 #pragma once
+
 #include "defs.h"
-#include <ydb/core/base/blobstorage.h>
 #include "vdisk_config.h"
 
+#include <ydb/core/base/blobstorage.h>
+
 namespace NKikimr {
 
     ////////////////////////////////////////////////////////////////////////////
@@ -31,43 +33,52 @@ namespace NKikimr {
         }
 
         TPDiskErrorState() {
-            SetPrivate(Good);
+            SetPrivate(Good, "");
         }
 
         EState GetState() const {
-            return static_cast<EState>(AtomicGet(State));
+            return State;
+        }
+
+        const TString& GetErrorReason() const {
+            return ErrorReason;
         }
 
         // We call this function when PDisk returned ERROR and we pass pdiskFlags to set the correct state
-        EState Set(NKikimrProto::EReplyStatus status, ui32 pdiskFlags) {
+        EState Set(NKikimrProto::EReplyStatus status, ui32 pdiskFlags, const TString& errorReason) {
             switch (status) {
                 case NKikimrProto::CORRUPTED:
-                    return SetPrivate(NoWrites);
+                    return SetPrivate(NoWrites, errorReason);
                 case NKikimrProto::OUT_OF_SPACE:
                     // check flags additionally
                     Y_ABORT_UNLESS(pdiskFlags & NKikimrBlobStorage::StatusNotEnoughDiskSpaceForOperation);
-                    return SetPrivate(WriteOnlyLog);
+                    return SetPrivate(WriteOnlyLog, errorReason);
                 default:
                     Y_ABORT("Unexpected state# %s", NKikimrProto::EReplyStatus_Name(status).data());
             }
         }
 
-    private:
-        TAtomic State = 0;
-
-        EState SetPrivate(EState state) {
-            // make sure bad state increments (not decrements), use CAS for that
-            while (true) {
-                EState curState = GetState();
-                if (state > curState) {
-                    // if state is worse than curState:
-                    TAtomicBase newState = static_cast<TAtomicBase>(state);
-                    bool done = AtomicCas(&State, newState, curState);
-                    if (done)
-                        return state;
-                } else
-                    return curState;
+        TString ToString() const {
+            TStringStream ss;
+            ss << "State# " << StateToString(State);
+            if (!ErrorReason.empty()) {
+                ss << ", PDiskError# " << ErrorReason;
             }
+            return ss.Str();
+        }
+
+    private:
+        EState State = EState::Unspecified;
+
+        TString ErrorReason;
+
+        EState SetPrivate(EState state, const TString& errorReason) {
+            if (state > State) {
+                State = state;
+                ErrorReason = errorReason;
+                return state;
+            } else
+                return State;
         }
     };
 
@@ -80,10 +91,14 @@ namespace NKikimr {
     struct TEvPDiskErrorStateChange :
         public TEventLocal<TEvPDiskErrorStateChange, TEvBlobStorage::EvPDiskErrorStateChange>
     {
-        const TPDiskErrorState::EState State;
+        const NKikimrProto::EReplyStatus Status;
+        const ui32 PDiskFlags;
+        const TString ErrorReason;
 
-        TEvPDiskErrorStateChange(TPDiskErrorState::EState state)
-            : State(state)
+        TEvPDiskErrorStateChange(NKikimrProto::EReplyStatus status, ui32 pdiskFlags, const TString &errorReason)
+            : Status(status)
+            , PDiskFlags(pdiskFlags)
+            , ErrorReason(errorReason)
         {}
     };
 
diff --git a/ydb/core/blobstorage/vdisk/common/vdisk_pdisk_error_ut.cpp b/ydb/core/blobstorage/vdisk/common/vdisk_pdisk_error_ut.cpp
index 3b6103cfa3..f3eb58f0bf 100644
--- a/ydb/core/blobstorage/vdisk/common/vdisk_pdisk_error_ut.cpp
+++ b/ydb/core/blobstorage/vdisk/common/vdisk_pdisk_error_ut.cpp
@@ -16,20 +16,39 @@ namespace NKikimr {
         Y_UNIT_TEST(Basic) {
             TPDiskErrorState state;
             UNIT_ASSERT(state.GetState() == TPDiskErrorState::Good);
+            UNIT_ASSERT(state.GetErrorReason().Empty());
 
-            state.Set(NKikimrProto::CORRUPTED, 0);
+            state.Set(NKikimrProto::CORRUPTED, 0, "");
             UNIT_ASSERT(state.GetState() == TPDiskErrorState::NoWrites);
+            UNIT_ASSERT(state.GetErrorReason().Empty());
         }
 
         Y_UNIT_TEST(Basic2) {
             TPDiskErrorState state;
             UNIT_ASSERT(state.GetState() == TPDiskErrorState::Good);
+            UNIT_ASSERT(state.GetErrorReason().Empty());
 
-            state.Set(NKikimrProto::OUT_OF_SPACE, NKikimrBlobStorage::StatusNotEnoughDiskSpaceForOperation);
+            state.Set(NKikimrProto::OUT_OF_SPACE, NKikimrBlobStorage::StatusNotEnoughDiskSpaceForOperation, "");
             UNIT_ASSERT(state.GetState() == TPDiskErrorState::WriteOnlyLog);
+            UNIT_ASSERT(state.GetErrorReason().Empty());
 
-            state.Set(NKikimrProto::CORRUPTED, 0);
+            state.Set(NKikimrProto::CORRUPTED, 0, "");
             UNIT_ASSERT(state.GetState() == TPDiskErrorState::NoWrites);
+            UNIT_ASSERT(state.GetErrorReason().Empty());
+        }
+
+        Y_UNIT_TEST(BasicErrorReason) {
+            TPDiskErrorState state;
+            UNIT_ASSERT(state.GetState() == TPDiskErrorState::Good);
+            UNIT_ASSERT(state.GetErrorReason().Empty());
+
+            state.Set(NKikimrProto::OUT_OF_SPACE, NKikimrBlobStorage::StatusNotEnoughDiskSpaceForOperation, "Foo");
+            UNIT_ASSERT(state.GetState() == TPDiskErrorState::WriteOnlyLog);
+            UNIT_ASSERT(state.GetErrorReason() == "Foo");
+
+            state.Set(NKikimrProto::CORRUPTED, 0, "Bar");
+            UNIT_ASSERT(state.GetState() == TPDiskErrorState::NoWrites);
+            UNIT_ASSERT(state.GetErrorReason() == "Bar");
         }
 
     }
diff --git a/ydb/core/blobstorage/vdisk/skeleton/blobstorage_skeletonfront.cpp b/ydb/core/blobstorage/vdisk/skeleton/blobstorage_skeletonfront.cpp
index 1ef0e41a60..f360860dc4 100644
--- a/ydb/core/blobstorage/vdisk/skeleton/blobstorage_skeletonfront.cpp
+++ b/ydb/core/blobstorage/vdisk/skeleton/blobstorage_skeletonfront.cpp
@@ -636,6 +636,7 @@ namespace NKikimr {
         friend class TIntQueueClass;
 
         TVDiskContextPtr VCtx;
+        TPDiskErrorState PDiskErrorState;
         TIntrusivePtr<TVDiskConfig> Config;
         TIntrusivePtr<TBlobStorageGroupInfo> GInfo;
         std::shared_ptr<TBlobStorageGroupInfo::TTopology> Top;
@@ -934,7 +935,7 @@ namespace NKikimr {
                                         TABLED() {str << "Error Details";}
                                         TABLED() {
                                             str << "PDisk reported error: "
-                                                << TPDiskErrorState::StateToString(VCtx->GetPDiskErrorState());
+                                                << PDiskErrorState.ToString();
                                         }
                                     }
                                 } else if (VCtx->LocalRecoveryErrorStr) {
@@ -1663,12 +1664,15 @@ namespace NKikimr {
         }
 
         void Handle(TEvPDiskErrorStateChange::TPtr &ev, const TActorContext &ctx) {
+            auto errorStateChange = ev->Get();
+            
+            PDiskErrorState.Set(errorStateChange->Status, errorStateChange->PDiskFlags, errorStateChange->ErrorReason);
+
             LOG_ERROR_S(ctx, NKikimrServices::BS_SKELETON, VCtx->VDiskLogPrefix
                     << "SkeletonFront: got TEvPDiskErrorStateChange;"
-                    << " state# " << TPDiskErrorState::StateToString(ev->Get()->State)
+                    << PDiskErrorState.ToString()
                     << " Marker# BSVSF03");
 
-
             // switch skeleton state to PDiskError
             SkeletonFrontGroup->ResetCounters();
             VDiskMonGroup.VDiskState(NKikimrWhiteboard::EVDiskState::PDiskError);