Merge pull request #1327 from judemars/automated-cherry-pick-of-#1298-upstream-release-1.8

mattcary · web-flow · commit 754d0962b26c · 2023-09-27T12:07:49.000-07:00
[release-1.8] Automated cherry pick of #1298: Use original error code when backing off
diff --git a/pkg/common/utils.go b/pkg/common/utils.go
@@ -79,6 +79,14 @@ var (
 	// Full or partial URL of the machine type resource, in the format:
 	//   zones/zone/machineTypes/machine-type
 	machineTypeRegex = regexp.MustCompile(machineTypePattern)
+
+	// userErrorCodeMap tells how API error types are translated to error codes.
+	userErrorCodeMap = map[int]codes.Code{
+		http.StatusForbidden:       codes.PermissionDenied,
+		http.StatusBadRequest:      codes.InvalidArgument,
+		http.StatusTooManyRequests: codes.ResourceExhausted,
+		http.StatusNotFound:        codes.NotFound,
+	}
 )
 
 func BytesToGbRoundDown(bytes int64) int64 {
@@ -296,82 +304,63 @@ func ParseMachineType(machineTypeUrl string) (string, error) {
 	return machineType[1], nil
 }
 
-// CodeForError returns a pointer to the grpc error code that maps to the http
-// error code for the passed in user googleapi error or context error. Returns
-// codes.Internal if the given error is not a googleapi error caused by the user.
-// The following http error codes are considered user errors:
-// (1) http 400 Bad Request, returns grpc InvalidArgument,
-// (2) http 403 Forbidden, returns grpc PermissionDenied,
-// (3) http 404 Not Found, returns grpc NotFound
-// (4) http 429 Too Many Requests, returns grpc ResourceExhausted
-// The following errors are considered context errors:
-// (1) "context deadline exceeded", returns grpc DeadlineExceeded,
-// (2) "context canceled", returns grpc Canceled
-func CodeForError(err error) *codes.Code {
-	if err == nil {
-		return nil
+// CodeForError returns the grpc error code that maps to the http error code for the
+// passed in user googleapi error or context error. Returns codes.Internal if the given
+// error is not a googleapi error caused by the user. userErrorCodeMap is used for
+// encoding most errors.
+func CodeForError(sourceError error) codes.Code {
+	if sourceError == nil {
+		return codes.Internal
 	}
 
-	if errCode := existingErrorCode(err); errCode != nil {
-		return errCode
+	if code, err := existingErrorCode(sourceError); err == nil {
+		return code
 	}
-	if code := isContextError(err); code != nil {
+	if code, err := isContextError(sourceError); err == nil {
 		return code
 	}
 
-	internalErrorCode := codes.Internal
-	// Upwrap the error
 	var apiErr *googleapi.Error
-	if !errors.As(err, &apiErr) {
-		return &internalErrorCode
-	}
-
-	userErrors := map[int]codes.Code{
-		http.StatusForbidden:       codes.PermissionDenied,
-		http.StatusBadRequest:      codes.InvalidArgument,
-		http.StatusTooManyRequests: codes.ResourceExhausted,
-		http.StatusNotFound:        codes.NotFound,
+	if !errors.As(sourceError, &apiErr) {
+		return codes.Internal
 	}
-	if code, ok := userErrors[apiErr.Code]; ok {
-		return &code
+	if code, ok := userErrorCodeMap[apiErr.Code]; ok {
+		return code
 	}
 
-	return &internalErrorCode
+	return codes.Internal
 }
 
-// isContextError returns a pointer to the grpc error code DeadlineExceeded
-// if the passed in error contains the "context deadline exceeded" string and returns
-// the grpc error code Canceled if the error contains the "context canceled" string.
-func isContextError(err error) *codes.Code {
+// isContextError returns the grpc error code DeadlineExceeded if the passed in error
+// contains the "context deadline exceeded" string and returns the grpc error code
+// Canceled if the error contains the "context canceled" string. It returns and error if
+// err isn't a context error.
+func isContextError(err error) (codes.Code, error) {
 	if err == nil {
-		return nil
+		return codes.Unknown, fmt.Errorf("null error")
 	}
 
 	errStr := err.Error()
 	if strings.Contains(errStr, context.DeadlineExceeded.Error()) {
-		return errCodePtr(codes.DeadlineExceeded)
+		return codes.DeadlineExceeded, nil
 	}
 	if strings.Contains(errStr, context.Canceled.Error()) {
-		return errCodePtr(codes.Canceled)
+		return codes.Canceled, nil
 	}
-	return nil
+	return codes.Unknown, fmt.Errorf("Not a context error: %w", err)
 }
 
-func existingErrorCode(err error) *codes.Code {
+func existingErrorCode(err error) (codes.Code, error) {
 	if err == nil {
-		return nil
+		return codes.Unknown, fmt.Errorf("null error")
 	}
 	if status, ok := status.FromError(err); ok {
-		return errCodePtr(status.Code())
+		return status.Code(), nil
 	}
-	return nil
-}
-
-func errCodePtr(code codes.Code) *codes.Code {
-	return &code
+	return codes.Unknown, fmt.Errorf("no existing error code for %w", err)
 }
 
 func LoggedError(msg string, err error) error {
 	klog.Errorf(msg+"%v", err.Error())
-	return status.Errorf(*CodeForError(err), msg+"%v", err.Error())
+	return status.Errorf(CodeForError(err), msg+"%v", err.Error())
 }
diff --git a/pkg/common/utils_test.go b/pkg/common/utils_test.go
@@ -867,57 +867,51 @@ func TestParseMachineType(t *testing.T) {
 }
 
 func TestCodeForError(t *testing.T) {
-	internalErrorCode := codes.Internal
-	userErrorCode := codes.InvalidArgument
 	testCases := []struct {
 		name     string
 		inputErr error
-		expCode  *codes.Code
+		expCode  codes.Code
 	}{
 		{
 			name:     "Not googleapi.Error",
 			inputErr: errors.New("I am not a googleapi.Error"),
-			expCode:  &internalErrorCode,
+			expCode:  codes.Internal,
 		},
 		{
 			name:     "User error",
 			inputErr: &googleapi.Error{Code: http.StatusBadRequest, Message: "User error with bad request"},
-			expCode:  &userErrorCode,
+			expCode:  codes.InvalidArgument,
 		},
 		{
 			name:     "googleapi.Error but not a user error",
 			inputErr: &googleapi.Error{Code: http.StatusInternalServerError, Message: "Internal error"},
-			expCode:  &internalErrorCode,
+			expCode:  codes.Internal,
 		},
 		{
 			name:     "context canceled error",
 			inputErr: context.Canceled,
-			expCode:  errCodePtr(codes.Canceled),
+			expCode:  codes.Canceled,
 		},
 		{
 			name:     "context deadline exceeded error",
 			inputErr: context.DeadlineExceeded,
-			expCode:  errCodePtr(codes.DeadlineExceeded),
+			expCode:  codes.DeadlineExceeded,
 		},
 		{
 			name:     "status error with Aborted error code",
 			inputErr: status.Error(codes.Aborted, "aborted error"),
-			expCode:  errCodePtr(codes.Aborted),
+			expCode:  codes.Aborted,
 		},
 		{
 			name:     "nil error",
 			inputErr: nil,
-			expCode:  nil,
+			expCode:  codes.Internal,
 		},
 	}
 
 	for _, tc := range testCases {
-		t.Logf("Running test: %v", tc.name)
 		errCode := CodeForError(tc.inputErr)
-		if (tc.expCode == nil) != (errCode == nil) {
-			t.Errorf("test %v failed: got %v, expected %v", tc.name, errCode, tc.expCode)
-		}
-		if tc.expCode != nil && *errCode != *tc.expCode {
+		if errCode != tc.expCode {
 			t.Errorf("test %v failed: got %v, expected %v", tc.name, errCode, tc.expCode)
 		}
 	}
@@ -927,46 +921,48 @@ func TestIsContextError(t *testing.T) {
 	cases := []struct {
 		name            string
 		err             error
-		expectedErrCode *codes.Code
+		expectedErrCode codes.Code
+		expectError     bool
 	}{
 		{
 			name:            "deadline exceeded error",
 			err:             context.DeadlineExceeded,
-			expectedErrCode: errCodePtr(codes.DeadlineExceeded),
+			expectedErrCode: codes.DeadlineExceeded,
 		},
 		{
 			name:            "contains 'context deadline exceeded'",
 			err:             fmt.Errorf("got error: %w", context.DeadlineExceeded),
-			expectedErrCode: errCodePtr(codes.DeadlineExceeded),
+			expectedErrCode: codes.DeadlineExceeded,
 		},
 		{
 			name:            "context canceled error",
 			err:             context.Canceled,
-			expectedErrCode: errCodePtr(codes.Canceled),
+			expectedErrCode: codes.Canceled,
 		},
 		{
 			name:            "contains 'context canceled'",
 			err:             fmt.Errorf("got error: %w", context.Canceled),
-			expectedErrCode: errCodePtr(codes.Canceled),
+			expectedErrCode: codes.Canceled,
 		},
 		{
-			name:            "does not contain 'context canceled' or 'context deadline exceeded'",
-			err:             fmt.Errorf("unknown error"),
-			expectedErrCode: nil,
+			name:        "does not contain 'context canceled' or 'context deadline exceeded'",
+			err:         fmt.Errorf("unknown error"),
+			expectError: true,
 		},
 		{
-			name:            "nil error",
-			err:             nil,
-			expectedErrCode: nil,
+			name:        "nil error",
+			err:         nil,
+			expectError: true,
 		},
 	}
 
 	for _, test := range cases {
-		errCode := isContextError(test.err)
-		if (test.expectedErrCode == nil) != (errCode == nil) {
-			t.Errorf("test %v failed: got %v, expected %v", test.name, errCode, test.expectedErrCode)
-		}
-		if test.expectedErrCode != nil && *errCode != *test.expectedErrCode {
+		errCode, err := isContextError(test.err)
+		if test.expectError {
+			if err == nil {
+				t.Errorf("test %v failed, expected error, got %v", test.name, errCode)
+			}
+		} else if errCode != test.expectedErrCode {
 			t.Errorf("test %v failed: got %v, expected %v", test.name, errCode, test.expectedErrCode)
 		}
 	}
diff --git a/pkg/gce-pd-csi-driver/controller.go b/pkg/gce-pd-csi-driver/controller.go
@@ -91,10 +91,12 @@ type GCEControllerServer struct {
 	errorBackoff *csiErrorBackoff
 }
 
+type csiErrorBackoffId string
+
 type csiErrorBackoff struct {
-	backoff *flowcontrol.Backoff
+	backoff    *flowcontrol.Backoff
+	errorCodes map[csiErrorBackoffId]codes.Code
 }
-type csiErrorBackoffId string
 
 type workItem struct {
 	ctx          context.Context
@@ -482,13 +484,13 @@ func (gceCS *GCEControllerServer) ControllerPublishVolume(ctx context.Context, r
 
 	backoffId := gceCS.errorBackoff.backoffId(req.NodeId, req.VolumeId)
 	if gceCS.errorBackoff.blocking(backoffId) {
-		return nil, status.Errorf(codes.Unavailable, "ControllerPublish not permitted on node %q due to backoff condition", req.NodeId)
+		return nil, status.Errorf(gceCS.errorBackoff.code(backoffId), "ControllerPublish not permitted on node %q due to backoff condition", req.NodeId)
 	}
 
 	resp, err, diskTypeForMetric := gceCS.executeControllerPublishVolume(ctx, req)
 	if err != nil {
-		klog.Infof("For node %s adding backoff due to error for volume %s: %v", req.NodeId, req.VolumeId, err.Error())
-		gceCS.errorBackoff.next(backoffId)
+		klog.Infof("For node %s adding backoff due to error for volume %s: %v", req.NodeId, req.VolumeId, err)
+		gceCS.errorBackoff.next(backoffId, common.CodeForError(err))
 	} else {
 		klog.Infof("For node %s clear backoff due to successful publish of volume %v", req.NodeId, req.VolumeId)
 		gceCS.errorBackoff.reset(backoffId)
@@ -642,12 +644,12 @@ func (gceCS *GCEControllerServer) ControllerUnpublishVolume(ctx context.Context,
 	// Only valid requests will be queued
 	backoffId := gceCS.errorBackoff.backoffId(req.NodeId, req.VolumeId)
 	if gceCS.errorBackoff.blocking(backoffId) {
-		return nil, status.Errorf(codes.Unavailable, "ControllerUnpublish not permitted on node %q due to backoff condition", req.NodeId)
+		return nil, status.Errorf(gceCS.errorBackoff.code(backoffId), "ControllerUnpublish not permitted on node %q due to backoff condition", req.NodeId)
 	}
 	resp, err, diskTypeForMetric := gceCS.executeControllerUnpublishVolume(ctx, req)
 	if err != nil {
-		klog.Infof("For node %s adding backoff due to error for volume %s", req.NodeId, req.VolumeId)
-		gceCS.errorBackoff.next(backoffId)
+		klog.Infof("For node %s adding backoff due to error for volume %s: %v", req.NodeId, req.VolumeId, err)
+		gceCS.errorBackoff.next(backoffId, common.CodeForError(err))
 	} else {
 		klog.Infof("For node %s clear backoff due to successful unpublish of volume %v", req.NodeId, req.VolumeId)
 		gceCS.errorBackoff.reset(backoffId)
@@ -1775,7 +1777,7 @@ func pickRandAndConsecutive(slice []string, n int) ([]string, error) {
 }
 
 func newCsiErrorBackoff(initialDuration, errorBackoffMaxDuration time.Duration) *csiErrorBackoff {
-	return &csiErrorBackoff{flowcontrol.NewBackOff(initialDuration, errorBackoffMaxDuration)}
+	return &csiErrorBackoff{flowcontrol.NewBackOff(initialDuration, errorBackoffMaxDuration), make(map[csiErrorBackoffId]codes.Code)}
 }
 
 func (_ *csiErrorBackoff) backoffId(nodeId, volumeId string) csiErrorBackoffId {
@@ -1787,10 +1789,22 @@ func (b *csiErrorBackoff) blocking(id csiErrorBackoffId) bool {
 	return blk
 }
 
-func (b *csiErrorBackoff) next(id csiErrorBackoffId) {
+func (b *csiErrorBackoff) code(id csiErrorBackoffId) codes.Code {
+	if code, ok := b.errorCodes[id]; ok {
+		return code
+	}
+	// If we haven't recorded a code, return unavailable, which signals a problem with the driver
+	// (ie, next() wasn't called correctly).
+	klog.Errorf("using default code for %s", id)
+	return codes.Unavailable
+}
+
+func (b *csiErrorBackoff) next(id csiErrorBackoffId, code codes.Code) {
 	b.backoff.Next(string(id), b.backoff.Clock.Now())
+	b.errorCodes[id] = code
 }
 
 func (b *csiErrorBackoff) reset(id csiErrorBackoffId) {
 	b.backoff.Reset(string(id))
+	delete(b.errorCodes, id)
 }
diff --git a/pkg/gce-pd-csi-driver/controller_test.go b/pkg/gce-pd-csi-driver/controller_test.go
