From a558f943d0c69dbbdf41543b9991201ed0d672c9 Mon Sep 17 00:00:00 2001 From: Madhu Rajanna Date: Wed, 20 Nov 2024 10:07:38 +0100 Subject: [PATCH 1/4] rbd: take lock on targetpath during node operation We should not be dependent on the CO to ensure that it will serialize the request instead of that we need to have own internal locks to ensure that we dont do concurrent operations for same request. Signed-off-by: Madhu Rajanna (cherry picked from commit 38c0e64307d4fb108b95498d2220fe870f8440c8) --- internal/rbd/nodeserver.go | 18 ++++++++++++++---- internal/util/idlocker.go | 3 +++ 2 files changed, 17 insertions(+), 4 deletions(-) diff --git a/internal/rbd/nodeserver.go b/internal/rbd/nodeserver.go index 5f4dd000739..bb5c8deae1c 100644 --- a/internal/rbd/nodeserver.go +++ b/internal/rbd/nodeserver.go @@ -709,8 +709,12 @@ func (ns *NodeServer) NodePublishVolume( volID := req.GetVolumeId() stagingPath += "/" + volID - // Considering kubelet make sure the stage and publish operations - // are serialized, we dont need any extra locking in nodePublish + if acquired := ns.VolumeLocks.TryAcquire(targetPath); !acquired { + log.ErrorLog(ctx, util.TargetPathOperationAlreadyExistsFmt, targetPath) + + return nil, status.Errorf(codes.Aborted, util.TargetPathOperationAlreadyExistsFmt, targetPath) + } + defer ns.VolumeLocks.Release(targetPath) // Check if that target path exists properly notMnt, err := ns.createTargetMountPath(ctx, targetPath, isBlock) @@ -914,8 +918,14 @@ func (ns *NodeServer) NodeUnpublishVolume( } targetPath := req.GetTargetPath() - // considering kubelet make sure node operations like unpublish/unstage...etc can not be called - // at same time, an explicit locking at time of nodeunpublish is not required. + + if acquired := ns.VolumeLocks.TryAcquire(targetPath); !acquired { + log.ErrorLog(ctx, util.TargetPathOperationAlreadyExistsFmt, targetPath) + + return nil, status.Errorf(codes.Aborted, util.TargetPathOperationAlreadyExistsFmt, targetPath) + } + defer ns.VolumeLocks.Release(targetPath) + isMnt, err := ns.Mounter.IsMountPoint(targetPath) if err != nil { if os.IsNotExist(err) { diff --git a/internal/util/idlocker.go b/internal/util/idlocker.go index 92733c19c9c..211081a1372 100644 --- a/internal/util/idlocker.go +++ b/internal/util/idlocker.go @@ -28,6 +28,9 @@ const ( // SnapshotOperationAlreadyExistsFmt string format to return for concurrent operation. SnapshotOperationAlreadyExistsFmt = "an operation with the given Snapshot ID %s already exists" + + // TargetPathOperationAlreadyExistsFmt string format to return for concurrent operation on target path. + TargetPathOperationAlreadyExistsFmt = "an operation with the given target path %s already exists" ) // VolumeLocks implements a map with atomic operations. It stores a set of all volume IDs From fd1fcb7725f1b0851caae4f69b717ce9b6d32c35 Mon Sep 17 00:00:00 2001 From: Madhu Rajanna Date: Wed, 20 Nov 2024 10:12:59 +0100 Subject: [PATCH 2/4] cephfs: take lock on targetpath on node operation We should not be dependent on the CO to ensure that it will serialize the request instead of that we need to have own internal locks to ensure that we dont do concurrent operations for same request. Signed-off-by: Madhu Rajanna (cherry picked from commit 38b0a4cbadfb20e818e76cc0d3b603274513ec8e) --- internal/cephfs/nodeserver.go | 21 +++++++++++++++------ 1 file changed, 15 insertions(+), 6 deletions(-) diff --git a/internal/cephfs/nodeserver.go b/internal/cephfs/nodeserver.go index 44c06860153..a97cc09196b 100644 --- a/internal/cephfs/nodeserver.go +++ b/internal/cephfs/nodeserver.go @@ -494,6 +494,13 @@ func (ns *NodeServer) NodePublishVolume( targetPath := req.GetTargetPath() volID := fsutil.VolumeID(req.GetVolumeId()) + if acquired := ns.VolumeLocks.TryAcquire(targetPath); !acquired { + log.ErrorLog(ctx, util.TargetPathOperationAlreadyExistsFmt, targetPath) + + return nil, status.Errorf(codes.Aborted, util.TargetPathOperationAlreadyExistsFmt, targetPath) + } + defer ns.VolumeLocks.Release(targetPath) + volOptions := &store.VolumeOptions{} defer volOptions.Destroy() @@ -506,9 +513,6 @@ func (ns *NodeServer) NodePublishVolume( return nil, status.Errorf(codes.Internal, "failed to create mounter for volume %s: %v", volID, err.Error()) } - // Considering kubelet make sure the stage and publish operations - // are serialized, we dont need any extra locking in nodePublish - if err = util.CreateMountPoint(targetPath); err != nil { log.ErrorLog(ctx, "failed to create mount point at %s: %v", targetPath, err) @@ -599,12 +603,17 @@ func (ns *NodeServer) NodeUnpublishVolume( return nil, err } - // considering kubelet make sure node operations like unpublish/unstage...etc can not be called - // at same time, an explicit locking at time of nodeunpublish is not required. targetPath := req.GetTargetPath() + volID := req.GetVolumeId() + if acquired := ns.VolumeLocks.TryAcquire(targetPath); !acquired { + log.ErrorLog(ctx, util.TargetPathOperationAlreadyExistsFmt, targetPath) + + return nil, status.Errorf(codes.Aborted, util.TargetPathOperationAlreadyExistsFmt, targetPath) + } + defer ns.VolumeLocks.Release(targetPath) // stop the health-checker that may have been started in NodeGetVolumeStats() - ns.healthChecker.StopChecker(req.GetVolumeId(), targetPath) + ns.healthChecker.StopChecker(volID, targetPath) isMnt, err := util.IsMountPoint(ns.Mounter, targetPath) if err != nil { From 4e9753a672332c6c0fa6f813ad90731562bd82be Mon Sep 17 00:00:00 2001 From: Madhu Rajanna Date: Wed, 20 Nov 2024 10:14:26 +0100 Subject: [PATCH 3/4] cephfs: use os.Remove to remove directory using os.RemoveAll will remove everything in the director after the Umount we should be using os.Remove only to remove the empty directory Signed-off-by: Madhu Rajanna (cherry picked from commit ffa8eaf5ddc3a6f3041bf010f76e7d71c0a08b8d) --- internal/cephfs/nodeserver.go | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/internal/cephfs/nodeserver.go b/internal/cephfs/nodeserver.go index a97cc09196b..b3a02ea321c 100644 --- a/internal/cephfs/nodeserver.go +++ b/internal/cephfs/nodeserver.go @@ -636,7 +636,7 @@ func (ns *NodeServer) NodeUnpublishVolume( isMnt = true } if !isMnt { - if err = os.RemoveAll(targetPath); err != nil { + if err = os.Remove(targetPath); err != nil { return nil, status.Error(codes.Internal, err.Error()) } From a58ee1c3602199bd8c9925d3de602f299ce5a395 Mon Sep 17 00:00:00 2001 From: Madhu Rajanna Date: Wed, 20 Nov 2024 10:15:40 +0100 Subject: [PATCH 4/4] rbd: use os.Remove to remove directory using os.RemoveAll will remove everything in the director after the Umount we should be using os.Remove only to remove the empty directory Signed-off-by: Madhu Rajanna (cherry picked from commit 39cc628adf1da1ab78f3aacfb7fe8f53e9a1fd29) --- internal/rbd/nodeserver.go | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/internal/rbd/nodeserver.go b/internal/rbd/nodeserver.go index bb5c8deae1c..92a92a561be 100644 --- a/internal/rbd/nodeserver.go +++ b/internal/rbd/nodeserver.go @@ -938,7 +938,7 @@ func (ns *NodeServer) NodeUnpublishVolume( return nil, status.Error(codes.NotFound, err.Error()) } if !isMnt { - if err = os.RemoveAll(targetPath); err != nil { + if err = os.Remove(targetPath); err != nil { return nil, status.Error(codes.Internal, err.Error()) } @@ -949,7 +949,7 @@ func (ns *NodeServer) NodeUnpublishVolume( return nil, status.Error(codes.Internal, err.Error()) } - if err = os.RemoveAll(targetPath); err != nil { + if err = os.Remove(targetPath); err != nil { return nil, status.Error(codes.Internal, err.Error()) }