dynamic host volumes: account for other claims in capability check

When we feasibility check a dynamic host volume against a volume request, we
check the attachment mode and access mode. This only ensures that the
capabilities match, but doesn't enforce the semantics of the capabilities
against other claims that may be made on the allocation.

Add support for checking the requested capability against other allocations that
the volume claimed.

Ref: #24479

nomad/state/state_store_host_volumes.go

@@ -30,6 +30,7 @@ func (s *StateStore) HostVolumeByID(ws memdb.WatchSet, ns, id string, withAllocs
 
 	vol = vol.Copy()
 	vol.Allocations = []*structs.AllocListStub{}
+	vol.Writers = []string{}
 
 	// we can't use AllocsByNodeTerminal because we only want to filter out
 	// allocs that are client-terminal, not server-terminal
@@ -43,6 +44,9 @@ func (s *StateStore) HostVolumeByID(ws memdb.WatchSet, ns, id string, withAllocs
 		}
 		for _, volClaim := range alloc.Job.LookupTaskGroup(alloc.TaskGroup).Volumes {
 			if volClaim.Type == structs.VolumeTypeHost && volClaim.Source == vol.Name {
+				if !volClaim.ReadOnly {
+					vol.Writers = append(vol.Writers, alloc.ID)
+				}
 				vol.Allocations = append(vol.Allocations, alloc.Stub(nil))
 			}
 		}
@@ -101,6 +105,7 @@ func (s *StateStore) UpsertHostVolume(index uint64, vol *structs.HostVolume) err
 	// Allocations are denormalized on read, so we don't want these to be
 	// written to the state store.
 	vol.Allocations = nil
+	vol.Writers = nil
 	vol.ModifyIndex = index
 
 	err = txn.Insert(TableHostVolumes, vol)

nomad/structs/host_volumes.go

@@ -84,6 +84,11 @@ type HostVolume struct {
 	// this host volume. They are denormalized on read and this field will be
 	// never written to Raft
 	Allocations []*AllocListStub `json:",omitempty"`
+
+	// Writers and is a list of alloc IDs from the Allocations field that can
+	// write to this volume. This count is denormalized on read, never written
+	// to Raft, and never returned by the API.
+	Writers []string `json:"-"`
 }
 
 type HostVolumeState string

scheduler/feasible.go

@@ -138,10 +138,11 @@ func NewRandomIterator(ctx Context, nodes []*structs.Node) *StaticIterator {
 // HostVolumeChecker is a FeasibilityChecker which returns whether a node has
 // the host volumes necessary to schedule a task group.
 type HostVolumeChecker struct {
-	ctx           Context
-	volumeReqs    []*structs.VolumeRequest
-	hostVolumeIDs []string
-	namespace     string
+	ctx                  Context
+	volumeReqs           []*structs.VolumeRequest
+	hostVolumeIDs        []string
+	namespace            string
+	previousAllocationID string
 }
 
 // NewHostVolumeChecker creates a HostVolumeChecker from a set of volumes
@@ -155,11 +156,11 @@ func NewHostVolumeChecker(ctx Context) *HostVolumeChecker {
 
 // SetVolumes takes the volumes required by a task group and updates the checker.
 func (h *HostVolumeChecker) SetVolumes(
-	allocName, ns string, volumes map[string]*structs.VolumeRequest, allocHostVolumeIDs []string,
-) {
+	allocName, ns string, volumes map[string]*structs.VolumeRequest, allocHostVolumeIDs []string, prevAllocID string) {
 	h.namespace = ns
 	h.volumeReqs = []*structs.VolumeRequest{}
 	h.hostVolumeIDs = allocHostVolumeIDs
+	h.previousAllocationID = prevAllocID
 	for _, req := range volumes {
 		if req.Type != structs.VolumeTypeHost {
 			continue // filter CSI volumes
@@ -192,33 +193,33 @@ func (h *HostVolumeChecker) hasVolumes(n *structs.Node) bool {
 		return true
 	}
 
+	proposed, err := h.ctx.ProposedAllocs(n.ID)
+	if err != nil {
+		return false // only hit this on state store invariant failure
+	}
+
 	for _, req := range h.volumeReqs {
 		volCfg, ok := n.HostVolumes[req.Source]
 		if !ok {
 			return false
 		}
 
 		if volCfg.ID != "" { // dynamic host volume
-			vol, err := h.ctx.State().HostVolumeByID(nil, h.namespace, volCfg.ID, false)
+			vol, err := h.ctx.State().HostVolumeByID(nil, h.namespace, volCfg.ID, true)
 			if err != nil || vol == nil {
 				// node fingerprint has a dynamic volume that's no longer in the
 				// state store; this is only possible if the batched fingerprint
 				// update from a delete RPC is written before the delete RPC's
 				// raft entry completes
 				return false
 			}
-			if vol.State != structs.HostVolumeStateReady {
-				return false
-			}
-			var capOk bool
-			for _, cap := range vol.RequestedCapabilities {
-				if req.AccessMode == structs.CSIVolumeAccessMode(cap.AccessMode) &&
-					req.AttachmentMode == structs.CSIVolumeAttachmentMode(cap.AttachmentMode) {
-					capOk = true
-					break
-				}
-			}
-			if !capOk {
+			if !hostVolumeIsAvailable(vol,
+				structs.HostVolumeAccessMode(req.AccessMode),
+				structs.HostVolumeAttachmentMode(req.AttachmentMode),
+				req.ReadOnly,
+				h.previousAllocationID,
+				proposed,
+			) {
 				return false
 			}
 
@@ -242,6 +243,81 @@ func (h *HostVolumeChecker) hasVolumes(n *structs.Node) bool {
 	return true
 }
 
+// hostVolumeIsAvailable determines if a dynamic host volume is available for a request
+func hostVolumeIsAvailable(hv *structs.HostVolume, reqAccess structs.HostVolumeAccessMode, reqAttach structs.HostVolumeAttachmentMode, readOnly bool, previousAllocID string, proposed []*structs.Allocation) bool {
+
+	if hv.State != structs.HostVolumeStateReady {
+		return false
+	}
+
+	// pick a default capability based on the read-only flag. this happens here
+	// in the scheduler rather than job submit because we don't know whether a
+	// host volume is dynamic or not until we try to schedule it (ex. the same
+	// name could be static on one node and dynamic on another)
+	if reqAccess == structs.HostVolumeAccessModeUnknown {
+		if readOnly {
+			reqAccess = structs.HostVolumeAccessModeSingleNodeReader
+		} else {
+			reqAccess = structs.HostVolumeAccessModeSingleNodeWriter
+		}
+	}
+	if reqAttach == structs.HostVolumeAttachmentModeUnknown {
+		reqAttach = structs.HostVolumeAttachmentModeFilesystem
+	}
+
+	// check that the volume has the requested capability at all
+	var capOk bool
+	for _, cap := range hv.RequestedCapabilities {
+		if reqAccess == cap.AccessMode &&
+			reqAttach == cap.AttachmentMode {
+			capOk = true
+			break
+		}
+	}
+	if !capOk {
+		return false
+	}
+
+	// if no other allocations claim the volume, it's first-come-first-serve to
+	// determine how subsequent allocs can claim it
+	if len(hv.Allocations) == 0 && len(proposed) == 0 {
+		return true
+	}
+
+	switch reqAccess {
+	case structs.HostVolumeAccessModeSingleNodeReader:
+		return readOnly
+	case structs.HostVolumeAccessModeSingleNodeWriter:
+		return !readOnly
+	case structs.HostVolumeAccessModeSingleNodeSingleWriter:
+		// make a cheap check that no other persisted alloc has claimed writes
+		// if its not being replaced by this alloc
+		if len(hv.Writers) > 0 && hv.Writers[0] != previousAllocID {
+			return false
+		}
+
+		// examine all allocs on the node, including those that might not have
+		// yet been persisted
+		for _, alloc := range proposed {
+			tg := alloc.Job.LookupTaskGroup(alloc.TaskGroup)
+			for _, req := range tg.Volumes {
+				if req.Type == structs.VolumeTypeHost && req.Source == hv.Name {
+					if !req.ReadOnly {
+						return false
+					}
+				}
+			}
+
+		}
+		// return len(hv.Writers) == 0 ||
+		// 	(len(hv.Writers) == 1 && hv.Writers[0] == previousAllocID)
+	case structs.HostVolumeAccessModeSingleNodeMultiWriter:
+		// no contraint
+	}
+
+	return true
+}
+
 type CSIVolumeChecker struct {
 	ctx       Context
 	namespace string