fix: capacity filter correctly accounts for multi-VM CR reservation slots (#784)

Fixes capacity blocking for CommittedResource reservation slots that
contain multiple VMs at different confirmation stages.

Author: mblos

internal/scheduling/nova/plugins/filters/filter_has_enough_capacity.go

@@ -41,6 +41,16 @@ type FilterHasEnoughCapacity struct {
 //
 // In case the project and flavor match, space reserved is unlocked (slotting).
 //
+// Capacity accounting uses two sources: hv.Status.Allocation (aggregate real-time usage of
+// all running VMs) and Reservation.Status.Allocations (which VMs are confirmed on a slot,
+// maintained by the reservation controller with a one-reconcile-cycle lag). During the window
+// between a VM starting and the reservation controller reconciling, a VM appears in both
+// sources — a conservative transient over-count that self-corrects on the next reconcile.
+//
+// During a CR reservation migration (TargetHost != Status.Host), both the source and target
+// host are blocked with the full slot. The source block is intentionally conservative to
+// preserve rollback capacity if the migration fails.
+//
 // Please note that, if num_instances is larger than 1, there needs to be enough
 // capacity to place all instances on the same host. This limitation is necessary
 // because we can't spread out instances, as the final set of valid hosts is not
@@ -170,41 +180,61 @@ func (s *FilterHasEnoughCapacity) Run(traceLog *slog.Logger, request api.Externa
 			continue
 		}
 
-		// For CR reservations with allocations, calculate remaining (unallocated) resources to block.
-		// This prevents double-blocking of resources already consumed by running instances.
+		// For CR reservations with allocations, compute the effective block:
+		//   confirmed = sum of resources for VMs present in both Spec and Status allocations
+		//   specOnly  = sum of resources for VMs present in Spec but not yet in Status
+		//   remaining = max(0, Spec.Resources - confirmed)  [clamped: never negative]
+		//   block     = max(remaining, specOnly)            [spec-only VM must be fully covered]
+		//
+		// Clamping: if confirmed VMs exceed slot size (e.g. after resize), block = 0.
+		// Oversize spec-only: if a pending VM is larger than the remaining slot, block its full size.
 		var resourcesToBlock map[hv1.ResourceName]resource.Quantity
 		if reservation.Spec.Type == v1alpha1.ReservationTypeCommittedResource &&
 			// if the reservation is not being migrated, block only unused resources
 			reservation.Spec.TargetHost == reservation.Status.Host &&
 			reservation.Spec.CommittedResourceReservation != nil &&
-			reservation.Status.CommittedResourceReservation != nil &&
-			len(reservation.Spec.CommittedResourceReservation.Allocations) > 0 &&
-			len(reservation.Status.CommittedResourceReservation.Allocations) > 0 {
-			// Start with full reservation resources
-			resourcesToBlock = make(map[hv1.ResourceName]resource.Quantity)
-			for k, v := range reservation.Spec.Resources {
-				resourcesToBlock[k] = v.DeepCopy()
+			len(reservation.Spec.CommittedResourceReservation.Allocations) > 0 {
+			confirmedResources := make(map[hv1.ResourceName]resource.Quantity)
+			specOnlyResources := make(map[hv1.ResourceName]resource.Quantity)
+
+			statusAllocs := map[string]string{}
+			if reservation.Status.CommittedResourceReservation != nil {
+				statusAllocs = reservation.Status.CommittedResourceReservation.Allocations
 			}
 
-			// Subtract already-allocated resources because those consume already resources on the host
 			for instanceUUID, allocation := range reservation.Spec.CommittedResourceReservation.Allocations {
-				// Only subtract if allocation is already present in status (VM is actually running)
-				if _, isRunning := reservation.Status.CommittedResourceReservation.Allocations[instanceUUID]; !isRunning {
-					continue
-				}
-
+				_, isConfirmed := statusAllocs[instanceUUID]
 				for resourceName, quantity := range allocation.Resources {
-					if current, ok := resourcesToBlock[resourceName]; ok {
-						current.Sub(quantity)
-						resourcesToBlock[resourceName] = current
-						traceLog.Debug("subtracting allocated resources from reservation",
-							"reservation", reservation.Name,
-							"instanceUUID", instanceUUID,
-							"resource", resourceName,
-							"quantity", quantity.String())
+					if isConfirmed {
+						existing := confirmedResources[resourceName]
+						existing.Add(quantity)
+						confirmedResources[resourceName] = existing
+					} else {
+						existing := specOnlyResources[resourceName]
+						existing.Add(quantity)
+						specOnlyResources[resourceName] = existing
 					}
 				}
 			}
+
+			resourcesToBlock = make(map[hv1.ResourceName]resource.Quantity)
+			zero := resource.Quantity{}
+			for resourceName, slotSize := range reservation.Spec.Resources {
+				confirmed := confirmedResources[resourceName]
+				specOnly := specOnlyResources[resourceName]
+
+				remaining := slotSize.DeepCopy()
+				remaining.Sub(confirmed)
+				if remaining.Cmp(zero) < 0 {
+					remaining = zero.DeepCopy()
+				}
+
+				if specOnly.Cmp(remaining) > 0 {
+					resourcesToBlock[resourceName] = specOnly.DeepCopy()
+				} else {
+					resourcesToBlock[resourceName] = remaining
+				}
+			}
 		} else {
 			// For other reservation types or CR without allocations, block full resources
 			resourcesToBlock = reservation.Spec.Resources
@@ -229,7 +259,7 @@ func (s *FilterHasEnoughCapacity) Run(traceLog *slog.Logger, request api.Externa
 							"reservationType", reservation.Spec.Type,
 							"freeCPU", freeCPU.String(),
 							"blocked", cpu.String())
-						freeCPU = resource.MustParse("0")
+						freeCPU = resource.Quantity{}
 					}
 					freeResourcesByHost[host]["cpu"] = freeCPU
 				}
@@ -244,7 +274,7 @@ func (s *FilterHasEnoughCapacity) Run(traceLog *slog.Logger, request api.Externa
 							"reservationType", reservation.Spec.Type,
 							"freeMemory", freeMemory.String(),
 							"blocked", memory.String())
-						freeMemory = resource.MustParse("0")
+						freeMemory = resource.Quantity{}
 					}
 					freeResourcesByHost[host]["memory"] = freeMemory
 				}