Fix remaining calculation

Author: sjmiller609

integration/vgpu_test.go

@@ -128,9 +128,23 @@ func TestVGPU(t *testing.T) {
 	require.Equal(t, images.StatusReady, img.Status, "Image should be ready")
 	t.Log("Image ready")
 
-	// Step 3: Create instance with vGPU using QEMU hypervisor
+	// Step 3: Check GPU resources BEFORE creating instance
+	t.Log("Step 3: Recording GPU resources before instance creation...")
+	profilesBefore, err := devices.ListGPUProfiles()
+	require.NoError(t, err, "should list GPU profiles")
+	var availableBefore int
+	for _, p := range profilesBefore {
+		if p.Name == profile {
+			availableBefore = p.Available
+			break
+		}
+	}
+	require.Greater(t, availableBefore, 0, "profile should have availability before creation")
+	t.Logf("Profile %q available instances before: %d", profile, availableBefore)
+
+	// Step 4: Create instance with vGPU using QEMU hypervisor
 	// QEMU is required for vGPU/mdev passthrough with NVIDIA's vGPU manager
-	t.Log("Step 3: Creating instance with vGPU profile:", profile)
+	t.Log("Step 4: Creating instance with vGPU profile:", profile)
 	inst, err := instanceManager.CreateInstance(ctx, instances.CreateInstanceRequest{
 		Name:           "vgpu-test",
 		Image:          imageName,
@@ -152,20 +166,37 @@ func TestVGPU(t *testing.T) {
 	require.NotEmpty(t, inst.GPUMdevUUID, "Instance should have mdev UUID assigned")
 	t.Logf("mdev UUID: %s", inst.GPUMdevUUID)
 
-	// Step 4: Verify mdev was created in sysfs
+	// Step 5: Check GPU resources AFTER creating instance
+	t.Run("ResourcesDecrementedAfterCreation", func(t *testing.T) {
+		profilesAfter, err := devices.ListGPUProfiles()
+		require.NoError(t, err, "should list GPU profiles after creation")
+
+		var availableAfter int
+		for _, p := range profilesAfter {
+			if p.Name == profile {
+				availableAfter = p.Available
+				break
+			}
+		}
+
+		t.Logf("Profile %q available instances after: %d (was %d)", profile, availableAfter, availableBefore)
+		assert.Less(t, availableAfter, availableBefore, "available instances should decrease after creating VM")
+	})
+
+	// Step 6: Verify mdev was created in sysfs
 	t.Run("MdevCreated", func(t *testing.T) {
 		mdevPath := "/sys/bus/mdev/devices/" + inst.GPUMdevUUID
 		_, err := os.Stat(mdevPath)
 		assert.NoError(t, err, "mdev device should exist at %s", mdevPath)
 		t.Logf("mdev exists at: %s", mdevPath)
 	})
 
-	// Step 5: Wait for guest agent to be ready
-	t.Log("Step 5: Waiting for guest agent...")
+	// Step 7: Wait for guest agent to be ready
+	t.Log("Step 7: Waiting for guest agent...")
 	err = waitForGuestAgent(ctx, instanceManager, inst.Id, 60*time.Second)
 	require.NoError(t, err, "guest agent should be ready")
 
-	// Step 6: Verify GPU is visible inside VM via PCI
+	// Step 8: Verify GPU is visible inside VM via PCI
 	t.Run("GPUVisibleInVM", func(t *testing.T) {
 		actualInst, err := instanceManager.GetInstance(ctx, inst.Id)
 		require.NoError(t, err)
@@ -191,7 +222,7 @@ func TestVGPU(t *testing.T) {
 		assert.Contains(t, output, "NVIDIA_FOUND", "NVIDIA GPU (vendor 0x10de) should be visible in guest")
 	})
 
-	// Step 7: Check instance GPU info is correct
+	// Step 9: Check instance GPU info is correct
 	t.Run("InstanceGPUInfo", func(t *testing.T) {
 		actualInst, err := instanceManager.GetInstance(ctx, inst.Id)
 		require.NoError(t, err)

lib/devices/mdev.go

@@ -237,7 +237,7 @@ func countAvailableForSingleProfile(freeVFsByParent map[string][]VirtualFunction
 		if len(parentVFs) == 0 {
 			continue
 		}
-		// Sample just ONE VF per parent - all VFs on same parent have same profiles
+		// Sample just ONE VF per parent - available_instances reflects GPU-wide capacity
 		sampleVF := parentVFs[0]
 		availPath := filepath.Join(mdevBusPath, sampleVF.PCIAddress, "mdev_supported_types", profileType, "available_instances")
 		data, err := os.ReadFile(availPath)
@@ -248,8 +248,8 @@ func countAvailableForSingleProfile(freeVFsByParent map[string][]VirtualFunction
 		if err != nil || instances < 1 {
 			continue
 		}
-		// Profile is available - count all free VFs on this parent
-		count += len(parentVFs)
+		// available_instances reflects GPU-wide remaining capacity, not per-VF
+		count += instances
 	}
 	return count
 }