Add support to configure vmi disk I/O mode options

QEMU supports 3 disk I/O modes:
- threads
- native (kernel async i/o)
- default

(io_uring offers a better performance than traditional async i/o but it
was introduced in a more recent version of QEMU which we don't use yet
for virt-launcher)

This patch allows users to configure the I/O mode that fits most to
their use-case.

Signed-off-by: Daniel Belenky <[email protected]>

api/openapi-spec/swagger.json

@@ -6846,6 +6846,10 @@
       "description": "Attach a volume as a floppy to the vmi.",
       "$ref": "#/definitions/v1.FloppyTarget"
      },
+     "io": {
+      "description": "IO specifies which QEMU disk IO mode should be used. Supported values are: native, default, threads.",
+      "type": "string"
+     },
      "lun": {
       "description": "Attach a volume as a LUN to the vmi.",
       "$ref": "#/definitions/v1.LunTarget"

pkg/virt-api/webhooks/validating-webhook/admitters/vmi-create-admitter.go

@@ -1629,6 +1629,15 @@ func validateDisks(field *k8sfield.Path, disks []v1.Disk) []metav1.StatusCause {
 			})
 		}
 
+		if disk.IO != "" && disk.IO != v1.IODefault && disk.IO != v1.IONative && disk.IO != v1.IOThreads {
+			field := field.Child("domain", "devices", "disks").Index(idx).Child("io").String()
+			causes = append(causes, metav1.StatusCause{
+				Type:    metav1.CauseTypeFieldValueNotSupported,
+				Message: fmt.Sprintf("Disk IO mode for %s is not supported. Supported modes are: native, threads, default.", field),
+				Field:   field,
+			})
+		}
+
 		// Verify disk and volume name can be a valid container name since disk
 		// name can become a container name which will fail to schedule if invalid
 		errs := validation.IsDNS1123Label(disk.Name)

pkg/virt-api/webhooks/validating-webhook/admitters/vmi-create-admitter_test.go

@@ -2226,6 +2226,23 @@ var _ = Describe("Validating VMICreate Admitter", func() {
 			Expect(causes[1].Field).To(Equal("fake[1].lun.bus"))
 		})
 
+		It("should reject disks with unsupported I/O modes", func() {
+			vmi := v1.NewMinimalVMI("testvmi")
+
+			vmi.Spec.Domain.Devices.Disks = append(vmi.Spec.Domain.Devices.Disks, v1.Disk{
+				Name: "testdisk1",
+				IO:   "native",
+			})
+			vmi.Spec.Domain.Devices.Disks = append(vmi.Spec.Domain.Devices.Disks, v1.Disk{
+				Name: "testdisk2",
+				IO:   "unsupported",
+			})
+
+			causes := validateDisks(k8sfield.NewPath("fake"), vmi.Spec.Domain.Devices.Disks)
+			Expect(len(causes)).To(Equal(1))
+			Expect(causes[0].Field).To(Equal("fake.domain.devices.disks[1].io"))
+		})
+
 		It("should reject disk with invalid cache mode", func() {
 			vmi := v1.NewMinimalVMI("testvmi")
 			vmi.Spec.Domain.Devices.Disks = append(vmi.Spec.Domain.Devices.Disks, v1.Disk{

pkg/virt-launcher/virtwrap/api/converter.go

@@ -120,6 +120,7 @@ func Convert_v1_Disk_To_api_Disk(diskDevice *v1.Disk, disk *Disk, devicePerBus m
 	disk.Driver = &DiskDriver{
 		Name:  "qemu",
 		Cache: string(diskDevice.Cache),
+		IO:    string(diskDevice.IO),
 	}
 	if numQueues != nil && disk.Target.Bus == "virtio" {
 		disk.Driver.Queues = numQueues

pkg/virt-launcher/virtwrap/api/converter_test.go

@@ -64,6 +64,32 @@ var _ = Describe("Converter", func() {
 			Expect(xml).To(Equal(convertedDisk))
 		})
 
+		It("should set disk I/O mode if requested", func() {
+			v1Disk := &v1.Disk{
+				IO: "native",
+			}
+			xml := diskToDiskXML(v1Disk)
+			expectedXML := `<Disk device="" type="">
+  <source></source>
+  <target></target>
+  <driver io="native" name="qemu" type=""></driver>
+  <alias name="ua-"></alias>
+</Disk>`
+			Expect(xml).To(Equal(expectedXML))
+		})
+
+		It("should not set disk I/O mode if not requested", func() {
+			v1Disk := &v1.Disk{}
+			xml := diskToDiskXML(v1Disk)
+			expectedXML := `<Disk device="" type="">
+  <source></source>
+  <target></target>
+  <driver name="qemu" type=""></driver>
+  <alias name="ua-"></alias>
+</Disk>`
+			Expect(xml).To(Equal(expectedXML))
+		})
+
 		It("Should omit boot order when not provided", func() {
 			kubevirtDisk := &v1.Disk{
 				Name: "mydisk",

staging/src/kubevirt.io/client-go/api/v1/schema.go

@@ -420,6 +420,10 @@ type Disk struct {
 	// Cache specifies which kvm disk cache mode should be used.
 	// +optional
 	Cache DriverCache `json:"cache,omitempty"`
+	// IO specifies which QEMU disk IO mode should be used.
+	// Supported values are: native, default, threads.
+	// +optional
+	IO DriverIO `json:"io,omitempty"`
 	// If specified, disk address and its tag will be provided to the guest via config drive metadata
 	// +optional
 	Tag string `json:"tag,omitempty"`

staging/src/kubevirt.io/client-go/api/v1/types.go

@@ -1006,11 +1006,25 @@ const (
 // +k8s:openapi-gen=true
 type DriverCache string
 
+//
+// +k8s:openapi-gen=true
+type DriverIO string
+
 const (
 	// CacheNone - I/O from the guest is not cached on the host, but may be kept in a writeback disk cache.
 	CacheNone DriverCache = "none"
 	// CacheWriteThrough - I/O from the guest is cached on the host but written through to the physical medium.
 	CacheWriteThrough DriverCache = "writethrough"
+
+	// IOThreads - User mode based threads with a shared lock that perform I/O tasks. Can impact performance but offers
+	// more predictable behaviour. This method is also takes fewer CPU cycles to submit I/O requests.
+	IOThreads DriverIO = "threads"
+	// IONative - Kernel native I/O tasks (AIO) offer a better performance but can block the VM if the file is not fully
+	// allocated so this method recommended only when the backing file/disk/etc is fully preallocated.
+	IONative DriverIO = "native"
+	// IODefault - Fallback to the default value from the kernel. With recent Kernel versions (for example RHEL-7) the
+	// default is AIO.
+	IODefault DriverIO = "default"
 )
 
 // Handler defines a specific action that should be taken