CP-44320 scaffolding for NVidia Virtual Compute Service (VCS)

NVidia is supporting vGPUs that can run as compute vGPUs. These are
marked in their vgpuConfig.xml with a new value "Compute" for the
existing "class" attribute:

<vgpuType id="307" name="GRID P100C-12C" class="Compute">

We don't parse and use this attribute so far but need to in order to
support VCS. This commit extends the parser to recognise the class
attribute and to pass it on under the internal name "vclass" to xenopsd.
Xenopsd's job is to start device emulators accordingly. Selecting the
correct emulators is future work. So far, xenopsd logs the vclass it
receives.

Unit tests have been updated.

This commit should have no impact on current functionality, be safe to
merge, and lay groundwork for extending xenopsd to support VCS.

Signed-off-by: Christian Lindig <[email protected]>

ocaml/tests/common/test_vgpu_common.ml

@@ -33,6 +33,7 @@ let k100 =
           ; vdev_id= 0x0fe7
           ; vsubdev_id= 0x101e
           ; sriov= false
+          ; vclass= "NVS"
           }
       )
   ; experimental= false
@@ -59,6 +60,7 @@ let k140q =
           ; vdev_id= 0x0ff7
           ; vsubdev_id= 0x1037
           ; sriov= false
+          ; vclass= "NVS"
           }
       )
   ; experimental= false
@@ -85,6 +87,7 @@ let k200 =
           ; vdev_id= 0x118d
           ; vsubdev_id= 0x101d
           ; sriov= false
+          ; vclass= "NVS"
           }
       )
   ; experimental= false
@@ -111,6 +114,7 @@ let k240q =
           ; vdev_id= 0x11b0
           ; vsubdev_id= 0x101a
           ; sriov= false
+          ; vclass= "NVS"
           }
       )
   ; experimental= false
@@ -137,6 +141,7 @@ let k260q =
           ; vdev_id= 0x11b0
           ; vsubdev_id= 0x101b
           ; sriov= false
+          ; vclass= "NVS"
           }
       )
   ; experimental= false

ocaml/tests/test_vgpu_type.ml

@@ -67,6 +67,7 @@ module NvidiaTest = struct
                       ; vdev_id= 0x1111
                       ; vsubdev_id= 0x2222
                       ; sriov= false
+                      ; vclass= "NVS"
                       }
                 ; framebufferlength= 0x10000000L
                 ; num_heads= 2L
@@ -92,6 +93,7 @@ module NvidiaTest = struct
                       ; vdev_id= 0x1111
                       ; vsubdev_id= 0x2222
                       ; sriov= false
+                      ; vclass= "NVS"
                       }
                 ; framebufferlength= 0x10000000L
                 ; num_heads= 2L
@@ -117,6 +119,7 @@ module NvidiaTest = struct
                       ; vdev_id= 0x1112
                       ; vsubdev_id= 0x2223
                       ; sriov= false
+                      ; vclass= "NVS"
                       }
                 ; framebufferlength= 0x20000000L
                 ; num_heads= 4L
@@ -138,6 +141,7 @@ module NvidiaTest = struct
                       ; vdev_id= 0x1111
                       ; vsubdev_id= 0x2222
                       ; sriov= false
+                      ; vclass= "NVS"
                       }
                 ; framebufferlength= 0x10000000L
                 ; num_heads= 2L

ocaml/tests/test_xenopsd_metadata.ml

@@ -280,6 +280,7 @@ module GenerateVGPUMetadata = Generic.MakeStateful (struct
                 ; type_id= Some "type_id_1"
                 ; uuid= Some (uuid_with_index 0)
                 ; extra_args= ""
+                ; vclass= None
                 }
             )
           ]
@@ -364,6 +365,7 @@ module GenerateMultiVGPUMetadata = Generic.MakeStateful (struct
                 ; type_id= Some "type_id_1"
                 ; uuid= Some (uuid_with_index 0)
                 ; extra_args= ""
+                ; vclass= None
                 }
             )
           ; Xenops_interface.Vgpu.(
@@ -375,6 +377,7 @@ module GenerateMultiVGPUMetadata = Generic.MakeStateful (struct
                 ; type_id= Some "type_id_1"
                 ; uuid= Some (uuid_with_index 1)
                 ; extra_args= ""
+                ; vclass= None
                 }
             )
           ]
@@ -398,6 +401,7 @@ module GenerateMultiVGPUMetadata = Generic.MakeStateful (struct
                 ; type_id= Some "type_id_1"
                 ; uuid= Some (uuid_with_index 0)
                 ; extra_args= ""
+                ; vclass= None
                 }
             )
           ; Xenops_interface.Vgpu.(
@@ -409,6 +413,7 @@ module GenerateMultiVGPUMetadata = Generic.MakeStateful (struct
                 ; type_id= Some "type_id_1"
                 ; uuid= Some (uuid_with_index 1)
                 ; extra_args= ""
+                ; vclass= None
                 }
             )
           ; Xenops_interface.Vgpu.(
@@ -420,6 +425,7 @@ module GenerateMultiVGPUMetadata = Generic.MakeStateful (struct
                 ; type_id= Some "type_id_1"
                 ; uuid= Some (uuid_with_index 2)
                 ; extra_args= ""
+                ; vclass= None
                 }
             )
           ; Xenops_interface.Vgpu.(
@@ -431,6 +437,7 @@ module GenerateMultiVGPUMetadata = Generic.MakeStateful (struct
                 ; type_id= Some "type_id_1"
                 ; uuid= Some (uuid_with_index 3)
                 ; extra_args= ""
+                ; vclass= None
                 }
             )
           ]

ocaml/xapi-idl/xen/xenops_types.ml

@@ -32,6 +32,7 @@ module Vgpu = struct
           [@default {domain= 0000; bus= 0; dev= 11; fn= 0}]
     ; type_id: string option
     ; uuid: string option
+    ; vclass: string option  (** from vgpu: Compute, NVS, Quadro *)
     ; extra_args: string [@default ""]
           (** string is passed on as is and no structure is assumed *)
   }

ocaml/xapi/xapi_globs.ml

@@ -536,6 +536,9 @@ let pass_through_pif_carrier = ref false
 
 let vgpu_type_id = "type_id"
 
+(** key for NVidia vgpu "class" attribute in vgpuConfig.xml *)
+let vgpu_type_vclass = "vclass"
+
 let igd_passthru_key = "igd_passthrough"
 
 let vgt_low_gm_sz = "vgt_low_gm_sz"

ocaml/xapi/xapi_vgpu_type.ml

@@ -39,6 +39,7 @@ module Identifier = struct
     ; vdev_id: int  (** vgpuTYpe/deviceId in XML *)
     ; vsubdev_id: int  (** vgpuTYpe/subsystemId in XML *)
     ; sriov: bool  (** true if SRIOV mode to be used *)
+    ; vclass: string  (** vgpuType/class in XML: Quadro, NVS, Compute *)
   }
 
   type gvt_g_id = {
@@ -399,6 +400,7 @@ let read_whitelist_line_by_line ~whitelist ~device_id ~parse_line
     []
 
 module Vendor_nvidia = struct
+  (* represents a vgpuType declaration found in vgpuConfig.xml *)
   type vgpu_conf = {
       identifier: Identifier.nvidia_id
     ; framebufferlength: int64
@@ -652,6 +654,7 @@ module Vendor_nvidia = struct
             )
           in
           let devid = find_one_by_name "devId" vgpu_type in
+          let vclass = get_attr "class" vgpu_type in
           let identifier =
             Identifier.
               {
@@ -669,6 +672,7 @@ module Vendor_nvidia = struct
                       false
                   )
                   (* don't use SRIOV *)
+              ; vclass
               }
           in
 
@@ -760,7 +764,11 @@ module Vendor_nvidia = struct
       ; max_resolution_x= conf.max_x
       ; max_resolution_y= conf.max_y
       ; size= Int64.div Constants.pgpu_default_size conf.max_instance
-      ; internal_config= [(Xapi_globs.vgpu_type_id, conf.type_id)]
+      ; internal_config=
+          [
+            (Xapi_globs.vgpu_type_id, conf.type_id)
+          ; (Xapi_globs.vgpu_type_vclass, conf.identifier.vclass)
+          ]
       ; identifier= Nvidia conf.identifier
       ; experimental= false
       ; compatible_model_names_in_vm= conf.compatible_model_names_in_vm
@@ -999,7 +1007,15 @@ module Nvidia_compat = struct
       Scanf.sscanf (List.assoc "plugin0.vdev_id" args) {|"0x%x:0x%x"|}
         (fun vdev_id vsubdev_id ->
           Identifier.(
-            Nvidia {pdev_id; psubdev_id; vdev_id; vsubdev_id; sriov= false}
+            Nvidia
+              {
+                pdev_id
+              ; psubdev_id
+              ; vdev_id
+              ; vsubdev_id
+              ; sriov= false
+              ; vclass= ""
+              }
           )
       )
     with e -> raise (Parse_error e)

ocaml/xapi/xapi_xenops.ml

@@ -946,10 +946,11 @@ module MD = struct
     let physical_pci_address = get_target_pci_address ~__context vgpu in
     let virtual_pci_address = get_virtual_pci_address ~__context vgpu in
     let vgpu_type = vgpu.Db_actions.vGPU_type in
-    let type_id, config_file =
-      Db.VGPU_type.get_internal_config ~__context ~self:vgpu_type |> fun x ->
-      ( List.assoc_opt Xapi_globs.vgpu_type_id x
-      , List.assoc_opt Xapi_globs.nvidia_compat_config_file_key x
+    let type_id, config_file, vclass =
+      Db.VGPU_type.get_internal_config ~__context ~self:vgpu_type |> fun kv ->
+      ( List.assoc_opt Xapi_globs.vgpu_type_id kv
+      , List.assoc_opt Xapi_globs.nvidia_compat_config_file_key kv
+      , List.assoc_opt Xapi_globs.vgpu_type_vclass kv
       )
     in
     let uuid = vgpu.Db_actions.vGPU_uuid in
@@ -964,6 +965,7 @@ module MD = struct
         ; type_id
         ; uuid= Some uuid
         ; extra_args
+        ; vclass (* from vgpuType class attribute in vgpuConfig.xml *)
         }
     in
     {

ocaml/xenopsd/xc/device.ml

@@ -3635,7 +3635,9 @@ module Dm = struct
   let start_vgpu ~xc:_ ~xs task ?(restore = false) domid vgpus vcpus profile =
     let open Xenops_interface.Vgpu in
     match vgpus with
-    | {implementation= Nvidia _; _} :: _ ->
+    | {implementation= Nvidia {vclass; _}; _} :: _ ->
+        let vclass = Option.value ~default:"unknown" vclass in
+        info "NVidia vgpu vclass=%s" vclass ;
         (* Start DEMU and wait until it has reached the desired state *)
         if not (Service.Vgpu.is_running ~xs domid) then (
           let pcis = List.map (fun x -> x.physical_pci_address) vgpus in