CP-310822: upstream patchqueue entries for XS9

ocaml/libs/xenctrl-ext/xenctrlext.ml

@@ -124,7 +124,6 @@ end
 exception Not_available
 
 let domain_claim_pages handle domid ?(numa_node = NumaNode.none) nr_pages =
-  if numa_node <> NumaNode.none then raise Not_available ;
   stub_domain_claim_pages handle domid numa_node nr_pages
 
 let get_nr_nodes handle =

ocaml/libs/xenctrl-ext/xenctrlext.mli

@@ -109,6 +109,14 @@ val get_nr_nodes : handle -> int
 (** Returns the count of NUMA nodes available in the system. *)
 
 module DomainNuma : sig
+  type domain_numainfo_node_pages = {
+      tot_pages_per_node: int64 array (* page=4k bytes *)
+  }
+
+  external domain_get_numa_info_node_pages :
+    handle -> int -> domain_numainfo_node_pages
+    = "stub_xc_domain_numa_get_node_pages_wrapper"
+
   type t = {optimised: bool; nodes: int; memory: int64 array (* bytes *)}
 
   val state : handle -> domid:int -> t

ocaml/libs/xenctrl-ext/xenctrlext_stubs.c

@@ -678,14 +678,15 @@ CAMLprim value stub_xenctrlext_domain_claim_pages(value xch_val, value domid_val
         value numa_node_val, value nr_pages_val)
 {
         CAMLparam4(xch_val, domid_val, numa_node_val, nr_pages_val);
+#ifdef XEN_DOMCTL_NUMA_OP_GET_NODE_PAGES
         int retval, the_errno;
         xc_interface* xch = xch_of_val(xch_val);
         uint32_t domid = Int_val(domid_val);
-        // unsigned int numa_node = Int_val(numa_node_val);
+        unsigned int numa_node = Int_val(numa_node_val);
         unsigned long nr_pages = Long_val(nr_pages_val);
 
         caml_release_runtime_system();
-        retval = xc_domain_claim_pages(xch, domid, /*numa_node,*/ nr_pages);
+        retval = xc_domain_claim_pages_node(xch, domid, numa_node, nr_pages);
         the_errno = errno;
         caml_acquire_runtime_system();
 
@@ -694,6 +695,9 @@ CAMLprim value stub_xenctrlext_domain_claim_pages(value xch_val, value domid_val
                         "Error when trying to claim memory pages");
         }
         CAMLreturn(Val_unit);
+#else
+        raise_unix_errno_msg(ENOSYS, "xc_domain_claim_pages_node");
+#endif
 }
 
 #ifdef XEN_DOMCTL_NUMA_OP_GET_NODE_PAGES

ocaml/xcp-rrdd/bin/rrdp-cpu/rrdp_cpu.ml

@@ -80,6 +80,28 @@ let dss_vcpus xc doms =
                   )
                 ]
           in
+          let nonaffine_vcpus_ds =
+            match ri.Xenctrl.Runstateinfo.V2.running with
+            | 0L ->
+                []
+            | _ ->
+                [
+                  ( Rrd.VM uuid
+                  , Ds.ds_make ~name:"numa_node_nonaffine_vcpus"
+                      ~units:"(fraction)"
+                      ~value:
+                        (Rrd.VT_Float
+                           (Int64.to_float ri.Xenctrl.Runstateinfo.V2.nonaffine
+                           /. 1.0e9
+                           )
+                        )
+                      ~description:
+                        "Fraction of vCPU time running outside of vCPU \
+                         soft-affinity"
+                      ~ty:Rrd.Derive ~default:false ~min:0.0 ~max:1.0 ()
+                  )
+                ]
+          in
           ( Rrd.VM uuid
           , Ds.ds_make ~name:"runstate_fullrun" ~units:"(fraction)"
               ~value:
@@ -172,6 +194,7 @@ let dss_vcpus xc doms =
              )
           :: dss
           @ runnable_vcpus_ds
+          @ nonaffine_vcpus_ds
         with _ -> dss
       in
       try cpus 0 dss with _ -> dss

ocaml/xcp-rrdd/bin/rrdp-squeezed/dune

@@ -13,6 +13,7 @@
     xapi-log
     xapi-rrd
     xenctrl
+    xenctrl_ext
     xenstore
     xenstore.unix
     xenstore_transport

ocaml/xcp-rrdd/bin/rrdp-squeezed/rrdp_squeezed.ml

@@ -252,6 +252,62 @@ let free_other uuid free =
         ~value:(Rrd.VT_Int64 free) ~ty:Rrd.Gauge ~min:0.0 ~default:true ()
     )
 
+let dss_numa_info uuid domid =
+  try
+    let handle = Xenctrlext.get_handle () in
+    let host_nr_nodes = Xenctrlext.get_nr_nodes handle in
+    let vm_nodes =
+      Xenctrlext.DomainNuma.domain_get_numa_info_node_pages handle domid
+    in
+    let dss_memory_numa_nodes_of_vm (dss, nr_nodes_used_by_vm)
+        (node_id, tot_pages_per_node) =
+      (*
+        for each numa node used by the host, show the
+        corresponding amount of memory used by the vm
+      *)
+      let is_node_used_by_vm = tot_pages_per_node > 4096L in
+      let is_node_used_by_host = node_id < host_nr_nodes in
+      if is_node_used_by_host then
+        ( ( Rrd.VM uuid
+          , Ds.ds_make
+              ~name:(Printf.sprintf "memory_numa_node_%d" node_id)
+              ~units:"B"
+              ~description:
+                (Printf.sprintf "Memory from NUMA node %d used by VM" node_id)
+              ~value:(Rrd.VT_Int64 (Int64.mul tot_pages_per_node 4096L))
+              ~min:0.0 ~ty:Rrd.Gauge ~default:false ()
+          )
+          :: dss
+          (* remember the number of nodes used by vm *)
+        , nr_nodes_used_by_vm
+          +
+          if is_node_used_by_vm then
+            1
+          else
+            0
+        )
+      else
+        (dss, nr_nodes_used_by_vm)
+    in
+    let dss_numa_nodes_of_vm (dss, nr_nodes_used_by_vm) =
+      ( Rrd.VM uuid
+      , Ds.ds_make
+          ~name:(Printf.sprintf "numa_nodes")
+          ~units:"count"
+          ~description:(Printf.sprintf "Number of NUMA nodes used by VM")
+          ~value:(Rrd.VT_Int64 (Int64.of_int nr_nodes_used_by_vm))
+          ~min:0.0 ~ty:Rrd.Gauge ~default:false ()
+      )
+      :: List.rev dss
+    in
+    vm_nodes.Xenctrlext.DomainNuma.tot_pages_per_node
+    |> Array.mapi (fun i x -> (i, x))
+    |> Array.fold_left dss_memory_numa_nodes_of_vm ([], 0)
+    |> dss_numa_nodes_of_vm
+  with e ->
+    D.debug "dss_numa_info: %s" (Printexc.to_string e) ;
+    []
+
 let get_list f = Option.to_list (f ())
 
 let generate_vm_sources domains =
@@ -288,14 +344,15 @@ let generate_vm_sources domains =
             ~value:(Rrd.VT_Int64 memory) ~ty:Rrd.Gauge ~min:0.0 ~default:true ()
         )
     in
+    let get_list_numa_info = dss_numa_info uuid domid in
     (* CA-34383: Memory updates from paused domains serve no useful purpose.
        During a migrate such updates can also cause undesirable
        discontinuities in the observed value of memory_actual. Hence, we
        ignore changes from paused domains: *)
     if dom.Xenctrl.paused then
       []
     else
-      get_list target @ get_list free @ get_list total
+      get_list target @ get_list free @ get_list total @ get_list_numa_info
   in
 
   List.concat_map metrics_of domains

ocaml/xenopsd/xc/domain.ml

@@ -987,8 +987,6 @@ let numa_hierarchy =
 
 let numa_mutex = Mutex.create ()
 
-let numa_resources = ref None
-
 let numa_init () =
   let xcext = Xenctrlext.get_handle () in
   let host = Lazy.force numa_hierarchy in
@@ -1022,14 +1020,7 @@ let numa_placement domid ~vcpus ~cores ~memory affinity =
           (NUMA.nodes host) numa_meminfo
       in
       let vm = NUMARequest.make ~memory ~vcpus ~cores in
-      let nodea =
-        match !numa_resources with
-        | None ->
-            Array.of_seq nodes
-        | Some a ->
-            Array.map2 NUMAResource.min_memory (Array.of_seq nodes) a
-      in
-      numa_resources := Some nodea ;
+      let nodea = Array.of_seq nodes in
       let cpu_affinity, memory_plan =
         match Softaffinity.plan ~vm host nodea with
         | None ->