CP-53642: change default NUMA placement policy to best-effort

We've seen that using the policy can be up to 10% faster than using any is some
workflows, while not observing workflows that were negatively affected. The
policy per VM can always be change if need be.

Note that currently sometime the best-effort falls back to the same behaviour,
especially when restarting on starting more than one VM at a time. This needs
xen patches to be fixed:
https://lore.kernel.org/xen-devel/[email protected]/T/#ma1246e352ea3cce71c7ddc26d1329a368548b3b2

Also fix the legacy numa-placement configuration option for xenopsd. It was
always deciding the setting, even when not used, now it only takes effect when
it's present, otherwise it leaves the default option untouched.

Signed-off-by: Pau Ruiz Safont <[email protected]>

Author: psafont

ocaml/xenopsd/lib/xenops_server.ml

@@ -3472,9 +3472,9 @@ module VIF = struct
       ()
 end
 
-let default_numa_affinity_policy = ref Xenops_interface.Host.Any
+let default_numa_affinity_policy = ref Xenops_interface.Host.Best_effort
 
-let numa_placement = ref Xenops_interface.Host.Any
+let numa_placement = ref !default_numa_affinity_policy
 
 let string_of_numa_affinity_policy =
   Xenops_interface.Host.(function Any -> "any" | Best_effort -> "best-effort")

ocaml/xenopsd/lib/xenopsd.ml

@@ -59,7 +59,17 @@ let feature_flags_path = ref "/etc/xenserver/features.d"
 
 let pvinpvh_xen_cmdline = ref "pv-shim console=xen"
 
-let numa_placement_compat = ref false
+(* This is a deprecated argument. If it's not set, use the default numa
+   placement policy, otherwise true means to use Best effort, and false to use
+   Any *)
+let numa_placement_compat = ref None
+
+let set_default_numa_placement () =
+  let change_default numa_enabled =
+    Xenops_server.default_numa_affinity_policy :=
+      if numa_enabled then Best_effort else Any
+  in
+  Option.iter change_default !numa_placement_compat
 
 (* O(N^2) operations, until we get a xenstore cache, so use a small number here *)
 let vm_guest_agent_xenstore_quota = ref 128
@@ -242,8 +252,10 @@ let options =
     , "Command line for the inner-xen for PV-in-PVH guests"
     )
   ; ( "numa-placement"
-    , Arg.Bool (fun x -> numa_placement_compat := x)
-    , (fun () -> string_of_bool !numa_placement_compat)
+    , Arg.Bool (fun x -> numa_placement_compat := Some x)
+    , (fun () ->
+        Option.fold ~none:"true" ~some:string_of_bool !numa_placement_compat
+      )
     , "NUMA-aware placement of VMs (deprecated, use XAPI setting)"
     )
   ; ( "pci-quarantine"

ocaml/xenopsd/xc/xenops_server_xen.ml

@@ -5294,8 +5294,7 @@ let init () =
         {Xs_protocol.ACL.owner= 0; other= Xs_protocol.ACL.READ; acl= []}
   ) ;
   Device.Backend.init () ;
-  Xenops_server.default_numa_affinity_policy :=
-    if !Xenopsd.numa_placement_compat then Best_effort else Any ;
+  Xenopsd.set_default_numa_placement () ;
   info "Default NUMA affinity policy is '%s'"
     Xenops_server.(string_of_numa_affinity_policy !default_numa_affinity_policy) ;
   Xenops_server.numa_placement := !Xenops_server.default_numa_affinity_policy ;