NUMA claim handling improvements

doc/content/design/numa.md

@@ -112,7 +112,7 @@ This function receives as arguments a domain ID and the number of nodes
 this domain is using (acquired using `domain_get_numa_info_node_pages`)
 
 The number of NUMA nodes of the host (not domain) is reported by
-`Xenctrl.physinfo` which returns a value of type `physinfo`.
+`Xenctrlext.physinfo` which returns a value of type `physinfo`.
 
 ```diff
     index b4579862ff..491bd3fc73 100644

doc/content/xenopsd/walkthroughs/VM.build/Domain.build.md

@@ -64,7 +64,7 @@ to call:
     [wait_xen_free_mem](https://github.com/xapi-project/xen-api/blob/master/ocaml/xenopsd/xc/domain.ml#L236-L272)
     to wait (if necessary), for the Xen memory scrubber to catch up reclaiming memory.
     It
-    1. calls `Xenctrl.physinfo` which returns:
+    1. calls `Xenctrlext.physinfo` which returns:
        - `hostinfo.free_pages` - the free and already scrubbed pages (available)
        - `host.scrub_pages` - the not yet scrubbed pages (not yet available)
     2. repeats this until a timeout as long as `free_pages` is *lower*

ocaml/libs/xenctrl-ext/xenctrlext.ml

@@ -131,3 +131,8 @@ let domain_claim_pages handle domid ?(numa_node = NumaNode.none) nr_pages =
 let get_nr_nodes handle =
   let info = numainfo handle in
   Array.length info.memory
+
+let physinfo xc =
+  let info = Xenctrl.physinfo xc in
+  let emergency_reserve_pages = Nativeint.shift_left 1n 16 in
+  {info with free_pages= Nativeint.sub info.free_pages emergency_reserve_pages}

ocaml/libs/xenctrl-ext/xenctrlext.mli

@@ -107,3 +107,5 @@ val domain_claim_pages : handle -> domid -> ?numa_node:NumaNode.t -> int -> unit
 
 val get_nr_nodes : handle -> int
 (** Returns the count of NUMA nodes available in the system. *)
+
+val physinfo : Xenctrl.handle -> Xenctrl.physinfo

ocaml/squeezed/src/dune

@@ -13,6 +13,7 @@
     threads.posix
     unix
     xenctrl
+    xenctrl_ext
     xenstore
     xenstore.unix
     xenstore_transport

ocaml/squeezed/src/squeeze_xen.ml

@@ -579,7 +579,7 @@ let make_host ~verbose ~xc =
      pages -- this might cause something else to fail (eg domain builder?) *)
   while
     Int64.div
-      ((Xenctrl.physinfo xc).Xenctrl.scrub_pages |> Int64.of_nativeint)
+      ((Xenctrlext.physinfo xc).Xenctrl.scrub_pages |> Int64.of_nativeint)
       1024L
     <> 0L
   do
@@ -762,7 +762,7 @@ let make_host ~verbose ~xc =
   (* For the host free memory we sum the free pages and the pages needing
      scrubbing: we don't want to adjust targets simply because the scrubber is
      slow. *)
-  let physinfo = Xenctrl.physinfo xc in
+  let physinfo = Xenctrlext.physinfo xc in
   let free_pages_kib =
     Xenctrl.pages_to_kib (Int64.of_nativeint physinfo.Xenctrl.free_pages)
   and scrub_pages_kib =

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

@@ -11,6 +11,7 @@
     xapi-rrd
     xapi-stdext-unix
     xenctrl
+    xenctrl_ext
   )
 )
 

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

@@ -184,7 +184,7 @@ let dss_pcpus xc =
   let len = Array.length !physcpus in
   let newinfos =
     if len = 0 then (
-      let physinfo = Xenctrl.physinfo xc in
+      let physinfo = Xenctrlext.physinfo xc in
       let pcpus = physinfo.Xenctrl.nr_cpus in
       physcpus := if pcpus > 0 then Array.make pcpus 0L else [||] ;
       Xenctrl.pcpu_info xc pcpus
@@ -237,7 +237,7 @@ let count_power_state_running_domains domains =
     0 domains
 
 let dss_hostload xc domains =
-  let physinfo = Xenctrl.physinfo xc in
+  let physinfo = Xenctrlext.physinfo xc in
   let pcpus = physinfo.Xenctrl.nr_cpus in
   let rec sum acc n f =
     match n with n when n >= 0 -> sum (acc + f n) (n - 1) f | _ -> acc
@@ -298,7 +298,7 @@ let _ =
       let _, domains, _ = Xenctrl_lib.domain_snapshot xc in
       Process.initialise () ;
       (* Share one page per PCPU and dom each *)
-      let physinfo = Xenctrl.physinfo xc in
+      let physinfo = Xenctrlext.physinfo xc in
       let shared_page_count =
         physinfo.Xenctrl.nr_cpus
         + Int.max Rrd_interface.max_supported_vms (List.length domains)

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

@@ -169,7 +169,7 @@ let generate_host_sources xc counters =
   in
   let memory_reclaimed = bytes_of_kib memory_reclaimed in
   let memory_possibly_reclaimed = bytes_of_kib memory_possibly_reclaimed in
-  let physinfo = Xenctrl.physinfo xc in
+  let physinfo = Xenctrlext.physinfo xc in
   let total_kib =
     Xenctrl.pages_to_kib (Int64.of_nativeint physinfo.Xenctrl.total_pages)
   in

ocaml/xenopsd/xc/domain.ml

@@ -235,7 +235,7 @@ let wait_xen_free_mem ~xc ?(maximum_wait_time_seconds = 64) required_memory_kib
     : bool =
   let open Memory in
   let rec wait accumulated_wait_time_seconds =
-    let host_info = Xenctrl.physinfo xc in
+    let host_info = Xenctrlext.physinfo xc in
     let free_memory_kib =
       kib_of_pages (Int64.of_nativeint host_info.Xenctrl.free_pages)
     in
@@ -244,7 +244,10 @@ let wait_xen_free_mem ~xc ?(maximum_wait_time_seconds = 64) required_memory_kib
     in
     (* At exponentially increasing intervals, write  *)
     (* a debug message saying how long we've waited: *)
-    if is_power_of_2 accumulated_wait_time_seconds then
+    if
+      accumulated_wait_time_seconds = 0
+      || is_power_of_2 accumulated_wait_time_seconds
+    then
       debug
         "Waited %i second(s) for memory to become available: %Ld KiB free, %Ld \
          KiB scrub, %Ld KiB required"
@@ -272,7 +275,7 @@ let wait_xen_free_mem ~xc ?(maximum_wait_time_seconds = 64) required_memory_kib
 let make ~xc ~xs vm_info vcpus domain_config uuid final_uuid no_sharept
     num_of_vbds num_of_vifs =
   let open Xenctrl in
-  let host_info = Xenctrl.physinfo xc in
+  let host_info = Xenctrlext.physinfo xc in
 
   (* Confirm that the running hypervisor supports a specific capability. *)
   let assert_capability cap ~on_error =
@@ -1000,17 +1003,21 @@ let numa_placement domid ~vcpus ~cores ~memory affinity =
               __FUNCTION__ domid ;
             None
       in
-      let nr_pages = Int64.div memory 4096L |> Int64.to_int in
+      let nr_pages = (Int64.div memory 4096L |> Int64.to_int) - 32 in
       try
+        D.debug "NUMAClaim domid %d: local claim on node %d: %d pages" domid
+          node nr_pages ;
         Xenctrlext.domain_claim_pages xcext domid ~numa_node nr_pages ;
         set_vcpu_affinity cpu_affinity ;
         Some (node, memory)
       with
       | Xenctrlext.Not_available ->
           (* Xen does not provide the interface to claim pages from a single NUMA
              node, ignore the error and continue. *)
+          D.debug "NUMAClaim domid %d: local claim not available" domid ;
+          set_vcpu_affinity cpu_affinity ;
           None
-      | Xenctrlext.Unix_error (errno, _) ->
+      | Xenctrlext.Unix_error ((Unix.ENOMEM as errno), _) ->
           D.info
             "%s: unable to claim enough memory, domain %d won't be hosted in a \
              single NUMA node. (error %s)"
@@ -1109,10 +1116,33 @@ let build_pre ~xc ~xs ~vcpus ~memory ~hard_affinity domid =
               and cores =
                 Xenops_server.cores_of_numa_affinity_policy pin ~vcpus
               in
-              numa_placement domid ~vcpus ~cores
-                ~memory:(Int64.mul memory.xen_max_mib 1048576L)
-                affinity
-              |> Option.map fst
+              let memory =
+                Int64.(mul memory.build_start_mib (shift_left 1L 20))
+              in
+              match numa_placement domid ~vcpus ~cores ~memory affinity with
+              | None ->
+                  (* Always perform a global claim when NUMA placement is
+                     enabled, and single node claims failed or were
+                     unavailable:
+                     This tries to ensures that memory allocated for this
+                     domain won't use up memory claimed by other domains.
+                     If claims are mixed with non-claims then Xen can't
+                     currently guarantee that it would honour the existing
+                     claims.
+                     A failure here is a hard failure: we'd fail allocating
+                     memory later anyway
+                  *)
+                  let nr_pages =
+                    (Int64.div memory 4096L |> Int64.to_int) - 32
+                  in
+                  let xcext = Xenctrlext.get_handle () in
+                  D.debug "NUMAClaim domid %d: global claim: %d pages" domid
+                    nr_pages ;
+                  Xenctrlext.domain_claim_pages xcext domid
+                    ~numa_node:Xenctrlext.NumaNode.none nr_pages ;
+                  None
+              | Some (plan, _) ->
+                  Some plan
         )
   in
   let store_chan, console_chan = create_channels ~xc uuid domid in

ocaml/xenopsd/xc/emu_manager.ml

@@ -205,7 +205,7 @@ let non_debug_receive ?debug_callback cnx =
         let open Memory in
         let open Int64 in
         let open Xenctrl in
-        let p = Xenctrl.physinfo xc in
+        let p = Xenctrlext.physinfo xc in
         error "Memory F %Ld KiB S %Ld KiB T %Ld MiB"
           (p.free_pages |> of_nativeint |> kib_of_pages)
           (p.scrub_pages |> of_nativeint |> kib_of_pages)

ocaml/xenopsd/xc/memory_breakdown.ml

@@ -217,7 +217,7 @@ let print_memory_field_names () =
 
 (** Prints memory field values to the console. *)
 let print_memory_field_values xc xs =
-  let host = Xenctrl.physinfo xc in
+  let host = Xenctrlext.physinfo xc in
   let control_domain_info = Xenctrl.domain_getinfo xc 0 in
   let control_domain_id = control_domain_info.Xenctrl.handle in
   let guests =

ocaml/xenopsd/xc/memory_summary.ml

@@ -38,7 +38,7 @@ let _ =
     finished := !delay < 0. ;
     if !delay > 0. then Unix.sleepf !delay ;
     flush stdout ;
-    let physinfo = Xenctrl.physinfo xc in
+    let physinfo = Xenctrlext.physinfo xc in
     let one_page = 4096L in
     let total_pages = Int64.of_nativeint physinfo.Xenctrl.total_pages in
     let free_pages =

ocaml/xenopsd/xc/xenguestHelper.ml

@@ -205,7 +205,7 @@ let non_debug_receive ?debug_callback cnx =
         let open Memory in
         let open Int64 in
         let open Xenctrl in
-        let p = Xenctrl.physinfo xc in
+        let p = Xenctrlext.physinfo xc in
         (match log_type with Syslog.Debug -> debug | _ -> error)
           "Memory F %Ld KiB S %Ld KiB T %Ld MiB"
           (p.free_pages |> of_nativeint |> kib_of_pages)

ocaml/xenopsd/xc/xenops_server_xen.ml

@@ -1064,7 +1064,7 @@ module HOST = struct
         let pages_per_mib = 256L in
         Int64.(
           div
-            ((Xenctrl.physinfo xc).Xenctrl.total_pages |> of_nativeint)
+            ((Xenctrlext.physinfo xc).Xenctrl.total_pages |> of_nativeint)
             pages_per_mib
         )
     )