[Qemu-devel] [PATCH v3 for-2.3 1/2] numa: introduce machine callback for VCPU to node mapping

[Igor Mammedov]

Current default round-robin way of distributing VCPUs among
NUMA nodes might be wrong in case on multi-core/threads
CPUs. Making guests confused wrt topology where cores from
the same socket are on different nodes.

Allow a machine to override default mapping by providing
 MachineClass->cpu_index_to_socket_id()
callback which would allow it group VCPUs from a socket
on the same NUMA node.

Signed-off-by: Igor Mammedov <address@hidden>
---
v3:
  - split out numa/machine change into a separate patch
---
 include/hw/boards.h   |  5 +++++
 include/sysemu/numa.h |  3 ++-
 numa.c                | 18 +++++++++++++-----
 vl.c                  |  2 +-
 4 files changed, 21 insertions(+), 7 deletions(-)

diff --git a/include/hw/boards.h b/include/hw/boards.h
index 1feea2b..78838d1 100644
--- a/include/hw/boards.h
+++ b/include/hw/boards.h
@@ -82,6 +82,10 @@ bool machine_mem_merge(MachineState *machine);
  *    of HotplugHandler object, which handles hotplug operation
  *    for a given @dev. It may return NULL if @dev doesn't require
  *    any actions to be performed by hotplug handler.
+ * @cpu_index_to_socket_id:
+ *    used to provide @cpu_index to socket number mapping, allowing
+ *    a machine to group CPU threads belonging to the same socket/package
+ *    Returns: socket number given cpu_index belongs to.
  */
 struct MachineClass {
     /*< private >*/
@@ -118,6 +122,7 @@ struct MachineClass {
 
     HotplugHandler *(*get_hotplug_handler)(MachineState *machine,
                                            DeviceState *dev);
+    unsigned (*cpu_index_to_socket_id)(unsigned cpu_index);
 };
 
 /**
diff --git a/include/sysemu/numa.h b/include/sysemu/numa.h
index 5633b85..6523b4d 100644
--- a/include/sysemu/numa.h
+++ b/include/sysemu/numa.h
@@ -6,6 +6,7 @@
 #include "qemu/option.h"
 #include "sysemu/sysemu.h"
 #include "sysemu/hostmem.h"
+#include "hw/boards.h"
 
 extern int nb_numa_nodes;   /* Number of NUMA nodes */
 
@@ -16,7 +17,7 @@ typedef struct node_info {
     bool present;
 } NodeInfo;
 extern NodeInfo numa_info[MAX_NODES];
-void parse_numa_opts(void);
+void parse_numa_opts(MachineClass *mc);
 void numa_post_machine_init(void);
 void query_numa_node_mem(uint64_t node_mem[]);
 extern QemuOptsList qemu_numa_opts;
diff --git a/numa.c b/numa.c
index ffbec68..f1f571a 100644
--- a/numa.c
+++ b/numa.c
@@ -165,7 +165,7 @@ error:
     return -1;
 }
 
-void parse_numa_opts(void)
+void parse_numa_opts(MachineClass *mc)
 {
     int i;
 
@@ -233,13 +233,21 @@ void parse_numa_opts(void)
                 break;
             }
         }
-        /* assigning the VCPUs round-robin is easier to implement, guest OSes
-         * must cope with this anyway, because there are BIOSes out there in
-         * real machines which also use this scheme.
+        /* Historically VCPUs were assigned in round-robin order to NUMA
+         * nodes. However it causes issues with guest not handling it nice
+         * in case where cores/threads from a multicore CPU appear on
+         * different nodes. So allow boards to override default distribution
+         * rule grouping VCPUs by socket so that VCPUs from the same socket
+         * would be on the same node.
          */
         if (i == nb_numa_nodes) {
             for (i = 0; i < max_cpus; i++) {
-                set_bit(i, numa_info[i % nb_numa_nodes].node_cpu);
+                unsigned node_id = i % nb_numa_nodes;
+                if (mc->cpu_index_to_socket_id) {
+                    node_id = mc->cpu_index_to_socket_id(i) % nb_numa_nodes;
+                }
+
+                set_bit(i, numa_info[node_id].node_cpu);
             }
         }
     }
diff --git a/vl.c b/vl.c
index 69617d6..75ec292 100644
--- a/vl.c
+++ b/vl.c
@@ -4170,7 +4170,7 @@ int main(int argc, char **argv, char **envp)
     default_drive(default_floppy, snapshot, IF_FLOPPY, 0, FD_OPTS);
     default_drive(default_sdcard, snapshot, IF_SD, 0, SD_OPTS);
 
-    parse_numa_opts();
+    parse_numa_opts(machine_class);
 
     if (qemu_opts_foreach(qemu_find_opts("mon"), mon_init_func, NULL, 1) != 0) 
{
         exit(1);
-- 
1.8.3.1