node_memory: Add new parameter field to tune the new sysfs knob

Upstream kernel introduced new sysfs knob "merge_across_nodes" to
specify if pages from different numa nodes can be merged. When set
to 0, only pages which physically reside in the memory area of
same NUMA node can be merged. When set to 1, pages from all nodes
can be merged.

This patch supports the tuning by adding new param field
"shm_merge_across_nodes".

include/libvirt/libvirt.h.in

@@ -4490,6 +4490,18 @@ typedef virMemoryParameter *virMemoryParameterPtr;
  */
 # define VIR_NODE_MEMORY_SHARED_FULL_SCANS         "shm_full_scans"
 
+/* VIR_NODE_MEMORY_SHARED_MERGE_ACROSS_NODES:
+ *
+ * Macro for typed parameter that represents whether pages from
+ * different NUMA nodes can be merged. The parameter has type int,
+ * when its value is 0, only pages which physically reside in the
+ * memory area of same NUMA node are merged; When its value is 1,
+ * pages from all nodes can be merged. Other values are reserved
+ * for future use.
+ */
+# define VIR_NODE_MEMORY_SHARED_MERGE_ACROSS_NODES "shm_merge_across_nodes"
+
+
 int virNodeGetMemoryParameters(virConnectPtr conn,
                                virTypedParameterPtr params,
                                int *nparams,

src/nodeinfo.c

@@ -1005,6 +1005,8 @@ nodeSetMemoryParameters(virConnectPtr conn ATTRIBUTE_UNUSED,
                                        VIR_TYPED_PARAM_UINT,
                                        VIR_NODE_MEMORY_SHARED_SLEEP_MILLISECS,
                                        VIR_TYPED_PARAM_UINT,
+                                       VIR_NODE_MEMORY_SHARED_MERGE_ACROSS_NODES,
+                                       VIR_TYPED_PARAM_UINT,
                                        NULL) < 0)
         return -1;
 
@@ -1022,6 +1024,13 @@ nodeSetMemoryParameters(virConnectPtr conn ATTRIBUTE_UNUSED,
                          VIR_NODE_MEMORY_SHARED_SLEEP_MILLISECS)) {
             ret = nodeSetMemoryParameterValue("sleep_millisecs", param);
 
+            /* Out of memory */
+            if (ret == -2)
+                return -1;
+        } else if (STREQ(param->field,
+                         VIR_NODE_MEMORY_SHARED_MERGE_ACROSS_NODES)) {
+            ret = nodeSetMemoryParameterValue("merge_across_nodes", param);
+
             /* Out of memory */
             if (ret == -2)
                 return -1;
@@ -1061,7 +1070,8 @@ nodeGetMemoryParameterValue(const char *field,
         *tmp = '\0';
 
     if (STREQ(field, "pages_to_scan")   ||
-        STREQ(field, "sleep_millisecs"))
+        STREQ(field, "sleep_millisecs") ||
+        STREQ(field, "merge_across_nodes"))
         rc = virStrToLong_ui(buf, NULL, 10, (unsigned int *)value);
     else if (STREQ(field, "pages_shared")    ||
              STREQ(field, "pages_sharing")   ||
@@ -1084,7 +1094,7 @@ cleanup:
 }
 #endif
 
-#define NODE_MEMORY_PARAMETERS_NUM 7
+#define NODE_MEMORY_PARAMETERS_NUM 8
 int
 nodeGetMemoryParameters(virConnectPtr conn ATTRIBUTE_UNUSED,
                         virTypedParameterPtr params ATTRIBUTE_UNUSED,
@@ -1096,6 +1106,7 @@ nodeGetMemoryParameters(virConnectPtr conn ATTRIBUTE_UNUSED,
 #ifdef __linux__
     unsigned int pages_to_scan;
     unsigned int sleep_millisecs;
+    unsigned int merge_across_nodes;
     unsigned long long pages_shared;
     unsigned long long pages_sharing;
     unsigned long long pages_unshared;
@@ -1189,6 +1200,17 @@ nodeGetMemoryParameters(virConnectPtr conn ATTRIBUTE_UNUSED,
 
             break;
 
+        case 7:
+            if (nodeGetMemoryParameterValue("merge_across_nodes",
+                                            &merge_across_nodes) < 0)
+                return -1;
+
+            if (virTypedParameterAssign(param, VIR_NODE_MEMORY_SHARED_MERGE_ACROSS_NODES,
+                                        VIR_TYPED_PARAM_UINT, merge_across_nodes) < 0)
+                return -1;
+
+            break;
+
         default:
             break;
         }

tools/virsh-host.c

@@ -920,6 +920,8 @@ static const vshCmdOptDef opts_node_memory_tune[] = {
     {"shm-sleep-millisecs", VSH_OT_INT, VSH_OFLAG_NONE,
       N_("number of millisecs the shared memory service should "
          "sleep before next scan")},
+    {"shm-merge-across-nodes", VSH_OT_INT, VSH_OFLAG_NONE,
+      N_("Specifies if pages from different numa nodes can be merged")},
     {NULL, 0, 0, NULL}
 };
 
@@ -931,6 +933,7 @@ cmdNodeMemoryTune(vshControl *ctl, const vshCmd *cmd)
     unsigned int flags = 0;
     unsigned int shm_pages_to_scan = 0;
     unsigned int shm_sleep_millisecs = 0;
+    unsigned int shm_merge_across_nodes = 0;
     bool ret = false;
     int i = 0;
 
@@ -946,12 +949,21 @@ cmdNodeMemoryTune(vshControl *ctl, const vshCmd *cmd)
         return false;
     }
 
+    if (vshCommandOptUInt(cmd, "shm-merge-across-nodes",
+                          &shm_merge_across_nodes) < 0) {
+        vshError(ctl, "%s", _("invalid shm-merge-across-nodes number"));
+        return false;
+    }
+
     if (shm_pages_to_scan)
         nparams++;
 
     if (shm_sleep_millisecs)
         nparams++;
 
+    if (shm_merge_across_nodes)
+        nparams++;
+
     if (nparams == 0) {
         /* Get the number of memory parameters */
         if (virNodeGetMemoryParameters(ctl->conn, NULL, &nparams, flags) != 0) {
@@ -1003,6 +1015,14 @@ cmdNodeMemoryTune(vshControl *ctl, const vshCmd *cmd)
                 goto error;
         }
 
+        if (i < nparams && shm_merge_across_nodes) {
+            if (virTypedParameterAssign(&params[i++],
+                                        VIR_NODE_MEMORY_SHARED_MERGE_ACROSS_NODES,
+                                        VIR_TYPED_PARAM_UINT,
+                                        shm_merge_across_nodes) < 0)
+                goto error;
+        }
+
         if (virNodeSetMemoryParameters(ctl->conn, params, nparams, flags) != 0)
             goto error;
         else

tools/virsh.pod

@@ -303,7 +303,10 @@ Allows you to display or set the node memory parameters.
 I<shm-pages-to-scan> can be used to set the number of pages to scan
 before the shared memory service goes to sleep; I<shm-sleep-millisecs>
 can be used to set the number of millisecs the shared memory service should
-sleep before next scan.
+sleep before next scan; I<shm-merge-across-nodes> specifies if pages from
+different numa nodes can be merged. When set to 0, only pages which physically
+reside in the memory area of same NUMA node can be merged. When set to 1,
+pages from all nodes can be merged. Default to 1.
 
 =item B<capabilities>