WARNING - OLD ARCHIVES

This is an archived copy of the Xen.org mailing list, which we have preserved to ensure that existing links to archives are not broken. The live archive, which contains the latest emails, can be found at http://lists.xen.org/
   
 
 
Xen 
 
Home Products Support Community News
 
   
 

xen-devel

Re: [Xen-devel] [RFC][PATCH 1/2] export NUMA topology from xen

* Ryan Harper <ryanh@xxxxxxxxxx> [2007-04-24 10:30]:
> * Keir Fraser <keir@xxxxxxxxxxxxx> [2007-04-10 04:13]:
> > On 10/4/07 02:09, "Ryan Harper" <ryanh@xxxxxxxxxx> wrote:
> > 
> > > nr_nodes               : 4
> > > mem_chunks             : node0:0x0000000000000000-0x0000000190000000
> > >                          node1:0x0000000190000000-0x0000000300000000
> > >                          node2:0x0000000300000000-0x0000000470000000
> > >                          node3:0x0000000470000000-0x0000000640000000
> > > node_to_cpu            : node0:0-7
> > >                          node1:8-15
> > >                          node2:16-23
> > >                          node3:24-31
> > > 
> > > I've also reworked the the physinfo call to contain an array of
> > > cpu_to_node elements rather than node_to_cpu to support machines larger
> > > than 64-way.  I convert the array back to node_to_cpu for brevity in
> > > xm info display.
> > 
> > The same would make sense for memory regions (i.e., have a list of
> > memory-regions and include a node identifier for each one, rather than
> > mapping node-id to memory-region) as this would allow to have multiple
> > memory regions per node quite easily. But actually I'm not convinced that
> > allowing dom0 to read out the physical addresses of memory regions is at all
> > useful -- why would anyone care which particular physical address ranges
> > belong to a particular node? The hypercall to find amount of free memory per
> > node seems more useful, and probably sufficient by itself.
> 
> Updated.
> 
> - Dropped mem_chunks (removed that from existing ia64 NUMA physinfo)
> - Fixed up ia64 cpu_to_node_map array size (was MAX_NUMNODES, now
>   NR_CPUS)
> - Fixed sockets_per_node calculation (was bogus on Opteron systems)
> - Updated all arches physinfo call to use num_online_nodes() and new
>   sockets_per_node calculation
> 
> Untested on ia64, ppc.

Refreshed to changeset:   15200:bd3d6b4c52ec

-- 
Ryan Harper
Software Engineer; Linux Technology Center
IBM Corp., Austin, Tx
(512) 838-9253   T/L: 678-9253
ryanh@xxxxxxxxxx


diffstat output:
 tools/libxc/xc_misc.c                               |    3 
 tools/libxc/xenctrl.h                               |    1 
 tools/python/xen/lowlevel/xc/xc.c                   |   61 ++++++++++++++++----
 tools/python/xen/xend/XendNode.py                   |   50 ++++++++++++++++
 tools/xenmon/xenbaked.c                             |    3 
 tools/xenstat/libxenstat/src/xenstat.c              |    3 
 tools/xentrace/xentrace.c                           |    3 
 tools/xm-test/tests/info/02_info_compiledata_pos.py |    4 -
 xen/arch/ia64/xen/dom0_ops.c                        |   46 +--------------
 xen/arch/powerpc/sysctl.c                           |    6 -
 xen/arch/x86/sysctl.c                               |   33 +++++++++-
 xen/include/public/sysctl.h                         |    1 
 12 files changed, 152 insertions(+), 62 deletions(-)

Signed-off-by: Ryan Harper <ryanh@xxxxxxxxxx>
---
This patch modifies the physinfo hcall to export NUMA cpu_to_node
topology information.  The new physinfo hcall is integrated into libxc
and xend (xm info specifically).  Included in this patch is a minor
tweak to xm-test's xm info testcase.  I've also fixed the
sockets_per_node calculation.  The new fields in xm info are:


nr_cpus                : 32
nr_nodes               : 4
sockets_per_node       : 4
cores_per_socket       : 1
threads_per_core       : 2
...
node_to_cpu            : node0:0-7
                         node1:8-15
                         node2:16-23
                         node3:24-31

I've also reworked the the physinfo call to contain an array of
cpu_to_node elements rather than node_to_cpu to support machines larger
than 64-ways.  I convert the array back to node_to_cpu for brevity in
xm info display.

Signed-off-by: Ryan Harper <ryanh@xxxxxxxxxx>

diff -r 400a3dca237e tools/libxc/xc_misc.c
--- a/tools/libxc/xc_misc.c     Mon Apr 09 12:05:26 2007 +0100
+++ b/tools/libxc/xc_misc.c     Fri Apr 13 13:04:24 2007 -0500
@@ -59,6 +59,9 @@ int xc_physinfo(int xc_handle,
     DECLARE_SYSCTL;
 
     sysctl.cmd = XEN_SYSCTL_physinfo;
+
+    /* set pointers to caller's so memcpy doesn't clobber them */
+    sysctl.u.physinfo.cpu_to_node = put_info->cpu_to_node;
 
     if ( (ret = do_sysctl(xc_handle, &sysctl)) != 0 )
         return ret;
diff -r 400a3dca237e tools/libxc/xenctrl.h
--- a/tools/libxc/xenctrl.h     Mon Apr 09 12:05:26 2007 +0100
+++ b/tools/libxc/xenctrl.h     Fri Apr 13 13:04:24 2007 -0500
@@ -473,6 +473,7 @@ int xc_send_debug_keys(int xc_handle, ch
 int xc_send_debug_keys(int xc_handle, char *keys);
 
 typedef xen_sysctl_physinfo_t xc_physinfo_t;
+typedef uint32_t xc_cpu_to_node_t;
 int xc_physinfo(int xc_handle,
                 xc_physinfo_t *info);
 
diff -r 400a3dca237e tools/python/xen/lowlevel/xc/xc.c
--- a/tools/python/xen/lowlevel/xc/xc.c Mon Apr 09 12:05:26 2007 +0100
+++ b/tools/python/xen/lowlevel/xc/xc.c Fri Apr 13 15:41:39 2007 -0500
@@ -640,14 +640,26 @@ static PyObject *pyxc_pages_to_kib(XcObj
 }
 
 
+#define MAX_NR_CPUS 256
 static PyObject *pyxc_physinfo(XcObject *self)
 {
     xc_physinfo_t info;
     char cpu_cap[128], *p=cpu_cap, *q=cpu_cap;
-    int i;
+    int i,j, nr_cpus;
+    PyObject *ret_obj, *node_to_cpu_obj;
+    xc_cpu_to_node_t *map;
+
+    /* make space for cpu_to_node mapping, up to MAX_NR_CPUS cpus */
+    map = (xc_cpu_to_node_t *)malloc( sizeof(xc_cpu_to_node_t) * MAX_NR_CPUS);
+                                      
+    set_xen_guest_handle(info.cpu_to_node, map);
     
     if ( xc_physinfo(self->xc_handle, &info) != 0 )
         return pyxc_error_to_exception();
+
+    /* calc number of cpus */
+    nr_cpus = info.threads_per_core * info.cores_per_socket *
+              info.sockets_per_node * info.nr_nodes;
 
     *q=0;
     for(i=0;i<sizeof(info.hw_cap)/4;i++)
@@ -659,16 +671,43 @@ static PyObject *pyxc_physinfo(XcObject 
     if(q>cpu_cap)
         *(q-1)=0;
 
-    return Py_BuildValue("{s:i,s:i,s:i,s:i,s:l,s:l,s:l,s:i,s:s}",
-                         "threads_per_core", info.threads_per_core,
-                         "cores_per_socket", info.cores_per_socket,
-                         "sockets_per_node", info.sockets_per_node,
-                         "nr_nodes",         info.nr_nodes,
-                         "total_memory",     pages_to_kib(info.total_pages),
-                         "free_memory",      pages_to_kib(info.free_pages),
-                         "scrub_memory",     pages_to_kib(info.scrub_pages),
-                         "cpu_khz",          info.cpu_khz,
-                         "hw_caps",          cpu_cap);
+    ret_obj = Py_BuildValue("{s:i,s:i,s:i,s:l,s:l,s:l,s:i,s:s}",
+                            "threads_per_core", info.threads_per_core,
+                            "cores_per_socket", info.cores_per_socket,
+                            "sockets_per_node", info.sockets_per_node,
+                            "total_memory",     pages_to_kib(info.total_pages),
+                            "free_memory",      pages_to_kib(info.free_pages),
+                            "scrub_memory",     pages_to_kib(info.scrub_pages),
+                            "cpu_khz",          info.cpu_khz,
+                            "hw_caps",          cpu_cap);
+ 
+    /* node to cpu mappings */
+    node_to_cpu_obj = PyList_New(0);
+
+    /* make a list for each node */
+    for ( i=0; i<info.nr_nodes; i++)
+    {
+        PyObject *cpus = PyList_New(0);
+ 
+        /* walk the cpu_to_node array, for each cpu
+           which maps to node i, add to cpus list */
+        for ( j=0; j<nr_cpus; j++)
+        {
+            /* this cpu j maps to node i */
+            if ( i == (uint32_t)map[j])
+                PyList_Append(cpus, PyInt_FromLong(j));
+        }
+        PyList_Append(node_to_cpu_obj, cpus); 
+    }
+    /* add list of node to cpu mappings and nr_nodes to physinfo dictionary */
+    PyDict_SetItemString(ret_obj, "node_to_cpu",  node_to_cpu_obj);
+    PyDict_SetItemString(ret_obj, "nr_nodes", 
+                         Py_BuildValue("i", info.nr_nodes));
+
+    /* free malloc'd memory */
+    free(map);
+ 
+    return ret_obj;
 }
 
 static PyObject *pyxc_xeninfo(XcObject *self)
diff -r 400a3dca237e tools/python/xen/xend/XendNode.py
--- a/tools/python/xen/xend/XendNode.py Mon Apr 09 12:05:26 2007 +0100
+++ b/tools/python/xen/xend/XendNode.py Fri Apr 13 13:04:24 2007 -0500
@@ -547,6 +547,54 @@ class XendNode:
                 ['version', ver],
                 ['machine', mch]]
 
+    def list_to_rangepairs(self,cmap):
+            cmap.sort()
+            pairs = []
+            x = y = 0
+            for i in range(0,len(cmap)):
+                try:
+                    if ((cmap[y+1] - cmap[i]) > 1):
+                        pairs.append((cmap[x],cmap[y]))
+                        x = y = i+1
+                    else:
+                        y = y + 1
+                # if we go off the end, then just add x to y
+                except IndexError:
+                    pairs.append((cmap[x],cmap[y]))
+
+            return pairs
+
+    def format_pairs(self,pairs):
+            if not pairs:
+                return "no cpus"
+            out = ""
+            for f,s in pairs:
+                if (f==s):
+                    out += '%d'%f
+                else:
+                    out += '%d-%d'%(f,s)
+                out += ','
+            # trim trailing ','
+            return out[:-1]
+
+    def list_to_strrange(self,list):
+        return self.format_pairs(self.list_to_rangepairs(list))
+
+    def format_node_to_cpu(self, pinfo):
+        str=''
+        whitespace=''
+        try:
+            node_to_cpu=pinfo['node_to_cpu']
+            for i in range(0, pinfo['nr_nodes']):
+                str+='%snode%d:%s\n' % (whitespace,
+                                        i, 
+                                      self.list_to_strrange(node_to_cpu[i]))
+                whitespace='%25s' % ''        
+        except:
+            str='none\n'
+        return str[:-1];
+
+
     def physinfo(self):
         info = self.xc.physinfo()
 
@@ -559,6 +607,7 @@ class XendNode:
         # physinfo is in KiB, need it in MiB
         info['total_memory'] = info['total_memory'] / 1024
         info['free_memory']  = info['free_memory'] / 1024
+        info['node_to_cpu']  = self.format_node_to_cpu(info)
 
         ITEM_ORDER = ['nr_cpus',
                       'nr_nodes',
@@ -569,6 +618,7 @@ class XendNode:
                       'hw_caps',
                       'total_memory',
                       'free_memory',
+                      'node_to_cpu'
                       ]
 
         return [[k, info[k]] for k in ITEM_ORDER]
diff -r 400a3dca237e tools/xenmon/xenbaked.c
--- a/tools/xenmon/xenbaked.c   Mon Apr 09 12:05:26 2007 +0100
+++ b/tools/xenmon/xenbaked.c   Fri Apr 13 13:04:24 2007 -0500
@@ -448,6 +448,9 @@ unsigned int get_num_cpus(void)
     int xc_handle = xc_interface_open();
     int ret;
 
+    /* ensure node_to_cpu is NULL */
+    memset(&physinfo, 0, sizeof(physinfo));
+
     ret = xc_physinfo(xc_handle, &physinfo);
 
     if ( ret != 0 )
diff -r 400a3dca237e tools/xenstat/libxenstat/src/xenstat.c
--- a/tools/xenstat/libxenstat/src/xenstat.c    Mon Apr 09 12:05:26 2007 +0100
+++ b/tools/xenstat/libxenstat/src/xenstat.c    Fri Apr 13 13:04:24 2007 -0500
@@ -147,6 +147,9 @@ xenstat_node *xenstat_get_node(xenstat_h
 
        /* Store the handle in the node for later access */
        node->handle = handle;
+
+       /* ensure node_to_cpu is NULL */
+       memset(&physinfo, 0, sizeof(physinfo));
 
        /* Get information about the physical system */
        if (xc_physinfo(handle->xc_handle, &physinfo) < 0) {
diff -r 400a3dca237e tools/xentrace/xentrace.c
--- a/tools/xentrace/xentrace.c Mon Apr 09 12:05:26 2007 +0100
+++ b/tools/xentrace/xentrace.c Fri Apr 13 13:04:24 2007 -0500
@@ -260,6 +260,9 @@ unsigned int get_num_cpus(void)
     int xc_handle = xc_interface_open();
     int ret;
     
+    /* ensure node_to_cpu is NULL */
+    memset(&physinfo, 0, sizeof(physinfo));
+
     ret = xc_physinfo(xc_handle, &physinfo);
     
     if ( ret != 0 )
diff -r 400a3dca237e tools/xm-test/tests/info/02_info_compiledata_pos.py
--- a/tools/xm-test/tests/info/02_info_compiledata_pos.py       Mon Apr 09 
12:05:26 2007 +0100
+++ b/tools/xm-test/tests/info/02_info_compiledata_pos.py       Fri Apr 13 
13:04:24 2007 -0500
@@ -18,9 +18,7 @@ for line in lines:
 for line in lines:
     pieces = line.split(" : ", 1)
 
-    if len(pieces) < 2:
-        FAIL("Found invalid line: [%s]" % line)
-    else:
+    if len(pieces) > 1:
         map[pieces[0]] = pieces[1]
 
 for field in ["cores_per_socket", "threads_per_core", "cpu_mhz",
diff -r 400a3dca237e xen/arch/ia64/xen/dom0_ops.c
--- a/xen/arch/ia64/xen/dom0_ops.c      Mon Apr 09 12:05:26 2007 +0100
+++ b/xen/arch/ia64/xen/dom0_ops.c      Fri Apr 13 13:20:38 2007 -0500
@@ -239,8 +239,7 @@ long arch_do_sysctl(xen_sysctl_t *op, XE
     {
 #ifdef IA64_NUMA_PHYSINFO
         int i;
-        node_data_t *chunks;
-        u64 *map, cpu_to_node_map[MAX_NUMNODES];
+        uint32_t *map, cpu_to_node_map[NR_CPUS];
 #endif
 
         xen_sysctl_physinfo_t *pi = &op->u.physinfo;
@@ -249,11 +248,9 @@ long arch_do_sysctl(xen_sysctl_t *op, XE
             cpus_weight(cpu_sibling_map[0]);
         pi->cores_per_socket =
             cpus_weight(cpu_core_map[0]) / pi->threads_per_core;
-        pi->sockets_per_node = 
-            num_online_cpus() / cpus_weight(cpu_core_map[0]);
-#ifndef IA64_NUMA_PHYSINFO
-        pi->nr_nodes         = 1; 
-#endif
+        pi->nr_nodes         = num_online_nodes();
+        pi->sockets_per_node = num_online_cpus() / 
+            (pi->nr_nodes * pi->cores_per_socket * pi->threads_per_core);
         pi->total_pages      = total_pages; 
         pi->free_pages       = avail_domheap_pages();
         pi->scrub_pages      = avail_scrub_pages();
@@ -263,41 +260,6 @@ long arch_do_sysctl(xen_sysctl_t *op, XE
         ret = 0;
 
 #ifdef IA64_NUMA_PHYSINFO
-        /* fetch memory_chunk pointer from guest */
-        get_xen_guest_handle(chunks, pi->memory_chunks);
-
-        printk("chunks=%p, num_node_memblks=%u\n", chunks, num_node_memblks);
-        /* if it is set, fill out memory chunk array */
-        if (chunks != NULL) {
-            if (num_node_memblks == 0) {
-                /* Non-NUMA machine.  Put pseudo-values.  */
-                node_data_t data;
-                data.node_start_pfn = 0;
-                data.node_spanned_pages = total_pages;
-                data.node_id = 0;
-                /* copy memory chunk structs to guest */
-                if (copy_to_guest_offset(pi->memory_chunks, 0, &data, 1)) {
-                    ret = -EFAULT;
-                    break;
-                }
-            } else {
-                for (i = 0; i < num_node_memblks && i < PUBLIC_MAXCHUNKS; i++) 
{
-                    node_data_t data;
-                    data.node_start_pfn = node_memblk[i].start_paddr >>
-                                          PAGE_SHIFT;
-                    data.node_spanned_pages = node_memblk[i].size >> 
PAGE_SHIFT;
-                    data.node_id = node_memblk[i].nid;
-                    /* copy memory chunk structs to guest */
-                    if (copy_to_guest_offset(pi->memory_chunks, i, &data, 1)) {
-                        ret = -EFAULT;
-                        break;
-                    }
-                }
-            }
-        }
-        /* set number of notes */
-        pi->nr_nodes = num_online_nodes();
-
         /* fetch cpu_to_node pointer from guest */
         get_xen_guest_handle(map, pi->cpu_to_node);
 
diff -r 400a3dca237e xen/arch/powerpc/sysctl.c
--- a/xen/arch/powerpc/sysctl.c Mon Apr 09 12:05:26 2007 +0100
+++ b/xen/arch/powerpc/sysctl.c Fri Apr 13 13:09:31 2007 -0500
@@ -45,10 +45,10 @@ long arch_do_sysctl(struct xen_sysctl *s
             cpus_weight(cpu_sibling_map[0]);
         pi->cores_per_socket =
             cpus_weight(cpu_core_map[0]) / pi->threads_per_core;
-        pi->sockets_per_node = 
-            num_online_cpus() / cpus_weight(cpu_core_map[0]);
+        pi->sockets_per_node = num_online_cpus() / 
+            (num_online_nodes() * pi->cores_per_socket * pi->threads_per_core);
 
-        pi->nr_nodes         = 1;
+        pi->nr_nodes         = num_online_nodes();
         pi->total_pages      = total_pages;
         pi->free_pages       = avail_domheap_pages();
         pi->cpu_khz          = cpu_khz;
diff -r 400a3dca237e xen/arch/x86/sysctl.c
--- a/xen/arch/x86/sysctl.c     Mon Apr 09 12:05:26 2007 +0100
+++ b/xen/arch/x86/sysctl.c     Fri Apr 13 13:11:15 2007 -0500
@@ -23,6 +23,10 @@
 #include <asm/hvm/hvm.h>
 #include <asm/hvm/support.h>
 #include <asm/processor.h>
+#include <asm/numa.h>
+#include <xen/nodemask.h>
+
+#define get_xen_guest_handle(val, hnd)  do { val = (hnd).p; } while (0)
 
 long arch_do_sysctl(
     struct xen_sysctl *sysctl, XEN_GUEST_HANDLE(xen_sysctl_t) u_sysctl)
@@ -34,16 +38,19 @@ long arch_do_sysctl(
 
     case XEN_SYSCTL_physinfo:
     {
+        int i;
+        uint32_t *map, cpu_to_node_map[NR_CPUS];
+
         xen_sysctl_physinfo_t *pi = &sysctl->u.physinfo;
 
         pi->threads_per_core =
             cpus_weight(cpu_sibling_map[0]);
         pi->cores_per_socket =
             cpus_weight(cpu_core_map[0]) / pi->threads_per_core;
-        pi->sockets_per_node = 
-            num_online_cpus() / cpus_weight(cpu_core_map[0]);
+        pi->nr_nodes = num_online_nodes();
+        pi->sockets_per_node = num_online_cpus() / 
+            (pi->nr_nodes * pi->cores_per_socket * pi->threads_per_core);
 
-        pi->nr_nodes         = 1;
         pi->total_pages      = total_pages;
         pi->free_pages       = avail_domheap_pages();
         pi->scrub_pages      = avail_scrub_pages();
@@ -51,6 +58,26 @@ long arch_do_sysctl(
         memset(pi->hw_cap, 0, sizeof(pi->hw_cap));
         memcpy(pi->hw_cap, boot_cpu_data.x86_capability, NCAPINTS*4);
         ret = 0;
+
+        /* fetch cpu_to_node pointer from guest */
+        get_xen_guest_handle(map, sysctl->u.physinfo.cpu_to_node);
+
+        /* if set, fill out cpu_to_node array */
+        if ( map != NULL )
+        {
+            /* for each cpu, mark in which node the cpu belongs */
+            memset(cpu_to_node_map, 0, sizeof(cpu_to_node_map));
+            for ( i = 0; i < num_online_cpus(); i++)
+            {
+                cpu_to_node_map[i]=cpu_to_node(i);
+                if ( copy_to_guest_offset(sysctl->u.physinfo.cpu_to_node,
+                                          i, &(cpu_to_node_map[i]), 1) ) {
+                    ret = -EFAULT;
+                    break;
+                }
+            }
+        }
+
         if ( copy_to_guest(u_sysctl, sysctl, 1) )
             ret = -EFAULT;
     }
diff -r 400a3dca237e xen/include/public/sysctl.h
--- a/xen/include/public/sysctl.h       Mon Apr 09 12:05:26 2007 +0100
+++ b/xen/include/public/sysctl.h       Fri Apr 13 13:04:24 2007 -0500
@@ -85,6 +85,7 @@ struct xen_sysctl_physinfo {
     uint64_aligned_t free_pages;
     uint64_aligned_t scrub_pages;
     uint32_t hw_cap[8];
+    XEN_GUEST_HANDLE(uint32_t) cpu_to_node;
 };
 typedef struct xen_sysctl_physinfo xen_sysctl_physinfo_t;
 DEFINE_XEN_GUEST_HANDLE(xen_sysctl_physinfo_t);

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel

<Prev in Thread] Current Thread [Next in Thread>
  • Re: [Xen-devel] [RFC][PATCH 1/2] export NUMA topology from xen, Ryan Harper <=