[PATCH RFC v2 1/2] linux/vnuma: vNUMA for PV domu guest

[Elena Ufimtseva <ufimtseva@gmail.com>]

Requests NUMA topology info from Xen by issuing subop
hypercall. Initializes NUMA nodes, sets number of CPUs,
distance table and NUMA nodes memory ranges during boot.
vNUMA topology defined by user in VM config file. Memory
ranges are represented by structure vnuma_topology_info
where start and end of memory area are defined in guests
pfns numbers, constructed and aligned accordingly to
e820 domain map.
In case the received structure has errors, will fail to
dummy numa init.
Requires XEN with applied patches from vnuma patchset;

Changes since v1:
- moved the test for xen_pv_domain() into xen_numa_init;
- replaced memory block search/allocation by single memblock_alloc;
- moved xen_numa_init to vnuma.c from enlighten.c;
- moved memblock structure to public interface memory.h;
- specified signedness of vnuma topology structure members;
- removed excessive debug output;

TODO:
- consider common interface for Dom0, HVM and PV guests to provide
vNUMA topology;
- dynamic numa balancing at the time of this patch (kernel 3.11
6e4664525b1db28f8c4e1130957f70a94c19213e with boot parameter
numa_balancing=true that is such by default) crashes numa-enabled
guest. Investigate further.

Signed-off-by: Elena Ufimtseva <ufimtseva@gmail.com>
---
 arch/x86/include/asm/xen/vnuma.h |   12 +++++
 arch/x86/mm/numa.c               |    5 +++
 arch/x86/xen/Makefile            |    2 +-
 arch/x86/xen/vnuma.c             |   92 ++++++++++++++++++++++++++++++++++++++
 include/xen/interface/memory.h   |   27 +++++++++++
 5 files changed, 137 insertions(+), 1 deletion(-)
 create mode 100644 arch/x86/include/asm/xen/vnuma.h
 create mode 100644 arch/x86/xen/vnuma.c

diff --git a/arch/x86/include/asm/xen/vnuma.h b/arch/x86/include/asm/xen/vnuma.h
new file mode 100644
index 0000000..1bf4cae
--- /dev/null
+++ b/arch/x86/include/asm/xen/vnuma.h
@@ -0,0 +1,12 @@
+#ifndef _ASM_X86_VNUMA_H
+#define _ASM_X86_VNUMA_H
+
+#ifdef CONFIG_XEN
+int xen_vnuma_support(void);
+int xen_numa_init(void);
+#else
+int xen_vnuma_support(void) { return 0; };
+int xen_numa_init(void) {};
+#endif
+
+#endif /* _ASM_X86_VNUMA_H */
diff --git a/arch/x86/mm/numa.c b/arch/x86/mm/numa.c
index 8bf93ba..a95fadf 100644
--- a/arch/x86/mm/numa.c
+++ b/arch/x86/mm/numa.c
@@ -19,6 +19,7 @@
 #include <asm/amd_nb.h>
 
 #include "numa_internal.h"
+#include "asm/xen/vnuma.h"
 
 int __initdata numa_off;
 nodemask_t numa_nodes_parsed __initdata;
@@ -621,6 +622,10 @@ static int __init dummy_numa_init(void)
 void __init x86_numa_init(void)
 {
 	if (!numa_off) {
+#ifdef CONFIG_XEN
+		if (!numa_init(xen_numa_init))
+			return;
+#endif
 #ifdef CONFIG_X86_NUMAQ
 		if (!numa_init(numaq_numa_init))
 			return;
diff --git a/arch/x86/xen/Makefile b/arch/x86/xen/Makefile
index 96ab2c0..de9deab 100644
--- a/arch/x86/xen/Makefile
+++ b/arch/x86/xen/Makefile
@@ -13,7 +13,7 @@ CFLAGS_mmu.o			:= $(nostackp)
 obj-y		:= enlighten.o setup.o multicalls.o mmu.o irq.o \
 			time.o xen-asm.o xen-asm_$(BITS).o \
 			grant-table.o suspend.o platform-pci-unplug.o \
-			p2m.o
+			p2m.o vnuma.o
 
 obj-$(CONFIG_EVENT_TRACING) += trace.o
 
diff --git a/arch/x86/xen/vnuma.c b/arch/x86/xen/vnuma.c
new file mode 100644
index 0000000..3c6c73f
--- /dev/null
+++ b/arch/x86/xen/vnuma.c
@@ -0,0 +1,92 @@
+#include <linux/err.h>
+#include <linux/memblock.h>
+#include <xen/interface/xen.h>
+#include <xen/interface/memory.h>
+#include <asm/xen/interface.h>
+#include <asm/xen/hypercall.h>
+#include <asm/xen/vnuma.h>
+#ifdef CONFIG_NUMA
+/* Xen PV NUMA topology initialization */
+static unsigned int xen_vnuma_init = 0;
+int xen_vnuma_support()
+{
+	return xen_vnuma_init;
+}
+int __init xen_numa_init(void)
+{
+	int rc;
+	unsigned int i, j, cpu, idx, pcpus;
+	u64 phys, physd, physc;
+	unsigned int *vdistance, *cpu_to_node;
+	unsigned long mem_size, dist_size, cpu_to_node_size;
+	struct vnuma_memarea *varea;
+
+	struct vnuma_topology_info numa_topo = {
+		.domid = DOMID_SELF
+	};
+	rc = -EINVAL;
+	if (!xen_pv_domain())
+		return rc;
+	pcpus = num_possible_cpus();
+	mem_size =  pcpus * sizeof(struct vnuma_memarea);
+	dist_size = pcpus * pcpus * sizeof(*numa_topo.vdistance);
+	cpu_to_node_size = pcpus * sizeof(*numa_topo.cpu_to_node);
+	phys = memblock_alloc(mem_size, PAGE_SIZE);
+	physd = memblock_alloc(dist_size, PAGE_SIZE);
+	physc = memblock_alloc(cpu_to_node_size, PAGE_SIZE);
+	if (!phys || !physc || !physd)
+		goto vnumaout;
+	varea = __va(phys);
+	vdistance  = __va(physd);
+	cpu_to_node  = __va(physc);
+	set_xen_guest_handle(numa_topo.vmemarea, varea);
+	set_xen_guest_handle(numa_topo.vdistance, vdistance);
+	set_xen_guest_handle(numa_topo.cpu_to_node, cpu_to_node);
+	rc = HYPERVISOR_memory_op(XENMEM_get_vnuma_info, &numa_topo);
+	if (rc < 0)
+		goto vnumaout;
+	rc = -EINVAL;
+	if (numa_topo.nr_nodes == 0) {
+		/* will pass to dummy_numa_init */
+		goto vnumaout;
+	}
+	if (numa_topo.nr_nodes > num_possible_cpus()) {
+		pr_debug("vNUMA: Node without cpu is not supported in this version.\n");
+		goto vnumaout;
+	}
+	/*
+	 * NUMA nodes memory ranges are in pfns, constructed and
+	 * aligned based on e820 ram domain map
+	*/
+	for (i = 0; i < numa_topo.nr_nodes; i++) {
+		if (numa_add_memblk(i, varea[i].start, varea[i].end))
+			/* pass to numa_dummy_init */
+			goto vnumaout;
+		node_set(i, numa_nodes_parsed);
+	}
+	setup_nr_node_ids();
+	/* Setting the cpu, apicid to node */
+	for_each_cpu(cpu, cpu_possible_mask) {
+		set_apicid_to_node(cpu, cpu_to_node[cpu]);
+		numa_set_node(cpu, cpu_to_node[cpu]);
+		__apicid_to_node[cpu] = cpu_to_node[cpu];
+		cpumask_set_cpu(cpu, node_to_cpumask_map[cpu_to_node[cpu]]);
+	}
+	for (i = 0; i < numa_topo.nr_nodes; i++) {
+		for (j = 0; j < numa_topo.nr_nodes; j++) {
+			idx = (j * numa_topo.nr_nodes) + i;
+			numa_set_distance(i, j, *(vdistance + idx));
+		}
+	}
+	rc = 0;
+	xen_vnuma_init = 1;
+vnumaout:
+	if (phys)
+		memblock_free(__pa(phys), mem_size);
+	if (physd)
+		memblock_free(__pa(physd), dist_size);
+	if (physc)
+		memblock_free(__pa(physc), cpu_to_node_size);
+	return rc;
+}
+#endif
diff --git a/include/xen/interface/memory.h b/include/xen/interface/memory.h
index 2ecfe4f..4237f51 100644
--- a/include/xen/interface/memory.h
+++ b/include/xen/interface/memory.h
@@ -263,4 +263,31 @@ struct xen_remove_from_physmap {
 };
 DEFINE_GUEST_HANDLE_STRUCT(xen_remove_from_physmap);
 
+/* vNUMA structures */
+struct vnuma_memarea {
+	uint64_t start, end;
+};
+DEFINE_GUEST_HANDLE_STRUCT(vnuma_memarea);
+
+struct vnuma_topology_info {
+	/* OUT */
+	domid_t domid;
+	/* IN */
+	uint16_t nr_nodes; /* number of virtual numa nodes */
+	uint32_t _pad;
+	/* distance table */
+	GUEST_HANDLE(uint) vdistance;
+	/* cpu mapping to vnodes */
+	GUEST_HANDLE(uint) cpu_to_node;
+	/*
+	* array of numa memory areas constructed by Xen
+	* where start and end are pfn numbers of the area
+	* Xen takes into account domains e820 map
+	*/
+	GUEST_HANDLE(vnuma_memarea) vmemarea;
+};
+DEFINE_GUEST_HANDLE_STRUCT(vnuma_topology_info);
+
+#define XENMEM_get_vnuma_info	25
+
 #endif /* __XEN_PUBLIC_MEMORY_H__ */
-- 
1.7.10.4