2 * NUMA parameter parsing routines
4 * Copyright (c) 2014 Fujitsu Ltd.
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to deal
8 * in the Software without restriction, including without limitation the rights
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10 * copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
25 #include "sysemu/numa.h"
26 #include "exec/cpu-common.h"
27 #include "qemu/bitmap.h"
29 #include "qemu/error-report.h"
30 #include "include/exec/cpu-common.h" /* for RAM_ADDR_FMT */
31 #include "qapi-visit.h"
32 #include "qapi/opts-visitor.h"
33 #include "qapi/dealloc-visitor.h"
34 #include "hw/boards.h"
35 #include "sysemu/hostmem.h"
36 #include "qmp-commands.h"
37 #include "hw/mem/pc-dimm.h"
38 #include "qemu/option.h"
39 #include "qemu/config-file.h"
41 QemuOptsList qemu_numa_opts = {
43 .implied_opt_name = "type",
44 .head = QTAILQ_HEAD_INITIALIZER(qemu_numa_opts.head),
45 .desc = { { 0 } } /* validated with OptsVisitor */
48 static int have_memdevs = -1;
49 static int max_numa_nodeid; /* Highest specified NUMA node ID, plus one.
50 * For all nodes, nodeid < max_numa_nodeid
53 NodeInfo numa_info[MAX_NODES];
55 static void numa_node_parse(NumaNodeOptions *node, QemuOpts *opts, Error **errp)
58 uint16List *cpus = NULL;
60 if (node->has_nodeid) {
61 nodenr = node->nodeid;
63 nodenr = nb_numa_nodes;
66 if (nodenr >= MAX_NODES) {
67 error_setg(errp, "Max number of NUMA nodes reached: %"
72 if (numa_info[nodenr].present) {
73 error_setg(errp, "Duplicate NUMA nodeid: %" PRIu16, nodenr);
77 for (cpus = node->cpus; cpus; cpus = cpus->next) {
78 if (cpus->value >= max_cpus) {
80 "CPU index (%" PRIu16 ")"
81 " should be smaller than maxcpus (%d)",
82 cpus->value, max_cpus);
85 bitmap_set(numa_info[nodenr].node_cpu, cpus->value, 1);
88 if (node->has_mem && node->has_memdev) {
89 error_setg(errp, "qemu: cannot specify both mem= and memdev=");
93 if (have_memdevs == -1) {
94 have_memdevs = node->has_memdev;
96 if (node->has_memdev != have_memdevs) {
97 error_setg(errp, "qemu: memdev option must be specified for either "
103 uint64_t mem_size = node->mem;
104 const char *mem_str = qemu_opt_get(opts, "mem");
105 /* Fix up legacy suffix-less format */
106 if (g_ascii_isdigit(mem_str[strlen(mem_str) - 1])) {
109 numa_info[nodenr].node_mem = mem_size;
111 if (node->has_memdev) {
113 o = object_resolve_path_type(node->memdev, TYPE_MEMORY_BACKEND, NULL);
115 error_setg(errp, "memdev=%s is ambiguous", node->memdev);
120 numa_info[nodenr].node_mem = object_property_get_int(o, "size", NULL);
121 numa_info[nodenr].node_memdev = MEMORY_BACKEND(o);
123 numa_info[nodenr].present = true;
124 max_numa_nodeid = MAX(max_numa_nodeid, nodenr + 1);
127 static int parse_numa(void *opaque, QemuOpts *opts, Error **errp)
129 NumaOptions *object = NULL;
133 OptsVisitor *ov = opts_visitor_new(opts);
134 visit_type_NumaOptions(opts_get_visitor(ov), &object, NULL, &err);
135 opts_visitor_cleanup(ov);
142 switch (object->kind) {
143 case NUMA_OPTIONS_KIND_NODE:
144 numa_node_parse(object->node, opts, &err);
157 error_report_err(err);
160 QapiDeallocVisitor *dv = qapi_dealloc_visitor_new();
161 visit_type_NumaOptions(qapi_dealloc_get_visitor(dv),
162 &object, NULL, NULL);
163 qapi_dealloc_visitor_cleanup(dv);
169 static char *enumerate_cpus(unsigned long *cpus, int max_cpus)
173 GString *s = g_string_new(NULL);
175 for (cpu = find_first_bit(cpus, max_cpus);
177 cpu = find_next_bit(cpus, max_cpus, cpu + 1)) {
178 g_string_append_printf(s, "%s%d", first ? "" : " ", cpu);
181 return g_string_free(s, FALSE);
184 static void validate_numa_cpus(void)
187 DECLARE_BITMAP(seen_cpus, MAX_CPUMASK_BITS);
189 bitmap_zero(seen_cpus, MAX_CPUMASK_BITS);
190 for (i = 0; i < nb_numa_nodes; i++) {
191 if (bitmap_intersects(seen_cpus, numa_info[i].node_cpu,
193 bitmap_and(seen_cpus, seen_cpus,
194 numa_info[i].node_cpu, MAX_CPUMASK_BITS);
195 error_report("CPU(s) present in multiple NUMA nodes: %s",
196 enumerate_cpus(seen_cpus, max_cpus));;
199 bitmap_or(seen_cpus, seen_cpus,
200 numa_info[i].node_cpu, MAX_CPUMASK_BITS);
203 if (!bitmap_full(seen_cpus, max_cpus)) {
205 bitmap_complement(seen_cpus, seen_cpus, max_cpus);
206 msg = enumerate_cpus(seen_cpus, max_cpus);
207 error_report("warning: CPU(s) not present in any NUMA nodes: %s", msg);
208 error_report("warning: All CPU(s) up to maxcpus should be described "
214 void parse_numa_opts(MachineClass *mc)
218 if (qemu_opts_foreach(qemu_find_opts("numa"), parse_numa, NULL, NULL)) {
222 assert(max_numa_nodeid <= MAX_NODES);
224 /* No support for sparse NUMA node IDs yet: */
225 for (i = max_numa_nodeid - 1; i >= 0; i--) {
226 /* Report large node IDs first, to make mistakes easier to spot */
227 if (!numa_info[i].present) {
228 error_report("numa: Node ID missing: %d", i);
233 /* This must be always true if all nodes are present: */
234 assert(nb_numa_nodes == max_numa_nodeid);
236 if (nb_numa_nodes > 0) {
239 if (nb_numa_nodes > MAX_NODES) {
240 nb_numa_nodes = MAX_NODES;
243 /* If no memory size is given for any node, assume the default case
244 * and distribute the available memory equally across all nodes
246 for (i = 0; i < nb_numa_nodes; i++) {
247 if (numa_info[i].node_mem != 0) {
251 if (i == nb_numa_nodes) {
252 uint64_t usedmem = 0;
254 /* On Linux, each node's border has to be 8MB aligned,
255 * the final node gets the rest.
257 for (i = 0; i < nb_numa_nodes - 1; i++) {
258 numa_info[i].node_mem = (ram_size / nb_numa_nodes) &
260 usedmem += numa_info[i].node_mem;
262 numa_info[i].node_mem = ram_size - usedmem;
266 for (i = 0; i < nb_numa_nodes; i++) {
267 numa_total += numa_info[i].node_mem;
269 if (numa_total != ram_size) {
270 error_report("total memory for NUMA nodes (0x%" PRIx64 ")"
271 " should equal RAM size (0x" RAM_ADDR_FMT ")",
272 numa_total, ram_size);
276 for (i = 0; i < nb_numa_nodes; i++) {
277 if (!bitmap_empty(numa_info[i].node_cpu, MAX_CPUMASK_BITS)) {
281 /* Historically VCPUs were assigned in round-robin order to NUMA
282 * nodes. However it causes issues with guest not handling it nice
283 * in case where cores/threads from a multicore CPU appear on
284 * different nodes. So allow boards to override default distribution
285 * rule grouping VCPUs by socket so that VCPUs from the same socket
286 * would be on the same node.
288 if (i == nb_numa_nodes) {
289 for (i = 0; i < max_cpus; i++) {
290 unsigned node_id = i % nb_numa_nodes;
291 if (mc->cpu_index_to_socket_id) {
292 node_id = mc->cpu_index_to_socket_id(i) % nb_numa_nodes;
295 set_bit(i, numa_info[node_id].node_cpu);
299 validate_numa_cpus();
303 void numa_post_machine_init(void)
309 for (i = 0; i < nb_numa_nodes; i++) {
310 if (test_bit(cpu->cpu_index, numa_info[i].node_cpu)) {
317 static void allocate_system_memory_nonnuma(MemoryRegion *mr, Object *owner,
324 memory_region_init_ram_from_file(mr, owner, name, ram_size, false,
327 /* Legacy behavior: if allocation failed, fall back to
328 * regular RAM allocation.
331 error_report_err(err);
332 memory_region_init_ram(mr, owner, name, ram_size, &error_abort);
335 fprintf(stderr, "-mem-path not supported on this host\n");
339 memory_region_init_ram(mr, owner, name, ram_size, &error_abort);
341 vmstate_register_ram_global(mr);
344 void memory_region_allocate_system_memory(MemoryRegion *mr, Object *owner,
351 if (nb_numa_nodes == 0 || !have_memdevs) {
352 allocate_system_memory_nonnuma(mr, owner, name, ram_size);
356 memory_region_init(mr, owner, name, ram_size);
357 for (i = 0; i < MAX_NODES; i++) {
358 Error *local_err = NULL;
359 uint64_t size = numa_info[i].node_mem;
360 HostMemoryBackend *backend = numa_info[i].node_memdev;
364 MemoryRegion *seg = host_memory_backend_get_memory(backend, &local_err);
366 error_report_err(local_err);
370 if (memory_region_is_mapped(seg)) {
371 char *path = object_get_canonical_path_component(OBJECT(backend));
372 error_report("memory backend %s is used multiple times. Each "
373 "-numa option must use a different memdev value.",
378 memory_region_add_subregion(mr, addr, seg);
379 vmstate_register_ram_global(seg);
384 static void numa_stat_memory_devices(uint64_t node_mem[])
386 MemoryDeviceInfoList *info_list = NULL;
387 MemoryDeviceInfoList **prev = &info_list;
388 MemoryDeviceInfoList *info;
390 qmp_pc_dimm_device_list(qdev_get_machine(), &prev);
391 for (info = info_list; info; info = info->next) {
392 MemoryDeviceInfo *value = info->value;
395 switch (value->kind) {
396 case MEMORY_DEVICE_INFO_KIND_DIMM:
397 node_mem[value->dimm->node] += value->dimm->size;
404 qapi_free_MemoryDeviceInfoList(info_list);
407 void query_numa_node_mem(uint64_t node_mem[])
411 if (nb_numa_nodes <= 0) {
415 numa_stat_memory_devices(node_mem);
416 for (i = 0; i < nb_numa_nodes; i++) {
417 node_mem[i] += numa_info[i].node_mem;
421 static int query_memdev(Object *obj, void *opaque)
423 MemdevList **list = opaque;
424 MemdevList *m = NULL;
427 if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) {
428 m = g_malloc0(sizeof(*m));
430 m->value = g_malloc0(sizeof(*m->value));
432 m->value->size = object_property_get_int(obj, "size",
438 m->value->merge = object_property_get_bool(obj, "merge",
444 m->value->dump = object_property_get_bool(obj, "dump",
450 m->value->prealloc = object_property_get_bool(obj,
456 m->value->policy = object_property_get_enum(obj,
464 object_property_get_uint16List(obj, "host-nodes",
465 &m->value->host_nodes, &err);
482 MemdevList *qmp_query_memdev(Error **errp)
485 MemdevList *list = NULL;
487 obj = object_get_objects_root();
492 if (object_child_foreach(obj, query_memdev, &list) != 0) {
499 qapi_free_MemdevList(list);