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 "qemu/osdep.h"
26 #include "qemu/units.h"
27 #include "sysemu/hostmem.h"
28 #include "sysemu/numa.h"
29 #include "sysemu/sysemu.h"
30 #include "exec/cpu-common.h"
31 #include "exec/ramlist.h"
32 #include "qemu/bitmap.h"
33 #include "qemu/error-report.h"
34 #include "qapi/error.h"
35 #include "qapi/opts-visitor.h"
36 #include "qapi/qapi-visit-machine.h"
37 #include "sysemu/qtest.h"
38 #include "hw/core/cpu.h"
39 #include "hw/mem/pc-dimm.h"
40 #include "migration/vmstate.h"
41 #include "hw/boards.h"
42 #include "hw/mem/memory-device.h"
43 #include "qemu/option.h"
44 #include "qemu/config-file.h"
45 #include "qemu/cutils.h"
47 QemuOptsList qemu_numa_opts = {
49 .implied_opt_name = "type",
50 .head = QTAILQ_HEAD_INITIALIZER(qemu_numa_opts.head),
51 .desc = { { 0 } } /* validated with OptsVisitor */
54 static int have_memdevs;
55 bool numa_uses_legacy_mem(void)
61 static int max_numa_nodeid; /* Highest specified NUMA node ID, plus one.
62 * For all nodes, nodeid < max_numa_nodeid
65 static void parse_numa_node(MachineState *ms, NumaNodeOptions *node,
70 uint16List *cpus = NULL;
71 MachineClass *mc = MACHINE_GET_CLASS(ms);
72 unsigned int max_cpus = ms->smp.max_cpus;
73 NodeInfo *numa_info = ms->numa_state->nodes;
75 if (node->has_nodeid) {
76 nodenr = node->nodeid;
78 nodenr = ms->numa_state->num_nodes;
81 if (nodenr >= MAX_NODES) {
82 error_setg(errp, "Max number of NUMA nodes reached: %"
87 if (numa_info[nodenr].present) {
88 error_setg(errp, "Duplicate NUMA nodeid: %" PRIu16, nodenr);
92 for (cpus = node->cpus; cpus; cpus = cpus->next) {
93 CpuInstanceProperties props;
94 if (cpus->value >= max_cpus) {
96 "CPU index (%" PRIu16 ")"
97 " should be smaller than maxcpus (%d)",
98 cpus->value, max_cpus);
101 props = mc->cpu_index_to_instance_props(ms, cpus->value);
102 props.node_id = nodenr;
103 props.has_node_id = true;
104 machine_set_cpu_numa_node(ms, &props, &err);
106 error_propagate(errp, err);
111 have_memdevs = have_memdevs ? : node->has_memdev;
112 have_mem = have_mem ? : node->has_mem;
113 if ((node->has_mem && have_memdevs) || (node->has_memdev && have_mem)) {
114 error_setg(errp, "numa configuration should use either mem= or memdev=,"
115 "mixing both is not allowed");
120 if (!mc->numa_mem_supported) {
121 error_setg(errp, "Parameter -numa node,mem is not supported by this"
123 error_append_hint(errp, "Use -numa node,memdev instead\n");
127 numa_info[nodenr].node_mem = node->mem;
128 if (!qtest_enabled()) {
129 warn_report("Parameter -numa node,mem is deprecated,"
130 " use -numa node,memdev instead");
133 if (node->has_memdev) {
135 o = object_resolve_path_type(node->memdev, TYPE_MEMORY_BACKEND, NULL);
137 error_setg(errp, "memdev=%s is ambiguous", node->memdev);
142 numa_info[nodenr].node_mem = object_property_get_uint(o, "size", NULL);
143 numa_info[nodenr].node_memdev = MEMORY_BACKEND(o);
147 * If not set the initiator, set it to MAX_NODES. And if
148 * HMAT is enabled and this node has no cpus, QEMU will raise error.
150 numa_info[nodenr].initiator = MAX_NODES;
151 if (node->has_initiator) {
152 if (!ms->numa_state->hmat_enabled) {
153 error_setg(errp, "ACPI Heterogeneous Memory Attribute Table "
154 "(HMAT) is disabled, enable it with -machine hmat=on "
155 "before using any of hmat specific options");
159 if (node->initiator >= MAX_NODES) {
160 error_report("The initiator id %" PRIu16 " expects an integer "
161 "between 0 and %d", node->initiator,
166 numa_info[nodenr].initiator = node->initiator;
168 numa_info[nodenr].present = true;
169 max_numa_nodeid = MAX(max_numa_nodeid, nodenr + 1);
170 ms->numa_state->num_nodes++;
174 void parse_numa_distance(MachineState *ms, NumaDistOptions *dist, Error **errp)
176 uint16_t src = dist->src;
177 uint16_t dst = dist->dst;
178 uint8_t val = dist->val;
179 NodeInfo *numa_info = ms->numa_state->nodes;
181 if (src >= MAX_NODES || dst >= MAX_NODES) {
182 error_setg(errp, "Parameter '%s' expects an integer between 0 and %d",
183 src >= MAX_NODES ? "src" : "dst", MAX_NODES - 1);
187 if (!numa_info[src].present || !numa_info[dst].present) {
188 error_setg(errp, "Source/Destination NUMA node is missing. "
189 "Please use '-numa node' option to declare it first.");
193 if (val < NUMA_DISTANCE_MIN) {
194 error_setg(errp, "NUMA distance (%" PRIu8 ") is invalid, "
195 "it shouldn't be less than %d.",
196 val, NUMA_DISTANCE_MIN);
200 if (src == dst && val != NUMA_DISTANCE_MIN) {
201 error_setg(errp, "Local distance of node %d should be %d.",
202 src, NUMA_DISTANCE_MIN);
206 numa_info[src].distance[dst] = val;
207 ms->numa_state->have_numa_distance = true;
210 void parse_numa_hmat_lb(NumaState *numa_state, NumaHmatLBOptions *node,
213 int i, first_bit, last_bit;
214 uint64_t max_entry, temp_base, bitmap_copy;
215 NodeInfo *numa_info = numa_state->nodes;
216 HMAT_LB_Info *hmat_lb =
217 numa_state->hmat_lb[node->hierarchy][node->data_type];
218 HMAT_LB_Data lb_data = {};
219 HMAT_LB_Data *lb_temp;
222 if (node->initiator > numa_state->num_nodes) {
223 error_setg(errp, "Invalid initiator=%d, it should be less than %d",
224 node->initiator, numa_state->num_nodes);
227 if (node->target > numa_state->num_nodes) {
228 error_setg(errp, "Invalid target=%d, it should be less than %d",
229 node->target, numa_state->num_nodes);
232 if (!numa_info[node->initiator].has_cpu) {
233 error_setg(errp, "Invalid initiator=%d, it isn't an "
234 "initiator proximity domain", node->initiator);
237 if (!numa_info[node->target].present) {
238 error_setg(errp, "The target=%d should point to an existing node",
244 hmat_lb = g_malloc0(sizeof(*hmat_lb));
245 numa_state->hmat_lb[node->hierarchy][node->data_type] = hmat_lb;
246 hmat_lb->list = g_array_new(false, true, sizeof(HMAT_LB_Data));
248 hmat_lb->hierarchy = node->hierarchy;
249 hmat_lb->data_type = node->data_type;
250 lb_data.initiator = node->initiator;
251 lb_data.target = node->target;
253 if (node->data_type <= HMATLB_DATA_TYPE_WRITE_LATENCY) {
254 /* Input latency data */
256 if (!node->has_latency) {
257 error_setg(errp, "Missing 'latency' option");
260 if (node->has_bandwidth) {
261 error_setg(errp, "Invalid option 'bandwidth' since "
262 "the data type is latency");
266 /* Detect duplicate configuration */
267 for (i = 0; i < hmat_lb->list->len; i++) {
268 lb_temp = &g_array_index(hmat_lb->list, HMAT_LB_Data, i);
270 if (node->initiator == lb_temp->initiator &&
271 node->target == lb_temp->target) {
272 error_setg(errp, "Duplicate configuration of the latency for "
273 "initiator=%d and target=%d", node->initiator,
279 hmat_lb->base = hmat_lb->base ? hmat_lb->base : UINT64_MAX;
282 /* Calculate the temporary base and compressed latency */
283 max_entry = node->latency;
285 while (QEMU_IS_ALIGNED(max_entry, 10)) {
290 /* Calculate the max compressed latency */
291 temp_base = MIN(hmat_lb->base, temp_base);
292 max_entry = node->latency / hmat_lb->base;
293 max_entry = MAX(hmat_lb->range_bitmap, max_entry);
296 * For latency hmat_lb->range_bitmap record the max compressed
297 * latency which should be less than 0xFFFF (UINT16_MAX)
299 if (max_entry >= UINT16_MAX) {
300 error_setg(errp, "Latency %" PRIu64 " between initiator=%d and "
301 "target=%d should not differ from previously entered "
302 "min or max values on more than %d", node->latency,
303 node->initiator, node->target, UINT16_MAX - 1);
306 hmat_lb->base = temp_base;
307 hmat_lb->range_bitmap = max_entry;
311 * Set lb_info_provided bit 0 as 1,
312 * latency information is provided
314 numa_info[node->target].lb_info_provided |= BIT(0);
316 lb_data.data = node->latency;
317 } else if (node->data_type >= HMATLB_DATA_TYPE_ACCESS_BANDWIDTH) {
318 /* Input bandwidth data */
319 if (!node->has_bandwidth) {
320 error_setg(errp, "Missing 'bandwidth' option");
323 if (node->has_latency) {
324 error_setg(errp, "Invalid option 'latency' since "
325 "the data type is bandwidth");
328 if (!QEMU_IS_ALIGNED(node->bandwidth, MiB)) {
329 error_setg(errp, "Bandwidth %" PRIu64 " between initiator=%d and "
330 "target=%d should be 1MB aligned", node->bandwidth,
331 node->initiator, node->target);
335 /* Detect duplicate configuration */
336 for (i = 0; i < hmat_lb->list->len; i++) {
337 lb_temp = &g_array_index(hmat_lb->list, HMAT_LB_Data, i);
339 if (node->initiator == lb_temp->initiator &&
340 node->target == lb_temp->target) {
341 error_setg(errp, "Duplicate configuration of the bandwidth for "
342 "initiator=%d and target=%d", node->initiator,
348 hmat_lb->base = hmat_lb->base ? hmat_lb->base : 1;
350 if (node->bandwidth) {
351 /* Keep bitmap unchanged when bandwidth out of range */
352 bitmap_copy = hmat_lb->range_bitmap;
353 bitmap_copy |= node->bandwidth;
354 first_bit = ctz64(bitmap_copy);
355 temp_base = UINT64_C(1) << first_bit;
356 max_entry = node->bandwidth / temp_base;
357 last_bit = 64 - clz64(bitmap_copy);
360 * For bandwidth, first_bit record the base unit of bandwidth bits,
361 * last_bit record the last bit of the max bandwidth. The max
362 * compressed bandwidth should be less than 0xFFFF (UINT16_MAX)
364 if ((last_bit - first_bit) > UINT16_BITS ||
365 max_entry >= UINT16_MAX) {
366 error_setg(errp, "Bandwidth %" PRIu64 " between initiator=%d "
367 "and target=%d should not differ from previously "
368 "entered values on more than %d", node->bandwidth,
369 node->initiator, node->target, UINT16_MAX - 1);
372 hmat_lb->base = temp_base;
373 hmat_lb->range_bitmap = bitmap_copy;
377 * Set lb_info_provided bit 1 as 1,
378 * bandwidth information is provided
380 numa_info[node->target].lb_info_provided |= BIT(1);
382 lb_data.data = node->bandwidth;
387 g_array_append_val(hmat_lb->list, lb_data);
390 void parse_numa_hmat_cache(MachineState *ms, NumaHmatCacheOptions *node,
393 int nb_numa_nodes = ms->numa_state->num_nodes;
394 NodeInfo *numa_info = ms->numa_state->nodes;
395 NumaHmatCacheOptions *hmat_cache = NULL;
397 if (node->node_id >= nb_numa_nodes) {
398 error_setg(errp, "Invalid node-id=%" PRIu32 ", it should be less "
399 "than %d", node->node_id, nb_numa_nodes);
403 if (numa_info[node->node_id].lb_info_provided != (BIT(0) | BIT(1))) {
404 error_setg(errp, "The latency and bandwidth information of "
405 "node-id=%" PRIu32 " should be provided before memory side "
406 "cache attributes", node->node_id);
410 if (node->level < 1 || node->level >= HMAT_LB_LEVELS) {
411 error_setg(errp, "Invalid level=%" PRIu8 ", it should be larger than 0 "
412 "and less than or equal to %d", node->level,
417 assert(node->associativity < HMAT_CACHE_ASSOCIATIVITY__MAX);
418 assert(node->policy < HMAT_CACHE_WRITE_POLICY__MAX);
419 if (ms->numa_state->hmat_cache[node->node_id][node->level]) {
420 error_setg(errp, "Duplicate configuration of the side cache for "
421 "node-id=%" PRIu32 " and level=%" PRIu8,
422 node->node_id, node->level);
426 if ((node->level > 1) &&
427 ms->numa_state->hmat_cache[node->node_id][node->level - 1] &&
429 ms->numa_state->hmat_cache[node->node_id][node->level - 1]->size)) {
430 error_setg(errp, "Invalid size=%" PRIu64 ", the size of level=%" PRIu8
431 " should be less than the size(%" PRIu64 ") of "
432 "level=%u", node->size, node->level,
433 ms->numa_state->hmat_cache[node->node_id]
434 [node->level - 1]->size,
439 if ((node->level < HMAT_LB_LEVELS - 1) &&
440 ms->numa_state->hmat_cache[node->node_id][node->level + 1] &&
442 ms->numa_state->hmat_cache[node->node_id][node->level + 1]->size)) {
443 error_setg(errp, "Invalid size=%" PRIu64 ", the size of level=%" PRIu8
444 " should be larger than the size(%" PRIu64 ") of "
445 "level=%u", node->size, node->level,
446 ms->numa_state->hmat_cache[node->node_id]
447 [node->level + 1]->size,
452 hmat_cache = g_malloc0(sizeof(*hmat_cache));
453 memcpy(hmat_cache, node, sizeof(*hmat_cache));
454 ms->numa_state->hmat_cache[node->node_id][node->level] = hmat_cache;
457 void set_numa_options(MachineState *ms, NumaOptions *object, Error **errp)
461 if (!ms->numa_state) {
462 error_setg(errp, "NUMA is not supported by this machine-type");
466 switch (object->type) {
467 case NUMA_OPTIONS_TYPE_NODE:
468 parse_numa_node(ms, &object->u.node, &err);
473 case NUMA_OPTIONS_TYPE_DIST:
474 parse_numa_distance(ms, &object->u.dist, &err);
479 case NUMA_OPTIONS_TYPE_CPU:
480 if (!object->u.cpu.has_node_id) {
481 error_setg(&err, "Missing mandatory node-id property");
484 if (!ms->numa_state->nodes[object->u.cpu.node_id].present) {
485 error_setg(&err, "Invalid node-id=%" PRId64 ", NUMA node must be "
486 "defined with -numa node,nodeid=ID before it's used with "
487 "-numa cpu,node-id=ID", object->u.cpu.node_id);
491 machine_set_cpu_numa_node(ms, qapi_NumaCpuOptions_base(&object->u.cpu),
494 case NUMA_OPTIONS_TYPE_HMAT_LB:
495 if (!ms->numa_state->hmat_enabled) {
496 error_setg(errp, "ACPI Heterogeneous Memory Attribute Table "
497 "(HMAT) is disabled, enable it with -machine hmat=on "
498 "before using any of hmat specific options");
502 parse_numa_hmat_lb(ms->numa_state, &object->u.hmat_lb, &err);
507 case NUMA_OPTIONS_TYPE_HMAT_CACHE:
508 if (!ms->numa_state->hmat_enabled) {
509 error_setg(errp, "ACPI Heterogeneous Memory Attribute Table "
510 "(HMAT) is disabled, enable it with -machine hmat=on "
511 "before using any of hmat specific options");
515 parse_numa_hmat_cache(ms, &object->u.hmat_cache, &err);
525 error_propagate(errp, err);
528 static int parse_numa(void *opaque, QemuOpts *opts, Error **errp)
530 NumaOptions *object = NULL;
531 MachineState *ms = MACHINE(opaque);
533 Visitor *v = opts_visitor_new(opts);
535 visit_type_NumaOptions(v, NULL, &object, &err);
541 /* Fix up legacy suffix-less format */
542 if ((object->type == NUMA_OPTIONS_TYPE_NODE) && object->u.node.has_mem) {
543 const char *mem_str = qemu_opt_get(opts, "mem");
544 qemu_strtosz_MiB(mem_str, NULL, &object->u.node.mem);
547 set_numa_options(ms, object, &err);
550 qapi_free_NumaOptions(object);
552 error_propagate(errp, err);
559 /* If all node pair distances are symmetric, then only distances
560 * in one direction are enough. If there is even one asymmetric
561 * pair, though, then all distances must be provided. The
562 * distance from a node to itself is always NUMA_DISTANCE_MIN,
563 * so providing it is never necessary.
565 static void validate_numa_distance(MachineState *ms)
568 bool is_asymmetrical = false;
569 int nb_numa_nodes = ms->numa_state->num_nodes;
570 NodeInfo *numa_info = ms->numa_state->nodes;
572 for (src = 0; src < nb_numa_nodes; src++) {
573 for (dst = src; dst < nb_numa_nodes; dst++) {
574 if (numa_info[src].distance[dst] == 0 &&
575 numa_info[dst].distance[src] == 0) {
577 error_report("The distance between node %d and %d is "
578 "missing, at least one distance value "
579 "between each nodes should be provided.",
585 if (numa_info[src].distance[dst] != 0 &&
586 numa_info[dst].distance[src] != 0 &&
587 numa_info[src].distance[dst] !=
588 numa_info[dst].distance[src]) {
589 is_asymmetrical = true;
594 if (is_asymmetrical) {
595 for (src = 0; src < nb_numa_nodes; src++) {
596 for (dst = 0; dst < nb_numa_nodes; dst++) {
597 if (src != dst && numa_info[src].distance[dst] == 0) {
598 error_report("At least one asymmetrical pair of "
599 "distances is given, please provide distances "
600 "for both directions of all node pairs.");
608 static void complete_init_numa_distance(MachineState *ms)
611 NodeInfo *numa_info = ms->numa_state->nodes;
613 /* Fixup NUMA distance by symmetric policy because if it is an
614 * asymmetric distance table, it should be a complete table and
615 * there would not be any missing distance except local node, which
616 * is verified by validate_numa_distance above.
618 for (src = 0; src < ms->numa_state->num_nodes; src++) {
619 for (dst = 0; dst < ms->numa_state->num_nodes; dst++) {
620 if (numa_info[src].distance[dst] == 0) {
622 numa_info[src].distance[dst] = NUMA_DISTANCE_MIN;
624 numa_info[src].distance[dst] = numa_info[dst].distance[src];
631 void numa_legacy_auto_assign_ram(MachineClass *mc, NodeInfo *nodes,
632 int nb_nodes, ram_addr_t size)
635 uint64_t usedmem = 0;
637 /* Align each node according to the alignment
638 * requirements of the machine class
641 for (i = 0; i < nb_nodes - 1; i++) {
642 nodes[i].node_mem = (size / nb_nodes) &
643 ~((1 << mc->numa_mem_align_shift) - 1);
644 usedmem += nodes[i].node_mem;
646 nodes[i].node_mem = size - usedmem;
649 void numa_default_auto_assign_ram(MachineClass *mc, NodeInfo *nodes,
650 int nb_nodes, ram_addr_t size)
653 uint64_t usedmem = 0, node_mem;
654 uint64_t granularity = size / nb_nodes;
655 uint64_t propagate = 0;
657 for (i = 0; i < nb_nodes - 1; i++) {
658 node_mem = (granularity + propagate) &
659 ~((1 << mc->numa_mem_align_shift) - 1);
660 propagate = granularity + propagate - node_mem;
661 nodes[i].node_mem = node_mem;
664 nodes[i].node_mem = size - usedmem;
667 static void numa_init_memdev_container(MachineState *ms, MemoryRegion *ram)
672 for (i = 0; i < ms->numa_state->num_nodes; i++) {
673 uint64_t size = ms->numa_state->nodes[i].node_mem;
674 HostMemoryBackend *backend = ms->numa_state->nodes[i].node_memdev;
678 MemoryRegion *seg = machine_consume_memdev(ms, backend);
679 memory_region_add_subregion(ram, addr, seg);
684 void numa_complete_configuration(MachineState *ms)
687 MachineClass *mc = MACHINE_GET_CLASS(ms);
688 NodeInfo *numa_info = ms->numa_state->nodes;
691 * If memory hotplug is enabled (slot > 0) or memory devices are enabled
692 * (ms->maxram_size > ram_size) but without '-numa' options explicitly on
693 * CLI, guests will break.
695 * Windows: won't enable memory hotplug without SRAT table at all
697 * Linux: if QEMU is started with initial memory all below 4Gb
698 * and no SRAT table present, guest kernel will use nommu DMA ops,
699 * which breaks 32bit hw drivers when memory is hotplugged and
700 * guest tries to use it with that drivers.
702 * Enable NUMA implicitly by adding a new NUMA node automatically.
704 * Or if MachineClass::auto_enable_numa is true and no NUMA nodes,
705 * assume there is just one node with whole RAM.
707 if (ms->numa_state->num_nodes == 0 &&
708 ((ms->ram_slots && mc->auto_enable_numa_with_memhp) ||
709 (ms->maxram_size > ms->ram_size && mc->auto_enable_numa_with_memdev) ||
710 mc->auto_enable_numa)) {
711 NumaNodeOptions node = { };
712 parse_numa_node(ms, &node, &error_abort);
713 numa_info[0].node_mem = ram_size;
716 assert(max_numa_nodeid <= MAX_NODES);
718 /* No support for sparse NUMA node IDs yet: */
719 for (i = max_numa_nodeid - 1; i >= 0; i--) {
720 /* Report large node IDs first, to make mistakes easier to spot */
721 if (!numa_info[i].present) {
722 error_report("numa: Node ID missing: %d", i);
727 /* This must be always true if all nodes are present: */
728 assert(ms->numa_state->num_nodes == max_numa_nodeid);
730 if (ms->numa_state->num_nodes > 0) {
733 if (ms->numa_state->num_nodes > MAX_NODES) {
734 ms->numa_state->num_nodes = MAX_NODES;
737 /* If no memory size is given for any node, assume the default case
738 * and distribute the available memory equally across all nodes
740 for (i = 0; i < ms->numa_state->num_nodes; i++) {
741 if (numa_info[i].node_mem != 0) {
745 if (i == ms->numa_state->num_nodes) {
746 assert(mc->numa_auto_assign_ram);
747 mc->numa_auto_assign_ram(mc, numa_info,
748 ms->numa_state->num_nodes, ram_size);
749 if (!qtest_enabled()) {
750 warn_report("Default splitting of RAM between nodes is deprecated,"
751 " Use '-numa node,memdev' to explictly define RAM"
752 " allocation per node");
757 for (i = 0; i < ms->numa_state->num_nodes; i++) {
758 numa_total += numa_info[i].node_mem;
760 if (numa_total != ram_size) {
761 error_report("total memory for NUMA nodes (0x%" PRIx64 ")"
762 " should equal RAM size (0x" RAM_ADDR_FMT ")",
763 numa_total, ram_size);
767 if (!numa_uses_legacy_mem() && mc->default_ram_id) {
768 if (ms->ram_memdev_id) {
769 error_report("'-machine memory-backend' and '-numa memdev'"
770 " properties are mutually exclusive");
773 ms->ram = g_new(MemoryRegion, 1);
774 memory_region_init(ms->ram, OBJECT(ms), mc->default_ram_id,
776 numa_init_memdev_container(ms, ms->ram);
778 /* QEMU needs at least all unique node pair distances to build
779 * the whole NUMA distance table. QEMU treats the distance table
780 * as symmetric by default, i.e. distance A->B == distance B->A.
781 * Thus, QEMU is able to complete the distance table
782 * initialization even though only distance A->B is provided and
783 * distance B->A is not. QEMU knows the distance of a node to
784 * itself is always 10, so A->A distances may be omitted. When
785 * the distances of two nodes of a pair differ, i.e. distance
786 * A->B != distance B->A, then that means the distance table is
787 * asymmetric. In this case, the distances for both directions
788 * of all node pairs are required.
790 if (ms->numa_state->have_numa_distance) {
791 /* Validate enough NUMA distance information was provided. */
792 validate_numa_distance(ms);
794 /* Validation succeeded, now fill in any missing distances. */
795 complete_init_numa_distance(ms);
800 void parse_numa_opts(MachineState *ms)
802 qemu_opts_foreach(qemu_find_opts("numa"), parse_numa, ms, &error_fatal);
805 void numa_cpu_pre_plug(const CPUArchId *slot, DeviceState *dev, Error **errp)
807 int node_id = object_property_get_int(OBJECT(dev), "node-id", &error_abort);
809 if (node_id == CPU_UNSET_NUMA_NODE_ID) {
810 /* due to bug in libvirt, it doesn't pass node-id from props on
811 * device_add as expected, so we have to fix it up here */
812 if (slot->props.has_node_id) {
813 object_property_set_int(OBJECT(dev), slot->props.node_id,
816 } else if (node_id != slot->props.node_id) {
817 error_setg(errp, "invalid node-id, must be %"PRId64,
818 slot->props.node_id);
822 static void numa_stat_memory_devices(NumaNodeMem node_mem[])
824 MemoryDeviceInfoList *info_list = qmp_memory_device_list();
825 MemoryDeviceInfoList *info;
826 PCDIMMDeviceInfo *pcdimm_info;
827 VirtioPMEMDeviceInfo *vpi;
828 VirtioMEMDeviceInfo *vmi;
830 for (info = info_list; info; info = info->next) {
831 MemoryDeviceInfo *value = info->value;
834 switch (value->type) {
835 case MEMORY_DEVICE_INFO_KIND_DIMM:
836 case MEMORY_DEVICE_INFO_KIND_NVDIMM:
837 pcdimm_info = value->type == MEMORY_DEVICE_INFO_KIND_DIMM ?
838 value->u.dimm.data : value->u.nvdimm.data;
839 node_mem[pcdimm_info->node].node_mem += pcdimm_info->size;
840 node_mem[pcdimm_info->node].node_plugged_mem +=
843 case MEMORY_DEVICE_INFO_KIND_VIRTIO_PMEM:
844 vpi = value->u.virtio_pmem.data;
845 /* TODO: once we support numa, assign to right node */
846 node_mem[0].node_mem += vpi->size;
847 node_mem[0].node_plugged_mem += vpi->size;
849 case MEMORY_DEVICE_INFO_KIND_VIRTIO_MEM:
850 vmi = value->u.virtio_mem.data;
851 node_mem[vmi->node].node_mem += vmi->size;
852 node_mem[vmi->node].node_plugged_mem += vmi->size;
855 g_assert_not_reached();
859 qapi_free_MemoryDeviceInfoList(info_list);
862 void query_numa_node_mem(NumaNodeMem node_mem[], MachineState *ms)
866 if (ms->numa_state == NULL || ms->numa_state->num_nodes <= 0) {
870 numa_stat_memory_devices(node_mem);
871 for (i = 0; i < ms->numa_state->num_nodes; i++) {
872 node_mem[i].node_mem += ms->numa_state->nodes[i].node_mem;
876 void ram_block_notifier_add(RAMBlockNotifier *n)
878 QLIST_INSERT_HEAD(&ram_list.ramblock_notifiers, n, next);
881 void ram_block_notifier_remove(RAMBlockNotifier *n)
883 QLIST_REMOVE(n, next);
886 void ram_block_notify_add(void *host, size_t size)
888 RAMBlockNotifier *notifier;
890 QLIST_FOREACH(notifier, &ram_list.ramblock_notifiers, next) {
891 notifier->ram_block_added(notifier, host, size);
895 void ram_block_notify_remove(void *host, size_t size)
897 RAMBlockNotifier *notifier;
899 QLIST_FOREACH(notifier, &ram_list.ramblock_notifiers, next) {
900 notifier->ram_block_removed(notifier, host, size);