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 "sysemu/hostmem.h"
27 #include "sysemu/numa.h"
28 #include "sysemu/sysemu.h"
29 #include "exec/cpu-common.h"
30 #include "exec/ramlist.h"
31 #include "qemu/bitmap.h"
32 #include "qemu/error-report.h"
33 #include "qapi/error.h"
34 #include "qapi/opts-visitor.h"
35 #include "qapi/qapi-visit-machine.h"
36 #include "sysemu/qtest.h"
37 #include "hw/core/cpu.h"
38 #include "hw/mem/pc-dimm.h"
39 #include "migration/vmstate.h"
40 #include "hw/boards.h"
41 #include "hw/mem/memory-device.h"
42 #include "qemu/option.h"
43 #include "qemu/config-file.h"
44 #include "qemu/cutils.h"
46 QemuOptsList qemu_numa_opts
= {
48 .implied_opt_name
= "type",
49 .head
= QTAILQ_HEAD_INITIALIZER(qemu_numa_opts
.head
),
50 .desc
= { { 0 } } /* validated with OptsVisitor */
53 static int have_memdevs
;
55 static int max_numa_nodeid
; /* Highest specified NUMA node ID, plus one.
56 * For all nodes, nodeid < max_numa_nodeid
59 static void parse_numa_node(MachineState
*ms
, NumaNodeOptions
*node
,
64 uint16List
*cpus
= NULL
;
65 MachineClass
*mc
= MACHINE_GET_CLASS(ms
);
66 unsigned int max_cpus
= ms
->smp
.max_cpus
;
67 NodeInfo
*numa_info
= ms
->numa_state
->nodes
;
69 if (node
->has_nodeid
) {
70 nodenr
= node
->nodeid
;
72 nodenr
= ms
->numa_state
->num_nodes
;
75 if (nodenr
>= MAX_NODES
) {
76 error_setg(errp
, "Max number of NUMA nodes reached: %"
81 if (numa_info
[nodenr
].present
) {
82 error_setg(errp
, "Duplicate NUMA nodeid: %" PRIu16
, nodenr
);
86 if (!mc
->cpu_index_to_instance_props
|| !mc
->get_default_cpu_node_id
) {
87 error_setg(errp
, "NUMA is not supported by this machine-type");
90 for (cpus
= node
->cpus
; cpus
; cpus
= cpus
->next
) {
91 CpuInstanceProperties props
;
92 if (cpus
->value
>= max_cpus
) {
94 "CPU index (%" PRIu16
")"
95 " should be smaller than maxcpus (%d)",
96 cpus
->value
, max_cpus
);
99 props
= mc
->cpu_index_to_instance_props(ms
, cpus
->value
);
100 props
.node_id
= nodenr
;
101 props
.has_node_id
= true;
102 machine_set_cpu_numa_node(ms
, &props
, &err
);
104 error_propagate(errp
, err
);
109 have_memdevs
= have_memdevs
? : node
->has_memdev
;
110 have_mem
= have_mem
? : node
->has_mem
;
111 if ((node
->has_mem
&& have_memdevs
) || (node
->has_memdev
&& have_mem
)) {
112 error_setg(errp
, "numa configuration should use either mem= or memdev=,"
113 "mixing both is not allowed");
118 numa_info
[nodenr
].node_mem
= node
->mem
;
119 if (!qtest_enabled()) {
120 warn_report("Parameter -numa node,mem is deprecated,"
121 " use -numa node,memdev instead");
124 if (node
->has_memdev
) {
126 o
= object_resolve_path_type(node
->memdev
, TYPE_MEMORY_BACKEND
, NULL
);
128 error_setg(errp
, "memdev=%s is ambiguous", node
->memdev
);
133 numa_info
[nodenr
].node_mem
= object_property_get_uint(o
, "size", NULL
);
134 numa_info
[nodenr
].node_memdev
= MEMORY_BACKEND(o
);
136 numa_info
[nodenr
].present
= true;
137 max_numa_nodeid
= MAX(max_numa_nodeid
, nodenr
+ 1);
138 ms
->numa_state
->num_nodes
++;
142 void parse_numa_distance(MachineState
*ms
, NumaDistOptions
*dist
, Error
**errp
)
144 uint16_t src
= dist
->src
;
145 uint16_t dst
= dist
->dst
;
146 uint8_t val
= dist
->val
;
147 NodeInfo
*numa_info
= ms
->numa_state
->nodes
;
149 if (src
>= MAX_NODES
|| dst
>= MAX_NODES
) {
150 error_setg(errp
, "Parameter '%s' expects an integer between 0 and %d",
151 src
>= MAX_NODES
? "src" : "dst", MAX_NODES
- 1);
155 if (!numa_info
[src
].present
|| !numa_info
[dst
].present
) {
156 error_setg(errp
, "Source/Destination NUMA node is missing. "
157 "Please use '-numa node' option to declare it first.");
161 if (val
< NUMA_DISTANCE_MIN
) {
162 error_setg(errp
, "NUMA distance (%" PRIu8
") is invalid, "
163 "it shouldn't be less than %d.",
164 val
, NUMA_DISTANCE_MIN
);
168 if (src
== dst
&& val
!= NUMA_DISTANCE_MIN
) {
169 error_setg(errp
, "Local distance of node %d should be %d.",
170 src
, NUMA_DISTANCE_MIN
);
174 numa_info
[src
].distance
[dst
] = val
;
175 ms
->numa_state
->have_numa_distance
= true;
178 void set_numa_options(MachineState
*ms
, NumaOptions
*object
, Error
**errp
)
181 MachineClass
*mc
= MACHINE_GET_CLASS(ms
);
183 if (!mc
->numa_mem_supported
) {
184 error_setg(errp
, "NUMA is not supported by this machine-type");
188 switch (object
->type
) {
189 case NUMA_OPTIONS_TYPE_NODE
:
190 parse_numa_node(ms
, &object
->u
.node
, &err
);
195 case NUMA_OPTIONS_TYPE_DIST
:
196 parse_numa_distance(ms
, &object
->u
.dist
, &err
);
201 case NUMA_OPTIONS_TYPE_CPU
:
202 if (!object
->u
.cpu
.has_node_id
) {
203 error_setg(&err
, "Missing mandatory node-id property");
206 if (!ms
->numa_state
->nodes
[object
->u
.cpu
.node_id
].present
) {
207 error_setg(&err
, "Invalid node-id=%" PRId64
", NUMA node must be "
208 "defined with -numa node,nodeid=ID before it's used with "
209 "-numa cpu,node-id=ID", object
->u
.cpu
.node_id
);
213 machine_set_cpu_numa_node(ms
, qapi_NumaCpuOptions_base(&object
->u
.cpu
),
221 error_propagate(errp
, err
);
224 static int parse_numa(void *opaque
, QemuOpts
*opts
, Error
**errp
)
226 NumaOptions
*object
= NULL
;
227 MachineState
*ms
= MACHINE(opaque
);
229 Visitor
*v
= opts_visitor_new(opts
);
231 visit_type_NumaOptions(v
, NULL
, &object
, &err
);
237 /* Fix up legacy suffix-less format */
238 if ((object
->type
== NUMA_OPTIONS_TYPE_NODE
) && object
->u
.node
.has_mem
) {
239 const char *mem_str
= qemu_opt_get(opts
, "mem");
240 qemu_strtosz_MiB(mem_str
, NULL
, &object
->u
.node
.mem
);
243 set_numa_options(ms
, object
, &err
);
246 qapi_free_NumaOptions(object
);
248 error_propagate(errp
, err
);
255 /* If all node pair distances are symmetric, then only distances
256 * in one direction are enough. If there is even one asymmetric
257 * pair, though, then all distances must be provided. The
258 * distance from a node to itself is always NUMA_DISTANCE_MIN,
259 * so providing it is never necessary.
261 static void validate_numa_distance(MachineState
*ms
)
264 bool is_asymmetrical
= false;
265 int nb_numa_nodes
= ms
->numa_state
->num_nodes
;
266 NodeInfo
*numa_info
= ms
->numa_state
->nodes
;
268 for (src
= 0; src
< nb_numa_nodes
; src
++) {
269 for (dst
= src
; dst
< nb_numa_nodes
; dst
++) {
270 if (numa_info
[src
].distance
[dst
] == 0 &&
271 numa_info
[dst
].distance
[src
] == 0) {
273 error_report("The distance between node %d and %d is "
274 "missing, at least one distance value "
275 "between each nodes should be provided.",
281 if (numa_info
[src
].distance
[dst
] != 0 &&
282 numa_info
[dst
].distance
[src
] != 0 &&
283 numa_info
[src
].distance
[dst
] !=
284 numa_info
[dst
].distance
[src
]) {
285 is_asymmetrical
= true;
290 if (is_asymmetrical
) {
291 for (src
= 0; src
< nb_numa_nodes
; src
++) {
292 for (dst
= 0; dst
< nb_numa_nodes
; dst
++) {
293 if (src
!= dst
&& numa_info
[src
].distance
[dst
] == 0) {
294 error_report("At least one asymmetrical pair of "
295 "distances is given, please provide distances "
296 "for both directions of all node pairs.");
304 static void complete_init_numa_distance(MachineState
*ms
)
307 NodeInfo
*numa_info
= ms
->numa_state
->nodes
;
309 /* Fixup NUMA distance by symmetric policy because if it is an
310 * asymmetric distance table, it should be a complete table and
311 * there would not be any missing distance except local node, which
312 * is verified by validate_numa_distance above.
314 for (src
= 0; src
< ms
->numa_state
->num_nodes
; src
++) {
315 for (dst
= 0; dst
< ms
->numa_state
->num_nodes
; dst
++) {
316 if (numa_info
[src
].distance
[dst
] == 0) {
318 numa_info
[src
].distance
[dst
] = NUMA_DISTANCE_MIN
;
320 numa_info
[src
].distance
[dst
] = numa_info
[dst
].distance
[src
];
327 void numa_legacy_auto_assign_ram(MachineClass
*mc
, NodeInfo
*nodes
,
328 int nb_nodes
, ram_addr_t size
)
331 uint64_t usedmem
= 0;
333 /* Align each node according to the alignment
334 * requirements of the machine class
337 for (i
= 0; i
< nb_nodes
- 1; i
++) {
338 nodes
[i
].node_mem
= (size
/ nb_nodes
) &
339 ~((1 << mc
->numa_mem_align_shift
) - 1);
340 usedmem
+= nodes
[i
].node_mem
;
342 nodes
[i
].node_mem
= size
- usedmem
;
345 void numa_default_auto_assign_ram(MachineClass
*mc
, NodeInfo
*nodes
,
346 int nb_nodes
, ram_addr_t size
)
349 uint64_t usedmem
= 0, node_mem
;
350 uint64_t granularity
= size
/ nb_nodes
;
351 uint64_t propagate
= 0;
353 for (i
= 0; i
< nb_nodes
- 1; i
++) {
354 node_mem
= (granularity
+ propagate
) &
355 ~((1 << mc
->numa_mem_align_shift
) - 1);
356 propagate
= granularity
+ propagate
- node_mem
;
357 nodes
[i
].node_mem
= node_mem
;
360 nodes
[i
].node_mem
= size
- usedmem
;
363 void numa_complete_configuration(MachineState
*ms
)
366 MachineClass
*mc
= MACHINE_GET_CLASS(ms
);
367 NodeInfo
*numa_info
= ms
->numa_state
->nodes
;
370 * If memory hotplug is enabled (slots > 0) but without '-numa'
371 * options explicitly on CLI, guestes will break.
373 * Windows: won't enable memory hotplug without SRAT table at all
375 * Linux: if QEMU is started with initial memory all below 4Gb
376 * and no SRAT table present, guest kernel will use nommu DMA ops,
377 * which breaks 32bit hw drivers when memory is hotplugged and
378 * guest tries to use it with that drivers.
380 * Enable NUMA implicitly by adding a new NUMA node automatically.
382 * Or if MachineClass::auto_enable_numa is true and no NUMA nodes,
383 * assume there is just one node with whole RAM.
385 if (ms
->numa_state
->num_nodes
== 0 &&
386 ((ms
->ram_slots
> 0 &&
387 mc
->auto_enable_numa_with_memhp
) ||
388 mc
->auto_enable_numa
)) {
389 NumaNodeOptions node
= { };
390 parse_numa_node(ms
, &node
, &error_abort
);
391 numa_info
[0].node_mem
= ram_size
;
394 assert(max_numa_nodeid
<= MAX_NODES
);
396 /* No support for sparse NUMA node IDs yet: */
397 for (i
= max_numa_nodeid
- 1; i
>= 0; i
--) {
398 /* Report large node IDs first, to make mistakes easier to spot */
399 if (!numa_info
[i
].present
) {
400 error_report("numa: Node ID missing: %d", i
);
405 /* This must be always true if all nodes are present: */
406 assert(ms
->numa_state
->num_nodes
== max_numa_nodeid
);
408 if (ms
->numa_state
->num_nodes
> 0) {
411 if (ms
->numa_state
->num_nodes
> MAX_NODES
) {
412 ms
->numa_state
->num_nodes
= MAX_NODES
;
415 /* If no memory size is given for any node, assume the default case
416 * and distribute the available memory equally across all nodes
418 for (i
= 0; i
< ms
->numa_state
->num_nodes
; i
++) {
419 if (numa_info
[i
].node_mem
!= 0) {
423 if (i
== ms
->numa_state
->num_nodes
) {
424 assert(mc
->numa_auto_assign_ram
);
425 mc
->numa_auto_assign_ram(mc
, numa_info
,
426 ms
->numa_state
->num_nodes
, ram_size
);
427 if (!qtest_enabled()) {
428 warn_report("Default splitting of RAM between nodes is deprecated,"
429 " Use '-numa node,memdev' to explictly define RAM"
430 " allocation per node");
435 for (i
= 0; i
< ms
->numa_state
->num_nodes
; i
++) {
436 numa_total
+= numa_info
[i
].node_mem
;
438 if (numa_total
!= ram_size
) {
439 error_report("total memory for NUMA nodes (0x%" PRIx64
")"
440 " should equal RAM size (0x" RAM_ADDR_FMT
")",
441 numa_total
, ram_size
);
445 /* QEMU needs at least all unique node pair distances to build
446 * the whole NUMA distance table. QEMU treats the distance table
447 * as symmetric by default, i.e. distance A->B == distance B->A.
448 * Thus, QEMU is able to complete the distance table
449 * initialization even though only distance A->B is provided and
450 * distance B->A is not. QEMU knows the distance of a node to
451 * itself is always 10, so A->A distances may be omitted. When
452 * the distances of two nodes of a pair differ, i.e. distance
453 * A->B != distance B->A, then that means the distance table is
454 * asymmetric. In this case, the distances for both directions
455 * of all node pairs are required.
457 if (ms
->numa_state
->have_numa_distance
) {
458 /* Validate enough NUMA distance information was provided. */
459 validate_numa_distance(ms
);
461 /* Validation succeeded, now fill in any missing distances. */
462 complete_init_numa_distance(ms
);
467 void parse_numa_opts(MachineState
*ms
)
469 qemu_opts_foreach(qemu_find_opts("numa"), parse_numa
, ms
, &error_fatal
);
472 void numa_cpu_pre_plug(const CPUArchId
*slot
, DeviceState
*dev
, Error
**errp
)
474 int node_id
= object_property_get_int(OBJECT(dev
), "node-id", &error_abort
);
476 if (node_id
== CPU_UNSET_NUMA_NODE_ID
) {
477 /* due to bug in libvirt, it doesn't pass node-id from props on
478 * device_add as expected, so we have to fix it up here */
479 if (slot
->props
.has_node_id
) {
480 object_property_set_int(OBJECT(dev
), slot
->props
.node_id
,
483 } else if (node_id
!= slot
->props
.node_id
) {
484 error_setg(errp
, "invalid node-id, must be %"PRId64
,
485 slot
->props
.node_id
);
489 static void allocate_system_memory_nonnuma(MemoryRegion
*mr
, Object
*owner
,
496 memory_region_init_ram_from_file(mr
, owner
, name
, ram_size
, 0, 0,
499 error_report_err(err
);
503 warn_report("falling back to regular RAM allocation");
504 error_printf("This is deprecated. Make sure that -mem-path "
505 " specified path has sufficient resources to allocate"
506 " -m specified RAM amount\n");
507 /* Legacy behavior: if allocation failed, fall back to
508 * regular RAM allocation.
511 memory_region_init_ram_nomigrate(mr
, owner
, name
, ram_size
, &error_fatal
);
514 fprintf(stderr
, "-mem-path not supported on this host\n");
518 memory_region_init_ram_nomigrate(mr
, owner
, name
, ram_size
, &error_fatal
);
520 vmstate_register_ram_global(mr
);
523 void memory_region_allocate_system_memory(MemoryRegion
*mr
, Object
*owner
,
529 MachineState
*ms
= MACHINE(qdev_get_machine());
531 if (ms
->numa_state
== NULL
||
532 ms
->numa_state
->num_nodes
== 0 || !have_memdevs
) {
533 allocate_system_memory_nonnuma(mr
, owner
, name
, ram_size
);
537 memory_region_init(mr
, owner
, name
, ram_size
);
538 for (i
= 0; i
< ms
->numa_state
->num_nodes
; i
++) {
539 uint64_t size
= ms
->numa_state
->nodes
[i
].node_mem
;
540 HostMemoryBackend
*backend
= ms
->numa_state
->nodes
[i
].node_memdev
;
544 MemoryRegion
*seg
= host_memory_backend_get_memory(backend
);
546 if (memory_region_is_mapped(seg
)) {
547 char *path
= object_get_canonical_path_component(OBJECT(backend
));
548 error_report("memory backend %s is used multiple times. Each "
549 "-numa option must use a different memdev value.",
555 host_memory_backend_set_mapped(backend
, true);
556 memory_region_add_subregion(mr
, addr
, seg
);
557 vmstate_register_ram_global(seg
);
562 static void numa_stat_memory_devices(NumaNodeMem node_mem
[])
564 MemoryDeviceInfoList
*info_list
= qmp_memory_device_list();
565 MemoryDeviceInfoList
*info
;
566 PCDIMMDeviceInfo
*pcdimm_info
;
567 VirtioPMEMDeviceInfo
*vpi
;
569 for (info
= info_list
; info
; info
= info
->next
) {
570 MemoryDeviceInfo
*value
= info
->value
;
573 switch (value
->type
) {
574 case MEMORY_DEVICE_INFO_KIND_DIMM
:
575 case MEMORY_DEVICE_INFO_KIND_NVDIMM
:
576 pcdimm_info
= value
->type
== MEMORY_DEVICE_INFO_KIND_DIMM
?
577 value
->u
.dimm
.data
: value
->u
.nvdimm
.data
;
578 node_mem
[pcdimm_info
->node
].node_mem
+= pcdimm_info
->size
;
579 node_mem
[pcdimm_info
->node
].node_plugged_mem
+=
582 case MEMORY_DEVICE_INFO_KIND_VIRTIO_PMEM
:
583 vpi
= value
->u
.virtio_pmem
.data
;
584 /* TODO: once we support numa, assign to right node */
585 node_mem
[0].node_mem
+= vpi
->size
;
586 node_mem
[0].node_plugged_mem
+= vpi
->size
;
589 g_assert_not_reached();
593 qapi_free_MemoryDeviceInfoList(info_list
);
596 void query_numa_node_mem(NumaNodeMem node_mem
[], MachineState
*ms
)
600 if (ms
->numa_state
== NULL
|| ms
->numa_state
->num_nodes
<= 0) {
604 numa_stat_memory_devices(node_mem
);
605 for (i
= 0; i
< ms
->numa_state
->num_nodes
; i
++) {
606 node_mem
[i
].node_mem
+= ms
->numa_state
->nodes
[i
].node_mem
;
610 void ram_block_notifier_add(RAMBlockNotifier
*n
)
612 QLIST_INSERT_HEAD(&ram_list
.ramblock_notifiers
, n
, next
);
615 void ram_block_notifier_remove(RAMBlockNotifier
*n
)
617 QLIST_REMOVE(n
, next
);
620 void ram_block_notify_add(void *host
, size_t size
)
622 RAMBlockNotifier
*notifier
;
624 QLIST_FOREACH(notifier
, &ram_list
.ramblock_notifiers
, next
) {
625 notifier
->ram_block_added(notifier
, host
, size
);
629 void ram_block_notify_remove(void *host
, size_t size
)
631 RAMBlockNotifier
*notifier
;
633 QLIST_FOREACH(notifier
, &ram_list
.ramblock_notifiers
, next
) {
634 notifier
->ram_block_removed(notifier
, host
, size
);