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/numa.h"
27 #include "exec/cpu-common.h"
28 #include "exec/ramlist.h"
29 #include "qemu/bitmap.h"
31 #include "qemu/error-report.h"
32 #include "include/exec/cpu-common.h" /* for RAM_ADDR_FMT */
33 #include "qapi-visit.h"
34 #include "qapi/opts-visitor.h"
35 #include "hw/boards.h"
36 #include "sysemu/hostmem.h"
37 #include "qmp-commands.h"
38 #include "hw/mem/pc-dimm.h"
39 #include "qemu/option.h"
40 #include "qemu/config-file.h"
42 QemuOptsList qemu_numa_opts
= {
44 .implied_opt_name
= "type",
45 .head
= QTAILQ_HEAD_INITIALIZER(qemu_numa_opts
.head
),
46 .desc
= { { 0 } } /* validated with OptsVisitor */
49 static int have_memdevs
= -1;
50 static int max_numa_nodeid
; /* Highest specified NUMA node ID, plus one.
51 * For all nodes, nodeid < max_numa_nodeid
54 bool have_numa_distance
;
55 NodeInfo numa_info
[MAX_NODES
];
57 void numa_set_mem_node_id(ram_addr_t addr
, uint64_t size
, uint32_t node
)
59 struct numa_addr_range
*range
;
62 * Memory-less nodes can come here with 0 size in which case,
63 * there is nothing to do.
69 range
= g_malloc0(sizeof(*range
));
70 range
->mem_start
= addr
;
71 range
->mem_end
= addr
+ size
- 1;
72 QLIST_INSERT_HEAD(&numa_info
[node
].addr
, range
, entry
);
75 void numa_unset_mem_node_id(ram_addr_t addr
, uint64_t size
, uint32_t node
)
77 struct numa_addr_range
*range
, *next
;
79 QLIST_FOREACH_SAFE(range
, &numa_info
[node
].addr
, entry
, next
) {
80 if (addr
== range
->mem_start
&& (addr
+ size
- 1) == range
->mem_end
) {
81 QLIST_REMOVE(range
, entry
);
88 static void numa_set_mem_ranges(void)
91 ram_addr_t mem_start
= 0;
94 * Deduce start address of each node and use it to store
95 * the address range info in numa_info address range list
97 for (i
= 0; i
< nb_numa_nodes
; i
++) {
98 numa_set_mem_node_id(mem_start
, numa_info
[i
].node_mem
, i
);
99 mem_start
+= numa_info
[i
].node_mem
;
104 * Check if @addr falls under NUMA @node.
106 static bool numa_addr_belongs_to_node(ram_addr_t addr
, uint32_t node
)
108 struct numa_addr_range
*range
;
110 QLIST_FOREACH(range
, &numa_info
[node
].addr
, entry
) {
111 if (addr
>= range
->mem_start
&& addr
<= range
->mem_end
) {
119 * Given an address, return the index of the NUMA node to which the
120 * address belongs to.
122 uint32_t numa_get_node(ram_addr_t addr
, Error
**errp
)
126 /* For non NUMA configurations, check if the addr falls under node 0 */
127 if (!nb_numa_nodes
) {
128 if (numa_addr_belongs_to_node(addr
, 0)) {
133 for (i
= 0; i
< nb_numa_nodes
; i
++) {
134 if (numa_addr_belongs_to_node(addr
, i
)) {
139 error_setg(errp
, "Address 0x" RAM_ADDR_FMT
" doesn't belong to any "
144 static void parse_numa_node(MachineState
*ms
, NumaNodeOptions
*node
,
145 QemuOpts
*opts
, Error
**errp
)
148 uint16List
*cpus
= NULL
;
149 MachineClass
*mc
= MACHINE_GET_CLASS(ms
);
151 if (node
->has_nodeid
) {
152 nodenr
= node
->nodeid
;
154 nodenr
= nb_numa_nodes
;
157 if (nodenr
>= MAX_NODES
) {
158 error_setg(errp
, "Max number of NUMA nodes reached: %"
163 if (numa_info
[nodenr
].present
) {
164 error_setg(errp
, "Duplicate NUMA nodeid: %" PRIu16
, nodenr
);
168 if (!mc
->cpu_index_to_instance_props
) {
169 error_report("NUMA is not supported by this machine-type");
172 for (cpus
= node
->cpus
; cpus
; cpus
= cpus
->next
) {
173 if (cpus
->value
>= max_cpus
) {
175 "CPU index (%" PRIu16
")"
176 " should be smaller than maxcpus (%d)",
177 cpus
->value
, max_cpus
);
180 bitmap_set(numa_info
[nodenr
].node_cpu
, cpus
->value
, 1);
183 if (node
->has_mem
&& node
->has_memdev
) {
184 error_setg(errp
, "qemu: cannot specify both mem= and memdev=");
188 if (have_memdevs
== -1) {
189 have_memdevs
= node
->has_memdev
;
191 if (node
->has_memdev
!= have_memdevs
) {
192 error_setg(errp
, "qemu: memdev option must be specified for either "
198 uint64_t mem_size
= node
->mem
;
199 const char *mem_str
= qemu_opt_get(opts
, "mem");
200 /* Fix up legacy suffix-less format */
201 if (g_ascii_isdigit(mem_str
[strlen(mem_str
) - 1])) {
204 numa_info
[nodenr
].node_mem
= mem_size
;
206 if (node
->has_memdev
) {
208 o
= object_resolve_path_type(node
->memdev
, TYPE_MEMORY_BACKEND
, NULL
);
210 error_setg(errp
, "memdev=%s is ambiguous", node
->memdev
);
215 numa_info
[nodenr
].node_mem
= object_property_get_int(o
, "size", NULL
);
216 numa_info
[nodenr
].node_memdev
= MEMORY_BACKEND(o
);
218 numa_info
[nodenr
].present
= true;
219 max_numa_nodeid
= MAX(max_numa_nodeid
, nodenr
+ 1);
222 static void parse_numa_distance(NumaDistOptions
*dist
, Error
**errp
)
224 uint16_t src
= dist
->src
;
225 uint16_t dst
= dist
->dst
;
226 uint8_t val
= dist
->val
;
228 if (src
>= MAX_NODES
|| dst
>= MAX_NODES
) {
230 "Invalid node %" PRIu16
231 ", max possible could be %" PRIu16
,
232 MAX(src
, dst
), MAX_NODES
);
236 if (!numa_info
[src
].present
|| !numa_info
[dst
].present
) {
237 error_setg(errp
, "Source/Destination NUMA node is missing. "
238 "Please use '-numa node' option to declare it first.");
242 if (val
< NUMA_DISTANCE_MIN
) {
243 error_setg(errp
, "NUMA distance (%" PRIu8
") is invalid, "
244 "it shouldn't be less than %d.",
245 val
, NUMA_DISTANCE_MIN
);
249 if (src
== dst
&& val
!= NUMA_DISTANCE_MIN
) {
250 error_setg(errp
, "Local distance of node %d should be %d.",
251 src
, NUMA_DISTANCE_MIN
);
255 numa_info
[src
].distance
[dst
] = val
;
256 have_numa_distance
= true;
259 static int parse_numa(void *opaque
, QemuOpts
*opts
, Error
**errp
)
261 NumaOptions
*object
= NULL
;
262 MachineState
*ms
= opaque
;
266 Visitor
*v
= opts_visitor_new(opts
);
267 visit_type_NumaOptions(v
, NULL
, &object
, &err
);
275 switch (object
->type
) {
276 case NUMA_OPTIONS_TYPE_NODE
:
277 parse_numa_node(ms
, &object
->u
.node
, opts
, &err
);
283 case NUMA_OPTIONS_TYPE_DIST
:
284 parse_numa_distance(&object
->u
.dist
, &err
);
294 qapi_free_NumaOptions(object
);
296 error_report_err(err
);
303 static char *enumerate_cpus(unsigned long *cpus
, int max_cpus
)
307 GString
*s
= g_string_new(NULL
);
309 for (cpu
= find_first_bit(cpus
, max_cpus
);
311 cpu
= find_next_bit(cpus
, max_cpus
, cpu
+ 1)) {
312 g_string_append_printf(s
, "%s%d", first
? "" : " ", cpu
);
315 return g_string_free(s
, FALSE
);
318 static void validate_numa_cpus(void)
321 unsigned long *seen_cpus
= bitmap_new(max_cpus
);
323 for (i
= 0; i
< nb_numa_nodes
; i
++) {
324 if (bitmap_intersects(seen_cpus
, numa_info
[i
].node_cpu
, max_cpus
)) {
325 bitmap_and(seen_cpus
, seen_cpus
,
326 numa_info
[i
].node_cpu
, max_cpus
);
327 error_report("CPU(s) present in multiple NUMA nodes: %s",
328 enumerate_cpus(seen_cpus
, max_cpus
));
332 bitmap_or(seen_cpus
, seen_cpus
,
333 numa_info
[i
].node_cpu
, max_cpus
);
336 if (!bitmap_full(seen_cpus
, max_cpus
)) {
338 bitmap_complement(seen_cpus
, seen_cpus
, max_cpus
);
339 msg
= enumerate_cpus(seen_cpus
, max_cpus
);
340 error_report("warning: CPU(s) not present in any NUMA nodes: %s", msg
);
341 error_report("warning: All CPU(s) up to maxcpus should be described "
348 /* If all node pair distances are symmetric, then only distances
349 * in one direction are enough. If there is even one asymmetric
350 * pair, though, then all distances must be provided. The
351 * distance from a node to itself is always NUMA_DISTANCE_MIN,
352 * so providing it is never necessary.
354 static void validate_numa_distance(void)
357 bool is_asymmetrical
= false;
359 for (src
= 0; src
< nb_numa_nodes
; src
++) {
360 for (dst
= src
; dst
< nb_numa_nodes
; dst
++) {
361 if (numa_info
[src
].distance
[dst
] == 0 &&
362 numa_info
[dst
].distance
[src
] == 0) {
364 error_report("The distance between node %d and %d is "
365 "missing, at least one distance value "
366 "between each nodes should be provided.",
372 if (numa_info
[src
].distance
[dst
] != 0 &&
373 numa_info
[dst
].distance
[src
] != 0 &&
374 numa_info
[src
].distance
[dst
] !=
375 numa_info
[dst
].distance
[src
]) {
376 is_asymmetrical
= true;
381 if (is_asymmetrical
) {
382 for (src
= 0; src
< nb_numa_nodes
; src
++) {
383 for (dst
= 0; dst
< nb_numa_nodes
; dst
++) {
384 if (src
!= dst
&& numa_info
[src
].distance
[dst
] == 0) {
385 error_report("At least one asymmetrical pair of "
386 "distances is given, please provide distances "
387 "for both directions of all node pairs.");
395 static void complete_init_numa_distance(void)
399 /* Fixup NUMA distance by symmetric policy because if it is an
400 * asymmetric distance table, it should be a complete table and
401 * there would not be any missing distance except local node, which
402 * is verified by validate_numa_distance above.
404 for (src
= 0; src
< nb_numa_nodes
; src
++) {
405 for (dst
= 0; dst
< nb_numa_nodes
; dst
++) {
406 if (numa_info
[src
].distance
[dst
] == 0) {
408 numa_info
[src
].distance
[dst
] = NUMA_DISTANCE_MIN
;
410 numa_info
[src
].distance
[dst
] = numa_info
[dst
].distance
[src
];
417 void numa_legacy_auto_assign_ram(MachineClass
*mc
, NodeInfo
*nodes
,
418 int nb_nodes
, ram_addr_t size
)
421 uint64_t usedmem
= 0;
423 /* Align each node according to the alignment
424 * requirements of the machine class
427 for (i
= 0; i
< nb_nodes
- 1; i
++) {
428 nodes
[i
].node_mem
= (size
/ nb_nodes
) &
429 ~((1 << mc
->numa_mem_align_shift
) - 1);
430 usedmem
+= nodes
[i
].node_mem
;
432 nodes
[i
].node_mem
= size
- usedmem
;
435 void numa_default_auto_assign_ram(MachineClass
*mc
, NodeInfo
*nodes
,
436 int nb_nodes
, ram_addr_t size
)
439 uint64_t usedmem
= 0, node_mem
;
440 uint64_t granularity
= size
/ nb_nodes
;
441 uint64_t propagate
= 0;
443 for (i
= 0; i
< nb_nodes
- 1; i
++) {
444 node_mem
= (granularity
+ propagate
) &
445 ~((1 << mc
->numa_mem_align_shift
) - 1);
446 propagate
= granularity
+ propagate
- node_mem
;
447 nodes
[i
].node_mem
= node_mem
;
450 nodes
[i
].node_mem
= size
- usedmem
;
453 void parse_numa_opts(MachineState
*ms
)
456 MachineClass
*mc
= MACHINE_GET_CLASS(ms
);
458 for (i
= 0; i
< MAX_NODES
; i
++) {
459 numa_info
[i
].node_cpu
= bitmap_new(max_cpus
);
462 if (qemu_opts_foreach(qemu_find_opts("numa"), parse_numa
, ms
, NULL
)) {
466 assert(max_numa_nodeid
<= MAX_NODES
);
468 /* No support for sparse NUMA node IDs yet: */
469 for (i
= max_numa_nodeid
- 1; i
>= 0; i
--) {
470 /* Report large node IDs first, to make mistakes easier to spot */
471 if (!numa_info
[i
].present
) {
472 error_report("numa: Node ID missing: %d", i
);
477 /* This must be always true if all nodes are present: */
478 assert(nb_numa_nodes
== max_numa_nodeid
);
480 if (nb_numa_nodes
> 0) {
483 if (nb_numa_nodes
> MAX_NODES
) {
484 nb_numa_nodes
= MAX_NODES
;
487 /* If no memory size is given for any node, assume the default case
488 * and distribute the available memory equally across all nodes
490 for (i
= 0; i
< nb_numa_nodes
; i
++) {
491 if (numa_info
[i
].node_mem
!= 0) {
495 if (i
== nb_numa_nodes
) {
496 assert(mc
->numa_auto_assign_ram
);
497 mc
->numa_auto_assign_ram(mc
, numa_info
, nb_numa_nodes
, ram_size
);
501 for (i
= 0; i
< nb_numa_nodes
; i
++) {
502 numa_total
+= numa_info
[i
].node_mem
;
504 if (numa_total
!= ram_size
) {
505 error_report("total memory for NUMA nodes (0x%" PRIx64
")"
506 " should equal RAM size (0x" RAM_ADDR_FMT
")",
507 numa_total
, ram_size
);
511 for (i
= 0; i
< nb_numa_nodes
; i
++) {
512 QLIST_INIT(&numa_info
[i
].addr
);
515 numa_set_mem_ranges();
517 for (i
= 0; i
< nb_numa_nodes
; i
++) {
518 if (!bitmap_empty(numa_info
[i
].node_cpu
, max_cpus
)) {
523 /* assign CPUs to nodes using board provided default mapping */
524 if (!mc
->cpu_index_to_instance_props
) {
525 error_report("default CPUs to NUMA node mapping isn't supported");
528 if (i
== nb_numa_nodes
) {
529 for (i
= 0; i
< max_cpus
; i
++) {
530 CpuInstanceProperties props
;
531 props
= mc
->cpu_index_to_instance_props(ms
, i
);
533 set_bit(i
, numa_info
[props
.node_id
].node_cpu
);
537 validate_numa_cpus();
539 /* QEMU needs at least all unique node pair distances to build
540 * the whole NUMA distance table. QEMU treats the distance table
541 * as symmetric by default, i.e. distance A->B == distance B->A.
542 * Thus, QEMU is able to complete the distance table
543 * initialization even though only distance A->B is provided and
544 * distance B->A is not. QEMU knows the distance of a node to
545 * itself is always 10, so A->A distances may be omitted. When
546 * the distances of two nodes of a pair differ, i.e. distance
547 * A->B != distance B->A, then that means the distance table is
548 * asymmetric. In this case, the distances for both directions
549 * of all node pairs are required.
551 if (have_numa_distance
) {
552 /* Validate enough NUMA distance information was provided. */
553 validate_numa_distance();
555 /* Validation succeeded, now fill in any missing distances. */
556 complete_init_numa_distance();
559 numa_set_mem_node_id(0, ram_size
, 0);
563 void numa_post_machine_init(void)
569 for (i
= 0; i
< nb_numa_nodes
; i
++) {
570 assert(cpu
->cpu_index
< max_cpus
);
571 if (test_bit(cpu
->cpu_index
, numa_info
[i
].node_cpu
)) {
578 static void allocate_system_memory_nonnuma(MemoryRegion
*mr
, Object
*owner
,
585 memory_region_init_ram_from_file(mr
, owner
, name
, ram_size
, false,
588 error_report_err(err
);
593 /* Legacy behavior: if allocation failed, fall back to
594 * regular RAM allocation.
596 memory_region_init_ram(mr
, owner
, name
, ram_size
, &error_fatal
);
599 fprintf(stderr
, "-mem-path not supported on this host\n");
603 memory_region_init_ram(mr
, owner
, name
, ram_size
, &error_fatal
);
605 vmstate_register_ram_global(mr
);
608 void memory_region_allocate_system_memory(MemoryRegion
*mr
, Object
*owner
,
615 if (nb_numa_nodes
== 0 || !have_memdevs
) {
616 allocate_system_memory_nonnuma(mr
, owner
, name
, ram_size
);
620 memory_region_init(mr
, owner
, name
, ram_size
);
621 for (i
= 0; i
< MAX_NODES
; i
++) {
622 uint64_t size
= numa_info
[i
].node_mem
;
623 HostMemoryBackend
*backend
= numa_info
[i
].node_memdev
;
627 MemoryRegion
*seg
= host_memory_backend_get_memory(backend
,
630 if (memory_region_is_mapped(seg
)) {
631 char *path
= object_get_canonical_path_component(OBJECT(backend
));
632 error_report("memory backend %s is used multiple times. Each "
633 "-numa option must use a different memdev value.",
638 host_memory_backend_set_mapped(backend
, true);
639 memory_region_add_subregion(mr
, addr
, seg
);
640 vmstate_register_ram_global(seg
);
645 static void numa_stat_memory_devices(uint64_t node_mem
[])
647 MemoryDeviceInfoList
*info_list
= NULL
;
648 MemoryDeviceInfoList
**prev
= &info_list
;
649 MemoryDeviceInfoList
*info
;
651 qmp_pc_dimm_device_list(qdev_get_machine(), &prev
);
652 for (info
= info_list
; info
; info
= info
->next
) {
653 MemoryDeviceInfo
*value
= info
->value
;
656 switch (value
->type
) {
657 case MEMORY_DEVICE_INFO_KIND_DIMM
:
658 node_mem
[value
->u
.dimm
.data
->node
] += value
->u
.dimm
.data
->size
;
665 qapi_free_MemoryDeviceInfoList(info_list
);
668 void query_numa_node_mem(uint64_t node_mem
[])
672 if (nb_numa_nodes
<= 0) {
676 numa_stat_memory_devices(node_mem
);
677 for (i
= 0; i
< nb_numa_nodes
; i
++) {
678 node_mem
[i
] += numa_info
[i
].node_mem
;
682 static int query_memdev(Object
*obj
, void *opaque
)
684 MemdevList
**list
= opaque
;
685 MemdevList
*m
= NULL
;
687 if (object_dynamic_cast(obj
, TYPE_MEMORY_BACKEND
)) {
688 m
= g_malloc0(sizeof(*m
));
690 m
->value
= g_malloc0(sizeof(*m
->value
));
692 m
->value
->id
= object_property_get_str(obj
, "id", NULL
);
693 m
->value
->has_id
= !!m
->value
->id
;
695 m
->value
->size
= object_property_get_int(obj
, "size",
697 m
->value
->merge
= object_property_get_bool(obj
, "merge",
699 m
->value
->dump
= object_property_get_bool(obj
, "dump",
701 m
->value
->prealloc
= object_property_get_bool(obj
,
704 m
->value
->policy
= object_property_get_enum(obj
,
708 object_property_get_uint16List(obj
, "host-nodes",
709 &m
->value
->host_nodes
,
719 MemdevList
*qmp_query_memdev(Error
**errp
)
721 Object
*obj
= object_get_objects_root();
722 MemdevList
*list
= NULL
;
724 object_child_foreach(obj
, query_memdev
, &list
);
728 int numa_get_node_for_cpu(int idx
)
732 assert(idx
< max_cpus
);
734 for (i
= 0; i
< nb_numa_nodes
; i
++) {
735 if (test_bit(idx
, numa_info
[i
].node_cpu
)) {
742 void ram_block_notifier_add(RAMBlockNotifier
*n
)
744 QLIST_INSERT_HEAD(&ram_list
.ramblock_notifiers
, n
, next
);
747 void ram_block_notifier_remove(RAMBlockNotifier
*n
)
749 QLIST_REMOVE(n
, next
);
752 void ram_block_notify_add(void *host
, size_t size
)
754 RAMBlockNotifier
*notifier
;
756 QLIST_FOREACH(notifier
, &ram_list
.ramblock_notifiers
, next
) {
757 notifier
->ram_block_added(notifier
, host
, size
);
761 void ram_block_notify_remove(void *host
, size_t size
)
763 RAMBlockNotifier
*notifier
;
765 QLIST_FOREACH(notifier
, &ram_list
.ramblock_notifiers
, next
) {
766 notifier
->ram_block_removed(notifier
, host
, size
);