1/*
2 * NUMA parameter parsing routines
3 *
4 * Copyright (c) 2014 Fujitsu Ltd.
5 *
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:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
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
22 * THE SOFTWARE.
23 */
24
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"
45
46QemuOptsList qemu_numa_opts = {
47 .name = "numa",
48 .implied_opt_name = "type",
49 .head = QTAILQ_HEAD_INITIALIZER(qemu_numa_opts.head),
50 .desc = { { 0 } } /* validated with OptsVisitor */
51};
52
53static int have_memdevs;
54static int have_mem;
55static int max_numa_nodeid; /* Highest specified NUMA node ID, plus one.
56 * For all nodes, nodeid < max_numa_nodeid
57 */
58
59static void parse_numa_node(MachineState *ms, NumaNodeOptions *node,
60 Error **errp)
61{
62 Error *err = NULL;
63 uint16_t nodenr;
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;
68
69 if (node->has_nodeid) {
70 nodenr = node->nodeid;
71 } else {
72 nodenr = ms->numa_state->num_nodes;
73 }
74
75 if (nodenr >= MAX_NODES) {
76 error_setg(errp, "Max number of NUMA nodes reached: %"
77 PRIu16 "", nodenr);
78 return;
79 }
80
81 if (numa_info[nodenr].present) {
82 error_setg(errp, "Duplicate NUMA nodeid: %" PRIu16, nodenr);
83 return;
84 }
85
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");
88 return;
89 }
90 for (cpus = node->cpus; cpus; cpus = cpus->next) {
91 CpuInstanceProperties props;
92 if (cpus->value >= max_cpus) {
93 error_setg(errp,
94 "CPU index (%" PRIu16 ")"
95 " should be smaller than maxcpus (%d)",
96 cpus->value, max_cpus);
97 return;
98 }
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);
103 if (err) {
104 error_propagate(errp, err);
105 return;
106 }
107 }
108
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");
114 return;
115 }
116
117 if (node->has_mem) {
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");
122 }
123 }
124 if (node->has_memdev) {
125 Object *o;
126 o = object_resolve_path_type(node->memdev, TYPE_MEMORY_BACKEND, NULL);
127 if (!o) {
128 error_setg(errp, "memdev=%s is ambiguous", node->memdev);
129 return;
130 }
131
132 object_ref(o);
133 numa_info[nodenr].node_mem = object_property_get_uint(o, "size", NULL);
134 numa_info[nodenr].node_memdev = MEMORY_BACKEND(o);
135 }
136 numa_info[nodenr].present = true;
137 max_numa_nodeid = MAX(max_numa_nodeid, nodenr + 1);
138 ms->numa_state->num_nodes++;
139}
140
141static
142void parse_numa_distance(MachineState *ms, NumaDistOptions *dist, Error **errp)
143{
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;
148
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);
152 return;
153 }
154
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.");
158 return;
159 }
160
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);
165 return;
166 }
167
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);
171 return;
172 }
173
174 numa_info[src].distance[dst] = val;
175 ms->numa_state->have_numa_distance = true;
176}
177
178void set_numa_options(MachineState *ms, NumaOptions *object, Error **errp)
179{
180 Error *err = NULL;
181 MachineClass *mc = MACHINE_GET_CLASS(ms);
182
183 if (!mc->numa_mem_supported) {
184 error_setg(errp, "NUMA is not supported by this machine-type");
185 goto end;
186 }
187
188 switch (object->type) {
189 case NUMA_OPTIONS_TYPE_NODE:
190 parse_numa_node(ms, &object->u.node, &err);
191 if (err) {
192 goto end;
193 }
194 break;
195 case NUMA_OPTIONS_TYPE_DIST:
196 parse_numa_distance(ms, &object->u.dist, &err);
197 if (err) {
198 goto end;
199 }
200 break;
201 case NUMA_OPTIONS_TYPE_CPU:
202 if (!object->u.cpu.has_node_id) {
203 error_setg(&err, "Missing mandatory node-id property");
204 goto end;
205 }
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);
210 goto end;
211 }
212
213 machine_set_cpu_numa_node(ms, qapi_NumaCpuOptions_base(&object->u.cpu),
214 &err);
215 break;
216 default:
217 abort();
218 }
219
220end:
221 error_propagate(errp, err);
222}
223
224static int parse_numa(void *opaque, QemuOpts *opts, Error **errp)
225{
226 NumaOptions *object = NULL;
227 MachineState *ms = MACHINE(opaque);
228 Error *err = NULL;
229 Visitor *v = opts_visitor_new(opts);
230
231 visit_type_NumaOptions(v, NULL, &object, &err);
232 visit_free(v);
233 if (err) {
234 goto end;
235 }
236
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);
241 }
242
243 set_numa_options(ms, object, &err);
244
245end:
246 qapi_free_NumaOptions(object);
247 if (err) {
248 error_propagate(errp, err);
249 return -1;
250 }
251
252 return 0;
253}
254
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.
260 */
261static void validate_numa_distance(MachineState *ms)
262{
263 int src, dst;
264 bool is_asymmetrical = false;
265 int nb_numa_nodes = ms->numa_state->num_nodes;
266 NodeInfo *numa_info = ms->numa_state->nodes;
267
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) {
272 if (src != dst) {
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.",
276 src, dst);
277 exit(EXIT_FAILURE);
278 }
279 }
280
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;
286 }
287 }
288 }
289
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.");
297 exit(EXIT_FAILURE);
298 }
299 }
300 }
301 }
302}
303
304static void complete_init_numa_distance(MachineState *ms)
305{
306 int src, dst;
307 NodeInfo *numa_info = ms->numa_state->nodes;
308
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.
313 */
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) {
317 if (src == dst) {
318 numa_info[src].distance[dst] = NUMA_DISTANCE_MIN;
319 } else {
320 numa_info[src].distance[dst] = numa_info[dst].distance[src];
321 }
322 }
323 }
324 }
325}
326
327void numa_legacy_auto_assign_ram(MachineClass *mc, NodeInfo *nodes,
328 int nb_nodes, ram_addr_t size)
329{
330 int i;
331 uint64_t usedmem = 0;
332
333 /* Align each node according to the alignment
334 * requirements of the machine class
335 */
336
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;
341 }
342 nodes[i].node_mem = size - usedmem;
343}
344
345void numa_default_auto_assign_ram(MachineClass *mc, NodeInfo *nodes,
346 int nb_nodes, ram_addr_t size)
347{
348 int i;
349 uint64_t usedmem = 0, node_mem;
350 uint64_t granularity = size / nb_nodes;
351 uint64_t propagate = 0;
352
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;
358 usedmem += node_mem;
359 }
360 nodes[i].node_mem = size - usedmem;
361}
362
363void numa_complete_configuration(MachineState *ms)
364{
365 int i;
366 MachineClass *mc = MACHINE_GET_CLASS(ms);
367 NodeInfo *numa_info = ms->numa_state->nodes;
368
369 /*
370 * If memory hotplug is enabled (slots > 0) but without '-numa'
371 * options explicitly on CLI, guestes will break.
372 *
373 * Windows: won't enable memory hotplug without SRAT table at all
374 *
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.
379 *
380 * Enable NUMA implicitly by adding a new NUMA node automatically.
381 */
382 if (ms->ram_slots > 0 && ms->numa_state->num_nodes == 0 &&
383 mc->auto_enable_numa_with_memhp) {
384 NumaNodeOptions node = { };
385 parse_numa_node(ms, &node, &error_abort);
386 }
387
388 assert(max_numa_nodeid <= MAX_NODES);
389
390 /* No support for sparse NUMA node IDs yet: */
391 for (i = max_numa_nodeid - 1; i >= 0; i--) {
392 /* Report large node IDs first, to make mistakes easier to spot */
393 if (!numa_info[i].present) {
394 error_report("numa: Node ID missing: %d", i);
395 exit(1);
396 }
397 }
398
399 /* This must be always true if all nodes are present: */
400 assert(ms->numa_state->num_nodes == max_numa_nodeid);
401
402 if (ms->numa_state->num_nodes > 0) {
403 uint64_t numa_total;
404
405 if (ms->numa_state->num_nodes > MAX_NODES) {
406 ms->numa_state->num_nodes = MAX_NODES;
407 }
408
409 /* If no memory size is given for any node, assume the default case
410 * and distribute the available memory equally across all nodes
411 */
412 for (i = 0; i < ms->numa_state->num_nodes; i++) {
413 if (numa_info[i].node_mem != 0) {
414 break;
415 }
416 }
417 if (i == ms->numa_state->num_nodes) {
418 assert(mc->numa_auto_assign_ram);
419 mc->numa_auto_assign_ram(mc, numa_info,
420 ms->numa_state->num_nodes, ram_size);
421 if (!qtest_enabled()) {
422 warn_report("Default splitting of RAM between nodes is deprecated,"
423 " Use '-numa node,memdev' to explictly define RAM"
424 " allocation per node");
425 }
426 }
427
428 numa_total = 0;
429 for (i = 0; i < ms->numa_state->num_nodes; i++) {
430 numa_total += numa_info[i].node_mem;
431 }
432 if (numa_total != ram_size) {
433 error_report("total memory for NUMA nodes (0x%" PRIx64 ")"
434 " should equal RAM size (0x" RAM_ADDR_FMT ")",
435 numa_total, ram_size);
436 exit(1);
437 }
438
439 /* QEMU needs at least all unique node pair distances to build
440 * the whole NUMA distance table. QEMU treats the distance table
441 * as symmetric by default, i.e. distance A->B == distance B->A.
442 * Thus, QEMU is able to complete the distance table
443 * initialization even though only distance A->B is provided and
444 * distance B->A is not. QEMU knows the distance of a node to
445 * itself is always 10, so A->A distances may be omitted. When
446 * the distances of two nodes of a pair differ, i.e. distance
447 * A->B != distance B->A, then that means the distance table is
448 * asymmetric. In this case, the distances for both directions
449 * of all node pairs are required.
450 */
451 if (ms->numa_state->have_numa_distance) {
452 /* Validate enough NUMA distance information was provided. */
453 validate_numa_distance(ms);
454
455 /* Validation succeeded, now fill in any missing distances. */
456 complete_init_numa_distance(ms);
457 }
458 }
459}
460
461void parse_numa_opts(MachineState *ms)
462{
463 qemu_opts_foreach(qemu_find_opts("numa"), parse_numa, ms, &error_fatal);
464}
465
466void numa_cpu_pre_plug(const CPUArchId *slot, DeviceState *dev, Error **errp)
467{
468 int node_id = object_property_get_int(OBJECT(dev), "node-id", &error_abort);
469
470 if (node_id == CPU_UNSET_NUMA_NODE_ID) {
471 /* due to bug in libvirt, it doesn't pass node-id from props on
472 * device_add as expected, so we have to fix it up here */
473 if (slot->props.has_node_id) {
474 object_property_set_int(OBJECT(dev), slot->props.node_id,
475 "node-id", errp);
476 }
477 } else if (node_id != slot->props.node_id) {
478 error_setg(errp, "invalid node-id, must be %"PRId64,
479 slot->props.node_id);
480 }
481}
482
483static void allocate_system_memory_nonnuma(MemoryRegion *mr, Object *owner,
484 const char *name,
485 uint64_t ram_size)
486{
487 if (mem_path) {
488#ifdef __linux__
489 Error *err = NULL;
490 memory_region_init_ram_from_file(mr, owner, name, ram_size, 0, 0,
491 mem_path, &err);
492 if (err) {
493 error_report_err(err);
494 if (mem_prealloc) {
495 exit(1);
496 }
497 warn_report("falling back to regular RAM allocation");
498 error_printf("This is deprecated. Make sure that -mem-path "
499 " specified path has sufficient resources to allocate"
500 " -m specified RAM amount");
501 /* Legacy behavior: if allocation failed, fall back to
502 * regular RAM allocation.
503 */
504 mem_path = NULL;
505 memory_region_init_ram_nomigrate(mr, owner, name, ram_size, &error_fatal);
506 }
507#else
508 fprintf(stderr, "-mem-path not supported on this host\n");
509 exit(1);
510#endif
511 } else {
512 memory_region_init_ram_nomigrate(mr, owner, name, ram_size, &error_fatal);
513 }
514 vmstate_register_ram_global(mr);
515}
516
517void memory_region_allocate_system_memory(MemoryRegion *mr, Object *owner,
518 const char *name,
519 uint64_t ram_size)
520{
521 uint64_t addr = 0;
522 int i;
523 MachineState *ms = MACHINE(qdev_get_machine());
524
525 if (ms->numa_state == NULL ||
526 ms->numa_state->num_nodes == 0 || !have_memdevs) {
527 allocate_system_memory_nonnuma(mr, owner, name, ram_size);
528 return;
529 }
530
531 memory_region_init(mr, owner, name, ram_size);
532 for (i = 0; i < ms->numa_state->num_nodes; i++) {
533 uint64_t size = ms->numa_state->nodes[i].node_mem;
534 HostMemoryBackend *backend = ms->numa_state->nodes[i].node_memdev;
535 if (!backend) {
536 continue;
537 }
538 MemoryRegion *seg = host_memory_backend_get_memory(backend);
539
540 if (memory_region_is_mapped(seg)) {
541 char *path = object_get_canonical_path_component(OBJECT(backend));
542 error_report("memory backend %s is used multiple times. Each "
543 "-numa option must use a different memdev value.",
544 path);
545 g_free(path);
546 exit(1);
547 }
548
549 host_memory_backend_set_mapped(backend, true);
550 memory_region_add_subregion(mr, addr, seg);
551 vmstate_register_ram_global(seg);
552 addr += size;
553 }
554}
555
556static void numa_stat_memory_devices(NumaNodeMem node_mem[])
557{
558 MemoryDeviceInfoList *info_list = qmp_memory_device_list();
559 MemoryDeviceInfoList *info;
560 PCDIMMDeviceInfo *pcdimm_info;
561 VirtioPMEMDeviceInfo *vpi;
562
563 for (info = info_list; info; info = info->next) {
564 MemoryDeviceInfo *value = info->value;
565
566 if (value) {
567 switch (value->type) {
568 case MEMORY_DEVICE_INFO_KIND_DIMM:
569 case MEMORY_DEVICE_INFO_KIND_NVDIMM:
570 pcdimm_info = value->type == MEMORY_DEVICE_INFO_KIND_DIMM ?
571 value->u.dimm.data : value->u.nvdimm.data;
572 node_mem[pcdimm_info->node].node_mem += pcdimm_info->size;
573 node_mem[pcdimm_info->node].node_plugged_mem +=
574 pcdimm_info->size;
575 break;
576 case MEMORY_DEVICE_INFO_KIND_VIRTIO_PMEM:
577 vpi = value->u.virtio_pmem.data;
578 /* TODO: once we support numa, assign to right node */
579 node_mem[0].node_mem += vpi->size;
580 node_mem[0].node_plugged_mem += vpi->size;
581 break;
582 default:
583 g_assert_not_reached();
584 }
585 }
586 }
587 qapi_free_MemoryDeviceInfoList(info_list);
588}
589
590void query_numa_node_mem(NumaNodeMem node_mem[], MachineState *ms)
591{
592 int i;
593
594 if (ms->numa_state == NULL || ms->numa_state->num_nodes <= 0) {
595 return;
596 }
597
598 numa_stat_memory_devices(node_mem);
599 for (i = 0; i < ms->numa_state->num_nodes; i++) {
600 node_mem[i].node_mem += ms->numa_state->nodes[i].node_mem;
601 }
602}
603
604void ram_block_notifier_add(RAMBlockNotifier *n)
605{
606 QLIST_INSERT_HEAD(&ram_list.ramblock_notifiers, n, next);
607}
608
609void ram_block_notifier_remove(RAMBlockNotifier *n)
610{
611 QLIST_REMOVE(n, next);
612}
613
614void ram_block_notify_add(void *host, size_t size)
615{
616 RAMBlockNotifier *notifier;
617
618 QLIST_FOREACH(notifier, &ram_list.ramblock_notifiers, next) {
619 notifier->ram_block_added(notifier, host, size);
620 }
621}
622
623void ram_block_notify_remove(void *host, size_t size)
624{
625 RAMBlockNotifier *notifier;
626
627 QLIST_FOREACH(notifier, &ram_list.ramblock_notifiers, next) {
628 notifier->ram_block_removed(notifier, host, size);
629 }
630}
631