]>
Commit | Line | Data |
---|---|---|
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/numa.h" | |
27 | #include "exec/cpu-common.h" | |
28 | #include "qemu/bitmap.h" | |
29 | #include "qom/cpu.h" | |
30 | #include "qemu/error-report.h" | |
31 | #include "include/exec/cpu-common.h" /* for RAM_ADDR_FMT */ | |
32 | #include "qapi-visit.h" | |
33 | #include "qapi/opts-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" | |
40 | ||
41 | QemuOptsList qemu_numa_opts = { | |
42 | .name = "numa", | |
43 | .implied_opt_name = "type", | |
44 | .head = QTAILQ_HEAD_INITIALIZER(qemu_numa_opts.head), | |
45 | .desc = { { 0 } } /* validated with OptsVisitor */ | |
46 | }; | |
47 | ||
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 | |
51 | */ | |
52 | int nb_numa_nodes; | |
53 | NodeInfo numa_info[MAX_NODES]; | |
54 | ||
55 | void numa_set_mem_node_id(ram_addr_t addr, uint64_t size, uint32_t node) | |
56 | { | |
57 | struct numa_addr_range *range; | |
58 | ||
59 | /* | |
60 | * Memory-less nodes can come here with 0 size in which case, | |
61 | * there is nothing to do. | |
62 | */ | |
63 | if (!size) { | |
64 | return; | |
65 | } | |
66 | ||
67 | range = g_malloc0(sizeof(*range)); | |
68 | range->mem_start = addr; | |
69 | range->mem_end = addr + size - 1; | |
70 | QLIST_INSERT_HEAD(&numa_info[node].addr, range, entry); | |
71 | } | |
72 | ||
73 | void numa_unset_mem_node_id(ram_addr_t addr, uint64_t size, uint32_t node) | |
74 | { | |
75 | struct numa_addr_range *range, *next; | |
76 | ||
77 | QLIST_FOREACH_SAFE(range, &numa_info[node].addr, entry, next) { | |
78 | if (addr == range->mem_start && (addr + size - 1) == range->mem_end) { | |
79 | QLIST_REMOVE(range, entry); | |
80 | g_free(range); | |
81 | return; | |
82 | } | |
83 | } | |
84 | } | |
85 | ||
86 | static void numa_set_mem_ranges(void) | |
87 | { | |
88 | int i; | |
89 | ram_addr_t mem_start = 0; | |
90 | ||
91 | /* | |
92 | * Deduce start address of each node and use it to store | |
93 | * the address range info in numa_info address range list | |
94 | */ | |
95 | for (i = 0; i < nb_numa_nodes; i++) { | |
96 | numa_set_mem_node_id(mem_start, numa_info[i].node_mem, i); | |
97 | mem_start += numa_info[i].node_mem; | |
98 | } | |
99 | } | |
100 | ||
101 | /* | |
102 | * Check if @addr falls under NUMA @node. | |
103 | */ | |
104 | static bool numa_addr_belongs_to_node(ram_addr_t addr, uint32_t node) | |
105 | { | |
106 | struct numa_addr_range *range; | |
107 | ||
108 | QLIST_FOREACH(range, &numa_info[node].addr, entry) { | |
109 | if (addr >= range->mem_start && addr <= range->mem_end) { | |
110 | return true; | |
111 | } | |
112 | } | |
113 | return false; | |
114 | } | |
115 | ||
116 | /* | |
117 | * Given an address, return the index of the NUMA node to which the | |
118 | * address belongs to. | |
119 | */ | |
120 | uint32_t numa_get_node(ram_addr_t addr, Error **errp) | |
121 | { | |
122 | uint32_t i; | |
123 | ||
124 | /* For non NUMA configurations, check if the addr falls under node 0 */ | |
125 | if (!nb_numa_nodes) { | |
126 | if (numa_addr_belongs_to_node(addr, 0)) { | |
127 | return 0; | |
128 | } | |
129 | } | |
130 | ||
131 | for (i = 0; i < nb_numa_nodes; i++) { | |
132 | if (numa_addr_belongs_to_node(addr, i)) { | |
133 | return i; | |
134 | } | |
135 | } | |
136 | ||
137 | error_setg(errp, "Address 0x" RAM_ADDR_FMT " doesn't belong to any " | |
138 | "NUMA node", addr); | |
139 | return -1; | |
140 | } | |
141 | ||
142 | static void numa_node_parse(NumaNodeOptions *node, QemuOpts *opts, Error **errp) | |
143 | { | |
144 | uint16_t nodenr; | |
145 | uint16List *cpus = NULL; | |
146 | ||
147 | if (node->has_nodeid) { | |
148 | nodenr = node->nodeid; | |
149 | } else { | |
150 | nodenr = nb_numa_nodes; | |
151 | } | |
152 | ||
153 | if (nodenr >= MAX_NODES) { | |
154 | error_setg(errp, "Max number of NUMA nodes reached: %" | |
155 | PRIu16 "", nodenr); | |
156 | return; | |
157 | } | |
158 | ||
159 | if (numa_info[nodenr].present) { | |
160 | error_setg(errp, "Duplicate NUMA nodeid: %" PRIu16, nodenr); | |
161 | return; | |
162 | } | |
163 | ||
164 | for (cpus = node->cpus; cpus; cpus = cpus->next) { | |
165 | if (cpus->value >= max_cpus) { | |
166 | error_setg(errp, | |
167 | "CPU index (%" PRIu16 ")" | |
168 | " should be smaller than maxcpus (%d)", | |
169 | cpus->value, max_cpus); | |
170 | return; | |
171 | } | |
172 | bitmap_set(numa_info[nodenr].node_cpu, cpus->value, 1); | |
173 | } | |
174 | ||
175 | if (node->has_mem && node->has_memdev) { | |
176 | error_setg(errp, "qemu: cannot specify both mem= and memdev="); | |
177 | return; | |
178 | } | |
179 | ||
180 | if (have_memdevs == -1) { | |
181 | have_memdevs = node->has_memdev; | |
182 | } | |
183 | if (node->has_memdev != have_memdevs) { | |
184 | error_setg(errp, "qemu: memdev option must be specified for either " | |
185 | "all or no nodes"); | |
186 | return; | |
187 | } | |
188 | ||
189 | if (node->has_mem) { | |
190 | uint64_t mem_size = node->mem; | |
191 | const char *mem_str = qemu_opt_get(opts, "mem"); | |
192 | /* Fix up legacy suffix-less format */ | |
193 | if (g_ascii_isdigit(mem_str[strlen(mem_str) - 1])) { | |
194 | mem_size <<= 20; | |
195 | } | |
196 | numa_info[nodenr].node_mem = mem_size; | |
197 | } | |
198 | if (node->has_memdev) { | |
199 | Object *o; | |
200 | o = object_resolve_path_type(node->memdev, TYPE_MEMORY_BACKEND, NULL); | |
201 | if (!o) { | |
202 | error_setg(errp, "memdev=%s is ambiguous", node->memdev); | |
203 | return; | |
204 | } | |
205 | ||
206 | object_ref(o); | |
207 | numa_info[nodenr].node_mem = object_property_get_int(o, "size", NULL); | |
208 | numa_info[nodenr].node_memdev = MEMORY_BACKEND(o); | |
209 | } | |
210 | numa_info[nodenr].present = true; | |
211 | max_numa_nodeid = MAX(max_numa_nodeid, nodenr + 1); | |
212 | } | |
213 | ||
214 | static int parse_numa(void *opaque, QemuOpts *opts, Error **errp) | |
215 | { | |
216 | NumaOptions *object = NULL; | |
217 | Error *err = NULL; | |
218 | ||
219 | { | |
220 | OptsVisitor *ov = opts_visitor_new(opts); | |
221 | visit_type_NumaOptions(opts_get_visitor(ov), NULL, &object, &err); | |
222 | opts_visitor_cleanup(ov); | |
223 | } | |
224 | ||
225 | if (err) { | |
226 | goto error; | |
227 | } | |
228 | ||
229 | switch (object->type) { | |
230 | case NUMA_OPTIONS_KIND_NODE: | |
231 | numa_node_parse(object->u.node.data, opts, &err); | |
232 | if (err) { | |
233 | goto error; | |
234 | } | |
235 | nb_numa_nodes++; | |
236 | break; | |
237 | default: | |
238 | abort(); | |
239 | } | |
240 | ||
241 | return 0; | |
242 | ||
243 | error: | |
244 | error_report_err(err); | |
245 | qapi_free_NumaOptions(object); | |
246 | ||
247 | return -1; | |
248 | } | |
249 | ||
250 | static char *enumerate_cpus(unsigned long *cpus, int max_cpus) | |
251 | { | |
252 | int cpu; | |
253 | bool first = true; | |
254 | GString *s = g_string_new(NULL); | |
255 | ||
256 | for (cpu = find_first_bit(cpus, max_cpus); | |
257 | cpu < max_cpus; | |
258 | cpu = find_next_bit(cpus, max_cpus, cpu + 1)) { | |
259 | g_string_append_printf(s, "%s%d", first ? "" : " ", cpu); | |
260 | first = false; | |
261 | } | |
262 | return g_string_free(s, FALSE); | |
263 | } | |
264 | ||
265 | static void validate_numa_cpus(void) | |
266 | { | |
267 | int i; | |
268 | DECLARE_BITMAP(seen_cpus, MAX_CPUMASK_BITS); | |
269 | ||
270 | bitmap_zero(seen_cpus, MAX_CPUMASK_BITS); | |
271 | for (i = 0; i < nb_numa_nodes; i++) { | |
272 | if (bitmap_intersects(seen_cpus, numa_info[i].node_cpu, | |
273 | MAX_CPUMASK_BITS)) { | |
274 | bitmap_and(seen_cpus, seen_cpus, | |
275 | numa_info[i].node_cpu, MAX_CPUMASK_BITS); | |
276 | error_report("CPU(s) present in multiple NUMA nodes: %s", | |
277 | enumerate_cpus(seen_cpus, max_cpus)); | |
278 | exit(EXIT_FAILURE); | |
279 | } | |
280 | bitmap_or(seen_cpus, seen_cpus, | |
281 | numa_info[i].node_cpu, MAX_CPUMASK_BITS); | |
282 | } | |
283 | ||
284 | if (!bitmap_full(seen_cpus, max_cpus)) { | |
285 | char *msg; | |
286 | bitmap_complement(seen_cpus, seen_cpus, max_cpus); | |
287 | msg = enumerate_cpus(seen_cpus, max_cpus); | |
288 | error_report("warning: CPU(s) not present in any NUMA nodes: %s", msg); | |
289 | error_report("warning: All CPU(s) up to maxcpus should be described " | |
290 | "in NUMA config"); | |
291 | g_free(msg); | |
292 | } | |
293 | } | |
294 | ||
295 | void parse_numa_opts(MachineClass *mc) | |
296 | { | |
297 | int i; | |
298 | ||
299 | if (qemu_opts_foreach(qemu_find_opts("numa"), parse_numa, NULL, NULL)) { | |
300 | exit(1); | |
301 | } | |
302 | ||
303 | assert(max_numa_nodeid <= MAX_NODES); | |
304 | ||
305 | /* No support for sparse NUMA node IDs yet: */ | |
306 | for (i = max_numa_nodeid - 1; i >= 0; i--) { | |
307 | /* Report large node IDs first, to make mistakes easier to spot */ | |
308 | if (!numa_info[i].present) { | |
309 | error_report("numa: Node ID missing: %d", i); | |
310 | exit(1); | |
311 | } | |
312 | } | |
313 | ||
314 | /* This must be always true if all nodes are present: */ | |
315 | assert(nb_numa_nodes == max_numa_nodeid); | |
316 | ||
317 | if (nb_numa_nodes > 0) { | |
318 | uint64_t numa_total; | |
319 | ||
320 | if (nb_numa_nodes > MAX_NODES) { | |
321 | nb_numa_nodes = MAX_NODES; | |
322 | } | |
323 | ||
324 | /* If no memory size is given for any node, assume the default case | |
325 | * and distribute the available memory equally across all nodes | |
326 | */ | |
327 | for (i = 0; i < nb_numa_nodes; i++) { | |
328 | if (numa_info[i].node_mem != 0) { | |
329 | break; | |
330 | } | |
331 | } | |
332 | if (i == nb_numa_nodes) { | |
333 | uint64_t usedmem = 0; | |
334 | ||
335 | /* On Linux, each node's border has to be 8MB aligned, | |
336 | * the final node gets the rest. | |
337 | */ | |
338 | for (i = 0; i < nb_numa_nodes - 1; i++) { | |
339 | numa_info[i].node_mem = (ram_size / nb_numa_nodes) & | |
340 | ~((1 << 23UL) - 1); | |
341 | usedmem += numa_info[i].node_mem; | |
342 | } | |
343 | numa_info[i].node_mem = ram_size - usedmem; | |
344 | } | |
345 | ||
346 | numa_total = 0; | |
347 | for (i = 0; i < nb_numa_nodes; i++) { | |
348 | numa_total += numa_info[i].node_mem; | |
349 | } | |
350 | if (numa_total != ram_size) { | |
351 | error_report("total memory for NUMA nodes (0x%" PRIx64 ")" | |
352 | " should equal RAM size (0x" RAM_ADDR_FMT ")", | |
353 | numa_total, ram_size); | |
354 | exit(1); | |
355 | } | |
356 | ||
357 | for (i = 0; i < nb_numa_nodes; i++) { | |
358 | QLIST_INIT(&numa_info[i].addr); | |
359 | } | |
360 | ||
361 | numa_set_mem_ranges(); | |
362 | ||
363 | for (i = 0; i < nb_numa_nodes; i++) { | |
364 | if (!bitmap_empty(numa_info[i].node_cpu, MAX_CPUMASK_BITS)) { | |
365 | break; | |
366 | } | |
367 | } | |
368 | /* Historically VCPUs were assigned in round-robin order to NUMA | |
369 | * nodes. However it causes issues with guest not handling it nice | |
370 | * in case where cores/threads from a multicore CPU appear on | |
371 | * different nodes. So allow boards to override default distribution | |
372 | * rule grouping VCPUs by socket so that VCPUs from the same socket | |
373 | * would be on the same node. | |
374 | */ | |
375 | if (i == nb_numa_nodes) { | |
376 | for (i = 0; i < max_cpus; i++) { | |
377 | unsigned node_id = i % nb_numa_nodes; | |
378 | if (mc->cpu_index_to_socket_id) { | |
379 | node_id = mc->cpu_index_to_socket_id(i) % nb_numa_nodes; | |
380 | } | |
381 | ||
382 | set_bit(i, numa_info[node_id].node_cpu); | |
383 | } | |
384 | } | |
385 | ||
386 | validate_numa_cpus(); | |
387 | } else { | |
388 | numa_set_mem_node_id(0, ram_size, 0); | |
389 | } | |
390 | } | |
391 | ||
392 | void numa_post_machine_init(void) | |
393 | { | |
394 | CPUState *cpu; | |
395 | int i; | |
396 | ||
397 | CPU_FOREACH(cpu) { | |
398 | for (i = 0; i < nb_numa_nodes; i++) { | |
399 | if (test_bit(cpu->cpu_index, numa_info[i].node_cpu)) { | |
400 | cpu->numa_node = i; | |
401 | } | |
402 | } | |
403 | } | |
404 | } | |
405 | ||
406 | static void allocate_system_memory_nonnuma(MemoryRegion *mr, Object *owner, | |
407 | const char *name, | |
408 | uint64_t ram_size) | |
409 | { | |
410 | if (mem_path) { | |
411 | #ifdef __linux__ | |
412 | Error *err = NULL; | |
413 | memory_region_init_ram_from_file(mr, owner, name, ram_size, false, | |
414 | mem_path, &err); | |
415 | if (err) { | |
416 | error_report_err(err); | |
417 | if (mem_prealloc) { | |
418 | exit(1); | |
419 | } | |
420 | ||
421 | /* Legacy behavior: if allocation failed, fall back to | |
422 | * regular RAM allocation. | |
423 | */ | |
424 | memory_region_init_ram(mr, owner, name, ram_size, &error_fatal); | |
425 | } | |
426 | #else | |
427 | fprintf(stderr, "-mem-path not supported on this host\n"); | |
428 | exit(1); | |
429 | #endif | |
430 | } else { | |
431 | memory_region_init_ram(mr, owner, name, ram_size, &error_fatal); | |
432 | } | |
433 | vmstate_register_ram_global(mr); | |
434 | } | |
435 | ||
436 | void memory_region_allocate_system_memory(MemoryRegion *mr, Object *owner, | |
437 | const char *name, | |
438 | uint64_t ram_size) | |
439 | { | |
440 | uint64_t addr = 0; | |
441 | int i; | |
442 | ||
443 | if (nb_numa_nodes == 0 || !have_memdevs) { | |
444 | allocate_system_memory_nonnuma(mr, owner, name, ram_size); | |
445 | return; | |
446 | } | |
447 | ||
448 | memory_region_init(mr, owner, name, ram_size); | |
449 | for (i = 0; i < MAX_NODES; i++) { | |
450 | uint64_t size = numa_info[i].node_mem; | |
451 | HostMemoryBackend *backend = numa_info[i].node_memdev; | |
452 | if (!backend) { | |
453 | continue; | |
454 | } | |
455 | MemoryRegion *seg = host_memory_backend_get_memory(backend, | |
456 | &error_fatal); | |
457 | ||
458 | if (memory_region_is_mapped(seg)) { | |
459 | char *path = object_get_canonical_path_component(OBJECT(backend)); | |
460 | error_report("memory backend %s is used multiple times. Each " | |
461 | "-numa option must use a different memdev value.", | |
462 | path); | |
463 | exit(1); | |
464 | } | |
465 | ||
466 | memory_region_add_subregion(mr, addr, seg); | |
467 | vmstate_register_ram_global(seg); | |
468 | addr += size; | |
469 | } | |
470 | } | |
471 | ||
472 | static void numa_stat_memory_devices(uint64_t node_mem[]) | |
473 | { | |
474 | MemoryDeviceInfoList *info_list = NULL; | |
475 | MemoryDeviceInfoList **prev = &info_list; | |
476 | MemoryDeviceInfoList *info; | |
477 | ||
478 | qmp_pc_dimm_device_list(qdev_get_machine(), &prev); | |
479 | for (info = info_list; info; info = info->next) { | |
480 | MemoryDeviceInfo *value = info->value; | |
481 | ||
482 | if (value) { | |
483 | switch (value->type) { | |
484 | case MEMORY_DEVICE_INFO_KIND_DIMM: | |
485 | node_mem[value->u.dimm.data->node] += value->u.dimm.data->size; | |
486 | break; | |
487 | default: | |
488 | break; | |
489 | } | |
490 | } | |
491 | } | |
492 | qapi_free_MemoryDeviceInfoList(info_list); | |
493 | } | |
494 | ||
495 | void query_numa_node_mem(uint64_t node_mem[]) | |
496 | { | |
497 | int i; | |
498 | ||
499 | if (nb_numa_nodes <= 0) { | |
500 | return; | |
501 | } | |
502 | ||
503 | numa_stat_memory_devices(node_mem); | |
504 | for (i = 0; i < nb_numa_nodes; i++) { | |
505 | node_mem[i] += numa_info[i].node_mem; | |
506 | } | |
507 | } | |
508 | ||
509 | static int query_memdev(Object *obj, void *opaque) | |
510 | { | |
511 | MemdevList **list = opaque; | |
512 | MemdevList *m = NULL; | |
513 | ||
514 | if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) { | |
515 | m = g_malloc0(sizeof(*m)); | |
516 | ||
517 | m->value = g_malloc0(sizeof(*m->value)); | |
518 | ||
519 | m->value->size = object_property_get_int(obj, "size", | |
520 | &error_abort); | |
521 | m->value->merge = object_property_get_bool(obj, "merge", | |
522 | &error_abort); | |
523 | m->value->dump = object_property_get_bool(obj, "dump", | |
524 | &error_abort); | |
525 | m->value->prealloc = object_property_get_bool(obj, | |
526 | "prealloc", | |
527 | &error_abort); | |
528 | m->value->policy = object_property_get_enum(obj, | |
529 | "policy", | |
530 | "HostMemPolicy", | |
531 | &error_abort); | |
532 | object_property_get_uint16List(obj, "host-nodes", | |
533 | &m->value->host_nodes, | |
534 | &error_abort); | |
535 | ||
536 | m->next = *list; | |
537 | *list = m; | |
538 | } | |
539 | ||
540 | return 0; | |
541 | } | |
542 | ||
543 | MemdevList *qmp_query_memdev(Error **errp) | |
544 | { | |
545 | Object *obj = object_get_objects_root(); | |
546 | MemdevList *list = NULL; | |
547 | ||
548 | object_child_foreach(obj, query_memdev, &list); | |
549 | return list; | |
550 | } |