numa.c

   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 "sysemu/numa.h"
  26 #include "exec/cpu-common.h"
  27 #include "qemu/bitmap.h"
  28 #include "qom/cpu.h"
  29 #include "qemu/error-report.h"
  30 #include "include/exec/cpu-common.h" /* for RAM_ADDR_FMT */
  31 #include "qapi-visit.h"
  32 #include "qapi/opts-visitor.h"
  33 #include "qapi/dealloc-visitor.h"
  34 #include "qapi/qmp/qerror.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"
  41
  42 QemuOptsList qemu_numa_opts = {
  43     .name = "numa",
  44     .implied_opt_name = "type",
  45     .head = QTAILQ_HEAD_INITIALIZER(qemu_numa_opts.head),
  46     .desc = { { 0 } } /* validated with OptsVisitor */
  47 };
  48
  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
  52                              */
  53 int nb_numa_nodes;
  54 NodeInfo numa_info[MAX_NODES];
  55
  56 static void numa_node_parse(NumaNodeOptions *node, QemuOpts *opts, Error **errp)
  57 {
  58     uint16_t nodenr;
  59     uint16List *cpus = NULL;
  60
  61     if (node->has_nodeid) {
  62         nodenr = node->nodeid;
  63     } else {
  64         nodenr = nb_numa_nodes;
  65     }
  66
  67     if (nodenr >= MAX_NODES) {
  68         error_setg(errp, "Max number of NUMA nodes reached: %"
  69                    PRIu16 "\n", nodenr);
  70         return;
  71     }
  72
  73     if (numa_info[nodenr].present) {
  74         error_setg(errp, "Duplicate NUMA nodeid: %" PRIu16, nodenr);
  75         return;
  76     }
  77
  78     for (cpus = node->cpus; cpus; cpus = cpus->next) {
  79         if (cpus->value > MAX_CPUMASK_BITS) {
  80             error_setg(errp, "CPU number %" PRIu16 " is bigger than %d",
  81                        cpus->value, MAX_CPUMASK_BITS);
  82             return;
  83         }
  84         bitmap_set(numa_info[nodenr].node_cpu, cpus->value, 1);
  85     }
  86
  87     if (node->has_mem && node->has_memdev) {
  88         error_setg(errp, "qemu: cannot specify both mem= and memdev=\n");
  89         return;
  90     }
  91
  92     if (have_memdevs == -1) {
  93         have_memdevs = node->has_memdev;
  94     }
  95     if (node->has_memdev != have_memdevs) {
  96         error_setg(errp, "qemu: memdev option must be specified for either "
  97                    "all or no nodes\n");
  98         return;
  99     }
 100
 101     if (node->has_mem) {
 102         uint64_t mem_size = node->mem;
 103         const char *mem_str = qemu_opt_get(opts, "mem");
 104         /* Fix up legacy suffix-less format */
 105         if (g_ascii_isdigit(mem_str[strlen(mem_str) - 1])) {
 106             mem_size <<= 20;
 107         }
 108         numa_info[nodenr].node_mem = mem_size;
 109     }
 110     if (node->has_memdev) {
 111         Object *o;
 112         o = object_resolve_path_type(node->memdev, TYPE_MEMORY_BACKEND, NULL);
 113         if (!o) {
 114             error_setg(errp, "memdev=%s is ambiguous", node->memdev);
 115             return;
 116         }
 117
 118         object_ref(o);
 119         numa_info[nodenr].node_mem = object_property_get_int(o, "size", NULL);
 120         numa_info[nodenr].node_memdev = MEMORY_BACKEND(o);
 121     }
 122     numa_info[nodenr].present = true;
 123     max_numa_nodeid = MAX(max_numa_nodeid, nodenr + 1);
 124 }
 125
 126 static int parse_numa(QemuOpts *opts, void *opaque)
 127 {
 128     NumaOptions *object = NULL;
 129     Error *err = NULL;
 130
 131     {
 132         OptsVisitor *ov = opts_visitor_new(opts);
 133         visit_type_NumaOptions(opts_get_visitor(ov), &object, NULL, &err);
 134         opts_visitor_cleanup(ov);
 135     }
 136
 137     if (err) {
 138         goto error;
 139     }
 140
 141     switch (object->kind) {
 142     case NUMA_OPTIONS_KIND_NODE:
 143         numa_node_parse(object->node, opts, &err);
 144         if (err) {
 145             goto error;
 146         }
 147         nb_numa_nodes++;
 148         break;
 149     default:
 150         abort();
 151     }
 152
 153     return 0;
 154
 155 error:
 156     qerror_report_err(err);
 157     error_free(err);
 158
 159     if (object) {
 160         QapiDeallocVisitor *dv = qapi_dealloc_visitor_new();
 161         visit_type_NumaOptions(qapi_dealloc_get_visitor(dv),
 162                                &object, NULL, NULL);
 163         qapi_dealloc_visitor_cleanup(dv);
 164     }
 165
 166     return -1;
 167 }
 168
 169 void parse_numa_opts(void)
 170 {
 171     int i;
 172
 173     if (qemu_opts_foreach(qemu_find_opts("numa"), parse_numa,
 174                           NULL, 1) != 0) {
 175         exit(1);
 176     }
 177
 178     assert(max_numa_nodeid <= MAX_NODES);
 179
 180     /* No support for sparse NUMA node IDs yet: */
 181     for (i = max_numa_nodeid - 1; i >= 0; i--) {
 182         /* Report large node IDs first, to make mistakes easier to spot */
 183         if (!numa_info[i].present) {
 184             error_report("numa: Node ID missing: %d", i);
 185             exit(1);
 186         }
 187     }
 188
 189     /* This must be always true if all nodes are present: */
 190     assert(nb_numa_nodes == max_numa_nodeid);
 191
 192     if (nb_numa_nodes > 0) {
 193         uint64_t numa_total;
 194
 195         if (nb_numa_nodes > MAX_NODES) {
 196             nb_numa_nodes = MAX_NODES;
 197         }
 198
 199         /* If no memory size is given for any node, assume the default case
 200          * and distribute the available memory equally across all nodes
 201          */
 202         for (i = 0; i < nb_numa_nodes; i++) {
 203             if (numa_info[i].node_mem != 0) {
 204                 break;
 205             }
 206         }
 207         if (i == nb_numa_nodes) {
 208             uint64_t usedmem = 0;
 209
 210             /* On Linux, each node's border has to be 8MB aligned,
 211              * the final node gets the rest.
 212              */
 213             for (i = 0; i < nb_numa_nodes - 1; i++) {
 214                 numa_info[i].node_mem = (ram_size / nb_numa_nodes) &
 215                                         ~((1 << 23UL) - 1);
 216                 usedmem += numa_info[i].node_mem;
 217             }
 218             numa_info[i].node_mem = ram_size - usedmem;
 219         }
 220
 221         numa_total = 0;
 222         for (i = 0; i < nb_numa_nodes; i++) {
 223             numa_total += numa_info[i].node_mem;
 224         }
 225         if (numa_total != ram_size) {
 226             error_report("total memory for NUMA nodes (0x%" PRIx64 ")"
 227                          " should equal RAM size (0x" RAM_ADDR_FMT ")",
 228                          numa_total, ram_size);
 229             exit(1);
 230         }
 231
 232         for (i = 0; i < nb_numa_nodes; i++) {
 233             if (!bitmap_empty(numa_info[i].node_cpu, MAX_CPUMASK_BITS)) {
 234                 break;
 235             }
 236         }
 237         /* assigning the VCPUs round-robin is easier to implement, guest OSes
 238          * must cope with this anyway, because there are BIOSes out there in
 239          * real machines which also use this scheme.
 240          */
 241         if (i == nb_numa_nodes) {
 242             for (i = 0; i < max_cpus; i++) {
 243                 set_bit(i, numa_info[i % nb_numa_nodes].node_cpu);
 244             }
 245         }
 246     }
 247 }
 248
 249 void set_numa_modes(void)
 250 {
 251     CPUState *cpu;
 252     int i;
 253
 254     CPU_FOREACH(cpu) {
 255         for (i = 0; i < nb_numa_nodes; i++) {
 256             if (test_bit(cpu->cpu_index, numa_info[i].node_cpu)) {
 257                 cpu->numa_node = i;
 258             }
 259         }
 260     }
 261 }
 262
 263 static void allocate_system_memory_nonnuma(MemoryRegion *mr, Object *owner,
 264                                            const char *name,
 265                                            uint64_t ram_size)
 266 {
 267     if (mem_path) {
 268 #ifdef __linux__
 269         Error *err = NULL;
 270         memory_region_init_ram_from_file(mr, owner, name, ram_size, false,
 271                                          mem_path, &err);
 272
 273         /* Legacy behavior: if allocation failed, fall back to
 274          * regular RAM allocation.
 275          */
 276         if (err) {
 277             qerror_report_err(err);
 278             error_free(err);
 279             memory_region_init_ram(mr, owner, name, ram_size, &error_abort);
 280         }
 281 #else
 282         fprintf(stderr, "-mem-path not supported on this host\n");
 283         exit(1);
 284 #endif
 285     } else {
 286         memory_region_init_ram(mr, owner, name, ram_size, &error_abort);
 287     }
 288     vmstate_register_ram_global(mr);
 289 }
 290
 291 void memory_region_allocate_system_memory(MemoryRegion *mr, Object *owner,
 292                                           const char *name,
 293                                           uint64_t ram_size)
 294 {
 295     uint64_t addr = 0;
 296     int i;
 297
 298     if (nb_numa_nodes == 0 || !have_memdevs) {
 299         allocate_system_memory_nonnuma(mr, owner, name, ram_size);
 300         return;
 301     }
 302
 303     memory_region_init(mr, owner, name, ram_size);
 304     for (i = 0; i < MAX_NODES; i++) {
 305         Error *local_err = NULL;
 306         uint64_t size = numa_info[i].node_mem;
 307         HostMemoryBackend *backend = numa_info[i].node_memdev;
 308         if (!backend) {
 309             continue;
 310         }
 311         MemoryRegion *seg = host_memory_backend_get_memory(backend, &local_err);
 312         if (local_err) {
 313             qerror_report_err(local_err);
 314             exit(1);
 315         }
 316
 317         if (memory_region_is_mapped(seg)) {
 318             char *path = object_get_canonical_path_component(OBJECT(backend));
 319             error_report("memory backend %s is used multiple times. Each "
 320                          "-numa option must use a different memdev value.",
 321                          path);
 322             exit(1);
 323         }
 324
 325         memory_region_add_subregion(mr, addr, seg);
 326         vmstate_register_ram_global(seg);
 327         addr += size;
 328     }
 329 }
 330
 331 static void numa_stat_memory_devices(uint64_t node_mem[])
 332 {
 333     MemoryDeviceInfoList *info_list = NULL;
 334     MemoryDeviceInfoList **prev = &info_list;
 335     MemoryDeviceInfoList *info;
 336
 337     qmp_pc_dimm_device_list(qdev_get_machine(), &prev);
 338     for (info = info_list; info; info = info->next) {
 339         MemoryDeviceInfo *value = info->value;
 340
 341         if (value) {
 342             switch (value->kind) {
 343             case MEMORY_DEVICE_INFO_KIND_DIMM:
 344                 node_mem[value->dimm->node] += value->dimm->size;
 345                 break;
 346             default:
 347                 break;
 348             }
 349         }
 350     }
 351     qapi_free_MemoryDeviceInfoList(info_list);
 352 }
 353
 354 void query_numa_node_mem(uint64_t node_mem[])
 355 {
 356     int i;
 357
 358     if (nb_numa_nodes <= 0) {
 359         return;
 360     }
 361
 362     numa_stat_memory_devices(node_mem);
 363     for (i = 0; i < nb_numa_nodes; i++) {
 364         node_mem[i] += numa_info[i].node_mem;
 365     }
 366 }
 367
 368 static int query_memdev(Object *obj, void *opaque)
 369 {
 370     MemdevList **list = opaque;
 371     MemdevList *m = NULL;
 372     Error *err = NULL;
 373
 374     if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) {
 375         m = g_malloc0(sizeof(*m));
 376
 377         m->value = g_malloc0(sizeof(*m->value));
 378
 379         m->value->size = object_property_get_int(obj, "size",
 380                                                  &err);
 381         if (err) {
 382             goto error;
 383         }
 384
 385         m->value->merge = object_property_get_bool(obj, "merge",
 386                                                    &err);
 387         if (err) {
 388             goto error;
 389         }
 390
 391         m->value->dump = object_property_get_bool(obj, "dump",
 392                                                   &err);
 393         if (err) {
 394             goto error;
 395         }
 396
 397         m->value->prealloc = object_property_get_bool(obj,
 398                                                       "prealloc", &err);
 399         if (err) {
 400             goto error;
 401         }
 402
 403         m->value->policy = object_property_get_enum(obj,
 404                                                     "policy",
 405                                                     HostMemPolicy_lookup,
 406                                                     &err);
 407         if (err) {
 408             goto error;
 409         }
 410
 411         object_property_get_uint16List(obj, "host-nodes",
 412                                        &m->value->host_nodes, &err);
 413         if (err) {
 414             goto error;
 415         }
 416
 417         m->next = *list;
 418         *list = m;
 419     }
 420
 421     return 0;
 422 error:
 423     g_free(m->value);
 424     g_free(m);
 425
 426     return -1;
 427 }
 428
 429 MemdevList *qmp_query_memdev(Error **errp)
 430 {
 431     Object *obj;
 432     MemdevList *list = NULL;
 433
 434     obj = object_resolve_path("/objects", NULL);
 435     if (obj == NULL) {
 436         return NULL;
 437     }
 438
 439     if (object_child_foreach(obj, query_memdev, &list) != 0) {
 440         goto error;
 441     }
 442
 443     return list;
 444
 445 error:
 446     qapi_free_MemdevList(list);
 447     return NULL;
 448 }