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 "", 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=");
  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");
  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     error_report_err(err);
 157
 158     if (object) {
 159         QapiDeallocVisitor *dv = qapi_dealloc_visitor_new();
 160         visit_type_NumaOptions(qapi_dealloc_get_visitor(dv),
 161                                &object, NULL, NULL);
 162         qapi_dealloc_visitor_cleanup(dv);
 163     }
 164
 165     return -1;
 166 }
 167
 168 void parse_numa_opts(void)
 169 {
 170     int i;
 171
 172     if (qemu_opts_foreach(qemu_find_opts("numa"), parse_numa,
 173                           NULL, 1) != 0) {
 174         exit(1);
 175     }
 176
 177     assert(max_numa_nodeid <= MAX_NODES);
 178
 179     /* No support for sparse NUMA node IDs yet: */
 180     for (i = max_numa_nodeid - 1; i >= 0; i--) {
 181         /* Report large node IDs first, to make mistakes easier to spot */
 182         if (!numa_info[i].present) {
 183             error_report("numa: Node ID missing: %d", i);
 184             exit(1);
 185         }
 186     }
 187
 188     /* This must be always true if all nodes are present: */
 189     assert(nb_numa_nodes == max_numa_nodeid);
 190
 191     if (nb_numa_nodes > 0) {
 192         uint64_t numa_total;
 193
 194         if (nb_numa_nodes > MAX_NODES) {
 195             nb_numa_nodes = MAX_NODES;
 196         }
 197
 198         /* If no memory size is given for any node, assume the default case
 199          * and distribute the available memory equally across all nodes
 200          */
 201         for (i = 0; i < nb_numa_nodes; i++) {
 202             if (numa_info[i].node_mem != 0) {
 203                 break;
 204             }
 205         }
 206         if (i == nb_numa_nodes) {
 207             uint64_t usedmem = 0;
 208
 209             /* On Linux, each node's border has to be 8MB aligned,
 210              * the final node gets the rest.
 211              */
 212             for (i = 0; i < nb_numa_nodes - 1; i++) {
 213                 numa_info[i].node_mem = (ram_size / nb_numa_nodes) &
 214                                         ~((1 << 23UL) - 1);
 215                 usedmem += numa_info[i].node_mem;
 216             }
 217             numa_info[i].node_mem = ram_size - usedmem;
 218         }
 219
 220         numa_total = 0;
 221         for (i = 0; i < nb_numa_nodes; i++) {
 222             numa_total += numa_info[i].node_mem;
 223         }
 224         if (numa_total != ram_size) {
 225             error_report("total memory for NUMA nodes (0x%" PRIx64 ")"
 226                          " should equal RAM size (0x" RAM_ADDR_FMT ")",
 227                          numa_total, ram_size);
 228             exit(1);
 229         }
 230
 231         for (i = 0; i < nb_numa_nodes; i++) {
 232             if (!bitmap_empty(numa_info[i].node_cpu, MAX_CPUMASK_BITS)) {
 233                 break;
 234             }
 235         }
 236         /* assigning the VCPUs round-robin is easier to implement, guest OSes
 237          * must cope with this anyway, because there are BIOSes out there in
 238          * real machines which also use this scheme.
 239          */
 240         if (i == nb_numa_nodes) {
 241             for (i = 0; i < max_cpus; i++) {
 242                 set_bit(i, numa_info[i % nb_numa_nodes].node_cpu);
 243             }
 244         }
 245     }
 246 }
 247
 248 void numa_post_machine_init(void)
 249 {
 250     CPUState *cpu;
 251     int i;
 252
 253     CPU_FOREACH(cpu) {
 254         for (i = 0; i < nb_numa_nodes; i++) {
 255             if (test_bit(cpu->cpu_index, numa_info[i].node_cpu)) {
 256                 cpu->numa_node = i;
 257             }
 258         }
 259     }
 260 }
 261
 262 static void allocate_system_memory_nonnuma(MemoryRegion *mr, Object *owner,
 263                                            const char *name,
 264                                            uint64_t ram_size)
 265 {
 266     if (mem_path) {
 267 #ifdef __linux__
 268         Error *err = NULL;
 269         memory_region_init_ram_from_file(mr, owner, name, ram_size, false,
 270                                          mem_path, &err);
 271
 272         /* Legacy behavior: if allocation failed, fall back to
 273          * regular RAM allocation.
 274          */
 275         if (err) {
 276             error_report_err(err);
 277             memory_region_init_ram(mr, owner, name, ram_size, &error_abort);
 278         }
 279 #else
 280         fprintf(stderr, "-mem-path not supported on this host\n");
 281         exit(1);
 282 #endif
 283     } else {
 284         memory_region_init_ram(mr, owner, name, ram_size, &error_abort);
 285     }
 286     vmstate_register_ram_global(mr);
 287 }
 288
 289 void memory_region_allocate_system_memory(MemoryRegion *mr, Object *owner,
 290                                           const char *name,
 291                                           uint64_t ram_size)
 292 {
 293     uint64_t addr = 0;
 294     int i;
 295
 296     if (nb_numa_nodes == 0 || !have_memdevs) {
 297         allocate_system_memory_nonnuma(mr, owner, name, ram_size);
 298         return;
 299     }
 300
 301     memory_region_init(mr, owner, name, ram_size);
 302     for (i = 0; i < MAX_NODES; i++) {
 303         Error *local_err = NULL;
 304         uint64_t size = numa_info[i].node_mem;
 305         HostMemoryBackend *backend = numa_info[i].node_memdev;
 306         if (!backend) {
 307             continue;
 308         }
 309         MemoryRegion *seg = host_memory_backend_get_memory(backend, &local_err);
 310         if (local_err) {
 311             error_report_err(local_err);
 312             exit(1);
 313         }
 314
 315         if (memory_region_is_mapped(seg)) {
 316             char *path = object_get_canonical_path_component(OBJECT(backend));
 317             error_report("memory backend %s is used multiple times. Each "
 318                          "-numa option must use a different memdev value.",
 319                          path);
 320             exit(1);
 321         }
 322
 323         memory_region_add_subregion(mr, addr, seg);
 324         vmstate_register_ram_global(seg);
 325         addr += size;
 326     }
 327 }
 328
 329 static void numa_stat_memory_devices(uint64_t node_mem[])
 330 {
 331     MemoryDeviceInfoList *info_list = NULL;
 332     MemoryDeviceInfoList **prev = &info_list;
 333     MemoryDeviceInfoList *info;
 334
 335     qmp_pc_dimm_device_list(qdev_get_machine(), &prev);
 336     for (info = info_list; info; info = info->next) {
 337         MemoryDeviceInfo *value = info->value;
 338
 339         if (value) {
 340             switch (value->kind) {
 341             case MEMORY_DEVICE_INFO_KIND_DIMM:
 342                 node_mem[value->dimm->node] += value->dimm->size;
 343                 break;
 344             default:
 345                 break;
 346             }
 347         }
 348     }
 349     qapi_free_MemoryDeviceInfoList(info_list);
 350 }
 351
 352 void query_numa_node_mem(uint64_t node_mem[])
 353 {
 354     int i;
 355
 356     if (nb_numa_nodes <= 0) {
 357         return;
 358     }
 359
 360     numa_stat_memory_devices(node_mem);
 361     for (i = 0; i < nb_numa_nodes; i++) {
 362         node_mem[i] += numa_info[i].node_mem;
 363     }
 364 }
 365
 366 static int query_memdev(Object *obj, void *opaque)
 367 {
 368     MemdevList **list = opaque;
 369     MemdevList *m = NULL;
 370     Error *err = NULL;
 371
 372     if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) {
 373         m = g_malloc0(sizeof(*m));
 374
 375         m->value = g_malloc0(sizeof(*m->value));
 376
 377         m->value->size = object_property_get_int(obj, "size",
 378                                                  &err);
 379         if (err) {
 380             goto error;
 381         }
 382
 383         m->value->merge = object_property_get_bool(obj, "merge",
 384                                                    &err);
 385         if (err) {
 386             goto error;
 387         }
 388
 389         m->value->dump = object_property_get_bool(obj, "dump",
 390                                                   &err);
 391         if (err) {
 392             goto error;
 393         }
 394
 395         m->value->prealloc = object_property_get_bool(obj,
 396                                                       "prealloc", &err);
 397         if (err) {
 398             goto error;
 399         }
 400
 401         m->value->policy = object_property_get_enum(obj,
 402                                                     "policy",
 403                                                     HostMemPolicy_lookup,
 404                                                     &err);
 405         if (err) {
 406             goto error;
 407         }
 408
 409         object_property_get_uint16List(obj, "host-nodes",
 410                                        &m->value->host_nodes, &err);
 411         if (err) {
 412             goto error;
 413         }
 414
 415         m->next = *list;
 416         *list = m;
 417     }
 418
 419     return 0;
 420 error:
 421     g_free(m->value);
 422     g_free(m);
 423
 424     return -1;
 425 }
 426
 427 MemdevList *qmp_query_memdev(Error **errp)
 428 {
 429     Object *obj;
 430     MemdevList *list = NULL;
 431
 432     obj = object_resolve_path("/objects", NULL);
 433     if (obj == NULL) {
 434         return NULL;
 435     }
 436
 437     if (object_child_foreach(obj, query_memdev, &list) != 0) {
 438         goto error;
 439     }
 440
 441     return list;
 442
 443 error:
 444     qapi_free_MemdevList(list);
 445     return NULL;
 446 }