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_cpus) {
  80             error_setg(errp,
  81                        "CPU index (%" PRIu16 ")"
  82                        " should be smaller than maxcpus (%d)",
  83                        cpus->value, max_cpus);
  84             return;
  85         }
  86         bitmap_set(numa_info[nodenr].node_cpu, cpus->value, 1);
  87     }
  88
  89     if (node->has_mem && node->has_memdev) {
  90         error_setg(errp, "qemu: cannot specify both mem= and memdev=");
  91         return;
  92     }
  93
  94     if (have_memdevs == -1) {
  95         have_memdevs = node->has_memdev;
  96     }
  97     if (node->has_memdev != have_memdevs) {
  98         error_setg(errp, "qemu: memdev option must be specified for either "
  99                    "all or no nodes");
 100         return;
 101     }
 102
 103     if (node->has_mem) {
 104         uint64_t mem_size = node->mem;
 105         const char *mem_str = qemu_opt_get(opts, "mem");
 106         /* Fix up legacy suffix-less format */
 107         if (g_ascii_isdigit(mem_str[strlen(mem_str) - 1])) {
 108             mem_size <<= 20;
 109         }
 110         numa_info[nodenr].node_mem = mem_size;
 111     }
 112     if (node->has_memdev) {
 113         Object *o;
 114         o = object_resolve_path_type(node->memdev, TYPE_MEMORY_BACKEND, NULL);
 115         if (!o) {
 116             error_setg(errp, "memdev=%s is ambiguous", node->memdev);
 117             return;
 118         }
 119
 120         object_ref(o);
 121         numa_info[nodenr].node_mem = object_property_get_int(o, "size", NULL);
 122         numa_info[nodenr].node_memdev = MEMORY_BACKEND(o);
 123     }
 124     numa_info[nodenr].present = true;
 125     max_numa_nodeid = MAX(max_numa_nodeid, nodenr + 1);
 126 }
 127
 128 static int parse_numa(QemuOpts *opts, void *opaque)
 129 {
 130     NumaOptions *object = NULL;
 131     Error *err = NULL;
 132
 133     {
 134         OptsVisitor *ov = opts_visitor_new(opts);
 135         visit_type_NumaOptions(opts_get_visitor(ov), &object, NULL, &err);
 136         opts_visitor_cleanup(ov);
 137     }
 138
 139     if (err) {
 140         goto error;
 141     }
 142
 143     switch (object->kind) {
 144     case NUMA_OPTIONS_KIND_NODE:
 145         numa_node_parse(object->node, opts, &err);
 146         if (err) {
 147             goto error;
 148         }
 149         nb_numa_nodes++;
 150         break;
 151     default:
 152         abort();
 153     }
 154
 155     return 0;
 156
 157 error:
 158     error_report_err(err);
 159
 160     if (object) {
 161         QapiDeallocVisitor *dv = qapi_dealloc_visitor_new();
 162         visit_type_NumaOptions(qapi_dealloc_get_visitor(dv),
 163                                &object, NULL, NULL);
 164         qapi_dealloc_visitor_cleanup(dv);
 165     }
 166
 167     return -1;
 168 }
 169
 170 static char *enumerate_cpus(unsigned long *cpus, int max_cpus)
 171 {
 172     int cpu;
 173     bool first = true;
 174     GString *s = g_string_new(NULL);
 175
 176     for (cpu = find_first_bit(cpus, max_cpus);
 177         cpu < max_cpus;
 178         cpu = find_next_bit(cpus, max_cpus, cpu + 1)) {
 179         g_string_append_printf(s, "%s%d", first ? "" : " ", cpu);
 180         first = false;
 181     }
 182     return g_string_free(s, FALSE);
 183 }
 184
 185 static void validate_numa_cpus(void)
 186 {
 187     int i;
 188     DECLARE_BITMAP(seen_cpus, MAX_CPUMASK_BITS);
 189
 190     bitmap_zero(seen_cpus, MAX_CPUMASK_BITS);
 191     for (i = 0; i < nb_numa_nodes; i++) {
 192         if (bitmap_intersects(seen_cpus, numa_info[i].node_cpu,
 193                               MAX_CPUMASK_BITS)) {
 194             bitmap_and(seen_cpus, seen_cpus,
 195                        numa_info[i].node_cpu, MAX_CPUMASK_BITS);
 196             error_report("CPU(s) present in multiple NUMA nodes: %s",
 197                          enumerate_cpus(seen_cpus, max_cpus));;
 198             exit(EXIT_FAILURE);
 199         }
 200         bitmap_or(seen_cpus, seen_cpus,
 201                   numa_info[i].node_cpu, MAX_CPUMASK_BITS);
 202     }
 203 }
 204
 205 void parse_numa_opts(MachineClass *mc)
 206 {
 207     int i;
 208
 209     if (qemu_opts_foreach(qemu_find_opts("numa"), parse_numa,
 210                           NULL, 1) != 0) {
 211         exit(1);
 212     }
 213
 214     assert(max_numa_nodeid <= MAX_NODES);
 215
 216     /* No support for sparse NUMA node IDs yet: */
 217     for (i = max_numa_nodeid - 1; i >= 0; i--) {
 218         /* Report large node IDs first, to make mistakes easier to spot */
 219         if (!numa_info[i].present) {
 220             error_report("numa: Node ID missing: %d", i);
 221             exit(1);
 222         }
 223     }
 224
 225     /* This must be always true if all nodes are present: */
 226     assert(nb_numa_nodes == max_numa_nodeid);
 227
 228     if (nb_numa_nodes > 0) {
 229         uint64_t numa_total;
 230
 231         if (nb_numa_nodes > MAX_NODES) {
 232             nb_numa_nodes = MAX_NODES;
 233         }
 234
 235         /* If no memory size is given for any node, assume the default case
 236          * and distribute the available memory equally across all nodes
 237          */
 238         for (i = 0; i < nb_numa_nodes; i++) {
 239             if (numa_info[i].node_mem != 0) {
 240                 break;
 241             }
 242         }
 243         if (i == nb_numa_nodes) {
 244             uint64_t usedmem = 0;
 245
 246             /* On Linux, each node's border has to be 8MB aligned,
 247              * the final node gets the rest.
 248              */
 249             for (i = 0; i < nb_numa_nodes - 1; i++) {
 250                 numa_info[i].node_mem = (ram_size / nb_numa_nodes) &
 251                                         ~((1 << 23UL) - 1);
 252                 usedmem += numa_info[i].node_mem;
 253             }
 254             numa_info[i].node_mem = ram_size - usedmem;
 255         }
 256
 257         numa_total = 0;
 258         for (i = 0; i < nb_numa_nodes; i++) {
 259             numa_total += numa_info[i].node_mem;
 260         }
 261         if (numa_total != ram_size) {
 262             error_report("total memory for NUMA nodes (0x%" PRIx64 ")"
 263                          " should equal RAM size (0x" RAM_ADDR_FMT ")",
 264                          numa_total, ram_size);
 265             exit(1);
 266         }
 267
 268         for (i = 0; i < nb_numa_nodes; i++) {
 269             if (!bitmap_empty(numa_info[i].node_cpu, MAX_CPUMASK_BITS)) {
 270                 break;
 271             }
 272         }
 273         /* Historically VCPUs were assigned in round-robin order to NUMA
 274          * nodes. However it causes issues with guest not handling it nice
 275          * in case where cores/threads from a multicore CPU appear on
 276          * different nodes. So allow boards to override default distribution
 277          * rule grouping VCPUs by socket so that VCPUs from the same socket
 278          * would be on the same node.
 279          */
 280         if (i == nb_numa_nodes) {
 281             for (i = 0; i < max_cpus; i++) {
 282                 unsigned node_id = i % nb_numa_nodes;
 283                 if (mc->cpu_index_to_socket_id) {
 284                     node_id = mc->cpu_index_to_socket_id(i) % nb_numa_nodes;
 285                 }
 286
 287                 set_bit(i, numa_info[node_id].node_cpu);
 288             }
 289         }
 290
 291         validate_numa_cpus();
 292     }
 293 }
 294
 295 void numa_post_machine_init(void)
 296 {
 297     CPUState *cpu;
 298     int i;
 299
 300     CPU_FOREACH(cpu) {
 301         for (i = 0; i < nb_numa_nodes; i++) {
 302             if (test_bit(cpu->cpu_index, numa_info[i].node_cpu)) {
 303                 cpu->numa_node = i;
 304             }
 305         }
 306     }
 307 }
 308
 309 static void allocate_system_memory_nonnuma(MemoryRegion *mr, Object *owner,
 310                                            const char *name,
 311                                            uint64_t ram_size)
 312 {
 313     if (mem_path) {
 314 #ifdef __linux__
 315         Error *err = NULL;
 316         memory_region_init_ram_from_file(mr, owner, name, ram_size, false,
 317                                          mem_path, &err);
 318
 319         /* Legacy behavior: if allocation failed, fall back to
 320          * regular RAM allocation.
 321          */
 322         if (err) {
 323             error_report_err(err);
 324             memory_region_init_ram(mr, owner, name, ram_size, &error_abort);
 325         }
 326 #else
 327         fprintf(stderr, "-mem-path not supported on this host\n");
 328         exit(1);
 329 #endif
 330     } else {
 331         memory_region_init_ram(mr, owner, name, ram_size, &error_abort);
 332     }
 333     vmstate_register_ram_global(mr);
 334 }
 335
 336 void memory_region_allocate_system_memory(MemoryRegion *mr, Object *owner,
 337                                           const char *name,
 338                                           uint64_t ram_size)
 339 {
 340     uint64_t addr = 0;
 341     int i;
 342
 343     if (nb_numa_nodes == 0 || !have_memdevs) {
 344         allocate_system_memory_nonnuma(mr, owner, name, ram_size);
 345         return;
 346     }
 347
 348     memory_region_init(mr, owner, name, ram_size);
 349     for (i = 0; i < MAX_NODES; i++) {
 350         Error *local_err = NULL;
 351         uint64_t size = numa_info[i].node_mem;
 352         HostMemoryBackend *backend = numa_info[i].node_memdev;
 353         if (!backend) {
 354             continue;
 355         }
 356         MemoryRegion *seg = host_memory_backend_get_memory(backend, &local_err);
 357         if (local_err) {
 358             error_report_err(local_err);
 359             exit(1);
 360         }
 361
 362         if (memory_region_is_mapped(seg)) {
 363             char *path = object_get_canonical_path_component(OBJECT(backend));
 364             error_report("memory backend %s is used multiple times. Each "
 365                          "-numa option must use a different memdev value.",
 366                          path);
 367             exit(1);
 368         }
 369
 370         memory_region_add_subregion(mr, addr, seg);
 371         vmstate_register_ram_global(seg);
 372         addr += size;
 373     }
 374 }
 375
 376 static void numa_stat_memory_devices(uint64_t node_mem[])
 377 {
 378     MemoryDeviceInfoList *info_list = NULL;
 379     MemoryDeviceInfoList **prev = &info_list;
 380     MemoryDeviceInfoList *info;
 381
 382     qmp_pc_dimm_device_list(qdev_get_machine(), &prev);
 383     for (info = info_list; info; info = info->next) {
 384         MemoryDeviceInfo *value = info->value;
 385
 386         if (value) {
 387             switch (value->kind) {
 388             case MEMORY_DEVICE_INFO_KIND_DIMM:
 389                 node_mem[value->dimm->node] += value->dimm->size;
 390                 break;
 391             default:
 392                 break;
 393             }
 394         }
 395     }
 396     qapi_free_MemoryDeviceInfoList(info_list);
 397 }
 398
 399 void query_numa_node_mem(uint64_t node_mem[])
 400 {
 401     int i;
 402
 403     if (nb_numa_nodes <= 0) {
 404         return;
 405     }
 406
 407     numa_stat_memory_devices(node_mem);
 408     for (i = 0; i < nb_numa_nodes; i++) {
 409         node_mem[i] += numa_info[i].node_mem;
 410     }
 411 }
 412
 413 static int query_memdev(Object *obj, void *opaque)
 414 {
 415     MemdevList **list = opaque;
 416     MemdevList *m = NULL;
 417     Error *err = NULL;
 418
 419     if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) {
 420         m = g_malloc0(sizeof(*m));
 421
 422         m->value = g_malloc0(sizeof(*m->value));
 423
 424         m->value->size = object_property_get_int(obj, "size",
 425                                                  &err);
 426         if (err) {
 427             goto error;
 428         }
 429
 430         m->value->merge = object_property_get_bool(obj, "merge",
 431                                                    &err);
 432         if (err) {
 433             goto error;
 434         }
 435
 436         m->value->dump = object_property_get_bool(obj, "dump",
 437                                                   &err);
 438         if (err) {
 439             goto error;
 440         }
 441
 442         m->value->prealloc = object_property_get_bool(obj,
 443                                                       "prealloc", &err);
 444         if (err) {
 445             goto error;
 446         }
 447
 448         m->value->policy = object_property_get_enum(obj,
 449                                                     "policy",
 450                                                     HostMemPolicy_lookup,
 451                                                     &err);
 452         if (err) {
 453             goto error;
 454         }
 455
 456         object_property_get_uint16List(obj, "host-nodes",
 457                                        &m->value->host_nodes, &err);
 458         if (err) {
 459             goto error;
 460         }
 461
 462         m->next = *list;
 463         *list = m;
 464     }
 465
 466     return 0;
 467 error:
 468     g_free(m->value);
 469     g_free(m);
 470
 471     return -1;
 472 }
 473
 474 MemdevList *qmp_query_memdev(Error **errp)
 475 {
 476     Object *obj;
 477     MemdevList *list = NULL;
 478
 479     obj = object_resolve_path("/objects", NULL);
 480     if (obj == NULL) {
 481         return NULL;
 482     }
 483
 484     if (object_child_foreach(obj, query_memdev, &list) != 0) {
 485         goto error;
 486     }
 487
 488     return list;
 489
 490 error:
 491     qapi_free_MemdevList(list);
 492     return NULL;
 493 }