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     if (!bitmap_full(seen_cpus, max_cpus)) {
 205         char *msg;
 206         bitmap_complement(seen_cpus, seen_cpus, max_cpus);
 207         msg = enumerate_cpus(seen_cpus, max_cpus);
 208         error_report("warning: CPU(s) not present in any NUMA nodes: %s", msg);
 209         error_report("warning: All CPU(s) up to maxcpus should be described "
 210                      "in NUMA config");
 211         g_free(msg);
 212     }
 213 }
 214
 215 void parse_numa_opts(MachineClass *mc)
 216 {
 217     int i;
 218
 219     if (qemu_opts_foreach(qemu_find_opts("numa"), parse_numa,
 220                           NULL, 1) != 0) {
 221         exit(1);
 222     }
 223
 224     assert(max_numa_nodeid <= MAX_NODES);
 225
 226     /* No support for sparse NUMA node IDs yet: */
 227     for (i = max_numa_nodeid - 1; i >= 0; i--) {
 228         /* Report large node IDs first, to make mistakes easier to spot */
 229         if (!numa_info[i].present) {
 230             error_report("numa: Node ID missing: %d", i);
 231             exit(1);
 232         }
 233     }
 234
 235     /* This must be always true if all nodes are present: */
 236     assert(nb_numa_nodes == max_numa_nodeid);
 237
 238     if (nb_numa_nodes > 0) {
 239         uint64_t numa_total;
 240
 241         if (nb_numa_nodes > MAX_NODES) {
 242             nb_numa_nodes = MAX_NODES;
 243         }
 244
 245         /* If no memory size is given for any node, assume the default case
 246          * and distribute the available memory equally across all nodes
 247          */
 248         for (i = 0; i < nb_numa_nodes; i++) {
 249             if (numa_info[i].node_mem != 0) {
 250                 break;
 251             }
 252         }
 253         if (i == nb_numa_nodes) {
 254             uint64_t usedmem = 0;
 255
 256             /* On Linux, each node's border has to be 8MB aligned,
 257              * the final node gets the rest.
 258              */
 259             for (i = 0; i < nb_numa_nodes - 1; i++) {
 260                 numa_info[i].node_mem = (ram_size / nb_numa_nodes) &
 261                                         ~((1 << 23UL) - 1);
 262                 usedmem += numa_info[i].node_mem;
 263             }
 264             numa_info[i].node_mem = ram_size - usedmem;
 265         }
 266
 267         numa_total = 0;
 268         for (i = 0; i < nb_numa_nodes; i++) {
 269             numa_total += numa_info[i].node_mem;
 270         }
 271         if (numa_total != ram_size) {
 272             error_report("total memory for NUMA nodes (0x%" PRIx64 ")"
 273                          " should equal RAM size (0x" RAM_ADDR_FMT ")",
 274                          numa_total, ram_size);
 275             exit(1);
 276         }
 277
 278         for (i = 0; i < nb_numa_nodes; i++) {
 279             if (!bitmap_empty(numa_info[i].node_cpu, MAX_CPUMASK_BITS)) {
 280                 break;
 281             }
 282         }
 283         /* Historically VCPUs were assigned in round-robin order to NUMA
 284          * nodes. However it causes issues with guest not handling it nice
 285          * in case where cores/threads from a multicore CPU appear on
 286          * different nodes. So allow boards to override default distribution
 287          * rule grouping VCPUs by socket so that VCPUs from the same socket
 288          * would be on the same node.
 289          */
 290         if (i == nb_numa_nodes) {
 291             for (i = 0; i < max_cpus; i++) {
 292                 unsigned node_id = i % nb_numa_nodes;
 293                 if (mc->cpu_index_to_socket_id) {
 294                     node_id = mc->cpu_index_to_socket_id(i) % nb_numa_nodes;
 295                 }
 296
 297                 set_bit(i, numa_info[node_id].node_cpu);
 298             }
 299         }
 300
 301         validate_numa_cpus();
 302     }
 303 }
 304
 305 void numa_post_machine_init(void)
 306 {
 307     CPUState *cpu;
 308     int i;
 309
 310     CPU_FOREACH(cpu) {
 311         for (i = 0; i < nb_numa_nodes; i++) {
 312             if (test_bit(cpu->cpu_index, numa_info[i].node_cpu)) {
 313                 cpu->numa_node = i;
 314             }
 315         }
 316     }
 317 }
 318
 319 static void allocate_system_memory_nonnuma(MemoryRegion *mr, Object *owner,
 320                                            const char *name,
 321                                            uint64_t ram_size)
 322 {
 323     if (mem_path) {
 324 #ifdef __linux__
 325         Error *err = NULL;
 326         memory_region_init_ram_from_file(mr, owner, name, ram_size, false,
 327                                          mem_path, &err);
 328
 329         /* Legacy behavior: if allocation failed, fall back to
 330          * regular RAM allocation.
 331          */
 332         if (err) {
 333             error_report_err(err);
 334             memory_region_init_ram(mr, owner, name, ram_size, &error_abort);
 335         }
 336 #else
 337         fprintf(stderr, "-mem-path not supported on this host\n");
 338         exit(1);
 339 #endif
 340     } else {
 341         memory_region_init_ram(mr, owner, name, ram_size, &error_abort);
 342     }
 343     vmstate_register_ram_global(mr);
 344 }
 345
 346 void memory_region_allocate_system_memory(MemoryRegion *mr, Object *owner,
 347                                           const char *name,
 348                                           uint64_t ram_size)
 349 {
 350     uint64_t addr = 0;
 351     int i;
 352
 353     if (nb_numa_nodes == 0 || !have_memdevs) {
 354         allocate_system_memory_nonnuma(mr, owner, name, ram_size);
 355         return;
 356     }
 357
 358     memory_region_init(mr, owner, name, ram_size);
 359     for (i = 0; i < MAX_NODES; i++) {
 360         Error *local_err = NULL;
 361         uint64_t size = numa_info[i].node_mem;
 362         HostMemoryBackend *backend = numa_info[i].node_memdev;
 363         if (!backend) {
 364             continue;
 365         }
 366         MemoryRegion *seg = host_memory_backend_get_memory(backend, &local_err);
 367         if (local_err) {
 368             error_report_err(local_err);
 369             exit(1);
 370         }
 371
 372         if (memory_region_is_mapped(seg)) {
 373             char *path = object_get_canonical_path_component(OBJECT(backend));
 374             error_report("memory backend %s is used multiple times. Each "
 375                          "-numa option must use a different memdev value.",
 376                          path);
 377             exit(1);
 378         }
 379
 380         memory_region_add_subregion(mr, addr, seg);
 381         vmstate_register_ram_global(seg);
 382         addr += size;
 383     }
 384 }
 385
 386 static void numa_stat_memory_devices(uint64_t node_mem[])
 387 {
 388     MemoryDeviceInfoList *info_list = NULL;
 389     MemoryDeviceInfoList **prev = &info_list;
 390     MemoryDeviceInfoList *info;
 391
 392     qmp_pc_dimm_device_list(qdev_get_machine(), &prev);
 393     for (info = info_list; info; info = info->next) {
 394         MemoryDeviceInfo *value = info->value;
 395
 396         if (value) {
 397             switch (value->kind) {
 398             case MEMORY_DEVICE_INFO_KIND_DIMM:
 399                 node_mem[value->dimm->node] += value->dimm->size;
 400                 break;
 401             default:
 402                 break;
 403             }
 404         }
 405     }
 406     qapi_free_MemoryDeviceInfoList(info_list);
 407 }
 408
 409 void query_numa_node_mem(uint64_t node_mem[])
 410 {
 411     int i;
 412
 413     if (nb_numa_nodes <= 0) {
 414         return;
 415     }
 416
 417     numa_stat_memory_devices(node_mem);
 418     for (i = 0; i < nb_numa_nodes; i++) {
 419         node_mem[i] += numa_info[i].node_mem;
 420     }
 421 }
 422
 423 static int query_memdev(Object *obj, void *opaque)
 424 {
 425     MemdevList **list = opaque;
 426     MemdevList *m = NULL;
 427     Error *err = NULL;
 428
 429     if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) {
 430         m = g_malloc0(sizeof(*m));
 431
 432         m->value = g_malloc0(sizeof(*m->value));
 433
 434         m->value->size = object_property_get_int(obj, "size",
 435                                                  &err);
 436         if (err) {
 437             goto error;
 438         }
 439
 440         m->value->merge = object_property_get_bool(obj, "merge",
 441                                                    &err);
 442         if (err) {
 443             goto error;
 444         }
 445
 446         m->value->dump = object_property_get_bool(obj, "dump",
 447                                                   &err);
 448         if (err) {
 449             goto error;
 450         }
 451
 452         m->value->prealloc = object_property_get_bool(obj,
 453                                                       "prealloc", &err);
 454         if (err) {
 455             goto error;
 456         }
 457
 458         m->value->policy = object_property_get_enum(obj,
 459                                                     "policy",
 460                                                     HostMemPolicy_lookup,
 461                                                     &err);
 462         if (err) {
 463             goto error;
 464         }
 465
 466         object_property_get_uint16List(obj, "host-nodes",
 467                                        &m->value->host_nodes, &err);
 468         if (err) {
 469             goto error;
 470         }
 471
 472         m->next = *list;
 473         *list = m;
 474     }
 475
 476     return 0;
 477 error:
 478     g_free(m->value);
 479     g_free(m);
 480
 481     return -1;
 482 }
 483
 484 MemdevList *qmp_query_memdev(Error **errp)
 485 {
 486     Object *obj;
 487     MemdevList *list = NULL;
 488
 489     obj = object_resolve_path("/objects", NULL);
 490     if (obj == NULL) {
 491         return NULL;
 492     }
 493
 494     if (object_child_foreach(obj, query_memdev, &list) != 0) {
 495         goto error;
 496     }
 497
 498     return list;
 499
 500 error:
 501     qapi_free_MemdevList(list);
 502     return NULL;
 503 }