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