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 void numa_set_mem_node_id(ram_addr_t addr, uint64_t size, uint32_t node)
  56 {
  57     struct numa_addr_range *range = g_malloc0(sizeof(*range));
  58
  59     range->mem_start = addr;
  60     range->mem_end = addr + size - 1;
  61     QLIST_INSERT_HEAD(&numa_info[node].addr, range, entry);
  62 }
  63
  64 void numa_unset_mem_node_id(ram_addr_t addr, uint64_t size, uint32_t node)
  65 {
  66     struct numa_addr_range *range, *next;
  67
  68     QLIST_FOREACH_SAFE(range, &numa_info[node].addr, entry, next) {
  69         if (addr == range->mem_start && (addr + size - 1) == range->mem_end) {
  70             QLIST_REMOVE(range, entry);
  71             g_free(range);
  72             return;
  73         }
  74     }
  75 }
  76
  77 static void numa_node_parse(NumaNodeOptions *node, QemuOpts *opts, Error **errp)
  78 {
  79     uint16_t nodenr;
  80     uint16List *cpus = NULL;
  81
  82     if (node->has_nodeid) {
  83         nodenr = node->nodeid;
  84     } else {
  85         nodenr = nb_numa_nodes;
  86     }
  87
  88     if (nodenr >= MAX_NODES) {
  89         error_setg(errp, "Max number of NUMA nodes reached: %"
  90                    PRIu16 "", nodenr);
  91         return;
  92     }
  93
  94     if (numa_info[nodenr].present) {
  95         error_setg(errp, "Duplicate NUMA nodeid: %" PRIu16, nodenr);
  96         return;
  97     }
  98
  99     for (cpus = node->cpus; cpus; cpus = cpus->next) {
 100         if (cpus->value >= max_cpus) {
 101             error_setg(errp,
 102                        "CPU index (%" PRIu16 ")"
 103                        " should be smaller than maxcpus (%d)",
 104                        cpus->value, max_cpus);
 105             return;
 106         }
 107         bitmap_set(numa_info[nodenr].node_cpu, cpus->value, 1);
 108     }
 109
 110     if (node->has_mem && node->has_memdev) {
 111         error_setg(errp, "qemu: cannot specify both mem= and memdev=");
 112         return;
 113     }
 114
 115     if (have_memdevs == -1) {
 116         have_memdevs = node->has_memdev;
 117     }
 118     if (node->has_memdev != have_memdevs) {
 119         error_setg(errp, "qemu: memdev option must be specified for either "
 120                    "all or no nodes");
 121         return;
 122     }
 123
 124     if (node->has_mem) {
 125         uint64_t mem_size = node->mem;
 126         const char *mem_str = qemu_opt_get(opts, "mem");
 127         /* Fix up legacy suffix-less format */
 128         if (g_ascii_isdigit(mem_str[strlen(mem_str) - 1])) {
 129             mem_size <<= 20;
 130         }
 131         numa_info[nodenr].node_mem = mem_size;
 132     }
 133     if (node->has_memdev) {
 134         Object *o;
 135         o = object_resolve_path_type(node->memdev, TYPE_MEMORY_BACKEND, NULL);
 136         if (!o) {
 137             error_setg(errp, "memdev=%s is ambiguous", node->memdev);
 138             return;
 139         }
 140
 141         object_ref(o);
 142         numa_info[nodenr].node_mem = object_property_get_int(o, "size", NULL);
 143         numa_info[nodenr].node_memdev = MEMORY_BACKEND(o);
 144     }
 145     numa_info[nodenr].present = true;
 146     max_numa_nodeid = MAX(max_numa_nodeid, nodenr + 1);
 147 }
 148
 149 static int parse_numa(void *opaque, QemuOpts *opts, Error **errp)
 150 {
 151     NumaOptions *object = NULL;
 152     Error *err = NULL;
 153
 154     {
 155         OptsVisitor *ov = opts_visitor_new(opts);
 156         visit_type_NumaOptions(opts_get_visitor(ov), &object, NULL, &err);
 157         opts_visitor_cleanup(ov);
 158     }
 159
 160     if (err) {
 161         goto error;
 162     }
 163
 164     switch (object->kind) {
 165     case NUMA_OPTIONS_KIND_NODE:
 166         numa_node_parse(object->node, opts, &err);
 167         if (err) {
 168             goto error;
 169         }
 170         nb_numa_nodes++;
 171         break;
 172     default:
 173         abort();
 174     }
 175
 176     return 0;
 177
 178 error:
 179     error_report_err(err);
 180
 181     if (object) {
 182         QapiDeallocVisitor *dv = qapi_dealloc_visitor_new();
 183         visit_type_NumaOptions(qapi_dealloc_get_visitor(dv),
 184                                &object, NULL, NULL);
 185         qapi_dealloc_visitor_cleanup(dv);
 186     }
 187
 188     return -1;
 189 }
 190
 191 static char *enumerate_cpus(unsigned long *cpus, int max_cpus)
 192 {
 193     int cpu;
 194     bool first = true;
 195     GString *s = g_string_new(NULL);
 196
 197     for (cpu = find_first_bit(cpus, max_cpus);
 198         cpu < max_cpus;
 199         cpu = find_next_bit(cpus, max_cpus, cpu + 1)) {
 200         g_string_append_printf(s, "%s%d", first ? "" : " ", cpu);
 201         first = false;
 202     }
 203     return g_string_free(s, FALSE);
 204 }
 205
 206 static void validate_numa_cpus(void)
 207 {
 208     int i;
 209     DECLARE_BITMAP(seen_cpus, MAX_CPUMASK_BITS);
 210
 211     bitmap_zero(seen_cpus, MAX_CPUMASK_BITS);
 212     for (i = 0; i < nb_numa_nodes; i++) {
 213         if (bitmap_intersects(seen_cpus, numa_info[i].node_cpu,
 214                               MAX_CPUMASK_BITS)) {
 215             bitmap_and(seen_cpus, seen_cpus,
 216                        numa_info[i].node_cpu, MAX_CPUMASK_BITS);
 217             error_report("CPU(s) present in multiple NUMA nodes: %s",
 218                          enumerate_cpus(seen_cpus, max_cpus));;
 219             exit(EXIT_FAILURE);
 220         }
 221         bitmap_or(seen_cpus, seen_cpus,
 222                   numa_info[i].node_cpu, MAX_CPUMASK_BITS);
 223     }
 224
 225     if (!bitmap_full(seen_cpus, max_cpus)) {
 226         char *msg;
 227         bitmap_complement(seen_cpus, seen_cpus, max_cpus);
 228         msg = enumerate_cpus(seen_cpus, max_cpus);
 229         error_report("warning: CPU(s) not present in any NUMA nodes: %s", msg);
 230         error_report("warning: All CPU(s) up to maxcpus should be described "
 231                      "in NUMA config");
 232         g_free(msg);
 233     }
 234 }
 235
 236 void parse_numa_opts(MachineClass *mc)
 237 {
 238     int i;
 239
 240     if (qemu_opts_foreach(qemu_find_opts("numa"), parse_numa, NULL, NULL)) {
 241         exit(1);
 242     }
 243
 244     assert(max_numa_nodeid <= MAX_NODES);
 245
 246     /* No support for sparse NUMA node IDs yet: */
 247     for (i = max_numa_nodeid - 1; i >= 0; i--) {
 248         /* Report large node IDs first, to make mistakes easier to spot */
 249         if (!numa_info[i].present) {
 250             error_report("numa: Node ID missing: %d", i);
 251             exit(1);
 252         }
 253     }
 254
 255     /* This must be always true if all nodes are present: */
 256     assert(nb_numa_nodes == max_numa_nodeid);
 257
 258     if (nb_numa_nodes > 0) {
 259         uint64_t numa_total;
 260
 261         if (nb_numa_nodes > MAX_NODES) {
 262             nb_numa_nodes = MAX_NODES;
 263         }
 264
 265         /* If no memory size is given for any node, assume the default case
 266          * and distribute the available memory equally across all nodes
 267          */
 268         for (i = 0; i < nb_numa_nodes; i++) {
 269             if (numa_info[i].node_mem != 0) {
 270                 break;
 271             }
 272         }
 273         if (i == nb_numa_nodes) {
 274             uint64_t usedmem = 0;
 275
 276             /* On Linux, each node's border has to be 8MB aligned,
 277              * the final node gets the rest.
 278              */
 279             for (i = 0; i < nb_numa_nodes - 1; i++) {
 280                 numa_info[i].node_mem = (ram_size / nb_numa_nodes) &
 281                                         ~((1 << 23UL) - 1);
 282                 usedmem += numa_info[i].node_mem;
 283             }
 284             numa_info[i].node_mem = ram_size - usedmem;
 285         }
 286
 287         numa_total = 0;
 288         for (i = 0; i < nb_numa_nodes; i++) {
 289             numa_total += numa_info[i].node_mem;
 290         }
 291         if (numa_total != ram_size) {
 292             error_report("total memory for NUMA nodes (0x%" PRIx64 ")"
 293                          " should equal RAM size (0x" RAM_ADDR_FMT ")",
 294                          numa_total, ram_size);
 295             exit(1);
 296         }
 297
 298         for (i = 0; i < nb_numa_nodes; i++) {
 299             QLIST_INIT(&numa_info[i].addr);
 300         }
 301
 302         for (i = 0; i < nb_numa_nodes; i++) {
 303             if (!bitmap_empty(numa_info[i].node_cpu, MAX_CPUMASK_BITS)) {
 304                 break;
 305             }
 306         }
 307         /* Historically VCPUs were assigned in round-robin order to NUMA
 308          * nodes. However it causes issues with guest not handling it nice
 309          * in case where cores/threads from a multicore CPU appear on
 310          * different nodes. So allow boards to override default distribution
 311          * rule grouping VCPUs by socket so that VCPUs from the same socket
 312          * would be on the same node.
 313          */
 314         if (i == nb_numa_nodes) {
 315             for (i = 0; i < max_cpus; i++) {
 316                 unsigned node_id = i % nb_numa_nodes;
 317                 if (mc->cpu_index_to_socket_id) {
 318                     node_id = mc->cpu_index_to_socket_id(i) % nb_numa_nodes;
 319                 }
 320
 321                 set_bit(i, numa_info[node_id].node_cpu);
 322             }
 323         }
 324
 325         validate_numa_cpus();
 326     }
 327 }
 328
 329 void numa_post_machine_init(void)
 330 {
 331     CPUState *cpu;
 332     int i;
 333
 334     CPU_FOREACH(cpu) {
 335         for (i = 0; i < nb_numa_nodes; i++) {
 336             if (test_bit(cpu->cpu_index, numa_info[i].node_cpu)) {
 337                 cpu->numa_node = i;
 338             }
 339         }
 340     }
 341 }
 342
 343 static void allocate_system_memory_nonnuma(MemoryRegion *mr, Object *owner,
 344                                            const char *name,
 345                                            uint64_t ram_size)
 346 {
 347     if (mem_path) {
 348 #ifdef __linux__
 349         Error *err = NULL;
 350         memory_region_init_ram_from_file(mr, owner, name, ram_size, false,
 351                                          mem_path, &err);
 352
 353         /* Legacy behavior: if allocation failed, fall back to
 354          * regular RAM allocation.
 355          */
 356         if (err) {
 357             error_report_err(err);
 358             memory_region_init_ram(mr, owner, name, ram_size, &error_abort);
 359         }
 360 #else
 361         fprintf(stderr, "-mem-path not supported on this host\n");
 362         exit(1);
 363 #endif
 364     } else {
 365         memory_region_init_ram(mr, owner, name, ram_size, &error_abort);
 366     }
 367     vmstate_register_ram_global(mr);
 368 }
 369
 370 void memory_region_allocate_system_memory(MemoryRegion *mr, Object *owner,
 371                                           const char *name,
 372                                           uint64_t ram_size)
 373 {
 374     uint64_t addr = 0;
 375     int i;
 376
 377     if (nb_numa_nodes == 0 || !have_memdevs) {
 378         allocate_system_memory_nonnuma(mr, owner, name, ram_size);
 379         return;
 380     }
 381
 382     memory_region_init(mr, owner, name, ram_size);
 383     for (i = 0; i < MAX_NODES; i++) {
 384         Error *local_err = NULL;
 385         uint64_t size = numa_info[i].node_mem;
 386         HostMemoryBackend *backend = numa_info[i].node_memdev;
 387         if (!backend) {
 388             continue;
 389         }
 390         MemoryRegion *seg = host_memory_backend_get_memory(backend, &local_err);
 391         if (local_err) {
 392             error_report_err(local_err);
 393             exit(1);
 394         }
 395
 396         if (memory_region_is_mapped(seg)) {
 397             char *path = object_get_canonical_path_component(OBJECT(backend));
 398             error_report("memory backend %s is used multiple times. Each "
 399                          "-numa option must use a different memdev value.",
 400                          path);
 401             exit(1);
 402         }
 403
 404         memory_region_add_subregion(mr, addr, seg);
 405         vmstate_register_ram_global(seg);
 406         addr += size;
 407     }
 408 }
 409
 410 static void numa_stat_memory_devices(uint64_t node_mem[])
 411 {
 412     MemoryDeviceInfoList *info_list = NULL;
 413     MemoryDeviceInfoList **prev = &info_list;
 414     MemoryDeviceInfoList *info;
 415
 416     qmp_pc_dimm_device_list(qdev_get_machine(), &prev);
 417     for (info = info_list; info; info = info->next) {
 418         MemoryDeviceInfo *value = info->value;
 419
 420         if (value) {
 421             switch (value->kind) {
 422             case MEMORY_DEVICE_INFO_KIND_DIMM:
 423                 node_mem[value->dimm->node] += value->dimm->size;
 424                 break;
 425             default:
 426                 break;
 427             }
 428         }
 429     }
 430     qapi_free_MemoryDeviceInfoList(info_list);
 431 }
 432
 433 void query_numa_node_mem(uint64_t node_mem[])
 434 {
 435     int i;
 436
 437     if (nb_numa_nodes <= 0) {
 438         return;
 439     }
 440
 441     numa_stat_memory_devices(node_mem);
 442     for (i = 0; i < nb_numa_nodes; i++) {
 443         node_mem[i] += numa_info[i].node_mem;
 444     }
 445 }
 446
 447 static int query_memdev(Object *obj, void *opaque)
 448 {
 449     MemdevList **list = opaque;
 450     MemdevList *m = NULL;
 451     Error *err = NULL;
 452
 453     if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) {
 454         m = g_malloc0(sizeof(*m));
 455
 456         m->value = g_malloc0(sizeof(*m->value));
 457
 458         m->value->size = object_property_get_int(obj, "size",
 459                                                  &err);
 460         if (err) {
 461             goto error;
 462         }
 463
 464         m->value->merge = object_property_get_bool(obj, "merge",
 465                                                    &err);
 466         if (err) {
 467             goto error;
 468         }
 469
 470         m->value->dump = object_property_get_bool(obj, "dump",
 471                                                   &err);
 472         if (err) {
 473             goto error;
 474         }
 475
 476         m->value->prealloc = object_property_get_bool(obj,
 477                                                       "prealloc", &err);
 478         if (err) {
 479             goto error;
 480         }
 481
 482         m->value->policy = object_property_get_enum(obj,
 483                                                     "policy",
 484                                                     "HostMemPolicy",
 485                                                     &err);
 486         if (err) {
 487             goto error;
 488         }
 489
 490         object_property_get_uint16List(obj, "host-nodes",
 491                                        &m->value->host_nodes, &err);
 492         if (err) {
 493             goto error;
 494         }
 495
 496         m->next = *list;
 497         *list = m;
 498     }
 499
 500     return 0;
 501 error:
 502     g_free(m->value);
 503     g_free(m);
 504
 505     return -1;
 506 }
 507
 508 MemdevList *qmp_query_memdev(Error **errp)
 509 {
 510     Object *obj;
 511     MemdevList *list = NULL;
 512
 513     obj = object_get_objects_root();
 514     if (obj == NULL) {
 515         return NULL;
 516     }
 517
 518     if (object_child_foreach(obj, query_memdev, &list) != 0) {
 519         goto error;
 520     }
 521
 522     return list;
 523
 524 error:
 525     qapi_free_MemdevList(list);
 526     return NULL;
 527 }