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     /*
  60      * Memory-less nodes can come here with 0 size in which case,
  61      * there is nothing to do.
  62      */
  63     if (!size) {
  64         return;
  65     }
  66
  67     range->mem_start = addr;
  68     range->mem_end = addr + size - 1;
  69     QLIST_INSERT_HEAD(&numa_info[node].addr, range, entry);
  70 }
  71
  72 void numa_unset_mem_node_id(ram_addr_t addr, uint64_t size, uint32_t node)
  73 {
  74     struct numa_addr_range *range, *next;
  75
  76     QLIST_FOREACH_SAFE(range, &numa_info[node].addr, entry, next) {
  77         if (addr == range->mem_start && (addr + size - 1) == range->mem_end) {
  78             QLIST_REMOVE(range, entry);
  79             g_free(range);
  80             return;
  81         }
  82     }
  83 }
  84
  85 static void numa_set_mem_ranges(void)
  86 {
  87     int i;
  88     ram_addr_t mem_start = 0;
  89
  90     /*
  91      * Deduce start address of each node and use it to store
  92      * the address range info in numa_info address range list
  93      */
  94     for (i = 0; i < nb_numa_nodes; i++) {
  95         numa_set_mem_node_id(mem_start, numa_info[i].node_mem, i);
  96         mem_start += numa_info[i].node_mem;
  97     }
  98 }
  99
 100 /*
 101  * Check if @addr falls under NUMA @node.
 102  */
 103 static bool numa_addr_belongs_to_node(ram_addr_t addr, uint32_t node)
 104 {
 105     struct numa_addr_range *range;
 106
 107     QLIST_FOREACH(range, &numa_info[node].addr, entry) {
 108         if (addr >= range->mem_start && addr <= range->mem_end) {
 109             return true;
 110         }
 111     }
 112     return false;
 113 }
 114
 115 /*
 116  * Given an address, return the index of the NUMA node to which the
 117  * address belongs to.
 118  */
 119 uint32_t numa_get_node(ram_addr_t addr, Error **errp)
 120 {
 121     uint32_t i;
 122
 123     /* For non NUMA configurations, check if the addr falls under node 0 */
 124     if (!nb_numa_nodes) {
 125         if (numa_addr_belongs_to_node(addr, 0)) {
 126             return 0;
 127         }
 128     }
 129
 130     for (i = 0; i < nb_numa_nodes; i++) {
 131         if (numa_addr_belongs_to_node(addr, i)) {
 132             return i;
 133         }
 134     }
 135
 136     error_setg(errp, "Address 0x" RAM_ADDR_FMT " doesn't belong to any "
 137                 "NUMA node", addr);
 138     return -1;
 139 }
 140
 141 static void numa_node_parse(NumaNodeOptions *node, QemuOpts *opts, Error **errp)
 142 {
 143     uint16_t nodenr;
 144     uint16List *cpus = NULL;
 145
 146     if (node->has_nodeid) {
 147         nodenr = node->nodeid;
 148     } else {
 149         nodenr = nb_numa_nodes;
 150     }
 151
 152     if (nodenr >= MAX_NODES) {
 153         error_setg(errp, "Max number of NUMA nodes reached: %"
 154                    PRIu16 "", nodenr);
 155         return;
 156     }
 157
 158     if (numa_info[nodenr].present) {
 159         error_setg(errp, "Duplicate NUMA nodeid: %" PRIu16, nodenr);
 160         return;
 161     }
 162
 163     for (cpus = node->cpus; cpus; cpus = cpus->next) {
 164         if (cpus->value >= max_cpus) {
 165             error_setg(errp,
 166                        "CPU index (%" PRIu16 ")"
 167                        " should be smaller than maxcpus (%d)",
 168                        cpus->value, max_cpus);
 169             return;
 170         }
 171         bitmap_set(numa_info[nodenr].node_cpu, cpus->value, 1);
 172     }
 173
 174     if (node->has_mem && node->has_memdev) {
 175         error_setg(errp, "qemu: cannot specify both mem= and memdev=");
 176         return;
 177     }
 178
 179     if (have_memdevs == -1) {
 180         have_memdevs = node->has_memdev;
 181     }
 182     if (node->has_memdev != have_memdevs) {
 183         error_setg(errp, "qemu: memdev option must be specified for either "
 184                    "all or no nodes");
 185         return;
 186     }
 187
 188     if (node->has_mem) {
 189         uint64_t mem_size = node->mem;
 190         const char *mem_str = qemu_opt_get(opts, "mem");
 191         /* Fix up legacy suffix-less format */
 192         if (g_ascii_isdigit(mem_str[strlen(mem_str) - 1])) {
 193             mem_size <<= 20;
 194         }
 195         numa_info[nodenr].node_mem = mem_size;
 196     }
 197     if (node->has_memdev) {
 198         Object *o;
 199         o = object_resolve_path_type(node->memdev, TYPE_MEMORY_BACKEND, NULL);
 200         if (!o) {
 201             error_setg(errp, "memdev=%s is ambiguous", node->memdev);
 202             return;
 203         }
 204
 205         object_ref(o);
 206         numa_info[nodenr].node_mem = object_property_get_int(o, "size", NULL);
 207         numa_info[nodenr].node_memdev = MEMORY_BACKEND(o);
 208     }
 209     numa_info[nodenr].present = true;
 210     max_numa_nodeid = MAX(max_numa_nodeid, nodenr + 1);
 211 }
 212
 213 static int parse_numa(void *opaque, QemuOpts *opts, Error **errp)
 214 {
 215     NumaOptions *object = NULL;
 216     Error *err = NULL;
 217
 218     {
 219         OptsVisitor *ov = opts_visitor_new(opts);
 220         visit_type_NumaOptions(opts_get_visitor(ov), &object, NULL, &err);
 221         opts_visitor_cleanup(ov);
 222     }
 223
 224     if (err) {
 225         goto error;
 226     }
 227
 228     switch (object->kind) {
 229     case NUMA_OPTIONS_KIND_NODE:
 230         numa_node_parse(object->node, opts, &err);
 231         if (err) {
 232             goto error;
 233         }
 234         nb_numa_nodes++;
 235         break;
 236     default:
 237         abort();
 238     }
 239
 240     return 0;
 241
 242 error:
 243     error_report_err(err);
 244
 245     if (object) {
 246         QapiDeallocVisitor *dv = qapi_dealloc_visitor_new();
 247         visit_type_NumaOptions(qapi_dealloc_get_visitor(dv),
 248                                &object, NULL, NULL);
 249         qapi_dealloc_visitor_cleanup(dv);
 250     }
 251
 252     return -1;
 253 }
 254
 255 static char *enumerate_cpus(unsigned long *cpus, int max_cpus)
 256 {
 257     int cpu;
 258     bool first = true;
 259     GString *s = g_string_new(NULL);
 260
 261     for (cpu = find_first_bit(cpus, max_cpus);
 262         cpu < max_cpus;
 263         cpu = find_next_bit(cpus, max_cpus, cpu + 1)) {
 264         g_string_append_printf(s, "%s%d", first ? "" : " ", cpu);
 265         first = false;
 266     }
 267     return g_string_free(s, FALSE);
 268 }
 269
 270 static void validate_numa_cpus(void)
 271 {
 272     int i;
 273     DECLARE_BITMAP(seen_cpus, MAX_CPUMASK_BITS);
 274
 275     bitmap_zero(seen_cpus, MAX_CPUMASK_BITS);
 276     for (i = 0; i < nb_numa_nodes; i++) {
 277         if (bitmap_intersects(seen_cpus, numa_info[i].node_cpu,
 278                               MAX_CPUMASK_BITS)) {
 279             bitmap_and(seen_cpus, seen_cpus,
 280                        numa_info[i].node_cpu, MAX_CPUMASK_BITS);
 281             error_report("CPU(s) present in multiple NUMA nodes: %s",
 282                          enumerate_cpus(seen_cpus, max_cpus));;
 283             exit(EXIT_FAILURE);
 284         }
 285         bitmap_or(seen_cpus, seen_cpus,
 286                   numa_info[i].node_cpu, MAX_CPUMASK_BITS);
 287     }
 288
 289     if (!bitmap_full(seen_cpus, max_cpus)) {
 290         char *msg;
 291         bitmap_complement(seen_cpus, seen_cpus, max_cpus);
 292         msg = enumerate_cpus(seen_cpus, max_cpus);
 293         error_report("warning: CPU(s) not present in any NUMA nodes: %s", msg);
 294         error_report("warning: All CPU(s) up to maxcpus should be described "
 295                      "in NUMA config");
 296         g_free(msg);
 297     }
 298 }
 299
 300 void parse_numa_opts(MachineClass *mc)
 301 {
 302     int i;
 303
 304     if (qemu_opts_foreach(qemu_find_opts("numa"), parse_numa, NULL, NULL)) {
 305         exit(1);
 306     }
 307
 308     assert(max_numa_nodeid <= MAX_NODES);
 309
 310     /* No support for sparse NUMA node IDs yet: */
 311     for (i = max_numa_nodeid - 1; i >= 0; i--) {
 312         /* Report large node IDs first, to make mistakes easier to spot */
 313         if (!numa_info[i].present) {
 314             error_report("numa: Node ID missing: %d", i);
 315             exit(1);
 316         }
 317     }
 318
 319     /* This must be always true if all nodes are present: */
 320     assert(nb_numa_nodes == max_numa_nodeid);
 321
 322     if (nb_numa_nodes > 0) {
 323         uint64_t numa_total;
 324
 325         if (nb_numa_nodes > MAX_NODES) {
 326             nb_numa_nodes = MAX_NODES;
 327         }
 328
 329         /* If no memory size is given for any node, assume the default case
 330          * and distribute the available memory equally across all nodes
 331          */
 332         for (i = 0; i < nb_numa_nodes; i++) {
 333             if (numa_info[i].node_mem != 0) {
 334                 break;
 335             }
 336         }
 337         if (i == nb_numa_nodes) {
 338             uint64_t usedmem = 0;
 339
 340             /* On Linux, each node's border has to be 8MB aligned,
 341              * the final node gets the rest.
 342              */
 343             for (i = 0; i < nb_numa_nodes - 1; i++) {
 344                 numa_info[i].node_mem = (ram_size / nb_numa_nodes) &
 345                                         ~((1 << 23UL) - 1);
 346                 usedmem += numa_info[i].node_mem;
 347             }
 348             numa_info[i].node_mem = ram_size - usedmem;
 349         }
 350
 351         numa_total = 0;
 352         for (i = 0; i < nb_numa_nodes; i++) {
 353             numa_total += numa_info[i].node_mem;
 354         }
 355         if (numa_total != ram_size) {
 356             error_report("total memory for NUMA nodes (0x%" PRIx64 ")"
 357                          " should equal RAM size (0x" RAM_ADDR_FMT ")",
 358                          numa_total, ram_size);
 359             exit(1);
 360         }
 361
 362         for (i = 0; i < nb_numa_nodes; i++) {
 363             QLIST_INIT(&numa_info[i].addr);
 364         }
 365
 366         numa_set_mem_ranges();
 367
 368         for (i = 0; i < nb_numa_nodes; i++) {
 369             if (!bitmap_empty(numa_info[i].node_cpu, MAX_CPUMASK_BITS)) {
 370                 break;
 371             }
 372         }
 373         /* Historically VCPUs were assigned in round-robin order to NUMA
 374          * nodes. However it causes issues with guest not handling it nice
 375          * in case where cores/threads from a multicore CPU appear on
 376          * different nodes. So allow boards to override default distribution
 377          * rule grouping VCPUs by socket so that VCPUs from the same socket
 378          * would be on the same node.
 379          */
 380         if (i == nb_numa_nodes) {
 381             for (i = 0; i < max_cpus; i++) {
 382                 unsigned node_id = i % nb_numa_nodes;
 383                 if (mc->cpu_index_to_socket_id) {
 384                     node_id = mc->cpu_index_to_socket_id(i) % nb_numa_nodes;
 385                 }
 386
 387                 set_bit(i, numa_info[node_id].node_cpu);
 388             }
 389         }
 390
 391         validate_numa_cpus();
 392     } else {
 393         numa_set_mem_node_id(0, ram_size, 0);
 394     }
 395 }
 396
 397 void numa_post_machine_init(void)
 398 {
 399     CPUState *cpu;
 400     int i;
 401
 402     CPU_FOREACH(cpu) {
 403         for (i = 0; i < nb_numa_nodes; i++) {
 404             if (test_bit(cpu->cpu_index, numa_info[i].node_cpu)) {
 405                 cpu->numa_node = i;
 406             }
 407         }
 408     }
 409 }
 410
 411 static void allocate_system_memory_nonnuma(MemoryRegion *mr, Object *owner,
 412                                            const char *name,
 413                                            uint64_t ram_size)
 414 {
 415     if (mem_path) {
 416 #ifdef __linux__
 417         Error *err = NULL;
 418         memory_region_init_ram_from_file(mr, owner, name, ram_size, false,
 419                                          mem_path, &err);
 420
 421         /* Legacy behavior: if allocation failed, fall back to
 422          * regular RAM allocation.
 423          */
 424         if (err) {
 425             error_report_err(err);
 426             memory_region_init_ram(mr, owner, name, ram_size, &error_abort);
 427         }
 428 #else
 429         fprintf(stderr, "-mem-path not supported on this host\n");
 430         exit(1);
 431 #endif
 432     } else {
 433         memory_region_init_ram(mr, owner, name, ram_size, &error_abort);
 434     }
 435     vmstate_register_ram_global(mr);
 436 }
 437
 438 void memory_region_allocate_system_memory(MemoryRegion *mr, Object *owner,
 439                                           const char *name,
 440                                           uint64_t ram_size)
 441 {
 442     uint64_t addr = 0;
 443     int i;
 444
 445     if (nb_numa_nodes == 0 || !have_memdevs) {
 446         allocate_system_memory_nonnuma(mr, owner, name, ram_size);
 447         return;
 448     }
 449
 450     memory_region_init(mr, owner, name, ram_size);
 451     for (i = 0; i < MAX_NODES; i++) {
 452         Error *local_err = NULL;
 453         uint64_t size = numa_info[i].node_mem;
 454         HostMemoryBackend *backend = numa_info[i].node_memdev;
 455         if (!backend) {
 456             continue;
 457         }
 458         MemoryRegion *seg = host_memory_backend_get_memory(backend, &local_err);
 459         if (local_err) {
 460             error_report_err(local_err);
 461             exit(1);
 462         }
 463
 464         if (memory_region_is_mapped(seg)) {
 465             char *path = object_get_canonical_path_component(OBJECT(backend));
 466             error_report("memory backend %s is used multiple times. Each "
 467                          "-numa option must use a different memdev value.",
 468                          path);
 469             exit(1);
 470         }
 471
 472         memory_region_add_subregion(mr, addr, seg);
 473         vmstate_register_ram_global(seg);
 474         addr += size;
 475     }
 476 }
 477
 478 static void numa_stat_memory_devices(uint64_t node_mem[])
 479 {
 480     MemoryDeviceInfoList *info_list = NULL;
 481     MemoryDeviceInfoList **prev = &info_list;
 482     MemoryDeviceInfoList *info;
 483
 484     qmp_pc_dimm_device_list(qdev_get_machine(), &prev);
 485     for (info = info_list; info; info = info->next) {
 486         MemoryDeviceInfo *value = info->value;
 487
 488         if (value) {
 489             switch (value->kind) {
 490             case MEMORY_DEVICE_INFO_KIND_DIMM:
 491                 node_mem[value->dimm->node] += value->dimm->size;
 492                 break;
 493             default:
 494                 break;
 495             }
 496         }
 497     }
 498     qapi_free_MemoryDeviceInfoList(info_list);
 499 }
 500
 501 void query_numa_node_mem(uint64_t node_mem[])
 502 {
 503     int i;
 504
 505     if (nb_numa_nodes <= 0) {
 506         return;
 507     }
 508
 509     numa_stat_memory_devices(node_mem);
 510     for (i = 0; i < nb_numa_nodes; i++) {
 511         node_mem[i] += numa_info[i].node_mem;
 512     }
 513 }
 514
 515 static int query_memdev(Object *obj, void *opaque)
 516 {
 517     MemdevList **list = opaque;
 518     MemdevList *m = NULL;
 519     Error *err = NULL;
 520
 521     if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) {
 522         m = g_malloc0(sizeof(*m));
 523
 524         m->value = g_malloc0(sizeof(*m->value));
 525
 526         m->value->size = object_property_get_int(obj, "size",
 527                                                  &err);
 528         if (err) {
 529             goto error;
 530         }
 531
 532         m->value->merge = object_property_get_bool(obj, "merge",
 533                                                    &err);
 534         if (err) {
 535             goto error;
 536         }
 537
 538         m->value->dump = object_property_get_bool(obj, "dump",
 539                                                   &err);
 540         if (err) {
 541             goto error;
 542         }
 543
 544         m->value->prealloc = object_property_get_bool(obj,
 545                                                       "prealloc", &err);
 546         if (err) {
 547             goto error;
 548         }
 549
 550         m->value->policy = object_property_get_enum(obj,
 551                                                     "policy",
 552                                                     "HostMemPolicy",
 553                                                     &err);
 554         if (err) {
 555             goto error;
 556         }
 557
 558         object_property_get_uint16List(obj, "host-nodes",
 559                                        &m->value->host_nodes, &err);
 560         if (err) {
 561             goto error;
 562         }
 563
 564         m->next = *list;
 565         *list = m;
 566     }
 567
 568     return 0;
 569 error:
 570     g_free(m->value);
 571     g_free(m);
 572
 573     return -1;
 574 }
 575
 576 MemdevList *qmp_query_memdev(Error **errp)
 577 {
 578     Object *obj;
 579     MemdevList *list = NULL;
 580
 581     obj = object_get_objects_root();
 582     if (obj == NULL) {
 583         return NULL;
 584     }
 585
 586     if (object_child_foreach(obj, query_memdev, &list) != 0) {
 587         goto error;
 588     }
 589
 590     return list;
 591
 592 error:
 593     qapi_free_MemdevList(list);
 594     return NULL;
 595 }