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 "qemu/osdep.h"
  26 #include "sysemu/numa.h"
  27 #include "exec/cpu-common.h"
  28 #include "qemu/bitmap.h"
  29 #include "qom/cpu.h"
  30 #include "qemu/error-report.h"
  31 #include "include/exec/cpu-common.h" /* for RAM_ADDR_FMT */
  32 #include "qapi-visit.h"
  33 #include "qapi/opts-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;
  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 = g_malloc0(sizeof(*range));
  68     range->mem_start = addr;
  69     range->mem_end = addr + size - 1;
  70     QLIST_INSERT_HEAD(&numa_info[node].addr, range, entry);
  71 }
  72
  73 void numa_unset_mem_node_id(ram_addr_t addr, uint64_t size, uint32_t node)
  74 {
  75     struct numa_addr_range *range, *next;
  76
  77     QLIST_FOREACH_SAFE(range, &numa_info[node].addr, entry, next) {
  78         if (addr == range->mem_start && (addr + size - 1) == range->mem_end) {
  79             QLIST_REMOVE(range, entry);
  80             g_free(range);
  81             return;
  82         }
  83     }
  84 }
  85
  86 static void numa_set_mem_ranges(void)
  87 {
  88     int i;
  89     ram_addr_t mem_start = 0;
  90
  91     /*
  92      * Deduce start address of each node and use it to store
  93      * the address range info in numa_info address range list
  94      */
  95     for (i = 0; i < nb_numa_nodes; i++) {
  96         numa_set_mem_node_id(mem_start, numa_info[i].node_mem, i);
  97         mem_start += numa_info[i].node_mem;
  98     }
  99 }
 100
 101 /*
 102  * Check if @addr falls under NUMA @node.
 103  */
 104 static bool numa_addr_belongs_to_node(ram_addr_t addr, uint32_t node)
 105 {
 106     struct numa_addr_range *range;
 107
 108     QLIST_FOREACH(range, &numa_info[node].addr, entry) {
 109         if (addr >= range->mem_start && addr <= range->mem_end) {
 110             return true;
 111         }
 112     }
 113     return false;
 114 }
 115
 116 /*
 117  * Given an address, return the index of the NUMA node to which the
 118  * address belongs to.
 119  */
 120 uint32_t numa_get_node(ram_addr_t addr, Error **errp)
 121 {
 122     uint32_t i;
 123
 124     /* For non NUMA configurations, check if the addr falls under node 0 */
 125     if (!nb_numa_nodes) {
 126         if (numa_addr_belongs_to_node(addr, 0)) {
 127             return 0;
 128         }
 129     }
 130
 131     for (i = 0; i < nb_numa_nodes; i++) {
 132         if (numa_addr_belongs_to_node(addr, i)) {
 133             return i;
 134         }
 135     }
 136
 137     error_setg(errp, "Address 0x" RAM_ADDR_FMT " doesn't belong to any "
 138                 "NUMA node", addr);
 139     return -1;
 140 }
 141
 142 static void numa_node_parse(NumaNodeOptions *node, QemuOpts *opts, Error **errp)
 143 {
 144     uint16_t nodenr;
 145     uint16List *cpus = NULL;
 146
 147     if (node->has_nodeid) {
 148         nodenr = node->nodeid;
 149     } else {
 150         nodenr = nb_numa_nodes;
 151     }
 152
 153     if (nodenr >= MAX_NODES) {
 154         error_setg(errp, "Max number of NUMA nodes reached: %"
 155                    PRIu16 "", nodenr);
 156         return;
 157     }
 158
 159     if (numa_info[nodenr].present) {
 160         error_setg(errp, "Duplicate NUMA nodeid: %" PRIu16, nodenr);
 161         return;
 162     }
 163
 164     for (cpus = node->cpus; cpus; cpus = cpus->next) {
 165         if (cpus->value >= max_cpus) {
 166             error_setg(errp,
 167                        "CPU index (%" PRIu16 ")"
 168                        " should be smaller than maxcpus (%d)",
 169                        cpus->value, max_cpus);
 170             return;
 171         }
 172         bitmap_set(numa_info[nodenr].node_cpu, cpus->value, 1);
 173     }
 174
 175     if (node->has_mem && node->has_memdev) {
 176         error_setg(errp, "qemu: cannot specify both mem= and memdev=");
 177         return;
 178     }
 179
 180     if (have_memdevs == -1) {
 181         have_memdevs = node->has_memdev;
 182     }
 183     if (node->has_memdev != have_memdevs) {
 184         error_setg(errp, "qemu: memdev option must be specified for either "
 185                    "all or no nodes");
 186         return;
 187     }
 188
 189     if (node->has_mem) {
 190         uint64_t mem_size = node->mem;
 191         const char *mem_str = qemu_opt_get(opts, "mem");
 192         /* Fix up legacy suffix-less format */
 193         if (g_ascii_isdigit(mem_str[strlen(mem_str) - 1])) {
 194             mem_size <<= 20;
 195         }
 196         numa_info[nodenr].node_mem = mem_size;
 197     }
 198     if (node->has_memdev) {
 199         Object *o;
 200         o = object_resolve_path_type(node->memdev, TYPE_MEMORY_BACKEND, NULL);
 201         if (!o) {
 202             error_setg(errp, "memdev=%s is ambiguous", node->memdev);
 203             return;
 204         }
 205
 206         object_ref(o);
 207         numa_info[nodenr].node_mem = object_property_get_int(o, "size", NULL);
 208         numa_info[nodenr].node_memdev = MEMORY_BACKEND(o);
 209     }
 210     numa_info[nodenr].present = true;
 211     max_numa_nodeid = MAX(max_numa_nodeid, nodenr + 1);
 212 }
 213
 214 static int parse_numa(void *opaque, QemuOpts *opts, Error **errp)
 215 {
 216     NumaOptions *object = NULL;
 217     Error *err = NULL;
 218
 219     {
 220         Visitor *v = opts_visitor_new(opts);
 221         visit_type_NumaOptions(v, NULL, &object, &err);
 222         visit_free(v);
 223     }
 224
 225     if (err) {
 226         goto end;
 227     }
 228
 229     switch (object->type) {
 230     case NUMA_OPTIONS_KIND_NODE:
 231         numa_node_parse(object->u.node.data, opts, &err);
 232         if (err) {
 233             goto end;
 234         }
 235         nb_numa_nodes++;
 236         break;
 237     default:
 238         abort();
 239     }
 240
 241 end:
 242     qapi_free_NumaOptions(object);
 243     if (err) {
 244         error_report_err(err);
 245         return -1;
 246     }
 247
 248     return 0;
 249 }
 250
 251 static char *enumerate_cpus(unsigned long *cpus, int max_cpus)
 252 {
 253     int cpu;
 254     bool first = true;
 255     GString *s = g_string_new(NULL);
 256
 257     for (cpu = find_first_bit(cpus, max_cpus);
 258         cpu < max_cpus;
 259         cpu = find_next_bit(cpus, max_cpus, cpu + 1)) {
 260         g_string_append_printf(s, "%s%d", first ? "" : " ", cpu);
 261         first = false;
 262     }
 263     return g_string_free(s, FALSE);
 264 }
 265
 266 static void validate_numa_cpus(void)
 267 {
 268     int i;
 269     unsigned long *seen_cpus = bitmap_new(max_cpus);
 270
 271     for (i = 0; i < nb_numa_nodes; i++) {
 272         if (bitmap_intersects(seen_cpus, numa_info[i].node_cpu, max_cpus)) {
 273             bitmap_and(seen_cpus, seen_cpus,
 274                        numa_info[i].node_cpu, max_cpus);
 275             error_report("CPU(s) present in multiple NUMA nodes: %s",
 276                          enumerate_cpus(seen_cpus, max_cpus));
 277             g_free(seen_cpus);
 278             exit(EXIT_FAILURE);
 279         }
 280         bitmap_or(seen_cpus, seen_cpus,
 281                   numa_info[i].node_cpu, max_cpus);
 282     }
 283
 284     if (!bitmap_full(seen_cpus, max_cpus)) {
 285         char *msg;
 286         bitmap_complement(seen_cpus, seen_cpus, max_cpus);
 287         msg = enumerate_cpus(seen_cpus, max_cpus);
 288         error_report("warning: CPU(s) not present in any NUMA nodes: %s", msg);
 289         error_report("warning: All CPU(s) up to maxcpus should be described "
 290                      "in NUMA config");
 291         g_free(msg);
 292     }
 293     g_free(seen_cpus);
 294 }
 295
 296 void parse_numa_opts(MachineClass *mc)
 297 {
 298     int i;
 299
 300     for (i = 0; i < MAX_NODES; i++) {
 301         numa_info[i].node_cpu = bitmap_new(max_cpus);
 302     }
 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_cpus)) {
 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             assert(cpu->cpu_index < max_cpus);
 405             if (test_bit(cpu->cpu_index, numa_info[i].node_cpu)) {
 406                 cpu->numa_node = i;
 407             }
 408         }
 409     }
 410 }
 411
 412 static void allocate_system_memory_nonnuma(MemoryRegion *mr, Object *owner,
 413                                            const char *name,
 414                                            uint64_t ram_size)
 415 {
 416     if (mem_path) {
 417 #ifdef __linux__
 418         Error *err = NULL;
 419         memory_region_init_ram_from_file(mr, owner, name, ram_size, false,
 420                                          mem_path, &err);
 421         if (err) {
 422             error_report_err(err);
 423             if (mem_prealloc) {
 424                 exit(1);
 425             }
 426
 427             /* Legacy behavior: if allocation failed, fall back to
 428              * regular RAM allocation.
 429              */
 430             memory_region_init_ram(mr, owner, name, ram_size, &error_fatal);
 431         }
 432 #else
 433         fprintf(stderr, "-mem-path not supported on this host\n");
 434         exit(1);
 435 #endif
 436     } else {
 437         memory_region_init_ram(mr, owner, name, ram_size, &error_fatal);
 438     }
 439     vmstate_register_ram_global(mr);
 440 }
 441
 442 void memory_region_allocate_system_memory(MemoryRegion *mr, Object *owner,
 443                                           const char *name,
 444                                           uint64_t ram_size)
 445 {
 446     uint64_t addr = 0;
 447     int i;
 448
 449     if (nb_numa_nodes == 0 || !have_memdevs) {
 450         allocate_system_memory_nonnuma(mr, owner, name, ram_size);
 451         return;
 452     }
 453
 454     memory_region_init(mr, owner, name, ram_size);
 455     for (i = 0; i < MAX_NODES; i++) {
 456         uint64_t size = numa_info[i].node_mem;
 457         HostMemoryBackend *backend = numa_info[i].node_memdev;
 458         if (!backend) {
 459             continue;
 460         }
 461         MemoryRegion *seg = host_memory_backend_get_memory(backend,
 462                                                            &error_fatal);
 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         host_memory_backend_set_mapped(backend, true);
 473         memory_region_add_subregion(mr, addr, seg);
 474         vmstate_register_ram_global(seg);
 475         addr += size;
 476     }
 477 }
 478
 479 static void numa_stat_memory_devices(uint64_t node_mem[])
 480 {
 481     MemoryDeviceInfoList *info_list = NULL;
 482     MemoryDeviceInfoList **prev = &info_list;
 483     MemoryDeviceInfoList *info;
 484
 485     qmp_pc_dimm_device_list(qdev_get_machine(), &prev);
 486     for (info = info_list; info; info = info->next) {
 487         MemoryDeviceInfo *value = info->value;
 488
 489         if (value) {
 490             switch (value->type) {
 491             case MEMORY_DEVICE_INFO_KIND_DIMM:
 492                 node_mem[value->u.dimm.data->node] += value->u.dimm.data->size;
 493                 break;
 494             default:
 495                 break;
 496             }
 497         }
 498     }
 499     qapi_free_MemoryDeviceInfoList(info_list);
 500 }
 501
 502 void query_numa_node_mem(uint64_t node_mem[])
 503 {
 504     int i;
 505
 506     if (nb_numa_nodes <= 0) {
 507         return;
 508     }
 509
 510     numa_stat_memory_devices(node_mem);
 511     for (i = 0; i < nb_numa_nodes; i++) {
 512         node_mem[i] += numa_info[i].node_mem;
 513     }
 514 }
 515
 516 static int query_memdev(Object *obj, void *opaque)
 517 {
 518     MemdevList **list = opaque;
 519     MemdevList *m = 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->id = object_property_get_str(obj, "id", NULL);
 527         m->value->has_id = !!m->value->id;
 528
 529         m->value->size = object_property_get_int(obj, "size",
 530                                                  &error_abort);
 531         m->value->merge = object_property_get_bool(obj, "merge",
 532                                                    &error_abort);
 533         m->value->dump = object_property_get_bool(obj, "dump",
 534                                                   &error_abort);
 535         m->value->prealloc = object_property_get_bool(obj,
 536                                                       "prealloc",
 537                                                       &error_abort);
 538         m->value->policy = object_property_get_enum(obj,
 539                                                     "policy",
 540                                                     "HostMemPolicy",
 541                                                     &error_abort);
 542         object_property_get_uint16List(obj, "host-nodes",
 543                                        &m->value->host_nodes,
 544                                        &error_abort);
 545
 546         m->next = *list;
 547         *list = m;
 548     }
 549
 550     return 0;
 551 }
 552
 553 MemdevList *qmp_query_memdev(Error **errp)
 554 {
 555     Object *obj = object_get_objects_root();
 556     MemdevList *list = NULL;
 557
 558     object_child_foreach(obj, query_memdev, &list);
 559     return list;
 560 }
 561
 562 int numa_get_node_for_cpu(int idx)
 563 {
 564     int i;
 565
 566     assert(idx < max_cpus);
 567
 568     for (i = 0; i < nb_numa_nodes; i++) {
 569         if (test_bit(idx, numa_info[i].node_cpu)) {
 570             break;
 571         }
 572     }
 573     return i;
 574 }