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;
  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         OptsVisitor *ov = opts_visitor_new(opts);
 221         visit_type_NumaOptions(opts_get_visitor(ov), &object, NULL, &err);
 222         opts_visitor_cleanup(ov);
 223     }
 224
 225     if (err) {
 226         goto error;
 227     }
 228
 229     switch (object->kind) {
 230     case NUMA_OPTIONS_KIND_NODE:
 231         numa_node_parse(object->node, opts, &err);
 232         if (err) {
 233             goto error;
 234         }
 235         nb_numa_nodes++;
 236         break;
 237     default:
 238         abort();
 239     }
 240
 241     return 0;
 242
 243 error:
 244     error_report_err(err);
 245
 246     if (object) {
 247         QapiDeallocVisitor *dv = qapi_dealloc_visitor_new();
 248         visit_type_NumaOptions(qapi_dealloc_get_visitor(dv),
 249                                &object, NULL, NULL);
 250         qapi_dealloc_visitor_cleanup(dv);
 251     }
 252
 253     return -1;
 254 }
 255
 256 static char *enumerate_cpus(unsigned long *cpus, int max_cpus)
 257 {
 258     int cpu;
 259     bool first = true;
 260     GString *s = g_string_new(NULL);
 261
 262     for (cpu = find_first_bit(cpus, max_cpus);
 263         cpu < max_cpus;
 264         cpu = find_next_bit(cpus, max_cpus, cpu + 1)) {
 265         g_string_append_printf(s, "%s%d", first ? "" : " ", cpu);
 266         first = false;
 267     }
 268     return g_string_free(s, FALSE);
 269 }
 270
 271 static void validate_numa_cpus(void)
 272 {
 273     int i;
 274     DECLARE_BITMAP(seen_cpus, MAX_CPUMASK_BITS);
 275
 276     bitmap_zero(seen_cpus, MAX_CPUMASK_BITS);
 277     for (i = 0; i < nb_numa_nodes; i++) {
 278         if (bitmap_intersects(seen_cpus, numa_info[i].node_cpu,
 279                               MAX_CPUMASK_BITS)) {
 280             bitmap_and(seen_cpus, seen_cpus,
 281                        numa_info[i].node_cpu, MAX_CPUMASK_BITS);
 282             error_report("CPU(s) present in multiple NUMA nodes: %s",
 283                          enumerate_cpus(seen_cpus, max_cpus));;
 284             exit(EXIT_FAILURE);
 285         }
 286         bitmap_or(seen_cpus, seen_cpus,
 287                   numa_info[i].node_cpu, MAX_CPUMASK_BITS);
 288     }
 289
 290     if (!bitmap_full(seen_cpus, max_cpus)) {
 291         char *msg;
 292         bitmap_complement(seen_cpus, seen_cpus, max_cpus);
 293         msg = enumerate_cpus(seen_cpus, max_cpus);
 294         error_report("warning: CPU(s) not present in any NUMA nodes: %s", msg);
 295         error_report("warning: All CPU(s) up to maxcpus should be described "
 296                      "in NUMA config");
 297         g_free(msg);
 298     }
 299 }
 300
 301 void parse_numa_opts(MachineClass *mc)
 302 {
 303     int i;
 304
 305     if (qemu_opts_foreach(qemu_find_opts("numa"), parse_numa, NULL, NULL)) {
 306         exit(1);
 307     }
 308
 309     assert(max_numa_nodeid <= MAX_NODES);
 310
 311     /* No support for sparse NUMA node IDs yet: */
 312     for (i = max_numa_nodeid - 1; i >= 0; i--) {
 313         /* Report large node IDs first, to make mistakes easier to spot */
 314         if (!numa_info[i].present) {
 315             error_report("numa: Node ID missing: %d", i);
 316             exit(1);
 317         }
 318     }
 319
 320     /* This must be always true if all nodes are present: */
 321     assert(nb_numa_nodes == max_numa_nodeid);
 322
 323     if (nb_numa_nodes > 0) {
 324         uint64_t numa_total;
 325
 326         if (nb_numa_nodes > MAX_NODES) {
 327             nb_numa_nodes = MAX_NODES;
 328         }
 329
 330         /* If no memory size is given for any node, assume the default case
 331          * and distribute the available memory equally across all nodes
 332          */
 333         for (i = 0; i < nb_numa_nodes; i++) {
 334             if (numa_info[i].node_mem != 0) {
 335                 break;
 336             }
 337         }
 338         if (i == nb_numa_nodes) {
 339             uint64_t usedmem = 0;
 340
 341             /* On Linux, each node's border has to be 8MB aligned,
 342              * the final node gets the rest.
 343              */
 344             for (i = 0; i < nb_numa_nodes - 1; i++) {
 345                 numa_info[i].node_mem = (ram_size / nb_numa_nodes) &
 346                                         ~((1 << 23UL) - 1);
 347                 usedmem += numa_info[i].node_mem;
 348             }
 349             numa_info[i].node_mem = ram_size - usedmem;
 350         }
 351
 352         numa_total = 0;
 353         for (i = 0; i < nb_numa_nodes; i++) {
 354             numa_total += numa_info[i].node_mem;
 355         }
 356         if (numa_total != ram_size) {
 357             error_report("total memory for NUMA nodes (0x%" PRIx64 ")"
 358                          " should equal RAM size (0x" RAM_ADDR_FMT ")",
 359                          numa_total, ram_size);
 360             exit(1);
 361         }
 362
 363         for (i = 0; i < nb_numa_nodes; i++) {
 364             QLIST_INIT(&numa_info[i].addr);
 365         }
 366
 367         numa_set_mem_ranges();
 368
 369         for (i = 0; i < nb_numa_nodes; i++) {
 370             if (!bitmap_empty(numa_info[i].node_cpu, MAX_CPUMASK_BITS)) {
 371                 break;
 372             }
 373         }
 374         /* Historically VCPUs were assigned in round-robin order to NUMA
 375          * nodes. However it causes issues with guest not handling it nice
 376          * in case where cores/threads from a multicore CPU appear on
 377          * different nodes. So allow boards to override default distribution
 378          * rule grouping VCPUs by socket so that VCPUs from the same socket
 379          * would be on the same node.
 380          */
 381         if (i == nb_numa_nodes) {
 382             for (i = 0; i < max_cpus; i++) {
 383                 unsigned node_id = i % nb_numa_nodes;
 384                 if (mc->cpu_index_to_socket_id) {
 385                     node_id = mc->cpu_index_to_socket_id(i) % nb_numa_nodes;
 386                 }
 387
 388                 set_bit(i, numa_info[node_id].node_cpu);
 389             }
 390         }
 391
 392         validate_numa_cpus();
 393     } else {
 394         numa_set_mem_node_id(0, ram_size, 0);
 395     }
 396 }
 397
 398 void numa_post_machine_init(void)
 399 {
 400     CPUState *cpu;
 401     int i;
 402
 403     CPU_FOREACH(cpu) {
 404         for (i = 0; i < nb_numa_nodes; i++) {
 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
 422         /* Legacy behavior: if allocation failed, fall back to
 423          * regular RAM allocation.
 424          */
 425         if (err) {
 426             error_report_err(err);
 427             memory_region_init_ram(mr, owner, name, ram_size, &error_abort);
 428         }
 429 #else
 430         fprintf(stderr, "-mem-path not supported on this host\n");
 431         exit(1);
 432 #endif
 433     } else {
 434         memory_region_init_ram(mr, owner, name, ram_size, &error_abort);
 435     }
 436     vmstate_register_ram_global(mr);
 437 }
 438
 439 void memory_region_allocate_system_memory(MemoryRegion *mr, Object *owner,
 440                                           const char *name,
 441                                           uint64_t ram_size)
 442 {
 443     uint64_t addr = 0;
 444     int i;
 445
 446     if (nb_numa_nodes == 0 || !have_memdevs) {
 447         allocate_system_memory_nonnuma(mr, owner, name, ram_size);
 448         return;
 449     }
 450
 451     memory_region_init(mr, owner, name, ram_size);
 452     for (i = 0; i < MAX_NODES; i++) {
 453         Error *local_err = NULL;
 454         uint64_t size = numa_info[i].node_mem;
 455         HostMemoryBackend *backend = numa_info[i].node_memdev;
 456         if (!backend) {
 457             continue;
 458         }
 459         MemoryRegion *seg = host_memory_backend_get_memory(backend, &local_err);
 460         if (local_err) {
 461             error_report_err(local_err);
 462             exit(1);
 463         }
 464
 465         if (memory_region_is_mapped(seg)) {
 466             char *path = object_get_canonical_path_component(OBJECT(backend));
 467             error_report("memory backend %s is used multiple times. Each "
 468                          "-numa option must use a different memdev value.",
 469                          path);
 470             exit(1);
 471         }
 472
 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->kind) {
 491             case MEMORY_DEVICE_INFO_KIND_DIMM:
 492                 node_mem[value->dimm->node] += value->dimm->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     Error *err = NULL;
 521
 522     if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) {
 523         m = g_malloc0(sizeof(*m));
 524
 525         m->value = g_malloc0(sizeof(*m->value));
 526
 527         m->value->size = object_property_get_int(obj, "size",
 528                                                  &err);
 529         if (err) {
 530             goto error;
 531         }
 532
 533         m->value->merge = object_property_get_bool(obj, "merge",
 534                                                    &err);
 535         if (err) {
 536             goto error;
 537         }
 538
 539         m->value->dump = object_property_get_bool(obj, "dump",
 540                                                   &err);
 541         if (err) {
 542             goto error;
 543         }
 544
 545         m->value->prealloc = object_property_get_bool(obj,
 546                                                       "prealloc", &err);
 547         if (err) {
 548             goto error;
 549         }
 550
 551         m->value->policy = object_property_get_enum(obj,
 552                                                     "policy",
 553                                                     "HostMemPolicy",
 554                                                     &err);
 555         if (err) {
 556             goto error;
 557         }
 558
 559         object_property_get_uint16List(obj, "host-nodes",
 560                                        &m->value->host_nodes, &err);
 561         if (err) {
 562             goto error;
 563         }
 564
 565         m->next = *list;
 566         *list = m;
 567     }
 568
 569     return 0;
 570 error:
 571     g_free(m->value);
 572     g_free(m);
 573
 574     return -1;
 575 }
 576
 577 MemdevList *qmp_query_memdev(Error **errp)
 578 {
 579     Object *obj;
 580     MemdevList *list = NULL;
 581
 582     obj = object_get_objects_root();
 583     if (obj == NULL) {
 584         return NULL;
 585     }
 586
 587     if (object_child_foreach(obj, query_memdev, &list) != 0) {
 588         goto error;
 589     }
 590
 591     return list;
 592
 593 error:
 594     qapi_free_MemdevList(list);
 595     return NULL;
 596 }