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 "qapi/dealloc-visitor.h"
  35 #include "hw/boards.h"
  36 #include "sysemu/hostmem.h"
  37 #include "qmp-commands.h"
  38 #include "hw/mem/pc-dimm.h"
  39 #include "qemu/option.h"
  40 #include "qemu/config-file.h"
  41
  42 QemuOptsList qemu_numa_opts = {
  43     .name = "numa",
  44     .implied_opt_name = "type",
  45     .head = QTAILQ_HEAD_INITIALIZER(qemu_numa_opts.head),
  46     .desc = { { 0 } } /* validated with OptsVisitor */
  47 };
  48
  49 static int have_memdevs = -1;
  50 static int max_numa_nodeid; /* Highest specified NUMA node ID, plus one.
  51                              * For all nodes, nodeid < max_numa_nodeid
  52                              */
  53 int nb_numa_nodes;
  54 NodeInfo numa_info[MAX_NODES];
  55
  56 void numa_set_mem_node_id(ram_addr_t addr, uint64_t size, uint32_t node)
  57 {
  58     struct numa_addr_range *range;
  59
  60     /*
  61      * Memory-less nodes can come here with 0 size in which case,
  62      * there is nothing to do.
  63      */
  64     if (!size) {
  65         return;
  66     }
  67
  68     range = g_malloc0(sizeof(*range));
  69     range->mem_start = addr;
  70     range->mem_end = addr + size - 1;
  71     QLIST_INSERT_HEAD(&numa_info[node].addr, range, entry);
  72 }
  73
  74 void numa_unset_mem_node_id(ram_addr_t addr, uint64_t size, uint32_t node)
  75 {
  76     struct numa_addr_range *range, *next;
  77
  78     QLIST_FOREACH_SAFE(range, &numa_info[node].addr, entry, next) {
  79         if (addr == range->mem_start && (addr + size - 1) == range->mem_end) {
  80             QLIST_REMOVE(range, entry);
  81             g_free(range);
  82             return;
  83         }
  84     }
  85 }
  86
  87 static void numa_set_mem_ranges(void)
  88 {
  89     int i;
  90     ram_addr_t mem_start = 0;
  91
  92     /*
  93      * Deduce start address of each node and use it to store
  94      * the address range info in numa_info address range list
  95      */
  96     for (i = 0; i < nb_numa_nodes; i++) {
  97         numa_set_mem_node_id(mem_start, numa_info[i].node_mem, i);
  98         mem_start += numa_info[i].node_mem;
  99     }
 100 }
 101
 102 /*
 103  * Check if @addr falls under NUMA @node.
 104  */
 105 static bool numa_addr_belongs_to_node(ram_addr_t addr, uint32_t node)
 106 {
 107     struct numa_addr_range *range;
 108
 109     QLIST_FOREACH(range, &numa_info[node].addr, entry) {
 110         if (addr >= range->mem_start && addr <= range->mem_end) {
 111             return true;
 112         }
 113     }
 114     return false;
 115 }
 116
 117 /*
 118  * Given an address, return the index of the NUMA node to which the
 119  * address belongs to.
 120  */
 121 uint32_t numa_get_node(ram_addr_t addr, Error **errp)
 122 {
 123     uint32_t i;
 124
 125     /* For non NUMA configurations, check if the addr falls under node 0 */
 126     if (!nb_numa_nodes) {
 127         if (numa_addr_belongs_to_node(addr, 0)) {
 128             return 0;
 129         }
 130     }
 131
 132     for (i = 0; i < nb_numa_nodes; i++) {
 133         if (numa_addr_belongs_to_node(addr, i)) {
 134             return i;
 135         }
 136     }
 137
 138     error_setg(errp, "Address 0x" RAM_ADDR_FMT " doesn't belong to any "
 139                 "NUMA node", addr);
 140     return -1;
 141 }
 142
 143 static void numa_node_parse(NumaNodeOptions *node, QemuOpts *opts, Error **errp)
 144 {
 145     uint16_t nodenr;
 146     uint16List *cpus = NULL;
 147
 148     if (node->has_nodeid) {
 149         nodenr = node->nodeid;
 150     } else {
 151         nodenr = nb_numa_nodes;
 152     }
 153
 154     if (nodenr >= MAX_NODES) {
 155         error_setg(errp, "Max number of NUMA nodes reached: %"
 156                    PRIu16 "", nodenr);
 157         return;
 158     }
 159
 160     if (numa_info[nodenr].present) {
 161         error_setg(errp, "Duplicate NUMA nodeid: %" PRIu16, nodenr);
 162         return;
 163     }
 164
 165     for (cpus = node->cpus; cpus; cpus = cpus->next) {
 166         if (cpus->value >= max_cpus) {
 167             error_setg(errp,
 168                        "CPU index (%" PRIu16 ")"
 169                        " should be smaller than maxcpus (%d)",
 170                        cpus->value, max_cpus);
 171             return;
 172         }
 173         bitmap_set(numa_info[nodenr].node_cpu, cpus->value, 1);
 174     }
 175
 176     if (node->has_mem && node->has_memdev) {
 177         error_setg(errp, "qemu: cannot specify both mem= and memdev=");
 178         return;
 179     }
 180
 181     if (have_memdevs == -1) {
 182         have_memdevs = node->has_memdev;
 183     }
 184     if (node->has_memdev != have_memdevs) {
 185         error_setg(errp, "qemu: memdev option must be specified for either "
 186                    "all or no nodes");
 187         return;
 188     }
 189
 190     if (node->has_mem) {
 191         uint64_t mem_size = node->mem;
 192         const char *mem_str = qemu_opt_get(opts, "mem");
 193         /* Fix up legacy suffix-less format */
 194         if (g_ascii_isdigit(mem_str[strlen(mem_str) - 1])) {
 195             mem_size <<= 20;
 196         }
 197         numa_info[nodenr].node_mem = mem_size;
 198     }
 199     if (node->has_memdev) {
 200         Object *o;
 201         o = object_resolve_path_type(node->memdev, TYPE_MEMORY_BACKEND, NULL);
 202         if (!o) {
 203             error_setg(errp, "memdev=%s is ambiguous", node->memdev);
 204             return;
 205         }
 206
 207         object_ref(o);
 208         numa_info[nodenr].node_mem = object_property_get_int(o, "size", NULL);
 209         numa_info[nodenr].node_memdev = MEMORY_BACKEND(o);
 210     }
 211     numa_info[nodenr].present = true;
 212     max_numa_nodeid = MAX(max_numa_nodeid, nodenr + 1);
 213 }
 214
 215 static int parse_numa(void *opaque, QemuOpts *opts, Error **errp)
 216 {
 217     NumaOptions *object = NULL;
 218     Error *err = NULL;
 219
 220     {
 221         OptsVisitor *ov = opts_visitor_new(opts);
 222         visit_type_NumaOptions(opts_get_visitor(ov), NULL, &object, &err);
 223         opts_visitor_cleanup(ov);
 224     }
 225
 226     if (err) {
 227         goto error;
 228     }
 229
 230     switch (object->type) {
 231     case NUMA_OPTIONS_KIND_NODE:
 232         numa_node_parse(object->u.node, opts, &err);
 233         if (err) {
 234             goto error;
 235         }
 236         nb_numa_nodes++;
 237         break;
 238     default:
 239         abort();
 240     }
 241
 242     return 0;
 243
 244 error:
 245     error_report_err(err);
 246
 247     if (object) {
 248         QapiDeallocVisitor *dv = qapi_dealloc_visitor_new();
 249         visit_type_NumaOptions(qapi_dealloc_get_visitor(dv), NULL, &object,
 250                                NULL);
 251         qapi_dealloc_visitor_cleanup(dv);
 252     }
 253
 254     return -1;
 255 }
 256
 257 static char *enumerate_cpus(unsigned long *cpus, int max_cpus)
 258 {
 259     int cpu;
 260     bool first = true;
 261     GString *s = g_string_new(NULL);
 262
 263     for (cpu = find_first_bit(cpus, max_cpus);
 264         cpu < max_cpus;
 265         cpu = find_next_bit(cpus, max_cpus, cpu + 1)) {
 266         g_string_append_printf(s, "%s%d", first ? "" : " ", cpu);
 267         first = false;
 268     }
 269     return g_string_free(s, FALSE);
 270 }
 271
 272 static void validate_numa_cpus(void)
 273 {
 274     int i;
 275     DECLARE_BITMAP(seen_cpus, MAX_CPUMASK_BITS);
 276
 277     bitmap_zero(seen_cpus, MAX_CPUMASK_BITS);
 278     for (i = 0; i < nb_numa_nodes; i++) {
 279         if (bitmap_intersects(seen_cpus, numa_info[i].node_cpu,
 280                               MAX_CPUMASK_BITS)) {
 281             bitmap_and(seen_cpus, seen_cpus,
 282                        numa_info[i].node_cpu, MAX_CPUMASK_BITS);
 283             error_report("CPU(s) present in multiple NUMA nodes: %s",
 284                          enumerate_cpus(seen_cpus, max_cpus));
 285             exit(EXIT_FAILURE);
 286         }
 287         bitmap_or(seen_cpus, seen_cpus,
 288                   numa_info[i].node_cpu, MAX_CPUMASK_BITS);
 289     }
 290
 291     if (!bitmap_full(seen_cpus, max_cpus)) {
 292         char *msg;
 293         bitmap_complement(seen_cpus, seen_cpus, max_cpus);
 294         msg = enumerate_cpus(seen_cpus, max_cpus);
 295         error_report("warning: CPU(s) not present in any NUMA nodes: %s", msg);
 296         error_report("warning: All CPU(s) up to maxcpus should be described "
 297                      "in NUMA config");
 298         g_free(msg);
 299     }
 300 }
 301
 302 void parse_numa_opts(MachineClass *mc)
 303 {
 304     int i;
 305
 306     if (qemu_opts_foreach(qemu_find_opts("numa"), parse_numa, NULL, NULL)) {
 307         exit(1);
 308     }
 309
 310     assert(max_numa_nodeid <= MAX_NODES);
 311
 312     /* No support for sparse NUMA node IDs yet: */
 313     for (i = max_numa_nodeid - 1; i >= 0; i--) {
 314         /* Report large node IDs first, to make mistakes easier to spot */
 315         if (!numa_info[i].present) {
 316             error_report("numa: Node ID missing: %d", i);
 317             exit(1);
 318         }
 319     }
 320
 321     /* This must be always true if all nodes are present: */
 322     assert(nb_numa_nodes == max_numa_nodeid);
 323
 324     if (nb_numa_nodes > 0) {
 325         uint64_t numa_total;
 326
 327         if (nb_numa_nodes > MAX_NODES) {
 328             nb_numa_nodes = MAX_NODES;
 329         }
 330
 331         /* If no memory size is given for any node, assume the default case
 332          * and distribute the available memory equally across all nodes
 333          */
 334         for (i = 0; i < nb_numa_nodes; i++) {
 335             if (numa_info[i].node_mem != 0) {
 336                 break;
 337             }
 338         }
 339         if (i == nb_numa_nodes) {
 340             uint64_t usedmem = 0;
 341
 342             /* On Linux, each node's border has to be 8MB aligned,
 343              * the final node gets the rest.
 344              */
 345             for (i = 0; i < nb_numa_nodes - 1; i++) {
 346                 numa_info[i].node_mem = (ram_size / nb_numa_nodes) &
 347                                         ~((1 << 23UL) - 1);
 348                 usedmem += numa_info[i].node_mem;
 349             }
 350             numa_info[i].node_mem = ram_size - usedmem;
 351         }
 352
 353         numa_total = 0;
 354         for (i = 0; i < nb_numa_nodes; i++) {
 355             numa_total += numa_info[i].node_mem;
 356         }
 357         if (numa_total != ram_size) {
 358             error_report("total memory for NUMA nodes (0x%" PRIx64 ")"
 359                          " should equal RAM size (0x" RAM_ADDR_FMT ")",
 360                          numa_total, ram_size);
 361             exit(1);
 362         }
 363
 364         for (i = 0; i < nb_numa_nodes; i++) {
 365             QLIST_INIT(&numa_info[i].addr);
 366         }
 367
 368         numa_set_mem_ranges();
 369
 370         for (i = 0; i < nb_numa_nodes; i++) {
 371             if (!bitmap_empty(numa_info[i].node_cpu, MAX_CPUMASK_BITS)) {
 372                 break;
 373             }
 374         }
 375         /* Historically VCPUs were assigned in round-robin order to NUMA
 376          * nodes. However it causes issues with guest not handling it nice
 377          * in case where cores/threads from a multicore CPU appear on
 378          * different nodes. So allow boards to override default distribution
 379          * rule grouping VCPUs by socket so that VCPUs from the same socket
 380          * would be on the same node.
 381          */
 382         if (i == nb_numa_nodes) {
 383             for (i = 0; i < max_cpus; i++) {
 384                 unsigned node_id = i % nb_numa_nodes;
 385                 if (mc->cpu_index_to_socket_id) {
 386                     node_id = mc->cpu_index_to_socket_id(i) % nb_numa_nodes;
 387                 }
 388
 389                 set_bit(i, numa_info[node_id].node_cpu);
 390             }
 391         }
 392
 393         validate_numa_cpus();
 394     } else {
 395         numa_set_mem_node_id(0, ram_size, 0);
 396     }
 397 }
 398
 399 void numa_post_machine_init(void)
 400 {
 401     CPUState *cpu;
 402     int i;
 403
 404     CPU_FOREACH(cpu) {
 405         for (i = 0; i < nb_numa_nodes; i++) {
 406             if (test_bit(cpu->cpu_index, numa_info[i].node_cpu)) {
 407                 cpu->numa_node = i;
 408             }
 409         }
 410     }
 411 }
 412
 413 static void allocate_system_memory_nonnuma(MemoryRegion *mr, Object *owner,
 414                                            const char *name,
 415                                            uint64_t ram_size)
 416 {
 417     if (mem_path) {
 418 #ifdef __linux__
 419         Error *err = NULL;
 420         memory_region_init_ram_from_file(mr, owner, name, ram_size, false,
 421                                          mem_path, &err);
 422         if (err) {
 423             error_report_err(err);
 424             if (mem_prealloc) {
 425                 exit(1);
 426             }
 427
 428             /* Legacy behavior: if allocation failed, fall back to
 429              * regular RAM allocation.
 430              */
 431             memory_region_init_ram(mr, owner, name, ram_size, &error_fatal);
 432         }
 433 #else
 434         fprintf(stderr, "-mem-path not supported on this host\n");
 435         exit(1);
 436 #endif
 437     } else {
 438         memory_region_init_ram(mr, owner, name, ram_size, &error_fatal);
 439     }
 440     vmstate_register_ram_global(mr);
 441 }
 442
 443 void memory_region_allocate_system_memory(MemoryRegion *mr, Object *owner,
 444                                           const char *name,
 445                                           uint64_t ram_size)
 446 {
 447     uint64_t addr = 0;
 448     int i;
 449
 450     if (nb_numa_nodes == 0 || !have_memdevs) {
 451         allocate_system_memory_nonnuma(mr, owner, name, ram_size);
 452         return;
 453     }
 454
 455     memory_region_init(mr, owner, name, ram_size);
 456     for (i = 0; i < MAX_NODES; i++) {
 457         uint64_t size = numa_info[i].node_mem;
 458         HostMemoryBackend *backend = numa_info[i].node_memdev;
 459         if (!backend) {
 460             continue;
 461         }
 462         MemoryRegion *seg = host_memory_backend_get_memory(backend,
 463                                                            &error_fatal);
 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->type) {
 491             case MEMORY_DEVICE_INFO_KIND_DIMM:
 492                 node_mem[value->u.dimm->node] += value->u.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
 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                                                  &error_abort);
 528         m->value->merge = object_property_get_bool(obj, "merge",
 529                                                    &error_abort);
 530         m->value->dump = object_property_get_bool(obj, "dump",
 531                                                   &error_abort);
 532         m->value->prealloc = object_property_get_bool(obj,
 533                                                       "prealloc",
 534                                                       &error_abort);
 535         m->value->policy = object_property_get_enum(obj,
 536                                                     "policy",
 537                                                     "HostMemPolicy",
 538                                                     &error_abort);
 539         object_property_get_uint16List(obj, "host-nodes",
 540                                        &m->value->host_nodes,
 541                                        &error_abort);
 542
 543         m->next = *list;
 544         *list = m;
 545     }
 546
 547     return 0;
 548 }
 549
 550 MemdevList *qmp_query_memdev(Error **errp)
 551 {
 552     Object *obj = object_get_objects_root();
 553     MemdevList *list = NULL;
 554
 555     object_child_foreach(obj, query_memdev, &list);
 556     return list;
 557 }