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     DECLARE_BITMAP(seen_cpus, MAX_CPUMASK_BITS);
 270
 271     bitmap_zero(seen_cpus, MAX_CPUMASK_BITS);
 272     for (i = 0; i < nb_numa_nodes; i++) {
 273         if (bitmap_intersects(seen_cpus, numa_info[i].node_cpu,
 274                               MAX_CPUMASK_BITS)) {
 275             bitmap_and(seen_cpus, seen_cpus,
 276                        numa_info[i].node_cpu, MAX_CPUMASK_BITS);
 277             error_report("CPU(s) present in multiple NUMA nodes: %s",
 278                          enumerate_cpus(seen_cpus, max_cpus));
 279             exit(EXIT_FAILURE);
 280         }
 281         bitmap_or(seen_cpus, seen_cpus,
 282                   numa_info[i].node_cpu, MAX_CPUMASK_BITS);
 283     }
 284
 285     if (!bitmap_full(seen_cpus, max_cpus)) {
 286         char *msg;
 287         bitmap_complement(seen_cpus, seen_cpus, max_cpus);
 288         msg = enumerate_cpus(seen_cpus, max_cpus);
 289         error_report("warning: CPU(s) not present in any NUMA nodes: %s", msg);
 290         error_report("warning: All CPU(s) up to maxcpus should be described "
 291                      "in NUMA config");
 292         g_free(msg);
 293     }
 294 }
 295
 296 void parse_numa_opts(MachineClass *mc)
 297 {
 298     int i;
 299
 300     if (qemu_opts_foreach(qemu_find_opts("numa"), parse_numa, NULL, NULL)) {
 301         exit(1);
 302     }
 303
 304     assert(max_numa_nodeid <= MAX_NODES);
 305
 306     /* No support for sparse NUMA node IDs yet: */
 307     for (i = max_numa_nodeid - 1; i >= 0; i--) {
 308         /* Report large node IDs first, to make mistakes easier to spot */
 309         if (!numa_info[i].present) {
 310             error_report("numa: Node ID missing: %d", i);
 311             exit(1);
 312         }
 313     }
 314
 315     /* This must be always true if all nodes are present: */
 316     assert(nb_numa_nodes == max_numa_nodeid);
 317
 318     if (nb_numa_nodes > 0) {
 319         uint64_t numa_total;
 320
 321         if (nb_numa_nodes > MAX_NODES) {
 322             nb_numa_nodes = MAX_NODES;
 323         }
 324
 325         /* If no memory size is given for any node, assume the default case
 326          * and distribute the available memory equally across all nodes
 327          */
 328         for (i = 0; i < nb_numa_nodes; i++) {
 329             if (numa_info[i].node_mem != 0) {
 330                 break;
 331             }
 332         }
 333         if (i == nb_numa_nodes) {
 334             uint64_t usedmem = 0;
 335
 336             /* On Linux, each node's border has to be 8MB aligned,
 337              * the final node gets the rest.
 338              */
 339             for (i = 0; i < nb_numa_nodes - 1; i++) {
 340                 numa_info[i].node_mem = (ram_size / nb_numa_nodes) &
 341                                         ~((1 << 23UL) - 1);
 342                 usedmem += numa_info[i].node_mem;
 343             }
 344             numa_info[i].node_mem = ram_size - usedmem;
 345         }
 346
 347         numa_total = 0;
 348         for (i = 0; i < nb_numa_nodes; i++) {
 349             numa_total += numa_info[i].node_mem;
 350         }
 351         if (numa_total != ram_size) {
 352             error_report("total memory for NUMA nodes (0x%" PRIx64 ")"
 353                          " should equal RAM size (0x" RAM_ADDR_FMT ")",
 354                          numa_total, ram_size);
 355             exit(1);
 356         }
 357
 358         for (i = 0; i < nb_numa_nodes; i++) {
 359             QLIST_INIT(&numa_info[i].addr);
 360         }
 361
 362         numa_set_mem_ranges();
 363
 364         for (i = 0; i < nb_numa_nodes; i++) {
 365             if (!bitmap_empty(numa_info[i].node_cpu, MAX_CPUMASK_BITS)) {
 366                 break;
 367             }
 368         }
 369         /* Historically VCPUs were assigned in round-robin order to NUMA
 370          * nodes. However it causes issues with guest not handling it nice
 371          * in case where cores/threads from a multicore CPU appear on
 372          * different nodes. So allow boards to override default distribution
 373          * rule grouping VCPUs by socket so that VCPUs from the same socket
 374          * would be on the same node.
 375          */
 376         if (i == nb_numa_nodes) {
 377             for (i = 0; i < max_cpus; i++) {
 378                 unsigned node_id = i % nb_numa_nodes;
 379                 if (mc->cpu_index_to_socket_id) {
 380                     node_id = mc->cpu_index_to_socket_id(i) % nb_numa_nodes;
 381                 }
 382
 383                 set_bit(i, numa_info[node_id].node_cpu);
 384             }
 385         }
 386
 387         validate_numa_cpus();
 388     } else {
 389         numa_set_mem_node_id(0, ram_size, 0);
 390     }
 391 }
 392
 393 void numa_post_machine_init(void)
 394 {
 395     CPUState *cpu;
 396     int i;
 397
 398     CPU_FOREACH(cpu) {
 399         for (i = 0; i < nb_numa_nodes; i++) {
 400             if (test_bit(cpu->cpu_index, numa_info[i].node_cpu)) {
 401                 cpu->numa_node = i;
 402             }
 403         }
 404     }
 405 }
 406
 407 static void allocate_system_memory_nonnuma(MemoryRegion *mr, Object *owner,
 408                                            const char *name,
 409                                            uint64_t ram_size)
 410 {
 411     if (mem_path) {
 412 #ifdef __linux__
 413         Error *err = NULL;
 414         memory_region_init_ram_from_file(mr, owner, name, ram_size, false,
 415                                          mem_path, &err);
 416         if (err) {
 417             error_report_err(err);
 418             if (mem_prealloc) {
 419                 exit(1);
 420             }
 421
 422             /* Legacy behavior: if allocation failed, fall back to
 423              * regular RAM allocation.
 424              */
 425             memory_region_init_ram(mr, owner, name, ram_size, &error_fatal);
 426         }
 427 #else
 428         fprintf(stderr, "-mem-path not supported on this host\n");
 429         exit(1);
 430 #endif
 431     } else {
 432         memory_region_init_ram(mr, owner, name, ram_size, &error_fatal);
 433     }
 434     vmstate_register_ram_global(mr);
 435 }
 436
 437 void memory_region_allocate_system_memory(MemoryRegion *mr, Object *owner,
 438                                           const char *name,
 439                                           uint64_t ram_size)
 440 {
 441     uint64_t addr = 0;
 442     int i;
 443
 444     if (nb_numa_nodes == 0 || !have_memdevs) {
 445         allocate_system_memory_nonnuma(mr, owner, name, ram_size);
 446         return;
 447     }
 448
 449     memory_region_init(mr, owner, name, ram_size);
 450     for (i = 0; i < MAX_NODES; i++) {
 451         uint64_t size = numa_info[i].node_mem;
 452         HostMemoryBackend *backend = numa_info[i].node_memdev;
 453         if (!backend) {
 454             continue;
 455         }
 456         MemoryRegion *seg = host_memory_backend_get_memory(backend,
 457                                                            &error_fatal);
 458
 459         if (memory_region_is_mapped(seg)) {
 460             char *path = object_get_canonical_path_component(OBJECT(backend));
 461             error_report("memory backend %s is used multiple times. Each "
 462                          "-numa option must use a different memdev value.",
 463                          path);
 464             exit(1);
 465         }
 466
 467         host_memory_backend_set_mapped(backend, true);
 468         memory_region_add_subregion(mr, addr, seg);
 469         vmstate_register_ram_global(seg);
 470         addr += size;
 471     }
 472 }
 473
 474 static void numa_stat_memory_devices(uint64_t node_mem[])
 475 {
 476     MemoryDeviceInfoList *info_list = NULL;
 477     MemoryDeviceInfoList **prev = &info_list;
 478     MemoryDeviceInfoList *info;
 479
 480     qmp_pc_dimm_device_list(qdev_get_machine(), &prev);
 481     for (info = info_list; info; info = info->next) {
 482         MemoryDeviceInfo *value = info->value;
 483
 484         if (value) {
 485             switch (value->type) {
 486             case MEMORY_DEVICE_INFO_KIND_DIMM:
 487                 node_mem[value->u.dimm.data->node] += value->u.dimm.data->size;
 488                 break;
 489             default:
 490                 break;
 491             }
 492         }
 493     }
 494     qapi_free_MemoryDeviceInfoList(info_list);
 495 }
 496
 497 void query_numa_node_mem(uint64_t node_mem[])
 498 {
 499     int i;
 500
 501     if (nb_numa_nodes <= 0) {
 502         return;
 503     }
 504
 505     numa_stat_memory_devices(node_mem);
 506     for (i = 0; i < nb_numa_nodes; i++) {
 507         node_mem[i] += numa_info[i].node_mem;
 508     }
 509 }
 510
 511 static int query_memdev(Object *obj, void *opaque)
 512 {
 513     MemdevList **list = opaque;
 514     MemdevList *m = NULL;
 515
 516     if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) {
 517         m = g_malloc0(sizeof(*m));
 518
 519         m->value = g_malloc0(sizeof(*m->value));
 520
 521         m->value->size = object_property_get_int(obj, "size",
 522                                                  &error_abort);
 523         m->value->merge = object_property_get_bool(obj, "merge",
 524                                                    &error_abort);
 525         m->value->dump = object_property_get_bool(obj, "dump",
 526                                                   &error_abort);
 527         m->value->prealloc = object_property_get_bool(obj,
 528                                                       "prealloc",
 529                                                       &error_abort);
 530         m->value->policy = object_property_get_enum(obj,
 531                                                     "policy",
 532                                                     "HostMemPolicy",
 533                                                     &error_abort);
 534         object_property_get_uint16List(obj, "host-nodes",
 535                                        &m->value->host_nodes,
 536                                        &error_abort);
 537
 538         m->next = *list;
 539         *list = m;
 540     }
 541
 542     return 0;
 543 }
 544
 545 MemdevList *qmp_query_memdev(Error **errp)
 546 {
 547     Object *obj = object_get_objects_root();
 548     MemdevList *list = NULL;
 549
 550     object_child_foreach(obj, query_memdev, &list);
 551     return list;
 552 }
 553
 554 int numa_get_node_for_cpu(int idx)
 555 {
 556     int i;
 557
 558     for (i = 0; i < nb_numa_nodes; i++) {
 559         if (test_bit(idx, numa_info[i].node_cpu)) {
 560             break;
 561         }
 562     }
 563     return i;
 564 }