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