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/sysemu.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 "qapi/qmp/qerror.h"
  35 #include "hw/boards.h"
  36 #include "sysemu/hostmem.h"
  37 #include "qmp-commands.h"
  38 #include "hw/mem/pc-dimm.h"
  39
  40 QemuOptsList qemu_numa_opts = {
  41     .name = "numa",
  42     .implied_opt_name = "type",
  43     .head = QTAILQ_HEAD_INITIALIZER(qemu_numa_opts.head),
  44     .desc = { { 0 } } /* validated with OptsVisitor */
  45 };
  46
  47 static int have_memdevs = -1;
  48
  49 static void numa_node_parse(NumaNodeOptions *node, QemuOpts *opts, Error **errp)
  50 {
  51     uint16_t nodenr;
  52     uint16List *cpus = NULL;
  53
  54     if (node->has_nodeid) {
  55         nodenr = node->nodeid;
  56     } else {
  57         nodenr = nb_numa_nodes;
  58     }
  59
  60     if (nodenr >= MAX_NODES) {
  61         error_setg(errp, "Max number of NUMA nodes reached: %"
  62                    PRIu16 "\n", nodenr);
  63         return;
  64     }
  65
  66     if (numa_info[nodenr].present) {
  67         error_setg(errp, "Duplicate NUMA nodeid: %" PRIu16, nodenr);
  68         return;
  69     }
  70
  71     for (cpus = node->cpus; cpus; cpus = cpus->next) {
  72         if (cpus->value > MAX_CPUMASK_BITS) {
  73             error_setg(errp, "CPU number %" PRIu16 " is bigger than %d",
  74                        cpus->value, MAX_CPUMASK_BITS);
  75             return;
  76         }
  77         bitmap_set(numa_info[nodenr].node_cpu, cpus->value, 1);
  78     }
  79
  80     if (node->has_mem && node->has_memdev) {
  81         error_setg(errp, "qemu: cannot specify both mem= and memdev=\n");
  82         return;
  83     }
  84
  85     if (have_memdevs == -1) {
  86         have_memdevs = node->has_memdev;
  87     }
  88     if (node->has_memdev != have_memdevs) {
  89         error_setg(errp, "qemu: memdev option must be specified for either "
  90                    "all or no nodes\n");
  91         return;
  92     }
  93
  94     if (node->has_mem) {
  95         uint64_t mem_size = node->mem;
  96         const char *mem_str = qemu_opt_get(opts, "mem");
  97         /* Fix up legacy suffix-less format */
  98         if (g_ascii_isdigit(mem_str[strlen(mem_str) - 1])) {
  99             mem_size <<= 20;
 100         }
 101         numa_info[nodenr].node_mem = mem_size;
 102     }
 103     if (node->has_memdev) {
 104         Object *o;
 105         o = object_resolve_path_type(node->memdev, TYPE_MEMORY_BACKEND, NULL);
 106         if (!o) {
 107             error_setg(errp, "memdev=%s is ambiguous", node->memdev);
 108             return;
 109         }
 110
 111         object_ref(o);
 112         numa_info[nodenr].node_mem = object_property_get_int(o, "size", NULL);
 113         numa_info[nodenr].node_memdev = MEMORY_BACKEND(o);
 114     }
 115     numa_info[nodenr].present = true;
 116     max_numa_nodeid = MAX(max_numa_nodeid, nodenr + 1);
 117 }
 118
 119 int numa_init_func(QemuOpts *opts, void *opaque)
 120 {
 121     NumaOptions *object = NULL;
 122     Error *err = NULL;
 123
 124     {
 125         OptsVisitor *ov = opts_visitor_new(opts);
 126         visit_type_NumaOptions(opts_get_visitor(ov), &object, NULL, &err);
 127         opts_visitor_cleanup(ov);
 128     }
 129
 130     if (err) {
 131         goto error;
 132     }
 133
 134     switch (object->kind) {
 135     case NUMA_OPTIONS_KIND_NODE:
 136         numa_node_parse(object->node, opts, &err);
 137         if (err) {
 138             goto error;
 139         }
 140         nb_numa_nodes++;
 141         break;
 142     default:
 143         abort();
 144     }
 145
 146     return 0;
 147
 148 error:
 149     qerror_report_err(err);
 150     error_free(err);
 151
 152     if (object) {
 153         QapiDeallocVisitor *dv = qapi_dealloc_visitor_new();
 154         visit_type_NumaOptions(qapi_dealloc_get_visitor(dv),
 155                                &object, NULL, NULL);
 156         qapi_dealloc_visitor_cleanup(dv);
 157     }
 158
 159     return -1;
 160 }
 161
 162 void set_numa_nodes(void)
 163 {
 164     int i;
 165
 166     assert(max_numa_nodeid <= MAX_NODES);
 167
 168     /* No support for sparse NUMA node IDs yet: */
 169     for (i = max_numa_nodeid - 1; i >= 0; i--) {
 170         /* Report large node IDs first, to make mistakes easier to spot */
 171         if (!numa_info[i].present) {
 172             error_report("numa: Node ID missing: %d", i);
 173             exit(1);
 174         }
 175     }
 176
 177     /* This must be always true if all nodes are present: */
 178     assert(nb_numa_nodes == max_numa_nodeid);
 179
 180     if (nb_numa_nodes > 0) {
 181         uint64_t numa_total;
 182
 183         if (nb_numa_nodes > MAX_NODES) {
 184             nb_numa_nodes = MAX_NODES;
 185         }
 186
 187         /* If no memory size is given for any node, assume the default case
 188          * and distribute the available memory equally across all nodes
 189          */
 190         for (i = 0; i < nb_numa_nodes; i++) {
 191             if (numa_info[i].node_mem != 0) {
 192                 break;
 193             }
 194         }
 195         if (i == nb_numa_nodes) {
 196             uint64_t usedmem = 0;
 197
 198             /* On Linux, each node's border has to be 8MB aligned,
 199              * the final node gets the rest.
 200              */
 201             for (i = 0; i < nb_numa_nodes - 1; i++) {
 202                 numa_info[i].node_mem = (ram_size / nb_numa_nodes) &
 203                                         ~((1 << 23UL) - 1);
 204                 usedmem += numa_info[i].node_mem;
 205             }
 206             numa_info[i].node_mem = ram_size - usedmem;
 207         }
 208
 209         numa_total = 0;
 210         for (i = 0; i < nb_numa_nodes; i++) {
 211             numa_total += numa_info[i].node_mem;
 212         }
 213         if (numa_total != ram_size) {
 214             error_report("total memory for NUMA nodes (0x%" PRIx64 ")"
 215                          " should equal RAM size (0x" RAM_ADDR_FMT ")",
 216                          numa_total, ram_size);
 217             exit(1);
 218         }
 219
 220         for (i = 0; i < nb_numa_nodes; i++) {
 221             if (!bitmap_empty(numa_info[i].node_cpu, MAX_CPUMASK_BITS)) {
 222                 break;
 223             }
 224         }
 225         /* assigning the VCPUs round-robin is easier to implement, guest OSes
 226          * must cope with this anyway, because there are BIOSes out there in
 227          * real machines which also use this scheme.
 228          */
 229         if (i == nb_numa_nodes) {
 230             for (i = 0; i < max_cpus; i++) {
 231                 set_bit(i, numa_info[i % nb_numa_nodes].node_cpu);
 232             }
 233         }
 234     }
 235 }
 236
 237 void set_numa_modes(void)
 238 {
 239     CPUState *cpu;
 240     int i;
 241
 242     CPU_FOREACH(cpu) {
 243         for (i = 0; i < nb_numa_nodes; i++) {
 244             if (test_bit(cpu->cpu_index, numa_info[i].node_cpu)) {
 245                 cpu->numa_node = i;
 246             }
 247         }
 248     }
 249 }
 250
 251 static void allocate_system_memory_nonnuma(MemoryRegion *mr, Object *owner,
 252                                            const char *name,
 253                                            uint64_t ram_size)
 254 {
 255     if (mem_path) {
 256 #ifdef __linux__
 257         Error *err = NULL;
 258         memory_region_init_ram_from_file(mr, owner, name, ram_size, false,
 259                                          mem_path, &err);
 260
 261         /* Legacy behavior: if allocation failed, fall back to
 262          * regular RAM allocation.
 263          */
 264         if (err) {
 265             qerror_report_err(err);
 266             error_free(err);
 267             memory_region_init_ram(mr, owner, name, ram_size, &error_abort);
 268         }
 269 #else
 270         fprintf(stderr, "-mem-path not supported on this host\n");
 271         exit(1);
 272 #endif
 273     } else {
 274         memory_region_init_ram(mr, owner, name, ram_size, &error_abort);
 275     }
 276     vmstate_register_ram_global(mr);
 277 }
 278
 279 void memory_region_allocate_system_memory(MemoryRegion *mr, Object *owner,
 280                                           const char *name,
 281                                           uint64_t ram_size)
 282 {
 283     uint64_t addr = 0;
 284     int i;
 285
 286     if (nb_numa_nodes == 0 || !have_memdevs) {
 287         allocate_system_memory_nonnuma(mr, owner, name, ram_size);
 288         return;
 289     }
 290
 291     memory_region_init(mr, owner, name, ram_size);
 292     for (i = 0; i < MAX_NODES; i++) {
 293         Error *local_err = NULL;
 294         uint64_t size = numa_info[i].node_mem;
 295         HostMemoryBackend *backend = numa_info[i].node_memdev;
 296         if (!backend) {
 297             continue;
 298         }
 299         MemoryRegion *seg = host_memory_backend_get_memory(backend, &local_err);
 300         if (local_err) {
 301             qerror_report_err(local_err);
 302             exit(1);
 303         }
 304
 305         if (memory_region_is_mapped(seg)) {
 306             char *path = object_get_canonical_path_component(OBJECT(backend));
 307             error_report("memory backend %s is used multiple times. Each "
 308                          "-numa option must use a different memdev value.",
 309                          path);
 310             exit(1);
 311         }
 312
 313         memory_region_add_subregion(mr, addr, seg);
 314         vmstate_register_ram_global(seg);
 315         addr += size;
 316     }
 317 }
 318
 319 static void numa_stat_memory_devices(uint64_t node_mem[])
 320 {
 321     MemoryDeviceInfoList *info_list = NULL;
 322     MemoryDeviceInfoList **prev = &info_list;
 323     MemoryDeviceInfoList *info;
 324
 325     qmp_pc_dimm_device_list(qdev_get_machine(), &prev);
 326     for (info = info_list; info; info = info->next) {
 327         MemoryDeviceInfo *value = info->value;
 328
 329         if (value) {
 330             switch (value->kind) {
 331             case MEMORY_DEVICE_INFO_KIND_DIMM:
 332                 node_mem[value->dimm->node] += value->dimm->size;
 333                 break;
 334             default:
 335                 break;
 336             }
 337         }
 338     }
 339     qapi_free_MemoryDeviceInfoList(info_list);
 340 }
 341
 342 void query_numa_node_mem(uint64_t node_mem[])
 343 {
 344     int i;
 345
 346     if (nb_numa_nodes <= 0) {
 347         return;
 348     }
 349
 350     numa_stat_memory_devices(node_mem);
 351     for (i = 0; i < nb_numa_nodes; i++) {
 352         node_mem[i] += numa_info[i].node_mem;
 353     }
 354 }
 355
 356 static int query_memdev(Object *obj, void *opaque)
 357 {
 358     MemdevList **list = opaque;
 359     MemdevList *m = NULL;
 360     Error *err = NULL;
 361
 362     if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) {
 363         m = g_malloc0(sizeof(*m));
 364
 365         m->value = g_malloc0(sizeof(*m->value));
 366
 367         m->value->size = object_property_get_int(obj, "size",
 368                                                  &err);
 369         if (err) {
 370             goto error;
 371         }
 372
 373         m->value->merge = object_property_get_bool(obj, "merge",
 374                                                    &err);
 375         if (err) {
 376             goto error;
 377         }
 378
 379         m->value->dump = object_property_get_bool(obj, "dump",
 380                                                   &err);
 381         if (err) {
 382             goto error;
 383         }
 384
 385         m->value->prealloc = object_property_get_bool(obj,
 386                                                       "prealloc", &err);
 387         if (err) {
 388             goto error;
 389         }
 390
 391         m->value->policy = object_property_get_enum(obj,
 392                                                     "policy",
 393                                                     HostMemPolicy_lookup,
 394                                                     &err);
 395         if (err) {
 396             goto error;
 397         }
 398
 399         object_property_get_uint16List(obj, "host-nodes",
 400                                        &m->value->host_nodes, &err);
 401         if (err) {
 402             goto error;
 403         }
 404
 405         m->next = *list;
 406         *list = m;
 407     }
 408
 409     return 0;
 410 error:
 411     g_free(m->value);
 412     g_free(m);
 413
 414     return -1;
 415 }
 416
 417 MemdevList *qmp_query_memdev(Error **errp)
 418 {
 419     Object *obj;
 420     MemdevList *list = NULL;
 421
 422     obj = object_resolve_path("/objects", NULL);
 423     if (obj == NULL) {
 424         return NULL;
 425     }
 426
 427     if (object_child_foreach(obj, query_memdev, &list) != 0) {
 428         goto error;
 429     }
 430
 431     return list;
 432
 433 error:
 434     qapi_free_MemdevList(list);
 435     return NULL;
 436 }