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
  39 QemuOptsList qemu_numa_opts = {
  40     .name = "numa",
  41     .implied_opt_name = "type",
  42     .head = QTAILQ_HEAD_INITIALIZER(qemu_numa_opts.head),
  43     .desc = { { 0 } } /* validated with OptsVisitor */
  44 };
  45
  46 static int have_memdevs = -1;
  47
  48 static void numa_node_parse(NumaNodeOptions *node, QemuOpts *opts, Error **errp)
  49 {
  50     uint16_t nodenr;
  51     uint16List *cpus = NULL;
  52
  53     if (node->has_nodeid) {
  54         nodenr = node->nodeid;
  55     } else {
  56         nodenr = nb_numa_nodes;
  57     }
  58
  59     if (nodenr >= MAX_NODES) {
  60         error_setg(errp, "Max number of NUMA nodes reached: %"
  61                    PRIu16 "\n", nodenr);
  62         return;
  63     }
  64
  65     if (numa_info[nodenr].present) {
  66         error_setg(errp, "Duplicate NUMA nodeid: %" PRIu16, nodenr);
  67         return;
  68     }
  69
  70     for (cpus = node->cpus; cpus; cpus = cpus->next) {
  71         if (cpus->value > MAX_CPUMASK_BITS) {
  72             error_setg(errp, "CPU number %" PRIu16 " is bigger than %d",
  73                        cpus->value, MAX_CPUMASK_BITS);
  74             return;
  75         }
  76         bitmap_set(numa_info[nodenr].node_cpu, cpus->value, 1);
  77     }
  78
  79     if (node->has_mem && node->has_memdev) {
  80         error_setg(errp, "qemu: cannot specify both mem= and memdev=\n");
  81         return;
  82     }
  83
  84     if (have_memdevs == -1) {
  85         have_memdevs = node->has_memdev;
  86     }
  87     if (node->has_memdev != have_memdevs) {
  88         error_setg(errp, "qemu: memdev option must be specified for either "
  89                    "all or no nodes\n");
  90         return;
  91     }
  92
  93     if (node->has_mem) {
  94         uint64_t mem_size = node->mem;
  95         const char *mem_str = qemu_opt_get(opts, "mem");
  96         /* Fix up legacy suffix-less format */
  97         if (g_ascii_isdigit(mem_str[strlen(mem_str) - 1])) {
  98             mem_size <<= 20;
  99         }
 100         numa_info[nodenr].node_mem = mem_size;
 101     }
 102     if (node->has_memdev) {
 103         Object *o;
 104         o = object_resolve_path_type(node->memdev, TYPE_MEMORY_BACKEND, NULL);
 105         if (!o) {
 106             error_setg(errp, "memdev=%s is ambiguous", node->memdev);
 107             return;
 108         }
 109
 110         object_ref(o);
 111         numa_info[nodenr].node_mem = object_property_get_int(o, "size", NULL);
 112         numa_info[nodenr].node_memdev = MEMORY_BACKEND(o);
 113     }
 114     numa_info[nodenr].present = true;
 115     max_numa_nodeid = MAX(max_numa_nodeid, nodenr + 1);
 116 }
 117
 118 int numa_init_func(QemuOpts *opts, void *opaque)
 119 {
 120     NumaOptions *object = NULL;
 121     Error *err = NULL;
 122
 123     {
 124         OptsVisitor *ov = opts_visitor_new(opts);
 125         visit_type_NumaOptions(opts_get_visitor(ov), &object, NULL, &err);
 126         opts_visitor_cleanup(ov);
 127     }
 128
 129     if (err) {
 130         goto error;
 131     }
 132
 133     switch (object->kind) {
 134     case NUMA_OPTIONS_KIND_NODE:
 135         numa_node_parse(object->node, opts, &err);
 136         if (err) {
 137             goto error;
 138         }
 139         nb_numa_nodes++;
 140         break;
 141     default:
 142         abort();
 143     }
 144
 145     return 0;
 146
 147 error:
 148     qerror_report_err(err);
 149     error_free(err);
 150
 151     if (object) {
 152         QapiDeallocVisitor *dv = qapi_dealloc_visitor_new();
 153         visit_type_NumaOptions(qapi_dealloc_get_visitor(dv),
 154                                &object, NULL, NULL);
 155         qapi_dealloc_visitor_cleanup(dv);
 156     }
 157
 158     return -1;
 159 }
 160
 161 void set_numa_nodes(void)
 162 {
 163     int i;
 164
 165     assert(max_numa_nodeid <= MAX_NODES);
 166
 167     /* No support for sparse NUMA node IDs yet: */
 168     for (i = max_numa_nodeid - 1; i >= 0; i--) {
 169         /* Report large node IDs first, to make mistakes easier to spot */
 170         if (!numa_info[i].present) {
 171             error_report("numa: Node ID missing: %d", i);
 172             exit(1);
 173         }
 174     }
 175
 176     /* This must be always true if all nodes are present: */
 177     assert(nb_numa_nodes == max_numa_nodeid);
 178
 179     if (nb_numa_nodes > 0) {
 180         uint64_t numa_total;
 181
 182         if (nb_numa_nodes > MAX_NODES) {
 183             nb_numa_nodes = MAX_NODES;
 184         }
 185
 186         /* If no memory size is given for any node, assume the default case
 187          * and distribute the available memory equally across all nodes
 188          */
 189         for (i = 0; i < nb_numa_nodes; i++) {
 190             if (numa_info[i].node_mem != 0) {
 191                 break;
 192             }
 193         }
 194         if (i == nb_numa_nodes) {
 195             uint64_t usedmem = 0;
 196
 197             /* On Linux, each node's border has to be 8MB aligned,
 198              * the final node gets the rest.
 199              */
 200             for (i = 0; i < nb_numa_nodes - 1; i++) {
 201                 numa_info[i].node_mem = (ram_size / nb_numa_nodes) &
 202                                         ~((1 << 23UL) - 1);
 203                 usedmem += numa_info[i].node_mem;
 204             }
 205             numa_info[i].node_mem = ram_size - usedmem;
 206         }
 207
 208         numa_total = 0;
 209         for (i = 0; i < nb_numa_nodes; i++) {
 210             numa_total += numa_info[i].node_mem;
 211         }
 212         if (numa_total != ram_size) {
 213             error_report("total memory for NUMA nodes (%" PRIu64 ")"
 214                          " should equal RAM size (" RAM_ADDR_FMT ")",
 215                          numa_total, ram_size);
 216             exit(1);
 217         }
 218
 219         for (i = 0; i < nb_numa_nodes; i++) {
 220             if (!bitmap_empty(numa_info[i].node_cpu, MAX_CPUMASK_BITS)) {
 221                 break;
 222             }
 223         }
 224         /* assigning the VCPUs round-robin is easier to implement, guest OSes
 225          * must cope with this anyway, because there are BIOSes out there in
 226          * real machines which also use this scheme.
 227          */
 228         if (i == nb_numa_nodes) {
 229             for (i = 0; i < max_cpus; i++) {
 230                 set_bit(i, numa_info[i % nb_numa_nodes].node_cpu);
 231             }
 232         }
 233     }
 234 }
 235
 236 void set_numa_modes(void)
 237 {
 238     CPUState *cpu;
 239     int i;
 240
 241     CPU_FOREACH(cpu) {
 242         for (i = 0; i < nb_numa_nodes; i++) {
 243             if (test_bit(cpu->cpu_index, numa_info[i].node_cpu)) {
 244                 cpu->numa_node = i;
 245             }
 246         }
 247     }
 248 }
 249
 250 static void allocate_system_memory_nonnuma(MemoryRegion *mr, Object *owner,
 251                                            const char *name,
 252                                            uint64_t ram_size)
 253 {
 254     if (mem_path) {
 255 #ifdef __linux__
 256         Error *err = NULL;
 257         memory_region_init_ram_from_file(mr, owner, name, ram_size, false,
 258                                          mem_path, &err);
 259
 260         /* Legacy behavior: if allocation failed, fall back to
 261          * regular RAM allocation.
 262          */
 263         if (err) {
 264             qerror_report_err(err);
 265             error_free(err);
 266             memory_region_init_ram(mr, owner, name, ram_size);
 267         }
 268 #else
 269         fprintf(stderr, "-mem-path not supported on this host\n");
 270         exit(1);
 271 #endif
 272     } else {
 273         memory_region_init_ram(mr, owner, name, ram_size);
 274     }
 275     vmstate_register_ram_global(mr);
 276 }
 277
 278 void memory_region_allocate_system_memory(MemoryRegion *mr, Object *owner,
 279                                           const char *name,
 280                                           uint64_t ram_size)
 281 {
 282     uint64_t addr = 0;
 283     int i;
 284
 285     if (nb_numa_nodes == 0 || !have_memdevs) {
 286         allocate_system_memory_nonnuma(mr, owner, name, ram_size);
 287         return;
 288     }
 289
 290     memory_region_init(mr, owner, name, ram_size);
 291     for (i = 0; i < MAX_NODES; i++) {
 292         Error *local_err = NULL;
 293         uint64_t size = numa_info[i].node_mem;
 294         HostMemoryBackend *backend = numa_info[i].node_memdev;
 295         if (!backend) {
 296             continue;
 297         }
 298         MemoryRegion *seg = host_memory_backend_get_memory(backend, &local_err);
 299         if (local_err) {
 300             qerror_report_err(local_err);
 301             exit(1);
 302         }
 303
 304         if (memory_region_is_mapped(seg)) {
 305             char *path = object_get_canonical_path_component(OBJECT(backend));
 306             error_report("memory backend %s is used multiple times. Each "
 307                          "-numa option must use a different memdev value.",
 308                          path);
 309             exit(1);
 310         }
 311
 312         memory_region_add_subregion(mr, addr, seg);
 313         vmstate_register_ram_global(seg);
 314         addr += size;
 315     }
 316 }
 317
 318 static int query_memdev(Object *obj, void *opaque)
 319 {
 320     MemdevList **list = opaque;
 321     Error *err = NULL;
 322
 323     if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) {
 324         MemdevList *m = g_malloc0(sizeof(*m));
 325
 326         m->value = g_malloc0(sizeof(*m->value));
 327
 328         m->value->size = object_property_get_int(obj, "size",
 329                                                  &err);
 330         if (err) {
 331             goto error;
 332         }
 333
 334         m->value->merge = object_property_get_bool(obj, "merge",
 335                                                    &err);
 336         if (err) {
 337             goto error;
 338         }
 339
 340         m->value->dump = object_property_get_bool(obj, "dump",
 341                                                   &err);
 342         if (err) {
 343             goto error;
 344         }
 345
 346         m->value->prealloc = object_property_get_bool(obj,
 347                                                       "prealloc", &err);
 348         if (err) {
 349             goto error;
 350         }
 351
 352         m->value->policy = object_property_get_enum(obj,
 353                                                     "policy",
 354                                                     HostMemPolicy_lookup,
 355                                                     &err);
 356         if (err) {
 357             goto error;
 358         }
 359
 360         object_property_get_uint16List(obj, "host-nodes",
 361                                        &m->value->host_nodes, &err);
 362         if (err) {
 363             goto error;
 364         }
 365
 366         m->next = *list;
 367         *list = m;
 368     }
 369
 370     return 0;
 371 error:
 372     return -1;
 373 }
 374
 375 MemdevList *qmp_query_memdev(Error **errp)
 376 {
 377     Object *obj;
 378     MemdevList *list = NULL, *m;
 379
 380     obj = object_resolve_path("/objects", NULL);
 381     if (obj == NULL) {
 382         return NULL;
 383     }
 384
 385     if (object_child_foreach(obj, query_memdev, &list) != 0) {
 386         goto error;
 387     }
 388
 389     return list;
 390
 391 error:
 392     while (list) {
 393         m = list;
 394         list = list->next;
 395         g_free(m->value);
 396         g_free(m);
 397     }
 398     return NULL;
 399 }