backends/hostmem.c

   1 /*
   2  * QEMU Host Memory Backend
   3  *
   4  * Copyright (C) 2013-2014 Red Hat Inc
   5  *
   6  * Authors:
   7  *   Igor Mammedov <imammedo@redhat.com>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2 or later.
  10  * See the COPYING file in the top-level directory.
  11  */
  12
  13 #include "qemu/osdep.h"
  14 #include "sysemu/hostmem.h"
  15 #include "sysemu/sysemu.h"
  16 #include "hw/boards.h"
  17 #include "qapi/error.h"
  18 #include "qapi/qapi-builtin-visit.h"
  19 #include "qapi/visitor.h"
  20 #include "qemu/config-file.h"
  21 #include "qom/object_interfaces.h"
  22 #include "qemu/mmap-alloc.h"
  23
  24 #ifdef CONFIG_NUMA
  25 #include <numaif.h>
  26 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_DEFAULT != MPOL_DEFAULT);
  27 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_PREFERRED != MPOL_PREFERRED);
  28 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_BIND != MPOL_BIND);
  29 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_INTERLEAVE != MPOL_INTERLEAVE);
  30 #endif
  31
  32 char *
  33 host_memory_backend_get_name(HostMemoryBackend *backend)
  34 {
  35     if (!backend->use_canonical_path) {
  36         return object_get_canonical_path_component(OBJECT(backend));
  37     }
  38
  39     return object_get_canonical_path(OBJECT(backend));
  40 }
  41
  42 static void
  43 host_memory_backend_get_size(Object *obj, Visitor *v, const char *name,
  44                              void *opaque, Error **errp)
  45 {
  46     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
  47     uint64_t value = backend->size;
  48
  49     visit_type_size(v, name, &value, errp);
  50 }
  51
  52 static void
  53 host_memory_backend_set_size(Object *obj, Visitor *v, const char *name,
  54                              void *opaque, Error **errp)
  55 {
  56     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
  57     uint64_t value;
  58
  59     if (host_memory_backend_mr_inited(backend)) {
  60         error_setg(errp, "cannot change property %s of %s ", name,
  61                    object_get_typename(obj));
  62         return;
  63     }
  64
  65     if (!visit_type_size(v, name, &value, errp)) {
  66         return;
  67     }
  68     if (!value) {
  69         error_setg(errp,
  70                    "property '%s' of %s doesn't take value '%" PRIu64 "'",
  71                    name, object_get_typename(obj), value);
  72         return;
  73     }
  74     backend->size = value;
  75 }
  76
  77 static void
  78 host_memory_backend_get_host_nodes(Object *obj, Visitor *v, const char *name,
  79                                    void *opaque, Error **errp)
  80 {
  81     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
  82     uint16List *host_nodes = NULL;
  83     uint16List **node = &host_nodes;
  84     unsigned long value;
  85
  86     value = find_first_bit(backend->host_nodes, MAX_NODES);
  87     if (value == MAX_NODES) {
  88         goto ret;
  89     }
  90
  91     *node = g_malloc0(sizeof(**node));
  92     (*node)->value = value;
  93     node = &(*node)->next;
  94
  95     do {
  96         value = find_next_bit(backend->host_nodes, MAX_NODES, value + 1);
  97         if (value == MAX_NODES) {
  98             break;
  99         }
 100
 101         *node = g_malloc0(sizeof(**node));
 102         (*node)->value = value;
 103         node = &(*node)->next;
 104     } while (true);
 105
 106 ret:
 107     visit_type_uint16List(v, name, &host_nodes, errp);
 108 }
 109
 110 static void
 111 host_memory_backend_set_host_nodes(Object *obj, Visitor *v, const char *name,
 112                                    void *opaque, Error **errp)
 113 {
 114 #ifdef CONFIG_NUMA
 115     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 116     uint16List *l, *host_nodes = NULL;
 117
 118     visit_type_uint16List(v, name, &host_nodes, errp);
 119
 120     for (l = host_nodes; l; l = l->next) {
 121         if (l->value >= MAX_NODES) {
 122             error_setg(errp, "Invalid host-nodes value: %d", l->value);
 123             goto out;
 124         }
 125     }
 126
 127     for (l = host_nodes; l; l = l->next) {
 128         bitmap_set(backend->host_nodes, l->value, 1);
 129     }
 130
 131 out:
 132     qapi_free_uint16List(host_nodes);
 133 #else
 134     error_setg(errp, "NUMA node binding are not supported by this QEMU");
 135 #endif
 136 }
 137
 138 static int
 139 host_memory_backend_get_policy(Object *obj, Error **errp G_GNUC_UNUSED)
 140 {
 141     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 142     return backend->policy;
 143 }
 144
 145 static void
 146 host_memory_backend_set_policy(Object *obj, int policy, Error **errp)
 147 {
 148     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 149     backend->policy = policy;
 150
 151 #ifndef CONFIG_NUMA
 152     if (policy != HOST_MEM_POLICY_DEFAULT) {
 153         error_setg(errp, "NUMA policies are not supported by this QEMU");
 154     }
 155 #endif
 156 }
 157
 158 static bool host_memory_backend_get_merge(Object *obj, Error **errp)
 159 {
 160     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 161
 162     return backend->merge;
 163 }
 164
 165 static void host_memory_backend_set_merge(Object *obj, bool value, Error **errp)
 166 {
 167     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 168
 169     if (!host_memory_backend_mr_inited(backend)) {
 170         backend->merge = value;
 171         return;
 172     }
 173
 174     if (value != backend->merge) {
 175         void *ptr = memory_region_get_ram_ptr(&backend->mr);
 176         uint64_t sz = memory_region_size(&backend->mr);
 177
 178         qemu_madvise(ptr, sz,
 179                      value ? QEMU_MADV_MERGEABLE : QEMU_MADV_UNMERGEABLE);
 180         backend->merge = value;
 181     }
 182 }
 183
 184 static bool host_memory_backend_get_dump(Object *obj, Error **errp)
 185 {
 186     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 187
 188     return backend->dump;
 189 }
 190
 191 static void host_memory_backend_set_dump(Object *obj, bool value, Error **errp)
 192 {
 193     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 194
 195     if (!host_memory_backend_mr_inited(backend)) {
 196         backend->dump = value;
 197         return;
 198     }
 199
 200     if (value != backend->dump) {
 201         void *ptr = memory_region_get_ram_ptr(&backend->mr);
 202         uint64_t sz = memory_region_size(&backend->mr);
 203
 204         qemu_madvise(ptr, sz,
 205                      value ? QEMU_MADV_DODUMP : QEMU_MADV_DONTDUMP);
 206         backend->dump = value;
 207     }
 208 }
 209
 210 static bool host_memory_backend_get_prealloc(Object *obj, Error **errp)
 211 {
 212     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 213
 214     return backend->prealloc;
 215 }
 216
 217 static void host_memory_backend_set_prealloc(Object *obj, bool value,
 218                                              Error **errp)
 219 {
 220     Error *local_err = NULL;
 221     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 222
 223     if (!host_memory_backend_mr_inited(backend)) {
 224         backend->prealloc = value;
 225         return;
 226     }
 227
 228     if (value && !backend->prealloc) {
 229         int fd = memory_region_get_fd(&backend->mr);
 230         void *ptr = memory_region_get_ram_ptr(&backend->mr);
 231         uint64_t sz = memory_region_size(&backend->mr);
 232
 233         os_mem_prealloc(fd, ptr, sz, backend->prealloc_threads, &local_err);
 234         if (local_err) {
 235             error_propagate(errp, local_err);
 236             return;
 237         }
 238         backend->prealloc = true;
 239     }
 240 }
 241
 242 static void host_memory_backend_get_prealloc_threads(Object *obj, Visitor *v,
 243     const char *name, void *opaque, Error **errp)
 244 {
 245     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 246     visit_type_uint32(v, name, &backend->prealloc_threads, errp);
 247 }
 248
 249 static void host_memory_backend_set_prealloc_threads(Object *obj, Visitor *v,
 250     const char *name, void *opaque, Error **errp)
 251 {
 252     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 253     uint32_t value;
 254
 255     if (!visit_type_uint32(v, name, &value, errp)) {
 256         return;
 257     }
 258     if (value <= 0) {
 259         error_setg(errp, "property '%s' of %s doesn't take value '%d'", name,
 260                    object_get_typename(obj), value);
 261         return;
 262     }
 263     backend->prealloc_threads = value;
 264 }
 265
 266 static void host_memory_backend_init(Object *obj)
 267 {
 268     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 269     MachineState *machine = MACHINE(qdev_get_machine());
 270
 271     /* TODO: convert access to globals to compat properties */
 272     backend->merge = machine_mem_merge(machine);
 273     backend->dump = machine_dump_guest_core(machine);
 274     backend->prealloc_threads = 1;
 275 }
 276
 277 static void host_memory_backend_post_init(Object *obj)
 278 {
 279     object_apply_compat_props(obj);
 280 }
 281
 282 bool host_memory_backend_mr_inited(HostMemoryBackend *backend)
 283 {
 284     /*
 285      * NOTE: We forbid zero-length memory backend, so here zero means
 286      * "we haven't inited the backend memory region yet".
 287      */
 288     return memory_region_size(&backend->mr) != 0;
 289 }
 290
 291 MemoryRegion *host_memory_backend_get_memory(HostMemoryBackend *backend)
 292 {
 293     return host_memory_backend_mr_inited(backend) ? &backend->mr : NULL;
 294 }
 295
 296 void host_memory_backend_set_mapped(HostMemoryBackend *backend, bool mapped)
 297 {
 298     backend->is_mapped = mapped;
 299 }
 300
 301 bool host_memory_backend_is_mapped(HostMemoryBackend *backend)
 302 {
 303     return backend->is_mapped;
 304 }
 305
 306 #ifdef __linux__
 307 size_t host_memory_backend_pagesize(HostMemoryBackend *memdev)
 308 {
 309     Object *obj = OBJECT(memdev);
 310     char *path = object_property_get_str(obj, "mem-path", NULL);
 311     size_t pagesize = qemu_mempath_getpagesize(path);
 312
 313     g_free(path);
 314     return pagesize;
 315 }
 316 #else
 317 size_t host_memory_backend_pagesize(HostMemoryBackend *memdev)
 318 {
 319     return qemu_real_host_page_size;
 320 }
 321 #endif
 322
 323 static void
 324 host_memory_backend_memory_complete(UserCreatable *uc, Error **errp)
 325 {
 326     HostMemoryBackend *backend = MEMORY_BACKEND(uc);
 327     HostMemoryBackendClass *bc = MEMORY_BACKEND_GET_CLASS(uc);
 328     Error *local_err = NULL;
 329     void *ptr;
 330     uint64_t sz;
 331
 332     if (bc->alloc) {
 333         bc->alloc(backend, &local_err);
 334         if (local_err) {
 335             goto out;
 336         }
 337
 338         ptr = memory_region_get_ram_ptr(&backend->mr);
 339         sz = memory_region_size(&backend->mr);
 340
 341         if (backend->merge) {
 342             qemu_madvise(ptr, sz, QEMU_MADV_MERGEABLE);
 343         }
 344         if (!backend->dump) {
 345             qemu_madvise(ptr, sz, QEMU_MADV_DONTDUMP);
 346         }
 347 #ifdef CONFIG_NUMA
 348         unsigned long lastbit = find_last_bit(backend->host_nodes, MAX_NODES);
 349         /* lastbit == MAX_NODES means maxnode = 0 */
 350         unsigned long maxnode = (lastbit + 1) % (MAX_NODES + 1);
 351         /* ensure policy won't be ignored in case memory is preallocated
 352          * before mbind(). note: MPOL_MF_STRICT is ignored on hugepages so
 353          * this doesn't catch hugepage case. */
 354         unsigned flags = MPOL_MF_STRICT | MPOL_MF_MOVE;
 355
 356         /* check for invalid host-nodes and policies and give more verbose
 357          * error messages than mbind(). */
 358         if (maxnode && backend->policy == MPOL_DEFAULT) {
 359             error_setg(errp, "host-nodes must be empty for policy default,"
 360                        " or you should explicitly specify a policy other"
 361                        " than default");
 362             return;
 363         } else if (maxnode == 0 && backend->policy != MPOL_DEFAULT) {
 364             error_setg(errp, "host-nodes must be set for policy %s",
 365                        HostMemPolicy_str(backend->policy));
 366             return;
 367         }
 368
 369         /* We can have up to MAX_NODES nodes, but we need to pass maxnode+1
 370          * as argument to mbind() due to an old Linux bug (feature?) which
 371          * cuts off the last specified node. This means backend->host_nodes
 372          * must have MAX_NODES+1 bits available.
 373          */
 374         assert(sizeof(backend->host_nodes) >=
 375                BITS_TO_LONGS(MAX_NODES + 1) * sizeof(unsigned long));
 376         assert(maxnode <= MAX_NODES);
 377
 378         if (maxnode &&
 379             mbind(ptr, sz, backend->policy, backend->host_nodes, maxnode + 1,
 380                   flags)) {
 381             if (backend->policy != MPOL_DEFAULT || errno != ENOSYS) {
 382                 error_setg_errno(errp, errno,
 383                                  "cannot bind memory to host NUMA nodes");
 384                 return;
 385             }
 386         }
 387 #endif
 388         /* Preallocate memory after the NUMA policy has been instantiated.
 389          * This is necessary to guarantee memory is allocated with
 390          * specified NUMA policy in place.
 391          */
 392         if (backend->prealloc) {
 393             os_mem_prealloc(memory_region_get_fd(&backend->mr), ptr, sz,
 394                             backend->prealloc_threads, &local_err);
 395             if (local_err) {
 396                 goto out;
 397             }
 398         }
 399     }
 400 out:
 401     error_propagate(errp, local_err);
 402 }
 403
 404 static bool
 405 host_memory_backend_can_be_deleted(UserCreatable *uc)
 406 {
 407     if (host_memory_backend_is_mapped(MEMORY_BACKEND(uc))) {
 408         return false;
 409     } else {
 410         return true;
 411     }
 412 }
 413
 414 static bool host_memory_backend_get_share(Object *o, Error **errp)
 415 {
 416     HostMemoryBackend *backend = MEMORY_BACKEND(o);
 417
 418     return backend->share;
 419 }
 420
 421 static void host_memory_backend_set_share(Object *o, bool value, Error **errp)
 422 {
 423     HostMemoryBackend *backend = MEMORY_BACKEND(o);
 424
 425     if (host_memory_backend_mr_inited(backend)) {
 426         error_setg(errp, "cannot change property value");
 427         return;
 428     }
 429     backend->share = value;
 430 }
 431
 432 static bool
 433 host_memory_backend_get_use_canonical_path(Object *obj, Error **errp)
 434 {
 435     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 436
 437     return backend->use_canonical_path;
 438 }
 439
 440 static void
 441 host_memory_backend_set_use_canonical_path(Object *obj, bool value,
 442                                            Error **errp)
 443 {
 444     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 445
 446     backend->use_canonical_path = value;
 447 }
 448
 449 static void
 450 host_memory_backend_class_init(ObjectClass *oc, void *data)
 451 {
 452     UserCreatableClass *ucc = USER_CREATABLE_CLASS(oc);
 453
 454     ucc->complete = host_memory_backend_memory_complete;
 455     ucc->can_be_deleted = host_memory_backend_can_be_deleted;
 456
 457     object_class_property_add_bool(oc, "merge",
 458         host_memory_backend_get_merge,
 459         host_memory_backend_set_merge);
 460     object_class_property_set_description(oc, "merge",
 461         "Mark memory as mergeable");
 462     object_class_property_add_bool(oc, "dump",
 463         host_memory_backend_get_dump,
 464         host_memory_backend_set_dump);
 465     object_class_property_set_description(oc, "dump",
 466         "Set to 'off' to exclude from core dump");
 467     object_class_property_add_bool(oc, "prealloc",
 468         host_memory_backend_get_prealloc,
 469         host_memory_backend_set_prealloc);
 470     object_class_property_set_description(oc, "prealloc",
 471         "Preallocate memory");
 472     object_class_property_add(oc, "prealloc-threads", "int",
 473         host_memory_backend_get_prealloc_threads,
 474         host_memory_backend_set_prealloc_threads,
 475         NULL, NULL);
 476     object_class_property_set_description(oc, "prealloc-threads",
 477         "Number of CPU threads to use for prealloc");
 478     object_class_property_add(oc, "size", "int",
 479         host_memory_backend_get_size,
 480         host_memory_backend_set_size,
 481         NULL, NULL);
 482     object_class_property_set_description(oc, "size",
 483         "Size of the memory region (ex: 500M)");
 484     object_class_property_add(oc, "host-nodes", "int",
 485         host_memory_backend_get_host_nodes,
 486         host_memory_backend_set_host_nodes,
 487         NULL, NULL);
 488     object_class_property_set_description(oc, "host-nodes",
 489         "Binds memory to the list of NUMA host nodes");
 490     object_class_property_add_enum(oc, "policy", "HostMemPolicy",
 491         &HostMemPolicy_lookup,
 492         host_memory_backend_get_policy,
 493         host_memory_backend_set_policy);
 494     object_class_property_set_description(oc, "policy",
 495         "Set the NUMA policy");
 496     object_class_property_add_bool(oc, "share",
 497         host_memory_backend_get_share, host_memory_backend_set_share);
 498     object_class_property_set_description(oc, "share",
 499         "Mark the memory as private to QEMU or shared");
 500     object_class_property_add_bool(oc, "x-use-canonical-path-for-ramblock-id",
 501         host_memory_backend_get_use_canonical_path,
 502         host_memory_backend_set_use_canonical_path);
 503 }
 504
 505 static const TypeInfo host_memory_backend_info = {
 506     .name = TYPE_MEMORY_BACKEND,
 507     .parent = TYPE_OBJECT,
 508     .abstract = true,
 509     .class_size = sizeof(HostMemoryBackendClass),
 510     .class_init = host_memory_backend_class_init,
 511     .instance_size = sizeof(HostMemoryBackend),
 512     .instance_init = host_memory_backend_init,
 513     .instance_post_init = host_memory_backend_post_init,
 514     .interfaces = (InterfaceInfo[]) {
 515         { TYPE_USER_CREATABLE },
 516         { }
 517     }
 518 };
 519
 520 static void register_types(void)
 521 {
 522     type_register_static(&host_memory_backend_info);
 523 }
 524
 525 type_init(register_types);