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 "hw/boards.h"
  16 #include "qapi/error.h"
  17 #include "qapi/qapi-builtin-visit.h"
  18 #include "qapi/visitor.h"
  19 #include "qemu/config-file.h"
  20 #include "qom/object_interfaces.h"
  21 #include "qemu/mmap-alloc.h"
  22 #include "qemu/madvise.h"
  23
  24 #ifdef CONFIG_NUMA
  25 #include <numaif.h>
  26 #include <numa.h>
  27 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_DEFAULT != MPOL_DEFAULT);
  28 /*
  29  * HOST_MEM_POLICY_PREFERRED may either translate to MPOL_PREFERRED or
  30  * MPOL_PREFERRED_MANY, see comments further below.
  31  */
  32 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_PREFERRED != MPOL_PREFERRED);
  33 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_BIND != MPOL_BIND);
  34 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_INTERLEAVE != MPOL_INTERLEAVE);
  35 #endif
  36
  37 char *
  38 host_memory_backend_get_name(HostMemoryBackend *backend)
  39 {
  40     if (!backend->use_canonical_path) {
  41         return g_strdup(object_get_canonical_path_component(OBJECT(backend)));
  42     }
  43
  44     return object_get_canonical_path(OBJECT(backend));
  45 }
  46
  47 static void
  48 host_memory_backend_get_size(Object *obj, Visitor *v, const char *name,
  49                              void *opaque, Error **errp)
  50 {
  51     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
  52     uint64_t value = backend->size;
  53
  54     visit_type_size(v, name, &value, errp);
  55 }
  56
  57 static void
  58 host_memory_backend_set_size(Object *obj, Visitor *v, const char *name,
  59                              void *opaque, Error **errp)
  60 {
  61     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
  62     uint64_t value;
  63
  64     if (host_memory_backend_mr_inited(backend)) {
  65         error_setg(errp, "cannot change property %s of %s ", name,
  66                    object_get_typename(obj));
  67         return;
  68     }
  69
  70     if (!visit_type_size(v, name, &value, errp)) {
  71         return;
  72     }
  73     if (!value) {
  74         error_setg(errp,
  75                    "property '%s' of %s doesn't take value '%" PRIu64 "'",
  76                    name, object_get_typename(obj), value);
  77         return;
  78     }
  79     backend->size = value;
  80 }
  81
  82 static void
  83 host_memory_backend_get_host_nodes(Object *obj, Visitor *v, const char *name,
  84                                    void *opaque, Error **errp)
  85 {
  86     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
  87     uint16List *host_nodes = NULL;
  88     uint16List **tail = &host_nodes;
  89     unsigned long value;
  90
  91     value = find_first_bit(backend->host_nodes, MAX_NODES);
  92     if (value == MAX_NODES) {
  93         goto ret;
  94     }
  95
  96     QAPI_LIST_APPEND(tail, value);
  97
  98     do {
  99         value = find_next_bit(backend->host_nodes, MAX_NODES, value + 1);
 100         if (value == MAX_NODES) {
 101             break;
 102         }
 103
 104         QAPI_LIST_APPEND(tail, value);
 105     } while (true);
 106
 107 ret:
 108     visit_type_uint16List(v, name, &host_nodes, errp);
 109     qapi_free_uint16List(host_nodes);
 110 }
 111
 112 static void
 113 host_memory_backend_set_host_nodes(Object *obj, Visitor *v, const char *name,
 114                                    void *opaque, Error **errp)
 115 {
 116 #ifdef CONFIG_NUMA
 117     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 118     uint16List *l, *host_nodes = NULL;
 119
 120     visit_type_uint16List(v, name, &host_nodes, errp);
 121
 122     for (l = host_nodes; l; l = l->next) {
 123         if (l->value >= MAX_NODES) {
 124             error_setg(errp, "Invalid host-nodes value: %d", l->value);
 125             goto out;
 126         }
 127     }
 128
 129     for (l = host_nodes; l; l = l->next) {
 130         bitmap_set(backend->host_nodes, l->value, 1);
 131     }
 132
 133 out:
 134     qapi_free_uint16List(host_nodes);
 135 #else
 136     error_setg(errp, "NUMA node binding are not supported by this QEMU");
 137 #endif
 138 }
 139
 140 static int
 141 host_memory_backend_get_policy(Object *obj, Error **errp G_GNUC_UNUSED)
 142 {
 143     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 144     return backend->policy;
 145 }
 146
 147 static void
 148 host_memory_backend_set_policy(Object *obj, int policy, Error **errp)
 149 {
 150     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 151     backend->policy = policy;
 152
 153 #ifndef CONFIG_NUMA
 154     if (policy != HOST_MEM_POLICY_DEFAULT) {
 155         error_setg(errp, "NUMA policies are not supported by this QEMU");
 156     }
 157 #endif
 158 }
 159
 160 static bool host_memory_backend_get_merge(Object *obj, Error **errp)
 161 {
 162     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 163
 164     return backend->merge;
 165 }
 166
 167 static void host_memory_backend_set_merge(Object *obj, bool value, Error **errp)
 168 {
 169     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 170
 171     if (!host_memory_backend_mr_inited(backend)) {
 172         backend->merge = value;
 173         return;
 174     }
 175
 176     if (value != backend->merge) {
 177         void *ptr = memory_region_get_ram_ptr(&backend->mr);
 178         uint64_t sz = memory_region_size(&backend->mr);
 179
 180         qemu_madvise(ptr, sz,
 181                      value ? QEMU_MADV_MERGEABLE : QEMU_MADV_UNMERGEABLE);
 182         backend->merge = value;
 183     }
 184 }
 185
 186 static bool host_memory_backend_get_dump(Object *obj, Error **errp)
 187 {
 188     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 189
 190     return backend->dump;
 191 }
 192
 193 static void host_memory_backend_set_dump(Object *obj, bool value, Error **errp)
 194 {
 195     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 196
 197     if (!host_memory_backend_mr_inited(backend)) {
 198         backend->dump = value;
 199         return;
 200     }
 201
 202     if (value != backend->dump) {
 203         void *ptr = memory_region_get_ram_ptr(&backend->mr);
 204         uint64_t sz = memory_region_size(&backend->mr);
 205
 206         qemu_madvise(ptr, sz,
 207                      value ? QEMU_MADV_DODUMP : QEMU_MADV_DONTDUMP);
 208         backend->dump = value;
 209     }
 210 }
 211
 212 static bool host_memory_backend_get_prealloc(Object *obj, Error **errp)
 213 {
 214     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 215
 216     return backend->prealloc;
 217 }
 218
 219 static void host_memory_backend_set_prealloc(Object *obj, bool value,
 220                                              Error **errp)
 221 {
 222     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 223
 224     if (!backend->reserve && value) {
 225         error_setg(errp, "'prealloc=on' and 'reserve=off' are incompatible");
 226         return;
 227     }
 228
 229     if (!host_memory_backend_mr_inited(backend)) {
 230         backend->prealloc = value;
 231         return;
 232     }
 233
 234     if (value && !backend->prealloc) {
 235         int fd = memory_region_get_fd(&backend->mr);
 236         void *ptr = memory_region_get_ram_ptr(&backend->mr);
 237         uint64_t sz = memory_region_size(&backend->mr);
 238
 239         if (!qemu_prealloc_mem(fd, ptr, sz, backend->prealloc_threads,
 240                                backend->prealloc_context, errp)) {
 241             return;
 242         }
 243         backend->prealloc = true;
 244     }
 245 }
 246
 247 static void host_memory_backend_get_prealloc_threads(Object *obj, Visitor *v,
 248     const char *name, void *opaque, Error **errp)
 249 {
 250     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 251     visit_type_uint32(v, name, &backend->prealloc_threads, errp);
 252 }
 253
 254 static void host_memory_backend_set_prealloc_threads(Object *obj, Visitor *v,
 255     const char *name, void *opaque, Error **errp)
 256 {
 257     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 258     uint32_t value;
 259
 260     if (!visit_type_uint32(v, name, &value, errp)) {
 261         return;
 262     }
 263     if (value <= 0) {
 264         error_setg(errp, "property '%s' of %s doesn't take value '%d'", name,
 265                    object_get_typename(obj), value);
 266         return;
 267     }
 268     backend->prealloc_threads = value;
 269 }
 270
 271 static void host_memory_backend_init(Object *obj)
 272 {
 273     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 274     MachineState *machine = MACHINE(qdev_get_machine());
 275
 276     /* TODO: convert access to globals to compat properties */
 277     backend->merge = machine_mem_merge(machine);
 278     backend->dump = machine_dump_guest_core(machine);
 279     backend->reserve = true;
 280     backend->prealloc_threads = machine->smp.cpus;
 281 }
 282
 283 static void host_memory_backend_post_init(Object *obj)
 284 {
 285     object_apply_compat_props(obj);
 286 }
 287
 288 bool host_memory_backend_mr_inited(HostMemoryBackend *backend)
 289 {
 290     /*
 291      * NOTE: We forbid zero-length memory backend, so here zero means
 292      * "we haven't inited the backend memory region yet".
 293      */
 294     return memory_region_size(&backend->mr) != 0;
 295 }
 296
 297 MemoryRegion *host_memory_backend_get_memory(HostMemoryBackend *backend)
 298 {
 299     return host_memory_backend_mr_inited(backend) ? &backend->mr : NULL;
 300 }
 301
 302 void host_memory_backend_set_mapped(HostMemoryBackend *backend, bool mapped)
 303 {
 304     backend->is_mapped = mapped;
 305 }
 306
 307 bool host_memory_backend_is_mapped(HostMemoryBackend *backend)
 308 {
 309     return backend->is_mapped;
 310 }
 311
 312 size_t host_memory_backend_pagesize(HostMemoryBackend *memdev)
 313 {
 314     size_t pagesize = qemu_ram_pagesize(memdev->mr.ram_block);
 315     g_assert(pagesize >= qemu_real_host_page_size());
 316     return pagesize;
 317 }
 318
 319 static void
 320 host_memory_backend_memory_complete(UserCreatable *uc, Error **errp)
 321 {
 322     HostMemoryBackend *backend = MEMORY_BACKEND(uc);
 323     HostMemoryBackendClass *bc = MEMORY_BACKEND_GET_CLASS(uc);
 324     void *ptr;
 325     uint64_t sz;
 326
 327     if (!bc->alloc) {
 328         return;
 329     }
 330     if (!bc->alloc(backend, errp)) {
 331         return;
 332     }
 333
 334     ptr = memory_region_get_ram_ptr(&backend->mr);
 335     sz = memory_region_size(&backend->mr);
 336
 337     if (backend->merge) {
 338         qemu_madvise(ptr, sz, QEMU_MADV_MERGEABLE);
 339     }
 340     if (!backend->dump) {
 341         qemu_madvise(ptr, sz, QEMU_MADV_DONTDUMP);
 342     }
 343 #ifdef CONFIG_NUMA
 344     unsigned long lastbit = find_last_bit(backend->host_nodes, MAX_NODES);
 345     /* lastbit == MAX_NODES means maxnode = 0 */
 346     unsigned long maxnode = (lastbit + 1) % (MAX_NODES + 1);
 347     /*
 348      * Ensure policy won't be ignored in case memory is preallocated
 349      * before mbind(). note: MPOL_MF_STRICT is ignored on hugepages so
 350      * this doesn't catch hugepage case.
 351      */
 352     unsigned flags = MPOL_MF_STRICT | MPOL_MF_MOVE;
 353     int mode = backend->policy;
 354
 355     /* check for invalid host-nodes and policies and give more verbose
 356      * error messages than mbind(). */
 357     if (maxnode && backend->policy == MPOL_DEFAULT) {
 358         error_setg(errp, "host-nodes must be empty for policy default,"
 359                    " or you should explicitly specify a policy other"
 360                    " than default");
 361         return;
 362     } else if (maxnode == 0 && backend->policy != MPOL_DEFAULT) {
 363         error_setg(errp, "host-nodes must be set for policy %s",
 364                    HostMemPolicy_str(backend->policy));
 365         return;
 366     }
 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 #ifdef HAVE_NUMA_HAS_PREFERRED_MANY
 379     if (mode == MPOL_PREFERRED && numa_has_preferred_many() > 0) {
 380         /*
 381          * Replace with MPOL_PREFERRED_MANY otherwise the mbind() below
 382          * silently picks the first node.
 383          */
 384         mode = MPOL_PREFERRED_MANY;
 385     }
 386 #endif
 387
 388     if (maxnode &&
 389         mbind(ptr, sz, mode, backend->host_nodes, maxnode + 1, flags)) {
 390         if (backend->policy != MPOL_DEFAULT || errno != ENOSYS) {
 391             error_setg_errno(errp, errno,
 392                              "cannot bind memory to host NUMA nodes");
 393             return;
 394         }
 395     }
 396 #endif
 397     /*
 398      * Preallocate memory after the NUMA policy has been instantiated.
 399      * This is necessary to guarantee memory is allocated with
 400      * specified NUMA policy in place.
 401      */
 402     if (backend->prealloc && !qemu_prealloc_mem(memory_region_get_fd(&backend->mr),
 403                                                 ptr, sz,
 404                                                 backend->prealloc_threads,
 405                                                 backend->prealloc_context, errp)) {
 406         return;
 407     }
 408 }
 409
 410 static bool
 411 host_memory_backend_can_be_deleted(UserCreatable *uc)
 412 {
 413     if (host_memory_backend_is_mapped(MEMORY_BACKEND(uc))) {
 414         return false;
 415     } else {
 416         return true;
 417     }
 418 }
 419
 420 static bool host_memory_backend_get_share(Object *o, Error **errp)
 421 {
 422     HostMemoryBackend *backend = MEMORY_BACKEND(o);
 423
 424     return backend->share;
 425 }
 426
 427 static void host_memory_backend_set_share(Object *o, bool value, Error **errp)
 428 {
 429     HostMemoryBackend *backend = MEMORY_BACKEND(o);
 430
 431     if (host_memory_backend_mr_inited(backend)) {
 432         error_setg(errp, "cannot change property value");
 433         return;
 434     }
 435     backend->share = value;
 436 }
 437
 438 #ifdef CONFIG_LINUX
 439 static bool host_memory_backend_get_reserve(Object *o, Error **errp)
 440 {
 441     HostMemoryBackend *backend = MEMORY_BACKEND(o);
 442
 443     return backend->reserve;
 444 }
 445
 446 static void host_memory_backend_set_reserve(Object *o, bool value, Error **errp)
 447 {
 448     HostMemoryBackend *backend = MEMORY_BACKEND(o);
 449
 450     if (host_memory_backend_mr_inited(backend)) {
 451         error_setg(errp, "cannot change property value");
 452         return;
 453     }
 454     if (backend->prealloc && !value) {
 455         error_setg(errp, "'prealloc=on' and 'reserve=off' are incompatible");
 456         return;
 457     }
 458     backend->reserve = value;
 459 }
 460 #endif /* CONFIG_LINUX */
 461
 462 static bool
 463 host_memory_backend_get_use_canonical_path(Object *obj, Error **errp)
 464 {
 465     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 466
 467     return backend->use_canonical_path;
 468 }
 469
 470 static void
 471 host_memory_backend_set_use_canonical_path(Object *obj, bool value,
 472                                            Error **errp)
 473 {
 474     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 475
 476     backend->use_canonical_path = value;
 477 }
 478
 479 static void
 480 host_memory_backend_class_init(ObjectClass *oc, void *data)
 481 {
 482     UserCreatableClass *ucc = USER_CREATABLE_CLASS(oc);
 483
 484     ucc->complete = host_memory_backend_memory_complete;
 485     ucc->can_be_deleted = host_memory_backend_can_be_deleted;
 486
 487     object_class_property_add_bool(oc, "merge",
 488         host_memory_backend_get_merge,
 489         host_memory_backend_set_merge);
 490     object_class_property_set_description(oc, "merge",
 491         "Mark memory as mergeable");
 492     object_class_property_add_bool(oc, "dump",
 493         host_memory_backend_get_dump,
 494         host_memory_backend_set_dump);
 495     object_class_property_set_description(oc, "dump",
 496         "Set to 'off' to exclude from core dump");
 497     object_class_property_add_bool(oc, "prealloc",
 498         host_memory_backend_get_prealloc,
 499         host_memory_backend_set_prealloc);
 500     object_class_property_set_description(oc, "prealloc",
 501         "Preallocate memory");
 502     object_class_property_add(oc, "prealloc-threads", "int",
 503         host_memory_backend_get_prealloc_threads,
 504         host_memory_backend_set_prealloc_threads,
 505         NULL, NULL);
 506     object_class_property_set_description(oc, "prealloc-threads",
 507         "Number of CPU threads to use for prealloc");
 508     object_class_property_add_link(oc, "prealloc-context",
 509         TYPE_THREAD_CONTEXT, offsetof(HostMemoryBackend, prealloc_context),
 510         object_property_allow_set_link, OBJ_PROP_LINK_STRONG);
 511     object_class_property_set_description(oc, "prealloc-context",
 512         "Context to use for creating CPU threads for preallocation");
 513     object_class_property_add(oc, "size", "int",
 514         host_memory_backend_get_size,
 515         host_memory_backend_set_size,
 516         NULL, NULL);
 517     object_class_property_set_description(oc, "size",
 518         "Size of the memory region (ex: 500M)");
 519     object_class_property_add(oc, "host-nodes", "int",
 520         host_memory_backend_get_host_nodes,
 521         host_memory_backend_set_host_nodes,
 522         NULL, NULL);
 523     object_class_property_set_description(oc, "host-nodes",
 524         "Binds memory to the list of NUMA host nodes");
 525     object_class_property_add_enum(oc, "policy", "HostMemPolicy",
 526         &HostMemPolicy_lookup,
 527         host_memory_backend_get_policy,
 528         host_memory_backend_set_policy);
 529     object_class_property_set_description(oc, "policy",
 530         "Set the NUMA policy");
 531     object_class_property_add_bool(oc, "share",
 532         host_memory_backend_get_share, host_memory_backend_set_share);
 533     object_class_property_set_description(oc, "share",
 534         "Mark the memory as private to QEMU or shared");
 535 #ifdef CONFIG_LINUX
 536     object_class_property_add_bool(oc, "reserve",
 537         host_memory_backend_get_reserve, host_memory_backend_set_reserve);
 538     object_class_property_set_description(oc, "reserve",
 539         "Reserve swap space (or huge pages) if applicable");
 540 #endif /* CONFIG_LINUX */
 541     /*
 542      * Do not delete/rename option. This option must be considered stable
 543      * (as if it didn't have the 'x-' prefix including deprecation period) as
 544      * long as 4.0 and older machine types exists.
 545      * Option will be used by upper layers to override (disable) canonical path
 546      * for ramblock-id set by compat properties on old machine types ( <= 4.0),
 547      * to keep migration working when backend is used for main RAM with
 548      * -machine memory-backend= option (main RAM historically used prefix-less
 549      * ramblock-id).
 550      */
 551     object_class_property_add_bool(oc, "x-use-canonical-path-for-ramblock-id",
 552         host_memory_backend_get_use_canonical_path,
 553         host_memory_backend_set_use_canonical_path);
 554 }
 555
 556 static const TypeInfo host_memory_backend_info = {
 557     .name = TYPE_MEMORY_BACKEND,
 558     .parent = TYPE_OBJECT,
 559     .abstract = true,
 560     .class_size = sizeof(HostMemoryBackendClass),
 561     .class_init = host_memory_backend_class_init,
 562     .instance_size = sizeof(HostMemoryBackend),
 563     .instance_init = host_memory_backend_init,
 564     .instance_post_init = host_memory_backend_post_init,
 565     .interfaces = (InterfaceInfo[]) {
 566         { TYPE_USER_CREATABLE },
 567         { }
 568     }
 569 };
 570
 571 static void register_types(void)
 572 {
 573     type_register_static(&host_memory_backend_info);
 574 }
 575
 576 type_init(register_types);