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 #include "hw/qdev-core.h"
  24
  25 #ifdef CONFIG_NUMA
  26 #include <numaif.h>
  27 #include <numa.h>
  28 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_DEFAULT != MPOL_DEFAULT);
  29 /*
  30  * HOST_MEM_POLICY_PREFERRED may either translate to MPOL_PREFERRED or
  31  * MPOL_PREFERRED_MANY, see comments further below.
  32  */
  33 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_PREFERRED != MPOL_PREFERRED);
  34 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_BIND != MPOL_BIND);
  35 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_INTERLEAVE != MPOL_INTERLEAVE);
  36 #endif
  37
  38 char *
  39 host_memory_backend_get_name(HostMemoryBackend *backend)
  40 {
  41     if (!backend->use_canonical_path) {
  42         return g_strdup(object_get_canonical_path_component(OBJECT(backend)));
  43     }
  44
  45     return object_get_canonical_path(OBJECT(backend));
  46 }
  47
  48 static void
  49 host_memory_backend_get_size(Object *obj, Visitor *v, const char *name,
  50                              void *opaque, Error **errp)
  51 {
  52     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
  53     uint64_t value = backend->size;
  54
  55     visit_type_size(v, name, &value, errp);
  56 }
  57
  58 static void
  59 host_memory_backend_set_size(Object *obj, Visitor *v, const char *name,
  60                              void *opaque, Error **errp)
  61 {
  62     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
  63     uint64_t value;
  64
  65     if (host_memory_backend_mr_inited(backend)) {
  66         error_setg(errp, "cannot change property %s of %s ", name,
  67                    object_get_typename(obj));
  68         return;
  69     }
  70
  71     if (!visit_type_size(v, name, &value, errp)) {
  72         return;
  73     }
  74     if (!value) {
  75         error_setg(errp,
  76                    "property '%s' of %s doesn't take value '%" PRIu64 "'",
  77                    name, object_get_typename(obj), value);
  78         return;
  79     }
  80     backend->size = value;
  81 }
  82
  83 static void
  84 host_memory_backend_get_host_nodes(Object *obj, Visitor *v, const char *name,
  85                                    void *opaque, Error **errp)
  86 {
  87     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
  88     uint16List *host_nodes = NULL;
  89     uint16List **tail = &host_nodes;
  90     unsigned long value;
  91
  92     value = find_first_bit(backend->host_nodes, MAX_NODES);
  93     if (value == MAX_NODES) {
  94         goto ret;
  95     }
  96
  97     QAPI_LIST_APPEND(tail, value);
  98
  99     do {
 100         value = find_next_bit(backend->host_nodes, MAX_NODES, value + 1);
 101         if (value == MAX_NODES) {
 102             break;
 103         }
 104
 105         QAPI_LIST_APPEND(tail, value);
 106     } while (true);
 107
 108 ret:
 109     visit_type_uint16List(v, name, &host_nodes, errp);
 110     qapi_free_uint16List(host_nodes);
 111 }
 112
 113 static void
 114 host_memory_backend_set_host_nodes(Object *obj, Visitor *v, const char *name,
 115                                    void *opaque, Error **errp)
 116 {
 117 #ifdef CONFIG_NUMA
 118     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 119     uint16List *l, *host_nodes = NULL;
 120
 121     visit_type_uint16List(v, name, &host_nodes, errp);
 122
 123     for (l = host_nodes; l; l = l->next) {
 124         if (l->value >= MAX_NODES) {
 125             error_setg(errp, "Invalid host-nodes value: %d", l->value);
 126             goto out;
 127         }
 128     }
 129
 130     for (l = host_nodes; l; l = l->next) {
 131         bitmap_set(backend->host_nodes, l->value, 1);
 132     }
 133
 134 out:
 135     qapi_free_uint16List(host_nodes);
 136 #else
 137     error_setg(errp, "NUMA node binding are not supported by this QEMU");
 138 #endif
 139 }
 140
 141 static int
 142 host_memory_backend_get_policy(Object *obj, Error **errp G_GNUC_UNUSED)
 143 {
 144     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 145     return backend->policy;
 146 }
 147
 148 static void
 149 host_memory_backend_set_policy(Object *obj, int policy, Error **errp)
 150 {
 151     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 152     backend->policy = policy;
 153
 154 #ifndef CONFIG_NUMA
 155     if (policy != HOST_MEM_POLICY_DEFAULT) {
 156         error_setg(errp, "NUMA policies are not supported by this QEMU");
 157     }
 158 #endif
 159 }
 160
 161 static bool host_memory_backend_get_merge(Object *obj, Error **errp)
 162 {
 163     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 164
 165     return backend->merge;
 166 }
 167
 168 static void host_memory_backend_set_merge(Object *obj, bool value, Error **errp)
 169 {
 170     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 171
 172     if (!host_memory_backend_mr_inited(backend)) {
 173         backend->merge = value;
 174         return;
 175     }
 176
 177     if (value != backend->merge) {
 178         void *ptr = memory_region_get_ram_ptr(&backend->mr);
 179         uint64_t sz = memory_region_size(&backend->mr);
 180
 181         qemu_madvise(ptr, sz,
 182                      value ? QEMU_MADV_MERGEABLE : QEMU_MADV_UNMERGEABLE);
 183         backend->merge = value;
 184     }
 185 }
 186
 187 static bool host_memory_backend_get_dump(Object *obj, Error **errp)
 188 {
 189     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 190
 191     return backend->dump;
 192 }
 193
 194 static void host_memory_backend_set_dump(Object *obj, bool value, Error **errp)
 195 {
 196     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 197
 198     if (!host_memory_backend_mr_inited(backend)) {
 199         backend->dump = value;
 200         return;
 201     }
 202
 203     if (value != backend->dump) {
 204         void *ptr = memory_region_get_ram_ptr(&backend->mr);
 205         uint64_t sz = memory_region_size(&backend->mr);
 206
 207         qemu_madvise(ptr, sz,
 208                      value ? QEMU_MADV_DODUMP : QEMU_MADV_DONTDUMP);
 209         backend->dump = value;
 210     }
 211 }
 212
 213 static bool host_memory_backend_get_prealloc(Object *obj, Error **errp)
 214 {
 215     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 216
 217     return backend->prealloc;
 218 }
 219
 220 static void host_memory_backend_set_prealloc(Object *obj, bool value,
 221                                              Error **errp)
 222 {
 223     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 224
 225     if (!backend->reserve && value) {
 226         error_setg(errp, "'prealloc=on' and 'reserve=off' are incompatible");
 227         return;
 228     }
 229
 230     if (!host_memory_backend_mr_inited(backend)) {
 231         backend->prealloc = value;
 232         return;
 233     }
 234
 235     if (value && !backend->prealloc) {
 236         int fd = memory_region_get_fd(&backend->mr);
 237         void *ptr = memory_region_get_ram_ptr(&backend->mr);
 238         uint64_t sz = memory_region_size(&backend->mr);
 239
 240         if (!qemu_prealloc_mem(fd, ptr, sz, backend->prealloc_threads,
 241                                backend->prealloc_context, false, errp)) {
 242             return;
 243         }
 244         backend->prealloc = true;
 245     }
 246 }
 247
 248 static void host_memory_backend_get_prealloc_threads(Object *obj, Visitor *v,
 249     const char *name, void *opaque, Error **errp)
 250 {
 251     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 252     visit_type_uint32(v, name, &backend->prealloc_threads, errp);
 253 }
 254
 255 static void host_memory_backend_set_prealloc_threads(Object *obj, Visitor *v,
 256     const char *name, void *opaque, Error **errp)
 257 {
 258     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 259     uint32_t value;
 260
 261     if (!visit_type_uint32(v, name, &value, errp)) {
 262         return;
 263     }
 264     if (value <= 0) {
 265         error_setg(errp, "property '%s' of %s doesn't take value '%d'", name,
 266                    object_get_typename(obj), value);
 267         return;
 268     }
 269     backend->prealloc_threads = value;
 270 }
 271
 272 static void host_memory_backend_init(Object *obj)
 273 {
 274     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 275     MachineState *machine = MACHINE(qdev_get_machine());
 276
 277     /* TODO: convert access to globals to compat properties */
 278     backend->merge = machine_mem_merge(machine);
 279     backend->dump = machine_dump_guest_core(machine);
 280     backend->guest_memfd = machine_require_guest_memfd(machine);
 281     backend->reserve = true;
 282     backend->prealloc_threads = machine->smp.cpus;
 283 }
 284
 285 static void host_memory_backend_post_init(Object *obj)
 286 {
 287     object_apply_compat_props(obj);
 288 }
 289
 290 bool host_memory_backend_mr_inited(HostMemoryBackend *backend)
 291 {
 292     /*
 293      * NOTE: We forbid zero-length memory backend, so here zero means
 294      * "we haven't inited the backend memory region yet".
 295      */
 296     return memory_region_size(&backend->mr) != 0;
 297 }
 298
 299 MemoryRegion *host_memory_backend_get_memory(HostMemoryBackend *backend)
 300 {
 301     return host_memory_backend_mr_inited(backend) ? &backend->mr : NULL;
 302 }
 303
 304 void host_memory_backend_set_mapped(HostMemoryBackend *backend, bool mapped)
 305 {
 306     backend->is_mapped = mapped;
 307 }
 308
 309 bool host_memory_backend_is_mapped(HostMemoryBackend *backend)
 310 {
 311     return backend->is_mapped;
 312 }
 313
 314 size_t host_memory_backend_pagesize(HostMemoryBackend *memdev)
 315 {
 316     size_t pagesize = qemu_ram_pagesize(memdev->mr.ram_block);
 317     g_assert(pagesize >= qemu_real_host_page_size());
 318     return pagesize;
 319 }
 320
 321 static void
 322 host_memory_backend_memory_complete(UserCreatable *uc, Error **errp)
 323 {
 324     HostMemoryBackend *backend = MEMORY_BACKEND(uc);
 325     HostMemoryBackendClass *bc = MEMORY_BACKEND_GET_CLASS(uc);
 326     void *ptr;
 327     uint64_t sz;
 328     bool async = !phase_check(PHASE_LATE_BACKENDS_CREATED);
 329
 330     if (!bc->alloc) {
 331         return;
 332     }
 333     if (!bc->alloc(backend, errp)) {
 334         return;
 335     }
 336
 337     ptr = memory_region_get_ram_ptr(&backend->mr);
 338     sz = memory_region_size(&backend->mr);
 339
 340     if (backend->merge) {
 341         qemu_madvise(ptr, sz, QEMU_MADV_MERGEABLE);
 342     }
 343     if (!backend->dump) {
 344         qemu_madvise(ptr, sz, QEMU_MADV_DONTDUMP);
 345     }
 346 #ifdef CONFIG_NUMA
 347     unsigned long lastbit = find_last_bit(backend->host_nodes, MAX_NODES);
 348     /* lastbit == MAX_NODES means maxnode = 0 */
 349     unsigned long maxnode = (lastbit + 1) % (MAX_NODES + 1);
 350     /*
 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      */
 355     unsigned flags = MPOL_MF_STRICT | MPOL_MF_MOVE;
 356     int mode = backend->policy;
 357
 358     /* check for invalid host-nodes and policies and give more verbose
 359      * error messages than mbind(). */
 360     if (maxnode && backend->policy == MPOL_DEFAULT) {
 361         error_setg(errp, "host-nodes must be empty for policy default,"
 362                    " or you should explicitly specify a policy other"
 363                    " than default");
 364         return;
 365     } else if (maxnode == 0 && backend->policy != MPOL_DEFAULT) {
 366         error_setg(errp, "host-nodes must be set for policy %s",
 367                    HostMemPolicy_str(backend->policy));
 368         return;
 369     }
 370
 371     /*
 372      * We can have up to MAX_NODES nodes, but we need to pass maxnode+1
 373      * as argument to mbind() due to an old Linux bug (feature?) which
 374      * cuts off the last specified node. This means backend->host_nodes
 375      * must have MAX_NODES+1 bits available.
 376      */
 377     assert(sizeof(backend->host_nodes) >=
 378            BITS_TO_LONGS(MAX_NODES + 1) * sizeof(unsigned long));
 379     assert(maxnode <= MAX_NODES);
 380
 381 #ifdef HAVE_NUMA_HAS_PREFERRED_MANY
 382     if (mode == MPOL_PREFERRED && numa_has_preferred_many() > 0) {
 383         /*
 384          * Replace with MPOL_PREFERRED_MANY otherwise the mbind() below
 385          * silently picks the first node.
 386          */
 387         mode = MPOL_PREFERRED_MANY;
 388     }
 389 #endif
 390
 391     if (maxnode &&
 392         mbind(ptr, sz, mode, backend->host_nodes, maxnode + 1, flags)) {
 393         if (backend->policy != MPOL_DEFAULT || errno != ENOSYS) {
 394             error_setg_errno(errp, errno,
 395                              "cannot bind memory to host NUMA nodes");
 396             return;
 397         }
 398     }
 399 #endif
 400     /*
 401      * Preallocate memory after the NUMA policy has been instantiated.
 402      * This is necessary to guarantee memory is allocated with
 403      * specified NUMA policy in place.
 404      */
 405     if (backend->prealloc && !qemu_prealloc_mem(memory_region_get_fd(&backend->mr),
 406                                                 ptr, sz,
 407                                                 backend->prealloc_threads,
 408                                                 backend->prealloc_context,
 409                                                 async, errp)) {
 410         return;
 411     }
 412 }
 413
 414 static bool
 415 host_memory_backend_can_be_deleted(UserCreatable *uc)
 416 {
 417     if (host_memory_backend_is_mapped(MEMORY_BACKEND(uc))) {
 418         return false;
 419     } else {
 420         return true;
 421     }
 422 }
 423
 424 static bool host_memory_backend_get_share(Object *o, Error **errp)
 425 {
 426     HostMemoryBackend *backend = MEMORY_BACKEND(o);
 427
 428     return backend->share;
 429 }
 430
 431 static void host_memory_backend_set_share(Object *o, bool value, Error **errp)
 432 {
 433     HostMemoryBackend *backend = MEMORY_BACKEND(o);
 434
 435     if (host_memory_backend_mr_inited(backend)) {
 436         error_setg(errp, "cannot change property value");
 437         return;
 438     }
 439     backend->share = value;
 440 }
 441
 442 #ifdef CONFIG_LINUX
 443 static bool host_memory_backend_get_reserve(Object *o, Error **errp)
 444 {
 445     HostMemoryBackend *backend = MEMORY_BACKEND(o);
 446
 447     return backend->reserve;
 448 }
 449
 450 static void host_memory_backend_set_reserve(Object *o, bool value, Error **errp)
 451 {
 452     HostMemoryBackend *backend = MEMORY_BACKEND(o);
 453
 454     if (host_memory_backend_mr_inited(backend)) {
 455         error_setg(errp, "cannot change property value");
 456         return;
 457     }
 458     if (backend->prealloc && !value) {
 459         error_setg(errp, "'prealloc=on' and 'reserve=off' are incompatible");
 460         return;
 461     }
 462     backend->reserve = value;
 463 }
 464 #endif /* CONFIG_LINUX */
 465
 466 static bool
 467 host_memory_backend_get_use_canonical_path(Object *obj, Error **errp)
 468 {
 469     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 470
 471     return backend->use_canonical_path;
 472 }
 473
 474 static void
 475 host_memory_backend_set_use_canonical_path(Object *obj, bool value,
 476                                            Error **errp)
 477 {
 478     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 479
 480     backend->use_canonical_path = value;
 481 }
 482
 483 static void
 484 host_memory_backend_class_init(ObjectClass *oc, void *data)
 485 {
 486     UserCreatableClass *ucc = USER_CREATABLE_CLASS(oc);
 487
 488     ucc->complete = host_memory_backend_memory_complete;
 489     ucc->can_be_deleted = host_memory_backend_can_be_deleted;
 490
 491     object_class_property_add_bool(oc, "merge",
 492         host_memory_backend_get_merge,
 493         host_memory_backend_set_merge);
 494     object_class_property_set_description(oc, "merge",
 495         "Mark memory as mergeable");
 496     object_class_property_add_bool(oc, "dump",
 497         host_memory_backend_get_dump,
 498         host_memory_backend_set_dump);
 499     object_class_property_set_description(oc, "dump",
 500         "Set to 'off' to exclude from core dump");
 501     object_class_property_add_bool(oc, "prealloc",
 502         host_memory_backend_get_prealloc,
 503         host_memory_backend_set_prealloc);
 504     object_class_property_set_description(oc, "prealloc",
 505         "Preallocate memory");
 506     object_class_property_add(oc, "prealloc-threads", "int",
 507         host_memory_backend_get_prealloc_threads,
 508         host_memory_backend_set_prealloc_threads,
 509         NULL, NULL);
 510     object_class_property_set_description(oc, "prealloc-threads",
 511         "Number of CPU threads to use for prealloc");
 512     object_class_property_add_link(oc, "prealloc-context",
 513         TYPE_THREAD_CONTEXT, offsetof(HostMemoryBackend, prealloc_context),
 514         object_property_allow_set_link, OBJ_PROP_LINK_STRONG);
 515     object_class_property_set_description(oc, "prealloc-context",
 516         "Context to use for creating CPU threads for preallocation");
 517     object_class_property_add(oc, "size", "int",
 518         host_memory_backend_get_size,
 519         host_memory_backend_set_size,
 520         NULL, NULL);
 521     object_class_property_set_description(oc, "size",
 522         "Size of the memory region (ex: 500M)");
 523     object_class_property_add(oc, "host-nodes", "int",
 524         host_memory_backend_get_host_nodes,
 525         host_memory_backend_set_host_nodes,
 526         NULL, NULL);
 527     object_class_property_set_description(oc, "host-nodes",
 528         "Binds memory to the list of NUMA host nodes");
 529     object_class_property_add_enum(oc, "policy", "HostMemPolicy",
 530         &HostMemPolicy_lookup,
 531         host_memory_backend_get_policy,
 532         host_memory_backend_set_policy);
 533     object_class_property_set_description(oc, "policy",
 534         "Set the NUMA policy");
 535     object_class_property_add_bool(oc, "share",
 536         host_memory_backend_get_share, host_memory_backend_set_share);
 537     object_class_property_set_description(oc, "share",
 538         "Mark the memory as private to QEMU or shared");
 539 #ifdef CONFIG_LINUX
 540     object_class_property_add_bool(oc, "reserve",
 541         host_memory_backend_get_reserve, host_memory_backend_set_reserve);
 542     object_class_property_set_description(oc, "reserve",
 543         "Reserve swap space (or huge pages) if applicable");
 544 #endif /* CONFIG_LINUX */
 545     /*
 546      * Do not delete/rename option. This option must be considered stable
 547      * (as if it didn't have the 'x-' prefix including deprecation period) as
 548      * long as 4.0 and older machine types exists.
 549      * Option will be used by upper layers to override (disable) canonical path
 550      * for ramblock-id set by compat properties on old machine types ( <= 4.0),
 551      * to keep migration working when backend is used for main RAM with
 552      * -machine memory-backend= option (main RAM historically used prefix-less
 553      * ramblock-id).
 554      */
 555     object_class_property_add_bool(oc, "x-use-canonical-path-for-ramblock-id",
 556         host_memory_backend_get_use_canonical_path,
 557         host_memory_backend_set_use_canonical_path);
 558 }
 559
 560 static const TypeInfo host_memory_backend_info = {
 561     .name = TYPE_MEMORY_BACKEND,
 562     .parent = TYPE_OBJECT,
 563     .abstract = true,
 564     .class_size = sizeof(HostMemoryBackendClass),
 565     .class_init = host_memory_backend_class_init,
 566     .instance_size = sizeof(HostMemoryBackend),
 567     .instance_init = host_memory_backend_init,
 568     .instance_post_init = host_memory_backend_post_init,
 569     .interfaces = (InterfaceInfo[]) {
 570         { TYPE_USER_CREATABLE },
 571         { }
 572     }
 573 };
 574
 575 static void register_types(void)
 576 {
 577     type_register_static(&host_memory_backend_info);
 578 }
 579
 580 type_init(register_types);