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
  23 #ifdef CONFIG_NUMA
  24 #include <numaif.h>
  25 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_DEFAULT != MPOL_DEFAULT);
  26 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_PREFERRED != MPOL_PREFERRED);
  27 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_BIND != MPOL_BIND);
  28 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_INTERLEAVE != MPOL_INTERLEAVE);
  29 #endif
  30
  31 char *
  32 host_memory_backend_get_name(HostMemoryBackend *backend)
  33 {
  34     if (!backend->use_canonical_path) {
  35         return g_strdup(object_get_canonical_path_component(OBJECT(backend)));
  36     }
  37
  38     return object_get_canonical_path(OBJECT(backend));
  39 }
  40
  41 static void
  42 host_memory_backend_get_size(Object *obj, Visitor *v, const char *name,
  43                              void *opaque, Error **errp)
  44 {
  45     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
  46     uint64_t value = backend->size;
  47
  48     visit_type_size(v, name, &value, errp);
  49 }
  50
  51 static void
  52 host_memory_backend_set_size(Object *obj, Visitor *v, const char *name,
  53                              void *opaque, Error **errp)
  54 {
  55     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
  56     uint64_t value;
  57
  58     if (host_memory_backend_mr_inited(backend)) {
  59         error_setg(errp, "cannot change property %s of %s ", name,
  60                    object_get_typename(obj));
  61         return;
  62     }
  63
  64     if (!visit_type_size(v, name, &value, errp)) {
  65         return;
  66     }
  67     if (!value) {
  68         error_setg(errp,
  69                    "property '%s' of %s doesn't take value '%" PRIu64 "'",
  70                    name, object_get_typename(obj), value);
  71         return;
  72     }
  73     backend->size = value;
  74 }
  75
  76 static void
  77 host_memory_backend_get_host_nodes(Object *obj, Visitor *v, const char *name,
  78                                    void *opaque, Error **errp)
  79 {
  80     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
  81     uint16List *host_nodes = NULL;
  82     uint16List **tail = &host_nodes;
  83     unsigned long value;
  84
  85     value = find_first_bit(backend->host_nodes, MAX_NODES);
  86     if (value == MAX_NODES) {
  87         goto ret;
  88     }
  89
  90     QAPI_LIST_APPEND(tail, value);
  91
  92     do {
  93         value = find_next_bit(backend->host_nodes, MAX_NODES, value + 1);
  94         if (value == MAX_NODES) {
  95             break;
  96         }
  97
  98         QAPI_LIST_APPEND(tail, value);
  99     } while (true);
 100
 101 ret:
 102     visit_type_uint16List(v, name, &host_nodes, errp);
 103     qapi_free_uint16List(host_nodes);
 104 }
 105
 106 static void
 107 host_memory_backend_set_host_nodes(Object *obj, Visitor *v, const char *name,
 108                                    void *opaque, Error **errp)
 109 {
 110 #ifdef CONFIG_NUMA
 111     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 112     uint16List *l, *host_nodes = NULL;
 113
 114     visit_type_uint16List(v, name, &host_nodes, errp);
 115
 116     for (l = host_nodes; l; l = l->next) {
 117         if (l->value >= MAX_NODES) {
 118             error_setg(errp, "Invalid host-nodes value: %d", l->value);
 119             goto out;
 120         }
 121     }
 122
 123     for (l = host_nodes; l; l = l->next) {
 124         bitmap_set(backend->host_nodes, l->value, 1);
 125     }
 126
 127 out:
 128     qapi_free_uint16List(host_nodes);
 129 #else
 130     error_setg(errp, "NUMA node binding are not supported by this QEMU");
 131 #endif
 132 }
 133
 134 static int
 135 host_memory_backend_get_policy(Object *obj, Error **errp G_GNUC_UNUSED)
 136 {
 137     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 138     return backend->policy;
 139 }
 140
 141 static void
 142 host_memory_backend_set_policy(Object *obj, int policy, Error **errp)
 143 {
 144     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 145     backend->policy = policy;
 146
 147 #ifndef CONFIG_NUMA
 148     if (policy != HOST_MEM_POLICY_DEFAULT) {
 149         error_setg(errp, "NUMA policies are not supported by this QEMU");
 150     }
 151 #endif
 152 }
 153
 154 static bool host_memory_backend_get_merge(Object *obj, Error **errp)
 155 {
 156     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 157
 158     return backend->merge;
 159 }
 160
 161 static void host_memory_backend_set_merge(Object *obj, bool value, Error **errp)
 162 {
 163     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 164
 165     if (!host_memory_backend_mr_inited(backend)) {
 166         backend->merge = value;
 167         return;
 168     }
 169
 170     if (value != backend->merge) {
 171         void *ptr = memory_region_get_ram_ptr(&backend->mr);
 172         uint64_t sz = memory_region_size(&backend->mr);
 173
 174         qemu_madvise(ptr, sz,
 175                      value ? QEMU_MADV_MERGEABLE : QEMU_MADV_UNMERGEABLE);
 176         backend->merge = value;
 177     }
 178 }
 179
 180 static bool host_memory_backend_get_dump(Object *obj, Error **errp)
 181 {
 182     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 183
 184     return backend->dump;
 185 }
 186
 187 static void host_memory_backend_set_dump(Object *obj, bool value, Error **errp)
 188 {
 189     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 190
 191     if (!host_memory_backend_mr_inited(backend)) {
 192         backend->dump = value;
 193         return;
 194     }
 195
 196     if (value != backend->dump) {
 197         void *ptr = memory_region_get_ram_ptr(&backend->mr);
 198         uint64_t sz = memory_region_size(&backend->mr);
 199
 200         qemu_madvise(ptr, sz,
 201                      value ? QEMU_MADV_DODUMP : QEMU_MADV_DONTDUMP);
 202         backend->dump = value;
 203     }
 204 }
 205
 206 static bool host_memory_backend_get_prealloc(Object *obj, Error **errp)
 207 {
 208     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 209
 210     return backend->prealloc;
 211 }
 212
 213 static void host_memory_backend_set_prealloc(Object *obj, bool value,
 214                                              Error **errp)
 215 {
 216     Error *local_err = NULL;
 217     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 218
 219     if (!host_memory_backend_mr_inited(backend)) {
 220         backend->prealloc = value;
 221         return;
 222     }
 223
 224     if (value && !backend->prealloc) {
 225         int fd = memory_region_get_fd(&backend->mr);
 226         void *ptr = memory_region_get_ram_ptr(&backend->mr);
 227         uint64_t sz = memory_region_size(&backend->mr);
 228
 229         os_mem_prealloc(fd, ptr, sz, backend->prealloc_threads, &local_err);
 230         if (local_err) {
 231             error_propagate(errp, local_err);
 232             return;
 233         }
 234         backend->prealloc = true;
 235     }
 236 }
 237
 238 static void host_memory_backend_get_prealloc_threads(Object *obj, Visitor *v,
 239     const char *name, void *opaque, Error **errp)
 240 {
 241     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 242     visit_type_uint32(v, name, &backend->prealloc_threads, errp);
 243 }
 244
 245 static void host_memory_backend_set_prealloc_threads(Object *obj, Visitor *v,
 246     const char *name, void *opaque, Error **errp)
 247 {
 248     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 249     uint32_t value;
 250
 251     if (!visit_type_uint32(v, name, &value, errp)) {
 252         return;
 253     }
 254     if (value <= 0) {
 255         error_setg(errp, "property '%s' of %s doesn't take value '%d'", name,
 256                    object_get_typename(obj), value);
 257         return;
 258     }
 259     backend->prealloc_threads = value;
 260 }
 261
 262 static void host_memory_backend_init(Object *obj)
 263 {
 264     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 265     MachineState *machine = MACHINE(qdev_get_machine());
 266
 267     /* TODO: convert access to globals to compat properties */
 268     backend->merge = machine_mem_merge(machine);
 269     backend->dump = machine_dump_guest_core(machine);
 270     backend->prealloc_threads = 1;
 271 }
 272
 273 static void host_memory_backend_post_init(Object *obj)
 274 {
 275     object_apply_compat_props(obj);
 276 }
 277
 278 bool host_memory_backend_mr_inited(HostMemoryBackend *backend)
 279 {
 280     /*
 281      * NOTE: We forbid zero-length memory backend, so here zero means
 282      * "we haven't inited the backend memory region yet".
 283      */
 284     return memory_region_size(&backend->mr) != 0;
 285 }
 286
 287 MemoryRegion *host_memory_backend_get_memory(HostMemoryBackend *backend)
 288 {
 289     return host_memory_backend_mr_inited(backend) ? &backend->mr : NULL;
 290 }
 291
 292 void host_memory_backend_set_mapped(HostMemoryBackend *backend, bool mapped)
 293 {
 294     backend->is_mapped = mapped;
 295 }
 296
 297 bool host_memory_backend_is_mapped(HostMemoryBackend *backend)
 298 {
 299     return backend->is_mapped;
 300 }
 301
 302 #ifdef __linux__
 303 size_t host_memory_backend_pagesize(HostMemoryBackend *memdev)
 304 {
 305     Object *obj = OBJECT(memdev);
 306     char *path = object_property_get_str(obj, "mem-path", NULL);
 307     size_t pagesize = qemu_mempath_getpagesize(path);
 308
 309     g_free(path);
 310     return pagesize;
 311 }
 312 #else
 313 size_t host_memory_backend_pagesize(HostMemoryBackend *memdev)
 314 {
 315     return qemu_real_host_page_size;
 316 }
 317 #endif
 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     Error *local_err = NULL;
 325     void *ptr;
 326     uint64_t sz;
 327
 328     if (bc->alloc) {
 329         bc->alloc(backend, &local_err);
 330         if (local_err) {
 331             goto out;
 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         /* ensure policy won't be ignored in case memory is preallocated
 348          * before mbind(). note: MPOL_MF_STRICT is ignored on hugepages so
 349          * this doesn't catch hugepage case. */
 350         unsigned flags = MPOL_MF_STRICT | MPOL_MF_MOVE;
 351
 352         /* check for invalid host-nodes and policies and give more verbose
 353          * error messages than mbind(). */
 354         if (maxnode && backend->policy == MPOL_DEFAULT) {
 355             error_setg(errp, "host-nodes must be empty for policy default,"
 356                        " or you should explicitly specify a policy other"
 357                        " than default");
 358             return;
 359         } else if (maxnode == 0 && backend->policy != MPOL_DEFAULT) {
 360             error_setg(errp, "host-nodes must be set for policy %s",
 361                        HostMemPolicy_str(backend->policy));
 362             return;
 363         }
 364
 365         /* We can have up to MAX_NODES nodes, but we need to pass maxnode+1
 366          * as argument to mbind() due to an old Linux bug (feature?) which
 367          * cuts off the last specified node. This means backend->host_nodes
 368          * must have MAX_NODES+1 bits available.
 369          */
 370         assert(sizeof(backend->host_nodes) >=
 371                BITS_TO_LONGS(MAX_NODES + 1) * sizeof(unsigned long));
 372         assert(maxnode <= MAX_NODES);
 373
 374         if (maxnode &&
 375             mbind(ptr, sz, backend->policy, backend->host_nodes, maxnode + 1,
 376                   flags)) {
 377             if (backend->policy != MPOL_DEFAULT || errno != ENOSYS) {
 378                 error_setg_errno(errp, errno,
 379                                  "cannot bind memory to host NUMA nodes");
 380                 return;
 381             }
 382         }
 383 #endif
 384         /* Preallocate memory after the NUMA policy has been instantiated.
 385          * This is necessary to guarantee memory is allocated with
 386          * specified NUMA policy in place.
 387          */
 388         if (backend->prealloc) {
 389             os_mem_prealloc(memory_region_get_fd(&backend->mr), ptr, sz,
 390                             backend->prealloc_threads, &local_err);
 391             if (local_err) {
 392                 goto out;
 393             }
 394         }
 395     }
 396 out:
 397     error_propagate(errp, local_err);
 398 }
 399
 400 static bool
 401 host_memory_backend_can_be_deleted(UserCreatable *uc)
 402 {
 403     if (host_memory_backend_is_mapped(MEMORY_BACKEND(uc))) {
 404         return false;
 405     } else {
 406         return true;
 407     }
 408 }
 409
 410 static bool host_memory_backend_get_share(Object *o, Error **errp)
 411 {
 412     HostMemoryBackend *backend = MEMORY_BACKEND(o);
 413
 414     return backend->share;
 415 }
 416
 417 static void host_memory_backend_set_share(Object *o, bool value, Error **errp)
 418 {
 419     HostMemoryBackend *backend = MEMORY_BACKEND(o);
 420
 421     if (host_memory_backend_mr_inited(backend)) {
 422         error_setg(errp, "cannot change property value");
 423         return;
 424     }
 425     backend->share = value;
 426 }
 427
 428 static bool
 429 host_memory_backend_get_use_canonical_path(Object *obj, Error **errp)
 430 {
 431     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 432
 433     return backend->use_canonical_path;
 434 }
 435
 436 static void
 437 host_memory_backend_set_use_canonical_path(Object *obj, bool value,
 438                                            Error **errp)
 439 {
 440     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 441
 442     backend->use_canonical_path = value;
 443 }
 444
 445 static void
 446 host_memory_backend_class_init(ObjectClass *oc, void *data)
 447 {
 448     UserCreatableClass *ucc = USER_CREATABLE_CLASS(oc);
 449
 450     ucc->complete = host_memory_backend_memory_complete;
 451     ucc->can_be_deleted = host_memory_backend_can_be_deleted;
 452
 453     object_class_property_add_bool(oc, "merge",
 454         host_memory_backend_get_merge,
 455         host_memory_backend_set_merge);
 456     object_class_property_set_description(oc, "merge",
 457         "Mark memory as mergeable");
 458     object_class_property_add_bool(oc, "dump",
 459         host_memory_backend_get_dump,
 460         host_memory_backend_set_dump);
 461     object_class_property_set_description(oc, "dump",
 462         "Set to 'off' to exclude from core dump");
 463     object_class_property_add_bool(oc, "prealloc",
 464         host_memory_backend_get_prealloc,
 465         host_memory_backend_set_prealloc);
 466     object_class_property_set_description(oc, "prealloc",
 467         "Preallocate memory");
 468     object_class_property_add(oc, "prealloc-threads", "int",
 469         host_memory_backend_get_prealloc_threads,
 470         host_memory_backend_set_prealloc_threads,
 471         NULL, NULL);
 472     object_class_property_set_description(oc, "prealloc-threads",
 473         "Number of CPU threads to use for prealloc");
 474     object_class_property_add(oc, "size", "int",
 475         host_memory_backend_get_size,
 476         host_memory_backend_set_size,
 477         NULL, NULL);
 478     object_class_property_set_description(oc, "size",
 479         "Size of the memory region (ex: 500M)");
 480     object_class_property_add(oc, "host-nodes", "int",
 481         host_memory_backend_get_host_nodes,
 482         host_memory_backend_set_host_nodes,
 483         NULL, NULL);
 484     object_class_property_set_description(oc, "host-nodes",
 485         "Binds memory to the list of NUMA host nodes");
 486     object_class_property_add_enum(oc, "policy", "HostMemPolicy",
 487         &HostMemPolicy_lookup,
 488         host_memory_backend_get_policy,
 489         host_memory_backend_set_policy);
 490     object_class_property_set_description(oc, "policy",
 491         "Set the NUMA policy");
 492     object_class_property_add_bool(oc, "share",
 493         host_memory_backend_get_share, host_memory_backend_set_share);
 494     object_class_property_set_description(oc, "share",
 495         "Mark the memory as private to QEMU or shared");
 496     /*
 497      * Do not delete/rename option. This option must be considered stable
 498      * (as if it didn't have the 'x-' prefix including deprecation period) as
 499      * long as 4.0 and older machine types exists.
 500      * Option will be used by upper layers to override (disable) canonical path
 501      * for ramblock-id set by compat properties on old machine types ( <= 4.0),
 502      * to keep migration working when backend is used for main RAM with
 503      * -machine memory-backend= option (main RAM historically used prefix-less
 504      * ramblock-id).
 505      */
 506     object_class_property_add_bool(oc, "x-use-canonical-path-for-ramblock-id",
 507         host_memory_backend_get_use_canonical_path,
 508         host_memory_backend_set_use_canonical_path);
 509 }
 510
 511 static const TypeInfo host_memory_backend_info = {
 512     .name = TYPE_MEMORY_BACKEND,
 513     .parent = TYPE_OBJECT,
 514     .abstract = true,
 515     .class_size = sizeof(HostMemoryBackendClass),
 516     .class_init = host_memory_backend_class_init,
 517     .instance_size = sizeof(HostMemoryBackend),
 518     .instance_init = host_memory_backend_init,
 519     .instance_post_init = host_memory_backend_post_init,
 520     .interfaces = (InterfaceInfo[]) {
 521         { TYPE_USER_CREATABLE },
 522         { }
 523     }
 524 };
 525
 526 static void register_types(void)
 527 {
 528     type_register_static(&host_memory_backend_info);
 529 }
 530
 531 type_init(register_types);