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     Error *local_err = NULL;
  58     uint64_t value;
  59
  60     if (host_memory_backend_mr_inited(backend)) {
  61         error_setg(errp, "cannot change property %s of %s ", name,
  62                    object_get_typename(obj));
  63         return;
  64     }
  65
  66     if (!visit_type_size(v, name, &value, &local_err)) {
  67         error_propagate(errp, local_err);
  68         return;
  69     }
  70     if (!value) {
  71         error_setg(errp,
  72                    "property '%s' of %s doesn't take value '%" PRIu64 "'",
  73                    name, object_get_typename(obj), value);
  74         return;
  75     }
  76     backend->size = value;
  77 }
  78
  79 static void
  80 host_memory_backend_get_host_nodes(Object *obj, Visitor *v, const char *name,
  81                                    void *opaque, Error **errp)
  82 {
  83     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
  84     uint16List *host_nodes = NULL;
  85     uint16List **node = &host_nodes;
  86     unsigned long value;
  87
  88     value = find_first_bit(backend->host_nodes, MAX_NODES);
  89     if (value == MAX_NODES) {
  90         goto ret;
  91     }
  92
  93     *node = g_malloc0(sizeof(**node));
  94     (*node)->value = value;
  95     node = &(*node)->next;
  96
  97     do {
  98         value = find_next_bit(backend->host_nodes, MAX_NODES, value + 1);
  99         if (value == MAX_NODES) {
 100             break;
 101         }
 102
 103         *node = g_malloc0(sizeof(**node));
 104         (*node)->value = value;
 105         node = &(*node)->next;
 106     } while (true);
 107
 108 ret:
 109     visit_type_uint16List(v, name, &host_nodes, errp);
 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     Error *local_err = NULL;
 223     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 224
 225     if (!host_memory_backend_mr_inited(backend)) {
 226         backend->prealloc = value;
 227         return;
 228     }
 229
 230     if (value && !backend->prealloc) {
 231         int fd = memory_region_get_fd(&backend->mr);
 232         void *ptr = memory_region_get_ram_ptr(&backend->mr);
 233         uint64_t sz = memory_region_size(&backend->mr);
 234
 235         os_mem_prealloc(fd, ptr, sz, backend->prealloc_threads, &local_err);
 236         if (local_err) {
 237             error_propagate(errp, local_err);
 238             return;
 239         }
 240         backend->prealloc = true;
 241     }
 242 }
 243
 244 static void host_memory_backend_get_prealloc_threads(Object *obj, Visitor *v,
 245     const char *name, void *opaque, Error **errp)
 246 {
 247     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 248     visit_type_uint32(v, name, &backend->prealloc_threads, errp);
 249 }
 250
 251 static void host_memory_backend_set_prealloc_threads(Object *obj, Visitor *v,
 252     const char *name, void *opaque, Error **errp)
 253 {
 254     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 255     Error *local_err = NULL;
 256     uint32_t value;
 257
 258     if (!visit_type_uint32(v, name, &value, &local_err)) {
 259         error_propagate(errp, local_err);
 260         return;
 261     }
 262     if (value <= 0) {
 263         error_setg(errp, "property '%s' of %s doesn't take value '%d'", name,
 264                    object_get_typename(obj), value);
 265         return;
 266     }
 267     backend->prealloc_threads = value;
 268 }
 269
 270 static void host_memory_backend_init(Object *obj)
 271 {
 272     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 273     MachineState *machine = MACHINE(qdev_get_machine());
 274
 275     /* TODO: convert access to globals to compat properties */
 276     backend->merge = machine_mem_merge(machine);
 277     backend->dump = machine_dump_guest_core(machine);
 278     backend->prealloc_threads = 1;
 279 }
 280
 281 static void host_memory_backend_post_init(Object *obj)
 282 {
 283     object_apply_compat_props(obj);
 284 }
 285
 286 bool host_memory_backend_mr_inited(HostMemoryBackend *backend)
 287 {
 288     /*
 289      * NOTE: We forbid zero-length memory backend, so here zero means
 290      * "we haven't inited the backend memory region yet".
 291      */
 292     return memory_region_size(&backend->mr) != 0;
 293 }
 294
 295 MemoryRegion *host_memory_backend_get_memory(HostMemoryBackend *backend)
 296 {
 297     return host_memory_backend_mr_inited(backend) ? &backend->mr : NULL;
 298 }
 299
 300 void host_memory_backend_set_mapped(HostMemoryBackend *backend, bool mapped)
 301 {
 302     backend->is_mapped = mapped;
 303 }
 304
 305 bool host_memory_backend_is_mapped(HostMemoryBackend *backend)
 306 {
 307     return backend->is_mapped;
 308 }
 309
 310 #ifdef __linux__
 311 size_t host_memory_backend_pagesize(HostMemoryBackend *memdev)
 312 {
 313     Object *obj = OBJECT(memdev);
 314     char *path = object_property_get_str(obj, "mem-path", NULL);
 315     size_t pagesize = qemu_mempath_getpagesize(path);
 316
 317     g_free(path);
 318     return pagesize;
 319 }
 320 #else
 321 size_t host_memory_backend_pagesize(HostMemoryBackend *memdev)
 322 {
 323     return qemu_real_host_page_size;
 324 }
 325 #endif
 326
 327 static void
 328 host_memory_backend_memory_complete(UserCreatable *uc, Error **errp)
 329 {
 330     HostMemoryBackend *backend = MEMORY_BACKEND(uc);
 331     HostMemoryBackendClass *bc = MEMORY_BACKEND_GET_CLASS(uc);
 332     Error *local_err = NULL;
 333     void *ptr;
 334     uint64_t sz;
 335
 336     if (bc->alloc) {
 337         bc->alloc(backend, &local_err);
 338         if (local_err) {
 339             goto out;
 340         }
 341
 342         ptr = memory_region_get_ram_ptr(&backend->mr);
 343         sz = memory_region_size(&backend->mr);
 344
 345         if (backend->merge) {
 346             qemu_madvise(ptr, sz, QEMU_MADV_MERGEABLE);
 347         }
 348         if (!backend->dump) {
 349             qemu_madvise(ptr, sz, QEMU_MADV_DONTDUMP);
 350         }
 351 #ifdef CONFIG_NUMA
 352         unsigned long lastbit = find_last_bit(backend->host_nodes, MAX_NODES);
 353         /* lastbit == MAX_NODES means maxnode = 0 */
 354         unsigned long maxnode = (lastbit + 1) % (MAX_NODES + 1);
 355         /* ensure policy won't be ignored in case memory is preallocated
 356          * before mbind(). note: MPOL_MF_STRICT is ignored on hugepages so
 357          * this doesn't catch hugepage case. */
 358         unsigned flags = MPOL_MF_STRICT | MPOL_MF_MOVE;
 359
 360         /* check for invalid host-nodes and policies and give more verbose
 361          * error messages than mbind(). */
 362         if (maxnode && backend->policy == MPOL_DEFAULT) {
 363             error_setg(errp, "host-nodes must be empty for policy default,"
 364                        " or you should explicitly specify a policy other"
 365                        " than default");
 366             return;
 367         } else if (maxnode == 0 && backend->policy != MPOL_DEFAULT) {
 368             error_setg(errp, "host-nodes must be set for policy %s",
 369                        HostMemPolicy_str(backend->policy));
 370             return;
 371         }
 372
 373         /* We can have up to MAX_NODES nodes, but we need to pass maxnode+1
 374          * as argument to mbind() due to an old Linux bug (feature?) which
 375          * cuts off the last specified node. This means backend->host_nodes
 376          * must have MAX_NODES+1 bits available.
 377          */
 378         assert(sizeof(backend->host_nodes) >=
 379                BITS_TO_LONGS(MAX_NODES + 1) * sizeof(unsigned long));
 380         assert(maxnode <= MAX_NODES);
 381
 382         if (maxnode &&
 383             mbind(ptr, sz, backend->policy, backend->host_nodes, maxnode + 1,
 384                   flags)) {
 385             if (backend->policy != MPOL_DEFAULT || errno != ENOSYS) {
 386                 error_setg_errno(errp, errno,
 387                                  "cannot bind memory to host NUMA nodes");
 388                 return;
 389             }
 390         }
 391 #endif
 392         /* Preallocate memory after the NUMA policy has been instantiated.
 393          * This is necessary to guarantee memory is allocated with
 394          * specified NUMA policy in place.
 395          */
 396         if (backend->prealloc) {
 397             os_mem_prealloc(memory_region_get_fd(&backend->mr), ptr, sz,
 398                             backend->prealloc_threads, &local_err);
 399             if (local_err) {
 400                 goto out;
 401             }
 402         }
 403     }
 404 out:
 405     error_propagate(errp, local_err);
 406 }
 407
 408 static bool
 409 host_memory_backend_can_be_deleted(UserCreatable *uc)
 410 {
 411     if (host_memory_backend_is_mapped(MEMORY_BACKEND(uc))) {
 412         return false;
 413     } else {
 414         return true;
 415     }
 416 }
 417
 418 static bool host_memory_backend_get_share(Object *o, Error **errp)
 419 {
 420     HostMemoryBackend *backend = MEMORY_BACKEND(o);
 421
 422     return backend->share;
 423 }
 424
 425 static void host_memory_backend_set_share(Object *o, bool value, Error **errp)
 426 {
 427     HostMemoryBackend *backend = MEMORY_BACKEND(o);
 428
 429     if (host_memory_backend_mr_inited(backend)) {
 430         error_setg(errp, "cannot change property value");
 431         return;
 432     }
 433     backend->share = value;
 434 }
 435
 436 static bool
 437 host_memory_backend_get_use_canonical_path(Object *obj, Error **errp)
 438 {
 439     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 440
 441     return backend->use_canonical_path;
 442 }
 443
 444 static void
 445 host_memory_backend_set_use_canonical_path(Object *obj, bool value,
 446                                            Error **errp)
 447 {
 448     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 449
 450     backend->use_canonical_path = value;
 451 }
 452
 453 static void
 454 host_memory_backend_class_init(ObjectClass *oc, void *data)
 455 {
 456     UserCreatableClass *ucc = USER_CREATABLE_CLASS(oc);
 457
 458     ucc->complete = host_memory_backend_memory_complete;
 459     ucc->can_be_deleted = host_memory_backend_can_be_deleted;
 460
 461     object_class_property_add_bool(oc, "merge",
 462         host_memory_backend_get_merge,
 463         host_memory_backend_set_merge);
 464     object_class_property_set_description(oc, "merge",
 465         "Mark memory as mergeable");
 466     object_class_property_add_bool(oc, "dump",
 467         host_memory_backend_get_dump,
 468         host_memory_backend_set_dump);
 469     object_class_property_set_description(oc, "dump",
 470         "Set to 'off' to exclude from core dump");
 471     object_class_property_add_bool(oc, "prealloc",
 472         host_memory_backend_get_prealloc,
 473         host_memory_backend_set_prealloc);
 474     object_class_property_set_description(oc, "prealloc",
 475         "Preallocate memory");
 476     object_class_property_add(oc, "prealloc-threads", "int",
 477         host_memory_backend_get_prealloc_threads,
 478         host_memory_backend_set_prealloc_threads,
 479         NULL, NULL);
 480     object_class_property_set_description(oc, "prealloc-threads",
 481         "Number of CPU threads to use for prealloc");
 482     object_class_property_add(oc, "size", "int",
 483         host_memory_backend_get_size,
 484         host_memory_backend_set_size,
 485         NULL, NULL);
 486     object_class_property_set_description(oc, "size",
 487         "Size of the memory region (ex: 500M)");
 488     object_class_property_add(oc, "host-nodes", "int",
 489         host_memory_backend_get_host_nodes,
 490         host_memory_backend_set_host_nodes,
 491         NULL, NULL);
 492     object_class_property_set_description(oc, "host-nodes",
 493         "Binds memory to the list of NUMA host nodes");
 494     object_class_property_add_enum(oc, "policy", "HostMemPolicy",
 495         &HostMemPolicy_lookup,
 496         host_memory_backend_get_policy,
 497         host_memory_backend_set_policy);
 498     object_class_property_set_description(oc, "policy",
 499         "Set the NUMA policy");
 500     object_class_property_add_bool(oc, "share",
 501         host_memory_backend_get_share, host_memory_backend_set_share);
 502     object_class_property_set_description(oc, "share",
 503         "Mark the memory as private to QEMU or shared");
 504     object_class_property_add_bool(oc, "x-use-canonical-path-for-ramblock-id",
 505         host_memory_backend_get_use_canonical_path,
 506         host_memory_backend_set_use_canonical_path);
 507 }
 508
 509 static const TypeInfo host_memory_backend_info = {
 510     .name = TYPE_MEMORY_BACKEND,
 511     .parent = TYPE_OBJECT,
 512     .abstract = true,
 513     .class_size = sizeof(HostMemoryBackendClass),
 514     .class_init = host_memory_backend_class_init,
 515     .instance_size = sizeof(HostMemoryBackend),
 516     .instance_init = host_memory_backend_init,
 517     .instance_post_init = host_memory_backend_post_init,
 518     .interfaces = (InterfaceInfo[]) {
 519         { TYPE_USER_CREATABLE },
 520         { }
 521     }
 522 };
 523
 524 static void register_types(void)
 525 {
 526     type_register_static(&host_memory_backend_info);
 527 }
 528
 529 type_init(register_types);