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 "qemu/cutils.h"
  24 #include "hw/qdev-core.h"
  25
  26 #ifdef CONFIG_NUMA
  27 #include <numaif.h>
  28 #include <numa.h>
  29 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_DEFAULT != MPOL_DEFAULT);
  30 /*
  31  * HOST_MEM_POLICY_PREFERRED may either translate to MPOL_PREFERRED or
  32  * MPOL_PREFERRED_MANY, see comments further below.
  33  */
  34 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_PREFERRED != MPOL_PREFERRED);
  35 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_BIND != MPOL_BIND);
  36 QEMU_BUILD_BUG_ON(HOST_MEM_POLICY_INTERLEAVE != MPOL_INTERLEAVE);
  37 #endif
  38
  39 char *
  40 host_memory_backend_get_name(HostMemoryBackend *backend)
  41 {
  42     if (!backend->use_canonical_path) {
  43         return g_strdup(object_get_canonical_path_component(OBJECT(backend)));
  44     }
  45
  46     return object_get_canonical_path(OBJECT(backend));
  47 }
  48
  49 static void
  50 host_memory_backend_get_size(Object *obj, Visitor *v, const char *name,
  51                              void *opaque, Error **errp)
  52 {
  53     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
  54     uint64_t value = backend->size;
  55
  56     visit_type_size(v, name, &value, errp);
  57 }
  58
  59 static void
  60 host_memory_backend_set_size(Object *obj, Visitor *v, const char *name,
  61                              void *opaque, Error **errp)
  62 {
  63     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
  64     uint64_t value;
  65
  66     if (host_memory_backend_mr_inited(backend)) {
  67         error_setg(errp, "cannot change property %s of %s ", name,
  68                    object_get_typename(obj));
  69         return;
  70     }
  71
  72     if (!visit_type_size(v, name, &value, errp)) {
  73         return;
  74     }
  75     if (!value) {
  76         error_setg(errp,
  77                    "property '%s' of %s doesn't take value '%" PRIu64 "'",
  78                    name, object_get_typename(obj), value);
  79         return;
  80     }
  81     backend->size = value;
  82 }
  83
  84 static void
  85 host_memory_backend_get_host_nodes(Object *obj, Visitor *v, const char *name,
  86                                    void *opaque, Error **errp)
  87 {
  88     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
  89     uint16List *host_nodes = NULL;
  90     uint16List **tail = &host_nodes;
  91     unsigned long value;
  92
  93     value = find_first_bit(backend->host_nodes, MAX_NODES);
  94     if (value == MAX_NODES) {
  95         goto ret;
  96     }
  97
  98     QAPI_LIST_APPEND(tail, value);
  99
 100     do {
 101         value = find_next_bit(backend->host_nodes, MAX_NODES, value + 1);
 102         if (value == MAX_NODES) {
 103             break;
 104         }
 105
 106         QAPI_LIST_APPEND(tail, value);
 107     } while (true);
 108
 109 ret:
 110     visit_type_uint16List(v, name, &host_nodes, errp);
 111     qapi_free_uint16List(host_nodes);
 112 }
 113
 114 static void
 115 host_memory_backend_set_host_nodes(Object *obj, Visitor *v, const char *name,
 116                                    void *opaque, Error **errp)
 117 {
 118 #ifdef CONFIG_NUMA
 119     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 120     uint16List *l, *host_nodes = NULL;
 121
 122     visit_type_uint16List(v, name, &host_nodes, errp);
 123
 124     for (l = host_nodes; l; l = l->next) {
 125         if (l->value >= MAX_NODES) {
 126             error_setg(errp, "Invalid host-nodes value: %d", l->value);
 127             goto out;
 128         }
 129     }
 130
 131     for (l = host_nodes; l; l = l->next) {
 132         bitmap_set(backend->host_nodes, l->value, 1);
 133     }
 134
 135 out:
 136     qapi_free_uint16List(host_nodes);
 137 #else
 138     error_setg(errp, "NUMA node binding are not supported by this QEMU");
 139 #endif
 140 }
 141
 142 static int
 143 host_memory_backend_get_policy(Object *obj, Error **errp G_GNUC_UNUSED)
 144 {
 145     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 146     return backend->policy;
 147 }
 148
 149 static void
 150 host_memory_backend_set_policy(Object *obj, int policy, Error **errp)
 151 {
 152     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 153     backend->policy = policy;
 154
 155 #ifndef CONFIG_NUMA
 156     if (policy != HOST_MEM_POLICY_DEFAULT) {
 157         error_setg(errp, "NUMA policies are not supported by this QEMU");
 158     }
 159 #endif
 160 }
 161
 162 static bool host_memory_backend_get_merge(Object *obj, Error **errp)
 163 {
 164     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 165
 166     return backend->merge;
 167 }
 168
 169 static void host_memory_backend_set_merge(Object *obj, bool value, Error **errp)
 170 {
 171     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 172
 173     if (QEMU_MADV_MERGEABLE == QEMU_MADV_INVALID) {
 174         if (value) {
 175             error_setg(errp, "Memory merging is not supported on this host");
 176         }
 177         assert(!backend->merge);
 178         return;
 179     }
 180
 181     if (!host_memory_backend_mr_inited(backend) &&
 182         value != backend->merge) {
 183         void *ptr = memory_region_get_ram_ptr(&backend->mr);
 184         uint64_t sz = memory_region_size(&backend->mr);
 185
 186         qemu_madvise(ptr, sz,
 187                      value ? QEMU_MADV_MERGEABLE : QEMU_MADV_UNMERGEABLE);
 188     }
 189
 190     backend->merge = value;
 191 }
 192
 193 static bool host_memory_backend_get_dump(Object *obj, Error **errp)
 194 {
 195     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 196
 197     return backend->dump;
 198 }
 199
 200 static void host_memory_backend_set_dump(Object *obj, bool value, Error **errp)
 201 {
 202     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 203
 204     if (QEMU_MADV_DONTDUMP == QEMU_MADV_INVALID) {
 205         if (!value) {
 206             error_setg(errp, "Dumping guest memory cannot be disabled on this host");
 207         }
 208         assert(backend->dump);
 209         return;
 210     }
 211
 212     if (host_memory_backend_mr_inited(backend) &&
 213         value != backend->dump) {
 214         void *ptr = memory_region_get_ram_ptr(&backend->mr);
 215         uint64_t sz = memory_region_size(&backend->mr);
 216
 217         qemu_madvise(ptr, sz,
 218                      value ? QEMU_MADV_DODUMP : QEMU_MADV_DONTDUMP);
 219     }
 220
 221     backend->dump = value;
 222 }
 223
 224 static bool host_memory_backend_get_prealloc(Object *obj, Error **errp)
 225 {
 226     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 227
 228     return backend->prealloc;
 229 }
 230
 231 static void host_memory_backend_set_prealloc(Object *obj, bool value,
 232                                              Error **errp)
 233 {
 234     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 235
 236     if (!backend->reserve && value) {
 237         error_setg(errp, "'prealloc=on' and 'reserve=off' are incompatible");
 238         return;
 239     }
 240
 241     if (!host_memory_backend_mr_inited(backend)) {
 242         backend->prealloc = value;
 243         return;
 244     }
 245
 246     if (value && !backend->prealloc) {
 247         int fd = memory_region_get_fd(&backend->mr);
 248         void *ptr = memory_region_get_ram_ptr(&backend->mr);
 249         uint64_t sz = memory_region_size(&backend->mr);
 250
 251         if (!qemu_prealloc_mem(fd, ptr, sz, backend->prealloc_threads,
 252                                backend->prealloc_context, false, errp)) {
 253             return;
 254         }
 255         backend->prealloc = true;
 256     }
 257 }
 258
 259 static void host_memory_backend_get_prealloc_threads(Object *obj, Visitor *v,
 260     const char *name, void *opaque, Error **errp)
 261 {
 262     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 263     visit_type_uint32(v, name, &backend->prealloc_threads, errp);
 264 }
 265
 266 static void host_memory_backend_set_prealloc_threads(Object *obj, Visitor *v,
 267     const char *name, void *opaque, Error **errp)
 268 {
 269     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 270     uint32_t value;
 271
 272     if (!visit_type_uint32(v, name, &value, errp)) {
 273         return;
 274     }
 275     if (value <= 0) {
 276         error_setg(errp, "property '%s' of %s doesn't take value '%d'", name,
 277                    object_get_typename(obj), value);
 278         return;
 279     }
 280     backend->prealloc_threads = value;
 281 }
 282
 283 static void host_memory_backend_init(Object *obj)
 284 {
 285     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 286     MachineState *machine = MACHINE(qdev_get_machine());
 287
 288     /* TODO: convert access to globals to compat properties */
 289     backend->merge = machine_mem_merge(machine);
 290     backend->dump = machine_dump_guest_core(machine);
 291     backend->guest_memfd = machine_require_guest_memfd(machine);
 292     backend->reserve = true;
 293     backend->prealloc_threads = machine->smp.cpus;
 294 }
 295
 296 static void host_memory_backend_post_init(Object *obj)
 297 {
 298     object_apply_compat_props(obj);
 299 }
 300
 301 bool host_memory_backend_mr_inited(HostMemoryBackend *backend)
 302 {
 303     /*
 304      * NOTE: We forbid zero-length memory backend, so here zero means
 305      * "we haven't inited the backend memory region yet".
 306      */
 307     return memory_region_size(&backend->mr) != 0;
 308 }
 309
 310 MemoryRegion *host_memory_backend_get_memory(HostMemoryBackend *backend)
 311 {
 312     return host_memory_backend_mr_inited(backend) ? &backend->mr : NULL;
 313 }
 314
 315 void host_memory_backend_set_mapped(HostMemoryBackend *backend, bool mapped)
 316 {
 317     backend->is_mapped = mapped;
 318 }
 319
 320 bool host_memory_backend_is_mapped(HostMemoryBackend *backend)
 321 {
 322     return backend->is_mapped;
 323 }
 324
 325 size_t host_memory_backend_pagesize(HostMemoryBackend *memdev)
 326 {
 327     size_t pagesize = qemu_ram_pagesize(memdev->mr.ram_block);
 328     g_assert(pagesize >= qemu_real_host_page_size());
 329     return pagesize;
 330 }
 331
 332 static void
 333 host_memory_backend_memory_complete(UserCreatable *uc, Error **errp)
 334 {
 335     HostMemoryBackend *backend = MEMORY_BACKEND(uc);
 336     HostMemoryBackendClass *bc = MEMORY_BACKEND_GET_CLASS(uc);
 337     void *ptr;
 338     uint64_t sz;
 339     size_t pagesize;
 340     bool async = !phase_check(PHASE_LATE_BACKENDS_CREATED);
 341
 342     if (!bc->alloc) {
 343         return;
 344     }
 345     if (!bc->alloc(backend, errp)) {
 346         return;
 347     }
 348
 349     ptr = memory_region_get_ram_ptr(&backend->mr);
 350     sz = memory_region_size(&backend->mr);
 351     pagesize = qemu_ram_pagesize(backend->mr.ram_block);
 352
 353     if (backend->aligned && !QEMU_IS_ALIGNED(sz, pagesize)) {
 354         g_autofree char *pagesize_str = size_to_str(pagesize);
 355         error_setg(errp, "backend '%s' memory size must be multiple of %s",
 356                    object_get_typename(OBJECT(uc)), pagesize_str);
 357         return;
 358     }
 359
 360     if (backend->merge) {
 361         qemu_madvise(ptr, sz, QEMU_MADV_MERGEABLE);
 362     }
 363     if (!backend->dump) {
 364         qemu_madvise(ptr, sz, QEMU_MADV_DONTDUMP);
 365     }
 366 #ifdef CONFIG_NUMA
 367     unsigned long lastbit = find_last_bit(backend->host_nodes, MAX_NODES);
 368     /* lastbit == MAX_NODES means maxnode = 0 */
 369     unsigned long maxnode = (lastbit + 1) % (MAX_NODES + 1);
 370     /*
 371      * Ensure policy won't be ignored in case memory is preallocated
 372      * before mbind(). note: MPOL_MF_STRICT is ignored on hugepages so
 373      * this doesn't catch hugepage case.
 374      */
 375     unsigned flags = MPOL_MF_STRICT | MPOL_MF_MOVE;
 376     int mode = backend->policy;
 377
 378     /* check for invalid host-nodes and policies and give more verbose
 379      * error messages than mbind(). */
 380     if (maxnode && backend->policy == MPOL_DEFAULT) {
 381         error_setg(errp, "host-nodes must be empty for policy default,"
 382                    " or you should explicitly specify a policy other"
 383                    " than default");
 384         return;
 385     } else if (maxnode == 0 && backend->policy != MPOL_DEFAULT) {
 386         error_setg(errp, "host-nodes must be set for policy %s",
 387                    HostMemPolicy_str(backend->policy));
 388         return;
 389     }
 390
 391     /*
 392      * We can have up to MAX_NODES nodes, but we need to pass maxnode+1
 393      * as argument to mbind() due to an old Linux bug (feature?) which
 394      * cuts off the last specified node. This means backend->host_nodes
 395      * must have MAX_NODES+1 bits available.
 396      */
 397     assert(sizeof(backend->host_nodes) >=
 398            BITS_TO_LONGS(MAX_NODES + 1) * sizeof(unsigned long));
 399     assert(maxnode <= MAX_NODES);
 400
 401 #ifdef HAVE_NUMA_HAS_PREFERRED_MANY
 402     if (mode == MPOL_PREFERRED && numa_has_preferred_many() > 0) {
 403         /*
 404          * Replace with MPOL_PREFERRED_MANY otherwise the mbind() below
 405          * silently picks the first node.
 406          */
 407         mode = MPOL_PREFERRED_MANY;
 408     }
 409 #endif
 410
 411     if (maxnode &&
 412         mbind(ptr, sz, mode, backend->host_nodes, maxnode + 1, flags)) {
 413         if (backend->policy != MPOL_DEFAULT || errno != ENOSYS) {
 414             error_setg_errno(errp, errno,
 415                              "cannot bind memory to host NUMA nodes");
 416             return;
 417         }
 418     }
 419 #endif
 420     /*
 421      * Preallocate memory after the NUMA policy has been instantiated.
 422      * This is necessary to guarantee memory is allocated with
 423      * specified NUMA policy in place.
 424      */
 425     if (backend->prealloc && !qemu_prealloc_mem(memory_region_get_fd(&backend->mr),
 426                                                 ptr, sz,
 427                                                 backend->prealloc_threads,
 428                                                 backend->prealloc_context,
 429                                                 async, errp)) {
 430         return;
 431     }
 432 }
 433
 434 static bool
 435 host_memory_backend_can_be_deleted(UserCreatable *uc)
 436 {
 437     if (host_memory_backend_is_mapped(MEMORY_BACKEND(uc))) {
 438         return false;
 439     } else {
 440         return true;
 441     }
 442 }
 443
 444 static bool host_memory_backend_get_share(Object *o, Error **errp)
 445 {
 446     HostMemoryBackend *backend = MEMORY_BACKEND(o);
 447
 448     return backend->share;
 449 }
 450
 451 static void host_memory_backend_set_share(Object *o, bool value, Error **errp)
 452 {
 453     HostMemoryBackend *backend = MEMORY_BACKEND(o);
 454
 455     if (host_memory_backend_mr_inited(backend)) {
 456         error_setg(errp, "cannot change property value");
 457         return;
 458     }
 459     backend->share = value;
 460 }
 461
 462 #ifdef CONFIG_LINUX
 463 static bool host_memory_backend_get_reserve(Object *o, Error **errp)
 464 {
 465     HostMemoryBackend *backend = MEMORY_BACKEND(o);
 466
 467     return backend->reserve;
 468 }
 469
 470 static void host_memory_backend_set_reserve(Object *o, bool value, Error **errp)
 471 {
 472     HostMemoryBackend *backend = MEMORY_BACKEND(o);
 473
 474     if (host_memory_backend_mr_inited(backend)) {
 475         error_setg(errp, "cannot change property value");
 476         return;
 477     }
 478     if (backend->prealloc && !value) {
 479         error_setg(errp, "'prealloc=on' and 'reserve=off' are incompatible");
 480         return;
 481     }
 482     backend->reserve = value;
 483 }
 484 #endif /* CONFIG_LINUX */
 485
 486 static bool
 487 host_memory_backend_get_use_canonical_path(Object *obj, Error **errp)
 488 {
 489     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 490
 491     return backend->use_canonical_path;
 492 }
 493
 494 static void
 495 host_memory_backend_set_use_canonical_path(Object *obj, bool value,
 496                                            Error **errp)
 497 {
 498     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
 499
 500     backend->use_canonical_path = value;
 501 }
 502
 503 static void
 504 host_memory_backend_class_init(ObjectClass *oc, void *data)
 505 {
 506     UserCreatableClass *ucc = USER_CREATABLE_CLASS(oc);
 507
 508     ucc->complete = host_memory_backend_memory_complete;
 509     ucc->can_be_deleted = host_memory_backend_can_be_deleted;
 510
 511     object_class_property_add_bool(oc, "merge",
 512         host_memory_backend_get_merge,
 513         host_memory_backend_set_merge);
 514     object_class_property_set_description(oc, "merge",
 515         "Mark memory as mergeable");
 516     object_class_property_add_bool(oc, "dump",
 517         host_memory_backend_get_dump,
 518         host_memory_backend_set_dump);
 519     object_class_property_set_description(oc, "dump",
 520         "Set to 'off' to exclude from core dump");
 521     object_class_property_add_bool(oc, "prealloc",
 522         host_memory_backend_get_prealloc,
 523         host_memory_backend_set_prealloc);
 524     object_class_property_set_description(oc, "prealloc",
 525         "Preallocate memory");
 526     object_class_property_add(oc, "prealloc-threads", "int",
 527         host_memory_backend_get_prealloc_threads,
 528         host_memory_backend_set_prealloc_threads,
 529         NULL, NULL);
 530     object_class_property_set_description(oc, "prealloc-threads",
 531         "Number of CPU threads to use for prealloc");
 532     object_class_property_add_link(oc, "prealloc-context",
 533         TYPE_THREAD_CONTEXT, offsetof(HostMemoryBackend, prealloc_context),
 534         object_property_allow_set_link, OBJ_PROP_LINK_STRONG);
 535     object_class_property_set_description(oc, "prealloc-context",
 536         "Context to use for creating CPU threads for preallocation");
 537     object_class_property_add(oc, "size", "int",
 538         host_memory_backend_get_size,
 539         host_memory_backend_set_size,
 540         NULL, NULL);
 541     object_class_property_set_description(oc, "size",
 542         "Size of the memory region (ex: 500M)");
 543     object_class_property_add(oc, "host-nodes", "int",
 544         host_memory_backend_get_host_nodes,
 545         host_memory_backend_set_host_nodes,
 546         NULL, NULL);
 547     object_class_property_set_description(oc, "host-nodes",
 548         "Binds memory to the list of NUMA host nodes");
 549     object_class_property_add_enum(oc, "policy", "HostMemPolicy",
 550         &HostMemPolicy_lookup,
 551         host_memory_backend_get_policy,
 552         host_memory_backend_set_policy);
 553     object_class_property_set_description(oc, "policy",
 554         "Set the NUMA policy");
 555     object_class_property_add_bool(oc, "share",
 556         host_memory_backend_get_share, host_memory_backend_set_share);
 557     object_class_property_set_description(oc, "share",
 558         "Mark the memory as private to QEMU or shared");
 559 #ifdef CONFIG_LINUX
 560     object_class_property_add_bool(oc, "reserve",
 561         host_memory_backend_get_reserve, host_memory_backend_set_reserve);
 562     object_class_property_set_description(oc, "reserve",
 563         "Reserve swap space (or huge pages) if applicable");
 564 #endif /* CONFIG_LINUX */
 565     /*
 566      * Do not delete/rename option. This option must be considered stable
 567      * (as if it didn't have the 'x-' prefix including deprecation period) as
 568      * long as 4.0 and older machine types exists.
 569      * Option will be used by upper layers to override (disable) canonical path
 570      * for ramblock-id set by compat properties on old machine types ( <= 4.0),
 571      * to keep migration working when backend is used for main RAM with
 572      * -machine memory-backend= option (main RAM historically used prefix-less
 573      * ramblock-id).
 574      */
 575     object_class_property_add_bool(oc, "x-use-canonical-path-for-ramblock-id",
 576         host_memory_backend_get_use_canonical_path,
 577         host_memory_backend_set_use_canonical_path);
 578 }
 579
 580 static const TypeInfo host_memory_backend_info = {
 581     .name = TYPE_MEMORY_BACKEND,
 582     .parent = TYPE_OBJECT,
 583     .abstract = true,
 584     .class_size = sizeof(HostMemoryBackendClass),
 585     .class_init = host_memory_backend_class_init,
 586     .instance_size = sizeof(HostMemoryBackend),
 587     .instance_init = host_memory_backend_init,
 588     .instance_post_init = host_memory_backend_post_init,
 589     .interfaces = (InterfaceInfo[]) {
 590         { TYPE_USER_CREATABLE },
 591         { }
 592     }
 593 };
 594
 595 static void register_types(void)
 596 {
 597     type_register_static(&host_memory_backend_info);
 598 }
 599
 600 type_init(register_types);