xref: /openbmc/qemu/backends/hostmem.c (revision ea73f370) 
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 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     Error *local_err = NULL;
57     uint64_t value;
58 
59     if (host_memory_backend_mr_inited(backend)) {
60         error_setg(&local_err, "cannot change property %s of %s ",
61                    name, object_get_typename(obj));
62         goto out;
63     }
64 
65     visit_type_size(v, name, &value, &local_err);
66     if (local_err) {
67         goto out;
68     }
69     if (!value) {
70         error_setg(&local_err,
71                    "property '%s' of %s doesn't take value '%" PRIu64 "'",
72                    name, object_get_typename(obj), value);
73         goto out;
74     }
75     backend->size = value;
76 out:
77     error_propagate(errp, local_err);
78 }
79 
80 static void
81 host_memory_backend_get_host_nodes(Object *obj, Visitor *v, const char *name,
82                                    void *opaque, Error **errp)
83 {
84     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
85     uint16List *host_nodes = NULL;
86     uint16List **node = &host_nodes;
87     unsigned long value;
88 
89     value = find_first_bit(backend->host_nodes, MAX_NODES);
90     if (value == MAX_NODES) {
91         goto ret;
92     }
93 
94     *node = g_malloc0(sizeof(**node));
95     (*node)->value = value;
96     node = &(*node)->next;
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         *node = g_malloc0(sizeof(**node));
105         (*node)->value = value;
106         node = &(*node)->next;
107     } while (true);
108 
109 ret:
110     visit_type_uint16List(v, name, &host_nodes, errp);
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 || backend->force_prealloc;
218 }
219 
220 static void host_memory_backend_set_prealloc(Object *obj, bool value,
221                                              Error **errp)
222 {
223     Error *local_err = NULL;
224     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
225     MachineState *ms = MACHINE(qdev_get_machine());
226 
227     if (backend->force_prealloc) {
228         if (value) {
229             error_setg(errp,
230                        "remove -mem-prealloc to use the prealloc property");
231             return;
232         }
233     }
234 
235     if (!host_memory_backend_mr_inited(backend)) {
236         backend->prealloc = value;
237         return;
238     }
239 
240     if (value && !backend->prealloc) {
241         int fd = memory_region_get_fd(&backend->mr);
242         void *ptr = memory_region_get_ram_ptr(&backend->mr);
243         uint64_t sz = memory_region_size(&backend->mr);
244 
245         os_mem_prealloc(fd, ptr, sz, ms->smp.cpus, &local_err);
246         if (local_err) {
247             error_propagate(errp, local_err);
248             return;
249         }
250         backend->prealloc = true;
251     }
252 }
253 
254 static void host_memory_backend_init(Object *obj)
255 {
256     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
257     MachineState *machine = MACHINE(qdev_get_machine());
258 
259     backend->merge = machine_mem_merge(machine);
260     backend->dump = machine_dump_guest_core(machine);
261     backend->prealloc = mem_prealloc;
262 }
263 
264 static void host_memory_backend_post_init(Object *obj)
265 {
266     object_apply_compat_props(obj);
267 }
268 
269 bool host_memory_backend_mr_inited(HostMemoryBackend *backend)
270 {
271     /*
272      * NOTE: We forbid zero-length memory backend, so here zero means
273      * "we haven't inited the backend memory region yet".
274      */
275     return memory_region_size(&backend->mr) != 0;
276 }
277 
278 MemoryRegion *host_memory_backend_get_memory(HostMemoryBackend *backend)
279 {
280     return host_memory_backend_mr_inited(backend) ? &backend->mr : NULL;
281 }
282 
283 void host_memory_backend_set_mapped(HostMemoryBackend *backend, bool mapped)
284 {
285     backend->is_mapped = mapped;
286 }
287 
288 bool host_memory_backend_is_mapped(HostMemoryBackend *backend)
289 {
290     return backend->is_mapped;
291 }
292 
293 #ifdef __linux__
294 size_t host_memory_backend_pagesize(HostMemoryBackend *memdev)
295 {
296     Object *obj = OBJECT(memdev);
297     char *path = object_property_get_str(obj, "mem-path", NULL);
298     size_t pagesize = qemu_mempath_getpagesize(path);
299 
300     g_free(path);
301     return pagesize;
302 }
303 #else
304 size_t host_memory_backend_pagesize(HostMemoryBackend *memdev)
305 {
306     return getpagesize();
307 }
308 #endif
309 
310 static void
311 host_memory_backend_memory_complete(UserCreatable *uc, Error **errp)
312 {
313     HostMemoryBackend *backend = MEMORY_BACKEND(uc);
314     HostMemoryBackendClass *bc = MEMORY_BACKEND_GET_CLASS(uc);
315     MachineState *ms = MACHINE(qdev_get_machine());
316     Error *local_err = NULL;
317     void *ptr;
318     uint64_t sz;
319 
320     if (bc->alloc) {
321         bc->alloc(backend, &local_err);
322         if (local_err) {
323             goto out;
324         }
325 
326         ptr = memory_region_get_ram_ptr(&backend->mr);
327         sz = memory_region_size(&backend->mr);
328 
329         if (backend->merge) {
330             qemu_madvise(ptr, sz, QEMU_MADV_MERGEABLE);
331         }
332         if (!backend->dump) {
333             qemu_madvise(ptr, sz, QEMU_MADV_DONTDUMP);
334         }
335 #ifdef CONFIG_NUMA
336         unsigned long lastbit = find_last_bit(backend->host_nodes, MAX_NODES);
337         /* lastbit == MAX_NODES means maxnode = 0 */
338         unsigned long maxnode = (lastbit + 1) % (MAX_NODES + 1);
339         /* ensure policy won't be ignored in case memory is preallocated
340          * before mbind(). note: MPOL_MF_STRICT is ignored on hugepages so
341          * this doesn't catch hugepage case. */
342         unsigned flags = MPOL_MF_STRICT | MPOL_MF_MOVE;
343 
344         /* check for invalid host-nodes and policies and give more verbose
345          * error messages than mbind(). */
346         if (maxnode && backend->policy == MPOL_DEFAULT) {
347             error_setg(errp, "host-nodes must be empty for policy default,"
348                        " or you should explicitly specify a policy other"
349                        " than default");
350             return;
351         } else if (maxnode == 0 && backend->policy != MPOL_DEFAULT) {
352             error_setg(errp, "host-nodes must be set for policy %s",
353                        HostMemPolicy_str(backend->policy));
354             return;
355         }
356 
357         /* We can have up to MAX_NODES nodes, but we need to pass maxnode+1
358          * as argument to mbind() due to an old Linux bug (feature?) which
359          * cuts off the last specified node. This means backend->host_nodes
360          * must have MAX_NODES+1 bits available.
361          */
362         assert(sizeof(backend->host_nodes) >=
363                BITS_TO_LONGS(MAX_NODES + 1) * sizeof(unsigned long));
364         assert(maxnode <= MAX_NODES);
365         if (mbind(ptr, sz, backend->policy,
366                   maxnode ? backend->host_nodes : NULL, maxnode + 1, flags)) {
367             if (backend->policy != MPOL_DEFAULT || errno != ENOSYS) {
368                 error_setg_errno(errp, errno,
369                                  "cannot bind memory to host NUMA nodes");
370                 return;
371             }
372         }
373 #endif
374         /* Preallocate memory after the NUMA policy has been instantiated.
375          * This is necessary to guarantee memory is allocated with
376          * specified NUMA policy in place.
377          */
378         if (backend->prealloc) {
379             os_mem_prealloc(memory_region_get_fd(&backend->mr), ptr, sz,
380                             ms->smp.cpus, &local_err);
381             if (local_err) {
382                 goto out;
383             }
384         }
385     }
386 out:
387     error_propagate(errp, local_err);
388 }
389 
390 static bool
391 host_memory_backend_can_be_deleted(UserCreatable *uc)
392 {
393     if (host_memory_backend_is_mapped(MEMORY_BACKEND(uc))) {
394         return false;
395     } else {
396         return true;
397     }
398 }
399 
400 static bool host_memory_backend_get_share(Object *o, Error **errp)
401 {
402     HostMemoryBackend *backend = MEMORY_BACKEND(o);
403 
404     return backend->share;
405 }
406 
407 static void host_memory_backend_set_share(Object *o, bool value, Error **errp)
408 {
409     HostMemoryBackend *backend = MEMORY_BACKEND(o);
410 
411     if (host_memory_backend_mr_inited(backend)) {
412         error_setg(errp, "cannot change property value");
413         return;
414     }
415     backend->share = value;
416 }
417 
418 static bool
419 host_memory_backend_get_use_canonical_path(Object *obj, Error **errp)
420 {
421     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
422 
423     return backend->use_canonical_path;
424 }
425 
426 static void
427 host_memory_backend_set_use_canonical_path(Object *obj, bool value,
428                                            Error **errp)
429 {
430     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
431 
432     backend->use_canonical_path = value;
433 }
434 
435 static void
436 host_memory_backend_class_init(ObjectClass *oc, void *data)
437 {
438     UserCreatableClass *ucc = USER_CREATABLE_CLASS(oc);
439 
440     ucc->complete = host_memory_backend_memory_complete;
441     ucc->can_be_deleted = host_memory_backend_can_be_deleted;
442 
443     object_class_property_add_bool(oc, "merge",
444         host_memory_backend_get_merge,
445         host_memory_backend_set_merge, &error_abort);
446     object_class_property_set_description(oc, "merge",
447         "Mark memory as mergeable", &error_abort);
448     object_class_property_add_bool(oc, "dump",
449         host_memory_backend_get_dump,
450         host_memory_backend_set_dump, &error_abort);
451     object_class_property_set_description(oc, "dump",
452         "Set to 'off' to exclude from core dump", &error_abort);
453     object_class_property_add_bool(oc, "prealloc",
454         host_memory_backend_get_prealloc,
455         host_memory_backend_set_prealloc, &error_abort);
456     object_class_property_set_description(oc, "prealloc",
457         "Preallocate memory", &error_abort);
458     object_class_property_add(oc, "size", "int",
459         host_memory_backend_get_size,
460         host_memory_backend_set_size,
461         NULL, NULL, &error_abort);
462     object_class_property_set_description(oc, "size",
463         "Size of the memory region (ex: 500M)", &error_abort);
464     object_class_property_add(oc, "host-nodes", "int",
465         host_memory_backend_get_host_nodes,
466         host_memory_backend_set_host_nodes,
467         NULL, NULL, &error_abort);
468     object_class_property_set_description(oc, "host-nodes",
469         "Binds memory to the list of NUMA host nodes", &error_abort);
470     object_class_property_add_enum(oc, "policy", "HostMemPolicy",
471         &HostMemPolicy_lookup,
472         host_memory_backend_get_policy,
473         host_memory_backend_set_policy, &error_abort);
474     object_class_property_set_description(oc, "policy",
475         "Set the NUMA policy", &error_abort);
476     object_class_property_add_bool(oc, "share",
477         host_memory_backend_get_share, host_memory_backend_set_share,
478         &error_abort);
479     object_class_property_set_description(oc, "share",
480         "Mark the memory as private to QEMU or shared", &error_abort);
481     object_class_property_add_bool(oc, "x-use-canonical-path-for-ramblock-id",
482         host_memory_backend_get_use_canonical_path,
483         host_memory_backend_set_use_canonical_path, &error_abort);
484 }
485 
486 static const TypeInfo host_memory_backend_info = {
487     .name = TYPE_MEMORY_BACKEND,
488     .parent = TYPE_OBJECT,
489     .abstract = true,
490     .class_size = sizeof(HostMemoryBackendClass),
491     .class_init = host_memory_backend_class_init,
492     .instance_size = sizeof(HostMemoryBackend),
493     .instance_init = host_memory_backend_init,
494     .instance_post_init = host_memory_backend_post_init,
495     .interfaces = (InterfaceInfo[]) {
496         { TYPE_USER_CREATABLE },
497         { }
498     }
499 };
500 
501 static void register_types(void)
502 {
503     type_register_static(&host_memory_backend_info);
504 }
505 
506 type_init(register_types);
507