xref: /openbmc/qemu/backends/hostmem.c (revision 62a35aaa) 
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(&local_err, "cannot change property %s of %s ",
62                    name, object_get_typename(obj));
63         goto out;
64     }
65 
66     if (!visit_type_size(v, name, &value, &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;
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 
226     if (!host_memory_backend_mr_inited(backend)) {
227         backend->prealloc = value;
228         return;
229     }
230 
231     if (value && !backend->prealloc) {
232         int fd = memory_region_get_fd(&backend->mr);
233         void *ptr = memory_region_get_ram_ptr(&backend->mr);
234         uint64_t sz = memory_region_size(&backend->mr);
235 
236         os_mem_prealloc(fd, ptr, sz, backend->prealloc_threads, &local_err);
237         if (local_err) {
238             error_propagate(errp, local_err);
239             return;
240         }
241         backend->prealloc = true;
242     }
243 }
244 
245 static void host_memory_backend_get_prealloc_threads(Object *obj, Visitor *v,
246     const char *name, void *opaque, Error **errp)
247 {
248     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
249     visit_type_uint32(v, name, &backend->prealloc_threads, errp);
250 }
251 
252 static void host_memory_backend_set_prealloc_threads(Object *obj, Visitor *v,
253     const char *name, void *opaque, Error **errp)
254 {
255     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
256     Error *local_err = NULL;
257     uint32_t value;
258 
259     if (!visit_type_uint32(v, name, &value, &local_err)) {
260         goto out;
261     }
262     if (value <= 0) {
263         error_setg(&local_err,
264                    "property '%s' of %s doesn't take value '%d'",
265                    name, object_get_typename(obj), value);
266         goto out;
267     }
268     backend->prealloc_threads = value;
269 out:
270     error_propagate(errp, local_err);
271 }
272 
273 static void host_memory_backend_init(Object *obj)
274 {
275     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
276     MachineState *machine = MACHINE(qdev_get_machine());
277 
278     /* TODO: convert access to globals to compat properties */
279     backend->merge = machine_mem_merge(machine);
280     backend->dump = machine_dump_guest_core(machine);
281     backend->prealloc_threads = 1;
282 }
283 
284 static void host_memory_backend_post_init(Object *obj)
285 {
286     object_apply_compat_props(obj);
287 }
288 
289 bool host_memory_backend_mr_inited(HostMemoryBackend *backend)
290 {
291     /*
292      * NOTE: We forbid zero-length memory backend, so here zero means
293      * "we haven't inited the backend memory region yet".
294      */
295     return memory_region_size(&backend->mr) != 0;
296 }
297 
298 MemoryRegion *host_memory_backend_get_memory(HostMemoryBackend *backend)
299 {
300     return host_memory_backend_mr_inited(backend) ? &backend->mr : NULL;
301 }
302 
303 void host_memory_backend_set_mapped(HostMemoryBackend *backend, bool mapped)
304 {
305     backend->is_mapped = mapped;
306 }
307 
308 bool host_memory_backend_is_mapped(HostMemoryBackend *backend)
309 {
310     return backend->is_mapped;
311 }
312 
313 #ifdef __linux__
314 size_t host_memory_backend_pagesize(HostMemoryBackend *memdev)
315 {
316     Object *obj = OBJECT(memdev);
317     char *path = object_property_get_str(obj, "mem-path", NULL);
318     size_t pagesize = qemu_mempath_getpagesize(path);
319 
320     g_free(path);
321     return pagesize;
322 }
323 #else
324 size_t host_memory_backend_pagesize(HostMemoryBackend *memdev)
325 {
326     return qemu_real_host_page_size;
327 }
328 #endif
329 
330 static void
331 host_memory_backend_memory_complete(UserCreatable *uc, Error **errp)
332 {
333     HostMemoryBackend *backend = MEMORY_BACKEND(uc);
334     HostMemoryBackendClass *bc = MEMORY_BACKEND_GET_CLASS(uc);
335     Error *local_err = NULL;
336     void *ptr;
337     uint64_t sz;
338 
339     if (bc->alloc) {
340         bc->alloc(backend, &local_err);
341         if (local_err) {
342             goto out;
343         }
344 
345         ptr = memory_region_get_ram_ptr(&backend->mr);
346         sz = memory_region_size(&backend->mr);
347 
348         if (backend->merge) {
349             qemu_madvise(ptr, sz, QEMU_MADV_MERGEABLE);
350         }
351         if (!backend->dump) {
352             qemu_madvise(ptr, sz, QEMU_MADV_DONTDUMP);
353         }
354 #ifdef CONFIG_NUMA
355         unsigned long lastbit = find_last_bit(backend->host_nodes, MAX_NODES);
356         /* lastbit == MAX_NODES means maxnode = 0 */
357         unsigned long maxnode = (lastbit + 1) % (MAX_NODES + 1);
358         /* ensure policy won't be ignored in case memory is preallocated
359          * before mbind(). note: MPOL_MF_STRICT is ignored on hugepages so
360          * this doesn't catch hugepage case. */
361         unsigned flags = MPOL_MF_STRICT | MPOL_MF_MOVE;
362 
363         /* check for invalid host-nodes and policies and give more verbose
364          * error messages than mbind(). */
365         if (maxnode && backend->policy == MPOL_DEFAULT) {
366             error_setg(errp, "host-nodes must be empty for policy default,"
367                        " or you should explicitly specify a policy other"
368                        " than default");
369             return;
370         } else if (maxnode == 0 && backend->policy != MPOL_DEFAULT) {
371             error_setg(errp, "host-nodes must be set for policy %s",
372                        HostMemPolicy_str(backend->policy));
373             return;
374         }
375 
376         /* We can have up to MAX_NODES nodes, but we need to pass maxnode+1
377          * as argument to mbind() due to an old Linux bug (feature?) which
378          * cuts off the last specified node. This means backend->host_nodes
379          * must have MAX_NODES+1 bits available.
380          */
381         assert(sizeof(backend->host_nodes) >=
382                BITS_TO_LONGS(MAX_NODES + 1) * sizeof(unsigned long));
383         assert(maxnode <= MAX_NODES);
384 
385         if (maxnode &&
386             mbind(ptr, sz, backend->policy, backend->host_nodes, maxnode + 1,
387                   flags)) {
388             if (backend->policy != MPOL_DEFAULT || errno != ENOSYS) {
389                 error_setg_errno(errp, errno,
390                                  "cannot bind memory to host NUMA nodes");
391                 return;
392             }
393         }
394 #endif
395         /* Preallocate memory after the NUMA policy has been instantiated.
396          * This is necessary to guarantee memory is allocated with
397          * specified NUMA policy in place.
398          */
399         if (backend->prealloc) {
400             os_mem_prealloc(memory_region_get_fd(&backend->mr), ptr, sz,
401                             backend->prealloc_threads, &local_err);
402             if (local_err) {
403                 goto out;
404             }
405         }
406     }
407 out:
408     error_propagate(errp, local_err);
409 }
410 
411 static bool
412 host_memory_backend_can_be_deleted(UserCreatable *uc)
413 {
414     if (host_memory_backend_is_mapped(MEMORY_BACKEND(uc))) {
415         return false;
416     } else {
417         return true;
418     }
419 }
420 
421 static bool host_memory_backend_get_share(Object *o, Error **errp)
422 {
423     HostMemoryBackend *backend = MEMORY_BACKEND(o);
424 
425     return backend->share;
426 }
427 
428 static void host_memory_backend_set_share(Object *o, bool value, Error **errp)
429 {
430     HostMemoryBackend *backend = MEMORY_BACKEND(o);
431 
432     if (host_memory_backend_mr_inited(backend)) {
433         error_setg(errp, "cannot change property value");
434         return;
435     }
436     backend->share = value;
437 }
438 
439 static bool
440 host_memory_backend_get_use_canonical_path(Object *obj, Error **errp)
441 {
442     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
443 
444     return backend->use_canonical_path;
445 }
446 
447 static void
448 host_memory_backend_set_use_canonical_path(Object *obj, bool value,
449                                            Error **errp)
450 {
451     HostMemoryBackend *backend = MEMORY_BACKEND(obj);
452 
453     backend->use_canonical_path = value;
454 }
455 
456 static void
457 host_memory_backend_class_init(ObjectClass *oc, void *data)
458 {
459     UserCreatableClass *ucc = USER_CREATABLE_CLASS(oc);
460 
461     ucc->complete = host_memory_backend_memory_complete;
462     ucc->can_be_deleted = host_memory_backend_can_be_deleted;
463 
464     object_class_property_add_bool(oc, "merge",
465         host_memory_backend_get_merge,
466         host_memory_backend_set_merge);
467     object_class_property_set_description(oc, "merge",
468         "Mark memory as mergeable");
469     object_class_property_add_bool(oc, "dump",
470         host_memory_backend_get_dump,
471         host_memory_backend_set_dump);
472     object_class_property_set_description(oc, "dump",
473         "Set to 'off' to exclude from core dump");
474     object_class_property_add_bool(oc, "prealloc",
475         host_memory_backend_get_prealloc,
476         host_memory_backend_set_prealloc);
477     object_class_property_set_description(oc, "prealloc",
478         "Preallocate memory");
479     object_class_property_add(oc, "prealloc-threads", "int",
480         host_memory_backend_get_prealloc_threads,
481         host_memory_backend_set_prealloc_threads,
482         NULL, NULL);
483     object_class_property_set_description(oc, "prealloc-threads",
484         "Number of CPU threads to use for prealloc");
485     object_class_property_add(oc, "size", "int",
486         host_memory_backend_get_size,
487         host_memory_backend_set_size,
488         NULL, NULL);
489     object_class_property_set_description(oc, "size",
490         "Size of the memory region (ex: 500M)");
491     object_class_property_add(oc, "host-nodes", "int",
492         host_memory_backend_get_host_nodes,
493         host_memory_backend_set_host_nodes,
494         NULL, NULL);
495     object_class_property_set_description(oc, "host-nodes",
496         "Binds memory to the list of NUMA host nodes");
497     object_class_property_add_enum(oc, "policy", "HostMemPolicy",
498         &HostMemPolicy_lookup,
499         host_memory_backend_get_policy,
500         host_memory_backend_set_policy);
501     object_class_property_set_description(oc, "policy",
502         "Set the NUMA policy");
503     object_class_property_add_bool(oc, "share",
504         host_memory_backend_get_share, host_memory_backend_set_share);
505     object_class_property_set_description(oc, "share",
506         "Mark the memory as private to QEMU or shared");
507     object_class_property_add_bool(oc, "x-use-canonical-path-for-ramblock-id",
508         host_memory_backend_get_use_canonical_path,
509         host_memory_backend_set_use_canonical_path);
510 }
511 
512 static const TypeInfo host_memory_backend_info = {
513     .name = TYPE_MEMORY_BACKEND,
514     .parent = TYPE_OBJECT,
515     .abstract = true,
516     .class_size = sizeof(HostMemoryBackendClass),
517     .class_init = host_memory_backend_class_init,
518     .instance_size = sizeof(HostMemoryBackend),
519     .instance_init = host_memory_backend_init,
520     .instance_post_init = host_memory_backend_post_init,
521     .interfaces = (InterfaceInfo[]) {
522         { TYPE_USER_CREATABLE },
523         { }
524     }
525 };
526 
527 static void register_types(void)
528 {
529     type_register_static(&host_memory_backend_info);
530 }
531 
532 type_init(register_types);
533