| 1 | /* |
| 2 | * QEMU Thread Context |
| 3 | * |
| 4 | * Copyright Red Hat Inc., 2022 |
| 5 | * |
| 6 | * Authors: |
| 7 | * David Hildenbrand <david@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 "qemu/thread-context.h" |
| 15 | #include "qapi/error.h" |
| 16 | #include "qapi/qapi-builtin-visit.h" |
| 17 | #include "qapi/visitor.h" |
| 18 | #include "qemu/config-file.h" |
| 19 | #include "qapi/qapi-builtin-visit.h" |
| 20 | #include "qom/object_interfaces.h" |
| 21 | #include "qemu/module.h" |
| 22 | #include "qemu/bitmap.h" |
| 23 | |
| 24 | #ifdef CONFIG_NUMA |
| 25 | #include <numa.h> |
| 26 | #endif |
| 27 | |
| 28 | enum { |
| 29 | TC_CMD_NONE = 0, |
| 30 | TC_CMD_STOP, |
| 31 | TC_CMD_NEW, |
| 32 | }; |
| 33 | |
| 34 | typedef struct ThreadContextCmdNew { |
| 35 | QemuThread *thread; |
| 36 | const char *name; |
| 37 | void *(*start_routine)(void *); |
| 38 | void *arg; |
| 39 | int mode; |
| 40 | } ThreadContextCmdNew; |
| 41 | |
| 42 | static void *thread_context_run(void *opaque) |
| 43 | { |
| 44 | ThreadContext *tc = opaque; |
| 45 | |
| 46 | tc->thread_id = qemu_get_thread_id(); |
| 47 | qemu_sem_post(&tc->sem); |
| 48 | |
| 49 | while (true) { |
| 50 | /* |
| 51 | * Threads inherit the CPU affinity of the creating thread. For this |
| 52 | * reason, we create new (especially short-lived) threads from our |
| 53 | * persistent context thread. |
| 54 | * |
| 55 | * Especially when QEMU is not allowed to set the affinity itself, |
| 56 | * management tools can simply set the affinity of the context thread |
| 57 | * after creating the context, to have new threads created via |
| 58 | * the context inherit the CPU affinity automatically. |
| 59 | */ |
| 60 | switch (tc->thread_cmd) { |
| 61 | case TC_CMD_NONE: |
| 62 | break; |
| 63 | case TC_CMD_STOP: |
| 64 | tc->thread_cmd = TC_CMD_NONE; |
| 65 | qemu_sem_post(&tc->sem); |
| 66 | return NULL; |
| 67 | case TC_CMD_NEW: { |
| 68 | ThreadContextCmdNew *cmd_new = tc->thread_cmd_data; |
| 69 | |
| 70 | qemu_thread_create(cmd_new->thread, cmd_new->name, |
| 71 | cmd_new->start_routine, cmd_new->arg, |
| 72 | cmd_new->mode); |
| 73 | tc->thread_cmd = TC_CMD_NONE; |
| 74 | tc->thread_cmd_data = NULL; |
| 75 | qemu_sem_post(&tc->sem); |
| 76 | break; |
| 77 | } |
| 78 | default: |
| 79 | g_assert_not_reached(); |
| 80 | } |
| 81 | qemu_sem_wait(&tc->sem_thread); |
| 82 | } |
| 83 | } |
| 84 | |
| 85 | static void thread_context_set_cpu_affinity(Object *obj, Visitor *v, |
| 86 | const char *name, void *opaque, |
| 87 | Error **errp) |
| 88 | { |
| 89 | ThreadContext *tc = THREAD_CONTEXT(obj); |
| 90 | uint16List *l, *host_cpus = NULL; |
| 91 | unsigned long *bitmap = NULL; |
| 92 | int nbits = 0, ret; |
| 93 | |
| 94 | if (tc->init_cpu_bitmap) { |
| 95 | error_setg(errp, "Mixing CPU and node affinity not supported"); |
| 96 | return; |
| 97 | } |
| 98 | |
| 99 | if (!visit_type_uint16List(v, name, &host_cpus, errp)) { |
| 100 | return; |
| 101 | } |
| 102 | |
| 103 | if (!host_cpus) { |
| 104 | error_setg(errp, "CPU list is empty"); |
| 105 | goto out; |
| 106 | } |
| 107 | |
| 108 | for (l = host_cpus; l; l = l->next) { |
| 109 | nbits = MAX(nbits, l->value + 1); |
| 110 | } |
| 111 | bitmap = bitmap_new(nbits); |
| 112 | for (l = host_cpus; l; l = l->next) { |
| 113 | set_bit(l->value, bitmap); |
| 114 | } |
| 115 | |
| 116 | if (tc->thread_id != -1) { |
| 117 | /* |
| 118 | * Note: we won't be adjusting the affinity of any thread that is still |
| 119 | * around, but only the affinity of the context thread. |
| 120 | */ |
| 121 | ret = qemu_thread_set_affinity(&tc->thread, bitmap, nbits); |
| 122 | if (ret) { |
| 123 | error_setg(errp, "Setting CPU affinity failed: %s", strerror(ret)); |
| 124 | } |
| 125 | } else { |
| 126 | tc->init_cpu_bitmap = bitmap; |
| 127 | bitmap = NULL; |
| 128 | tc->init_cpu_nbits = nbits; |
| 129 | } |
| 130 | out: |
| 131 | g_free(bitmap); |
| 132 | qapi_free_uint16List(host_cpus); |
| 133 | } |
| 134 | |
| 135 | static void thread_context_get_cpu_affinity(Object *obj, Visitor *v, |
| 136 | const char *name, void *opaque, |
| 137 | Error **errp) |
| 138 | { |
| 139 | unsigned long *bitmap, nbits, value; |
| 140 | ThreadContext *tc = THREAD_CONTEXT(obj); |
| 141 | uint16List *host_cpus = NULL; |
| 142 | uint16List **tail = &host_cpus; |
| 143 | int ret; |
| 144 | |
| 145 | if (tc->thread_id == -1) { |
| 146 | error_setg(errp, "Object not initialized yet"); |
| 147 | return; |
| 148 | } |
| 149 | |
| 150 | ret = qemu_thread_get_affinity(&tc->thread, &bitmap, &nbits); |
| 151 | if (ret) { |
| 152 | error_setg(errp, "Getting CPU affinity failed: %s", strerror(ret)); |
| 153 | return; |
| 154 | } |
| 155 | |
| 156 | value = find_first_bit(bitmap, nbits); |
| 157 | while (value < nbits) { |
| 158 | QAPI_LIST_APPEND(tail, value); |
| 159 | |
| 160 | value = find_next_bit(bitmap, nbits, value + 1); |
| 161 | } |
| 162 | g_free(bitmap); |
| 163 | |
| 164 | visit_type_uint16List(v, name, &host_cpus, errp); |
| 165 | qapi_free_uint16List(host_cpus); |
| 166 | } |
| 167 | |
| 168 | static void thread_context_set_node_affinity(Object *obj, Visitor *v, |
| 169 | const char *name, void *opaque, |
| 170 | Error **errp) |
| 171 | { |
| 172 | #ifdef CONFIG_NUMA |
| 173 | const int nbits = numa_num_possible_cpus(); |
| 174 | ThreadContext *tc = THREAD_CONTEXT(obj); |
| 175 | uint16List *l, *host_nodes = NULL; |
| 176 | unsigned long *bitmap = NULL; |
| 177 | struct bitmask *tmp_cpus; |
| 178 | int ret, i; |
| 179 | |
| 180 | if (tc->init_cpu_bitmap) { |
| 181 | error_setg(errp, "Mixing CPU and node affinity not supported"); |
| 182 | return; |
| 183 | } |
| 184 | |
| 185 | if (!visit_type_uint16List(v, name, &host_nodes, errp)) { |
| 186 | return; |
| 187 | } |
| 188 | |
| 189 | if (!host_nodes) { |
| 190 | error_setg(errp, "Node list is empty"); |
| 191 | goto out; |
| 192 | } |
| 193 | |
| 194 | bitmap = bitmap_new(nbits); |
| 195 | tmp_cpus = numa_allocate_cpumask(); |
| 196 | for (l = host_nodes; l; l = l->next) { |
| 197 | numa_bitmask_clearall(tmp_cpus); |
| 198 | ret = numa_node_to_cpus(l->value, tmp_cpus); |
| 199 | if (ret) { |
| 200 | /* We ignore any errors, such as impossible nodes. */ |
| 201 | continue; |
| 202 | } |
| 203 | for (i = 0; i < nbits; i++) { |
| 204 | if (numa_bitmask_isbitset(tmp_cpus, i)) { |
| 205 | set_bit(i, bitmap); |
| 206 | } |
| 207 | } |
| 208 | } |
| 209 | numa_free_cpumask(tmp_cpus); |
| 210 | |
| 211 | if (bitmap_empty(bitmap, nbits)) { |
| 212 | error_setg(errp, "The nodes select no CPUs"); |
| 213 | goto out; |
| 214 | } |
| 215 | |
| 216 | if (tc->thread_id != -1) { |
| 217 | /* |
| 218 | * Note: we won't be adjusting the affinity of any thread that is still |
| 219 | * around for now, but only the affinity of the context thread. |
| 220 | */ |
| 221 | ret = qemu_thread_set_affinity(&tc->thread, bitmap, nbits); |
| 222 | if (ret) { |
| 223 | error_setg(errp, "Setting CPU affinity failed: %s", strerror(ret)); |
| 224 | } |
| 225 | } else { |
| 226 | tc->init_cpu_bitmap = bitmap; |
| 227 | bitmap = NULL; |
| 228 | tc->init_cpu_nbits = nbits; |
| 229 | } |
| 230 | out: |
| 231 | g_free(bitmap); |
| 232 | qapi_free_uint16List(host_nodes); |
| 233 | #else |
| 234 | error_setg(errp, "NUMA node affinity is not supported by this QEMU"); |
| 235 | #endif |
| 236 | } |
| 237 | |
| 238 | static void thread_context_get_thread_id(Object *obj, Visitor *v, |
| 239 | const char *name, void *opaque, |
| 240 | Error **errp) |
| 241 | { |
| 242 | ThreadContext *tc = THREAD_CONTEXT(obj); |
| 243 | uint64_t value = tc->thread_id; |
| 244 | |
| 245 | visit_type_uint64(v, name, &value, errp); |
| 246 | } |
| 247 | |
| 248 | static void thread_context_instance_complete(UserCreatable *uc, Error **errp) |
| 249 | { |
| 250 | ThreadContext *tc = THREAD_CONTEXT(uc); |
| 251 | char *thread_name; |
| 252 | int ret; |
| 253 | |
| 254 | thread_name = g_strdup_printf("TC %s", |
| 255 | object_get_canonical_path_component(OBJECT(uc))); |
| 256 | qemu_thread_create(&tc->thread, thread_name, thread_context_run, tc, |
| 257 | QEMU_THREAD_JOINABLE); |
| 258 | g_free(thread_name); |
| 259 | |
| 260 | /* Wait until initialization of the thread is done. */ |
| 261 | while (tc->thread_id == -1) { |
| 262 | qemu_sem_wait(&tc->sem); |
| 263 | } |
| 264 | |
| 265 | if (tc->init_cpu_bitmap) { |
| 266 | ret = qemu_thread_set_affinity(&tc->thread, tc->init_cpu_bitmap, |
| 267 | tc->init_cpu_nbits); |
| 268 | if (ret) { |
| 269 | error_setg(errp, "Setting CPU affinity failed: %s", strerror(ret)); |
| 270 | } |
| 271 | g_free(tc->init_cpu_bitmap); |
| 272 | tc->init_cpu_bitmap = NULL; |
| 273 | } |
| 274 | } |
| 275 | |
| 276 | static void thread_context_class_init(ObjectClass *oc, const void *data) |
| 277 | { |
| 278 | UserCreatableClass *ucc = USER_CREATABLE_CLASS(oc); |
| 279 | |
| 280 | ucc->complete = thread_context_instance_complete; |
| 281 | object_class_property_add(oc, "thread-id", "int", |
| 282 | thread_context_get_thread_id, NULL, NULL, |
| 283 | NULL); |
| 284 | object_class_property_add(oc, "cpu-affinity", "int", |
| 285 | thread_context_get_cpu_affinity, |
| 286 | thread_context_set_cpu_affinity, NULL, NULL); |
| 287 | object_class_property_add(oc, "node-affinity", "int", NULL, |
| 288 | thread_context_set_node_affinity, NULL, NULL); |
| 289 | } |
| 290 | |
| 291 | static void thread_context_instance_init(Object *obj) |
| 292 | { |
| 293 | ThreadContext *tc = THREAD_CONTEXT(obj); |
| 294 | |
| 295 | tc->thread_id = -1; |
| 296 | qemu_sem_init(&tc->sem, 0); |
| 297 | qemu_sem_init(&tc->sem_thread, 0); |
| 298 | qemu_mutex_init(&tc->mutex); |
| 299 | } |
| 300 | |
| 301 | static void thread_context_instance_finalize(Object *obj) |
| 302 | { |
| 303 | ThreadContext *tc = THREAD_CONTEXT(obj); |
| 304 | |
| 305 | if (tc->thread_id != -1) { |
| 306 | tc->thread_cmd = TC_CMD_STOP; |
| 307 | qemu_sem_post(&tc->sem_thread); |
| 308 | qemu_thread_join(&tc->thread); |
| 309 | } |
| 310 | qemu_sem_destroy(&tc->sem); |
| 311 | qemu_sem_destroy(&tc->sem_thread); |
| 312 | qemu_mutex_destroy(&tc->mutex); |
| 313 | } |
| 314 | |
| 315 | static const TypeInfo thread_context_info = { |
| 316 | .name = TYPE_THREAD_CONTEXT, |
| 317 | .parent = TYPE_OBJECT, |
| 318 | .class_init = thread_context_class_init, |
| 319 | .instance_size = sizeof(ThreadContext), |
| 320 | .instance_init = thread_context_instance_init, |
| 321 | .instance_finalize = thread_context_instance_finalize, |
| 322 | .interfaces = (const InterfaceInfo[]) { |
| 323 | { TYPE_USER_CREATABLE }, |
| 324 | { } |
| 325 | } |
| 326 | }; |
| 327 | |
| 328 | static void thread_context_register_types(void) |
| 329 | { |
| 330 | type_register_static(&thread_context_info); |
| 331 | } |
| 332 | type_init(thread_context_register_types) |
| 333 | |
| 334 | void thread_context_create_thread(ThreadContext *tc, QemuThread *thread, |
| 335 | const char *name, |
| 336 | void *(*start_routine)(void *), void *arg, |
| 337 | int mode) |
| 338 | { |
| 339 | ThreadContextCmdNew data = { |
| 340 | .thread = thread, |
| 341 | .name = name, |
| 342 | .start_routine = start_routine, |
| 343 | .arg = arg, |
| 344 | .mode = mode, |
| 345 | }; |
| 346 | |
| 347 | qemu_mutex_lock(&tc->mutex); |
| 348 | tc->thread_cmd = TC_CMD_NEW; |
| 349 | tc->thread_cmd_data = &data; |
| 350 | qemu_sem_post(&tc->sem_thread); |
| 351 | |
| 352 | while (tc->thread_cmd != TC_CMD_NONE) { |
| 353 | qemu_sem_wait(&tc->sem); |
| 354 | } |
| 355 | qemu_mutex_unlock(&tc->mutex); |
| 356 | } |