Update version for v5.1.0-rc3 release
[qemu.git] / cpus-common.c
1 /*
2 * CPU thread main loop - common bits for user and system mode emulation
3 *
4 * Copyright (c) 2003-2005 Fabrice Bellard
5 *
6 * This library is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2 of the License, or (at your option) any later version.
10 *
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
15 *
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with this library; if not, see <http://www.gnu.org/licenses/>.
18 */
19
20 #include "qemu/osdep.h"
21 #include "qemu/main-loop.h"
22 #include "exec/cpu-common.h"
23 #include "hw/core/cpu.h"
24 #include "sysemu/cpus.h"
25 #include "qemu/lockable.h"
26
27 static QemuMutex qemu_cpu_list_lock;
28 static QemuCond exclusive_cond;
29 static QemuCond exclusive_resume;
30 static QemuCond qemu_work_cond;
31
32 /* >= 1 if a thread is inside start_exclusive/end_exclusive. Written
33 * under qemu_cpu_list_lock, read with atomic operations.
34 */
35 static int pending_cpus;
36
37 void qemu_init_cpu_list(void)
38 {
39 /* This is needed because qemu_init_cpu_list is also called by the
40 * child process in a fork. */
41 pending_cpus = 0;
42
43 qemu_mutex_init(&qemu_cpu_list_lock);
44 qemu_cond_init(&exclusive_cond);
45 qemu_cond_init(&exclusive_resume);
46 qemu_cond_init(&qemu_work_cond);
47 }
48
49 void cpu_list_lock(void)
50 {
51 qemu_mutex_lock(&qemu_cpu_list_lock);
52 }
53
54 void cpu_list_unlock(void)
55 {
56 qemu_mutex_unlock(&qemu_cpu_list_lock);
57 }
58
59 static bool cpu_index_auto_assigned;
60
61 static int cpu_get_free_index(void)
62 {
63 CPUState *some_cpu;
64 int max_cpu_index = 0;
65
66 cpu_index_auto_assigned = true;
67 CPU_FOREACH(some_cpu) {
68 if (some_cpu->cpu_index >= max_cpu_index) {
69 max_cpu_index = some_cpu->cpu_index + 1;
70 }
71 }
72 return max_cpu_index;
73 }
74
75 CPUTailQ cpus = QTAILQ_HEAD_INITIALIZER(cpus);
76
77 void cpu_list_add(CPUState *cpu)
78 {
79 QEMU_LOCK_GUARD(&qemu_cpu_list_lock);
80 if (cpu->cpu_index == UNASSIGNED_CPU_INDEX) {
81 cpu->cpu_index = cpu_get_free_index();
82 assert(cpu->cpu_index != UNASSIGNED_CPU_INDEX);
83 } else {
84 assert(!cpu_index_auto_assigned);
85 }
86 QTAILQ_INSERT_TAIL_RCU(&cpus, cpu, node);
87 }
88
89 void cpu_list_remove(CPUState *cpu)
90 {
91 QEMU_LOCK_GUARD(&qemu_cpu_list_lock);
92 if (!QTAILQ_IN_USE(cpu, node)) {
93 /* there is nothing to undo since cpu_exec_init() hasn't been called */
94 return;
95 }
96
97 QTAILQ_REMOVE_RCU(&cpus, cpu, node);
98 cpu->cpu_index = UNASSIGNED_CPU_INDEX;
99 }
100
101 CPUState *qemu_get_cpu(int index)
102 {
103 CPUState *cpu;
104
105 CPU_FOREACH(cpu) {
106 if (cpu->cpu_index == index) {
107 return cpu;
108 }
109 }
110
111 return NULL;
112 }
113
114 /* current CPU in the current thread. It is only valid inside cpu_exec() */
115 __thread CPUState *current_cpu;
116
117 struct qemu_work_item {
118 QSIMPLEQ_ENTRY(qemu_work_item) node;
119 run_on_cpu_func func;
120 run_on_cpu_data data;
121 bool free, exclusive, done;
122 };
123
124 static void queue_work_on_cpu(CPUState *cpu, struct qemu_work_item *wi)
125 {
126 qemu_mutex_lock(&cpu->work_mutex);
127 QSIMPLEQ_INSERT_TAIL(&cpu->work_list, wi, node);
128 wi->done = false;
129 qemu_mutex_unlock(&cpu->work_mutex);
130
131 qemu_cpu_kick(cpu);
132 }
133
134 void do_run_on_cpu(CPUState *cpu, run_on_cpu_func func, run_on_cpu_data data,
135 QemuMutex *mutex)
136 {
137 struct qemu_work_item wi;
138
139 if (qemu_cpu_is_self(cpu)) {
140 func(cpu, data);
141 return;
142 }
143
144 wi.func = func;
145 wi.data = data;
146 wi.done = false;
147 wi.free = false;
148 wi.exclusive = false;
149
150 queue_work_on_cpu(cpu, &wi);
151 while (!atomic_mb_read(&wi.done)) {
152 CPUState *self_cpu = current_cpu;
153
154 qemu_cond_wait(&qemu_work_cond, mutex);
155 current_cpu = self_cpu;
156 }
157 }
158
159 void async_run_on_cpu(CPUState *cpu, run_on_cpu_func func, run_on_cpu_data data)
160 {
161 struct qemu_work_item *wi;
162
163 wi = g_malloc0(sizeof(struct qemu_work_item));
164 wi->func = func;
165 wi->data = data;
166 wi->free = true;
167
168 queue_work_on_cpu(cpu, wi);
169 }
170
171 /* Wait for pending exclusive operations to complete. The CPU list lock
172 must be held. */
173 static inline void exclusive_idle(void)
174 {
175 while (pending_cpus) {
176 qemu_cond_wait(&exclusive_resume, &qemu_cpu_list_lock);
177 }
178 }
179
180 /* Start an exclusive operation.
181 Must only be called from outside cpu_exec. */
182 void start_exclusive(void)
183 {
184 CPUState *other_cpu;
185 int running_cpus;
186
187 qemu_mutex_lock(&qemu_cpu_list_lock);
188 exclusive_idle();
189
190 /* Make all other cpus stop executing. */
191 atomic_set(&pending_cpus, 1);
192
193 /* Write pending_cpus before reading other_cpu->running. */
194 smp_mb();
195 running_cpus = 0;
196 CPU_FOREACH(other_cpu) {
197 if (atomic_read(&other_cpu->running)) {
198 other_cpu->has_waiter = true;
199 running_cpus++;
200 qemu_cpu_kick(other_cpu);
201 }
202 }
203
204 atomic_set(&pending_cpus, running_cpus + 1);
205 while (pending_cpus > 1) {
206 qemu_cond_wait(&exclusive_cond, &qemu_cpu_list_lock);
207 }
208
209 /* Can release mutex, no one will enter another exclusive
210 * section until end_exclusive resets pending_cpus to 0.
211 */
212 qemu_mutex_unlock(&qemu_cpu_list_lock);
213
214 current_cpu->in_exclusive_context = true;
215 }
216
217 /* Finish an exclusive operation. */
218 void end_exclusive(void)
219 {
220 current_cpu->in_exclusive_context = false;
221
222 qemu_mutex_lock(&qemu_cpu_list_lock);
223 atomic_set(&pending_cpus, 0);
224 qemu_cond_broadcast(&exclusive_resume);
225 qemu_mutex_unlock(&qemu_cpu_list_lock);
226 }
227
228 /* Wait for exclusive ops to finish, and begin cpu execution. */
229 void cpu_exec_start(CPUState *cpu)
230 {
231 atomic_set(&cpu->running, true);
232
233 /* Write cpu->running before reading pending_cpus. */
234 smp_mb();
235
236 /* 1. start_exclusive saw cpu->running == true and pending_cpus >= 1.
237 * After taking the lock we'll see cpu->has_waiter == true and run---not
238 * for long because start_exclusive kicked us. cpu_exec_end will
239 * decrement pending_cpus and signal the waiter.
240 *
241 * 2. start_exclusive saw cpu->running == false but pending_cpus >= 1.
242 * This includes the case when an exclusive item is running now.
243 * Then we'll see cpu->has_waiter == false and wait for the item to
244 * complete.
245 *
246 * 3. pending_cpus == 0. Then start_exclusive is definitely going to
247 * see cpu->running == true, and it will kick the CPU.
248 */
249 if (unlikely(atomic_read(&pending_cpus))) {
250 QEMU_LOCK_GUARD(&qemu_cpu_list_lock);
251 if (!cpu->has_waiter) {
252 /* Not counted in pending_cpus, let the exclusive item
253 * run. Since we have the lock, just set cpu->running to true
254 * while holding it; no need to check pending_cpus again.
255 */
256 atomic_set(&cpu->running, false);
257 exclusive_idle();
258 /* Now pending_cpus is zero. */
259 atomic_set(&cpu->running, true);
260 } else {
261 /* Counted in pending_cpus, go ahead and release the
262 * waiter at cpu_exec_end.
263 */
264 }
265 }
266 }
267
268 /* Mark cpu as not executing, and release pending exclusive ops. */
269 void cpu_exec_end(CPUState *cpu)
270 {
271 atomic_set(&cpu->running, false);
272
273 /* Write cpu->running before reading pending_cpus. */
274 smp_mb();
275
276 /* 1. start_exclusive saw cpu->running == true. Then it will increment
277 * pending_cpus and wait for exclusive_cond. After taking the lock
278 * we'll see cpu->has_waiter == true.
279 *
280 * 2. start_exclusive saw cpu->running == false but here pending_cpus >= 1.
281 * This includes the case when an exclusive item started after setting
282 * cpu->running to false and before we read pending_cpus. Then we'll see
283 * cpu->has_waiter == false and not touch pending_cpus. The next call to
284 * cpu_exec_start will run exclusive_idle if still necessary, thus waiting
285 * for the item to complete.
286 *
287 * 3. pending_cpus == 0. Then start_exclusive is definitely going to
288 * see cpu->running == false, and it can ignore this CPU until the
289 * next cpu_exec_start.
290 */
291 if (unlikely(atomic_read(&pending_cpus))) {
292 QEMU_LOCK_GUARD(&qemu_cpu_list_lock);
293 if (cpu->has_waiter) {
294 cpu->has_waiter = false;
295 atomic_set(&pending_cpus, pending_cpus - 1);
296 if (pending_cpus == 1) {
297 qemu_cond_signal(&exclusive_cond);
298 }
299 }
300 }
301 }
302
303 void async_safe_run_on_cpu(CPUState *cpu, run_on_cpu_func func,
304 run_on_cpu_data data)
305 {
306 struct qemu_work_item *wi;
307
308 wi = g_malloc0(sizeof(struct qemu_work_item));
309 wi->func = func;
310 wi->data = data;
311 wi->free = true;
312 wi->exclusive = true;
313
314 queue_work_on_cpu(cpu, wi);
315 }
316
317 void process_queued_cpu_work(CPUState *cpu)
318 {
319 struct qemu_work_item *wi;
320
321 qemu_mutex_lock(&cpu->work_mutex);
322 if (QSIMPLEQ_EMPTY(&cpu->work_list)) {
323 qemu_mutex_unlock(&cpu->work_mutex);
324 return;
325 }
326 while (!QSIMPLEQ_EMPTY(&cpu->work_list)) {
327 wi = QSIMPLEQ_FIRST(&cpu->work_list);
328 QSIMPLEQ_REMOVE_HEAD(&cpu->work_list, node);
329 qemu_mutex_unlock(&cpu->work_mutex);
330 if (wi->exclusive) {
331 /* Running work items outside the BQL avoids the following deadlock:
332 * 1) start_exclusive() is called with the BQL taken while another
333 * CPU is running; 2) cpu_exec in the other CPU tries to takes the
334 * BQL, so it goes to sleep; start_exclusive() is sleeping too, so
335 * neither CPU can proceed.
336 */
337 qemu_mutex_unlock_iothread();
338 start_exclusive();
339 wi->func(cpu, wi->data);
340 end_exclusive();
341 qemu_mutex_lock_iothread();
342 } else {
343 wi->func(cpu, wi->data);
344 }
345 qemu_mutex_lock(&cpu->work_mutex);
346 if (wi->free) {
347 g_free(wi);
348 } else {
349 atomic_mb_set(&wi->done, true);
350 }
351 }
352 qemu_mutex_unlock(&cpu->work_mutex);
353 qemu_cond_broadcast(&qemu_work_cond);
354 }