Merge remote-tracking branch 'remotes/stefanha/tags/block-pull-request' into staging
[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
26 static QemuMutex qemu_cpu_list_lock;
27 static QemuCond exclusive_cond;
28 static QemuCond exclusive_resume;
29 static QemuCond qemu_work_cond;
30
31 /* >= 1 if a thread is inside start_exclusive/end_exclusive. Written
32 * under qemu_cpu_list_lock, read with atomic operations.
33 */
34 static int pending_cpus;
35
36 void qemu_init_cpu_list(void)
37 {
38 /* This is needed because qemu_init_cpu_list is also called by the
39 * child process in a fork. */
40 pending_cpus = 0;
41
42 qemu_mutex_init(&qemu_cpu_list_lock);
43 qemu_cond_init(&exclusive_cond);
44 qemu_cond_init(&exclusive_resume);
45 qemu_cond_init(&qemu_work_cond);
46 }
47
48 void cpu_list_lock(void)
49 {
50 qemu_mutex_lock(&qemu_cpu_list_lock);
51 }
52
53 void cpu_list_unlock(void)
54 {
55 qemu_mutex_unlock(&qemu_cpu_list_lock);
56 }
57
58 static bool cpu_index_auto_assigned;
59
60 static int cpu_get_free_index(void)
61 {
62 CPUState *some_cpu;
63 int cpu_index = 0;
64
65 cpu_index_auto_assigned = true;
66 CPU_FOREACH(some_cpu) {
67 cpu_index++;
68 }
69 return cpu_index;
70 }
71
72 void cpu_list_add(CPUState *cpu)
73 {
74 qemu_mutex_lock(&qemu_cpu_list_lock);
75 if (cpu->cpu_index == UNASSIGNED_CPU_INDEX) {
76 cpu->cpu_index = cpu_get_free_index();
77 assert(cpu->cpu_index != UNASSIGNED_CPU_INDEX);
78 } else {
79 assert(!cpu_index_auto_assigned);
80 }
81 QTAILQ_INSERT_TAIL_RCU(&cpus, cpu, node);
82 qemu_mutex_unlock(&qemu_cpu_list_lock);
83 }
84
85 void cpu_list_remove(CPUState *cpu)
86 {
87 qemu_mutex_lock(&qemu_cpu_list_lock);
88 if (!QTAILQ_IN_USE(cpu, node)) {
89 /* there is nothing to undo since cpu_exec_init() hasn't been called */
90 qemu_mutex_unlock(&qemu_cpu_list_lock);
91 return;
92 }
93
94 assert(!(cpu_index_auto_assigned && cpu != QTAILQ_LAST(&cpus)));
95
96 QTAILQ_REMOVE_RCU(&cpus, cpu, node);
97 cpu->cpu_index = UNASSIGNED_CPU_INDEX;
98 qemu_mutex_unlock(&qemu_cpu_list_lock);
99 }
100
101 struct qemu_work_item {
102 struct qemu_work_item *next;
103 run_on_cpu_func func;
104 run_on_cpu_data data;
105 bool free, exclusive, done;
106 };
107
108 static void queue_work_on_cpu(CPUState *cpu, struct qemu_work_item *wi)
109 {
110 qemu_mutex_lock(&cpu->work_mutex);
111 if (cpu->queued_work_first == NULL) {
112 cpu->queued_work_first = wi;
113 } else {
114 cpu->queued_work_last->next = wi;
115 }
116 cpu->queued_work_last = wi;
117 wi->next = NULL;
118 wi->done = false;
119 qemu_mutex_unlock(&cpu->work_mutex);
120
121 qemu_cpu_kick(cpu);
122 }
123
124 void do_run_on_cpu(CPUState *cpu, run_on_cpu_func func, run_on_cpu_data data,
125 QemuMutex *mutex)
126 {
127 struct qemu_work_item wi;
128
129 if (qemu_cpu_is_self(cpu)) {
130 func(cpu, data);
131 return;
132 }
133
134 wi.func = func;
135 wi.data = data;
136 wi.done = false;
137 wi.free = false;
138 wi.exclusive = false;
139
140 queue_work_on_cpu(cpu, &wi);
141 while (!atomic_mb_read(&wi.done)) {
142 CPUState *self_cpu = current_cpu;
143
144 qemu_cond_wait(&qemu_work_cond, mutex);
145 current_cpu = self_cpu;
146 }
147 }
148
149 void async_run_on_cpu(CPUState *cpu, run_on_cpu_func func, run_on_cpu_data data)
150 {
151 struct qemu_work_item *wi;
152
153 wi = g_malloc0(sizeof(struct qemu_work_item));
154 wi->func = func;
155 wi->data = data;
156 wi->free = true;
157
158 queue_work_on_cpu(cpu, wi);
159 }
160
161 /* Wait for pending exclusive operations to complete. The CPU list lock
162 must be held. */
163 static inline void exclusive_idle(void)
164 {
165 while (pending_cpus) {
166 qemu_cond_wait(&exclusive_resume, &qemu_cpu_list_lock);
167 }
168 }
169
170 /* Start an exclusive operation.
171 Must only be called from outside cpu_exec. */
172 void start_exclusive(void)
173 {
174 CPUState *other_cpu;
175 int running_cpus;
176
177 qemu_mutex_lock(&qemu_cpu_list_lock);
178 exclusive_idle();
179
180 /* Make all other cpus stop executing. */
181 atomic_set(&pending_cpus, 1);
182
183 /* Write pending_cpus before reading other_cpu->running. */
184 smp_mb();
185 running_cpus = 0;
186 CPU_FOREACH(other_cpu) {
187 if (atomic_read(&other_cpu->running)) {
188 other_cpu->has_waiter = true;
189 running_cpus++;
190 qemu_cpu_kick(other_cpu);
191 }
192 }
193
194 atomic_set(&pending_cpus, running_cpus + 1);
195 while (pending_cpus > 1) {
196 qemu_cond_wait(&exclusive_cond, &qemu_cpu_list_lock);
197 }
198
199 /* Can release mutex, no one will enter another exclusive
200 * section until end_exclusive resets pending_cpus to 0.
201 */
202 qemu_mutex_unlock(&qemu_cpu_list_lock);
203
204 current_cpu->in_exclusive_context = true;
205 }
206
207 /* Finish an exclusive operation. */
208 void end_exclusive(void)
209 {
210 current_cpu->in_exclusive_context = false;
211
212 qemu_mutex_lock(&qemu_cpu_list_lock);
213 atomic_set(&pending_cpus, 0);
214 qemu_cond_broadcast(&exclusive_resume);
215 qemu_mutex_unlock(&qemu_cpu_list_lock);
216 }
217
218 /* Wait for exclusive ops to finish, and begin cpu execution. */
219 void cpu_exec_start(CPUState *cpu)
220 {
221 atomic_set(&cpu->running, true);
222
223 /* Write cpu->running before reading pending_cpus. */
224 smp_mb();
225
226 /* 1. start_exclusive saw cpu->running == true and pending_cpus >= 1.
227 * After taking the lock we'll see cpu->has_waiter == true and run---not
228 * for long because start_exclusive kicked us. cpu_exec_end will
229 * decrement pending_cpus and signal the waiter.
230 *
231 * 2. start_exclusive saw cpu->running == false but pending_cpus >= 1.
232 * This includes the case when an exclusive item is running now.
233 * Then we'll see cpu->has_waiter == false and wait for the item to
234 * complete.
235 *
236 * 3. pending_cpus == 0. Then start_exclusive is definitely going to
237 * see cpu->running == true, and it will kick the CPU.
238 */
239 if (unlikely(atomic_read(&pending_cpus))) {
240 qemu_mutex_lock(&qemu_cpu_list_lock);
241 if (!cpu->has_waiter) {
242 /* Not counted in pending_cpus, let the exclusive item
243 * run. Since we have the lock, just set cpu->running to true
244 * while holding it; no need to check pending_cpus again.
245 */
246 atomic_set(&cpu->running, false);
247 exclusive_idle();
248 /* Now pending_cpus is zero. */
249 atomic_set(&cpu->running, true);
250 } else {
251 /* Counted in pending_cpus, go ahead and release the
252 * waiter at cpu_exec_end.
253 */
254 }
255 qemu_mutex_unlock(&qemu_cpu_list_lock);
256 }
257 }
258
259 /* Mark cpu as not executing, and release pending exclusive ops. */
260 void cpu_exec_end(CPUState *cpu)
261 {
262 atomic_set(&cpu->running, false);
263
264 /* Write cpu->running before reading pending_cpus. */
265 smp_mb();
266
267 /* 1. start_exclusive saw cpu->running == true. Then it will increment
268 * pending_cpus and wait for exclusive_cond. After taking the lock
269 * we'll see cpu->has_waiter == true.
270 *
271 * 2. start_exclusive saw cpu->running == false but here pending_cpus >= 1.
272 * This includes the case when an exclusive item started after setting
273 * cpu->running to false and before we read pending_cpus. Then we'll see
274 * cpu->has_waiter == false and not touch pending_cpus. The next call to
275 * cpu_exec_start will run exclusive_idle if still necessary, thus waiting
276 * for the item to complete.
277 *
278 * 3. pending_cpus == 0. Then start_exclusive is definitely going to
279 * see cpu->running == false, and it can ignore this CPU until the
280 * next cpu_exec_start.
281 */
282 if (unlikely(atomic_read(&pending_cpus))) {
283 qemu_mutex_lock(&qemu_cpu_list_lock);
284 if (cpu->has_waiter) {
285 cpu->has_waiter = false;
286 atomic_set(&pending_cpus, pending_cpus - 1);
287 if (pending_cpus == 1) {
288 qemu_cond_signal(&exclusive_cond);
289 }
290 }
291 qemu_mutex_unlock(&qemu_cpu_list_lock);
292 }
293 }
294
295 void async_safe_run_on_cpu(CPUState *cpu, run_on_cpu_func func,
296 run_on_cpu_data data)
297 {
298 struct qemu_work_item *wi;
299
300 wi = g_malloc0(sizeof(struct qemu_work_item));
301 wi->func = func;
302 wi->data = data;
303 wi->free = true;
304 wi->exclusive = true;
305
306 queue_work_on_cpu(cpu, wi);
307 }
308
309 void process_queued_cpu_work(CPUState *cpu)
310 {
311 struct qemu_work_item *wi;
312
313 if (cpu->queued_work_first == NULL) {
314 return;
315 }
316
317 qemu_mutex_lock(&cpu->work_mutex);
318 while (cpu->queued_work_first != NULL) {
319 wi = cpu->queued_work_first;
320 cpu->queued_work_first = wi->next;
321 if (!cpu->queued_work_first) {
322 cpu->queued_work_last = NULL;
323 }
324 qemu_mutex_unlock(&cpu->work_mutex);
325 if (wi->exclusive) {
326 /* Running work items outside the BQL avoids the following deadlock:
327 * 1) start_exclusive() is called with the BQL taken while another
328 * CPU is running; 2) cpu_exec in the other CPU tries to takes the
329 * BQL, so it goes to sleep; start_exclusive() is sleeping too, so
330 * neither CPU can proceed.
331 */
332 qemu_mutex_unlock_iothread();
333 start_exclusive();
334 wi->func(cpu, wi->data);
335 end_exclusive();
336 qemu_mutex_lock_iothread();
337 } else {
338 wi->func(cpu, wi->data);
339 }
340 qemu_mutex_lock(&cpu->work_mutex);
341 if (wi->free) {
342 g_free(wi);
343 } else {
344 atomic_mb_set(&wi->done, true);
345 }
346 }
347 qemu_mutex_unlock(&cpu->work_mutex);
348 qemu_cond_broadcast(&qemu_work_cond);
349 }