quorum: Avoid bdrv_aio_writev() for rewrites
[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 "qom/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 static void finish_safe_work(CPUState *cpu)
73 {
74 cpu_exec_start(cpu);
75 cpu_exec_end(cpu);
76 }
77
78 void cpu_list_add(CPUState *cpu)
79 {
80 qemu_mutex_lock(&qemu_cpu_list_lock);
81 if (cpu->cpu_index == UNASSIGNED_CPU_INDEX) {
82 cpu->cpu_index = cpu_get_free_index();
83 assert(cpu->cpu_index != UNASSIGNED_CPU_INDEX);
84 } else {
85 assert(!cpu_index_auto_assigned);
86 }
87 QTAILQ_INSERT_TAIL(&cpus, cpu, node);
88 qemu_mutex_unlock(&qemu_cpu_list_lock);
89
90 finish_safe_work(cpu);
91 }
92
93 void cpu_list_remove(CPUState *cpu)
94 {
95 qemu_mutex_lock(&qemu_cpu_list_lock);
96 if (!QTAILQ_IN_USE(cpu, node)) {
97 /* there is nothing to undo since cpu_exec_init() hasn't been called */
98 qemu_mutex_unlock(&qemu_cpu_list_lock);
99 return;
100 }
101
102 assert(!(cpu_index_auto_assigned && cpu != QTAILQ_LAST(&cpus, CPUTailQ)));
103
104 QTAILQ_REMOVE(&cpus, cpu, node);
105 cpu->cpu_index = UNASSIGNED_CPU_INDEX;
106 qemu_mutex_unlock(&qemu_cpu_list_lock);
107 }
108
109 struct qemu_work_item {
110 struct qemu_work_item *next;
111 run_on_cpu_func func;
112 run_on_cpu_data data;
113 bool free, exclusive, done;
114 };
115
116 static void queue_work_on_cpu(CPUState *cpu, struct qemu_work_item *wi)
117 {
118 qemu_mutex_lock(&cpu->work_mutex);
119 if (cpu->queued_work_first == NULL) {
120 cpu->queued_work_first = wi;
121 } else {
122 cpu->queued_work_last->next = wi;
123 }
124 cpu->queued_work_last = wi;
125 wi->next = NULL;
126 wi->done = false;
127 qemu_mutex_unlock(&cpu->work_mutex);
128
129 qemu_cpu_kick(cpu);
130 }
131
132 void do_run_on_cpu(CPUState *cpu, run_on_cpu_func func, run_on_cpu_data data,
133 QemuMutex *mutex)
134 {
135 struct qemu_work_item wi;
136
137 if (qemu_cpu_is_self(cpu)) {
138 func(cpu, data);
139 return;
140 }
141
142 wi.func = func;
143 wi.data = data;
144 wi.done = false;
145 wi.free = false;
146 wi.exclusive = false;
147
148 queue_work_on_cpu(cpu, &wi);
149 while (!atomic_mb_read(&wi.done)) {
150 CPUState *self_cpu = current_cpu;
151
152 qemu_cond_wait(&qemu_work_cond, mutex);
153 current_cpu = self_cpu;
154 }
155 }
156
157 void async_run_on_cpu(CPUState *cpu, run_on_cpu_func func, run_on_cpu_data data)
158 {
159 struct qemu_work_item *wi;
160
161 wi = g_malloc0(sizeof(struct qemu_work_item));
162 wi->func = func;
163 wi->data = data;
164 wi->free = true;
165
166 queue_work_on_cpu(cpu, wi);
167 }
168
169 /* Wait for pending exclusive operations to complete. The CPU list lock
170 must be held. */
171 static inline void exclusive_idle(void)
172 {
173 while (pending_cpus) {
174 qemu_cond_wait(&exclusive_resume, &qemu_cpu_list_lock);
175 }
176 }
177
178 /* Start an exclusive operation.
179 Must only be called from outside cpu_exec. */
180 void start_exclusive(void)
181 {
182 CPUState *other_cpu;
183 int running_cpus;
184
185 qemu_mutex_lock(&qemu_cpu_list_lock);
186 exclusive_idle();
187
188 /* Make all other cpus stop executing. */
189 atomic_set(&pending_cpus, 1);
190
191 /* Write pending_cpus before reading other_cpu->running. */
192 smp_mb();
193 running_cpus = 0;
194 CPU_FOREACH(other_cpu) {
195 if (atomic_read(&other_cpu->running)) {
196 other_cpu->has_waiter = true;
197 running_cpus++;
198 qemu_cpu_kick(other_cpu);
199 }
200 }
201
202 atomic_set(&pending_cpus, running_cpus + 1);
203 while (pending_cpus > 1) {
204 qemu_cond_wait(&exclusive_cond, &qemu_cpu_list_lock);
205 }
206
207 /* Can release mutex, no one will enter another exclusive
208 * section until end_exclusive resets pending_cpus to 0.
209 */
210 qemu_mutex_unlock(&qemu_cpu_list_lock);
211 }
212
213 /* Finish an exclusive operation. */
214 void end_exclusive(void)
215 {
216 qemu_mutex_lock(&qemu_cpu_list_lock);
217 atomic_set(&pending_cpus, 0);
218 qemu_cond_broadcast(&exclusive_resume);
219 qemu_mutex_unlock(&qemu_cpu_list_lock);
220 }
221
222 /* Wait for exclusive ops to finish, and begin cpu execution. */
223 void cpu_exec_start(CPUState *cpu)
224 {
225 atomic_set(&cpu->running, true);
226
227 /* Write cpu->running before reading pending_cpus. */
228 smp_mb();
229
230 /* 1. start_exclusive saw cpu->running == true and pending_cpus >= 1.
231 * After taking the lock we'll see cpu->has_waiter == true and run---not
232 * for long because start_exclusive kicked us. cpu_exec_end will
233 * decrement pending_cpus and signal the waiter.
234 *
235 * 2. start_exclusive saw cpu->running == false but pending_cpus >= 1.
236 * This includes the case when an exclusive item is running now.
237 * Then we'll see cpu->has_waiter == false and wait for the item to
238 * complete.
239 *
240 * 3. pending_cpus == 0. Then start_exclusive is definitely going to
241 * see cpu->running == true, and it will kick the CPU.
242 */
243 if (unlikely(atomic_read(&pending_cpus))) {
244 qemu_mutex_lock(&qemu_cpu_list_lock);
245 if (!cpu->has_waiter) {
246 /* Not counted in pending_cpus, let the exclusive item
247 * run. Since we have the lock, just set cpu->running to true
248 * while holding it; no need to check pending_cpus again.
249 */
250 atomic_set(&cpu->running, false);
251 exclusive_idle();
252 /* Now pending_cpus is zero. */
253 atomic_set(&cpu->running, true);
254 } else {
255 /* Counted in pending_cpus, go ahead and release the
256 * waiter at cpu_exec_end.
257 */
258 }
259 qemu_mutex_unlock(&qemu_cpu_list_lock);
260 }
261 }
262
263 /* Mark cpu as not executing, and release pending exclusive ops. */
264 void cpu_exec_end(CPUState *cpu)
265 {
266 atomic_set(&cpu->running, false);
267
268 /* Write cpu->running before reading pending_cpus. */
269 smp_mb();
270
271 /* 1. start_exclusive saw cpu->running == true. Then it will increment
272 * pending_cpus and wait for exclusive_cond. After taking the lock
273 * we'll see cpu->has_waiter == true.
274 *
275 * 2. start_exclusive saw cpu->running == false but here pending_cpus >= 1.
276 * This includes the case when an exclusive item started after setting
277 * cpu->running to false and before we read pending_cpus. Then we'll see
278 * cpu->has_waiter == false and not touch pending_cpus. The next call to
279 * cpu_exec_start will run exclusive_idle if still necessary, thus waiting
280 * for the item to complete.
281 *
282 * 3. pending_cpus == 0. Then start_exclusive is definitely going to
283 * see cpu->running == false, and it can ignore this CPU until the
284 * next cpu_exec_start.
285 */
286 if (unlikely(atomic_read(&pending_cpus))) {
287 qemu_mutex_lock(&qemu_cpu_list_lock);
288 if (cpu->has_waiter) {
289 cpu->has_waiter = false;
290 atomic_set(&pending_cpus, pending_cpus - 1);
291 if (pending_cpus == 1) {
292 qemu_cond_signal(&exclusive_cond);
293 }
294 }
295 qemu_mutex_unlock(&qemu_cpu_list_lock);
296 }
297 }
298
299 void async_safe_run_on_cpu(CPUState *cpu, run_on_cpu_func func,
300 run_on_cpu_data data)
301 {
302 struct qemu_work_item *wi;
303
304 wi = g_malloc0(sizeof(struct qemu_work_item));
305 wi->func = func;
306 wi->data = data;
307 wi->free = true;
308 wi->exclusive = true;
309
310 queue_work_on_cpu(cpu, wi);
311 }
312
313 void process_queued_cpu_work(CPUState *cpu)
314 {
315 struct qemu_work_item *wi;
316
317 if (cpu->queued_work_first == NULL) {
318 return;
319 }
320
321 qemu_mutex_lock(&cpu->work_mutex);
322 while (cpu->queued_work_first != NULL) {
323 wi = cpu->queued_work_first;
324 cpu->queued_work_first = wi->next;
325 if (!cpu->queued_work_first) {
326 cpu->queued_work_last = NULL;
327 }
328 qemu_mutex_unlock(&cpu->work_mutex);
329 if (wi->exclusive) {
330 /* Running work items outside the BQL avoids the following deadlock:
331 * 1) start_exclusive() is called with the BQL taken while another
332 * CPU is running; 2) cpu_exec in the other CPU tries to takes the
333 * BQL, so it goes to sleep; start_exclusive() is sleeping too, so
334 * neither CPU can proceed.
335 */
336 qemu_mutex_unlock_iothread();
337 start_exclusive();
338 wi->func(cpu, wi->data);
339 end_exclusive();
340 qemu_mutex_lock_iothread();
341 } else {
342 wi->func(cpu, wi->data);
343 }
344 qemu_mutex_lock(&cpu->work_mutex);
345 if (wi->free) {
346 g_free(wi);
347 } else {
348 atomic_mb_set(&wi->done, true);
349 }
350 }
351 qemu_mutex_unlock(&cpu->work_mutex);
352 qemu_cond_broadcast(&qemu_work_cond);
353 }