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