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     QSIMPLEQ_ENTRY(qemu_work_item) node;
 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     QSIMPLEQ_INSERT_TAIL(&cpu->work_list, wi, node);
 110     wi->done = false;
 111     qemu_mutex_unlock(&cpu->work_mutex);
 112
 113     qemu_cpu_kick(cpu);
 114 }
 115
 116 void do_run_on_cpu(CPUState *cpu, run_on_cpu_func func, run_on_cpu_data data,
 117                    QemuMutex *mutex)
 118 {
 119     struct qemu_work_item wi;
 120
 121     if (qemu_cpu_is_self(cpu)) {
 122         func(cpu, data);
 123         return;
 124     }
 125
 126     wi.func = func;
 127     wi.data = data;
 128     wi.done = false;
 129     wi.free = false;
 130     wi.exclusive = false;
 131
 132     queue_work_on_cpu(cpu, &wi);
 133     while (!atomic_mb_read(&wi.done)) {
 134         CPUState *self_cpu = current_cpu;
 135
 136         qemu_cond_wait(&qemu_work_cond, mutex);
 137         current_cpu = self_cpu;
 138     }
 139 }
 140
 141 void async_run_on_cpu(CPUState *cpu, run_on_cpu_func func, run_on_cpu_data data)
 142 {
 143     struct qemu_work_item *wi;
 144
 145     wi = g_malloc0(sizeof(struct qemu_work_item));
 146     wi->func = func;
 147     wi->data = data;
 148     wi->free = true;
 149
 150     queue_work_on_cpu(cpu, wi);
 151 }
 152
 153 /* Wait for pending exclusive operations to complete.  The CPU list lock
 154    must be held.  */
 155 static inline void exclusive_idle(void)
 156 {
 157     while (pending_cpus) {
 158         qemu_cond_wait(&exclusive_resume, &qemu_cpu_list_lock);
 159     }
 160 }
 161
 162 /* Start an exclusive operation.
 163    Must only be called from outside cpu_exec.  */
 164 void start_exclusive(void)
 165 {
 166     CPUState *other_cpu;
 167     int running_cpus;
 168
 169     qemu_mutex_lock(&qemu_cpu_list_lock);
 170     exclusive_idle();
 171
 172     /* Make all other cpus stop executing.  */
 173     atomic_set(&pending_cpus, 1);
 174
 175     /* Write pending_cpus before reading other_cpu->running.  */
 176     smp_mb();
 177     running_cpus = 0;
 178     CPU_FOREACH(other_cpu) {
 179         if (atomic_read(&other_cpu->running)) {
 180             other_cpu->has_waiter = true;
 181             running_cpus++;
 182             qemu_cpu_kick(other_cpu);
 183         }
 184     }
 185
 186     atomic_set(&pending_cpus, running_cpus + 1);
 187     while (pending_cpus > 1) {
 188         qemu_cond_wait(&exclusive_cond, &qemu_cpu_list_lock);
 189     }
 190
 191     /* Can release mutex, no one will enter another exclusive
 192      * section until end_exclusive resets pending_cpus to 0.
 193      */
 194     qemu_mutex_unlock(&qemu_cpu_list_lock);
 195
 196     current_cpu->in_exclusive_context = true;
 197 }
 198
 199 /* Finish an exclusive operation.  */
 200 void end_exclusive(void)
 201 {
 202     current_cpu->in_exclusive_context = false;
 203
 204     qemu_mutex_lock(&qemu_cpu_list_lock);
 205     atomic_set(&pending_cpus, 0);
 206     qemu_cond_broadcast(&exclusive_resume);
 207     qemu_mutex_unlock(&qemu_cpu_list_lock);
 208 }
 209
 210 /* Wait for exclusive ops to finish, and begin cpu execution.  */
 211 void cpu_exec_start(CPUState *cpu)
 212 {
 213     atomic_set(&cpu->running, true);
 214
 215     /* Write cpu->running before reading pending_cpus.  */
 216     smp_mb();
 217
 218     /* 1. start_exclusive saw cpu->running == true and pending_cpus >= 1.
 219      * After taking the lock we'll see cpu->has_waiter == true and run---not
 220      * for long because start_exclusive kicked us.  cpu_exec_end will
 221      * decrement pending_cpus and signal the waiter.
 222      *
 223      * 2. start_exclusive saw cpu->running == false but pending_cpus >= 1.
 224      * This includes the case when an exclusive item is running now.
 225      * Then we'll see cpu->has_waiter == false and wait for the item to
 226      * complete.
 227      *
 228      * 3. pending_cpus == 0.  Then start_exclusive is definitely going to
 229      * see cpu->running == true, and it will kick the CPU.
 230      */
 231     if (unlikely(atomic_read(&pending_cpus))) {
 232         QEMU_LOCK_GUARD(&qemu_cpu_list_lock);
 233         if (!cpu->has_waiter) {
 234             /* Not counted in pending_cpus, let the exclusive item
 235              * run.  Since we have the lock, just set cpu->running to true
 236              * while holding it; no need to check pending_cpus again.
 237              */
 238             atomic_set(&cpu->running, false);
 239             exclusive_idle();
 240             /* Now pending_cpus is zero.  */
 241             atomic_set(&cpu->running, true);
 242         } else {
 243             /* Counted in pending_cpus, go ahead and release the
 244              * waiter at cpu_exec_end.
 245              */
 246         }
 247     }
 248 }
 249
 250 /* Mark cpu as not executing, and release pending exclusive ops.  */
 251 void cpu_exec_end(CPUState *cpu)
 252 {
 253     atomic_set(&cpu->running, false);
 254
 255     /* Write cpu->running before reading pending_cpus.  */
 256     smp_mb();
 257
 258     /* 1. start_exclusive saw cpu->running == true.  Then it will increment
 259      * pending_cpus and wait for exclusive_cond.  After taking the lock
 260      * we'll see cpu->has_waiter == true.
 261      *
 262      * 2. start_exclusive saw cpu->running == false but here pending_cpus >= 1.
 263      * This includes the case when an exclusive item started after setting
 264      * cpu->running to false and before we read pending_cpus.  Then we'll see
 265      * cpu->has_waiter == false and not touch pending_cpus.  The next call to
 266      * cpu_exec_start will run exclusive_idle if still necessary, thus waiting
 267      * for the item to complete.
 268      *
 269      * 3. pending_cpus == 0.  Then start_exclusive is definitely going to
 270      * see cpu->running == false, and it can ignore this CPU until the
 271      * next cpu_exec_start.
 272      */
 273     if (unlikely(atomic_read(&pending_cpus))) {
 274         QEMU_LOCK_GUARD(&qemu_cpu_list_lock);
 275         if (cpu->has_waiter) {
 276             cpu->has_waiter = false;
 277             atomic_set(&pending_cpus, pending_cpus - 1);
 278             if (pending_cpus == 1) {
 279                 qemu_cond_signal(&exclusive_cond);
 280             }
 281         }
 282     }
 283 }
 284
 285 void async_safe_run_on_cpu(CPUState *cpu, run_on_cpu_func func,
 286                            run_on_cpu_data data)
 287 {
 288     struct qemu_work_item *wi;
 289
 290     wi = g_malloc0(sizeof(struct qemu_work_item));
 291     wi->func = func;
 292     wi->data = data;
 293     wi->free = true;
 294     wi->exclusive = true;
 295
 296     queue_work_on_cpu(cpu, wi);
 297 }
 298
 299 void process_queued_cpu_work(CPUState *cpu)
 300 {
 301     struct qemu_work_item *wi;
 302
 303     qemu_mutex_lock(&cpu->work_mutex);
 304     if (QSIMPLEQ_EMPTY(&cpu->work_list)) {
 305         qemu_mutex_unlock(&cpu->work_mutex);
 306         return;
 307     }
 308     while (!QSIMPLEQ_EMPTY(&cpu->work_list)) {
 309         wi = QSIMPLEQ_FIRST(&cpu->work_list);
 310         QSIMPLEQ_REMOVE_HEAD(&cpu->work_list, node);
 311         qemu_mutex_unlock(&cpu->work_mutex);
 312         if (wi->exclusive) {
 313             /* Running work items outside the BQL avoids the following deadlock:
 314              * 1) start_exclusive() is called with the BQL taken while another
 315              * CPU is running; 2) cpu_exec in the other CPU tries to takes the
 316              * BQL, so it goes to sleep; start_exclusive() is sleeping too, so
 317              * neither CPU can proceed.
 318              */
 319             qemu_mutex_unlock_iothread();
 320             start_exclusive();
 321             wi->func(cpu, wi->data);
 322             end_exclusive();
 323             qemu_mutex_lock_iothread();
 324         } else {
 325             wi->func(cpu, wi->data);
 326         }
 327         qemu_mutex_lock(&cpu->work_mutex);
 328         if (wi->free) {
 329             g_free(wi);
 330         } else {
 331             atomic_mb_set(&wi->done, true);
 332         }
 333     }
 334     qemu_mutex_unlock(&cpu->work_mutex);
 335     qemu_cond_broadcast(&qemu_work_cond);
 336 }