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 2.1 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 static unsigned int cpu_list_generation_id;
  77
  78 unsigned int cpu_list_generation_id_get(void)
  79 {
  80     return cpu_list_generation_id;
  81 }
  82
  83 void cpu_list_add(CPUState *cpu)
  84 {
  85     QEMU_LOCK_GUARD(&qemu_cpu_list_lock);
  86     if (cpu->cpu_index == UNASSIGNED_CPU_INDEX) {
  87         cpu->cpu_index = cpu_get_free_index();
  88         assert(cpu->cpu_index != UNASSIGNED_CPU_INDEX);
  89     } else {
  90         assert(!cpu_index_auto_assigned);
  91     }
  92     QTAILQ_INSERT_TAIL_RCU(&cpus, cpu, node);
  93     cpu_list_generation_id++;
  94 }
  95
  96 void cpu_list_remove(CPUState *cpu)
  97 {
  98     QEMU_LOCK_GUARD(&qemu_cpu_list_lock);
  99     if (!QTAILQ_IN_USE(cpu, node)) {
 100         /* there is nothing to undo since cpu_exec_init() hasn't been called */
 101         return;
 102     }
 103
 104     QTAILQ_REMOVE_RCU(&cpus, cpu, node);
 105     cpu->cpu_index = UNASSIGNED_CPU_INDEX;
 106     cpu_list_generation_id++;
 107 }
 108
 109 CPUState *qemu_get_cpu(int index)
 110 {
 111     CPUState *cpu;
 112
 113     CPU_FOREACH(cpu) {
 114         if (cpu->cpu_index == index) {
 115             return cpu;
 116         }
 117     }
 118
 119     return NULL;
 120 }
 121
 122 /* current CPU in the current thread. It is only valid inside cpu_exec() */
 123 __thread CPUState *current_cpu;
 124
 125 struct qemu_work_item {
 126     QSIMPLEQ_ENTRY(qemu_work_item) node;
 127     run_on_cpu_func func;
 128     run_on_cpu_data data;
 129     bool free, exclusive, done;
 130 };
 131
 132 static void queue_work_on_cpu(CPUState *cpu, struct qemu_work_item *wi)
 133 {
 134     qemu_mutex_lock(&cpu->work_mutex);
 135     QSIMPLEQ_INSERT_TAIL(&cpu->work_list, wi, node);
 136     wi->done = false;
 137     qemu_mutex_unlock(&cpu->work_mutex);
 138
 139     qemu_cpu_kick(cpu);
 140 }
 141
 142 void do_run_on_cpu(CPUState *cpu, run_on_cpu_func func, run_on_cpu_data data,
 143                    QemuMutex *mutex)
 144 {
 145     struct qemu_work_item wi;
 146
 147     if (qemu_cpu_is_self(cpu)) {
 148         func(cpu, data);
 149         return;
 150     }
 151
 152     wi.func = func;
 153     wi.data = data;
 154     wi.done = false;
 155     wi.free = false;
 156     wi.exclusive = false;
 157
 158     queue_work_on_cpu(cpu, &wi);
 159     while (!qatomic_mb_read(&wi.done)) {
 160         CPUState *self_cpu = current_cpu;
 161
 162         qemu_cond_wait(&qemu_work_cond, mutex);
 163         current_cpu = self_cpu;
 164     }
 165 }
 166
 167 void async_run_on_cpu(CPUState *cpu, run_on_cpu_func func, run_on_cpu_data data)
 168 {
 169     struct qemu_work_item *wi;
 170
 171     wi = g_new0(struct qemu_work_item, 1);
 172     wi->func = func;
 173     wi->data = data;
 174     wi->free = true;
 175
 176     queue_work_on_cpu(cpu, wi);
 177 }
 178
 179 /* Wait for pending exclusive operations to complete.  The CPU list lock
 180    must be held.  */
 181 static inline void exclusive_idle(void)
 182 {
 183     while (pending_cpus) {
 184         qemu_cond_wait(&exclusive_resume, &qemu_cpu_list_lock);
 185     }
 186 }
 187
 188 /* Start an exclusive operation.
 189    Must only be called from outside cpu_exec.  */
 190 void start_exclusive(void)
 191 {
 192     CPUState *other_cpu;
 193     int running_cpus;
 194
 195     if (current_cpu->exclusive_context_count) {
 196         current_cpu->exclusive_context_count++;
 197         return;
 198     }
 199
 200     qemu_mutex_lock(&qemu_cpu_list_lock);
 201     exclusive_idle();
 202
 203     /* Make all other cpus stop executing.  */
 204     qatomic_set(&pending_cpus, 1);
 205
 206     /* Write pending_cpus before reading other_cpu->running.  */
 207     smp_mb();
 208     running_cpus = 0;
 209     CPU_FOREACH(other_cpu) {
 210         if (qatomic_read(&other_cpu->running)) {
 211             other_cpu->has_waiter = true;
 212             running_cpus++;
 213             qemu_cpu_kick(other_cpu);
 214         }
 215     }
 216
 217     qatomic_set(&pending_cpus, running_cpus + 1);
 218     while (pending_cpus > 1) {
 219         qemu_cond_wait(&exclusive_cond, &qemu_cpu_list_lock);
 220     }
 221
 222     /* Can release mutex, no one will enter another exclusive
 223      * section until end_exclusive resets pending_cpus to 0.
 224      */
 225     qemu_mutex_unlock(&qemu_cpu_list_lock);
 226
 227     current_cpu->exclusive_context_count = 1;
 228 }
 229
 230 /* Finish an exclusive operation.  */
 231 void end_exclusive(void)
 232 {
 233     current_cpu->exclusive_context_count--;
 234     if (current_cpu->exclusive_context_count) {
 235         return;
 236     }
 237
 238     qemu_mutex_lock(&qemu_cpu_list_lock);
 239     qatomic_set(&pending_cpus, 0);
 240     qemu_cond_broadcast(&exclusive_resume);
 241     qemu_mutex_unlock(&qemu_cpu_list_lock);
 242 }
 243
 244 /* Wait for exclusive ops to finish, and begin cpu execution.  */
 245 void cpu_exec_start(CPUState *cpu)
 246 {
 247     qatomic_set(&cpu->running, true);
 248
 249     /* Write cpu->running before reading pending_cpus.  */
 250     smp_mb();
 251
 252     /* 1. start_exclusive saw cpu->running == true and pending_cpus >= 1.
 253      * After taking the lock we'll see cpu->has_waiter == true and run---not
 254      * for long because start_exclusive kicked us.  cpu_exec_end will
 255      * decrement pending_cpus and signal the waiter.
 256      *
 257      * 2. start_exclusive saw cpu->running == false but pending_cpus >= 1.
 258      * This includes the case when an exclusive item is running now.
 259      * Then we'll see cpu->has_waiter == false and wait for the item to
 260      * complete.
 261      *
 262      * 3. pending_cpus == 0.  Then start_exclusive is definitely going to
 263      * see cpu->running == true, and it will kick the CPU.
 264      */
 265     if (unlikely(qatomic_read(&pending_cpus))) {
 266         QEMU_LOCK_GUARD(&qemu_cpu_list_lock);
 267         if (!cpu->has_waiter) {
 268             /* Not counted in pending_cpus, let the exclusive item
 269              * run.  Since we have the lock, just set cpu->running to true
 270              * while holding it; no need to check pending_cpus again.
 271              */
 272             qatomic_set(&cpu->running, false);
 273             exclusive_idle();
 274             /* Now pending_cpus is zero.  */
 275             qatomic_set(&cpu->running, true);
 276         } else {
 277             /* Counted in pending_cpus, go ahead and release the
 278              * waiter at cpu_exec_end.
 279              */
 280         }
 281     }
 282 }
 283
 284 /* Mark cpu as not executing, and release pending exclusive ops.  */
 285 void cpu_exec_end(CPUState *cpu)
 286 {
 287     qatomic_set(&cpu->running, false);
 288
 289     /* Write cpu->running before reading pending_cpus.  */
 290     smp_mb();
 291
 292     /* 1. start_exclusive saw cpu->running == true.  Then it will increment
 293      * pending_cpus and wait for exclusive_cond.  After taking the lock
 294      * we'll see cpu->has_waiter == true.
 295      *
 296      * 2. start_exclusive saw cpu->running == false but here pending_cpus >= 1.
 297      * This includes the case when an exclusive item started after setting
 298      * cpu->running to false and before we read pending_cpus.  Then we'll see
 299      * cpu->has_waiter == false and not touch pending_cpus.  The next call to
 300      * cpu_exec_start will run exclusive_idle if still necessary, thus waiting
 301      * for the item to complete.
 302      *
 303      * 3. pending_cpus == 0.  Then start_exclusive is definitely going to
 304      * see cpu->running == false, and it can ignore this CPU until the
 305      * next cpu_exec_start.
 306      */
 307     if (unlikely(qatomic_read(&pending_cpus))) {
 308         QEMU_LOCK_GUARD(&qemu_cpu_list_lock);
 309         if (cpu->has_waiter) {
 310             cpu->has_waiter = false;
 311             qatomic_set(&pending_cpus, pending_cpus - 1);
 312             if (pending_cpus == 1) {
 313                 qemu_cond_signal(&exclusive_cond);
 314             }
 315         }
 316     }
 317 }
 318
 319 void async_safe_run_on_cpu(CPUState *cpu, run_on_cpu_func func,
 320                            run_on_cpu_data data)
 321 {
 322     struct qemu_work_item *wi;
 323
 324     wi = g_new0(struct qemu_work_item, 1);
 325     wi->func = func;
 326     wi->data = data;
 327     wi->free = true;
 328     wi->exclusive = true;
 329
 330     queue_work_on_cpu(cpu, wi);
 331 }
 332
 333 void process_queued_cpu_work(CPUState *cpu)
 334 {
 335     struct qemu_work_item *wi;
 336
 337     qemu_mutex_lock(&cpu->work_mutex);
 338     if (QSIMPLEQ_EMPTY(&cpu->work_list)) {
 339         qemu_mutex_unlock(&cpu->work_mutex);
 340         return;
 341     }
 342     while (!QSIMPLEQ_EMPTY(&cpu->work_list)) {
 343         wi = QSIMPLEQ_FIRST(&cpu->work_list);
 344         QSIMPLEQ_REMOVE_HEAD(&cpu->work_list, node);
 345         qemu_mutex_unlock(&cpu->work_mutex);
 346         if (wi->exclusive) {
 347             /* Running work items outside the BQL avoids the following deadlock:
 348              * 1) start_exclusive() is called with the BQL taken while another
 349              * CPU is running; 2) cpu_exec in the other CPU tries to takes the
 350              * BQL, so it goes to sleep; start_exclusive() is sleeping too, so
 351              * neither CPU can proceed.
 352              */
 353             qemu_mutex_unlock_iothread();
 354             start_exclusive();
 355             wi->func(cpu, wi->data);
 356             end_exclusive();
 357             qemu_mutex_lock_iothread();
 358         } else {
 359             wi->func(cpu, wi->data);
 360         }
 361         qemu_mutex_lock(&cpu->work_mutex);
 362         if (wi->free) {
 363             g_free(wi);
 364         } else {
 365             qatomic_mb_set(&wi->done, true);
 366         }
 367     }
 368     qemu_mutex_unlock(&cpu->work_mutex);
 369     qemu_cond_broadcast(&qemu_work_cond);
 370 }