*/
#include "qemu/osdep.h"
+#include "qemu/main-loop.h"
#include "exec/cpu-common.h"
#include "qom/cpu.h"
#include "sysemu/cpus.h"
static QemuMutex qemu_cpu_list_lock;
+static QemuCond exclusive_cond;
+static QemuCond exclusive_resume;
static QemuCond qemu_work_cond;
+/* >= 1 if a thread is inside start_exclusive/end_exclusive. Written
+ * under qemu_cpu_list_lock, read with atomic operations.
+ */
+static int pending_cpus;
+
void qemu_init_cpu_list(void)
{
+ /* This is needed because qemu_init_cpu_list is also called by the
+ * child process in a fork. */
+ pending_cpus = 0;
+
qemu_mutex_init(&qemu_cpu_list_lock);
+ qemu_cond_init(&exclusive_cond);
+ qemu_cond_init(&exclusive_resume);
qemu_cond_init(&qemu_work_cond);
}
return cpu_index;
}
+static void finish_safe_work(CPUState *cpu)
+{
+ cpu_exec_start(cpu);
+ cpu_exec_end(cpu);
+}
+
void cpu_list_add(CPUState *cpu)
{
qemu_mutex_lock(&qemu_cpu_list_lock);
} else {
assert(!cpu_index_auto_assigned);
}
- QTAILQ_INSERT_TAIL(&cpus, cpu, node);
+ QTAILQ_INSERT_TAIL_RCU(&cpus, cpu, node);
qemu_mutex_unlock(&qemu_cpu_list_lock);
+
+ finish_safe_work(cpu);
}
void cpu_list_remove(CPUState *cpu)
return;
}
- assert(!(cpu_index_auto_assigned && cpu != QTAILQ_LAST(&cpus, CPUTailQ)));
+ assert(!(cpu_index_auto_assigned && cpu != QTAILQ_LAST(&cpus)));
- QTAILQ_REMOVE(&cpus, cpu, node);
+ QTAILQ_REMOVE_RCU(&cpus, cpu, node);
cpu->cpu_index = UNASSIGNED_CPU_INDEX;
qemu_mutex_unlock(&qemu_cpu_list_lock);
}
struct qemu_work_item {
struct qemu_work_item *next;
run_on_cpu_func func;
- void *data;
- bool free, done;
+ run_on_cpu_data data;
+ bool free, exclusive, done;
};
static void queue_work_on_cpu(CPUState *cpu, struct qemu_work_item *wi)
qemu_cpu_kick(cpu);
}
-void do_run_on_cpu(CPUState *cpu, run_on_cpu_func func, void *data,
+void do_run_on_cpu(CPUState *cpu, run_on_cpu_func func, run_on_cpu_data data,
QemuMutex *mutex)
{
struct qemu_work_item wi;
wi.data = data;
wi.done = false;
wi.free = false;
+ wi.exclusive = false;
queue_work_on_cpu(cpu, &wi);
while (!atomic_mb_read(&wi.done)) {
}
}
-void async_run_on_cpu(CPUState *cpu, run_on_cpu_func func, void *data)
+void async_run_on_cpu(CPUState *cpu, run_on_cpu_func func, run_on_cpu_data data)
{
struct qemu_work_item *wi;
- if (qemu_cpu_is_self(cpu)) {
- func(cpu, data);
- return;
+ wi = g_malloc0(sizeof(struct qemu_work_item));
+ wi->func = func;
+ wi->data = data;
+ wi->free = true;
+
+ queue_work_on_cpu(cpu, wi);
+}
+
+/* Wait for pending exclusive operations to complete. The CPU list lock
+ must be held. */
+static inline void exclusive_idle(void)
+{
+ while (pending_cpus) {
+ qemu_cond_wait(&exclusive_resume, &qemu_cpu_list_lock);
+ }
+}
+
+/* Start an exclusive operation.
+ Must only be called from outside cpu_exec. */
+void start_exclusive(void)
+{
+ CPUState *other_cpu;
+ int running_cpus;
+
+ qemu_mutex_lock(&qemu_cpu_list_lock);
+ exclusive_idle();
+
+ /* Make all other cpus stop executing. */
+ atomic_set(&pending_cpus, 1);
+
+ /* Write pending_cpus before reading other_cpu->running. */
+ smp_mb();
+ running_cpus = 0;
+ CPU_FOREACH(other_cpu) {
+ if (atomic_read(&other_cpu->running)) {
+ other_cpu->has_waiter = true;
+ running_cpus++;
+ qemu_cpu_kick(other_cpu);
+ }
+ }
+
+ atomic_set(&pending_cpus, running_cpus + 1);
+ while (pending_cpus > 1) {
+ qemu_cond_wait(&exclusive_cond, &qemu_cpu_list_lock);
}
+ /* Can release mutex, no one will enter another exclusive
+ * section until end_exclusive resets pending_cpus to 0.
+ */
+ qemu_mutex_unlock(&qemu_cpu_list_lock);
+}
+
+/* Finish an exclusive operation. */
+void end_exclusive(void)
+{
+ qemu_mutex_lock(&qemu_cpu_list_lock);
+ atomic_set(&pending_cpus, 0);
+ qemu_cond_broadcast(&exclusive_resume);
+ qemu_mutex_unlock(&qemu_cpu_list_lock);
+}
+
+/* Wait for exclusive ops to finish, and begin cpu execution. */
+void cpu_exec_start(CPUState *cpu)
+{
+ atomic_set(&cpu->running, true);
+
+ /* Write cpu->running before reading pending_cpus. */
+ smp_mb();
+
+ /* 1. start_exclusive saw cpu->running == true and pending_cpus >= 1.
+ * After taking the lock we'll see cpu->has_waiter == true and run---not
+ * for long because start_exclusive kicked us. cpu_exec_end will
+ * decrement pending_cpus and signal the waiter.
+ *
+ * 2. start_exclusive saw cpu->running == false but pending_cpus >= 1.
+ * This includes the case when an exclusive item is running now.
+ * Then we'll see cpu->has_waiter == false and wait for the item to
+ * complete.
+ *
+ * 3. pending_cpus == 0. Then start_exclusive is definitely going to
+ * see cpu->running == true, and it will kick the CPU.
+ */
+ if (unlikely(atomic_read(&pending_cpus))) {
+ qemu_mutex_lock(&qemu_cpu_list_lock);
+ if (!cpu->has_waiter) {
+ /* Not counted in pending_cpus, let the exclusive item
+ * run. Since we have the lock, just set cpu->running to true
+ * while holding it; no need to check pending_cpus again.
+ */
+ atomic_set(&cpu->running, false);
+ exclusive_idle();
+ /* Now pending_cpus is zero. */
+ atomic_set(&cpu->running, true);
+ } else {
+ /* Counted in pending_cpus, go ahead and release the
+ * waiter at cpu_exec_end.
+ */
+ }
+ qemu_mutex_unlock(&qemu_cpu_list_lock);
+ }
+}
+
+/* Mark cpu as not executing, and release pending exclusive ops. */
+void cpu_exec_end(CPUState *cpu)
+{
+ atomic_set(&cpu->running, false);
+
+ /* Write cpu->running before reading pending_cpus. */
+ smp_mb();
+
+ /* 1. start_exclusive saw cpu->running == true. Then it will increment
+ * pending_cpus and wait for exclusive_cond. After taking the lock
+ * we'll see cpu->has_waiter == true.
+ *
+ * 2. start_exclusive saw cpu->running == false but here pending_cpus >= 1.
+ * This includes the case when an exclusive item started after setting
+ * cpu->running to false and before we read pending_cpus. Then we'll see
+ * cpu->has_waiter == false and not touch pending_cpus. The next call to
+ * cpu_exec_start will run exclusive_idle if still necessary, thus waiting
+ * for the item to complete.
+ *
+ * 3. pending_cpus == 0. Then start_exclusive is definitely going to
+ * see cpu->running == false, and it can ignore this CPU until the
+ * next cpu_exec_start.
+ */
+ if (unlikely(atomic_read(&pending_cpus))) {
+ qemu_mutex_lock(&qemu_cpu_list_lock);
+ if (cpu->has_waiter) {
+ cpu->has_waiter = false;
+ atomic_set(&pending_cpus, pending_cpus - 1);
+ if (pending_cpus == 1) {
+ qemu_cond_signal(&exclusive_cond);
+ }
+ }
+ qemu_mutex_unlock(&qemu_cpu_list_lock);
+ }
+}
+
+void async_safe_run_on_cpu(CPUState *cpu, run_on_cpu_func func,
+ run_on_cpu_data data)
+{
+ struct qemu_work_item *wi;
+
wi = g_malloc0(sizeof(struct qemu_work_item));
wi->func = func;
wi->data = data;
wi->free = true;
+ wi->exclusive = true;
queue_work_on_cpu(cpu, wi);
}
cpu->queued_work_last = NULL;
}
qemu_mutex_unlock(&cpu->work_mutex);
- wi->func(cpu, wi->data);
+ if (wi->exclusive) {
+ /* Running work items outside the BQL avoids the following deadlock:
+ * 1) start_exclusive() is called with the BQL taken while another
+ * CPU is running; 2) cpu_exec in the other CPU tries to takes the
+ * BQL, so it goes to sleep; start_exclusive() is sleeping too, so
+ * neither CPU can proceed.
+ */
+ qemu_mutex_unlock_iothread();
+ start_exclusive();
+ wi->func(cpu, wi->data);
+ end_exclusive();
+ qemu_mutex_lock_iothread();
+ } else {
+ wi->func(cpu, wi->data);
+ }
qemu_mutex_lock(&cpu->work_mutex);
if (wi->free) {
g_free(wi);
projects / mirror_qemu.git / blobdiff