drain_call_rcu() triggers the force quiescent state, but it can be
delayed if the RCU thread is sleeping. Ensure the force quiescent state
is immediately triggered by waking the RCU thread up.
The logic to trigger the force quiescent state is decoupled as
force_rcu() so that it can be used independently.
Signed-off-by: Akihiko Odaki <odaki@rsg.ci.i.u-tokyo.ac.jp>
---
include/qemu/rcu.h | 1 +
util/rcu.c | 106 ++++++++++++++++++++++++++++++++---------------------
2 files changed, 65 insertions(+), 42 deletions(-)
diff --git a/include/qemu/rcu.h b/include/qemu/rcu.h
index 020dbe4d8b77..d6aa4e5854d3 100644
--- a/include/qemu/rcu.h
+++ b/include/qemu/rcu.h
@@ -118,6 +118,7 @@ static inline void rcu_read_unlock(void)
}
}
+void force_rcu(void);
void synchronize_rcu(void);
/*
diff --git a/util/rcu.c b/util/rcu.c
index 3c4af9d213c8..85f9333f5dff 100644
--- a/util/rcu.c
+++ b/util/rcu.c
@@ -49,10 +49,13 @@
unsigned long rcu_gp_ctr = RCU_GP_LOCKED;
QemuEvent rcu_gp_event;
-static int in_drain_call_rcu;
+static bool forced;
static int rcu_call_count;
static QemuMutex rcu_registry_lock;
+/* Set when the forced variable is set or rcu_call_count becomes non-zero. */
+static QemuEvent sync_event;
+
/*
* Check whether a quiescent state was crossed between the beginning of
* update_counter_and_wait and now.
@@ -74,36 +77,21 @@ QEMU_DEFINE_CO_TLS(struct rcu_reader_data, rcu_reader)
typedef QLIST_HEAD(, rcu_reader_data) ThreadList;
static ThreadList registry = QLIST_HEAD_INITIALIZER(registry);
+void force_rcu(void)
+{
+ qatomic_set(&forced, true);
+ qemu_event_set(&sync_event);
+}
+
/* Wait for previous parity/grace period to be empty of readers. */
-static void wait_for_readers(void)
+static void wait_for_readers(bool sleep)
{
ThreadList qsreaders = QLIST_HEAD_INITIALIZER(qsreaders);
struct rcu_reader_data *index, *tmp;
- int sleeps = 0;
- bool forced = false;
+ int sleeps = sleep ? 5 : 0;
+ bool waiting = false;
for (;;) {
- /*
- * Force the grace period to end and wait for it if any of the
- * following heuristical conditions are satisfied:
- * - A decent number of callbacks piled up.
- * - It timed out.
- * - It is in a drain_call_rcu() call.
- *
- * Otherwise, periodically poll the grace period, hoping it ends
- * promptly.
- */
- if (!forced &&
- (qatomic_read(&rcu_call_count) >= RCU_CALL_MIN_SIZE ||
- sleeps >= 5 || qatomic_read(&in_drain_call_rcu))) {
- forced = true;
-
- QLIST_FOREACH(index, ®istry, node) {
- notifier_list_notify(&index->force_rcu, NULL);
- qatomic_set(&index->waiting, true);
- }
- }
-
/* Here, order the stores to index->waiting before the loads of
* index->ctr. Pairs with smp_mb_placeholder() in rcu_read_unlock(),
* ensuring that the loads of index->ctr are sequentially consistent.
@@ -150,7 +138,8 @@ static void wait_for_readers(void)
*/
qemu_mutex_unlock(&rcu_registry_lock);
- if (forced) {
+ if (waiting) {
+ /* Wait for the forced quiescent state. */
qemu_event_wait(&rcu_gp_event);
/*
@@ -158,9 +147,25 @@ static void wait_for_readers(void)
* while we walk the list.
*/
qemu_event_reset(&rcu_gp_event);
+ } else if (qatomic_read(&rcu_call_count) >= RCU_CALL_MIN_SIZE ||
+ !sleeps || qemu_event_timedwait(&sync_event, 10)) {
+ /*
+ * Now one of the following heuristical conditions is satisfied:
+ * - A decent number of callbacks piled up.
+ * - It timed out.
+ * - force_rcu() was called.
+ *
+ * Force a quiescent state.
+ */
+ waiting = true;
+
+ QLIST_FOREACH(index, ®istry, node) {
+ notifier_list_notify(&index->force_rcu, NULL);
+ qatomic_set(&index->waiting, true);
+ }
} else {
- g_usleep(10000);
- sleeps++;
+ /* Try again. */
+ sleeps--;
}
qemu_mutex_lock(&rcu_registry_lock);
@@ -170,7 +175,7 @@ static void wait_for_readers(void)
QLIST_SWAP(®istry, &qsreaders, node);
}
-static void enter_qs(void)
+static void enter_qs(bool sleep)
{
/* Write RCU-protected pointers before reading p_rcu_reader->ctr.
* Pairs with smp_mb_placeholder() in rcu_read_lock().
@@ -189,14 +194,14 @@ static void enter_qs(void)
* Switch parity: 0 -> 1, 1 -> 0.
*/
qatomic_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
- wait_for_readers();
+ wait_for_readers(sleep);
qatomic_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
} else {
/* Increment current grace period. */
qatomic_set(&rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
}
- wait_for_readers();
+ wait_for_readers(sleep);
}
}
@@ -205,7 +210,6 @@ static void enter_qs(void)
*/
static struct rcu_head dummy;
static struct rcu_head *head = &dummy, **tail = &dummy.next;
-static QemuEvent rcu_call_ready_event;
static void enqueue(struct rcu_head *node)
{
@@ -282,6 +286,7 @@ static void *call_rcu_thread(void *opaque)
rcu_register_thread();
for (;;) {
+ bool sleep = true;
int n;
/*
@@ -289,7 +294,7 @@ static void *call_rcu_thread(void *opaque)
* added before enter_qs() starts.
*/
for (;;) {
- qemu_event_reset(&rcu_call_ready_event);
+ qemu_event_reset(&sync_event);
n = qatomic_read(&rcu_call_count);
if (n) {
break;
@@ -298,20 +303,36 @@ static void *call_rcu_thread(void *opaque)
#if defined(CONFIG_MALLOC_TRIM)
malloc_trim(4 * 1024 * 1024);
#endif
- qemu_event_wait(&rcu_call_ready_event);
+ qemu_event_wait(&sync_event);
+ }
+
+ /*
+ * Ensure that an event for a rcu_call_count change will not interrupt
+ * wait_for_readers().
+ */
+ qemu_event_reset(&sync_event);
+
+ /*
+ * Ensure that the forced variable has not been set after fetching
+ * rcu_call_count; otherwise we may get confused by a force quiescent
+ * state request for an element later than n.
+ */
+ while (qatomic_xchg(&forced, false)) {
+ sleep = false;
+ n = qatomic_read(&rcu_call_count);
}
- enter_qs();
+ enter_qs(sleep);
qatomic_sub(&rcu_call_count, n);
bql_lock();
while (n > 0) {
node = try_dequeue();
while (!node) {
bql_unlock();
- qemu_event_reset(&rcu_call_ready_event);
+ qemu_event_reset(&sync_event);
node = try_dequeue();
if (!node) {
- qemu_event_wait(&rcu_call_ready_event);
+ qemu_event_wait(&sync_event);
node = try_dequeue();
}
bql_lock();
@@ -329,8 +350,10 @@ void call_rcu1(struct rcu_head *node, void (*func)(struct rcu_head *node))
{
node->func = func;
enqueue(node);
- qatomic_inc(&rcu_call_count);
- qemu_event_set(&rcu_call_ready_event);
+
+ if (!qatomic_fetch_inc(&rcu_call_count)) {
+ qemu_event_set(&sync_event);
+ }
}
@@ -388,10 +411,9 @@ void drain_call_rcu(void)
* assumed.
*/
- qatomic_inc(&in_drain_call_rcu);
call_rcu(&sync, sync_rcu_callback, rcu);
+ force_rcu();
qemu_event_wait(&sync.complete_event);
- qatomic_dec(&in_drain_call_rcu);
if (locked) {
bql_lock();
@@ -435,7 +457,7 @@ static void rcu_init_complete(void)
qemu_mutex_init(&rcu_registry_lock);
qemu_event_init(&rcu_gp_event, true);
- qemu_event_init(&rcu_call_ready_event, false);
+ qemu_event_init(&sync_event, false);
/* The caller is assumed to have BQL, so the call_rcu thread
* must have been quiescent even after forking, just recreate it.
--
2.51.0On Wed, Oct 29, 2025 at 03:12:48PM +0900, Akihiko Odaki wrote:
> drain_call_rcu() triggers the force quiescent state, but it can be
> delayed if the RCU thread is sleeping. Ensure the force quiescent state
> is immediately triggered by waking the RCU thread up.
>
> The logic to trigger the force quiescent state is decoupled as
> force_rcu() so that it can be used independently.
>
> Signed-off-by: Akihiko Odaki <odaki@rsg.ci.i.u-tokyo.ac.jp>
> ---
> include/qemu/rcu.h | 1 +
> util/rcu.c | 106 ++++++++++++++++++++++++++++++++---------------------
> 2 files changed, 65 insertions(+), 42 deletions(-)
>
> diff --git a/include/qemu/rcu.h b/include/qemu/rcu.h
> index 020dbe4d8b77..d6aa4e5854d3 100644
> --- a/include/qemu/rcu.h
> +++ b/include/qemu/rcu.h
> @@ -118,6 +118,7 @@ static inline void rcu_read_unlock(void)
> }
> }
>
> +void force_rcu(void);
> void synchronize_rcu(void);
>
> /*
> diff --git a/util/rcu.c b/util/rcu.c
> index 3c4af9d213c8..85f9333f5dff 100644
> --- a/util/rcu.c
> +++ b/util/rcu.c
> @@ -49,10 +49,13 @@
> unsigned long rcu_gp_ctr = RCU_GP_LOCKED;
>
> QemuEvent rcu_gp_event;
> -static int in_drain_call_rcu;
> +static bool forced;
> static int rcu_call_count;
> static QemuMutex rcu_registry_lock;
>
> +/* Set when the forced variable is set or rcu_call_count becomes non-zero. */
> +static QemuEvent sync_event;
> +
> /*
> * Check whether a quiescent state was crossed between the beginning of
> * update_counter_and_wait and now.
> @@ -74,36 +77,21 @@ QEMU_DEFINE_CO_TLS(struct rcu_reader_data, rcu_reader)
> typedef QLIST_HEAD(, rcu_reader_data) ThreadList;
> static ThreadList registry = QLIST_HEAD_INITIALIZER(registry);
>
> +void force_rcu(void)
> +{
> + qatomic_set(&forced, true);
> + qemu_event_set(&sync_event);
> +}
> +
> /* Wait for previous parity/grace period to be empty of readers. */
> -static void wait_for_readers(void)
> +static void wait_for_readers(bool sleep)
> {
> ThreadList qsreaders = QLIST_HEAD_INITIALIZER(qsreaders);
> struct rcu_reader_data *index, *tmp;
> - int sleeps = 0;
> - bool forced = false;
> + int sleeps = sleep ? 5 : 0;
> + bool waiting = false;
>
> for (;;) {
> - /*
> - * Force the grace period to end and wait for it if any of the
> - * following heuristical conditions are satisfied:
> - * - A decent number of callbacks piled up.
> - * - It timed out.
> - * - It is in a drain_call_rcu() call.
> - *
> - * Otherwise, periodically poll the grace period, hoping it ends
> - * promptly.
> - */
> - if (!forced &&
> - (qatomic_read(&rcu_call_count) >= RCU_CALL_MIN_SIZE ||
> - sleeps >= 5 || qatomic_read(&in_drain_call_rcu))) {
> - forced = true;
> -
> - QLIST_FOREACH(index, ®istry, node) {
> - notifier_list_notify(&index->force_rcu, NULL);
> - qatomic_set(&index->waiting, true);
> - }
> - }
IIUC this is the part to set index->waiting first whenever necessary, then
that'll guarantee the wait(rcu_gp_event) will be notified in rcu unlock.
Now we removed this chunk, then could it happen if waiting=true and the
wait(rcu_gp_event) may wait for more than it should (as nobody will wake it
up if all threads have waiting=false)?
The other thing is, right below here there's the code and comment:
/* Here, order the stores to index->waiting before the loads of
* index->ctr. Pairs with smp_mb_placeholder() in rcu_read_unlock(),
* ensuring that the loads of index->ctr are sequentially consistent.
*
* If this is the last iteration, this barrier also prevents
* frees from seeping upwards, and orders the two wait phases
* on architectures with 32-bit longs; see enter_qs().
*/
smp_mb_global();
IIUC it explains the mb_global() to order the updates of waiting and the
reads of index->ctr. It doesn't look like applicable anymore. Said that,
I think we should indeed still need some barrier to make sure we read
index->ctr at least to be after we update global gp_ctr (done before
calling wait_for_readers()). I'm not sure if it means the mb is needed,
however maybe at least the comment is outdated if so.
> -
> /* Here, order the stores to index->waiting before the loads of
> * index->ctr. Pairs with smp_mb_placeholder() in rcu_read_unlock(),
> * ensuring that the loads of index->ctr are sequentially consistent.
> @@ -150,7 +138,8 @@ static void wait_for_readers(void)
> */
> qemu_mutex_unlock(&rcu_registry_lock);
>
> - if (forced) {
> + if (waiting) {
> + /* Wait for the forced quiescent state. */
> qemu_event_wait(&rcu_gp_event);
>
> /*
> @@ -158,9 +147,25 @@ static void wait_for_readers(void)
> * while we walk the list.
> */
> qemu_event_reset(&rcu_gp_event);
> + } else if (qatomic_read(&rcu_call_count) >= RCU_CALL_MIN_SIZE ||
> + !sleeps || qemu_event_timedwait(&sync_event, 10)) {
> + /*
> + * Now one of the following heuristical conditions is satisfied:
> + * - A decent number of callbacks piled up.
> + * - It timed out.
> + * - force_rcu() was called.
> + *
> + * Force a quiescent state.
> + */
> + waiting = true;
> +
> + QLIST_FOREACH(index, ®istry, node) {
> + notifier_list_notify(&index->force_rcu, NULL);
> + qatomic_set(&index->waiting, true);
> + }
> } else {
> - g_usleep(10000);
> - sleeps++;
> + /* Try again. */
> + sleeps--;
> }
>
> qemu_mutex_lock(&rcu_registry_lock);
> @@ -170,7 +175,7 @@ static void wait_for_readers(void)
> QLIST_SWAP(®istry, &qsreaders, node);
> }
>
> -static void enter_qs(void)
> +static void enter_qs(bool sleep)
> {
> /* Write RCU-protected pointers before reading p_rcu_reader->ctr.
> * Pairs with smp_mb_placeholder() in rcu_read_lock().
> @@ -189,14 +194,14 @@ static void enter_qs(void)
> * Switch parity: 0 -> 1, 1 -> 0.
> */
> qatomic_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
> - wait_for_readers();
> + wait_for_readers(sleep);
> qatomic_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
> } else {
> /* Increment current grace period. */
> qatomic_set(&rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
> }
>
> - wait_for_readers();
> + wait_for_readers(sleep);
> }
> }
>
> @@ -205,7 +210,6 @@ static void enter_qs(void)
> */
> static struct rcu_head dummy;
> static struct rcu_head *head = &dummy, **tail = &dummy.next;
> -static QemuEvent rcu_call_ready_event;
>
> static void enqueue(struct rcu_head *node)
> {
> @@ -282,6 +286,7 @@ static void *call_rcu_thread(void *opaque)
> rcu_register_thread();
>
> for (;;) {
> + bool sleep = true;
> int n;
>
> /*
> @@ -289,7 +294,7 @@ static void *call_rcu_thread(void *opaque)
> * added before enter_qs() starts.
> */
> for (;;) {
> - qemu_event_reset(&rcu_call_ready_event);
> + qemu_event_reset(&sync_event);
> n = qatomic_read(&rcu_call_count);
> if (n) {
> break;
> @@ -298,20 +303,36 @@ static void *call_rcu_thread(void *opaque)
> #if defined(CONFIG_MALLOC_TRIM)
> malloc_trim(4 * 1024 * 1024);
> #endif
> - qemu_event_wait(&rcu_call_ready_event);
> + qemu_event_wait(&sync_event);
> + }
> +
> + /*
> + * Ensure that an event for a rcu_call_count change will not interrupt
> + * wait_for_readers().
> + */
> + qemu_event_reset(&sync_event);
> +
> + /*
> + * Ensure that the forced variable has not been set after fetching
> + * rcu_call_count; otherwise we may get confused by a force quiescent
> + * state request for an element later than n.
> + */
> + while (qatomic_xchg(&forced, false)) {
> + sleep = false;
> + n = qatomic_read(&rcu_call_count);
> }
This is pretty tricky, and I wonder if it will make the code easier to read
if we convert the sync_event to be a semaphore instead. When as a sem, it
will take account of whatever kick to it, either a call_rcu1() or an
enforced rcu flush, so that we don't need to reset it. Meanwhile, we don't
worry on an slightly outdated "n" read because the 2nd round of sem_wait()
will catch that new "n".
Instead, worst case is rcu thread runs one more round without seeing
callbacks on the queue.
I'm not sure if that could help simplying code, maybe also make it less
error-prone.
>
> - enter_qs();
> + enter_qs(sleep);
> qatomic_sub(&rcu_call_count, n);
> bql_lock();
> while (n > 0) {
> node = try_dequeue();
> while (!node) {
I have a pure question here not relevant to your changes.. do you know when
this "if" will trigger? It seems to me the enqueue() should always happen
before the increment of rcu_call_count:
void call_rcu1(struct rcu_head *node, void (*func)(struct rcu_head *node))
{
node->func = func;
enqueue(node);
if (!qatomic_fetch_inc(&rcu_call_count)) {
qemu_event_set(&sync_event);
}
}
I believe qatomic_fetch_inc() is RMW so it's strong mb and order
guaranteed. Then here why the node can be null even if we're sure >=n have
been enqueued?
Thanks,
> bql_unlock();
> - qemu_event_reset(&rcu_call_ready_event);
> + qemu_event_reset(&sync_event);
> node = try_dequeue();
> if (!node) {
> - qemu_event_wait(&rcu_call_ready_event);
> + qemu_event_wait(&sync_event);
> node = try_dequeue();
> }
> bql_lock();
> @@ -329,8 +350,10 @@ void call_rcu1(struct rcu_head *node, void (*func)(struct rcu_head *node))
> {
> node->func = func;
> enqueue(node);
> - qatomic_inc(&rcu_call_count);
> - qemu_event_set(&rcu_call_ready_event);
> +
> + if (!qatomic_fetch_inc(&rcu_call_count)) {
> + qemu_event_set(&sync_event);
> + }
> }
>
>
> @@ -388,10 +411,9 @@ void drain_call_rcu(void)
> * assumed.
> */
>
> - qatomic_inc(&in_drain_call_rcu);
> call_rcu(&sync, sync_rcu_callback, rcu);
> + force_rcu();
> qemu_event_wait(&sync.complete_event);
> - qatomic_dec(&in_drain_call_rcu);
>
> if (locked) {
> bql_lock();
> @@ -435,7 +457,7 @@ static void rcu_init_complete(void)
> qemu_mutex_init(&rcu_registry_lock);
> qemu_event_init(&rcu_gp_event, true);
>
> - qemu_event_init(&rcu_call_ready_event, false);
> + qemu_event_init(&sync_event, false);
>
> /* The caller is assumed to have BQL, so the call_rcu thread
> * must have been quiescent even after forking, just recreate it.
>
> --
> 2.51.0
>
>
--
Peter Xu
On 2025/10/30 3:22, Peter Xu wrote:
> On Wed, Oct 29, 2025 at 03:12:48PM +0900, Akihiko Odaki wrote:
>> drain_call_rcu() triggers the force quiescent state, but it can be
>> delayed if the RCU thread is sleeping. Ensure the force quiescent state
>> is immediately triggered by waking the RCU thread up.
>>
>> The logic to trigger the force quiescent state is decoupled as
>> force_rcu() so that it can be used independently.
>>
>> Signed-off-by: Akihiko Odaki <odaki@rsg.ci.i.u-tokyo.ac.jp>
>> ---
>> include/qemu/rcu.h | 1 +
>> util/rcu.c | 106 ++++++++++++++++++++++++++++++++---------------------
>> 2 files changed, 65 insertions(+), 42 deletions(-)
>>
>> diff --git a/include/qemu/rcu.h b/include/qemu/rcu.h
>> index 020dbe4d8b77..d6aa4e5854d3 100644
>> --- a/include/qemu/rcu.h
>> +++ b/include/qemu/rcu.h
>> @@ -118,6 +118,7 @@ static inline void rcu_read_unlock(void)
>> }
>> }
>>
>> +void force_rcu(void);
>> void synchronize_rcu(void);
>>
>> /*
>> diff --git a/util/rcu.c b/util/rcu.c
>> index 3c4af9d213c8..85f9333f5dff 100644
>> --- a/util/rcu.c
>> +++ b/util/rcu.c
>> @@ -49,10 +49,13 @@
>> unsigned long rcu_gp_ctr = RCU_GP_LOCKED;
>>
>> QemuEvent rcu_gp_event;
>> -static int in_drain_call_rcu;
>> +static bool forced;
>> static int rcu_call_count;
>> static QemuMutex rcu_registry_lock;
>>
>> +/* Set when the forced variable is set or rcu_call_count becomes non-zero. */
>> +static QemuEvent sync_event;
>> +
>> /*
>> * Check whether a quiescent state was crossed between the beginning of
>> * update_counter_and_wait and now.
>> @@ -74,36 +77,21 @@ QEMU_DEFINE_CO_TLS(struct rcu_reader_data, rcu_reader)
>> typedef QLIST_HEAD(, rcu_reader_data) ThreadList;
>> static ThreadList registry = QLIST_HEAD_INITIALIZER(registry);
>>
>> +void force_rcu(void)
>> +{
>> + qatomic_set(&forced, true);
>> + qemu_event_set(&sync_event);
>> +}
>> +
>> /* Wait for previous parity/grace period to be empty of readers. */
>> -static void wait_for_readers(void)
>> +static void wait_for_readers(bool sleep)
>> {
>> ThreadList qsreaders = QLIST_HEAD_INITIALIZER(qsreaders);
>> struct rcu_reader_data *index, *tmp;
>> - int sleeps = 0;
>> - bool forced = false;
>> + int sleeps = sleep ? 5 : 0;
>> + bool waiting = false;
>>
>> for (;;) {
>> - /*
>> - * Force the grace period to end and wait for it if any of the
>> - * following heuristical conditions are satisfied:
>> - * - A decent number of callbacks piled up.
>> - * - It timed out.
>> - * - It is in a drain_call_rcu() call.
>> - *
>> - * Otherwise, periodically poll the grace period, hoping it ends
>> - * promptly.
>> - */
>> - if (!forced &&
>> - (qatomic_read(&rcu_call_count) >= RCU_CALL_MIN_SIZE ||
>> - sleeps >= 5 || qatomic_read(&in_drain_call_rcu))) {
>> - forced = true;
>> -
>> - QLIST_FOREACH(index, ®istry, node) {
>> - notifier_list_notify(&index->force_rcu, NULL);
>> - qatomic_set(&index->waiting, true);
>> - }
>> - }
>
> IIUC this is the part to set index->waiting first whenever necessary, then
> that'll guarantee the wait(rcu_gp_event) will be notified in rcu unlock.
>
> Now we removed this chunk, then could it happen if waiting=true and the
> wait(rcu_gp_event) may wait for more than it should (as nobody will wake it
> up if all threads have waiting=false)?
It is not removed, but it is moved along with the assignment of the
waiting variable. So index->waiting is still set whenever the waiting
variable is set and no hang up will happen.
>
> The other thing is, right below here there's the code and comment:
>
> /* Here, order the stores to index->waiting before the loads of
> * index->ctr. Pairs with smp_mb_placeholder() in rcu_read_unlock(),
> * ensuring that the loads of index->ctr are sequentially consistent.
> *
> * If this is the last iteration, this barrier also prevents
> * frees from seeping upwards, and orders the two wait phases
> * on architectures with 32-bit longs; see enter_qs().
> */
> smp_mb_global();
>
> IIUC it explains the mb_global() to order the updates of waiting and the
> reads of index->ctr. It doesn't look like applicable anymore. Said that,
> I think we should indeed still need some barrier to make sure we read
> index->ctr at least to be after we update global gp_ctr (done before
> calling wait_for_readers()). I'm not sure if it means the mb is needed,
> however maybe at least the comment is outdated if so.
It is still applicable. The stores to index->waiting is still present
and needs to be ordered before the loads of index->ctr.
>
>> -
>> /* Here, order the stores to index->waiting before the loads of
>> * index->ctr. Pairs with smp_mb_placeholder() in rcu_read_unlock(),
>> * ensuring that the loads of index->ctr are sequentially consistent.
>> @@ -150,7 +138,8 @@ static void wait_for_readers(void)
>> */
>> qemu_mutex_unlock(&rcu_registry_lock);
>>
>> - if (forced) {
>> + if (waiting) {
>> + /* Wait for the forced quiescent state. */
>> qemu_event_wait(&rcu_gp_event);
>>
>> /*
>> @@ -158,9 +147,25 @@ static void wait_for_readers(void)
>> * while we walk the list.
>> */
>> qemu_event_reset(&rcu_gp_event);
>> + } else if (qatomic_read(&rcu_call_count) >= RCU_CALL_MIN_SIZE ||
>> + !sleeps || qemu_event_timedwait(&sync_event, 10)) {
>> + /*
>> + * Now one of the following heuristical conditions is satisfied:
>> + * - A decent number of callbacks piled up.
>> + * - It timed out.
>> + * - force_rcu() was called.
>> + *
>> + * Force a quiescent state.
>> + */
>> + waiting = true;
>> +
>> + QLIST_FOREACH(index, ®istry, node) {
>> + notifier_list_notify(&index->force_rcu, NULL);
>> + qatomic_set(&index->waiting, true);
>> + }
>> } else {
>> - g_usleep(10000);
>> - sleeps++;
>> + /* Try again. */
>> + sleeps--;
>> }
>>
>> qemu_mutex_lock(&rcu_registry_lock);
>> @@ -170,7 +175,7 @@ static void wait_for_readers(void)
>> QLIST_SWAP(®istry, &qsreaders, node);
>> }
>>
>> -static void enter_qs(void)
>> +static void enter_qs(bool sleep)
>> {
>> /* Write RCU-protected pointers before reading p_rcu_reader->ctr.
>> * Pairs with smp_mb_placeholder() in rcu_read_lock().
>> @@ -189,14 +194,14 @@ static void enter_qs(void)
>> * Switch parity: 0 -> 1, 1 -> 0.
>> */
>> qatomic_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
>> - wait_for_readers();
>> + wait_for_readers(sleep);
>> qatomic_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
>> } else {
>> /* Increment current grace period. */
>> qatomic_set(&rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
>> }
>>
>> - wait_for_readers();
>> + wait_for_readers(sleep);
>> }
>> }
>>
>> @@ -205,7 +210,6 @@ static void enter_qs(void)
>> */
>> static struct rcu_head dummy;
>> static struct rcu_head *head = &dummy, **tail = &dummy.next;
>> -static QemuEvent rcu_call_ready_event;
>>
>> static void enqueue(struct rcu_head *node)
>> {
>> @@ -282,6 +286,7 @@ static void *call_rcu_thread(void *opaque)
>> rcu_register_thread();
>>
>> for (;;) {
>> + bool sleep = true;
>> int n;
>>
>> /*
>> @@ -289,7 +294,7 @@ static void *call_rcu_thread(void *opaque)
>> * added before enter_qs() starts.
>> */
>> for (;;) {
>> - qemu_event_reset(&rcu_call_ready_event);
>> + qemu_event_reset(&sync_event);
>> n = qatomic_read(&rcu_call_count);
>> if (n) {
>> break;
>> @@ -298,20 +303,36 @@ static void *call_rcu_thread(void *opaque)
>> #if defined(CONFIG_MALLOC_TRIM)
>> malloc_trim(4 * 1024 * 1024);
>> #endif
>> - qemu_event_wait(&rcu_call_ready_event);
>> + qemu_event_wait(&sync_event);
>> + }
>> +
>> + /*
>> + * Ensure that an event for a rcu_call_count change will not interrupt
>> + * wait_for_readers().
>> + */
>> + qemu_event_reset(&sync_event);
>> +
>> + /*
>> + * Ensure that the forced variable has not been set after fetching
>> + * rcu_call_count; otherwise we may get confused by a force quiescent
>> + * state request for an element later than n.
>> + */
>> + while (qatomic_xchg(&forced, false)) {
>> + sleep = false;
>> + n = qatomic_read(&rcu_call_count);
>> }
>
> This is pretty tricky, and I wonder if it will make the code easier to read
> if we convert the sync_event to be a semaphore instead. When as a sem, it
> will take account of whatever kick to it, either a call_rcu1() or an
> enforced rcu flush, so that we don't need to reset it. Meanwhile, we don't
> worry on an slightly outdated "n" read because the 2nd round of sem_wait()
> will catch that new "n".
>
> Instead, worst case is rcu thread runs one more round without seeing
> callbacks on the queue.
>
> I'm not sure if that could help simplying code, maybe also make it less
> error-prone.
Semaphore is not applicable here because it will not de-duplicate
concurrent kicks of RCU threads.
>
>>
>> - enter_qs();
>> + enter_qs(sleep);
>> qatomic_sub(&rcu_call_count, n);
>> bql_lock();
>> while (n > 0) {
>> node = try_dequeue();
>> while (!node) {
>
> I have a pure question here not relevant to your changes.. do you know when
> this "if" will trigger? It seems to me the enqueue() should always happen
> before the increment of rcu_call_count:
>
> void call_rcu1(struct rcu_head *node, void (*func)(struct rcu_head *node))
> {
> node->func = func;
> enqueue(node);
>
> if (!qatomic_fetch_inc(&rcu_call_count)) {
> qemu_event_set(&sync_event);
> }
> }>
> I believe qatomic_fetch_inc() is RMW so it's strong mb and order
> guaranteed. Then here why the node can be null even if we're sure >=n have
> been enqueued?
Indeed, enqueue() always happens before the increment of rcu_call_count
performed by the same thread.
The node can still be NULL when there are two concurrent call_rcu1()
executions. In the following example, rcu_call_count will be greater
than the number of visible nodes after (A) and before (B):
Thread T Thread U
call_rcu1(O)
enqueue(O)
Load N from tail
tail = O->next
call_rcu1(P)
enqueue(P)
Load O->next from tail
tail = P
O->next = P
rcu_call_count++ (A)
N->next = O (B)
rcu_call_count++
Regards,
Akihiko Odaki
>
> Thanks,
>
>> bql_unlock();
>> - qemu_event_reset(&rcu_call_ready_event);
>> + qemu_event_reset(&sync_event);
>> node = try_dequeue();
>> if (!node) {
>> - qemu_event_wait(&rcu_call_ready_event);
>> + qemu_event_wait(&sync_event);
>> node = try_dequeue();
>> }
>> bql_lock();
>> @@ -329,8 +350,10 @@ void call_rcu1(struct rcu_head *node, void (*func)(struct rcu_head *node))
>> {
>> node->func = func;
>> enqueue(node);
>> - qatomic_inc(&rcu_call_count);
>> - qemu_event_set(&rcu_call_ready_event);
>> +
>> + if (!qatomic_fetch_inc(&rcu_call_count)) {
>> + qemu_event_set(&sync_event);
>> + }
>> }
>>
>>
>> @@ -388,10 +411,9 @@ void drain_call_rcu(void)
>> * assumed.
>> */
>>
>> - qatomic_inc(&in_drain_call_rcu);
>> call_rcu(&sync, sync_rcu_callback, rcu);
>> + force_rcu();
>> qemu_event_wait(&sync.complete_event);
>> - qatomic_dec(&in_drain_call_rcu);
>>
>> if (locked) {
>> bql_lock();
>> @@ -435,7 +457,7 @@ static void rcu_init_complete(void)
>> qemu_mutex_init(&rcu_registry_lock);
>> qemu_event_init(&rcu_gp_event, true);
>>
>> - qemu_event_init(&rcu_call_ready_event, false);
>> + qemu_event_init(&sync_event, false);
>>
>> /* The caller is assumed to have BQL, so the call_rcu thread
>> * must have been quiescent even after forking, just recreate it.
>>
>> --
>> 2.51.0
>>
>>
>
On Mon, Nov 03, 2025 at 06:45:30PM +0900, Akihiko Odaki wrote:
> On 2025/10/30 3:22, Peter Xu wrote:
> > On Wed, Oct 29, 2025 at 03:12:48PM +0900, Akihiko Odaki wrote:
> > > drain_call_rcu() triggers the force quiescent state, but it can be
> > > delayed if the RCU thread is sleeping. Ensure the force quiescent state
> > > is immediately triggered by waking the RCU thread up.
> > >
> > > The logic to trigger the force quiescent state is decoupled as
> > > force_rcu() so that it can be used independently.
> > >
> > > Signed-off-by: Akihiko Odaki <odaki@rsg.ci.i.u-tokyo.ac.jp>
> > > ---
> > > include/qemu/rcu.h | 1 +
> > > util/rcu.c | 106 ++++++++++++++++++++++++++++++++---------------------
> > > 2 files changed, 65 insertions(+), 42 deletions(-)
> > >
> > > diff --git a/include/qemu/rcu.h b/include/qemu/rcu.h
> > > index 020dbe4d8b77..d6aa4e5854d3 100644
> > > --- a/include/qemu/rcu.h
> > > +++ b/include/qemu/rcu.h
> > > @@ -118,6 +118,7 @@ static inline void rcu_read_unlock(void)
> > > }
> > > }
> > > +void force_rcu(void);
> > > void synchronize_rcu(void);
> > > /*
> > > diff --git a/util/rcu.c b/util/rcu.c
> > > index 3c4af9d213c8..85f9333f5dff 100644
> > > --- a/util/rcu.c
> > > +++ b/util/rcu.c
> > > @@ -49,10 +49,13 @@
> > > unsigned long rcu_gp_ctr = RCU_GP_LOCKED;
> > > QemuEvent rcu_gp_event;
> > > -static int in_drain_call_rcu;
> > > +static bool forced;
> > > static int rcu_call_count;
> > > static QemuMutex rcu_registry_lock;
> > > +/* Set when the forced variable is set or rcu_call_count becomes non-zero. */
> > > +static QemuEvent sync_event;
> > > +
> > > /*
> > > * Check whether a quiescent state was crossed between the beginning of
> > > * update_counter_and_wait and now.
> > > @@ -74,36 +77,21 @@ QEMU_DEFINE_CO_TLS(struct rcu_reader_data, rcu_reader)
> > > typedef QLIST_HEAD(, rcu_reader_data) ThreadList;
> > > static ThreadList registry = QLIST_HEAD_INITIALIZER(registry);
> > > +void force_rcu(void)
> > > +{
> > > + qatomic_set(&forced, true);
> > > + qemu_event_set(&sync_event);
> > > +}
> > > +
> > > /* Wait for previous parity/grace period to be empty of readers. */
> > > -static void wait_for_readers(void)
> > > +static void wait_for_readers(bool sleep)
> > > {
> > > ThreadList qsreaders = QLIST_HEAD_INITIALIZER(qsreaders);
> > > struct rcu_reader_data *index, *tmp;
> > > - int sleeps = 0;
> > > - bool forced = false;
> > > + int sleeps = sleep ? 5 : 0;
> > > + bool waiting = false;
> > > for (;;) {
> > > - /*
> > > - * Force the grace period to end and wait for it if any of the
> > > - * following heuristical conditions are satisfied:
> > > - * - A decent number of callbacks piled up.
> > > - * - It timed out.
> > > - * - It is in a drain_call_rcu() call.
> > > - *
> > > - * Otherwise, periodically poll the grace period, hoping it ends
> > > - * promptly.
> > > - */
> > > - if (!forced &&
> > > - (qatomic_read(&rcu_call_count) >= RCU_CALL_MIN_SIZE ||
> > > - sleeps >= 5 || qatomic_read(&in_drain_call_rcu))) {
> > > - forced = true;
> > > -
> > > - QLIST_FOREACH(index, ®istry, node) {
> > > - notifier_list_notify(&index->force_rcu, NULL);
> > > - qatomic_set(&index->waiting, true);
> > > - }
> > > - }
> >
> > IIUC this is the part to set index->waiting first whenever necessary, then
> > that'll guarantee the wait(rcu_gp_event) will be notified in rcu unlock.
> >
> > Now we removed this chunk, then could it happen if waiting=true and the
> > wait(rcu_gp_event) may wait for more than it should (as nobody will wake it
> > up if all threads have waiting=false)?
>
> It is not removed, but it is moved along with the assignment of the waiting
> variable. So index->waiting is still set whenever the waiting variable is
> set and no hang up will happen.
Ah, I noticed the "waiting" is actually the global variable and I think
when I read it last time I somehow misread it with "sleep" (which was
passed down from the caller).
if (waiting) { <--------------------
qemu_event_wait(&rcu_gp_event);
qemu_event_reset(&rcu_gp_event);
} else if (qatomic_read(&rcu_call_count) >= RCU_CALL_MIN_SIZE ||
Yeah, I think it's non-issue. Please ignore my question.
>
> >
> > The other thing is, right below here there's the code and comment:
> >
> > /* Here, order the stores to index->waiting before the loads of
> > * index->ctr. Pairs with smp_mb_placeholder() in rcu_read_unlock(),
> > * ensuring that the loads of index->ctr are sequentially consistent.
> > *
> > * If this is the last iteration, this barrier also prevents
> > * frees from seeping upwards, and orders the two wait phases
> > * on architectures with 32-bit longs; see enter_qs().
> > */
> > smp_mb_global();
> >
> > IIUC it explains the mb_global() to order the updates of waiting and the
> > reads of index->ctr. It doesn't look like applicable anymore. Said that,
> > I think we should indeed still need some barrier to make sure we read
> > index->ctr at least to be after we update global gp_ctr (done before
> > calling wait_for_readers()). I'm not sure if it means the mb is needed,
> > however maybe at least the comment is outdated if so.
>
> It is still applicable. The stores to index->waiting is still present and
> needs to be ordered before the loads of index->ctr.
>
> >
> > > -
> > > /* Here, order the stores to index->waiting before the loads of
> > > * index->ctr. Pairs with smp_mb_placeholder() in rcu_read_unlock(),
> > > * ensuring that the loads of index->ctr are sequentially consistent.
> > > @@ -150,7 +138,8 @@ static void wait_for_readers(void)
> > > */
> > > qemu_mutex_unlock(&rcu_registry_lock);
> > > - if (forced) {
> > > + if (waiting) {
> > > + /* Wait for the forced quiescent state. */
> > > qemu_event_wait(&rcu_gp_event);
> > > /*
> > > @@ -158,9 +147,25 @@ static void wait_for_readers(void)
> > > * while we walk the list.
> > > */
> > > qemu_event_reset(&rcu_gp_event);
> > > + } else if (qatomic_read(&rcu_call_count) >= RCU_CALL_MIN_SIZE ||
> > > + !sleeps || qemu_event_timedwait(&sync_event, 10)) {
> > > + /*
> > > + * Now one of the following heuristical conditions is satisfied:
> > > + * - A decent number of callbacks piled up.
> > > + * - It timed out.
> > > + * - force_rcu() was called.
> > > + *
> > > + * Force a quiescent state.
> > > + */
> > > + waiting = true;
> > > +
> > > + QLIST_FOREACH(index, ®istry, node) {
> > > + notifier_list_notify(&index->force_rcu, NULL);
> > > + qatomic_set(&index->waiting, true);
> > > + }
> > > } else {
> > > - g_usleep(10000);
> > > - sleeps++;
> > > + /* Try again. */
> > > + sleeps--;
> > > }
> > > qemu_mutex_lock(&rcu_registry_lock);
> > > @@ -170,7 +175,7 @@ static void wait_for_readers(void)
> > > QLIST_SWAP(®istry, &qsreaders, node);
> > > }
> > > -static void enter_qs(void)
> > > +static void enter_qs(bool sleep)
> > > {
> > > /* Write RCU-protected pointers before reading p_rcu_reader->ctr.
> > > * Pairs with smp_mb_placeholder() in rcu_read_lock().
> > > @@ -189,14 +194,14 @@ static void enter_qs(void)
> > > * Switch parity: 0 -> 1, 1 -> 0.
> > > */
> > > qatomic_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
> > > - wait_for_readers();
> > > + wait_for_readers(sleep);
> > > qatomic_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
> > > } else {
> > > /* Increment current grace period. */
> > > qatomic_set(&rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
> > > }
> > > - wait_for_readers();
> > > + wait_for_readers(sleep);
> > > }
> > > }
> > > @@ -205,7 +210,6 @@ static void enter_qs(void)
> > > */
> > > static struct rcu_head dummy;
> > > static struct rcu_head *head = &dummy, **tail = &dummy.next;
> > > -static QemuEvent rcu_call_ready_event;
> > > static void enqueue(struct rcu_head *node)
> > > {
> > > @@ -282,6 +286,7 @@ static void *call_rcu_thread(void *opaque)
> > > rcu_register_thread();
> > > for (;;) {
> > > + bool sleep = true;
> > > int n;
> > > /*
> > > @@ -289,7 +294,7 @@ static void *call_rcu_thread(void *opaque)
> > > * added before enter_qs() starts.
> > > */
> > > for (;;) {
> > > - qemu_event_reset(&rcu_call_ready_event);
> > > + qemu_event_reset(&sync_event);
> > > n = qatomic_read(&rcu_call_count);
> > > if (n) {
> > > break;
> > > @@ -298,20 +303,36 @@ static void *call_rcu_thread(void *opaque)
> > > #if defined(CONFIG_MALLOC_TRIM)
> > > malloc_trim(4 * 1024 * 1024);
> > > #endif
> > > - qemu_event_wait(&rcu_call_ready_event);
> > > + qemu_event_wait(&sync_event);
> > > + }
> > > +
> > > + /*
> > > + * Ensure that an event for a rcu_call_count change will not interrupt
> > > + * wait_for_readers().
> > > + */
> > > + qemu_event_reset(&sync_event);
> > > +
> > > + /*
> > > + * Ensure that the forced variable has not been set after fetching
> > > + * rcu_call_count; otherwise we may get confused by a force quiescent
> > > + * state request for an element later than n.
> > > + */
> > > + while (qatomic_xchg(&forced, false)) {
> > > + sleep = false;
> > > + n = qatomic_read(&rcu_call_count);
> > > }
> >
> > This is pretty tricky, and I wonder if it will make the code easier to read
> > if we convert the sync_event to be a semaphore instead. When as a sem, it
> > will take account of whatever kick to it, either a call_rcu1() or an
> > enforced rcu flush, so that we don't need to reset it. Meanwhile, we don't
> > worry on an slightly outdated "n" read because the 2nd round of sem_wait()
> > will catch that new "n".
> >
> > Instead, worst case is rcu thread runs one more round without seeing
> > callbacks on the queue.
> >
> > I'm not sure if that could help simplying code, maybe also make it less
> > error-prone.
>
> Semaphore is not applicable here because it will not de-duplicate concurrent
> kicks of RCU threads.
Why concurrent kicks of rcu threads is a problem? QemuSemaphore is
thread-safe itself, meanwhile IIUC it only still causes call_rcu_thread()
loops some more rounds reading "n", which looks all safe. No?
>
> >
> > > - enter_qs();
> > > + enter_qs(sleep);
> > > qatomic_sub(&rcu_call_count, n);
> > > bql_lock();
> > > while (n > 0) {
> > > node = try_dequeue();
> > > while (!node) {
> >
> > I have a pure question here not relevant to your changes.. do you know when
> > this "if" will trigger? It seems to me the enqueue() should always happen
> > before the increment of rcu_call_count:
> >
> > void call_rcu1(struct rcu_head *node, void (*func)(struct rcu_head *node))
> > {
> > node->func = func;
> > enqueue(node);
> >
> > if (!qatomic_fetch_inc(&rcu_call_count)) {
> > qemu_event_set(&sync_event);
> > }
> > }>
> > I believe qatomic_fetch_inc() is RMW so it's strong mb and order
> > guaranteed. Then here why the node can be null even if we're sure >=n have
> > been enqueued?
>
> Indeed, enqueue() always happens before the increment of rcu_call_count
> performed by the same thread.
>
> The node can still be NULL when there are two concurrent call_rcu1()
> executions. In the following example, rcu_call_count will be greater than
> the number of visible nodes after (A) and before (B):
>
> Thread T Thread U
> call_rcu1(O)
> enqueue(O)
> Load N from tail
> tail = O->next
> call_rcu1(P)
> enqueue(P)
> Load O->next from tail
> tail = P
> O->next = P
> rcu_call_count++ (A)
> N->next = O (B)
> rcu_call_count++
Thanks, yeah it makes sense. If you think worthwhile, maybe we could add a
comment after the first try_dequeue().
--
Peter Xu
On 2025/11/06 5:43, Peter Xu wrote:
> On Mon, Nov 03, 2025 at 06:45:30PM +0900, Akihiko Odaki wrote:
>> On 2025/10/30 3:22, Peter Xu wrote:
>>> On Wed, Oct 29, 2025 at 03:12:48PM +0900, Akihiko Odaki wrote:
>>>> drain_call_rcu() triggers the force quiescent state, but it can be
>>>> delayed if the RCU thread is sleeping. Ensure the force quiescent state
>>>> is immediately triggered by waking the RCU thread up.
>>>>
>>>> The logic to trigger the force quiescent state is decoupled as
>>>> force_rcu() so that it can be used independently.
>>>>
>>>> Signed-off-by: Akihiko Odaki <odaki@rsg.ci.i.u-tokyo.ac.jp>
>>>> ---
>>>> include/qemu/rcu.h | 1 +
>>>> util/rcu.c | 106 ++++++++++++++++++++++++++++++++---------------------
>>>> 2 files changed, 65 insertions(+), 42 deletions(-)
>>>>
>>>> diff --git a/include/qemu/rcu.h b/include/qemu/rcu.h
>>>> index 020dbe4d8b77..d6aa4e5854d3 100644
>>>> --- a/include/qemu/rcu.h
>>>> +++ b/include/qemu/rcu.h
>>>> @@ -118,6 +118,7 @@ static inline void rcu_read_unlock(void)
>>>> }
>>>> }
>>>> +void force_rcu(void);
>>>> void synchronize_rcu(void);
>>>> /*
>>>> diff --git a/util/rcu.c b/util/rcu.c
>>>> index 3c4af9d213c8..85f9333f5dff 100644
>>>> --- a/util/rcu.c
>>>> +++ b/util/rcu.c
>>>> @@ -49,10 +49,13 @@
>>>> unsigned long rcu_gp_ctr = RCU_GP_LOCKED;
>>>> QemuEvent rcu_gp_event;
>>>> -static int in_drain_call_rcu;
>>>> +static bool forced;
>>>> static int rcu_call_count;
>>>> static QemuMutex rcu_registry_lock;
>>>> +/* Set when the forced variable is set or rcu_call_count becomes non-zero. */
>>>> +static QemuEvent sync_event;
>>>> +
>>>> /*
>>>> * Check whether a quiescent state was crossed between the beginning of
>>>> * update_counter_and_wait and now.
>>>> @@ -74,36 +77,21 @@ QEMU_DEFINE_CO_TLS(struct rcu_reader_data, rcu_reader)
>>>> typedef QLIST_HEAD(, rcu_reader_data) ThreadList;
>>>> static ThreadList registry = QLIST_HEAD_INITIALIZER(registry);
>>>> +void force_rcu(void)
>>>> +{
>>>> + qatomic_set(&forced, true);
>>>> + qemu_event_set(&sync_event);
>>>> +}
>>>> +
>>>> /* Wait for previous parity/grace period to be empty of readers. */
>>>> -static void wait_for_readers(void)
>>>> +static void wait_for_readers(bool sleep)
>>>> {
>>>> ThreadList qsreaders = QLIST_HEAD_INITIALIZER(qsreaders);
>>>> struct rcu_reader_data *index, *tmp;
>>>> - int sleeps = 0;
>>>> - bool forced = false;
>>>> + int sleeps = sleep ? 5 : 0;
>>>> + bool waiting = false;
>>>> for (;;) {
>>>> - /*
>>>> - * Force the grace period to end and wait for it if any of the
>>>> - * following heuristical conditions are satisfied:
>>>> - * - A decent number of callbacks piled up.
>>>> - * - It timed out.
>>>> - * - It is in a drain_call_rcu() call.
>>>> - *
>>>> - * Otherwise, periodically poll the grace period, hoping it ends
>>>> - * promptly.
>>>> - */
>>>> - if (!forced &&
>>>> - (qatomic_read(&rcu_call_count) >= RCU_CALL_MIN_SIZE ||
>>>> - sleeps >= 5 || qatomic_read(&in_drain_call_rcu))) {
>>>> - forced = true;
>>>> -
>>>> - QLIST_FOREACH(index, ®istry, node) {
>>>> - notifier_list_notify(&index->force_rcu, NULL);
>>>> - qatomic_set(&index->waiting, true);
>>>> - }
>>>> - }
>>>
>>> IIUC this is the part to set index->waiting first whenever necessary, then
>>> that'll guarantee the wait(rcu_gp_event) will be notified in rcu unlock.
>>>
>>> Now we removed this chunk, then could it happen if waiting=true and the
>>> wait(rcu_gp_event) may wait for more than it should (as nobody will wake it
>>> up if all threads have waiting=false)?
>>
>> It is not removed, but it is moved along with the assignment of the waiting
>> variable. So index->waiting is still set whenever the waiting variable is
>> set and no hang up will happen.
>
> Ah, I noticed the "waiting" is actually the global variable and I think
> when I read it last time I somehow misread it with "sleep" (which was
> passed down from the caller).
waiting is local. There are several variables and some are global and
the others are local so please be careful.
>
> if (waiting) { <--------------------
> qemu_event_wait(&rcu_gp_event);
> qemu_event_reset(&rcu_gp_event);
> } else if (qatomic_read(&rcu_call_count) >= RCU_CALL_MIN_SIZE ||
>
> Yeah, I think it's non-issue. Please ignore my question.
>
>>
>>>
>>> The other thing is, right below here there's the code and comment:
>>>
>>> /* Here, order the stores to index->waiting before the loads of
>>> * index->ctr. Pairs with smp_mb_placeholder() in rcu_read_unlock(),
>>> * ensuring that the loads of index->ctr are sequentially consistent.
>>> *
>>> * If this is the last iteration, this barrier also prevents
>>> * frees from seeping upwards, and orders the two wait phases
>>> * on architectures with 32-bit longs; see enter_qs().
>>> */
>>> smp_mb_global();
>>>
>>> IIUC it explains the mb_global() to order the updates of waiting and the
>>> reads of index->ctr. It doesn't look like applicable anymore. Said that,
>>> I think we should indeed still need some barrier to make sure we read
>>> index->ctr at least to be after we update global gp_ctr (done before
>>> calling wait_for_readers()). I'm not sure if it means the mb is needed,
>>> however maybe at least the comment is outdated if so.
>>
>> It is still applicable. The stores to index->waiting is still present and
>> needs to be ordered before the loads of index->ctr.
>>
>>>
>>>> -
>>>> /* Here, order the stores to index->waiting before the loads of
>>>> * index->ctr. Pairs with smp_mb_placeholder() in rcu_read_unlock(),
>>>> * ensuring that the loads of index->ctr are sequentially consistent.
>>>> @@ -150,7 +138,8 @@ static void wait_for_readers(void)
>>>> */
>>>> qemu_mutex_unlock(&rcu_registry_lock);
>>>> - if (forced) {
>>>> + if (waiting) {
>>>> + /* Wait for the forced quiescent state. */
>>>> qemu_event_wait(&rcu_gp_event);
>>>> /*
>>>> @@ -158,9 +147,25 @@ static void wait_for_readers(void)
>>>> * while we walk the list.
>>>> */
>>>> qemu_event_reset(&rcu_gp_event);
>>>> + } else if (qatomic_read(&rcu_call_count) >= RCU_CALL_MIN_SIZE ||
>>>> + !sleeps || qemu_event_timedwait(&sync_event, 10)) {
>>>> + /*
>>>> + * Now one of the following heuristical conditions is satisfied:
>>>> + * - A decent number of callbacks piled up.
>>>> + * - It timed out.
>>>> + * - force_rcu() was called.
>>>> + *
>>>> + * Force a quiescent state.
>>>> + */
>>>> + waiting = true;
>>>> +
>>>> + QLIST_FOREACH(index, ®istry, node) {
>>>> + notifier_list_notify(&index->force_rcu, NULL);
>>>> + qatomic_set(&index->waiting, true);
>>>> + }
>>>> } else {
>>>> - g_usleep(10000);
>>>> - sleeps++;
>>>> + /* Try again. */
>>>> + sleeps--;
>>>> }
>>>> qemu_mutex_lock(&rcu_registry_lock);
>>>> @@ -170,7 +175,7 @@ static void wait_for_readers(void)
>>>> QLIST_SWAP(®istry, &qsreaders, node);
>>>> }
>>>> -static void enter_qs(void)
>>>> +static void enter_qs(bool sleep)
>>>> {
>>>> /* Write RCU-protected pointers before reading p_rcu_reader->ctr.
>>>> * Pairs with smp_mb_placeholder() in rcu_read_lock().
>>>> @@ -189,14 +194,14 @@ static void enter_qs(void)
>>>> * Switch parity: 0 -> 1, 1 -> 0.
>>>> */
>>>> qatomic_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
>>>> - wait_for_readers();
>>>> + wait_for_readers(sleep);
>>>> qatomic_set(&rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
>>>> } else {
>>>> /* Increment current grace period. */
>>>> qatomic_set(&rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
>>>> }
>>>> - wait_for_readers();
>>>> + wait_for_readers(sleep);
>>>> }
>>>> }
>>>> @@ -205,7 +210,6 @@ static void enter_qs(void)
>>>> */
>>>> static struct rcu_head dummy;
>>>> static struct rcu_head *head = &dummy, **tail = &dummy.next;
>>>> -static QemuEvent rcu_call_ready_event;
>>>> static void enqueue(struct rcu_head *node)
>>>> {
>>>> @@ -282,6 +286,7 @@ static void *call_rcu_thread(void *opaque)
>>>> rcu_register_thread();
>>>> for (;;) {
>>>> + bool sleep = true;
>>>> int n;
>>>> /*
>>>> @@ -289,7 +294,7 @@ static void *call_rcu_thread(void *opaque)
>>>> * added before enter_qs() starts.
>>>> */
>>>> for (;;) {
>>>> - qemu_event_reset(&rcu_call_ready_event);
>>>> + qemu_event_reset(&sync_event);
>>>> n = qatomic_read(&rcu_call_count);
>>>> if (n) {
>>>> break;
>>>> @@ -298,20 +303,36 @@ static void *call_rcu_thread(void *opaque)
>>>> #if defined(CONFIG_MALLOC_TRIM)
>>>> malloc_trim(4 * 1024 * 1024);
>>>> #endif
>>>> - qemu_event_wait(&rcu_call_ready_event);
>>>> + qemu_event_wait(&sync_event);
>>>> + }
>>>> +
>>>> + /*
>>>> + * Ensure that an event for a rcu_call_count change will not interrupt
>>>> + * wait_for_readers().
>>>> + */
>>>> + qemu_event_reset(&sync_event);
>>>> +
>>>> + /*
>>>> + * Ensure that the forced variable has not been set after fetching
>>>> + * rcu_call_count; otherwise we may get confused by a force quiescent
>>>> + * state request for an element later than n.
>>>> + */
>>>> + while (qatomic_xchg(&forced, false)) {
>>>> + sleep = false;
>>>> + n = qatomic_read(&rcu_call_count);
>>>> }
>>>
>>> This is pretty tricky, and I wonder if it will make the code easier to read
>>> if we convert the sync_event to be a semaphore instead. When as a sem, it
>>> will take account of whatever kick to it, either a call_rcu1() or an
>>> enforced rcu flush, so that we don't need to reset it. Meanwhile, we don't
>>> worry on an slightly outdated "n" read because the 2nd round of sem_wait()
>>> will catch that new "n".
>>>
>>> Instead, worst case is rcu thread runs one more round without seeing
>>> callbacks on the queue.
>>>
>>> I'm not sure if that could help simplying code, maybe also make it less
>>> error-prone.
>>
>> Semaphore is not applicable here because it will not de-duplicate concurrent
>> kicks of RCU threads.
>
> Why concurrent kicks of rcu threads is a problem? QemuSemaphore is
> thread-safe itself, meanwhile IIUC it only still causes call_rcu_thread()
> loops some more rounds reading "n", which looks all safe. No?
It is safe but incurs overheads and confusing. QemuEvent represents the
boolean semantics better.
I also have difficulty to understand how converting sync_event to a
semaphore simplifies the code. Perhaps some (pseudo)code to show how the
code will look like may be useful.
>
>>
>>>
>>>> - enter_qs();
>>>> + enter_qs(sleep);
>>>> qatomic_sub(&rcu_call_count, n);
>>>> bql_lock();
>>>> while (n > 0) {
>>>> node = try_dequeue();
>>>> while (!node) {
>>>
>>> I have a pure question here not relevant to your changes.. do you know when
>>> this "if" will trigger? It seems to me the enqueue() should always happen
>>> before the increment of rcu_call_count:
>>>
>>> void call_rcu1(struct rcu_head *node, void (*func)(struct rcu_head *node))
>>> {
>>> node->func = func;
>>> enqueue(node);
>>>
>>> if (!qatomic_fetch_inc(&rcu_call_count)) {
>>> qemu_event_set(&sync_event);
>>> }
>>> }>
>>> I believe qatomic_fetch_inc() is RMW so it's strong mb and order
>>> guaranteed. Then here why the node can be null even if we're sure >=n have
>>> been enqueued?
>>
>> Indeed, enqueue() always happens before the increment of rcu_call_count
>> performed by the same thread.
>>
>> The node can still be NULL when there are two concurrent call_rcu1()
>> executions. In the following example, rcu_call_count will be greater than
>> the number of visible nodes after (A) and before (B):
>>
>> Thread T Thread U
>> call_rcu1(O)
>> enqueue(O)
>> Load N from tail
>> tail = O->next
>> call_rcu1(P)
>> enqueue(P)
>> Load O->next from tail
>> tail = P
>> O->next = P
>> rcu_call_count++ (A)
>> N->next = O (B)
>> rcu_call_count++
>
> Thanks, yeah it makes sense. If you think worthwhile, maybe we could add a
> comment after the first try_dequeue().
>
try_dequeue() and enqueue() do have comments that say that enqueue() is
asynchronous and explain why, but perhaps they can be expanded by adding
the example I showed.
Besides enqueue() may be renamed like enqueue_async() so that it will be
clear that enqueuing does not happen immediately when you look at
call_rcu1(). If you wonder enqueue_async() is asynchronous, you will
look into enqueue_async() and find the explanatory comments.
Regards,
Akihiko Odaki
On Thu, Nov 06, 2025 at 10:40:52AM +0900, Akihiko Odaki wrote:
> > > > > + /*
> > > > > + * Ensure that the forced variable has not been set after fetching
> > > > > + * rcu_call_count; otherwise we may get confused by a force quiescent
> > > > > + * state request for an element later than n.
> > > > > + */
> > > > > + while (qatomic_xchg(&forced, false)) {
> > > > > + sleep = false;
> > > > > + n = qatomic_read(&rcu_call_count);
> > > > > }
> > > >
> > > > This is pretty tricky, and I wonder if it will make the code easier to read
> > > > if we convert the sync_event to be a semaphore instead. When as a sem, it
> > > > will take account of whatever kick to it, either a call_rcu1() or an
> > > > enforced rcu flush, so that we don't need to reset it. Meanwhile, we don't
> > > > worry on an slightly outdated "n" read because the 2nd round of sem_wait()
> > > > will catch that new "n".
> > > >
> > > > Instead, worst case is rcu thread runs one more round without seeing
> > > > callbacks on the queue.
> > > >
> > > > I'm not sure if that could help simplying code, maybe also make it less
> > > > error-prone.
> > >
> > > Semaphore is not applicable here because it will not de-duplicate concurrent
> > > kicks of RCU threads.
> >
> > Why concurrent kicks of rcu threads is a problem? QemuSemaphore is
> > thread-safe itself, meanwhile IIUC it only still causes call_rcu_thread()
> > loops some more rounds reading "n", which looks all safe. No?
>
> It is safe but incurs overheads and confusing. QemuEvent represents the
> boolean semantics better.
>
> I also have difficulty to understand how converting sync_event to a
> semaphore simplifies the code. Perhaps some (pseudo)code to show how the
> code will look like may be useful.
I prepared a patch on top of your current patchset to show what I meant. I
also added comments and some test results showing why I think it might be
fine that the sem overhead should be small.
In short, I tested a VM with 8 vCPUs and 4G mem, booting Linux and properly
poweroff, I only saw <1000 rcu_call1 users in total. That should be the
max-bound of sem overhead on looping in rcu thread.
It's in patch format but still treat it as a comment instead to discuss
with you. Attaching it is just easier for me.
===8<===
From 71f15ed19050a973088352a8d71b6cc6b7b5f7cf Mon Sep 17 00:00:00 2001
From: Peter Xu <peterx@redhat.com>
Date: Thu, 6 Nov 2025 16:03:00 -0500
Subject: [PATCH] rcu: Make sync_event a semaphore
It could simply all reset logic, especially after enforced rcu is
introduced we'll also need a tweak to re-read "n", which can be avoided too
when with a sem.
However, the sem can introduce an overhead in high frequecy rcu frees.
This patch is drafted with the assumption that rcu free is at least very
rare in QEMU, hence it's not a problem.
When I tested with this command:
qemu-system-x86_64 -M q35,kernel-irqchip=split,suppress-vmdesc=on -smp 8 \
-m 4G -msg timestamp=on -name peter-vm,debug-threads=on -cpu Nehalem \
-accel kvm -qmp unix:/tmp/peter.sock,server,nowait -nographic \
-monitor telnet::6666,server,nowait -netdev user,id=net0,hostfwd=tcp::5555-:22
-device e1000,netdev=net0 -device virtio-balloon $DISK
I booted a pre-installed Linux, login and poweroff, wait until VM
completely shutdowns. I captured less than 1000 rcu_free1() calls in
summary. It means for the whole lifetime of such VM the max overhead of
the call_rcu_thread() loop reading rcu_call_count will be 1000 loops.
Signed-off-by: Peter Xu <peterx@redhat.com>
---
util/rcu.c | 36 ++++++++----------------------------
1 file changed, 8 insertions(+), 28 deletions(-)
diff --git a/util/rcu.c b/util/rcu.c
index 85f9333f5d..dfe031a5c9 100644
--- a/util/rcu.c
+++ b/util/rcu.c
@@ -54,7 +54,7 @@ static int rcu_call_count;
static QemuMutex rcu_registry_lock;
/* Set when the forced variable is set or rcu_call_count becomes non-zero. */
-static QemuEvent sync_event;
+static QemuSemaphore sync_event;
/*
* Check whether a quiescent state was crossed between the beginning of
@@ -80,7 +80,7 @@ static ThreadList registry = QLIST_HEAD_INITIALIZER(registry);
void force_rcu(void)
{
qatomic_set(&forced, true);
- qemu_event_set(&sync_event);
+ qemu_sem_post(&sync_event);
}
/* Wait for previous parity/grace period to be empty of readers. */
@@ -148,7 +148,7 @@ static void wait_for_readers(bool sleep)
*/
qemu_event_reset(&rcu_gp_event);
} else if (qatomic_read(&rcu_call_count) >= RCU_CALL_MIN_SIZE ||
- !sleeps || qemu_event_timedwait(&sync_event, 10)) {
+ !sleeps || qemu_sem_timedwait(&sync_event, 10)) {
/*
* Now one of the following heuristical conditions is satisfied:
* - A decent number of callbacks piled up.
@@ -286,7 +286,6 @@ static void *call_rcu_thread(void *opaque)
rcu_register_thread();
for (;;) {
- bool sleep = true;
int n;
/*
@@ -294,7 +293,6 @@ static void *call_rcu_thread(void *opaque)
* added before enter_qs() starts.
*/
for (;;) {
- qemu_event_reset(&sync_event);
n = qatomic_read(&rcu_call_count);
if (n) {
break;
@@ -303,36 +301,19 @@ static void *call_rcu_thread(void *opaque)
#if defined(CONFIG_MALLOC_TRIM)
malloc_trim(4 * 1024 * 1024);
#endif
- qemu_event_wait(&sync_event);
+ qemu_sem_wait(&sync_event);
}
- /*
- * Ensure that an event for a rcu_call_count change will not interrupt
- * wait_for_readers().
- */
- qemu_event_reset(&sync_event);
-
- /*
- * Ensure that the forced variable has not been set after fetching
- * rcu_call_count; otherwise we may get confused by a force quiescent
- * state request for an element later than n.
- */
- while (qatomic_xchg(&forced, false)) {
- sleep = false;
- n = qatomic_read(&rcu_call_count);
- }
-
- enter_qs(sleep);
+ enter_qs(!qatomic_xchg(&forced, false));
qatomic_sub(&rcu_call_count, n);
bql_lock();
while (n > 0) {
node = try_dequeue();
while (!node) {
bql_unlock();
- qemu_event_reset(&sync_event);
node = try_dequeue();
if (!node) {
- qemu_event_wait(&sync_event);
+ qemu_sem_wait(&sync_event);
node = try_dequeue();
}
bql_lock();
@@ -352,7 +333,7 @@ void call_rcu1(struct rcu_head *node, void (*func)(struct rcu_head *node))
enqueue(node);
if (!qatomic_fetch_inc(&rcu_call_count)) {
- qemu_event_set(&sync_event);
+ qemu_sem_post(&sync_event);
}
}
@@ -456,8 +437,7 @@ static void rcu_init_complete(void)
qemu_mutex_init(&rcu_registry_lock);
qemu_event_init(&rcu_gp_event, true);
-
- qemu_event_init(&sync_event, false);
+ qemu_sem_init(&sync_event, 0);
/* The caller is assumed to have BQL, so the call_rcu thread
* must have been quiescent even after forking, just recreate it.
--
2.50.1
===8<===
When I was having a closer look, I found some other issues, I'll list it
all here.
1. I found that rcu_gp_event was initialized as "true". I'm not sure
whether it should be false. This has nothing to do with your series as
well, only some observation of mine.
qemu_event_init(&rcu_gp_event, true);
2. It looks to me your patch here checked the wrong retval of
qemu_event_timedwait()..
} else if (qatomic_read(&rcu_call_count) >= RCU_CALL_MIN_SIZE ||
!sleeps || qemu_event_timedwait(&sync_event, 10)) {
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
should it be "!qemu_event_timedwait(&sync_event, 10)"? Side note: maybe
we should cleanup all _timedwait() for QEMU's eventfd, sem, cond,
... because they don't return the same retval.. but if you think sem is
good, then we don't need eventfd's timedwait() in this series (your
initial two patches).
3. I doubt if malloc_trim() feature is broken with your current patchset,
because now the loop looks like:
for (;;) {
qemu_event_reset(&sync_event);
n = qatomic_read(&rcu_call_count);
if (n) {
break;
}
#if defined(CONFIG_MALLOC_TRIM)
malloc_trim(4 * 1024 * 1024);
#endif
qemu_event_wait(&sync_event);
}
I don't know if n can be zero here at all.. if you use the sem change
this might trigger but it's not designed for it. When using sem we may
want to periodically trim. But I'd like to know how you think in general
on the sem idea first (e.g. do we need to be prepared for high freq rcu
frees).
Thanks,
--
Peter Xu
On 2025/11/07 6:52, Peter Xu wrote:
> On Thu, Nov 06, 2025 at 10:40:52AM +0900, Akihiko Odaki wrote:
>>>>>> + /*
>>>>>> + * Ensure that the forced variable has not been set after fetching
>>>>>> + * rcu_call_count; otherwise we may get confused by a force quiescent
>>>>>> + * state request for an element later than n.
>>>>>> + */
>>>>>> + while (qatomic_xchg(&forced, false)) {
>>>>>> + sleep = false;
>>>>>> + n = qatomic_read(&rcu_call_count);
>>>>>> }
>>>>>
>>>>> This is pretty tricky, and I wonder if it will make the code easier to read
>>>>> if we convert the sync_event to be a semaphore instead. When as a sem, it
>>>>> will take account of whatever kick to it, either a call_rcu1() or an
>>>>> enforced rcu flush, so that we don't need to reset it. Meanwhile, we don't
>>>>> worry on an slightly outdated "n" read because the 2nd round of sem_wait()
>>>>> will catch that new "n".
>>>>>
>>>>> Instead, worst case is rcu thread runs one more round without seeing
>>>>> callbacks on the queue.
>>>>>
>>>>> I'm not sure if that could help simplying code, maybe also make it less
>>>>> error-prone.
>>>>
>>>> Semaphore is not applicable here because it will not de-duplicate concurrent
>>>> kicks of RCU threads.
>>>
>>> Why concurrent kicks of rcu threads is a problem? QemuSemaphore is
>>> thread-safe itself, meanwhile IIUC it only still causes call_rcu_thread()
>>> loops some more rounds reading "n", which looks all safe. No?
>>
>> It is safe but incurs overheads and confusing. QemuEvent represents the
>> boolean semantics better.
>>
>> I also have difficulty to understand how converting sync_event to a
>> semaphore simplifies the code. Perhaps some (pseudo)code to show how the
>> code will look like may be useful.
>
> I prepared a patch on top of your current patchset to show what I meant. I
> also added comments and some test results showing why I think it might be
> fine that the sem overhead should be small.
>
> In short, I tested a VM with 8 vCPUs and 4G mem, booting Linux and properly
> poweroff, I only saw <1000 rcu_call1 users in total. That should be the
> max-bound of sem overhead on looping in rcu thread.
>
> It's in patch format but still treat it as a comment instead to discuss
> with you. Attaching it is just easier for me.
>
> ===8<===
> From 71f15ed19050a973088352a8d71b6cc6b7b5f7cf Mon Sep 17 00:00:00 2001
> From: Peter Xu <peterx@redhat.com>
> Date: Thu, 6 Nov 2025 16:03:00 -0500
> Subject: [PATCH] rcu: Make sync_event a semaphore
>
> It could simply all reset logic, especially after enforced rcu is
> introduced we'll also need a tweak to re-read "n", which can be avoided too
> when with a sem.
>
> However, the sem can introduce an overhead in high frequecy rcu frees.
> This patch is drafted with the assumption that rcu free is at least very
> rare in QEMU, hence it's not a problem.
>
> When I tested with this command:
>
> qemu-system-x86_64 -M q35,kernel-irqchip=split,suppress-vmdesc=on -smp 8 \
> -m 4G -msg timestamp=on -name peter-vm,debug-threads=on -cpu Nehalem \
> -accel kvm -qmp unix:/tmp/peter.sock,server,nowait -nographic \
> -monitor telnet::6666,server,nowait -netdev user,id=net0,hostfwd=tcp::5555-:22
> -device e1000,netdev=net0 -device virtio-balloon $DISK
>
> I booted a pre-installed Linux, login and poweroff, wait until VM
> completely shutdowns. I captured less than 1000 rcu_free1() calls in
> summary. It means for the whole lifetime of such VM the max overhead of
> the call_rcu_thread() loop reading rcu_call_count will be 1000 loops.
>
> Signed-off-by: Peter Xu <peterx@redhat.com>
> ---
> util/rcu.c | 36 ++++++++----------------------------
> 1 file changed, 8 insertions(+), 28 deletions(-)
>
> diff --git a/util/rcu.c b/util/rcu.c
> index 85f9333f5d..dfe031a5c9 100644
> --- a/util/rcu.c
> +++ b/util/rcu.c
> @@ -54,7 +54,7 @@ static int rcu_call_count;
> static QemuMutex rcu_registry_lock;
>
> /* Set when the forced variable is set or rcu_call_count becomes non-zero. */
> -static QemuEvent sync_event;
> +static QemuSemaphore sync_event;
>
> /*
> * Check whether a quiescent state was crossed between the beginning of
> @@ -80,7 +80,7 @@ static ThreadList registry = QLIST_HEAD_INITIALIZER(registry);
> void force_rcu(void)
> {
> qatomic_set(&forced, true);
> - qemu_event_set(&sync_event);
> + qemu_sem_post(&sync_event);
> }
>
> /* Wait for previous parity/grace period to be empty of readers. */
> @@ -148,7 +148,7 @@ static void wait_for_readers(bool sleep)
> */
> qemu_event_reset(&rcu_gp_event);
> } else if (qatomic_read(&rcu_call_count) >= RCU_CALL_MIN_SIZE ||
> - !sleeps || qemu_event_timedwait(&sync_event, 10)) {
> + !sleeps || qemu_sem_timedwait(&sync_event, 10)) {
> /*
> * Now one of the following heuristical conditions is satisfied:
> * - A decent number of callbacks piled up.
> @@ -286,7 +286,6 @@ static void *call_rcu_thread(void *opaque)
> rcu_register_thread();
>
> for (;;) {
> - bool sleep = true;
> int n;
>
> /*
> @@ -294,7 +293,6 @@ static void *call_rcu_thread(void *opaque)
> * added before enter_qs() starts.
> */
> for (;;) {
> - qemu_event_reset(&sync_event);
> n = qatomic_read(&rcu_call_count);
> if (n) {
> break;
> @@ -303,36 +301,19 @@ static void *call_rcu_thread(void *opaque)
> #if defined(CONFIG_MALLOC_TRIM)
> malloc_trim(4 * 1024 * 1024);
> #endif
> - qemu_event_wait(&sync_event);
> + qemu_sem_wait(&sync_event);
> }
>
> - /*
> - * Ensure that an event for a rcu_call_count change will not interrupt
> - * wait_for_readers().
> - */
> - qemu_event_reset(&sync_event);
> -
> - /*
> - * Ensure that the forced variable has not been set after fetching
> - * rcu_call_count; otherwise we may get confused by a force quiescent
> - * state request for an element later than n.
> - */
> - while (qatomic_xchg(&forced, false)) {
> - sleep = false;
> - n = qatomic_read(&rcu_call_count);
> - }
> -
> - enter_qs(sleep);
> + enter_qs(!qatomic_xchg(&forced, false));
This is not OK; the forced variable may be set after rcu_call_count is
fetched. In that case, we should avoid unsetting the force quiescent
state request for the elements later than "n" or refetch "n".
> qatomic_sub(&rcu_call_count, n);
> bql_lock();
> while (n > 0) {
> node = try_dequeue();
> while (!node) {
> bql_unlock();
> - qemu_event_reset(&sync_event);
> node = try_dequeue();
> if (!node) {
> - qemu_event_wait(&sync_event);
> + qemu_sem_wait(&sync_event);
> node = try_dequeue();
> }
> bql_lock();
> @@ -352,7 +333,7 @@ void call_rcu1(struct rcu_head *node, void (*func)(struct rcu_head *node))
> enqueue(node);
>
> if (!qatomic_fetch_inc(&rcu_call_count)) {
> - qemu_event_set(&sync_event);
> + qemu_sem_post(&sync_event);
> }
> }
>
> @@ -456,8 +437,7 @@ static void rcu_init_complete(void)
>
> qemu_mutex_init(&rcu_registry_lock);
> qemu_event_init(&rcu_gp_event, true);
> -
> - qemu_event_init(&sync_event, false);
> + qemu_sem_init(&sync_event, 0);
>
> /* The caller is assumed to have BQL, so the call_rcu thread
> * must have been quiescent even after forking, just recreate it.
On Fri, Nov 07, 2025 at 10:47:35AM +0900, Akihiko Odaki wrote:
> On 2025/11/07 6:52, Peter Xu wrote:
> > On Thu, Nov 06, 2025 at 10:40:52AM +0900, Akihiko Odaki wrote:
> > > > > > > + /*
> > > > > > > + * Ensure that the forced variable has not been set after fetching
> > > > > > > + * rcu_call_count; otherwise we may get confused by a force quiescent
> > > > > > > + * state request for an element later than n.
> > > > > > > + */
> > > > > > > + while (qatomic_xchg(&forced, false)) {
> > > > > > > + sleep = false;
> > > > > > > + n = qatomic_read(&rcu_call_count);
> > > > > > > }
> > > > > >
> > > > > > This is pretty tricky, and I wonder if it will make the code easier to read
> > > > > > if we convert the sync_event to be a semaphore instead. When as a sem, it
> > > > > > will take account of whatever kick to it, either a call_rcu1() or an
> > > > > > enforced rcu flush, so that we don't need to reset it. Meanwhile, we don't
> > > > > > worry on an slightly outdated "n" read because the 2nd round of sem_wait()
> > > > > > will catch that new "n".
> > > > > >
> > > > > > Instead, worst case is rcu thread runs one more round without seeing
> > > > > > callbacks on the queue.
> > > > > >
> > > > > > I'm not sure if that could help simplying code, maybe also make it less
> > > > > > error-prone.
> > > > >
> > > > > Semaphore is not applicable here because it will not de-duplicate concurrent
> > > > > kicks of RCU threads.
> > > >
> > > > Why concurrent kicks of rcu threads is a problem? QemuSemaphore is
> > > > thread-safe itself, meanwhile IIUC it only still causes call_rcu_thread()
> > > > loops some more rounds reading "n", which looks all safe. No?
> > >
> > > It is safe but incurs overheads and confusing. QemuEvent represents the
> > > boolean semantics better.
> > >
> > > I also have difficulty to understand how converting sync_event to a
> > > semaphore simplifies the code. Perhaps some (pseudo)code to show how the
> > > code will look like may be useful.
> >
> > I prepared a patch on top of your current patchset to show what I meant. I
> > also added comments and some test results showing why I think it might be
> > fine that the sem overhead should be small.
> >
> > In short, I tested a VM with 8 vCPUs and 4G mem, booting Linux and properly
> > poweroff, I only saw <1000 rcu_call1 users in total. That should be the
> > max-bound of sem overhead on looping in rcu thread.
> >
> > It's in patch format but still treat it as a comment instead to discuss
> > with you. Attaching it is just easier for me.
> >
> > ===8<===
> > From 71f15ed19050a973088352a8d71b6cc6b7b5f7cf Mon Sep 17 00:00:00 2001
> > From: Peter Xu <peterx@redhat.com>
> > Date: Thu, 6 Nov 2025 16:03:00 -0500
> > Subject: [PATCH] rcu: Make sync_event a semaphore
> >
> > It could simply all reset logic, especially after enforced rcu is
> > introduced we'll also need a tweak to re-read "n", which can be avoided too
> > when with a sem.
> >
> > However, the sem can introduce an overhead in high frequecy rcu frees.
> > This patch is drafted with the assumption that rcu free is at least very
> > rare in QEMU, hence it's not a problem.
> >
> > When I tested with this command:
> >
> > qemu-system-x86_64 -M q35,kernel-irqchip=split,suppress-vmdesc=on -smp 8 \
> > -m 4G -msg timestamp=on -name peter-vm,debug-threads=on -cpu Nehalem \
> > -accel kvm -qmp unix:/tmp/peter.sock,server,nowait -nographic \
> > -monitor telnet::6666,server,nowait -netdev user,id=net0,hostfwd=tcp::5555-:22
> > -device e1000,netdev=net0 -device virtio-balloon $DISK
> >
> > I booted a pre-installed Linux, login and poweroff, wait until VM
> > completely shutdowns. I captured less than 1000 rcu_free1() calls in
> > summary. It means for the whole lifetime of such VM the max overhead of
> > the call_rcu_thread() loop reading rcu_call_count will be 1000 loops.
> >
> > Signed-off-by: Peter Xu <peterx@redhat.com>
> > ---
> > util/rcu.c | 36 ++++++++----------------------------
> > 1 file changed, 8 insertions(+), 28 deletions(-)
> >
> > diff --git a/util/rcu.c b/util/rcu.c
> > index 85f9333f5d..dfe031a5c9 100644
> > --- a/util/rcu.c
> > +++ b/util/rcu.c
> > @@ -54,7 +54,7 @@ static int rcu_call_count;
> > static QemuMutex rcu_registry_lock;
> > /* Set when the forced variable is set or rcu_call_count becomes non-zero. */
> > -static QemuEvent sync_event;
> > +static QemuSemaphore sync_event;
> > /*
> > * Check whether a quiescent state was crossed between the beginning of
> > @@ -80,7 +80,7 @@ static ThreadList registry = QLIST_HEAD_INITIALIZER(registry);
> > void force_rcu(void)
> > {
> > qatomic_set(&forced, true);
> > - qemu_event_set(&sync_event);
> > + qemu_sem_post(&sync_event);
> > }
> > /* Wait for previous parity/grace period to be empty of readers. */
> > @@ -148,7 +148,7 @@ static void wait_for_readers(bool sleep)
> > */
> > qemu_event_reset(&rcu_gp_event);
> > } else if (qatomic_read(&rcu_call_count) >= RCU_CALL_MIN_SIZE ||
> > - !sleeps || qemu_event_timedwait(&sync_event, 10)) {
> > + !sleeps || qemu_sem_timedwait(&sync_event, 10)) {
> > /*
> > * Now one of the following heuristical conditions is satisfied:
> > * - A decent number of callbacks piled up.
> > @@ -286,7 +286,6 @@ static void *call_rcu_thread(void *opaque)
> > rcu_register_thread();
> > for (;;) {
> > - bool sleep = true;
> > int n;
> > /*
> > @@ -294,7 +293,6 @@ static void *call_rcu_thread(void *opaque)
> > * added before enter_qs() starts.
> > */
> > for (;;) {
> > - qemu_event_reset(&sync_event);
> > n = qatomic_read(&rcu_call_count);
> > if (n) {
> > break;
> > @@ -303,36 +301,19 @@ static void *call_rcu_thread(void *opaque)
> > #if defined(CONFIG_MALLOC_TRIM)
> > malloc_trim(4 * 1024 * 1024);
> > #endif
> > - qemu_event_wait(&sync_event);
> > + qemu_sem_wait(&sync_event);
> > }
> > - /*
> > - * Ensure that an event for a rcu_call_count change will not interrupt
> > - * wait_for_readers().
> > - */
> > - qemu_event_reset(&sync_event);
> > -
> > - /*
> > - * Ensure that the forced variable has not been set after fetching
> > - * rcu_call_count; otherwise we may get confused by a force quiescent
> > - * state request for an element later than n.
> > - */
> > - while (qatomic_xchg(&forced, false)) {
> > - sleep = false;
> > - n = qatomic_read(&rcu_call_count);
> > - }
> > -
> > - enter_qs(sleep);
> > + enter_qs(!qatomic_xchg(&forced, false));
>
> This is not OK; the forced variable may be set after rcu_call_count is
> fetched. In that case, we should avoid unsetting the force quiescent state
> request for the elements later than "n" or refetch "n".
Indeed I missed that part, but it should be trivial to fix, on top of my
previous patch:
===8<===
diff --git a/util/rcu.c b/util/rcu.c
index dfe031a5c9..aff98d9ee2 100644
--- a/util/rcu.c
+++ b/util/rcu.c
@@ -286,6 +286,7 @@ static void *call_rcu_thread(void *opaque)
rcu_register_thread();
for (;;) {
+ bool sleep;
int n;
/*
@@ -293,6 +294,7 @@ static void *call_rcu_thread(void *opaque)
* added before enter_qs() starts.
*/
for (;;) {
+ sleep = !qatomic_xchg(&forced, false);
n = qatomic_read(&rcu_call_count);
if (n) {
break;
@@ -304,7 +306,7 @@ static void *call_rcu_thread(void *opaque)
qemu_sem_wait(&sync_event);
}
- enter_qs(!qatomic_xchg(&forced, false));
+ enter_qs(sleep);
qatomic_sub(&rcu_call_count, n);
bql_lock();
while (n > 0) {
===8<===
The idea is still the same, using semaphore can avoid explicit resets and a
lot of other ordering constraints like reading call_count, etc.
E.g. even before this series, we still need to properly reset at explicit
time to make sure we can capture a set() correct. When with sem, all these
issues are gone simply because we won't miss post() when it's a counter not
boolean.
Also, would you please also have a look at other comments I left in the
same email (after the patch I attached)?
https://lore.kernel.org/qemu-devel/aQ0Ys09WtlSPoapm@x1.local/
Can search "When I was having a closer look, I found some other issues".
Thanks,
--
Peter Xu
On 2025/11/07 23:00, Peter Xu wrote:
> On Fri, Nov 07, 2025 at 10:47:35AM +0900, Akihiko Odaki wrote:
>> On 2025/11/07 6:52, Peter Xu wrote:
>>> On Thu, Nov 06, 2025 at 10:40:52AM +0900, Akihiko Odaki wrote:
>>>>>>>> + /*
>>>>>>>> + * Ensure that the forced variable has not been set after fetching
>>>>>>>> + * rcu_call_count; otherwise we may get confused by a force quiescent
>>>>>>>> + * state request for an element later than n.
>>>>>>>> + */
>>>>>>>> + while (qatomic_xchg(&forced, false)) {
>>>>>>>> + sleep = false;
>>>>>>>> + n = qatomic_read(&rcu_call_count);
>>>>>>>> }
>>>>>>>
>>>>>>> This is pretty tricky, and I wonder if it will make the code easier to read
>>>>>>> if we convert the sync_event to be a semaphore instead. When as a sem, it
>>>>>>> will take account of whatever kick to it, either a call_rcu1() or an
>>>>>>> enforced rcu flush, so that we don't need to reset it. Meanwhile, we don't
>>>>>>> worry on an slightly outdated "n" read because the 2nd round of sem_wait()
>>>>>>> will catch that new "n".
>>>>>>>
>>>>>>> Instead, worst case is rcu thread runs one more round without seeing
>>>>>>> callbacks on the queue.
>>>>>>>
>>>>>>> I'm not sure if that could help simplying code, maybe also make it less
>>>>>>> error-prone.
>>>>>>
>>>>>> Semaphore is not applicable here because it will not de-duplicate concurrent
>>>>>> kicks of RCU threads.
>>>>>
>>>>> Why concurrent kicks of rcu threads is a problem? QemuSemaphore is
>>>>> thread-safe itself, meanwhile IIUC it only still causes call_rcu_thread()
>>>>> loops some more rounds reading "n", which looks all safe. No?
>>>>
>>>> It is safe but incurs overheads and confusing. QemuEvent represents the
>>>> boolean semantics better.
>>>>
>>>> I also have difficulty to understand how converting sync_event to a
>>>> semaphore simplifies the code. Perhaps some (pseudo)code to show how the
>>>> code will look like may be useful.
>>>
>>> I prepared a patch on top of your current patchset to show what I meant. I
>>> also added comments and some test results showing why I think it might be
>>> fine that the sem overhead should be small.
>>>
>>> In short, I tested a VM with 8 vCPUs and 4G mem, booting Linux and properly
>>> poweroff, I only saw <1000 rcu_call1 users in total. That should be the
>>> max-bound of sem overhead on looping in rcu thread.
>>>
>>> It's in patch format but still treat it as a comment instead to discuss
>>> with you. Attaching it is just easier for me.
>>>
>>> ===8<===
>>> From 71f15ed19050a973088352a8d71b6cc6b7b5f7cf Mon Sep 17 00:00:00 2001
>>> From: Peter Xu <peterx@redhat.com>
>>> Date: Thu, 6 Nov 2025 16:03:00 -0500
>>> Subject: [PATCH] rcu: Make sync_event a semaphore
>>>
>>> It could simply all reset logic, especially after enforced rcu is
>>> introduced we'll also need a tweak to re-read "n", which can be avoided too
>>> when with a sem.
>>>
>>> However, the sem can introduce an overhead in high frequecy rcu frees.
>>> This patch is drafted with the assumption that rcu free is at least very
>>> rare in QEMU, hence it's not a problem.
>>>
>>> When I tested with this command:
>>>
>>> qemu-system-x86_64 -M q35,kernel-irqchip=split,suppress-vmdesc=on -smp 8 \
>>> -m 4G -msg timestamp=on -name peter-vm,debug-threads=on -cpu Nehalem \
>>> -accel kvm -qmp unix:/tmp/peter.sock,server,nowait -nographic \
>>> -monitor telnet::6666,server,nowait -netdev user,id=net0,hostfwd=tcp::5555-:22
>>> -device e1000,netdev=net0 -device virtio-balloon $DISK
>>>
>>> I booted a pre-installed Linux, login and poweroff, wait until VM
>>> completely shutdowns. I captured less than 1000 rcu_free1() calls in
>>> summary. It means for the whole lifetime of such VM the max overhead of
>>> the call_rcu_thread() loop reading rcu_call_count will be 1000 loops.
>>>
>>> Signed-off-by: Peter Xu <peterx@redhat.com>
>>> ---
>>> util/rcu.c | 36 ++++++++----------------------------
>>> 1 file changed, 8 insertions(+), 28 deletions(-)
>>>
>>> diff --git a/util/rcu.c b/util/rcu.c
>>> index 85f9333f5d..dfe031a5c9 100644
>>> --- a/util/rcu.c
>>> +++ b/util/rcu.c
>>> @@ -54,7 +54,7 @@ static int rcu_call_count;
>>> static QemuMutex rcu_registry_lock;
>>> /* Set when the forced variable is set or rcu_call_count becomes non-zero. */
>>> -static QemuEvent sync_event;
>>> +static QemuSemaphore sync_event;
>>> /*
>>> * Check whether a quiescent state was crossed between the beginning of
>>> @@ -80,7 +80,7 @@ static ThreadList registry = QLIST_HEAD_INITIALIZER(registry);
>>> void force_rcu(void)
>>> {
>>> qatomic_set(&forced, true);
>>> - qemu_event_set(&sync_event);
>>> + qemu_sem_post(&sync_event);
>>> }
>>> /* Wait for previous parity/grace period to be empty of readers. */
>>> @@ -148,7 +148,7 @@ static void wait_for_readers(bool sleep)
>>> */
>>> qemu_event_reset(&rcu_gp_event);
>>> } else if (qatomic_read(&rcu_call_count) >= RCU_CALL_MIN_SIZE ||
>>> - !sleeps || qemu_event_timedwait(&sync_event, 10)) {
>>> + !sleeps || qemu_sem_timedwait(&sync_event, 10)) {
>>> /*
>>> * Now one of the following heuristical conditions is satisfied:
>>> * - A decent number of callbacks piled up.
>>> @@ -286,7 +286,6 @@ static void *call_rcu_thread(void *opaque)
>>> rcu_register_thread();
>>> for (;;) {
>>> - bool sleep = true;
>>> int n;
>>> /*
>>> @@ -294,7 +293,6 @@ static void *call_rcu_thread(void *opaque)
>>> * added before enter_qs() starts.
>>> */
>>> for (;;) {
>>> - qemu_event_reset(&sync_event);
>>> n = qatomic_read(&rcu_call_count);
>>> if (n) {
>>> break;
>>> @@ -303,36 +301,19 @@ static void *call_rcu_thread(void *opaque)
>>> #if defined(CONFIG_MALLOC_TRIM)
>>> malloc_trim(4 * 1024 * 1024);
>>> #endif
>>> - qemu_event_wait(&sync_event);
>>> + qemu_sem_wait(&sync_event);
>>> }
>>> - /*
>>> - * Ensure that an event for a rcu_call_count change will not interrupt
>>> - * wait_for_readers().
>>> - */
>>> - qemu_event_reset(&sync_event);
>>> -
>>> - /*
>>> - * Ensure that the forced variable has not been set after fetching
>>> - * rcu_call_count; otherwise we may get confused by a force quiescent
>>> - * state request for an element later than n.
>>> - */
>>> - while (qatomic_xchg(&forced, false)) {
>>> - sleep = false;
>>> - n = qatomic_read(&rcu_call_count);
>>> - }
>>> -
>>> - enter_qs(sleep);
>>> + enter_qs(!qatomic_xchg(&forced, false));
>>
>> This is not OK; the forced variable may be set after rcu_call_count is
>> fetched. In that case, we should avoid unsetting the force quiescent state
>> request for the elements later than "n" or refetch "n".
>
> Indeed I missed that part, but it should be trivial to fix, on top of my
> previous patch:
>
> ===8<===
> diff --git a/util/rcu.c b/util/rcu.c
> index dfe031a5c9..aff98d9ee2 100644
> --- a/util/rcu.c
> +++ b/util/rcu.c
> @@ -286,6 +286,7 @@ static void *call_rcu_thread(void *opaque)
> rcu_register_thread();
>
> for (;;) {
> + bool sleep;
> int n;
>
> /*
> @@ -293,6 +294,7 @@ static void *call_rcu_thread(void *opaque)
> * added before enter_qs() starts.
> */
> for (;;) {
> + sleep = !qatomic_xchg(&forced, false);
This doesn't work either; the following sequence may happen (assume
forced is false at beginning):
qatomic_xchg(&forced, false) |
| call_rcu1()
| qatomic_fetch_inc(&rcu_call_count)
| force_rcu()
| qatomic_set(&forced, true)
qatomic_read(&rcu_call_count)
We need to enter the force quiescent state for the node added with the
call_rcu1() call in this sequence, but this code doesn't.
> n = qatomic_read(&rcu_call_count);
> if (n) {
> break;
> @@ -304,7 +306,7 @@ static void *call_rcu_thread(void *opaque)
> qemu_sem_wait(&sync_event);
> }
>
> - enter_qs(!qatomic_xchg(&forced, false));
> + enter_qs(sleep);
> qatomic_sub(&rcu_call_count, n);
> bql_lock();
> while (n > 0) {
> ===8<===
>
> The idea is still the same, using semaphore can avoid explicit resets and a
> lot of other ordering constraints like reading call_count, etc.
>
> E.g. even before this series, we still need to properly reset at explicit
> time to make sure we can capture a set() correct. When with sem, all these
> issues are gone simply because we won't miss post() when it's a counter not
> boolean.
>
> Also, would you please also have a look at other comments I left in the
> same email (after the patch I attached)?
>
> https://lore.kernel.org/qemu-devel/aQ0Ys09WtlSPoapm@x1.local/
>
> Can search "When I was having a closer look, I found some other issues".
I have just replied to the email. My mailer ignored the part after "--".
Regards,
Akihiko Odaki
© 2016 - 2025 Red Hat, Inc.