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 <[email protected]>
---
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
