MEDIUM: mux-h2: use task_kill() during h2_takeover() instead of task_wakeup()
task_wakeup() passes the task through the global run queue under the
global RQ lock, which is expensive when dealing with large amounts of
h2_takeover() calls. Let's use the new task_kill() instead to kill the
task.
diff --git a/src/mux_h2.c b/src/mux_h2.c
index 5fad272..0ad066b 100644
--- a/src/mux_h2.c
+++ b/src/mux_h2.c
@@ -6050,6 +6050,7 @@
static int h2_takeover(struct connection *conn)
{
struct h2c *h2c = conn->ctx;
+ struct task *task;
if (fd_takeover(conn->handle.fd, conn) != 0)
return -1;
@@ -6061,10 +6062,13 @@
*/
h2c->wait_event.tasklet->context = NULL;
tasklet_wakeup(h2c->wait_event.tasklet);
- if (h2c->task) {
- h2c->task->context = NULL;
- /* Wake the task, to let it free itself */
- task_wakeup(h2c->task, TASK_WOKEN_OTHER);
+
+ task = h2c->task;
+ if (task) {
+ task->context = NULL;
+ h2c->task = NULL;
+ __ha_barrier_store();
+ task_kill(task);
h2c->task = task_new(tid_bit);
if (!h2c->task) {