If a task exits normally, then the teardown of async workers happens
quickly. But if it simply closes the ring fd, then the async teardown
ends up waiting for workers to timeout if they are sleeping, which can
then take up to 5 seconds (by default). This isn't a big issue as this
happens off the workqueue path, but let's be nicer and ensure that we
exit as quick as possible.
Signed-off-by: Jens Axboe <axboe@xxxxxxxxx>
---
diff --git a/fs/io-wq.c b/fs/io-wq.c
index d80e4a735677..60cd841c6c57 100644
--- a/fs/io-wq.c
+++ b/fs/io-wq.c
@@ -1130,7 +1130,17 @@ static bool io_task_work_match(struct callback_head *cb, void *data)
void io_wq_exit_start(struct io_wq *wq)
{
+ int node;
+
set_bit(IO_WQ_BIT_EXIT, &wq->state);
+
+ rcu_read_lock();
+ for_each_node(node) {
+ struct io_wqe *wqe = wq->wqes[node];
+
+ io_wq_for_each_worker(wqe, io_wq_worker_wake, NULL);
+ }
+ rcu_read_unlock();
}
static void io_wq_exit_workers(struct io_wq *wq)
diff --git a/fs/io_uring.c b/fs/io_uring.c
index 2bde732a1183..9936ebaa8180 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -9158,6 +9158,7 @@ static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
{
+ struct io_uring_task *tctx = current->io_uring;
unsigned long index;
struct creds *creds;
@@ -9175,6 +9176,10 @@ static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
/* if we failed setting up the ctx, we might not have any rings */
io_iopoll_try_reap_events(ctx);
+ /* trigger exit, if it hasn't been done already */
+ if (tctx->io_wq)
+ io_wq_exit_start(tctx->io_wq);
+
INIT_WORK(&ctx->exit_work, io_ring_exit_work);
/*
* Use system_unbound_wq to avoid spawning tons of event kworkers
--
Jens Axboe
