[PATCH] io_uring: fix io_sq_thread no schedule when busy

[PATCH] io_uring: fix io_sq_thread no schedule when busy

[Xuan Zhuo]

When the user consumes and generates sqe at a fast rate,
io_sqring_entries can always get sqe, and ret will not be equal to -EBUSY,
so that io_sq_thread will never call cond_resched or schedule, and then
we will get the following system error prompt:

rcu: INFO: rcu_sched self-detected stall on CPU
or
watchdog: BUG: soft lockup-CPU#23 stuck for 112s! [io_uring-sq:1863]

This patch adds a check after io_submit_sqes. If io_sq_thread does not call
cond_resched or schedule for more than HZ/2, it will call them.

Signed-off-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
---
 fs/io_uring.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index a78201b..de92363 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -5983,7 +5983,7 @@ static int io_sq_thread(void *data)
 	struct io_ring_ctx *ctx = data;
 	const struct cred *old_cred;
 	DEFINE_WAIT(wait);
-	unsigned long timeout;
+	unsigned long timeout, timeout_busy;
 	int ret = 0;
 
 	complete(&ctx->sq_thread_comp);
@@ -5991,6 +5991,7 @@ static int io_sq_thread(void *data)
 	old_cred = override_creds(ctx->creds);
 
 	timeout = jiffies + ctx->sq_thread_idle;
+	timeout_busy = jiffies + HZ/2;
 	while (!kthread_should_park()) {
 		unsigned int to_submit;
 
@@ -6012,6 +6013,7 @@ static int io_sq_thread(void *data)
 		 * to enter the kernel to reap and flush events.
 		 */
 		if (!to_submit || ret == -EBUSY) {
+sched:
 			/*
 			 * Drop cur_mm before scheduling, we can't hold it for
 			 * long periods (or over schedule()). Do this before
@@ -6033,6 +6035,7 @@ static int io_sq_thread(void *data)
 				if (current->task_works)
 					task_work_run();
 				cond_resched();
+				timeout_busy = jiffies + HZ/2;
 				continue;
 			}
 
@@ -6073,6 +6076,8 @@ static int io_sq_thread(void *data)
 				schedule();
 				finish_wait(&ctx->sqo_wait, &wait);
 
+				timeout_busy = jiffies + HZ/2;
+
 				ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
 				ret = 0;
 				continue;
@@ -6087,6 +6092,9 @@ static int io_sq_thread(void *data)
 			ret = io_submit_sqes(ctx, to_submit, NULL, -1);
 		mutex_unlock(&ctx->uring_lock);
 		timeout = jiffies + ctx->sq_thread_idle;
+
+		if (time_after(jiffies, timeout_busy))
+			goto sched;
 	}
 
 	if (current->task_works)
-- 
1.8.3.1

Re: [PATCH] io_uring: fix io_sq_thread no schedule when busy

[Jens Axboe]

On 6/21/20 11:50 PM, Xuan Zhuo wrote:
> When the user consumes and generates sqe at a fast rate,
> io_sqring_entries can always get sqe, and ret will not be equal to -EBUSY,
> so that io_sq_thread will never call cond_resched or schedule, and then
> we will get the following system error prompt:
> 
> rcu: INFO: rcu_sched self-detected stall on CPU
> or
> watchdog: BUG: soft lockup-CPU#23 stuck for 112s! [io_uring-sq:1863]
> 
> This patch adds a check after io_submit_sqes. If io_sq_thread does not call
> cond_resched or schedule for more than HZ/2, it will call them.

This looks reasonable. It'd be easier if we could just cond_resched() after
the call unconditionally, but that would not drop the mm.

But maybe we can just drop the timeout and just rely on need_resched()
for this, ala the below?

diff --git a/fs/io_uring.c b/fs/io_uring.c
index c686061c3762..955d0765f302 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -6248,7 +6248,7 @@ static int io_sq_thread(void *data)
 		 * If submit got -EBUSY, flag us as needing the application
 		 * to enter the kernel to reap and flush events.
 		 */
-		if (!to_submit || ret == -EBUSY) {
+		if (!to_submit || ret == -EBUSY || need_resched()) {
 			/*
 			 * Drop cur_mm before scheduling, we can't hold it for
 			 * long periods (or over schedule()). Do this before
@@ -6264,7 +6264,7 @@ static int io_sq_thread(void *data)
 			 * more IO, we should wait for the application to
 			 * reap events and wake us up.
 			 */
-			if (!list_empty(&ctx->poll_list) ||
+			if (!list_empty(&ctx->poll_list) || need_resched() ||
 			    (!time_after(jiffies, timeout) && ret != -EBUSY &&
 			    !percpu_ref_is_dying(&ctx->refs))) {
 				if (current->task_works)

-- 
Jens Axboe