Skip to contentSkip to conversation

io_uring: fix IO hang in io wq exit

PATCH New Ming Lei 1 participant 3 messages Aug 25, 2023

Subsystems (2)
lore
Share & tools
lore

Ming Lei

author
#0 lore raw

cover io_uring: fix IO hang in io wq exit

Hello,

The 1st patch add one helper, and the 2nd one fixes IO hang issue[1] reported by David Howells.

[1] https://lore.kernel.org/linux-block/[email protected]/

Ming Lei (2):
io_uring: add one helper for reaping iopoll events
io_uring: reap iopoll events before exiting io wq

 io_uring/io_uring.c | 27 +++++++++++++-------
 io_uring/io_uring.h |  1 +
 io_uring/tctx.c     | 60 +++++++++++++++++++++++++++++++++++++++------
 3 files changed, 71 insertions(+), 17 deletions(-)
Signature · 1 line
2.40.1
author
#1 lore raw

PATCH 1/2 io_uring: add one helper for reaping iopoll events

Prepare for reaping iopoll events before exiting io_wq, which may depend on iopoll.

---
 io_uring/io_uring.c | 27 ++++++++++++++++++---------
 1 file changed, 18 insertions(+), 9 deletions(-)
io_uring/io_uring.c +18 -9
index 88599852af82..c4adb44f1aa4 100644
@@ -3214,6 +3214,23 @@ static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
 	return ret;
 }
 
+static bool iopoll_reap_events(struct io_ring_ctx *ctx, bool reap_all)
+{
+	bool reapped = false;
+
+	/* SQPOLL thread does its own polling */
+	if ((!(ctx->flags & IORING_SETUP_SQPOLL) && reap_all) ||
+	    (ctx->sq_data && ctx->sq_data->thread == current)) {
+		while (!wq_list_empty(&ctx->iopoll_list)) {
+			io_iopoll_try_reap_events(ctx);
+			reapped = true;
+			cond_resched();
+		}
+	}
+
+	return reapped;
+}
+
 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
 						struct task_struct *task,
 						bool cancel_all)
@@ -3245,15 +3262,7 @@ static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
 		ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
 	}
 
-	/* SQPOLL thread does its own polling */
-	if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
-	    (ctx->sq_data && ctx->sq_data->thread == current)) {
-		while (!wq_list_empty(&ctx->iopoll_list)) {
-			io_iopoll_try_reap_events(ctx);
-			ret = true;
-			cond_resched();
-		}
-	}
+	ret = iopoll_reap_events(ctx, cancel_all);
 
 	if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
 	    io_allowed_defer_tw_run(ctx))
Signature · 1 line
2.40.1
author
#2 lore raw

PATCH 2/2 io_uring: reap iopoll events before exiting io wq

In reply to Ming Lei

io_uring delays reaping iopoll events into exit work, which is scheduled in io_uring_release(). But io wq is exited inside do_exit(), and wq code path may share resource with iopoll code path, such as request. If iopoll events aren't reaped, io wq exit can't be done, and cause IO hang.

The issue can be triggered when terminating 't/io_uring -n4 /dev/nullb0' with default null_blk parameters.

Fix it by reaping iopoll events in io_uring_clean_tctx() and moving io wq exit into workqueue context.

---
 io_uring/io_uring.c |  2 +-
 io_uring/io_uring.h |  1 +
 io_uring/tctx.c     | 60 +++++++++++++++++++++++++++++++++++++++------
 3 files changed, 54 insertions(+), 9 deletions(-)
io_uring/io_uring.c +1 -1
index c4adb44f1aa4..ab7844c3380c 100644
@@ -3214,7 +3214,7 @@ static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
 	return ret;
 }
 
-static bool iopoll_reap_events(struct io_ring_ctx *ctx, bool reap_all)
+bool iopoll_reap_events(struct io_ring_ctx *ctx, bool reap_all)
 {
 	bool reapped = false;
 
io_uring/io_uring.h +1 -0
index 547c30582fb8..f1556666a064 100644
@@ -63,6 +63,7 @@ void io_req_task_queue_fail(struct io_kiocb *req, int ret);
 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts);
 void tctx_task_work(struct callback_head *cb);
 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
+bool iopoll_reap_events(struct io_ring_ctx *ctx, bool reap_all);
 int io_uring_alloc_task_context(struct task_struct *task,
 				struct io_ring_ctx *ctx);
 
io_uring/tctx.c +52 -8

82 lines not on this page · raw message

Signature · 1 line
2.40.1