io_uring: don't recurse on tsk->sighand->siglock with signalfd

If an application is doing reads on signalfd, and we arm the poll handler
because there's no data available, then the wakeup can recurse on the
tasks sighand->siglock as the signal delivery from task_work_add() will
use TWA_SIGNAL and that attempts to lock it again.

We can detect the signalfd case pretty easily by comparing the poll->head
wait_queue_head_t with the target task signalfd wait queue. Just use
normal task wakeup for this case.

Cc: stable@vger.kernel.org # v5.7+
Signed-off-by: Jens Axboe <axboe@kernel.dk>

1 files changed, 16 insertions, 6 deletions

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 91e2cc8414f9..c9d526ff55e0 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -1746,7 +1746,8 @@ static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
 	return __io_req_find_next(req);
 }
 
-static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb)
+static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb,
+				bool twa_signal_ok)
 {
 	struct task_struct *tsk = req->task;
 	struct io_ring_ctx *ctx = req->ctx;
@@ -1759,7 +1760,7 @@ static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb)
 	 * will do the job.
 	 */
 	notify = 0;
-	if (!(ctx->flags & IORING_SETUP_SQPOLL))
+	if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
 		notify = TWA_SIGNAL;
 
 	ret = task_work_add(tsk, cb, notify);
@@ -1819,7 +1820,7 @@ static void io_req_task_queue(struct io_kiocb *req)
 	init_task_work(&req->task_work, io_req_task_submit);
 	percpu_ref_get(&req->ctx->refs);
 
-	ret = io_req_task_work_add(req, &req->task_work);
+	ret = io_req_task_work_add(req, &req->task_work, true);
 	if (unlikely(ret)) {
 		struct task_struct *tsk;
 
@@ -2322,7 +2323,7 @@ static bool io_rw_reissue(struct io_kiocb *req, long res)
 	init_task_work(&req->task_work, io_rw_resubmit);
 	percpu_ref_get(&req->ctx->refs);
 
-	ret = io_req_task_work_add(req, &req->task_work);
+	ret = io_req_task_work_add(req, &req->task_work, true);
 	if (!ret)
 		return true;
 #endif
@@ -3044,7 +3045,7 @@ static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
 
 	/* submit ref gets dropped, acquire a new one */
 	refcount_inc(&req->refs);
-	ret = io_req_task_work_add(req, &req->task_work);
+	ret = io_req_task_work_add(req, &req->task_work, true);
 	if (unlikely(ret)) {
 		struct task_struct *tsk;
 
@@ -4566,6 +4567,7 @@ struct io_poll_table {
 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
 			   __poll_t mask, task_work_func_t func)
 {
+	bool twa_signal_ok;
 	int ret;
 
 	/* for instances that support it check for an event match first: */
@@ -4581,12 +4583,20 @@ static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
 	percpu_ref_get(&req->ctx->refs);
 
 	/*
+	 * If we using the signalfd wait_queue_head for this wakeup, then
+	 * it's not safe to use TWA_SIGNAL as we could be recursing on the
+	 * tsk->sighand->siglock on doing the wakeup. Should not be needed
+	 * either, as the normal wakeup will suffice.
+	 */
+	twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
+
+	/*
 	 * If this fails, then the task is exiting. When a task exits, the
 	 * work gets canceled, so just cancel this request as well instead
 	 * of executing it. We can't safely execute it anyway, as we may not
 	 * have the needed state needed for it anyway.
 	 */
-	ret = io_req_task_work_add(req, &req->task_work);
+	ret = io_req_task_work_add(req, &req->task_work, twa_signal_ok);
 	if (unlikely(ret)) {
 		struct task_struct *tsk;