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

[ Upstream commit fd7d6de ]

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>
Signed-off-by: Sasha Levin <sashal@kernel.org>

fs/io_uring.c

@@ -4114,7 +4114,8 @@ struct io_poll_table {
 	int error;
 };
 
-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;
@@ -4127,7 +4128,7 @@ 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);
@@ -4141,6 +4142,7 @@ static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
 			   __poll_t mask, task_work_func_t func)
 {
 	struct task_struct *tsk;
+	bool twa_signal_ok;
 	int ret;
 
 	/* for instances that support it check for an event match first: */
@@ -4156,13 +4158,21 @@ static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
 	init_task_work(&req->task_work, func);
 	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)) {
 		WRITE_ONCE(poll->canceled, true);
 		tsk = io_wq_get_task(req->ctx->io_wq);