io_uring/timeout: flush timeouts outside of the timeout lock

syzbot reports that a recent fix causes nesting issues between the (now)
raw timeoutlock and the eventfd locking:

=============================
[ BUG: Invalid wait context ]
6.13.0-rc4-00080-g9828a4c0901f #29 Not tainted
-----------------------------
kworker/u32:0/68094 is trying to lock:
ffff000014d7a520 (&ctx->wqh#2){..-.}-{3:3}, at: eventfd_signal_mask+0x64/0x180
other info that might help us debug this:
context-{5:5}
6 locks held by kworker/u32:0/68094:
 #0: ffff0000c1d98148 ((wq_completion)iou_exit){+.+.}-{0:0}, at: process_one_work+0x4e8/0xfc0
 #1: ffff80008d927c78 ((work_completion)(&ctx->exit_work)){+.+.}-{0:0}, at: process_one_work+0x53c/0xfc0
 #2: ffff0000c59bc3d8 (&ctx->completion_lock){+.+.}-{3:3}, at: io_kill_timeouts+0x40/0x180
 #3: ffff0000c59bc358 (&ctx->timeout_lock){-.-.}-{2:2}, at: io_kill_timeouts+0x48/0x180
 #4: ffff800085127aa0 (rcu_read_lock){....}-{1:3}, at: rcu_lock_acquire+0x8/0x38
 #5: ffff800085127aa0 (rcu_read_lock){....}-{1:3}, at: rcu_lock_acquire+0x8/0x38
stack backtrace:
CPU: 7 UID: 0 PID: 68094 Comm: kworker/u32:0 Not tainted 6.13.0-rc4-00080-g9828a4c0901f #29
Hardware name: linux,dummy-virt (DT)
Workqueue: iou_exit io_ring_exit_work
Call trace:
 show_stack+0x1c/0x30 (C)
 __dump_stack+0x24/0x30
 dump_stack_lvl+0x60/0x80
 dump_stack+0x14/0x20
 __lock_acquire+0x19f8/0x60c8
 lock_acquire+0x1a4/0x540
 _raw_spin_lock_irqsave+0x90/0xd0
 eventfd_signal_mask+0x64/0x180
 io_eventfd_signal+0x64/0x108
 io_req_local_work_add+0x294/0x430
 __io_req_task_work_add+0x1c0/0x270
 io_kill_timeout+0x1f0/0x288
 io_kill_timeouts+0xd4/0x180
 io_uring_try_cancel_requests+0x2e8/0x388
 io_ring_exit_work+0x150/0x550
 process_one_work+0x5e8/0xfc0
 worker_thread+0x7ec/0xc80
 kthread+0x24c/0x300
 ret_from_fork+0x10/0x20

because after the preempt-rt fix for the timeout lock nesting inside
the io-wq lock, we now have the eventfd spinlock nesting inside the
raw timeout spinlock.

Rather than play whack-a-mole with other nesting on the timeout lock,
split the deletion and killing of timeouts so queueing the task_work
for the timeout cancelations can get done outside of the timeout lock.

Reported-by: [email protected]
Fixes: 020b40f ("io_uring: make ctx->timeout_lock a raw spinlock")
Signed-off-by: Jens Axboe <[email protected]>

io_uring/timeout.c

@@ -85,29 +85,45 @@ static void io_timeout_complete(struct io_kiocb *req, struct io_tw_state *ts)
 	io_req_task_complete(req, ts);
 }
 
-static bool io_kill_timeout(struct io_kiocb *req, int status)
+static __cold bool io_flush_killed_timeouts(struct list_head *list, int err)
+{
+	if (list_empty(list))
+		return false;
+
+	while (!list_empty(list)) {
+		struct io_timeout *timeout;
+		struct io_kiocb *req;
+
+		timeout = list_first_entry(list, struct io_timeout, list);
+		list_del_init(&timeout->list);
+		req = cmd_to_io_kiocb(timeout);
+		if (err)
+			req_set_fail(req);
+		io_req_queue_tw_complete(req, err);
+	}
+
+	return true;
+}
+
+static void io_kill_timeout(struct io_kiocb *req, struct list_head *list)
 	__must_hold(&req->ctx->timeout_lock)
 {
 	struct io_timeout_data *io = req->async_data;
 
 	if (hrtimer_try_to_cancel(&io->timer) != -1) {
 		struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
 
-		if (status)
-			req_set_fail(req);
 		atomic_set(&req->ctx->cq_timeouts,
 			atomic_read(&req->ctx->cq_timeouts) + 1);
-		list_del_init(&timeout->list);
-		io_req_queue_tw_complete(req, status);
-		return true;
+		list_move_tail(&timeout->list, list);
 	}
-	return false;
 }
 
 __cold void io_flush_timeouts(struct io_ring_ctx *ctx)
 {
-	u32 seq;
 	struct io_timeout *timeout, *tmp;
+	LIST_HEAD(list);
+	u32 seq;
 
 	raw_spin_lock_irq(&ctx->timeout_lock);
 	seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
@@ -131,10 +147,11 @@ __cold void io_flush_timeouts(struct io_ring_ctx *ctx)
 		if (events_got < events_needed)
 			break;
 
-		io_kill_timeout(req, 0);
+		io_kill_timeout(req, &list);
 	}
 	ctx->cq_last_tm_flush = seq;
 	raw_spin_unlock_irq(&ctx->timeout_lock);
+	io_flush_killed_timeouts(&list, 0);
 }
 
 static void io_req_tw_fail_links(struct io_kiocb *link, struct io_tw_state *ts)
@@ -661,7 +678,7 @@ __cold bool io_kill_timeouts(struct io_ring_ctx *ctx, struct io_uring_task *tctx
 			     bool cancel_all)
 {
 	struct io_timeout *timeout, *tmp;
-	int canceled = 0;
+	LIST_HEAD(list);
 
 	/*
 	 * completion_lock is needed for io_match_task(). Take it before
@@ -672,11 +689,11 @@ __cold bool io_kill_timeouts(struct io_ring_ctx *ctx, struct io_uring_task *tctx
 	list_for_each_entry_safe(timeout, tmp, &ctx->timeout_list, list) {
 		struct io_kiocb *req = cmd_to_io_kiocb(timeout);
 
-		if (io_match_task(req, tctx, cancel_all) &&
-		    io_kill_timeout(req, -ECANCELED))
-			canceled++;
+		if (io_match_task(req, tctx, cancel_all))
+			io_kill_timeout(req, &list);
 	}
 	raw_spin_unlock_irq(&ctx->timeout_lock);
 	spin_unlock(&ctx->completion_lock);
-	return canceled != 0;
+
+	return io_flush_killed_timeouts(&list, -ECANCELED);
 }