gpio: do not double-check direction on sleeping chips

commit 60f8339 upstream.

When locking a GPIO line as IRQ, we go to lengths to
double-check that the line is really set as input before
marking it as used for IRQ. This is not good on GPIO chips
that can sleep, because this function is called in IRQ-safe
context. Just skip this if it can't be checked quickly.

Currently this happens on sleeping expanders such as STMPE
or TC3589x:

BUG: scheduling while atomic: swapper/1/0x00000002
Modules linked in:
CPU: 0 PID: 1 Comm: swapper Not tainted 4.9.0-rc1+ #38
Hardware name: Nomadik STn8815
[<c000f2e0>] (unwind_backtrace) from [<c000d244>] (show_stack+0x10/0x14)
[<c000d244>] (show_stack) from [<c0037b78>] (__schedule_bug+0x54/0x80)
[<c0037b78>] (__schedule_bug) from [<c042df14>] (__schedule+0x3a0/0x460)
[<c042df14>] (__schedule) from [<c042e028>] (schedule+0x54/0xb8)
(...)

This patch fixes that problem and relies on the direction
read from the chip when it was added.

Fixes: 9c10280 ("gpio: flush direction status in gpiochip_lock_as_irq()")
Cc: Patrice Chotard <[email protected]>
Signed-off-by: Linus Walleij <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>

drivers/gpio/gpiolib.c

@@ -2667,8 +2667,11 @@ int gpiochip_lock_as_irq(struct gpio_chip *chip, unsigned int offset)
 	if (IS_ERR(desc))
 		return PTR_ERR(desc);
 
-	/* Flush direction if something changed behind our back */
-	if (chip->get_direction) {
+	/*
+	 * If it's fast: flush the direction setting if something changed
+	 * behind our back
+	 */
+	if (!chip->can_sleep && chip->get_direction) {
 		int dir = chip->get_direction(chip, offset);
 
 		if (dir)