Re: [PATCH 07/25] serial: sh-sci: Convert from clk_get() to devm_clk_get()

[Laurent Pinchart <laurent.pinchart@ideasonboard.com>]

Hi Geert,

On Monday 14 December 2015 11:34:11 Geert Uytterhoeven wrote:
> On Sun, Dec 13, 2015 at 8:39 PM, Laurent Pinchart wrote:
> > On Thursday 10 December 2015 13:30:18 Geert Uytterhoeven wrote:
> >> On Thu, Nov 19, 2015 at 9:38 PM, Laurent Pinchart wrote:
> >> > On Thursday 19 November 2015 19:38:46 Geert Uytterhoeven wrote:
> >> >> Transfer clock cleanup handling to the core device management code.
> >> >> 
> >> >> Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
> >> >> ---
> >> >> 
> >> >>  drivers/tty/serial/sh-sci.c | 8 +++-----
> >> >>  1 file changed, 3 insertions(+), 5 deletions(-)
> >> >> 
> >> >> diff --git a/drivers/tty/serial/sh-sci.c b/drivers/tty/serial/sh-sci.c
> >> >> index cba51da604253db6..9442961a198378c7 100644
> >> >> --- a/drivers/tty/serial/sh-sci.c
> >> >> +++ b/drivers/tty/serial/sh-sci.c
> >> >> @@ -2216,7 +2216,7 @@ static struct uart_ops sci_uart_ops = {
> >> >> 
> >> >>  static int sci_init_clocks(struct sci_port *sci_port, struct device
> >> >>  *dev)
> >> >>  {
> >> >>       /* Get the SCI functional clock. It's called "fck" on ARM. */
> >> >> -     sci_port->fclk = clk_get(dev, "fck");
> >> >> +     sci_port->fclk = devm_clk_get(dev, "fck");
> >> > 
> >> > Have you tested what happens if you unbind the device from the driver
> >> > while userspace has the serial port open ?
> >> 
> >> Yes I have. And I didn't notice any user-visible behavioral differences.
> >> 
> >> There is a small ordering difference, though: with clk_get() and
> >> clk_put(), the explicit clk_put() is done before removing the device from
> >> its PM
> >> 
> >> Domain (which involces another call to clk_put() on the module clock):
> >>     device_release_driver
> >>         __device_release_driver
> >>             platform_drv_remove
> >>                 sci_remove
> >>                     sci_cleanup_single
> >>                         __clk_put
> >>                 genpd_dev_pm_detach
> >>                     pm_genpd_remove_device
> >>                         pm_clk_destroy
> >>                             __pm_clk_remove
> >>                                 __clk_put
> >> 
> >> When using devm_clk_get(), the managed cleanup is done after the device
> >> has been removed from its PM Domain:
> >>     device_release_driver
> >>         __device_release_driver
> >>             platform_drv_remove
> >>                 genpd_dev_pm_detach
> >>                     pm_genpd_remove_device
> >>                         pm_clk_destroy
> >>                             __pm_clk_remove
> >>                                 __clk_put
> >>             release_nodes
> >>                 __clk_put
> >> 
> >> This shouldn't make a difference, and applies to all other drivers using
> >> devm_*(), and devices that are part of a PM Domain.
> > 
> > I was more concerned about the clk_disable_unprepare() calls. The function
> > is called through the uart pm handler which is called when closing the
> > uart (uart_close()). It's a bit difficult to follow the dynamic code
> > paths as I'm not very familiar with the framework, but I was concerned
> > that the would end up calling clk_disable_unprepare() on a clock for
> > which we have no reference anymore. What prevents clk_disable_unprepare()
> > from being called in the close path after the device has been unbound ?
> > And what ensures that the clocks get unprepared when unbinding the device
> > if the port is held open by a userspace process ?
> 
> Both questions are answered by uart_hangup():
>   - uart_hangup() sets port->count to zero, hence uart_close() returns early
> (second return statement), long before it would call the uart pm handler.
> - uart_hangup() calls uart_change_pm(state, UART_PM_STATE_OFF);

Thank you for making me slightly wiser :-)

-- 
Regards,

Laurent Pinchart