On Tue, Jul 04, 2023 at 06:19:20PM +0300, Dmitry Rokosov wrote:
> On Tue, Jul 04, 2023 at 04:46:40PM +0200, neil.armstrong@xxxxxxxxxx wrote:
> > On 04/07/2023 15:59, Dmitry Rokosov wrote:
> > > Actually, the meson_uart module is already a platform_driver, but it is
> > > currently registered manually and the uart core registration is run
> > > outside the probe() scope, which results in some restrictions. For
> > > instance, it is not possible to communicate with the OF subsystem
> > > because it requires an initialized device object.
> > >
> > > To address this issue, apply module_platform_driver() instead of direct
> > > module init/exit routines. Additionally, move uart_register_driver() to
> > > the driver probe(), and destroy manual console registration because it's
> > > already run in the uart_register_driver() flow.
> > >
> > > Signed-off-by: Dmitry Rokosov <ddrokosov@xxxxxxxxxxxxxx>
> > > ---
> > > drivers/tty/serial/meson_uart.c | 46 +++++++--------------------------
> > > 1 file changed, 10 insertions(+), 36 deletions(-)
> > >
> > > diff --git a/drivers/tty/serial/meson_uart.c b/drivers/tty/serial/meson_uart.c
> > > index 169f028956ae..87c0eb5f2dba 100644
> > > --- a/drivers/tty/serial/meson_uart.c
> > > +++ b/drivers/tty/serial/meson_uart.c
> > > @@ -621,12 +621,6 @@ static struct console meson_serial_console = {
> > > .data = &meson_uart_driver,
> > > };
> > > -static int __init meson_serial_console_init(void)
> > > -{
> > > - register_console(&meson_serial_console);
> > > - return 0;
> > > -}
> > > -
> > > static void meson_serial_early_console_write(struct console *co,
> > > const char *s,
> > > u_int count)
> > > @@ -652,9 +646,6 @@ OF_EARLYCON_DECLARE(meson, "amlogic,meson-ao-uart",
> > > #define MESON_SERIAL_CONSOLE (&meson_serial_console)
> > > #else
> > > -static int __init meson_serial_console_init(void) {
> > > - return 0;
> > > -}
> > > #define MESON_SERIAL_CONSOLE NULL
> > > #endif
> > > @@ -738,6 +729,13 @@ static int meson_uart_probe(struct platform_device *pdev)
> > > if (ret)
> > > return ret;
> > > + if (!meson_uart_driver.state) {
> > > + ret = uart_register_driver(&meson_uart_driver);
> > > + if (ret)
> > > + return dev_err_probe(&pdev->dev, ret,
> > > + "failed to register meson-uart driver\n");
> > > + }
> >
> > PL010 protects this in a mutex, and I think you should do the same otherwise
> > if multiple uart probes at the same it will do weird stuff.
> >
>
> Looks like that not all drivers protect this location with a specialized
> mutex object. Firstly, I think it's important to verify parallel probe()
> calling and implementing mutex protection at the platform core level.
> For example, I've faced with the same problem during regmap mutex based
> protection.
>
Upon examining the core logic in drivers/base/dd.c, I have observed that
driver_probe_device() is consistently executed under the device_lock().
This lock is already based on a mutex, thus ensuring parallel execution
protection:
https://elixir.bootlin.com/linux/latest/source/include/linux/device.h#L835
> > > +
> > > port->iotype = UPIO_MEM;
> > > port->mapbase = res_mem->start;
> > > port->mapsize = resource_size(res_mem);
> > > @@ -776,6 +774,8 @@ static int meson_uart_remove(struct platform_device *pdev)
> > > uart_remove_one_port(&meson_uart_driver, port);
> > > meson_ports[pdev->id] = NULL;
> > > + uart_unregister_driver(&meson_uart_driver);
> > > +
> >
> > This is dangerous, it will remove the driver even if some uart are still attached to it.
> >
> > You should probably do like in pl010_remove() and remove only if the last one is removed.
> >
>
> Indeed... multiple ports can be registered...
>
> > > return 0;
> > > }
> > > @@ -809,33 +809,7 @@ static struct platform_driver meson_uart_platform_driver = {
> > > },
> > > };
> > > -static int __init meson_uart_init(void)
> > > -{
> > > - int ret;
> > > -
> > > - ret = meson_serial_console_init();
> > > - if (ret)
> > > - return ret;
> > > -
> > > - ret = uart_register_driver(&meson_uart_driver);
> > > - if (ret)
> > > - return ret;
> > > -
> > > - ret = platform_driver_register(&meson_uart_platform_driver);
> > > - if (ret)
> > > - uart_unregister_driver(&meson_uart_driver);
> > > -
> > > - return ret;
> > > -}
> > > -
> > > -static void __exit meson_uart_exit(void)
> > > -{
> > > - platform_driver_unregister(&meson_uart_platform_driver);
> > > - uart_unregister_driver(&meson_uart_driver);
> > > -}
> > > -
> > > -module_init(meson_uart_init);
> > > -module_exit(meson_uart_exit);
> > > +module_platform_driver(meson_uart_platform_driver);
> >
> > Only pl010 uses this scheme, and I don't know why... if it works then it's ok for me.
>
> From my point of view, the "scheme" is using uart driver registration
> from the probe() routine. Many drivers are based on such approach:
> samsung-tty, timbuart, sprd, max3100, etc. Some of them are platform
> drivers as well.
>
> > > MODULE_AUTHOR("Carlo Caione <carlo@xxxxxxxxxx>");
> > > MODULE_DESCRIPTION("Amlogic Meson serial port driver");
> >
>
> --
> Thank you,
> Dmitry
--
Thank you,
Dmitry
