Re: [Xenomai-help] rtdm_iomap_to_user question

[Jan Kiszka <jan.kiszka@domain.hid>]

[-- Attachment #1: Type: text/plain, Size: 2455 bytes --]

Zhang, Haobo wrote:
> Hi Jan,
> 
> Thank you so much for your reply!
> 
> In my driver, I wrote:
> 
> int map_io( struct gpiodrv_context *dev_context, rtdm_user_info_t
> *user_info, struct MAPPED_IO *io )
> {
> 	int ret;
> 
> 	if ((dev_context == 0) || (user_info == 0) || (io == 0))
> 		return -EFAULT;
> 
> 	if (dev_context->mapped_user_info != 0)
> 	{
> 		if ((ret = gpio_unmap_io( dev_context )) != 0)
> 			return ret;
> 	}
> 
> 	dev_context->mapped_user_info = user_info;
> 
> 	ret = rtdm_iomap_to_user( user_info, MPC52xx_PA(
> MPC52xx_GPIO_WKUP_OFFSET + 0x0C ),
> 		1, PROT_READ | PROT_WRITE, (void
> **)&(io->gpio_heart_beat), NULL, NULL );

Hmm, hmm. Looking at our code and the kernel again I would say we should
catch non-page-aligned requests like this and document this pitfall better.

The mapping code always assumes page-alignment and rounds your physical
address down. So what likely happened here is that you got a mapping,
but with unexpected offset (gpio_heart_beat should point to MPC52xx_PA).

> 	
> 	if (ret != 0)
> 		goto __ERROR;
> 	else
> 		dev_context->mapped_heart_beat = (void
> *)io->gpio_heart_beat;
> 
> 	return 0;
> 
> __ERROR:
> 
> 	gpio_unmap_io( dev_context );
> 	return ret;
> } 
> 
> In the application, I wrote:
> 
> int main(int argc, char *argv[]) 
> {
> 	int dev_gpio;
> 	struct MAPPED_IO gpio;
> 
> 	signal(SIGTERM, clean_exit);	
> 	signal(SIGINT, clean_exit);	
> 	
> 	mlockall( MCL_CURRENT | MCL_FUTURE );
> 
> 	rt_task_shadow( &mainapp, "mainapp", 1, 0 );
> 
> 	dev_gpio = rt_dev_open( "gpio", 0 );
> 	if (dev_gpio < 0)
> 	{
> 		return -1;
> 	}
> 	
> 	gpio.gpio_heart_beat = NULL;
> 
> 	if (rt_dev_ioctl( dev_gpio, 1, &gpio ) != 0)
> 	{
> 		return -1;
> 	}
> 
> 	gpio.gpio_heart_beat[0] = 0x01;
> 
> 	...
> }
> 
> I also tried another way by opening "/dev/mem" and then using "mmap" to
> map the I/O memory addresses in the application. It worked! 
> I am very curious about why "rtdm_iomap_to_user" did not work for me. 

When you do precisely the same odd-numbered mapping via mmap on
/dev/mem, you do get the right address at gpio_heart_beat (check if the
start address ends with 0x0b0c)?

We either have to improve rtdm_iomap/mmap_to_user in this regard or
catch unaligned requests. Probably it's easier to catch it so that the
driver writer also realises that this is always about mapping full pages...

Jan


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 250 bytes --]