use {readl|writel}_relaxed instead of readl/writel in i2c-designware-core ?

use {readl|writel}_relaxed instead of readl/writel in i2c-designware-core ?

[Jisheng Zhang]

Hi all,

The writel/readl is too expensive especially on Cortex A9 w/ outer L2 cache. This
introduce i2c read/write error on Marvell Berlin SoCs when there are L2 cache
maintenance operations at the same time.

In our internal berlin bsp, we just replaced readl/writel with the relaxed
version. But AFAIK, the "relaxed" version doesn't exist on all architectures. How
to handle this issue? 

Any suggestions are appreciated.

Thanks in advance,
Jisheng

use {readl|writel}_relaxed instead of readl/writel in i2c-designware-core ?

[Arnd Bergmann]

On Friday 14 February 2014 15:54:38 Jisheng Zhang wrote:
> Hi all,
> 
> The writel/readl is too expensive especially on Cortex A9 w/ outer L2 cache. This
> introduce i2c read/write error on Marvell Berlin SoCs when there are L2 cache
> maintenance operations at the same time.
> 
> In our internal berlin bsp, we just replaced readl/writel with the relaxed
> version. But AFAIK, the "relaxed" version doesn't exist on all architectures. How
> to handle this issue? 

In case of i2c-designware, this is safe because that driver does not perform
DMA. In other drivers, you may have to be more careful, to ensure that all MMIO
is serialized with DMA operations performed by the driver.

> Any suggestions are appreciated.

I would definitely welcome a patch that adds a default  _relaxed implementation
to include/linux/io.h, like this:

#ifndef readb_relaxed
#define readb_relaxed(p) readb(p)
#endif

and then adds "#define readb_relaxed(p) readb_relaxed(p)" etc. to all
architectures that have a non-macro definition for readb.

Alternatively, we could have a CONFIG_ARCH_MMIO_RELAXED configuration
symbol that gets selected by any architecture that provides the _relaxed
accessors, and get linux/io.h to define all of them for the other
architectures.

	Arnd

use {readl|writel}_relaxed instead of readl/writel in i2c-designware-core ?

[Jisheng Zhang]

Dear Arnd,

On Fri, 14 Feb 2014 01:09:44 -0800
Arnd Bergmann <arnd@arndb.de> wrote:

> On Friday 14 February 2014 15:54:38 Jisheng Zhang wrote:
> > Hi all,
> > 
> > The writel/readl is too expensive especially on Cortex A9 w/ outer L2
> > cache. This introduce i2c read/write error on Marvell Berlin SoCs when
> > there are L2 cache maintenance operations at the same time.
> > 
> > In our internal berlin bsp, we just replaced readl/writel with the relaxed
> > version. But AFAIK, the "relaxed" version doesn't exist on all
> > architectures. How to handle this issue? 
> 
> In case of i2c-designware, this is safe because that driver does not perform
> DMA. In other drivers, you may have to be more careful, to ensure that all
> MMIO is serialized with DMA operations performed by the driver.
> 
> > Any suggestions are appreciated.
> 
> I would definitely welcome a patch that adds a default  _relaxed
> implementation to include/linux/io.h, like this:
> 
> #ifndef readb_relaxed
> #define readb_relaxed(p) readb(p)
> #endif
> 
> and then adds "#define readb_relaxed(p) readb_relaxed(p)" etc. to all
> architectures that have a non-macro definition for readb.

Thanks for the suggestions. I'll try to send out one RFC patch.

Thanks,
Jisheng