Re: cache coherency in DMA and MMIO

[Samuel Thibault]

Da Zheng, le Sat 24 Apr 2010 18:29:50 +0800, a écrit :
> On 10-4-23 下午7:37, Samuel Thibault wrote:
> > Da Zheng, le Fri 23 Apr 2010 14:51:22 +0800, a écrit :
> >> They are not just one or two variables. They are an array of structure
> >> variables. Should I define all fields of the structure with volatile 
> >> qualifier?
> > 
> > You can qualify the whole structure
> > 
> >> It's strange. The Linux driver doesn't do so but it definitely works in the
> >> Linux kernel.
> > 
> > It does (or should do): such pointers have the __io qualifier.
> I checked the code of e1000 again. The data for MMIO is qualified by __iomem,

Right, I hadn't checked the precise name :)

> but the data in the transmission or receiving queues (which are accessed by 
> DMA)
> doesn't have __iomem or any similar qualifiers.

Because they are only accessed by DMA, aren't they?

> I guess the data in the transmission queue is very unlikely to be in
> registers when the driver accesses it.

You said yourself that DMAs take care of cache coherency :)
(and I indeed guess they do).

Samuel