Re: [PATCH] 2.6.1 tg3 DMA engine test failure

[Grant Grundler <grundler@parisc-linux.org>]

On Fri, Jan 23, 2004 at 09:00:23PM -0800, David S. Miller wrote:
>    From: Grant Grundler <grundler@parisc-linux.org>
>    Date: Fri, 23 Jan 2004 18:36:14 -0700
> 
>    3) Broadcom engineer noted the meaning of DMA_RWCTRL_ASSERT_ALL_BE
>       has changed for bcm570[34] and also advised against setting
>       it on BCM570[01] chips. I'm just implementing his advice.
>       Comment below spells out more details.
> 
> Setting this bit is absolutely required on many RISC PCI boxes, where
> streaming mappings must have cacheline sized DMA transactions done
> on them with all byte enables on.

My gut feeling is if linux aligns or pads things nicely for any reason,
then the bye enables don't get used or clobber padding.

> In fact, since the later chips don't allow controlling this, some of
> them cause streaming byte hole errors on sparc64 and other RISC
> systems when they do cacheline sized DMA to streaming DMA mappings
> with not all the byte enables on.

yup.

> So I'm not going to add this part of your changes.

No problem. My take is, if it hasn't caused a problem on x86 because
things are well aligned, then no reason to change the code.
Knowing the issues about other RISC archs 

Maybe keep a shorter note about the bit changed meaning in later models
just to document the issues.

thanks,
grant