Re: Streaming DMA mapping question

[Adam Kropelin <akropel1@rochester.rr.com>]

On Fri, Sep 13, 2002 at 12:36:41PM -0700, David S. Miller wrote:
>    From: Adam Kropelin <akropel1@rochester.rr.com>
>    Date: Fri, 13 Sep 2002 15:39:16 -0400
>    
>    According to the docs, you should either unmap or sync your DMA buffer before
>    touching it from the host. The i386 implementation of pci_unmap is empty --no
>    problem; there must not be any unmap work to do on this arch. But the 
>    implementation of pci_dma_sync does contain a flush_write_buffers() call. This
>    makes me think that perhaps if I'm going to modify the buffer before I submit it
>    back to the controller I need to do:
> 
> Actually, rather it appears that the i386 pci_unmap_*() routines need
> the write buffer flush as well.

Ah, a bug then. 

> The x86 implementation is a bad example to read if you're trying to
> see what the worst case scenerio is.

I was looking at the x86 implementation to help me narrow down the possible
source of a bug I'm seeing in the driver. I noticed the driver was examining a
DMA buffer without unmapping or syncing. Before I went to the work to fix the
code I wanted to see if those operations were no-ops on the platform I'm testing
on. I'll definitely fix the driver regardless, but I didn't want to blame the
lockup on it if unmapping won't actually change anything on this platform. Thus
my confusion upon seeing pci_unmap_*() and pci_dma_sync_*() differ.

> Just follow the document and your driver will work properly on all
> platforms.  DMA-mapping.txt was meant to be written in a way such
> that you should not ever need to look at an implementation of the
> interfaces to figure out how to use them.

Absolutely. It does that job very well and I am greatly appreciative for the
code groveling time that document has saved me. Kudos to you and others who
spent time writing it.

--Adam