svn commit: r188350 - in head/sys: amd64/amd64 arm/arm dev/usb2/core i386/i386 ia64/ia64 sys

[Scott Long]

Marcel Moolenaar wrote:
> 
> On Feb 9, 2009, at 10:42 AM, John Baldwin wrote:
> 
>>>> The 'align' parameter to bus_dma_tag_create().  If your hardware
>>>> needs buffers
>>>> to be aligned on a 4-byte boundary and you bus_dmamap_load() a buffer
>>>> where 'addr % 4 != 0', then the buffer is bounced.  Since by default
>>>> the new
>>>> buffer starts on a page boundary, it satifies the 'addr % 4'.
>>>
>>> But according to the man page, bounce buffering may not
>>> be implemented or not be applicable to a platform. It
>>> seems to me that you cannot depend on this side-effect
>>> in a generic driver. Are you guys talking only in terms
>>> of i386 or is this generally applicable?
>>
>> All of the platforms that do bounce buffering will honor this.  If a 
>> given
>> platform doesn't support bounce buffering then I imagine it will fail
>> misaligned requests with an error.  It would probably make it easier to
>> support bounce buffering on more platforms if some of the support code 
>> for
>> managing bounce zones could be moved into a subr_busdma.c or the like 
>> since
>> much of it is copy and pasted.
> 
> Very much agreed with the unification in subr_busdma.c.
> 
> If the shared implementation then grows a IOMMU interface,
> then platforms with a IOMMU can avoid bounce buffering
> when possible (i.e. no re-alignment is needed or enough
> free IOMMU mappings exist to avoid the bounce buffering).
> 

Yes, this work in progress.

Scott