Michael Buesch wrote:
> On Wednesday 28 March 2007 19:42, Larry Finger wrote:
>> Will Dyson wrote:
>>> On 3/24/07, Michael Buesch <[EMAIL PROTECTED]> wrote:
>>>
>>>> I _still_ think that this should be fixed in the arch DMA code
>>>> and _not_ in every single driver on earth. But I discussed this in
>>>> the past.
>>> That would be nice. What would that look like to you? A GFP flag that
>>> gets memory within the dma_mask?
>> To me that would mean honoring the dma_mask for _ALL_ DMA memory 
>> allocations. I have not followed
>> all the arguments as to why this was never fixed in the i386 code,
> 
> On i386 we are fine, as long as we don't use scatter-gather IO for the
> DMA operations. If we kmalloc() the DMA buffers, we are implicitely
> guaranteed (as long as nobody changed the user/kernel split) to be
> under the 1G barrier (actually the barrier is at 1G minus 128M).
> 
> Although, people want to implement SG-IO for mac80211. That
> would mean buffers are not copied to kernelspace, so they could
> be above 1G.
> 
>> but I know that the x86_64 
>> architecture was forced to produce the same (wrong) results as were obtained 
>> in the i386-specific
>> code. I agree that it should be fixed; however, it will never happen.
> 
> Explain why, please. I'm very interrested in knowing why this
> is so fucking broken and why nobody fixes it. ;)

John - do you know the history?

>> As a result, every driver for 
>> hardware with less than 32-bit DMA addressing has to address (pun intended) 
>> the problem. We will
>> probably be fighting the problem again when > 4 GB RAM computers become 
>> common and 32 bits are not
>> enough to address all of memory.
> 
> Nope, see Johannes' mail.

Yes, if an IOMMU is included, but why don't we have them now on our 64-bit 
processors?

Larry
_______________________________________________
Bcm43xx-dev mailing list
[email protected]
https://lists.berlios.de/mailman/listinfo/bcm43xx-dev

Reply via email to