Re: [PATCH 0/6] get rid of on-stack dma buffers

[Florian Mickler]

2011/3/22 James Bottomley <James.Bottomley@xxxxxxxxxxxxxxxxxxxxx>:
> On Mon, 2011-03-21 at 22:03 +0100, Florian Mickler wrote:
>> On Mon, 21 Mar 2011 15:26:43 -0400
>> Andy Walls <awalls@xxxxxxxxxxxxxxxx> wrote:
>>
>> > Florian Mickler <florian@xxxxxxxxxxx> wrote:
>>
>> To be blunt, I'm not shure I fully understand the requirements myself.
>> But as far as I grasped it, the main problem is that we need memory
>> which the processor can see as soon as the device has scribbled upon
>> it. (think caches and the like)
>>
>> Somewhere down the line, the buffer to usb_control_msg get's to be
>> a parameter to dma_map_single which is described as part of
>> the DMA API in Documentation/DMA-API.txt
>>
>> The main point I filter out from that is that the memory has to begin
>> exactly at a cache line boundary...
>
> The API will round up so that the correct region covers the API.
> However, if you have other structures packed into the space (as very
> often happens on stack), you get cache line interference in the CPU if
> they get accessed:  The act of accessing an adjacent object pulls in
> cache above your object and destroys DMA coherence.  This is the
> principle reason why DMA to stack is a bad idea.

Thanks, this was the missing piece of information to make sense of
 why it's bad for stack memory to be part of this.

>
>> I guess (not verified), that the dma api takes sufficient precautions
>> to abort the dma transfer if a timeout happens.  So freeing _should_
>> not be an issue. (At least, I would expect big fat warnings everywhere
>> if that were the case)

I did mean s/dma api/usb_control_msg/ in the above paragraph. As that is the
 ''dma api'' these drivers are using... sorry for the confusion there...

>
> No, it doesn't take any precautions like this.  the DMA API is just
> mapping (possibly via an IOMMU).  If the transfer times out, that's done
> in the DMA engine of the card, and must be cleaned up by the driver and
> unmapped.

ok.

> The general rule though is never DMA to stack.  On some processors, the
> way stack is allocated can actually make this not work.
>
> James

thanks,
Flo

p.s.: hope this message get's through to the list... I am on the road
at the moment,
 so I'm not shure that there won't be any html in it again :(
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/