Hi David,
On Tuesday, 26 September 2017 21:53:21 PDT David Miller wrote:
> From: Paul Burton <paul.burton@xxxxxxxxxx>
> Date: Tue, 26 Sep 2017 21:30:56 -0700
>
> > Nobody said that you are required to do anything, I suggested that
> > it would be beneficial if you were to suggest a change to the
> > documented DMA API such that it allows your usage where it currently
> > does not.
>
> Documentation is often wrong and it is here. What 200+ drivers
> actually do and depend upon trumps a simple text document.
Agreed - but if the documented API is wrong then it should change.
> The requirement is that the memory remains quiescent on the cpu side
> while the device messes with it. And that this quiescence requirement
> may or may not be on a cache line basis.
>
> There is absolutely no requirement that the buffers themselves are
> cache line aligned.
>
> In fact, receive buffers for networking are intentionally 2-byte
> aligned in order for the ipv4 headers to be naturally 32-bit aligned.
>
> Cache line aligning receive buffers will actually make some
> architectures trap because of the bad alignment.
>
> So see, this cache line alignment requirement is pure madness from
> just about any perspective whatsoever.
Thank you - that's more constructive.
I understand that the network code doesn't suffer from a problem with using
non-cacheline-aligned buffers, because you guarantee that the CPU will not be
writing to anything on either side of the memory mapped for DMA up to at least
a cache line boundary. That is all well & good (though still, I think that
ought to be documented somewhere even if just by a comment somewhere in linux/
sk_buff.h).
There is still a problem though in other cases which do not provide such a
guarantee - for example the MMC issue I pointed out previously - which it
would be useful to be able to catch rather than allowing silent memory
corruption which can be difficult to track down. Whilst the particular case of
mapping a struct sk_buff's data for DMA might not trigger this issue the issue
does still exist in other cases for which aligning data to a cache line
boundary is not always "pure madness".
So whilst it sounds like you'd happily just change or remove the paragraph
about cache-line alignment in Documentation/DMA-API.txt, and I agree that
would be a good start, I wonder whether we could do something better. One
option might be for the warning in the MIPS DMA code to be predicated on one
of the cache lines only partially covered by a DMA mapping actually being
dirty - though this would probably be a more expensive check than we'd want in
the general case so might have to be conditional upon some new debug entry in
Kconfig. I'll give this some thought.
Thanks,
PaulAttachment:
signature.asc
Description: This is a digitally signed message part.
