Re: [PATCH] EHCI Oops on CONFIG_NOT_COHERENT_CACHE system

On Aug 30, 2006, at 7:04 AM, Marcus Comstedt wrote:

"Gerhard Pircher" <gerhard_pircher at gmx.net> writes:

quoted

I guess this message should also be forwarded to
linux-usb-devel at lists.sourceforge.net. I hope the developers there
can make some comments.

Well, I figured the first thing to do would be to reach a consensus
here on whether or not atomic_t:s in DMA memory should be ruled as
unallowed.  Judging from the massive silence, there doesn't seem to be
any strong opinions either way though...

I think its a case of the people that know the architecture agree
that atomics and DMA do not mix, and that this is an issue for
usb-devel to fix.

The other question is: does the current access polute the cache by
reading things where the device is supposed to be writing?

I think some on that list are familiar with the issues for incoherent
DMA, as some of the other usb drivers have been cleaned up in that
regard.

My suggestion is take the issue over there, and only cc us here if
you get pushback over there.

milton

Milton Miller [off-list ref] writes:

I think its a case of the people that know the architecture agree
that atomics and DMA do not mix, and that this is an issue for
usb-devel to fix.

That was precisely the kind of agreement I was looking for.  :-)

The other question is: does the current access polute the cache by
reading things where the device is supposed to be writing?

Well, that shouldn't be an issue as the page is cache-inhibited.
And on systems with cache coherency, where the pages are not cache-
inhibited, this would be handled by bus snooping, right?

My suggestion is take the issue over there, and only cc us here if
you get pushback over there.

Ok, will do.  Thanks.


  // Marcus