Re: pci_set_mwi() and ath5k

["Luis R. Rodriguez" <mcgrof@gmail.com>]

On Wed, Nov 4, 2009 at 2:45 PM, Luis R. Rodriguez <mcgrof@gmail.com> wrote:
> On Wed, Nov 4, 2009 at 2:31 PM, Nick Kossifidis <mickflemm@gmail.com> wrote:
>> 2009/11/5 Luis R. Rodriguez <mcgrof@bombadil.infradead.org>:
>>> On Wed, Nov 04, 2009 at 02:04:11PM -0800, Luis R. Rodriguez wrote:
>>>> On Wed, Nov 4, 2009 at 2:00 PM, Matthew Wilcox <matthew@wil.cx> wrote:
>>>> > On Wed, Nov 04, 2009 at 01:52:30PM -0800, Luis R. Rodriguez wrote:
>>>> >> > Even better: I just confirmation from our systems team that our legacy
>>>> >> > devices and 11n PCI devices don't support MWR so I'll remove all that
>>>> >> > cruft crap.
>>>> >>
>>>> >> I meant MWI of course.
>>>> >
>>>> > Yes, but they don't necessarily just use cacheline size for MWI ... some
>>>> > devices use cacheline size for setting up data structures.  Might be
>>>> > worth just checking explicitly that they don't use the cacheline size
>>>> > register for anything.
>>>>
>>>> Oh right -- so the typical Atheros hack for this is to check the cache
>>>> line size, and if its 0 set it to L1_CACHE_BYTES. Then eventually read
>>>> from PCI_CACHE_LINE_SIZE pci config to align the skb data. So what I
>>>> was doing now is removing all this cruft and replacing it with a
>>>> generic allocator for atheros drivers that aligns simply to the
>>>> L1_CACHE_BYTES. Sound kosher?
>>>
>>> Something like this:
>>>
>>
>> According to comments inside MadWiFi AR5210 needs cache line align
>> else we get corruptions.
>
> For what though?
>
>> I don't know if this is correct for all
>> platforms or later cards but since we (plan to) support AR5210 i guess
>> we should leave it there. We need to test this a lot on various
>> archs/cards before applying it.
>
> There are two cases where we can use the PCI_CACHE_LINE_SIZE:
>
> 1) Hardware has been designed to use it on some block to align some data somehow
> 2) Software uses it to align skb->data for performance/hw purposes
>
> I believe the second case can be handled by using L1_CACHE_BYTES
> instead and I'd at least like to change our common skb allocator to
> use that.
>
> The first case is where it seems there may be some skepticism as to
> whether or not hw really did not rely on it and I agree its safer to
> keep the programming of the  PCI_CACHE_LINE_SIZE in case it has a
> bogus value.
>
> Does this seem reasonable?

I guess we can keep the second case allocator which relies on
PCI_CACHE_LINE_SIZE, but if so then USB will just define its own csz
statically always to L1_CACHE_BYTES which is fine too -- I was just
hoping we could remove the PCI_CACHE_LINE_SIZE programming completely
too and use a simple L1_CACHE_BYTES aligned allocator.

  Luis