On Feb 5, 12:46am, Ralf Baechle wrote:
> Subject: pci_* API
> It seems the consens is that they want to go for a second set of functions
> named pci64_* and safe the hand full of cycles it takes to check the
> flags.
>
Compared to the cost of doing a dma, the cycles spent on checking
flags are insignificant. This approach strikes me as very near
sighted, but I suppose that kind of argument doesn't go very far
in the Linux community. How about the kernel-bloat argument? Would
that sway them? What about the cost of an i-cache miss on account
of having these additional, highly redundant routines?
I think it's a mistake to introduce such a rigid interface when there
is opportunity to make it more flexible at less cost. If they are
truly interested in performance then they should introduce an
interface that allows different platforms to easily provide performance
features that tailored for the platform.
- leo
--
Leo Dagum SGI Mountain View, CA 94043 (650-933-2179)
|