On Thu, Oct 07, 2021 at 02:36:25PM +0200, Marek Behún wrote: > On Thu, 7 Oct 2021 12:51:57 +0100 > Lorenzo Pieralisi <lorenzo.pieralisi@xxxxxxx> wrote: > > > On Wed, Oct 06, 2021 at 03:13:05PM -0500, Bjorn Helgaas wrote: > > > > [...] > > > > > > We need a way for those PCI controllers to communicate to SW that > > > > they actually received a CRS completion (and that they don't retry > > > > in HW). > > > > > > AFAICT this would be controller-dependent. I think some controllers > > > have registers that control the number of retries, but of course > > > that's completely controller-dependent, too. > > > > > > > By implementing the logic in the aardvark controller that platform > > > > information is there so to the best of my knowledge this patch > > > > is sound. > > > > > > > > I assume that the HW retry is in the specs because there is no > > > > architected way if CRS Software Visibility is not enabled/present to > > > > report CRS completion in an architected PCI manner but I just > > > > don't know the entire background behind this. > > > > > > I assume an error bit would be set in the Status or Secondary Status > > > register when a PCIe transaction fails. I'm not sure anybody *looks* > > > at those, though. > > > > > > This still looks like a PCI controller band-aid for a device or driver > > > problem that will likely still exist on other platforms. > > > > Yes that's true. I believe we can merge this patch (?) but it would > > also be good if Pali/Marek have time to test on other HW and > > maybe generalize the concept. > > We are willing to try to implement a generic API if you propose should > such API look (at least some hints). That's the whole point -- CRS SV *is* the generic way for controllers to report CRS completions to software. If the controller doesn't support CRS SV, I don't think a generic API is possible. Or maybe I don't understand the sort of generic API you have in mind. What sort of information do you envision the API might provide? Bjorn
