On Tue, Jun 13, 2017 at 10:00:33AM -0700, Yinghai Lu wrote: > On Mon, Jun 12, 2017 at 2:48 PM, Bjorn Helgaas <helg...@kernel.org> wrote: > > On Fri, Jun 09, 2017 at 04:16:17PM -0700, Yinghai Lu wrote: > >> From: James Puthukattukaran <james.puthukattuka...@oracle.com> > >> > >> The IDT switch incorrectly flags an ACS source violation on a read config > >> request to an end point device on the completion (IDT 89H32H8G3-YC, > >> errata #36) even though the PCI Express spec states that completions are > >> never affected by ACS source violation (PCI Spec 3.1, Section 6.12.1.1). > ...
> > Have you considered ways to make this patch apply only to the affected > > IDT switches? Currently it applies to *all* devices. > > But we need to apply that workaround before we know vendorid/deviceid > under that root port or downstream port. If I understand correctly, the problem (the ACS source violation) occurs when we enumerate a device below an IDT switch. We enumerate a switch before we enumerate any devices under the switch, so I don't understand why this can't be made IDT-specific. > > The purpose of the pci_bus_read_dev_vendor_id() path is to support the > > Configuration Request Retry Status feature (see PCIe r3.1, sec 2.3.2), > > which works by special handling of config reads of the Vendor ID after > > a reset. Normally, that Vendor ID read would be the first access to > > a device when we enumerate it. > > > > This patch adds config reads and writes of the ACS capability *before* > > the Vendor ID read. At that point we don't even know whether the > > device exists. If it doesn't exist, pci_find_ext_capability() would > > read 0xffffffff data, and it probably fails reasonably gracefully. > > > > But if the device *does* exist, I think this patch breaks the CRS > > Software Visibility feature. Without this patch, we try to read > > Vendor ID, and the device may return a CRS Completion Status. If CRS > > visibility is enabled, the root complex may complete the request by > > returning 0x0001 for the Vendor ID, in which case we sleep and try > > again later. > > > > With this patch, we first try to read the ACS capability. If the > > device returns a CRS Completion Status, the root complex is required > > to reissue the config request. This is the required behavior > > regardless of whether CRS Software Visibility is enabled, so I think > > this effectively disables that feature. > > The workaround (acs reading etc) is applied to root port or downstream port. > and pci_bus_read_dev_vendor_id() is for reading vendorid of device > under that root port or downstream port. OK, I see what you're saying: pci_std_enable_acs_sv() twiddles the ACS capability of the upstream bridge when we're enumerating a device *below* the bridge. Since the bridge has already been enumerated, we've already read its Vendor ID, so looking up its ACS capability should not cause any CRS completions. Still, I think all this fiddling around with the ACS capability is too intrusive for non-IDT devices. Caching the ACS capability location would help a little bit, but it's still too much in my opinion. If ACS is broken on the IDT switch, one obvious possibility is a quirk to disable use of ACS on those switches. The current patch appears to penalize everybody for the sins of IDT, which I'd like to avoid. Bjorn
