On Fri, Sep 18, 2020 at 11:02:07AM +0000, Z.q. Hou wrote: > But now the SError is exactly caused by the first access of the > non-existent function, I dug into the kernel enumeration code and > found it will fire a 4Byte CFG read transaction to read the Vendor > ID and Device ID together, so I suspect the root cause is access the > Device ID of a non-existent function triggers SError. > > So the alternative solution seems to correct the PCIe enumeration, I > will submit a patch to let the first access only read the Vendor ID. If it is incorrect for the first access to be a 32-bit read of both the Vendor and the Device ID, please cite the relevant section of the spec in your patch. I don't like to make changes to generic code to accommodate specific pieces of hardware because then we restrict future changes based on some device that will soon be obsolete and forgotten. I'm pretty sure the spec language about CRS handling is careful to talk about "reads that *include* Vendor ID", not just "reads of Vendor ID", so the implication is that it covers 32-bit reads as well as 16-bit reads. Bjorn
