On Thu, Aug 25, 2022 at 11:04:43AM -0500, Bjorn Helgaas wrote: > On Thu, Aug 25, 2022 at 08:16:14AM +0300, Serge Semin wrote: > > On Wed, Aug 24, 2022 at 01:17:54PM -0500, Bjorn Helgaas wrote: > > > On Wed, Aug 24, 2022 at 09:13:19PM +0300, Serge Semin wrote: > > > > On Wed, Aug 24, 2022 at 11:51:18AM -0500, Bjorn Helgaas wrote: > > > > > On Mon, Aug 22, 2022 at 09:53:32PM +0300, Serge Semin wrote: > > > > > > > > > + val = dw_pcie_readl_dbi(pci, PCIE_DMA_VIEWPORT_BASE + PCIE_DMA_CTRL); > > > > > > + if (val == 0xFFFFFFFF && pci->edma.reg_base) { > > > > > > + pci->edma.mf = EDMA_MF_EDMA_UNROLL; > > > > > > + > > > > > > + val = dw_pcie_readl_dma(pci, PCIE_DMA_CTRL); > > > > > > + } else if (val != 0xFFFFFFFF) { > > > > > > > > > > > > > > Consider PCI_POSSIBLE_ERROR() as an annotation about the meaning of > > > > > 0xFFFFFFFF and something to grep for. > > > > > > > > In this case FFs don't mean an error but a special value, which > > > > indicates that the eDMA is mapped via the unrolled CSRs space. The > > > > similar approach has been implemented for the iATU legacy/unroll setup > > > > auto-detection. So I don't see much reasons to have it grepped, so as > > > > to have a macro-based parametrization since the special value will > > > > unluckily change while having the explicit literal utilized gives a > > > > better understanding of the way the algorithm works. > > > > > If 0xFFFFFFFF is the result of a successful PCIe Memory Read, > > > > Right. It is. > > > > > and not > > > something synthesized by the host bridge when it handles an > > > Unsupported Request completion, > > > > No it isn't. To be clear 0xFFs don't indicate some PCIe bus/controller > > malfunction, but they are a result of reading the > > DMA_CTRL_VIEWPORT_OFF register which doesn't exist. The manual > > explicitly says: "Note - When register does not exist, value is fixed > > to 32'hFFFF_FFFF". The register doesn't exist if either eDMA is > > unavailable or the eDMA CSRs are mapped via the unrolled state. > > OK. I don't think that's worded very well in the manual. A register > that does not exist does not have a value, and attempts to read it > should fail. No. The manual explicitly says that this particular CSR (DMA_CTRL_VIEWPORT_OFF) value is tied to 32'hFFFF_FFFF if the register doesn't exist. There is no such text mentioned for any other non-existing CSR. > If they want to say the register always exists and > contains 0xFFFFFFFF for versions earlier than X, that would make > sense. Wouldn't be the first time a manual is ambiguous ;) They say, that the register doesn't exist if either eDMA isn't available or it's mapped via the unrolled CSR space. There is no reference to the IP-core version. Anyway basically you are right they indeed imply that the register always exists. > > If the device itself, i.e., not the Root Complex, is fabricating this > 0xFFFFFFFF value, reading it should not cause any AER or other error > status bits to be set. > > If the Root Complex fabricates 0xFFFFFFFF upon receipt of a Completion > with Unsupported Request status, I would expect bits like Received > Master Abort to be set in the Root Port's Secondary Status register. The device CSRs are accessed by means of the controller DBI-interface. Even though the whole CSRs space do look as the extended PCIe config space, the DBI-based access isn't tracked by the standard PCIe capabilities like AER. So in case of the eDMA auto-detection procedure introduced in this patch (and iATU auto-detection, which is already available in the DW PCIe driver ) we won't have any bus error status raised. -Sergey