On Thu, Feb 15, 2024 at 11:21:45AM +0100, Konrad Dybcio wrote: > On 14.02.2024 23:28, Bjorn Helgaas wrote: > > On Wed, Feb 14, 2024 at 10:35:16PM +0100, Konrad Dybcio wrote: > >> On 12.02.2024 22:17, Bjorn Helgaas wrote: > >>> Maybe include the reason in the subject? "Read back" is literally > >>> what the diff says. > >>> > >>> On Sat, Feb 10, 2024 at 06:10:06PM +0100, Konrad Dybcio wrote: > >>>> To ensure write completion, read the PARF_LTSSM register after setting > >>>> the LTSSM enable bit before polling for "link up". > >>> > >>> The write will obviously complete *some* time; I assume the point is > >>> that it's important for it to complete before some other event, and it > >>> would be nice to know why that's important. > >> > >> Right, that's very much meaningful on non-total-store-ordering > >> architectures, like arm64, where the CPU receives a store instruction, > >> but that does not necessarily impact the memory/MMIO state immediately. > > > > I was hinting that maybe we could say what the other event is, or what > > problem this solves? E.g., maybe it's as simple as "there's no point > > in polling for link up until after the PARF_LTSSM store completes." > > > > But while the read of PARF_LTSSM might reduce the number of "is the > > link up" polls, it probably wouldn't speed anything up otherwise, so I > > suspect there's an actual functional reason for this patch, and that's > > what I'm getting at. > > So, the register containing the "enable switch" (PARF_LTSSM) can (due > to the armv8 memory model) be "written" but not "change the value of > memory/mmio from the perspective of other (non-CPU) memory-readers > (such as the MMIO-mapped PCI controller itself)". > > In that case, the CPU will happily continue calling qcom_pcie_link_up() > in a loop, waiting for the PCIe controller to bring the link up, however > the PCIE controller may have never received the PARF_LTSSM "enable link" > write by the time we decide to time out on checking the link status. > > It may also never happen for you, but that's exactly like a classic race > condition, where it may simply not manifest due to the code around the > problematic lines hiding it. It may also only manifest on certain CPU > cores that try to be smarter than you and keep reordering/delaying > instructions if they don't seem immediately necessary. Does this mean the register is mapped incorrectly, e.g., I see arm64 has many different kinds of mappings for cacheability, write-buffering, etc? Or, if it is already mapped correctly, are we confident that none of the *other* register writes need similar treatment? Is there a rule we can apply to know when the read-after-write is needed? Bjorn