On Fri, Oct 23, 2026 at 04:00:36PM -0300, Jason Gunthorpe wrote: > On Fri, Oct 23, 2026 at 01:27:15PM -0500, Bjorn Helgaas wrote: > > > > Yes, the ACS combination is selected to avoid ordering issues > > > > Does a user have any way to know whether the configuration is safe? > > No, unfortunately in PCIe so much is really difficult when it comes to > deployment :( > > > I'm wondering whether users might use "pci=config_acs=" to configure > > ACS so things appear to work as they prefer but may not provide the > > ordering guarantees drivers depend on. > > I imagine that could happen, but on most practical systems the root > complex will blow up if you do asymmetric routing so there is an early > indication. > > I know this because I've help debug such configurations at > customers. Some servers helpfully MCE if you do this. Lovely.
I'm still confused about this because earlier in this thread, I thought Leon said you do asymmetric routing: Bjorn> And I guess asymmetric fabrics must mean fabrics where Bjorn> Request Redirect and Completion Redirect are not set the same Bjorn> way? Leon> Yes, in some of our systems, deviceA to deviceB is going Leon> through different route than deviceB to deviceA. > > If that's a possibility, do we have any notes in dmesg that we are > > depending on fabric behavior that the OS can't discover and verify? > > Maybe, it could be interesting if the P2P logic found asymetric paths > it refuses to permit it. It seems like P2PDMA *could* check for asymmetric paths and at least warn about it. If it's practical to do so, it would make me feel a lot better about this.
