On Sat, Jun 4, 2011 at 7:14 PM, Benjamin Herrenschmidt <benh@xxxxxxxxxxxxxxxxxxx> wrote: > On Fri, 2011-06-03 at 11:00 -0500, Jon Mason wrote: >> >> You are correct. I was only thinking about the DMA write. The entire >> fabric needs have each device/bridge MPS set to the smallest MPS >> found, otherwise DMA reads could be too large for a device on the >> fabric and cause an PCI unrecoverable error. >> >> So the correct solution is to find the smallest MPS per fabric while >> walking the bus, then walk the bus again setting each device/bridge's >> MPS to the smallest MPS found. >> >> This patch just got much more complicated. Damn you crappy BIOS >> developers! > > Same with max read req. size except that for that one we don't even have > a register telling us what the bridge max is in the first place. Yes, I was going to look into that as well, since it makes sense to check and set it at the same time as the MPS. > > I have some in-house code to fix that up on powerpc at boot that I > haven't released yet (for a platform that we haven't published yet) but > it's only boot time, I haven't dealt with the hotplug case. > > The problem I see with hotplug is that potentially, one could have to > change the entire hierarchy below the PHB when a new device is plugged > with a smaller max payload size than what was used before. > > However, I didn't think it was "safe" to just change the size on a live > system, devices have to be quiescent. > > That makes it impossible to do reliably in my book, ie, anything with a > hotplug might have to be hard wired to the smallest size unless the > hotplug happens at the PHB level in which case there is no risk of > disrupting another device while changing the size. On the limited amount of systems I've access to, I see the PCI-E slots wired directly to the RC. So, this would not be an issue for these. Though, we might want to add a check for a bus between the RC and the hotpluggable PCI-E slot, and force it to the minimum in that case. > > Ben. > > > -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
