On Wednesday, November 12, 2008 1:15 pm Greg KH wrote: > On Wed, Nov 12, 2008 at 01:02:58PM -0800, Jesse Barnes wrote: > > On Wednesday, November 12, 2008 12:12 pm Greg KH wrote: > > > On Wed, Nov 12, 2008 at 11:51:04AM -0800, Jesse Barnes wrote: > > > > On Tuesday, November 11, 2008 9:37 pm Yuji Shimada wrote: > > > > > As I have not received a reply to my mail of November 6, 2008, I am > > > > > resending it herewith. > > > > > > > > > > This patch adds the function that reassigns page-aligned memory > > > > > resources to device, to linux. > > > > > > > > > > I created this patch for xen's dom0 linux. It have already been > > > > > included in xen's dom0 linux. It is useful when we assign I/O > > > > > device to HVM domain using pci passthrough, because page-aligned > > > > > memory resource is required for pci passthrough. It is also useful > > > > > for KVM. So I submit it to linux-pci ML. > > > > > > > > Ok, so the fundamental requirement here is to assign PCI devices to > > > > guests, right? That means PCI resources be aligned to a page > > > > boundary and take up at least a page so that no other resources will > > > > fall into the same page, right? > > > > > > > > So you add a quirk to unassign the resources of the device(s) in > > > > question and let the core reassign them according to your new > > > > constraints. So far, so good. > > > > > > > > Like Matthew said, reassigning at runtime should be possible, but can > > > > be a little trickier since your configuration has to disallow driver > > > > binding (or unbind everything) until after you've done your > > > > reassignments. For v11n setups that doesn't seem wholly > > > > unreasonable, but probably isn't as convenient as simply doing it at > > > > startup time. > > > > > > > > Anyway, comments on your patch below. > > > > > > > > > To reassign page-aligned memory resources to device, please add > > > > > boot parameter of linux as follows. > > > > > > > > > > reassigndev=00:1d.7,01:00.0 > > > > > > > > > > reassigndev= Specifies device to reassign page-aligned > > > > > memory resources. PCI-PCI bridge can be > > > > > specified, if resource windows need to be > > > > > expanded. > > > > > > > > I'd call this pagealignbars= or something instead, since you're not > > > > just reassigning things (that happens anyway for conflicting > > > > resources and for quirked devices), you're making sure that resources > > > > really are page aligned. It should probably also be part of the pci= > > > > parameter; besides your code for parsing out PCI device specifiers > > > > might be handy in the future (I was surprised my quick look didn't > > > > find some other boot option that already did it). > > > > > > As I said before, for a different patch recently, you can't use pci > > > device ids as a kernel command line option as they are not stable and > > > the user/distro never knows them ahead of time. > > > > > > Please do this dynamically while the kernel is running instead, if you > > > really have to do this. > > > > In the face of system configuration changes and kernel changes that's > > true. But for a specific configuration it's generally stable, and people > > definitely rely on that today. > > They do? In what way? > > It isn't stable at all. I have 2 laptops here that randomly decide to > reassign the pci bus ids depending on the phase of the moon. Caused > havoc with a xorg driver that was trying to "assume" that it always knew > what the pci bus id ordering was going to be. Well there's one example: config files often specify specific PCI bus IDs. > This is in a laptop model that is shipping today. The BIOS people > rightly said that there was nothing in the PCI spec that prevented this > from happening, and that Windows didn't have a problem with it at all. > > For other users, lots of laptops reassign pci devices when booting if > the laptop is plugged into a docking station or not. > > I've also seen pci hotplug systems that like to renumber them > semi-randomly every other boot depending on how long different PCI > "nodes" take to powerup in the system. > > So if anyone is relying on this in any way right now, that's wrong. Hey, I'm not saying they *should* rely on it, only that they *do* :). IME servers are more consistent here. SGI systems even had some firmware code to make sure the system came up with the same bus numbers on every boot, even in the face of node topology changes. So in the context of this patch relying on specific layouts may not be unreasonable, but like I said in my original reply, a runtime reassignment is probably better, even if a little more trouble. Jesse -- 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
