On Friday, May 18, 2018 2:57:07 PM CEST Bjorn Helgaas wrote: > On Fri, May 18, 2018 at 09:48:54AM +0200, Rafael J. Wysocki wrote: > > On Thursday, May 17, 2018 4:13:24 PM CEST Bjorn Helgaas wrote: > > > On Thu, May 17, 2018 at 12:28:54PM +0300, Mika Westerberg wrote: > > > > > host_bridge = to_pci_host_bridge(bus->bridge); > > > > if (!(root->osc_control_set & OSC_PCI_EXPRESS_NATIVE_HP_CONTROL)) > > > > - host_bridge->native_hotplug = 0; > > > > + host_bridge->native_pcie_hotplug = 0; > > > > if (!(root->osc_control_set & OSC_PCI_EXPRESS_AER_CONTROL)) > > > > host_bridge->native_aer = 0; > > > > if (!(root->osc_control_set & OSC_PCI_EXPRESS_PME_CONTROL)) > > > > > > Rafael, I noticed that the PCI Firmware Spec, r3.2, sec 4.5.2.2, says > > > we're required to evaluate _OSC when resuming from S4. I don't see a > > > path where we do that today. > > > > No, we don't do that today. In fact, I was not exactly right when I said that. We don't do that in the image kernel, but the restore kernel started in order to load the image, actually evaluates _OSC (as part of its initialization, which is the same as on a fresh boot). So we do evaluate _OSC on *wakeup* from S4 which is what the specifications require IMO. What may be problematic is that we silently assume that the responses from the firmware received in the process will always be the same last time. > > > Am I missing it? If not, what resume hook do we need to use to do > > > this? > > > > We could do that in the ->restore_noirq callback of the host bridge > > device I suppose. Basically, before resuming all of the PCI devices > > on the bus. > > > > However, question is on what conditions. Do the previous versions of the > > spec have this requirement or is it just a new thing? > > I think the requirement has been there since the beginning of _OSC. > The first reference I can find is ACPI 3.0, sec 6.2.9, which says we > should evaluate _OSC on wake from S4 and when a Notify(<device>, 8) is > delivered. AFAICT, this applies to all _OSC methods, not just those > for PCI host bridges. > > The OSHP description (PCI Firmware Spec r3.2, sec 4.8) talks about > interrupt reconfiguration that may be needed when the OS takes over > control of SHPC. I assume this kind of reconfiguration needs to be > done on wake from S4. The fact that in our case resume from hibernation involves two instances of the kernel doesn't make things more straightforward in that respect, so to speak. We just need to assume that whatever the restore kernel sees will match what the image kernel saw before hibernation - or we need a special protocol to pass that information between the two kernels which is not there ATM (and I would like to see clear evidence that it really is necessary before attempting to invent one).
