On Thu, Apr 01, 2021 at 06:45:02PM +0300, Andy Shevchenko wrote: > On Tue, Mar 09, 2021 at 09:42:52AM +0100, Henning Schild wrote: > > Am Mon, 8 Mar 2021 19:42:21 -0600 > > schrieb Bjorn Helgaas <helgaas@xxxxxxxxxx>: > > > On Mon, Mar 08, 2021 at 09:16:50PM +0200, Andy Shevchenko wrote: > > > > On Mon, Mar 08, 2021 at 12:52:12PM -0600, Bjorn Helgaas wrote: > > > > > On Mon, Mar 08, 2021 at 02:20:16PM +0200, Andy Shevchenko wrote: > > ... > > > > > > > + /* Read the first BAR of the device in question */ > > > > > > + __pci_bus_read_base(bus, devfn, pci_bar_unknown, mem, > > > > > > PCI_BASE_ADDRESS_0, true); > > > > > > > > > > I don't get this. Apparently this normally hidden device is > > > > > consuming PCI address space. The PCI core needs to know > > > > > about this. If it doesn't, the PCI core may assign this > > > > > space to another device. > > > > > > > > Right, it returns all 1:s to any request so PCI core *thinks* > > > > it's plugged off (like D3cold or so). > > > > > > I'm asking about the MMIO address space. The BAR is a register > > > in config space. AFAICT, clearing P2SBC_HIDE_BYTE makes that > > > BAR visible. The BAR describes a region of PCI address space. > > > It looks like setting P2SBC_HIDE_BIT makes the BAR disappear > > > from config space, but it sounds like the PCI address space > > > *described* by the BAR is still claimed by the device. If the > > > device didn't respond to that MMIO space, you would have no > > > reason to read the BAR at all. > > > > > > So what keeps the PCI core from assigning that MMIO space to > > > another device? > > > > The device will respond to MMIO while being hidden. I am afraid > > nothing stops a collision, except for the assumption that the BIOS > > is always right and PCI devices never get remapped. But just > > guessing here. > > > > I have seen devices with coreboot having the P2SB visible, and > > most likely relocatable. Making it visible in Linux and not hiding > > it again might work, but probably only as long as Linux will not > > relocate it. Which i am afraid might seriously upset the BIOS, > > depending on what a device does with those GPIOs and which parts > > are implemented in the BIOS. > > So the question is, do we have knobs in PCI core to mark device > fixes in terms of BARs, no relocation must be applied, no other > devices must have the region? I think the closest thing is the IORESOURCE_PCI_FIXED bit that we use for things that must not be moved. Generally PCI resources are associated with a pci_dev, and we set IORESOURCE_PCI_FIXED for BARs, e.g., dev->resource[n]. We do that for IDE legacy regions (see LEGACY_IO_RESOURCE), Langwell devices (pci_fixed_bar_fixup()), "enhanced allocation" (pci_ea_flags()), and some quirks (quirk_io()). In your case, the device is hidden so it doesn't respond to config accesses, so there is no pci_dev for it. Maybe you could do some sort of quirk that allocates its own struct resource, fills it in, sets IORESOURCE_PCI_FIXED, and does something similar to pci_claim_resource()? Bjorn
