Hi, On 10/11/22 19:40, Bjorn Helgaas wrote: > On Mon, Oct 10, 2022 at 05:02:06PM +0200, Hans de Goede wrote: >> Clevo NL4XLU barebones have the same E820 reservation covering >> the entire _CRS 32-bit window issue as the Lenovo *IIL* and >> Clevo X170KM-G models, relevant dmesg bits (with pci=no_e820): >> >> BIOS-e820: [mem 0x000000005bc50000-0x00000000cfffffff] reserved >> pci_bus 0000:00: root bus resource [mem 0x6d800000-0xbfffffff window] >> >> Note how the e820 reservation covers the entire PCI root mem window. >> >> Add a no_e820 quirk for these models to fix the touchpad not working >> (due to Linux being unable to assign a PCI BAR for the i2c-controller). > > I do plan to apply this, but a little food for thought below. > > I explored this issue a little bit with the ACPI/UEFI folks (see > https://members.uefi.org/wg/aswg/mail/thread/9265 if you have access). > > One aspect I had glossed over earlier is that on most recent machines, > the "E820 map" Linux uses is actually constructed internally by Linux > based on the UEFI memory map, and that construction conflates several > EFI types into E820_TYPE_RESERVED; see > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/firmware/efi/libstub/x86-stub.c?id=v5.19#n576 > > We don't have a dmesg log with "efi=debug" for this case, but in most > cases where E820 says the entire root mem window is reserved, I think > it's because an EfiMemoryMappedIO entry was converted to > E820_TYPE_RESERVED. > > From a response in the ACPI/UEFI discussion: > > The reason EfiMemoryMappedIO[1] exist in the UEFI memory map is to > request a virtual mapping for UEFI Runtime Services. > > For example if the EFI Runtime Service needed to write to FLASH then > the NOR FLASH would need a mapping. Also the NOR FLASH controller > might also need a mapping and this could include PCI Config Space if > needed registers existed in that space. > > Thus the EfiMemoryMappedIO entries just exist to pass up the > EFI_MEMORY_RUNTIME attribute in the UEFI Memory Map. This is the part > of the contract for UEFI Runtime Service to use virtual mappings > provided by the OS. So from an OS point of view EfiMemoryMappedIO has > no other purpose. > > [1] UEFI: Table 7-5 Memory Type Usage before ExitBootServices() "Used > by system firmware to request that a memory-mapped IO region be > mapped by the OS to a virtual address so it can be accessed by EFI > runtime services." > > So the point here is that Linux currently converts EfiMemoryMappedIO > to E820_TYPE_RESERVED, and that likely attaches more meaning to those > regions than firmware intended. > > I'm a little leery of changing that UEFI->E820 conversion because of > other possible implications, but it may be that omitting > EfiMemoryMappedIO entries from the E820 map and keeping the original > "avoid E820 regions" (4dc2287c1805) would also solve this problem. Actually during my many attempts to fix this I did write a patch adding a new E820_TYPE_MMIO to the generated e820-memmap which would only show up in the EFI -> E820 entry generation case and then used that to not exclude that E820 region, see this RFC series: https://lore.kernel.org/linux-pci/20220214151759.98267-1-hdegoede@xxxxxxxxxx/ Unfortunately this series caused the same X1 carbon gen2 suspend/resume not working problem as the earlier set pci=no_e820 based on a BIOS date cutoff attempt which I did earlier and which even briefly was in some -rc kernels through Rafael's tree. I also did another series which used the EfiMemoryMappedIO type as an input to heuristics to automatically set pci=no_e820, see: https://lore.kernel.org/linux-pci/20220228105259.230903-1-hdegoede@xxxxxxxxxx/ IIRC that patch eventually got replaced by a similar but simpler heuristic from you. Which IIRC eventually got dropped again because it was causing regressions on some models again. So we ended up with the current set pci=no_e820 using DMI based quirks + try to enable it for all BIOS-es with date >= 2023 approach, with the plan to do DMI quirks setting pci=use_e820 if any (buggy) 2023 BIOS-es show up which need this. Regards, Hans >> Link: https://bugzilla.kernel.org/show_bug.cgi?id=216565 >> Signed-off-by: Hans de Goede <hdegoede@xxxxxxxxxx> >> --- >> arch/x86/pci/acpi.c | 13 +++++++++++++ >> 1 file changed, 13 insertions(+) >> >> diff --git a/arch/x86/pci/acpi.c b/arch/x86/pci/acpi.c >> index 2f82480fd430..45ef65d31a40 100644 >> --- a/arch/x86/pci/acpi.c >> +++ b/arch/x86/pci/acpi.c >> @@ -189,6 +189,19 @@ static const struct dmi_system_id pci_crs_quirks[] __initconst = { >> DMI_MATCH(DMI_BOARD_NAME, "X170KM-G"), >> }, >> }, >> + >> + /* >> + * Clevo NL4XLU barebones have the same E820 reservation covering >> + * the entire _CRS 32-bit window issue as the Lenovo *IIL* models. >> + * See https://bugzilla.kernel.org/show_bug.cgi?id=216565 >> + */ >> + { >> + .callback = set_no_e820, >> + .ident = "Clevo NL4XLU Barebone", >> + .matches = { >> + DMI_MATCH(DMI_BOARD_NAME, "NL4XLU"), >> + }, >> + }, >> {} >> }; >> >> -- >> 2.37.3 >> >
