When running as an SNP or TDX guest under KVM, force the legacy PCI hole, i.e. memory between Top of Lower Usable DRAM and 4GiB, to be mapped as UC via a forced variable MTRR range. In most KVM-based setups, legacy devices such as the HPET and TPM are enumerated via ACPI. ACPI enumeration includes a Memory32Fixed entry, and optionally a SystemMemory descriptor for an OperationRegion, e.g. if the device needs to be accessed via a Control Method. If a SystemMemory entry is present, then the kernel's ACPI driver will auto-ioremap the region so that it can be accessed at will. However, the ACPI spec doesn't provide a way to enumerate the memory type of SystemMemory regions, i.e. there's no way to tell software that a region must be mapped as UC vs. WB, etc. As a result, Linux's ACPI driver always maps SystemMemory regions using ioremap_cache(), i.e. as WB on x86. The dedicated device drivers however, e.g. the HPET driver and TPM driver, want to map their associated memory as UC or WC, as accessing PCI devices using WB is unsupported. On bare metal and non-CoCO, the conflicting requirements "work" as firmware configures the PCI hole (and other device memory) to be UC in the MTRRs. So even though the ACPI mappings request WB, they are forced to UC due to the kernel properly handling the MTRR overrides, and thus are compatible with the drivers' requested WC/UC-. With force WB MTRRs on SNP and TDX guests, the ACPI mappings get their requested WB if the ACPI mappings are established before the dedicated driver code attempts to initialize the device. E.g. if acpi_init() runs before the corresponding device driver is probed, ACPI's WB mapping will "win", and result in the driver's ioremap() failing because the existing WB mapping isn't compatible with the requested WC/UC-. E.g. when a TPM is emulated by the hypervisor (ignoring the security implications of relying on what is allegedly an untrusted entity to store measurements), the TPM driver will request UC and fail: [ 1.730459] ioremap error for 0xfed40000-0xfed45000, requested 0x2, got 0x0 [ 1.732780] tpm_tis MSFT0101:00: probe with driver tpm_tis failed with error -12 Note, the '0x2' and '0x0' values refer to "enum page_cache_mode", not x86's memtypes (which frustratingly are an almost pure inversion; 2 == WB, 0 == UC). E.g. tracing mapping requests for TPM TIS yields: Mapping TPM TIS with req_type = 0 WARNING: CPU: 22 PID: 1 at arch/x86/mm/pat/memtype.c:530 memtype_reserve+0x2ab/0x460 Modules linked in: CPU: 22 UID: 0 PID: 1 Comm: swapper/0 Tainted: G W 6.16.0-rc7+ #2 VOLUNTARY Tainted: [W]=WARN Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 05/29/2025 RIP: 0010:memtype_reserve+0x2ab/0x460 __ioremap_caller+0x16d/0x3d0 ioremap_cache+0x17/0x30 x86_acpi_os_ioremap+0xe/0x20 acpi_os_map_iomem+0x1f3/0x240 acpi_os_map_memory+0xe/0x20 acpi_ex_system_memory_space_handler+0x273/0x440 acpi_ev_address_space_dispatch+0x176/0x4c0 acpi_ex_access_region+0x2ad/0x530 acpi_ex_field_datum_io+0xa2/0x4f0 acpi_ex_extract_from_field+0x296/0x3e0 acpi_ex_read_data_from_field+0xd1/0x460 acpi_ex_resolve_node_to_value+0x2ee/0x530 acpi_ex_resolve_to_value+0x1f2/0x540 acpi_ds_evaluate_name_path+0x11b/0x190 acpi_ds_exec_end_op+0x456/0x960 acpi_ps_parse_loop+0x27a/0xa50 acpi_ps_parse_aml+0x226/0x600 acpi_ps_execute_method+0x172/0x3e0 acpi_ns_evaluate+0x175/0x5f0 acpi_evaluate_object+0x213/0x490 acpi_evaluate_integer+0x6d/0x140 acpi_bus_get_status+0x93/0x150 acpi_add_single_object+0x43a/0x7c0 acpi_bus_check_add+0x149/0x3a0 acpi_bus_check_add_1+0x16/0x30 acpi_ns_walk_namespace+0x22c/0x360 acpi_walk_namespace+0x15c/0x170 acpi_bus_scan+0x1dd/0x200 acpi_scan_init+0xe5/0x2b0 acpi_init+0x264/0x5b0 do_one_initcall+0x5a/0x310 kernel_init_freeable+0x34f/0x4f0 kernel_init+0x1b/0x200 ret_from_fork+0x186/0x1b0 ret_from_fork_asm+0x1a/0x30 The above traces are from a Google-VMM based VM, but the same behavior happens with a QEMU based VM that is modified to add a SystemMemory range for the TPM TIS address space. The only reason this doesn't cause problems for HPET, which appears to require a SystemMemory region, is because HPET gets special treatment via x86_init.timers.timer_init(), and so gets a chance to create its UC- mapping before acpi_init() clobbers things. Disabling the early call to hpet_time_init() yields the same behavior for HPET: [ 0.318264] ioremap error for 0xfed00000-0xfed01000, requested 0x2, got 0x0 Hack around the ACPI gap by forcing the legacy PCI hole to UC when overriding the (virtual) MTRRs for CoCo guest, so that ioremap handling of MTRRs naturally kicks in and forces the ACPI mappings to be UC. Note, the requested/mapped memtype doesn't actually matter in terms of accessing the device. In practically every setup, legacy PCI devices are emulated by the hypervisor, and accesses are intercepted and handled as emulated MMIO, i.e. never access physical memory and thus don't have an effective memtype. Even in a theoretical setup where such devices are passed through by the host, i.e. point at real MMIO memory, it is KVM's (as the hypervisor) responsibility to force the memory to be WC/UC, e.g. via EPT memtype under TDX or real hardware MTRRs under SNP. Not doing so cannot work, and the hypervisor is highly motivated to do the right thing as letting the guest access hardware MMIO with WB would likely result in a variety of fatal #MCs. In other words, forcing the range to be UC is all about coercing the kernel's tracking into thinking that it has established UC mappings, so that the ioremap code doesn't reject mappings from e.g. the TPM driver and thus prevent the driver from loading and the device from functioning. Note #2, relying on guest firmware to handle this scenario, e.g. by setting virtual MTRRs and then consuming them in Linux, is not a viable option, as the virtual MTRR state is managed by the untrusted hypervisor, and because OVMF at least has stopped programming virtual MTRRs when running as a TDX guest. Link: https://lore.kernel.org/all/8137d98e-8825-415b-9282-1d2a115bb51a@linux.intel.com Fixes: 8e690b817e38 ("x86/kvm: Override default caching mode for SEV-SNP and TDX") Cc: stable@vger.kernel.org Cc: Peter Gonda Cc: Vitaly Kuznetsov Cc: Tom Lendacky Cc: Jürgen Groß Cc: Korakit Seemakhupt Cc: Jianxiong Gao Cc: Nikolay Borisov Suggested-by: Binbin Wu Signed-off-by: Sean Christopherson --- v2: Force the PCI hole to be UC via synthetic variable MTRR range instead of hijacking is_untracked_pat_range() (which was horrific and apparently didn't work). [Binbin] v1: - https://lore.kernel.org/all/20250201005048.657470-1-seanjc@google.com - https://lore.kernel.org/all/CAMGD6P1Q9tK89AjaPXAVvVNKtD77-zkDr0Kmrm29+e=i+R+33w@mail.gmail.com arch/x86/kernel/kvm.c | 21 +++++++++++++++++++-- 1 file changed, 19 insertions(+), 2 deletions(-) diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c index 8ae750cde0c6..57379698015e 100644 --- a/arch/x86/kernel/kvm.c +++ b/arch/x86/kernel/kvm.c @@ -933,6 +933,19 @@ static void kvm_sev_hc_page_enc_status(unsigned long pfn, int npages, bool enc) static void __init kvm_init_platform(void) { + u64 tolud = PFN_PHYS(e820__end_of_low_ram_pfn()); + /* + * Note, hardware requires variable MTRR ranges to be power-of-2 sized + * and naturally aligned. But when forcing guest MTRR state, Linux + * doesn't program the forced ranges into hardware. Don't bother doing + * the math to generate a technically-legal range. + */ + struct mtrr_var_range pci_hole = { + .base_lo = tolud | X86_MEMTYPE_UC, + .mask_lo = (u32)(~(SZ_4G - tolud - 1)) | MTRR_PHYSMASK_V, + .mask_hi = (BIT_ULL(boot_cpu_data.x86_phys_bits) - 1) >> 32, + }; + if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) && kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) { unsigned long nr_pages; @@ -982,8 +995,12 @@ static void __init kvm_init_platform(void) kvmclock_init(); x86_platform.apic_post_init = kvm_apic_init; - /* Set WB as the default cache mode for SEV-SNP and TDX */ - guest_force_mtrr_state(NULL, 0, MTRR_TYPE_WRBACK); + /* + * Set WB as the default cache mode for SEV-SNP and TDX, with a single + * UC range for the legacy PCI hole, e.g. so that devices that expect + * to get UC/WC mappings don't get surprised with WB. + */ + guest_force_mtrr_state(&pci_hole, 1, MTRR_TYPE_WRBACK); } #if defined(CONFIG_AMD_MEM_ENCRYPT) base-commit: 1b237f190eb3d36f52dffe07a40b5eb210280e00 -- 2.51.0.268.g9569e192d0-goog