- 26 Sep, 2022 33 commits
-
-
Vitaly Kuznetsov authored
SECONDARY_EXEC_ENCLS_EXITING is the only control which is conditionally added to the 'optional' checklist in setup_vmcs_config() but the special case can be avoided by always checking for its presence first and filtering out the result later. Note: the situation when SECONDARY_EXEC_ENCLS_EXITING is present but cpu_has_sgx() is false is possible when SGX is "soft-disabled", e.g. if software writes MCE control MSRs or there's an uncorrectable #MC. Reviewed-by:
Jim Mattson <jmattson@google.com> Reviewed-by:
Maxim Levitsky <mlevitsk@redhat.com> Signed-off-by:
Vitaly Kuznetsov <vkuznets@redhat.com> Signed-off-by:
Sean Christopherson <seanjc@google.com> Link: https://lore.kernel.org/r/20220830133737.1539624-22-vkuznets@redhat.comSigned-off-by:
Paolo Bonzini <pbonzini@redhat.com>
-
Vitaly Kuznetsov authored
CPU_BASED_{INTR,NMI}_WINDOW_EXITING controls are toggled dynamically by vmx_enable_{irq,nmi}_window, handle_interrupt_window(), handle_nmi_window() but setup_vmcs_config() doesn't check their existence. Add the check and filter the controls out in vmx_exec_control(). Note: KVM explicitly supports CPUs without VIRTUAL_NMIS and all these CPUs are supposedly lacking NMI_WINDOW_EXITING too. Adjust cpu_has_virtual_nmis() accordingly. No functional change intended. Reviewed-by: -
Vitaly Kuznetsov authored
VM_ENTRY_IA32E_MODE control is toggled dynamically by vmx_set_efer() and setup_vmcs_config() doesn't check its existence. On the contrary, nested_vmx_setup_ctls_msrs() doesn set it on x86_64. Add the missing check and filter the bit out in vmx_vmentry_ctrl(). No (real) functional change intended as all existing CPUs supporting long mode and VMX are supposed to have it. Reviewed-by:
-
Sean Christopherson authored
Advertise VM_{ENTRY,EXIT}_LOAD_IA32_PERF_GLOBAL_CTRL as being supported for nested VMs irrespective of hardware support. KVM fully emulates the controls, i.e. manually emulates MSR writes on entry/exit, and never propagates the guest settings directly to vmcs02. In addition to allowing L1 VMMs to use the controls on older hardware, unconditionally advertising the controls will also allow KVM to use its vmcs01 configuration as the basis for the nested VMX configuration without causing a regression (due the errata which causes KVM to "hide" the control from vmcs01 but not vmcs12). Signed-off-by: -
Sean Christopherson authored
Don't propagate vmcs12's VM_ENTRY_LOAD_IA32_PERF_GLOBAL_CTRL to vmcs02. KVM doesn't disallow L1 from using VM_ENTRY_LOAD_IA32_PERF_GLOBAL_CTRL even when KVM itself doesn't use the control, e.g. due to the various CPU errata that where the MSR can be corrupted on VM-Exit. Preserve KVM's (vmcs01) setting to hopefully avoid having to toggle the bit in vmcs02 at a later point. E.g. if KVM is loading PERF_GLOBAL_CTRL when running L1, then odds are good KVM will also load the MSR when running L2. Fixes: 8bf00a52 ("KVM: VMX: add support for switching of PERF_GLOBAL_CTRL") Cc: stable@vger.kernel.org Signed-off-by:
-
Vitaly Kuznetsov authored
With the updated eVMCSv1 definition, there's no known 'problematic' controls which are exposed in VMX control MSRs but are not present in eVMCSv1: all known Hyper-V versions either don't expose the new fields by not setting bits in the VMX feature controls or support the new eVMCS revision. Get rid of VMX control MSRs filtering for KVM on Hyper-V. Note: VMX control MSRs filtering for Hyper-V on KVM (nested_evmcs_filter_control_msr()) stays as even the updated eVMCSv1 definition doesn't have all the features implemented by KVM and some fields are still missing. Moreover, nested_evmcs_filter_control_msr() has to support the original eVMCSv1 version when VMM wishes so. Reviewed-by:
-
Vitaly Kuznetsov authored
Enlightened VMCS v1 got updated and now includes the required fields for loading PERF_GLOBAL_CTRL upon VMENTER/VMEXIT features. For KVM on Hyper-V enablement, KVM can just observe VMX control MSRs and use the features (with or without eVMCS) when possible. Hyper-V on KVM is messier as Windows 11 guests fail to boot if the controls are advertised and a new PV feature flag, CPUID.0x4000000A.EBX BIT(0), is not set. Honor the Hyper-V CPUID feature flag to play nice with Windows guests. Signed-off-by:
-
Sean Christopherson authored
WARN and reject nested VM-Enter if KVM is using eVMCS and manages to allow a non-zero value in the upper 32 bits of VM-function controls. The eVMCS code assumes all inputs are 32-bit values and subtly drops the upper bits. WARN instead of adding proper "support", it's unlikely the upper bits will be defined/used in the next decade. Signed-off-by:
-
Vitaly Kuznetsov authored
Update Enlightened VMCS definition in selftests from KVM. Reviewed-by:
-
Vitaly Kuznetsov authored
The updated Enlightened VMCS definition has 'encls_exiting_bitmap' field which needs mapping to VMCS, add the missing encoding. Reviewed-by:
Kai Huang <kai.huang@intel.com> Signed-off-by:
-
Vitaly Kuznetsov authored
KVM has to check guest visible HYPERV_CPUID_NESTED_FEATURES.EBX CPUID leaf to know which Enlightened VMCS definition to use (original or 2022 update). Cache the leaf along with other Hyper-V CPUID feature leaves to make the check quick. Reviewed-by:
-
Vitaly Kuznetsov authored
Enlightened VMCS v1 definition was updated with new fields, add support for them for Hyper-V on KVM. Note: SSP, CET and Guest LBR features are not supported by KVM yet and 'struct vmcs12' has no corresponding fields. Reviewed-by:
-
Vitaly Kuznetsov authored
Enlightened VMCS v1 definition was updated with new fields, support them in KVM by defining VMCS-to-EVMCS conversion. Note: SSP, CET and Guest LBR features are not supported by KVM yet and the corresponding fields are not defined in 'enum vmcs_field', leave them commented out for now. Reviewed-by:
-
Sean Christopherson authored
Locally #define and use the nested virtualization Consistency Check (CC) macro to handle eVMCS unsupported controls checks. Using the macro loses the existing printing of the unsupported controls, but that's a feature and not a bug. The existing approach is flawed because the @err param to trace_kvm_nested_vmenter_failed() is the error code, not the error value. The eVMCS trickery mostly works as __print_symbolic() falls back to printing the raw hex value, but that subtly relies on not having a match between the unsupported value and VMX_VMENTER_INSTRUCTION_ERRORS. If it's really truly necessary to snapshot the bad value, then the tracepoint can be extended in the future. Signed-off-by:
-
Vitaly Kuznetsov authored
Refactor the handling of unsupported eVMCS to use a 2-d array to store the set of unsupported controls. KVM's handling of eVMCS is completely broken as there is no way for userspace to query which features are unsupported, nor does KVM prevent userspace from attempting to enable unsupported features. A future commit will remedy that by filtering and enforcing unsupported features when eVMCS, but that needs to be opt-in from userspace to avoid breakage, i.e. KVM needs to maintain its legacy behavior by snapshotting the exact set of controls that are currently (un)supported by eVMCS. No functional change intended. Suggested-by:
-
Sean Christopherson authored
When querying whether or not eVMCS is enabled on behalf of the guest, treat eVMCS as enable if and only if Hyper-V is enabled/exposed to the guest. Note, flows that come from the host, e.g. KVM_SET_NESTED_STATE, must NOT check for Hyper-V being enabled as KVM doesn't require guest CPUID to be set before most ioctls(). Signed-off-by:
-
Sean Christopherson authored
Return -ENOMEM back to userspace if allocating the Hyper-V vCPU struct fails when enabling Hyper-V in guest CPUID. Silently ignoring failure means that KVM will not have an up-to-date CPUID cache if allocating the struct succeeds later on, e.g. when activating SynIC. Rejecting the CPUID operation also guarantess that vcpu->arch.hyperv is non-NULL if hyperv_enabled is true, which will allow for additional cleanup, e.g. in the eVMCS code. Note, the initialization needs to be done before CPUID is set, and more subtly before kvm_check_cpuid(), which potentially enables dynamic XFEATURES. Sadly, there's no easy way to avoid exposing Hyper-V details to CPUID or vice versa. Expose kvm_hv_vcpu_init() and the Hyper-V CPUID signature to CPUID instead of exposing cpuid_entry2_find() outside of CPUID code. It's hard to envision kvm_hv_vcpu_init() being misused, whereas cpuid_entry2_find() absolutely shouldn't be used outside of core CPUID code. Fixes: 10d7bf1e ("KVM: x86: hyper-v: Cache guest CPUID leaves determining features availability") Signed-off-by:
-
Sean Christopherson authored
When potentially allocating/initializing the Hyper-V vCPU struct, check for an existing instance in kvm_hv_vcpu_init() instead of requiring callers to perform the check. Relying on callers to do the check is risky as it's all too easy for KVM to overwrite vcpu->arch.hyperv and leak memory, and it adds additional burden on callers without much benefit. No functional change intended. Signed-off-by:
Wei Liu <wei.liu@kernel.org> Link: https://lore.kernel.org/r/20220830133737.1539624-5-vkuznets@redhat.comSigned-off-by:
-
Vitaly Kuznetsov authored
Wipe the whole 'hv_vcpu->cpuid_cache' with memset() instead of having to zero each particular member when the corresponding CPUID entry was not found. No functional change intended. Signed-off-by:
-
Vitaly Kuznetsov authored
Updated Hyper-V Enlightened VMCS specification lists several new fields for the following features: - PerfGlobalCtrl - EnclsExitingBitmap - Tsc Scaling - GuestLbrCtl - CET - SSP Update the definition. Note, the updated spec also provides an additional CPUID feature flag, CPUIDD.0x4000000A.EBX BIT(0), for PerfGlobalCtrl to workaround a Windows 11 quirk. Despite what the TLFS says: Indicates support for the GuestPerfGlobalCtrl and HostPerfGlobalCtrl fields in the enlightened VMCS. guests can safely use the fields if they are enumerated in the architectural VMX MSRs. I.e. KVM-on-HyperV doesn't need to check the CPUID bit, but KVM-as-HyperV must ensure the bit is set if PerfGlobalCtrl fields are exposed to L1. https://docs.microsoft.com/en-us/virtualization/hyper-v-on-windows/tlfs/tlfsSigned-off-by:
-
Vitaly Kuznetsov authored
Section 1.9 of TLFS v6.0b says: "All structures are padded in such a way that fields are aligned naturally (that is, an 8-byte field is aligned to an offset of 8 bytes and so on)". 'struct enlightened_vmcs' has a glitch: ... struct { u32 nested_flush_hypercall:1; /* 836: 0 4 */ u32 msr_bitmap:1; /* 836: 1 4 */ u32 reserved:30; /* 836: 2 4 */ } hv_enlightenments_control; /* 836 4 */ u32 hv_vp_id; /* 840 4 */ u64 hv_vm_id; /* 844 8 */ u64 partition_assist_page; /* 852 8 */ ... And the observed values in 'partition_assist_page' make no sense at all. Fix the layout by padding the structure properly. Fixes: 68d1eb72 ("x86/hyper-v: define struct hv_enlightened_vmcs and clean field bits") Reviewed-by:Michael Kelley <mikelley@microsoft.com> Signed-off-by:
-
Oliver Upton authored
Require KVM_CAP_VM_DISABLE_NX_HUGE_PAGES for the entire NX hugepage test instead of skipping the "disable" subtest if the capability isn't supported by the host kernel. While the "enable" subtest does provide value when the capability isn't supported, silently providing only half the promised coveraged is undesirable, i.e. it's better to skip the test so that the user knows something. Alternatively, the test could print something to alert the user instead of silently skipping the subtest, but that would encourage other tests to follow suit, and it's not clear that it's desirable to take selftests in that direction. And if selftests do head down the path of skipping subtests, such behavior needs first-class support in the framework. Opportunistically convert other test preconditions to TEST_REQUIRE(). Signed-off-by:
Oliver Upton <oliver.upton@linux.dev> Reviewed-by:
David Matlack <dmatlack@google.com> Link: https://lore.kernel.org/r/20220812175301.3915004-1-oliver.upton@linux.dev [sean: rewrote changelog to capture discussion about skipping the test] Signed-off-by:
-
Uros Bizjak authored
There is no need to declare vmread_error() asmlinkage, its arguments can be passed via registers for both 32-bit and 64-bit targets. Function argument registers are considered call-clobbered registers, they are saved in the trampoline just before the function call and restored afterwards. Dropping "asmlinkage" patch unifies trampoline function argument handling between 32-bit and 64-bit targets and improves generated code for 32-bit targets. Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: Sean Christopherson <seanjc@google.com> Signed-off-by:
Uros Bizjak <ubizjak@gmail.com> Link: https://lore.kernel.org/r/20220817144045.3206-1-ubizjak@gmail.comSigned-off-by:
-
Liam Ni authored
Refactor decode_register_operand() to get the ModR/M register if and only if the instruction uses a ModR/M encoding to make it more obvious how the register operand is retrieved. Signed-off-by:
Liam Ni <zhiguangni01@gmail.com> Link: https://lore.kernel.org/r/20220908141210.1375828-1-zhiguangni01@zhaoxin.comSigned-off-by:
-
Mingwei Zhang authored
Print guest pgd in kvm_nested_vmenter() to enrich the information for tracing. When tdp is enabled, print the value of tdp page table (EPT/NPT); when tdp is disabled, print the value of non-nested CR3. Suggested-by:
Mingwei Zhang <mizhang@google.com> Link: https://lore.kernel.org/r/20220825225755.907001-4-mizhang@google.com [sean: print nested_cr3 vs. nested_eptp vs. guest_cr3] Signed-off-by:
-
David Matlack authored
Call trace_kvm_nested_vmenter() during nested VMLAUNCH/VMRESUME to bring parity with nSVM's usage of the tracepoint during nested VMRUN. Attempt to use analagous VMCS fields to the VMCB fields that are reported in the SVM case: "int_ctl": 32-bit field of the VMCB that the CPU uses to deliver virtual interrupts. The analagous VMCS field is the 16-bit "guest interrupt status". "event_inj": 32-bit field of VMCB that is used to inject events (exceptions and interrupts) into the guest. The analagous VMCS field is the "VM-entry interruption-information field". "npt_enabled": 1 when the VCPU has enabled nested paging. The analagous VMCS field is the enable-EPT execution control. "npt_addr": 64-bit field when the VCPU has enabled nested paging. The analagous VMCS field is the ept_pointer. Signed-off-by:
-
Mingwei Zhang authored
Update trace function for nested VM entry to support VMX. Existing trace function only supports nested VMX and the information printed out is AMD specific. So, rename trace_kvm_nested_vmrun() to trace_kvm_nested_vmenter(), since 'vmenter' is generic. Add a new field 'isa' to recognize Intel and AMD; Update the output to print out VMX/SVM related naming respectively, eg., vmcb vs. vmcs; npt vs. ept. Opportunistically update the call site of trace_kvm_nested_vmenter() to make one line per parameter. Signed-off-by:
-
Sean Christopherson authored
Track the address and error code as 64-bit values in the page fault tracepoint. When TDP is enabled, the address is a GPA and thus can be a 64-bit value even on 32-bit hosts. And SVM's #NPF genereates 64-bit error codes. Opportunistically clean up the formatting. Signed-off-by:
-
Wonhyuk Yang authored
Currently, kvm_page_fault trace point provide fault_address and error code. However it is not enough to find which cpu and instruction cause kvm_page_faults. So add vcpu id and instruction pointer in kvm_page_fault trace point. Cc: Baik Song An <bsahn@etri.re.kr> Cc: Hong Yeon Kim <kimhy@etri.re.kr> Cc: Taeung Song <taeung@reallinux.co.kr> Cc: linuxgeek@linuxgeek.io Signed-off-by:
Wonhyuk Yang <vvghjk1234@gmail.com> Link: https://lore.kernel.org/r/20220510071001.87169-1-vvghjk1234@gmail.comSigned-off-by:
-
Aaron Lewis authored
Two copies of KVM_X86_SET_MSR_FILTER somehow managed to make it's way into the documentation. Remove one copy and merge the difference from the removed copy into the copy that's being kept. Fixes: fd49e8ee ("Merge branch 'kvm-sev-cgroup' into HEAD") Signed-off-by:
Aaron Lewis <aaronlewis@google.com> Link: https://lore.kernel.org/r/20220712001045.2364298-2-aaronlewis@google.comSigned-off-by:
-
Paolo Bonzini authored
Since svm_check_nested_events() is now handling INIT signals, there is no need to latch it until the VMEXIT is injected. The only condition under which INIT signals are latched is GIF=0. Suggested-by:
-
Uros Bizjak authored
Avoid instructions with explicit uses of the stack pointer between instructions that implicitly refer to it. The sequence of POP %reg; ADD $x, %RSP; POP %reg forces emission of synchronization uop to synchronize the value of the stack pointer in the stack engine and the out-of-order core. Using POP with the dummy register instead of ADD $x, %RSP results in a smaller code size and faster code. The patch also fixes the reference to the wrong register in the nearby comment. Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: Sean Christopherson <seanjc@google.com> Signed-off-by:
-
Miaohe Lin authored
When alloc_cpumask_var_node() fails for a certain cpu, there might be some allocated cpumasks for percpu cpu_kick_mask. We should free these cpumasks or memoryleak will occur. Fixes: baff59cc ("KVM: Pre-allocate cpumasks for kvm_make_all_cpus_request_except()") Signed-off-by:
Miaohe Lin <linmiaohe@huawei.com> Link: https://lore.kernel.org/r/20220823063414.59778-1-linmiaohe@huawei.comSigned-off-by:
-
- 30 Aug, 2022 2 commits
-
-
Yosry Ahmed authored
Count the pages used by KVM in arm64 for stage2 mmu in memory stats under secondary pagetable stats (e.g. "SecPageTables" in /proc/meminfo) to give better visibility into the memory consumption of KVM mmu in a similar way to how normal user page tables are accounted. Signed-off-by:
Yosry Ahmed <yosryahmed@google.com> Reviewed-by:
Marc Zyngier <maz@kernel.org> Link: https://lore.kernel.org/r/20220823004639.2387269-5-yosryahmed@google.comSigned-off-by:
-
Yosry Ahmed authored
Count the pages used by KVM mmu on x86 in memory stats under secondary pagetable stats (e.g. "SecPageTables" in /proc/meminfo) to give better visibility into the memory consumption of KVM mmu in a similar way to how normal user page tables are accounted. Add the inner helper in common KVM, ARM will also use it to count stats in a future commit. Signed-off-by:
-
- 24 Aug, 2022 4 commits
-
-
Yosry Ahmed authored
We keep track of several kernel memory stats (total kernel memory, page tables, stack, vmalloc, etc) on multiple levels (global, per-node, per-memcg, etc). These stats give insights to users to how much memory is used by the kernel and for what purposes. Currently, memory used by KVM mmu is not accounted in any of those kernel memory stats. This patch series accounts the memory pages used by KVM for page tables in those stats in a new NR_SECONDARY_PAGETABLE stat. This stat can be later extended to account for other types of secondary pages tables (e.g. iommu page tables). KVM has a decent number of large allocations that aren't for page tables, but for most of them, the number/size of those allocations scales linearly with either the number of vCPUs or the amount of memory assigned to the VM. KVM's secondary page table allocations do not scale linearly, especially when nested virtualization is in use. From a KVM perspective, NR_SECONDARY_PAGETABLE will scale with KVM's per-VM pages_{4k,2m,1g} stats unless the guest is doing something bizarre (e.g. accessing only 4kb chunks of 2mb pages so that KVM is forced to allocate a large number of page tables even though the guest isn't accessing that much memory). However, someone would need to either understand how KVM works to make that connection, or know (or be told) to go look at KVM's stats if they're running VMs to better decipher the stats. Furthermore, having NR_PAGETABLE side-by-side with NR_SECONDARY_PAGETABLE is informative. For example, when backing a VM with THP vs. HugeTLB, NR_SECONDARY_PAGETABLE is roughly the same, but NR_PAGETABLE is an order of magnitude higher with THP. So having this stat will at the very least prove to be useful for understanding tradeoffs between VM backing types, and likely even steer folks towards potential optimizations. The original discussion with more details about the rationale: https://lore.kernel.org/all/87ilqoi77b.wl-maz@kernel.org This stat will be used by subsequent patches to count KVM mmu memory usage. Signed-off-by:Shakeel Butt <shakeelb@google.com> Acked-by:
-
Miaohe Lin authored
When register_shrinker() fails, KVM doesn't release the percpu counter kvm_total_used_mmu_pages leading to memoryleak. Fix this issue by calling percpu_counter_destroy() when register_shrinker() fails. Fixes: ab271bd4 ("x86: kvm: propagate register_shrinker return code") Signed-off-by:
-
Michal Luczaj authored
The emulator checks the wrong variable while setting the CPU interruptibility state, the target segment is embedded in the instruction opcode, not the ModR/M register. Fix the condition. Signed-off-by:
Michal Luczaj <mhal@rbox.co> Fixes: a5457e7b ("KVM: emulate: POP SS triggers a MOV SS shadow too") Cc: stable@vger.kernel.org Link: https://lore.kernel.org/all/20220821215900.1419215-1-mhal@rbox.coSigned-off-by:
-
Junaid Shahid authored
If vm_init() fails [which can happen, for instance, if a memory allocation fails during avic_vm_init()], we need to cleanup some state in order to avoid resource leaks. Signed-off-by:
Junaid Shahid <junaids@google.com> Link: https://lore.kernel.org/r/20220729224329.323378-1-junaids@google.comSigned-off-by:
-
- 19 Aug, 2022 1 commit
-
-
David Matlack authored
Change the mov in KVM_ASM_SAFE() that zeroes @vector to a movb to make it unambiguous. This fixes a build failure with Clang since, unlike the GNU assembler, the LLVM integrated assembler rejects ambiguous X86 instructions that don't have suffixes: In file included from x86_64/hyperv_features.c:13: include/x86_64/processor.h:825:9: error: ambiguous instructions require an explicit suffix (could be 'movb', 'movw', 'movl', or 'movq') return kvm_asm_safe("wrmsr", "a"(val & -1u), "d"(val >> 32), "c"(msr)); ^ include/x86_64/processor.h:802:15: note: expanded from macro 'kvm_asm_safe' asm volatile(KVM_ASM_SAFE(insn) \ ^ include/x86_64/processor.h:788:16: note: expanded from macro 'KVM_ASM_SAFE' "1: " insn "\n\t" \ ^ <inline asm>:5:2: note: instantiated into assembly here mov $0, 15(%rsp) ^ It seems like this change could introduce undesirable behavior in the future, e.g. if someone used a type larger than a u8 for @vector, since KVM_ASM_SAFE() will only zero the bottom byte. I tried changing the type of @vector to an int to see what would happen. GCC failed to compile due to a size mismatch between `movb` and `%eax`. Clang succeeded in compiling, but the generated code looked correct, so perhaps it will not be an issue. That being said it seems like there could be a better solution to this issue that does not assume @vector is a u8. Fixes: 3b23054c ("KVM: selftests: Add x86-64 support for exception fixup") Signed-off-by:
-
