| Message ID | 20210504171734.1434054-1-seanjc@google.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | KVM: x86: RDPID/RDTSCP fixes and uret MSR cleanups | expand |
On 04/05/21 19:17, Sean Christopherson wrote: > This is a continuation of a less ambitious effort to unify MSR_TSC_AUX > handling across SVM and VMX. Reiji pointed out that MSR_TSC_AUX exists if > RDTSCP *or* RDPID is supported, and things went downhill from there. > > The first half of this series fixes a variety of RDTSCP and RDPID related > bugs. > > The second half of the series cleans up VMX's user return MSR framework > and consolidates more of the uret logic into common x86. > > The last two patches leverage the uret MSR cleanups to move MSR_TSC_AUX > handling to common x86 and add sanity checks to guard against misreporting > of RDPID and/or RDTSCP support. > > This will conflict with my vCPU RESET/INIT cleanup series. Feel free to > punt the conflicts to me. > > Other "fun" things to tackle: > > - The kernel proper also botches RDPID vs. RDTSCP, as MSR_TSC_AUX is > configured if RDTSCP is supported, but is consumed if RDPID is > supported. I'll send this fix separately. > > - Commit 844d69c26d83 ("KVM: SVM: Delay restoration of host MSR_TSC_AUX > until return to userspace") unwittingly fixed a bug where KVM would > write MSR_TSC_AUX with the guest's value when svm->guest_state_loaded > is false, which could lead to running the host with the guest's value. > The bug only exists in 5.12 (maybe 5.11 too?), so crafting a fix for > stable won't be too awful. > > Sean Christopherson (15): > KVM: VMX: Do not adverise RDPID if ENABLE_RDTSCP control is > unsupported > KVM: x86: Emulate RDPID only if RDTSCP is supported > KVM: SVM: Inject #UD on RDTSCP when it should be disabled in the guest > KVM: x86: Move RDPID emulation intercept to its own enum > KVM: VMX: Disable preemption when probing user return MSRs > KVM: SVM: Probe and load MSR_TSC_AUX regardless of RDTSCP support in > host > KVM: x86: Add support for RDPID without RDTSCP > KVM: VMX: Configure list of user return MSRs at module init > KVM: VMX: Use flag to indicate "active" uret MSRs instead of sorting > list > KVM: VMX: Use common x86's uret MSR list as the one true list > KVM: VMX: Disable loading of TSX_CTRL MSR the more conventional way > KVM: x86: Export the number of uret MSRs to vendor modules > KVM: x86: Move uret MSR slot management to common x86 > KVM: x86: Tie Intel and AMD behavior for MSR_TSC_AUX to guest CPU > model > KVM: x86: Hide RDTSCP and RDPID if MSR_TSC_AUX probing failed > > arch/x86/include/asm/kvm_host.h | 9 +- > arch/x86/kvm/cpuid.c | 18 ++- > arch/x86/kvm/emulate.c | 2 +- > arch/x86/kvm/kvm_emulate.h | 1 + > arch/x86/kvm/svm/svm.c | 50 +++----- > arch/x86/kvm/vmx/vmx.c | 217 ++++++++++++++++---------------- > arch/x86/kvm/vmx/vmx.h | 12 +- > arch/x86/kvm/x86.c | 101 ++++++++++++--- > 8 files changed, 245 insertions(+), 165 deletions(-) > Queued, thanks. Paolo
This is a continuation of a less ambitious effort to unify MSR_TSC_AUX handling across SVM and VMX. Reiji pointed out that MSR_TSC_AUX exists if RDTSCP *or* RDPID is supported, and things went downhill from there. The first half of this series fixes a variety of RDTSCP and RDPID related bugs. The second half of the series cleans up VMX's user return MSR framework and consolidates more of the uret logic into common x86. The last two patches leverage the uret MSR cleanups to move MSR_TSC_AUX handling to common x86 and add sanity checks to guard against misreporting of RDPID and/or RDTSCP support. This will conflict with my vCPU RESET/INIT cleanup series. Feel free to punt the conflicts to me. Other "fun" things to tackle: - The kernel proper also botches RDPID vs. RDTSCP, as MSR_TSC_AUX is configured if RDTSCP is supported, but is consumed if RDPID is supported. I'll send this fix separately. - Commit 844d69c26d83 ("KVM: SVM: Delay restoration of host MSR_TSC_AUX until return to userspace") unwittingly fixed a bug where KVM would write MSR_TSC_AUX with the guest's value when svm->guest_state_loaded is false, which could lead to running the host with the guest's value. The bug only exists in 5.12 (maybe 5.11 too?), so crafting a fix for stable won't be too awful. Sean Christopherson (15): KVM: VMX: Do not adverise RDPID if ENABLE_RDTSCP control is unsupported KVM: x86: Emulate RDPID only if RDTSCP is supported KVM: SVM: Inject #UD on RDTSCP when it should be disabled in the guest KVM: x86: Move RDPID emulation intercept to its own enum KVM: VMX: Disable preemption when probing user return MSRs KVM: SVM: Probe and load MSR_TSC_AUX regardless of RDTSCP support in host KVM: x86: Add support for RDPID without RDTSCP KVM: VMX: Configure list of user return MSRs at module init KVM: VMX: Use flag to indicate "active" uret MSRs instead of sorting list KVM: VMX: Use common x86's uret MSR list as the one true list KVM: VMX: Disable loading of TSX_CTRL MSR the more conventional way KVM: x86: Export the number of uret MSRs to vendor modules KVM: x86: Move uret MSR slot management to common x86 KVM: x86: Tie Intel and AMD behavior for MSR_TSC_AUX to guest CPU model KVM: x86: Hide RDTSCP and RDPID if MSR_TSC_AUX probing failed arch/x86/include/asm/kvm_host.h | 9 +- arch/x86/kvm/cpuid.c | 18 ++- arch/x86/kvm/emulate.c | 2 +- arch/x86/kvm/kvm_emulate.h | 1 + arch/x86/kvm/svm/svm.c | 50 +++----- arch/x86/kvm/vmx/vmx.c | 217 ++++++++++++++++---------------- arch/x86/kvm/vmx/vmx.h | 12 +- arch/x86/kvm/x86.c | 101 ++++++++++++--- 8 files changed, 245 insertions(+), 165 deletions(-)