Re: [PATCH v5 0/7] KVM: nSVM: avoid TOC/TOU race when checking vmcb12

[Paolo Bonzini]

On 11/3/21 15:05, Emanuele Giuseppe Esposito wrote:
> Currently there is a TOC/TOU race between the check of vmcb12's
> efer, cr0 and cr4 registers and the later save of their values in
> svm_set_*, because the guest could modify the values in the meanwhile.
>
> To solve this issue, this series introduces and uses svm->nested.save
> structure in enter_svm_guest_mode to save the current value of efer,
> cr0 and cr4 and later use these to set the vcpu->arch.* state.
>
> Similarly, svm->nested.ctl contains fields that are not used, so having
> a full vmcb_control_area means passing uninitialized fields.
>
> Patches 1,3 and 8 take care of renaming and refactoring code.
> Patches 2 and 6 introduce respectively vmcb_ctrl_area_cached and
> vmcb_save_area_cached.
> Patches 4 and 5 use vmcb_save_area_cached to avoid TOC/TOU, and patch
> 7 uses vmcb_ctrl_area_cached.
>
> Signed-off-by: Emanuele Giuseppe Esposito <eesposit@xxxxxxxxxx>

Queued (for 5.17).  I changed the helper functions to have a "__" prefix 
since that's more common in Linux.

Paolo

> ---
> v5:
> * rebased on kvm/queue branch
>
> v4:
> * introduce _* helpers (_nested_vmcb_check_save,
>    _nested_copy_vmcb_control_to_cache, _nested_copy_vmcb_save_to_cache)
>    that take care of additional parameters.
> * svm_set_nested_state: introduce {save, ctl}_cached variables
>    to not pollute svm->nested.{save,ctl} state, especially if the
>    check fails. remove also unnecessary memset added in previous versions.
> * svm_get_nested_state: change stack variable ctl introduced in this series
>   into a pointer that will be zeroed and freed after it has been copied to user
>
> v3:
> * merge this series with "KVM: nSVM: use vmcb_ctrl_area_cached instead
>    of vmcb_control_area in nested state"
> * rename "nested_load_save_from_vmcb12" in
>    "nested_copy_vmcb_save_to_cache"
> * rename "nested_load_control_from_vmcb12" in
>    "nested_copy_vmcb_control_to_cache"
> * change check functions (nested_vmcb_valid_sregs and nested_vmcb_valid_sregs)
>    to accept only the vcpu parameter, since we only check
>    nested state now
> * rename "vmcb_is_intercept_cached" in "vmcb12_is_intercept"
>    and duplicate the implementation instead of calling vmcb_is_intercept
>
> v2:
> * svm->nested.save is a separate struct vmcb_save_area_cached,
>    and not vmcb_save_area.
> * update also vmcb02->cr3 with svm->nested.save.cr3
>
> RFC:
> * use svm->nested.save instead of local variables.
> * not dependent anymore from "KVM: nSVM: remove useless kvm_clear_*_queue"
> * simplified patches, we just use the struct and not move the check
>    nearer to the TOU.
>
> Emanuele Giuseppe Esposito (7):
>    KVM: nSVM: move nested_vmcb_check_cr3_cr4 logic in
>      nested_vmcb_valid_sregs
>    nSVM: introduce smv->nested.save to cache save area fields
>    nSVM: rename nested_load_control_from_vmcb12 in
>      nested_copy_vmcb_control_to_cache
>    nSVM: use vmcb_save_area_cached in nested_vmcb_valid_sregs()
>    nSVM: use svm->nested.save to load vmcb12 registers and avoid TOC/TOU
>      races
>    nSVM: introduce struct vmcb_ctrl_area_cached
>    nSVM: use vmcb_ctrl_area_cached instead of vmcb_control_area in struct
>      svm_nested_state
>
>   arch/x86/kvm/svm/nested.c | 250 ++++++++++++++++++++++++--------------
>   arch/x86/kvm/svm/svm.c    |   7 +-
>   arch/x86/kvm/svm/svm.h    |  59 ++++++++-
>   3 files changed, 213 insertions(+), 103 deletions(-)