On Mon, Nov 15, 2021 at 4:03 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
>
> On Mon, Nov 15, 2021, Ben Gardon wrote:
> > When recursively clearing out disconnected pts, the range based TLB
> > flush in handle_removed_tdp_mmu_page uses the wrong starting GFN,
> > resulting in the flush mostly missing the affected range. Fix this by
> > using base_gfn for the flush.
> >
> > In response to feedback from David Matlack on the RFC version of this
> > patch, also move a few definitions into the for loop in the function to
> > prevent unintended references to them in the future.
>
> Rats, I didn't read David's feedback or I would've responded there.
>
> > Fixes: a066e61f13cf ("KVM: x86/mmu: Factor out handling of removed page tables")
> > CC: stable@xxxxxxxxxxxxxxx
> >
> > Signed-off-by: Ben Gardon <bgardon@xxxxxxxxxx>
> > ---
> > arch/x86/kvm/mmu/tdp_mmu.c | 10 ++++------
> > 1 file changed, 4 insertions(+), 6 deletions(-)
> >
> > diff --git a/arch/x86/kvm/mmu/tdp_mmu.c b/arch/x86/kvm/mmu/tdp_mmu.c
> > index 7c5dd83e52de..4bd541050d21 100644
> > --- a/arch/x86/kvm/mmu/tdp_mmu.c
> > +++ b/arch/x86/kvm/mmu/tdp_mmu.c
> > @@ -317,9 +317,6 @@ static void handle_removed_tdp_mmu_page(struct kvm *kvm, tdp_ptep_t pt,
> > struct kvm_mmu_page *sp = sptep_to_sp(rcu_dereference(pt));
> > int level = sp->role.level;
> > gfn_t base_gfn = sp->gfn;
> > - u64 old_child_spte;
> > - u64 *sptep;
> > - gfn_t gfn;
> > int i;
> >
> > trace_kvm_mmu_prepare_zap_page(sp);
> > @@ -327,8 +324,9 @@ static void handle_removed_tdp_mmu_page(struct kvm *kvm, tdp_ptep_t pt,
> > tdp_mmu_unlink_page(kvm, sp, shared);
> >
> > for (i = 0; i < PT64_ENT_PER_PAGE; i++) {
> > - sptep = rcu_dereference(pt) + i;
> > - gfn = base_gfn + i * KVM_PAGES_PER_HPAGE(level);
> > + u64 *sptep = rcu_dereference(pt) + i;
> > + gfn_t gfn = base_gfn + i * KVM_PAGES_PER_HPAGE(level);
> > + u64 old_child_spte;
>
> TL;DR: this type of optional refactoring doesn't belong in a patch Cc'd for stable,
> and my personal preference is to always declare variables at function scope (it's
> not a hard rule though, Paolo has overruled me at least once :-) ).
That makes sense. I don't have a preference either way. Paolo, if you
want the version without the refactor, the version I sent in the RFC
should be good. If the refactor is desired, I can separate it out into
another patch and send a v2 of this patch as a mini series, tagging
only the fix for stable.
I've generally preferred declaring variables at function scope too
since that seems like the overwhelming convention, but it's always
struck me as a bit of a waste to not make use of scoping rules more.
It does make it nice and clear how things should be laid out when
debugging the kernel with GDB or something though.
In any case, please let me know how you'd like the changes organized
and I can send up follow ups as needed, or we can just move forward
with the RFC version.
>
> Declaring variables in an inner scope is not always "better". In particular, it
> can lead to variable shadowing, which can lead to functional issues of a different
> sort. Most shadowing is fairly obvious, and truly egregious bugs will often result
> in the compiler complaining about consuming an uninitialized variable.
>
> But the worst-case scenario is if the inner scope shadows a function parameter, in
> which the case the compiler will not complain and will even consume an uninitialized
> variable without warning. IIRC, we actually had a Hyper-V bug of that nature
> where an incoming @vcpu was shadowed. Examples below.
>
> So yes, on one hand moving the declarations inside the loop avoid potential flavor
> of bug, but they create the possibility for an entirely different class of bugs.
> The main reason I prefer declaring at function scope is that I find it easier to
> visually detect using variables after a for loop, versus detecting that a variable
> is being shadowed, especially if the function is largish and the two declarations
> don't fit on the screen.
>
> There are of course counter-examples, e.g. commit 5c49d1850ddd ("KVM: VMX: Fix a
> TSX_CTRL_CPUID_CLEAR field mask issue") immediately jumps to mind, so there's
> certainly an element of personal preference.
>
> E.g. this will fail with "error: ‘sptep’ redeclared as different kind of symbol
>
> diff --git a/arch/x86/kvm/mmu/tdp_mmu.c b/arch/x86/kvm/mmu/tdp_mmu.c
> index 4e226cdb40d9..011639bf633c 100644
> --- a/arch/x86/kvm/mmu/tdp_mmu.c
> +++ b/arch/x86/kvm/mmu/tdp_mmu.c
> @@ -369,7 +369,7 @@ static void tdp_mmu_unlink_page(struct kvm *kvm, struct kvm_mmu_page *sp,
> * early rcu_dereferences in the function.
> */
> static void handle_removed_tdp_mmu_page(struct kvm *kvm, tdp_ptep_t pt,
> - bool shared)
> + bool shared, u64 *sptep)
> {
> struct kvm_mmu_page *sp = sptep_to_sp(rcu_dereference(pt));
> int level = sp->role.level;
> @@ -431,8 +431,9 @@ static void handle_removed_tdp_mmu_page(struct kvm *kvm, tdp_ptep_t pt,
> shared);
> }
>
> - kvm_flush_remote_tlbs_with_address(kvm, gfn,
> - KVM_PAGES_PER_HPAGE(level + 1));
> + if (sptep)
> + kvm_flush_remote_tlbs_with_address(kvm, gfn,
> + KVM_PAGES_PER_HPAGE(level + 1));
>
> call_rcu(&sp->rcu_head, tdp_mmu_free_sp_rcu_callback);
> }
> @@ -532,7 +533,7 @@ static void __handle_changed_spte(struct kvm *kvm, int as_id, gfn_t gfn,
> */
> if (was_present && !was_leaf && (is_leaf || !is_present))
> handle_removed_tdp_mmu_page(kvm,
> - spte_to_child_pt(old_spte, level), shared);
> + spte_to_child_pt(old_spte, level), shared, NULL);
> }
>
> static void handle_changed_spte(struct kvm *kvm, int as_id, gfn_t gfn,
>
>
> whereas moving the second declaration into the loop will compile happily.
>
> diff --git a/arch/x86/kvm/mmu/tdp_mmu.c b/arch/x86/kvm/mmu/tdp_mmu.c
> index 4e226cdb40d9..3e83fd66c0dc 100644
> --- a/arch/x86/kvm/mmu/tdp_mmu.c
> +++ b/arch/x86/kvm/mmu/tdp_mmu.c
> @@ -369,13 +369,12 @@ static void tdp_mmu_unlink_page(struct kvm *kvm, struct kvm_mmu_page *sp,
> * early rcu_dereferences in the function.
> */
> static void handle_removed_tdp_mmu_page(struct kvm *kvm, tdp_ptep_t pt,
> - bool shared)
> + bool shared, u64 *sptep)
> {
> struct kvm_mmu_page *sp = sptep_to_sp(rcu_dereference(pt));
> int level = sp->role.level;
> gfn_t base_gfn = sp->gfn;
> u64 old_child_spte;
> - u64 *sptep;
> gfn_t gfn;
> int i;
>
> @@ -384,7 +383,7 @@ static void handle_removed_tdp_mmu_page(struct kvm *kvm, tdp_ptep_t pt,
> tdp_mmu_unlink_page(kvm, sp, shared);
>
> for (i = 0; i < PT64_ENT_PER_PAGE; i++) {
> - sptep = rcu_dereference(pt) + i;
> + u64 *sptep = rcu_dereference(pt) + i;
> gfn = base_gfn + i * KVM_PAGES_PER_HPAGE(level);
>
> if (shared) {
> @@ -431,8 +430,9 @@ static void handle_removed_tdp_mmu_page(struct kvm *kvm, tdp_ptep_t pt,
> shared);
> }
>
> - kvm_flush_remote_tlbs_with_address(kvm, gfn,
> - KVM_PAGES_PER_HPAGE(level + 1));
> + if (sptep)
> + kvm_flush_remote_tlbs_with_address(kvm, gfn,
> + KVM_PAGES_PER_HPAGE(level + 1));
>
> call_rcu(&sp->rcu_head, tdp_mmu_free_sp_rcu_callback);
> }
> @@ -532,7 +532,7 @@ static void __handle_changed_spte(struct kvm *kvm, int as_id, gfn_t gfn,
> */
> if (was_present && !was_leaf && (is_leaf || !is_present))
> handle_removed_tdp_mmu_page(kvm,
> - spte_to_child_pt(old_spte, level), shared);
> + spte_to_child_pt(old_spte, level), shared, NULL);
> }
>
> static void handle_changed_spte(struct kvm *kvm, int as_id, gfn_t gfn,
