On Thu, Jul 04, 2019 at 08:32:52AM +0530, Viresh Kumar wrote:
> On 03-07-19, 18:42, Dave Martin wrote:
> > From: Zhang Lei <zhang.lei@xxxxxxxxxxxxxx>
> >
> > The original implementation of vq_present() relied on aggressive
> > inlining in order for the compiler to know that the code is
> > correct, due to some const-casting issues. This was causing sparse
> > and clang to complain, while GCC compiled cleanly.
> >
> > Commit 0c529ff789bc addressed this problem, but since vq_present()
> > is no longer a function, there is now no implicit casting of the
> > returned value to the return type (bool).
> >
> > In set_sve_vls(), this uncast bit value is compared against a bool,
> > and so may spuriously compare as unequal when both are nonzero. As
> > a result, KVM may reject valid SVE vector length configurations as
> > invalid, and vice versa.
> >
> > Fix it by forcing the returned value to a bool.
> >
> > Signed-off-by: Zhang Lei <zhang.lei@xxxxxxxxxxxxxx>
> > Fixes: 0c529ff789bc ("KVM: arm64: Implement vq_present() as a macro")
> > Signed-off-by: Dave Martin <Dave.Martin@xxxxxxx> [commit message rewrite]
> > Cc: Viresh Kumar <viresh.kumar@xxxxxxxxxx>
> >
> > ---
> >
> > Posting this under Zhang Lei's authorship, due to the need to turn this
> > fix around quickly. The fix is as per the original suggestion [1].
> >
> > Originally observed with the QEMU KVM SVE support under review:
> > https://lists.gnu.org/archive/html/qemu-devel/2019-06/msg04945.html
> >
> > Bug reproduced and fix tested on the Arm Fast Model, with
> > http://linux-arm.org/git?p=kvmtool-dm.git;a=shortlog;h=refs/heads/sve/v3/head
> > (After rerunning util/update_headers.sh.)
> >
> > (the --sve-vls command line argument was removed in v4 of the
> > kvmtool patches).
> >
> > [1] http://lists.infradead.org/pipermail/linux-arm-kernel/2019-July/664633.html
> > ---
> > arch/arm64/kvm/guest.c | 2 +-
> > 1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/arch/arm64/kvm/guest.c b/arch/arm64/kvm/guest.c
> > index c2afa79..dfd6264 100644
> > --- a/arch/arm64/kvm/guest.c
> > +++ b/arch/arm64/kvm/guest.c
> > @@ -208,7 +208,7 @@ static int set_core_reg(struct kvm_vcpu *vcpu, const struct kvm_one_reg *reg)
> >
> > #define vq_word(vq) (((vq) - SVE_VQ_MIN) / 64)
> > #define vq_mask(vq) ((u64)1 << ((vq) - SVE_VQ_MIN) % 64)
> > -#define vq_present(vqs, vq) ((vqs)[vq_word(vq)] & vq_mask(vq))
> > +#define vq_present(vqs, vq) (!!((vqs)[vq_word(vq)] & vq_mask(vq)))
> >
> > static int get_sve_vls(struct kvm_vcpu *vcpu, const struct kvm_one_reg *reg)
> > {
>
> It was a really nice bug :)
>
> Reviewed-by: Viresh Kumar <viresh.kumar@xxxxxxxxxx>
Thanks for the quick review!
Maybe it makes sense to write equality comparisons on bools as !x == !y
to be more defensive against this kind of thing. Anyway, probably best
to leave this as-is while the dust settles.
Cheers
---Dave
_______________________________________________
kvmarm mailing list
kvmarm@xxxxxxxxxxxxxxxxxxxxx
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
