On Tue, Apr 10, 2018 at 11:51:19AM +0100, Mark Rutland wrote: > On Mon, Apr 09, 2018 at 10:51:39PM +0200, Christoffer Dall wrote: > > On Mon, Apr 09, 2018 at 06:07:06PM +0100, Marc Zyngier wrote: > > > Before entering the guest, we check whether our VMID is still > > > part of the current generation. In order to avoid taking a lock, > > > we start with checking that the generation is still current, and > > > only if not current do we take the lock, recheck, and update the > > > generation and VMID. > > > > > > This leaves open a small race: A vcpu can bump up the global > > > generation number as well as the VM's, but has not updated > > > the VMID itself yet. > > > > > > At that point another vcpu from the same VM comes in, checks > > > the generation (and finds it not needing anything), and jumps > > > into the guest. At this point, we end-up with two vcpus belonging > > > to the same VM running with two different VMIDs. Eventually, the > > > VMID used by the second vcpu will get reassigned, and things will > > > really go wrong... > > > > > > A simple solution would be to drop this initial check, and always take > > > the lock. This is likely to cause performance issues. A middle ground > > > is to convert the spinlock to a rwlock, and only take the read lock > > > on the fast path. If the check fails at that point, drop it and > > > acquire the write lock, rechecking the condition. > > > > > > This ensures that the above scenario doesn't occur. > > > > > > Reported-by: Mark Rutland <mark.rutland@xxxxxxx> > > > Signed-off-by: Marc Zyngier <marc.zyngier@xxxxxxx> > > > --- > > > I haven't seen any reply from Shannon, so reposting this to > > > a slightly wider audience for feedback. > > > > > > virt/kvm/arm/arm.c | 15 ++++++++++----- > > > 1 file changed, 10 insertions(+), 5 deletions(-) > > > > > > diff --git a/virt/kvm/arm/arm.c b/virt/kvm/arm/arm.c > > > index dba629c5f8ac..a4c1b76240df 100644 > > > --- a/virt/kvm/arm/arm.c > > > +++ b/virt/kvm/arm/arm.c > > > @@ -63,7 +63,7 @@ static DEFINE_PER_CPU(struct kvm_vcpu *, kvm_arm_running_vcpu); > > > static atomic64_t kvm_vmid_gen = ATOMIC64_INIT(1); > > > static u32 kvm_next_vmid; > > > static unsigned int kvm_vmid_bits __read_mostly; > > > -static DEFINE_SPINLOCK(kvm_vmid_lock); > > > +static DEFINE_RWLOCK(kvm_vmid_lock); > > > > > > static bool vgic_present; > > > > > > @@ -473,11 +473,16 @@ static void update_vttbr(struct kvm *kvm) > > > { > > > phys_addr_t pgd_phys; > > > u64 vmid; > > > + bool new_gen; > > > > > > - if (!need_new_vmid_gen(kvm)) > > > + read_lock(&kvm_vmid_lock); > > > + new_gen = need_new_vmid_gen(kvm); > > > + read_unlock(&kvm_vmid_lock); > > > + > > > + if (!new_gen) > > > return; > > > > > > - spin_lock(&kvm_vmid_lock); > > > + write_lock(&kvm_vmid_lock); > > > > > > /* > > > * We need to re-check the vmid_gen here to ensure that if another vcpu > > > @@ -485,7 +490,7 @@ static void update_vttbr(struct kvm *kvm) > > > * use the same vmid. > > > */ > > > if (!need_new_vmid_gen(kvm)) { > > > - spin_unlock(&kvm_vmid_lock); > > > + write_unlock(&kvm_vmid_lock); > > > return; > > > } > > > > > > @@ -519,7 +524,7 @@ static void update_vttbr(struct kvm *kvm) > > > vmid = ((u64)(kvm->arch.vmid) << VTTBR_VMID_SHIFT) & VTTBR_VMID_MASK(kvm_vmid_bits); > > > kvm->arch.vttbr = kvm_phys_to_vttbr(pgd_phys) | vmid; > > > > > > - spin_unlock(&kvm_vmid_lock); > > > + write_unlock(&kvm_vmid_lock); > > > } > > > > > > static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu) > > > -- > > > 2.14.2 > > > > > > > The above looks correct to me. I am wondering if something like the > > following would also work, which may be slightly more efficient, > > although I doubt the difference can be measured: > > [...] > > I think we also need to update kvm->arch.vttbr before updating > kvm->arch.vmid_gen, otherwise another CPU can come in, see that the > vmid_gen is up-to-date, jump to hyp, and program a stale VTTBR (with the > old VMID). > > With the smp_wmb() and update of kvm->arch.vmid_gen moved to the end of > the critical section, I think that works, modulo using READ_ONCE() and > WRITE_ONCE() to ensure single-copy-atomicity of the fields we access > locklessly. Indeed, you're right. I would look something like this, then: diff --git a/virt/kvm/arm/arm.c b/virt/kvm/arm/arm.c index 2e43f9d42bd5..6cb08995e7ff 100644 --- a/virt/kvm/arm/arm.c +++ b/virt/kvm/arm/arm.c @@ -450,7 +450,9 @@ void force_vm_exit(const cpumask_t *mask) */ static bool need_new_vmid_gen(struct kvm *kvm) { - return unlikely(kvm->arch.vmid_gen != atomic64_read(&kvm_vmid_gen)); + u64 current_vmid_gen = atomic64_read(&kvm_vmid_gen); + smp_rmb(); /* Orders read of kvm_vmid_gen and kvm->arch.vmid */ + return unlikely(READ_ONCE(kvm->arch.vmid_gen) != current_vmid_gen); } /** @@ -500,7 +502,6 @@ static void update_vttbr(struct kvm *kvm) kvm_call_hyp(__kvm_flush_vm_context); } - kvm->arch.vmid_gen = atomic64_read(&kvm_vmid_gen); kvm->arch.vmid = kvm_next_vmid; kvm_next_vmid++; kvm_next_vmid &= (1 << kvm_vmid_bits) - 1; @@ -509,7 +510,10 @@ static void update_vttbr(struct kvm *kvm) pgd_phys = virt_to_phys(kvm->arch.pgd); BUG_ON(pgd_phys & ~VTTBR_BADDR_MASK); vmid = ((u64)(kvm->arch.vmid) << VTTBR_VMID_SHIFT) & VTTBR_VMID_MASK(kvm_vmid_bits); - kvm->arch.vttbr = pgd_phys | vmid; + WRITE_ONCE(kvm->arch.vttbr, pgd_phys | vmid); + + smp_wmb(); /* Ensure vttbr update is observed before vmid_gen update */ + kvm->arch.vmid_gen = atomic64_read(&kvm_vmid_gen); spin_unlock(&kvm_vmid_lock); } It's probably easier to convince ourselves about the correctness of Marc's code using a rwlock instead, though. Thoughts? Thanks, -Christoffer