Hi Halil,
Dne 31. 01. 22 v 12:53 Halil Pasic napsal(a):
On Mon, 31 Jan 2022 11:13:18 +0100
Petr Tesařík <ptesarik@xxxxxxx> wrote:
Hi Halil,
Dne 27. 08. 21 v 14:54 Halil Pasic napsal(a):
While in practice vcpu->vcpu_idx == vcpu->vcp_id is often true,
it may not always be, and we must not rely on this.
Currently kvm->arch.idle_mask is indexed by vcpu_id, which implies
that code like
for_each_set_bit(vcpu_id, kvm->arch.idle_mask, online_vcpus) {
vcpu = kvm_get_vcpu(kvm, vcpu_id);
do_stuff(vcpu);
}
is not legit. The trouble is, we do actually use kvm->arch.idle_mask
like this. To fix this problem we have two options. Either use
kvm_get_vcpu_by_id(vcpu_id), which would loop to find the right vcpu_id,
or switch to indexing via vcpu_idx. The latter is preferable for obvious
reasons.
I'm just backporting this fix to SLES 12 SP5, and I've noticed that
there is still this code in __floating_irq_kick():
/* find idle VCPUs first, then round robin */
sigcpu = find_first_bit(fi->idle_mask, online_vcpus);
/* ... round robin loop removed ...
dst_vcpu = kvm_get_vcpu(kvm, sigcpu);
It seems to me that this does exactly the thing that is not legit, but
I'm no expert. Did I miss something?
We made that legit by making the N-th bit in idle_mask correspond to the
vcpu whose vcpu_idx == N. The second argument of kvm_get_vcpu() is the
vcpu_idx. IMHO that ain't super-intuitive because it ain't spelled out.
So before this was a mismatch (with a vcpu_id based bitmap we would have
to use kvm_get_vcpu_by_id()), and with this patch applied this code
becomes legit because both idle_mask and kvm_get_vcpu() operate with
vcpu_idx.
Does that make sense?
Yes!
I'm sorry the commit message did not convey this clearly enough...
No, it's not your fault. I didn't pay enough attention to the change,
and with vcpu_id and vcpu_idx being so similar I got confused.
In short, there's no bug now, indeed. Thanks for your patience.
Petr T
