On Sat, Sep 18, 2021, Zeng Guang wrote: > On 9/11/2021 7:55 AM, Sean Christopherson wrote: > > On Fri, Sep 10, 2021, Sean Christopherson wrote: > > > On Mon, Aug 09, 2021, Zeng Guang wrote: > > > > + if (!pages) > > > > + return -ENOMEM; > > > > + > > > > + to_kvm_vmx(kvm)->pid_table = (void *)page_address(pages); > > > > + to_kvm_vmx(kvm)->pid_last_index = KVM_MAX_VCPU_ID; > > > I don't see the point of pid_last_index if we're hardcoding it to KVM_MAX_VCPU_ID. > > > If I understand the ucode pseudocode, there's no performance hit in the happy > > > case, i.e. it only guards against out-of-bounds accesses. > > > > > > And I wonder if we want to fail the build if this grows beyond an order-1 > > > allocation, e.g. > > > > > > BUILD_BUG_ON(PID_TABLE_ORDER > 1); > > > > > > Allocating two pages per VM isn't terrible, but 4+ starts to get painful when > > > considering the fact that most VMs aren't going to need more than one page. For > > > now I agree the simplicity of not dynamically growing the table is worth burning > > > a page. > > Ugh, Paolo has queued a series which bumps KVM_MAX_VCPU_ID to 4096[*]. That makes > > this an order-3 allocation, which is quite painful. One thought would be to let > > userspace declare the max vCPU it wants to create, not sure if that would work for > > xAPIC though. > > > > [*] https://lkml.kernel.org/r/1111efc8-b32f-bd50-2c0f-4c6f506b544b@xxxxxxxxxx > Thus we keep current design as no change. Not necessarily. I was pointing out that the current design is already problematic from a memory allocation perspective. Burning a few pages per vCPU isn't the end of the world, but 32kb of _contiguous_ memory is rough, especially when 28kb is unlikely to be used in many cases.
