Re: [PATCH 3/3] KVM: Avoid synchronize_srcu() in kvm_io_bus_register_dev()

[Sean Christopherson <seanjc@google.com>]

On Tue, Jul 15, 2025, Keir Fraser wrote:
> On Mon, Jul 07, 2025 at 11:49:34AM -0700, Sean Christopherson wrote:
> > For all intents and purposes, holding kvm->srcu across VM-Enter/VM-Exit is
> > disallowed (though I don't think this is formally documented), i.e. every
> > architecture is guaranteed to do srcu_read_lock() after a VM-Exit, prior to
> > reading kvm->buses.  And srcu_read_lock() contains a full smp_mb(), which ensures
> > KVM will get a fresh kvm->buses relative to the instruction that triggered the
> > exit.
> 
> I've got a new patch series ready to go, but thinking more about the
> one-off accesses after a VM-Exit: I think VM-Exit is a barrier on all
> architectures? That would mean the changes to include
> smp_mb__after_srcu_read_lock() are unnecessary and confusing. Maybe I
> can drop those hunks. What do you think?

It's not.

commit 65a4de0ffd975af7e2ffc9acb875b6a8ae7ee1aa
Author:     Yan Zhao <yan.y.zhao@intel.com>
AuthorDate: Fri Mar 8 17:09:28 2024 -0800
Commit:     Sean Christopherson <seanjc@google.com>
CommitDate: Fri Jun 7 07:18:02 2024 -0700

    KVM: x86: Ensure a full memory barrier is emitted in the VM-Exit path
    
    Ensure a full memory barrier is emitted in the VM-Exit path, as a full
    barrier is required on Intel CPUs to evict WC buffers.  This will allow
    unconditionally honoring guest PAT on Intel CPUs that support self-snoop.