On Fri, 6 Jul 2018, Paolo Bonzini wrote:
> On 06/07/2018 11:24, Thomas Gleixner wrote:
> >> The reason for this is to avoid wasting a lot of BSS memory when KVM is
> >> not in use. Thomas is going to send his take on this!
> > Got it working with per cpu variables, but there is a different subtle
> > issue with that.
> >
> > The pvclock data is mapped into the VDSO as well, i.e. as a full page.
> >
> > Right now with the linear array, which is forced to be page sized at least
> > this only maps pvclock data or zeroed data (after the last CPU) into the
> > VDSO.
> >
> > With PER CPU variables this would map arbitraty other per cpu data which
> > happens to be in the same page into the VDSO. Not really what we want.
> >
> > That means to utilize PER CPU data this requires to allocate page sized
> > pvclock data space for each CPU to prevent leaking arbitrary stuff.
> >
> > As this data is allocated on demand, i.e. only if kvmclock is used, this
> > might be tolerable, but I'm not so sure.
>
> One possibility is to introduce another layer of indirection: in
> addition to the percpu pvclock data, add a percpu pointer to the pvclock
> data and initialize it to point to a page-aligned variable in BSS. CPU0
> (used by vDSO) doesn't touch the pointer and keeps using the BSS
> variable, APs instead redirect the pointer to the percpu data.
Yeah, thought about that, but the extra indirection is ugly. Instead of
using per cpu data, I just can allocate the memory _after_ the allocators
are up and running and use a single page sized static __initdata for the
early boot.
Thanks,
tglx
