Re: [PATCH 04/15] ia64/pv_ops: introduce pv_info which describes

[Jes Sorensen <jes@sgi.com>]

Isaku Yamahata wrote:
> Our justification is as follows.
> The difference is its scope. pv_ops for virtualization and
> machine vector is for platform difference.
> 
> - pv_ops does cover the area which shouldn't belong to machine vector.
>   For example, ia64 intrinsics paravirtualization.
>   It shouldn't belong to the machine vector.
>   It must be initialized very early before platform detection.

Hi Isaku,

Ok this is a good point.

> - pv_ops covers some performance critical part (e.g. ia64 intrinsics)
>   so that in the future they should be optimized with binary patch like x86.
>   We had the experimental patch to do that, but they are dropped for
>   the merge. It reduced patch size greatly.
>   After merging the first patch series, we're planning to optimize
>   pv_ops with binary patch.
>   The optimization with binary patch is out of the machine vector scope.

Rather than making these binary patches, why not make them fast syscalls
and using a vdso page. Some of the priviledged instructions are simply
reads and we could have that information in a read-only data page, so
there is no need to do a context switch at all. Others could benefit
from a fast system call that doesn't do a full context switch.

It would be nice if we could come up with a generic implementation for
such a vdso style interface that could be shared between xen/kvm/lguest.


> - The current pv_ops implements only one for only domU, but in future
>   pv_ops will support dom0. It means dom0 linux would run with
>   the underlying machine vector + pv_ops, i.e.
>   {dig, hpzx1, hpzx1_swiotlb, ...} machine vector + xen pv_ops
> 
> Probably some hooks of pv_ops could be replaced with
> enhancing machine vector. But from the above separating pv_ops from
> machine vector looks reasonable.

Would it make sense to make the pv_ops pointer part of the machine
vector?

Cheers,
Jes