Re: [kvm PATCH v4 0/2] use vmalloc to allocate vmx vcpus

[Marc Orr <marcorr@google.com>]

On Fri, Oct 26, 2018 at 7:49 AM Dave Hansen <dave.hansen@intel.com> wrote:
>
> On 10/26/18 7:45 AM, Matthew Wilcox wrote:
> >         struct fpu                 user_fpu;             /*  2176  4160 */
> >         struct fpu                 guest_fpu;            /*  6336  4160 */
>
> Those are *not* supposed to be embedded in any other structures.  My bad
> for not documenting this better.
>
> It also seems really goofy that we need an xsave buffer in the
> task_struct for user fpu state, then another in the vcpu.  Isn't one for
> user state enough?
>
> In any case, I'd suggest getting rid of 'user_fpu', then either moving
> 'guest_fpu' to the bottom of the structure, or just make it a 'struct
> fpu *' and dynamically allocating it separately.

I've written a patch to get rid of user_fpu, as suggested here and
will be sending that out shortly.

>
> To do this, I'd take fpu__init_task_struct_size(), and break it apart a
> bit to tell you the size of the 'struct fpu' separately from the size of
> the 'task struct'.

I've written a 2nd patch to make guest_cpu  a 'struct fpu *' and
dynamically allocate it separately. The reason I went with this
suggestion, rather than moving 'struct fpu' to the bottom of
kvm_vcpu_arch is because I believe that solution would still expand
the kvm_vcpu_arch by the size of the fpu, according to which
fpregs_state was in use.