Re: [RFC PATCH 0/7] x86/entry: Atomic statck switching for IST

[Lai Jiangshan]

On Mon, Apr 3, 2023 at 10:23 PM Dave Hansen <dave.hansen@xxxxxxxxx> wrote:
>
> On 4/3/23 07:05, Lai Jiangshan wrote:
> > 2.3 #VE
> > -------
> >
> > The approach for fixing the kernel mode #VE recursion issue is to just
> > NOT use IST for #VE although #VE is also considered to be one of the
> > super exceptions and had raised some worries:
> > https://lore.kernel.org/lkml/YCEQiDNSHTGBXBcj@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/
> > https://lore.kernel.org/lkml/CALCETrU9XypKbj-TrXLB3CPW6=MZ__5ifLz0ckbB=c=Myegn9Q@xxxxxxxxxxxxxx/
> > https://lore.kernel.org/lkml/1843debc-05e8-4d10-73e4-7ddce3b3eae2@xxxxxxxxx/
> >
> > To remit the worries,  SEPT_VE_DISABLE is forced used currently and
> > also disables its abilities (accept-on-demand or memory balloon which
> > is critical to lightweight VMs like Kata Containers):
> > https://lore.kernel.org/lkml/YCb0%2FDg28uI7TRD%2F@xxxxxxxxxx/
>
> You don't need #VE for accept-on-demand.  Pages go through _very_
> well-defined software choke points before they get used *and* before
> they get ballooned.  Thus:
>
> > https://lore.kernel.org/lkml/20230330114956.20342-3-kirill.shutemov@xxxxxxxxxxxxxxx/
>

Thanks for the information.

I will have a look to see how it supports memory balloons.

And if accept-on-demand were supported, do we still need this
CONFIG_UNACCEPTED_MEMORY?

> BTW, _who_ considers #VE to be a "super exception"?  Can you explain how
> it is any more "super" than #PF?  #PF can recurse.  You can take #PF in
> the entry paths.
>
> I kinda don't think you should be using TDX and #VE as part of the
> justification for this series.

You are right, #VE is not a super exception anymore since SEPT_VE_DISABLE
is forced set in the Linux kernel and it is nothing to do with this series.

But #VE was once thought to be a super exception (I will correct the
sentence in the cover letter), so it is worth mentioning it.

And since SEPT_VE_DISABLE is configurable, it would allow some paranoids
to have a try with SEPT_VE_DISABLE=false even without FRED.
The paranoids can try it with IST #VE.

Thanks
Lai