Re: [PATCH 02/24] x86/traps: Move arch/x86/kernel/traps.c to arch/x86/entry/

[Lai Jiangshan]

On 2021/9/2 18:50, Peter Zijlstra wrote:
> On Thu, Sep 02, 2021 at 05:21:51PM +0800, Lai Jiangshan wrote:
> > On 2021/9/2 16:09, Joerg Roedel wrote:
> > > On Wed, Sep 01, 2021 at 01:50:03AM +0800, Lai Jiangshan wrote:
> > > >    arch/x86/entry/Makefile            | 5 ++++-
> > > >    arch/x86/{kernel => entry}/traps.c | 0
> > > >    arch/x86/kernel/Makefile           | 5 +----
> > > >    3 files changed, 5 insertions(+), 5 deletions(-)
> > > >    rename arch/x86/{kernel => entry}/traps.c (100%)
> > >
> > >   From looking over the patch-set I didn't spot the reason for putting the
> > > entry C code into traps.c. Can you explain that please? Otherwise I'd
> > > prefer to create a new file, like arch/x86/entry/entry64.c.
> >
> >
> > I agree, and I think Peter is handling it.
>
> I don't think I said that. I said I like the patches but there's a lot
> of scary code and details to review, which takes time.
>
> I've now done a second reading of the patches and provided some more
> feedback, but I'm very sure I didn't get to the scary details yet.
>
> One thing that got pointed out (by Andrew Cooper) is that by moving the
> whole SWAPGS trainwreck to C it becomes entirely 'trivial' to sneak in
> an 'accidental' per-cpu reference before having done the SWAPGS dance.
>
> I'm myself not (yet) convinced that's a good enough reason to leave it
> in ASM, but it does certainly need consideration.


It is real concern and it proves that my having put the C code in traps.c
was totally wrong.

To relieve the concern, I think the C code can be put into a single file, like
arch/x86/entry/entry64.c,  and be documented that it is as critical, dangerous as
entry_64.S and any one should take no less care on modifying/reviewing it than on
modifying/reviewing entry_64.S.

And all the other users of native_swapgs can be moved to this file too, such as
__[rd|wr]gsbase_inactive().

A noninstr function can sometimes have 'accidental' instrument to sneak in.
For example, stack-protector is instrumenting many noninstr functions now
if the CONFIG is yes.  It is normally Ok and gcc is adding per-function control
on it.

But the C code can not be instrumented by any way.  For example stack-protector
would add per-cpu reference before having done the SWAPGS dance.)  Entry C code
required a stronger limitation than noninstr code.

By the way, can objtool check the per-cpu reference?