Re: [PATCH RFC 4/4] x86/srso: Use CALL-based return thunks to reduce overhead

[Josh Poimboeuf]

On Tue, Aug 22, 2023 at 09:45:07AM +0300, Nikolay Borisov wrote:
> 
> 
> On 22.08.23 г. 5:22 ч., Josh Poimboeuf wrote:
> > On Tue, Aug 22, 2023 at 12:01:29AM +0100, Andrew Cooper wrote:
> > > On 21/08/2023 4:16 pm, Josh Poimboeuf wrote:
> > > > On Mon, Aug 21, 2023 at 12:27:23PM +0100, Andrew Cooper wrote:
> > > > > The SRSO safety depends on having a CALL to an {ADD,LEA}/RET sequence which
> > > > > has been made safe in the BTB.  Specifically, there needs to be no pertubance
> > > > > to the RAS between a correctly predicted CALL and the subsequent RET.
> > > > > 
> > > > > Use the new infrastructure to CALL to a return thunk.  Remove
> > > > > srso_fam1?_safe_ret() symbols and point srso_fam1?_return_thunk().
> > > > > 
> > > > > This removes one taken branch from every function return, which will reduce
> > > > > the overhead of the mitigation.  It also removes one of three moving pieces
> > > > > from the SRSO mess.
> > > > So, the address of whatever instruction comes after the 'CALL
> > > > srso_*_return_thunk' is added to the RSB/RAS, and that might be
> > > > speculated to when the thunk returns.  Is that a concern?
> > > 
> > > That is very intentional, and key to the safety.
> > > 
> > > Replacing a RET with a CALL/{ADD,LEA}/RET sequence is a form of
> > > retpoline thunk.  The only difference with regular retpolines is that
> > > the intended target is already on the stack, and not in a GPR.
> > > 
> > > 
> > > If the CALL mispredicts, it doesn't matter.  When decode catches up
> > > (allegedly either instantaneously on Fam19h, or a few cycles late on
> > > Fam17h), the top of the RAS is corrected will point at the INT3
> > > following the CALL instruction.
> > 
> > That's the thing though, at least with my kernel/compiler combo there's
> > no INT3 after the JMP __x86_return_thunk, and there's no room to patch
> > one in after the CALL, as the JMP and CALL are both 5 bytes.
> 
> FWIW gcc's mfunction-return=thunk-return only ever generates a jmp,
> thunk/thunk-inline OTOH generates a "full fledged" thunk with all the
> necessary speculation catching tricks.
> 
> For reference:
> 
> https://godbolt.org/z/M1avYc63b

The problem is the call-site, not the thunk.  Ideally we'd have an
option which adds an INT3 after the 'JMP __x86_return_thunk'.

-- 
Josh