Re: [RFC][PATCH] x86,ibt: Use UDB instead of 0xEA

[Peter Zijlstra <peterz@infradead.org>]

On Fri, Aug 15, 2025 at 12:28:39PM +0200, Peter Zijlstra wrote:
> On Fri, Aug 15, 2025 at 09:49:39AM +0200, Peter Zijlstra wrote:
> > On Thu, Aug 14, 2025 at 06:27:44PM -0700, H. Peter Anvin wrote:
> > > On 2025-08-14 04:17, Peter Zijlstra wrote:
> > > > Hi!
> > > > 
> > > > A while ago FineIBT started using the instruction 0xEA to generate #UD.
> > > > All existing parts will generate #UD in 64bit mode on that instruction.
> > > > 
> > > > However; Intel/AMD have not blessed using this instruction, it is on
> > > > their 'reserved' list for future use.
> > > > 
> > > > Peter Anvin worked the committees and got use of 0xD6 blessed, and it
> > > > will be called UDB (per the next SDM or so).
> > > > 
> > > > Reworking the FineIBT code to use UDB wasn't entirely trivial, and I've
> > > > had to switch the hash register to EAX in order to free up some bytes.
> > > > 
> > > > Per the x86_64 ABI, EAX is used to pass the number of vector registers
> > > > for varargs -- something that should not happen in the kernel. More so,
> > > > we build with -mskip-rax-setup, which should leave EAX completely unused
> > > > in the calling convention.
> > > > 
> > > > The code boots and passes the LKDTM CFI_FORWARD_PROTO test for various
> > > > combinations (non exhaustive so far).
> > > > 
> > > > Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> > > 
> > > Looks good to me (and using %eax will save one byte per call site as
> > > well), but as per our IRC discussion, *my understanding* is that the
> > > best possible performance (least branch predictor impact across
> > > implementations) is to use a forward branch with a 2E prefix (jcc,pn in
> > > GAS syntax) rather than a reverse branch, if space allows.
> > 
> > Oh right. I did see that comment on IRC and them promptly forgot about
> > it again :/ I'll have a poke. Scott, do you agree? You being responsible
> > for the backward jump and such.
> 
> On top of the other, to show the delta.
> 
> If we want a fwd branch, we can stick the D6 in the endbr poison nop.
> That frees up more bytes again, but also that matches what I already did
> for the bhi1 case, so less special cases is more better.
> 
> I've had to use cs prefixed jcc.d32, because our older toolchains don't
> like the ,pn notation.

And then I forgot to move that cs prefix around in the bhi1 case...
fixed that.