Re: [PATCH 2/4] objtool: Implement jump_assert for _static_cpu_has()

[Josh Poimboeuf]

On Mon, Jan 15, 2018 at 01:08:19PM -0600, Josh Poimboeuf wrote:
> On Mon, Jan 15, 2018 at 07:59:37PM +0100, Borislav Petkov wrote:
> > Right,
> > 
> > I've been putting away extending struct alt_instr for a long time now,
> > trying to be conservative about it but I guess this might be the right
> > time to change that. How about:
> > 
> > struct alt_instr {
> >         s32 instr_offset;       /* original instruction */
> >         s32 repl_offset;        /* offset to replacement instruction */
> >         u16 cpuid;              /* cpuid bit set for replacement */
> >         u8  instrlen;           /* length of original instruction */
> >         u8  replacementlen;     /* length of new instruction */
> >         u8  padlen;             /* length of build-time padding */
> > 	u64 flags;		/* alternative flags, see <some enum> */
> > } __packed;
> > 
> > This way we have 64 settings. So we could do:
> > 
> > ...
> > 	.flags 	= ALT_FLAGS_STATIC_CPU_HAS,
> > 
> > or something like that and then we can do additional processing/matching
> > for the alternatives.
> > 
> > Or, we can do
> > 
> > struct alt_instr {
> >         s32 instr_offset;       /* original instruction */
> >         s32 repl_offset;        /* offset to replacement instruction */
> >         u16 cpuid;              /* cpuid bit set for replacement */
> >         u8  instrlen;           /* length of original instruction */
> >         u8  replacementlen;     /* length of new instruction */
> >         u8  padlen;             /* length of build-time padding */
> > 	u8 type;		/* types */
> > } __packed;
> > 
> > and have 256 types but that would be limiting as we won't be able to set
> > more than one.
> > 
> > Hmmm?
> 
> That might be a good idea, but here we also need to annotate jump
> labels.  So unless you want to make alternatives broad enough to
> encompass jump labels, I don't think it solves this particular problem.

Well, to clarify, it would solve _some_ of the problem.  Maybe even most
of the problem.  We'd still need to special-case jump labels in objtool
(like in 1/4), but that's probably not a big deal.

So, contradicting my previous answer here... yes, it would help.

-- 
Josh