On Wed, 2018-01-10 at 14:15 -0600, Josh Poimboeuf wrote:
> On Wed, Jan 10, 2018 at 08:55:40PM +0100, Thomas Gleixner wrote:
> > On Wed, 10 Jan 2018, Linus Torvalds wrote:
> > 
> > > On Wed, Jan 10, 2018 at 3:28 AM, Borislav Petkov <bp@alien8.de> wrote:
> > > >
> > > > Make sure we scan all bytes before we decide to optimize the NOPs in
> > > > there.
> > > 
> > > Can we also add compile-time checking (presumably in objtool, but who
> > > knows) that there are no relocations in the alternative section?
> > 
> > Cc'ing the overlor^Haded objtool wizard
> > 
> > > Because that was the other "oops, this really doesn't work with
> > > altinstructions" issue, wasn't it?
> 
> I think .altinstruction relocations *do* work if they're for the first
> instruction, and it's a jump or a call.  There's some alternatives code
> which adjusts the jump/call offset in that case, and there are some
> users of alternatives who rely on that.
> 
> I think Boris had a patch floating around to add an instruction decoder
> to alternatives, so you can do a call/jmp anywhere.

Strictly speaking, it's not about the ELF relocs. Those get processed
in advance, and the altinstructions already have them applied.

What Borislav had was a patch to process the actual branch
instructions, then frob their targets by {&altinstr - &oldinstr} to
make them work again... except it only worked by chance for targets
*within* the altinstr.

On the whole I think we're better off not touching that right now for
fear that it will introduce new bugs. But yes, it *should* be fixed....
just not this week please ;)