Re: drivers/gpu/drm/i915/i915.prelink.o: warning: objtool: __intel_wait_for_register_fw.cold()+0xce: relocation to !ENDBR: vlv_allow_gt_wake.cold+0x0

[Peter Zijlstra <peterz@infradead.org>]

[-- Attachment #1: Type: text/plain, Size: 1400 bytes --]

On Wed, Apr 06, 2022 at 09:37:03AM -0700, Josh Poimboeuf wrote:
> On Wed, Apr 06, 2022 at 09:43:30AM +0200, Peter Zijlstra wrote:
> > On Tue, Apr 05, 2022 at 10:32:51PM -0700, Josh Poimboeuf wrote:
> > > More broadly, this issue could theoretically happen in some other places
> > > throughout the kernel tree, since _THIS_IP_ is fundamentally unreliable
> > > as currently written.
> > > 
> > > So we could look at making _THIS_IP_ more predictable.
> > > 
> > > Inline asm would work better ("lea 0(%rip), %[rip]"), but then you need
> > > an arch-dependent implementation...
> > 
> > Well, there's a ton of _THIS_IP_ instances all around, and it would be
> > unfortunate to have them grow into actual code :/
> 
> What do you mean by growing into actual code?  It's still just a single
> instruction, as was the immediate load before.

Aah, indeed. I was somehow thinking we'd get extra instructions.

> Though, you pasted this on irc:
> 
>   #define _THIS_IP_  ({ __label__ __here; __here: asm_volatile_goto ("":::: __here); (unsigned long)&&__here; })
> 
> which seems decent to me, though less than ideal because it grows an
> ENDBR.  But I like its arch-independence, so yeah, LGTM.

I did send hjl an email about that extra endbr, because I really don't
like that. And jump_label (also using asm-goto) doesn't grow those endbr
instructions, so something is weird.