Re: [RFC PATCH v3 11/12] powerpc: Remove unreachable() from WARN_ON()

[Peter Zijlstra]

On Mon, Jul 04, 2022 at 12:44:30PM +0000, Christophe Leroy wrote:
> 
> 
> Le 04/07/2022 à 14:05, Peter Zijlstra a écrit :
> > On Sat, Jun 25, 2022 at 06:46:54AM +0000, Christophe Leroy wrote:
> >>
> >>
> >> Le 24/06/2022 à 20:32, Sathvika Vasireddy a écrit :
> >>> objtool is throwing *unannotated intra-function call*
> >>> warnings with a few instructions that are marked
> >>> unreachable. Remove unreachable() from WARN_ON()
> >>> to fix these warnings, as the codegen remains same
> >>> with and without unreachable() in WARN_ON().
> >>
> >> Did you try the two exemples described in commit 1e688dd2a3d6
> >> ("powerpc/bug: Provide better flexibility to WARN_ON/__WARN_FLAGS() with
> >> asm goto") ?
> >>
> >> Without your patch:
> >>
> >> 00000640 <test>:
> >>    640:	81 23 00 84 	lwz     r9,132(r3)
> >>    644:	71 29 40 00 	andi.   r9,r9,16384
> >>    648:	40 82 00 0c 	bne     654 <test+0x14>
> >>    64c:	80 63 00 0c 	lwz     r3,12(r3)
> >>    650:	4e 80 00 20 	blr
> >>    654:	0f e0 00 00 	twui    r0,0
> >>
> >> 00000658 <test9w>:
> >>    658:	2c 04 00 00 	cmpwi   r4,0
> >>    65c:	41 82 00 0c 	beq     668 <test9w+0x10>
> >>    660:	7c 63 23 96 	divwu   r3,r3,r4
> >>    664:	4e 80 00 20 	blr
> >>    668:	0f e0 00 00 	twui    r0,0
> >>    66c:	38 60 00 00 	li      r3,0
> >>    670:	4e 80 00 20 	blr
> > 
> > Per this construct you should do as x86 does and assume twui terminates
> > control flow and explicitly annotate the WARN case. That is, given the
> > fact that BUG as no instructions following it, you can't very well
> > annotate that.
> 
> That exactly the problem I guess. I'm fine with replacing the 
> unreachable() by __builtin_unreachable() with our __WARN_FLAGS() and 
> BUG() but we will still have a problem with some of the unrachable() 
> that are in core parts of the kernel.
> 
> Even the ones in arch/powerpc/, they are valid and should remain. The 
> point seems that the generic annotate_unreachable() is wrong for powerpc 
> as is, and activating CONFIG_OBJTOOL lead to bad code generation.

Right; I'm not against making that depend on yet-another OBJTOOL_$config
thing.

> By the way, for which functionnalities of objtool is that analysis 
> necessary ? I understand it is not necessary to mcount accounting, so 
> maybe the not empty annotate_unreachable() should be limited to those 
> those functionnalities ?

For all the things where it needs to follow control flow, so stack
validation, ORC generation, unreachable instruction validation etc..

I'd need to double check code-gen on x86, but the way
__builtin_unreachable() makes code-gen stop dead, I'm not entirely sure
it's right for us either.

> > Alternatively, you can teach objtool to look at __bug_table to
> > distinguish these cases.
> 
> Isn't it enough to tell objtool that execution never go past twui, using 
> INSN_BUG ?

That should work I suppose.

> By the way, for __WARN_FLAGS, we use the __extable for the continuation. 
> Is objtools able to follow __extable ?

Yes.