Fail the build on objtool warnings

[PATCH 06/13] objtool: Remove --unret dependency on --rethunk

Posted by Josh Poimboeuf 9 months, 1 week ago

With unret validation enabled and IBT/LTO disabled, objtool runs on TUs
with --rethunk and on vmlinux.o with --unret.  So this dependency isn't
valid as they don't always run on the same object.

This error never triggered before because --unret is always coupled with
--noinstr, so the first conditional in opts_valid() returns early due to
opts.noinstr being true.

Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
---
 tools/objtool/builtin-check.c | 5 -----
 1 file changed, 5 deletions(-)

diff --git a/tools/objtool/builtin-check.c b/tools/objtool/builtin-check.c
index 387d56a7f5fb..c7275cf7641b 100644
--- a/tools/objtool/builtin-check.c
+++ b/tools/objtool/builtin-check.c
@@ -151,11 +151,6 @@ static bool opts_valid(void)
 		return true;
 	}
 
-	if (opts.unret && !opts.rethunk) {
-		ERROR("--unret requires --rethunk");
-		return false;
-	}
-
 	if (opts.dump_orc)
 		return true;
 
-- 
2.48.1

Re: [PATCH 06/13] objtool: Remove --unret dependency on --rethunk

Posted by Peter Zijlstra 9 months, 1 week ago

On Fri, Mar 14, 2025 at 12:29:04PM -0700, Josh Poimboeuf wrote:
> With unret validation enabled and IBT/LTO disabled, objtool runs on TUs
> with --rethunk and on vmlinux.o with --unret.  So this dependency isn't
> valid as they don't always run on the same object.
> 
> This error never triggered before because --unret is always coupled with
> --noinstr, so the first conditional in opts_valid() returns early due to
> opts.noinstr being true.
> 
> Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
> ---
>  tools/objtool/builtin-check.c | 5 -----
>  1 file changed, 5 deletions(-)
> 
> diff --git a/tools/objtool/builtin-check.c b/tools/objtool/builtin-check.c
> index 387d56a7f5fb..c7275cf7641b 100644
> --- a/tools/objtool/builtin-check.c
> +++ b/tools/objtool/builtin-check.c
> @@ -151,11 +151,6 @@ static bool opts_valid(void)
>  		return true;
>  	}
>  
> -	if (opts.unret && !opts.rethunk) {
> -		ERROR("--unret requires --rethunk");
> -		return false;
> -	}

But but but, the whole UNTRAIN_RET stuff relies on return thunks. Even
though objtool can do that VALUDATE_UNRET_BEGIN -> VALIDATE_UNRET_END
check without there being return thunks, it simply doesn't make sense.

Re: [PATCH 06/13] objtool: Remove --unret dependency on --rethunk

Posted by Josh Poimboeuf 9 months, 1 week ago

On Fri, Mar 14, 2025 at 08:38:48PM +0100, Peter Zijlstra wrote:
> On Fri, Mar 14, 2025 at 12:29:04PM -0700, Josh Poimboeuf wrote:
> > With unret validation enabled and IBT/LTO disabled, objtool runs on TUs
> > with --rethunk and on vmlinux.o with --unret.  So this dependency isn't
> > valid as they don't always run on the same object.
> > 
> > This error never triggered before because --unret is always coupled with
> > --noinstr, so the first conditional in opts_valid() returns early due to
> > opts.noinstr being true.
> > 
> > Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
> > ---
> >  tools/objtool/builtin-check.c | 5 -----
> >  1 file changed, 5 deletions(-)
> > 
> > diff --git a/tools/objtool/builtin-check.c b/tools/objtool/builtin-check.c
> > index 387d56a7f5fb..c7275cf7641b 100644
> > --- a/tools/objtool/builtin-check.c
> > +++ b/tools/objtool/builtin-check.c
> > @@ -151,11 +151,6 @@ static bool opts_valid(void)
> >  		return true;
> >  	}
> >  
> > -	if (opts.unret && !opts.rethunk) {
> > -		ERROR("--unret requires --rethunk");
> > -		return false;
> > -	}
> 
> But but but, the whole UNTRAIN_RET stuff relies on return thunks. Even
> though objtool can do that VALUDATE_UNRET_BEGIN -> VALIDATE_UNRET_END
> check without there being return thunks, it simply doesn't make sense.

That's at least enforced by kconfig dependencies as
MITIGATION_UNRET_ENTRY depends on MITIGATION_RETHUNK.

If you wanted to enforce it on the objtool level, we might need to do
some makefile acrobatics to defer --rethunk to vmlinux link time for the
above mentioned case.

-- 
Josh

Re: [PATCH 06/13] objtool: Remove --unret dependency on --rethunk

Posted by Miroslav Benes 9 months ago

Hi,

On Fri, 14 Mar 2025, Josh Poimboeuf wrote:

> On Fri, Mar 14, 2025 at 08:38:48PM +0100, Peter Zijlstra wrote:
> > On Fri, Mar 14, 2025 at 12:29:04PM -0700, Josh Poimboeuf wrote:
> > > With unret validation enabled and IBT/LTO disabled, objtool runs on TUs
> > > with --rethunk and on vmlinux.o with --unret.  So this dependency isn't
> > > valid as they don't always run on the same object.
> > > 
> > > This error never triggered before because --unret is always coupled with
> > > --noinstr, so the first conditional in opts_valid() returns early due to
> > > opts.noinstr being true.
> > > 
> > > Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
> > > ---
> > >  tools/objtool/builtin-check.c | 5 -----
> > >  1 file changed, 5 deletions(-)
> > > 
> > > diff --git a/tools/objtool/builtin-check.c b/tools/objtool/builtin-check.c
> > > index 387d56a7f5fb..c7275cf7641b 100644
> > > --- a/tools/objtool/builtin-check.c
> > > +++ b/tools/objtool/builtin-check.c
> > > @@ -151,11 +151,6 @@ static bool opts_valid(void)
> > >  		return true;
> > >  	}
> > >  
> > > -	if (opts.unret && !opts.rethunk) {
> > > -		ERROR("--unret requires --rethunk");
> > > -		return false;
> > > -	}
> > 
> > But but but, the whole UNTRAIN_RET stuff relies on return thunks. Even
> > though objtool can do that VALUDATE_UNRET_BEGIN -> VALIDATE_UNRET_END
> > check without there being return thunks, it simply doesn't make sense.
> 
> That's at least enforced by kconfig dependencies as
> MITIGATION_UNRET_ENTRY depends on MITIGATION_RETHUNK.

could you add it to the changelog, please? My future self would definitely 
welcome it.

Miroslav

Re: [PATCH 06/13] objtool: Remove --unret dependency on --rethunk

Posted by Peter Zijlstra 9 months, 1 week ago

On Fri, Mar 14, 2025 at 12:52:32PM -0700, Josh Poimboeuf wrote:
> On Fri, Mar 14, 2025 at 08:38:48PM +0100, Peter Zijlstra wrote:
> > On Fri, Mar 14, 2025 at 12:29:04PM -0700, Josh Poimboeuf wrote:
> > > With unret validation enabled and IBT/LTO disabled, objtool runs on TUs
> > > with --rethunk and on vmlinux.o with --unret.  So this dependency isn't
> > > valid as they don't always run on the same object.
> > > 
> > > This error never triggered before because --unret is always coupled with
> > > --noinstr, so the first conditional in opts_valid() returns early due to
> > > opts.noinstr being true.
> > > 
> > > Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
> > > ---
> > >  tools/objtool/builtin-check.c | 5 -----
> > >  1 file changed, 5 deletions(-)
> > > 
> > > diff --git a/tools/objtool/builtin-check.c b/tools/objtool/builtin-check.c
> > > index 387d56a7f5fb..c7275cf7641b 100644
> > > --- a/tools/objtool/builtin-check.c
> > > +++ b/tools/objtool/builtin-check.c
> > > @@ -151,11 +151,6 @@ static bool opts_valid(void)
> > >  		return true;
> > >  	}
> > >  
> > > -	if (opts.unret && !opts.rethunk) {
> > > -		ERROR("--unret requires --rethunk");
> > > -		return false;
> > > -	}
> > 
> > But but but, the whole UNTRAIN_RET stuff relies on return thunks. Even
> > though objtool can do that VALUDATE_UNRET_BEGIN -> VALIDATE_UNRET_END
> > check without there being return thunks, it simply doesn't make sense.
> 
> That's at least enforced by kconfig dependencies as
> MITIGATION_UNRET_ENTRY depends on MITIGATION_RETHUNK.
> 
> If you wanted to enforce it on the objtool level, we might need to do
> some makefile acrobatics to defer --rethunk to vmlinux link time for the
> above mentioned case.

Nah, I suppose it should all work like this. Its just weird seeing this.