x86: ASM based __bug_table and rethunks

[RFC][PATCH 2/2] x86: Clean up default rethunk warning

Posted by Peter Zijlstra 1 month, 3 weeks ago

Replace the funny __warn_thunk thing with a more regular
WARN_ON_ONCE(), and simplify the ifdeffery.

Notably this avoids RET from having recursive RETs (once from the
thunk and once from the C function) -- recursive RET makes my head
hurt for no good reason.

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
---
 arch/x86/entry/entry.S               |    3 ---
 arch/x86/include/asm/nospec-branch.h |    2 --
 arch/x86/kernel/cpu/bugs.c           |    5 -----
 arch/x86/lib/retpoline.S             |   20 ++++++++++++--------
 4 files changed, 12 insertions(+), 18 deletions(-)

--- a/arch/x86/entry/entry.S
+++ b/arch/x86/entry/entry.S
@@ -10,8 +10,6 @@
 #include <asm/segment.h>
 #include <asm/cache.h>
 
-#include "calling.h"
-
 .pushsection .noinstr.text, "ax"
 
 SYM_FUNC_START(entry_ibpb)
@@ -45,4 +43,3 @@ EXPORT_SYMBOL_GPL(mds_verw_sel);
 
 .popsection
 
-THUNK warn_thunk_thunk, __warn_thunk
--- a/arch/x86/include/asm/nospec-branch.h
+++ b/arch/x86/include/asm/nospec-branch.h
@@ -387,8 +387,6 @@ extern void clear_bhb_loop(void);
 
 extern void (*x86_return_thunk)(void);
 
-extern void __warn_thunk(void);
-
 #ifdef CONFIG_MITIGATION_CALL_DEPTH_TRACKING
 extern void call_depth_return_thunk(void);
 
--- a/arch/x86/kernel/cpu/bugs.c
+++ b/arch/x86/kernel/cpu/bugs.c
@@ -3025,8 +3025,3 @@ ssize_t cpu_show_reg_file_data_sampling(
 	return cpu_show_common(dev, attr, buf, X86_BUG_RFDS);
 }
 #endif
-
-void __warn_thunk(void)
-{
-	WARN_ONCE(1, "Unpatched return thunk in use. This should not happen!\n");
-}
--- a/arch/x86/lib/retpoline.S
+++ b/arch/x86/lib/retpoline.S
@@ -12,9 +12,14 @@
 #include <asm/percpu.h>
 #include <asm/frame.h>
 #include <asm/nops.h>
+#include <asm/bug.h>
 
-	.section .text..__x86.indirect_thunk
+#define WARN_ONCE							\
+	1: ALTERNATIVE "", "ud2", X86_FEATURE_ALWAYS ;			\
+	ASM_BUGTABLE_FLAGS(1b, 0, 0, BUGFLAG_WARNING | BUGFLAG_ONCE) ;	\
+	REACHABLE
 
+	.section .text..__x86.indirect_thunk
 
 .macro POLINE reg
 	ANNOTATE_INTRA_FUNCTION_CALL
@@ -382,16 +387,15 @@ SYM_FUNC_END(call_depth_return_thunk)
 SYM_CODE_START(__x86_return_thunk)
 	UNWIND_HINT_FUNC
 	ANNOTATE_NOENDBR
-#if defined(CONFIG_MITIGATION_UNRET_ENTRY) || \
-    defined(CONFIG_MITIGATION_SRSO) || \
-    defined(CONFIG_MITIGATION_CALL_DEPTH_TRACKING)
-	ALTERNATIVE __stringify(ANNOTATE_UNRET_SAFE; ret), \
-		   "jmp warn_thunk_thunk", X86_FEATURE_ALWAYS
-#else
+
+#ifdef CONFIG_X86_64
+	WARN_ONCE
+#endif
+
 	ANNOTATE_UNRET_SAFE
 	ret
-#endif
 	int3
+
 SYM_CODE_END(__x86_return_thunk)
 EXPORT_SYMBOL(__x86_return_thunk)

Re: [RFC][PATCH 2/2] x86: Clean up default rethunk warning

Posted by Borislav Petkov 3 weeks, 2 days ago

On Mon, Oct 07, 2024 at 10:32:12AM +0200, Peter Zijlstra wrote:
> -	.section .text..__x86.indirect_thunk
> +#define WARN_ONCE							\

This should be in the asm section of arch/x86/include/asm/bug.h so that other
asm code can use it. It will come in handy...

> +	1: ALTERNATIVE "", "ud2", X86_FEATURE_ALWAYS ;			\

... but uff, you can't because of this ALTERNATIVE. This is a conditional
WARN_ONCE.  Yuck.

I guess ALT_WARN_ONCE or so...

> +	ASM_BUGTABLE_FLAGS(1b, 0, 0, BUGFLAG_WARNING | BUGFLAG_ONCE) ;	\
> +	REACHABLE
>  
> +	.section .text..__x86.indirect_thunk
>  
>  .macro POLINE reg
>  	ANNOTATE_INTRA_FUNCTION_CALL
> @@ -382,16 +387,15 @@ SYM_FUNC_END(call_depth_return_thunk)
>  SYM_CODE_START(__x86_return_thunk)
>  	UNWIND_HINT_FUNC
>  	ANNOTATE_NOENDBR
> -#if defined(CONFIG_MITIGATION_UNRET_ENTRY) || \
> -    defined(CONFIG_MITIGATION_SRSO) || \
> -    defined(CONFIG_MITIGATION_CALL_DEPTH_TRACKING)
> -	ALTERNATIVE __stringify(ANNOTATE_UNRET_SAFE; ret), \
> -		   "jmp warn_thunk_thunk", X86_FEATURE_ALWAYS
> -#else
> +
> +#ifdef CONFIG_X86_64
> +	WARN_ONCE
> +#endif

And you can add an empty 32-bit WARN_ONCE macro so that we don't have this
ifdeffery here where ifdeffery gives the last drop of making this file totally
unreadable...

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette

Re: [RFC][PATCH 2/2] x86: Clean up default rethunk warning

Posted by Peter Zijlstra 3 weeks, 2 days ago

On Mon, Nov 04, 2024 at 12:47:28PM +0100, Borislav Petkov wrote:
> On Mon, Oct 07, 2024 at 10:32:12AM +0200, Peter Zijlstra wrote:
> > -	.section .text..__x86.indirect_thunk
> > +#define WARN_ONCE							\
> 
> This should be in the asm section of arch/x86/include/asm/bug.h so that other
> asm code can use it. It will come in handy...
> 
> > +	1: ALTERNATIVE "", "ud2", X86_FEATURE_ALWAYS ;			\
> 
> ... but uff, you can't because of this ALTERNATIVE. This is a conditional
> WARN_ONCE.  Yuck.
> 
> I guess ALT_WARN_ONCE or so...

Yeah, Josh already said similar things.

> > +	ASM_BUGTABLE_FLAGS(1b, 0, 0, BUGFLAG_WARNING | BUGFLAG_ONCE) ;	\
> > +	REACHABLE
> >  
> > +	.section .text..__x86.indirect_thunk
> >  
> >  .macro POLINE reg
> >  	ANNOTATE_INTRA_FUNCTION_CALL
> > @@ -382,16 +387,15 @@ SYM_FUNC_END(call_depth_return_thunk)
> >  SYM_CODE_START(__x86_return_thunk)
> >  	UNWIND_HINT_FUNC
> >  	ANNOTATE_NOENDBR
> > -#if defined(CONFIG_MITIGATION_UNRET_ENTRY) || \
> > -    defined(CONFIG_MITIGATION_SRSO) || \
> > -    defined(CONFIG_MITIGATION_CALL_DEPTH_TRACKING)
> > -	ALTERNATIVE __stringify(ANNOTATE_UNRET_SAFE; ret), \
> > -		   "jmp warn_thunk_thunk", X86_FEATURE_ALWAYS
> > -#else
> > +
> > +#ifdef CONFIG_X86_64
> > +	WARN_ONCE
> > +#endif
> 
> And you can add an empty 32-bit WARN_ONCE macro so that we don't have this
> ifdeffery here where ifdeffery gives the last drop of making this file totally
> unreadable...

I just realized all the rethunk crap is 64bit only anyway. So it don't
matter.

But the reason I did this is that we never rewrite thunk calls on 32bit
(really, we should just strip all mitigation shit from it and leave it
to rot).

Re: [RFC][PATCH 2/2] x86: Clean up default rethunk warning

Posted by Borislav Petkov 3 weeks, 2 days ago

On Mon, Nov 04, 2024 at 03:29:53PM +0100, Peter Zijlstra wrote:
> I just realized all the rethunk crap is 64bit only anyway. So it don't
> matter.
> 
> But the reason I did this is that we never rewrite thunk calls on 32bit
> (really, we should just strip all mitigation shit from it and leave it
> to rot).

Perfectly fine with me.

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette

Re: [RFC][PATCH 2/2] x86: Clean up default rethunk warning

Posted by Josh Poimboeuf 1 month, 3 weeks ago

On Mon, Oct 07, 2024 at 10:32:12AM +0200, Peter Zijlstra wrote:
> Replace the funny __warn_thunk thing with a more regular
> WARN_ON_ONCE(), and simplify the ifdeffery.
> 
> Notably this avoids RET from having recursive RETs (once from the
> thunk and once from the C function) -- recursive RET makes my head
> hurt for no good reason.

This could use an explanation for why the ifdefs can be removed and why
the alternative can be removed.

> +#define WARN_ONCE							\
> +	1: ALTERNATIVE "", "ud2", X86_FEATURE_ALWAYS ;			\
> +	ASM_BUGTABLE_FLAGS(1b, 0, 0, BUGFLAG_WARNING | BUGFLAG_ONCE) ;	\
> +	REACHABLE

Can we not use __FILE__ and __LINE__ here?  Also why not put this in
asm/bug.h?

>  SYM_CODE_START(__x86_return_thunk)
>  	UNWIND_HINT_FUNC
>  	ANNOTATE_NOENDBR
> -#if defined(CONFIG_MITIGATION_UNRET_ENTRY) || \
> -    defined(CONFIG_MITIGATION_SRSO) || \
> -    defined(CONFIG_MITIGATION_CALL_DEPTH_TRACKING)
> -	ALTERNATIVE __stringify(ANNOTATE_UNRET_SAFE; ret), \
> -		   "jmp warn_thunk_thunk", X86_FEATURE_ALWAYS
> -#else
> +
> +#ifdef CONFIG_X86_64
> +	WARN_ONCE
> +#endif

Isn't this return thunk used before apply_returns()?  How does that not
trigger the warning?

-- 
Josh

Re: [RFC][PATCH 2/2] x86: Clean up default rethunk warning

Posted by Peter Zijlstra 1 month, 2 weeks ago

On Mon, Oct 07, 2024 at 10:33:45AM -0700, Josh Poimboeuf wrote:
> On Mon, Oct 07, 2024 at 10:32:12AM +0200, Peter Zijlstra wrote:
> > Replace the funny __warn_thunk thing with a more regular
> > WARN_ON_ONCE(), and simplify the ifdeffery.
> > 
> > Notably this avoids RET from having recursive RETs (once from the
> > thunk and once from the C function) -- recursive RET makes my head
> > hurt for no good reason.
> 
> This could use an explanation for why the ifdefs can be removed and why
> the alternative can be removed.

The alternative is in the WARN_ONCE now.

> > +#define WARN_ONCE							\
> > +	1: ALTERNATIVE "", "ud2", X86_FEATURE_ALWAYS ;			\
> > +	ASM_BUGTABLE_FLAGS(1b, 0, 0, BUGFLAG_WARNING | BUGFLAG_ONCE) ;	\
> > +	REACHABLE
> 
> Can we not use __FILE__ and __LINE__ here?  

Because for asm, __FILE__ is spelled like:

#ifdef CONFIG_DEBUG_BUGVERBOSE
.pushsection .rodata.str1.1, "aMS",@progbits,1
.LC0:
	.string __FILE__
.popsection
#endif
	1: ALTERNATIVE "", "ud2", X86_FEATURE_ALWAYS
	ASM_BUGTABLE_FLAGS(1b, LC0b, __LINE__, BUGFLAG_WARNING | BUGFLAG_ONCE)
	REACHABLE

And I didn't feel the whole thing was worth the trouble, if NULL bug
will only print the symbol name and that should be clear enough.

> Also why not put this in asm/bug.h?

Because the ALTERNATIVE..

> >  SYM_CODE_START(__x86_return_thunk)
> >  	UNWIND_HINT_FUNC
> >  	ANNOTATE_NOENDBR
> > -#if defined(CONFIG_MITIGATION_UNRET_ENTRY) || \
> > -    defined(CONFIG_MITIGATION_SRSO) || \
> > -    defined(CONFIG_MITIGATION_CALL_DEPTH_TRACKING)
> > -	ALTERNATIVE __stringify(ANNOTATE_UNRET_SAFE; ret), \
> > -		   "jmp warn_thunk_thunk", X86_FEATURE_ALWAYS
> > -#else
> > +
> > +#ifdef CONFIG_X86_64
> > +	WARN_ONCE
> > +#endif
> 
> Isn't this return thunk used before apply_returns()?  How does that not
> trigger the warning?

You missed the ALTERNATIVE I squirreled away in the WARN thing :-)

Re: [RFC][PATCH 2/2] x86: Clean up default rethunk warning

Posted by Josh Poimboeuf 1 month, 2 weeks ago

On Tue, Oct 08, 2024 at 09:25:02AM +0200, Peter Zijlstra wrote:
> On Mon, Oct 07, 2024 at 10:33:45AM -0700, Josh Poimboeuf wrote:
> > On Mon, Oct 07, 2024 at 10:32:12AM +0200, Peter Zijlstra wrote:
> > > Replace the funny __warn_thunk thing with a more regular
> > > WARN_ON_ONCE(), and simplify the ifdeffery.
> > > 
> > > Notably this avoids RET from having recursive RETs (once from the
> > > thunk and once from the C function) -- recursive RET makes my head
> > > hurt for no good reason.
> > 
> > This could use an explanation for why the ifdefs can be removed and why
> > the alternative can be removed.
> 
> The alternative is in the WARN_ONCE now.

Ah, sneaky... It should really be called WARN_ONCE_AFTER_ALTERNATIVES or
something.

-- 
Josh

Re: [RFC][PATCH 2/2] x86: Clean up default rethunk warning

Posted by Peter Zijlstra 1 month, 2 weeks ago

On Tue, Oct 08, 2024 at 09:45:14AM -0700, Josh Poimboeuf wrote:
> On Tue, Oct 08, 2024 at 09:25:02AM +0200, Peter Zijlstra wrote:
> > On Mon, Oct 07, 2024 at 10:33:45AM -0700, Josh Poimboeuf wrote:
> > > On Mon, Oct 07, 2024 at 10:32:12AM +0200, Peter Zijlstra wrote:
> > > > Replace the funny __warn_thunk thing with a more regular
> > > > WARN_ON_ONCE(), and simplify the ifdeffery.
> > > > 
> > > > Notably this avoids RET from having recursive RETs (once from the
> > > > thunk and once from the C function) -- recursive RET makes my head
> > > > hurt for no good reason.
> > > 
> > > This could use an explanation for why the ifdefs can be removed and why
> > > the alternative can be removed.
> > 
> > The alternative is in the WARN_ONCE now.
> 
> Ah, sneaky... It should really be called WARN_ONCE_AFTER_ALTERNATIVES or
> something.

Yeah, I suppose it should.