On Wed, Mar 22, 2023 at 12:22:07PM +0000, Mark Rutland wrote:

quoted

The problem is that the __perf_guest_state() static call does an
indirect branch to __static_call_return0(), which isn't CFI-compliant.

IIUC that'd be broken even with the old CFI mechanism, since commit:

  87b940a0675e2526 ("perf/core: Use static_call to optimize perf_guest_info_callbacks")

If so, we probably want a Fixes tag?

Yeah, it should definitely get a Fixes tag.  I wasn't quite sure if this
bug started with the above commit or with the CFI_CLANG switch to kcfi.
And then I forgot to investigate.

quoted

+/* Generate a CFI-compliant static call NOP function */
+#define __ARCH_DEFINE_STATIC_CALL_CFI(name, insns)			\
+	asm(".align 4						\n"	\
+	    ".word __kcfi_typeid_" name "			\n"	\
+	    ".globl " name "					\n"	\
+	    name ":						\n"	\
+	    "bti c						\n"	\
+	    insns "						\n"	\
+	    "ret						\n"	\
+	    ".type " name ", @function				\n"	\
+	    ".size " name ", . - " name "			\n")
+
+#define __ARCH_DEFINE_STATIC_CALL_RET0_CFI(name)			\
+	GEN_CFI_SYM(STATIC_CALL_RET0_CFI(name));			\
+	__ARCH_DEFINE_STATIC_CALL_CFI(STATIC_CALL_RET0_CFI_STR(name), "mov x0, xzr")

This looks correct, but given we're generating a regular functions it's
unfortunate we can't have the compiler generate the actual code with something
like:

#define __ARCH_DEFINE_STATIC_CALL_RET0_CFI(rettype, name, args...)	\
rettype name(args)							\
{									\
	return (rettype)0;						\
}

... but I guess passing the rettype and args around is painful.

Hm, I hadn't considered that.  I'll play around with it.

Regardless, I gave this a spin atop v6.3-rc3 using LLVM 16.0.0 and CFI_CLANG,
and it does seem to work, so:

Tested-by: Mark Rutland <mark.rutland@arm.com>

Thanks!

-- 
Josh

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel