ARM: ftrace: use ADD not POP to counter PUSH at entry

The compiler emitted hook used for ftrace consists of a PUSH {LR} to
preserve the link register, followed by a branch-and-link (BL) to
__gnu_mount_nc. Dynamic ftrace patches away the latter to turn the
combined sequence into a NOP, using a POP {LR} instruction.

This is not necessary, since the link register does not get clobbered in
this case, and simply adding #4 to the stack pointer is sufficient, and
avoids a memory access that may take a few cycles to resolve depending
on the micro-architecture.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Nick Desaulniers <ndesaulniers@google.com>
Reviewed-by: Linus Walleij <linus.walleij@linaro.org>
Reviewed-by: Steven Rostedt (Google) <rostedt@goodmis.org>

1 files changed, 13 insertions, 2 deletions

diff --git a/arch/arm/kernel/ftrace.c b/arch/arm/kernel/ftrace.c
index a006585e1c09..811cadf7eefc 100644
--- a/arch/arm/kernel/ftrace.c
+++ b/arch/arm/kernel/ftrace.c
@@ -24,10 +24,21 @@
 #include <asm/set_memory.h>
 #include <asm/patch.h>
 
+/*
+ * The compiler emitted profiling hook consists of
+ *
+ *   PUSH    {LR}
+ *   BL	     __gnu_mcount_nc
+ *
+ * To turn this combined sequence into a NOP, we need to restore the value of
+ * SP before the PUSH. Let's use an ADD rather than a POP into LR, as LR is not
+ * modified anyway, and reloading LR from memory is highly likely to be less
+ * efficient.
+ */
 #ifdef CONFIG_THUMB2_KERNEL
-#define	NOP		0xf85deb04	/* pop.w {lr} */
+#define	NOP		0xf10d0d04	/* add.w sp, sp, #4 */
 #else
-#define	NOP		0xe8bd4000	/* pop {lr} */
+#define	NOP		0xe28dd004	/* add   sp, sp, #4 */
 #endif
 
 #ifdef CONFIG_DYNAMIC_FTRACE