On 2023/3/17 7:21, Russell King (Oracle) wrote:

On Wed, Mar 15, 2023 at 06:39:10AM -0700, Donglin Peng wrote:

quoted

diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
index e24a9820e12f..ad03fc868f34 100644
--- a/arch/arm/Kconfig
+++ b/arch/arm/Kconfig

@@ -99,6 +99,7 @@ config ARM
  	select HAVE_FTRACE_MCOUNT_RECORD if !XIP_KERNEL
  	select HAVE_FUNCTION_ERROR_INJECTION
  	select HAVE_FUNCTION_GRAPH_TRACER
+	select HAVE_FUNCTION_GRAPH_RETVAL if HAVE_FUNCTION_GRAPH_TRACER
  	select HAVE_FUNCTION_TRACER if !XIP_KERNEL
  	select HAVE_GCC_PLUGINS
  	select HAVE_HW_BREAKPOINT if PERF_EVENTS && (CPU_V6 || CPU_V6K || CPU_V7)

diff --git a/arch/arm/kernel/entry-ftrace.S b/arch/arm/kernel/entry-ftrace.S
index 3e7bcaca5e07..0151d2ce9958 100644
--- a/arch/arm/kernel/entry-ftrace.S
+++ b/arch/arm/kernel/entry-ftrace.S

@@ -258,6 +258,10 @@ ENDPROC(ftrace_graph_regs_caller)
  #ifdef CONFIG_FUNCTION_GRAPH_TRACER
  ENTRY(return_to_handler)
  	stmdb	sp!, {r0-r3}
+#ifdef CONFIG_FUNCTION_GRAPH_RETVAL
+	/* Pass the function return value to ftrace_return_to_handler */
+	mov	r1, r0

In a similar vein to Peter's comment, do we care about 64-bit return
values here, because the above only covers 32-bit values.

If we do care about 64-bit values, then we get into EABI/OABI
stickyness, because on EABI the 64-bit value would have to be passed
in r2,r3, and OABI would need r1,r2.

it would be better to have the 64-bit argument as the first argument
to ftrace_return_to_handler() which would eliminate that variability,
but I don't know what effect that would have for other architectures.

Things get more icky if we want 128-bit values. For EABI, we've
conveniently just stacked that. For OABI, that would need to be in
r1-r3 and the final high bits on the stack.

With a 128-bit argument as the first, that would be r0-r3 with the
existing stack pointer argument stored... on the stack.

So, really it depends what size of return value we want to report.
Also, please bear in mind that where a function returns a 32-bit
value, that will be in r0, and r1 will be whatever happened to be
in it at function exit - there's no defined value for r1.

Thank you. I will document this as a limitation of fgraph return value.
It can just cover most cases at present and I think the r0 is enough.