Re: [PATCH 13/13] powerpc/rtas: place tracepoints in do_enter_rtas()

Nathan Lynch [off-list ref] writes:

Michael Ellerman [off-list ref] writes:

quoted

Nathan Lynch [off-list ref] writes:

quoted

"Nicholas Piggin" [off-list ref] writes:

quoted

On Sat Nov 19, 2022 at 1:07 AM AEST, Nathan Lynch wrote:

quoted

Call the just-added rtas tracepoints in do_enter_rtas(), taking care
to avoid function name lookups in the CPU offline path.

Signed-off-by: Nathan Lynch <redacted>
---
 arch/powerpc/kernel/rtas.c | 23 +++++++++++++++++++++++
 1 file changed, 23 insertions(+)

diff --git a/arch/powerpc/kernel/rtas.c b/arch/powerpc/kernel/rtas.c
index 198366d641d0..3487b42cfbf7 100644
--- a/arch/powerpc/kernel/rtas.c
+++ b/arch/powerpc/kernel/rtas.c

@@ -38,6 +38,7 @@
 #include <asm/page.h>
 #include <asm/rtas.h>
 #include <asm/time.h>
+#include <asm/trace.h>
 #include <asm/udbg.h>
 
 enum rtas_function_flags {

@@ -525,6 +526,7 @@ void enter_rtas(unsigned long);
 static void do_enter_rtas(struct rtas_args *args)
 {
 	unsigned long msr;
+	const char *name = NULL;
 
 	/*
 	 * Make sure MSR[RI] is currently enabled as it will be forced later

@@ -537,9 +539,30 @@ static void do_enter_rtas(struct rtas_args *args)
 
 	hard_irq_disable(); /* Ensure MSR[EE] is disabled on PPC64 */
 
+	if ((trace_rtas_input_enabled() || trace_rtas_output_enabled())) {
+		/*
+		 * rtas_token_to_function() uses xarray which uses RCU,
+		 * but this code can run in the CPU offline path
+		 * (e.g. stop-self), after it's become invalid to call
+		 * RCU APIs.
+		 */

We can call this in real-mode via pseries_machine_check_realmode
-> fwnmi_release_errinfo, so tracing should be disabled for that
case too... Does this_cpu_set_ftrace_enabled(0) in the early
machine check handler cover that sufficiently?

I suspect so, but I'd like to verify. Do you know how I could exercise
this path in qemu or LPAR?

On a P9 or P10 LPAR you should be able to use tools/testing/selftests/powerpc/mce/inject-ra-err

Nice. Looks like I was too optimistic. From a P10 LPAR:

# trace-cmd record -T -e powerpc:rtas_input -- \
  sh -c 'sleep 10; ./inject-ra-err' && trace-cmd report
     kworker/7:1-73    [007]    72.882159: rtas_input:           event-scan arguments: 4294967295 0 80419368 2048
     kworker/7:1-73    [007]    72.882165: kernel_stack:         <stack trace >
=> do_enter_rtas (c000000000045180)
=> rtas_call (c000000000045da8)
=> rtas_event_scan (c000000000049458)
=> process_one_work (c0000000001c7618)
=> worker_thread (c0000000001c7bd8)
=> kthread (c0000000001d6858)
=> ret_from_kernel_thread (c00000000000cf5c)
   inject-ra-err-1080  [001]    78.386947: rtas_input:           ibm,nmi-interlock arguments: 
   inject-ra-err-1080  [001]    78.386950: kernel_stack:         <stack trace >
=> do_enter_rtas (c000000000045180)
=> rtas_call_unlocked (c000000000046ff4)
=> pseries_machine_check_realmode (c0000000000e8db8)
=> machine_check_early (c0000000000400c4)
=> machine_check_early_common (c00000000000836c)

So... that's bad. (right?)

Potentially. If you booted in hash mode it might crash, if the tracing
causes accesses outside the RMO, or to vmalloc etc.

I guess this patch needs something like this?

Yeah I think that's probably the easiest solution.

You could split do_enter_rtas() into a tracing and non-tracing version,
but I don't think that would be any cleaner.

cheers