Chandan reported that trying to enable function_graph tracer on ppc64 BE
now locks up the system. This is due to prepare_ftrace_return() using
ppc_function_entry() for resolving return_to_handler(), which in turn
invokes kernel_text_address(), which also gets traced resulting in a
loop.
We added a check for kernel_text_address() in ppc_function_entry() to
guard all users in case we were called with a function, rather than a
function descriptor. In hindsight, I feel that this is inefficient since
we usually only pass function descriptors to ppc_function_entry() (and
ppc_global_function_entry()). So, I am proposing that we revert the
previous patch and instead implement the necessary checks in the kprobes
subsystem.
The other way to fix this is to simply guard the call to
kernel_text_address() within [un]pause_graph_tracing(), if you think
it's useful to have the check in ppc_function_entry() for all users.
- Naveen
Naveen N. Rao (2):
Revert "powerpc64/elfv1: Only dereference function descriptor for
non-text symbols"
powerpc/kprobes: Dereference function pointers only if the address
does not belong to kernel text
arch/powerpc/include/asm/code-patching.h | 10 +---------
arch/powerpc/kernel/kprobes.c | 7 ++++++-
2 files changed, 7 insertions(+), 10 deletions(-)
--
2.14.2
