diff options
author | Petri Gynther <pgynther@google.com> | 2014-07-23 22:55:02 -0700 |
---|---|---|
committer | Ralf Baechle <ralf@linux-mips.org> | 2014-07-30 20:37:42 +0200 |
commit | dce0e7d54a710400c0056d86d0f0ed972133b25d (patch) | |
tree | 09c4e4a9ff99b1d11150e62c21f519264513254c /crypto | |
parent | b1442d39fac2fcfbe6a4814979020e993ca59c9e (diff) | |
download | linux-dce0e7d54a710400c0056d86d0f0ed972133b25d.tar.bz2 |
MIPS: ftrace: Fix dynamic tracing of kernel modules
Dynamic tracing of kernel modules is broken on 32-bit MIPS. When modules
are loaded, the kernel crashes when dynamic tracing is enabled with:
cd /sys/kernel/debug/tracing
echo > set_ftrace_filter
echo function > current_tracer
1) arch/mips/kernel/ftrace.c
When the kernel boots, or when a module is initialized, ftrace_make_nop()
modifies every _mcount call site to eliminate the ftrace overhead.
However, when ftrace is later enabled for a call site, ftrace_make_call()
does not currently restore the _mcount call correctly for module call sites.
Added ftrace_modify_code_2r() and modified ftrace_make_call() to fix this.
2) arch/mips/kernel/mcount.S
_mcount assembly routine is supposed to have the caller's _mcount call site
address in register a0. However, a0 is currently not calculated correctly for
module call sites. a0 should be (ra - 20) or (ra - 24), depending on whether
the kernel was built with KBUILD_MCOUNT_RA_ADDRESS or not.
This fix has been tested on Broadcom BMIPS5000 processor. Dynamic tracing
now works for both built-in functions and module functions.
Signed-off-by: Petri Gynther <pgynther@google.com>
Cc: linux-mips@linux-mips.org
Cc: rostedt@goodmis.org
Cc: alcooperx@gmail.com
Cc: cminyard@mvista.com
Patchwork: https://patchwork.linux-mips.org/patch/7476/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Diffstat (limited to 'crypto')
0 files changed, 0 insertions, 0 deletions