| Message ID | 1436765375-7119-2-git-send-email-takahiro.akashi@linaro.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Mon, 13 Jul 2015 14:29:33 +0900 AKASHI Takahiro <takahiro.akashi@linaro.org> wrote: > Ftace's stack tracer on arm64 returns wrong information about call stacks: > > Depth Size Location (50 entries) > ----- ---- -------- > 0) 5256 0 notifier_call_chain+0x30/0x94 > 1) 5256 0 ftrace_call+0x0/0x4 > 2) 5256 0 notifier_call_chain+0x2c/0x94 > 3) 5256 0 raw_notifier_call_chain+0x34/0x44 > 4) 5256 0 timekeeping_update.constprop.9+0xb8/0x114 > 5) 5256 0 update_wall_time+0x408/0x6dc > > Most of 'Size' fields are unexpectedly zero. > > This is because stack tracer fails to recognize each function's stack frame > in check_stack(). Stack tracer searches for a function's pc in the stack > based on the list returned by save_stack_trace(), but save_stack_trace() on > arm64 does not return the exact return address saved in a stack frame, but > a value decrmented by 4 (which means a branch instruction's address). > This behavior was introduced by > commit e306dfd06fcb ("ARM64: unwind: Fix PC calculation") > > So the matching doesn't succeed in most cases. > > This problem can be fixed either by > a) reverting the commit above > b) adding an arm64-specific hack to check_patch() > > This patch does b). > > Signed-off-by: AKASHI Takahiro <takahiro.akashi@linaro.org> > --- > kernel/trace/trace_stack.c | 4 ++++ > 1 file changed, 4 insertions(+) > > diff --git a/kernel/trace/trace_stack.c b/kernel/trace/trace_stack.c > index 3f34496..7086fc3 100644 > --- a/kernel/trace/trace_stack.c > +++ b/kernel/trace/trace_stack.c > @@ -143,7 +143,11 @@ check_stack(unsigned long ip, unsigned long *stack) > p = start; > > for (; p < top && i < max_stack_trace.nr_entries; p++) { > +#ifdef CONFIG_ARM64 > + if (*p == (stack_dump_trace[i] + 4)) { > +#else > if (*p == stack_dump_trace[i]) { > +#endif Instead of the ugly #ifdef in this code, please add a macro FTRACE_STACK_FRAME_OFFSET Then in include/linux/ftrace.h have: #ifndef FTRACE_STACK_FRAME_OFFSET # define FTRACE_STACK_FRAME_OFFSET 0 #endif And in arch/arm64/include/asm/ftrace.h #define FTRACE_STACK_FRAME_OFFSET 4 And then just do: if (*p == (stack_dump_trace[i] + FTRACE_STACK_FRAME_OFFSET)) { -- Steve > this_size = stack_dump_index[i++] = > (top - p) * sizeof(unsigned long); > found = 1;
On 07/14/2015 12:24 AM, Steven Rostedt wrote: > On Mon, 13 Jul 2015 14:29:33 +0900 > AKASHI Takahiro <takahiro.akashi@linaro.org> wrote: > >> Ftace's stack tracer on arm64 returns wrong information about call stacks: >> >> Depth Size Location (50 entries) >> ----- ---- -------- >> 0) 5256 0 notifier_call_chain+0x30/0x94 >> 1) 5256 0 ftrace_call+0x0/0x4 >> 2) 5256 0 notifier_call_chain+0x2c/0x94 >> 3) 5256 0 raw_notifier_call_chain+0x34/0x44 >> 4) 5256 0 timekeeping_update.constprop.9+0xb8/0x114 >> 5) 5256 0 update_wall_time+0x408/0x6dc >> >> Most of 'Size' fields are unexpectedly zero. >> >> This is because stack tracer fails to recognize each function's stack frame >> in check_stack(). Stack tracer searches for a function's pc in the stack >> based on the list returned by save_stack_trace(), but save_stack_trace() on >> arm64 does not return the exact return address saved in a stack frame, but >> a value decrmented by 4 (which means a branch instruction's address). >> This behavior was introduced by >> commit e306dfd06fcb ("ARM64: unwind: Fix PC calculation") >> >> So the matching doesn't succeed in most cases. >> >> This problem can be fixed either by >> a) reverting the commit above >> b) adding an arm64-specific hack to check_patch() >> >> This patch does b). >> >> Signed-off-by: AKASHI Takahiro <takahiro.akashi@linaro.org> >> --- >> kernel/trace/trace_stack.c | 4 ++++ >> 1 file changed, 4 insertions(+) >> >> diff --git a/kernel/trace/trace_stack.c b/kernel/trace/trace_stack.c >> index 3f34496..7086fc3 100644 >> --- a/kernel/trace/trace_stack.c >> +++ b/kernel/trace/trace_stack.c >> @@ -143,7 +143,11 @@ check_stack(unsigned long ip, unsigned long *stack) >> p = start; >> >> for (; p < top && i < max_stack_trace.nr_entries; p++) { >> +#ifdef CONFIG_ARM64 >> + if (*p == (stack_dump_trace[i] + 4)) { >> +#else >> if (*p == stack_dump_trace[i]) { >> +#endif > > Instead of the ugly #ifdef in this code, please add a macro > FTRACE_STACK_FRAME_OFFSET > > Then in include/linux/ftrace.h have: > > #ifndef FTRACE_STACK_FRAME_OFFSET > # define FTRACE_STACK_FRAME_OFFSET 0 > #endif > > And in arch/arm64/include/asm/ftrace.h > > #define FTRACE_STACK_FRAME_OFFSET 4 > > And then just do: > > if (*p == (stack_dump_trace[i] + FTRACE_STACK_FRAME_OFFSET)) { Yes. Thanks, -Takahiro AKASHI > -- Steve > >> this_size = stack_dump_index[i++] = >> (top - p) * sizeof(unsigned long); >> found = 1; >
diff --git a/kernel/trace/trace_stack.c b/kernel/trace/trace_stack.c index 3f34496..7086fc3 100644 --- a/kernel/trace/trace_stack.c +++ b/kernel/trace/trace_stack.c @@ -143,7 +143,11 @@ check_stack(unsigned long ip, unsigned long *stack) p = start; for (; p < top && i < max_stack_trace.nr_entries; p++) { +#ifdef CONFIG_ARM64 + if (*p == (stack_dump_trace[i] + 4)) { +#else if (*p == stack_dump_trace[i]) { +#endif this_size = stack_dump_index[i++] = (top - p) * sizeof(unsigned long); found = 1;
Ftace's stack tracer on arm64 returns wrong information about call stacks: Depth Size Location (50 entries) ----- ---- -------- 0) 5256 0 notifier_call_chain+0x30/0x94 1) 5256 0 ftrace_call+0x0/0x4 2) 5256 0 notifier_call_chain+0x2c/0x94 3) 5256 0 raw_notifier_call_chain+0x34/0x44 4) 5256 0 timekeeping_update.constprop.9+0xb8/0x114 5) 5256 0 update_wall_time+0x408/0x6dc Most of 'Size' fields are unexpectedly zero. This is because stack tracer fails to recognize each function's stack frame in check_stack(). Stack tracer searches for a function's pc in the stack based on the list returned by save_stack_trace(), but save_stack_trace() on arm64 does not return the exact return address saved in a stack frame, but a value decrmented by 4 (which means a branch instruction's address). This behavior was introduced by commit e306dfd06fcb ("ARM64: unwind: Fix PC calculation") So the matching doesn't succeed in most cases. This problem can be fixed either by a) reverting the commit above b) adding an arm64-specific hack to check_patch() This patch does b). Signed-off-by: AKASHI Takahiro <takahiro.akashi@linaro.org> --- kernel/trace/trace_stack.c | 4 ++++ 1 file changed, 4 insertions(+)