Message ID | 20090131190650.GA4284@oksana.dev.rtsoft.ru (mailing list archive) |
---|---|
State | Not Applicable, archived |
Headers | show |
On Sat, 2009-01-31 at 22:06 +0300, Anton Vorontsov wrote: > According to this discussion: > > http://lkml.org/lkml/2008/7/25/338 > http://lkml.org/lkml/2008/7/26/72 > > Frame pointers do nothing useful on PowerPC, so lib/Kconfig.debug > makes CONFIG_FRAME_POINTER unselectable on PPC targets. But ftrace.h > requires CONFIG_FRAME_POINTER for CALLER_ADDR macros. Therefore > tracing is completely useless on PowerPC: But we need them for -pg which ftrace uses no ? (ie, gcc forces you to have -fno-omit-frame-pointers with -fpg iirc). Now, regardless, I agree that on PPC, __builtin_return_address() should always work with our without that cruft, so we may as well apply that patch... > [...] > <idle>-0 0X.h3 2us+: 0:140:R + [000] 1733:120:S mvtsd > <idle>-0 0X.h3 9us+: 0 (0) > <idle>-0 0X..3 72us : 0 (0) > <idle>-0 0X..3 73us : 0:140:R ==> [000] 1733:120:R mvtsd > > On PPC we can safely use __builtin_return_address(1..6) w/o frame > pointers, and with this patch the trace output looks OK: > > [...] > <idle>-0 0X.h3 2us+: 0:140:R + [000] 1740:120:S mvtsd > <idle>-0 0X.h3 9us+: hrtimer_wakeup (__run_hrtimer) > <idle>-0 0X..3 87us : cpu_idle (__got2_end) > <idle>-0 0X..3 89us : 0:140:R ==> [000] 1740:120:R mvtsd > > Signed-off-by: Anton Vorontsov <avorontsov@ru.mvista.com> > --- > > Btw, on PowerPC tracing is also broken w/o "ring-buffer: fix alignment > problem" patch (currently collecting dust in the -tip tree, commit > 082605de5f82eb692cc90f7fda071cc01bb5ac34). Any chance the fix go into > Linus' tree, to not waste other people's time bisecting and debugging > the problem? ;-) > > For google: tracing, regression, "ring-buffer: move some metadata > into buffer page", commit abc9b56d66fbd4d93302ef4bf6fa726e1b8255f9, > answer is here. > > include/linux/ftrace.h | 2 +- > 1 files changed, 1 insertions(+), 1 deletions(-) > > diff --git a/include/linux/ftrace.h b/include/linux/ftrace.h > index 9c5bc6b..13eba02 100644 > --- a/include/linux/ftrace.h > +++ b/include/linux/ftrace.h > @@ -146,7 +146,7 @@ static inline void __ftrace_enabled_restore(int enabled) > #endif > } > > -#ifdef CONFIG_FRAME_POINTER > +#if defined(CONFIG_FRAME_POINTER) || defined(CONFIG_PPC) > /* TODO: need to fix this for ARM */ > # define CALLER_ADDR0 ((unsigned long)__builtin_return_address(0)) > # define CALLER_ADDR1 ((unsigned long)__builtin_return_address(1))
On Mon, 2009-02-02 at 11:35 +1100, Benjamin Herrenschmidt wrote: > On Sat, 2009-01-31 at 22:06 +0300, Anton Vorontsov wrote: > > According to this discussion: > > > > http://lkml.org/lkml/2008/7/25/338 > > http://lkml.org/lkml/2008/7/26/72 > > > > Frame pointers do nothing useful on PowerPC, so lib/Kconfig.debug > > makes CONFIG_FRAME_POINTER unselectable on PPC targets. But ftrace.h > > requires CONFIG_FRAME_POINTER for CALLER_ADDR macros. Therefore > > tracing is completely useless on PowerPC: > > But we need them for -pg which ftrace uses no ? (ie, gcc forces you to > have -fno-omit-frame-pointers with -fpg iirc). Yes, ftrace function tracing requires -pg which requires FRAME_POINTER turned on. > > Now, regardless, I agree that on PPC, __builtin_return_address() should > always work with our without that cruft, so we may as well apply that > patch... > > > [...] > > <idle>-0 0X.h3 2us+: 0:140:R + [000] 1733:120:S mvtsd > > <idle>-0 0X.h3 9us+: 0 (0) > > <idle>-0 0X..3 72us : 0 (0) > > <idle>-0 0X..3 73us : 0:140:R ==> [000] 1733:120:R mvtsd > > > > On PPC we can safely use __builtin_return_address(1..6) w/o frame > > pointers, and with this patch the trace output looks OK: > > > > [...] > > <idle>-0 0X.h3 2us+: 0:140:R + [000] 1740:120:S mvtsd > > <idle>-0 0X.h3 9us+: hrtimer_wakeup (__run_hrtimer) > > <idle>-0 0X..3 87us : cpu_idle (__got2_end) > > <idle>-0 0X..3 89us : 0:140:R ==> [000] 1740:120:R mvtsd > > > > Signed-off-by: Anton Vorontsov <avorontsov@ru.mvista.com> > > --- > > > > Btw, on PowerPC tracing is also broken w/o "ring-buffer: fix alignment > > problem" patch (currently collecting dust in the -tip tree, commit > > 082605de5f82eb692cc90f7fda071cc01bb5ac34). Any chance the fix go into > > Linus' tree, to not waste other people's time bisecting and debugging > > the problem? ;-) > > > > For google: tracing, regression, "ring-buffer: move some metadata > > into buffer page", commit abc9b56d66fbd4d93302ef4bf6fa726e1b8255f9, > > answer is here. > > > > include/linux/ftrace.h | 2 +- > > 1 files changed, 1 insertions(+), 1 deletions(-) > > > > diff --git a/include/linux/ftrace.h b/include/linux/ftrace.h > > index 9c5bc6b..13eba02 100644 > > --- a/include/linux/ftrace.h > > +++ b/include/linux/ftrace.h > > @@ -146,7 +146,7 @@ static inline void __ftrace_enabled_restore(int enabled) > > #endif > > } > > > > -#ifdef CONFIG_FRAME_POINTER > > +#if defined(CONFIG_FRAME_POINTER) || defined(CONFIG_PPC) Perhaps we should add a HAVE_NORMAL_FRAME_POINTERS in arch/powerpc/Kconfig under PPC and then we can change the above line to: #if defined(CONFIG_FRAME_POINTERS) || \ defined(CONFIG_HAVE_NORMAL_FRAME_POINTERS) This way when another arch wants to belong to this, we do not need to have a list of archs here. -- Steve > > /* TODO: need to fix this for ARM */ > > # define CALLER_ADDR0 ((unsigned long)__builtin_return_address(0)) > > # define CALLER_ADDR1 ((unsigned long)__builtin_return_address(1)) >
diff --git a/include/linux/ftrace.h b/include/linux/ftrace.h index 9c5bc6b..13eba02 100644 --- a/include/linux/ftrace.h +++ b/include/linux/ftrace.h @@ -146,7 +146,7 @@ static inline void __ftrace_enabled_restore(int enabled) #endif } -#ifdef CONFIG_FRAME_POINTER +#if defined(CONFIG_FRAME_POINTER) || defined(CONFIG_PPC) /* TODO: need to fix this for ARM */ # define CALLER_ADDR0 ((unsigned long)__builtin_return_address(0)) # define CALLER_ADDR1 ((unsigned long)__builtin_return_address(1))
According to this discussion: http://lkml.org/lkml/2008/7/25/338 http://lkml.org/lkml/2008/7/26/72 Frame pointers do nothing useful on PowerPC, so lib/Kconfig.debug makes CONFIG_FRAME_POINTER unselectable on PPC targets. But ftrace.h requires CONFIG_FRAME_POINTER for CALLER_ADDR macros. Therefore tracing is completely useless on PowerPC: [...] <idle>-0 0X.h3 2us+: 0:140:R + [000] 1733:120:S mvtsd <idle>-0 0X.h3 9us+: 0 (0) <idle>-0 0X..3 72us : 0 (0) <idle>-0 0X..3 73us : 0:140:R ==> [000] 1733:120:R mvtsd On PPC we can safely use __builtin_return_address(1..6) w/o frame pointers, and with this patch the trace output looks OK: [...] <idle>-0 0X.h3 2us+: 0:140:R + [000] 1740:120:S mvtsd <idle>-0 0X.h3 9us+: hrtimer_wakeup (__run_hrtimer) <idle>-0 0X..3 87us : cpu_idle (__got2_end) <idle>-0 0X..3 89us : 0:140:R ==> [000] 1740:120:R mvtsd Signed-off-by: Anton Vorontsov <avorontsov@ru.mvista.com> --- Btw, on PowerPC tracing is also broken w/o "ring-buffer: fix alignment problem" patch (currently collecting dust in the -tip tree, commit 082605de5f82eb692cc90f7fda071cc01bb5ac34). Any chance the fix go into Linus' tree, to not waste other people's time bisecting and debugging the problem? ;-) For google: tracing, regression, "ring-buffer: move some metadata into buffer page", commit abc9b56d66fbd4d93302ef4bf6fa726e1b8255f9, answer is here. include/linux/ftrace.h | 2 +- 1 files changed, 1 insertions(+), 1 deletions(-)