Patchwork [1/4] Allow arch-specific cleanup before breakpoint unregistration

login
register
mail settings
Submitter K.Prasad
Date May 25, 2010, 9:13 a.m.
Message ID <20100525091356.GB29003@in.ibm.com>
Download mbox | patch
Permalink /patch/53518/
State Superseded
Headers show

Comments

K.Prasad - May 25, 2010, 9:13 a.m.
Certain architectures (such as PowerPC Book III S) have a need to cleanup
data-structures before the breakpoint is unregistered. This patch introduces
an arch-specific hook in release_bp_slot() along with a weak definition in
the form of a stub funciton.

Signed-off-by: K.Prasad <prasad@linux.vnet.ibm.com>
---
 kernel/hw_breakpoint.c |   12 ++++++++++++
 1 file changed, 12 insertions(+)
Milton Miller - May 25, 2010, 11:39 a.m.
On Tue, 25 May 2010 at 14:43:56 +0530, K.Prasad wrote:
> Certain architectures (such as PowerPC Book III S) have a need to cleanup
> data-structures before the breakpoint is unregistered. This patch introduces
> an arch-specific hook in release_bp_slot() along with a weak definition in
> the form of a stub funciton.
> 
> Signed-off-by: K.Prasad <prasad@linux.vnet.ibm.com>
> ---
>  kernel/hw_breakpoint.c |   12 ++++++++++++
>  1 file changed, 12 insertions(+)


My understanding is weak function definitions must appear in a different C
file than their call sites to work on some toolchains.

Andrew, can you confirm the above statement?

> Index: linux-2.6.ppc64_test/kernel/hw_breakpoint.c
> ===================================================================
> --- linux-2.6.ppc64_test.orig/kernel/hw_breakpoint.c
> +++ linux-2.6.ppc64_test/kernel/hw_breakpoint.c
> @@ -242,6 +242,17 @@  toggle_bp_slot(struct perf_event *bp, bo
>  }
>  
>  /*
> + * Function to perform processor-specific cleanup during unregistration
> + */
> +__weak void arch_unregister_hw_breakpoint(struct perf_event *bp)
> +{
> +	/*
> +	 * A weak stub function here for those archs that don't define
> +	 * it inside arch/.../kernel/hw_breakpoint.c
> +	 */
> +}
> +
> +/*
>   * Contraints to check before allowing this new breakpoint counter:
>   *
>   *  == Non-pinned counter == (Considered as pinned for now)
> @@ -339,6 +350,7 @@  void release_bp_slot(struct perf_event *
>  {
>  	mutex_lock(&nr_bp_mutex);
>  
> +	arch_unregister_hw_breakpoint(bp);
>  	__release_bp_slot(bp);
>  
>  	mutex_unlock(&nr_bp_mutex);
> 


Since the weak version is empty, should it just be delcared (in
a header, put the comment there) and not defined?

milton
K.Prasad - May 26, 2010, 6:51 a.m.
On Tue, May 25, 2010 at 06:39:19AM -0500, Millton Miller wrote:
> On Tue, 25 May 2010 at 14:43:56 +0530, K.Prasad wrote:
> > Certain architectures (such as PowerPC Book III S) have a need to cleanup
> > data-structures before the breakpoint is unregistered. This patch introduces
> > an arch-specific hook in release_bp_slot() along with a weak definition in
> > the form of a stub funciton.
> > 
> > Signed-off-by: K.Prasad <prasad@linux.vnet.ibm.com>
> > ---
> >  kernel/hw_breakpoint.c |   12 ++++++++++++
> >  1 file changed, 12 insertions(+)
> 
> 
> My understanding is weak function definitions must appear in a different C
> file than their call sites to work on some toolchains.
> 

Atleast, there are quite a few precedents inside the Linux kernel for
__weak functions being invoked from the file in which they are defined
(arch_hwblk_init, arch_enable_nonboot_cpus_begin and hw_perf_disable to
name a few).
Moreover the online GCC docs haven't any such constraints mentioned.

> Andrew, can you confirm the above statement?
> 
> > Index: linux-2.6.ppc64_test/kernel/hw_breakpoint.c
> > ===================================================================
> > --- linux-2.6.ppc64_test.orig/kernel/hw_breakpoint.c
> > +++ linux-2.6.ppc64_test/kernel/hw_breakpoint.c
> > @@ -242,6 +242,17 @@  toggle_bp_slot(struct perf_event *bp, bo
> >  }
> >  
> >  /*
> > + * Function to perform processor-specific cleanup during unregistration
> > + */
> > +__weak void arch_unregister_hw_breakpoint(struct perf_event *bp)
> > +{
> > +	/*
> > +	 * A weak stub function here for those archs that don't define
> > +	 * it inside arch/.../kernel/hw_breakpoint.c
> > +	 */
> > +}
> > +
> > +/*
> >   * Contraints to check before allowing this new breakpoint counter:
> >   *
> >   *  == Non-pinned counter == (Considered as pinned for now)
> > @@ -339,6 +350,7 @@  void release_bp_slot(struct perf_event *
> >  {
> >  	mutex_lock(&nr_bp_mutex);
> >  
> > +	arch_unregister_hw_breakpoint(bp);
> >  	__release_bp_slot(bp);
> >  
> >  	mutex_unlock(&nr_bp_mutex);
> > 
> 
> 
> Since the weak version is empty, should it just be delcared (in
> a header, put the comment there) and not defined?
>

The initial thinking behind defining it in the .c file was, for one,
the function need not be moved (from .h to .c) when other architectures
have a need to populate them. Secondly, given that powerpc (which has a
'strong' definition for arch_unregister_hw_breakpoint()) includes the
header file (in which this can be moved to) I wasn't sure about
possible conflicts.

> milton
> _______________________________________________
> Linuxppc-dev mailing list
> Linuxppc-dev@lists.ozlabs.org
> https://lists.ozlabs.org/listinfo/linuxppc-dev

Thanks,
K.Prasad
David Howells - May 26, 2010, 9:54 a.m.
K.Prasad <prasad@linux.vnet.ibm.com> wrote:

> > My understanding is weak function definitions must appear in a different C
> > file than their call sites to work on some toolchains.
> > 
> 
> Atleast, there are quite a few precedents inside the Linux kernel for
> __weak functions being invoked from the file in which they are defined
> (arch_hwblk_init, arch_enable_nonboot_cpus_begin and hw_perf_disable to
> name a few).
> Moreover the online GCC docs haven't any such constraints mentioned.

I've seen problems in this area.  gcc sometimes inlines a weak function that's
in the same file as the call point.

David
Michael Ellerman - May 26, 2010, 3:13 p.m.
On Wed, 2010-05-26 at 10:54 +0100, David Howells wrote:
> K.Prasad <prasad@linux.vnet.ibm.com> wrote:
> 
> > > My understanding is weak function definitions must appear in a different C
> > > file than their call sites to work on some toolchains.
> > > 
> > 
> > Atleast, there are quite a few precedents inside the Linux kernel for
> > __weak functions being invoked from the file in which they are defined
> > (arch_hwblk_init, arch_enable_nonboot_cpus_begin and hw_perf_disable to
> > name a few).
> > Moreover the online GCC docs haven't any such constraints mentioned.
> 
> I've seen problems in this area.  gcc sometimes inlines a weak function that's
> in the same file as the call point.

See the functions in kernel/softirq.c for example, and commits 43a256322
and b2e2fe996 - though unhelpfully they don't mention the gcc version. A
bit of googling suggests it was probably "gcc version 4.1.1 20060525
(Red Hat 4.1.1-1)" in that case.

But the example of hw_perf_enable() (which is weak in the same unit),
suggests maybe this isn't a bug many people are hitting in practice
anymore.

Having said that the #define foo foo pattern is reasonably neat and
avoids the problem altogether, see eg. arch_setup_msi_irqs.

cheers
K.Prasad - May 26, 2010, 5:17 p.m.
On Wed, May 26, 2010 at 10:54:41AM +0100, David Howells wrote:
> K.Prasad <prasad@linux.vnet.ibm.com> wrote:
> 
> > > My understanding is weak function definitions must appear in a different C
> > > file than their call sites to work on some toolchains.
> > > 
> > 
> > Atleast, there are quite a few precedents inside the Linux kernel for
> > __weak functions being invoked from the file in which they are defined
> > (arch_hwblk_init, arch_enable_nonboot_cpus_begin and hw_perf_disable to
> > name a few).
> > Moreover the online GCC docs haven't any such constraints mentioned.
> 
> I've seen problems in this area.  gcc sometimes inlines a weak function that's
> in the same file as the call point.
> 

We've seen such behaviour even otherwise....even with noinline attribute
in place. I'm not sure if this gcc fix
(http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16922) helped correct the
behaviour, but the lesson has been to not trust a function to be
inlined/remain non-inline consistently.

> David

Thanks,
K.Prasad
Frédéric Weisbecker - May 26, 2010, 5:23 p.m.
On Wed, May 26, 2010 at 10:47:42PM +0530, K.Prasad wrote:
> On Wed, May 26, 2010 at 10:54:41AM +0100, David Howells wrote:
> > K.Prasad <prasad@linux.vnet.ibm.com> wrote:
> > 
> > > > My understanding is weak function definitions must appear in a different C
> > > > file than their call sites to work on some toolchains.
> > > > 
> > > 
> > > Atleast, there are quite a few precedents inside the Linux kernel for
> > > __weak functions being invoked from the file in which they are defined
> > > (arch_hwblk_init, arch_enable_nonboot_cpus_begin and hw_perf_disable to
> > > name a few).
> > > Moreover the online GCC docs haven't any such constraints mentioned.
> > 
> > I've seen problems in this area.  gcc sometimes inlines a weak function that's
> > in the same file as the call point.
> > 
> 
> We've seen such behaviour even otherwise....even with noinline attribute
> in place. I'm not sure if this gcc fix
> (http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16922) helped correct the
> behaviour, but the lesson has been to not trust a function to be
> inlined/remain non-inline consistently.


If we can't put the call to the function in the same file of its weak
definition, then perf is totally screwed.

And in fact it makes __weak basically useless and unusable. I guess
that happened in old gcc versions that have been fixed now.

Anyway, I'm personally fine with this patch (you can put my hack
if you want).

Thanks.
K.Prasad - May 26, 2010, 5:28 p.m.
On Wed, May 26, 2010 at 10:47:42PM +0530, K.Prasad wrote:
> On Wed, May 26, 2010 at 10:54:41AM +0100, David Howells wrote:
> > K.Prasad <prasad@linux.vnet.ibm.com> wrote:
> > 
> > > > My understanding is weak function definitions must appear in a different C
> > > > file than their call sites to work on some toolchains.
> > > > 
> > > 
> > > Atleast, there are quite a few precedents inside the Linux kernel for
> > > __weak functions being invoked from the file in which they are defined
> > > (arch_hwblk_init, arch_enable_nonboot_cpus_begin and hw_perf_disable to
> > > name a few).
> > > Moreover the online GCC docs haven't any such constraints mentioned.
> > 
> > I've seen problems in this area.  gcc sometimes inlines a weak function that's
> > in the same file as the call point.
> > 
> 
> We've seen such behaviour even otherwise....even with noinline attribute
> in place. I'm not sure if this gcc fix
> (http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16922) helped correct the

Looks like I cited the wrong bug. The appropriate one is
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34563.

Thanks,
K.Prasad
K.Prasad - May 26, 2010, 5:31 p.m.
On Wed, May 26, 2010 at 07:23:15PM +0200, Frederic Weisbecker wrote:
> On Wed, May 26, 2010 at 10:47:42PM +0530, K.Prasad wrote:
> > On Wed, May 26, 2010 at 10:54:41AM +0100, David Howells wrote:
> > > K.Prasad <prasad@linux.vnet.ibm.com> wrote:
> > > 
> > > > > My understanding is weak function definitions must appear in a different C
> > > > > file than their call sites to work on some toolchains.
> > > > > 
> > > > 
> > > > Atleast, there are quite a few precedents inside the Linux kernel for
> > > > __weak functions being invoked from the file in which they are defined
> > > > (arch_hwblk_init, arch_enable_nonboot_cpus_begin and hw_perf_disable to
> > > > name a few).
> > > > Moreover the online GCC docs haven't any such constraints mentioned.
> > > 
> > > I've seen problems in this area.  gcc sometimes inlines a weak function that's
> > > in the same file as the call point.
> > > 
> > 
> > We've seen such behaviour even otherwise....even with noinline attribute
> > in place. I'm not sure if this gcc fix
> > (http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16922) helped correct the
> > behaviour, but the lesson has been to not trust a function to be
> > inlined/remain non-inline consistently.
> 
> 
> If we can't put the call to the function in the same file of its weak
> definition, then perf is totally screwed.
> 
> And in fact it makes __weak basically useless and unusable. I guess
> that happened in old gcc versions that have been fixed now.
> 
> Anyway, I'm personally fine with this patch (you can put my hack
> if you want).
>

I guess you meant "Acked-by:" :-)

Thanks, I'll add the same.

--K.Prasad
Frédéric Weisbecker - May 26, 2010, 5:35 p.m.
On Wed, May 26, 2010 at 11:01:24PM +0530, K.Prasad wrote:
> On Wed, May 26, 2010 at 07:23:15PM +0200, Frederic Weisbecker wrote:
> > On Wed, May 26, 2010 at 10:47:42PM +0530, K.Prasad wrote:
> > > On Wed, May 26, 2010 at 10:54:41AM +0100, David Howells wrote:
> > > > K.Prasad <prasad@linux.vnet.ibm.com> wrote:
> > > > 
> > > > > > My understanding is weak function definitions must appear in a different C
> > > > > > file than their call sites to work on some toolchains.
> > > > > > 
> > > > > 
> > > > > Atleast, there are quite a few precedents inside the Linux kernel for
> > > > > __weak functions being invoked from the file in which they are defined
> > > > > (arch_hwblk_init, arch_enable_nonboot_cpus_begin and hw_perf_disable to
> > > > > name a few).
> > > > > Moreover the online GCC docs haven't any such constraints mentioned.
> > > > 
> > > > I've seen problems in this area.  gcc sometimes inlines a weak function that's
> > > > in the same file as the call point.
> > > > 
> > > 
> > > We've seen such behaviour even otherwise....even with noinline attribute
> > > in place. I'm not sure if this gcc fix
> > > (http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16922) helped correct the
> > > behaviour, but the lesson has been to not trust a function to be
> > > inlined/remain non-inline consistently.
> > 
> > 
> > If we can't put the call to the function in the same file of its weak
> > definition, then perf is totally screwed.
> > 
> > And in fact it makes __weak basically useless and unusable. I guess
> > that happened in old gcc versions that have been fixed now.
> > 
> > Anyway, I'm personally fine with this patch (you can put my hack
> > if you want).
> >
> 
> I guess you meant "Acked-by:" :-)



Oops, right :)

Patch

Index: linux-2.6.ppc64_test/kernel/hw_breakpoint.c
===================================================================
--- linux-2.6.ppc64_test.orig/kernel/hw_breakpoint.c
+++ linux-2.6.ppc64_test/kernel/hw_breakpoint.c
@@ -242,6 +242,17 @@  toggle_bp_slot(struct perf_event *bp, bo
 }
 
 /*
+ * Function to perform processor-specific cleanup during unregistration
+ */
+__weak void arch_unregister_hw_breakpoint(struct perf_event *bp)
+{
+	/*
+	 * A weak stub function here for those archs that don't define
+	 * it inside arch/.../kernel/hw_breakpoint.c
+	 */
+}
+
+/*
  * Contraints to check before allowing this new breakpoint counter:
  *
  *  == Non-pinned counter == (Considered as pinned for now)
@@ -339,6 +350,7 @@  void release_bp_slot(struct perf_event *
 {
 	mutex_lock(&nr_bp_mutex);
 
+	arch_unregister_hw_breakpoint(bp);
 	__release_bp_slot(bp);
 
 	mutex_unlock(&nr_bp_mutex);