Skip to contentSkip to conversation

Please pull arch perfmon update

DISCUSSION Applied Andi Kleen 2 participants 7 messages Sep 29, 2008 – Oct 14, 2008 x86

lore
Share & tools
lore

Andi Kleen

author
#0 lore raw

Robert,

Here are the arch perfmon changes ported to your latest oprofile tree. I didn't include the forcepmu=... change yet, that will come later. Right now the force option is just removed. Please pull.

Thanks,
-Andi

The following changes since commit 22d87103484983035cf46891428819573ec7508d:
Robert Richter (1):
Merge branch 'oprofile/powerpc-for-paul' into oprofile/master

are available in the git repository at:

git://git.kernel.org/pub/scm/linux/kernel/git/ak/linux-misc-2.6.git arch-perfmon2

Andi Kleen (4):
oprofile: drop const in num counters field
oprofile: Don't report Nehalem as core_2
oprofile: Implement Intel architectural perfmon support
oprofile: discover counters for op ppro too

 arch/x86/oprofile/nmi_int.c       |   26 ++++++---
 arch/x86/oprofile/op_model_ppro.c |  108 +++++++++++++++++++++++++++++--------
 arch/x86/oprofile/op_x86_model.h  |    9 ++-
 3 files changed, 108 insertions(+), 35 deletions(-)
Signature · 1 line
[email protected]
#1 lore raw
Quoted 15 lines · Andi Kleen

Robert,

Here are the arch perfmon changes ported to your latest oprofile tree. I didn't include the forcepmu=... change yet, that will come later. Right now the force option is just removed. Please pull.

Thanks,
-Andi

The following changes since commit 22d87103484983035cf46891428819573ec7508d:
Robert Richter (1):
Merge branch 'oprofile/powerpc-for-paul' into oprofile/master

are available in the git repository at:

Quoted 7 lines

git://git.kernel.org/pub/scm/linux/kernel/git/ak/linux-misc-2.6.git arch-perfmon2

Andi Kleen (4):
oprofile: drop const in num counters field
oprofile: Don't report Nehalem as core_2
oprofile: Implement Intel architectural perfmon support
oprofile: discover counters for op ppro too

Thanks Andi,

I applied your patches to the x86-oprofile-for-tip branch of
git://git.kernel.org/pub/scm/linux/kernel/git/rric/oprofile.git

I added a follow on patch (446223f) on branch arch-perfmon that changes the initialization so that it uses the init function in op_x86_model_spec (see below). Please give it a try, it is compile tested only. If it works for you, I will add it to the tip branch too.

There is one question to patch "oprofile: discover counters for op ppro too". arch_perfmon_setup_counters() is actually never called for ppro, so there is no code that changes the numbers in op_ppro_spec. The patch as it is has no effect (except then loading additional module code).

Thanks,

-Robert

commit 446223f8d1454e647b54160584c5dec8f83fdddc

Author: Robert Richter <[email protected]>
Date:   Sun Oct 12 15:12:34 2008 -0400

x86/oprofile: moving arch perfmon counter setup to op_x86_model_spec.init

arch/x86/oprofile/nmi_int.c +0 -1
index 12d6f85..fd902b8 100644
@@ -435,7 +435,6 @@ static int __init arch_perfmon_init(char **cpu_type)
                return 0;
        *cpu_type = "i386/arch_perfmon";
        model = &op_arch_perfmon_spec;
-       arch_perfmon_setup_counters();
        return 1;
 }
 
arch/x86/oprofile/op_model_ppro.c +4 -1
index f5a2268..3266795 100644
@@ -220,7 +220,7 @@ struct op_x86_model_spec op_ppro_spec = {
  * the specific CPU.
  */
 
-void arch_perfmon_setup_counters(void)
+static int arch_perfmon_setup_counters(struct oprofile_operations *ignore)
 {
        union cpuid10_eax eax;
 
@@ -240,9 +240,12 @@ void arch_perfmon_setup_counters(void)
        op_arch_perfmon_spec.num_controls = num_counters;
        op_ppro_spec.num_counters = num_counters;
        op_ppro_spec.num_controls = num_counters;
+
+       return 0;
 }
 
 struct op_x86_model_spec op_arch_perfmon_spec = {
+       .init = &arch_perfmon_setup_counters,
        /* num_counters/num_controls filled in at runtime */
        .fill_in_addresses = &ppro_fill_in_addresses,
        /* user space does the cpuid check for available events */
arch/x86/oprofile/op_x86_model.h +0 -2
index 596de7a..418c2d0 100644
@@ -51,6 +51,4 @@ extern struct op_x86_model_spec const op_p4_ht2_spec;
 extern struct op_x86_model_spec const op_amd_spec;
 extern struct op_x86_model_spec op_arch_perfmon_spec;
 
-extern void arch_perfmon_setup_counters(void);
-
 #endif /* OP_X86_MODEL_H */
Signature · 3 lines
Advanced Micro Devices, Inc.
Operating System Research Center
email: [email protected]
author
#2 lore raw
Quoted 4 lines

I applied your patches to the x86-oprofile-for-tip branch of
git://git.kernel.org/pub/scm/linux/kernel/git/rric/oprofile.git

Will that still make 2.6.28?

Quoted 1 line

I added a follow on patch (446223f) on branch arch-perfmon that

I didn't do that intentionally because it's called too late. The function really has to be called early, so that the fallback works.

Quoted 5 lines

There is one question to patch "oprofile: discover counters for op ppro too". arch_perfmon_setup_counters() is actually never called for ppro, so there is no code that changes the numbers in op_ppro_spec. The patch as it is has no effect (except then loading additional module code).

Ah true, i must have deleted one hunk too much while refactoring. It's ok to drop this patch for now.

-Andi

#3 lore raw
Quoted 5 lines · Andi Kleen
Quoted 1 line

I added a follow on patch (446223f) on branch arch-perfmon that

I didn't do that intentionally because it's called too late. The function really has to be called early, so that the fallback works.

The hook is in op_nmi_init() and directly called after arch_perfmon_init() and before init_sysfs(). Only register_cpu_notifier() and the setup of oprofile_operations are in between. This should work.

-Robert

Signature · 3 lines
Advanced Micro Devices, Inc.
Operating System Research Center
email: [email protected]
author
#4 lore raw
Quoted 11 lines · Robert Richter
Quoted 5 lines · Andi Kleen
Quoted 1 line

I added a follow on patch (446223f) on branch arch-perfmon that

I didn't do that intentionally because it's called too late. The function really has to be called early, so that the fallback works.

The hook is in op_nmi_init() and directly called after arch_perfmon_init() and before init_sysfs(). Only register_cpu_notifier() and the setup of oprofile_operations are in between. This should work.

The problem is that arch perfmon init should only be called after the other initialization function failed.

So you would need a chain of op_x86_model_spec for fallback.

It's simpler and cleaner to just write that out in explicit C.

-Andi

#5 lore raw
Quoted 13 lines · Andi Kleen
Quoted 11 lines · Robert Richter
Quoted 5 lines · Andi Kleen
Quoted 1 line

I added a follow on patch (446223f) on branch arch-perfmon that

I didn't do that intentionally because it's called too late. The function really has to be called early, so that the fallback works.

The hook is in op_nmi_init() and directly called after arch_perfmon_init() and before init_sysfs(). Only register_cpu_notifier() and the setup of oprofile_operations are in between. This should work.

Quoted 6 lines

The problem is that arch perfmon init should only be called after the other initialization function failed.

So you would need a chain of op_x86_model_spec for fallback.

It's simpler and cleaner to just write that out in explicit C.

The patch makes arch_perfmon_setup_counters() static, so the init code is bound directly to the model (it is model specific code). The interface is much cleaner now since the delaration as an external is not needed then.

The init function is only called, if this cpu type (i386/arch_perfmon) is selected. And this type is only selected after all other init funtions were failing.

Please test the patch, I don't see a reason why it should not work.

-Robert

Signature · 3 lines
Advanced Micro Devices, Inc.
Operating System Research Center
email: [email protected]
#6 lore raw
In reply to Andi Kleen
Quoted 8 lines · Andi Kleen
Quoted 5 lines

There is one question to patch "oprofile: discover counters for op ppro too". arch_perfmon_setup_counters() is actually never called for ppro, so there is no code that changes the numbers in op_ppro_spec. The patch as it is has no effect (except then loading additional module code).

Ah true, i must have deleted one hunk too much while refactoring. It's ok to drop this patch for now.

I sent it upstream anyway, please send an additional patch that fixes this.

Thanks,

-Robert

Signature · 3 lines
Advanced Micro Devices, Inc.
Operating System Research Center
email: [email protected]