Hi Christophe,
Thankyou for reviewing the patch.
Le 05/12/2019 à 06:25, Kajol Jain a écrit :
Many of the performance moniroting unit (PMU) SPRs are
exposed in the sysfs. "perf" API is the primary interface to program
PMU and collect counter data in the system. So expose these
PMU SPRs in the absence of CONFIG_PERF_EVENTS.
Patch adds a new CONFIG option 'CONFIG_PMU_SYSFS'. The new config
option used in kernel/sysfs.c for PMU SPRs sysfs file creation and
this new option is enabled only if 'CONFIG_PERF_EVENTS' option is
disabled.
Not sure this new unselectable option is worth it. See below.
By the way I also find the subject misleading, as you may believe when reading it that there is something to select.
Ok I wiil change the subject.
Tested this patch with enable/disable CONFIG_PERF_EVENTS option
in powernv and pseries machines.
Also did compilation testing for different architecture include:
x86, mips, mips64, alpha, arm. And with book3s_32.config option.
How do you use book3s_32.config exactly ? That's a portion of config, not a config by itself. You should use pmac32_defconfig I guess.
Yes you are right, its not a config option. Actually I was playing around with 'CONFIG_PPC_BOOK3S' option in .config file. As you suggested,
I will try to build with 'pmac32_defconfig' option.
Signed-off-by: Kajol Jain <kjain@linux.ibm.com>
Reviewed-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Tested-by: Nageswara R Sastry <nasastry@in.ibm.com>
Tested using the following different scenarios:
1. CONFIG_PERF_EVENT - enabled, CONFIG_PMU_SYSFS - disabled,
RESULT: not seen any sysfs files(mmrc*, pmc*) from /sys/bus/cpu/devices/cpu?/
2. CONFIG_PERF_EVENT - disabled, CONFIG_PMU_SYSFS - enabled,
RESULT: seen any sysfs files(mmrc*, pmc*) from /sys/bus/cpu/devices/cpu?/
3. CONFIG_PERF_EVENT -disabled, CONFIG_PMU_SYSFS - disabled,
RESULT: not possible, any one of the config options need to be enabled.
4. CONFIG_PERF_EVENT -enabled, CONFIG_PMU_SYSFS - enabled,
RESULT: not possible, any one of the config options need to be enabled.
---
arch/powerpc/kernel/sysfs.c | 21 +++++++++++++++++++++
arch/powerpc/platforms/Kconfig.cputype | 8 ++++++++
2 files changed, 29 insertions(+)
---
Changelog:
Resend v2
Added 'Reviewed-by' and 'Tested-by' tag along with test scenarios.
v1 -> v2
- Added new config option 'PMU_SYSFS' for PMU SPR's creation
rather than using PERF_EVENTS config option directly and make
sure SPR's file creation only if 'CONFIG_PERF_EVENTS' disabled.
---
diff --git a/arch/powerpc/kernel/sysfs.c b/arch/powerpc/kernel/sysfs.c
index 80a676da11cb..b7c01f1ef236 100644
--- a/arch/powerpc/kernel/sysfs.c
+++ b/arch/powerpc/kernel/sysfs.c
@@ -457,16 +457,21 @@ static ssize_t __used \
#if defined(CONFIG_PPC64)
#define HAS_PPC_PMC_CLASSIC 1
+#ifdef CONFIG_PMU_SYSFS
#define HAS_PPC_PMC_IBM 1
+#endif
#define HAS_PPC_PMC_PA6T 1
#elif defined(CONFIG_PPC_BOOK3S_32)
#define HAS_PPC_PMC_CLASSIC 1
+#ifdef CONFIG_PMU_SYSFS
#define HAS_PPC_PMC_IBM 1
#define HAS_PPC_PMC_G4 1
#endif
+#endif
#ifdef HAS_PPC_PMC_CLASSIC
+#ifdef CONFIG_PMU_SYSFS
You don't need this big forest of #ifdefs (this one and all the ones after). All the objets you are protecting with this are indeed static. So the only thing you have to do is to register them only when relevant, and GCC will get rid of the objects by itself when the config option is not enabled. See below.
And the advantage of doing that way is that you don't need to build it with both options to check the build. That's recommended by kernel codying style (Refer https://www.kernel.org/doc/html/latest/process/coding-style.html#conditional-compilation)
[...]
@@ -787,8 +804,10 @@ static int register_cpu_online(unsigned int cpu)
device_create_file(s, &pmc_attrs[i]);
#ifdef CONFIG_PPC64
+#ifdef CONFIG_PMU_SYSFS
Don't use #ifdef here, just do instead:
if (IS_ENABLED(CONFIG_PMU_SYSFS) && cpu_has_feature(CPU_FTR_MMCRA))
Thanks for the suggestion I will use IS_ENABLED here.
if (cpu_has_feature(CPU_FTR_MMCRA))
device_create_file(s, &dev_attr_mmcra);
+#endif /* CONFIG_PMU_SYSFS */
if (cpu_has_feature(CPU_FTR_PURR)) {
if (!firmware_has_feature(FW_FEATURE_LPAR))
@@ -876,8 +895,10 @@ static int unregister_cpu_online(unsigned int cpu)
device_remove_file(s, &pmc_attrs[i]);
#ifdef CONFIG_PPC64
+#ifdef CONFIG_PMU_SYSFS
Same, use IS_ENABLED() here as well.
if (cpu_has_feature(CPU_FTR_MMCRA))
device_remove_file(s, &dev_attr_mmcra);
+#endif /* CONFIG_PMU_SYSFS */
if (cpu_has_feature(CPU_FTR_PURR))
device_remove_file(s, &dev_attr_purr);
diff --git a/arch/powerpc/platforms/Kconfig.cputype b/arch/powerpc/platforms/Kconfig.cputype
index 12543e53fa96..f3ad579c559f 100644
--- a/arch/powerpc/platforms/Kconfig.cputype
+++ b/arch/powerpc/platforms/Kconfig.cputype
@@ -417,6 +417,14 @@ config PPC_MM_SLICES
config PPC_HAVE_PMU_SUPPORT
bool
+config PMU_SYSFS
+ bool
+ default y if !PERF_EVENTS
+ help
+ This option enables PMU SPR sysfs file creation. Since PMU SPRs are
+ intended to be used via "perf" interface, config option is enabled
+ only when CONFIG_PERF_EVENTS is disabled.
+
Not sure you need this at all. Once you have changed to just using IS_ENABLED() in the two places above, I think it is acceptable to use !IS_ENABLED(CONFIG_PERF_EVENTS) instead.
Actually with v1 of the patch, I tried with PERF_EVENT option, but it was getting bit messy to recreate the current arrangement in the file. So took a new config option path.
config PPC_PERF_CTRS
def_bool y
depends on PERF_EVENTS && PPC_HAVE_PMU_SUPPORT
Christophe
Thanks,
Kajol Jain