PROVE_LOCKING enables LOCKDEP, which causes big overhead on cache and
bus transactions.
On some ARM big.LITTLE architecutres (Exynos 5433) the overhead is really big.
The overhead can be measures using hackbench test which will speed up
by x3 times (11sec -> 3.4sec).
When you check transaction on cache or buses, the results are way higher
than normal for the same hackbench test:
L1d cache invalidations: 26mln vs 4mln
L2u cache invalidations: 42mln vs 12mln
bus cyc/access: 30cyc/access vs. 20cyc/access
context switch is x3 times cheaper
Enable this option only when you have some locking issue to investigate.
Signed-off-by: Lukasz Luba <[email protected]>
---
arch/arm/configs/exynos_defconfig | 1 -
1 file changed, 1 deletion(-)
diff --git a/arch/arm/configs/exynos_defconfig b/arch/arm/configs/exynos_defconfig
index 27ea6dfc..fbf52f3 100644
--- a/arch/arm/configs/exynos_defconfig
+++ b/arch/arm/configs/exynos_defconfig
@@ -324,7 +324,6 @@ CONFIG_MAGIC_SYSRQ=y
CONFIG_DEBUG_KERNEL=y
CONFIG_SOFTLOCKUP_DETECTOR=y
# CONFIG_DETECT_HUNG_TASK is not set
-CONFIG_PROVE_LOCKING=y
CONFIG_DEBUG_ATOMIC_SLEEP=y
CONFIG_DEBUG_USER=y
CONFIG_CRYPTO_RSA=m
--
2.7.4
On Tue, 9 Oct 2018 at 17:34, Lukasz Luba <[email protected]> wrote:
>
> PROVE_LOCKING enables LOCKDEP, which causes big overhead on cache and
> bus transactions.
>
> On some ARM big.LITTLE architecutres (Exynos 5433) the overhead is really big.
> The overhead can be measures using hackbench test which will speed up
> by x3 times (11sec -> 3.4sec).
> When you check transaction on cache or buses, the results are way higher
> than normal for the same hackbench test:
> L1d cache invalidations: 26mln vs 4mln
> L2u cache invalidations: 42mln vs 12mln
> bus cyc/access: 30cyc/access vs. 20cyc/access
> context switch is x3 times cheaper
>
> Enable this option only when you have some locking issue to investigate.
>
> Signed-off-by: Lukasz Luba <[email protected]>
> ---
> arch/arm/configs/exynos_defconfig | 1 -
> 1 file changed, 1 deletion(-)
>
> diff --git a/arch/arm/configs/exynos_defconfig b/arch/arm/configs/exynos_defconfig
> index 27ea6dfc..fbf52f3 100644
> --- a/arch/arm/configs/exynos_defconfig
> +++ b/arch/arm/configs/exynos_defconfig
> @@ -324,7 +324,6 @@ CONFIG_MAGIC_SYSRQ=y
> CONFIG_DEBUG_KERNEL=y
> CONFIG_SOFTLOCKUP_DETECTOR=y
> # CONFIG_DETECT_HUNG_TASK is not set
> -CONFIG_PROVE_LOCKING=y
You just disabled all debugging around locks (so also basic checks).
You did not describe this in the message... and it also makes me
wonder what were you testing? Kernel without any lock debugging (so
this patch) or kernel without lockdep only (not this patch).
As for the reason behind this change, I am not sure. We had locking
issues and this change really helped us to spot them. Most of the
issues are solved so maybe this is not needed anymore but on the other
hand - this is development/debugging config, not a distro-ready. It
might serve as a reference for distro configs but clearly it is not
for production.
Best regards,
Krzysztof
> CONFIG_DEBUG_ATOMIC_SLEEP=y
> CONFIG_DEBUG_USER=y
> CONFIG_CRYPTO_RSA=m
> --
> 2.7.4
>
Hi Krzysztof,
Is it still valid that the Exynos has locking issues?
Or is it enabled 'just-in-case'?
On arm64 the config misses that setting.
Should be enabled there as well?
You have introduced this entry a year ago, probably during bug
investigation.
The issue is in lockdep, the setting which enables it in Exynos
is prove_locking. For now, just get rid of the second.
My tests showed lockdep impact and not other minor debug features
enabled.
Regards,
Lukasz
On 10/10/2018 11:18 AM, Krzysztof Kozlowski wrote:
> On Tue, 9 Oct 2018 at 17:34, Lukasz Luba <[email protected]> wrote:
>>
>> PROVE_LOCKING enables LOCKDEP, which causes big overhead on cache and
>> bus transactions.
>>
>> On some ARM big.LITTLE architecutres (Exynos 5433) the overhead is really big.
>> The overhead can be measures using hackbench test which will speed up
>> by x3 times (11sec -> 3.4sec).
>> When you check transaction on cache or buses, the results are way higher
>> than normal for the same hackbench test:
>> L1d cache invalidations: 26mln vs 4mln
>> L2u cache invalidations: 42mln vs 12mln
>> bus cyc/access: 30cyc/access vs. 20cyc/access
>> context switch is x3 times cheaper
>>
>> Enable this option only when you have some locking issue to investigate.
>>
>> Signed-off-by: Lukasz Luba <[email protected]>
>> ---
>> arch/arm/configs/exynos_defconfig | 1 -
>> 1 file changed, 1 deletion(-)
>>
>> diff --git a/arch/arm/configs/exynos_defconfig b/arch/arm/configs/exynos_defconfig
>> index 27ea6dfc..fbf52f3 100644
>> --- a/arch/arm/configs/exynos_defconfig
>> +++ b/arch/arm/configs/exynos_defconfig
>> @@ -324,7 +324,6 @@ CONFIG_MAGIC_SYSRQ=y
>> CONFIG_DEBUG_KERNEL=y
>> CONFIG_SOFTLOCKUP_DETECTOR=y
>> # CONFIG_DETECT_HUNG_TASK is not set
>> -CONFIG_PROVE_LOCKING=y
>
> You just disabled all debugging around locks (so also basic checks).
> You did not describe this in the message... and it also makes me
> wonder what were you testing? Kernel without any lock debugging (so
> this patch) or kernel without lockdep only (not this patch).
>
> As for the reason behind this change, I am not sure. We had locking
> issues and this change really helped us to spot them. Most of the
> issues are solved so maybe this is not needed anymore but on the other
> hand - this is development/debugging config, not a distro-ready. It
> might serve as a reference for distro configs but clearly it is not
> for production.
>
> Best regards,
> Krzysztof
>
>> CONFIG_DEBUG_ATOMIC_SLEEP=y
>> CONFIG_DEBUG_USER=y
>> CONFIG_CRYPTO_RSA=m
>> --
>> 2.7.4
>>
>
>
On Wed, 10 Oct 2018 at 11:56, Lukasz Luba <[email protected]> wrote:
>
> Hi Krzysztof,
>
> Is it still valid that the Exynos has locking issues?
The prepare lock of clocks was not solved entirely. We just
implemented a set of workarounds.
> Or is it enabled 'just-in-case'?
Yes, all debugging features are implemented just in case to spot any
issue as early as possible.
> On arm64 the config misses that setting.
> Should be enabled there as well?
Because it is a shared defconfig among all platforms and apparently no
one wanted to convincing others to enable it.
> You have introduced this entry a year ago, probably during bug
> investigation.
Actually not, I enabled it much later after issues where discovered.
> The issue is in lockdep, the setting which enables it in Exynos
> is prove_locking. For now, just get rid of the second.
> My tests showed lockdep impact and not other minor debug features
> enabled.
You did not reply directly to my comment - you disabled all debug for
locking. It seems you are unaware of this so I question your test
results. If you run tests on this patch, then basically you compared
existing config with a config stripped out of debugging around locks.
I understand that this is not what you wanted...
Best regards,
Krzysztof