2018-10-09 15:40:48

by Lukasz Luba

[permalink] [raw]
Subject: [PATCH] lib/Kconfig.debug: add a comment to PROVE_LOCKING impact

This patch add some comment related to performance impact,
which can be really big (x3 times slower context switch).

Signed-off-by: Lukasz Luba <[email protected]>
---
lib/Kconfig.debug | 11 +++++++++++
1 file changed, 11 insertions(+)

diff --git a/lib/Kconfig.debug b/lib/Kconfig.debug
index 4966c4f..9e67a2a3 100644
--- a/lib/Kconfig.debug
+++ b/lib/Kconfig.debug
@@ -1088,6 +1088,9 @@ config PROVE_LOCKING
the proof of observed correctness is also maintained for an
arbitrary combination of these separate locking variants.

+ This feature enables LOCKDEP which can harm system performance
+ even x3 times.
+
For more details, see Documentation/locking/lockdep-design.txt.

config LOCK_STAT
@@ -1112,6 +1115,10 @@ config LOCK_STAT
CONFIG_LOCK_STAT defines "contended" and "acquired" lock events.
(CONFIG_LOCKDEP defines "acquire" and "release" events.)

+ This feature enables LOCKDEP which can harm system performance
+ even x3 times.
+ For more details, see Documentation/locking/lockdep-design.txt.
+
config DEBUG_RT_MUTEXES
bool "RT Mutex debugging, deadlock detection"
depends on DEBUG_KERNEL && RT_MUTEXES
@@ -1175,6 +1182,10 @@ config DEBUG_LOCK_ALLOC
spin_lock_init()/mutex_init()/etc., or whether there is any lock
held during task exit.

+ This feature enables LOCKDEP which can harm system performance
+ even x3 times.
+ For more details, see Documentation/locking/lockdep-design.txt.
+
config LOCKDEP
bool
depends on DEBUG_KERNEL && LOCK_DEBUGGING_SUPPORT
--
2.7.4



2018-10-09 15:54:34

by Nikolay Borisov

[permalink] [raw]
Subject: Re: [PATCH] lib/Kconfig.debug: add a comment to PROVE_LOCKING impact



On 9.10.2018 18:39, Lukasz Luba wrote:
> This patch add some comment related to performance impact,
> which can be really big (x3 times slower context switch).
>
> Signed-off-by: Lukasz Luba <[email protected]>

I don't think this brings any value. lockdep is a debugging aid aimed at
developers. A developer should be aware that it has a performance impact
and the classification x3 doesn't really bring anything.
> ---
> lib/Kconfig.debug | 11 +++++++++++
> 1 file changed, 11 insertions(+)
>
> diff --git a/lib/Kconfig.debug b/lib/Kconfig.debug
> index 4966c4f..9e67a2a3 100644
> --- a/lib/Kconfig.debug
> +++ b/lib/Kconfig.debug
> @@ -1088,6 +1088,9 @@ config PROVE_LOCKING
> the proof of observed correctness is also maintained for an
> arbitrary combination of these separate locking variants.
>
> + This feature enables LOCKDEP which can harm system performance
> + even x3 times.
> +
> For more details, see Documentation/locking/lockdep-design.txt.
>
> config LOCK_STAT
> @@ -1112,6 +1115,10 @@ config LOCK_STAT
> CONFIG_LOCK_STAT defines "contended" and "acquired" lock events.
> (CONFIG_LOCKDEP defines "acquire" and "release" events.)
>
> + This feature enables LOCKDEP which can harm system performance
> + even x3 times.
> + For more details, see Documentation/locking/lockdep-design.txt.
> +
> config DEBUG_RT_MUTEXES
> bool "RT Mutex debugging, deadlock detection"
> depends on DEBUG_KERNEL && RT_MUTEXES
> @@ -1175,6 +1182,10 @@ config DEBUG_LOCK_ALLOC
> spin_lock_init()/mutex_init()/etc., or whether there is any lock
> held during task exit.
>
> + This feature enables LOCKDEP which can harm system performance
> + even x3 times.
> + For more details, see Documentation/locking/lockdep-design.txt.
> +
> config LOCKDEP
> bool
> depends on DEBUG_KERNEL && LOCK_DEBUGGING_SUPPORT
>

2018-10-09 17:30:21

by Waiman Long

[permalink] [raw]
Subject: Re: [PATCH] lib/Kconfig.debug: add a comment to PROVE_LOCKING impact

On 10/09/2018 11:39 AM, Lukasz Luba wrote:
> This patch add some comment related to performance impact,
> which can be really big (x3 times slower context switch).
>
> Signed-off-by: Lukasz Luba <[email protected]>
> ---
> lib/Kconfig.debug | 11 +++++++++++
> 1 file changed, 11 insertions(+)
>
> diff --git a/lib/Kconfig.debug b/lib/Kconfig.debug
> index 4966c4f..9e67a2a3 100644
> --- a/lib/Kconfig.debug
> +++ b/lib/Kconfig.debug
> @@ -1088,6 +1088,9 @@ config PROVE_LOCKING
> the proof of observed correctness is also maintained for an
> arbitrary combination of these separate locking variants.
>
> + This feature enables LOCKDEP which can harm system performance
> + even x3 times.
> +
> For more details, see Documentation/locking/lockdep-design.txt.
>
> config LOCK_STAT
> @@ -1112,6 +1115,10 @@ config LOCK_STAT
> CONFIG_LOCK_STAT defines "contended" and "acquired" lock events.
> (CONFIG_LOCKDEP defines "acquire" and "release" events.)
>
> + This feature enables LOCKDEP which can harm system performance
> + even x3 times.
> + For more details, see Documentation/locking/lockdep-design.txt.
> +
> config DEBUG_RT_MUTEXES
> bool "RT Mutex debugging, deadlock detection"
> depends on DEBUG_KERNEL && RT_MUTEXES
> @@ -1175,6 +1182,10 @@ config DEBUG_LOCK_ALLOC
> spin_lock_init()/mutex_init()/etc., or whether there is any lock
> held during task exit.
>
> + This feature enables LOCKDEP which can harm system performance
> + even x3 times.
> + For more details, see Documentation/locking/lockdep-design.txt.
> +
> config LOCKDEP
> bool
> depends on DEBUG_KERNEL && LOCK_DEBUGGING_SUPPORT

I don't mind seeing warning message about the performance impact of
enabling lockdep. However, the actual slowdown depends on many factors
including the type of systems, workloads and so on. I am a bit hesitant
to quote an actual figure like x3 in the help text as it may not be
correct in many cases.

BTW, my recent lockdep patch (https://lkml.org/lkml/2018/10/2/1410)
should be able to reduce the performance impact of enabling lockdep.
Could you try to see if that help in your case?

Cheers,
Longman



2018-10-10 08:43:50

by Lukasz Luba

[permalink] [raw]
Subject: Re: [PATCH] lib/Kconfig.debug: add a comment to PROVE_LOCKING impact

Hi Longman,

Thanks for the patches. I have applied them and run on ARM.
In context switch test, there is ~2% improvement
(but still ~15us instead of 5-6us).
In dhrystone: test results are the same.
There is also 2% improvement in boot time.

You can add me on CC list if you have some patches related to
this feature. I will give it a try on arm platforms.

Regards,
Lukasz

On 10/09/2018 07:29 PM, Waiman Long wrote:
> On 10/09/2018 11:39 AM, Lukasz Luba wrote:
>> This patch add some comment related to performance impact,
>> which can be really big (x3 times slower context switch).
>>
>> Signed-off-by: Lukasz Luba <[email protected]>
>> ---
>> lib/Kconfig.debug | 11 +++++++++++
>> 1 file changed, 11 insertions(+)
>>
>> diff --git a/lib/Kconfig.debug b/lib/Kconfig.debug
>> index 4966c4f..9e67a2a3 100644
>> --- a/lib/Kconfig.debug
>> +++ b/lib/Kconfig.debug
>> @@ -1088,6 +1088,9 @@ config PROVE_LOCKING
>> the proof of observed correctness is also maintained for an
>> arbitrary combination of these separate locking variants.
>>
>> + This feature enables LOCKDEP which can harm system performance
>> + even x3 times.
>> +
>> For more details, see Documentation/locking/lockdep-design.txt.
>>
>> config LOCK_STAT
>> @@ -1112,6 +1115,10 @@ config LOCK_STAT
>> CONFIG_LOCK_STAT defines "contended" and "acquired" lock events.
>> (CONFIG_LOCKDEP defines "acquire" and "release" events.)
>>
>> + This feature enables LOCKDEP which can harm system performance
>> + even x3 times.
>> + For more details, see Documentation/locking/lockdep-design.txt.
>> +
>> config DEBUG_RT_MUTEXES
>> bool "RT Mutex debugging, deadlock detection"
>> depends on DEBUG_KERNEL && RT_MUTEXES
>> @@ -1175,6 +1182,10 @@ config DEBUG_LOCK_ALLOC
>> spin_lock_init()/mutex_init()/etc., or whether there is any lock
>> held during task exit.
>>
>> + This feature enables LOCKDEP which can harm system performance
>> + even x3 times.
>> + For more details, see Documentation/locking/lockdep-design.txt.
>> +
>> config LOCKDEP
>> bool
>> depends on DEBUG_KERNEL && LOCK_DEBUGGING_SUPPORT
>
> I don't mind seeing warning message about the performance impact of
> enabling lockdep. However, the actual slowdown depends on many factors
> including the type of systems, workloads and so on. I am a bit hesitant
> to quote an actual figure like x3 in the help text as it may not be
> correct in many cases.
>
> BTW, my recent lockdep patch (https://lkml.org/lkml/2018/10/2/1410)
> should be able to reduce the performance impact of enabling lockdep.
> Could you try to see if that help in your case?
>
> Cheers,
> Longman
>
>
>
>