From mboxrd@z Thu Jan  1 00:00:00 1970
From: zhaolei@cn.fujitsu.com (Zhaolei)
Date: Tue, 13 Jan 2009 15:59:49 +0800
Subject: [ltt-dev] Degression from "fix-x86_64-sysenter-trace-race"
In-Reply-To: <20090113042452.GA10924@Krystal>
References: <496BEA99.7030901@cn.fujitsu.com> <20090113021332.GA7852@Krystal>
	<496C02F2.3020801@cn.fujitsu.com> <20090113042452.GA10924@Krystal>
Message-ID: <496C49F5.9080300@cn.fujitsu.com>

Mathieu Desnoyers ??:
> * Zhaolei (zhaolei at cn.fujitsu.com) wrote:
>> * Mathieu Desnoyers Wrote:
>>> * Zhaolei (zhaolei at cn.fujitsu.com) wrote:
>>>> Hello,
>>>>
>>>> I tested lttng's performance by tbench program in x86_64 platform.
>>>> And I found patch named "fix-x86_64-sysenter-trace-race" makes large
>>>> degression(about 5%).
>>>> Here is test result:
>>>>
>>>> [----------PATCH----------]                     [TBENCH]
>>>> lttng-kernel-trace-thread-flag-s390		1317.08
>>>> lttng-kernel-trace-thread-flag-sh		1315.61
>>>> lttng-kernel-trace-thread-flag-sparc		1305.02
>>>> lttng-kernel-trace-thread-flag-sparc64		1314.06
>>>> lttng-kernel-trace-thread-flag-um		1316.35
>>>> fix-x86_64-sysenter-trace-race			1250.89
>>>> lttng-kernel-trace-thread-flag-x86		1259.07
>>>> lttng-kernel-trace-thread-flag-xtensa		1252.86
>>>> lttng-kernel-trace-thread-flag-api		1253.35
>>>> lttng-use-kernel-trace-thread-flag-api		1251.9
>>>> stringify-support-commas			1247.11
>>>>
>>>> And if I remove this patch, degression is not happened:
>>>> [----------PATCH----------]                     [TBENCH]
>>>> lttng-kernel-trace-thread-flag-s390		1307.73
>>>> lttng-kernel-trace-thread-flag-sh		1317.33
>>>> lttng-kernel-trace-thread-flag-sparc		1306.85
>>>> lttng-kernel-trace-thread-flag-sparc64		1303.62
>>>> lttng-kernel-trace-thread-flag-um		1312.08
>>>> lttng-kernel-trace-thread-flag-x86		1308.5
>>>> lttng-kernel-trace-thread-flag-xtensa		1310.77
>>>> lttng-kernel-trace-thread-flag-api		1309.05
>>>> lttng-use-kernel-trace-thread-flag-api		1310.26
>>>> stringify-support-commas			1307.69
>>>>
>>>> I want to know system's benchmark when lttng turn off,
>>>> so test environment is following:
>>>> - Compile kernel with lttng patch up to given one and default config,
>>>> - Install kernel, Reboot system,
>>>>   (don't load any lttng module, don't run ltt-arm or lttctl)
>>>> - Begin tbench test.
>>>>
>>>> Plarform is RHEL5.2 root filesystem with linux-2.6.27.7 and lttng-0.60 patchs.
>>>> Hardware is 4 * Intel(R) Xeon(R) X5260 at 3.33GHz with 4G memory.
>>>>
>>>> I think [fix-x86_64-sysenter-trace-race] should not leads so large degression,
>>>> But I tested several times and result is near.
>>>>
>>> Hrm, interesting. Can you give your full kernel .config ? I'd like to
>>> see if the _TIF_SYSCALL_AUDIT or _TIF_SECCOMP bits are set.
>> Hello, Mathieu
>>
>> kernel config is send as attachment.
>> I use RHEL's default config with ltt support.
>>
> 
> OK, I see that you have 
> 
> CONFIG_AUDIT=y
> CONFIG_AUDITSYSCALL=y
> 
> (therefore _TIF_SYSCALL_AUDIT is probably on)
> 
> One way we could lessen the impact on your system when tracing is not
> active is to target more specifically the bit which should be tested at
> this second test location (when returning to userspace). Can you try
> changing :
> 
> sysret_careful:
>         testl $(_TIF_SYSCALL_TRACE|_TIF_KERNEL_TRACE|_TIF_SYSCALL_AUDIT|_TIF_SECCOMP),%edx
>         jnz ret_from_sys_call_trace
> 
> (this is after the lttng-instrumentation-x86.patch patch has been
> applied)
> 
> for
> 
> sysret_careful:
>         testl $_TIF_KERNEL_TRACE,%edx
>         jnz ret_from_sys_call_trace
> 
> The logic is that only the _TIF_KERNEL_TRACE thread flag needs to be
> tested at syscall return because it is the only flag which may be
> changed concurrently wrt thread execution.
> 
Hello, Mathieu

I finished this test, result is following:
	[lttng-instrumentation-x86.patc]	[only test _TIF_KERNEL_TRACE]
step1	1250.55					1301.41
step2	1255.47					1300.72

It seems degression is gone when remove _TIF_SYSCALL_TRACE|_TIF_SYSCALL_AUDIT|_TIF_SECCOMP check.

I'll check which flag (_TIF_SYSCALL_TRACE or _TIF_SYSCALL_AUDIT or _TIF_SECCOMP) cause degression.

B.R.
Zhaolei

> If we see that this helps, we might want to dig further and try to find
> a way that would help not having the runtime penality cost even when
> tracing is enabled.
> 
> Mathieu
> 
> 
>> B.R.
>> Zhaolei
>>
>>> If you have a better way to support asynchronously changing the thread
>>> flags and still have the affected running threads running correctly, I'm
>>> willing to consider it.
>>>
>>> Mathieu
>>>
>>>> B.R.
>>>> Zhaolei
>>>>
>>>>
>>>> _______________________________________________
>>>> ltt-dev mailing list
>>>> ltt-dev at lists.casi.polymtl.ca
>>>> http://lists.casi.polymtl.ca/cgi-bin/mailman/listinfo/ltt-dev
>>>>
> 
>> #
>> # Automatically generated make config: don't edit
>> # Linux kernel version: 2.6.27.7
>> # Fri Dec 12 20:32:16 2008
>> #
>> CONFIG_64BIT=y
>> # CONFIG_X86_32 is not set
>> CONFIG_X86_64=y
>> CONFIG_X86=y
>> CONFIG_ARCH_DEFCONFIG="arch/x86/configs/x86_64_defconfig"
>> # CONFIG_GENERIC_LOCKBREAK is not set
>> CONFIG_GENERIC_TIME=y
>> CONFIG_GENERIC_CMOS_UPDATE=y
>> CONFIG_CLOCKSOURCE_WATCHDOG=y