[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <078f662d-a73f-766b-3a07-c82cd37026c5@linux.dev>
Date: Mon, 9 Oct 2023 18:58:27 +0800
From: Yajun Deng <yajun.deng@...ux.dev>
To: Eric Dumazet <edumazet@...gle.com>
Cc: rostedt@...dmis.org, mhiramat@...nel.org, dennis@...nel.org,
tj@...nel.org, cl@...ux.com, mark.rutland@....com,
davem@...emloft.net, kuba@...nel.org, pabeni@...hat.com,
netdev@...r.kernel.org, linux-kernel@...r.kernel.org,
Alexander Lobakin <aleksander.lobakin@...el.com>,
linux-trace-kernel@...r.kernel.org, linux-mm@...ck.org
Subject: Re: [PATCH net-next v7] net/core: Introduce netdev_core_stats_inc()
On 2023/10/9 18:16, Eric Dumazet wrote:
> On Mon, Oct 9, 2023 at 11:43 AM Yajun Deng <yajun.deng@...ux.dev> wrote:
>>
>> On 2023/10/9 17:30, Eric Dumazet wrote:
>>> On Mon, Oct 9, 2023 at 10:36 AM Yajun Deng <yajun.deng@...ux.dev> wrote:
>>>> On 2023/10/9 16:20, Eric Dumazet wrote:
>>>>> On Mon, Oct 9, 2023 at 10:14 AM Yajun Deng <yajun.deng@...ux.dev> wrote:
>>>>>> On 2023/10/9 15:53, Eric Dumazet wrote:
>>>>>>> On Mon, Oct 9, 2023 at 5:07 AM Yajun Deng <yajun.deng@...ux.dev> wrote:
>>>>>>>
>>>>>>>> 'this_cpu_read + this_cpu_write' and 'pr_info + this_cpu_inc' will make
>>>>>>>> the trace work well.
>>>>>>>>
>>>>>>>> They all have 'pop' instructions in them. This may be the key to making
>>>>>>>> the trace work well.
>>>>>>>>
>>>>>>>> Hi all,
>>>>>>>>
>>>>>>>> I need your help on percpu and ftrace.
>>>>>>>>
>>>>>>> I do not think you made sure netdev_core_stats_inc() was never inlined.
>>>>>>>
>>>>>>> Adding more code in it is simply changing how the compiler decides to
>>>>>>> inline or not.
>>>>>> Yes, you are right. It needs to add the 'noinline' prefix. The
>>>>>> disassembly code will have 'pop'
>>>>>>
>>>>>> instruction.
>>>>>>
>>>>> The function was fine, you do not need anything like push or pop.
>>>>>
>>>>> The only needed stuff was the call __fentry__.
>>>>>
>>>>> The fact that the function was inlined for some invocations was the
>>>>> issue, because the trace point
>>>>> is only planted in the out of line function.
>>>> But somehow the following code isn't inline? They didn't need to add the
>>>> 'noinline' prefix.
>>>>
>>>> + field = (unsigned long *)((void *)this_cpu_ptr(p) + offset);
>>>> + WRITE_ONCE(*field, READ_ONCE(*field) + 1);
>>>>
>>>> Or
>>>> + (*(unsigned long *)((void *)this_cpu_ptr(p) + offset))++;
>>>>
>>> I think you are very confused.
>>>
>>> You only want to trace netdev_core_stats_inc() entry point, not
>>> arbitrary pieces of it.
>>
>> Yes, I will trace netdev_core_stats_inc() entry point. I mean to replace
>>
>> + field = (__force unsigned long
>> __percpu *)((__force void *)p + offset);
>> + this_cpu_inc(*field);
>>
>> with
>>
>> + field = (unsigned long *)((void *)this_cpu_ptr(p) + offset);
>> + WRITE_ONCE(*field, READ_ONCE(*field) + 1);
>>
>> Or
>> + (*(unsigned long *)((void *)this_cpu_ptr(p) + offset))++;
>>
>> The netdev_core_stats_inc() entry point will work fine even if it doesn't
>> have 'noinline' prefix.
>>
>> I don't know why this code needs to add 'noinline' prefix.
>> + field = (__force unsigned long __percpu *)((__force void *)p + offset);
>> + this_cpu_inc(*field);
>>
> C compiler decides to inline or not, depending on various factors.
>
> The most efficient (and small) code is generated by this_cpu_inc()
> version, allowing the compiler to inline it.
>
> If you copy/paste this_cpu_inc() twenty times, then the compiler
> would not inline the function anymore.
Got it. Thank you.
Powered by blists - more mailing lists