Re: [PATCH bpf-next] selftests/bpf: Switch test_vmlinux to use hrtimer_range_start_ns.

From: Yonghong Song <yhs@fb.com>
To: Hao Luo <haoluo@google.com>
Cc: Networking <netdev@vger.kernel.org>, bpf <bpf@vger.kernel.org>,
	<linux-kernel@vger.kernel.org>,
	<clang-built-linux@googlegroups.com>,
	<linux-kselftest@vger.kernel.org>,
	Stanislav Fomichev <sdf@google.com>,
	Shuah Khan <shuah@kernel.org>,
	Alexei Starovoitov <ast@kernel.org>,
	Andrii Nakryiko <andriin@fb.com>,
	Daniel Borkmann <daniel@iogearbox.net>,
	Martin KaFai Lau <kafai@fb.com>, Song Liu <songliubraving@fb.com>,
	John Fastabend <john.fastabend@gmail.com>,
	KP Singh <kpsingh@chromium.org>, Bill Wendling <morbo@google.com>
Subject: Re: [PATCH bpf-next] selftests/bpf: Switch test_vmlinux to use hrtimer_range_start_ns.
Date: Tue, 30 Jun 2020 19:26:02 -0700	[thread overview]
Message-ID: <46fc8e13-fb3e-6464-b794-60cf90d16543@fb.com> (raw)
In-Reply-To: <CA+khW7jNqVMqq2dzf6Dy0pWCZYjHrG7Vm_sUEKKLS-L-ptzEtQ@mail.gmail.com>

On 6/30/20 5:10 PM, Hao Luo wrote:
> Ok, with the help of my colleague Ian Rogers, I think we solved the
> mystery. Clang actually inlined hrtimer_nanosleep() inside
> SyS_nanosleep(), so there is no call to that function throughout the
> path of the nanosleep syscall. I've been looking at the function body
> of hrtimer_nanosleep for quite some time, but clearly overlooked the
> caller of hrtimer_nanosleep. hrtimer_nanosleep is pretty short and
> there are many constants, inlining would not be too surprising.

Oh thanks for explanation. inlining makes sense. We have many other
instances like this in the past where kprobe won't work properly.

Could you reword your commit message then?

 > causing fentry and kprobe to not hook on this function properly on a
 > Clang build kernel.

The above is a little vague on what happens. What really happens is
fentry/kprobe does hook on this function but has no effect since
its caller has inlined the function.

> Sigh...
> 
> Hao
> 
> On Tue, Jun 30, 2020 at 3:48 PM Hao Luo <haoluo@google.com> wrote:
>>
>> On Tue, Jun 30, 2020 at 1:37 PM Yonghong Song <yhs@fb.com> wrote:
>>>
>>> On 6/30/20 11:49 AM, Hao Luo wrote:
>>>> The test_vmlinux test uses hrtimer_nanosleep as hook to test tracing
>>>> programs. But it seems Clang may have done an aggressive optimization,
>>>> causing fentry and kprobe to not hook on this function properly on a
>>>> Clang build kernel.
>>>
>>> Could you explain why it does not on clang built kernel? How did you
>>> build the kernel? Did you use [thin]lto?
>>>
>>> hrtimer_nanosleep is a global function who is called in several
>>> different files. I am curious how clang optimization can make
>>> function disappear, or make its function signature change, or
>>> rename the function?
>>>
>>
>> Yonghong,
>>
>> We didn't enable LTO. It also puzzled me. But I can confirm those
>> fentry/kprobe test failures via many different experiments I've done.
>> After talking to my colleague on kernel compiling tools (Bill, cc'ed),
>> we suspected this could be because of clang's aggressive inlining. We
>> also noticed that all the callsites of hrtimer_nanosleep() are tail
>> calls.
>>
>> For a better explanation, I can reach out to the people who are more
>> familiar to clang in the compiler team to see if they have any
>> insights. This may not be of high priority for them though.
>>
>> Hao