Re: [PATCH bpf-next 09/13] bpf: Add BPF_FUNC_jiffies

From: Eric Dumazet <eric.dumazet@gmail.com>
To: Martin Lau <kafai@fb.com>, Eric Dumazet <eric.dumazet@gmail.com>
Cc: "bpf@vger.kernel.org" <bpf@vger.kernel.org>,
	Alexei Starovoitov <ast@kernel.org>,
	Daniel Borkmann <daniel@iogearbox.net>,
	David Miller <davem@davemloft.net>,
	Kernel Team <Kernel-team@fb.com>,
	"netdev@vger.kernel.org" <netdev@vger.kernel.org>
Subject: Re: [PATCH bpf-next 09/13] bpf: Add BPF_FUNC_jiffies
Date: Mon, 16 Dec 2019 11:33:01 -0800	[thread overview]
Message-ID: <a2d8d888-16e1-243d-92c9-56ba3a3e1b18@gmail.com> (raw)
In-Reply-To: <20191216191357.ftadvchztbpggcus@kafai-mbp.dhcp.thefacebook.com>

On 12/16/19 11:14 AM, Martin Lau wrote:
> On Fri, Dec 13, 2019 at 05:59:54PM -0800, Eric Dumazet wrote:
>>
>>
>> On 12/13/19 4:47 PM, Martin KaFai Lau wrote:
>>> This patch adds a helper to handle jiffies.  Some of the
>>> tcp_sock's timing is stored in jiffies.  Although things
>>> could be deduced by CONFIG_HZ, having an easy way to get
>>> jiffies will make the later bpf-tcp-cc implementation easier.
>>>
>>
>> ...
>>
>>> +
>>> +BPF_CALL_2(bpf_jiffies, u64, in, u64, flags)
>>> +{
>>> +	if (!flags)
>>> +		return get_jiffies_64();
>>> +
>>> +	if (flags & BPF_F_NS_TO_JIFFIES) {
>>> +		return nsecs_to_jiffies(in);
>>> +	} else if (flags & BPF_F_JIFFIES_TO_NS) {
>>> +		if (!in)
>>> +			in = get_jiffies_64();
>>> +		return jiffies_to_nsecs(in);
>>> +	}
>>> +
>>> +	return 0;
>>> +}
>>
>> This looks a bit convoluted :)
>>
>> Note that we could possibly change net/ipv4/tcp_cubic.c to no longer use jiffies at all.
>>
>> We have in tp->tcp_mstamp an accurate timestamp (in usec) that can be converted to ms.
> Thanks for the feedbacks!
> 
> I have a few questions that need some helps.
> 
> Does it mean tp->tcp_mstamp can be used as the "now" in cubic?

TCP makes sure to update tp->tcp_mstamp at least once when handling
a particular packet. We did that to avoid calling possibly expensive
kernel time service (Some platforms do not have fast TSC) 

> or tcp_clock_ns() should still be called in cubic, e.g. to replace
> bictcp_clock() and tcp_jiffies32?

Yeah, there is this lsndtime and tcp_jiffies32 thing, but maybe
we can find a way to fetch jiffies32 without having to call a bpf helper ?

> 
> BPF currently has a helper calling ktime_get_mono_fast_ns() which looks
> different from tcp_clock_ns().

That's maybe because of NMI requirements. 

TCP in the other hand is in process or BH context.

But it should not matter, cubic could should not have to call them.