Message-ID: <672269c08bcd5_3c834029423@willemb.c.googlers.com.notmuch>
Date: Wed, 30 Oct 2024 13:15:44 -0400
From: Willem de Bruijn <willemdebruijn.kernel@...il.com>
To: Jason Xing <kerneljasonxing@...il.com>, 
 Martin KaFai Lau <martin.lau@...ux.dev>
Cc: Willem de Bruijn <willemdebruijn.kernel@...il.com>, 
 willemb@...gle.com, 
 davem@...emloft.net, 
 edumazet@...gle.com, 
 kuba@...nel.org, 
 pabeni@...hat.com, 
 dsahern@...nel.org, 
 ast@...nel.org, 
 daniel@...earbox.net, 
 andrii@...nel.org, 
 eddyz87@...il.com, 
 song@...nel.org, 
 yonghong.song@...ux.dev, 
 john.fastabend@...il.com, 
 kpsingh@...nel.org, 
 sdf@...ichev.me, 
 haoluo@...gle.com, 
 jolsa@...nel.org, 
 shuah@...nel.org, 
 ykolal@...com, 
 bpf@...r.kernel.org, 
 netdev@...r.kernel.org, 
 Jason Xing <kernelxing@...cent.com>
Subject: Re: [PATCH net-next v3 02/14] net-timestamp: allow two features to
 work parallelly

Jason Xing wrote:
> On Wed, Oct 30, 2024 at 1:37 PM Martin KaFai Lau <martin.lau@...ux.dev> wrote:
> >
> > On 10/29/24 8:04 PM, Jason Xing wrote:
> > >>>>>>>    static void skb_tstamp_tx_output(struct sk_buff *orig_skb,
> > >>>>>>>                                 const struct sk_buff *ack_skb,
> > >>>>>>>                                 struct skb_shared_hwtstamps *hwtstamps,
> > >>>>>>> @@ -5549,6 +5575,9 @@ static void skb_tstamp_tx_output(struct sk_buff *orig_skb,
> > >>>>>>>        u32 tsflags;
> > >>>>>>>
> > >>>>>>>        tsflags = READ_ONCE(sk->sk_tsflags);
> > >>>>>>> +     if (!sk_tstamp_tx_flags(sk, tsflags, tstype))
> > >>>>>>
> > >>>>>> I still don't get this part since v2. How does it work with cmsg only
> > >>>>>> SOF_TIMESTAMPING_TX_*?
> > >>>>>>
> > >>>>>> I tried with "./txtimestamp -6 -c 1 -C -N -L ::1" and it does not return any tx
> > >>>>>> time stamp after this patch.
> > >>>>>>
> > >>>>>> I am likely missing something
> > >>>>>> or v2 concluded that this behavior change is acceptable?
> > >>>>>
> > >>>>> Sorry, I submitted this series accidentally removing one important
> > >>>>> thing which is similar to what Vadim Fedorenko mentioned in the v1
> > >>>>> [1]:
> > >>>>> adding another member like sk_flags_bpf to handle the cmsg case.
> > >>>>>
> > >>>>> Willem, would it be acceptable to add another field in struct sock to
> > >>>>> help us recognise the case where BPF and cmsg works parallelly?
> > >>>>>
> > >>>>> [1]: https://lore.kernel.org/all/662873cb-a897-464e-bdb3-edf01363c3b2@linux.dev/
> > >>>>
> > >>>> The current timestamp flags don't need a u32. Maybe just reserve a bit
> > >>>> for this purpose?
> > >>>
> > >>> Sure. Good suggestion.
> > >>>
> > >>> But I think only using one bit to reflect whether the sk->sk_tsflags
> > >>> is used by normal or cmsg features is not enough. The existing
> > >>> implementation in tcp_sendmsg_locked() doesn't override the
> > >>> sk->sk_tsflags even the normal and cmsg features enabled parallelly.
> > >>> It only overrides sockc.tsflags in tcp_sendmsg_locked(). Based on
> > >>> that, even if at some point users suddenly remove the cmsg use and
> > >>> then the prior normal SO_TIMESTAMPING continues to work.
> > >>>
> > >>> How about this, please see below:
> > >>> For now, sk->sk_tsflags only uses 17 bits (see the last one
> > >>> SOF_TIMESTAMPING_OPT_RX_FILTER). The cmsg feature only uses 4 flags
> > >>> (see SOF_TIMESTAMPING_TX_RECORD_MASK in __sock_cmsg_send()). With that
> > >>> said, we could reserve the highest four bits for cmsg use for the
> > >>> moment. Four bits represents four points where we can record the
> > >>> timestamp in the tx case.
> > >>>
> > >>> Do you agree on this point?
> > >>
> > >> I don't follow.
> > >>
> > >> I probably miss the entire point.
> > >>
> > >> The goal for sockcm fields is to start with the sk field and
> > >> optionally override based on cmsg. This is what sockcm_init does for
> > >> tsflags.
> > >>
> > >> This information is for the skb, so these are recording flags.
> > >>
> > >> Why does the new datapath need to know whether features are enabled
> > >> through setsockopt or on a per-call basis with a cmsg?
> > >>
> > >> The goal was always to keep the reporting flags per socket, but make
> > >> the recording flag per packet, mainly for sampling.
> > >
> > > If a user uses 1) cmsg feature, 2) bpf feature at the same time, we
> > > allow each feature to work independently.
> > >
> > > How could it work? It relies on sk_tstamp_tx_flags() function in the
> > > current patch: when we are in __skb_tstamp_tx(), we cannot know which
> > > flags in each feature are set without fetching sk->sk_tsflags and
> > > sk->sk_tsflags_bpf. Then we are able to know what timestamp we want to
> > > record. To put it in a simple way, we're not sure if the user wants to
> > > see a SCHED timestamp by using the cmsg feature in __skb_tstamp_tx()
> > > if we hit this test statement "skb_shinfo(skb)->tx_flags &
> > > SKBTX_SCHED_TSTAMP)". So we need those two socket tsflag fields to
> > > help us.
> >
> > I also don't see how a new bit/integer in a sk can help to tell the per cmsg
> > on/off. This cmsg may have tx timestamp on while the next cmsg can have it off.
> 
> It's not hard to use it because we can clear every socket cmsg tsflags
> when we're done the check in tcp_sendmsg_locked() if the cmsg feature
> is not enabled. Then we can accurately know which timestamp should we
> print in the tx path.
> 
> >
> > There is still one bit in skb_shinfo(skb)->tx_flags. How about define a
> > SKBTX_BPF for everything. imo, the fine control on
> > SOF_TIMESTAMPING_TX_{SCHED,SOFTWARE} is not useful for bpf. Almost all of the
> > time the bpf program wants all available time stamps (sched, software, and
> > hwtstamp if the NIC has it).

I like the approach of just calling BPF on every hook. Assuming that
the call is very cheap, which AFAIK is true.

In that case we don't need complex branching in C to optionally skip
this step, as we do for reporting to userspace.

All the logic and complexity is in the BPF program itself.

We obviously then let go of the goal to model the BPF API close to the
existing SO_TIMESTAMPING API. Though I advocated for keeping them
aligned, I also think we should just tailor it to what makes most
sense in the BPF space.
 
> Sorry, I really doubt that we can lose the fine control. 

Since BPF is called at each reporting point, no control is lost,
actually.

> I still
> reckon that providing more options to users is a good way to go,
> especially for some latency sensitive applications, enabling one or
> two or three tx flags could lead to different performances. For the
> users of SO_TIMESTAMPING, they use the feature very differently. Not
> all users prefer to record everything.
> 
> > Since bpf is in the kernel, it is much cheaper
> > because it does not need to do skb_alloc/clone and queue to the error queue.
> >
> > I think the bpf prog needs to capture a timestamp at the sendmsg() time, so a
> > bpf prog needs to be called at sendmsg().
> 
> Agreed, I planned to implement this after this series.
> 
> > Then it may as well allow the bpf
> > prog@...dmsg() to decide if it needs to set the SKBTX_BPF bit in
> > skb_shinfo(skb)->tx_flags or not.
> >
> > TCP_SKB_CB(skb)->txstamp_ack can also work similarly. There is still unused bit
> > in "struct tcp_skb_cb", so may be adding TCP_SKB_CB(skb)->bpf_txstamp_ack
> >
> > Then there is no need to control SOF_TIMESTAMPING_TX_* through bpf_setsockopt().
> > It only needs one bpf specific socket option like bpf_setsockopt(SOL_SOCKET,
> > BPF_TX_TIMESTAMPING) to guard if the bpf-prog@...dmsg() needs to be called or
> > not. There are already other TCP_BPF_IW,TCP_BPF_SNDCWND_CLAMP,... specific
> > socket options.
> >
> > imo, this is a simpler interface and also gives the bpf prog per packet control
> > at the same time.
> 
> Very interesting idea, but the precondition is that we give up the
> fine control...
> 
> Thanks,
> Jason

Powered by blists - more mailing lists