Toke Høiland-Jørgensen wrote:
> Martin KaFai Lau <martin.lau@xxxxxxxxx> writes:
>
> > On 11/15/22 10:38 PM, John Fastabend wrote:
> >>>>>>> +static void veth_unroll_kfunc(const struct bpf_prog *prog, u32 func_id,
> >>>>>>> + struct bpf_patch *patch)
> >>>>>>> +{
> >>>>>>> + if (func_id == xdp_metadata_kfunc_id(XDP_METADATA_KFUNC_RX_TIMESTAMP_SUPPORTED)) {
> >>>>>>> + /* return true; */
> >>>>>>> + bpf_patch_append(patch, BPF_MOV64_IMM(BPF_REG_0, 1));
> >>>>>>> + } else if (func_id == xdp_metadata_kfunc_id(XDP_METADATA_KFUNC_RX_TIMESTAMP)) {
> >>>>>>> + /* return ktime_get_mono_fast_ns(); */
> >>>>>>> + bpf_patch_append(patch, BPF_EMIT_CALL(ktime_get_mono_fast_ns));
> >>>>>>> + }
> >>>>>>> +}
> >>>>>>
> >>>>>> So these look reasonable enough, but would be good to see some examples
> >>>>>> of kfunc implementations that don't just BPF_CALL to a kernel function
> >>>>>> (with those helper wrappers we were discussing before).
> >>>>>
> >>>>> Let's maybe add them if/when needed as we add more metadata support?
> >>>>> xdp_metadata_export_to_skb has an example, and rfc 1/2 have more
> >>>>> examples, so it shouldn't be a problem to resurrect them back at some
> >>>>> point?
> >>>>
> >>>> Well, the reason I asked for them is that I think having to maintain the
> >>>> BPF code generation in the drivers is probably the biggest drawback of
> >>>> the kfunc approach, so it would be good to be relatively sure that we
> >>>> can manage that complexity (via helpers) before we commit to this :)
> >>>
> >>> Right, and I've added a bunch of examples in v2 rfc so we can judge
> >>> whether that complexity is manageable or not :-)
> >>> Do you want me to add those wrappers you've back without any real users?
> >>> Because I had to remove my veth tstamp accessors due to John/Jesper
> >>> objections; I can maybe bring some of this back gated by some
> >>> static_branch to avoid the fastpath cost?
> >>
> >> I missed the context a bit what did you mean "would be good to see some
> >> examples of kfunc implementations that don't just BPF_CALL to a kernel
> >> function"? In this case do you mean BPF code directly without the call?
> >>
> >> Early on I thought we should just expose the rx_descriptor which would
> >> be roughly the same right? (difference being code embedded in driver vs
> >> a lib) Trouble I ran into is driver code using seqlock_t and mutexs
> >> which wasn't as straight forward as the simpler just read it from
> >> the descriptor. For example in mlx getting the ts would be easy from
> >> BPF with the mlx4_cqe struct exposed
> >>
> >> u64 mlx4_en_get_cqe_ts(struct mlx4_cqe *cqe)
> >> {
> >> u64 hi, lo;
> >> struct mlx4_ts_cqe *ts_cqe = (struct mlx4_ts_cqe *)cqe;
> >>
> >> lo = (u64)be16_to_cpu(ts_cqe->timestamp_lo);
> >> hi = ((u64)be32_to_cpu(ts_cqe->timestamp_hi) + !lo) << 16;
> >>
> >> return hi | lo;
> >> }
> >>
> >> but converting that to nsec is a bit annoying,
> >>
> >> void mlx4_en_fill_hwtstamps(struct mlx4_en_dev *mdev,
> >> struct skb_shared_hwtstamps *hwts,
> >> u64 timestamp)
> >> {
> >> unsigned int seq;
> >> u64 nsec;
> >>
> >> do {
> >> seq = read_seqbegin(&mdev->clock_lock);
> >> nsec = timecounter_cyc2time(&mdev->clock, timestamp);
> >> } while (read_seqretry(&mdev->clock_lock, seq));
> >>
> >> memset(hwts, 0, sizeof(struct skb_shared_hwtstamps));
> >> hwts->hwtstamp = ns_to_ktime(nsec);
> >> }
> >>
> >> I think the nsec is what you really want.
> >>
> >> With all the drivers doing slightly different ops we would have
> >> to create read_seqbegin, read_seqretry, mutex_lock, ... to get
> >> at least the mlx and ice drivers it looks like we would need some
> >> more BPF primitives/helpers. Looks like some more work is needed
> >> on ice driver though to get rx tstamps on all packets.
> >>
> >> Anyways this convinced me real devices will probably use BPF_CALL
> >> and not BPF insns directly.
> >
> > Some of the mlx5 path looks like this:
> >
> > #define REAL_TIME_TO_NS(hi, low) (((u64)hi) * NSEC_PER_SEC + ((u64)low))
> >
> > static inline ktime_t mlx5_real_time_cyc2time(struct mlx5_clock *clock,
> > u64 timestamp)
> > {
> > u64 time = REAL_TIME_TO_NS(timestamp >> 32, timestamp & 0xFFFFFFFF);
> >
> > return ns_to_ktime(time);
> > }
> >
> > If some hints are harder to get, then just doing a kfunc call is better.
>
> Sure, but if we end up having a full function call for every field in
> the metadata, that will end up having a significant performance impact
> on the XDP data path (thinking mostly about the skb metadata case here,
> which will collect several bits of metadata).
>
> > csum may have a better chance to inline?
>
> Yup, I agree. Including that also makes it possible to benchmark this
> series against Jesper's; which I think we should definitely be doing
> before merging this.
Good point I got sort of singularly focused on timestamp because I have
a use case for it now.
Also hash is often sitting in the rx descriptor.
>
> > Regardless, BPF in-lining is a well solved problem and used in many
> > bpf helpers already, so there are many examples in the kernel. I don't
> > think it is necessary to block this series because of missing some
> > helper wrappers for inlining. The driver can always start with the
> > simpler kfunc call first and optimize later if some hints from the
> > drivers allow it.
>
> Well, "solved" in the sense of "there are a few handfuls of core BPF
> people who know how to do it". My concern is that we'll end up with
> either the BPF devs having to maintain all these bits of BPF byte code
> in all the drivers; or drivers just punting to regular function calls
> because the inlining is too complicated, with sub-par performance as per
> the above. I don't think we should just hand-wave this away as "solved",
> but rather treat this as an integral part of the initial series.
This was my motivation for pushing the rx_descriptor into the xdp_buff.
At this point if I'm going to have a kfunc call into the driver and
have the driver rewrite the code into some BPF instructions I would
just assume maintain this as a library code where I can hook it
into my BPF program directly from user space. Maybe a few drivers
will support all the things I want to read, but we run on lots of
hardware (mlx, intel, eks, azure, gke, etc) and its been a lot of work
to just get the basic feature parity. I also don't want to run around
writing driver code for each vendor if I can avoid it. Having raw
access to the rx descriptor gets me the escape hatch where I can
just do it myself. And the last piece of info from my point of view
(Tetragon, Cilium) I can run whatever libs I want and freely upgrade
libbpf and cilium/ebpf but have a lot less ability to get users
to upgrade kernels outside the LTS they picked. Meaning I can
add new things much easier if its lifted into BPF code placed
by user space.
I appreciate that it means I import the problem of hardware detection
and BTF CO-RE on networking codes, but we've already solved these
problems for other reasons. For example just configuring the timestamp
is a bit of an exercise in does my hardware support timestamping
and does it support timestamping the packets I care about, e.g.
all pkts, just ptp pkts, etc.
I don't think they are mutual exclusive with this series though
because I can't see how to write these timestamping logic directly
in BPF. But for rxhash and csum it seems doable. My preference
is to have both the kfuncs and expose the descriptor directly.
.John
