[xdp-hints] Re: [PATCH bpf-next 05/11] veth: Support rx timestamp metadata for xdp

[John Fastabend <john.fastabend@gmail.com>]

Stanislav Fomichev wrote:
> On Wed, Nov 16, 2022 at 6:59 PM Alexei Starovoitov
> <alexei.starovoitov@gmail.com> wrote:
> >
> > On Wed, Nov 16, 2022 at 6:53 PM Stanislav Fomichev <sdf@google.com> wrote:
> > >
> > > On Wed, Nov 16, 2022 at 6:17 PM Alexei Starovoitov
> > > <alexei.starovoitov@gmail.com> wrote:
> > > >
> > > > On Wed, Nov 16, 2022 at 4:19 PM Stanislav Fomichev <sdf@google.com> wrote:
> > > > >
> > > > > On Wed, Nov 16, 2022 at 3:47 PM John Fastabend <john.fastabend@gmail.com> wrote:
> > > > > >
> > > > > > Stanislav Fomichev wrote:
> > > > > > > On Wed, Nov 16, 2022 at 11:03 AM John Fastabend
> > > > > > > <john.fastabend@gmail.com> wrote:
> > > > > > > >
> > > > > > > > Toke Høiland-Jørgensen wrote:
> > > > > > > > > Martin KaFai Lau <martin.lau@linux.dev> 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.
> > > > > > >
> > > > > > > Ack, let me try to add something else (that's more inline-able) on the
> > > > > > > rx side for a v2.
> > > > > >
> > > > > > If you go with in-kernel BPF kfunc approach (vs user space side) I think
> > > > > > you also need to add CO-RE to be friendly for driver developers? Otherwise
> > > > > > they have to keep that read in sync with the descriptors? Also need to
> > > > > > handle versioning of descriptors where depending on specific options
> > > > > > and firmware and chip being enabled the descriptor might be moving
> > > > > > around. Of course can push this all to developer, but seems not so
> > > > > > nice when we have the machinery to do this and we handle it for all
> > > > > > other structures.
> > > > > >
> > > > > > With CO-RE you can simply do the rx_desc->hash and rx_desc->csum and
> > > > > > expect CO-RE sorts it out based on currently running btf_id of the
> > > > > > descriptor. If you go through normal path you get this for free of
> > > > > > course.
> > > > >
> > > > > Doesn't look like the descriptors are as nice as you're trying to
> > > > > paint them (with clear hash/csum fields) :-) So not sure how much
> > > > > CO-RE would help.
> > > > > At least looking at mlx4 rx_csum, the driver consults three different
> > > > > sets of flags to figure out the hash_type. Or am I just unlucky with
> > > > > mlx4?
> > > >
> > > > Which part are you talking about ?
> > > >         hw_checksum = csum_unfold((__force __sum16)cqe->checksum);
> > > > is trivial enough for bpf prog to do if it has access to 'cqe' pointer
> > > > which is what John is proposing (I think).

Yeah this is what I've been considering. If you just get the 'cqe' pointer
walking the check_sum and rxhash should be straightforward.

There seems to be a real difference between timestamps and most other 
fields IMO. Timestamps require some extra logic to turn into ns when
using the NIC hw clock. And the hw clock requires some coordination to
keep in sync and stop from overflowing and may be done through other
protocols like PTP in my use case. In some cases I think the clock is
also shared amongst multiple phys. Seems mlx has a seqlock_t to protect
it and I'm less sure about this but seems intel nic maybe has a sideband
control channel.

Then there is everything else that I can think up that is per packet data
and requires no coordination with the driver. Its just reading fields in
the completion queue. This would be the csum, check_sum, vlan_header and
so on. Sure we could kfunc each one of those things, but could also just
write that directly in BPF and remove some abstractions and kernel
dependency by doing it directly in the BPF program. If you like that
abstraction seems to be the point of contention my opinion is the cost
of kernel depency is high and I can abstract it with a user library
anyways so burying it in the kernel only causes me support issues and
backwards compat problems.

Hopefully, my position is more clear.

> > >
> > > I'm talking about mlx4_en_process_rx_cq, the caller of that check_csum.
> > > In particular: if (likely(dev->features & NETIF_F_RXCSUM)) branch
> > > I'm assuming we want to have hash_type available to the progs?
> > >
> > > But also, check_csum handles other corner cases:
> > > - short_frame: we simply force all those small frames to skip checksum complete
> > > - get_fixed_ipv6_csum: In IPv6 packets, hw_checksum lacks 6 bytes from
> > > IPv6 header
> > > - get_fixed_ipv4_csum: Although the stack expects checksum which
> > > doesn't include the pseudo header, the HW adds it
> >
> > Of course, but kfunc won't be doing them either.
> > We're talking XDP fast path.
> > The mlx4 hw is old and incapable.
> > No amount of sw can help.

Doesn't this lend itself to letting the XDP BPF program write the
BPF code to read it out. Maybe someone cares about these details
for some cpumap thing, but the rest of us wont care we might just
want to read check_csum. Maybe we have an IPv6 only network or
IPv4 network so can make further shortcuts. If a driver dev does
this they will be forced to do the cactch all version because
they have no way to know such details.

> >
> > > So it doesn't look like we can just unconditionally use cqe->checksum?
> > > The driver does a lot of massaging around that field to make it
> > > palatable.
> >
> > Of course we can. cqe->checksum is still usable. the bpf prog
> > would need to know what it's reading.
> > There should be no attempt to present a unified state of hw bits.
> > That's what skb is for. XDP layer should not hide such hw details.
> > Otherwise it will become a mini skb layer with all that overhead.
> 
> I was hoping the kfunc could at least parse the flags and return some
> pkt_hash_types-like enum to indicate what this csum covers.
> So the users won't have to find the hardware manuals (not sure they
> are even available?) to decipher what numbers they've got.
> Regarding old mlx4: true, but mlx5's mlx5e_handle_csum doesn't look
> that much different :-(

The driver developers could still provide and ship the BPF libs
with their drivers. I think if someone is going to use their NIC
and lots of them and requires XDP it will get done. We could put
them by the driver code mlx4.bpf or something.

> 
> But going back a bit: I'm probably missing what John has been
> suggesting. How is CO-RE relevant for kfuncs? kfuncs are already doing
> a CO-RE-like functionality by rewriting some "public api" (kfunc) into
> the bytecode to access the relevant data.

This was maybe a bit of an aside. What I was pondering a bit out
loud perhaps is my recollection is there may be a few different
descriptor layouts depending features enabled, exact device loaded
and such. So in this case if I was a driver writer I might not want
to hardcode the offset of the check_sum field. If I could use CO-RE
then I don't have to care if in one version is the Nth field and later on
someone makes it the Mth field just like any normal kernel struct.
But through the kfunc interface I couldn't see how to get that.
So instead of having a bunch of kfunc implementations you could just
have one for all your device classes because you always name the
field the same thing.