[xdp-hints] Re: [PATCH bpf-next v2 2/8] bpf: XDP metadata RX kfuncs

[Stanislav Fomichev <sdf@google.com>]

On Wed, Nov 30, 2022 at 11:06 AM Stanislav Fomichev <sdf@google.com> wrote:
>
> On Wed, Nov 30, 2022 at 9:38 AM Larysa Zaremba <larysa.zaremba@intel.com> wrote:
> >
> > On Mon, Nov 21, 2022 at 10:25:46AM -0800, Stanislav Fomichev wrote:
> >
> > > diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> > > index 9528a066cfa5..315876fa9d30 100644
> > > --- a/kernel/bpf/verifier.c
> > > +++ b/kernel/bpf/verifier.c
> > > @@ -15171,6 +15171,25 @@ static int fixup_call_args(struct bpf_verifier_env *env)
> > >       return err;
> > >  }
> > >
> > > +static int fixup_xdp_kfunc_call(struct bpf_verifier_env *env, u32 func_id)
> > > +{
> > > +     struct bpf_prog_aux *aux = env->prog->aux;
> > > +     void *resolved = NULL;
> >
> > First I would like to say I really like the kfunc hints impementation.
> >
> > I am currently trying to test possible performace benefits of the unrolled
> > version in the ice driver. I was working on top of the RFC v2,
> > when I noticed a problem that also persists in this newer version.
> >
> > For debugging purposes, I have put the following logs in this place in code.
> >
> > printk(KERN_ERR "func_id=%u\n", func_id);
> > printk(KERN_ERR "XDP_METADATA_KFUNC_RX_TIMESTAMP_SUPPORTED=%u\n",
> >        xdp_metadata_kfunc_id(XDP_METADATA_KFUNC_RX_TIMESTAMP_SUPPORTED));
> > printk(KERN_ERR "XDP_METADATA_KFUNC_RX_TIMESTAMP=%u\n",
> >        xdp_metadata_kfunc_id(XDP_METADATA_KFUNC_RX_TIMESTAMP));
> > printk(KERN_ERR "XDP_METADATA_KFUNC_RX_HASH_SUPPORTED=%u\n",
> >        xdp_metadata_kfunc_id(XDP_METADATA_KFUNC_RX_HASH_SUPPORTED));
> > printk(KERN_ERR "XDP_METADATA_KFUNC_RX_HASH=%u\n",
> >        xdp_metadata_kfunc_id(XDP_METADATA_KFUNC_RX_HASH));
> >
> > Loading the program, which uses bpf_xdp_metadata_rx_timestamp_supported()
> > and bpf_xdp_metadata_rx_timestamp(), has resulted in such messages:
> >
> > [  412.611888] func_id=108131
> > [  412.611891] XDP_METADATA_KFUNC_RX_TIMESTAMP_SUPPORTED=108126
> > [  412.611892] XDP_METADATA_KFUNC_RX_TIMESTAMP=108128
> > [  412.611892] XDP_METADATA_KFUNC_RX_HASH_SUPPORTED=108130
> > [  412.611893] XDP_METADATA_KFUNC_RX_HASH=108131
> > [  412.611894] func_id=108130
> > [  412.611894] XDP_METADATA_KFUNC_RX_TIMESTAMP_SUPPORTED=108126
> > [  412.611895] XDP_METADATA_KFUNC_RX_TIMESTAMP=108128
> > [  412.611895] XDP_METADATA_KFUNC_RX_HASH_SUPPORTED=108130
> > [  412.611895] XDP_METADATA_KFUNC_RX_HASH=108131
> >
> > As you can see, I've got 108131 and 108130 IDs in program,
> > while 108126 and 108128 would be more reasonable.
> > It's hard to proceed with the implementation, when IDs cannot be sustainably
> > compared.
>
> Thanks for the report!
> Toke has reported a similar issue in [0], have you tried his patch?
> I've also tried to address it in v3 [1], could you retry on top of it?
> I'll try to insert your printk in my local build to see what happens
> with btf ids on my side. Will get back to you..
>
> 0: https://lore.kernel.org/bpf/87mt8e2a69.fsf@toke.dk/
> 1: https://lore.kernel.org/bpf/20221129193452.3448944-3-sdf@google.com/T/#u

Nope, even if I go back to v2, I still can't reproduce locally.
Somehow in my setup they are sorted properly :-/
Would appreciate it if you can test the v3 patch and confirm whether
it's fixed on your side or not.

> > Furthermore, dumped vmlinux BTF shows the IDs is in the exactly reversed
> > order:
> >
> > [108126] FUNC 'bpf_xdp_metadata_rx_hash' type_id=108125 linkage=static
> > [108128] FUNC 'bpf_xdp_metadata_rx_hash_supported' type_id=108127 linkage=static
> > [108130] FUNC 'bpf_xdp_metadata_rx_timestamp' type_id=108129 linkage=static
> > [108131] FUNC 'bpf_xdp_metadata_rx_timestamp_supported' type_id=108127 linkage=static
> >
> > > +
> > > +     if (func_id == xdp_metadata_kfunc_id(XDP_METADATA_KFUNC_RX_TIMESTAMP_SUPPORTED))
> > > +             resolved = aux->xdp_netdev->netdev_ops->ndo_xdp_rx_timestamp_supported;
> > > +     else if (func_id == xdp_metadata_kfunc_id(XDP_METADATA_KFUNC_RX_TIMESTAMP))
> > > +             resolved = aux->xdp_netdev->netdev_ops->ndo_xdp_rx_timestamp;
> > > +     else if (func_id == xdp_metadata_kfunc_id(XDP_METADATA_KFUNC_RX_HASH_SUPPORTED))
> > > +             resolved = aux->xdp_netdev->netdev_ops->ndo_xdp_rx_hash_supported;
> > > +     else if (func_id == xdp_metadata_kfunc_id(XDP_METADATA_KFUNC_RX_HASH))
> > > +             resolved = aux->xdp_netdev->netdev_ops->ndo_xdp_rx_hash;
> > > +
> > > +     if (resolved)
> > > +             return BPF_CALL_IMM(resolved);
> > > +     return 0;
> > > +}
> > > +
> >
> > My working tree (based on this version) is available on github [0]. Situation
> > is also described in the last commit message.
> > I would be great, if you could check, whether this behaviour can be reproduced
> > on your setup.
> >
> > [0] https://github.com/walking-machine/linux/tree/hints-v2