On Thu, Dec 1, 2022 at 10:18 PM Martin KaFai Lau <martin.lau@xxxxxxxxx> wrote:
>
> On 12/1/22 5:30 AM, Eyal Birger wrote:
> > Hi Martin,
> >
> > On Thu, Dec 1, 2022 at 7:55 AM Eyal Birger <eyal.birger@xxxxxxxxx> wrote:
> >>
> >> Hi Martin,
> >>
> >> On Wed, Nov 30, 2022 at 8:15 PM Martin KaFai Lau <martin.lau@xxxxxxxxx> wrote:
> >>>
> >>> On 11/29/22 5:20 AM, Eyal Birger wrote:
> >>>> diff --git a/net/xfrm/xfrm_interface_bpf.c b/net/xfrm/xfrm_interface_bpf.c
> >>>> new file mode 100644
> >>>> index 000000000000..757e15857dbf
> >>>> --- /dev/null
> >>>> +++ b/net/xfrm/xfrm_interface_bpf.c
> >>>> @@ -0,0 +1,100 @@
> >>>> +// SPDX-License-Identifier: GPL-2.0-only
> >>>> +/* Unstable XFRM Helpers for TC-BPF hook
> >>>> + *
> >>>> + * These are called from SCHED_CLS BPF programs. Note that it is
> >>>> + * allowed to break compatibility for these functions since the interface they
> >>>> + * are exposed through to BPF programs is explicitly unstable.
> >>>> + */
> >>>> +
> >>>> +#include <linux/bpf.h>
> >>>> +#include <linux/btf_ids.h>
> >>>> +
> >>>> +#include <net/dst_metadata.h>
> >>>> +#include <net/xfrm.h>
> >>>> +
> >>>> +struct bpf_xfrm_info {
> >>> No need to introduce a bpf variant of the "struct xfrm_md_info" (more on this
> >>> later).
> >>>
> >>>> + u32 if_id;
> >>>> + int link;
> >>>> +};
> >>>> +
> >>>> +static struct metadata_dst __percpu *xfrm_md_dst;
> >>>> +__diag_push();
> >>>> +__diag_ignore_all("-Wmissing-prototypes",
> >>>> + "Global functions as their definitions will be in xfrm_interface BTF");
> >>>> +
> >>>> +__used noinline
> >>>> +int bpf_skb_get_xfrm_info(struct __sk_buff *skb_ctx, struct bpf_xfrm_info *to)
> >>>
> >>> This kfunc is not needed. It only reads the skb->_skb_refdst. The new kfunc
> >>> bpf_rdonly_cast() can be used. Take a look at the bpf_rdonly_cast() usages in
> >>> the selftests/bpf/progs/type_cast.c. It was in bpf-next only but should also be
> >>> in net-next now.
> >>
> >> I'm somewhat concerned with this approach.
> >> Indeed it would remove the kfunc, and the API is declared "unstable", but
> >> still the implementation as dst isn't relevant to the user and would make
> >> the programs less readable.
> >>
> >> Also note that the helper can be also used as it is to get the xfrm info at
> >> egress from an lwt route (which stores the xfrm_info in the dst lwstate).
>
> Right, the whole skb_xfrm_md_info() can be implemented in bpf prog itself, like
> checking lwtstate.
>
> If adding a kfunc, how about directly expose skb_xfrm_md_info() itself as a
> kfunc to bpf prog and directly return a "struct xfrm_md_info *" instead. Then
> there is no need to copy if_id/link...etc. The bpf prog has no need to
> initialize the "to" also. Something like this:
>
> __used noinline
> const struct xfrm_md_info *bpf_skb_xfrm_md_info(const struct __sk_buff *skb) { ... }
>
> BTF_ID_FLAGS(func, bpf_skb_xfrm_md_info, KF_RET_NULL)
>
> >>
> >>>
> >>>> +{
> >>>> + struct sk_buff *skb = (struct sk_buff *)skb_ctx;
> >>>> + struct xfrm_md_info *info;
> >>>> +
> >>>> + memset(to, 0, sizeof(*to));
> >>>> +
> >>>> + info = skb_xfrm_md_info(skb);
> >>>> + if (!info)
> >>>> + return -EINVAL;
> >>>> +
> >>>> + to->if_id = info->if_id;
> >>>> + to->link = info->link;
> >>>> + return 0;
> >>>> +}
> >>>> +
> >>>> +__used noinline
> >>>> +int bpf_skb_set_xfrm_info(struct __sk_buff *skb_ctx,
> >>>> + const struct bpf_xfrm_info *from)
> >>>
> >>> Directly use "const struct xfrm_md_info *from" instead. This kfunc can check
> >>> from->dst_orig != NULL and return -EINVAL. It will then have a consistent API
> >>> with the bpf_rdonly_cast() mentioned above.
> >>
> >> See above.
> >
> > Also, when trying this approach with bpf_set_xfrm_info() accepting
> > "const struct xfrm_md_info *from" I fail to load the program:
> >
> > libbpf: prog 'set_xfrm_info': BPF program load failed: Invalid argument
> > libbpf: prog 'set_xfrm_info': -- BEGIN PROG LOAD LOG --
> > 0: R1=ctx(off=0,imm=0) R10=fp0
> > ; int set_xfrm_info(struct __sk_buff *skb)
> > 0: (bf) r6 = r1 ; R1=ctx(off=0,imm=0)
> > R6_w=ctx(off=0,imm=0)
> > 1: (b7) r1 = 0 ; R1_w=0
> > ; struct xfrm_md_info info = {};
> > 2: (7b) *(u64 *)(r10 -8) = r1 ; R1_w=0 R10=fp0 fp-8_w=00000000
> > 3: (7b) *(u64 *)(r10 -16) = r1 ; R1_w=0 R10=fp0 fp-16_w=00000000
> > 4: (b4) w1 = 0 ; R1_w=0
> > ; __u32 index = 0;
> > 5: (63) *(u32 *)(r10 -20) = r1 ; R1_w=0 R10=fp0 fp-24=0000????
> > 6: (bf) r2 = r10 ; R2_w=fp0 R10=fp0
> > ;
> > 7: (07) r2 += -20 ; R2_w=fp-20
> > ; if_id = bpf_map_lookup_elem(&dst_if_id_map, &index);
> > 8: (18) r1 = 0xffff888006751c00 ; R1_w=map_ptr(off=0,ks=4,vs=4,imm=0)
> > 10: (85) call bpf_map_lookup_elem#1 ;
> > R0_w=map_value_or_null(id=1,off=0,ks=4,vs=4,imm=0)
> > 11: (bf) r1 = r0 ;
> > R0_w=map_value_or_null(id=1,off=0,ks=4,vs=4,imm=0)
> > R1_w=map_value_or_null(id=1,off=0,ks=4,vs=4,imm=0)
> > 12: (b4) w0 = 2 ; R0_w=2
> > ; if (!if_id)
> > 13: (15) if r1 == 0x0 goto pc+10 ; R1_w=map_value(off=0,ks=4,vs=4,imm=0)
> > 14: (bf) r2 = r10 ; R2_w=fp0 R10=fp0
> > ;
> > 15: (07) r2 += -16 ; R2_w=fp-16
> > ; info.if_id = *if_id;
> > 16: (61) r1 = *(u32 *)(r1 +0) ;
> > R1_w=scalar(umax=4294967295,var_off=(0x0; 0xffffffff))
> > ; info.if_id = *if_id;
> > 17: (63) *(u32 *)(r2 +0) = r1 ;
> > R1_w=scalar(umax=4294967295,var_off=(0x0; 0xffffffff)) R2_w=fp-16
> > fp-16_w=
> > ; ret = bpf_skb_set_xfrm_info(skb, &info);
> > 18: (bf) r1 = r6 ; R1_w=ctx(off=0,imm=0)
> > R6_w=ctx(off=0,imm=0)
> > 19: (85) call bpf_skb_set_xfrm_info#81442
> > arg#1 pointer type STRUCT xfrm_md_info must point to scalar, or struct
> > with scalar
> >
> > Is there some registration I need to do for this struct?
>
> Ah, thanks for trying!
> hmm... it will need a change to the verifier. likely tag the param with
> something like "const struct xfrm_md_info *from__nonscalar_ok".
>
> The reason of my earlier suggestion was to avoid the need to duplicate future
> changes in xfrm_md_info to bpf_xfrm_info and more importantly avoid creating
> another __sk_buff vs sk_buff like usage confusion.
>
> For now, lets stay with bpf_xfrm_info. This can be changed later.
Ok.
I suggest that in that case we use struct bpf_xfrm_info for both the getter
and setter for now, as it would be more confusing to use different structs.
Btw, I did try using vmlinux.h and importing struct xfrm_md_info, however,
it seems vmlinux.h and linux/pkt_cls.h don't play well together, and the
latter is needed for the TC_ACT_* macro definitions. So at least for the
time being it's another reason to define a single struct in the test program.
Eyal.
