Joanne Koong <joannelkoong@xxxxxxxxx> writes:
> On Mon, Aug 22, 2022 at 7:31 PM Kumar Kartikeya Dwivedi
> <memxor@xxxxxxxxx> wrote:
>>
>> +Cc XDP folks
>>
>> On Tue, 23 Aug 2022 at 02:12, Joanne Koong <joannelkoong@xxxxxxxxx> wrote:
>> >
>> > Add xdp dynptrs, which are dynptrs whose underlying pointer points
>> > to a xdp_buff. The dynptr acts on xdp data. xdp dynptrs have two main
>> > benefits. One is that they allow operations on sizes that are not
>> > statically known at compile-time (eg variable-sized accesses).
>> > Another is that parsing the packet data through dynptrs (instead of
>> > through direct access of xdp->data and xdp->data_end) can be more
>> > ergonomic and less brittle (eg does not need manual if checking for
>> > being within bounds of data_end).
>> >
>> > For reads and writes on the dynptr, this includes reading/writing
>> > from/to and across fragments. For data slices, direct access to
>>
>> It's a bit awkward to have such a difference between xdp and skb
>> dynptr's read/write. I understand why it is the way it is, but it
>> still doesn't feel right. I'm not sure if we can reconcile the
>> differences, but it makes writing common code for both xdp and tc
>> harder as it needs to be aware of the differences (and then the flags
>> for dynptr_write would differ too). So we're 90% there but not the
>> whole way...
>
> Yeah, it'd be great if the behavior for skb/xdp progs could be the
> same, but I'm not seeing a better solution here (unless we invalidate
> data slices on writes in xdp progs, just to make it match more :P).
>
> Regarding having 2 different interfaces bpf_dynptr_from_{skb/xdp}, I'm
> not convinced this is much of a problem - xdp and skb programs already
> have different interfaces for doing things (eg
> bpf_{skb/xdp}_{store/load}_bytes).
This is true, but it's quite possible to paper over these differences
and write BPF code that works for both TC and XDP. Subtle semantic
differences in otherwise identical functions makes this harder.
Today you can write a function like:
static inline int parse_pkt(void *data, void* data_end)
{
/* parse data */
}
And call it like:
SEC("xdp")
int parse_xdp(struct xdp_md *ctx)
{
return parse_pkt(ctx->data, ctx->data_end);
}
SEC("tc")
int parse_tc(struct __sk_buff *skb)
{
return parse_pkt(skb->data, skb->data_end);
}
IMO the goal should be to be able to do the equivalent for dynptrs, like:
static inline int parse_pkt(struct bpf_dynptr *ptr)
{
__u64 *data;
data = bpf_dynptr_data(ptr, 0, sizeof(*data));
if (!data)
return 0;
/* parse data */
}
SEC("xdp")
int parse_xdp(struct xdp_md *ctx)
{
struct bpf_dynptr ptr;
bpf_dynptr_from_xdp(ctx, 0, &ptr);
return parse_pkt(&ptr);
}
SEC("tc")
int parse_tc(struct __sk_buff *skb)
{
struct bpf_dynptr ptr;
bpf_dynptr_from_skb(skb, 0, &ptr);
return parse_pkt(&ptr);
}
If the dynptr-based parse_pkt() function has to take special care to
figure out where the dynptr comes from, it makes it a lot more difficult
to write reusable packet parsing functions. So I'd be in favour of
restricting the dynptr interface to the lowest common denominator of the
skb and xdp interfaces even if that makes things slightly more awkward
in the specialised cases...
-Toke
