mbox series

[bpf-next,0/2] bpf: add csum/ip_summed fields to __sk_buff

Message ID 20231229081409.1276386-1-menglong8.dong@gmail.com (mailing list archive)
Headers show
Series bpf: add csum/ip_summed fields to __sk_buff | expand

Message

Menglong Dong Dec. 29, 2023, 8:14 a.m. UTC
For now, we have to call some helpers when we need to update the csum,
such as bpf_l4_csum_replace, bpf_l3_csum_replace, etc. These helpers are
not inlined, which causes poor performance.

In fact, we can define our own csum update functions in BPF program
instead of bpf_l3_csum_replace, which is totally inlined and efficient.
However, we can't do this for bpf_l4_csum_replace for now, as we can't
update skb->csum, which can cause skb->csum invalid in the rx path with
CHECKSUM_COMPLETE mode.

What's more, we can't use the direct data access and have to use
skb_store_bytes() with the BPF_F_RECOMPUTE_CSUM flag in some case, such
as modifing the vni in the vxlan header and the underlay udp header has
no checksum.

In the first patch, we make skb->csum readable and writable, and we make
skb->ip_summed readable. For now, for tc only. With these 2 fields, we
don't need to call bpf helpers for csum update any more.

In the second patch, we add some testcases for the read/write testing for
skb->csum and skb->ip_summed.

If this series is acceptable, we can define the inlined functions for csum
update in libbpf in the next step.

Menglong Dong (2):
  bpf: add csum/ip_summed fields to __sk_buff
  testcases/bpf: add testcases for skb->csum to ctx_skb.c

 include/linux/skbuff.h                        |  2 +
 include/uapi/linux/bpf.h                      |  2 +
 net/core/filter.c                             | 22 ++++++++++
 tools/include/uapi/linux/bpf.h                |  2 +
 .../testing/selftests/bpf/verifier/ctx_skb.c  | 43 +++++++++++++++++++
 5 files changed, 71 insertions(+)

Comments

Stanislav Fomichev Jan. 2, 2024, 6:11 p.m. UTC | #1
On 12/29, Menglong Dong wrote:
> For now, we have to call some helpers when we need to update the csum,
> such as bpf_l4_csum_replace, bpf_l3_csum_replace, etc. These helpers are
> not inlined, which causes poor performance.
> 
> In fact, we can define our own csum update functions in BPF program
> instead of bpf_l3_csum_replace, which is totally inlined and efficient.
> However, we can't do this for bpf_l4_csum_replace for now, as we can't
> update skb->csum, which can cause skb->csum invalid in the rx path with
> CHECKSUM_COMPLETE mode.
> 
> What's more, we can't use the direct data access and have to use
> skb_store_bytes() with the BPF_F_RECOMPUTE_CSUM flag in some case, such
> as modifing the vni in the vxlan header and the underlay udp header has
> no checksum.
> 
> In the first patch, we make skb->csum readable and writable, and we make
> skb->ip_summed readable. For now, for tc only. With these 2 fields, we
> don't need to call bpf helpers for csum update any more.
> 
> In the second patch, we add some testcases for the read/write testing for
> skb->csum and skb->ip_summed.
> 
> If this series is acceptable, we can define the inlined functions for csum
> update in libbpf in the next step.

One downside of exposing those as __sk_buff fields is that all this
skb internal csum stuff now becomes a UAPI. And I'm not sure we want
that :-) Should we add a lightweight kfunc to reset the fields instead?
Or will it still have an unacceptable overhead?
Martin KaFai Lau Jan. 3, 2024, 12:52 a.m. UTC | #2
On 1/2/24 10:11 AM, Stanislav Fomichev wrote:
> On 12/29, Menglong Dong wrote:
>> For now, we have to call some helpers when we need to update the csum,
>> such as bpf_l4_csum_replace, bpf_l3_csum_replace, etc. These helpers are
>> not inlined, which causes poor performance.
>>
>> In fact, we can define our own csum update functions in BPF program
>> instead of bpf_l3_csum_replace, which is totally inlined and efficient.
>> However, we can't do this for bpf_l4_csum_replace for now, as we can't
>> update skb->csum, which can cause skb->csum invalid in the rx path with
>> CHECKSUM_COMPLETE mode.
>>
>> What's more, we can't use the direct data access and have to use
>> skb_store_bytes() with the BPF_F_RECOMPUTE_CSUM flag in some case, such
>> as modifing the vni in the vxlan header and the underlay udp header has
>> no checksum.

There is bpf_csum_update(), does it work?
A helper call should be acceptable comparing with the csum calculation itself.

>>
>> In the first patch, we make skb->csum readable and writable, and we make
>> skb->ip_summed readable. For now, for tc only. With these 2 fields, we
>> don't need to call bpf helpers for csum update any more.
>>
>> In the second patch, we add some testcases for the read/write testing for
>> skb->csum and skb->ip_summed.
>>
>> If this series is acceptable, we can define the inlined functions for csum
>> update in libbpf in the next step.
> 
> One downside of exposing those as __sk_buff fields is that all this
> skb internal csum stuff now becomes a UAPI. And I'm not sure we want

+1. Please no new __sk_buff extension and no new conversion in 
bpf_convert_ctx_access().

> that :-) Should we add a lightweight kfunc to reset the fields instead?
> Or will it still have an unacceptable overhead?
Menglong Dong Jan. 3, 2024, 2:54 a.m. UTC | #3
On Wed, Jan 3, 2024 at 8:52 AM Martin KaFai Lau <martin.lau@linux.dev> wrote:
>
> On 1/2/24 10:11 AM, Stanislav Fomichev wrote:
> > On 12/29, Menglong Dong wrote:
> >> For now, we have to call some helpers when we need to update the csum,
> >> such as bpf_l4_csum_replace, bpf_l3_csum_replace, etc. These helpers are
> >> not inlined, which causes poor performance.
> >>
> >> In fact, we can define our own csum update functions in BPF program
> >> instead of bpf_l3_csum_replace, which is totally inlined and efficient.
> >> However, we can't do this for bpf_l4_csum_replace for now, as we can't
> >> update skb->csum, which can cause skb->csum invalid in the rx path with
> >> CHECKSUM_COMPLETE mode.
> >>
> >> What's more, we can't use the direct data access and have to use
> >> skb_store_bytes() with the BPF_F_RECOMPUTE_CSUM flag in some case, such
> >> as modifing the vni in the vxlan header and the underlay udp header has
> >> no checksum.
>
> There is bpf_csum_update(), does it work?
> A helper call should be acceptable comparing with the csum calculation itself.

Yeah, this helper works in this case! Now we miss the last
piece for the tx path: ip_summed. We need to know if it is
CHECKSUM_PARTIAL to decide if we should update the
csum in the packet. In the tx path, the csum in the L4 is the
pseudo header only if skb->ip_summed is CHECKSUM_PARTIAL.

Maybe we can introduce a lightweight kfunc to get its
value? Such as bpf_skb_csum_mode(). As we need only call
it once, there shouldn't be overhead on it.

Thanks!
Menglong Dong

>
> >>
> >> In the first patch, we make skb->csum readable and writable, and we make
> >> skb->ip_summed readable. For now, for tc only. With these 2 fields, we
> >> don't need to call bpf helpers for csum update any more.
> >>
> >> In the second patch, we add some testcases for the read/write testing for
> >> skb->csum and skb->ip_summed.
> >>
> >> If this series is acceptable, we can define the inlined functions for csum
> >> update in libbpf in the next step.
> >
> > One downside of exposing those as __sk_buff fields is that all this
> > skb internal csum stuff now becomes a UAPI. And I'm not sure we want
>
> +1. Please no new __sk_buff extension and no new conversion in
> bpf_convert_ctx_access().
>
> > that :-) Should we add a lightweight kfunc to reset the fields instead?
> > Or will it still have an unacceptable overhead?
>
Yonghong Song Jan. 3, 2024, 3:55 a.m. UTC | #4
On 1/2/24 6:54 PM, Menglong Dong wrote:
> On Wed, Jan 3, 2024 at 8:52 AM Martin KaFai Lau <martin.lau@linux.dev> wrote:
>> On 1/2/24 10:11 AM, Stanislav Fomichev wrote:
>>> On 12/29, Menglong Dong wrote:
>>>> For now, we have to call some helpers when we need to update the csum,
>>>> such as bpf_l4_csum_replace, bpf_l3_csum_replace, etc. These helpers are
>>>> not inlined, which causes poor performance.
>>>>
>>>> In fact, we can define our own csum update functions in BPF program
>>>> instead of bpf_l3_csum_replace, which is totally inlined and efficient.
>>>> However, we can't do this for bpf_l4_csum_replace for now, as we can't
>>>> update skb->csum, which can cause skb->csum invalid in the rx path with
>>>> CHECKSUM_COMPLETE mode.
>>>>
>>>> What's more, we can't use the direct data access and have to use
>>>> skb_store_bytes() with the BPF_F_RECOMPUTE_CSUM flag in some case, such
>>>> as modifing the vni in the vxlan header and the underlay udp header has
>>>> no checksum.
>> There is bpf_csum_update(), does it work?
>> A helper call should be acceptable comparing with the csum calculation itself.
> Yeah, this helper works in this case! Now we miss the last
> piece for the tx path: ip_summed. We need to know if it is
> CHECKSUM_PARTIAL to decide if we should update the
> csum in the packet. In the tx path, the csum in the L4 is the
> pseudo header only if skb->ip_summed is CHECKSUM_PARTIAL.
>
> Maybe we can introduce a lightweight kfunc to get its
> value? Such as bpf_skb_csum_mode(). As we need only call
> it once, there shouldn't be overhead on it.

You don't need kfunc, you can do checking like
   struct sk_buff *kskb = bpf_cast_to_kern_ctx(skb);
   if (kskb->ip_summed == CHECKSUM_PARTIAL) ...
   ...
   

>
> Thanks!
> Menglong Dong
>
>>>> In the first patch, we make skb->csum readable and writable, and we make
>>>> skb->ip_summed readable. For now, for tc only. With these 2 fields, we
>>>> don't need to call bpf helpers for csum update any more.
>>>>
>>>> In the second patch, we add some testcases for the read/write testing for
>>>> skb->csum and skb->ip_summed.
>>>>
>>>> If this series is acceptable, we can define the inlined functions for csum
>>>> update in libbpf in the next step.
>>> One downside of exposing those as __sk_buff fields is that all this
>>> skb internal csum stuff now becomes a UAPI. And I'm not sure we want
>> +1. Please no new __sk_buff extension and no new conversion in
>> bpf_convert_ctx_access().
>>
>>> that :-) Should we add a lightweight kfunc to reset the fields instead?
>>> Or will it still have an unacceptable overhead?
Menglong Dong Jan. 3, 2024, 6:03 a.m. UTC | #5
On Wed, Jan 3, 2024 at 11:55 AM Yonghong Song <yonghong.song@linux.dev> wrote:
>
>
> On 1/2/24 6:54 PM, Menglong Dong wrote:
> > On Wed, Jan 3, 2024 at 8:52 AM Martin KaFai Lau <martin.lau@linux.dev> wrote:
> >> On 1/2/24 10:11 AM, Stanislav Fomichev wrote:
> >>> On 12/29, Menglong Dong wrote:
> >>>> For now, we have to call some helpers when we need to update the csum,
> >>>> such as bpf_l4_csum_replace, bpf_l3_csum_replace, etc. These helpers are
> >>>> not inlined, which causes poor performance.
> >>>>
> >>>> In fact, we can define our own csum update functions in BPF program
> >>>> instead of bpf_l3_csum_replace, which is totally inlined and efficient.
> >>>> However, we can't do this for bpf_l4_csum_replace for now, as we can't
> >>>> update skb->csum, which can cause skb->csum invalid in the rx path with
> >>>> CHECKSUM_COMPLETE mode.
> >>>>
> >>>> What's more, we can't use the direct data access and have to use
> >>>> skb_store_bytes() with the BPF_F_RECOMPUTE_CSUM flag in some case, such
> >>>> as modifing the vni in the vxlan header and the underlay udp header has
> >>>> no checksum.
> >> There is bpf_csum_update(), does it work?
> >> A helper call should be acceptable comparing with the csum calculation itself.
> > Yeah, this helper works in this case! Now we miss the last
> > piece for the tx path: ip_summed. We need to know if it is
> > CHECKSUM_PARTIAL to decide if we should update the
> > csum in the packet. In the tx path, the csum in the L4 is the
> > pseudo header only if skb->ip_summed is CHECKSUM_PARTIAL.
> >
> > Maybe we can introduce a lightweight kfunc to get its
> > value? Such as bpf_skb_csum_mode(). As we need only call
> > it once, there shouldn't be overhead on it.
>
> You don't need kfunc, you can do checking like
>    struct sk_buff *kskb = bpf_cast_to_kern_ctx(skb);
>    if (kskb->ip_summed == CHECKSUM_PARTIAL) ...
>    ...
>

Great, this is exactly what I need! Thanks~