BPF_PROG_TYPE_SK_REUSEPORT / sk_reuseport ebpf programs do not have
access to the queue_mapping or napi_id of the incoming skb. Having
this information can help ebpf progs determine which listen socket to
select.
This patch exposes both queue_mapping and napi_id so that
sk_reuseport ebpf programs can use this information to direct incoming
connections to the correct listen socket in the SOCKMAP.
For example:
A multi-threaded userland program with several threads accepting client
connections via a reuseport listen socket group might want to direct
incoming connections from specific receive queues (or NAPI IDs) to specific
listen sockets to maximize locality or for use with epoll busy-poll.
Signed-off-by: Joe Damato <[email protected]>
---
include/uapi/linux/bpf.h | 2 ++
net/core/filter.c | 10 ++++++++++
2 files changed, 12 insertions(+)
diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
index 9273c654743c..31560b506535 100644
--- a/include/uapi/linux/bpf.h
+++ b/include/uapi/linux/bpf.h
@@ -6286,6 +6286,8 @@ struct sk_reuseport_md {
*/
__u32 eth_protocol;
__u32 ip_protocol; /* IP protocol. e.g. IPPROTO_TCP, IPPROTO_UDP */
+ __u32 rx_queue_mapping; /* Rx queue associated with the skb */
+ __u32 napi_id; /* napi id associated with the skb */
__u32 bind_inany; /* Is sock bound to an INANY address? */
__u32 hash; /* A hash of the packet 4 tuples */
/* When reuse->migrating_sk is NULL, it is selecting a sk for the
diff --git a/net/core/filter.c b/net/core/filter.c
index 968139f4a1ac..71826e1ef7dc 100644
--- a/net/core/filter.c
+++ b/net/core/filter.c
@@ -11134,6 +11134,8 @@ sk_reuseport_is_valid_access(int off, int size,
case bpf_ctx_range(struct sk_reuseport_md, ip_protocol):
case bpf_ctx_range(struct sk_reuseport_md, bind_inany):
case bpf_ctx_range(struct sk_reuseport_md, len):
+ case bpf_ctx_range(struct sk_reuseport_md, rx_queue_mapping):
+ case bpf_ctx_range(struct sk_reuseport_md, napi_id):
bpf_ctx_record_field_size(info, size_default);
return bpf_ctx_narrow_access_ok(off, size, size_default);
@@ -11183,6 +11185,14 @@ static u32 sk_reuseport_convert_ctx_access(enum bpf_access_type type,
SK_REUSEPORT_LOAD_SKB_FIELD(protocol);
break;
+ case offsetof(struct sk_reuseport_md, rx_queue_mapping):
+ SK_REUSEPORT_LOAD_SKB_FIELD(queue_mapping);
+ break;
+
+ case offsetof(struct sk_reuseport_md, napi_id):
+ SK_REUSEPORT_LOAD_SKB_FIELD(napi_id);
+ break;
+
case offsetof(struct sk_reuseport_md, ip_protocol):
SK_REUSEPORT_LOAD_SK_FIELD(sk_protocol);
break;
--
2.25.1
On 5/24/23 8:37 PM, Joe Damato wrote:
> BPF_PROG_TYPE_SK_REUSEPORT / sk_reuseport ebpf programs do not have
> access to the queue_mapping or napi_id of the incoming skb. Having
> this information can help ebpf progs determine which listen socket to
> select.
>
> This patch exposes both queue_mapping and napi_id so that
> sk_reuseport ebpf programs can use this information to direct incoming
> connections to the correct listen socket in the SOCKMAP.
>
> For example:
>
> A multi-threaded userland program with several threads accepting client
> connections via a reuseport listen socket group might want to direct
> incoming connections from specific receive queues (or NAPI IDs) to specific
> listen sockets to maximize locality or for use with epoll busy-poll.
>
> Signed-off-by: Joe Damato <[email protected]>
> ---
> include/uapi/linux/bpf.h | 2 ++
> net/core/filter.c | 10 ++++++++++
> 2 files changed, 12 insertions(+)
>
> diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> index 9273c654743c..31560b506535 100644
> --- a/include/uapi/linux/bpf.h
> +++ b/include/uapi/linux/bpf.h
> @@ -6286,6 +6286,8 @@ struct sk_reuseport_md {
> */
> __u32 eth_protocol;
> __u32 ip_protocol; /* IP protocol. e.g. IPPROTO_TCP, IPPROTO_UDP */
> + __u32 rx_queue_mapping; /* Rx queue associated with the skb */
> + __u32 napi_id; /* napi id associated with the skb */
> __u32 bind_inany; /* Is sock bound to an INANY address? */
> __u32 hash; /* A hash of the packet 4 tuples */
This won't work. You will need to append to the end of data structure
to keep it backward compatibility.
Also, recent kernel has a kfunc bpf_cast_to_kern_ctx() which converts
a ctx to a kernel ctx and you can then use tracing-coding-style to
access those fields. In this particular case, you can do
struct sk_reuseport_kern *kctx = bpf_cast_to_kern_ctx(ctx);
We have
struct sk_reuseport_kern {
struct sk_buff *skb;
struct sock *sk;
struct sock *selected_sk;
struct sock *migrating_sk;
void *data_end;
u32 hash;
u32 reuseport_id;
bool bind_inany;
};
through sk and skb pointer, you should be access the fields presented in
this patch. You can access more fields too.
So using bpf_cast_to_kern_ctx(), there is no need for more uapi changes.
Please give a try.
> /* When reuse->migrating_sk is NULL, it is selecting a sk for the
> diff --git a/net/core/filter.c b/net/core/filter.c
> index 968139f4a1ac..71826e1ef7dc 100644
> --- a/net/core/filter.c
> +++ b/net/core/filter.c
> @@ -11134,6 +11134,8 @@ sk_reuseport_is_valid_access(int off, int size,
> case bpf_ctx_range(struct sk_reuseport_md, ip_protocol):
> case bpf_ctx_range(struct sk_reuseport_md, bind_inany):
> case bpf_ctx_range(struct sk_reuseport_md, len):
> + case bpf_ctx_range(struct sk_reuseport_md, rx_queue_mapping):
> + case bpf_ctx_range(struct sk_reuseport_md, napi_id):
> bpf_ctx_record_field_size(info, size_default);
> return bpf_ctx_narrow_access_ok(off, size, size_default);
>
> @@ -11183,6 +11185,14 @@ static u32 sk_reuseport_convert_ctx_access(enum bpf_access_type type,
> SK_REUSEPORT_LOAD_SKB_FIELD(protocol);
> break;
>
> + case offsetof(struct sk_reuseport_md, rx_queue_mapping):
> + SK_REUSEPORT_LOAD_SKB_FIELD(queue_mapping);
> + break;
> +
> + case offsetof(struct sk_reuseport_md, napi_id):
> + SK_REUSEPORT_LOAD_SKB_FIELD(napi_id);
> + break;
> +
> case offsetof(struct sk_reuseport_md, ip_protocol):
> SK_REUSEPORT_LOAD_SK_FIELD(sk_protocol);
> break;
On Wed, May 24, 2023 at 10:26:32PM -0700, Yonghong Song wrote:
>
>
> On 5/24/23 8:37 PM, Joe Damato wrote:
> >BPF_PROG_TYPE_SK_REUSEPORT / sk_reuseport ebpf programs do not have
> >access to the queue_mapping or napi_id of the incoming skb. Having
> >this information can help ebpf progs determine which listen socket to
> >select.
> >
> >This patch exposes both queue_mapping and napi_id so that
> >sk_reuseport ebpf programs can use this information to direct incoming
> >connections to the correct listen socket in the SOCKMAP.
> >
> >For example:
> >
> >A multi-threaded userland program with several threads accepting client
> >connections via a reuseport listen socket group might want to direct
> >incoming connections from specific receive queues (or NAPI IDs) to specific
> >listen sockets to maximize locality or for use with epoll busy-poll.
> >
> >Signed-off-by: Joe Damato <[email protected]>
> >---
> > include/uapi/linux/bpf.h | 2 ++
> > net/core/filter.c | 10 ++++++++++
> > 2 files changed, 12 insertions(+)
> >
> >diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> >index 9273c654743c..31560b506535 100644
> >--- a/include/uapi/linux/bpf.h
> >+++ b/include/uapi/linux/bpf.h
> >@@ -6286,6 +6286,8 @@ struct sk_reuseport_md {
> > */
> > __u32 eth_protocol;
> > __u32 ip_protocol; /* IP protocol. e.g. IPPROTO_TCP, IPPROTO_UDP */
> >+ __u32 rx_queue_mapping; /* Rx queue associated with the skb */
> >+ __u32 napi_id; /* napi id associated with the skb */
> > __u32 bind_inany; /* Is sock bound to an INANY address? */
> > __u32 hash; /* A hash of the packet 4 tuples */
>
> This won't work. You will need to append to the end of data structure
> to keep it backward compatibility.
>
> Also, recent kernel has a kfunc bpf_cast_to_kern_ctx() which converts
> a ctx to a kernel ctx and you can then use tracing-coding-style to
> access those fields. In this particular case, you can do
>
> struct sk_reuseport_kern *kctx = bpf_cast_to_kern_ctx(ctx);
>
> We have
>
> struct sk_reuseport_kern {
> struct sk_buff *skb;
> struct sock *sk;
> struct sock *selected_sk;
> struct sock *migrating_sk;
> void *data_end;
> u32 hash;
> u32 reuseport_id;
> bool bind_inany;
> };
>
> through sk and skb pointer, you should be access the fields presented in
> this patch. You can access more fields too.
>
> So using bpf_cast_to_kern_ctx(), there is no need for more uapi changes.
> Please give a try.
Thanks! I was looking at an LTS kernel tree that didn't have
bpf_cast_to_kern_ctx; this is very helpful and definitely a better way to
go.
Sorry for the noise.