| Message ID | 20230525033757.47483-1-jdamato@fastly.com (mailing list archive) |
|---|---|
| State | Changes Requested |
| Delegated to: | BPF |
| Headers | show |
| Series | [bpf-next] bpf: Export rx queue info for reuseport ebpf prog | expand |
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 <jdamato@fastly.com> > --- > 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 <jdamato@fastly.com> > >--- > > 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.
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;
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 <jdamato@fastly.com> --- include/uapi/linux/bpf.h | 2 ++ net/core/filter.c | 10 ++++++++++ 2 files changed, 12 insertions(+)