This patch adds struct_ops support to the bpftool.
To recap a bit on the recent bpf_struct_ops feature on the kernel side:
It currently supports "struct tcp_congestion_ops" to be implemented
in bpf. At a high level, bpf_struct_ops is struct_ops map populated
with a number of bpf progs. bpf_struct_ops currently supports the
"struct tcp_congestion_ops". However, the bpf_struct_ops design is
generic enough that other kernel struct ops can be supported in
the future.
Although struct_ops is map+progs at a high lever, there are differences
in details. For example,
1) After registering a struct_ops, the struct_ops is held by the kernel
subsystem (e.g. tcp-cc). Thus, there is no need to pin a
struct_ops map or its progs in order to keep them around.
2) To iterate all struct_ops in a system, it iterates all maps
in type BPF_MAP_TYPE_STRUCT_OPS. BPF_MAP_TYPE_STRUCT_OPS is
the current usual filter. In the future, it may need to
filter by other struct_ops specific properties. e.g. filter by
tcp_congestion_ops or other kernel subsystem ops in the future.
3) struct_ops requires the running kernel having BTF info. That allows
more flexibility in handling other kernel structs. e.g. it can
always dump the latest bpf_map_info.
4) Also, "struct_ops" command is not intended to repeat all features
already provided by "map" or "prog". For example, if there really
is a need to pin the struct_ops map, the user can use the "map" cmd
to do that.
While the first attempt was to reuse parts from map/prog.c, it ended up
not a lot to share. The only obvious item is the map_parse_fds() but
that still requires modifications to accommodate struct_ops map specific
filtering (for the immediate and the future needs). Together with the
earlier mentioned differences, it is better to part away from map/prog.c.
The initial set of subcmds are, register, unregister, show, and dump.
For register, it registers all struct_ops maps that can be found in an
obj file. Option can be added in the future to specify a particular
struct_ops map. Also, the common bpf_tcp_cc is stateless (e.g.
bpf_cubic.c and bpf_dctcp.c). The "reuse map" feature is not
implemented in this patch and it can be considered later also.
For other subcmds, please see the man doc for details.
A sample output of dump:
[root@arch-fb-vm1 bpf]# bpftool struct_ops dump name cubic
[{
"bpf_map_info": {
"type": 26,
"id": 64,
"key_size": 4,
"value_size": 256,
"max_entries": 1,
"map_flags": 0,
"name": "cubic",
"ifindex": 0,
"btf_vmlinux_value_type_id": 18452,
"netns_dev": 0,
"netns_ino": 0,
"btf_id": 52,
"btf_key_type_id": 0,
"btf_value_type_id": 0
}
},{
"bpf_struct_ops_tcp_congestion_ops": {
"refcnt": {
"refs": {
"counter": 1
}
},
"state": "BPF_STRUCT_OPS_STATE_INUSE",
"data": {
"list": {
"next": 0,
"prev": 0
},
"key": 0,
"flags": 0,
"init": "void (struct sock *) bictcp_init/prog_id:138",
"release": "void (struct sock *) 0",
"ssthresh": "u32 (struct sock *) bictcp_recalc_ssthresh/prog_id:141",
"cong_avoid": "void (struct sock *, u32, u32) bictcp_cong_avoid/prog_id:140",
"set_state": "void (struct sock *, u8) bictcp_state/prog_id:142",
"cwnd_event": "void (struct sock *, enum tcp_ca_event) bictcp_cwnd_event/prog_id:139",
"in_ack_event": "void (struct sock *, u32) 0",
"undo_cwnd": "u32 (struct sock *) tcp_reno_undo_cwnd/prog_id:144",
"pkts_acked": "void (struct sock *, const struct ack_sample *) bictcp_acked/prog_id:143",
"min_tso_segs": "u32 (struct sock *) 0",
"sndbuf_expand": "u32 (struct sock *) 0",
"cong_control": "void (struct sock *, const struct rate_sample *) 0",
"get_info": "size_t (struct sock *, u32, int *, union tcp_cc_info *) 0",
"name": "bpf_cubic",
"owner": 0
}
}
}
]
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Quentin Monnet <quentin@isovalent.com>
Link: https://lore.kernel.org/bpf/20200318171656.129650-1-kafai@fb.com