[RFC,iproute2,v6,0/3] ss: pretty-printing BPF socket-local storage

Message

Quentin Deslandes Jan. 18, 2024, 3:15 a.m. UTC

BPF allows programs to store socket-specific data using
BPF_MAP_TYPE_SK_STORAGE maps. The data is attached to the socket itself,
and Martin added INET_DIAG_REQ_SK_BPF_STORAGES, so it can be fetched
using the INET_DIAG mechanism.

Currently, ss doesn't request the socket-local data, this patch aims to
fix this.

The first patch requests the socket-local data for the requested map ID
(--bpf-map-id=) or all the maps (--bpf-maps). It then prints the map_id
in a dedicated column.

Patch #2 uses libbpf and BTF to pretty print the map's content, like
`bpftool map dump` would do.

Patch #3 updates ss' man page to explain new options.

While I think it makes sense for ss to provide the socket-local storage
content for the sockets, it's difficult to conciliate the column-based
output of ss and having readable socket-local data. Hence, the
socket-local data is printed in a readable fashion over multiple lines
under its socket statistics, independently of the column-based approach.

Here is an example of ss' output with --bpf-maps:
[...]
ESTAB                  340116             0 [...]
    map_id: 114 [
        (struct my_sk_storage){
            .field_hh = (char)3,
            (union){
                .a = (int)17,
                .b = (int)17,
            },
        }
    ]

Changed this series to an RFC as the merging window for net-next is
closed.

Changes from v5:
* Add support for --oneline when printing socket-local data.
* Use \t to indent instead of "  " to be consistent with other columns.
* Removed Martin's ack on patch #2 due to amount of lines changed.
Changes from v4:
* Fix return code for 2 calls.
* Fix issue when inet_show_netlink() retries a request.
* BPF dump object is created in bpf_map_opts_load_info().
Changes from v3:
* Minor refactoring to reduce number of HAVE_LIBBF usage.
* Update ss' man page.
* btf_dump structure created to print the socket-local data is cached
  in bpf_map_opts. Creation of the btf_dump structure is performed if
  needed, before printing the data.
* If a map can't be pretty-printed, print its ID and a message instead
  of skipping it.
* If show_all=true, send an empty message to the kernel to retrieve all
  the maps (as Martin suggested).
Changes from v2:
* bpf_map_opts_is_enabled is not inline anymore.
* Add more #ifdef HAVE_LIBBPF to prevent compilation error if
  libbpf support is disabled.
* Fix erroneous usage of args instead of _args in vout().
* Add missing btf__free() and close(fd).
Changes from v1:
* Remove the first patch from the series (fix) and submit it separately.
* Remove double allocation of struct rtattr.
* Close BPF map FDs on exit.
* If bpf_map_get_fd_by_id() fails with ENOENT, print an error message
  and continue to the next map ID.
* Fix typo in new command line option documentation.
* Only use bpf_map_info.btf_value_type_id and ignore
  bpf_map_info.btf_vmlinux_value_type_id (unused for socket-local storage).
* Use btf_dump__dump_type_data() instead of manually using BTF to
  pretty-print socket-local storage data. This change alone divides the size
  of the patch series by 2.

Quentin Deslandes (3):
  ss: add support for BPF socket-local storage
  ss: pretty-print BPF socket-local storage
  ss: update man page to document --bpf-maps and --bpf-map-id=

 man/man8/ss.8 |   6 +
 misc/ss.c     | 413 ++++++++++++++++++++++++++++++++++++++++++++++++--
 2 files changed, 410 insertions(+), 9 deletions(-)

--
2.43.0