[PATCH bpf-next v3 00/15] bpf: A common way to attach struct_ops to a cgroup

[Amery Hung <ameryhung@gmail.com>]

Hi,

I am continuing Martin's work to support attaching struct_ops to cgroup.

At LSF/MM/BPF 2025, Martin presented [1] the need for a new interface
to extend tcp_sock operations instead of adding more BPF_SOCK_OPS_*CB
enum values. The need for predictable ordering when attaching struct_ops
to a cgroup was also briefly discussed.

At LSF/MM/BPF 2026, additional use cases were raised, in particular
OOM and memcg use cases that also need to attach struct_ops to a
cgroup.

BPF already has a common bpf_link-based API for attaching different
BPF program types to a cgroup. It provides common attach, detach,
update, ordering, and query semantics across those program types.

This series extends the same model to struct_ops. Conceptually,
struct_ops is a group of BPF programs, so using similar
attachment/detachment/update/query APIs and ordering semantics for
cgroup attachment keeps the interface consistent with existing cgroup
BPF links.

This series uses a new struct bpf_tcp_ops as the first user. The
struct_ops mirrors the TCP-related sockops hooks except
BPF_SOCK_OPS_NEEDS_ECN and BPF_SOCK_OPS_BASE_RTT, which are
intentionally left out.

The selftests cover attach, query, update, ordering, before/after
placement, retval chaining, the header option hooks, and inheritance
across a multi-level cgroup hierarchy.

The map_free_pre_rcu addition in patch 2 is not very ideal; it will
need some thought too.

[1] page 13: https://drive.google.com/file/d/1wjKZth6T0llLJ_ONPAL_6Q_jbxbAjByp/view?usp=sharing


Changelog

v2 -> v3
  - Patch 3: Remove redundant bpf_struct_ops_kdata_map_id (Sashiko)
  - Patch 11: Reject sleepable programs (Sashiko)
  - Patch 12: Silence warning when casting ctx to arg pointer
  - Patch 13: Fix incorrect libbpf API addition (Andrii)
  - Patch 14: Fix selftest cgroup cleanup (Sashiko)
  - Patch 15: Use network_helpers.c

RFC v1 -> v2
  - Fix UAF of cfi_stubs
  - Fix retval: use bpf_tramp_run_ctx instead of bpf_cg_run_ctx and
    expose bpf_get_retval() to bpf_tcp_ops
  - Add selftests
    - struct_ops cgroup attachment
      - Test bpf_get_retval()
      - Test before/after order
      - Test cgroup hierarchy and inheritance
    - Test TCP header option hooks and helpers
  - Move bpf_tcp_ops out of legacy BPF_SOCK_OPS_TEST_FLAG guard
  - Complete bpf_tcp_ops (make it comparable to legacy sockops tcp)

---

Amery Hung (4):
  bpf: Allow all struct_ops to use bpf_dynptr_from_skb()
  bpf: tcp: Support selected sock_ops callbacks as struct_ops
  bpf: tcp: Support parse/len/write header option hooks in bpf_tcp_ops
  selftests/bpf: Add test for bpf_tcp_ops header option hooks

Martin KaFai Lau (11):
  bpf: Remove __rcu tagging in st_link->map
  bpf: Make struct_ops tasks_rcu grace period optional
  bpf: Add bpf_struct_ops accessor helpers
  bpf: Remove unnecessary prog_list_prog() check
  bpf: Replace prog_list_prog() check with direct pl->prog and pl->link
    check
  bpf: Add prog_list_init_item(), prog_list_replace_item(), and
    prog_list_id()
  bpf: Move LSM trampoline unlink into bpf_cgroup_link_auto_detach()
  bpf: Add a few bpf_cgroup_array_* helper functions
  bpf: Add infrastructure to support attaching struct_ops to cgroups
  libbpf: Support attaching struct_ops to a cgroup
  selftests/bpf: Test attaching struct_ops to a cgroup

 include/linux/bpf-cgroup-defs.h               |   1 +
 include/linux/bpf-cgroup.h                    |  28 +
 include/linux/bpf.h                           |  56 +-
 include/linux/filter.h                        |   5 +
 include/net/tcp.h                             | 153 ++++-
 include/uapi/linux/bpf.h                      |  39 +-
 kernel/bpf/bpf_struct_ops.c                   | 142 +++--
 kernel/bpf/btf.c                              |  23 +-
 kernel/bpf/cgroup.c                           | 472 ++++++++++++++-
 kernel/bpf/core.c                             |   5 +
 kernel/bpf/syscall.c                          |   4 +
 net/core/filter.c                             |  33 +-
 net/ipv4/Makefile                             |   1 +
 net/ipv4/af_inet.c                            |   1 +
 net/ipv4/bpf_tcp_ca.c                         |  16 +
 net/ipv4/bpf_tcp_ops.c                        | 324 ++++++++++
 net/ipv4/tcp.c                                |   1 +
 net/ipv4/tcp_input.c                          |  17 +
 net/ipv4/tcp_output.c                         |  48 ++
 net/ipv4/tcp_timer.c                          |   1 +
 net/sched/bpf_qdisc.c                         |   2 -
 tools/include/uapi/linux/bpf.h                |  39 +-
 tools/lib/bpf/bpf.c                           |   2 +
 tools/lib/bpf/bpf.h                           |   3 +-
 tools/lib/bpf/libbpf.c                        |  64 ++
 tools/lib/bpf/libbpf.h                        |   3 +
 tools/lib/bpf/libbpf.map                      |   1 +
 .../selftests/bpf/prog_tests/bpf_tcp_ops.c    | 560 ++++++++++++++++++
 .../bpf/prog_tests/bpf_tcp_ops_hdr.c          |  77 +++
 .../testing/selftests/bpf/progs/bpf_tcp_ops.c | 141 +++++
 .../selftests/bpf/progs/bpf_tcp_ops_hdr.c     |  83 +++
 31 files changed, 2224 insertions(+), 121 deletions(-)
 create mode 100644 net/ipv4/bpf_tcp_ops.c
 create mode 100644 tools/testing/selftests/bpf/prog_tests/bpf_tcp_ops.c
 create mode 100644 tools/testing/selftests/bpf/prog_tests/bpf_tcp_ops_hdr.c
 create mode 100644 tools/testing/selftests/bpf/progs/bpf_tcp_ops.c
 create mode 100644 tools/testing/selftests/bpf/progs/bpf_tcp_ops_hdr.c

-- 
2.52.0