[v2,bpf-next,0/5] bpf: Remove recursion check for struct_ops prog

Message

Martin KaFai Lau Sept. 23, 2022, 10:44 p.m. UTC

From: Martin KaFai Lau <[e-mail]>

The struct_ops is sharing the tracing-trampoline's enter/exit
function which tracks prog->active to avoid recursion.  It turns
out the struct_ops bpf prog will hit this prog->active and
unnecessarily skipped running the struct_ops prog.  eg.  The
'.ssthresh' may run in_task() and then interrupted by softirq
that runs the same '.ssthresh'.

The kernel does not call the tcp-cc's ops in a recursive way,
so this set is to remove the recursion check for struct_ops prog.

v2:
v1 [0] turned into a long discussion on a few cases and also
whether it needs to follow the bpf_run_ctx chain if there is
tracing bpf_run_ctx (kprobe/trace/trampoline) running in between.

It is a good signal that it is not obvious enough to reason
about it and needs a tradeoff for a more straight forward approach.

This revision uses one bit out of an existing 1 byte hole
in the tcp_sock.  It is in Patch 4.

[0]: https://lore.kernel.org/bpf/[e-mail]/T/#md98d40ac5ec295fdadef476c227a3401b2b6b911

Martin KaFai Lau (5):
  bpf: Add __bpf_prog_{enter,exit}_struct_ops for struct_ops trampoline
  bpf: Move the "cdg" tcp-cc check to the common sol_tcp_sockopt()
  bpf: Refactor bpf_setsockopt(TCP_CONGESTION) handling into another
    function
  bpf: tcp: Stop bpf_setsockopt(TCP_CONGESTION) in init ops to recur
    itself
  selftests/bpf: Check -EBUSY for the recurred
    bpf_setsockopt(TCP_CONGESTION)

 arch/x86/net/bpf_jit_comp.c                   |  3 +
 include/linux/bpf.h                           |  4 ++
 include/linux/tcp.h                           |  6 ++
 kernel/bpf/trampoline.c                       | 23 ++++++
 net/core/filter.c                             | 70 ++++++++++++++-----
 .../selftests/bpf/prog_tests/bpf_tcp_ca.c     |  4 ++
 tools/testing/selftests/bpf/progs/bpf_dctcp.c | 25 ++++---
 7 files changed, 111 insertions(+), 24 deletions(-)