Skip to content

Perf-based event notification for sock_ops

Sowmini Varadhan netdev 2 versions – New

Get series with b4
b4 am [email protected]

Versions newest first · linked by author and title

  1. v2 Latest2 patches 0 repliesNew
  2. v12 patches 1 replySuperseded

Patches in v2 No applies recorded yet

#Subject and filesSizeRepliesApplied
1/2bpf: add perf-event notificaton support for sock_opsnet/core/filter.c+19 -00 replies
2/2selftests/bpf: add a test case for sock_ops perf-event notificationtools/testing/selftests/bpf/Makefile tools/testing/selftests/bpf/test_tcpnotify.h tools/testing/selftests/bpf/test_tcpnotify_kern.c +1 more+303 -10 replies

Reviews trailers per patch

No Reviewed-by, Acked-by or Tested-by trailers for v2 yet.

Read the v2 thread

Cover letter [PATCH V2 bpf-next 0/2] Perf-based event notification for sock_ops

This patchset uses eBPF perf-event based notification mechanism to solve the problem described in
https://marc.info/?l=linux-netdev&m=154022219423571&w=2.
Thanks to Daniel Borkmann for feedback/input.

V2: inlined the call to sys_perf_event_open() following the style
of existing code in kselftests/bpf

The problem statement is
We would like to monitor some subset of TCP sockets in user-space, (the monitoring application would define 4-tuples it wants to monitor) using TCP_INFO stats to analyze reported problems. The idea is to use those stats to see where the bottlenecks are likely to be ("is it application-limited?" or "is there evidence of BufferBloat in the path?" etc)

Today we can do this by periodically polling for tcp_info, but this could be made more efficient if the kernel would asynchronously notify the application via tcp_info when some "interesting" thresholds (e.g., "RTT variance > X", or "total_retrans > Y" etc) are reached. And to make this effective, it is better if we could apply the threshold check *before* constructing the tcp_info netlink notification, so that we don't waste resources constructing notifications that will be discarded by the filter.

This patchset solves the problem by adding perf-event based notification support for sock_ops (Patch1). The eBPF kernel module can thus be designed to apply any desired filters to the bpf_sock_ops and trigger a perf-event notification based on the verdict from the filter. The uspace component can use these perf-event notifications to either read any state managed by the eBPF kernel module, or issue a TCP_INFO netlink call if desired.

Patch 2 provides a simple example that shows how to use this infra (and also provides a test case for it)

Sowmini Varadhan (2):
bpf: add perf-event notificaton support for sock_ops selftests/bpf: add a test case for sock_ops perf-event notification

 net/core/filter.c                                 |   19 ++
 tools/testing/selftests/bpf/Makefile              |    4 +-
 tools/testing/selftests/bpf/test_tcpnotify.h      |   19 ++
 tools/testing/selftests/bpf/test_tcpnotify_kern.c |   95 +++++++++++
 tools/testing/selftests/bpf/test_tcpnotify_user.c |  186 +++++++++++++++++++++
 5 files changed, 322 insertions(+), 1 deletions(-)
 create mode 100644 tools/testing/selftests/bpf/test_tcpnotify.h
 create mode 100644 tools/testing/selftests/bpf/test_tcpnotify_kern.c
 create mode 100644 tools/testing/selftests/bpf/test_tcpnotify_user.c