On 20/04/2023 19:42, Stanislav Fomichev wrote:
On 04/20, Gilad Sever wrote:
When calling socket lookup from L2 (tc, xdp), VRF boundaries aren't respected. This patchset fixes this by regarding the incoming device's VRF attachment when performing the socket lookups from tc/xdp.
The first two patches are coding changes which facilitate this fix by factoring out the tc helper's logic which was shared with cg/sk_skb (which operate correctly).
Why is not relevant for cgroup/egress? Is it already running with the correct device?
Yes.
Also, do we really need all this refactoring and separate paths? Can we just add that bpf_l2_sdif part to the existing code? It will trigger for tc, but I'm assuming it will be a no-op for cgroup path?
The reason we preferred the refactoring is to avoid triggering `inet_sdif()` from tc/xdp. This is because in our understanding, the IPCB is undefined before IP processing so it seems incorrect to use `inet_sdif()` from tc/xdp.
We did come up with a different option which could spare most of the refactoring and still partially separate the two paths:
Pass sdif to __bpf_skc_lookup() but instead of using different functions for tc, calculate sdif by calling `dev_sdif()` in bpf_skc_lookup() only when netif_is_l3_master() is false. In other words:
- xdp callers would check the device's l3 enslaved state using the new `dev_sdif()` - sock_addr callers would use inet{,6}_sdif() as they did before - cg/tc share the same code path, so when netif_is_l3_master() is true use inet{,6}_sdif() and when it is false use dev_sdif(). this relies on the following assumptions: - tc programs don't run on l3 master devices - cgroup callers never see l3 enslaved devices - inet{,6}_sdif() isn't relevant for non l3 master devices
In our opinion, it's safer to factor out the tc flow as in the patchset, similar to XDP which has its own functions.
What do you think?
And regarding bpf_l2_sdif: seems like it's really generic and should probably be called something like dev_sdif?
Agreed. I'll rename in the next patch.
The third patch contains the actual bugfix.
The fourth patch adds bpf tests for these lookup functions.
v2: Fixed uninitialized var in test patch (4).
Gilad Sever (4): bpf: factor out socket lookup functions for the TC hookpoint. bpf: Call __bpf_sk_lookup()/__bpf_skc_lookup() directly via TC hookpoint bpf: fix bpf socket lookup from tc/xdp to respect socket VRF bindings selftests/bpf: Add tc_socket_lookup tests
net/core/filter.c | 132 +++++-- .../bpf/prog_tests/tc_socket_lookup.c | 341 ++++++++++++++++++ .../selftests/bpf/progs/tc_socket_lookup.c | 73 ++++ 3 files changed, 525 insertions(+), 21 deletions(-) create mode 100644 tools/testing/selftests/bpf/prog_tests/tc_socket_lookup.c create mode 100644 tools/testing/selftests/bpf/progs/tc_socket_lookup.c
-- 2.34.1