Signal delivery during connect() may disconnect an already established socket. Problem is that such socket might have been placed in a sockmap before the connection was closed.
PATCH 1 ensures this race won't lead to an unconnected vsock staying in the sockmap. PATCH 2 selftests it.
PATCH 3 fixes a related race. Note that here the race window is rather difficult to hit and I can't think of an easy way of testing it.
Signed-off-by: Michal Luczaj mhal@rbox.co --- Changes in v2: - Handle one more path of tripping the warning - Add a selftest - Collect R-b [Stefano] - Link to v1: https://lore.kernel.org/r/20250307-vsock-trans-signal-race-v1-1-3aca3f771fbd...
--- Michal Luczaj (3): vsock/bpf: Fix EINTR connect() racing sockmap update selftest/bpf: Add test for AF_VSOCK connect() racing sockmap update vsock/bpf: Fix bpf recvmsg() racing transport reassignment
net/vmw_vsock/af_vsock.c | 10 +- net/vmw_vsock/vsock_bpf.c | 24 +++-- .../selftests/bpf/prog_tests/sockmap_basic.c | 111 +++++++++++++++++++++ 3 files changed, 136 insertions(+), 9 deletions(-) --- base-commit: da9e8efe7ee10e8425dc356a9fc593502c8e3933 change-id: 20250305-vsock-trans-signal-race-d62f7718d099
Best regards,