On Tue, Sep 28, 2021 at 10:13:45PM +0800, Hou Tao wrote: > Hi Martin, > > During the testing of bpf_tcp_ca, I found that if the test program > aborts before calling bpf_link__detach_struct_ops(), the registered > bpf_dctcp will not be unregistered, and running bpf_tcp_ca test again > will fail with -EEXIST error as shown below: > > test_dctcp:PASS:bpf_dctcp__open_and_load 0 nsec > test_dctcp:FAIL:bpf_map__attach_struct_ops unexpected error: -17 > > The root cause is that the release of BPF_MAP_TYPE_STRUCT_OPS fd > neither put struct_ops programs in maps nor unregister the struct_ops > from kernel. Was the implementation intentional, or was it an oversight ? > If it is an oversight, I will post a patch to fix it. The original use case for the struct_ops is to register and unregister by command like "bpftool struct_ops (un)register" and then the kernel sub-system (tcp here) owns it, instead of finding where the maps may be pinned (could be in multiple locations). More like the insmod/rmmod. I think you meant to unregister it in ".map_release_uref"? but it will break the existing contract above. I believe a better way to do the auto-unreg is to create a real bpf_link for struct_ops. bpf_link has been added to the xdp and cgroup attachments since then. I did not give too much thoughts on this yet. What bpf-tcp-cc are you implementing or it is for running the test_progs purpose only ?
