On Fri, 22 Feb 2019 18:28:53 -0800 Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> wrote: > First we introduce bpf_probe_kernel_read and bpf_probe_user_read and > introduce clang/gcc tooling to catch the mistakes. > Going over this 400+ places and manually grepping kernel sources > for __user keyword is not a great proposal if we want to keep those users. > Once we have this working we can remove bpf_probe_read() altogether. > Rejecting bpf prog at load time is a clear signal that user has to fix it > (instead of changing run-time behavior). > When the verifier gets even smarter it could potentially replace prob_read > with probe_kernel_read and probe_user_read when it has that type info. I was about to suggest this approach. Document that bpf_probe_read() is known to be buggy and will be deprecated in the future, and that all new bpf scripts should start using bpf_probe_kernel/user_read() instead (after they have been implemented of course). And give time for people to fix their current scripts. Perhaps in the near future, trigger some kind of warning for users that use bpf_probe_read(). -- Steve
