I sent out this RFC to get an idea if the approach suggested here would be something other people would also like to see. In addition, this cover letter mentions some concerns and questions that need answers before we can move to an acceptable implementation. This patch adds support for tracing eBPF XDP programs that get executed using the __BPF_PROG_RUN syscall. This is done by switching from JIT (if enabled) to executing the program using the interpreter and record each executed instruction. For now, the execution history is printed to the kernel ring buffer using pr_info(), the final version should have enough data stored in a user-supplied buffer to reconstruct this output. This should probably be part of bpftool, i.e. dump a similar output, and the ability to store all this in an elf-like format for dumping/analyzing/replaying at a later stage. This patch does not dump the XDP packet content before and after execution, however, this data is available to the caller of the API. The __bpf_prog_run_trace() interpreter is a copy of __bpf_prog_run() and we probably need a smarter way to re-use the code rather than a blind copy with some changes. Enabling the interpreter opens up the kernel for spectre variant 2, guess that's why the BPF_JIT_ALWAYS_ON option was introduced (commit 290af86629b2). Enabling it for debugging in the field does not sound like an option (talking to people doing kernel distributions). Any idea how to work around this (lfence before any call this will slow down, but I guess for debugging this does not matter)? I need to research this more as I'm no expert in this area. But I think this needs to be solved as I see this as a show stopper. So any input is welcome. To allow bpf_call support for tracing currently the general interpreter is enabled. See the fixup_call_args() function for why this is needed. We might need to find a way to fix this (see the above section on spectre). Signed-off-by: Eelco Chaudron <echaudro@xxxxxxxxxx>
