On Sat, Apr 8, 2023 at 6:34 AM andrea terzolo <andreaterzolo3@xxxxxxxxx> wrote: > > Il giorno mer 5 apr 2023 alle ore 00:32 Andrii Nakryiko > <andrii.nakryiko@xxxxxxxxx> ha scritto: > > > > On Thu, Mar 30, 2023 at 10:21 AM andrea terzolo > > <andreaterzolo3@xxxxxxxxx> wrote: > > > > > > Hi all! > > > > > > If I can I would like to ask one question about the > > > `libbpf_probe_bpf_prog_type` API. The idea is to use `fentry/fexit` > > > bpf progs only if they are available and fall back to simple `kprobes` > > > when they are not. Is there a way to probe `BPF_TRACE_FENTRY` support > > > with `libbpf` APIs? I was looking at `libbpf_probe_bpf_prog_type` API > > > but it seems to check the `prog_type` rather than the `attach_type`, > > > when I call it `libbpf_probe_bpf_prog_type(BPF_PROG_TYPE_TRACING, > > > NULL);` it returns `1` even if `fentry/fexit` progs are not supported > > > on my machine. Is there a way to probe this feature with other > > > `libbpf` APIs? > > > > > > > looking at libbpf probing code, for BPF_PROG_TYPE_TRACING we choose > > BPF_TRACE_FENTRY attach type automatically (because it doesn't really > > matter whether its BPF_TRACE_FEXIT or BPF_MODIFY_RETURN, they all are > > either supported or none is). We then expect that verifier will > > complain with "attach_btf_id 1 is not a function" error. If we do see > > that error, we know that verifier supports fentry/fexit programs *in > > principle*, which is what we are checking with > > libbpf_probe_bpf_prog_type(). > > Ok got it, thank you. My issue is that in my project I need to use > BPF_TRACE_RAW_TP programs that AFAIK don't require the support for bpf > trampoline, so they could be supported even if > BPF_TRACE_FENTRY/BPF_MODIFY_RETURN are not supported. This is what > happens on arm64 kernels where we have BPF_TRACE_RAW_TP but > BPF_TRACE_FENTRY/BPF_MODIFY_RETURN are still not supported... Right > now I'm using libbpf_probe_bpf_prog_type() to check the support for > BPF_TRACE_RAW_TP but this is just an approximation, probably the best > way to do that is to inject a small > BPF_TRACE_RAW_TP prog and check that it is correctly loaded. It seems > that libbpf doesn't provide APIs to do that, is it right? BPF_TRACE_RAW_TP is not a program type, so not sure what you are checking. fentry/fexit is BPF_PROG_TYPE_TRACING, while raw tracepoint is BPF_PROG_TYPE_RAW_TRACEPOINT, there shouldn't be any problem for you. > > > If kernel doesn't support fentry/fexit attachment for some specific > > function you'd like to attach to, that's a different matter. This > > would be equivalent to BPF_PROG_TYPE_KPROBE check -- we check if > > kprobes in general are supported, but not whether kprobing specific > > kernel function works. > > > > I assume by "not supported on my machine" you mean that you can't > > attach fentry/fexit to some function? If not, let me know, and we'd > > have to debug this further. > > Sorry, probably I was not so clear, with this statement I mean that > libbpf_probe_bpf_prog_type() returns 1 even if BPF_TRACE_FENTRY progs > cannot be attached into the kernel. [0] is an example of what I'm > doing. > 1. Check fentry support with libbpf_probe_bpf_prog_type > 2. Check fentry support with an approach similar to libbpf-tools (as > you suggested) > 3. Try to inject my real BPF programs. > > (2) (libbpf-tool check) is correctly able to detect that > BPF_TRACE_FENTRY progs are not supported, when we call > `bpf_raw_tracepoint_open` to attach the fentry prog, `524` is returned > so we understand that this program is not supported. On the other > side, (1) is not able to detect that programs are not supported, the > API returns `1` as if they were supported. Now I have to highlight > that this API is called libbpf_probe_bpf_prog_type and not > libbpf_probe_bpf_attach_type, so 1 could be the right return value > since BPF_PROG_TYPE_TRACING progs are effectively supported, for > example, attach_type BPF_TRACE_RAW_TP is supported, but some other > attach types like BPF_TRACE_FENTRY/BPF_MODIFY_RETURN are not. If this > API just checks for BPF_PROG_TYPE_TRACING support, probably the best > way I have to check if a specific attach type is supported is to > directly inject a small prog of this type, as libbpf-tool does. WDYT? Could it be that you are mixing up enum bpf_prog_type with enum bpf_attach_type when calling libbpf_probe_bpf_prog_type()? > > [0]: https://github.com/Andreagit97/BPF-perf-tests/blob/main/templates/fentry_attach.c > > > If you want to know if some function can be traced with fentry/fexit, > > check below helper function from libbpf-tools ([0]) > > > > bool fentry_can_attach(const char *name, const char *mod) > > > > > > [0] https://github.com/iovisor/bcc/blob/master/libbpf-tools/trace_helpers.c#LL1043-L1043C58 > > > Thank you for the pointer! > > > > > Thank you in advance for your time, > > > Andrea
