Skip to contentSkip to conversation

Best way to check for fentry attach support

DISCUSSION Martin Kelly 2 participants 3 messages Sep 12, 2023 – Sep 14, 2023 BPF

lore
Share & tools
lore
#0 lore raw

Hi all,

I'm trying to figure out the best way to handle the fact that fentry/fexit trampolines are not fully supported on all architectures and kernel versions. As an example, I want to be able to load an fentry if the kernel supports it, and a kprobe otherwise.

It's tempting to use libbpf_probe_bpf_prog_type for this, but on ARM64 kernels >= 5.5 (when BPF trampolines were introduced) but before the most recent ones, loading an fentry program will pass, but attaching it will still fail. This also means that libbpf_probe_bpf_prog_type will return true even if the program can't be attached, so that can't be used to test for attachability.

I can work around this by attempting to attach a dummy fentry program in my application, but I'm wondering if this is something that should be done more generally by libbpf. Some possible ways to do this are:

- Extend the libbpf_probe API to add libbpf_probe_trampoline or similar, attempting attach to a known-exported function, such as the BPF syscall, or to a user-specified symbol.

- Extend the libbpf_probe API to add a generic libbpf_probe_attach API to check if a given function is attachable. However, as attach code is different depending on the hook, this might be very complex and require a ton of parameters.

- Maybe there are other options that I haven't thought of.

I have a patch I could send for libbpf_probe_trampoline, but I wanted to first check if this is a good idea or if it's preferred to simply have applications probe this themselves.

Thanks,

Martin

#1 lore raw
Quoted 8 lines · Martin Kelly

Hi all,

I'm trying to figure out the best way to handle the fact that fentry/fexit trampolines are not fully supported on all architectures and kernel versions. As an example, I want to be able to load an fentry if the kernel supports it, and a kprobe otherwise.

Quoted 6 lines

It's tempting to use libbpf_probe_bpf_prog_type for this, but on ARM64 kernels >= 5.5 (when BPF trampolines were introduced) but before the most recent ones, loading an fentry program will pass, but attaching it will still fail. This also means that libbpf_probe_bpf_prog_type will return true even if the program can't be attached, so that can't be used to test for attachability.

Right, because libbpf_probe_bpf_prog_type() is testing whether given program type can be loaded, not attached.

Quoted 14 lines

I can work around this by attempting to attach a dummy fentry program in my application, but I'm wondering if this is something that should be done more generally by libbpf. Some possible ways to do this are:

- Extend the libbpf_probe API to add libbpf_probe_trampoline or similar, attempting attach to a known-exported function, such as the BPF syscall, or to a user-specified symbol.

- Extend the libbpf_probe API to add a generic libbpf_probe_attach API to check if a given function is attachable. However, as attach code is different depending on the hook, this might be very complex and require a ton of parameters.

Quoted 5 lines

- Maybe there are other options that I haven't thought of.

I have a patch I could send for libbpf_probe_trampoline, but I wanted to first check if this is a good idea or if it's preferred to simply have applications probe this themselves.

It doesn't seem too hard for an application to try to attach and if attachment fails fallback to attaching kprobe-based program. So I'd prefer that over much more maintenance burden of keeping this "can attach" generic API. At least for now.

Quoted 6 lines

Thanks,

Martin

#2 lore raw
Quoted 33 lines · Andrii Nakryiko
Quoted 7 lines · Martin Kelly

Hi all,

I'm trying to figure out the best way to handle the fact that fentry/fexit trampolines are not fully supported on all architectures and kernel versions. As an example, I want to be able to load an fentry if the kernel supports it, and a kprobe otherwise.

Quoted 6 lines

It's tempting to use libbpf_probe_bpf_prog_type for this, but on ARM64 kernels >= 5.5 (when BPF trampolines were introduced) but before the most recent ones, loading an fentry program will pass, but attaching it will still fail. This also means that libbpf_probe_bpf_prog_type will return true even if the program can't be attached, so that can't be used to test for attachability.

Right, because libbpf_probe_bpf_prog_type() is testing whether given program type can be loaded, not attached.

Quoted 15 lines

I can work around this by attempting to attach a dummy fentry program in my application, but I'm wondering if this is something that should be done more generally by libbpf. Some possible ways to do this are:

- Extend the libbpf_probe API to add libbpf_probe_trampoline or similar, attempting attach to a known-exported function, such as the BPF syscall, or to a user-specified symbol.

- Extend the libbpf_probe API to add a generic libbpf_probe_attach API to check if a given function is attachable. However, as attach code is different depending on the hook, this might be very complex and require a ton of parameters.

- Maybe there are other options that I haven't thought of.

Quoted 7 lines
Quoted 3 lines

I have a patch I could send for libbpf_probe_trampoline, but I wanted to first check if this is a good idea or if it's preferred to simply have applications probe this themselves.

It doesn't seem too hard for an application to try to attach and if attachment fails fallback to attaching kprobe-based program. So I'd prefer that over much more maintenance burden of keeping this "can attach" generic API. At least for now.

OK, it's easy enough to do it in the application, so we'll do that. Thanks!

Quoted 5 lines
Quoted 5 lines

Thanks,

Martin