Re: fentry/fexit attach to EXT type XDP program does not work

On Wed, Jul 29, 2020 at 08:23:56AM +0200, Eelco Chaudron wrote:

SNIP

quoted

quoted

quoted

quoted

a patch
that would nice.
You can also send it to me before bpf-next opens and I can verify
it, and
clean up the self-test so it can be included as well.

hi,
it seems that you cannot exten fentry/fexit programs,
but it's possible to attach fentry/fexit to ext program.

   /* Program extensions can extend all program types
    * except fentry/fexit. The reason is the following.
    * The fentry/fexit programs are used for performance
    * analysis, stats and can be attached to any program
    * type except themselves. When extension program is
    * replacing XDP function it is necessary to allow
    * performance analysis of all functions. Both original
    * XDP program and its program extension. Hence
    * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
    * allowed. If extending of fentry/fexit was allowed it
    * would be possible to create long call chain
    * fentry->extension->fentry->extension beyond
    * reasonable stack size. Hence extending fentry is not
    * allowed.
    */

I changed fexit_bpf2bpf.c test just to do a quick check
and it seems to work:

Hi Jiri this is exactly what I’m trying, however when you do this
where the
first argument is a pointer to some context data which you are
accessing
it’s failing in the verifier.
This is a link to the original email, which has a test patch
attached that
will show the failure when trying to load/attach the fentry function
and
access the context:

https://lore.kernel.org/bpf/159162546868.10791.12432342618156330247.stgit@ebuild/ (local)

ok, I tried to trace ext program with __sk_buff argument and I can see
the issue as well.. can't acess the skb argument

patch below fixes it for me, I can access the skb pointer and its data
via probe read, like:

	SEC("fexit/new_get_skb_ifindex")
	int BPF_PROG(fexit_new_get_skb_ifindex, int val, struct __sk_buff *skb,
int var, int ret)
	{
		__u32 data;
		int err;

		bpf_printk("EXIT skb %p", skb);
		bpf_probe_read_kernel(&data, sizeof(data), &skb->data);
		bpf_printk("EXIT ret %d, data %p", err, data);
		return 0;
	}

I think it should fix the xdp_md acess as well

Excellent patch ;) It works with xdp_md as well, and even better it does not
require the bpf_probe_read_kernel(), so the test_xdp_bpf2bpf.c code just
works.

great ;-) will check on xdp_md

Are you planning to send the patch upstream?

yep, I'll add some test for that and send it

thanks,
jirka