On 8 Jun 2020, at 18:58, Yonghong Song wrote:
On 6/8/20 7:11 AM, Eelco Chaudron wrote:
I'm trying for a while to do a fentry/fexit trace an EXT program
attached to an XDP program. To make it easier to explain I've
created a test case (see patch below) to show the issue.
Without the changes to test_xdp_bpf2bpf.c I'll get the following
error:
libbpf: -- BEGIN DUMP LOG ---
libbpf:
arg#0 type is not a struct
Unrecognized arg#0 type PTR
; int BPF_PROG(trace_on_entry, struct xdp_buff *xdp)
0: (79) r6 = *(u64 *)(r1 +0)
invalid bpf_context access off=0 size=8
processed 1 insns (limit 1000000) max_states_per_insn 0
total_states 0 peak_states 0 mark_read 0
libbpf: -- END LOG --
libbpf: failed to load program 'fentry/FUNC'
libbpf: failed to load object 'test_xdp_bpf2bpf'
libbpf: failed to load BPF skeleton 'test_xdp_bpf2bpf': -4007
test_xdp_fentry_ext:FAIL:__load ftrace skeleton failed
#91 xdp_fentry_ext:FAIL
Summary: 0/0 PASSED, 0 SKIPPED, 1 FAILED
With the change I get the following (but I do feel this change
should not be needed):
libbpf: -- BEGIN DUMP LOG ---
libbpf:
Unrecognized arg#0 type PTR
; int trace_on_entry(struct xdp_buff *xdp)
0: (bf) r6 = r1
; void *data = (void *)(long)xdp->data;
1: (79) r1 = *(u64 *)(r6 +0)
invalid bpf_context access off=0 size=8
processed 2 insns (limit 1000000) max_states_per_insn 0
total_states 0 peak_states 0 mark_read 0
libbpf: -- END LOG --
libbpf: failed to load program 'fentry/FUNC'
libbpf: failed to load object 'test_xdp_bpf2bpf'
libbpf: failed to load BPF skeleton 'test_xdp_bpf2bpf': -4007
test_xdp_fentry_ext:FAIL:__load ftrace skeleton failed
#91 xdp_fentry_ext:FAIL
Summary: 0/0 PASSED, 0 SKIPPED, 1 FAILED
Any idea what could be the case here? The same fentry/fexit attach
code works fine in the xdp_bpf2bpf.c tests case.
<SNIP>
I think this is not supported now. That is, you cannot attach a fentry
trace
to the EXT program. The current implementation for fentry program
simply
trying to find and match the signature of freplace program which by
default
is a pointer to void.
It is doable in that in kernel we could recognize to-be-attached
program is
a freplace and further trace down to find the real signature. The
related
kernel function is btf_get_prog_ctx_type(). You can try to implement
by yourself
or I can have a patch for this once bpf-next opens.
I’m not familiar with this area of the code, so if you could prepare 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.
