Can eBPF's perf_submit() be used in a socket_filter program as well?

0

So I was trying to send some data from the kernel space program to the user space program using perf_submit.

I've done some studies and here(https://github.com/iovisor/bcc/issues/2423), yonghong-song answered(the last comment) that a socket_filter program can not access bpf_perf_event_output helper and therefore it can only be used for tracing program types.

However, on BCC reference site(https://github.com/iovisor/bcc/blob/master/docs/reference_guide.md#2-bpf_perf_output), if you ctrl+f and search for : 3. perf_submit() , it says on the fifth line that "for SOCKET_FILTER programs, the struct __sk_buff *skb must be used instead." I believe this infers that perf_submit() can be used for socket_filter programs as well?

So I have hard time figuring out if perf_submit() can indeed be used for a socket filter program. Maybe some functionalities have been added since Yonghong-song answered the question above?

I'm checking if perf_submit() would work with a socket filter and there's not really a line of code that grabs the data output by perf_submit because just addint perf_submit() in the kernel program already omitted an error.

Here's the code for my program :

from bcc import BPF

# Network interface to be monoitored
INTERFACE = "br-netrome"

bpf_text = """

#include <uapi/linux/ptrace.h>
#include <net/sock.h>
#include <bcc/proto.h>
#include <linux/bpf.h>

#define IP_TCP 6
#define IP_UDP 17
#define IP_ICMP 1
#define ETH_HLEN 14

BPF_PERF_OUTPUT(events);    // has to be delcared outside any function

int packet_monitor(struct __sk_buff *skb) {
    u8 *cursor = 0;
    u64 saddr;
    u64 daddr;
    u64 ttl;
    u64 hchecksum;

    struct ethernet_t *ethernet = cursor_advance(cursor, sizeof(*ethernet));
    if (!(ethernet -> type == 0x0800)) {
        return 0; // drop
    }

    struct ip_t *ip = cursor_advance(cursor, sizeof(*ip));
    /*
    if (ip->nextp != IP_TCP) 
    {
        if (ip -> nextp != IP_UDP) 
        {
            if (ip -> nextp != IP_ICMP) 
                return 0; 
        }
    }
    */

    saddr = ip -> src;
    daddr = ip -> dst;
    ttl = ip -> ttl;
    hchecksum = ip -> hchecksum;

    events.perf_submit(skb, &saddr, sizeof(saddr));

//    bpf_trace_printk("saddr = %llu, daddr = %llu, ttl = %llu", saddr, daddr, ttl); // only three arguments can be passed using printk

//    bpf_trace_printk("Incoming packet!!\\n");
    return -1;
}

and here is the error code :

 R0=inv2048 R6=ctx(id=0,off=0,imm=0) R7=inv0 R10=fp0,call_-1
4: (20) r0 = *(u32 *)skb[26]
5: (7b) *(u64 *)(r10 -8) = r0
6: (18) r2 = 0xffff9bde204ffa00
8: (18) r7 = 0xffffffff
10: (bf) r4 = r10
11: (07) r4 += -8
12: (bf) r1 = r6
13: (18) r3 = 0xffffffff
15: (b7) r5 = 8
16: (85) call bpf_perf_event_output#25
unknown func bpf_perf_event_output#25

Traceback (most recent call last):
  File "packet_monitor.py", line 68, in <module>
    function_skb_matching = bpf.load_func("packet_monitor", BPF.SOCKET_FILTER)
  File "/usr/lib/python2.7/dist-packages/bcc/__init__.py", line 397, in load_func
    (func_name, errstr))
bpf
ebpf
asked on Stack Overflow Jan 7, 2020 by Rosè • edited Jan 7, 2020 by pchaigno

1 Answer

2

TL;DR. BPF programs of type BPF_PROG_TYPE_SOCKET_FILTER can use bpf_perf_event_output only starting with Linux 5.4.


Which helpers a given BPF program has access to is defined by the get_func_proto member of objects struct bpf_verifier_ops. You can find which bpf_verifier_ops object corresponds to which program type by reading function find_prog_type() and file bpf_types.h. In the case of BPF_PROG_TYPE_SOCKET_FILTER, the corresponding function is sk_filter_func_proto().

If you git blame that function on recent kernel sources, you will get something like the following (you can do the same with GitHub's blame feature):

$ git blame net/core/filter.c
[...]
2492d3b867043 (Daniel Borkmann          2017-01-24 01:06:27 +0100 6080) static const struct bpf_func_proto *
5e43f899b03a3 (Andrey Ignatov           2018-03-30 15:08:00 -0700 6081) sk_filter_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
2492d3b867043 (Daniel Borkmann          2017-01-24 01:06:27 +0100 6082) {
2492d3b867043 (Daniel Borkmann          2017-01-24 01:06:27 +0100 6083)         switch (func_id) {
2492d3b867043 (Daniel Borkmann          2017-01-24 01:06:27 +0100 6084)         case BPF_FUNC_skb_load_bytes:
2492d3b867043 (Daniel Borkmann          2017-01-24 01:06:27 +0100 6085)                 return &bpf_skb_load_bytes_proto;
4e1ec56cdc597 (Daniel Borkmann          2018-05-04 01:08:15 +0200 6086)         case BPF_FUNC_skb_load_bytes_relative:
4e1ec56cdc597 (Daniel Borkmann          2018-05-04 01:08:15 +0200 6087)                 return &bpf_skb_load_bytes_relative_proto;
91b8270f2a4d1 (Chenbo Feng              2017-03-22 17:27:34 -0700 6088)         case BPF_FUNC_get_socket_cookie:
91b8270f2a4d1 (Chenbo Feng              2017-03-22 17:27:34 -0700 6089)                 return &bpf_get_socket_cookie_proto;
6acc5c2910689 (Chenbo Feng              2017-03-22 17:27:35 -0700 6090)         case BPF_FUNC_get_socket_uid:
6acc5c2910689 (Chenbo Feng              2017-03-22 17:27:35 -0700 6091)                 return &bpf_get_socket_uid_proto;
7c4b90d79d0f4 (Allan Zhang              2019-07-23 17:07:24 -0700 6092)         case BPF_FUNC_perf_event_output:
7c4b90d79d0f4 (Allan Zhang              2019-07-23 17:07:24 -0700 6093)                 return &bpf_skb_event_output_proto;
2492d3b867043 (Daniel Borkmann          2017-01-24 01:06:27 +0100 6094)         default:
2492d3b867043 (Daniel Borkmann          2017-01-24 01:06:27 +0100 6095)                 return bpf_base_func_proto(func_id);
2492d3b867043 (Daniel Borkmann          2017-01-24 01:06:27 +0100 6096)         }
2492d3b867043 (Daniel Borkmann          2017-01-24 01:06:27 +0100 6097) }
[...]

As you can see, BPF_FUNC_perf_event_output was only recently added to the list of helpers these BPF programs can call. The commit which added this support, 7c4b90d79d0f4, was merged in Linux v5.4:

$ git describe --contains 7c4b90d79d0f4
v5.4-rc1~131^2~248^2~20
answered on Stack Overflow Jan 7, 2020 by pchaigno

User contributions licensed under CC BY-SA 3.0