Incorrect global variables usage for outputting skb and shinfo

[Asphaltt]

In PR #484 , when I confirm whether it's necesssary to share .bss map between different collections, I confirm that the global variables are used incorrectly.

What global variables in .bss map?

# bpftool m d n .bss
[{
        "value": {
            ".bss": [{
                    "set_skb_btf.p": {
                        "ptr": "(nil)",
                        "type_id": 0,
                        "flags": 0
                    }
                },{
                    "set_skb_btf.v": {
                        "len": 0,
                        "str": ""
                    }
                },{
                    "set_shinfo_btf.p": {
                        "ptr": "(nil)",
                        "type_id": 0,
                        "flags": 0
                    }
                },{
                    "set_shinfo_btf.v": {
                        "len": 0,
                        "str": ""
                    }
                }
            ]
        }
    }
]

But wait, what are they?

// bpf/kprobe_pwru.c

static __always_inline void
set_skb_btf(struct sk_buff *skb, u64 *event_id) {
	static struct btf_ptr p = {};
	static struct print_skb_value v = {};
	u64 id;

	p.type_id = cfg->skb_btf_id;
	p.ptr = skb;
	*event_id = sync_fetch_and_add(&print_skb_id_map);

	v.len = bpf_snprintf_btf(v.str, PRINT_SKB_STR_SIZE, &p, sizeof(p), 0);
	if (v.len < 0) {
		return;
	}

	bpf_map_update_elem(&print_skb_map, event_id, &v, BPF_ANY);
}

static __always_inline void
set_shinfo_btf(struct sk_buff *skb, u64 *event_id) {
	struct skb_shared_info *shinfo;
	static struct btf_ptr p = {};
	static struct print_shinfo_value v = {};
	unsigned char *head;
	unsigned int end;

        /* skb_shared_info is located at the end of skb data.
         * When CONFIG_NET_SKBUFF_DATA_USES_OFFSET is enabled, skb->end
         * is an offset from skb->head to the end of skb data. If not,
         * skb->end is a pointer to the end of skb data. For amd64 and
         * arm64 (in 64bit arch in general), CONFIG_NET_SKBUFF_DATA_USES_OFFSET
	 * is enabled by default.
         */
        head = BPF_CORE_READ(skb, head);
	end = BPF_CORE_READ(skb, end);
	shinfo = (struct skb_shared_info *)(head + end);

	p.type_id = cfg->shinfo_btf_id;
	p.ptr = shinfo;

	*event_id = sync_fetch_and_add(&print_shinfo_id_map);

	v.len = bpf_snprintf_btf(v.str, PRINT_SHINFO_STR_SIZE, &p, sizeof(p), 0);
	if (v.len < 0) {
		return;
	}

	bpf_map_update_elem(&print_shinfo_map, event_id, &v, BPF_ANY);
}

Aha, they are used for outputting skb and shinfo.

But when there are more-than-1 running pwru bpf progs on different CPUs, they will share the same globals at the same time without memory protection.

To resolve the memory-race issue, these variables should be set as global percpu variables by utilizing percpu_array map.