Skip to contentSkip to conversation

bpf: Add rcu ptr in btf_id_sock_common_types

PATCH In review Philo Lu 2 participants 5 messages Oct 8, 2024 – Oct 11, 2024

Subsystems (2)
lore
Share & tools
lore

Philo Lu

author
#0 lore raw

PATCH 1/1 bpf: Add rcu ptr in btf_id_sock_common_types

Sometimes sk is dereferenced as an rcu ptr, such as skb->sk in tp_btf, which is a valid type of sock common. Then helpers like bpf_skc_to_*() can be used with skb->sk.

For example, the following prog will be rejected without this patch:

```
SEC("tp_btf/tcp_bad_csum")
int BPF_PROG(tcp_bad_csum, struct sk_buff* skb)
{
	struct sock *sk = skb->sk;
	struct tcp_sock *tp;

	if (!sk)
		return 0;
	tp = bpf_skc_to_tcp_sock(sk);

	return 0;
}
```
---
 kernel/bpf/verifier.c | 1 +
 1 file changed, 1 insertion(+)
kernel/bpf/verifier.c +1 -0
index 9a7ed527e47e..3e7ce448ae03 100644
@@ -8362,6 +8362,7 @@ static const struct bpf_reg_types btf_id_sock_common_types = {
 		PTR_TO_XDP_SOCK,
 		PTR_TO_BTF_ID,
 		PTR_TO_BTF_ID | PTR_TRUSTED,
+		PTR_TO_BTF_ID | MEM_RCU,
 	},
 	.btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
 };
Signature · 1 line
2.32.0.3.g01195cf9f
#1 lore raw
Quoted 8 lines · Philo Lu

Sometimes sk is dereferenced as an rcu ptr, such as skb->sk in tp_btf, which is a valid type of sock common. Then helpers like bpf_skc_to_*() can be used with skb->sk.

For example, the following prog will be rejected without this patch:
```
SEC("tp_btf/tcp_bad_csum")
int BPF_PROG(tcp_bad_csum, struct sk_buff* skb)

Quoted 7 lines
{
	struct sock *sk = skb->sk;
	struct tcp_sock *tp;

	if (!sk)
		return 0;
	tp = bpf_skc_to_tcp_sock(sk);

If the use case is for reading the fields in tp, please use the bpf_core_cast from the libbpf's bpf_core_read.h. bpf_core_cast is using the bpf_rdonly_cast kfunc underneath.

pw-bot: cr

author
#2 lore raw
Quoted 13 lines · Martin KaFai Lau
Quoted 12 lines · Philo Lu

Sometimes sk is dereferenced as an rcu ptr, such as skb->sk in tp_btf, which is a valid type of sock common. Then helpers like bpf_skc_to_*() can be used with skb->sk.

For example, the following prog will be rejected without this patch:
```
SEC("tp_btf/tcp_bad_csum")
int BPF_PROG(tcp_bad_csum, struct sk_buff* skb)
{
    struct sock *sk = skb->sk;
    struct tcp_sock *tp;

Quoted 6 lines
Quoted 3 lines

    if (!sk)
        return 0;
    tp = bpf_skc_to_tcp_sock(sk);

If the use case is for reading the fields in tp, please use the bpf_core_cast from the libbpf's bpf_core_read.h. bpf_core_cast is using the bpf_rdonly_cast kfunc underneath.

Thank you! This works for me so this patch is unnecessary then.

Just curious is there any technical issue to include rcu_ptr into btf_id_sock_common_types? AFAICT rcu_ptr should also be a valid ptr type, and then btf_id_sock_common_types will behave like (PTR_TO_BTF_ID + &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON]) in bpf_func_proto.

Thanks.

Signature · 1 line
Philo
#3 lore raw
Quoted 20 lines · Philo Lu
Quoted 17 lines · Martin KaFai Lau
Quoted 15 lines · Philo Lu

Sometimes sk is dereferenced as an rcu ptr, such as skb->sk in tp_btf, which is a valid type of sock common. Then helpers like bpf_skc_to_*() can be used with skb->sk.

For example, the following prog will be rejected without this patch:
```
SEC("tp_btf/tcp_bad_csum")
int BPF_PROG(tcp_bad_csum, struct sk_buff* skb)
{
    struct sock *sk = skb->sk;
    struct tcp_sock *tp;

    if (!sk)
        return 0;
    tp = bpf_skc_to_tcp_sock(sk);

Quoted 6 lines
Quoted 2 lines

If the use case is for reading the fields in tp, please use the bpf_core_cast from the libbpf's bpf_core_read.h. bpf_core_cast is using the bpf_rdonly_cast kfunc underneath.

Thank you! This works for me so this patch is unnecessary then.

Just curious is there any technical issue to include rcu_ptr into btf_id_sock_common_types? AFAICT rcu_ptr should also be a valid ptr type, and then btf_id_sock_common_types will behave like (PTR_TO_BTF_ID + &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON]) in bpf_func_proto.

bpf_skc_to_*() returns a PTR_TO_BTF_ID which can be passed into other helpers that takes ARG_PTR_TO_BTF_ID_SOCK_COMMON. There are helpers that change the sk. e.g. bpf_setsockopt() changes the sk and needs sk to be locked. Other non tracing hooks do have a hold on the skb also. I did take a quick look at the bpf_setsockopt situation and looks ok. I am positive there are other helpers that need to audit first.

Tracing use case should only read the sk. bpf_core_cast() is the correct one to use. The bpf_sk_storage_{get,delete}() should be the only allowed helper that can change the sk.

author
#4 lore raw
Quoted 24 lines · Martin KaFai Lau
Quoted 23 lines · Philo Lu
Quoted 19 lines · Martin KaFai Lau
Quoted 12 lines · Philo Lu

Sometimes sk is dereferenced as an rcu ptr, such as skb->sk in tp_btf, which is a valid type of sock common. Then helpers like bpf_skc_to_*() can be used with skb->sk.

For example, the following prog will be rejected without this patch:
```
SEC("tp_btf/tcp_bad_csum")
int BPF_PROG(tcp_bad_csum, struct sk_buff* skb)
{
    struct sock *sk = skb->sk;
    struct tcp_sock *tp;

Quoted 3 lines

    if (!sk)
        return 0;
    tp = bpf_skc_to_tcp_sock(sk);

If the use case is for reading the fields in tp, please use the bpf_core_cast from the libbpf's bpf_core_read.h. bpf_core_cast is using the bpf_rdonly_cast kfunc underneath.

Quoted 7 lines
Quoted 3 lines

Thank you! This works for me so this patch is unnecessary then.

Just curious is there any technical issue to include rcu_ptr into btf_id_sock_common_types? AFAICT rcu_ptr should also be a valid ptr type, and then btf_id_sock_common_types will behave like (PTR_TO_BTF_ID + &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON]) in bpf_func_proto.

bpf_skc_to_*() returns a PTR_TO_BTF_ID which can be passed into other helpers that takes ARG_PTR_TO_BTF_ID_SOCK_COMMON. There are helpers that change the sk. e.g. bpf_setsockopt() changes the sk and needs sk to be locked. Other non tracing hooks do have a hold on the skb also. I did take a quick look at the bpf_setsockopt situation and looks ok. I am positive there are other helpers that need to audit first.

Tracing use case should only read the sk. bpf_core_cast() is the correct one to use. The bpf_sk_storage_{get,delete}() should be the only allowed helper that can change the sk.

Thank you for explanation, Martin. This helps me a lot.

Signature · 1 line
Philo