On 2021-10-20 06:28, John Fastabend wrote:
Maxim Mikityanskiy wrote:
bpf_tcp_check_syncookie returns errors when SYN cookie generation is
disabled (EINVAL) or when no cookies were recently generated (ENOENT).
The same error codes are used for other kinds of errors: invalid
parameters (EINVAL), invalid packet (EINVAL, ENOENT), bad cookie
(ENOENT). Such an overlap makes it impossible for a BPF program to
distinguish different cases that may require different handling.
I'm not sure we can change these errors now. They are embedded in
the helper API. I think a BPF program could uncover the meaning
of the error anyways with some error path handling?
Anyways even if we do change these most of us who run programs
on multiple kernel versions would not be able to rely on them
being one way or the other easily.
The thing is, the error codes aren't really documented:
* 0 if *iph* and *th* are a valid SYN cookie ACK, or a negative
* error otherwise.
My patch doesn't break this assumption.
Practically speaking, there are two use cases of bpf_tcp_check_syncookie
that I know about: traffic classification (find NEW ACK packets with the
right cookie) and SYN flood protection.
For traffic classification, it's not important what error code we get.
The logic for ACK packets is as follows:
1. Connection established => ESTABLISHED. Otherwise,
2. bpf_tcp_check_syncookie returns 0 => NEW. Otherwise,
3. INVALID (regardless of the specific error code).
My patch doesn't break this use case.
For a BPF program that accelerates generating and checking SYN cookies,
typical logic looks like this (with current error codes annotated):
1. Drop invalid packets (EINVAL, ENOENT).
2. Drop packets with bad cookies (ENOENT).
3. Pass packets with good cookies (0).
4. Pass all packets when cookies are not in use (EINVAL, ENOENT).
Now that I'm reflecting on it again, it would make more sense to drop
packets in case 4: it's a new packet, it's an ACK, and we don't expect
any cookies.
The last point also matches the behavior of cookie_v4_check and
cookie_v6_check that skip all checks if cookie generation is disabled or
no cookies were recently generated. Overlapping error codes, however,
make it impossible to distinguish case 4 from cases 1 and 2.
If so, we don't strictly need to distinguish case 4 from 1 and 2. The
logic for ACK packets is similar:
1. Connection established => XDP_PASS. Otherwise,
2. bpf_tcp_check_syncookie returns 0 => XDP_PASS. Otherwise,
3. XDP_DROP.
So, on one hand, it looks like both use cases can be implemented without
this patch. On the other hand, changing error codes to more meaningful
shouldn't break existing programs and can have its benefits, for
example, in debugging or in statistic counting.
The original commit message of commit 399040847084 ("bpf: add helper to
check for a valid SYN cookie") mentions another use case, though:
traffic classification, where it's important to distinguish new
connections from existing ones, and case 4 should be distinguishable
from case 3.
To match the requirements of both use cases, this patch reassigns error
codes of bpf_tcp_check_syncookie and adds missing documentation:
1. EINVAL: Invalid packets.
2. EACCES: Packets with bad cookies.
3. 0: Packets with good cookies.
4. ENOENT: Cookies are not in use.
This way all four cases are easily distinguishable.
Signed-off-by: Maxim Mikityanskiy <maximmi@xxxxxxxxxx>
Reviewed-by: Tariq Toukan <tariqt@xxxxxxxxxx>
At very leasst this would need a fixes tag and should be backported
as a bug. Then we at least have a chance stable and LTS kernels
report the same thing.
That's a good idea.
[...]
--- a/net/core/filter.c
+++ b/net/core/filter.c
I'll take a stab at how a program can learn the error cause today.
BPF_CALL_5(bpf_tcp_check_syncookie, struct sock *, sk, void *, iph, u32, iph_len,
struct tcphdr *, th, u32, th_len)
{
#ifdef CONFIG_SYN_COOKIES
u32 cookie;
int ret;
// BPF program should know it pass bad values and can check
if (unlikely(!sk || th_len < sizeof(*th)))
return -EINVAL;
// sk_protocol and sk_state are exposed in sk and can be read directly
/* sk_listener() allows TCP_NEW_SYN_RECV, which makes no sense here. */
if (sk->sk_protocol != IPPROTO_TCP || sk->sk_state != TCP_LISTEN)
return -EINVAL;
// This is a user space knob right? I think this is a misconfig user can
// check before loading a program with check_syncookie?
bpf_tcp_check_syncookie was initially introduced for the classification
use case, to be able to classify new ACK packets with the right cookie
as NEW. The XDP program classifies traffic regardless of whether SYN
cookies are enabled. If we need to check the sysctl in userspace, it
means we need two XDP programs (or additional trickery passing this
value through a map).
if (!sock_net(sk)->ipv4.sysctl_tcp_syncookies)
return -EINVAL;
// We have th pointer can't we just check?
Yes, most of the checks can be repeated in BPF, but it's obvious it's
slower to do all the checks twice.
if (!th->ack || th->rst || th->syn)
return -ENOENT;
if (tcp_synq_no_recent_overflow(sk))
return -ENOENT;
This specific check can't be done in BPF.
cookie = ntohl(th->ack_seq) - 1;
switch (sk->sk_family) {
case AF_INET:
// misconfiguration but can be checked.
if (unlikely(iph_len < sizeof(struct iphdr)))
return -EINVAL;
ret = __cookie_v4_check((struct iphdr *)iph, th, cookie);
break;
#if IS_BUILTIN(CONFIG_IPV6)
case AF_INET6:
// misconfiguration can check as well
if (unlikely(iph_len < sizeof(struct ipv6hdr)))
return -EINVAL;
ret = __cookie_v6_check((struct ipv6hdr *)iph, th, cookie);
break;
#endif /* CONFIG_IPV6 */
default:
return -EPROTONOSUPPORT;
}
if (ret > 0)
return 0;
return -ENOENT;
#else
return -ENOTSUPP;
#endif
}
So I guess my point is we have all the fields we could write a bit
of BPF to find the error cause if necessary. Might be better than
dealing with changing the error code and having to deal with the
differences in kernels. I do see how it would have been better
to get errors correct on the first patch though :/
By the way I haven't got to the next set of patches with the
actual features, but why not push everything above this patch
as fixes in its own series. Then the fixes can get going why
we review the feature.
OK, I'll respin the fixes separately, while the discussion on the
approach to expose conntrack is going on.
Thanks for reviewing!
Thanks,
John
