netdev - Re: [PATCH bpf-next 4/9] bpf: Add bpf helper bpf_tcp_check_probe

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Date:   Tue, 19 Feb 2019 11:56:24 +0100
From:   Daniel Borkmann <daniel@...earbox.net>
To:     brakmo <brakmo@...com>, netdev <netdev@...r.kernel.org>
Cc:     Martin Lau <kafai@...com>, Alexei Starovoitov <ast@...com>,
        Daniel Borkmann --cc=Kernel Team 
        <"daniel@...earbox.netKernel-team"@fb.com>
Subject: Re: [PATCH bpf-next 4/9] bpf: Add bpf helper
 bpf_tcp_check_probe_timer

On 02/19/2019 06:38 AM, brakmo wrote:
> This patch adds a new bpf helper BPF_FUNC_tcp_check_probe_timer
> "int bpf_check_tcp_probe_timer(struct tcp_bpf_sock *tp, u32 when_us)".
> It is added to BPF_PROG_TYPE_CGROUP_SKB typed bpf_prog which currently
> can be attached to the ingress and egress path.
> 
> The function forces when_us to be at least TCP_TIMEOUT_MIN (currently
> 2 jiffies) and no more than TCP_RTO_MIN (currently 200ms).
> 
> When using a bpf_prog to limit the egress bandwidth of a cgroup,
> it can happen that we drop a packet of a connection that has no
> packets out. In this case, the connection may not retry sending
> the packet until the probe timer fires. Since the default value
> of the probe timer is at least 200ms, this can introduce link
> underutiliation (i.e. the cgroup egress bandwidth being smaller
> than the specified rate) thus increased tail latency.
> This helper function allows for setting a smaller probe timer.
> 
> Signed-off-by: Lawrence Brakmo <brakmo@...com>
> ---
>  include/uapi/linux/bpf.h | 12 +++++++++++-
>  net/core/filter.c        | 27 +++++++++++++++++++++++++++
>  2 files changed, 38 insertions(+), 1 deletion(-)
> 
> diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> index 5daf404511f7..a78936acccae 100644
> --- a/include/uapi/linux/bpf.h
> +++ b/include/uapi/linux/bpf.h
> @@ -2372,6 +2372,15 @@ union bpf_attr {
>   *		val should be one of 0, 1, 2, 3.
>   *	Return
>   *		-EINVAL on error (e.g. val > 3), 0 otherwise.
> + *
> + * int bpf_tcp_check_probe_timer(struct bpf_tcp_sock *tp, int when_us)
> + *	Description
> + *		Checks that there are no packets out and there is no pending
> + *		timer. If both of these are true, it bounds when_us by
> + *		TCP_TIMEOUT_MIN (2 jiffies) or TCP_RTO_MIN (200ms) and
> + *		sets the probe timer.
> + *	Return
> + *		0
>   */
>  #define __BPF_FUNC_MAPPER(FN)		\
>  	FN(unspec),			\
> @@ -2472,7 +2481,8 @@ union bpf_attr {
>  	FN(sk_fullsock),		\
>  	FN(tcp_sock),			\
>  	FN(tcp_enter_cwr),		\
> -	FN(skb_set_ecn),
> +	FN(skb_set_ecn),		\
> +	FN(tcp_check_probe_timer),
>  
>  /* integer value in 'imm' field of BPF_CALL instruction selects which helper
>   * function eBPF program intends to call
> diff --git a/net/core/filter.c b/net/core/filter.c
> index 275acfb2117d..2b975e651a04 100644
> --- a/net/core/filter.c
> +++ b/net/core/filter.c
> @@ -5465,6 +5465,31 @@ static const struct bpf_func_proto bpf_skb_set_ecn_proto = {
>  	.arg1_type	= ARG_PTR_TO_CTX,
>  	.arg2_type	= ARG_ANYTHING,
>  };
> +
> +BPF_CALL_2(bpf_tcp_check_probe_timer, struct tcp_sock *, tp, u32, when_us)
> +{
> +	struct sock *sk = (struct sock *) tp;
> +	unsigned long when = usecs_to_jiffies(when_us);
> +
> +	if (!tp->packets_out && !inet_csk(sk)->icsk_pending) {
> +		if (when < TCP_TIMEOUT_MIN)
> +			when = TCP_TIMEOUT_MIN;
> +		else if (when > TCP_RTO_MIN)
> +			when = TCP_RTO_MIN;
> +
> +		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
> +					  when, TCP_RTO_MAX);

Should this be using tcp_reset_xmit_timer() instead to take pacing
into account? (If not, would be good to have a comment explaining
why it's okay to use directly here.)

> +	}
> +	return 0;
> +}
> +
> +static const struct bpf_func_proto bpf_tcp_check_probe_timer_proto = {
> +	.func		= bpf_tcp_check_probe_timer,
> +	.gpl_only	= false,
> +	.ret_type	= RET_INTEGER,
> +	.arg1_type	= ARG_PTR_TO_TCP_SOCK,
> +	.arg2_type	= ARG_ANYTHING,
> +};
>  #endif /* CONFIG_INET */
>  
>  bool bpf_helper_changes_pkt_data(void *func)
> @@ -5628,6 +5653,8 @@ cg_skb_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
>  		return &bpf_tcp_enter_cwr_proto;
>  	case BPF_FUNC_skb_set_ecn:
>  		return &bpf_skb_set_ecn_proto;
> +	case BPF_FUNC_tcp_check_probe_timer:
> +		return &bpf_tcp_check_probe_timer_proto;
>  #endif
>  	default:
>  		return sk_filter_func_proto(func_id, prog);
>