On 4/24/24 1:37 PM, Miao Xu wrote:
+#define USEC_PER_SEC 1000000UL
+#define TCP_PACING_SS_RATIO (200)
+#define TCP_PACING_CA_RATIO (120)
+#define TCP_REORDERING (12)
+#define likely(x) (__builtin_expect(!!(x), 1))
+
+static __always_inline __u64 div64_u64(__u64 dividend, __u64 divisor)
+{
+ return dividend / divisor;
+}
+
+static void tcp_update_pacing_rate(struct sock *sk)
+{
+ const struct tcp_sock *tp = tcp_sk(sk);
+ __u64 rate;
+
+ /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
+ rate = (__u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
+
+ /* current rate is (cwnd * mss) / srtt
+ * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
+ * In Congestion Avoidance phase, set it to 120 % the current rate.
+ *
+ * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
+ * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
+ * end of slow start and should slow down.
+ */
+ if (tp->snd_cwnd < tp->snd_ssthresh / 2)
+ rate *= TCP_PACING_SS_RATIO;
+ else
+ rate *= TCP_PACING_CA_RATIO;
+
+ rate *= max(tp->snd_cwnd, tp->packets_out);
+
+ if (likely(tp->srtt_us))
+ rate = div64_u64(rate, (__u64)tp->srtt_us);
+
+ sk->sk_pacing_rate = min(rate, (__u64)sk->sk_max_pacing_rate);
+}
+
+static __always_inline void tcp_cwnd_reduction(
+ struct sock *sk,
+ int newly_acked_sacked,
+ int newly_lost,
+ int flag) {
+ struct tcp_sock *tp = tcp_sk(sk);
+ int sndcnt = 0;
+ __u32 pkts_in_flight = tp->packets_out - (tp->sacked_out + tp->lost_out) + tp->retrans_out;
+ int delta = tp->snd_ssthresh - pkts_in_flight;
+
+ if (newly_acked_sacked <= 0 || !tp->prior_cwnd)
+ return;
+
+ __u32 prr_delivered = tp->prr_delivered + newly_acked_sacked;
+
+ if (delta < 0) {
+ __u64 dividend =
+ (__u64)tp->snd_ssthresh * prr_delivered + tp->prior_cwnd - 1;
+ sndcnt = (__u32)div64_u64(dividend, (__u64)tp->prior_cwnd) - tp->prr_out;
+ } else {
+ sndcnt = max(prr_delivered - tp->prr_out, newly_acked_sacked);
+ if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost)
+ sndcnt++;
+ sndcnt = min(delta, sndcnt);
+ }
+ /* Force a fast retransmit upon entering fast recovery */
+ sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
+ tp->snd_cwnd = pkts_in_flight + sndcnt;
+}
+
+/* Decide wheather to run the increase function of congestion control. */
+static __always_inline bool tcp_may_raise_cwnd(
+ const struct sock *sk,
+ const int flag) {
+ if (tcp_sk(sk)->reordering > TCP_REORDERING)
+ return flag & FLAG_FORWARD_PROGRESS;
+
+ return flag & FLAG_DATA_ACKED;
+}
+
+void BPF_STRUCT_OPS(bpf_cubic_cong_control, struct sock *sk, __u32 ack, int flag,
+ const struct rate_sample *rs)
+{
+ struct tcp_sock *tp = tcp_sk(sk);
+
+ if (((1<<TCP_CA_CWR) | (1<<TCP_CA_Recovery)) &
+ (1 << inet_csk(sk)->icsk_ca_state)) {
+ /* Reduce cwnd if state mandates */
+ tcp_cwnd_reduction(sk, rs->acked_sacked, rs->losses, flag);
+
+ if (!before(tp->snd_una, tp->high_seq)) {
+ /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
+ if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
+ inet_csk(sk)->icsk_ca_state == TCP_CA_CWR) {
+ tp->snd_cwnd = tp->snd_ssthresh;
+ tp->snd_cwnd_stamp = tcp_jiffies32;
+ }
+ // __cwnd_event(sk, CA_EVENT_COMPLETE_CWR);
+ }
+ } else if (tcp_may_raise_cwnd(sk, flag)) {
+ /* Advance cwnd if state allows */
+ cubictcp_cong_avoid(sk, ack, rs->acked_sacked);
+ tp->snd_cwnd_stamp = tcp_jiffies32;
+ }
+
+ tcp_update_pacing_rate(sk);
It will be useful to highlight what you want to do differently from the kernel's
tcp_cong_control()+tcp_cong_avoid() here. or it is something that I missed from
the above example?
