1 // SPDX-License-Identifier: GPL-2.0-only 2 3 /* Highlights: 4 * 1. The major difference between this bpf program and tcp_cubic.c 5 * is that this bpf program relies on `cong_control` rather than 6 * `cong_avoid` in the struct tcp_congestion_ops. 7 * 2. Logic such as tcp_cwnd_reduction, tcp_cong_avoid, and 8 * tcp_update_pacing_rate is bypassed when `cong_control` is 9 * defined, so moving these logic to `cong_control`. 10 * 3. WARNING: This bpf program is NOT the same as tcp_cubic.c. 11 * The main purpose is to show use cases of the arguments in 12 * `cong_control`. For simplicity's sake, it reuses tcp cubic's 13 * kernel functions. 14 */ 15 16 #include "bpf_tracing_net.h" 17 #include <bpf/bpf_helpers.h> 18 #include <bpf/bpf_tracing.h> 19 20 #define USEC_PER_SEC 1000000UL 21 #define TCP_PACING_SS_RATIO (200) 22 #define TCP_PACING_CA_RATIO (120) 23 #define TCP_REORDERING (12) 24 25 extern void cubictcp_init(struct sock *sk) __ksym; 26 extern void cubictcp_cwnd_event(struct sock *sk, enum tcp_ca_event event) __ksym; 27 extern __u32 cubictcp_recalc_ssthresh(struct sock *sk) __ksym; 28 extern void cubictcp_state(struct sock *sk, __u8 new_state) __ksym; 29 extern __u32 tcp_reno_undo_cwnd(struct sock *sk) __ksym; 30 extern void cubictcp_acked(struct sock *sk, const struct ack_sample *sample) __ksym; 31 extern void cubictcp_cong_avoid(struct sock *sk, __u32 ack, __u32 acked) __ksym; 32 33 static __u64 div64_u64(__u64 dividend, __u64 divisor) 34 { 35 return dividend / divisor; 36 } 37 38 static void tcp_update_pacing_rate(struct sock *sk) 39 { 40 const struct tcp_sock *tp = tcp_sk(sk); 41 __u64 rate; 42 43 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ 44 rate = (__u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3); 45 46 /* current rate is (cwnd * mss) / srtt 47 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate. 48 * In Congestion Avoidance phase, set it to 120 % the current rate. 49 * 50 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh) 51 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching 52 * end of slow start and should slow down. 53 */ 54 if (tp->snd_cwnd < tp->snd_ssthresh / 2) 55 rate *= TCP_PACING_SS_RATIO; 56 else 57 rate *= TCP_PACING_CA_RATIO; 58 59 rate *= max(tp->snd_cwnd, tp->packets_out); 60 61 if (tp->srtt_us) 62 rate = div64_u64(rate, (__u64)tp->srtt_us); 63 64 sk->sk_pacing_rate = min(rate, sk->sk_max_pacing_rate); 65 } 66 67 static void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, 68 int newly_lost, int flag) 69 { 70 struct tcp_sock *tp = tcp_sk(sk); 71 int sndcnt = 0; 72 __u32 pkts_in_flight = tp->packets_out - (tp->sacked_out + tp->lost_out) + tp->retrans_out; 73 int delta = tp->snd_ssthresh - pkts_in_flight; 74 75 if (newly_acked_sacked <= 0 || !tp->prior_cwnd) 76 return; 77 78 __u32 prr_delivered = tp->prr_delivered + newly_acked_sacked; 79 80 if (delta < 0) { 81 __u64 dividend = 82 (__u64)tp->snd_ssthresh * prr_delivered + tp->prior_cwnd - 1; 83 sndcnt = (__u32)div64_u64(dividend, (__u64)tp->prior_cwnd) - tp->prr_out; 84 } else { 85 sndcnt = max(prr_delivered - tp->prr_out, newly_acked_sacked); 86 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost) 87 sndcnt++; 88 sndcnt = min(delta, sndcnt); 89 } 90 /* Force a fast retransmit upon entering fast recovery */ 91 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1)); 92 tp->snd_cwnd = pkts_in_flight + sndcnt; 93 } 94 95 /* Decide whether to run the increase function of congestion control. */ 96 static bool tcp_may_raise_cwnd(const struct sock *sk, const int flag) 97 { 98 if (tcp_sk(sk)->reordering > TCP_REORDERING) 99 return flag & FLAG_FORWARD_PROGRESS; 100 101 return flag & FLAG_DATA_ACKED; 102 } 103 104 SEC("struct_ops") 105 void BPF_PROG(bpf_cubic_init, struct sock *sk) 106 { 107 cubictcp_init(sk); 108 } 109 110 SEC("struct_ops") 111 void BPF_PROG(bpf_cubic_cwnd_event, struct sock *sk, enum tcp_ca_event event) 112 { 113 cubictcp_cwnd_event(sk, event); 114 } 115 116 SEC("struct_ops") 117 void BPF_PROG(bpf_cubic_cong_control, struct sock *sk, __u32 ack, int flag, 118 const struct rate_sample *rs) 119 { 120 struct tcp_sock *tp = tcp_sk(sk); 121 122 if (((1<<TCP_CA_CWR) | (1<<TCP_CA_Recovery)) & 123 (1 << inet_csk(sk)->icsk_ca_state)) { 124 /* Reduce cwnd if state mandates */ 125 tcp_cwnd_reduction(sk, rs->acked_sacked, rs->losses, flag); 126 127 if (!before(tp->snd_una, tp->high_seq)) { 128 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */ 129 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH && 130 inet_csk(sk)->icsk_ca_state == TCP_CA_CWR) { 131 tp->snd_cwnd = tp->snd_ssthresh; 132 tp->snd_cwnd_stamp = tcp_jiffies32; 133 } 134 } 135 } else if (tcp_may_raise_cwnd(sk, flag)) { 136 /* Advance cwnd if state allows */ 137 cubictcp_cong_avoid(sk, ack, rs->acked_sacked); 138 tp->snd_cwnd_stamp = tcp_jiffies32; 139 } 140 141 tcp_update_pacing_rate(sk); 142 } 143 144 SEC("struct_ops") 145 __u32 BPF_PROG(bpf_cubic_recalc_ssthresh, struct sock *sk) 146 { 147 return cubictcp_recalc_ssthresh(sk); 148 } 149 150 SEC("struct_ops") 151 void BPF_PROG(bpf_cubic_state, struct sock *sk, __u8 new_state) 152 { 153 cubictcp_state(sk, new_state); 154 } 155 156 SEC("struct_ops") 157 void BPF_PROG(bpf_cubic_acked, struct sock *sk, const struct ack_sample *sample) 158 { 159 cubictcp_acked(sk, sample); 160 } 161 162 SEC("struct_ops") 163 __u32 BPF_PROG(bpf_cubic_undo_cwnd, struct sock *sk) 164 { 165 return tcp_reno_undo_cwnd(sk); 166 } 167 168 SEC(".struct_ops") 169 struct tcp_congestion_ops cc_cubic = { 170 .init = (void *)bpf_cubic_init, 171 .ssthresh = (void *)bpf_cubic_recalc_ssthresh, 172 .cong_control = (void *)bpf_cubic_cong_control, 173 .set_state = (void *)bpf_cubic_state, 174 .undo_cwnd = (void *)bpf_cubic_undo_cwnd, 175 .cwnd_event = (void *)bpf_cubic_cwnd_event, 176 .pkts_acked = (void *)bpf_cubic_acked, 177 .name = "bpf_cc_cubic", 178 }; 179 180 char _license[] SEC("license") = "GPL"; 181