[PATCH 1/2] [PATCH net-next,1/2] Add new args for cong_control in tcp_congestion_ops

[PATCH 1/2] [PATCH net-next,1/2] Add new args for cong_control in tcp_congestion_ops

[Miao Xu]

This patch adds two new arguments for cong_control of struct
tcp_congestion_ops:
 - ack
 - flag
These two arguments are inherited from the caller tcp_cong_control in
tcp_intput.c. One use case of them is to update cwnd and pacing rate
inside cong_control based on the info they provide. For example, the
flag can be used to decide if it is the right time to raise or reduce a
sender's cwnd.

Another change in this patch is to allow the write of tp->snd_cwnd_stamp
for a bpf tcp ca program. An use case of writing this field is to keep
track of the time whenever tp->snd_cwnd is raised or reduced inside the
cong_control callback.

Signed-off-by: Miao Xu <miaxu@meta.com>
---
 include/net/tcp.h     | 2 +-
 net/ipv4/bpf_tcp_ca.c | 6 +++++-
 net/ipv4/tcp_bbr.c    | 2 +-
 net/ipv4/tcp_input.c  | 2 +-
 4 files changed, 8 insertions(+), 4 deletions(-)

diff --git a/include/net/tcp.h b/include/net/tcp.h
index b935e1ae4caf..b37b8219060a 100644
--- a/include/net/tcp.h
+++ b/include/net/tcp.h
@@ -1167,7 +1167,7 @@ struct tcp_congestion_ops {
 	/* call when packets are delivered to update cwnd and pacing rate,
 	 * after all the ca_state processing. (optional)
 	 */
-	void (*cong_control)(struct sock *sk, const struct rate_sample *rs);
+	void (*cong_control)(struct sock *sk, u32 ack, int flag, const struct rate_sample *rs);
 
 
 	/* new value of cwnd after loss (required) */
diff --git a/net/ipv4/bpf_tcp_ca.c b/net/ipv4/bpf_tcp_ca.c
index 7f518ea5f4ac..18227757ec0c 100644
--- a/net/ipv4/bpf_tcp_ca.c
+++ b/net/ipv4/bpf_tcp_ca.c
@@ -107,6 +107,9 @@ static int bpf_tcp_ca_btf_struct_access(struct bpf_verifier_log *log,
 	case offsetof(struct tcp_sock, snd_cwnd_cnt):
 		end = offsetofend(struct tcp_sock, snd_cwnd_cnt);
 		break;
+	case offsetof(struct tcp_sock, snd_cwnd_stamp):
+		end = offsetofend(struct tcp_sock, snd_cwnd_stamp);
+		break;
 	case offsetof(struct tcp_sock, snd_ssthresh):
 		end = offsetofend(struct tcp_sock, snd_ssthresh);
 		break;
@@ -307,7 +310,8 @@ static u32 bpf_tcp_ca_min_tso_segs(struct sock *sk)
 	return 0;
 }
 
-static void bpf_tcp_ca_cong_control(struct sock *sk, const struct rate_sample *rs)
+static void bpf_tcp_ca_cong_control(struct sock *sk, u32 ack, int flag,
+				    const struct rate_sample *rs)
 {
 }
 
diff --git a/net/ipv4/tcp_bbr.c b/net/ipv4/tcp_bbr.c
index 05dc2d05bc7c..c13d263dae06 100644
--- a/net/ipv4/tcp_bbr.c
+++ b/net/ipv4/tcp_bbr.c
@@ -1024,7 +1024,7 @@ static void bbr_update_model(struct sock *sk, const struct rate_sample *rs)
 	bbr_update_gains(sk);
 }
 
-__bpf_kfunc static void bbr_main(struct sock *sk, const struct rate_sample *rs)
+__bpf_kfunc static void bbr_main(struct sock *sk, u32 ack, int flag, const struct rate_sample *rs)
 {
 	struct bbr *bbr = inet_csk_ca(sk);
 	u32 bw;
diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index 384fa5e2f065..661dca9e3895 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -3541,7 +3541,7 @@ static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
 	const struct inet_connection_sock *icsk = inet_csk(sk);
 
 	if (icsk->icsk_ca_ops->cong_control) {
-		icsk->icsk_ca_ops->cong_control(sk, rs);
+		icsk->icsk_ca_ops->cong_control(sk, ack, flag, rs);
 		return;
 	}
 
-- 
2.43.0

[PATCH 2/2] [PATCH net-next,2/2] Add test for the use of new args in cong_control

[Miao Xu]

This patch adds a selftest to show the usage of the new arguments in
cong_control. For simplicity's sake, the testing example reuses cubic's
kernel functions.

Signed-off-by: Miao Xu <miaxu@meta.com>
---
 tools/testing/selftests/bpf/bpf_tcp_helpers.h |  23 +++
 .../bpf/progs/bpf_cubic_cong_control.c        | 176 ++++++++++++++++++
 2 files changed, 199 insertions(+)
 create mode 100644 tools/testing/selftests/bpf/progs/bpf_cubic_cong_control.c

diff --git a/tools/testing/selftests/bpf/bpf_tcp_helpers.h b/tools/testing/selftests/bpf/bpf_tcp_helpers.h
index 82a7c9de95f9..3115bc80280e 100644
--- a/tools/testing/selftests/bpf/bpf_tcp_helpers.h
+++ b/tools/testing/selftests/bpf/bpf_tcp_helpers.h
@@ -21,6 +21,15 @@ BPF_PROG(name, args)
 #endif
 
 #define tcp_jiffies32 ((__u32)bpf_jiffies64())
+#define TCP_INFINITE_SSTHRESH 0x7fffffff
+
+#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data.		*/
+#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN.		*/
+#define FLAG_DATA_SACKED 0x20 /* New SACK.				*/
+#define FLAG_SND_UNA_ADVANCED \
+	0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
+#define FLAG_ACKED (FLAG_DATA_ACKED | FLAG_SYN_ACKED)
+#define FLAG_FORWARD_PROGRESS (FLAG_ACKED | FLAG_DATA_SACKED)
 
 struct sock_common {
 	unsigned char	skc_state;
@@ -37,6 +46,7 @@ struct sock {
 	struct sock_common	__sk_common;
 #define sk_state		__sk_common.skc_state
 	unsigned long		sk_pacing_rate;
+	unsigned long		sk_max_pacing_rate;
 	__u32			sk_pacing_status; /* see enum sk_pacing */
 } __attribute__((preserve_access_index));
 
@@ -86,6 +96,19 @@ struct tcp_sock {
 	__u32	prior_cwnd;
 	__u64	tcp_mstamp;	/* most recent packet received/sent */
 	bool	is_mptcp;
+	__u32	snd_cwnd_stamp;
+	__u32	mss_cache;	/* Cached effective mss, not including SACKS */
+	__u32	high_seq;	/* snd_nxt at onset of congestion	*/
+	__u32	packets_out;	/* Packets which are "in flight"	*/
+	__u32	srtt_us;	/* smoothed round trip time << 3 in usecs */
+	__u32	retrans_out;	/* Retransmitted packets out */
+	__u32	lost_out;	/* Lost packets */
+	__u32	sacked_out;	/* SACK'd packets */
+	__u32	prr_delivered;	/* Number of newly delivered packets to
+				 * receiver in Recovery.
+				 */
+	__u32	prr_out;	/* Total number of pkts sent during Recovery. */
+	__u32	reordering;	/* Packet reordering metric. */
 } __attribute__((preserve_access_index));
 
 static __always_inline struct inet_connection_sock *inet_csk(const struct sock *sk)
diff --git a/tools/testing/selftests/bpf/progs/bpf_cubic_cong_control.c b/tools/testing/selftests/bpf/progs/bpf_cubic_cong_control.c
new file mode 100644
index 000000000000..698964df1f33
--- /dev/null
+++ b/tools/testing/selftests/bpf/progs/bpf_cubic_cong_control.c
@@ -0,0 +1,176 @@
+// SPDX-License-Identifier: GPL-2.0-only
+
+/* WARNING: This implementation is NOT the same as the tcp_cubic.c.
+ * The purpose is mainly to show use cases of the new arguments in
+ * cong_control.
+ */
+
+#include <linux/bpf.h>
+#include <linux/stddef.h>
+#include <linux/tcp.h>
+#include "bpf_tcp_helpers.h"
+
+char _license[] SEC("license") = "GPL";
+
+extern void cubictcp_init(struct sock *sk) __ksym;
+extern void cubictcp_cwnd_event(struct sock *sk, enum tcp_ca_event event)
+	__ksym;
+	extern __u32 cubictcp_recalc_ssthresh(struct sock *sk) __ksym;
+	extern void cubictcp_state(struct sock *sk, __u8 new_state) __ksym;
+	extern __u32 tcp_reno_undo_cwnd(struct sock *sk) __ksym;
+extern void cubictcp_acked(struct sock *sk, const struct ack_sample *sample)
+	__ksym;
+	extern void cubictcp_cong_avoid(struct sock *sk, __u32 ack, __u32 acked) __ksym;
+
+
+void BPF_STRUCT_OPS(bpf_cubic_init, struct sock *sk)
+{
+	cubictcp_init(sk);
+}
+
+void BPF_STRUCT_OPS(bpf_cubic_cwnd_event, struct sock *sk, enum tcp_ca_event event)
+{
+	cubictcp_cwnd_event(sk, event);
+}
+
+#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);
+}
+
+__u32 BPF_STRUCT_OPS(bpf_cubic_recalc_ssthresh, struct sock *sk)
+{
+	return cubictcp_recalc_ssthresh(sk);
+}
+
+void BPF_STRUCT_OPS(bpf_cubic_state, struct sock *sk, __u8 new_state)
+{
+	cubictcp_state(sk, new_state);
+}
+
+void BPF_STRUCT_OPS(bpf_cubic_acked, struct sock *sk,
+		const struct ack_sample *sample)
+{
+	cubictcp_acked(sk, sample);
+}
+
+__u32 BPF_STRUCT_OPS(bpf_cubic_undo_cwnd, struct sock *sk)
+{
+	return tcp_reno_undo_cwnd(sk);
+}
+
+
+SEC(".struct_ops")
+struct tcp_congestion_ops cubic = {
+	.init		= (void *)bpf_cubic_init,
+	.ssthresh	= (void *)bpf_cubic_recalc_ssthresh,
+	.cong_control	= (void *)bpf_cubic_cong_control,
+	.set_state	= (void *)bpf_cubic_state,
+	.undo_cwnd	= (void *)bpf_cubic_undo_cwnd,
+	.cwnd_event	= (void *)bpf_cubic_cwnd_event,
+	.pkts_acked     = (void *)bpf_cubic_acked,
+	.name		= "bpf_cubic",
+};
-- 
2.43.0

Re: [PATCH 2/2] [PATCH net-next,2/2] Add test for the use of new args in cong_control

[Martin KaFai Lau]

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?

Re: [PATCH 1/2] [PATCH net-next,1/2] Add new args for cong_control in tcp_congestion_ops

[Eric Dumazet]

On Wed, Apr 24, 2024 at 10:38 PM Miao Xu <miaxu@meta.com> wrote:
>
> This patch adds two new arguments for cong_control of struct
> tcp_congestion_ops:
>  - ack
>  - flag
> These two arguments are inherited from the caller tcp_cong_control in
> tcp_intput.c. One use case of them is to update cwnd and pacing rate
> inside cong_control based on the info they provide. For example, the
> flag can be used to decide if it is the right time to raise or reduce a
> sender's cwnd.
>
> Another change in this patch is to allow the write of tp->snd_cwnd_stamp
> for a bpf tcp ca program. An use case of writing this field is to keep
> track of the time whenever tp->snd_cwnd is raised or reduced inside the
> cong_control callback.
>
> Signed-off-by: Miao Xu <miaxu@meta.com>

Reviewed-by: Eric Dumazet <edumazet@google.com>