Merge branch 'bpf-qdisc'

Amery Hung says: ==================== bpf qdisc Hi all, This patchset aims to support implementing qdisc using bpf struct_ops. This version takes a step back and only implements the minimum support for bpf qdisc. 1) support of adding skb to bpf_list and bpf_rbtree directly and 2) classful qdisc are deferred to future patchsets. In addition, we only allow attaching bpf qdisc to root or mq for now. This is to prevent accidentally breaking exisiting classful qdiscs that rely on data in a child qdisc. This limit may be lifted in the future after careful inspection. * Overview * This series supports implementing qdisc using bpf struct_ops. bpf qdisc aims to be a flexible and easy-to-use infrastructure that allows users to quickly experiment with different scheduling algorithms/policies. It only requires users to implement core qdisc logic using bpf and implements the mundane part for them. In addition, the ability to easily communicate between qdisc and other components will also bring new opportunities for new applications and optimizations. * Performance of bpf qdisc * This patchset includes two qdisc examples, bpf_fifo and bpf_fq, for __testing__ purposes. For performance test, we compare selftests and their kernel counterparts to give you a sense of the performance of qdisc implemented in bpf. The implementation of bpf_fq is fairly complex and slightly different from fq so later we only compare the two fifo qdiscs. bpf_fq implements a scheduling algorithm similar to fq before commit 29f834aa326e ("net_sched: sch_fq: add 3 bands and WRR scheduling") was introduced. bpf_fifo uses a single bpf_list as a queue instead of three queues for different priorities in pfifo_fast. The time complexity of fifo however should be similar since the queue selection time is negligible. Test setup: client -> qdisc -------------> server ~~~~~~~~~~~~~~~ ~~~~~~ nested VM1 @ DC1 VM2 @ DC2 Throghput: iperf3 -t 600, 5 times Qdisc Average (GBits/sec) ---------- ------------------- pfifo_fast 12.52 ± 0.26 bpf_fifo 11.72 ± 0.32 fq 10.24 ± 0.13 bpf_fq 11.92 ± 0.64 Latency: sockperf pp --tcp -t 600, 5 times Qdisc Average (usec) ---------- -------------- pfifo_fast 244.58 ± 7.93 bpf_fifo 244.92 ± 15.22 fq 234.30 ± 19.25 bpf_fq 221.34 ± 10.76 Looking at the two fifo qdiscs, the 6.4% drop in throughput in the bpf implementatioin is consistent with previous observation (v8 throughput test on a loopback device). This should be able to be mitigated by supporting adding skb to bpf_list or bpf_rbtree directly in the future. * Clean up skb in bpf qdisc during reset * The current implementation relies on bpf qdisc implementors to correctly release skbs in queues (bpf graphs or maps) in .reset, which might not be a safe thing to do. The solution as Martin has suggested would be supporting private data in struct_ops. This can also help simplifying implementation of qdisc that works with mq. For examples, qdiscs in the selftest mostly use global data. Therefore, even if user add multiple qdisc instances under mq, they would still share the same queue. ==================== Link: https://patch.msgid.link/20250409214606.2000194-1-ameryhung@gmail.com Signed-off-by: Martin KaFai Lau <martin.lau@kernel.org>
author: Martin KaFai Lau <martin.lau@kernel.org> 2025-04-17 10:50:56 -0700
committer: Martin KaFai Lau <martin.lau@kernel.org> 2025-04-17 10:54:41 -0700
commit: fd23ce3eb4a1005bd109977856d12ec0fde7ef75 (patch)
tree: eed77cf13f3a6402654d6d66300f4198f48bb046 /tools
parent: ab734b43c81f8696bf89ff11b6c29a3dd884030a (diff)
parent: 2b7b5b7f100e82ca314e76214626c82b608a1d8d (diff)
download: ath-fd23ce3eb4a1005bd109977856d12ec0fde7ef75.tar.gz
7 files changed, 1101 insertions, 4 deletions
diff --git a/tools/lib/bpf/libbpf.h b/tools/lib/bpf/libbpf.h
index e0605403f9773..fdcee6a71e0fa 100644
--- a/tools/lib/bpf/libbpf.h
+++ b/tools/lib/bpf/libbpf.h
@@ -1283,6 +1283,7 @@ enum bpf_tc_attach_point {
 	BPF_TC_INGRESS = 1 << 0,
 	BPF_TC_EGRESS  = 1 << 1,
 	BPF_TC_CUSTOM  = 1 << 2,
+	BPF_TC_QDISC   = 1 << 3,
 };
 
 #define BPF_TC_PARENT(a, b) 	\
@@ -1297,9 +1298,11 @@ struct bpf_tc_hook {
 	int ifindex;
 	enum bpf_tc_attach_point attach_point;
 	__u32 parent;
+	__u32 handle;
+	const char *qdisc;
 	size_t :0;
 };
-#define bpf_tc_hook__last_field parent
+#define bpf_tc_hook__last_field qdisc
 
 struct bpf_tc_opts {
 	size_t sz;
diff --git a/tools/lib/bpf/netlink.c b/tools/lib/bpf/netlink.c
index 68a2def171751..c997e69d507fe 100644
--- a/tools/lib/bpf/netlink.c
+++ b/tools/lib/bpf/netlink.c
@@ -529,9 +529,9 @@ int bpf_xdp_query_id(int ifindex, int flags, __u32 *prog_id)
 }
 
 
-typedef int (*qdisc_config_t)(struct libbpf_nla_req *req);
+typedef int (*qdisc_config_t)(struct libbpf_nla_req *req, const struct bpf_tc_hook *hook);
 
-static int clsact_config(struct libbpf_nla_req *req)
+static int clsact_config(struct libbpf_nla_req *req, const struct bpf_tc_hook *hook)
 {
 	req->tc.tcm_parent = TC_H_CLSACT;
 	req->tc.tcm_handle = TC_H_MAKE(TC_H_CLSACT, 0);
@@ -539,6 +539,16 @@ static int clsact_config(struct libbpf_nla_req *req)
 	return nlattr_add(req, TCA_KIND, "clsact", sizeof("clsact"));
 }
 
+static int qdisc_config(struct libbpf_nla_req *req, const struct bpf_tc_hook *hook)
+{
+	const char *qdisc = OPTS_GET(hook, qdisc, NULL);
+
+	req->tc.tcm_parent = OPTS_GET(hook, parent, TC_H_ROOT);
+	req->tc.tcm_handle = OPTS_GET(hook, handle, 0);
+
+	return nlattr_add(req, TCA_KIND, qdisc, strlen(qdisc) + 1);
+}
+
 static int attach_point_to_config(struct bpf_tc_hook *hook,
 				  qdisc_config_t *config)
 {
@@ -552,6 +562,9 @@ static int attach_point_to_config(struct bpf_tc_hook *hook,
 		return 0;
 	case BPF_TC_CUSTOM:
 		return -EOPNOTSUPP;
+	case BPF_TC_QDISC:
+		*config = &qdisc_config;
+		return 0;
 	default:
 		return -EINVAL;
 	}
@@ -596,7 +609,7 @@ static int tc_qdisc_modify(struct bpf_tc_hook *hook, int cmd, int flags)
 	req.tc.tcm_family  = AF_UNSPEC;
 	req.tc.tcm_ifindex = OPTS_GET(hook, ifindex, 0);
 
-	ret = config(&req);
+	ret = config(&req, hook);
 	if (ret < 0)
 		return ret;
 
@@ -639,6 +652,7 @@ int bpf_tc_hook_destroy(struct bpf_tc_hook *hook)
 	case BPF_TC_INGRESS:
 	case BPF_TC_EGRESS:
 		return libbpf_err(__bpf_tc_detach(hook, NULL, true));
+	case BPF_TC_QDISC:
 	case BPF_TC_INGRESS | BPF_TC_EGRESS:
 		return libbpf_err(tc_qdisc_delete(hook));
 	case BPF_TC_CUSTOM:
diff --git a/tools/testing/selftests/bpf/config b/tools/testing/selftests/bpf/config
index c378d5d07e029..3201a962b3dcd 100644
--- a/tools/testing/selftests/bpf/config
+++ b/tools/testing/selftests/bpf/config
@@ -71,8 +71,10 @@ CONFIG_NET_IPGRE=y
 CONFIG_NET_IPGRE_DEMUX=y
 CONFIG_NET_IPIP=y
 CONFIG_NET_MPLS_GSO=y
+CONFIG_NET_SCH_BPF=y
 CONFIG_NET_SCH_FQ=y
 CONFIG_NET_SCH_INGRESS=y
+CONFIG_NET_SCH_HTB=y
 CONFIG_NET_SCHED=y
 CONFIG_NETDEVSIM=y
 CONFIG_NETFILTER=y
diff --git a/tools/testing/selftests/bpf/prog_tests/bpf_qdisc.c b/tools/testing/selftests/bpf/prog_tests/bpf_qdisc.c
new file mode 100644
index 0000000000000..c9a54177c84ed
--- /dev/null
+++ b/tools/testing/selftests/bpf/prog_tests/bpf_qdisc.c
@@ -0,0 +1,180 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#include <linux/pkt_sched.h>
+#include <linux/rtnetlink.h>
+#include <test_progs.h>
+
+#include "network_helpers.h"
+#include "bpf_qdisc_fifo.skel.h"
+#include "bpf_qdisc_fq.skel.h"
+
+#define LO_IFINDEX 1
+
+static const unsigned int total_bytes = 10 * 1024 * 1024;
+
+static void do_test(char *qdisc)
+{
+	DECLARE_LIBBPF_OPTS(bpf_tc_hook, hook, .ifindex = LO_IFINDEX,
+			    .attach_point = BPF_TC_QDISC,
+			    .parent = TC_H_ROOT,
+			    .handle = 0x8000000,
+			    .qdisc = qdisc);
+	int srv_fd = -1, cli_fd = -1;
+	int err;
+
+	err = bpf_tc_hook_create(&hook);
+	if (!ASSERT_OK(err, "attach qdisc"))
+		return;
+
+	srv_fd = start_server(AF_INET6, SOCK_STREAM, NULL, 0, 0);
+	if (!ASSERT_OK_FD(srv_fd, "start server"))
+		goto done;
+
+	cli_fd = connect_to_fd(srv_fd, 0);
+	if (!ASSERT_OK_FD(cli_fd, "connect to client"))
+		goto done;
+
+	err = send_recv_data(srv_fd, cli_fd, total_bytes);
+	ASSERT_OK(err, "send_recv_data");
+
+done:
+	if (srv_fd != -1)
+		close(srv_fd);
+	if (cli_fd != -1)
+		close(cli_fd);
+
+	bpf_tc_hook_destroy(&hook);
+}
+
+static void test_fifo(void)
+{
+	struct bpf_qdisc_fifo *fifo_skel;
+	struct bpf_link *link;
+
+	fifo_skel = bpf_qdisc_fifo__open_and_load();
+	if (!ASSERT_OK_PTR(fifo_skel, "bpf_qdisc_fifo__open_and_load"))
+		return;
+
+	link = bpf_map__attach_struct_ops(fifo_skel->maps.fifo);
+	if (!ASSERT_OK_PTR(link, "bpf_map__attach_struct_ops")) {
+		bpf_qdisc_fifo__destroy(fifo_skel);
+		return;
+	}
+
+	do_test("bpf_fifo");
+
+	bpf_link__destroy(link);
+	bpf_qdisc_fifo__destroy(fifo_skel);
+}
+
+static void test_fq(void)
+{
+	struct bpf_qdisc_fq *fq_skel;
+	struct bpf_link *link;
+
+	fq_skel = bpf_qdisc_fq__open_and_load();
+	if (!ASSERT_OK_PTR(fq_skel, "bpf_qdisc_fq__open_and_load"))
+		return;
+
+	link = bpf_map__attach_struct_ops(fq_skel->maps.fq);
+	if (!ASSERT_OK_PTR(link, "bpf_map__attach_struct_ops")) {
+		bpf_qdisc_fq__destroy(fq_skel);
+		return;
+	}
+
+	do_test("bpf_fq");
+
+	bpf_link__destroy(link);
+	bpf_qdisc_fq__destroy(fq_skel);
+}
+
+static void test_qdisc_attach_to_mq(void)
+{
+	DECLARE_LIBBPF_OPTS(bpf_tc_hook, hook,
+			    .attach_point = BPF_TC_QDISC,
+			    .parent = TC_H_MAKE(1 << 16, 1),
+			    .handle = 0x11 << 16,
+			    .qdisc = "bpf_fifo");
+	struct bpf_qdisc_fifo *fifo_skel;
+	struct bpf_link *link;
+	int err;
+
+	fifo_skel = bpf_qdisc_fifo__open_and_load();
+	if (!ASSERT_OK_PTR(fifo_skel, "bpf_qdisc_fifo__open_and_load"))
+		return;
+
+	link = bpf_map__attach_struct_ops(fifo_skel->maps.fifo);
+	if (!ASSERT_OK_PTR(link, "bpf_map__attach_struct_ops")) {
+		bpf_qdisc_fifo__destroy(fifo_skel);
+		return;
+	}
+
+	SYS(out, "ip link add veth0 type veth peer veth1");
+	hook.ifindex = if_nametoindex("veth0");
+	SYS(out, "tc qdisc add dev veth0 root handle 1: mq");
+
+	err = bpf_tc_hook_create(&hook);
+	ASSERT_OK(err, "attach qdisc");
+
+	bpf_tc_hook_destroy(&hook);
+
+	SYS(out, "tc qdisc delete dev veth0 root mq");
+out:
+	bpf_link__destroy(link);
+	bpf_qdisc_fifo__destroy(fifo_skel);
+}
+
+static void test_qdisc_attach_to_non_root(void)
+{
+	DECLARE_LIBBPF_OPTS(bpf_tc_hook, hook, .ifindex = LO_IFINDEX,
+			    .attach_point = BPF_TC_QDISC,
+			    .parent = TC_H_MAKE(1 << 16, 1),
+			    .handle = 0x11 << 16,
+			    .qdisc = "bpf_fifo");
+	struct bpf_qdisc_fifo *fifo_skel;
+	struct bpf_link *link;
+	int err;
+
+	fifo_skel = bpf_qdisc_fifo__open_and_load();
+	if (!ASSERT_OK_PTR(fifo_skel, "bpf_qdisc_fifo__open_and_load"))
+		return;
+
+	link = bpf_map__attach_struct_ops(fifo_skel->maps.fifo);
+	if (!ASSERT_OK_PTR(link, "bpf_map__attach_struct_ops")) {
+		bpf_qdisc_fifo__destroy(fifo_skel);
+		return;
+	}
+
+	SYS(out, "tc qdisc add dev lo root handle 1: htb");
+	SYS(out_del_htb, "tc class add dev lo parent 1: classid 1:1 htb rate 75Kbit");
+
+	err = bpf_tc_hook_create(&hook);
+	if (!ASSERT_ERR(err, "attach qdisc"))
+		bpf_tc_hook_destroy(&hook);
+
+out_del_htb:
+	SYS(out, "tc qdisc delete dev lo root htb");
+out:
+	bpf_link__destroy(link);
+	bpf_qdisc_fifo__destroy(fifo_skel);
+}
+
+void test_bpf_qdisc(void)
+{
+	struct netns_obj *netns;
+
+	netns = netns_new("bpf_qdisc_ns", true);
+	if (!ASSERT_OK_PTR(netns, "netns_new"))
+		return;
+
+	if (test__start_subtest("fifo"))
+		test_fifo();
+	if (test__start_subtest("fq"))
+		test_fq();
+	if (test__start_subtest("attach to mq"))
+		test_qdisc_attach_to_mq();
+	if (test__start_subtest("attach to non root"))
+		test_qdisc_attach_to_non_root();
+
+	netns_free(netns);
+}
diff --git a/tools/testing/selftests/bpf/progs/bpf_qdisc_common.h b/tools/testing/selftests/bpf/progs/bpf_qdisc_common.h
new file mode 100644
index 0000000000000..65a2c561c0bb5
--- /dev/null
+++ b/tools/testing/selftests/bpf/progs/bpf_qdisc_common.h
@@ -0,0 +1,31 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+
+#ifndef _BPF_QDISC_COMMON_H
+#define _BPF_QDISC_COMMON_H
+
+#define NET_XMIT_SUCCESS        0x00
+#define NET_XMIT_DROP           0x01    /* skb dropped                  */
+#define NET_XMIT_CN             0x02    /* congestion notification      */
+
+#define TC_PRIO_CONTROL  7
+#define TC_PRIO_MAX      15
+
+#define private(name) SEC(".data." #name) __hidden __attribute__((aligned(8)))
+
+u32 bpf_skb_get_hash(struct sk_buff *p) __ksym;
+void bpf_kfree_skb(struct sk_buff *p) __ksym;
+void bpf_qdisc_skb_drop(struct sk_buff *p, struct bpf_sk_buff_ptr *to_free) __ksym;
+void bpf_qdisc_watchdog_schedule(struct Qdisc *sch, u64 expire, u64 delta_ns) __ksym;
+void bpf_qdisc_bstats_update(struct Qdisc *sch, const struct sk_buff *skb) __ksym;
+
+static struct qdisc_skb_cb *qdisc_skb_cb(const struct sk_buff *skb)
+{
+	return (struct qdisc_skb_cb *)skb->cb;
+}
+
+static inline unsigned int qdisc_pkt_len(const struct sk_buff *skb)
+{
+	return qdisc_skb_cb(skb)->pkt_len;
+}
+
+#endif
diff --git a/tools/testing/selftests/bpf/progs/bpf_qdisc_fifo.c b/tools/testing/selftests/bpf/progs/bpf_qdisc_fifo.c
new file mode 100644
index 0000000000000..0c7cfb82dae19
--- /dev/null
+++ b/tools/testing/selftests/bpf/progs/bpf_qdisc_fifo.c
@@ -0,0 +1,117 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#include <vmlinux.h>
+#include "bpf_experimental.h"
+#include "bpf_qdisc_common.h"
+
+char _license[] SEC("license") = "GPL";
+
+struct skb_node {
+	struct sk_buff __kptr * skb;
+	struct bpf_list_node node;
+};
+
+private(A) struct bpf_spin_lock q_fifo_lock;
+private(A) struct bpf_list_head q_fifo __contains(skb_node, node);
+
+SEC("struct_ops/bpf_fifo_enqueue")
+int BPF_PROG(bpf_fifo_enqueue, struct sk_buff *skb, struct Qdisc *sch,
+	     struct bpf_sk_buff_ptr *to_free)
+{
+	struct skb_node *skbn;
+	u32 pkt_len;
+
+	if (sch->q.qlen == sch->limit)
+		goto drop;
+
+	skbn = bpf_obj_new(typeof(*skbn));
+	if (!skbn)
+		goto drop;
+
+	pkt_len = qdisc_pkt_len(skb);
+
+	sch->q.qlen++;
+	skb = bpf_kptr_xchg(&skbn->skb, skb);
+	if (skb)
+		bpf_qdisc_skb_drop(skb, to_free);
+
+	bpf_spin_lock(&q_fifo_lock);
+	bpf_list_push_back(&q_fifo, &skbn->node);
+	bpf_spin_unlock(&q_fifo_lock);
+
+	sch->qstats.backlog += pkt_len;
+	return NET_XMIT_SUCCESS;
+drop:
+	bpf_qdisc_skb_drop(skb, to_free);
+	return NET_XMIT_DROP;
+}
+
+SEC("struct_ops/bpf_fifo_dequeue")
+struct sk_buff *BPF_PROG(bpf_fifo_dequeue, struct Qdisc *sch)
+{
+	struct bpf_list_node *node;
+	struct sk_buff *skb = NULL;
+	struct skb_node *skbn;
+
+	bpf_spin_lock(&q_fifo_lock);
+	node = bpf_list_pop_front(&q_fifo);
+	bpf_spin_unlock(&q_fifo_lock);
+	if (!node)
+		return NULL;
+
+	skbn = container_of(node, struct skb_node, node);
+	skb = bpf_kptr_xchg(&skbn->skb, skb);
+	bpf_obj_drop(skbn);
+	if (!skb)
+		return NULL;
+
+	sch->qstats.backlog -= qdisc_pkt_len(skb);
+	bpf_qdisc_bstats_update(sch, skb);
+	sch->q.qlen--;
+
+	return skb;
+}
+
+SEC("struct_ops/bpf_fifo_init")
+int BPF_PROG(bpf_fifo_init, struct Qdisc *sch, struct nlattr *opt,
+	     struct netlink_ext_ack *extack)
+{
+	sch->limit = 1000;
+	return 0;
+}
+
+SEC("struct_ops/bpf_fifo_reset")
+void BPF_PROG(bpf_fifo_reset, struct Qdisc *sch)
+{
+	struct bpf_list_node *node;
+	struct skb_node *skbn;
+	int i;
+
+	bpf_for(i, 0, sch->q.qlen) {
+		struct sk_buff *skb = NULL;
+
+		bpf_spin_lock(&q_fifo_lock);
+		node = bpf_list_pop_front(&q_fifo);
+		bpf_spin_unlock(&q_fifo_lock);
+
+		if (!node)
+			break;
+
+		skbn = container_of(node, struct skb_node, node);
+		skb = bpf_kptr_xchg(&skbn->skb, skb);
+		if (skb)
+			bpf_kfree_skb(skb);
+		bpf_obj_drop(skbn);
+	}
+	sch->q.qlen = 0;
+}
+
+SEC(".struct_ops")
+struct Qdisc_ops fifo = {
+	.enqueue   = (void *)bpf_fifo_enqueue,
+	.dequeue   = (void *)bpf_fifo_dequeue,
+	.init      = (void *)bpf_fifo_init,
+	.reset     = (void *)bpf_fifo_reset,
+	.id        = "bpf_fifo",
+};
+
diff --git a/tools/testing/selftests/bpf/progs/bpf_qdisc_fq.c b/tools/testing/selftests/bpf/progs/bpf_qdisc_fq.c
new file mode 100644
index 0000000000000..7c110a1562247
--- /dev/null
+++ b/tools/testing/selftests/bpf/progs/bpf_qdisc_fq.c
@@ -0,0 +1,750 @@
+// SPDX-License-Identifier: GPL-2.0
+
+/* bpf_fq is intended for testing the bpf qdisc infrastructure and not a direct
+ * copy of sch_fq. bpf_fq implements the scheduling algorithm of sch_fq before
+ * 29f834aa326e ("net_sched: sch_fq: add 3 bands and WRR scheduling") was
+ * introduced. It gives each flow a fair chance to transmit packets in a
+ * round-robin fashion. Note that for flow pacing, bpf_fq currently only
+ * respects skb->tstamp but not skb->sk->sk_pacing_rate. In addition, if there
+ * are multiple bpf_fq instances, they will have a shared view of flows and
+ * configuration since some key data structure such as fq_prio_flows,
+ * fq_nonprio_flows, and fq_bpf_data are global.
+ *
+ * To use bpf_fq alone without running selftests, use the following commands.
+ *
+ * 1. Register bpf_fq to the kernel
+ *     bpftool struct_ops register bpf_qdisc_fq.bpf.o /sys/fs/bpf
+ * 2. Add bpf_fq to an interface
+ *     tc qdisc add dev <interface name> root handle <handle> bpf_fq
+ * 3. Delete bpf_fq attached to the interface
+ *     tc qdisc delete dev <interface name> root
+ * 4. Unregister bpf_fq
+ *     bpftool struct_ops unregister name fq
+ *
+ * The qdisc name, bpf_fq, used in tc commands is defined by Qdisc_ops.id.
+ * The struct_ops_map_name, fq, used in the bpftool command is the name of the
+ * Qdisc_ops.
+ *
+ * SEC(".struct_ops")
+ * struct Qdisc_ops fq = {
+ *         ...
+ *         .id        = "bpf_fq",
+ * };
+ */
+
+#include <vmlinux.h>
+#include <errno.h>
+#include <bpf/bpf_helpers.h>
+#include "bpf_experimental.h"
+#include "bpf_qdisc_common.h"
+
+char _license[] SEC("license") = "GPL";
+
+#define NSEC_PER_USEC 1000L
+#define NSEC_PER_SEC 1000000000L
+
+#define NUM_QUEUE (1 << 20)
+
+struct fq_bpf_data {
+	u32 quantum;
+	u32 initial_quantum;
+	u32 flow_refill_delay;
+	u32 flow_plimit;
+	u64 horizon;
+	u32 orphan_mask;
+	u32 timer_slack;
+	u64 time_next_delayed_flow;
+	u64 unthrottle_latency_ns;
+	u8 horizon_drop;
+	u32 new_flow_cnt;
+	u32 old_flow_cnt;
+	u64 ktime_cache;
+};
+
+enum {
+	CLS_RET_PRIO	= 0,
+	CLS_RET_NONPRIO = 1,
+	CLS_RET_ERR	= 2,
+};
+
+struct skb_node {
+	u64 tstamp;
+	struct sk_buff __kptr * skb;
+	struct bpf_rb_node node;
+};
+
+struct fq_flow_node {
+	int credit;
+	u32 qlen;
+	u64 age;
+	u64 time_next_packet;
+	struct bpf_list_node list_node;
+	struct bpf_rb_node rb_node;
+	struct bpf_rb_root queue __contains(skb_node, node);
+	struct bpf_spin_lock lock;
+	struct bpf_refcount refcount;
+};
+
+struct dequeue_nonprio_ctx {
+	bool stop_iter;
+	u64 expire;
+	u64 now;
+};
+
+struct remove_flows_ctx {
+	bool gc_only;
+	u32 reset_cnt;
+	u32 reset_max;
+};
+
+struct unset_throttled_flows_ctx {
+	bool unset_all;
+	u64 now;
+};
+
+struct fq_stashed_flow {
+	struct fq_flow_node __kptr * flow;
+};
+
+struct {
+	__uint(type, BPF_MAP_TYPE_HASH);
+	__type(key, __u64);
+	__type(value, struct fq_stashed_flow);
+	__uint(max_entries, NUM_QUEUE);
+} fq_nonprio_flows SEC(".maps");
+
+struct {
+	__uint(type, BPF_MAP_TYPE_HASH);
+	__type(key, __u64);
+	__type(value, struct fq_stashed_flow);
+	__uint(max_entries, 1);
+} fq_prio_flows SEC(".maps");
+
+private(A) struct bpf_spin_lock fq_delayed_lock;
+private(A) struct bpf_rb_root fq_delayed __contains(fq_flow_node, rb_node);
+
+private(B) struct bpf_spin_lock fq_new_flows_lock;
+private(B) struct bpf_list_head fq_new_flows __contains(fq_flow_node, list_node);
+
+private(C) struct bpf_spin_lock fq_old_flows_lock;
+private(C) struct bpf_list_head fq_old_flows __contains(fq_flow_node, list_node);
+
+private(D) struct fq_bpf_data q;
+
+/* Wrapper for bpf_kptr_xchg that expects NULL dst */
+static void bpf_kptr_xchg_back(void *map_val, void *ptr)
+{
+	void *ret;
+
+	ret = bpf_kptr_xchg(map_val, ptr);
+	if (ret)
+		bpf_obj_drop(ret);
+}
+
+static bool skbn_tstamp_less(struct bpf_rb_node *a, const struct bpf_rb_node *b)
+{
+	struct skb_node *skbn_a;
+	struct skb_node *skbn_b;
+
+	skbn_a = container_of(a, struct skb_node, node);
+	skbn_b = container_of(b, struct skb_node, node);
+
+	return skbn_a->tstamp < skbn_b->tstamp;
+}
+
+static bool fn_time_next_packet_less(struct bpf_rb_node *a, const struct bpf_rb_node *b)
+{
+	struct fq_flow_node *flow_a;
+	struct fq_flow_node *flow_b;
+
+	flow_a = container_of(a, struct fq_flow_node, rb_node);
+	flow_b = container_of(b, struct fq_flow_node, rb_node);
+
+	return flow_a->time_next_packet < flow_b->time_next_packet;
+}
+
+static void
+fq_flows_add_head(struct bpf_list_head *head, struct bpf_spin_lock *lock,
+		  struct fq_flow_node *flow, u32 *flow_cnt)
+{
+	bpf_spin_lock(lock);
+	bpf_list_push_front(head, &flow->list_node);
+	bpf_spin_unlock(lock);
+	*flow_cnt += 1;
+}
+
+static void
+fq_flows_add_tail(struct bpf_list_head *head, struct bpf_spin_lock *lock,
+		  struct fq_flow_node *flow, u32 *flow_cnt)
+{
+	bpf_spin_lock(lock);
+	bpf_list_push_back(head, &flow->list_node);
+	bpf_spin_unlock(lock);
+	*flow_cnt += 1;
+}
+
+static void
+fq_flows_remove_front(struct bpf_list_head *head, struct bpf_spin_lock *lock,
+		      struct bpf_list_node **node, u32 *flow_cnt)
+{
+	bpf_spin_lock(lock);
+	*node = bpf_list_pop_front(head);
+	bpf_spin_unlock(lock);
+	*flow_cnt -= 1;
+}
+
+static bool
+fq_flows_is_empty(struct bpf_list_head *head, struct bpf_spin_lock *lock)
+{
+	struct bpf_list_node *node;
+
+	bpf_spin_lock(lock);
+	node = bpf_list_pop_front(head);
+	if (node) {
+		bpf_list_push_front(head, node);
+		bpf_spin_unlock(lock);
+		return false;
+	}
+	bpf_spin_unlock(lock);
+
+	return true;
+}
+
+/* flow->age is used to denote the state of the flow (not-detached, detached, throttled)
+ * as well as the timestamp when the flow is detached.
+ *
+ * 0: not-detached
+ * 1 - (~0ULL-1): detached
+ * ~0ULL: throttled
+ */
+static void fq_flow_set_detached(struct fq_flow_node *flow)
+{
+	flow->age = bpf_jiffies64();
+}
+
+static bool fq_flow_is_detached(struct fq_flow_node *flow)
+{
+	return flow->age != 0 && flow->age != ~0ULL;
+}
+
+static bool sk_listener(struct sock *sk)
+{
+	return (1 << sk->__sk_common.skc_state) & (TCPF_LISTEN | TCPF_NEW_SYN_RECV);
+}
+
+static void fq_gc(void);
+
+static int fq_new_flow(void *flow_map, struct fq_stashed_flow **sflow, u64 hash)
+{
+	struct fq_stashed_flow tmp = {};
+	struct fq_flow_node *flow;
+	int ret;
+
+	flow = bpf_obj_new(typeof(*flow));
+	if (!flow)
+		return -ENOMEM;
+
+	flow->credit = q.initial_quantum,
+	flow->qlen = 0,
+	flow->age = 1,
+	flow->time_next_packet = 0,
+
+	ret = bpf_map_update_elem(flow_map, &hash, &tmp, 0);
+	if (ret == -ENOMEM || ret == -E2BIG) {
+		fq_gc();
+		bpf_map_update_elem(&fq_nonprio_flows, &hash, &tmp, 0);
+	}
+
+	*sflow = bpf_map_lookup_elem(flow_map, &hash);
+	if (!*sflow) {
+		bpf_obj_drop(flow);
+		return -ENOMEM;
+	}
+
+	bpf_kptr_xchg_back(&(*sflow)->flow, flow);
+	return 0;
+}
+
+static int
+fq_classify(struct sk_buff *skb, struct fq_stashed_flow **sflow)
+{
+	struct sock *sk = skb->sk;
+	int ret = CLS_RET_NONPRIO;
+	u64 hash = 0;
+
+	if ((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL) {
+		*sflow = bpf_map_lookup_elem(&fq_prio_flows, &hash);
+		ret = CLS_RET_PRIO;
+	} else {
+		if (!sk || sk_listener(sk)) {
+			hash = bpf_skb_get_hash(skb) & q.orphan_mask;
+			/* Avoid collision with an existing flow hash, which
+			 * only uses the lower 32 bits of hash, by setting the
+			 * upper half of hash to 1.
+			 */
+			hash |= (1ULL << 32);
+		} else if (sk->__sk_common.skc_state == TCP_CLOSE) {
+			hash = bpf_skb_get_hash(skb) & q.orphan_mask;
+			hash |= (1ULL << 32);
+		} else {
+			hash = sk->__sk_common.skc_hash;
+		}
+		*sflow = bpf_map_lookup_elem(&fq_nonprio_flows, &hash);
+	}
+
+	if (!*sflow)
+		ret = fq_new_flow(&fq_nonprio_flows, sflow, hash) < 0 ?
+		      CLS_RET_ERR : CLS_RET_NONPRIO;
+
+	return ret;
+}
+
+static bool fq_packet_beyond_horizon(struct sk_buff *skb)
+{
+	return (s64)skb->tstamp > (s64)(q.ktime_cache + q.horizon);
+}
+
+SEC("struct_ops/bpf_fq_enqueue")
+int BPF_PROG(bpf_fq_enqueue, struct sk_buff *skb, struct Qdisc *sch,
+	     struct bpf_sk_buff_ptr *to_free)
+{
+	struct fq_flow_node *flow = NULL, *flow_copy;
+	struct fq_stashed_flow *sflow;
+	u64 time_to_send, jiffies;
+	struct skb_node *skbn;
+	int ret;
+
+	if (sch->q.qlen >= sch->limit)
+		goto drop;
+
+	if (!skb->tstamp) {
+		time_to_send = q.ktime_cache = bpf_ktime_get_ns();
+	} else {
+		if (fq_packet_beyond_horizon(skb)) {
+			q.ktime_cache = bpf_ktime_get_ns();
+			if (fq_packet_beyond_horizon(skb)) {
+				if (q.horizon_drop)
+					goto drop;
+
+				skb->tstamp = q.ktime_cache + q.horizon;
+			}
+		}
+		time_to_send = skb->tstamp;
+	}
+
+	ret = fq_classify(skb, &sflow);
+	if (ret == CLS_RET_ERR)
+		goto drop;
+
+	flow = bpf_kptr_xchg(&sflow->flow, flow);
+	if (!flow)
+		goto drop;
+
+	if (ret == CLS_RET_NONPRIO) {
+		if (flow->qlen >= q.flow_plimit) {
+			bpf_kptr_xchg_back(&sflow->flow, flow);
+			goto drop;
+		}
+
+		if (fq_flow_is_detached(flow)) {
+			flow_copy = bpf_refcount_acquire(flow);
+
+			jiffies = bpf_jiffies64();
+			if ((s64)(jiffies - (flow_copy->age + q.flow_refill_delay)) > 0) {
+				if (flow_copy->credit < q.quantum)
+					flow_copy->credit = q.quantum;
+			}
+			flow_copy->age = 0;
+			fq_flows_add_tail(&fq_new_flows, &fq_new_flows_lock, flow_copy,
+					  &q.new_flow_cnt);
+		}
+	}
+
+	skbn = bpf_obj_new(typeof(*skbn));
+	if (!skbn) {
+		bpf_kptr_xchg_back(&sflow->flow, flow);
+		goto drop;
+	}
+
+	skbn->tstamp = skb->tstamp = time_to_send;
+
+	sch->qstats.backlog += qdisc_pkt_len(skb);
+
+	skb = bpf_kptr_xchg(&skbn->skb, skb);
+	if (skb)
+		bpf_qdisc_skb_drop(skb, to_free);
+
+	bpf_spin_lock(&flow->lock);
+	bpf_rbtree_add(&flow->queue, &skbn->node, skbn_tstamp_less);
+	bpf_spin_unlock(&flow->lock);
+
+	flow->qlen++;
+	bpf_kptr_xchg_back(&sflow->flow, flow);
+
+	sch->q.qlen++;
+	return NET_XMIT_SUCCESS;
+
+drop:
+	bpf_qdisc_skb_drop(skb, to_free);
+	sch->qstats.drops++;
+	return NET_XMIT_DROP;
+}
+
+static int fq_unset_throttled_flows(u32 index, struct unset_throttled_flows_ctx *ctx)
+{
+	struct bpf_rb_node *node = NULL;
+	struct fq_flow_node *flow;
+
+	bpf_spin_lock(&fq_delayed_lock);
+
+	node = bpf_rbtree_first(&fq_delayed);
+	if (!node) {
+		bpf_spin_unlock(&fq_delayed_lock);
+		return 1;
+	}
+
+	flow = container_of(node, struct fq_flow_node, rb_node);
+	if (!ctx->unset_all && flow->time_next_packet > ctx->now) {
+		q.time_next_delayed_flow = flow->time_next_packet;
+		bpf_spin_unlock(&fq_delayed_lock);
+		return 1;
+	}
+
+	node = bpf_rbtree_remove(&fq_delayed, &flow->rb_node);
+
+	bpf_spin_unlock(&fq_delayed_lock);
+
+	if (!node)
+		return 1;
+
+	flow = container_of(node, struct fq_flow_node, rb_node);
+	flow->age = 0;
+	fq_flows_add_tail(&fq_old_flows, &fq_old_flows_lock, flow, &q.old_flow_cnt);
+
+	return 0;
+}
+
+static void fq_flow_set_throttled(struct fq_flow_node *flow)
+{
+	flow->age = ~0ULL;
+
+	if (q.time_next_delayed_flow > flow->time_next_packet)
+		q.time_next_delayed_flow = flow->time_next_packet;
+
+	bpf_spin_lock(&fq_delayed_lock);
+	bpf_rbtree_add(&fq_delayed, &flow->rb_node, fn_time_next_packet_less);
+	bpf_spin_unlock(&fq_delayed_lock);
+}
+
+static void fq_check_throttled(u64 now)
+{
+	struct unset_throttled_flows_ctx ctx = {
+		.unset_all = false,
+		.now = now,
+	};
+	unsigned long sample;
+
+	if (q.time_next_delayed_flow > now)
+		return;
+
+	sample = (unsigned long)(now - q.time_next_delayed_flow);
+	q.unthrottle_latency_ns -= q.unthrottle_latency_ns >> 3;
+	q.unthrottle_latency_ns += sample >> 3;
+
+	q.time_next_delayed_flow = ~0ULL;
+	bpf_loop(NUM_QUEUE, fq_unset_throttled_flows, &ctx, 0);
+}
+
+static struct sk_buff*
+fq_dequeue_nonprio_flows(u32 index, struct dequeue_nonprio_ctx *ctx)
+{
+	u64 time_next_packet, time_to_send;
+	struct bpf_rb_node *rb_node;
+	struct sk_buff *skb = NULL;
+	struct bpf_list_head *head;
+	struct bpf_list_node *node;
+	struct bpf_spin_lock *lock;
+	struct fq_flow_node *flow;
+	struct skb_node *skbn;
+	bool is_empty;
+	u32 *cnt;
+
+	if (q.new_flow_cnt) {
+		head = &fq_new_flows;
+		lock = &fq_new_flows_lock;
+		cnt = &q.new_flow_cnt;
+	} else if (q.old_flow_cnt) {
+		head = &fq_old_flows;
+		lock = &fq_old_flows_lock;
+		cnt = &q.old_flow_cnt;
+	} else {
+		if (q.time_next_delayed_flow != ~0ULL)
+			ctx->expire = q.time_next_delayed_flow;
+		goto break_loop;
+	}
+
+	fq_flows_remove_front(head, lock, &node, cnt);
+	if (!node)
+		goto break_loop;
+
+	flow = container_of(node, struct fq_flow_node, list_node);
+	if (flow->credit <= 0) {
+		flow->credit += q.quantum;
+		fq_flows_add_tail(&fq_old_flows, &fq_old_flows_lock, flow, &q.old_flow_cnt);
+		return NULL;
+	}
+
+	bpf_spin_lock(&flow->lock);
+	rb_node = bpf_rbtree_first(&flow->queue);
+	if (!rb_node) {
+		bpf_spin_unlock(&flow->lock);
+		is_empty = fq_flows_is_empty(&fq_old_flows, &fq_old_flows_lock);
+		if (head == &fq_new_flows && !is_empty) {
+			fq_flows_add_tail(&fq_old_flows, &fq_old_flows_lock, flow, &q.old_flow_cnt);
+		} else {
+			fq_flow_set_detached(flow);
+			bpf_obj_drop(flow);
+		}
+		return NULL;
+	}
+
+	skbn = container_of(rb_node, struct skb_node, node);
+	time_to_send = skbn->tstamp;
+
+	time_next_packet = (time_to_send > flow->time_next_packet) ?
+		time_to_send : flow->time_next_packet;
+	if (ctx->now < time_next_packet) {
+		bpf_spin_unlock(&flow->lock);
+		flow->time_next_packet = time_next_packet;
+		fq_flow_set_throttled(flow);
+		return NULL;
+	}
+
+	rb_node = bpf_rbtree_remove(&flow->queue, rb_node);
+	bpf_spin_unlock(&flow->lock);
+
+	if (!rb_node)
+		goto add_flow_and_break;
+
+	skbn = container_of(rb_node, struct skb_node, node);
+	skb = bpf_kptr_xchg(&skbn->skb, skb);
+	bpf_obj_drop(skbn);
+
+	if (!skb)
+		goto add_flow_and_break;
+
+	flow->credit -= qdisc_skb_cb(skb)->pkt_len;
+	flow->qlen--;
+
+add_flow_and_break:
+	fq_flows_add_head(head, lock, flow, cnt);
+
+break_loop:
+	ctx->stop_iter = true;
+	return skb;
+}
+
+static struct sk_buff *fq_dequeue_prio(void)
+{
+	struct fq_flow_node *flow = NULL;
+	struct fq_stashed_flow *sflow;
+	struct bpf_rb_node *rb_node;
+	struct sk_buff *skb = NULL;
+	struct skb_node *skbn;
+	u64 hash = 0;
+
+	sflow = bpf_map_lookup_elem(&fq_prio_flows, &hash);
+	if (!sflow)
+		return NULL;
+
+	flow = bpf_kptr_xchg(&sflow->flow, flow);
+	if (!flow)
+		return NULL;
+
+	bpf_spin_lock(&flow->lock);
+	rb_node = bpf_rbtree_first(&flow->queue);
+	if (!rb_node) {
+		bpf_spin_unlock(&flow->lock);
+		goto out;
+	}
+
+	skbn = container_of(rb_node, struct skb_node, node);
+	rb_node = bpf_rbtree_remove(&flow->queue, &skbn->node);
+	bpf_spin_unlock(&flow->lock);
+
+	if (!rb_node)
+		goto out;
+
+	skbn = container_of(rb_node, struct skb_node, node);
+	skb = bpf_kptr_xchg(&skbn->skb, skb);
+	bpf_obj_drop(skbn);
+
+out:
+	bpf_kptr_xchg_back(&sflow->flow, flow);
+
+	return skb;
+}
+
+SEC("struct_ops/bpf_fq_dequeue")
+struct sk_buff *BPF_PROG(bpf_fq_dequeue, struct Qdisc *sch)
+{
+	struct dequeue_nonprio_ctx cb_ctx = {};
+	struct sk_buff *skb = NULL;
+	int i;
+
+	if (!sch->q.qlen)
+		goto out;
+
+	skb = fq_dequeue_prio();
+	if (skb)
+		goto dequeue;
+
+	q.ktime_cache = cb_ctx.now = bpf_ktime_get_ns();
+	fq_check_throttled(q.ktime_cache);
+	bpf_for(i, 0, sch->limit) {
+		skb = fq_dequeue_nonprio_flows(i, &cb_ctx);
+		if (cb_ctx.stop_iter)
+			break;
+	};
+
+	if (skb) {
+dequeue:
+		sch->q.qlen--;
+		sch->qstats.backlog -= qdisc_pkt_len(skb);
+		bpf_qdisc_bstats_update(sch, skb);
+		return skb;
+	}
+
+	if (cb_ctx.expire)
+		bpf_qdisc_watchdog_schedule(sch, cb_ctx.expire, q.timer_slack);
+out:
+	return NULL;
+}
+
+static int fq_remove_flows_in_list(u32 index, void *ctx)
+{
+	struct bpf_list_node *node;
+	struct fq_flow_node *flow;
+
+	bpf_spin_lock(&fq_new_flows_lock);
+	node = bpf_list_pop_front(&fq_new_flows);
+	bpf_spin_unlock(&fq_new_flows_lock);
+	if (!node) {
+		bpf_spin_lock(&fq_old_flows_lock);
+		node = bpf_list_pop_front(&fq_old_flows);
+		bpf_spin_unlock(&fq_old_flows_lock);
+		if (!node)
+			return 1;
+	}
+
+	flow = container_of(node, struct fq_flow_node, list_node);
+	bpf_obj_drop(flow);
+
+	return 0;
+}
+
+extern unsigned CONFIG_HZ __kconfig;
+
+/* limit number of collected flows per round */
+#define FQ_GC_MAX 8
+#define FQ_GC_AGE (3*CONFIG_HZ)
+
+static bool fq_gc_candidate(struct fq_flow_node *flow)
+{
+	u64 jiffies = bpf_jiffies64();
+
+	return fq_flow_is_detached(flow) &&
+	       ((s64)(jiffies - (flow->age + FQ_GC_AGE)) > 0);
+}
+
+static int
+fq_remove_flows(struct bpf_map *flow_map, u64 *hash,
+		struct fq_stashed_flow *sflow, struct remove_flows_ctx *ctx)
+{
+	if (sflow->flow &&
+	    (!ctx->gc_only || fq_gc_candidate(sflow->flow))) {
+		bpf_map_delete_elem(flow_map, hash);
+		ctx->reset_cnt++;
+	}
+
+	return ctx->reset_cnt < ctx->reset_max ? 0 : 1;
+}
+
+static void fq_gc(void)
+{
+	struct remove_flows_ctx cb_ctx = {
+		.gc_only = true,
+		.reset_cnt = 0,
+		.reset_max = FQ_GC_MAX,
+	};
+
+	bpf_for_each_map_elem(&fq_nonprio_flows, fq_remove_flows, &cb_ctx, 0);
+}
+
+SEC("struct_ops/bpf_fq_reset")
+void BPF_PROG(bpf_fq_reset, struct Qdisc *sch)
+{
+	struct unset_throttled_flows_ctx utf_ctx = {
+		.unset_all = true,
+	};
+	struct remove_flows_ctx rf_ctx = {
+		.gc_only = false,
+		.reset_cnt = 0,
+		.reset_max = NUM_QUEUE,
+	};
+	struct fq_stashed_flow *sflow;
+	u64 hash = 0;
+
+	sch->q.qlen = 0;
+	sch->qstats.backlog = 0;
+
+	bpf_for_each_map_elem(&fq_nonprio_flows, fq_remove_flows, &rf_ctx, 0);
+
+	rf_ctx.reset_cnt = 0;
+	bpf_for_each_map_elem(&fq_prio_flows, fq_remove_flows, &rf_ctx, 0);
+	fq_new_flow(&fq_prio_flows, &sflow, hash);
+
+	bpf_loop(NUM_QUEUE, fq_remove_flows_in_list, NULL, 0);
+	q.new_flow_cnt = 0;
+	q.old_flow_cnt = 0;
+
+	bpf_loop(NUM_QUEUE, fq_unset_throttled_flows, &utf_ctx, 0);
+}
+
+SEC("struct_ops/bpf_fq_init")
+int BPF_PROG(bpf_fq_init, struct Qdisc *sch, struct nlattr *opt,
+	     struct netlink_ext_ack *extack)
+{
+	struct net_device *dev = sch->dev_queue->dev;
+	u32 psched_mtu = dev->mtu + dev->hard_header_len;
+	struct fq_stashed_flow *sflow;
+	u64 hash = 0;
+
+	if (fq_new_flow(&fq_prio_flows, &sflow, hash) < 0)
+		return -ENOMEM;
+
+	sch->limit = 10000;
+	q.initial_quantum = 10 * psched_mtu;
+	q.quantum = 2 * psched_mtu;
+	q.flow_refill_delay = 40;
+	q.flow_plimit = 100;
+	q.horizon = 10ULL * NSEC_PER_SEC;
+	q.horizon_drop = 1;
+	q.orphan_mask = 1024 - 1;
+	q.timer_slack = 10 * NSEC_PER_USEC;
+	q.time_next_delayed_flow = ~0ULL;
+	q.unthrottle_latency_ns = 0ULL;
+	q.new_flow_cnt = 0;
+	q.old_flow_cnt = 0;
+
+	return 0;
+}
+
+SEC(".struct_ops")
+struct Qdisc_ops fq = {
+	.enqueue   = (void *)bpf_fq_enqueue,
+	.dequeue   = (void *)bpf_fq_dequeue,
+	.reset     = (void *)bpf_fq_reset,
+	.init      = (void *)bpf_fq_init,
+	.id        = "bpf_fq",
+};
author	Martin KaFai Lau <martin.lau@kernel.org>	2025-04-17 10:50:56 -0700
committer	Martin KaFai Lau <martin.lau@kernel.org>	2025-04-17 10:54:41 -0700
commit	fd23ce3eb4a1005bd109977856d12ec0fde7ef75 (patch)
tree	eed77cf13f3a6402654d6d66300f4198f48bb046 /tools
parent	ab734b43c81f8696bf89ff11b6c29a3dd884030a (diff)
parent	2b7b5b7f100e82ca314e76214626c82b608a1d8d (diff)
download	ath-fd23ce3eb4a1005bd109977856d12ec0fde7ef75.tar.gz