bpf: Sample HBM BPF program to limit egress bw

A cgroup skb BPF program to limit cgroup output bandwidth.
It uses a modified virtual token bucket queue to limit average
egress bandwidth. The implementation uses credits instead of tokens.
Negative credits imply that queueing would have happened (this is
a virtual queue, so no queueing is done by it. However, queueing may
occur at the actual qdisc (which is not used for rate limiting).

This implementation uses 3 thresholds, one to start marking packets and
the other two to drop packets:
                                 CREDIT
       - <--------------------------|------------------------> +
             |    |          |      0
             |  Large pkt    |
             |  drop thresh  |
  Small pkt drop             Mark threshold
      thresh

The effect of marking depends on the type of packet:
a) If the packet is ECN enabled, then the packet is ECN ce marked.
   The current mark threshold is tuned for DCTCP.
c) Else, it is dropped if it is a large packet.

If the credit is below the drop threshold, the packet is dropped.
Note that dropping a packet through the BPF program does not trigger CWR
(Congestion Window Reduction) in TCP packets. A future patch will add
support for triggering CWR.

This BPF program actually uses 2 drop thresholds, one threshold
for larger packets (>= 120 bytes) and another for smaller packets. This
protects smaller packets such as SYNs, ACKs, etc.

The default bandwidth limit is set at 1Gbps but this can be changed by
a user program through a shared BPF map. In addition, by default this BPF
program does not limit connections using loopback. This behavior can be
overwritten by the user program. There is also an option to calculate
some statistics, such as percent of packets marked or dropped, which
the user program can access.

A latter patch provides such a program (hbm.c)

Signed-off-by: Lawrence Brakmo <brakmo@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

4 files changed, 327 insertions, 0 deletions

diff --git a/samples/bpf/Makefile b/samples/bpf/Makefile
index 0c62ac39c697..e1bdc96486f6 100644
--- a/samples/bpf/Makefile
+++ b/samples/bpf/Makefile
@@ -164,6 +164,7 @@ always += xdp_adjust_tail_kern.o
 always += xdp_fwd_kern.o
 always += task_fd_query_kern.o
 always += xdp_sample_pkts_kern.o
+always += hbm_out_kern.o
 
 KBUILD_HOSTCFLAGS += -I$(objtree)/usr/include
 KBUILD_HOSTCFLAGS += -I$(srctree)/tools/lib/
@@ -263,6 +264,7 @@ $(BPF_SAMPLES_PATH)/*.c: verify_target_bpf $(LIBBPF)
 $(src)/*.c: verify_target_bpf $(LIBBPF)
 
 $(obj)/tracex5_kern.o: $(obj)/syscall_nrs.h
+$(obj)/hbm_out_kern.o: $(src)/hbm.h $(src)/hbm_kern.h
 
 # asm/sysreg.h - inline assembly used by it is incompatible with llvm.
 # But, there is no easy way to fix it, so just exclude it since it is
diff --git a/samples/bpf/hbm.h b/samples/bpf/hbm.h
new file mode 100644
index 000000000000..518e8147d084
--- /dev/null
+++ b/samples/bpf/hbm.h
@@ -0,0 +1,31 @@
+/* SPDX-License-Identifier: GPL-2.0
+ *
+ * Copyright (c) 2019 Facebook
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * Include file for Host Bandwidth Management (HBM) programs
+ */
+struct hbm_vqueue {
+        struct bpf_spin_lock lock;
+        /* 4 byte hole */
+        unsigned long long lasttime;    /* In ns */
+        int credit;                     /* In bytes */
+        unsigned int rate;              /* In bytes per NS << 20 */
+};
+
+struct hbm_queue_stats {
+        unsigned long rate;             /* in Mbps*/
+        unsigned long stats:1,          /* get HBM stats (marked, dropped,..) */
+                loopback:1;             /* also limit flows using loopback */
+        unsigned long long pkts_marked;
+        unsigned long long bytes_marked;
+        unsigned long long pkts_dropped;
+        unsigned long long bytes_dropped;
+        unsigned long long pkts_total;
+        unsigned long long bytes_total;
+        unsigned long long firstPacketTime;
+        unsigned long long lastPacketTime;
+};
diff --git a/samples/bpf/hbm_kern.h b/samples/bpf/hbm_kern.h
new file mode 100644
index 000000000000..c5635d924193
--- /dev/null
+++ b/samples/bpf/hbm_kern.h
@@ -0,0 +1,137 @@
+/* SPDX-License-Identifier: GPL-2.0
+ *
+ * Copyright (c) 2019 Facebook
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * Include file for sample Host Bandwidth Manager (HBM) BPF programs
+ */
+#define KBUILD_MODNAME "foo"
+#include <stddef.h>
+#include <stdbool.h>
+#include <uapi/linux/bpf.h>
+#include <uapi/linux/if_ether.h>
+#include <uapi/linux/if_packet.h>
+#include <uapi/linux/ip.h>
+#include <uapi/linux/ipv6.h>
+#include <uapi/linux/in.h>
+#include <uapi/linux/tcp.h>
+#include <uapi/linux/filter.h>
+#include <uapi/linux/pkt_cls.h>
+#include <net/ipv6.h>
+#include <net/inet_ecn.h>
+#include "bpf_endian.h"
+#include "bpf_helpers.h"
+#include "hbm.h"
+
+#define DROP_PKT        0
+#define ALLOW_PKT       1
+#define TCP_ECN_OK      1
+
+#define HBM_DEBUG 0  // Set to 1 to enable debugging
+#if HBM_DEBUG
+#define bpf_printk(fmt, ...)                                    \
+({                                                              \
+        char ____fmt[] = fmt;                                   \
+        bpf_trace_printk(____fmt, sizeof(____fmt),              \
+                         ##__VA_ARGS__);                        \
+})
+#else
+#define bpf_printk(fmt, ...)
+#endif
+
+#define INITIAL_CREDIT_PACKETS  100
+#define MAX_BYTES_PER_PACKET    1500
+#define MARK_THRESH             (40 * MAX_BYTES_PER_PACKET)
+#define DROP_THRESH             (80 * 5 * MAX_BYTES_PER_PACKET)
+#define LARGE_PKT_DROP_THRESH   (DROP_THRESH - (15 * MAX_BYTES_PER_PACKET))
+#define MARK_REGION_SIZE        (LARGE_PKT_DROP_THRESH - MARK_THRESH)
+#define LARGE_PKT_THRESH        120
+#define MAX_CREDIT              (100 * MAX_BYTES_PER_PACKET)
+#define INIT_CREDIT             (INITIAL_CREDIT_PACKETS * MAX_BYTES_PER_PACKET)
+
+// rate in bytes per ns << 20
+#define CREDIT_PER_NS(delta, rate) ((((u64)(delta)) * (rate)) >> 20)
+
+struct bpf_map_def SEC("maps") queue_state = {
+        .type = BPF_MAP_TYPE_CGROUP_STORAGE,
+        .key_size = sizeof(struct bpf_cgroup_storage_key),
+        .value_size = sizeof(struct hbm_vqueue),
+};
+BPF_ANNOTATE_KV_PAIR(queue_state, struct bpf_cgroup_storage_key,
+                     struct hbm_vqueue);
+
+struct bpf_map_def SEC("maps") queue_stats = {
+        .type = BPF_MAP_TYPE_ARRAY,
+        .key_size = sizeof(u32),
+        .value_size = sizeof(struct hbm_queue_stats),
+        .max_entries = 1,
+};
+BPF_ANNOTATE_KV_PAIR(queue_stats, int, struct hbm_queue_stats);
+
+struct hbm_pkt_info {
+        bool    is_ip;
+        bool    is_tcp;
+        short   ecn;
+};
+
+static __always_inline void hbm_get_pkt_info(struct __sk_buff *skb,
+                                             struct hbm_pkt_info *pkti)
+{
+        struct iphdr iph;
+        struct ipv6hdr *ip6h;
+
+        bpf_skb_load_bytes(skb, 0, &iph, 12);
+        if (iph.version == 6) {
+                ip6h = (struct ipv6hdr *)&iph;
+                pkti->is_ip = true;
+                pkti->is_tcp = (ip6h->nexthdr == 6);
+                pkti->ecn = (ip6h->flow_lbl[0] >> 4) & INET_ECN_MASK;
+        } else if (iph.version == 4) {
+                pkti->is_ip = true;
+                pkti->is_tcp = (iph.protocol == 6);
+                pkti->ecn = iph.tos & INET_ECN_MASK;
+        } else {
+                pkti->is_ip = false;
+                pkti->is_tcp = false;
+                pkti->ecn = 0;
+        }
+}
+
+static __always_inline void hbm_init_vqueue(struct hbm_vqueue *qdp, int rate)
+{
+                bpf_printk("Initializing queue_state, rate:%d\n", rate * 128);
+                qdp->lasttime = bpf_ktime_get_ns();
+                qdp->credit = INIT_CREDIT;
+                qdp->rate = rate * 128;
+}
+
+static __always_inline void hbm_update_stats(struct hbm_queue_stats *qsp,
+                                             int len,
+                                             unsigned long long curtime,
+                                             bool congestion_flag,
+                                             bool drop_flag)
+{
+        if (qsp != NULL) {
+                // Following is needed for work conserving
+                __sync_add_and_fetch(&(qsp->bytes_total), len);
+                if (qsp->stats) {
+                        // Optionally update statistics
+                        if (qsp->firstPacketTime == 0)
+                                qsp->firstPacketTime = curtime;
+                        qsp->lastPacketTime = curtime;
+                        __sync_add_and_fetch(&(qsp->pkts_total), 1);
+                        if (congestion_flag || drop_flag) {
+                                __sync_add_and_fetch(&(qsp->pkts_marked), 1);
+                                __sync_add_and_fetch(&(qsp->bytes_marked), len);
+                        }
+                        if (drop_flag) {
+                                __sync_add_and_fetch(&(qsp->pkts_dropped), 1);
+                                __sync_add_and_fetch(&(qsp->bytes_dropped),
+                                                     len);
+                        }
+                }
+        }
+}
diff --git a/samples/bpf/hbm_out_kern.c b/samples/bpf/hbm_out_kern.c
new file mode 100644
index 000000000000..f806863d0b79
--- /dev/null
+++ b/samples/bpf/hbm_out_kern.c
@@ -0,0 +1,157 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2019 Facebook
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ *
+ * Sample Host Bandwidth Manager (HBM) BPF program.
+ *
+ * A cgroup skb BPF egress program to limit cgroup output bandwidth.
+ * It uses a modified virtual token bucket queue to limit average
+ * egress bandwidth. The implementation uses credits instead of tokens.
+ * Negative credits imply that queueing would have happened (this is
+ * a virtual queue, so no queueing is done by it. However, queueing may
+ * occur at the actual qdisc (which is not used for rate limiting).
+ *
+ * This implementation uses 3 thresholds, one to start marking packets and
+ * the other two to drop packets:
+ *                                  CREDIT
+ *        - <--------------------------|------------------------> +
+ *              |    |          |      0
+ *              |  Large pkt    |
+ *              |  drop thresh  |
+ *   Small pkt drop             Mark threshold
+ *       thresh
+ *
+ * The effect of marking depends on the type of packet:
+ * a) If the packet is ECN enabled and it is a TCP packet, then the packet
+ *    is ECN marked.
+ * b) If the packet is a TCP packet, then we probabilistically call tcp_cwr
+ *    to reduce the congestion window. The current implementation uses a linear
+ *    distribution (0% probability at marking threshold, 100% probability
+ *    at drop threshold).
+ * c) If the packet is not a TCP packet, then it is dropped.
+ *
+ * If the credit is below the drop threshold, the packet is dropped. If it
+ * is a TCP packet, then it also calls tcp_cwr since packets dropped by
+ * by a cgroup skb BPF program do not automatically trigger a call to
+ * tcp_cwr in the current kernel code.
+ *
+ * This BPF program actually uses 2 drop thresholds, one threshold
+ * for larger packets (>= 120 bytes) and another for smaller packets. This
+ * protects smaller packets such as SYNs, ACKs, etc.
+ *
+ * The default bandwidth limit is set at 1Gbps but this can be changed by
+ * a user program through a shared BPF map. In addition, by default this BPF
+ * program does not limit connections using loopback. This behavior can be
+ * overwritten by the user program. There is also an option to calculate
+ * some statistics, such as percent of packets marked or dropped, which
+ * the user program can access.
+ *
+ * A latter patch provides such a program (hbm.c)
+ */
+
+#include "hbm_kern.h"
+
+SEC("cgroup_skb/egress")
+int _hbm_out_cg(struct __sk_buff *skb)
+{
+        struct hbm_pkt_info pkti;
+        int len = skb->len;
+        unsigned int queue_index = 0;
+        unsigned long long curtime;
+        int credit;
+        signed long long delta = 0, zero = 0;
+        int max_credit = MAX_CREDIT;
+        bool congestion_flag = false;
+        bool drop_flag = false;
+        bool cwr_flag = false;
+        struct hbm_vqueue *qdp;
+        struct hbm_queue_stats *qsp = NULL;
+        int rv = ALLOW_PKT;
+
+        qsp = bpf_map_lookup_elem(&queue_stats, &queue_index);
+        if (qsp != NULL && !qsp->loopback && (skb->ifindex == 1))
+                return ALLOW_PKT;
+
+        hbm_get_pkt_info(skb, &pkti);
+
+        // We may want to account for the length of headers in len
+        // calculation, like ETH header + overhead, specially if it
+        // is a gso packet. But I am not doing it right now.
+
+        qdp = bpf_get_local_storage(&queue_state, 0);
+        if (!qdp)
+                return ALLOW_PKT;
+        else if (qdp->lasttime == 0)
+                hbm_init_vqueue(qdp, 1024);
+
+        curtime = bpf_ktime_get_ns();
+
+        // Begin critical section
+        bpf_spin_lock(&qdp->lock);
+        credit = qdp->credit;
+        delta = curtime - qdp->lasttime;
+        /* delta < 0 implies that another process with a curtime greater
+         * than ours beat us to the critical section and already added
+         * the new credit, so we should not add it ourselves
+         */
+        if (delta > 0) {
+                qdp->lasttime = curtime;
+                credit += CREDIT_PER_NS(delta, qdp->rate);
+                if (credit > MAX_CREDIT)
+                        credit = MAX_CREDIT;
+        }
+        credit -= len;
+        qdp->credit = credit;
+        bpf_spin_unlock(&qdp->lock);
+        // End critical section
+
+        // Check if we should update rate
+        if (qsp != NULL && (qsp->rate * 128) != qdp->rate) {
+                qdp->rate = qsp->rate * 128;
+                bpf_printk("Updating rate: %d (1sec:%llu bits)\n",
+                           (int)qdp->rate,
+                           CREDIT_PER_NS(1000000000, qdp->rate) * 8);
+        }
+
+        // Set flags (drop, congestion, cwr)
+        // Dropping => we are congested, so ignore congestion flag
+        if (credit < -DROP_THRESH ||
+            (len > LARGE_PKT_THRESH &&
+             credit < -LARGE_PKT_DROP_THRESH)) {
+                // Very congested, set drop flag
+                drop_flag = true;
+        } else if (credit < 0) {
+                // Congested, set congestion flag
+                if (pkti.ecn) {
+                        if (credit < -MARK_THRESH)
+                                congestion_flag = true;
+                        else
+                                congestion_flag = false;
+                } else {
+                        congestion_flag = true;
+                }
+        }
+
+        if (congestion_flag) {
+                if (!bpf_skb_ecn_set_ce(skb)) {
+                        if (len > LARGE_PKT_THRESH) {
+                                // Problem if too many small packets?
+                                drop_flag = true;
+                        }
+                }
+        }
+
+        if (drop_flag)
+                rv = DROP_PKT;
+
+        hbm_update_stats(qsp, len, curtime, congestion_flag, drop_flag);
+
+        if (rv == DROP_PKT)
+                __sync_add_and_fetch(&(qdp->credit), len);
+
+        return rv;
+}
+char _license[] SEC("license") = "GPL";