187d0738 |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* Copyright (c) 2019 Facebook |
3 | * |
4 | * This program is free software; you can redistribute it and/or |
5 | * modify it under the terms of version 2 of the GNU General Public |
6 | * License as published by the Free Software Foundation. |
7 | * |
8 | * Sample Host Bandwidth Manager (HBM) BPF program. |
9 | * |
10 | * A cgroup skb BPF egress program to limit cgroup output bandwidth. |
11 | * It uses a modified virtual token bucket queue to limit average |
12 | * egress bandwidth. The implementation uses credits instead of tokens. |
13 | * Negative credits imply that queueing would have happened (this is |
14 | * a virtual queue, so no queueing is done by it. However, queueing may |
15 | * occur at the actual qdisc (which is not used for rate limiting). |
16 | * |
17 | * This implementation uses 3 thresholds, one to start marking packets and |
18 | * the other two to drop packets: |
19 | * CREDIT |
20 | * - <--------------------------|------------------------> + |
21 | * | | | 0 |
22 | * | Large pkt | |
23 | * | drop thresh | |
24 | * Small pkt drop Mark threshold |
25 | * thresh |
26 | * |
27 | * The effect of marking depends on the type of packet: |
28 | * a) If the packet is ECN enabled and it is a TCP packet, then the packet |
29 | * is ECN marked. |
30 | * b) If the packet is a TCP packet, then we probabilistically call tcp_cwr |
31 | * to reduce the congestion window. The current implementation uses a linear |
32 | * distribution (0% probability at marking threshold, 100% probability |
33 | * at drop threshold). |
34 | * c) If the packet is not a TCP packet, then it is dropped. |
35 | * |
36 | * If the credit is below the drop threshold, the packet is dropped. If it |
37 | * is a TCP packet, then it also calls tcp_cwr since packets dropped by |
38 | * by a cgroup skb BPF program do not automatically trigger a call to |
39 | * tcp_cwr in the current kernel code. |
40 | * |
41 | * This BPF program actually uses 2 drop thresholds, one threshold |
42 | * for larger packets (>= 120 bytes) and another for smaller packets. This |
43 | * protects smaller packets such as SYNs, ACKs, etc. |
44 | * |
45 | * The default bandwidth limit is set at 1Gbps but this can be changed by |
46 | * a user program through a shared BPF map. In addition, by default this BPF |
47 | * program does not limit connections using loopback. This behavior can be |
48 | * overwritten by the user program. There is also an option to calculate |
49 | * some statistics, such as percent of packets marked or dropped, which |
50 | * the user program can access. |
51 | * |
52 | * A latter patch provides such a program (hbm.c) |
53 | */ |
54 | |
55 | #include "hbm_kern.h" |
56 | |
57 | SEC("cgroup_skb/egress") |
58 | int _hbm_out_cg(struct __sk_buff *skb) |
59 | { |
60 | struct hbm_pkt_info pkti; |
61 | int len = skb->len; |
62 | unsigned int queue_index = 0; |
63 | unsigned long long curtime; |
64 | int credit; |
d58c6f72 |
65 | signed long long delta = 0, new_credit; |
187d0738 |
66 | int max_credit = MAX_CREDIT; |
67 | bool congestion_flag = false; |
68 | bool drop_flag = false; |
69 | bool cwr_flag = false; |
d58c6f72 |
70 | bool ecn_ce_flag = false; |
187d0738 |
71 | struct hbm_vqueue *qdp; |
72 | struct hbm_queue_stats *qsp = NULL; |
73 | int rv = ALLOW_PKT; |
74 | |
75 | qsp = bpf_map_lookup_elem(&queue_stats, &queue_index); |
76 | if (qsp != NULL && !qsp->loopback && (skb->ifindex == 1)) |
77 | return ALLOW_PKT; |
78 | |
79 | hbm_get_pkt_info(skb, &pkti); |
80 | |
81 | // We may want to account for the length of headers in len |
82 | // calculation, like ETH header + overhead, specially if it |
83 | // is a gso packet. But I am not doing it right now. |
84 | |
85 | qdp = bpf_get_local_storage(&queue_state, 0); |
86 | if (!qdp) |
87 | return ALLOW_PKT; |
88 | else if (qdp->lasttime == 0) |
89 | hbm_init_vqueue(qdp, 1024); |
90 | |
91 | curtime = bpf_ktime_get_ns(); |
92 | |
93 | // Begin critical section |
94 | bpf_spin_lock(&qdp->lock); |
95 | credit = qdp->credit; |
96 | delta = curtime - qdp->lasttime; |
97 | /* delta < 0 implies that another process with a curtime greater |
98 | * than ours beat us to the critical section and already added |
99 | * the new credit, so we should not add it ourselves |
100 | */ |
101 | if (delta > 0) { |
102 | qdp->lasttime = curtime; |
d58c6f72 |
103 | new_credit = credit + CREDIT_PER_NS(delta, qdp->rate); |
104 | if (new_credit > MAX_CREDIT) |
187d0738 |
105 | credit = MAX_CREDIT; |
d58c6f72 |
106 | else |
107 | credit = new_credit; |
187d0738 |
108 | } |
109 | credit -= len; |
110 | qdp->credit = credit; |
111 | bpf_spin_unlock(&qdp->lock); |
112 | // End critical section |
113 | |
114 | // Check if we should update rate |
115 | if (qsp != NULL && (qsp->rate * 128) != qdp->rate) { |
116 | qdp->rate = qsp->rate * 128; |
117 | bpf_printk("Updating rate: %d (1sec:%llu bits)\n", |
118 | (int)qdp->rate, |
119 | CREDIT_PER_NS(1000000000, qdp->rate) * 8); |
120 | } |
121 | |
122 | // Set flags (drop, congestion, cwr) |
123 | // Dropping => we are congested, so ignore congestion flag |
124 | if (credit < -DROP_THRESH || |
ffd81558 |
125 | (len > LARGE_PKT_THRESH && credit < -LARGE_PKT_DROP_THRESH)) { |
126 | // Very congested, set drop packet |
187d0738 |
127 | drop_flag = true; |
ffd81558 |
128 | if (pkti.ecn) |
129 | congestion_flag = true; |
130 | else if (pkti.is_tcp) |
131 | cwr_flag = true; |
187d0738 |
132 | } else if (credit < 0) { |
133 | // Congested, set congestion flag |
ffd81558 |
134 | if (pkti.ecn || pkti.is_tcp) { |
187d0738 |
135 | if (credit < -MARK_THRESH) |
136 | congestion_flag = true; |
137 | else |
138 | congestion_flag = false; |
139 | } else { |
140 | congestion_flag = true; |
141 | } |
142 | } |
143 | |
144 | if (congestion_flag) { |
d58c6f72 |
145 | if (bpf_skb_ecn_set_ce(skb)) { |
146 | ecn_ce_flag = true; |
147 | } else { |
ffd81558 |
148 | if (pkti.is_tcp) { |
149 | unsigned int rand = bpf_get_prandom_u32(); |
150 | |
151 | if (-credit >= MARK_THRESH + |
152 | (rand % MARK_REGION_SIZE)) { |
153 | // Do congestion control |
154 | cwr_flag = true; |
155 | } |
156 | } else if (len > LARGE_PKT_THRESH) { |
187d0738 |
157 | // Problem if too many small packets? |
158 | drop_flag = true; |
159 | } |
160 | } |
161 | } |
162 | |
ffd81558 |
163 | if (qsp != NULL) |
164 | if (qsp->no_cn) |
165 | cwr_flag = false; |
187d0738 |
166 | |
d58c6f72 |
167 | hbm_update_stats(qsp, len, curtime, congestion_flag, drop_flag, |
168 | cwr_flag, ecn_ce_flag, &pkti, credit); |
187d0738 |
169 | |
d58c6f72 |
170 | if (drop_flag) { |
187d0738 |
171 | __sync_add_and_fetch(&(qdp->credit), len); |
d58c6f72 |
172 | rv = DROP_PKT; |
173 | } |
187d0738 |
174 | |
ffd81558 |
175 | if (cwr_flag) |
176 | rv |= 2; |
187d0738 |
177 | return rv; |
178 | } |
179 | char _license[] SEC("license") = "GPL"; |