1 | // SPDX-License-Identifier: GPL-2.0-or-later |
2 | /* DataCenter TCP (DCTCP) congestion control. |
3 | * |
4 | * http://simula.stanford.edu/~alizade/Site/DCTCP.html |
5 | * |
6 | * This is an implementation of DCTCP over Reno, an enhancement to the |
7 | * TCP congestion control algorithm designed for data centers. DCTCP |
8 | * leverages Explicit Congestion Notification (ECN) in the network to |
9 | * provide multi-bit feedback to the end hosts. DCTCP's goal is to meet |
10 | * the following three data center transport requirements: |
11 | * |
12 | * - High burst tolerance (incast due to partition/aggregate) |
13 | * - Low latency (short flows, queries) |
14 | * - High throughput (continuous data updates, large file transfers) |
15 | * with commodity shallow buffered switches |
16 | * |
17 | * The algorithm is described in detail in the following two papers: |
18 | * |
19 | * 1) Mohammad Alizadeh, Albert Greenberg, David A. Maltz, Jitendra Padhye, |
20 | * Parveen Patel, Balaji Prabhakar, Sudipta Sengupta, and Murari Sridharan: |
21 | * "Data Center TCP (DCTCP)", Data Center Networks session |
22 | * Proc. ACM SIGCOMM, New Delhi, 2010. |
23 | * http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp-final.pdf |
24 | * |
25 | * 2) Mohammad Alizadeh, Adel Javanmard, and Balaji Prabhakar: |
26 | * "Analysis of DCTCP: Stability, Convergence, and Fairness" |
27 | * Proc. ACM SIGMETRICS, San Jose, 2011. |
28 | * http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp_analysis-full.pdf |
29 | * |
30 | * Initial prototype from Abdul Kabbani, Masato Yasuda and Mohammad Alizadeh. |
31 | * |
32 | * Authors: |
33 | * |
34 | * Daniel Borkmann <dborkman@redhat.com> |
35 | * Florian Westphal <fw@strlen.de> |
36 | * Glenn Judd <glenn.judd@morganstanley.com> |
37 | */ |
38 | |
39 | #include <linux/btf.h> |
40 | #include <linux/btf_ids.h> |
41 | #include <linux/module.h> |
42 | #include <linux/mm.h> |
43 | #include <net/tcp.h> |
44 | #include <linux/inet_diag.h> |
45 | #include "tcp_dctcp.h" |
46 | |
47 | #define DCTCP_MAX_ALPHA 1024U |
48 | |
49 | struct dctcp { |
50 | u32 old_delivered; |
51 | u32 old_delivered_ce; |
52 | u32 prior_rcv_nxt; |
53 | u32 dctcp_alpha; |
54 | u32 next_seq; |
55 | u32 ce_state; |
56 | u32 loss_cwnd; |
57 | struct tcp_plb_state plb; |
58 | }; |
59 | |
60 | static unsigned int dctcp_shift_g __read_mostly = 4; /* g = 1/2^4 */ |
61 | module_param(dctcp_shift_g, uint, 0644); |
62 | MODULE_PARM_DESC(dctcp_shift_g, "parameter g for updating dctcp_alpha" ); |
63 | |
64 | static unsigned int dctcp_alpha_on_init __read_mostly = DCTCP_MAX_ALPHA; |
65 | module_param(dctcp_alpha_on_init, uint, 0644); |
66 | MODULE_PARM_DESC(dctcp_alpha_on_init, "parameter for initial alpha value" ); |
67 | |
68 | static struct tcp_congestion_ops dctcp_reno; |
69 | |
70 | static void dctcp_reset(const struct tcp_sock *tp, struct dctcp *ca) |
71 | { |
72 | ca->next_seq = tp->snd_nxt; |
73 | |
74 | ca->old_delivered = tp->delivered; |
75 | ca->old_delivered_ce = tp->delivered_ce; |
76 | } |
77 | |
78 | __bpf_kfunc static void dctcp_init(struct sock *sk) |
79 | { |
80 | const struct tcp_sock *tp = tcp_sk(sk); |
81 | |
82 | if ((tp->ecn_flags & TCP_ECN_OK) || |
83 | (sk->sk_state == TCP_LISTEN || |
84 | sk->sk_state == TCP_CLOSE)) { |
85 | struct dctcp *ca = inet_csk_ca(sk); |
86 | |
87 | ca->prior_rcv_nxt = tp->rcv_nxt; |
88 | |
89 | ca->dctcp_alpha = min(dctcp_alpha_on_init, DCTCP_MAX_ALPHA); |
90 | |
91 | ca->loss_cwnd = 0; |
92 | ca->ce_state = 0; |
93 | |
94 | dctcp_reset(tp, ca); |
95 | tcp_plb_init(sk, plb: &ca->plb); |
96 | |
97 | return; |
98 | } |
99 | |
100 | /* No ECN support? Fall back to Reno. Also need to clear |
101 | * ECT from sk since it is set during 3WHS for DCTCP. |
102 | */ |
103 | inet_csk(sk)->icsk_ca_ops = &dctcp_reno; |
104 | INET_ECN_dontxmit(sk); |
105 | } |
106 | |
107 | __bpf_kfunc static u32 dctcp_ssthresh(struct sock *sk) |
108 | { |
109 | struct dctcp *ca = inet_csk_ca(sk); |
110 | struct tcp_sock *tp = tcp_sk(sk); |
111 | |
112 | ca->loss_cwnd = tcp_snd_cwnd(tp); |
113 | return max(tcp_snd_cwnd(tp) - ((tcp_snd_cwnd(tp) * ca->dctcp_alpha) >> 11U), 2U); |
114 | } |
115 | |
116 | __bpf_kfunc static void dctcp_update_alpha(struct sock *sk, u32 flags) |
117 | { |
118 | const struct tcp_sock *tp = tcp_sk(sk); |
119 | struct dctcp *ca = inet_csk_ca(sk); |
120 | |
121 | /* Expired RTT */ |
122 | if (!before(seq1: tp->snd_una, seq2: ca->next_seq)) { |
123 | u32 delivered = tp->delivered - ca->old_delivered; |
124 | u32 delivered_ce = tp->delivered_ce - ca->old_delivered_ce; |
125 | u32 alpha = ca->dctcp_alpha; |
126 | u32 ce_ratio = 0; |
127 | |
128 | if (delivered > 0) { |
129 | /* dctcp_alpha keeps EWMA of fraction of ECN marked |
130 | * packets. Because of EWMA smoothing, PLB reaction can |
131 | * be slow so we use ce_ratio which is an instantaneous |
132 | * measure of congestion. ce_ratio is the fraction of |
133 | * ECN marked packets in the previous RTT. |
134 | */ |
135 | if (delivered_ce > 0) |
136 | ce_ratio = (delivered_ce << TCP_PLB_SCALE) / delivered; |
137 | tcp_plb_update_state(sk, plb: &ca->plb, cong_ratio: (int)ce_ratio); |
138 | tcp_plb_check_rehash(sk, plb: &ca->plb); |
139 | } |
140 | |
141 | /* alpha = (1 - g) * alpha + g * F */ |
142 | |
143 | alpha -= min_not_zero(alpha, alpha >> dctcp_shift_g); |
144 | if (delivered_ce) { |
145 | |
146 | /* If dctcp_shift_g == 1, a 32bit value would overflow |
147 | * after 8 M packets. |
148 | */ |
149 | delivered_ce <<= (10 - dctcp_shift_g); |
150 | delivered_ce /= max(1U, delivered); |
151 | |
152 | alpha = min(alpha + delivered_ce, DCTCP_MAX_ALPHA); |
153 | } |
154 | /* dctcp_alpha can be read from dctcp_get_info() without |
155 | * synchro, so we ask compiler to not use dctcp_alpha |
156 | * as a temporary variable in prior operations. |
157 | */ |
158 | WRITE_ONCE(ca->dctcp_alpha, alpha); |
159 | dctcp_reset(tp, ca); |
160 | } |
161 | } |
162 | |
163 | static void dctcp_react_to_loss(struct sock *sk) |
164 | { |
165 | struct dctcp *ca = inet_csk_ca(sk); |
166 | struct tcp_sock *tp = tcp_sk(sk); |
167 | |
168 | ca->loss_cwnd = tcp_snd_cwnd(tp); |
169 | tp->snd_ssthresh = max(tcp_snd_cwnd(tp) >> 1U, 2U); |
170 | } |
171 | |
172 | __bpf_kfunc static void dctcp_state(struct sock *sk, u8 new_state) |
173 | { |
174 | if (new_state == TCP_CA_Recovery && |
175 | new_state != inet_csk(sk)->icsk_ca_state) |
176 | dctcp_react_to_loss(sk); |
177 | /* We handle RTO in dctcp_cwnd_event to ensure that we perform only |
178 | * one loss-adjustment per RTT. |
179 | */ |
180 | } |
181 | |
182 | __bpf_kfunc static void dctcp_cwnd_event(struct sock *sk, enum tcp_ca_event ev) |
183 | { |
184 | struct dctcp *ca = inet_csk_ca(sk); |
185 | |
186 | switch (ev) { |
187 | case CA_EVENT_ECN_IS_CE: |
188 | case CA_EVENT_ECN_NO_CE: |
189 | dctcp_ece_ack_update(sk, evt: ev, prior_rcv_nxt: &ca->prior_rcv_nxt, ce_state: &ca->ce_state); |
190 | break; |
191 | case CA_EVENT_LOSS: |
192 | tcp_plb_update_state_upon_rto(sk, plb: &ca->plb); |
193 | dctcp_react_to_loss(sk); |
194 | break; |
195 | case CA_EVENT_TX_START: |
196 | tcp_plb_check_rehash(sk, plb: &ca->plb); /* Maybe rehash when inflight is 0 */ |
197 | break; |
198 | default: |
199 | /* Don't care for the rest. */ |
200 | break; |
201 | } |
202 | } |
203 | |
204 | static size_t dctcp_get_info(struct sock *sk, u32 ext, int *attr, |
205 | union tcp_cc_info *info) |
206 | { |
207 | const struct dctcp *ca = inet_csk_ca(sk); |
208 | const struct tcp_sock *tp = tcp_sk(sk); |
209 | |
210 | /* Fill it also in case of VEGASINFO due to req struct limits. |
211 | * We can still correctly retrieve it later. |
212 | */ |
213 | if (ext & (1 << (INET_DIAG_DCTCPINFO - 1)) || |
214 | ext & (1 << (INET_DIAG_VEGASINFO - 1))) { |
215 | memset(&info->dctcp, 0, sizeof(info->dctcp)); |
216 | if (inet_csk(sk)->icsk_ca_ops != &dctcp_reno) { |
217 | info->dctcp.dctcp_enabled = 1; |
218 | info->dctcp.dctcp_ce_state = (u16) ca->ce_state; |
219 | info->dctcp.dctcp_alpha = ca->dctcp_alpha; |
220 | info->dctcp.dctcp_ab_ecn = tp->mss_cache * |
221 | (tp->delivered_ce - ca->old_delivered_ce); |
222 | info->dctcp.dctcp_ab_tot = tp->mss_cache * |
223 | (tp->delivered - ca->old_delivered); |
224 | } |
225 | |
226 | *attr = INET_DIAG_DCTCPINFO; |
227 | return sizeof(info->dctcp); |
228 | } |
229 | return 0; |
230 | } |
231 | |
232 | __bpf_kfunc static u32 dctcp_cwnd_undo(struct sock *sk) |
233 | { |
234 | const struct dctcp *ca = inet_csk_ca(sk); |
235 | struct tcp_sock *tp = tcp_sk(sk); |
236 | |
237 | return max(tcp_snd_cwnd(tp), ca->loss_cwnd); |
238 | } |
239 | |
240 | static struct tcp_congestion_ops dctcp __read_mostly = { |
241 | .init = dctcp_init, |
242 | .in_ack_event = dctcp_update_alpha, |
243 | .cwnd_event = dctcp_cwnd_event, |
244 | .ssthresh = dctcp_ssthresh, |
245 | .cong_avoid = tcp_reno_cong_avoid, |
246 | .undo_cwnd = dctcp_cwnd_undo, |
247 | .set_state = dctcp_state, |
248 | .get_info = dctcp_get_info, |
249 | .flags = TCP_CONG_NEEDS_ECN, |
250 | .owner = THIS_MODULE, |
251 | .name = "dctcp" , |
252 | }; |
253 | |
254 | static struct tcp_congestion_ops dctcp_reno __read_mostly = { |
255 | .ssthresh = tcp_reno_ssthresh, |
256 | .cong_avoid = tcp_reno_cong_avoid, |
257 | .undo_cwnd = tcp_reno_undo_cwnd, |
258 | .get_info = dctcp_get_info, |
259 | .owner = THIS_MODULE, |
260 | .name = "dctcp-reno" , |
261 | }; |
262 | |
263 | BTF_SET8_START(tcp_dctcp_check_kfunc_ids) |
264 | #ifdef CONFIG_X86 |
265 | #ifdef CONFIG_DYNAMIC_FTRACE |
266 | BTF_ID_FLAGS(func, dctcp_init) |
267 | BTF_ID_FLAGS(func, dctcp_update_alpha) |
268 | BTF_ID_FLAGS(func, dctcp_cwnd_event) |
269 | BTF_ID_FLAGS(func, dctcp_ssthresh) |
270 | BTF_ID_FLAGS(func, dctcp_cwnd_undo) |
271 | BTF_ID_FLAGS(func, dctcp_state) |
272 | #endif |
273 | #endif |
274 | BTF_SET8_END(tcp_dctcp_check_kfunc_ids) |
275 | |
276 | static const struct btf_kfunc_id_set tcp_dctcp_kfunc_set = { |
277 | .owner = THIS_MODULE, |
278 | .set = &tcp_dctcp_check_kfunc_ids, |
279 | }; |
280 | |
281 | static int __init dctcp_register(void) |
282 | { |
283 | int ret; |
284 | |
285 | BUILD_BUG_ON(sizeof(struct dctcp) > ICSK_CA_PRIV_SIZE); |
286 | |
287 | ret = register_btf_kfunc_id_set(prog_type: BPF_PROG_TYPE_STRUCT_OPS, s: &tcp_dctcp_kfunc_set); |
288 | if (ret < 0) |
289 | return ret; |
290 | return tcp_register_congestion_control(type: &dctcp); |
291 | } |
292 | |
293 | static void __exit dctcp_unregister(void) |
294 | { |
295 | tcp_unregister_congestion_control(type: &dctcp); |
296 | } |
297 | |
298 | module_init(dctcp_register); |
299 | module_exit(dctcp_unregister); |
300 | |
301 | MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>" ); |
302 | MODULE_AUTHOR("Florian Westphal <fw@strlen.de>" ); |
303 | MODULE_AUTHOR("Glenn Judd <glenn.judd@morganstanley.com>" ); |
304 | |
305 | MODULE_LICENSE("GPL v2" ); |
306 | MODULE_DESCRIPTION("DataCenter TCP (DCTCP)" ); |
307 | |