Commit | Line | Data |
---|---|---|
2874c5fd | 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
e3118e83 DB |
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> | |
e3118e83 DB |
37 | */ |
38 | ||
0e32dfc8 KKD |
39 | #include <linux/btf.h> |
40 | #include <linux/btf_ids.h> | |
e3118e83 DB |
41 | #include <linux/module.h> |
42 | #include <linux/mm.h> | |
43 | #include <net/tcp.h> | |
44 | #include <linux/inet_diag.h> | |
ffd177de | 45 | #include "tcp_dctcp.h" |
e3118e83 DB |
46 | |
47 | #define DCTCP_MAX_ALPHA 1024U | |
48 | ||
49 | struct dctcp { | |
e3058450 ED |
50 | u32 old_delivered; |
51 | u32 old_delivered_ce; | |
e3118e83 DB |
52 | u32 prior_rcv_nxt; |
53 | u32 dctcp_alpha; | |
54 | u32 next_seq; | |
55 | u32 ce_state; | |
ce6dd233 | 56 | u32 loss_cwnd; |
c30f8e0b | 57 | struct tcp_plb_state plb; |
e3118e83 DB |
58 | }; |
59 | ||
60 | static unsigned int dctcp_shift_g __read_mostly = 4; /* g = 1/2^4 */ | |
3ebc46ca KI |
61 | |
62 | static int dctcp_shift_g_set(const char *val, const struct kernel_param *kp) | |
63 | { | |
64 | return param_set_uint_minmax(val, kp, 0, 10); | |
65 | } | |
66 | ||
67 | static const struct kernel_param_ops dctcp_shift_g_ops = { | |
68 | .set = dctcp_shift_g_set, | |
69 | .get = param_get_uint, | |
70 | }; | |
71 | ||
72 | module_param_cb(dctcp_shift_g, &dctcp_shift_g_ops, &dctcp_shift_g, 0644); | |
e3118e83 DB |
73 | MODULE_PARM_DESC(dctcp_shift_g, "parameter g for updating dctcp_alpha"); |
74 | ||
75 | static unsigned int dctcp_alpha_on_init __read_mostly = DCTCP_MAX_ALPHA; | |
76 | module_param(dctcp_alpha_on_init, uint, 0644); | |
77 | MODULE_PARM_DESC(dctcp_alpha_on_init, "parameter for initial alpha value"); | |
78 | ||
e3118e83 DB |
79 | static struct tcp_congestion_ops dctcp_reno; |
80 | ||
81 | static void dctcp_reset(const struct tcp_sock *tp, struct dctcp *ca) | |
82 | { | |
83 | ca->next_seq = tp->snd_nxt; | |
84 | ||
e3058450 ED |
85 | ca->old_delivered = tp->delivered; |
86 | ca->old_delivered_ce = tp->delivered_ce; | |
e3118e83 DB |
87 | } |
88 | ||
400031e0 | 89 | __bpf_kfunc static void dctcp_init(struct sock *sk) |
e3118e83 DB |
90 | { |
91 | const struct tcp_sock *tp = tcp_sk(sk); | |
92 | ||
93 | if ((tp->ecn_flags & TCP_ECN_OK) || | |
94 | (sk->sk_state == TCP_LISTEN || | |
95 | sk->sk_state == TCP_CLOSE)) { | |
96 | struct dctcp *ca = inet_csk_ca(sk); | |
97 | ||
e3118e83 DB |
98 | ca->prior_rcv_nxt = tp->rcv_nxt; |
99 | ||
100 | ca->dctcp_alpha = min(dctcp_alpha_on_init, DCTCP_MAX_ALPHA); | |
101 | ||
ce6dd233 | 102 | ca->loss_cwnd = 0; |
e3118e83 DB |
103 | ca->ce_state = 0; |
104 | ||
105 | dctcp_reset(tp, ca); | |
c30f8e0b MAQ |
106 | tcp_plb_init(sk, &ca->plb); |
107 | ||
e3118e83 DB |
108 | return; |
109 | } | |
110 | ||
111 | /* No ECN support? Fall back to Reno. Also need to clear | |
112 | * ECT from sk since it is set during 3WHS for DCTCP. | |
113 | */ | |
114 | inet_csk(sk)->icsk_ca_ops = &dctcp_reno; | |
115 | INET_ECN_dontxmit(sk); | |
116 | } | |
117 | ||
400031e0 | 118 | __bpf_kfunc static u32 dctcp_ssthresh(struct sock *sk) |
e3118e83 | 119 | { |
ce6dd233 | 120 | struct dctcp *ca = inet_csk_ca(sk); |
e3118e83 DB |
121 | struct tcp_sock *tp = tcp_sk(sk); |
122 | ||
40570375 ED |
123 | ca->loss_cwnd = tcp_snd_cwnd(tp); |
124 | return max(tcp_snd_cwnd(tp) - ((tcp_snd_cwnd(tp) * ca->dctcp_alpha) >> 11U), 2U); | |
e3118e83 DB |
125 | } |
126 | ||
400031e0 | 127 | __bpf_kfunc static void dctcp_update_alpha(struct sock *sk, u32 flags) |
e3118e83 | 128 | { |
343dfaa1 | 129 | const struct tcp_sock *tp = tcp_sk(sk); |
e3118e83 | 130 | struct dctcp *ca = inet_csk_ca(sk); |
e3118e83 DB |
131 | |
132 | /* Expired RTT */ | |
133 | if (!before(tp->snd_una, ca->next_seq)) { | |
c30f8e0b | 134 | u32 delivered = tp->delivered - ca->old_delivered; |
e3058450 | 135 | u32 delivered_ce = tp->delivered_ce - ca->old_delivered_ce; |
f9c2ff22 | 136 | u32 alpha = ca->dctcp_alpha; |
c30f8e0b MAQ |
137 | u32 ce_ratio = 0; |
138 | ||
139 | if (delivered > 0) { | |
140 | /* dctcp_alpha keeps EWMA of fraction of ECN marked | |
141 | * packets. Because of EWMA smoothing, PLB reaction can | |
142 | * be slow so we use ce_ratio which is an instantaneous | |
143 | * measure of congestion. ce_ratio is the fraction of | |
144 | * ECN marked packets in the previous RTT. | |
145 | */ | |
146 | if (delivered_ce > 0) | |
147 | ce_ratio = (delivered_ce << TCP_PLB_SCALE) / delivered; | |
148 | tcp_plb_update_state(sk, &ca->plb, (int)ce_ratio); | |
149 | tcp_plb_check_rehash(sk, &ca->plb); | |
150 | } | |
e3118e83 DB |
151 | |
152 | /* alpha = (1 - g) * alpha + g * F */ | |
e3118e83 | 153 | |
c80dbe04 | 154 | alpha -= min_not_zero(alpha, alpha >> dctcp_shift_g); |
e3058450 | 155 | if (delivered_ce) { |
e3058450 | 156 | |
f9c2ff22 | 157 | /* If dctcp_shift_g == 1, a 32bit value would overflow |
e3058450 | 158 | * after 8 M packets. |
f9c2ff22 | 159 | */ |
e3058450 ED |
160 | delivered_ce <<= (10 - dctcp_shift_g); |
161 | delivered_ce /= max(1U, delivered); | |
f9c2ff22 | 162 | |
e3058450 | 163 | alpha = min(alpha + delivered_ce, DCTCP_MAX_ALPHA); |
f9c2ff22 ED |
164 | } |
165 | /* dctcp_alpha can be read from dctcp_get_info() without | |
166 | * synchro, so we ask compiler to not use dctcp_alpha | |
167 | * as a temporary variable in prior operations. | |
168 | */ | |
169 | WRITE_ONCE(ca->dctcp_alpha, alpha); | |
e3118e83 DB |
170 | dctcp_reset(tp, ca); |
171 | } | |
172 | } | |
173 | ||
aecfde23 | 174 | static void dctcp_react_to_loss(struct sock *sk) |
e3118e83 | 175 | { |
aecfde23 KDS |
176 | struct dctcp *ca = inet_csk_ca(sk); |
177 | struct tcp_sock *tp = tcp_sk(sk); | |
e3118e83 | 178 | |
40570375 ED |
179 | ca->loss_cwnd = tcp_snd_cwnd(tp); |
180 | tp->snd_ssthresh = max(tcp_snd_cwnd(tp) >> 1U, 2U); | |
aecfde23 KDS |
181 | } |
182 | ||
400031e0 | 183 | __bpf_kfunc static void dctcp_state(struct sock *sk, u8 new_state) |
aecfde23 KDS |
184 | { |
185 | if (new_state == TCP_CA_Recovery && | |
186 | new_state != inet_csk(sk)->icsk_ca_state) | |
187 | dctcp_react_to_loss(sk); | |
188 | /* We handle RTO in dctcp_cwnd_event to ensure that we perform only | |
189 | * one loss-adjustment per RTT. | |
190 | */ | |
e3118e83 DB |
191 | } |
192 | ||
400031e0 | 193 | __bpf_kfunc static void dctcp_cwnd_event(struct sock *sk, enum tcp_ca_event ev) |
e3118e83 | 194 | { |
ffd177de YC |
195 | struct dctcp *ca = inet_csk_ca(sk); |
196 | ||
e3118e83 DB |
197 | switch (ev) { |
198 | case CA_EVENT_ECN_IS_CE: | |
e3118e83 | 199 | case CA_EVENT_ECN_NO_CE: |
ffd177de | 200 | dctcp_ece_ack_update(sk, ev, &ca->prior_rcv_nxt, &ca->ce_state); |
e3118e83 | 201 | break; |
aecfde23 | 202 | case CA_EVENT_LOSS: |
c30f8e0b | 203 | tcp_plb_update_state_upon_rto(sk, &ca->plb); |
aecfde23 KDS |
204 | dctcp_react_to_loss(sk); |
205 | break; | |
c30f8e0b MAQ |
206 | case CA_EVENT_TX_START: |
207 | tcp_plb_check_rehash(sk, &ca->plb); /* Maybe rehash when inflight is 0 */ | |
208 | break; | |
e3118e83 DB |
209 | default: |
210 | /* Don't care for the rest. */ | |
211 | break; | |
212 | } | |
213 | } | |
214 | ||
64f40ff5 ED |
215 | static size_t dctcp_get_info(struct sock *sk, u32 ext, int *attr, |
216 | union tcp_cc_info *info) | |
e3118e83 DB |
217 | { |
218 | const struct dctcp *ca = inet_csk_ca(sk); | |
e3058450 | 219 | const struct tcp_sock *tp = tcp_sk(sk); |
e3118e83 DB |
220 | |
221 | /* Fill it also in case of VEGASINFO due to req struct limits. | |
222 | * We can still correctly retrieve it later. | |
223 | */ | |
224 | if (ext & (1 << (INET_DIAG_DCTCPINFO - 1)) || | |
225 | ext & (1 << (INET_DIAG_VEGASINFO - 1))) { | |
dcf1158b | 226 | memset(&info->dctcp, 0, sizeof(info->dctcp)); |
e3118e83 | 227 | if (inet_csk(sk)->icsk_ca_ops != &dctcp_reno) { |
64f40ff5 ED |
228 | info->dctcp.dctcp_enabled = 1; |
229 | info->dctcp.dctcp_ce_state = (u16) ca->ce_state; | |
230 | info->dctcp.dctcp_alpha = ca->dctcp_alpha; | |
e3058450 ED |
231 | info->dctcp.dctcp_ab_ecn = tp->mss_cache * |
232 | (tp->delivered_ce - ca->old_delivered_ce); | |
233 | info->dctcp.dctcp_ab_tot = tp->mss_cache * | |
234 | (tp->delivered - ca->old_delivered); | |
e3118e83 DB |
235 | } |
236 | ||
64f40ff5 | 237 | *attr = INET_DIAG_DCTCPINFO; |
dcf1158b | 238 | return sizeof(info->dctcp); |
e3118e83 | 239 | } |
521f1cf1 | 240 | return 0; |
e3118e83 DB |
241 | } |
242 | ||
400031e0 | 243 | __bpf_kfunc static u32 dctcp_cwnd_undo(struct sock *sk) |
ce6dd233 FW |
244 | { |
245 | const struct dctcp *ca = inet_csk_ca(sk); | |
40570375 | 246 | struct tcp_sock *tp = tcp_sk(sk); |
ce6dd233 | 247 | |
40570375 | 248 | return max(tcp_snd_cwnd(tp), ca->loss_cwnd); |
ce6dd233 FW |
249 | } |
250 | ||
e3118e83 DB |
251 | static struct tcp_congestion_ops dctcp __read_mostly = { |
252 | .init = dctcp_init, | |
253 | .in_ack_event = dctcp_update_alpha, | |
254 | .cwnd_event = dctcp_cwnd_event, | |
255 | .ssthresh = dctcp_ssthresh, | |
256 | .cong_avoid = tcp_reno_cong_avoid, | |
ce6dd233 | 257 | .undo_cwnd = dctcp_cwnd_undo, |
e3118e83 DB |
258 | .set_state = dctcp_state, |
259 | .get_info = dctcp_get_info, | |
260 | .flags = TCP_CONG_NEEDS_ECN, | |
261 | .owner = THIS_MODULE, | |
262 | .name = "dctcp", | |
263 | }; | |
264 | ||
265 | static struct tcp_congestion_ops dctcp_reno __read_mostly = { | |
266 | .ssthresh = tcp_reno_ssthresh, | |
267 | .cong_avoid = tcp_reno_cong_avoid, | |
e9799183 | 268 | .undo_cwnd = tcp_reno_undo_cwnd, |
e3118e83 DB |
269 | .get_info = dctcp_get_info, |
270 | .owner = THIS_MODULE, | |
271 | .name = "dctcp-reno", | |
272 | }; | |
273 | ||
6f3189f3 | 274 | BTF_KFUNCS_START(tcp_dctcp_check_kfunc_ids) |
a4703e31 KKD |
275 | BTF_ID_FLAGS(func, dctcp_init) |
276 | BTF_ID_FLAGS(func, dctcp_update_alpha) | |
277 | BTF_ID_FLAGS(func, dctcp_cwnd_event) | |
278 | BTF_ID_FLAGS(func, dctcp_ssthresh) | |
279 | BTF_ID_FLAGS(func, dctcp_cwnd_undo) | |
280 | BTF_ID_FLAGS(func, dctcp_state) | |
6f3189f3 | 281 | BTF_KFUNCS_END(tcp_dctcp_check_kfunc_ids) |
0e32dfc8 | 282 | |
b202d844 | 283 | static const struct btf_kfunc_id_set tcp_dctcp_kfunc_set = { |
a4703e31 KKD |
284 | .owner = THIS_MODULE, |
285 | .set = &tcp_dctcp_check_kfunc_ids, | |
b202d844 | 286 | }; |
0e32dfc8 | 287 | |
e3118e83 DB |
288 | static int __init dctcp_register(void) |
289 | { | |
0e32dfc8 KKD |
290 | int ret; |
291 | ||
e3118e83 | 292 | BUILD_BUG_ON(sizeof(struct dctcp) > ICSK_CA_PRIV_SIZE); |
b202d844 KKD |
293 | |
294 | ret = register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &tcp_dctcp_kfunc_set); | |
295 | if (ret < 0) | |
0e32dfc8 | 296 | return ret; |
b202d844 | 297 | return tcp_register_congestion_control(&dctcp); |
e3118e83 DB |
298 | } |
299 | ||
300 | static void __exit dctcp_unregister(void) | |
301 | { | |
302 | tcp_unregister_congestion_control(&dctcp); | |
303 | } | |
304 | ||
305 | module_init(dctcp_register); | |
306 | module_exit(dctcp_unregister); | |
307 | ||
308 | MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>"); | |
309 | MODULE_AUTHOR("Florian Westphal <fw@strlen.de>"); | |
310 | MODULE_AUTHOR("Glenn Judd <glenn.judd@morganstanley.com>"); | |
311 | ||
312 | MODULE_LICENSE("GPL v2"); | |
313 | MODULE_DESCRIPTION("DataCenter TCP (DCTCP)"); |