aboutsummaryrefslogtreecommitdiffstats
path: root/net
diff options
context:
space:
mode:
Diffstat (limited to 'net')
-rw-r--r--net/ipv4/tcp_vegas.c80
1 files changed, 10 insertions, 70 deletions
diff --git a/net/ipv4/tcp_vegas.c b/net/ipv4/tcp_vegas.c
index 7cd22262de3a..a453aac91bd3 100644
--- a/net/ipv4/tcp_vegas.c
+++ b/net/ipv4/tcp_vegas.c
@@ -40,18 +40,14 @@
40 40
41#include "tcp_vegas.h" 41#include "tcp_vegas.h"
42 42
43/* Default values of the Vegas variables, in fixed-point representation 43static int alpha = 2;
44 * with V_PARAM_SHIFT bits to the right of the binary point. 44static int beta = 4;
45 */ 45static int gamma = 1;
46#define V_PARAM_SHIFT 1
47static int alpha = 2<<V_PARAM_SHIFT;
48static int beta = 4<<V_PARAM_SHIFT;
49static int gamma = 1<<V_PARAM_SHIFT;
50 46
51module_param(alpha, int, 0644); 47module_param(alpha, int, 0644);
52MODULE_PARM_DESC(alpha, "lower bound of packets in network (scale by 2)"); 48MODULE_PARM_DESC(alpha, "lower bound of packets in network");
53module_param(beta, int, 0644); 49module_param(beta, int, 0644);
54MODULE_PARM_DESC(beta, "upper bound of packets in network (scale by 2)"); 50MODULE_PARM_DESC(beta, "upper bound of packets in network");
55module_param(gamma, int, 0644); 51module_param(gamma, int, 0644);
56MODULE_PARM_DESC(gamma, "limit on increase (scale by 2)"); 52MODULE_PARM_DESC(gamma, "limit on increase (scale by 2)");
57 53
@@ -172,49 +168,13 @@ static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
172 return; 168 return;
173 } 169 }
174 170
175 /* The key players are v_beg_snd_una and v_beg_snd_nxt.
176 *
177 * These are so named because they represent the approximate values
178 * of snd_una and snd_nxt at the beginning of the current RTT. More
179 * precisely, they represent the amount of data sent during the RTT.
180 * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
181 * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding
182 * bytes of data have been ACKed during the course of the RTT, giving
183 * an "actual" rate of:
184 *
185 * (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration)
186 *
187 * Unfortunately, v_beg_snd_una is not exactly equal to snd_una,
188 * because delayed ACKs can cover more than one segment, so they
189 * don't line up nicely with the boundaries of RTTs.
190 *
191 * Another unfortunate fact of life is that delayed ACKs delay the
192 * advance of the left edge of our send window, so that the number
193 * of bytes we send in an RTT is often less than our cwnd will allow.
194 * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
195 */
196
197 if (after(ack, vegas->beg_snd_nxt)) { 171 if (after(ack, vegas->beg_snd_nxt)) {
198 /* Do the Vegas once-per-RTT cwnd adjustment. */ 172 /* Do the Vegas once-per-RTT cwnd adjustment. */
199 u32 old_wnd, old_snd_cwnd;
200
201
202 /* Here old_wnd is essentially the window of data that was
203 * sent during the previous RTT, and has all
204 * been acknowledged in the course of the RTT that ended
205 * with the ACK we just received. Likewise, old_snd_cwnd
206 * is the cwnd during the previous RTT.
207 */
208 old_wnd = (vegas->beg_snd_nxt - vegas->beg_snd_una) /
209 tp->mss_cache;
210 old_snd_cwnd = vegas->beg_snd_cwnd;
211 173
212 /* Save the extent of the current window so we can use this 174 /* Save the extent of the current window so we can use this
213 * at the end of the next RTT. 175 * at the end of the next RTT.
214 */ 176 */
215 vegas->beg_snd_una = vegas->beg_snd_nxt;
216 vegas->beg_snd_nxt = tp->snd_nxt; 177 vegas->beg_snd_nxt = tp->snd_nxt;
217 vegas->beg_snd_cwnd = tp->snd_cwnd;
218 178
219 /* We do the Vegas calculations only if we got enough RTT 179 /* We do the Vegas calculations only if we got enough RTT
220 * samples that we can be reasonably sure that we got 180 * samples that we can be reasonably sure that we got
@@ -252,22 +212,14 @@ static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
252 * 212 *
253 * This is: 213 * This is:
254 * (actual rate in segments) * baseRTT 214 * (actual rate in segments) * baseRTT
255 * We keep it as a fixed point number with
256 * V_PARAM_SHIFT bits to the right of the binary point.
257 */ 215 */
258 target_cwnd = ((u64)old_wnd * vegas->baseRTT); 216 target_cwnd = tp->snd_cwnd * vegas->baseRTT / rtt;
259 target_cwnd <<= V_PARAM_SHIFT;
260 do_div(target_cwnd, rtt);
261 217
262 /* Calculate the difference between the window we had, 218 /* Calculate the difference between the window we had,
263 * and the window we would like to have. This quantity 219 * and the window we would like to have. This quantity
264 * is the "Diff" from the Arizona Vegas papers. 220 * is the "Diff" from the Arizona Vegas papers.
265 *
266 * Again, this is a fixed point number with
267 * V_PARAM_SHIFT bits to the right of the binary
268 * point.
269 */ 221 */
270 diff = (old_wnd << V_PARAM_SHIFT) - target_cwnd; 222 diff = tp->snd_cwnd * (rtt-vegas->baseRTT) / vegas->baseRTT;
271 223
272 if (diff > gamma && tp->snd_ssthresh > 2 ) { 224 if (diff > gamma && tp->snd_ssthresh > 2 ) {
273 /* Going too fast. Time to slow down 225 /* Going too fast. Time to slow down
@@ -282,16 +234,13 @@ static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
282 * truncation robs us of full link 234 * truncation robs us of full link
283 * utilization. 235 * utilization.
284 */ 236 */
285 tp->snd_cwnd = min(tp->snd_cwnd, 237 tp->snd_cwnd = min(tp->snd_cwnd, (u32)target_cwnd+1);
286 ((u32)target_cwnd >>
287 V_PARAM_SHIFT)+1);
288 238
289 } else if (tp->snd_cwnd <= tp->snd_ssthresh) { 239 } else if (tp->snd_cwnd <= tp->snd_ssthresh) {
290 /* Slow start. */ 240 /* Slow start. */
291 tcp_slow_start(tp); 241 tcp_slow_start(tp);
292 } else { 242 } else {
293 /* Congestion avoidance. */ 243 /* Congestion avoidance. */
294 u32 next_snd_cwnd;
295 244
296 /* Figure out where we would like cwnd 245 /* Figure out where we would like cwnd
297 * to be. 246 * to be.
@@ -300,26 +249,17 @@ static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
300 /* The old window was too fast, so 249 /* The old window was too fast, so
301 * we slow down. 250 * we slow down.
302 */ 251 */
303 next_snd_cwnd = old_snd_cwnd - 1; 252 tp->snd_cwnd--;
304 } else if (diff < alpha) { 253 } else if (diff < alpha) {
305 /* We don't have enough extra packets 254 /* We don't have enough extra packets
306 * in the network, so speed up. 255 * in the network, so speed up.
307 */ 256 */
308 next_snd_cwnd = old_snd_cwnd + 1; 257 tp->snd_cwnd++;
309 } else { 258 } else {
310 /* Sending just as fast as we 259 /* Sending just as fast as we
311 * should be. 260 * should be.
312 */ 261 */
313 next_snd_cwnd = old_snd_cwnd;
314 } 262 }
315
316 /* Adjust cwnd upward or downward, toward the
317 * desired value.
318 */
319 if (next_snd_cwnd > tp->snd_cwnd)
320 tp->snd_cwnd++;
321 else if (next_snd_cwnd < tp->snd_cwnd)
322 tp->snd_cwnd--;
323 } 263 }
324 264
325 if (tp->snd_cwnd < 2) 265 if (tp->snd_cwnd < 2)