diff options
author | Arnaldo Carvalho de Melo <acme@mandriva.com> | 2005-11-05 19:30:29 -0500 |
---|---|---|
committer | Arnaldo Carvalho de Melo <acme@mandriva.com> | 2005-11-05 19:30:29 -0500 |
commit | 2d43f1128a4282fbe8442f40b4cbbac05d8f10aa (patch) | |
tree | 788e5a970c3efb090b73cef0de32aae25444b734 /include/net | |
parent | 6df716340da3a6fdd33d73d7ed4c6f7590ca1c42 (diff) | |
parent | bdc450a0bb1d48144ced1f899cc8366ec8e85024 (diff) |
Merge branch 'red' of 84.73.165.173:/home/tgr/repos/net-2.6
Diffstat (limited to 'include/net')
-rw-r--r-- | include/net/inet_ecn.h | 28 | ||||
-rw-r--r-- | include/net/red.h | 325 |
2 files changed, 349 insertions, 4 deletions
diff --git a/include/net/inet_ecn.h b/include/net/inet_ecn.h index f87845e2e965..b0c47e2eccf1 100644 --- a/include/net/inet_ecn.h +++ b/include/net/inet_ecn.h | |||
@@ -2,6 +2,7 @@ | |||
2 | #define _INET_ECN_H_ | 2 | #define _INET_ECN_H_ |
3 | 3 | ||
4 | #include <linux/ip.h> | 4 | #include <linux/ip.h> |
5 | #include <linux/skbuff.h> | ||
5 | #include <net/dsfield.h> | 6 | #include <net/dsfield.h> |
6 | 7 | ||
7 | enum { | 8 | enum { |
@@ -48,7 +49,7 @@ static inline __u8 INET_ECN_encapsulate(__u8 outer, __u8 inner) | |||
48 | (label) |= __constant_htons(INET_ECN_ECT_0 << 4); \ | 49 | (label) |= __constant_htons(INET_ECN_ECT_0 << 4); \ |
49 | } while (0) | 50 | } while (0) |
50 | 51 | ||
51 | static inline void IP_ECN_set_ce(struct iphdr *iph) | 52 | static inline int IP_ECN_set_ce(struct iphdr *iph) |
52 | { | 53 | { |
53 | u32 check = iph->check; | 54 | u32 check = iph->check; |
54 | u32 ecn = (iph->tos + 1) & INET_ECN_MASK; | 55 | u32 ecn = (iph->tos + 1) & INET_ECN_MASK; |
@@ -61,7 +62,7 @@ static inline void IP_ECN_set_ce(struct iphdr *iph) | |||
61 | * INET_ECN_CE => 00 | 62 | * INET_ECN_CE => 00 |
62 | */ | 63 | */ |
63 | if (!(ecn & 2)) | 64 | if (!(ecn & 2)) |
64 | return; | 65 | return !ecn; |
65 | 66 | ||
66 | /* | 67 | /* |
67 | * The following gives us: | 68 | * The following gives us: |
@@ -72,6 +73,7 @@ static inline void IP_ECN_set_ce(struct iphdr *iph) | |||
72 | 73 | ||
73 | iph->check = check + (check>=0xFFFF); | 74 | iph->check = check + (check>=0xFFFF); |
74 | iph->tos |= INET_ECN_CE; | 75 | iph->tos |= INET_ECN_CE; |
76 | return 1; | ||
75 | } | 77 | } |
76 | 78 | ||
77 | static inline void IP_ECN_clear(struct iphdr *iph) | 79 | static inline void IP_ECN_clear(struct iphdr *iph) |
@@ -87,11 +89,12 @@ static inline void ipv4_copy_dscp(struct iphdr *outer, struct iphdr *inner) | |||
87 | 89 | ||
88 | struct ipv6hdr; | 90 | struct ipv6hdr; |
89 | 91 | ||
90 | static inline void IP6_ECN_set_ce(struct ipv6hdr *iph) | 92 | static inline int IP6_ECN_set_ce(struct ipv6hdr *iph) |
91 | { | 93 | { |
92 | if (INET_ECN_is_not_ect(ipv6_get_dsfield(iph))) | 94 | if (INET_ECN_is_not_ect(ipv6_get_dsfield(iph))) |
93 | return; | 95 | return 0; |
94 | *(u32*)iph |= htonl(INET_ECN_CE << 20); | 96 | *(u32*)iph |= htonl(INET_ECN_CE << 20); |
97 | return 1; | ||
95 | } | 98 | } |
96 | 99 | ||
97 | static inline void IP6_ECN_clear(struct ipv6hdr *iph) | 100 | static inline void IP6_ECN_clear(struct ipv6hdr *iph) |
@@ -105,4 +108,21 @@ static inline void ipv6_copy_dscp(struct ipv6hdr *outer, struct ipv6hdr *inner) | |||
105 | ipv6_change_dsfield(inner, INET_ECN_MASK, dscp); | 108 | ipv6_change_dsfield(inner, INET_ECN_MASK, dscp); |
106 | } | 109 | } |
107 | 110 | ||
111 | static inline int INET_ECN_set_ce(struct sk_buff *skb) | ||
112 | { | ||
113 | switch (skb->protocol) { | ||
114 | case __constant_htons(ETH_P_IP): | ||
115 | if (skb->nh.raw + sizeof(struct iphdr) <= skb->tail) | ||
116 | return IP_ECN_set_ce(skb->nh.iph); | ||
117 | break; | ||
118 | |||
119 | case __constant_htons(ETH_P_IPV6): | ||
120 | if (skb->nh.raw + sizeof(struct ipv6hdr) <= skb->tail) | ||
121 | return IP6_ECN_set_ce(skb->nh.ipv6h); | ||
122 | break; | ||
123 | } | ||
124 | |||
125 | return 0; | ||
126 | } | ||
127 | |||
108 | #endif | 128 | #endif |
diff --git a/include/net/red.h b/include/net/red.h new file mode 100644 index 000000000000..2ed4358e3295 --- /dev/null +++ b/include/net/red.h | |||
@@ -0,0 +1,325 @@ | |||
1 | #ifndef __NET_SCHED_RED_H | ||
2 | #define __NET_SCHED_RED_H | ||
3 | |||
4 | #include <linux/config.h> | ||
5 | #include <linux/types.h> | ||
6 | #include <net/pkt_sched.h> | ||
7 | #include <net/inet_ecn.h> | ||
8 | #include <net/dsfield.h> | ||
9 | |||
10 | /* Random Early Detection (RED) algorithm. | ||
11 | ======================================= | ||
12 | |||
13 | Source: Sally Floyd and Van Jacobson, "Random Early Detection Gateways | ||
14 | for Congestion Avoidance", 1993, IEEE/ACM Transactions on Networking. | ||
15 | |||
16 | This file codes a "divisionless" version of RED algorithm | ||
17 | as written down in Fig.17 of the paper. | ||
18 | |||
19 | Short description. | ||
20 | ------------------ | ||
21 | |||
22 | When a new packet arrives we calculate the average queue length: | ||
23 | |||
24 | avg = (1-W)*avg + W*current_queue_len, | ||
25 | |||
26 | W is the filter time constant (chosen as 2^(-Wlog)), it controls | ||
27 | the inertia of the algorithm. To allow larger bursts, W should be | ||
28 | decreased. | ||
29 | |||
30 | if (avg > th_max) -> packet marked (dropped). | ||
31 | if (avg < th_min) -> packet passes. | ||
32 | if (th_min < avg < th_max) we calculate probability: | ||
33 | |||
34 | Pb = max_P * (avg - th_min)/(th_max-th_min) | ||
35 | |||
36 | and mark (drop) packet with this probability. | ||
37 | Pb changes from 0 (at avg==th_min) to max_P (avg==th_max). | ||
38 | max_P should be small (not 1), usually 0.01..0.02 is good value. | ||
39 | |||
40 | max_P is chosen as a number, so that max_P/(th_max-th_min) | ||
41 | is a negative power of two in order arithmetics to contain | ||
42 | only shifts. | ||
43 | |||
44 | |||
45 | Parameters, settable by user: | ||
46 | ----------------------------- | ||
47 | |||
48 | qth_min - bytes (should be < qth_max/2) | ||
49 | qth_max - bytes (should be at least 2*qth_min and less limit) | ||
50 | Wlog - bits (<32) log(1/W). | ||
51 | Plog - bits (<32) | ||
52 | |||
53 | Plog is related to max_P by formula: | ||
54 | |||
55 | max_P = (qth_max-qth_min)/2^Plog; | ||
56 | |||
57 | F.e. if qth_max=128K and qth_min=32K, then Plog=22 | ||
58 | corresponds to max_P=0.02 | ||
59 | |||
60 | Scell_log | ||
61 | Stab | ||
62 | |||
63 | Lookup table for log((1-W)^(t/t_ave). | ||
64 | |||
65 | |||
66 | NOTES: | ||
67 | |||
68 | Upper bound on W. | ||
69 | ----------------- | ||
70 | |||
71 | If you want to allow bursts of L packets of size S, | ||
72 | you should choose W: | ||
73 | |||
74 | L + 1 - th_min/S < (1-(1-W)^L)/W | ||
75 | |||
76 | th_min/S = 32 th_min/S = 4 | ||
77 | |||
78 | log(W) L | ||
79 | -1 33 | ||
80 | -2 35 | ||
81 | -3 39 | ||
82 | -4 46 | ||
83 | -5 57 | ||
84 | -6 75 | ||
85 | -7 101 | ||
86 | -8 135 | ||
87 | -9 190 | ||
88 | etc. | ||
89 | */ | ||
90 | |||
91 | #define RED_STAB_SIZE 256 | ||
92 | #define RED_STAB_MASK (RED_STAB_SIZE - 1) | ||
93 | |||
94 | struct red_stats | ||
95 | { | ||
96 | u32 prob_drop; /* Early probability drops */ | ||
97 | u32 prob_mark; /* Early probability marks */ | ||
98 | u32 forced_drop; /* Forced drops, qavg > max_thresh */ | ||
99 | u32 forced_mark; /* Forced marks, qavg > max_thresh */ | ||
100 | u32 pdrop; /* Drops due to queue limits */ | ||
101 | u32 other; /* Drops due to drop() calls */ | ||
102 | u32 backlog; | ||
103 | }; | ||
104 | |||
105 | struct red_parms | ||
106 | { | ||
107 | /* Parameters */ | ||
108 | u32 qth_min; /* Min avg length threshold: A scaled */ | ||
109 | u32 qth_max; /* Max avg length threshold: A scaled */ | ||
110 | u32 Scell_max; | ||
111 | u32 Rmask; /* Cached random mask, see red_rmask */ | ||
112 | u8 Scell_log; | ||
113 | u8 Wlog; /* log(W) */ | ||
114 | u8 Plog; /* random number bits */ | ||
115 | u8 Stab[RED_STAB_SIZE]; | ||
116 | |||
117 | /* Variables */ | ||
118 | int qcount; /* Number of packets since last random | ||
119 | number generation */ | ||
120 | u32 qR; /* Cached random number */ | ||
121 | |||
122 | unsigned long qavg; /* Average queue length: A scaled */ | ||
123 | psched_time_t qidlestart; /* Start of current idle period */ | ||
124 | }; | ||
125 | |||
126 | static inline u32 red_rmask(u8 Plog) | ||
127 | { | ||
128 | return Plog < 32 ? ((1 << Plog) - 1) : ~0UL; | ||
129 | } | ||
130 | |||
131 | static inline void red_set_parms(struct red_parms *p, | ||
132 | u32 qth_min, u32 qth_max, u8 Wlog, u8 Plog, | ||
133 | u8 Scell_log, u8 *stab) | ||
134 | { | ||
135 | /* Reset average queue length, the value is strictly bound | ||
136 | * to the parameters below, reseting hurts a bit but leaving | ||
137 | * it might result in an unreasonable qavg for a while. --TGR | ||
138 | */ | ||
139 | p->qavg = 0; | ||
140 | |||
141 | p->qcount = -1; | ||
142 | p->qth_min = qth_min << Wlog; | ||
143 | p->qth_max = qth_max << Wlog; | ||
144 | p->Wlog = Wlog; | ||
145 | p->Plog = Plog; | ||
146 | p->Rmask = red_rmask(Plog); | ||
147 | p->Scell_log = Scell_log; | ||
148 | p->Scell_max = (255 << Scell_log); | ||
149 | |||
150 | memcpy(p->Stab, stab, sizeof(p->Stab)); | ||
151 | } | ||
152 | |||
153 | static inline int red_is_idling(struct red_parms *p) | ||
154 | { | ||
155 | return !PSCHED_IS_PASTPERFECT(p->qidlestart); | ||
156 | } | ||
157 | |||
158 | static inline void red_start_of_idle_period(struct red_parms *p) | ||
159 | { | ||
160 | PSCHED_GET_TIME(p->qidlestart); | ||
161 | } | ||
162 | |||
163 | static inline void red_end_of_idle_period(struct red_parms *p) | ||
164 | { | ||
165 | PSCHED_SET_PASTPERFECT(p->qidlestart); | ||
166 | } | ||
167 | |||
168 | static inline void red_restart(struct red_parms *p) | ||
169 | { | ||
170 | red_end_of_idle_period(p); | ||
171 | p->qavg = 0; | ||
172 | p->qcount = -1; | ||
173 | } | ||
174 | |||
175 | static inline unsigned long red_calc_qavg_from_idle_time(struct red_parms *p) | ||
176 | { | ||
177 | psched_time_t now; | ||
178 | long us_idle; | ||
179 | int shift; | ||
180 | |||
181 | PSCHED_GET_TIME(now); | ||
182 | us_idle = PSCHED_TDIFF_SAFE(now, p->qidlestart, p->Scell_max); | ||
183 | |||
184 | /* | ||
185 | * The problem: ideally, average length queue recalcultion should | ||
186 | * be done over constant clock intervals. This is too expensive, so | ||
187 | * that the calculation is driven by outgoing packets. | ||
188 | * When the queue is idle we have to model this clock by hand. | ||
189 | * | ||
190 | * SF+VJ proposed to "generate": | ||
191 | * | ||
192 | * m = idletime / (average_pkt_size / bandwidth) | ||
193 | * | ||
194 | * dummy packets as a burst after idle time, i.e. | ||
195 | * | ||
196 | * p->qavg *= (1-W)^m | ||
197 | * | ||
198 | * This is an apparently overcomplicated solution (f.e. we have to | ||
199 | * precompute a table to make this calculation in reasonable time) | ||
200 | * I believe that a simpler model may be used here, | ||
201 | * but it is field for experiments. | ||
202 | */ | ||
203 | |||
204 | shift = p->Stab[(us_idle >> p->Scell_log) & RED_STAB_MASK]; | ||
205 | |||
206 | if (shift) | ||
207 | return p->qavg >> shift; | ||
208 | else { | ||
209 | /* Approximate initial part of exponent with linear function: | ||
210 | * | ||
211 | * (1-W)^m ~= 1-mW + ... | ||
212 | * | ||
213 | * Seems, it is the best solution to | ||
214 | * problem of too coarse exponent tabulation. | ||
215 | */ | ||
216 | us_idle = (p->qavg * us_idle) >> p->Scell_log; | ||
217 | |||
218 | if (us_idle < (p->qavg >> 1)) | ||
219 | return p->qavg - us_idle; | ||
220 | else | ||
221 | return p->qavg >> 1; | ||
222 | } | ||
223 | } | ||
224 | |||
225 | static inline unsigned long red_calc_qavg_no_idle_time(struct red_parms *p, | ||
226 | unsigned int backlog) | ||
227 | { | ||
228 | /* | ||
229 | * NOTE: p->qavg is fixed point number with point at Wlog. | ||
230 | * The formula below is equvalent to floating point | ||
231 | * version: | ||
232 | * | ||
233 | * qavg = qavg*(1-W) + backlog*W; | ||
234 | * | ||
235 | * --ANK (980924) | ||
236 | */ | ||
237 | return p->qavg + (backlog - (p->qavg >> p->Wlog)); | ||
238 | } | ||
239 | |||
240 | static inline unsigned long red_calc_qavg(struct red_parms *p, | ||
241 | unsigned int backlog) | ||
242 | { | ||
243 | if (!red_is_idling(p)) | ||
244 | return red_calc_qavg_no_idle_time(p, backlog); | ||
245 | else | ||
246 | return red_calc_qavg_from_idle_time(p); | ||
247 | } | ||
248 | |||
249 | static inline u32 red_random(struct red_parms *p) | ||
250 | { | ||
251 | return net_random() & p->Rmask; | ||
252 | } | ||
253 | |||
254 | static inline int red_mark_probability(struct red_parms *p, unsigned long qavg) | ||
255 | { | ||
256 | /* The formula used below causes questions. | ||
257 | |||
258 | OK. qR is random number in the interval 0..Rmask | ||
259 | i.e. 0..(2^Plog). If we used floating point | ||
260 | arithmetics, it would be: (2^Plog)*rnd_num, | ||
261 | where rnd_num is less 1. | ||
262 | |||
263 | Taking into account, that qavg have fixed | ||
264 | point at Wlog, and Plog is related to max_P by | ||
265 | max_P = (qth_max-qth_min)/2^Plog; two lines | ||
266 | below have the following floating point equivalent: | ||
267 | |||
268 | max_P*(qavg - qth_min)/(qth_max-qth_min) < rnd/qcount | ||
269 | |||
270 | Any questions? --ANK (980924) | ||
271 | */ | ||
272 | return !(((qavg - p->qth_min) >> p->Wlog) * p->qcount < p->qR); | ||
273 | } | ||
274 | |||
275 | enum { | ||
276 | RED_BELOW_MIN_THRESH, | ||
277 | RED_BETWEEN_TRESH, | ||
278 | RED_ABOVE_MAX_TRESH, | ||
279 | }; | ||
280 | |||
281 | static inline int red_cmp_thresh(struct red_parms *p, unsigned long qavg) | ||
282 | { | ||
283 | if (qavg < p->qth_min) | ||
284 | return RED_BELOW_MIN_THRESH; | ||
285 | else if (qavg >= p->qth_max) | ||
286 | return RED_ABOVE_MAX_TRESH; | ||
287 | else | ||
288 | return RED_BETWEEN_TRESH; | ||
289 | } | ||
290 | |||
291 | enum { | ||
292 | RED_DONT_MARK, | ||
293 | RED_PROB_MARK, | ||
294 | RED_HARD_MARK, | ||
295 | }; | ||
296 | |||
297 | static inline int red_action(struct red_parms *p, unsigned long qavg) | ||
298 | { | ||
299 | switch (red_cmp_thresh(p, qavg)) { | ||
300 | case RED_BELOW_MIN_THRESH: | ||
301 | p->qcount = -1; | ||
302 | return RED_DONT_MARK; | ||
303 | |||
304 | case RED_BETWEEN_TRESH: | ||
305 | if (++p->qcount) { | ||
306 | if (red_mark_probability(p, qavg)) { | ||
307 | p->qcount = 0; | ||
308 | p->qR = red_random(p); | ||
309 | return RED_PROB_MARK; | ||
310 | } | ||
311 | } else | ||
312 | p->qR = red_random(p); | ||
313 | |||
314 | return RED_DONT_MARK; | ||
315 | |||
316 | case RED_ABOVE_MAX_TRESH: | ||
317 | p->qcount = -1; | ||
318 | return RED_HARD_MARK; | ||
319 | } | ||
320 | |||
321 | BUG(); | ||
322 | return RED_DONT_MARK; | ||
323 | } | ||
324 | |||
325 | #endif | ||