1 From: Felix Fietkau <nbd@nbd.name>
2 Date: Sat, 10 Feb 2018 12:45:47 +0100
3 Subject: [PATCH] mac80211: minstrel: reduce minstrel_mcs_groups size
5 By storing a shift value for all duration values of a group, we can
6 reduce precision by a neglegible amount to make it fit into a u16 value.
7 This improves cache footprint and reduces size:
10 text data bss dec hex filename
11 10024 116 0 10140 279c rc80211_minstrel_ht.o
14 text data bss dec hex filename
15 9368 116 0 9484 250c rc80211_minstrel_ht.o
17 Signed-off-by: Felix Fietkau <nbd@nbd.name>
20 --- a/net/mac80211/rc80211_minstrel_ht.c
21 +++ b/net/mac80211/rc80211_minstrel_ht.c
25 /* MCS rate information for an MCS group */
26 -#define MCS_GROUP(_streams, _sgi, _ht40) \
27 +#define MCS_GROUP(_streams, _sgi, _ht40, _s) \
28 [GROUP_IDX(_streams, _sgi, _ht40)] = { \
29 .streams = _streams, \
32 IEEE80211_TX_RC_MCS | \
33 (_sgi ? IEEE80211_TX_RC_SHORT_GI : 0) | \
34 (_ht40 ? IEEE80211_TX_RC_40_MHZ_WIDTH : 0), \
36 - MCS_DURATION(_streams, _sgi, _ht40 ? 54 : 26), \
37 - MCS_DURATION(_streams, _sgi, _ht40 ? 108 : 52), \
38 - MCS_DURATION(_streams, _sgi, _ht40 ? 162 : 78), \
39 - MCS_DURATION(_streams, _sgi, _ht40 ? 216 : 104), \
40 - MCS_DURATION(_streams, _sgi, _ht40 ? 324 : 156), \
41 - MCS_DURATION(_streams, _sgi, _ht40 ? 432 : 208), \
42 - MCS_DURATION(_streams, _sgi, _ht40 ? 486 : 234), \
43 - MCS_DURATION(_streams, _sgi, _ht40 ? 540 : 260) \
44 + MCS_DURATION(_streams, _sgi, _ht40 ? 54 : 26) >> _s, \
45 + MCS_DURATION(_streams, _sgi, _ht40 ? 108 : 52) >> _s, \
46 + MCS_DURATION(_streams, _sgi, _ht40 ? 162 : 78) >> _s, \
47 + MCS_DURATION(_streams, _sgi, _ht40 ? 216 : 104) >> _s, \
48 + MCS_DURATION(_streams, _sgi, _ht40 ? 324 : 156) >> _s, \
49 + MCS_DURATION(_streams, _sgi, _ht40 ? 432 : 208) >> _s, \
50 + MCS_DURATION(_streams, _sgi, _ht40 ? 486 : 234) >> _s, \
51 + MCS_DURATION(_streams, _sgi, _ht40 ? 540 : 260) >> _s \
56 #define BW2VBPS(_bw, r3, r2, r1) \
57 (_bw == BW_80 ? r3 : _bw == BW_40 ? r2 : r1)
59 -#define VHT_GROUP(_streams, _sgi, _bw) \
60 +#define VHT_GROUP(_streams, _sgi, _bw, _s) \
61 [VHT_GROUP_IDX(_streams, _sgi, _bw)] = { \
62 .streams = _streams, \
65 IEEE80211_TX_RC_VHT_MCS | \
66 (_sgi ? IEEE80211_TX_RC_SHORT_GI : 0) | \
68 _bw == BW_40 ? IEEE80211_TX_RC_40_MHZ_WIDTH : 0), \
70 MCS_DURATION(_streams, _sgi, \
71 - BW2VBPS(_bw, 117, 54, 26)), \
72 + BW2VBPS(_bw, 117, 54, 26)) >> _s, \
73 MCS_DURATION(_streams, _sgi, \
74 - BW2VBPS(_bw, 234, 108, 52)), \
75 + BW2VBPS(_bw, 234, 108, 52)) >> _s, \
76 MCS_DURATION(_streams, _sgi, \
77 - BW2VBPS(_bw, 351, 162, 78)), \
78 + BW2VBPS(_bw, 351, 162, 78)) >> _s, \
79 MCS_DURATION(_streams, _sgi, \
80 - BW2VBPS(_bw, 468, 216, 104)), \
81 + BW2VBPS(_bw, 468, 216, 104)) >> _s, \
82 MCS_DURATION(_streams, _sgi, \
83 - BW2VBPS(_bw, 702, 324, 156)), \
84 + BW2VBPS(_bw, 702, 324, 156)) >> _s, \
85 MCS_DURATION(_streams, _sgi, \
86 - BW2VBPS(_bw, 936, 432, 208)), \
87 + BW2VBPS(_bw, 936, 432, 208)) >> _s, \
88 MCS_DURATION(_streams, _sgi, \
89 - BW2VBPS(_bw, 1053, 486, 234)), \
90 + BW2VBPS(_bw, 1053, 486, 234)) >> _s, \
91 MCS_DURATION(_streams, _sgi, \
92 - BW2VBPS(_bw, 1170, 540, 260)), \
93 + BW2VBPS(_bw, 1170, 540, 260)) >> _s, \
94 MCS_DURATION(_streams, _sgi, \
95 - BW2VBPS(_bw, 1404, 648, 312)), \
96 + BW2VBPS(_bw, 1404, 648, 312)) >> _s, \
97 MCS_DURATION(_streams, _sgi, \
98 - BW2VBPS(_bw, 1560, 720, 346)) \
99 + BW2VBPS(_bw, 1560, 720, 346)) >> _s \
103 @@ -121,19 +123,20 @@
104 (CCK_DURATION((_bitrate > 10 ? 20 : 10), false, 60) + \
105 CCK_DURATION(_bitrate, _short, AVG_PKT_SIZE))
107 -#define CCK_DURATION_LIST(_short) \
108 - CCK_ACK_DURATION(10, _short), \
109 - CCK_ACK_DURATION(20, _short), \
110 - CCK_ACK_DURATION(55, _short), \
111 - CCK_ACK_DURATION(110, _short)
112 +#define CCK_DURATION_LIST(_short, _s) \
113 + CCK_ACK_DURATION(10, _short) >> _s, \
114 + CCK_ACK_DURATION(20, _short) >> _s, \
115 + CCK_ACK_DURATION(55, _short) >> _s, \
116 + CCK_ACK_DURATION(110, _short) >> _s
119 +#define CCK_GROUP(_s) \
120 [MINSTREL_CCK_GROUP] = { \
125 - CCK_DURATION_LIST(false), \
126 - CCK_DURATION_LIST(true) \
127 + CCK_DURATION_LIST(false, _s), \
128 + CCK_DURATION_LIST(true, _s) \
132 @@ -151,47 +154,47 @@ MODULE_PARM_DESC(minstrel_vht_only,
133 * BW -> SGI -> #streams
135 const struct mcs_group minstrel_mcs_groups[] = {
136 - MCS_GROUP(1, 0, BW_20),
137 - MCS_GROUP(2, 0, BW_20),
138 - MCS_GROUP(3, 0, BW_20),
140 - MCS_GROUP(1, 1, BW_20),
141 - MCS_GROUP(2, 1, BW_20),
142 - MCS_GROUP(3, 1, BW_20),
144 - MCS_GROUP(1, 0, BW_40),
145 - MCS_GROUP(2, 0, BW_40),
146 - MCS_GROUP(3, 0, BW_40),
148 - MCS_GROUP(1, 1, BW_40),
149 - MCS_GROUP(2, 1, BW_40),
150 - MCS_GROUP(3, 1, BW_40),
154 - VHT_GROUP(1, 0, BW_20),
155 - VHT_GROUP(2, 0, BW_20),
156 - VHT_GROUP(3, 0, BW_20),
158 - VHT_GROUP(1, 1, BW_20),
159 - VHT_GROUP(2, 1, BW_20),
160 - VHT_GROUP(3, 1, BW_20),
162 - VHT_GROUP(1, 0, BW_40),
163 - VHT_GROUP(2, 0, BW_40),
164 - VHT_GROUP(3, 0, BW_40),
166 - VHT_GROUP(1, 1, BW_40),
167 - VHT_GROUP(2, 1, BW_40),
168 - VHT_GROUP(3, 1, BW_40),
170 - VHT_GROUP(1, 0, BW_80),
171 - VHT_GROUP(2, 0, BW_80),
172 - VHT_GROUP(3, 0, BW_80),
174 - VHT_GROUP(1, 1, BW_80),
175 - VHT_GROUP(2, 1, BW_80),
176 - VHT_GROUP(3, 1, BW_80),
177 + MCS_GROUP(1, 0, BW_20, 5),
178 + MCS_GROUP(2, 0, BW_20, 4),
179 + MCS_GROUP(3, 0, BW_20, 4),
181 + MCS_GROUP(1, 1, BW_20, 5),
182 + MCS_GROUP(2, 1, BW_20, 4),
183 + MCS_GROUP(3, 1, BW_20, 4),
185 + MCS_GROUP(1, 0, BW_40, 4),
186 + MCS_GROUP(2, 0, BW_40, 4),
187 + MCS_GROUP(3, 0, BW_40, 4),
189 + MCS_GROUP(1, 1, BW_40, 4),
190 + MCS_GROUP(2, 1, BW_40, 4),
191 + MCS_GROUP(3, 1, BW_40, 4),
195 + VHT_GROUP(1, 0, BW_20, 5),
196 + VHT_GROUP(2, 0, BW_20, 4),
197 + VHT_GROUP(3, 0, BW_20, 4),
199 + VHT_GROUP(1, 1, BW_20, 5),
200 + VHT_GROUP(2, 1, BW_20, 4),
201 + VHT_GROUP(3, 1, BW_20, 4),
203 + VHT_GROUP(1, 0, BW_40, 4),
204 + VHT_GROUP(2, 0, BW_40, 4),
205 + VHT_GROUP(3, 0, BW_40, 4),
207 + VHT_GROUP(1, 1, BW_40, 4),
208 + VHT_GROUP(2, 1, BW_40, 4),
209 + VHT_GROUP(3, 1, BW_40, 4),
211 + VHT_GROUP(1, 0, BW_80, 4),
212 + VHT_GROUP(2, 0, BW_80, 4),
213 + VHT_GROUP(3, 0, BW_80, 4),
215 + VHT_GROUP(1, 1, BW_80, 4),
216 + VHT_GROUP(2, 1, BW_80, 4),
217 + VHT_GROUP(3, 1, BW_80, 4),
220 static u8 sample_table[SAMPLE_COLUMNS][MCS_GROUP_RATES] __read_mostly;
221 @@ -307,7 +310,8 @@ minstrel_ht_get_tp_avg(struct minstrel_h
222 if (group != MINSTREL_CCK_GROUP)
223 nsecs = 1000 * mi->overhead / MINSTREL_TRUNC(mi->avg_ampdu_len);
225 - nsecs += minstrel_mcs_groups[group].duration[rate];
226 + nsecs += minstrel_mcs_groups[group].duration[rate] <<
227 + minstrel_mcs_groups[group].shift;
230 * For the throughput calculation, limit the probability value to 90% to
231 @@ -755,12 +759,19 @@ minstrel_ht_tx_status(void *priv, struct
232 minstrel_ht_update_rates(mp, mi);
236 +minstrel_get_duration(int index)
238 + const struct mcs_group *group = &minstrel_mcs_groups[index / MCS_GROUP_RATES];
239 + unsigned int duration = group->duration[index % MCS_GROUP_RATES];
240 + return duration << group->shift;
244 minstrel_calc_retransmit(struct minstrel_priv *mp, struct minstrel_ht_sta *mi,
247 struct minstrel_rate_stats *mrs;
248 - const struct mcs_group *group;
249 unsigned int tx_time, tx_time_rtscts, tx_time_data;
250 unsigned int cw = mp->cw_min;
251 unsigned int ctime = 0;
252 @@ -779,8 +790,7 @@ minstrel_calc_retransmit(struct minstrel
253 mrs->retry_count_rtscts = 2;
254 mrs->retry_updated = true;
256 - group = &minstrel_mcs_groups[index / MCS_GROUP_RATES];
257 - tx_time_data = group->duration[index % MCS_GROUP_RATES] * ampdu_len / 1000;
258 + tx_time_data = minstrel_get_duration(index) * ampdu_len / 1000;
260 /* Contention time for first 2 tries */
261 ctime = (t_slot * cw) >> 1;
262 @@ -874,20 +884,24 @@ minstrel_ht_get_max_amsdu_len(struct min
263 int group = mi->max_prob_rate / MCS_GROUP_RATES;
264 const struct mcs_group *g = &minstrel_mcs_groups[group];
265 int rate = mi->max_prob_rate % MCS_GROUP_RATES;
266 + unsigned int duration;
268 /* Disable A-MSDU if max_prob_rate is bad */
269 if (mi->groups[group].rates[rate].prob_ewma < MINSTREL_FRAC(50, 100))
272 + duration = g->duration[rate];
273 + duration <<= g->shift;
275 /* If the rate is slower than single-stream MCS1, make A-MSDU limit small */
276 - if (g->duration[rate] > MCS_DURATION(1, 0, 52))
277 + if (duration > MCS_DURATION(1, 0, 52))
281 * If the rate is slower than single-stream MCS4, limit A-MSDU to usual
284 - if (g->duration[rate] > MCS_DURATION(1, 0, 104))
285 + if (duration > MCS_DURATION(1, 0, 104))
289 @@ -895,7 +909,7 @@ minstrel_ht_get_max_amsdu_len(struct min
290 * rate success probability is less than 75%, limit A-MSDU to twice the usual
293 - if (g->duration[rate] > MCS_DURATION(1, 0, 260) ||
294 + if (duration > MCS_DURATION(1, 0, 260) ||
295 (minstrel_ht_get_prob_ewma(mi, mi->max_tp_rate[0]) <
296 MINSTREL_FRAC(75, 100)))
298 @@ -942,13 +956,6 @@ minstrel_ht_update_rates(struct minstrel
299 rate_control_set_rates(mp->hw, mi->sta, rates);
303 -minstrel_get_duration(int index)
305 - const struct mcs_group *group = &minstrel_mcs_groups[index / MCS_GROUP_RATES];
306 - return group->duration[index % MCS_GROUP_RATES];
310 minstrel_get_sample_rate(struct minstrel_priv *mp, struct minstrel_ht_sta *mi)
312 --- a/net/mac80211/rc80211_minstrel_ht.h
313 +++ b/net/mac80211/rc80211_minstrel_ht.h
315 #define MCS_GROUP_RATES 10
319 - unsigned int streams;
320 - unsigned int duration[MCS_GROUP_RATES];
324 + u16 duration[MCS_GROUP_RATES];
327 extern const struct mcs_group minstrel_mcs_groups[];
328 --- a/net/mac80211/rc80211_minstrel_ht_debugfs.c
329 +++ b/net/mac80211/rc80211_minstrel_ht_debugfs.c
330 @@ -58,6 +58,7 @@ minstrel_ht_stats_dump(struct minstrel_h
331 static const int bitrates[4] = { 10, 20, 55, 110 };
332 int idx = i * MCS_GROUP_RATES + j;
333 unsigned int prob_ewmsd;
334 + unsigned int duration;
336 if (!(mi->supported[i] & BIT(j)))
338 @@ -95,7 +96,9 @@ minstrel_ht_stats_dump(struct minstrel_h
339 p += sprintf(p, " %3u ", idx);
341 /* tx_time[rate(i)] in usec */
342 - tx_time = DIV_ROUND_CLOSEST(mg->duration[j], 1000);
343 + duration = mg->duration[j];
344 + duration <<= mg->shift;
345 + tx_time = DIV_ROUND_CLOSEST(duration, 1000);
346 p += sprintf(p, "%6u ", tx_time);
348 tp_max = minstrel_ht_get_tp_avg(mi, i, j, MINSTREL_FRAC(100, 100));
349 @@ -204,6 +207,7 @@ minstrel_ht_stats_csv_dump(struct minstr
350 static const int bitrates[4] = { 10, 20, 55, 110 };
351 int idx = i * MCS_GROUP_RATES + j;
352 unsigned int prob_ewmsd;
353 + unsigned int duration;
355 if (!(mi->supported[i] & BIT(j)))
357 @@ -238,7 +242,10 @@ minstrel_ht_stats_csv_dump(struct minstr
360 p += sprintf(p, "%u,", idx);
361 - tx_time = DIV_ROUND_CLOSEST(mg->duration[j], 1000);
363 + duration = mg->duration[j];
364 + duration <<= mg->shift;
365 + tx_time = DIV_ROUND_CLOSEST(duration, 1000);
366 p += sprintf(p, "%u,", tx_time);
368 tp_max = minstrel_ht_get_tp_avg(mi, i, j, MINSTREL_FRAC(100, 100));