2 * Copyright (c) 2016, Mellanox Technologies. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
35 /* Adaptive moderation profiles */
36 #define MLX5E_AM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE 256
37 #define MLX5E_RX_AM_DEF_PROFILE_CQE 1
38 #define MLX5E_RX_AM_DEF_PROFILE_EQE 1
39 #define MLX5E_PARAMS_AM_NUM_PROFILES 5
41 /* All profiles sizes must be MLX5E_PARAMS_AM_NUM_PROFILES */
42 #define MLX5_AM_EQE_PROFILES { \
43 {1, MLX5E_AM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
44 {8, MLX5E_AM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
45 {64, MLX5E_AM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
46 {128, MLX5E_AM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
47 {256, MLX5E_AM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
50 #define MLX5_AM_CQE_PROFILES { \
58 static const struct mlx5e_cq_moder
59 profile[MLX5_CQ_PERIOD_NUM_MODES][MLX5E_PARAMS_AM_NUM_PROFILES] = {
64 static inline struct mlx5e_cq_moder mlx5e_am_get_profile(u8 cq_period_mode, int ix)
66 return profile[cq_period_mode][ix];
69 struct mlx5e_cq_moder mlx5e_am_get_def_profile(u8 rx_cq_period_mode)
71 int default_profile_ix;
73 if (rx_cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE)
74 default_profile_ix = MLX5E_RX_AM_DEF_PROFILE_CQE;
75 else /* MLX5_CQ_PERIOD_MODE_START_FROM_EQE */
76 default_profile_ix = MLX5E_RX_AM_DEF_PROFILE_EQE;
78 return profile[rx_cq_period_mode][default_profile_ix];
81 /* Adaptive moderation logic */
83 MLX5E_AM_START_MEASURE,
84 MLX5E_AM_MEASURE_IN_PROGRESS,
85 MLX5E_AM_APPLY_NEW_PROFILE,
89 MLX5E_AM_PARKING_ON_TOP,
90 MLX5E_AM_PARKING_TIRED,
98 MLX5E_AM_STATS_BETTER,
107 static bool mlx5e_am_on_top(struct mlx5e_rx_am *am)
109 switch (am->tune_state) {
110 case MLX5E_AM_PARKING_ON_TOP:
111 case MLX5E_AM_PARKING_TIRED:
113 case MLX5E_AM_GOING_RIGHT:
114 return (am->steps_left > 1) && (am->steps_right == 1);
115 default: /* MLX5E_AM_GOING_LEFT */
116 return (am->steps_right > 1) && (am->steps_left == 1);
120 static void mlx5e_am_turn(struct mlx5e_rx_am *am)
122 switch (am->tune_state) {
123 case MLX5E_AM_PARKING_ON_TOP:
124 case MLX5E_AM_PARKING_TIRED:
126 case MLX5E_AM_GOING_RIGHT:
127 am->tune_state = MLX5E_AM_GOING_LEFT;
130 case MLX5E_AM_GOING_LEFT:
131 am->tune_state = MLX5E_AM_GOING_RIGHT;
137 static int mlx5e_am_step(struct mlx5e_rx_am *am)
139 if (am->tired == (MLX5E_PARAMS_AM_NUM_PROFILES * 2))
140 return MLX5E_AM_TOO_TIRED;
142 switch (am->tune_state) {
143 case MLX5E_AM_PARKING_ON_TOP:
144 case MLX5E_AM_PARKING_TIRED:
146 case MLX5E_AM_GOING_RIGHT:
147 if (am->profile_ix == (MLX5E_PARAMS_AM_NUM_PROFILES - 1))
148 return MLX5E_AM_ON_EDGE;
152 case MLX5E_AM_GOING_LEFT:
153 if (am->profile_ix == 0)
154 return MLX5E_AM_ON_EDGE;
161 return MLX5E_AM_STEPPED;
164 static void mlx5e_am_park_on_top(struct mlx5e_rx_am *am)
169 am->tune_state = MLX5E_AM_PARKING_ON_TOP;
172 static void mlx5e_am_park_tired(struct mlx5e_rx_am *am)
176 am->tune_state = MLX5E_AM_PARKING_TIRED;
179 static void mlx5e_am_exit_parking(struct mlx5e_rx_am *am)
181 am->tune_state = am->profile_ix ? MLX5E_AM_GOING_LEFT :
182 MLX5E_AM_GOING_RIGHT;
186 #define IS_SIGNIFICANT_DIFF(val, ref) \
187 (((100 * abs((val) - (ref))) / (ref)) > 10) /* more than 10% difference */
189 static int mlx5e_am_stats_compare(struct mlx5e_rx_am_stats *curr,
190 struct mlx5e_rx_am_stats *prev)
193 return curr->bpms ? MLX5E_AM_STATS_BETTER :
196 if (IS_SIGNIFICANT_DIFF(curr->bpms, prev->bpms))
197 return (curr->bpms > prev->bpms) ? MLX5E_AM_STATS_BETTER :
198 MLX5E_AM_STATS_WORSE;
200 if (IS_SIGNIFICANT_DIFF(curr->ppms, prev->ppms))
201 return (curr->ppms > prev->ppms) ? MLX5E_AM_STATS_BETTER :
202 MLX5E_AM_STATS_WORSE;
204 if (IS_SIGNIFICANT_DIFF(curr->epms, prev->epms))
205 return (curr->epms < prev->epms) ? MLX5E_AM_STATS_BETTER :
206 MLX5E_AM_STATS_WORSE;
208 return MLX5E_AM_STATS_SAME;
211 static bool mlx5e_am_decision(struct mlx5e_rx_am_stats *curr_stats,
212 struct mlx5e_rx_am *am)
214 int prev_state = am->tune_state;
215 int prev_ix = am->profile_ix;
219 switch (am->tune_state) {
220 case MLX5E_AM_PARKING_ON_TOP:
221 stats_res = mlx5e_am_stats_compare(curr_stats, &am->prev_stats);
222 if (stats_res != MLX5E_AM_STATS_SAME)
223 mlx5e_am_exit_parking(am);
226 case MLX5E_AM_PARKING_TIRED:
229 mlx5e_am_exit_parking(am);
232 case MLX5E_AM_GOING_RIGHT:
233 case MLX5E_AM_GOING_LEFT:
234 stats_res = mlx5e_am_stats_compare(curr_stats, &am->prev_stats);
235 if (stats_res != MLX5E_AM_STATS_BETTER)
238 if (mlx5e_am_on_top(am)) {
239 mlx5e_am_park_on_top(am);
243 step_res = mlx5e_am_step(am);
245 case MLX5E_AM_ON_EDGE:
246 mlx5e_am_park_on_top(am);
248 case MLX5E_AM_TOO_TIRED:
249 mlx5e_am_park_tired(am);
256 if ((prev_state != MLX5E_AM_PARKING_ON_TOP) ||
257 (am->tune_state != MLX5E_AM_PARKING_ON_TOP))
258 am->prev_stats = *curr_stats;
260 return am->profile_ix != prev_ix;
263 static void mlx5e_am_sample(struct mlx5e_rq *rq,
264 struct mlx5e_rx_am_sample *s)
266 s->time = ktime_get();
267 s->pkt_ctr = rq->stats.packets;
268 s->byte_ctr = rq->stats.bytes;
269 s->event_ctr = rq->cq.event_ctr;
272 #define MLX5E_AM_NEVENTS 64
273 #define BITS_PER_TYPE(type) (sizeof(type) * BITS_PER_BYTE)
274 #define BIT_GAP(bits, end, start) ((((end) - (start)) + BIT_ULL(bits)) & (BIT_ULL(bits) - 1))
276 static void mlx5e_am_calc_stats(struct mlx5e_rx_am_sample *start,
277 struct mlx5e_rx_am_sample *end,
278 struct mlx5e_rx_am_stats *curr_stats)
280 /* u32 holds up to 71 minutes, should be enough */
281 u32 delta_us = ktime_us_delta(end->time, start->time);
282 u32 npkts = BIT_GAP(BITS_PER_TYPE(u32), end->pkt_ctr, start->pkt_ctr);
283 u32 nbytes = BIT_GAP(BITS_PER_TYPE(u32), end->byte_ctr,
289 curr_stats->ppms = DIV_ROUND_UP(npkts * USEC_PER_MSEC, delta_us);
290 curr_stats->bpms = DIV_ROUND_UP(nbytes * USEC_PER_MSEC, delta_us);
291 curr_stats->epms = DIV_ROUND_UP(MLX5E_AM_NEVENTS * USEC_PER_MSEC,
295 void mlx5e_rx_am_work(struct work_struct *work)
297 struct mlx5e_rx_am *am = container_of(work, struct mlx5e_rx_am,
299 struct mlx5e_rq *rq = container_of(am, struct mlx5e_rq, am);
300 struct mlx5e_cq_moder cur_profile = profile[am->mode][am->profile_ix];
302 mlx5_core_modify_cq_moderation(rq->mdev, &rq->cq.mcq,
303 cur_profile.usec, cur_profile.pkts);
305 am->state = MLX5E_AM_START_MEASURE;
308 void mlx5e_rx_am(struct mlx5e_rq *rq)
310 struct mlx5e_rx_am *am = &rq->am;
311 struct mlx5e_rx_am_sample end_sample;
312 struct mlx5e_rx_am_stats curr_stats;
316 case MLX5E_AM_MEASURE_IN_PROGRESS:
317 nevents = BIT_GAP(BITS_PER_TYPE(u16), rq->cq.event_ctr,
318 am->start_sample.event_ctr);
319 if (nevents < MLX5E_AM_NEVENTS)
321 mlx5e_am_sample(rq, &end_sample);
322 mlx5e_am_calc_stats(&am->start_sample, &end_sample,
324 if (mlx5e_am_decision(&curr_stats, am)) {
325 am->state = MLX5E_AM_APPLY_NEW_PROFILE;
326 schedule_work(&am->work);
330 case MLX5E_AM_START_MEASURE:
331 mlx5e_am_sample(rq, &am->start_sample);
332 am->state = MLX5E_AM_MEASURE_IN_PROGRESS;
334 case MLX5E_AM_APPLY_NEW_PROFILE: