1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2018 Intel Corporation
8 #include <rte_cycles.h>
9 #include <rte_atomic.h>
10 #include <rte_malloc.h>
13 #include "rte_power.h"
14 #include "rte_power_empty_poll.h"
16 #define INTERVALS_PER_SECOND 100 /* (10ms) */
17 #define SECONDS_TO_TRAIN_FOR 2
18 #define DEFAULT_MED_TO_HIGH_PERCENT_THRESHOLD 70
19 #define DEFAULT_HIGH_TO_MED_PERCENT_THRESHOLD 30
20 #define DEFAULT_CYCLES_PER_PACKET 800
22 static struct ep_params *ep_params;
23 static uint32_t med_to_high_threshold = DEFAULT_MED_TO_HIGH_PERCENT_THRESHOLD;
24 static uint32_t high_to_med_threshold = DEFAULT_HIGH_TO_MED_PERCENT_THRESHOLD;
26 static uint32_t avail_freqs[RTE_MAX_LCORE][NUM_FREQS];
28 static uint32_t total_avail_freqs[RTE_MAX_LCORE];
30 static uint32_t freq_index[NUM_FREQ];
33 get_freq_index(enum freq_val index)
35 return freq_index[index];
40 set_power_freq(int lcore_id, enum freq_val freq, bool specific_freq)
43 uint32_t power_freq_index;
45 power_freq_index = get_freq_index(freq);
47 power_freq_index = freq;
49 err = rte_power_set_freq(lcore_id, power_freq_index);
55 static __rte_always_inline void
56 exit_training_state(struct priority_worker *poll_stats)
58 RTE_SET_USED(poll_stats);
61 static __rte_always_inline void
62 enter_training_state(struct priority_worker *poll_stats)
64 poll_stats->iter_counter = 0;
65 poll_stats->cur_freq = LOW;
66 poll_stats->queue_state = TRAINING;
69 static __rte_always_inline void
70 enter_normal_state(struct priority_worker *poll_stats)
72 /* Clear the averages arrays and strs */
73 memset(poll_stats->edpi_av, 0, sizeof(poll_stats->edpi_av));
76 poll_stats->cur_freq = MED;
77 poll_stats->iter_counter = 0;
78 poll_stats->threshold_ctr = 0;
79 poll_stats->queue_state = MED_NORMAL;
80 RTE_LOG(INFO, POWER, "Set the power freq to MED\n");
81 set_power_freq(poll_stats->lcore_id, MED, false);
83 poll_stats->thresh[MED].threshold_percent = med_to_high_threshold;
84 poll_stats->thresh[HGH].threshold_percent = high_to_med_threshold;
87 static __rte_always_inline void
88 enter_busy_state(struct priority_worker *poll_stats)
90 memset(poll_stats->edpi_av, 0, sizeof(poll_stats->edpi_av));
93 poll_stats->cur_freq = HGH;
94 poll_stats->iter_counter = 0;
95 poll_stats->threshold_ctr = 0;
96 poll_stats->queue_state = HGH_BUSY;
97 set_power_freq(poll_stats->lcore_id, HGH, false);
100 static __rte_always_inline void
101 enter_purge_state(struct priority_worker *poll_stats)
103 poll_stats->iter_counter = 0;
104 poll_stats->queue_state = LOW_PURGE;
107 static __rte_always_inline void
108 set_state(struct priority_worker *poll_stats,
109 enum queue_state new_state)
111 enum queue_state old_state = poll_stats->queue_state;
112 if (old_state != new_state) {
114 /* Call any old state exit functions */
115 if (old_state == TRAINING)
116 exit_training_state(poll_stats);
118 /* Call any new state entry functions */
119 if (new_state == TRAINING)
120 enter_training_state(poll_stats);
121 if (new_state == MED_NORMAL)
122 enter_normal_state(poll_stats);
123 if (new_state == HGH_BUSY)
124 enter_busy_state(poll_stats);
125 if (new_state == LOW_PURGE)
126 enter_purge_state(poll_stats);
130 static __rte_always_inline void
131 set_policy(struct priority_worker *poll_stats,
132 struct ep_policy *policy)
134 set_state(poll_stats, policy->state);
136 if (policy->state == TRAINING)
139 poll_stats->thresh[MED_NORMAL].base_edpi = policy->med_base_edpi;
140 poll_stats->thresh[HGH_BUSY].base_edpi = policy->hgh_base_edpi;
142 poll_stats->thresh[MED_NORMAL].trained = true;
143 poll_stats->thresh[HGH_BUSY].trained = true;
148 update_training_stats(struct priority_worker *poll_stats,
151 uint32_t max_train_iter)
153 RTE_SET_USED(specific_freq);
155 uint64_t p0_empty_deq;
157 if (poll_stats->cur_freq == freq &&
158 poll_stats->thresh[freq].trained == false) {
159 if (poll_stats->thresh[freq].cur_train_iter == 0) {
161 set_power_freq(poll_stats->lcore_id,
162 freq, specific_freq);
164 poll_stats->empty_dequeues_prev =
165 poll_stats->empty_dequeues;
167 poll_stats->thresh[freq].cur_train_iter++;
170 } else if (poll_stats->thresh[freq].cur_train_iter
173 p0_empty_deq = poll_stats->empty_dequeues -
174 poll_stats->empty_dequeues_prev;
176 poll_stats->empty_dequeues_prev =
177 poll_stats->empty_dequeues;
179 poll_stats->thresh[freq].base_edpi += p0_empty_deq;
180 poll_stats->thresh[freq].cur_train_iter++;
183 if (poll_stats->thresh[freq].trained == false) {
184 poll_stats->thresh[freq].base_edpi =
185 poll_stats->thresh[freq].base_edpi /
188 /* Add on a factor of 0.05%
189 * this should remove any
190 * false negatives when the system is 0% busy
192 poll_stats->thresh[freq].base_edpi +=
193 poll_stats->thresh[freq].base_edpi / 2000;
195 poll_stats->thresh[freq].trained = true;
196 poll_stats->cur_freq++;
203 static __rte_always_inline uint32_t
204 update_stats(struct priority_worker *poll_stats)
206 uint64_t tot_edpi = 0;
209 struct priority_worker *s = poll_stats;
211 uint64_t cur_edpi = s->empty_dequeues - s->empty_dequeues_prev;
213 s->empty_dequeues_prev = s->empty_dequeues;
215 if (s->thresh[s->cur_freq].base_edpi < cur_edpi) {
217 /* edpi mean empty poll counter difference per interval */
218 RTE_LOG(DEBUG, POWER, "cur_edpi is too large "
219 "cur edpi %"PRId64" "
220 "base edpi %"PRId64"\n",
222 s->thresh[s->cur_freq].base_edpi);
223 /* Value to make us fail need debug log*/
227 s->edpi_av[s->ec++ % BINS_AV] = cur_edpi;
229 for (j = 0; j < BINS_AV; j++) {
230 tot_edpi += s->edpi_av[j];
233 tot_edpi = tot_edpi / BINS_AV;
235 percent = 100 - (uint32_t)(((float)tot_edpi /
236 (float)s->thresh[s->cur_freq].base_edpi) * 100);
238 return (uint32_t)percent;
242 static __rte_always_inline void
243 update_stats_normal(struct priority_worker *poll_stats)
247 if (poll_stats->thresh[poll_stats->cur_freq].base_edpi == 0) {
249 enum freq_val cur_freq = poll_stats->cur_freq;
251 /* edpi mean empty poll counter difference per interval */
252 RTE_LOG(DEBUG, POWER, "cure freq is %d, edpi is %"PRIu64"\n",
254 poll_stats->thresh[cur_freq].base_edpi);
258 percent = update_stats(poll_stats);
261 /* edpi mean empty poll counter difference per interval */
262 RTE_LOG(DEBUG, POWER, "Edpi is bigger than threshold\n");
266 if (poll_stats->cur_freq == LOW)
267 RTE_LOG(INFO, POWER, "Purge Mode is not currently supported\n");
268 else if (poll_stats->cur_freq == MED) {
271 poll_stats->thresh[MED].threshold_percent) {
273 if (poll_stats->threshold_ctr < INTERVALS_PER_SECOND)
274 poll_stats->threshold_ctr++;
276 set_state(poll_stats, HGH_BUSY);
277 RTE_LOG(INFO, POWER, "MOVE to HGH\n");
282 poll_stats->threshold_ctr = 0;
285 } else if (poll_stats->cur_freq == HGH) {
288 poll_stats->thresh[HGH].threshold_percent) {
290 if (poll_stats->threshold_ctr < INTERVALS_PER_SECOND)
291 poll_stats->threshold_ctr++;
293 set_state(poll_stats, MED_NORMAL);
294 RTE_LOG(INFO, POWER, "MOVE to MED\n");
298 poll_stats->threshold_ctr = 0;
305 empty_poll_training(struct priority_worker *poll_stats,
306 uint32_t max_train_iter)
309 if (poll_stats->iter_counter < INTERVALS_PER_SECOND) {
310 poll_stats->iter_counter++;
315 update_training_stats(poll_stats,
320 update_training_stats(poll_stats,
325 update_training_stats(poll_stats,
331 if (poll_stats->thresh[LOW].trained == true
332 && poll_stats->thresh[MED].trained == true
333 && poll_stats->thresh[HGH].trained == true) {
335 set_state(poll_stats, MED_NORMAL);
337 RTE_LOG(INFO, POWER, "LOW threshold is %"PRIu64"\n",
338 poll_stats->thresh[LOW].base_edpi);
340 RTE_LOG(INFO, POWER, "MED threshold is %"PRIu64"\n",
341 poll_stats->thresh[MED].base_edpi);
344 RTE_LOG(INFO, POWER, "HIGH threshold is %"PRIu64"\n",
345 poll_stats->thresh[HGH].base_edpi);
347 RTE_LOG(INFO, POWER, "Training is Complete for %d\n",
348 poll_stats->lcore_id);
355 rte_empty_poll_detection(struct rte_timer *tim, void *arg)
360 struct priority_worker *poll_stats;
366 for (i = 0; i < NUM_NODES; i++) {
368 poll_stats = &(ep_params->wrk_data.wrk_stats[i]);
370 if (rte_lcore_is_enabled(poll_stats->lcore_id) == 0)
373 switch (poll_stats->queue_state) {
375 empty_poll_training(poll_stats,
376 ep_params->max_train_iter);
381 update_stats_normal(poll_stats);
396 rte_power_empty_poll_stat_init(struct ep_params **eptr, uint8_t *freq_tlb,
397 struct ep_policy *policy)
400 /* Allocate the ep_params structure */
401 ep_params = rte_zmalloc_socket(NULL,
402 sizeof(struct ep_params),
409 if (freq_tlb == NULL) {
410 freq_index[LOW] = 14;
414 freq_index[LOW] = freq_tlb[LOW];
415 freq_index[MED] = freq_tlb[MED];
416 freq_index[HGH] = freq_tlb[HGH];
419 RTE_LOG(INFO, POWER, "Initialize the Empty Poll\n");
421 /* Train for pre-defined period */
422 ep_params->max_train_iter = INTERVALS_PER_SECOND * SECONDS_TO_TRAIN_FOR;
424 struct stats_data *w = &ep_params->wrk_data;
428 /* initialize all wrk_stats state */
429 for (i = 0; i < NUM_NODES; i++) {
431 if (rte_lcore_is_enabled(i) == 0)
433 /*init the freqs table */
434 total_avail_freqs[i] = rte_power_freqs(i,
438 RTE_LOG(INFO, POWER, "total avail freq is %d , lcoreid %d\n",
439 total_avail_freqs[i],
442 if (get_freq_index(LOW) > total_avail_freqs[i])
445 if (rte_get_main_lcore() != i) {
446 w->wrk_stats[i].lcore_id = i;
447 set_policy(&w->wrk_stats[i], policy);
455 rte_power_empty_poll_stat_free(void)
458 RTE_LOG(INFO, POWER, "Close the Empty Poll\n");
460 if (ep_params != NULL)
465 rte_power_empty_poll_stat_update(unsigned int lcore_id)
467 struct priority_worker *poll_stats;
469 if (lcore_id >= NUM_NODES)
472 poll_stats = &(ep_params->wrk_data.wrk_stats[lcore_id]);
474 if (poll_stats->lcore_id == 0)
475 poll_stats->lcore_id = lcore_id;
477 poll_stats->empty_dequeues++;
483 rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt)
486 struct priority_worker *poll_stats;
488 if (lcore_id >= NUM_NODES)
491 poll_stats = &(ep_params->wrk_data.wrk_stats[lcore_id]);
493 if (poll_stats->lcore_id == 0)
494 poll_stats->lcore_id = lcore_id;
496 poll_stats->num_dequeue_pkts += nb_pkt;
503 rte_power_empty_poll_stat_fetch(unsigned int lcore_id)
505 struct priority_worker *poll_stats;
507 if (lcore_id >= NUM_NODES)
510 poll_stats = &(ep_params->wrk_data.wrk_stats[lcore_id]);
512 if (poll_stats->lcore_id == 0)
513 poll_stats->lcore_id = lcore_id;
515 return poll_stats->empty_dequeues;
519 rte_power_poll_stat_fetch(unsigned int lcore_id)
521 struct priority_worker *poll_stats;
523 if (lcore_id >= NUM_NODES)
526 poll_stats = &(ep_params->wrk_data.wrk_stats[lcore_id]);
528 if (poll_stats->lcore_id == 0)
529 poll_stats->lcore_id = lcore_id;
531 return poll_stats->num_dequeue_pkts;