1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2020 Intel Corporation
6 #include <rte_cycles.h>
7 #include <rte_cpuflags.h>
8 #include <rte_malloc.h>
9 #include <rte_ethdev.h>
10 #include <rte_power_intrinsics.h>
12 #include "rte_power_pmd_mgmt.h"
14 #define EMPTYPOLL_MAX 512
16 /* store some internal state */
17 static struct pmd_conf_data {
18 /** what do we support? */
19 struct rte_cpu_intrinsics intrinsics_support;
20 /** pre-calculated tsc diff for 1us */
22 /** how many rte_pause can we fit in a microsecond? */
23 uint64_t pause_per_us;
27 * Possible power management states of an ethdev port.
30 /** Device power management is disabled. */
31 PMD_MGMT_DISABLED = 0,
32 /** Device power management is enabled. */
44 struct queue_list_entry {
45 TAILQ_ENTRY(queue_list_entry) next;
47 uint64_t n_empty_polls;
49 const struct rte_eth_rxtx_callback *cb;
53 TAILQ_HEAD(queue_list_head, queue_list_entry) head;
54 /**< List of queues associated with this lcore */
56 /**< How many queues are in the list? */
57 volatile enum pmd_mgmt_state pwr_mgmt_state;
58 /**< State of power management for this queue */
59 enum rte_power_pmd_mgmt_type cb_mode;
60 /**< Callback mode for this queue */
61 uint64_t n_queues_ready_to_sleep;
62 /**< Number of queues ready to enter power optimized state */
63 uint64_t sleep_target;
64 /**< Prevent a queue from triggering sleep multiple times */
65 } __rte_cache_aligned;
66 static struct pmd_core_cfg lcore_cfgs[RTE_MAX_LCORE];
69 queue_equal(const union queue *l, const union queue *r)
71 return l->val == r->val;
75 queue_copy(union queue *dst, const union queue *src)
80 static struct queue_list_entry *
81 queue_list_find(const struct pmd_core_cfg *cfg, const union queue *q)
83 struct queue_list_entry *cur;
85 TAILQ_FOREACH(cur, &cfg->head, next) {
86 if (queue_equal(&cur->queue, q))
93 queue_list_add(struct pmd_core_cfg *cfg, const union queue *q)
95 struct queue_list_entry *qle;
97 /* is it already in the list? */
98 if (queue_list_find(cfg, q) != NULL)
101 qle = malloc(sizeof(*qle));
104 memset(qle, 0, sizeof(*qle));
106 queue_copy(&qle->queue, q);
107 TAILQ_INSERT_TAIL(&cfg->head, qle, next);
113 static struct queue_list_entry *
114 queue_list_take(struct pmd_core_cfg *cfg, const union queue *q)
116 struct queue_list_entry *found;
118 found = queue_list_find(cfg, q);
122 TAILQ_REMOVE(&cfg->head, found, next);
125 /* freeing is responsibility of the caller */
132 const uint64_t hz = rte_get_timer_hz();
133 const uint64_t tsc_per_us = hz / US_PER_S; /* 1us */
135 global_data.tsc_per_us = tsc_per_us;
137 /* only do this if we don't have tpause */
138 if (!global_data.intrinsics_support.power_pause) {
139 const uint64_t start = rte_rdtsc_precise();
140 const uint32_t n_pauses = 10000;
141 double us, us_per_pause;
145 /* estimate number of rte_pause() calls per us*/
146 for (i = 0; i < n_pauses; i++)
149 end = rte_rdtsc_precise();
150 us = (end - start) / (double)tsc_per_us;
151 us_per_pause = us / n_pauses;
153 global_data.pause_per_us = (uint64_t)(1.0 / us_per_pause);
158 queue_reset(struct pmd_core_cfg *cfg, struct queue_list_entry *qcfg)
160 const bool is_ready_to_sleep = qcfg->n_sleeps == cfg->sleep_target;
162 /* reset empty poll counter for this queue */
163 qcfg->n_empty_polls = 0;
164 /* reset the queue sleep counter as well */
166 /* remove the queue from list of queues ready to sleep */
167 if (is_ready_to_sleep)
168 cfg->n_queues_ready_to_sleep--;
170 * no need change the lcore sleep target counter because this lcore will
171 * reach the n_sleeps anyway, and the other cores are already counted so
172 * there's no need to do anything else.
177 queue_can_sleep(struct pmd_core_cfg *cfg, struct queue_list_entry *qcfg)
179 /* this function is called - that means we have an empty poll */
180 qcfg->n_empty_polls++;
182 /* if we haven't reached threshold for empty polls, we can't sleep */
183 if (qcfg->n_empty_polls <= EMPTYPOLL_MAX)
187 * we've reached a point where we are able to sleep, but we still need
188 * to check if this queue has already been marked for sleeping.
190 if (qcfg->n_sleeps == cfg->sleep_target)
193 /* mark this queue as ready for sleep */
194 qcfg->n_sleeps = cfg->sleep_target;
195 cfg->n_queues_ready_to_sleep++;
201 lcore_can_sleep(struct pmd_core_cfg *cfg)
203 /* are all queues ready to sleep? */
204 if (cfg->n_queues_ready_to_sleep != cfg->n_queues)
207 /* we've reached an iteration where we can sleep, reset sleep counter */
208 cfg->n_queues_ready_to_sleep = 0;
211 * we do not reset any individual queue empty poll counters, because
212 * we want to keep sleeping on every poll until we actually get traffic.
219 clb_umwait(uint16_t port_id, uint16_t qidx, struct rte_mbuf **pkts __rte_unused,
220 uint16_t nb_rx, uint16_t max_pkts __rte_unused, void *arg)
222 struct queue_list_entry *queue_conf = arg;
224 /* this callback can't do more than one queue, omit multiqueue logic */
225 if (unlikely(nb_rx == 0)) {
226 queue_conf->n_empty_polls++;
227 if (unlikely(queue_conf->n_empty_polls > EMPTYPOLL_MAX)) {
228 struct rte_power_monitor_cond pmc;
231 /* use monitoring condition to sleep */
232 ret = rte_eth_get_monitor_addr(port_id, qidx,
235 rte_power_monitor(&pmc, UINT64_MAX);
238 queue_conf->n_empty_polls = 0;
244 clb_pause(uint16_t port_id __rte_unused, uint16_t qidx __rte_unused,
245 struct rte_mbuf **pkts __rte_unused, uint16_t nb_rx,
246 uint16_t max_pkts __rte_unused, void *arg)
248 const unsigned int lcore = rte_lcore_id();
249 struct queue_list_entry *queue_conf = arg;
250 struct pmd_core_cfg *lcore_conf;
251 const bool empty = nb_rx == 0;
253 lcore_conf = &lcore_cfgs[lcore];
257 queue_reset(lcore_conf, queue_conf);
259 /* can this queue sleep? */
260 if (!queue_can_sleep(lcore_conf, queue_conf))
263 /* can this lcore sleep? */
264 if (!lcore_can_sleep(lcore_conf))
267 /* sleep for 1 microsecond, use tpause if we have it */
268 if (global_data.intrinsics_support.power_pause) {
269 const uint64_t cur = rte_rdtsc();
270 const uint64_t wait_tsc =
271 cur + global_data.tsc_per_us;
272 rte_power_pause(wait_tsc);
275 for (i = 0; i < global_data.pause_per_us; i++)
284 clb_scale_freq(uint16_t port_id __rte_unused, uint16_t qidx __rte_unused,
285 struct rte_mbuf **pkts __rte_unused, uint16_t nb_rx,
286 uint16_t max_pkts __rte_unused, void *arg)
288 const unsigned int lcore = rte_lcore_id();
289 const bool empty = nb_rx == 0;
290 struct pmd_core_cfg *lcore_conf = &lcore_cfgs[lcore];
291 struct queue_list_entry *queue_conf = arg;
293 if (likely(!empty)) {
295 queue_reset(lcore_conf, queue_conf);
297 /* scale up freq immediately */
298 rte_power_freq_max(rte_lcore_id());
300 /* can this queue sleep? */
301 if (!queue_can_sleep(lcore_conf, queue_conf))
304 /* can this lcore sleep? */
305 if (!lcore_can_sleep(lcore_conf))
308 rte_power_freq_min(rte_lcore_id());
315 queue_stopped(const uint16_t port_id, const uint16_t queue_id)
317 struct rte_eth_rxq_info qinfo;
319 if (rte_eth_rx_queue_info_get(port_id, queue_id, &qinfo) < 0)
322 return qinfo.queue_state == RTE_ETH_QUEUE_STATE_STOPPED;
326 cfg_queues_stopped(struct pmd_core_cfg *queue_cfg)
328 const struct queue_list_entry *entry;
330 TAILQ_FOREACH(entry, &queue_cfg->head, next) {
331 const union queue *q = &entry->queue;
332 int ret = queue_stopped(q->portid, q->qid);
340 check_scale(unsigned int lcore)
342 enum power_management_env env;
344 /* only PSTATE and ACPI modes are supported */
345 if (!rte_power_check_env_supported(PM_ENV_ACPI_CPUFREQ) &&
346 !rte_power_check_env_supported(PM_ENV_PSTATE_CPUFREQ)) {
347 RTE_LOG(DEBUG, POWER, "Neither ACPI nor PSTATE modes are supported\n");
350 /* ensure we could initialize the power library */
351 if (rte_power_init(lcore))
354 /* ensure we initialized the correct env */
355 env = rte_power_get_env();
356 if (env != PM_ENV_ACPI_CPUFREQ && env != PM_ENV_PSTATE_CPUFREQ) {
357 RTE_LOG(DEBUG, POWER, "Neither ACPI nor PSTATE modes were initialized\n");
366 check_monitor(struct pmd_core_cfg *cfg, const union queue *qdata)
368 struct rte_power_monitor_cond dummy;
370 /* check if rte_power_monitor is supported */
371 if (!global_data.intrinsics_support.power_monitor) {
372 RTE_LOG(DEBUG, POWER, "Monitoring intrinsics are not supported\n");
376 if (cfg->n_queues > 0) {
377 RTE_LOG(DEBUG, POWER, "Monitoring multiple queues is not supported\n");
381 /* check if the device supports the necessary PMD API */
382 if (rte_eth_get_monitor_addr(qdata->portid, qdata->qid,
383 &dummy) == -ENOTSUP) {
384 RTE_LOG(DEBUG, POWER, "The device does not support rte_eth_get_monitor_addr\n");
393 rte_power_ethdev_pmgmt_queue_enable(unsigned int lcore_id, uint16_t port_id,
394 uint16_t queue_id, enum rte_power_pmd_mgmt_type mode)
396 const union queue qdata = {.portid = port_id, .qid = queue_id};
397 struct pmd_core_cfg *lcore_cfg;
398 struct queue_list_entry *queue_cfg;
399 struct rte_eth_dev_info info;
400 rte_rx_callback_fn clb;
403 RTE_ETH_VALID_PORTID_OR_ERR_RET(port_id, -EINVAL);
405 if (queue_id >= RTE_MAX_QUEUES_PER_PORT || lcore_id >= RTE_MAX_LCORE) {
410 if (rte_eth_dev_info_get(port_id, &info) < 0) {
415 /* check if queue id is valid */
416 if (queue_id >= info.nb_rx_queues) {
421 /* check if the queue is stopped */
422 ret = queue_stopped(port_id, queue_id);
424 /* error means invalid queue, 0 means queue wasn't stopped */
425 ret = ret < 0 ? -EINVAL : -EBUSY;
429 lcore_cfg = &lcore_cfgs[lcore_id];
431 /* check if other queues are stopped as well */
432 ret = cfg_queues_stopped(lcore_cfg);
434 /* error means invalid queue, 0 means queue wasn't stopped */
435 ret = ret < 0 ? -EINVAL : -EBUSY;
439 /* if callback was already enabled, check current callback type */
440 if (lcore_cfg->pwr_mgmt_state != PMD_MGMT_DISABLED &&
441 lcore_cfg->cb_mode != mode) {
446 /* we need this in various places */
447 rte_cpu_get_intrinsics_support(&global_data.intrinsics_support);
450 case RTE_POWER_MGMT_TYPE_MONITOR:
451 /* check if we can add a new queue */
452 ret = check_monitor(lcore_cfg, &qdata);
458 case RTE_POWER_MGMT_TYPE_SCALE:
459 /* check if we can add a new queue */
460 ret = check_scale(lcore_id);
463 clb = clb_scale_freq;
465 case RTE_POWER_MGMT_TYPE_PAUSE:
466 /* figure out various time-to-tsc conversions */
467 if (global_data.tsc_per_us == 0)
473 RTE_LOG(DEBUG, POWER, "Invalid power management type\n");
477 /* add this queue to the list */
478 ret = queue_list_add(lcore_cfg, &qdata);
480 RTE_LOG(DEBUG, POWER, "Failed to add queue to list: %s\n",
484 /* new queue is always added last */
485 queue_cfg = TAILQ_LAST(&lcore_cfg->head, queue_list_head);
487 /* when enabling first queue, ensure sleep target is not 0 */
488 if (lcore_cfg->n_queues == 1 && lcore_cfg->sleep_target == 0)
489 lcore_cfg->sleep_target = 1;
491 /* initialize data before enabling the callback */
492 if (lcore_cfg->n_queues == 1) {
493 lcore_cfg->cb_mode = mode;
494 lcore_cfg->pwr_mgmt_state = PMD_MGMT_ENABLED;
496 queue_cfg->cb = rte_eth_add_rx_callback(port_id, queue_id,
505 rte_power_ethdev_pmgmt_queue_disable(unsigned int lcore_id,
506 uint16_t port_id, uint16_t queue_id)
508 const union queue qdata = {.portid = port_id, .qid = queue_id};
509 struct pmd_core_cfg *lcore_cfg;
510 struct queue_list_entry *queue_cfg;
513 RTE_ETH_VALID_PORTID_OR_ERR_RET(port_id, -EINVAL);
515 if (lcore_id >= RTE_MAX_LCORE || queue_id >= RTE_MAX_QUEUES_PER_PORT)
518 /* check if the queue is stopped */
519 ret = queue_stopped(port_id, queue_id);
521 /* error means invalid queue, 0 means queue wasn't stopped */
522 return ret < 0 ? -EINVAL : -EBUSY;
525 /* no need to check queue id as wrong queue id would not be enabled */
526 lcore_cfg = &lcore_cfgs[lcore_id];
528 /* check if other queues are stopped as well */
529 ret = cfg_queues_stopped(lcore_cfg);
531 /* error means invalid queue, 0 means queue wasn't stopped */
532 return ret < 0 ? -EINVAL : -EBUSY;
535 if (lcore_cfg->pwr_mgmt_state != PMD_MGMT_ENABLED)
539 * There is no good/easy way to do this without race conditions, so we
540 * are just going to throw our hands in the air and hope that the user
541 * has read the documentation and has ensured that ports are stopped at
542 * the time we enter the API functions.
544 queue_cfg = queue_list_take(lcore_cfg, &qdata);
545 if (queue_cfg == NULL)
548 /* if we've removed all queues from the lists, set state to disabled */
549 if (lcore_cfg->n_queues == 0)
550 lcore_cfg->pwr_mgmt_state = PMD_MGMT_DISABLED;
552 switch (lcore_cfg->cb_mode) {
553 case RTE_POWER_MGMT_TYPE_MONITOR: /* fall-through */
554 case RTE_POWER_MGMT_TYPE_PAUSE:
555 rte_eth_remove_rx_callback(port_id, queue_id, queue_cfg->cb);
557 case RTE_POWER_MGMT_TYPE_SCALE:
558 rte_power_freq_max(lcore_id);
559 rte_eth_remove_rx_callback(port_id, queue_id, queue_cfg->cb);
560 rte_power_exit(lcore_id);
564 * the API doc mandates that the user stops all processing on affected
565 * ports before calling any of these API's, so we can assume that the
566 * callbacks can be freed. we're intentionally casting away const-ness.
568 rte_free((void *)queue_cfg->cb);
574 RTE_INIT(rte_power_ethdev_pmgmt_init) {
577 /* initialize all tailqs */
578 for (i = 0; i < RTE_DIM(lcore_cfgs); i++) {
579 struct pmd_core_cfg *cfg = &lcore_cfgs[i];
580 TAILQ_INIT(&cfg->head);