1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2018 Intel Corporation
11 #include <sys/queue.h>
19 #include <rte_common.h>
20 #include <rte_byteorder.h>
22 #include <rte_malloc.h>
23 #include <rte_memory.h>
24 #include <rte_memcpy.h>
26 #include <rte_launch.h>
27 #include <rte_atomic.h>
28 #include <rte_cycles.h>
29 #include <rte_prefetch.h>
30 #include <rte_lcore.h>
31 #include <rte_per_lcore.h>
32 #include <rte_branch_prediction.h>
33 #include <rte_interrupts.h>
34 #include <rte_random.h>
35 #include <rte_debug.h>
36 #include <rte_ether.h>
37 #include <rte_ethdev.h>
38 #include <rte_mempool.h>
43 #include <rte_string_fns.h>
44 #include <rte_timer.h>
45 #include <rte_power.h>
46 #include <rte_spinlock.h>
47 #include <rte_power_empty_poll.h>
48 #include <rte_metrics.h>
49 #include <rte_telemetry.h>
50 #include <rte_power_pmd_mgmt.h>
52 #include "perf_core.h"
55 #define RTE_LOGTYPE_L3FWD_POWER RTE_LOGTYPE_USER1
57 #define MAX_PKT_BURST 32
59 #define MIN_ZERO_POLL_COUNT 10
62 #define TIMER_NUMBER_PER_SECOND 10
64 #define INTERVALS_PER_SECOND 100
66 #define SCALING_PERIOD (1000000/TIMER_NUMBER_PER_SECOND)
67 #define SCALING_DOWN_TIME_RATIO_THRESHOLD 0.25
69 #define APP_LOOKUP_EXACT_MATCH 0
70 #define APP_LOOKUP_LPM 1
71 #define DO_RFC_1812_CHECKS
73 #ifndef APP_LOOKUP_METHOD
74 #define APP_LOOKUP_METHOD APP_LOOKUP_LPM
77 #if (APP_LOOKUP_METHOD == APP_LOOKUP_EXACT_MATCH)
79 #elif (APP_LOOKUP_METHOD == APP_LOOKUP_LPM)
82 #error "APP_LOOKUP_METHOD set to incorrect value"
86 #define IPv6_BYTES_FMT "%02x%02x:%02x%02x:%02x%02x:%02x%02x:"\
87 "%02x%02x:%02x%02x:%02x%02x:%02x%02x"
88 #define IPv6_BYTES(addr) \
89 addr[0], addr[1], addr[2], addr[3], \
90 addr[4], addr[5], addr[6], addr[7], \
91 addr[8], addr[9], addr[10], addr[11],\
92 addr[12], addr[13],addr[14], addr[15]
95 #define MAX_JUMBO_PKT_LEN 9600
97 #define IPV6_ADDR_LEN 16
99 #define MEMPOOL_CACHE_SIZE 256
102 * This expression is used to calculate the number of mbufs needed depending on
103 * user input, taking into account memory for rx and tx hardware rings, cache
104 * per lcore and mtable per port per lcore. RTE_MAX is used to ensure that
105 * NB_MBUF never goes below a minimum value of 8192.
108 #define NB_MBUF RTE_MAX ( \
109 (nb_ports*nb_rx_queue*nb_rxd + \
110 nb_ports*nb_lcores*MAX_PKT_BURST + \
111 nb_ports*n_tx_queue*nb_txd + \
112 nb_lcores*MEMPOOL_CACHE_SIZE), \
115 #define BURST_TX_DRAIN_US 100 /* TX drain every ~100us */
119 /* Configure how many packets ahead to prefetch, when reading packets */
120 #define PREFETCH_OFFSET 3
123 * Configurable number of RX/TX ring descriptors
125 #define RTE_TEST_RX_DESC_DEFAULT 1024
126 #define RTE_TEST_TX_DESC_DEFAULT 1024
129 * These two thresholds were decided on by running the training algorithm on
130 * a 2.5GHz Xeon. These defaults can be overridden by supplying non-zero values
131 * for the med_threshold and high_threshold parameters on the command line.
133 #define EMPTY_POLL_MED_THRESHOLD 350000UL
134 #define EMPTY_POLL_HGH_THRESHOLD 580000UL
136 #define NUM_TELSTATS RTE_DIM(telstats_strings)
138 static uint16_t nb_rxd = RTE_TEST_RX_DESC_DEFAULT;
139 static uint16_t nb_txd = RTE_TEST_TX_DESC_DEFAULT;
141 /* ethernet addresses of ports */
142 static struct rte_ether_addr ports_eth_addr[RTE_MAX_ETHPORTS];
144 /* ethernet addresses of ports */
145 static rte_spinlock_t locks[RTE_MAX_ETHPORTS];
147 /* mask of enabled ports */
148 static uint32_t enabled_port_mask = 0;
149 /* Ports set in promiscuous mode off by default. */
150 static int promiscuous_on = 0;
151 /* NUMA is enabled by default. */
152 static int numa_on = 1;
153 static bool empty_poll_stop;
154 static bool empty_poll_train;
155 volatile bool quit_signal;
156 static struct ep_params *ep_params;
157 static struct ep_policy policy;
158 static long ep_med_edpi, ep_hgh_edpi;
159 /* timer to update telemetry every 500ms */
160 static struct rte_timer telemetry_timer;
162 /* stats index returned by metrics lib */
165 struct telstats_name {
166 char name[RTE_ETH_XSTATS_NAME_SIZE];
169 /* telemetry stats to be reported */
170 const struct telstats_name telstats_strings[] = {
176 /* core busyness in percentage */
183 /* reference poll count to measure core busyness */
184 #define DEFAULT_COUNT 10000
186 * reference CYCLES to be used to
187 * measure core busyness based on poll count
189 #define MIN_CYCLES 1500000ULL
190 #define MAX_CYCLES 22000000ULL
193 #define TELEMETRY_INTERVALS_PER_SEC 2
195 static int parse_ptype; /**< Parse packet type using rx callback, and */
196 /**< disabled by default */
199 APP_MODE_DEFAULT = 0,
207 enum appmode app_mode;
209 static enum rte_power_pmd_mgmt_type pmgmt_type;
211 enum freq_scale_hint_t
219 struct lcore_rx_queue {
222 enum freq_scale_hint_t freq_up_hint;
223 uint32_t zero_rx_packet_count;
225 } __rte_cache_aligned;
227 #define MAX_RX_QUEUE_PER_LCORE 16
228 #define MAX_TX_QUEUE_PER_PORT RTE_MAX_ETHPORTS
229 #define MAX_RX_QUEUE_PER_PORT 128
231 #define MAX_RX_QUEUE_INTERRUPT_PER_PORT 16
234 struct lcore_params lcore_params_array[MAX_LCORE_PARAMS];
235 static struct lcore_params lcore_params_array_default[] = {
247 struct lcore_params *lcore_params = lcore_params_array_default;
248 uint16_t nb_lcore_params = RTE_DIM(lcore_params_array_default);
250 static struct rte_eth_conf port_conf = {
252 .mq_mode = ETH_MQ_RX_RSS,
253 .max_rx_pkt_len = RTE_ETHER_MAX_LEN,
255 .offloads = DEV_RX_OFFLOAD_CHECKSUM,
260 .rss_hf = ETH_RSS_UDP,
264 .mq_mode = ETH_MQ_TX_NONE,
268 static struct rte_mempool * pktmbuf_pool[NB_SOCKETS];
271 #if (APP_LOOKUP_METHOD == APP_LOOKUP_EXACT_MATCH)
274 #include <rte_hash_crc.h>
275 #define DEFAULT_HASH_FUNC rte_hash_crc
277 #include <rte_jhash.h>
278 #define DEFAULT_HASH_FUNC rte_jhash
290 uint8_t ip_dst[IPV6_ADDR_LEN];
291 uint8_t ip_src[IPV6_ADDR_LEN];
297 struct ipv4_l3fwd_route {
298 struct ipv4_5tuple key;
302 struct ipv6_l3fwd_route {
303 struct ipv6_5tuple key;
307 static struct ipv4_l3fwd_route ipv4_l3fwd_route_array[] = {
308 {{RTE_IPV4(100,10,0,1), RTE_IPV4(200,10,0,1), 101, 11, IPPROTO_TCP}, 0},
309 {{RTE_IPV4(100,20,0,2), RTE_IPV4(200,20,0,2), 102, 12, IPPROTO_TCP}, 1},
310 {{RTE_IPV4(100,30,0,3), RTE_IPV4(200,30,0,3), 103, 13, IPPROTO_TCP}, 2},
311 {{RTE_IPV4(100,40,0,4), RTE_IPV4(200,40,0,4), 104, 14, IPPROTO_TCP}, 3},
314 static struct ipv6_l3fwd_route ipv6_l3fwd_route_array[] = {
317 {0xfe, 0x80, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
318 0x02, 0x1b, 0x21, 0xff, 0xfe, 0x91, 0x38, 0x05},
319 {0xfe, 0x80, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
320 0x02, 0x1e, 0x67, 0xff, 0xfe, 0x0d, 0xb6, 0x0a},
326 typedef struct rte_hash lookup_struct_t;
327 static lookup_struct_t *ipv4_l3fwd_lookup_struct[NB_SOCKETS];
328 static lookup_struct_t *ipv6_l3fwd_lookup_struct[NB_SOCKETS];
330 #define L3FWD_HASH_ENTRIES 1024
332 static uint16_t ipv4_l3fwd_out_if[L3FWD_HASH_ENTRIES] __rte_cache_aligned;
333 static uint16_t ipv6_l3fwd_out_if[L3FWD_HASH_ENTRIES] __rte_cache_aligned;
336 #if (APP_LOOKUP_METHOD == APP_LOOKUP_LPM)
337 struct ipv4_l3fwd_route {
343 static struct ipv4_l3fwd_route ipv4_l3fwd_route_array[] = {
344 {RTE_IPV4(1,1,1,0), 24, 0},
345 {RTE_IPV4(2,1,1,0), 24, 1},
346 {RTE_IPV4(3,1,1,0), 24, 2},
347 {RTE_IPV4(4,1,1,0), 24, 3},
348 {RTE_IPV4(5,1,1,0), 24, 4},
349 {RTE_IPV4(6,1,1,0), 24, 5},
350 {RTE_IPV4(7,1,1,0), 24, 6},
351 {RTE_IPV4(8,1,1,0), 24, 7},
354 #define IPV4_L3FWD_LPM_MAX_RULES 1024
356 typedef struct rte_lpm lookup_struct_t;
357 static lookup_struct_t *ipv4_l3fwd_lookup_struct[NB_SOCKETS];
362 struct lcore_rx_queue rx_queue_list[MAX_RX_QUEUE_PER_LCORE];
364 uint16_t tx_port_id[RTE_MAX_ETHPORTS];
365 uint16_t tx_queue_id[RTE_MAX_ETHPORTS];
366 struct rte_eth_dev_tx_buffer *tx_buffer[RTE_MAX_ETHPORTS];
367 lookup_struct_t * ipv4_lookup_struct;
368 lookup_struct_t * ipv6_lookup_struct;
369 } __rte_cache_aligned;
372 /* total sleep time in ms since last frequency scaling down */
374 /* number of long sleep recently */
375 uint32_t nb_long_sleep;
376 /* freq. scaling up trend */
378 /* total packet processed recently */
379 uint64_t nb_rx_processed;
380 /* total iterations looped recently */
381 uint64_t nb_iteration_looped;
383 * Represents empty and non empty polls
384 * of rte_eth_rx_burst();
385 * ep_nep[0] holds non empty polls
386 * i.e. 0 < nb_rx <= MAX_BURST
387 * ep_nep[1] holds empty polls.
392 * Represents full and empty+partial
393 * polls of rte_eth_rx_burst();
394 * ep_nep[0] holds empty+partial polls.
395 * i.e. 0 <= nb_rx < MAX_BURST
396 * ep_nep[1] holds full polls
397 * i.e. nb_rx == MAX_BURST
401 rte_spinlock_t telemetry_lock;
402 } __rte_cache_aligned;
404 static struct lcore_conf lcore_conf[RTE_MAX_LCORE] __rte_cache_aligned;
405 static struct lcore_stats stats[RTE_MAX_LCORE] __rte_cache_aligned;
406 static struct rte_timer power_timers[RTE_MAX_LCORE];
408 static inline uint32_t power_idle_heuristic(uint32_t zero_rx_packet_count);
409 static inline enum freq_scale_hint_t power_freq_scaleup_heuristic( \
410 unsigned int lcore_id, uint16_t port_id, uint16_t queue_id);
414 * These defaults are using the max frequency index (1), a medium index (9)
415 * and a typical low frequency index (14). These can be adjusted to use
416 * different indexes using the relevant command line parameters.
418 static uint8_t freq_tlb[] = {14, 9, 1};
420 static int is_done(void)
425 /* exit signal handler */
427 signal_exit_now(int sigtype)
430 if (sigtype == SIGINT)
435 /* Freqency scale down timer callback */
437 power_timer_cb(__rte_unused struct rte_timer *tim,
438 __rte_unused void *arg)
441 float sleep_time_ratio;
442 unsigned lcore_id = rte_lcore_id();
444 /* accumulate total execution time in us when callback is invoked */
445 sleep_time_ratio = (float)(stats[lcore_id].sleep_time) /
446 (float)SCALING_PERIOD;
448 * check whether need to scale down frequency a step if it sleep a lot.
450 if (sleep_time_ratio >= SCALING_DOWN_TIME_RATIO_THRESHOLD) {
451 if (rte_power_freq_down)
452 rte_power_freq_down(lcore_id);
454 else if ( (unsigned)(stats[lcore_id].nb_rx_processed /
455 stats[lcore_id].nb_iteration_looped) < MAX_PKT_BURST) {
457 * scale down a step if average packet per iteration less
460 if (rte_power_freq_down)
461 rte_power_freq_down(lcore_id);
465 * initialize another timer according to current frequency to ensure
466 * timer interval is relatively fixed.
468 hz = rte_get_timer_hz();
469 rte_timer_reset(&power_timers[lcore_id], hz/TIMER_NUMBER_PER_SECOND,
470 SINGLE, lcore_id, power_timer_cb, NULL);
472 stats[lcore_id].nb_rx_processed = 0;
473 stats[lcore_id].nb_iteration_looped = 0;
475 stats[lcore_id].sleep_time = 0;
478 /* Enqueue a single packet, and send burst if queue is filled */
480 send_single_packet(struct rte_mbuf *m, uint16_t port)
483 struct lcore_conf *qconf;
485 lcore_id = rte_lcore_id();
486 qconf = &lcore_conf[lcore_id];
488 rte_eth_tx_buffer(port, qconf->tx_queue_id[port],
489 qconf->tx_buffer[port], m);
494 #ifdef DO_RFC_1812_CHECKS
496 is_valid_ipv4_pkt(struct rte_ipv4_hdr *pkt, uint32_t link_len)
498 /* From http://www.rfc-editor.org/rfc/rfc1812.txt section 5.2.2 */
500 * 1. The packet length reported by the Link Layer must be large
501 * enough to hold the minimum length legal IP datagram (20 bytes).
503 if (link_len < sizeof(struct rte_ipv4_hdr))
506 /* 2. The IP checksum must be correct. */
507 /* this is checked in H/W */
510 * 3. The IP version number must be 4. If the version number is not 4
511 * then the packet may be another version of IP, such as IPng or
514 if (((pkt->version_ihl) >> 4) != 4)
517 * 4. The IP header length field must be large enough to hold the
518 * minimum length legal IP datagram (20 bytes = 5 words).
520 if ((pkt->version_ihl & 0xf) < 5)
524 * 5. The IP total length field must be large enough to hold the IP
525 * datagram header, whose length is specified in the IP header length
528 if (rte_cpu_to_be_16(pkt->total_length) < sizeof(struct rte_ipv4_hdr))
535 #if (APP_LOOKUP_METHOD == APP_LOOKUP_EXACT_MATCH)
537 print_ipv4_key(struct ipv4_5tuple key)
539 printf("IP dst = %08x, IP src = %08x, port dst = %d, port src = %d, "
540 "proto = %d\n", (unsigned)key.ip_dst, (unsigned)key.ip_src,
541 key.port_dst, key.port_src, key.proto);
544 print_ipv6_key(struct ipv6_5tuple key)
546 printf( "IP dst = " IPv6_BYTES_FMT ", IP src = " IPv6_BYTES_FMT ", "
547 "port dst = %d, port src = %d, proto = %d\n",
548 IPv6_BYTES(key.ip_dst), IPv6_BYTES(key.ip_src),
549 key.port_dst, key.port_src, key.proto);
552 static inline uint16_t
553 get_ipv4_dst_port(struct rte_ipv4_hdr *ipv4_hdr, uint16_t portid,
554 lookup_struct_t * ipv4_l3fwd_lookup_struct)
556 struct ipv4_5tuple key;
557 struct rte_tcp_hdr *tcp;
558 struct rte_udp_hdr *udp;
561 key.ip_dst = rte_be_to_cpu_32(ipv4_hdr->dst_addr);
562 key.ip_src = rte_be_to_cpu_32(ipv4_hdr->src_addr);
563 key.proto = ipv4_hdr->next_proto_id;
565 switch (ipv4_hdr->next_proto_id) {
567 tcp = (struct rte_tcp_hdr *)((unsigned char *)ipv4_hdr +
568 sizeof(struct rte_ipv4_hdr));
569 key.port_dst = rte_be_to_cpu_16(tcp->dst_port);
570 key.port_src = rte_be_to_cpu_16(tcp->src_port);
574 udp = (struct rte_udp_hdr *)((unsigned char *)ipv4_hdr +
575 sizeof(struct rte_ipv4_hdr));
576 key.port_dst = rte_be_to_cpu_16(udp->dst_port);
577 key.port_src = rte_be_to_cpu_16(udp->src_port);
586 /* Find destination port */
587 ret = rte_hash_lookup(ipv4_l3fwd_lookup_struct, (const void *)&key);
588 return ((ret < 0) ? portid : ipv4_l3fwd_out_if[ret]);
591 static inline uint16_t
592 get_ipv6_dst_port(struct rte_ipv6_hdr *ipv6_hdr, uint16_t portid,
593 lookup_struct_t *ipv6_l3fwd_lookup_struct)
595 struct ipv6_5tuple key;
596 struct rte_tcp_hdr *tcp;
597 struct rte_udp_hdr *udp;
600 memcpy(key.ip_dst, ipv6_hdr->dst_addr, IPV6_ADDR_LEN);
601 memcpy(key.ip_src, ipv6_hdr->src_addr, IPV6_ADDR_LEN);
603 key.proto = ipv6_hdr->proto;
605 switch (ipv6_hdr->proto) {
607 tcp = (struct rte_tcp_hdr *)((unsigned char *) ipv6_hdr +
608 sizeof(struct rte_ipv6_hdr));
609 key.port_dst = rte_be_to_cpu_16(tcp->dst_port);
610 key.port_src = rte_be_to_cpu_16(tcp->src_port);
614 udp = (struct rte_udp_hdr *)((unsigned char *) ipv6_hdr +
615 sizeof(struct rte_ipv6_hdr));
616 key.port_dst = rte_be_to_cpu_16(udp->dst_port);
617 key.port_src = rte_be_to_cpu_16(udp->src_port);
626 /* Find destination port */
627 ret = rte_hash_lookup(ipv6_l3fwd_lookup_struct, (const void *)&key);
628 return ((ret < 0) ? portid : ipv6_l3fwd_out_if[ret]);
632 #if (APP_LOOKUP_METHOD == APP_LOOKUP_LPM)
633 static inline uint16_t
634 get_ipv4_dst_port(struct rte_ipv4_hdr *ipv4_hdr, uint16_t portid,
635 lookup_struct_t *ipv4_l3fwd_lookup_struct)
639 return ((rte_lpm_lookup(ipv4_l3fwd_lookup_struct,
640 rte_be_to_cpu_32(ipv4_hdr->dst_addr), &next_hop) == 0)?
646 parse_ptype_one(struct rte_mbuf *m)
648 struct rte_ether_hdr *eth_hdr;
649 uint32_t packet_type = RTE_PTYPE_UNKNOWN;
652 eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
653 ether_type = eth_hdr->ether_type;
654 if (ether_type == rte_cpu_to_be_16(RTE_ETHER_TYPE_IPV4))
655 packet_type |= RTE_PTYPE_L3_IPV4_EXT_UNKNOWN;
656 else if (ether_type == rte_cpu_to_be_16(RTE_ETHER_TYPE_IPV6))
657 packet_type |= RTE_PTYPE_L3_IPV6_EXT_UNKNOWN;
659 m->packet_type = packet_type;
663 cb_parse_ptype(uint16_t port __rte_unused, uint16_t queue __rte_unused,
664 struct rte_mbuf *pkts[], uint16_t nb_pkts,
665 uint16_t max_pkts __rte_unused,
666 void *user_param __rte_unused)
670 for (i = 0; i < nb_pkts; ++i)
671 parse_ptype_one(pkts[i]);
677 add_cb_parse_ptype(uint16_t portid, uint16_t queueid)
679 printf("Port %d: softly parse packet type info\n", portid);
680 if (rte_eth_add_rx_callback(portid, queueid, cb_parse_ptype, NULL))
683 printf("Failed to add rx callback: port=%d\n", portid);
688 l3fwd_simple_forward(struct rte_mbuf *m, uint16_t portid,
689 struct lcore_conf *qconf)
691 struct rte_ether_hdr *eth_hdr;
692 struct rte_ipv4_hdr *ipv4_hdr;
696 eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
698 if (RTE_ETH_IS_IPV4_HDR(m->packet_type)) {
699 /* Handle IPv4 headers.*/
701 rte_pktmbuf_mtod_offset(m, struct rte_ipv4_hdr *,
702 sizeof(struct rte_ether_hdr));
704 #ifdef DO_RFC_1812_CHECKS
705 /* Check to make sure the packet is valid (RFC1812) */
706 if (is_valid_ipv4_pkt(ipv4_hdr, m->pkt_len) < 0) {
712 dst_port = get_ipv4_dst_port(ipv4_hdr, portid,
713 qconf->ipv4_lookup_struct);
714 if (dst_port >= RTE_MAX_ETHPORTS ||
715 (enabled_port_mask & 1 << dst_port) == 0)
718 /* 02:00:00:00:00:xx */
719 d_addr_bytes = ð_hdr->d_addr.addr_bytes[0];
720 *((uint64_t *)d_addr_bytes) =
721 0x000000000002 + ((uint64_t)dst_port << 40);
723 #ifdef DO_RFC_1812_CHECKS
724 /* Update time to live and header checksum */
725 --(ipv4_hdr->time_to_live);
726 ++(ipv4_hdr->hdr_checksum);
730 rte_ether_addr_copy(&ports_eth_addr[dst_port],
733 send_single_packet(m, dst_port);
734 } else if (RTE_ETH_IS_IPV6_HDR(m->packet_type)) {
735 /* Handle IPv6 headers.*/
736 #if (APP_LOOKUP_METHOD == APP_LOOKUP_EXACT_MATCH)
737 struct rte_ipv6_hdr *ipv6_hdr;
740 rte_pktmbuf_mtod_offset(m, struct rte_ipv6_hdr *,
741 sizeof(struct rte_ether_hdr));
743 dst_port = get_ipv6_dst_port(ipv6_hdr, portid,
744 qconf->ipv6_lookup_struct);
746 if (dst_port >= RTE_MAX_ETHPORTS ||
747 (enabled_port_mask & 1 << dst_port) == 0)
750 /* 02:00:00:00:00:xx */
751 d_addr_bytes = ð_hdr->d_addr.addr_bytes[0];
752 *((uint64_t *)d_addr_bytes) =
753 0x000000000002 + ((uint64_t)dst_port << 40);
756 rte_ether_addr_copy(&ports_eth_addr[dst_port],
759 send_single_packet(m, dst_port);
761 /* We don't currently handle IPv6 packets in LPM mode. */
769 #define MINIMUM_SLEEP_TIME 1
770 #define SUSPEND_THRESHOLD 300
772 static inline uint32_t
773 power_idle_heuristic(uint32_t zero_rx_packet_count)
775 /* If zero count is less than 100, sleep 1us */
776 if (zero_rx_packet_count < SUSPEND_THRESHOLD)
777 return MINIMUM_SLEEP_TIME;
778 /* If zero count is less than 1000, sleep 100 us which is the
779 minimum latency switching from C3/C6 to C0
782 return SUSPEND_THRESHOLD;
785 static inline enum freq_scale_hint_t
786 power_freq_scaleup_heuristic(unsigned lcore_id,
790 uint32_t rxq_count = rte_eth_rx_queue_count(port_id, queue_id);
792 * HW Rx queue size is 128 by default, Rx burst read at maximum 32 entries
795 #define FREQ_GEAR1_RX_PACKET_THRESHOLD MAX_PKT_BURST
796 #define FREQ_GEAR2_RX_PACKET_THRESHOLD (MAX_PKT_BURST*2)
797 #define FREQ_GEAR3_RX_PACKET_THRESHOLD (MAX_PKT_BURST*3)
798 #define FREQ_UP_TREND1_ACC 1
799 #define FREQ_UP_TREND2_ACC 100
800 #define FREQ_UP_THRESHOLD 10000
802 if (likely(rxq_count > FREQ_GEAR3_RX_PACKET_THRESHOLD)) {
803 stats[lcore_id].trend = 0;
805 } else if (likely(rxq_count > FREQ_GEAR2_RX_PACKET_THRESHOLD))
806 stats[lcore_id].trend += FREQ_UP_TREND2_ACC;
807 else if (likely(rxq_count > FREQ_GEAR1_RX_PACKET_THRESHOLD))
808 stats[lcore_id].trend += FREQ_UP_TREND1_ACC;
810 if (likely(stats[lcore_id].trend > FREQ_UP_THRESHOLD)) {
811 stats[lcore_id].trend = 0;
819 * force polling thread sleep until one-shot rx interrupt triggers
828 sleep_until_rx_interrupt(int num, int lcore)
831 * we want to track when we are woken up by traffic so that we can go
832 * back to sleep again without log spamming. Avoid cache line sharing
833 * to prevent threads stepping on each others' toes.
837 } __rte_cache_aligned status[RTE_MAX_LCORE];
838 struct rte_epoll_event event[num];
844 if (status[lcore].wakeup) {
845 RTE_LOG(INFO, L3FWD_POWER,
846 "lcore %u sleeps until interrupt triggers\n",
850 n = rte_epoll_wait(RTE_EPOLL_PER_THREAD, event, num, 10);
851 for (i = 0; i < n; i++) {
852 data = event[i].epdata.data;
853 port_id = ((uintptr_t)data) >> CHAR_BIT;
854 queue_id = ((uintptr_t)data) &
855 RTE_LEN2MASK(CHAR_BIT, uint8_t);
856 RTE_LOG(INFO, L3FWD_POWER,
857 "lcore %u is waked up from rx interrupt on"
858 " port %d queue %d\n",
859 rte_lcore_id(), port_id, queue_id);
861 status[lcore].wakeup = n != 0;
866 static void turn_on_off_intr(struct lcore_conf *qconf, bool on)
869 struct lcore_rx_queue *rx_queue;
873 for (i = 0; i < qconf->n_rx_queue; ++i) {
874 rx_queue = &(qconf->rx_queue_list[i]);
875 port_id = rx_queue->port_id;
876 queue_id = rx_queue->queue_id;
878 rte_spinlock_lock(&(locks[port_id]));
880 rte_eth_dev_rx_intr_enable(port_id, queue_id);
882 rte_eth_dev_rx_intr_disable(port_id, queue_id);
883 rte_spinlock_unlock(&(locks[port_id]));
887 static int event_register(struct lcore_conf *qconf)
889 struct lcore_rx_queue *rx_queue;
896 for (i = 0; i < qconf->n_rx_queue; ++i) {
897 rx_queue = &(qconf->rx_queue_list[i]);
898 portid = rx_queue->port_id;
899 queueid = rx_queue->queue_id;
900 data = portid << CHAR_BIT | queueid;
902 ret = rte_eth_dev_rx_intr_ctl_q(portid, queueid,
903 RTE_EPOLL_PER_THREAD,
905 (void *)((uintptr_t)data));
913 /* main processing loop */
914 static int main_intr_loop(__rte_unused void *dummy)
916 struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
917 unsigned int lcore_id;
918 uint64_t prev_tsc, diff_tsc, cur_tsc;
922 struct lcore_conf *qconf;
923 struct lcore_rx_queue *rx_queue;
924 uint32_t lcore_rx_idle_count = 0;
925 uint32_t lcore_idle_hint = 0;
928 const uint64_t drain_tsc = (rte_get_tsc_hz() + US_PER_S - 1) /
929 US_PER_S * BURST_TX_DRAIN_US;
933 lcore_id = rte_lcore_id();
934 qconf = &lcore_conf[lcore_id];
936 if (qconf->n_rx_queue == 0) {
937 RTE_LOG(INFO, L3FWD_POWER, "lcore %u has nothing to do\n",
942 RTE_LOG(INFO, L3FWD_POWER, "entering main interrupt loop on lcore %u\n",
945 for (i = 0; i < qconf->n_rx_queue; i++) {
946 portid = qconf->rx_queue_list[i].port_id;
947 queueid = qconf->rx_queue_list[i].queue_id;
948 RTE_LOG(INFO, L3FWD_POWER,
949 " -- lcoreid=%u portid=%u rxqueueid=%hhu\n",
950 lcore_id, portid, queueid);
953 /* add into event wait list */
954 if (event_register(qconf) == 0)
957 RTE_LOG(INFO, L3FWD_POWER, "RX interrupt won't enable.\n");
960 stats[lcore_id].nb_iteration_looped++;
962 cur_tsc = rte_rdtsc();
965 * TX burst queue drain
967 diff_tsc = cur_tsc - prev_tsc;
968 if (unlikely(diff_tsc > drain_tsc)) {
969 for (i = 0; i < qconf->n_tx_port; ++i) {
970 portid = qconf->tx_port_id[i];
971 rte_eth_tx_buffer_flush(portid,
972 qconf->tx_queue_id[portid],
973 qconf->tx_buffer[portid]);
980 * Read packet from RX queues
982 lcore_rx_idle_count = 0;
983 for (i = 0; i < qconf->n_rx_queue; ++i) {
984 rx_queue = &(qconf->rx_queue_list[i]);
985 rx_queue->idle_hint = 0;
986 portid = rx_queue->port_id;
987 queueid = rx_queue->queue_id;
989 nb_rx = rte_eth_rx_burst(portid, queueid, pkts_burst,
992 stats[lcore_id].nb_rx_processed += nb_rx;
993 if (unlikely(nb_rx == 0)) {
995 * no packet received from rx queue, try to
996 * sleep for a while forcing CPU enter deeper
999 rx_queue->zero_rx_packet_count++;
1001 if (rx_queue->zero_rx_packet_count <=
1002 MIN_ZERO_POLL_COUNT)
1005 rx_queue->idle_hint = power_idle_heuristic(
1006 rx_queue->zero_rx_packet_count);
1007 lcore_rx_idle_count++;
1009 rx_queue->zero_rx_packet_count = 0;
1012 /* Prefetch first packets */
1013 for (j = 0; j < PREFETCH_OFFSET && j < nb_rx; j++) {
1014 rte_prefetch0(rte_pktmbuf_mtod(
1015 pkts_burst[j], void *));
1018 /* Prefetch and forward already prefetched packets */
1019 for (j = 0; j < (nb_rx - PREFETCH_OFFSET); j++) {
1020 rte_prefetch0(rte_pktmbuf_mtod(
1021 pkts_burst[j + PREFETCH_OFFSET],
1023 l3fwd_simple_forward(
1024 pkts_burst[j], portid, qconf);
1027 /* Forward remaining prefetched packets */
1028 for (; j < nb_rx; j++) {
1029 l3fwd_simple_forward(
1030 pkts_burst[j], portid, qconf);
1034 if (unlikely(lcore_rx_idle_count == qconf->n_rx_queue)) {
1036 * All Rx queues empty in recent consecutive polls,
1037 * sleep in a conservative manner, meaning sleep as
1041 lcore_idle_hint = qconf->rx_queue_list[0].idle_hint;
1042 i < qconf->n_rx_queue; ++i) {
1043 rx_queue = &(qconf->rx_queue_list[i]);
1044 if (rx_queue->idle_hint < lcore_idle_hint)
1045 lcore_idle_hint = rx_queue->idle_hint;
1048 if (lcore_idle_hint < SUSPEND_THRESHOLD)
1050 * execute "pause" instruction to avoid context
1051 * switch which generally take hundred of
1052 * microseconds for short sleep.
1054 rte_delay_us(lcore_idle_hint);
1056 /* suspend until rx interrupt triggers */
1058 turn_on_off_intr(qconf, 1);
1059 sleep_until_rx_interrupt(
1062 turn_on_off_intr(qconf, 0);
1064 * start receiving packets immediately
1066 if (likely(!is_done()))
1070 stats[lcore_id].sleep_time += lcore_idle_hint;
1077 /* main processing loop */
1079 main_telemetry_loop(__rte_unused void *dummy)
1081 struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
1082 unsigned int lcore_id;
1083 uint64_t prev_tsc, diff_tsc, cur_tsc, prev_tel_tsc;
1087 struct lcore_conf *qconf;
1088 struct lcore_rx_queue *rx_queue;
1089 uint64_t ep_nep[2] = {0}, fp_nfp[2] = {0};
1090 uint64_t poll_count;
1093 const uint64_t drain_tsc = (rte_get_tsc_hz() + US_PER_S - 1) /
1094 US_PER_S * BURST_TX_DRAIN_US;
1100 lcore_id = rte_lcore_id();
1101 qconf = &lcore_conf[lcore_id];
1103 if (qconf->n_rx_queue == 0) {
1104 RTE_LOG(INFO, L3FWD_POWER, "lcore %u has nothing to do\n",
1109 RTE_LOG(INFO, L3FWD_POWER, "entering main telemetry loop on lcore %u\n",
1112 for (i = 0; i < qconf->n_rx_queue; i++) {
1113 portid = qconf->rx_queue_list[i].port_id;
1114 queueid = qconf->rx_queue_list[i].queue_id;
1115 RTE_LOG(INFO, L3FWD_POWER, " -- lcoreid=%u portid=%u "
1116 "rxqueueid=%hhu\n", lcore_id, portid, queueid);
1119 while (!is_done()) {
1121 cur_tsc = rte_rdtsc();
1123 * TX burst queue drain
1125 diff_tsc = cur_tsc - prev_tsc;
1126 if (unlikely(diff_tsc > drain_tsc)) {
1127 for (i = 0; i < qconf->n_tx_port; ++i) {
1128 portid = qconf->tx_port_id[i];
1129 rte_eth_tx_buffer_flush(portid,
1130 qconf->tx_queue_id[portid],
1131 qconf->tx_buffer[portid]);
1137 * Read packet from RX queues
1139 for (i = 0; i < qconf->n_rx_queue; ++i) {
1140 rx_queue = &(qconf->rx_queue_list[i]);
1141 portid = rx_queue->port_id;
1142 queueid = rx_queue->queue_id;
1144 nb_rx = rte_eth_rx_burst(portid, queueid, pkts_burst,
1146 ep_nep[nb_rx == 0]++;
1147 fp_nfp[nb_rx == MAX_PKT_BURST]++;
1149 if (unlikely(nb_rx == 0))
1152 /* Prefetch first packets */
1153 for (j = 0; j < PREFETCH_OFFSET && j < nb_rx; j++) {
1154 rte_prefetch0(rte_pktmbuf_mtod(
1155 pkts_burst[j], void *));
1158 /* Prefetch and forward already prefetched packets */
1159 for (j = 0; j < (nb_rx - PREFETCH_OFFSET); j++) {
1160 rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[
1161 j + PREFETCH_OFFSET], void *));
1162 l3fwd_simple_forward(pkts_burst[j], portid,
1166 /* Forward remaining prefetched packets */
1167 for (; j < nb_rx; j++) {
1168 l3fwd_simple_forward(pkts_burst[j], portid,
1172 if (unlikely(poll_count >= DEFAULT_COUNT)) {
1173 diff_tsc = cur_tsc - prev_tel_tsc;
1174 if (diff_tsc >= MAX_CYCLES) {
1176 } else if (diff_tsc > MIN_CYCLES &&
1177 diff_tsc < MAX_CYCLES) {
1178 br = (diff_tsc * 100) / MAX_CYCLES;
1183 prev_tel_tsc = cur_tsc;
1184 /* update stats for telemetry */
1185 rte_spinlock_lock(&stats[lcore_id].telemetry_lock);
1186 stats[lcore_id].ep_nep[0] = ep_nep[0];
1187 stats[lcore_id].ep_nep[1] = ep_nep[1];
1188 stats[lcore_id].fp_nfp[0] = fp_nfp[0];
1189 stats[lcore_id].fp_nfp[1] = fp_nfp[1];
1190 stats[lcore_id].br = br;
1191 rte_spinlock_unlock(&stats[lcore_id].telemetry_lock);
1197 /* main processing loop */
1199 main_empty_poll_loop(__rte_unused void *dummy)
1201 struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
1202 unsigned int lcore_id;
1203 uint64_t prev_tsc, diff_tsc, cur_tsc;
1207 struct lcore_conf *qconf;
1208 struct lcore_rx_queue *rx_queue;
1210 const uint64_t drain_tsc =
1211 (rte_get_tsc_hz() + US_PER_S - 1) /
1212 US_PER_S * BURST_TX_DRAIN_US;
1216 lcore_id = rte_lcore_id();
1217 qconf = &lcore_conf[lcore_id];
1219 if (qconf->n_rx_queue == 0) {
1220 RTE_LOG(INFO, L3FWD_POWER, "lcore %u has nothing to do\n",
1225 for (i = 0; i < qconf->n_rx_queue; i++) {
1226 portid = qconf->rx_queue_list[i].port_id;
1227 queueid = qconf->rx_queue_list[i].queue_id;
1228 RTE_LOG(INFO, L3FWD_POWER, " -- lcoreid=%u portid=%u "
1229 "rxqueueid=%hhu\n", lcore_id, portid, queueid);
1232 while (!is_done()) {
1233 stats[lcore_id].nb_iteration_looped++;
1235 cur_tsc = rte_rdtsc();
1237 * TX burst queue drain
1239 diff_tsc = cur_tsc - prev_tsc;
1240 if (unlikely(diff_tsc > drain_tsc)) {
1241 for (i = 0; i < qconf->n_tx_port; ++i) {
1242 portid = qconf->tx_port_id[i];
1243 rte_eth_tx_buffer_flush(portid,
1244 qconf->tx_queue_id[portid],
1245 qconf->tx_buffer[portid]);
1251 * Read packet from RX queues
1253 for (i = 0; i < qconf->n_rx_queue; ++i) {
1254 rx_queue = &(qconf->rx_queue_list[i]);
1255 rx_queue->idle_hint = 0;
1256 portid = rx_queue->port_id;
1257 queueid = rx_queue->queue_id;
1259 nb_rx = rte_eth_rx_burst(portid, queueid, pkts_burst,
1262 stats[lcore_id].nb_rx_processed += nb_rx;
1266 rte_power_empty_poll_stat_update(lcore_id);
1270 rte_power_poll_stat_update(lcore_id, nb_rx);
1274 /* Prefetch first packets */
1275 for (j = 0; j < PREFETCH_OFFSET && j < nb_rx; j++) {
1276 rte_prefetch0(rte_pktmbuf_mtod(
1277 pkts_burst[j], void *));
1280 /* Prefetch and forward already prefetched packets */
1281 for (j = 0; j < (nb_rx - PREFETCH_OFFSET); j++) {
1282 rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[
1283 j + PREFETCH_OFFSET],
1285 l3fwd_simple_forward(pkts_burst[j], portid,
1289 /* Forward remaining prefetched packets */
1290 for (; j < nb_rx; j++) {
1291 l3fwd_simple_forward(pkts_burst[j], portid,
1301 /* main processing loop */
1303 main_legacy_loop(__rte_unused void *dummy)
1305 struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
1307 uint64_t prev_tsc, diff_tsc, cur_tsc, tim_res_tsc, hz;
1308 uint64_t prev_tsc_power = 0, cur_tsc_power, diff_tsc_power;
1312 struct lcore_conf *qconf;
1313 struct lcore_rx_queue *rx_queue;
1314 enum freq_scale_hint_t lcore_scaleup_hint;
1315 uint32_t lcore_rx_idle_count = 0;
1316 uint32_t lcore_idle_hint = 0;
1319 const uint64_t drain_tsc = (rte_get_tsc_hz() + US_PER_S - 1) / US_PER_S * BURST_TX_DRAIN_US;
1322 hz = rte_get_timer_hz();
1323 tim_res_tsc = hz/TIMER_NUMBER_PER_SECOND;
1325 lcore_id = rte_lcore_id();
1326 qconf = &lcore_conf[lcore_id];
1328 if (qconf->n_rx_queue == 0) {
1329 RTE_LOG(INFO, L3FWD_POWER, "lcore %u has nothing to do\n", lcore_id);
1333 RTE_LOG(INFO, L3FWD_POWER, "entering main loop on lcore %u\n", lcore_id);
1335 for (i = 0; i < qconf->n_rx_queue; i++) {
1336 portid = qconf->rx_queue_list[i].port_id;
1337 queueid = qconf->rx_queue_list[i].queue_id;
1338 RTE_LOG(INFO, L3FWD_POWER, " -- lcoreid=%u portid=%u "
1339 "rxqueueid=%hhu\n", lcore_id, portid, queueid);
1342 /* add into event wait list */
1343 if (event_register(qconf) == 0)
1346 RTE_LOG(INFO, L3FWD_POWER, "RX interrupt won't enable.\n");
1348 while (!is_done()) {
1349 stats[lcore_id].nb_iteration_looped++;
1351 cur_tsc = rte_rdtsc();
1352 cur_tsc_power = cur_tsc;
1355 * TX burst queue drain
1357 diff_tsc = cur_tsc - prev_tsc;
1358 if (unlikely(diff_tsc > drain_tsc)) {
1359 for (i = 0; i < qconf->n_tx_port; ++i) {
1360 portid = qconf->tx_port_id[i];
1361 rte_eth_tx_buffer_flush(portid,
1362 qconf->tx_queue_id[portid],
1363 qconf->tx_buffer[portid]);
1368 diff_tsc_power = cur_tsc_power - prev_tsc_power;
1369 if (diff_tsc_power > tim_res_tsc) {
1371 prev_tsc_power = cur_tsc_power;
1376 * Read packet from RX queues
1378 lcore_scaleup_hint = FREQ_CURRENT;
1379 lcore_rx_idle_count = 0;
1380 for (i = 0; i < qconf->n_rx_queue; ++i) {
1381 rx_queue = &(qconf->rx_queue_list[i]);
1382 rx_queue->idle_hint = 0;
1383 portid = rx_queue->port_id;
1384 queueid = rx_queue->queue_id;
1386 nb_rx = rte_eth_rx_burst(portid, queueid, pkts_burst,
1389 stats[lcore_id].nb_rx_processed += nb_rx;
1390 if (unlikely(nb_rx == 0)) {
1392 * no packet received from rx queue, try to
1393 * sleep for a while forcing CPU enter deeper
1396 rx_queue->zero_rx_packet_count++;
1398 if (rx_queue->zero_rx_packet_count <=
1399 MIN_ZERO_POLL_COUNT)
1402 rx_queue->idle_hint = power_idle_heuristic(\
1403 rx_queue->zero_rx_packet_count);
1404 lcore_rx_idle_count++;
1406 rx_queue->zero_rx_packet_count = 0;
1409 * do not scale up frequency immediately as
1410 * user to kernel space communication is costly
1411 * which might impact packet I/O for received
1414 rx_queue->freq_up_hint =
1415 power_freq_scaleup_heuristic(lcore_id,
1419 /* Prefetch first packets */
1420 for (j = 0; j < PREFETCH_OFFSET && j < nb_rx; j++) {
1421 rte_prefetch0(rte_pktmbuf_mtod(
1422 pkts_burst[j], void *));
1425 /* Prefetch and forward already prefetched packets */
1426 for (j = 0; j < (nb_rx - PREFETCH_OFFSET); j++) {
1427 rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[
1428 j + PREFETCH_OFFSET], void *));
1429 l3fwd_simple_forward(pkts_burst[j], portid,
1433 /* Forward remaining prefetched packets */
1434 for (; j < nb_rx; j++) {
1435 l3fwd_simple_forward(pkts_burst[j], portid,
1440 if (likely(lcore_rx_idle_count != qconf->n_rx_queue)) {
1441 for (i = 1, lcore_scaleup_hint =
1442 qconf->rx_queue_list[0].freq_up_hint;
1443 i < qconf->n_rx_queue; ++i) {
1444 rx_queue = &(qconf->rx_queue_list[i]);
1445 if (rx_queue->freq_up_hint >
1447 lcore_scaleup_hint =
1448 rx_queue->freq_up_hint;
1451 if (lcore_scaleup_hint == FREQ_HIGHEST) {
1452 if (rte_power_freq_max)
1453 rte_power_freq_max(lcore_id);
1454 } else if (lcore_scaleup_hint == FREQ_HIGHER) {
1455 if (rte_power_freq_up)
1456 rte_power_freq_up(lcore_id);
1460 * All Rx queues empty in recent consecutive polls,
1461 * sleep in a conservative manner, meaning sleep as
1464 for (i = 1, lcore_idle_hint =
1465 qconf->rx_queue_list[0].idle_hint;
1466 i < qconf->n_rx_queue; ++i) {
1467 rx_queue = &(qconf->rx_queue_list[i]);
1468 if (rx_queue->idle_hint < lcore_idle_hint)
1469 lcore_idle_hint = rx_queue->idle_hint;
1472 if (lcore_idle_hint < SUSPEND_THRESHOLD)
1474 * execute "pause" instruction to avoid context
1475 * switch which generally take hundred of
1476 * microseconds for short sleep.
1478 rte_delay_us(lcore_idle_hint);
1480 /* suspend until rx interrupt triggers */
1482 turn_on_off_intr(qconf, 1);
1483 sleep_until_rx_interrupt(
1486 turn_on_off_intr(qconf, 0);
1488 * start receiving packets immediately
1490 if (likely(!is_done()))
1494 stats[lcore_id].sleep_time += lcore_idle_hint;
1502 check_lcore_params(void)
1504 uint8_t queue, lcore;
1508 for (i = 0; i < nb_lcore_params; ++i) {
1509 queue = lcore_params[i].queue_id;
1510 if (queue >= MAX_RX_QUEUE_PER_PORT) {
1511 printf("invalid queue number: %hhu\n", queue);
1514 lcore = lcore_params[i].lcore_id;
1515 if (!rte_lcore_is_enabled(lcore)) {
1516 printf("error: lcore %hhu is not enabled in lcore "
1520 if ((socketid = rte_lcore_to_socket_id(lcore) != 0) &&
1522 printf("warning: lcore %hhu is on socket %d with numa "
1523 "off\n", lcore, socketid);
1525 if (app_mode == APP_MODE_TELEMETRY && lcore == rte_lcore_id()) {
1526 printf("cannot enable main core %d in config for telemetry mode\n",
1535 check_port_config(void)
1540 for (i = 0; i < nb_lcore_params; ++i) {
1541 portid = lcore_params[i].port_id;
1542 if ((enabled_port_mask & (1 << portid)) == 0) {
1543 printf("port %u is not enabled in port mask\n",
1547 if (!rte_eth_dev_is_valid_port(portid)) {
1548 printf("port %u is not present on the board\n",
1557 get_port_n_rx_queues(const uint16_t port)
1562 for (i = 0; i < nb_lcore_params; ++i) {
1563 if (lcore_params[i].port_id == port &&
1564 lcore_params[i].queue_id > queue)
1565 queue = lcore_params[i].queue_id;
1567 return (uint8_t)(++queue);
1571 init_lcore_rx_queues(void)
1573 uint16_t i, nb_rx_queue;
1576 for (i = 0; i < nb_lcore_params; ++i) {
1577 lcore = lcore_params[i].lcore_id;
1578 nb_rx_queue = lcore_conf[lcore].n_rx_queue;
1579 if (nb_rx_queue >= MAX_RX_QUEUE_PER_LCORE) {
1580 printf("error: too many queues (%u) for lcore: %u\n",
1581 (unsigned)nb_rx_queue + 1, (unsigned)lcore);
1584 lcore_conf[lcore].rx_queue_list[nb_rx_queue].port_id =
1585 lcore_params[i].port_id;
1586 lcore_conf[lcore].rx_queue_list[nb_rx_queue].queue_id =
1587 lcore_params[i].queue_id;
1588 lcore_conf[lcore].n_rx_queue++;
1596 print_usage(const char *prgname)
1598 printf ("%s [EAL options] -- -p PORTMASK -P"
1599 " [--config (port,queue,lcore)[,(port,queue,lcore]]"
1600 " [--high-perf-cores CORELIST"
1601 " [--perf-config (port,queue,hi_perf,lcore_index)[,(port,queue,hi_perf,lcore_index]]"
1602 " [--enable-jumbo [--max-pkt-len PKTLEN]]\n"
1603 " -p PORTMASK: hexadecimal bitmask of ports to configure\n"
1604 " -P : enable promiscuous mode\n"
1605 " --config (port,queue,lcore): rx queues configuration\n"
1606 " --high-perf-cores CORELIST: list of high performance cores\n"
1607 " --perf-config: similar as config, cores specified as indices"
1608 " for bins containing high or regular performance cores\n"
1609 " --no-numa: optional, disable numa awareness\n"
1610 " --enable-jumbo: enable jumbo frame"
1611 " which max packet len is PKTLEN in decimal (64-9600)\n"
1612 " --parse-ptype: parse packet type by software\n"
1613 " --legacy: use legacy interrupt-based scaling\n"
1614 " --empty-poll: enable empty poll detection"
1615 " follow (training_flag, high_threshold, med_threshold)\n"
1616 " --telemetry: enable telemetry mode, to update"
1617 " empty polls, full polls, and core busyness to telemetry\n"
1618 " --interrupt-only: enable interrupt-only mode\n"
1619 " --pmd-mgmt MODE: enable PMD power management mode. "
1620 "Currently supported modes: monitor, pause, scale\n",
1624 static int parse_max_pkt_len(const char *pktlen)
1629 /* parse decimal string */
1630 len = strtoul(pktlen, &end, 10);
1631 if ((pktlen[0] == '\0') || (end == NULL) || (*end != '\0'))
1641 parse_portmask(const char *portmask)
1646 /* parse hexadecimal string */
1647 pm = strtoul(portmask, &end, 16);
1648 if ((portmask[0] == '\0') || (end == NULL) || (*end != '\0'))
1655 parse_config(const char *q_arg)
1658 const char *p, *p0 = q_arg;
1666 unsigned long int_fld[_NUM_FLD];
1667 char *str_fld[_NUM_FLD];
1671 nb_lcore_params = 0;
1673 while ((p = strchr(p0,'(')) != NULL) {
1675 if((p0 = strchr(p,')')) == NULL)
1679 if(size >= sizeof(s))
1682 snprintf(s, sizeof(s), "%.*s", size, p);
1683 if (rte_strsplit(s, sizeof(s), str_fld, _NUM_FLD, ',') !=
1686 for (i = 0; i < _NUM_FLD; i++){
1688 int_fld[i] = strtoul(str_fld[i], &end, 0);
1689 if (errno != 0 || end == str_fld[i] || int_fld[i] >
1693 if (nb_lcore_params >= MAX_LCORE_PARAMS) {
1694 printf("exceeded max number of lcore params: %hu\n",
1698 lcore_params_array[nb_lcore_params].port_id =
1699 (uint8_t)int_fld[FLD_PORT];
1700 lcore_params_array[nb_lcore_params].queue_id =
1701 (uint8_t)int_fld[FLD_QUEUE];
1702 lcore_params_array[nb_lcore_params].lcore_id =
1703 (uint8_t)int_fld[FLD_LCORE];
1706 lcore_params = lcore_params_array;
1712 parse_pmd_mgmt_config(const char *name)
1714 #define PMD_MGMT_MONITOR "monitor"
1715 #define PMD_MGMT_PAUSE "pause"
1716 #define PMD_MGMT_SCALE "scale"
1718 if (strncmp(PMD_MGMT_MONITOR, name, sizeof(PMD_MGMT_MONITOR)) == 0) {
1719 pmgmt_type = RTE_POWER_MGMT_TYPE_MONITOR;
1723 if (strncmp(PMD_MGMT_PAUSE, name, sizeof(PMD_MGMT_PAUSE)) == 0) {
1724 pmgmt_type = RTE_POWER_MGMT_TYPE_PAUSE;
1728 if (strncmp(PMD_MGMT_SCALE, name, sizeof(PMD_MGMT_SCALE)) == 0) {
1729 pmgmt_type = RTE_POWER_MGMT_TYPE_SCALE;
1732 /* unknown PMD power management mode */
1737 parse_ep_config(const char *q_arg)
1740 const char *p = q_arg;
1750 ep_med_edpi = EMPTY_POLL_MED_THRESHOLD;
1751 ep_hgh_edpi = EMPTY_POLL_HGH_THRESHOLD;
1753 strlcpy(s, p, sizeof(s));
1755 num_arg = rte_strsplit(s, sizeof(s), str_fld, 3, ',');
1757 empty_poll_train = false;
1764 training_flag = strtoul(str_fld[0], &end, 0);
1765 med_edpi = strtoul(str_fld[1], &end, 0);
1766 hgh_edpi = strtoul(str_fld[2], &end, 0);
1768 if (training_flag == 1)
1769 empty_poll_train = true;
1772 ep_med_edpi = med_edpi;
1775 ep_hgh_edpi = hgh_edpi;
1785 #define CMD_LINE_OPT_PARSE_PTYPE "parse-ptype"
1786 #define CMD_LINE_OPT_LEGACY "legacy"
1787 #define CMD_LINE_OPT_EMPTY_POLL "empty-poll"
1788 #define CMD_LINE_OPT_INTERRUPT_ONLY "interrupt-only"
1789 #define CMD_LINE_OPT_TELEMETRY "telemetry"
1790 #define CMD_LINE_OPT_PMD_MGMT "pmd-mgmt"
1792 /* Parse the argument given in the command line of the application */
1794 parse_args(int argc, char **argv)
1800 char *prgname = argv[0];
1801 static struct option lgopts[] = {
1802 {"config", 1, 0, 0},
1803 {"perf-config", 1, 0, 0},
1804 {"high-perf-cores", 1, 0, 0},
1805 {"no-numa", 0, 0, 0},
1806 {"enable-jumbo", 0, 0, 0},
1807 {CMD_LINE_OPT_EMPTY_POLL, 1, 0, 0},
1808 {CMD_LINE_OPT_PARSE_PTYPE, 0, 0, 0},
1809 {CMD_LINE_OPT_LEGACY, 0, 0, 0},
1810 {CMD_LINE_OPT_TELEMETRY, 0, 0, 0},
1811 {CMD_LINE_OPT_INTERRUPT_ONLY, 0, 0, 0},
1812 {CMD_LINE_OPT_PMD_MGMT, 1, 0, 0},
1818 while ((opt = getopt_long(argc, argvopt, "p:l:m:h:P",
1819 lgopts, &option_index)) != EOF) {
1824 enabled_port_mask = parse_portmask(optarg);
1825 if (enabled_port_mask == 0) {
1826 printf("invalid portmask\n");
1827 print_usage(prgname);
1832 printf("Promiscuous mode selected\n");
1836 limit = parse_max_pkt_len(optarg);
1837 freq_tlb[LOW] = limit;
1840 limit = parse_max_pkt_len(optarg);
1841 freq_tlb[MED] = limit;
1844 limit = parse_max_pkt_len(optarg);
1845 freq_tlb[HGH] = limit;
1849 if (!strncmp(lgopts[option_index].name, "config", 6)) {
1850 ret = parse_config(optarg);
1852 printf("invalid config\n");
1853 print_usage(prgname);
1858 if (!strncmp(lgopts[option_index].name,
1859 "perf-config", 11)) {
1860 ret = parse_perf_config(optarg);
1862 printf("invalid perf-config\n");
1863 print_usage(prgname);
1868 if (!strncmp(lgopts[option_index].name,
1869 "high-perf-cores", 15)) {
1870 ret = parse_perf_core_list(optarg);
1872 printf("invalid high-perf-cores\n");
1873 print_usage(prgname);
1878 if (!strncmp(lgopts[option_index].name,
1880 printf("numa is disabled \n");
1884 if (!strncmp(lgopts[option_index].name,
1885 CMD_LINE_OPT_LEGACY,
1886 sizeof(CMD_LINE_OPT_LEGACY))) {
1887 if (app_mode != APP_MODE_DEFAULT) {
1888 printf(" legacy mode is mutually exclusive with other modes\n");
1891 app_mode = APP_MODE_LEGACY;
1892 printf("legacy mode is enabled\n");
1895 if (!strncmp(lgopts[option_index].name,
1896 CMD_LINE_OPT_EMPTY_POLL, 10)) {
1897 if (app_mode != APP_MODE_DEFAULT) {
1898 printf(" empty-poll mode is mutually exclusive with other modes\n");
1901 app_mode = APP_MODE_EMPTY_POLL;
1902 ret = parse_ep_config(optarg);
1905 printf("invalid empty poll config\n");
1906 print_usage(prgname);
1909 printf("empty-poll is enabled\n");
1912 if (!strncmp(lgopts[option_index].name,
1913 CMD_LINE_OPT_TELEMETRY,
1914 sizeof(CMD_LINE_OPT_TELEMETRY))) {
1915 if (app_mode != APP_MODE_DEFAULT) {
1916 printf(" telemetry mode is mutually exclusive with other modes\n");
1919 app_mode = APP_MODE_TELEMETRY;
1920 printf("telemetry mode is enabled\n");
1923 if (!strncmp(lgopts[option_index].name,
1924 CMD_LINE_OPT_PMD_MGMT,
1925 sizeof(CMD_LINE_OPT_PMD_MGMT))) {
1926 if (app_mode != APP_MODE_DEFAULT) {
1927 printf(" power mgmt mode is mutually exclusive with other modes\n");
1930 if (parse_pmd_mgmt_config(optarg) < 0) {
1931 printf(" Invalid PMD power management mode: %s\n",
1935 app_mode = APP_MODE_PMD_MGMT;
1936 printf("PMD power mgmt mode is enabled\n");
1938 if (!strncmp(lgopts[option_index].name,
1939 CMD_LINE_OPT_INTERRUPT_ONLY,
1940 sizeof(CMD_LINE_OPT_INTERRUPT_ONLY))) {
1941 if (app_mode != APP_MODE_DEFAULT) {
1942 printf(" interrupt-only mode is mutually exclusive with other modes\n");
1945 app_mode = APP_MODE_INTERRUPT;
1946 printf("interrupt-only mode is enabled\n");
1949 if (!strncmp(lgopts[option_index].name,
1950 "enable-jumbo", 12)) {
1951 struct option lenopts =
1952 {"max-pkt-len", required_argument, \
1955 printf("jumbo frame is enabled \n");
1956 port_conf.rxmode.offloads |=
1957 DEV_RX_OFFLOAD_JUMBO_FRAME;
1958 port_conf.txmode.offloads |=
1959 DEV_TX_OFFLOAD_MULTI_SEGS;
1962 * if no max-pkt-len set, use the default value
1965 if (0 == getopt_long(argc, argvopt, "",
1966 &lenopts, &option_index)) {
1967 ret = parse_max_pkt_len(optarg);
1969 (ret > MAX_JUMBO_PKT_LEN)){
1970 printf("invalid packet "
1972 print_usage(prgname);
1975 port_conf.rxmode.max_rx_pkt_len = ret;
1977 printf("set jumbo frame "
1978 "max packet length to %u\n",
1979 (unsigned int)port_conf.rxmode.max_rx_pkt_len);
1982 if (!strncmp(lgopts[option_index].name,
1983 CMD_LINE_OPT_PARSE_PTYPE,
1984 sizeof(CMD_LINE_OPT_PARSE_PTYPE))) {
1985 printf("soft parse-ptype is enabled\n");
1992 print_usage(prgname);
1998 argv[optind-1] = prgname;
2001 optind = 1; /* reset getopt lib */
2006 print_ethaddr(const char *name, const struct rte_ether_addr *eth_addr)
2008 char buf[RTE_ETHER_ADDR_FMT_SIZE];
2009 rte_ether_format_addr(buf, RTE_ETHER_ADDR_FMT_SIZE, eth_addr);
2010 printf("%s%s", name, buf);
2013 #if (APP_LOOKUP_METHOD == APP_LOOKUP_EXACT_MATCH)
2015 setup_hash(int socketid)
2017 struct rte_hash_parameters ipv4_l3fwd_hash_params = {
2019 .entries = L3FWD_HASH_ENTRIES,
2020 .key_len = sizeof(struct ipv4_5tuple),
2021 .hash_func = DEFAULT_HASH_FUNC,
2022 .hash_func_init_val = 0,
2025 struct rte_hash_parameters ipv6_l3fwd_hash_params = {
2027 .entries = L3FWD_HASH_ENTRIES,
2028 .key_len = sizeof(struct ipv6_5tuple),
2029 .hash_func = DEFAULT_HASH_FUNC,
2030 .hash_func_init_val = 0,
2037 /* create ipv4 hash */
2038 snprintf(s, sizeof(s), "ipv4_l3fwd_hash_%d", socketid);
2039 ipv4_l3fwd_hash_params.name = s;
2040 ipv4_l3fwd_hash_params.socket_id = socketid;
2041 ipv4_l3fwd_lookup_struct[socketid] =
2042 rte_hash_create(&ipv4_l3fwd_hash_params);
2043 if (ipv4_l3fwd_lookup_struct[socketid] == NULL)
2044 rte_exit(EXIT_FAILURE, "Unable to create the l3fwd hash on "
2045 "socket %d\n", socketid);
2047 /* create ipv6 hash */
2048 snprintf(s, sizeof(s), "ipv6_l3fwd_hash_%d", socketid);
2049 ipv6_l3fwd_hash_params.name = s;
2050 ipv6_l3fwd_hash_params.socket_id = socketid;
2051 ipv6_l3fwd_lookup_struct[socketid] =
2052 rte_hash_create(&ipv6_l3fwd_hash_params);
2053 if (ipv6_l3fwd_lookup_struct[socketid] == NULL)
2054 rte_exit(EXIT_FAILURE, "Unable to create the l3fwd hash on "
2055 "socket %d\n", socketid);
2058 /* populate the ipv4 hash */
2059 for (i = 0; i < RTE_DIM(ipv4_l3fwd_route_array); i++) {
2060 ret = rte_hash_add_key (ipv4_l3fwd_lookup_struct[socketid],
2061 (void *) &ipv4_l3fwd_route_array[i].key);
2063 rte_exit(EXIT_FAILURE, "Unable to add entry %u to the"
2064 "l3fwd hash on socket %d\n", i, socketid);
2066 ipv4_l3fwd_out_if[ret] = ipv4_l3fwd_route_array[i].if_out;
2067 printf("Hash: Adding key\n");
2068 print_ipv4_key(ipv4_l3fwd_route_array[i].key);
2071 /* populate the ipv6 hash */
2072 for (i = 0; i < RTE_DIM(ipv6_l3fwd_route_array); i++) {
2073 ret = rte_hash_add_key (ipv6_l3fwd_lookup_struct[socketid],
2074 (void *) &ipv6_l3fwd_route_array[i].key);
2076 rte_exit(EXIT_FAILURE, "Unable to add entry %u to the"
2077 "l3fwd hash on socket %d\n", i, socketid);
2079 ipv6_l3fwd_out_if[ret] = ipv6_l3fwd_route_array[i].if_out;
2080 printf("Hash: Adding key\n");
2081 print_ipv6_key(ipv6_l3fwd_route_array[i].key);
2086 #if (APP_LOOKUP_METHOD == APP_LOOKUP_LPM)
2088 setup_lpm(int socketid)
2094 /* create the LPM table */
2095 struct rte_lpm_config lpm_ipv4_config;
2097 lpm_ipv4_config.max_rules = IPV4_L3FWD_LPM_MAX_RULES;
2098 lpm_ipv4_config.number_tbl8s = 256;
2099 lpm_ipv4_config.flags = 0;
2101 snprintf(s, sizeof(s), "IPV4_L3FWD_LPM_%d", socketid);
2102 ipv4_l3fwd_lookup_struct[socketid] =
2103 rte_lpm_create(s, socketid, &lpm_ipv4_config);
2104 if (ipv4_l3fwd_lookup_struct[socketid] == NULL)
2105 rte_exit(EXIT_FAILURE, "Unable to create the l3fwd LPM table"
2106 " on socket %d\n", socketid);
2108 /* populate the LPM table */
2109 for (i = 0; i < RTE_DIM(ipv4_l3fwd_route_array); i++) {
2110 ret = rte_lpm_add(ipv4_l3fwd_lookup_struct[socketid],
2111 ipv4_l3fwd_route_array[i].ip,
2112 ipv4_l3fwd_route_array[i].depth,
2113 ipv4_l3fwd_route_array[i].if_out);
2116 rte_exit(EXIT_FAILURE, "Unable to add entry %u to the "
2117 "l3fwd LPM table on socket %d\n",
2121 printf("LPM: Adding route 0x%08x / %d (%d)\n",
2122 (unsigned)ipv4_l3fwd_route_array[i].ip,
2123 ipv4_l3fwd_route_array[i].depth,
2124 ipv4_l3fwd_route_array[i].if_out);
2130 init_mem(unsigned nb_mbuf)
2132 struct lcore_conf *qconf;
2137 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
2138 if (rte_lcore_is_enabled(lcore_id) == 0)
2142 socketid = rte_lcore_to_socket_id(lcore_id);
2146 if (socketid >= NB_SOCKETS) {
2147 rte_exit(EXIT_FAILURE, "Socket %d of lcore %u is "
2148 "out of range %d\n", socketid,
2149 lcore_id, NB_SOCKETS);
2151 if (pktmbuf_pool[socketid] == NULL) {
2152 snprintf(s, sizeof(s), "mbuf_pool_%d", socketid);
2153 pktmbuf_pool[socketid] =
2154 rte_pktmbuf_pool_create(s, nb_mbuf,
2155 MEMPOOL_CACHE_SIZE, 0,
2156 RTE_MBUF_DEFAULT_BUF_SIZE,
2158 if (pktmbuf_pool[socketid] == NULL)
2159 rte_exit(EXIT_FAILURE,
2160 "Cannot init mbuf pool on socket %d\n",
2163 printf("Allocated mbuf pool on socket %d\n",
2166 #if (APP_LOOKUP_METHOD == APP_LOOKUP_LPM)
2167 setup_lpm(socketid);
2169 setup_hash(socketid);
2172 qconf = &lcore_conf[lcore_id];
2173 qconf->ipv4_lookup_struct = ipv4_l3fwd_lookup_struct[socketid];
2174 #if (APP_LOOKUP_METHOD == APP_LOOKUP_EXACT_MATCH)
2175 qconf->ipv6_lookup_struct = ipv6_l3fwd_lookup_struct[socketid];
2181 /* Check the link status of all ports in up to 9s, and print them finally */
2183 check_all_ports_link_status(uint32_t port_mask)
2185 #define CHECK_INTERVAL 100 /* 100ms */
2186 #define MAX_CHECK_TIME 90 /* 9s (90 * 100ms) in total */
2187 uint8_t count, all_ports_up, print_flag = 0;
2189 struct rte_eth_link link;
2191 char link_status_text[RTE_ETH_LINK_MAX_STR_LEN];
2193 printf("\nChecking link status");
2195 for (count = 0; count <= MAX_CHECK_TIME; count++) {
2197 RTE_ETH_FOREACH_DEV(portid) {
2198 if ((port_mask & (1 << portid)) == 0)
2200 memset(&link, 0, sizeof(link));
2201 ret = rte_eth_link_get_nowait(portid, &link);
2204 if (print_flag == 1)
2205 printf("Port %u link get failed: %s\n",
2206 portid, rte_strerror(-ret));
2209 /* print link status if flag set */
2210 if (print_flag == 1) {
2211 rte_eth_link_to_str(link_status_text,
2212 sizeof(link_status_text), &link);
2213 printf("Port %d %s\n", portid,
2217 /* clear all_ports_up flag if any link down */
2218 if (link.link_status == ETH_LINK_DOWN) {
2223 /* after finally printing all link status, get out */
2224 if (print_flag == 1)
2227 if (all_ports_up == 0) {
2230 rte_delay_ms(CHECK_INTERVAL);
2233 /* set the print_flag if all ports up or timeout */
2234 if (all_ports_up == 1 || count == (MAX_CHECK_TIME - 1)) {
2241 static int check_ptype(uint16_t portid)
2244 int ptype_l3_ipv4 = 0;
2245 #if (APP_LOOKUP_METHOD == APP_LOOKUP_EXACT_MATCH)
2246 int ptype_l3_ipv6 = 0;
2248 uint32_t ptype_mask = RTE_PTYPE_L3_MASK;
2250 ret = rte_eth_dev_get_supported_ptypes(portid, ptype_mask, NULL, 0);
2254 uint32_t ptypes[ret];
2256 ret = rte_eth_dev_get_supported_ptypes(portid, ptype_mask, ptypes, ret);
2257 for (i = 0; i < ret; ++i) {
2258 if (ptypes[i] & RTE_PTYPE_L3_IPV4)
2260 #if (APP_LOOKUP_METHOD == APP_LOOKUP_EXACT_MATCH)
2261 if (ptypes[i] & RTE_PTYPE_L3_IPV6)
2266 if (ptype_l3_ipv4 == 0)
2267 printf("port %d cannot parse RTE_PTYPE_L3_IPV4\n", portid);
2269 #if (APP_LOOKUP_METHOD == APP_LOOKUP_EXACT_MATCH)
2270 if (ptype_l3_ipv6 == 0)
2271 printf("port %d cannot parse RTE_PTYPE_L3_IPV6\n", portid);
2274 #if (APP_LOOKUP_METHOD == APP_LOOKUP_LPM)
2276 #else /* APP_LOOKUP_EXACT_MATCH */
2277 if (ptype_l3_ipv4 && ptype_l3_ipv6)
2286 init_power_library(void)
2288 enum power_management_env env;
2289 unsigned int lcore_id;
2292 RTE_LCORE_FOREACH(lcore_id) {
2293 /* init power management library */
2294 ret = rte_power_init(lcore_id);
2297 "Library initialization failed on core %u\n",
2301 /* we're not supporting the VM channel mode */
2302 env = rte_power_get_env();
2303 if (env != PM_ENV_ACPI_CPUFREQ &&
2304 env != PM_ENV_PSTATE_CPUFREQ) {
2306 "Only ACPI and PSTATE mode are supported\n");
2314 deinit_power_library(void)
2316 unsigned int lcore_id;
2319 RTE_LCORE_FOREACH(lcore_id) {
2320 /* deinit power management library */
2321 ret = rte_power_exit(lcore_id);
2324 "Library deinitialization failed on core %u\n",
2333 get_current_stat_values(uint64_t *values)
2335 unsigned int lcore_id = rte_lcore_id();
2336 struct lcore_conf *qconf;
2337 uint64_t app_eps = 0, app_fps = 0, app_br = 0;
2340 RTE_LCORE_FOREACH_WORKER(lcore_id) {
2341 qconf = &lcore_conf[lcore_id];
2342 if (qconf->n_rx_queue == 0)
2345 rte_spinlock_lock(&stats[lcore_id].telemetry_lock);
2346 app_eps += stats[lcore_id].ep_nep[1];
2347 app_fps += stats[lcore_id].fp_nfp[1];
2348 app_br += stats[lcore_id].br;
2349 rte_spinlock_unlock(&stats[lcore_id].telemetry_lock);
2353 values[0] = app_eps/count;
2354 values[1] = app_fps/count;
2355 values[2] = app_br/count;
2357 memset(values, 0, sizeof(uint64_t) * NUM_TELSTATS);
2362 update_telemetry(__rte_unused struct rte_timer *tim,
2363 __rte_unused void *arg)
2366 uint64_t values[NUM_TELSTATS] = {0};
2368 get_current_stat_values(values);
2369 ret = rte_metrics_update_values(RTE_METRICS_GLOBAL, telstats_index,
2370 values, RTE_DIM(values));
2372 RTE_LOG(WARNING, POWER, "failed to update metrcis\n");
2376 handle_app_stats(const char *cmd __rte_unused,
2377 const char *params __rte_unused,
2378 struct rte_tel_data *d)
2380 uint64_t values[NUM_TELSTATS] = {0};
2383 rte_tel_data_start_dict(d);
2384 get_current_stat_values(values);
2385 for (i = 0; i < NUM_TELSTATS; i++)
2386 rte_tel_data_add_dict_u64(d, telstats_strings[i].name,
2392 telemetry_setup_timer(void)
2394 int lcore_id = rte_lcore_id();
2395 uint64_t hz = rte_get_timer_hz();
2398 ticks = hz / TELEMETRY_INTERVALS_PER_SEC;
2399 rte_timer_reset_sync(&telemetry_timer,
2407 empty_poll_setup_timer(void)
2409 int lcore_id = rte_lcore_id();
2410 uint64_t hz = rte_get_timer_hz();
2412 struct ep_params *ep_ptr = ep_params;
2414 ep_ptr->interval_ticks = hz / INTERVALS_PER_SECOND;
2416 rte_timer_reset_sync(&ep_ptr->timer0,
2417 ep_ptr->interval_ticks,
2420 rte_empty_poll_detection,
2425 launch_timer(unsigned int lcore_id)
2427 int64_t prev_tsc = 0, cur_tsc, diff_tsc, cycles_10ms;
2429 RTE_SET_USED(lcore_id);
2432 if (rte_get_main_lcore() != lcore_id) {
2433 rte_panic("timer on lcore:%d which is not main core:%d\n",
2435 rte_get_main_lcore());
2438 RTE_LOG(INFO, POWER, "Bring up the Timer\n");
2440 if (app_mode == APP_MODE_EMPTY_POLL)
2441 empty_poll_setup_timer();
2443 telemetry_setup_timer();
2445 cycles_10ms = rte_get_timer_hz() / 100;
2447 while (!is_done()) {
2448 cur_tsc = rte_rdtsc();
2449 diff_tsc = cur_tsc - prev_tsc;
2450 if (diff_tsc > cycles_10ms) {
2453 cycles_10ms = rte_get_timer_hz() / 100;
2457 RTE_LOG(INFO, POWER, "Timer_subsystem is done\n");
2463 autodetect_mode(void)
2465 RTE_LOG(NOTICE, L3FWD_POWER, "Operating mode not specified, probing frequency scaling support...\n");
2468 * Empty poll and telemetry modes have to be specifically requested to
2469 * be enabled, but we can auto-detect between interrupt mode with or
2470 * without frequency scaling. Both ACPI and pstate can be used.
2472 if (rte_power_check_env_supported(PM_ENV_ACPI_CPUFREQ))
2473 return APP_MODE_LEGACY;
2474 if (rte_power_check_env_supported(PM_ENV_PSTATE_CPUFREQ))
2475 return APP_MODE_LEGACY;
2477 RTE_LOG(NOTICE, L3FWD_POWER, "Frequency scaling not supported, selecting interrupt-only mode\n");
2479 return APP_MODE_INTERRUPT;
2483 mode_to_str(enum appmode mode)
2486 case APP_MODE_LEGACY:
2488 case APP_MODE_EMPTY_POLL:
2489 return "empty poll";
2490 case APP_MODE_TELEMETRY:
2492 case APP_MODE_INTERRUPT:
2493 return "interrupt-only";
2494 case APP_MODE_PMD_MGMT:
2502 main(int argc, char **argv)
2504 struct lcore_conf *qconf;
2505 struct rte_eth_dev_info dev_info;
2506 struct rte_eth_txconf *txconf;
2512 uint32_t n_tx_queue, nb_lcores;
2513 uint32_t dev_rxq_num, dev_txq_num;
2514 uint8_t nb_rx_queue, queue, socketid;
2516 const char *ptr_strings[NUM_TELSTATS];
2518 /* catch SIGINT and restore cpufreq governor to ondemand */
2519 signal(SIGINT, signal_exit_now);
2522 ret = rte_eal_init(argc, argv);
2524 rte_exit(EXIT_FAILURE, "Invalid EAL parameters\n");
2528 /* init RTE timer library to be used late */
2529 rte_timer_subsystem_init();
2531 /* parse application arguments (after the EAL ones) */
2532 ret = parse_args(argc, argv);
2534 rte_exit(EXIT_FAILURE, "Invalid L3FWD parameters\n");
2536 if (app_mode == APP_MODE_DEFAULT)
2537 app_mode = autodetect_mode();
2539 RTE_LOG(INFO, L3FWD_POWER, "Selected operation mode: %s\n",
2540 mode_to_str(app_mode));
2542 /* only legacy and empty poll mode rely on power library */
2543 if ((app_mode == APP_MODE_LEGACY || app_mode == APP_MODE_EMPTY_POLL) &&
2544 init_power_library())
2545 rte_exit(EXIT_FAILURE, "init_power_library failed\n");
2547 if (update_lcore_params() < 0)
2548 rte_exit(EXIT_FAILURE, "update_lcore_params failed\n");
2550 if (check_lcore_params() < 0)
2551 rte_exit(EXIT_FAILURE, "check_lcore_params failed\n");
2553 ret = init_lcore_rx_queues();
2555 rte_exit(EXIT_FAILURE, "init_lcore_rx_queues failed\n");
2557 nb_ports = rte_eth_dev_count_avail();
2559 if (check_port_config() < 0)
2560 rte_exit(EXIT_FAILURE, "check_port_config failed\n");
2562 nb_lcores = rte_lcore_count();
2564 /* initialize all ports */
2565 RTE_ETH_FOREACH_DEV(portid) {
2566 struct rte_eth_conf local_port_conf = port_conf;
2567 /* not all app modes need interrupts */
2568 bool need_intr = app_mode == APP_MODE_LEGACY ||
2569 app_mode == APP_MODE_INTERRUPT;
2571 /* skip ports that are not enabled */
2572 if ((enabled_port_mask & (1 << portid)) == 0) {
2573 printf("\nSkipping disabled port %d\n", portid);
2578 printf("Initializing port %d ... ", portid );
2581 ret = rte_eth_dev_info_get(portid, &dev_info);
2583 rte_exit(EXIT_FAILURE,
2584 "Error during getting device (port %u) info: %s\n",
2585 portid, strerror(-ret));
2587 dev_rxq_num = dev_info.max_rx_queues;
2588 dev_txq_num = dev_info.max_tx_queues;
2590 nb_rx_queue = get_port_n_rx_queues(portid);
2591 if (nb_rx_queue > dev_rxq_num)
2592 rte_exit(EXIT_FAILURE,
2593 "Cannot configure not existed rxq: "
2594 "port=%d\n", portid);
2596 n_tx_queue = nb_lcores;
2597 if (n_tx_queue > dev_txq_num)
2598 n_tx_queue = dev_txq_num;
2599 printf("Creating queues: nb_rxq=%d nb_txq=%u... ",
2600 nb_rx_queue, (unsigned)n_tx_queue );
2601 /* If number of Rx queue is 0, no need to enable Rx interrupt */
2602 if (nb_rx_queue == 0)
2606 local_port_conf.intr_conf.rxq = 1;
2608 ret = rte_eth_dev_info_get(portid, &dev_info);
2610 rte_exit(EXIT_FAILURE,
2611 "Error during getting device (port %u) info: %s\n",
2612 portid, strerror(-ret));
2614 if (dev_info.tx_offload_capa & DEV_TX_OFFLOAD_MBUF_FAST_FREE)
2615 local_port_conf.txmode.offloads |=
2616 DEV_TX_OFFLOAD_MBUF_FAST_FREE;
2618 local_port_conf.rx_adv_conf.rss_conf.rss_hf &=
2619 dev_info.flow_type_rss_offloads;
2620 if (local_port_conf.rx_adv_conf.rss_conf.rss_hf !=
2621 port_conf.rx_adv_conf.rss_conf.rss_hf) {
2622 printf("Port %u modified RSS hash function based on hardware support,"
2623 "requested:%#"PRIx64" configured:%#"PRIx64"\n",
2625 port_conf.rx_adv_conf.rss_conf.rss_hf,
2626 local_port_conf.rx_adv_conf.rss_conf.rss_hf);
2629 ret = rte_eth_dev_configure(portid, nb_rx_queue,
2630 (uint16_t)n_tx_queue, &local_port_conf);
2632 rte_exit(EXIT_FAILURE, "Cannot configure device: "
2633 "err=%d, port=%d\n", ret, portid);
2635 ret = rte_eth_dev_adjust_nb_rx_tx_desc(portid, &nb_rxd,
2638 rte_exit(EXIT_FAILURE,
2639 "Cannot adjust number of descriptors: err=%d, port=%d\n",
2642 ret = rte_eth_macaddr_get(portid, &ports_eth_addr[portid]);
2644 rte_exit(EXIT_FAILURE,
2645 "Cannot get MAC address: err=%d, port=%d\n",
2648 print_ethaddr(" Address:", &ports_eth_addr[portid]);
2652 ret = init_mem(NB_MBUF);
2654 rte_exit(EXIT_FAILURE, "init_mem failed\n");
2656 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
2657 if (rte_lcore_is_enabled(lcore_id) == 0)
2660 /* Initialize TX buffers */
2661 qconf = &lcore_conf[lcore_id];
2662 qconf->tx_buffer[portid] = rte_zmalloc_socket("tx_buffer",
2663 RTE_ETH_TX_BUFFER_SIZE(MAX_PKT_BURST), 0,
2664 rte_eth_dev_socket_id(portid));
2665 if (qconf->tx_buffer[portid] == NULL)
2666 rte_exit(EXIT_FAILURE, "Can't allocate tx buffer for port %u\n",
2669 rte_eth_tx_buffer_init(qconf->tx_buffer[portid], MAX_PKT_BURST);
2672 /* init one TX queue per couple (lcore,port) */
2674 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
2675 if (rte_lcore_is_enabled(lcore_id) == 0)
2678 if (queueid >= dev_txq_num)
2683 (uint8_t)rte_lcore_to_socket_id(lcore_id);
2687 printf("txq=%u,%d,%d ", lcore_id, queueid, socketid);
2690 txconf = &dev_info.default_txconf;
2691 txconf->offloads = local_port_conf.txmode.offloads;
2692 ret = rte_eth_tx_queue_setup(portid, queueid, nb_txd,
2695 rte_exit(EXIT_FAILURE,
2696 "rte_eth_tx_queue_setup: err=%d, "
2697 "port=%d\n", ret, portid);
2699 qconf = &lcore_conf[lcore_id];
2700 qconf->tx_queue_id[portid] = queueid;
2703 qconf->tx_port_id[qconf->n_tx_port] = portid;
2709 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
2710 if (rte_lcore_is_enabled(lcore_id) == 0)
2713 if (app_mode == APP_MODE_LEGACY) {
2714 /* init timer structures for each enabled lcore */
2715 rte_timer_init(&power_timers[lcore_id]);
2716 hz = rte_get_timer_hz();
2717 rte_timer_reset(&power_timers[lcore_id],
2718 hz/TIMER_NUMBER_PER_SECOND,
2720 power_timer_cb, NULL);
2722 qconf = &lcore_conf[lcore_id];
2723 printf("\nInitializing rx queues on lcore %u ... ", lcore_id );
2726 /* PMD power management mode can only do 1 queue per core */
2727 if (app_mode == APP_MODE_PMD_MGMT && qconf->n_rx_queue > 1) {
2728 rte_exit(EXIT_FAILURE,
2729 "In PMD power management mode, only one queue per lcore is allowed\n");
2732 /* init RX queues */
2733 for(queue = 0; queue < qconf->n_rx_queue; ++queue) {
2734 struct rte_eth_rxconf rxq_conf;
2736 portid = qconf->rx_queue_list[queue].port_id;
2737 queueid = qconf->rx_queue_list[queue].queue_id;
2741 (uint8_t)rte_lcore_to_socket_id(lcore_id);
2745 printf("rxq=%d,%d,%d ", portid, queueid, socketid);
2748 ret = rte_eth_dev_info_get(portid, &dev_info);
2750 rte_exit(EXIT_FAILURE,
2751 "Error during getting device (port %u) info: %s\n",
2752 portid, strerror(-ret));
2754 rxq_conf = dev_info.default_rxconf;
2755 rxq_conf.offloads = port_conf.rxmode.offloads;
2756 ret = rte_eth_rx_queue_setup(portid, queueid, nb_rxd,
2757 socketid, &rxq_conf,
2758 pktmbuf_pool[socketid]);
2760 rte_exit(EXIT_FAILURE,
2761 "rte_eth_rx_queue_setup: err=%d, "
2762 "port=%d\n", ret, portid);
2765 if (add_cb_parse_ptype(portid, queueid) < 0)
2766 rte_exit(EXIT_FAILURE,
2767 "Fail to add ptype cb\n");
2770 if (app_mode == APP_MODE_PMD_MGMT) {
2771 ret = rte_power_ethdev_pmgmt_queue_enable(
2772 lcore_id, portid, queueid,
2775 rte_exit(EXIT_FAILURE,
2776 "rte_power_ethdev_pmgmt_queue_enable: err=%d, port=%d\n",
2785 RTE_ETH_FOREACH_DEV(portid) {
2786 if ((enabled_port_mask & (1 << portid)) == 0) {
2790 ret = rte_eth_dev_start(portid);
2792 rte_exit(EXIT_FAILURE, "rte_eth_dev_start: err=%d, "
2793 "port=%d\n", ret, portid);
2795 * If enabled, put device in promiscuous mode.
2796 * This allows IO forwarding mode to forward packets
2797 * to itself through 2 cross-connected ports of the
2800 if (promiscuous_on) {
2801 ret = rte_eth_promiscuous_enable(portid);
2803 rte_exit(EXIT_FAILURE,
2804 "rte_eth_promiscuous_enable: err=%s, port=%u\n",
2805 rte_strerror(-ret), portid);
2807 /* initialize spinlock for each port */
2808 rte_spinlock_init(&(locks[portid]));
2811 if (!check_ptype(portid))
2812 rte_exit(EXIT_FAILURE,
2813 "PMD can not provide needed ptypes\n");
2816 check_all_ports_link_status(enabled_port_mask);
2818 if (app_mode == APP_MODE_EMPTY_POLL) {
2820 if (empty_poll_train) {
2821 policy.state = TRAINING;
2823 policy.state = MED_NORMAL;
2824 policy.med_base_edpi = ep_med_edpi;
2825 policy.hgh_base_edpi = ep_hgh_edpi;
2828 ret = rte_power_empty_poll_stat_init(&ep_params,
2832 rte_exit(EXIT_FAILURE, "empty poll init failed");
2836 /* launch per-lcore init on every lcore */
2837 if (app_mode == APP_MODE_LEGACY) {
2838 rte_eal_mp_remote_launch(main_legacy_loop, NULL, CALL_MAIN);
2839 } else if (app_mode == APP_MODE_EMPTY_POLL) {
2840 empty_poll_stop = false;
2841 rte_eal_mp_remote_launch(main_empty_poll_loop, NULL,
2843 } else if (app_mode == APP_MODE_TELEMETRY) {
2846 /* Init metrics library */
2847 rte_metrics_init(rte_socket_id());
2848 /** Register stats with metrics library */
2849 for (i = 0; i < NUM_TELSTATS; i++)
2850 ptr_strings[i] = telstats_strings[i].name;
2852 ret = rte_metrics_reg_names(ptr_strings, NUM_TELSTATS);
2854 telstats_index = ret;
2856 rte_exit(EXIT_FAILURE, "failed to register metrics names");
2858 RTE_LCORE_FOREACH_WORKER(lcore_id) {
2859 rte_spinlock_init(&stats[lcore_id].telemetry_lock);
2861 rte_timer_init(&telemetry_timer);
2862 rte_telemetry_register_cmd("/l3fwd-power/stats",
2864 "Returns global power stats. Parameters: None");
2865 rte_eal_mp_remote_launch(main_telemetry_loop, NULL,
2867 } else if (app_mode == APP_MODE_INTERRUPT) {
2868 rte_eal_mp_remote_launch(main_intr_loop, NULL, CALL_MAIN);
2869 } else if (app_mode == APP_MODE_PMD_MGMT) {
2870 /* reuse telemetry loop for PMD power management mode */
2871 rte_eal_mp_remote_launch(main_telemetry_loop, NULL, CALL_MAIN);
2874 if (app_mode == APP_MODE_EMPTY_POLL || app_mode == APP_MODE_TELEMETRY)
2875 launch_timer(rte_lcore_id());
2877 RTE_LCORE_FOREACH_WORKER(lcore_id) {
2878 if (rte_eal_wait_lcore(lcore_id) < 0)
2882 if (app_mode == APP_MODE_PMD_MGMT) {
2883 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
2884 if (rte_lcore_is_enabled(lcore_id) == 0)
2886 qconf = &lcore_conf[lcore_id];
2887 for (queue = 0; queue < qconf->n_rx_queue; ++queue) {
2888 portid = qconf->rx_queue_list[queue].port_id;
2889 queueid = qconf->rx_queue_list[queue].queue_id;
2891 rte_power_ethdev_pmgmt_queue_disable(lcore_id,
2897 RTE_ETH_FOREACH_DEV(portid)
2899 if ((enabled_port_mask & (1 << portid)) == 0)
2902 ret = rte_eth_dev_stop(portid);
2904 RTE_LOG(ERR, L3FWD_POWER, "rte_eth_dev_stop: err=%d, port=%u\n",
2907 rte_eth_dev_close(portid);
2910 if (app_mode == APP_MODE_EMPTY_POLL)
2911 rte_power_empty_poll_stat_free();
2913 if ((app_mode == APP_MODE_LEGACY || app_mode == APP_MODE_EMPTY_POLL) &&
2914 deinit_power_library())
2915 rte_exit(EXIT_FAILURE, "deinit_power_library failed\n");
2917 if (rte_eal_cleanup() < 0)
2918 RTE_LOG(ERR, L3FWD_POWER, "EAL cleanup failed\n");