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_cycles.h>
28 #include <rte_prefetch.h>
29 #include <rte_lcore.h>
30 #include <rte_per_lcore.h>
31 #include <rte_branch_prediction.h>
32 #include <rte_interrupts.h>
33 #include <rte_random.h>
34 #include <rte_debug.h>
35 #include <rte_ether.h>
36 #include <rte_ethdev.h>
37 #include <rte_mempool.h>
42 #include <rte_string_fns.h>
43 #include <rte_timer.h>
44 #include <rte_power.h>
45 #include <rte_spinlock.h>
46 #include <rte_power_empty_poll.h>
47 #include <rte_metrics.h>
48 #include <rte_telemetry.h>
49 #include <rte_power_pmd_mgmt.h>
51 #include "perf_core.h"
54 #define RTE_LOGTYPE_L3FWD_POWER RTE_LOGTYPE_USER1
56 #define MAX_PKT_BURST 32
58 #define MIN_ZERO_POLL_COUNT 10
61 #define TIMER_NUMBER_PER_SECOND 10
63 #define INTERVALS_PER_SECOND 100
65 #define SCALING_PERIOD (1000000/TIMER_NUMBER_PER_SECOND)
66 #define SCALING_DOWN_TIME_RATIO_THRESHOLD 0.25
68 #define APP_LOOKUP_EXACT_MATCH 0
69 #define APP_LOOKUP_LPM 1
70 #define DO_RFC_1812_CHECKS
72 #ifndef APP_LOOKUP_METHOD
73 #define APP_LOOKUP_METHOD APP_LOOKUP_LPM
76 #if (APP_LOOKUP_METHOD == APP_LOOKUP_EXACT_MATCH)
78 #elif (APP_LOOKUP_METHOD == APP_LOOKUP_LPM)
81 #error "APP_LOOKUP_METHOD set to incorrect value"
85 #define IPv6_BYTES_FMT "%02x%02x:%02x%02x:%02x%02x:%02x%02x:"\
86 "%02x%02x:%02x%02x:%02x%02x:%02x%02x"
87 #define IPv6_BYTES(addr) \
88 addr[0], addr[1], addr[2], addr[3], \
89 addr[4], addr[5], addr[6], addr[7], \
90 addr[8], addr[9], addr[10], addr[11],\
91 addr[12], addr[13],addr[14], addr[15]
94 #define MAX_JUMBO_PKT_LEN 9600
96 #define IPV6_ADDR_LEN 16
98 #define MEMPOOL_CACHE_SIZE 256
101 * This expression is used to calculate the number of mbufs needed depending on
102 * user input, taking into account memory for rx and tx hardware rings, cache
103 * per lcore and mtable per port per lcore. RTE_MAX is used to ensure that
104 * NB_MBUF never goes below a minimum value of 8192.
107 #define NB_MBUF RTE_MAX ( \
108 (nb_ports*nb_rx_queue*nb_rxd + \
109 nb_ports*nb_lcores*MAX_PKT_BURST + \
110 nb_ports*n_tx_queue*nb_txd + \
111 nb_lcores*MEMPOOL_CACHE_SIZE), \
114 #define BURST_TX_DRAIN_US 100 /* TX drain every ~100us */
118 /* Configure how many packets ahead to prefetch, when reading packets */
119 #define PREFETCH_OFFSET 3
122 * Configurable number of RX/TX ring descriptors
124 #define RTE_TEST_RX_DESC_DEFAULT 1024
125 #define RTE_TEST_TX_DESC_DEFAULT 1024
128 * These two thresholds were decided on by running the training algorithm on
129 * a 2.5GHz Xeon. These defaults can be overridden by supplying non-zero values
130 * for the med_threshold and high_threshold parameters on the command line.
132 #define EMPTY_POLL_MED_THRESHOLD 350000UL
133 #define EMPTY_POLL_HGH_THRESHOLD 580000UL
135 #define NUM_TELSTATS RTE_DIM(telstats_strings)
137 static uint16_t nb_rxd = RTE_TEST_RX_DESC_DEFAULT;
138 static uint16_t nb_txd = RTE_TEST_TX_DESC_DEFAULT;
140 /* ethernet addresses of ports */
141 static struct rte_ether_addr ports_eth_addr[RTE_MAX_ETHPORTS];
143 /* ethernet addresses of ports */
144 static rte_spinlock_t locks[RTE_MAX_ETHPORTS];
146 /* mask of enabled ports */
147 static uint32_t enabled_port_mask = 0;
148 /* Ports set in promiscuous mode off by default. */
149 static int promiscuous_on = 0;
150 /* NUMA is enabled by default. */
151 static int numa_on = 1;
152 static bool empty_poll_stop;
153 static bool empty_poll_train;
154 volatile bool quit_signal;
155 static struct ep_params *ep_params;
156 static struct ep_policy policy;
157 static long ep_med_edpi, ep_hgh_edpi;
158 /* timer to update telemetry every 500ms */
159 static struct rte_timer telemetry_timer;
161 /* stats index returned by metrics lib */
164 struct telstats_name {
165 char name[RTE_ETH_XSTATS_NAME_SIZE];
168 /* telemetry stats to be reported */
169 const struct telstats_name telstats_strings[] = {
175 /* core busyness in percentage */
182 /* reference poll count to measure core busyness */
183 #define DEFAULT_COUNT 10000
185 * reference CYCLES to be used to
186 * measure core busyness based on poll count
188 #define MIN_CYCLES 1500000ULL
189 #define MAX_CYCLES 22000000ULL
192 #define TELEMETRY_INTERVALS_PER_SEC 2
194 static int parse_ptype; /**< Parse packet type using rx callback, and */
195 /**< disabled by default */
198 APP_MODE_DEFAULT = 0,
206 enum appmode app_mode;
208 static enum rte_power_pmd_mgmt_type pmgmt_type;
209 bool baseline_enabled;
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 = RTE_ETH_MQ_RX_RSS,
254 .offloads = RTE_ETH_RX_OFFLOAD_CHECKSUM,
259 .rss_hf = RTE_ETH_RSS_UDP,
263 .mq_mode = RTE_ETH_MQ_TX_NONE,
267 static uint32_t max_pkt_len;
268 static uint32_t max_empty_polls = 512;
269 static uint32_t pause_duration = 1;
270 static uint32_t scale_freq_min;
271 static uint32_t scale_freq_max;
273 static struct rte_mempool * pktmbuf_pool[NB_SOCKETS];
276 #if (APP_LOOKUP_METHOD == APP_LOOKUP_EXACT_MATCH)
279 #include <rte_hash_crc.h>
280 #define DEFAULT_HASH_FUNC rte_hash_crc
282 #include <rte_jhash.h>
283 #define DEFAULT_HASH_FUNC rte_jhash
295 uint8_t ip_dst[IPV6_ADDR_LEN];
296 uint8_t ip_src[IPV6_ADDR_LEN];
302 struct ipv4_l3fwd_route {
303 struct ipv4_5tuple key;
307 struct ipv6_l3fwd_route {
308 struct ipv6_5tuple key;
312 static struct ipv4_l3fwd_route ipv4_l3fwd_route_array[] = {
313 {{RTE_IPV4(100,10,0,1), RTE_IPV4(200,10,0,1), 101, 11, IPPROTO_TCP}, 0},
314 {{RTE_IPV4(100,20,0,2), RTE_IPV4(200,20,0,2), 102, 12, IPPROTO_TCP}, 1},
315 {{RTE_IPV4(100,30,0,3), RTE_IPV4(200,30,0,3), 103, 13, IPPROTO_TCP}, 2},
316 {{RTE_IPV4(100,40,0,4), RTE_IPV4(200,40,0,4), 104, 14, IPPROTO_TCP}, 3},
319 static struct ipv6_l3fwd_route ipv6_l3fwd_route_array[] = {
322 {0xfe, 0x80, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
323 0x02, 0x1b, 0x21, 0xff, 0xfe, 0x91, 0x38, 0x05},
324 {0xfe, 0x80, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
325 0x02, 0x1e, 0x67, 0xff, 0xfe, 0x0d, 0xb6, 0x0a},
331 typedef struct rte_hash lookup_struct_t;
332 static lookup_struct_t *ipv4_l3fwd_lookup_struct[NB_SOCKETS];
333 static lookup_struct_t *ipv6_l3fwd_lookup_struct[NB_SOCKETS];
335 #define L3FWD_HASH_ENTRIES 1024
337 static uint16_t ipv4_l3fwd_out_if[L3FWD_HASH_ENTRIES] __rte_cache_aligned;
338 static uint16_t ipv6_l3fwd_out_if[L3FWD_HASH_ENTRIES] __rte_cache_aligned;
341 #if (APP_LOOKUP_METHOD == APP_LOOKUP_LPM)
342 struct ipv4_l3fwd_route {
348 static struct ipv4_l3fwd_route ipv4_l3fwd_route_array[] = {
349 {RTE_IPV4(1,1,1,0), 24, 0},
350 {RTE_IPV4(2,1,1,0), 24, 1},
351 {RTE_IPV4(3,1,1,0), 24, 2},
352 {RTE_IPV4(4,1,1,0), 24, 3},
353 {RTE_IPV4(5,1,1,0), 24, 4},
354 {RTE_IPV4(6,1,1,0), 24, 5},
355 {RTE_IPV4(7,1,1,0), 24, 6},
356 {RTE_IPV4(8,1,1,0), 24, 7},
359 #define IPV4_L3FWD_LPM_MAX_RULES 1024
361 typedef struct rte_lpm lookup_struct_t;
362 static lookup_struct_t *ipv4_l3fwd_lookup_struct[NB_SOCKETS];
367 struct lcore_rx_queue rx_queue_list[MAX_RX_QUEUE_PER_LCORE];
369 uint16_t tx_port_id[RTE_MAX_ETHPORTS];
370 uint16_t tx_queue_id[RTE_MAX_ETHPORTS];
371 struct rte_eth_dev_tx_buffer *tx_buffer[RTE_MAX_ETHPORTS];
372 lookup_struct_t * ipv4_lookup_struct;
373 lookup_struct_t * ipv6_lookup_struct;
374 } __rte_cache_aligned;
377 /* total sleep time in ms since last frequency scaling down */
379 /* number of long sleep recently */
380 uint32_t nb_long_sleep;
381 /* freq. scaling up trend */
383 /* total packet processed recently */
384 uint64_t nb_rx_processed;
385 /* total iterations looped recently */
386 uint64_t nb_iteration_looped;
388 * Represents empty and non empty polls
389 * of rte_eth_rx_burst();
390 * ep_nep[0] holds non empty polls
391 * i.e. 0 < nb_rx <= MAX_BURST
392 * ep_nep[1] holds empty polls.
397 * Represents full and empty+partial
398 * polls of rte_eth_rx_burst();
399 * ep_nep[0] holds empty+partial polls.
400 * i.e. 0 <= nb_rx < MAX_BURST
401 * ep_nep[1] holds full polls
402 * i.e. nb_rx == MAX_BURST
406 rte_spinlock_t telemetry_lock;
407 } __rte_cache_aligned;
409 static struct lcore_conf lcore_conf[RTE_MAX_LCORE] __rte_cache_aligned;
410 static struct lcore_stats stats[RTE_MAX_LCORE] __rte_cache_aligned;
411 static struct rte_timer power_timers[RTE_MAX_LCORE];
413 static inline uint32_t power_idle_heuristic(uint32_t zero_rx_packet_count);
414 static inline enum freq_scale_hint_t power_freq_scaleup_heuristic( \
415 unsigned int lcore_id, uint16_t port_id, uint16_t queue_id);
419 * These defaults are using the max frequency index (1), a medium index (9)
420 * and a typical low frequency index (14). These can be adjusted to use
421 * different indexes using the relevant command line parameters.
423 static uint8_t freq_tlb[] = {14, 9, 1};
425 static int is_done(void)
430 /* exit signal handler */
432 signal_exit_now(int sigtype)
435 if (sigtype == SIGINT)
440 /* Frequency scale down timer callback */
442 power_timer_cb(__rte_unused struct rte_timer *tim,
443 __rte_unused void *arg)
446 float sleep_time_ratio;
447 unsigned lcore_id = rte_lcore_id();
449 /* accumulate total execution time in us when callback is invoked */
450 sleep_time_ratio = (float)(stats[lcore_id].sleep_time) /
451 (float)SCALING_PERIOD;
453 * check whether need to scale down frequency a step if it sleep a lot.
455 if (sleep_time_ratio >= SCALING_DOWN_TIME_RATIO_THRESHOLD) {
456 if (rte_power_freq_down)
457 rte_power_freq_down(lcore_id);
459 else if ( (unsigned)(stats[lcore_id].nb_rx_processed /
460 stats[lcore_id].nb_iteration_looped) < MAX_PKT_BURST) {
462 * scale down a step if average packet per iteration less
465 if (rte_power_freq_down)
466 rte_power_freq_down(lcore_id);
470 * initialize another timer according to current frequency to ensure
471 * timer interval is relatively fixed.
473 hz = rte_get_timer_hz();
474 rte_timer_reset(&power_timers[lcore_id], hz/TIMER_NUMBER_PER_SECOND,
475 SINGLE, lcore_id, power_timer_cb, NULL);
477 stats[lcore_id].nb_rx_processed = 0;
478 stats[lcore_id].nb_iteration_looped = 0;
480 stats[lcore_id].sleep_time = 0;
483 /* Enqueue a single packet, and send burst if queue is filled */
485 send_single_packet(struct rte_mbuf *m, uint16_t port)
488 struct lcore_conf *qconf;
490 lcore_id = rte_lcore_id();
491 qconf = &lcore_conf[lcore_id];
493 rte_eth_tx_buffer(port, qconf->tx_queue_id[port],
494 qconf->tx_buffer[port], m);
499 #ifdef DO_RFC_1812_CHECKS
501 is_valid_ipv4_pkt(struct rte_ipv4_hdr *pkt, uint32_t link_len)
503 /* From http://www.rfc-editor.org/rfc/rfc1812.txt section 5.2.2 */
505 * 1. The packet length reported by the Link Layer must be large
506 * enough to hold the minimum length legal IP datagram (20 bytes).
508 if (link_len < sizeof(struct rte_ipv4_hdr))
511 /* 2. The IP checksum must be correct. */
512 /* if this is not checked in H/W, check it. */
513 if ((port_conf.rxmode.offloads & RTE_ETH_RX_OFFLOAD_IPV4_CKSUM) == 0) {
514 uint16_t actual_cksum, expected_cksum;
515 actual_cksum = pkt->hdr_checksum;
516 pkt->hdr_checksum = 0;
517 expected_cksum = rte_ipv4_cksum(pkt);
518 if (actual_cksum != expected_cksum)
523 * 3. The IP version number must be 4. If the version number is not 4
524 * then the packet may be another version of IP, such as IPng or
527 if (((pkt->version_ihl) >> 4) != 4)
530 * 4. The IP header length field must be large enough to hold the
531 * minimum length legal IP datagram (20 bytes = 5 words).
533 if ((pkt->version_ihl & 0xf) < 5)
537 * 5. The IP total length field must be large enough to hold the IP
538 * datagram header, whose length is specified in the IP header length
541 if (rte_cpu_to_be_16(pkt->total_length) < sizeof(struct rte_ipv4_hdr))
548 #if (APP_LOOKUP_METHOD == APP_LOOKUP_EXACT_MATCH)
550 print_ipv4_key(struct ipv4_5tuple key)
552 printf("IP dst = %08x, IP src = %08x, port dst = %d, port src = %d, "
553 "proto = %d\n", (unsigned)key.ip_dst, (unsigned)key.ip_src,
554 key.port_dst, key.port_src, key.proto);
557 print_ipv6_key(struct ipv6_5tuple key)
559 printf( "IP dst = " IPv6_BYTES_FMT ", IP src = " IPv6_BYTES_FMT ", "
560 "port dst = %d, port src = %d, proto = %d\n",
561 IPv6_BYTES(key.ip_dst), IPv6_BYTES(key.ip_src),
562 key.port_dst, key.port_src, key.proto);
565 static inline uint16_t
566 get_ipv4_dst_port(struct rte_ipv4_hdr *ipv4_hdr, uint16_t portid,
567 lookup_struct_t * ipv4_l3fwd_lookup_struct)
569 struct ipv4_5tuple key;
570 struct rte_tcp_hdr *tcp;
571 struct rte_udp_hdr *udp;
574 key.ip_dst = rte_be_to_cpu_32(ipv4_hdr->dst_addr);
575 key.ip_src = rte_be_to_cpu_32(ipv4_hdr->src_addr);
576 key.proto = ipv4_hdr->next_proto_id;
578 switch (ipv4_hdr->next_proto_id) {
580 tcp = (struct rte_tcp_hdr *)((unsigned char *)ipv4_hdr +
581 sizeof(struct rte_ipv4_hdr));
582 key.port_dst = rte_be_to_cpu_16(tcp->dst_port);
583 key.port_src = rte_be_to_cpu_16(tcp->src_port);
587 udp = (struct rte_udp_hdr *)((unsigned char *)ipv4_hdr +
588 sizeof(struct rte_ipv4_hdr));
589 key.port_dst = rte_be_to_cpu_16(udp->dst_port);
590 key.port_src = rte_be_to_cpu_16(udp->src_port);
599 /* Find destination port */
600 ret = rte_hash_lookup(ipv4_l3fwd_lookup_struct, (const void *)&key);
601 return ((ret < 0) ? portid : ipv4_l3fwd_out_if[ret]);
604 static inline uint16_t
605 get_ipv6_dst_port(struct rte_ipv6_hdr *ipv6_hdr, uint16_t portid,
606 lookup_struct_t *ipv6_l3fwd_lookup_struct)
608 struct ipv6_5tuple key;
609 struct rte_tcp_hdr *tcp;
610 struct rte_udp_hdr *udp;
613 memcpy(key.ip_dst, ipv6_hdr->dst_addr, IPV6_ADDR_LEN);
614 memcpy(key.ip_src, ipv6_hdr->src_addr, IPV6_ADDR_LEN);
616 key.proto = ipv6_hdr->proto;
618 switch (ipv6_hdr->proto) {
620 tcp = (struct rte_tcp_hdr *)((unsigned char *) ipv6_hdr +
621 sizeof(struct rte_ipv6_hdr));
622 key.port_dst = rte_be_to_cpu_16(tcp->dst_port);
623 key.port_src = rte_be_to_cpu_16(tcp->src_port);
627 udp = (struct rte_udp_hdr *)((unsigned char *) ipv6_hdr +
628 sizeof(struct rte_ipv6_hdr));
629 key.port_dst = rte_be_to_cpu_16(udp->dst_port);
630 key.port_src = rte_be_to_cpu_16(udp->src_port);
639 /* Find destination port */
640 ret = rte_hash_lookup(ipv6_l3fwd_lookup_struct, (const void *)&key);
641 return ((ret < 0) ? portid : ipv6_l3fwd_out_if[ret]);
645 #if (APP_LOOKUP_METHOD == APP_LOOKUP_LPM)
646 static inline uint16_t
647 get_ipv4_dst_port(struct rte_ipv4_hdr *ipv4_hdr, uint16_t portid,
648 lookup_struct_t *ipv4_l3fwd_lookup_struct)
652 return ((rte_lpm_lookup(ipv4_l3fwd_lookup_struct,
653 rte_be_to_cpu_32(ipv4_hdr->dst_addr), &next_hop) == 0)?
659 parse_ptype_one(struct rte_mbuf *m)
661 struct rte_ether_hdr *eth_hdr;
662 uint32_t packet_type = RTE_PTYPE_UNKNOWN;
665 eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
666 ether_type = eth_hdr->ether_type;
667 if (ether_type == rte_cpu_to_be_16(RTE_ETHER_TYPE_IPV4))
668 packet_type |= RTE_PTYPE_L3_IPV4_EXT_UNKNOWN;
669 else if (ether_type == rte_cpu_to_be_16(RTE_ETHER_TYPE_IPV6))
670 packet_type |= RTE_PTYPE_L3_IPV6_EXT_UNKNOWN;
672 m->packet_type = packet_type;
676 cb_parse_ptype(uint16_t port __rte_unused, uint16_t queue __rte_unused,
677 struct rte_mbuf *pkts[], uint16_t nb_pkts,
678 uint16_t max_pkts __rte_unused,
679 void *user_param __rte_unused)
683 for (i = 0; i < nb_pkts; ++i)
684 parse_ptype_one(pkts[i]);
690 add_cb_parse_ptype(uint16_t portid, uint16_t queueid)
692 printf("Port %d: softly parse packet type info\n", portid);
693 if (rte_eth_add_rx_callback(portid, queueid, cb_parse_ptype, NULL))
696 printf("Failed to add rx callback: port=%d\n", portid);
701 l3fwd_simple_forward(struct rte_mbuf *m, uint16_t portid,
702 struct lcore_conf *qconf)
704 struct rte_ether_hdr *eth_hdr;
705 struct rte_ipv4_hdr *ipv4_hdr;
709 eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
711 if (RTE_ETH_IS_IPV4_HDR(m->packet_type)) {
712 /* Handle IPv4 headers.*/
714 rte_pktmbuf_mtod_offset(m, struct rte_ipv4_hdr *,
715 sizeof(struct rte_ether_hdr));
717 #ifdef DO_RFC_1812_CHECKS
718 /* Check to make sure the packet is valid (RFC1812) */
719 if (is_valid_ipv4_pkt(ipv4_hdr, m->pkt_len) < 0) {
725 dst_port = get_ipv4_dst_port(ipv4_hdr, portid,
726 qconf->ipv4_lookup_struct);
727 if (dst_port >= RTE_MAX_ETHPORTS ||
728 (enabled_port_mask & 1 << dst_port) == 0)
731 /* 02:00:00:00:00:xx */
732 d_addr_bytes = ð_hdr->dst_addr.addr_bytes[0];
733 *((uint64_t *)d_addr_bytes) =
734 0x000000000002 + ((uint64_t)dst_port << 40);
736 #ifdef DO_RFC_1812_CHECKS
737 /* Update time to live and header checksum */
738 --(ipv4_hdr->time_to_live);
739 ++(ipv4_hdr->hdr_checksum);
743 rte_ether_addr_copy(&ports_eth_addr[dst_port],
746 send_single_packet(m, dst_port);
747 } else if (RTE_ETH_IS_IPV6_HDR(m->packet_type)) {
748 /* Handle IPv6 headers.*/
749 #if (APP_LOOKUP_METHOD == APP_LOOKUP_EXACT_MATCH)
750 struct rte_ipv6_hdr *ipv6_hdr;
753 rte_pktmbuf_mtod_offset(m, struct rte_ipv6_hdr *,
754 sizeof(struct rte_ether_hdr));
756 dst_port = get_ipv6_dst_port(ipv6_hdr, portid,
757 qconf->ipv6_lookup_struct);
759 if (dst_port >= RTE_MAX_ETHPORTS ||
760 (enabled_port_mask & 1 << dst_port) == 0)
763 /* 02:00:00:00:00:xx */
764 d_addr_bytes = ð_hdr->dst_addr.addr_bytes[0];
765 *((uint64_t *)d_addr_bytes) =
766 0x000000000002 + ((uint64_t)dst_port << 40);
769 rte_ether_addr_copy(&ports_eth_addr[dst_port],
772 send_single_packet(m, dst_port);
774 /* We don't currently handle IPv6 packets in LPM mode. */
782 #define MINIMUM_SLEEP_TIME 1
783 #define SUSPEND_THRESHOLD 300
785 static inline uint32_t
786 power_idle_heuristic(uint32_t zero_rx_packet_count)
788 /* If zero count is less than 100, sleep 1us */
789 if (zero_rx_packet_count < SUSPEND_THRESHOLD)
790 return MINIMUM_SLEEP_TIME;
791 /* If zero count is less than 1000, sleep 100 us which is the
792 minimum latency switching from C3/C6 to C0
795 return SUSPEND_THRESHOLD;
798 static inline enum freq_scale_hint_t
799 power_freq_scaleup_heuristic(unsigned lcore_id,
803 uint32_t rxq_count = rte_eth_rx_queue_count(port_id, queue_id);
805 * HW Rx queue size is 128 by default, Rx burst read at maximum 32 entries
808 #define FREQ_GEAR1_RX_PACKET_THRESHOLD MAX_PKT_BURST
809 #define FREQ_GEAR2_RX_PACKET_THRESHOLD (MAX_PKT_BURST*2)
810 #define FREQ_GEAR3_RX_PACKET_THRESHOLD (MAX_PKT_BURST*3)
811 #define FREQ_UP_TREND1_ACC 1
812 #define FREQ_UP_TREND2_ACC 100
813 #define FREQ_UP_THRESHOLD 10000
815 if (likely(rxq_count > FREQ_GEAR3_RX_PACKET_THRESHOLD)) {
816 stats[lcore_id].trend = 0;
818 } else if (likely(rxq_count > FREQ_GEAR2_RX_PACKET_THRESHOLD))
819 stats[lcore_id].trend += FREQ_UP_TREND2_ACC;
820 else if (likely(rxq_count > FREQ_GEAR1_RX_PACKET_THRESHOLD))
821 stats[lcore_id].trend += FREQ_UP_TREND1_ACC;
823 if (likely(stats[lcore_id].trend > FREQ_UP_THRESHOLD)) {
824 stats[lcore_id].trend = 0;
832 * force polling thread sleep until one-shot rx interrupt triggers
841 sleep_until_rx_interrupt(int num, int lcore)
844 * we want to track when we are woken up by traffic so that we can go
845 * back to sleep again without log spamming. Avoid cache line sharing
846 * to prevent threads stepping on each others' toes.
850 } __rte_cache_aligned status[RTE_MAX_LCORE];
851 struct rte_epoll_event event[num];
857 if (status[lcore].wakeup) {
858 RTE_LOG(INFO, L3FWD_POWER,
859 "lcore %u sleeps until interrupt triggers\n",
863 n = rte_epoll_wait(RTE_EPOLL_PER_THREAD, event, num, 10);
864 for (i = 0; i < n; i++) {
865 data = event[i].epdata.data;
866 port_id = ((uintptr_t)data) >> CHAR_BIT;
867 queue_id = ((uintptr_t)data) &
868 RTE_LEN2MASK(CHAR_BIT, uint8_t);
869 RTE_LOG(INFO, L3FWD_POWER,
870 "lcore %u is waked up from rx interrupt on"
871 " port %d queue %d\n",
872 rte_lcore_id(), port_id, queue_id);
874 status[lcore].wakeup = n != 0;
879 static void turn_on_off_intr(struct lcore_conf *qconf, bool on)
882 struct lcore_rx_queue *rx_queue;
886 for (i = 0; i < qconf->n_rx_queue; ++i) {
887 rx_queue = &(qconf->rx_queue_list[i]);
888 port_id = rx_queue->port_id;
889 queue_id = rx_queue->queue_id;
891 rte_spinlock_lock(&(locks[port_id]));
893 rte_eth_dev_rx_intr_enable(port_id, queue_id);
895 rte_eth_dev_rx_intr_disable(port_id, queue_id);
896 rte_spinlock_unlock(&(locks[port_id]));
900 static int event_register(struct lcore_conf *qconf)
902 struct lcore_rx_queue *rx_queue;
909 for (i = 0; i < qconf->n_rx_queue; ++i) {
910 rx_queue = &(qconf->rx_queue_list[i]);
911 portid = rx_queue->port_id;
912 queueid = rx_queue->queue_id;
913 data = portid << CHAR_BIT | queueid;
915 ret = rte_eth_dev_rx_intr_ctl_q(portid, queueid,
916 RTE_EPOLL_PER_THREAD,
918 (void *)((uintptr_t)data));
926 /* Main processing loop. 8< */
927 static int main_intr_loop(__rte_unused void *dummy)
929 struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
930 unsigned int lcore_id;
931 uint64_t prev_tsc, diff_tsc, cur_tsc;
935 struct lcore_conf *qconf;
936 struct lcore_rx_queue *rx_queue;
937 uint32_t lcore_rx_idle_count = 0;
938 uint32_t lcore_idle_hint = 0;
941 const uint64_t drain_tsc = (rte_get_tsc_hz() + US_PER_S - 1) /
942 US_PER_S * BURST_TX_DRAIN_US;
946 lcore_id = rte_lcore_id();
947 qconf = &lcore_conf[lcore_id];
949 if (qconf->n_rx_queue == 0) {
950 RTE_LOG(INFO, L3FWD_POWER, "lcore %u has nothing to do\n",
955 RTE_LOG(INFO, L3FWD_POWER, "entering main interrupt loop on lcore %u\n",
958 for (i = 0; i < qconf->n_rx_queue; i++) {
959 portid = qconf->rx_queue_list[i].port_id;
960 queueid = qconf->rx_queue_list[i].queue_id;
961 RTE_LOG(INFO, L3FWD_POWER,
962 " -- lcoreid=%u portid=%u rxqueueid=%hhu\n",
963 lcore_id, portid, queueid);
966 /* add into event wait list */
967 if (event_register(qconf) == 0)
970 RTE_LOG(INFO, L3FWD_POWER, "RX interrupt won't enable.\n");
973 stats[lcore_id].nb_iteration_looped++;
975 cur_tsc = rte_rdtsc();
978 * TX burst queue drain
980 diff_tsc = cur_tsc - prev_tsc;
981 if (unlikely(diff_tsc > drain_tsc)) {
982 for (i = 0; i < qconf->n_tx_port; ++i) {
983 portid = qconf->tx_port_id[i];
984 rte_eth_tx_buffer_flush(portid,
985 qconf->tx_queue_id[portid],
986 qconf->tx_buffer[portid]);
993 * Read packet from RX queues
995 lcore_rx_idle_count = 0;
996 for (i = 0; i < qconf->n_rx_queue; ++i) {
997 rx_queue = &(qconf->rx_queue_list[i]);
998 rx_queue->idle_hint = 0;
999 portid = rx_queue->port_id;
1000 queueid = rx_queue->queue_id;
1002 nb_rx = rte_eth_rx_burst(portid, queueid, pkts_burst,
1005 stats[lcore_id].nb_rx_processed += nb_rx;
1006 if (unlikely(nb_rx == 0)) {
1008 * no packet received from rx queue, try to
1009 * sleep for a while forcing CPU enter deeper
1012 rx_queue->zero_rx_packet_count++;
1014 if (rx_queue->zero_rx_packet_count <=
1015 MIN_ZERO_POLL_COUNT)
1018 rx_queue->idle_hint = power_idle_heuristic(
1019 rx_queue->zero_rx_packet_count);
1020 lcore_rx_idle_count++;
1022 rx_queue->zero_rx_packet_count = 0;
1025 /* Prefetch first packets */
1026 for (j = 0; j < PREFETCH_OFFSET && j < nb_rx; j++) {
1027 rte_prefetch0(rte_pktmbuf_mtod(
1028 pkts_burst[j], void *));
1031 /* Prefetch and forward already prefetched packets */
1032 for (j = 0; j < (nb_rx - PREFETCH_OFFSET); j++) {
1033 rte_prefetch0(rte_pktmbuf_mtod(
1034 pkts_burst[j + PREFETCH_OFFSET],
1036 l3fwd_simple_forward(
1037 pkts_burst[j], portid, qconf);
1040 /* Forward remaining prefetched packets */
1041 for (; j < nb_rx; j++) {
1042 l3fwd_simple_forward(
1043 pkts_burst[j], portid, qconf);
1047 if (unlikely(lcore_rx_idle_count == qconf->n_rx_queue)) {
1049 * All Rx queues empty in recent consecutive polls,
1050 * sleep in a conservative manner, meaning sleep as
1054 lcore_idle_hint = qconf->rx_queue_list[0].idle_hint;
1055 i < qconf->n_rx_queue; ++i) {
1056 rx_queue = &(qconf->rx_queue_list[i]);
1057 if (rx_queue->idle_hint < lcore_idle_hint)
1058 lcore_idle_hint = rx_queue->idle_hint;
1061 if (lcore_idle_hint < SUSPEND_THRESHOLD)
1063 * execute "pause" instruction to avoid context
1064 * switch which generally take hundred of
1065 * microseconds for short sleep.
1067 rte_delay_us(lcore_idle_hint);
1069 /* suspend until rx interrupt triggers */
1071 turn_on_off_intr(qconf, 1);
1072 sleep_until_rx_interrupt(
1075 turn_on_off_intr(qconf, 0);
1077 * start receiving packets immediately
1079 if (likely(!is_done()))
1083 stats[lcore_id].sleep_time += lcore_idle_hint;
1089 /* >8 End of main processing loop. */
1091 /* main processing loop */
1093 main_telemetry_loop(__rte_unused void *dummy)
1095 struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
1096 unsigned int lcore_id;
1097 uint64_t prev_tsc, diff_tsc, cur_tsc, prev_tel_tsc;
1101 struct lcore_conf *qconf;
1102 struct lcore_rx_queue *rx_queue;
1103 uint64_t ep_nep[2] = {0}, fp_nfp[2] = {0};
1104 uint64_t poll_count;
1107 const uint64_t drain_tsc = (rte_get_tsc_hz() + US_PER_S - 1) /
1108 US_PER_S * BURST_TX_DRAIN_US;
1114 lcore_id = rte_lcore_id();
1115 qconf = &lcore_conf[lcore_id];
1117 if (qconf->n_rx_queue == 0) {
1118 RTE_LOG(INFO, L3FWD_POWER, "lcore %u has nothing to do\n",
1123 RTE_LOG(INFO, L3FWD_POWER, "entering main telemetry loop on lcore %u\n",
1126 for (i = 0; i < qconf->n_rx_queue; i++) {
1127 portid = qconf->rx_queue_list[i].port_id;
1128 queueid = qconf->rx_queue_list[i].queue_id;
1129 RTE_LOG(INFO, L3FWD_POWER, " -- lcoreid=%u portid=%u "
1130 "rxqueueid=%hhu\n", lcore_id, portid, queueid);
1133 while (!is_done()) {
1135 cur_tsc = rte_rdtsc();
1137 * TX burst queue drain
1139 diff_tsc = cur_tsc - prev_tsc;
1140 if (unlikely(diff_tsc > drain_tsc)) {
1141 for (i = 0; i < qconf->n_tx_port; ++i) {
1142 portid = qconf->tx_port_id[i];
1143 rte_eth_tx_buffer_flush(portid,
1144 qconf->tx_queue_id[portid],
1145 qconf->tx_buffer[portid]);
1151 * Read packet from RX queues
1153 for (i = 0; i < qconf->n_rx_queue; ++i) {
1154 rx_queue = &(qconf->rx_queue_list[i]);
1155 portid = rx_queue->port_id;
1156 queueid = rx_queue->queue_id;
1158 nb_rx = rte_eth_rx_burst(portid, queueid, pkts_burst,
1160 ep_nep[nb_rx == 0]++;
1161 fp_nfp[nb_rx == MAX_PKT_BURST]++;
1163 if (unlikely(nb_rx == 0))
1166 /* Prefetch first packets */
1167 for (j = 0; j < PREFETCH_OFFSET && j < nb_rx; j++) {
1168 rte_prefetch0(rte_pktmbuf_mtod(
1169 pkts_burst[j], void *));
1172 /* Prefetch and forward already prefetched packets */
1173 for (j = 0; j < (nb_rx - PREFETCH_OFFSET); j++) {
1174 rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[
1175 j + PREFETCH_OFFSET], void *));
1176 l3fwd_simple_forward(pkts_burst[j], portid,
1180 /* Forward remaining prefetched packets */
1181 for (; j < nb_rx; j++) {
1182 l3fwd_simple_forward(pkts_burst[j], portid,
1186 if (unlikely(poll_count >= DEFAULT_COUNT)) {
1187 diff_tsc = cur_tsc - prev_tel_tsc;
1188 if (diff_tsc >= MAX_CYCLES) {
1190 } else if (diff_tsc > MIN_CYCLES &&
1191 diff_tsc < MAX_CYCLES) {
1192 br = (diff_tsc * 100) / MAX_CYCLES;
1197 prev_tel_tsc = cur_tsc;
1198 /* update stats for telemetry */
1199 rte_spinlock_lock(&stats[lcore_id].telemetry_lock);
1200 stats[lcore_id].ep_nep[0] = ep_nep[0];
1201 stats[lcore_id].ep_nep[1] = ep_nep[1];
1202 stats[lcore_id].fp_nfp[0] = fp_nfp[0];
1203 stats[lcore_id].fp_nfp[1] = fp_nfp[1];
1204 stats[lcore_id].br = br;
1205 rte_spinlock_unlock(&stats[lcore_id].telemetry_lock);
1211 /* main processing loop */
1213 main_empty_poll_loop(__rte_unused void *dummy)
1215 struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
1216 unsigned int lcore_id;
1217 uint64_t prev_tsc, diff_tsc, cur_tsc;
1221 struct lcore_conf *qconf;
1222 struct lcore_rx_queue *rx_queue;
1224 const uint64_t drain_tsc =
1225 (rte_get_tsc_hz() + US_PER_S - 1) /
1226 US_PER_S * BURST_TX_DRAIN_US;
1230 lcore_id = rte_lcore_id();
1231 qconf = &lcore_conf[lcore_id];
1233 if (qconf->n_rx_queue == 0) {
1234 RTE_LOG(INFO, L3FWD_POWER, "lcore %u has nothing to do\n",
1239 for (i = 0; i < qconf->n_rx_queue; i++) {
1240 portid = qconf->rx_queue_list[i].port_id;
1241 queueid = qconf->rx_queue_list[i].queue_id;
1242 RTE_LOG(INFO, L3FWD_POWER, " -- lcoreid=%u portid=%u "
1243 "rxqueueid=%hhu\n", lcore_id, portid, queueid);
1246 while (!is_done()) {
1247 stats[lcore_id].nb_iteration_looped++;
1249 cur_tsc = rte_rdtsc();
1251 * TX burst queue drain
1253 diff_tsc = cur_tsc - prev_tsc;
1254 if (unlikely(diff_tsc > drain_tsc)) {
1255 for (i = 0; i < qconf->n_tx_port; ++i) {
1256 portid = qconf->tx_port_id[i];
1257 rte_eth_tx_buffer_flush(portid,
1258 qconf->tx_queue_id[portid],
1259 qconf->tx_buffer[portid]);
1265 * Read packet from RX queues
1267 for (i = 0; i < qconf->n_rx_queue; ++i) {
1268 rx_queue = &(qconf->rx_queue_list[i]);
1269 rx_queue->idle_hint = 0;
1270 portid = rx_queue->port_id;
1271 queueid = rx_queue->queue_id;
1273 nb_rx = rte_eth_rx_burst(portid, queueid, pkts_burst,
1276 stats[lcore_id].nb_rx_processed += nb_rx;
1280 rte_power_empty_poll_stat_update(lcore_id);
1284 rte_power_poll_stat_update(lcore_id, nb_rx);
1288 /* Prefetch first packets */
1289 for (j = 0; j < PREFETCH_OFFSET && j < nb_rx; j++) {
1290 rte_prefetch0(rte_pktmbuf_mtod(
1291 pkts_burst[j], void *));
1294 /* Prefetch and forward already prefetched packets */
1295 for (j = 0; j < (nb_rx - PREFETCH_OFFSET); j++) {
1296 rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[
1297 j + PREFETCH_OFFSET],
1299 l3fwd_simple_forward(pkts_burst[j], portid,
1303 /* Forward remaining prefetched packets */
1304 for (; j < nb_rx; j++) {
1305 l3fwd_simple_forward(pkts_burst[j], portid,
1315 /* main processing loop */
1317 main_legacy_loop(__rte_unused void *dummy)
1319 struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
1321 uint64_t prev_tsc, diff_tsc, cur_tsc, tim_res_tsc, hz;
1322 uint64_t prev_tsc_power = 0, cur_tsc_power, diff_tsc_power;
1326 struct lcore_conf *qconf;
1327 struct lcore_rx_queue *rx_queue;
1328 enum freq_scale_hint_t lcore_scaleup_hint;
1329 uint32_t lcore_rx_idle_count = 0;
1330 uint32_t lcore_idle_hint = 0;
1333 const uint64_t drain_tsc = (rte_get_tsc_hz() + US_PER_S - 1) / US_PER_S * BURST_TX_DRAIN_US;
1336 hz = rte_get_timer_hz();
1337 tim_res_tsc = hz/TIMER_NUMBER_PER_SECOND;
1339 lcore_id = rte_lcore_id();
1340 qconf = &lcore_conf[lcore_id];
1342 if (qconf->n_rx_queue == 0) {
1343 RTE_LOG(INFO, L3FWD_POWER, "lcore %u has nothing to do\n", lcore_id);
1347 RTE_LOG(INFO, L3FWD_POWER, "entering main loop on lcore %u\n", lcore_id);
1349 for (i = 0; i < qconf->n_rx_queue; i++) {
1350 portid = qconf->rx_queue_list[i].port_id;
1351 queueid = qconf->rx_queue_list[i].queue_id;
1352 RTE_LOG(INFO, L3FWD_POWER, " -- lcoreid=%u portid=%u "
1353 "rxqueueid=%hhu\n", lcore_id, portid, queueid);
1356 /* add into event wait list */
1357 if (event_register(qconf) == 0)
1360 RTE_LOG(INFO, L3FWD_POWER, "RX interrupt won't enable.\n");
1362 while (!is_done()) {
1363 stats[lcore_id].nb_iteration_looped++;
1365 cur_tsc = rte_rdtsc();
1366 cur_tsc_power = cur_tsc;
1369 * TX burst queue drain
1371 diff_tsc = cur_tsc - prev_tsc;
1372 if (unlikely(diff_tsc > drain_tsc)) {
1373 for (i = 0; i < qconf->n_tx_port; ++i) {
1374 portid = qconf->tx_port_id[i];
1375 rte_eth_tx_buffer_flush(portid,
1376 qconf->tx_queue_id[portid],
1377 qconf->tx_buffer[portid]);
1382 diff_tsc_power = cur_tsc_power - prev_tsc_power;
1383 if (diff_tsc_power > tim_res_tsc) {
1385 prev_tsc_power = cur_tsc_power;
1390 * Read packet from RX queues
1392 lcore_scaleup_hint = FREQ_CURRENT;
1393 lcore_rx_idle_count = 0;
1394 for (i = 0; i < qconf->n_rx_queue; ++i) {
1395 rx_queue = &(qconf->rx_queue_list[i]);
1396 rx_queue->idle_hint = 0;
1397 portid = rx_queue->port_id;
1398 queueid = rx_queue->queue_id;
1400 nb_rx = rte_eth_rx_burst(portid, queueid, pkts_burst,
1403 stats[lcore_id].nb_rx_processed += nb_rx;
1404 if (unlikely(nb_rx == 0)) {
1406 * no packet received from rx queue, try to
1407 * sleep for a while forcing CPU enter deeper
1410 rx_queue->zero_rx_packet_count++;
1412 if (rx_queue->zero_rx_packet_count <=
1413 MIN_ZERO_POLL_COUNT)
1416 rx_queue->idle_hint = power_idle_heuristic(\
1417 rx_queue->zero_rx_packet_count);
1418 lcore_rx_idle_count++;
1420 rx_queue->zero_rx_packet_count = 0;
1423 * do not scale up frequency immediately as
1424 * user to kernel space communication is costly
1425 * which might impact packet I/O for received
1428 rx_queue->freq_up_hint =
1429 power_freq_scaleup_heuristic(lcore_id,
1433 /* Prefetch first packets */
1434 for (j = 0; j < PREFETCH_OFFSET && j < nb_rx; j++) {
1435 rte_prefetch0(rte_pktmbuf_mtod(
1436 pkts_burst[j], void *));
1439 /* Prefetch and forward already prefetched packets */
1440 for (j = 0; j < (nb_rx - PREFETCH_OFFSET); j++) {
1441 rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[
1442 j + PREFETCH_OFFSET], void *));
1443 l3fwd_simple_forward(pkts_burst[j], portid,
1447 /* Forward remaining prefetched packets */
1448 for (; j < nb_rx; j++) {
1449 l3fwd_simple_forward(pkts_burst[j], portid,
1454 if (likely(lcore_rx_idle_count != qconf->n_rx_queue)) {
1455 for (i = 1, lcore_scaleup_hint =
1456 qconf->rx_queue_list[0].freq_up_hint;
1457 i < qconf->n_rx_queue; ++i) {
1458 rx_queue = &(qconf->rx_queue_list[i]);
1459 if (rx_queue->freq_up_hint >
1461 lcore_scaleup_hint =
1462 rx_queue->freq_up_hint;
1465 if (lcore_scaleup_hint == FREQ_HIGHEST) {
1466 if (rte_power_freq_max)
1467 rte_power_freq_max(lcore_id);
1468 } else if (lcore_scaleup_hint == FREQ_HIGHER) {
1469 if (rte_power_freq_up)
1470 rte_power_freq_up(lcore_id);
1474 * All Rx queues empty in recent consecutive polls,
1475 * sleep in a conservative manner, meaning sleep as
1478 for (i = 1, lcore_idle_hint =
1479 qconf->rx_queue_list[0].idle_hint;
1480 i < qconf->n_rx_queue; ++i) {
1481 rx_queue = &(qconf->rx_queue_list[i]);
1482 if (rx_queue->idle_hint < lcore_idle_hint)
1483 lcore_idle_hint = rx_queue->idle_hint;
1486 if (lcore_idle_hint < SUSPEND_THRESHOLD)
1488 * execute "pause" instruction to avoid context
1489 * switch which generally take hundred of
1490 * microseconds for short sleep.
1492 rte_delay_us(lcore_idle_hint);
1494 /* suspend until rx interrupt triggers */
1496 turn_on_off_intr(qconf, 1);
1497 sleep_until_rx_interrupt(
1500 turn_on_off_intr(qconf, 0);
1502 * start receiving packets immediately
1504 if (likely(!is_done()))
1508 stats[lcore_id].sleep_time += lcore_idle_hint;
1516 check_lcore_params(void)
1518 uint8_t queue, lcore;
1522 for (i = 0; i < nb_lcore_params; ++i) {
1523 queue = lcore_params[i].queue_id;
1524 if (queue >= MAX_RX_QUEUE_PER_PORT) {
1525 printf("invalid queue number: %hhu\n", queue);
1528 lcore = lcore_params[i].lcore_id;
1529 if (!rte_lcore_is_enabled(lcore)) {
1530 printf("error: lcore %hhu is not enabled in lcore "
1534 if ((socketid = rte_lcore_to_socket_id(lcore) != 0) &&
1536 printf("warning: lcore %hhu is on socket %d with numa "
1537 "off\n", lcore, socketid);
1539 if (app_mode == APP_MODE_TELEMETRY && lcore == rte_lcore_id()) {
1540 printf("cannot enable main core %d in config for telemetry mode\n",
1549 check_port_config(void)
1554 for (i = 0; i < nb_lcore_params; ++i) {
1555 portid = lcore_params[i].port_id;
1556 if ((enabled_port_mask & (1 << portid)) == 0) {
1557 printf("port %u is not enabled in port mask\n",
1561 if (!rte_eth_dev_is_valid_port(portid)) {
1562 printf("port %u is not present on the board\n",
1571 get_port_n_rx_queues(const uint16_t port)
1576 for (i = 0; i < nb_lcore_params; ++i) {
1577 if (lcore_params[i].port_id == port &&
1578 lcore_params[i].queue_id > queue)
1579 queue = lcore_params[i].queue_id;
1581 return (uint8_t)(++queue);
1585 init_lcore_rx_queues(void)
1587 uint16_t i, nb_rx_queue;
1590 for (i = 0; i < nb_lcore_params; ++i) {
1591 lcore = lcore_params[i].lcore_id;
1592 nb_rx_queue = lcore_conf[lcore].n_rx_queue;
1593 if (nb_rx_queue >= MAX_RX_QUEUE_PER_LCORE) {
1594 printf("error: too many queues (%u) for lcore: %u\n",
1595 (unsigned)nb_rx_queue + 1, (unsigned)lcore);
1598 lcore_conf[lcore].rx_queue_list[nb_rx_queue].port_id =
1599 lcore_params[i].port_id;
1600 lcore_conf[lcore].rx_queue_list[nb_rx_queue].queue_id =
1601 lcore_params[i].queue_id;
1602 lcore_conf[lcore].n_rx_queue++;
1610 print_usage(const char *prgname)
1612 printf ("%s [EAL options] -- -p PORTMASK -P"
1613 " [--config (port,queue,lcore)[,(port,queue,lcore]]"
1614 " [--high-perf-cores CORELIST"
1615 " [--perf-config (port,queue,hi_perf,lcore_index)[,(port,queue,hi_perf,lcore_index]]"
1616 " [--max-pkt-len PKTLEN]\n"
1617 " -p PORTMASK: hexadecimal bitmask of ports to configure\n"
1618 " -P: enable promiscuous mode\n"
1619 " --config (port,queue,lcore): rx queues configuration\n"
1620 " --high-perf-cores CORELIST: list of high performance cores\n"
1621 " --perf-config: similar as config, cores specified as indices"
1622 " for bins containing high or regular performance cores\n"
1623 " --no-numa: optional, disable numa awareness\n"
1624 " --max-pkt-len PKTLEN: maximum packet length in decimal (64-9600)\n"
1625 " --parse-ptype: parse packet type by software\n"
1626 " --legacy: use legacy interrupt-based scaling\n"
1627 " --empty-poll: enable empty poll detection"
1628 " follow (training_flag, high_threshold, med_threshold)\n"
1629 " --telemetry: enable telemetry mode, to update"
1630 " empty polls, full polls, and core busyness to telemetry\n"
1631 " --interrupt-only: enable interrupt-only mode\n"
1632 " --pmd-mgmt MODE: enable PMD power management mode. "
1633 "Currently supported modes: baseline, monitor, pause, scale\n"
1634 " --max-empty-polls MAX_EMPTY_POLLS: number of empty polls to"
1635 " wait before entering sleep state\n"
1636 " --pause-duration DURATION: set the duration, in microseconds,"
1637 " of the pause callback\n"
1638 " --scale-freq-min FREQ_MIN: set minimum frequency for scaling mode for"
1639 " all application lcores (FREQ_MIN must be in kHz, in increments of 100MHz)\n"
1640 " --scale-freq-max FREQ_MAX: set maximum frequency for scaling mode for"
1641 " all application lcores (FREQ_MAX must be in kHz, in increments of 100MHz)\n",
1646 parse_int(const char *opt)
1651 /* parse integer string */
1652 val = strtoul(opt, &end, 10);
1653 if ((opt[0] == '\0') || (end == NULL) || (*end != '\0'))
1659 static int parse_max_pkt_len(const char *pktlen)
1664 /* parse decimal string */
1665 len = strtoul(pktlen, &end, 10);
1666 if ((pktlen[0] == '\0') || (end == NULL) || (*end != '\0'))
1676 parse_portmask(const char *portmask)
1681 /* parse hexadecimal string */
1682 pm = strtoul(portmask, &end, 16);
1683 if ((portmask[0] == '\0') || (end == NULL) || (*end != '\0'))
1690 parse_config(const char *q_arg)
1693 const char *p, *p0 = q_arg;
1701 unsigned long int_fld[_NUM_FLD];
1702 char *str_fld[_NUM_FLD];
1706 nb_lcore_params = 0;
1708 while ((p = strchr(p0,'(')) != NULL) {
1710 if((p0 = strchr(p,')')) == NULL)
1714 if(size >= sizeof(s))
1717 snprintf(s, sizeof(s), "%.*s", size, p);
1718 if (rte_strsplit(s, sizeof(s), str_fld, _NUM_FLD, ',') !=
1721 for (i = 0; i < _NUM_FLD; i++){
1723 int_fld[i] = strtoul(str_fld[i], &end, 0);
1724 if (errno != 0 || end == str_fld[i] || int_fld[i] >
1728 if (nb_lcore_params >= MAX_LCORE_PARAMS) {
1729 printf("exceeded max number of lcore params: %hu\n",
1733 lcore_params_array[nb_lcore_params].port_id =
1734 (uint8_t)int_fld[FLD_PORT];
1735 lcore_params_array[nb_lcore_params].queue_id =
1736 (uint8_t)int_fld[FLD_QUEUE];
1737 lcore_params_array[nb_lcore_params].lcore_id =
1738 (uint8_t)int_fld[FLD_LCORE];
1741 lcore_params = lcore_params_array;
1747 parse_pmd_mgmt_config(const char *name)
1749 #define PMD_MGMT_MONITOR "monitor"
1750 #define PMD_MGMT_PAUSE "pause"
1751 #define PMD_MGMT_SCALE "scale"
1752 #define PMD_MGMT_BASELINE "baseline"
1754 if (strncmp(PMD_MGMT_MONITOR, name, sizeof(PMD_MGMT_MONITOR)) == 0) {
1755 pmgmt_type = RTE_POWER_MGMT_TYPE_MONITOR;
1759 if (strncmp(PMD_MGMT_PAUSE, name, sizeof(PMD_MGMT_PAUSE)) == 0) {
1760 pmgmt_type = RTE_POWER_MGMT_TYPE_PAUSE;
1764 if (strncmp(PMD_MGMT_SCALE, name, sizeof(PMD_MGMT_SCALE)) == 0) {
1765 pmgmt_type = RTE_POWER_MGMT_TYPE_SCALE;
1768 if (strncmp(PMD_MGMT_BASELINE, name, sizeof(PMD_MGMT_BASELINE)) == 0) {
1769 baseline_enabled = true;
1772 /* unknown PMD power management mode */
1777 parse_ep_config(const char *q_arg)
1780 const char *p = q_arg;
1790 ep_med_edpi = EMPTY_POLL_MED_THRESHOLD;
1791 ep_hgh_edpi = EMPTY_POLL_HGH_THRESHOLD;
1793 strlcpy(s, p, sizeof(s));
1795 num_arg = rte_strsplit(s, sizeof(s), str_fld, 3, ',');
1797 empty_poll_train = false;
1804 training_flag = strtoul(str_fld[0], &end, 0);
1805 med_edpi = strtoul(str_fld[1], &end, 0);
1806 hgh_edpi = strtoul(str_fld[2], &end, 0);
1808 if (training_flag == 1)
1809 empty_poll_train = true;
1812 ep_med_edpi = med_edpi;
1815 ep_hgh_edpi = hgh_edpi;
1825 #define CMD_LINE_OPT_PARSE_PTYPE "parse-ptype"
1826 #define CMD_LINE_OPT_LEGACY "legacy"
1827 #define CMD_LINE_OPT_EMPTY_POLL "empty-poll"
1828 #define CMD_LINE_OPT_INTERRUPT_ONLY "interrupt-only"
1829 #define CMD_LINE_OPT_TELEMETRY "telemetry"
1830 #define CMD_LINE_OPT_PMD_MGMT "pmd-mgmt"
1831 #define CMD_LINE_OPT_MAX_PKT_LEN "max-pkt-len"
1832 #define CMD_LINE_OPT_MAX_EMPTY_POLLS "max-empty-polls"
1833 #define CMD_LINE_OPT_PAUSE_DURATION "pause-duration"
1834 #define CMD_LINE_OPT_SCALE_FREQ_MIN "scale-freq-min"
1835 #define CMD_LINE_OPT_SCALE_FREQ_MAX "scale-freq-max"
1837 /* Parse the argument given in the command line of the application */
1839 parse_args(int argc, char **argv)
1845 char *prgname = argv[0];
1846 static struct option lgopts[] = {
1847 {"config", 1, 0, 0},
1848 {"perf-config", 1, 0, 0},
1849 {"high-perf-cores", 1, 0, 0},
1850 {"no-numa", 0, 0, 0},
1851 {CMD_LINE_OPT_MAX_PKT_LEN, 1, 0, 0},
1852 {CMD_LINE_OPT_EMPTY_POLL, 1, 0, 0},
1853 {CMD_LINE_OPT_PARSE_PTYPE, 0, 0, 0},
1854 {CMD_LINE_OPT_LEGACY, 0, 0, 0},
1855 {CMD_LINE_OPT_TELEMETRY, 0, 0, 0},
1856 {CMD_LINE_OPT_INTERRUPT_ONLY, 0, 0, 0},
1857 {CMD_LINE_OPT_PMD_MGMT, 1, 0, 0},
1858 {CMD_LINE_OPT_MAX_EMPTY_POLLS, 1, 0, 0},
1859 {CMD_LINE_OPT_PAUSE_DURATION, 1, 0, 0},
1860 {CMD_LINE_OPT_SCALE_FREQ_MIN, 1, 0, 0},
1861 {CMD_LINE_OPT_SCALE_FREQ_MAX, 1, 0, 0},
1867 while ((opt = getopt_long(argc, argvopt, "p:l:m:h:P",
1868 lgopts, &option_index)) != EOF) {
1873 enabled_port_mask = parse_portmask(optarg);
1874 if (enabled_port_mask == 0) {
1875 printf("invalid portmask\n");
1876 print_usage(prgname);
1881 printf("Promiscuous mode selected\n");
1885 limit = parse_max_pkt_len(optarg);
1886 freq_tlb[LOW] = limit;
1889 limit = parse_max_pkt_len(optarg);
1890 freq_tlb[MED] = limit;
1893 limit = parse_max_pkt_len(optarg);
1894 freq_tlb[HGH] = limit;
1898 if (!strncmp(lgopts[option_index].name, "config", 6)) {
1899 ret = parse_config(optarg);
1901 printf("invalid config\n");
1902 print_usage(prgname);
1907 if (!strncmp(lgopts[option_index].name,
1908 "perf-config", 11)) {
1909 ret = parse_perf_config(optarg);
1911 printf("invalid perf-config\n");
1912 print_usage(prgname);
1917 if (!strncmp(lgopts[option_index].name,
1918 "high-perf-cores", 15)) {
1919 ret = parse_perf_core_list(optarg);
1921 printf("invalid high-perf-cores\n");
1922 print_usage(prgname);
1927 if (!strncmp(lgopts[option_index].name,
1929 printf("numa is disabled \n");
1933 if (!strncmp(lgopts[option_index].name,
1934 CMD_LINE_OPT_LEGACY,
1935 sizeof(CMD_LINE_OPT_LEGACY))) {
1936 if (app_mode != APP_MODE_DEFAULT) {
1937 printf(" legacy mode is mutually exclusive with other modes\n");
1940 app_mode = APP_MODE_LEGACY;
1941 printf("legacy mode is enabled\n");
1944 if (!strncmp(lgopts[option_index].name,
1945 CMD_LINE_OPT_EMPTY_POLL, 10)) {
1946 if (app_mode != APP_MODE_DEFAULT) {
1947 printf(" empty-poll mode is mutually exclusive with other modes\n");
1950 app_mode = APP_MODE_EMPTY_POLL;
1951 ret = parse_ep_config(optarg);
1954 printf("invalid empty poll config\n");
1955 print_usage(prgname);
1958 printf("empty-poll is enabled\n");
1961 if (!strncmp(lgopts[option_index].name,
1962 CMD_LINE_OPT_TELEMETRY,
1963 sizeof(CMD_LINE_OPT_TELEMETRY))) {
1964 if (app_mode != APP_MODE_DEFAULT) {
1965 printf(" telemetry mode is mutually exclusive with other modes\n");
1968 app_mode = APP_MODE_TELEMETRY;
1969 printf("telemetry mode is enabled\n");
1972 if (!strncmp(lgopts[option_index].name,
1973 CMD_LINE_OPT_PMD_MGMT,
1974 sizeof(CMD_LINE_OPT_PMD_MGMT))) {
1975 if (app_mode != APP_MODE_DEFAULT) {
1976 printf(" power mgmt mode is mutually exclusive with other modes\n");
1979 if (parse_pmd_mgmt_config(optarg) < 0) {
1980 printf(" Invalid PMD power management mode: %s\n",
1984 app_mode = APP_MODE_PMD_MGMT;
1985 printf("PMD power mgmt mode is enabled\n");
1987 if (!strncmp(lgopts[option_index].name,
1988 CMD_LINE_OPT_INTERRUPT_ONLY,
1989 sizeof(CMD_LINE_OPT_INTERRUPT_ONLY))) {
1990 if (app_mode != APP_MODE_DEFAULT) {
1991 printf(" interrupt-only mode is mutually exclusive with other modes\n");
1994 app_mode = APP_MODE_INTERRUPT;
1995 printf("interrupt-only mode is enabled\n");
1998 if (!strncmp(lgopts[option_index].name,
1999 CMD_LINE_OPT_MAX_PKT_LEN,
2000 sizeof(CMD_LINE_OPT_MAX_PKT_LEN))) {
2001 printf("Custom frame size is configured\n");
2002 max_pkt_len = parse_max_pkt_len(optarg);
2005 if (!strncmp(lgopts[option_index].name,
2006 CMD_LINE_OPT_PARSE_PTYPE,
2007 sizeof(CMD_LINE_OPT_PARSE_PTYPE))) {
2008 printf("soft parse-ptype is enabled\n");
2012 if (!strncmp(lgopts[option_index].name,
2013 CMD_LINE_OPT_MAX_EMPTY_POLLS,
2014 sizeof(CMD_LINE_OPT_MAX_EMPTY_POLLS))) {
2015 printf("Maximum empty polls configured\n");
2016 max_empty_polls = parse_int(optarg);
2019 if (!strncmp(lgopts[option_index].name,
2020 CMD_LINE_OPT_PAUSE_DURATION,
2021 sizeof(CMD_LINE_OPT_PAUSE_DURATION))) {
2022 printf("Pause duration configured\n");
2023 pause_duration = parse_int(optarg);
2026 if (!strncmp(lgopts[option_index].name,
2027 CMD_LINE_OPT_SCALE_FREQ_MIN,
2028 sizeof(CMD_LINE_OPT_SCALE_FREQ_MIN))) {
2029 printf("Scaling frequency minimum configured\n");
2030 scale_freq_min = parse_int(optarg);
2033 if (!strncmp(lgopts[option_index].name,
2034 CMD_LINE_OPT_SCALE_FREQ_MAX,
2035 sizeof(CMD_LINE_OPT_SCALE_FREQ_MAX))) {
2036 printf("Scaling frequency maximum configured\n");
2037 scale_freq_max = parse_int(optarg);
2043 print_usage(prgname);
2049 argv[optind-1] = prgname;
2052 optind = 1; /* reset getopt lib */
2057 print_ethaddr(const char *name, const struct rte_ether_addr *eth_addr)
2059 char buf[RTE_ETHER_ADDR_FMT_SIZE];
2060 rte_ether_format_addr(buf, RTE_ETHER_ADDR_FMT_SIZE, eth_addr);
2061 printf("%s%s", name, buf);
2064 #if (APP_LOOKUP_METHOD == APP_LOOKUP_EXACT_MATCH)
2066 setup_hash(int socketid)
2068 struct rte_hash_parameters ipv4_l3fwd_hash_params = {
2070 .entries = L3FWD_HASH_ENTRIES,
2071 .key_len = sizeof(struct ipv4_5tuple),
2072 .hash_func = DEFAULT_HASH_FUNC,
2073 .hash_func_init_val = 0,
2076 struct rte_hash_parameters ipv6_l3fwd_hash_params = {
2078 .entries = L3FWD_HASH_ENTRIES,
2079 .key_len = sizeof(struct ipv6_5tuple),
2080 .hash_func = DEFAULT_HASH_FUNC,
2081 .hash_func_init_val = 0,
2088 /* create ipv4 hash */
2089 snprintf(s, sizeof(s), "ipv4_l3fwd_hash_%d", socketid);
2090 ipv4_l3fwd_hash_params.name = s;
2091 ipv4_l3fwd_hash_params.socket_id = socketid;
2092 ipv4_l3fwd_lookup_struct[socketid] =
2093 rte_hash_create(&ipv4_l3fwd_hash_params);
2094 if (ipv4_l3fwd_lookup_struct[socketid] == NULL)
2095 rte_exit(EXIT_FAILURE, "Unable to create the l3fwd hash on "
2096 "socket %d\n", socketid);
2098 /* create ipv6 hash */
2099 snprintf(s, sizeof(s), "ipv6_l3fwd_hash_%d", socketid);
2100 ipv6_l3fwd_hash_params.name = s;
2101 ipv6_l3fwd_hash_params.socket_id = socketid;
2102 ipv6_l3fwd_lookup_struct[socketid] =
2103 rte_hash_create(&ipv6_l3fwd_hash_params);
2104 if (ipv6_l3fwd_lookup_struct[socketid] == NULL)
2105 rte_exit(EXIT_FAILURE, "Unable to create the l3fwd hash on "
2106 "socket %d\n", socketid);
2109 /* populate the ipv4 hash */
2110 for (i = 0; i < RTE_DIM(ipv4_l3fwd_route_array); i++) {
2111 ret = rte_hash_add_key (ipv4_l3fwd_lookup_struct[socketid],
2112 (void *) &ipv4_l3fwd_route_array[i].key);
2114 rte_exit(EXIT_FAILURE, "Unable to add entry %u to the"
2115 "l3fwd hash on socket %d\n", i, socketid);
2117 ipv4_l3fwd_out_if[ret] = ipv4_l3fwd_route_array[i].if_out;
2118 printf("Hash: Adding key\n");
2119 print_ipv4_key(ipv4_l3fwd_route_array[i].key);
2122 /* populate the ipv6 hash */
2123 for (i = 0; i < RTE_DIM(ipv6_l3fwd_route_array); i++) {
2124 ret = rte_hash_add_key (ipv6_l3fwd_lookup_struct[socketid],
2125 (void *) &ipv6_l3fwd_route_array[i].key);
2127 rte_exit(EXIT_FAILURE, "Unable to add entry %u to the"
2128 "l3fwd hash on socket %d\n", i, socketid);
2130 ipv6_l3fwd_out_if[ret] = ipv6_l3fwd_route_array[i].if_out;
2131 printf("Hash: Adding key\n");
2132 print_ipv6_key(ipv6_l3fwd_route_array[i].key);
2137 #if (APP_LOOKUP_METHOD == APP_LOOKUP_LPM)
2139 setup_lpm(int socketid)
2145 /* create the LPM table */
2146 struct rte_lpm_config lpm_ipv4_config;
2148 lpm_ipv4_config.max_rules = IPV4_L3FWD_LPM_MAX_RULES;
2149 lpm_ipv4_config.number_tbl8s = 256;
2150 lpm_ipv4_config.flags = 0;
2152 snprintf(s, sizeof(s), "IPV4_L3FWD_LPM_%d", socketid);
2153 ipv4_l3fwd_lookup_struct[socketid] =
2154 rte_lpm_create(s, socketid, &lpm_ipv4_config);
2155 if (ipv4_l3fwd_lookup_struct[socketid] == NULL)
2156 rte_exit(EXIT_FAILURE, "Unable to create the l3fwd LPM table"
2157 " on socket %d\n", socketid);
2159 /* populate the LPM table */
2160 for (i = 0; i < RTE_DIM(ipv4_l3fwd_route_array); i++) {
2161 ret = rte_lpm_add(ipv4_l3fwd_lookup_struct[socketid],
2162 ipv4_l3fwd_route_array[i].ip,
2163 ipv4_l3fwd_route_array[i].depth,
2164 ipv4_l3fwd_route_array[i].if_out);
2167 rte_exit(EXIT_FAILURE, "Unable to add entry %u to the "
2168 "l3fwd LPM table on socket %d\n",
2172 printf("LPM: Adding route 0x%08x / %d (%d)\n",
2173 (unsigned)ipv4_l3fwd_route_array[i].ip,
2174 ipv4_l3fwd_route_array[i].depth,
2175 ipv4_l3fwd_route_array[i].if_out);
2181 init_mem(unsigned nb_mbuf)
2183 struct lcore_conf *qconf;
2188 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
2189 if (rte_lcore_is_enabled(lcore_id) == 0)
2193 socketid = rte_lcore_to_socket_id(lcore_id);
2197 if (socketid >= NB_SOCKETS) {
2198 rte_exit(EXIT_FAILURE, "Socket %d of lcore %u is "
2199 "out of range %d\n", socketid,
2200 lcore_id, NB_SOCKETS);
2202 if (pktmbuf_pool[socketid] == NULL) {
2203 snprintf(s, sizeof(s), "mbuf_pool_%d", socketid);
2204 pktmbuf_pool[socketid] =
2205 rte_pktmbuf_pool_create(s, nb_mbuf,
2206 MEMPOOL_CACHE_SIZE, 0,
2207 RTE_MBUF_DEFAULT_BUF_SIZE,
2209 if (pktmbuf_pool[socketid] == NULL)
2210 rte_exit(EXIT_FAILURE,
2211 "Cannot init mbuf pool on socket %d\n",
2214 printf("Allocated mbuf pool on socket %d\n",
2217 #if (APP_LOOKUP_METHOD == APP_LOOKUP_LPM)
2218 setup_lpm(socketid);
2220 setup_hash(socketid);
2223 qconf = &lcore_conf[lcore_id];
2224 qconf->ipv4_lookup_struct = ipv4_l3fwd_lookup_struct[socketid];
2225 #if (APP_LOOKUP_METHOD == APP_LOOKUP_EXACT_MATCH)
2226 qconf->ipv6_lookup_struct = ipv6_l3fwd_lookup_struct[socketid];
2232 /* Check the link status of all ports in up to 9s, and print them finally */
2234 check_all_ports_link_status(uint32_t port_mask)
2236 #define CHECK_INTERVAL 100 /* 100ms */
2237 #define MAX_CHECK_TIME 90 /* 9s (90 * 100ms) in total */
2238 uint8_t count, all_ports_up, print_flag = 0;
2240 struct rte_eth_link link;
2242 char link_status_text[RTE_ETH_LINK_MAX_STR_LEN];
2244 printf("\nChecking link status");
2246 for (count = 0; count <= MAX_CHECK_TIME; count++) {
2248 RTE_ETH_FOREACH_DEV(portid) {
2249 if ((port_mask & (1 << portid)) == 0)
2251 memset(&link, 0, sizeof(link));
2252 ret = rte_eth_link_get_nowait(portid, &link);
2255 if (print_flag == 1)
2256 printf("Port %u link get failed: %s\n",
2257 portid, rte_strerror(-ret));
2260 /* print link status if flag set */
2261 if (print_flag == 1) {
2262 rte_eth_link_to_str(link_status_text,
2263 sizeof(link_status_text), &link);
2264 printf("Port %d %s\n", portid,
2268 /* clear all_ports_up flag if any link down */
2269 if (link.link_status == RTE_ETH_LINK_DOWN) {
2274 /* after finally printing all link status, get out */
2275 if (print_flag == 1)
2278 if (all_ports_up == 0) {
2281 rte_delay_ms(CHECK_INTERVAL);
2284 /* set the print_flag if all ports up or timeout */
2285 if (all_ports_up == 1 || count == (MAX_CHECK_TIME - 1)) {
2292 static int check_ptype(uint16_t portid)
2295 int ptype_l3_ipv4 = 0;
2296 #if (APP_LOOKUP_METHOD == APP_LOOKUP_EXACT_MATCH)
2297 int ptype_l3_ipv6 = 0;
2299 uint32_t ptype_mask = RTE_PTYPE_L3_MASK;
2301 ret = rte_eth_dev_get_supported_ptypes(portid, ptype_mask, NULL, 0);
2305 uint32_t ptypes[ret];
2307 ret = rte_eth_dev_get_supported_ptypes(portid, ptype_mask, ptypes, ret);
2308 for (i = 0; i < ret; ++i) {
2309 if (ptypes[i] & RTE_PTYPE_L3_IPV4)
2311 #if (APP_LOOKUP_METHOD == APP_LOOKUP_EXACT_MATCH)
2312 if (ptypes[i] & RTE_PTYPE_L3_IPV6)
2317 if (ptype_l3_ipv4 == 0)
2318 printf("port %d cannot parse RTE_PTYPE_L3_IPV4\n", portid);
2320 #if (APP_LOOKUP_METHOD == APP_LOOKUP_EXACT_MATCH)
2321 if (ptype_l3_ipv6 == 0)
2322 printf("port %d cannot parse RTE_PTYPE_L3_IPV6\n", portid);
2325 #if (APP_LOOKUP_METHOD == APP_LOOKUP_LPM)
2327 #else /* APP_LOOKUP_EXACT_MATCH */
2328 if (ptype_l3_ipv4 && ptype_l3_ipv6)
2337 init_power_library(void)
2339 enum power_management_env env;
2340 unsigned int lcore_id;
2343 RTE_LCORE_FOREACH(lcore_id) {
2344 /* init power management library */
2345 ret = rte_power_init(lcore_id);
2348 "Library initialization failed on core %u\n",
2352 /* we're not supporting the VM channel mode */
2353 env = rte_power_get_env();
2354 if (env != PM_ENV_ACPI_CPUFREQ &&
2355 env != PM_ENV_PSTATE_CPUFREQ) {
2357 "Only ACPI and PSTATE mode are supported\n");
2365 deinit_power_library(void)
2367 unsigned int lcore_id;
2370 RTE_LCORE_FOREACH(lcore_id) {
2371 /* deinit power management library */
2372 ret = rte_power_exit(lcore_id);
2375 "Library deinitialization failed on core %u\n",
2384 get_current_stat_values(uint64_t *values)
2386 unsigned int lcore_id = rte_lcore_id();
2387 struct lcore_conf *qconf;
2388 uint64_t app_eps = 0, app_fps = 0, app_br = 0;
2391 RTE_LCORE_FOREACH_WORKER(lcore_id) {
2392 qconf = &lcore_conf[lcore_id];
2393 if (qconf->n_rx_queue == 0)
2396 rte_spinlock_lock(&stats[lcore_id].telemetry_lock);
2397 app_eps += stats[lcore_id].ep_nep[1];
2398 app_fps += stats[lcore_id].fp_nfp[1];
2399 app_br += stats[lcore_id].br;
2400 rte_spinlock_unlock(&stats[lcore_id].telemetry_lock);
2404 values[0] = app_eps/count;
2405 values[1] = app_fps/count;
2406 values[2] = app_br/count;
2408 memset(values, 0, sizeof(uint64_t) * NUM_TELSTATS);
2413 update_telemetry(__rte_unused struct rte_timer *tim,
2414 __rte_unused void *arg)
2417 uint64_t values[NUM_TELSTATS] = {0};
2419 get_current_stat_values(values);
2420 ret = rte_metrics_update_values(RTE_METRICS_GLOBAL, telstats_index,
2421 values, RTE_DIM(values));
2423 RTE_LOG(WARNING, POWER, "failed to update metrics\n");
2427 handle_app_stats(const char *cmd __rte_unused,
2428 const char *params __rte_unused,
2429 struct rte_tel_data *d)
2431 uint64_t values[NUM_TELSTATS] = {0};
2434 rte_tel_data_start_dict(d);
2435 get_current_stat_values(values);
2436 for (i = 0; i < NUM_TELSTATS; i++)
2437 rte_tel_data_add_dict_u64(d, telstats_strings[i].name,
2443 telemetry_setup_timer(void)
2445 int lcore_id = rte_lcore_id();
2446 uint64_t hz = rte_get_timer_hz();
2449 ticks = hz / TELEMETRY_INTERVALS_PER_SEC;
2450 rte_timer_reset_sync(&telemetry_timer,
2458 empty_poll_setup_timer(void)
2460 int lcore_id = rte_lcore_id();
2461 uint64_t hz = rte_get_timer_hz();
2463 struct ep_params *ep_ptr = ep_params;
2465 ep_ptr->interval_ticks = hz / INTERVALS_PER_SECOND;
2467 rte_timer_reset_sync(&ep_ptr->timer0,
2468 ep_ptr->interval_ticks,
2471 rte_empty_poll_detection,
2476 launch_timer(unsigned int lcore_id)
2478 int64_t prev_tsc = 0, cur_tsc, diff_tsc, cycles_10ms;
2480 RTE_SET_USED(lcore_id);
2483 if (rte_get_main_lcore() != lcore_id) {
2484 rte_panic("timer on lcore:%d which is not main core:%d\n",
2486 rte_get_main_lcore());
2489 RTE_LOG(INFO, POWER, "Bring up the Timer\n");
2491 if (app_mode == APP_MODE_EMPTY_POLL)
2492 empty_poll_setup_timer();
2494 telemetry_setup_timer();
2496 cycles_10ms = rte_get_timer_hz() / 100;
2498 while (!is_done()) {
2499 cur_tsc = rte_rdtsc();
2500 diff_tsc = cur_tsc - prev_tsc;
2501 if (diff_tsc > cycles_10ms) {
2504 cycles_10ms = rte_get_timer_hz() / 100;
2508 RTE_LOG(INFO, POWER, "Timer_subsystem is done\n");
2514 autodetect_mode(void)
2516 RTE_LOG(NOTICE, L3FWD_POWER, "Operating mode not specified, probing frequency scaling support...\n");
2519 * Empty poll and telemetry modes have to be specifically requested to
2520 * be enabled, but we can auto-detect between interrupt mode with or
2521 * without frequency scaling. Both ACPI and pstate can be used.
2523 if (rte_power_check_env_supported(PM_ENV_ACPI_CPUFREQ))
2524 return APP_MODE_LEGACY;
2525 if (rte_power_check_env_supported(PM_ENV_PSTATE_CPUFREQ))
2526 return APP_MODE_LEGACY;
2528 RTE_LOG(NOTICE, L3FWD_POWER, "Frequency scaling not supported, selecting interrupt-only mode\n");
2530 return APP_MODE_INTERRUPT;
2534 mode_to_str(enum appmode mode)
2537 case APP_MODE_LEGACY:
2539 case APP_MODE_EMPTY_POLL:
2540 return "empty poll";
2541 case APP_MODE_TELEMETRY:
2543 case APP_MODE_INTERRUPT:
2544 return "interrupt-only";
2545 case APP_MODE_PMD_MGMT:
2553 eth_dev_get_overhead_len(uint32_t max_rx_pktlen, uint16_t max_mtu)
2555 uint32_t overhead_len;
2557 if (max_mtu != UINT16_MAX && max_rx_pktlen > max_mtu)
2558 overhead_len = max_rx_pktlen - max_mtu;
2560 overhead_len = RTE_ETHER_HDR_LEN + RTE_ETHER_CRC_LEN;
2562 return overhead_len;
2566 config_port_max_pkt_len(struct rte_eth_conf *conf,
2567 struct rte_eth_dev_info *dev_info)
2569 uint32_t overhead_len;
2571 if (max_pkt_len == 0)
2574 if (max_pkt_len < RTE_ETHER_MIN_LEN || max_pkt_len > MAX_JUMBO_PKT_LEN)
2577 overhead_len = eth_dev_get_overhead_len(dev_info->max_rx_pktlen,
2579 conf->rxmode.mtu = max_pkt_len - overhead_len;
2581 if (conf->rxmode.mtu > RTE_ETHER_MTU)
2582 conf->txmode.offloads |= RTE_ETH_TX_OFFLOAD_MULTI_SEGS;
2587 /* Power library initialized in the main routine. 8< */
2589 main(int argc, char **argv)
2591 struct lcore_conf *qconf;
2592 struct rte_eth_dev_info dev_info;
2593 struct rte_eth_txconf *txconf;
2599 uint32_t n_tx_queue, nb_lcores;
2600 uint32_t dev_rxq_num, dev_txq_num;
2601 uint8_t nb_rx_queue, queue, socketid;
2603 const char *ptr_strings[NUM_TELSTATS];
2606 ret = rte_eal_init(argc, argv);
2608 rte_exit(EXIT_FAILURE, "Invalid EAL parameters\n");
2612 /* catch SIGINT and restore cpufreq governor to ondemand */
2613 signal(SIGINT, signal_exit_now);
2615 /* init RTE timer library to be used late */
2616 rte_timer_subsystem_init();
2618 /* if we're running pmd-mgmt mode, don't default to baseline mode */
2619 baseline_enabled = false;
2621 /* parse application arguments (after the EAL ones) */
2622 ret = parse_args(argc, argv);
2624 rte_exit(EXIT_FAILURE, "Invalid L3FWD parameters\n");
2626 if (app_mode == APP_MODE_DEFAULT)
2627 app_mode = autodetect_mode();
2629 RTE_LOG(INFO, L3FWD_POWER, "Selected operation mode: %s\n",
2630 mode_to_str(app_mode));
2632 /* only legacy and empty poll mode rely on power library */
2633 if ((app_mode == APP_MODE_LEGACY || app_mode == APP_MODE_EMPTY_POLL) &&
2634 init_power_library())
2635 rte_exit(EXIT_FAILURE, "init_power_library failed\n");
2637 if (update_lcore_params() < 0)
2638 rte_exit(EXIT_FAILURE, "update_lcore_params failed\n");
2640 if (check_lcore_params() < 0)
2641 rte_exit(EXIT_FAILURE, "check_lcore_params failed\n");
2643 ret = init_lcore_rx_queues();
2645 rte_exit(EXIT_FAILURE, "init_lcore_rx_queues failed\n");
2647 nb_ports = rte_eth_dev_count_avail();
2649 if (check_port_config() < 0)
2650 rte_exit(EXIT_FAILURE, "check_port_config failed\n");
2652 nb_lcores = rte_lcore_count();
2654 /* initialize all ports */
2655 RTE_ETH_FOREACH_DEV(portid) {
2656 struct rte_eth_conf local_port_conf = port_conf;
2657 /* not all app modes need interrupts */
2658 bool need_intr = app_mode == APP_MODE_LEGACY ||
2659 app_mode == APP_MODE_INTERRUPT;
2661 /* skip ports that are not enabled */
2662 if ((enabled_port_mask & (1 << portid)) == 0) {
2663 printf("\nSkipping disabled port %d\n", portid);
2668 printf("Initializing port %d ... ", portid );
2671 ret = rte_eth_dev_info_get(portid, &dev_info);
2673 rte_exit(EXIT_FAILURE,
2674 "Error during getting device (port %u) info: %s\n",
2675 portid, strerror(-ret));
2677 dev_rxq_num = dev_info.max_rx_queues;
2678 dev_txq_num = dev_info.max_tx_queues;
2680 nb_rx_queue = get_port_n_rx_queues(portid);
2681 if (nb_rx_queue > dev_rxq_num)
2682 rte_exit(EXIT_FAILURE,
2683 "Cannot configure not existed rxq: "
2684 "port=%d\n", portid);
2686 n_tx_queue = nb_lcores;
2687 if (n_tx_queue > dev_txq_num)
2688 n_tx_queue = dev_txq_num;
2689 printf("Creating queues: nb_rxq=%d nb_txq=%u... ",
2690 nb_rx_queue, (unsigned)n_tx_queue );
2691 /* If number of Rx queue is 0, no need to enable Rx interrupt */
2692 if (nb_rx_queue == 0)
2696 local_port_conf.intr_conf.rxq = 1;
2698 ret = rte_eth_dev_info_get(portid, &dev_info);
2700 rte_exit(EXIT_FAILURE,
2701 "Error during getting device (port %u) info: %s\n",
2702 portid, strerror(-ret));
2704 ret = config_port_max_pkt_len(&local_port_conf, &dev_info);
2706 rte_exit(EXIT_FAILURE,
2707 "Invalid max packet length: %u (port %u)\n",
2708 max_pkt_len, portid);
2710 if (dev_info.tx_offload_capa & RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE)
2711 local_port_conf.txmode.offloads |=
2712 RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE;
2714 local_port_conf.rx_adv_conf.rss_conf.rss_hf &=
2715 dev_info.flow_type_rss_offloads;
2716 if (local_port_conf.rx_adv_conf.rss_conf.rss_hf !=
2717 port_conf.rx_adv_conf.rss_conf.rss_hf) {
2718 printf("Port %u modified RSS hash function based on hardware support,"
2719 "requested:%#"PRIx64" configured:%#"PRIx64"\n",
2721 port_conf.rx_adv_conf.rss_conf.rss_hf,
2722 local_port_conf.rx_adv_conf.rss_conf.rss_hf);
2725 if (local_port_conf.rx_adv_conf.rss_conf.rss_hf == 0)
2726 local_port_conf.rxmode.mq_mode = RTE_ETH_MQ_RX_NONE;
2727 local_port_conf.rxmode.offloads &= dev_info.rx_offload_capa;
2728 port_conf.rxmode.offloads = local_port_conf.rxmode.offloads;
2730 ret = rte_eth_dev_configure(portid, nb_rx_queue,
2731 (uint16_t)n_tx_queue, &local_port_conf);
2733 rte_exit(EXIT_FAILURE, "Cannot configure device: "
2734 "err=%d, port=%d\n", ret, portid);
2736 ret = rte_eth_dev_adjust_nb_rx_tx_desc(portid, &nb_rxd,
2739 rte_exit(EXIT_FAILURE,
2740 "Cannot adjust number of descriptors: err=%d, port=%d\n",
2743 ret = rte_eth_macaddr_get(portid, &ports_eth_addr[portid]);
2745 rte_exit(EXIT_FAILURE,
2746 "Cannot get MAC address: err=%d, port=%d\n",
2749 print_ethaddr(" Address:", &ports_eth_addr[portid]);
2753 ret = init_mem(NB_MBUF);
2755 rte_exit(EXIT_FAILURE, "init_mem failed\n");
2757 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
2758 if (rte_lcore_is_enabled(lcore_id) == 0)
2761 /* Initialize TX buffers */
2762 qconf = &lcore_conf[lcore_id];
2763 qconf->tx_buffer[portid] = rte_zmalloc_socket("tx_buffer",
2764 RTE_ETH_TX_BUFFER_SIZE(MAX_PKT_BURST), 0,
2765 rte_eth_dev_socket_id(portid));
2766 if (qconf->tx_buffer[portid] == NULL)
2767 rte_exit(EXIT_FAILURE, "Can't allocate tx buffer for port %u\n",
2770 rte_eth_tx_buffer_init(qconf->tx_buffer[portid], MAX_PKT_BURST);
2773 /* init one TX queue per couple (lcore,port) */
2775 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
2776 if (rte_lcore_is_enabled(lcore_id) == 0)
2779 if (queueid >= dev_txq_num)
2784 (uint8_t)rte_lcore_to_socket_id(lcore_id);
2788 printf("txq=%u,%d,%d ", lcore_id, queueid, socketid);
2791 txconf = &dev_info.default_txconf;
2792 txconf->offloads = local_port_conf.txmode.offloads;
2793 ret = rte_eth_tx_queue_setup(portid, queueid, nb_txd,
2796 rte_exit(EXIT_FAILURE,
2797 "rte_eth_tx_queue_setup: err=%d, "
2798 "port=%d\n", ret, portid);
2800 qconf = &lcore_conf[lcore_id];
2801 qconf->tx_queue_id[portid] = queueid;
2804 qconf->tx_port_id[qconf->n_tx_port] = portid;
2810 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
2811 if (rte_lcore_is_enabled(lcore_id) == 0)
2814 if (app_mode == APP_MODE_LEGACY) {
2815 /* init timer structures for each enabled lcore */
2816 rte_timer_init(&power_timers[lcore_id]);
2817 hz = rte_get_timer_hz();
2818 rte_timer_reset(&power_timers[lcore_id],
2819 hz/TIMER_NUMBER_PER_SECOND,
2821 power_timer_cb, NULL);
2823 qconf = &lcore_conf[lcore_id];
2824 printf("\nInitializing rx queues on lcore %u ... ", lcore_id );
2827 /* init RX queues */
2828 for(queue = 0; queue < qconf->n_rx_queue; ++queue) {
2829 struct rte_eth_rxconf rxq_conf;
2831 portid = qconf->rx_queue_list[queue].port_id;
2832 queueid = qconf->rx_queue_list[queue].queue_id;
2836 (uint8_t)rte_lcore_to_socket_id(lcore_id);
2840 printf("rxq=%d,%d,%d ", portid, queueid, socketid);
2843 ret = rte_eth_dev_info_get(portid, &dev_info);
2845 rte_exit(EXIT_FAILURE,
2846 "Error during getting device (port %u) info: %s\n",
2847 portid, strerror(-ret));
2849 rxq_conf = dev_info.default_rxconf;
2850 rxq_conf.offloads = port_conf.rxmode.offloads;
2851 ret = rte_eth_rx_queue_setup(portid, queueid, nb_rxd,
2852 socketid, &rxq_conf,
2853 pktmbuf_pool[socketid]);
2855 rte_exit(EXIT_FAILURE,
2856 "rte_eth_rx_queue_setup: err=%d, "
2857 "port=%d\n", ret, portid);
2860 if (add_cb_parse_ptype(portid, queueid) < 0)
2861 rte_exit(EXIT_FAILURE,
2862 "Fail to add ptype cb\n");
2865 if (app_mode == APP_MODE_PMD_MGMT && !baseline_enabled) {
2866 /* Set power_pmd_mgmt configs passed by user */
2867 rte_power_pmd_mgmt_set_emptypoll_max(max_empty_polls);
2868 ret = rte_power_pmd_mgmt_set_pause_duration(pause_duration);
2870 rte_exit(EXIT_FAILURE,
2871 "Error setting pause_duration: err=%d, lcore=%d\n",
2874 ret = rte_power_pmd_mgmt_set_scaling_freq_min(lcore_id,
2877 rte_exit(EXIT_FAILURE,
2878 "Error setting scaling freq min: err=%d, lcore=%d\n",
2881 ret = rte_power_pmd_mgmt_set_scaling_freq_max(lcore_id,
2884 rte_exit(EXIT_FAILURE,
2885 "Error setting scaling freq max: err=%d, lcore %d\n",
2888 ret = rte_power_ethdev_pmgmt_queue_enable(
2889 lcore_id, portid, queueid,
2892 rte_exit(EXIT_FAILURE,
2893 "rte_power_ethdev_pmgmt_queue_enable: err=%d, port=%d\n",
2898 /* >8 End of power library initialization. */
2903 RTE_ETH_FOREACH_DEV(portid) {
2904 if ((enabled_port_mask & (1 << portid)) == 0) {
2908 ret = rte_eth_dev_start(portid);
2910 rte_exit(EXIT_FAILURE, "rte_eth_dev_start: err=%d, "
2911 "port=%d\n", ret, portid);
2913 * If enabled, put device in promiscuous mode.
2914 * This allows IO forwarding mode to forward packets
2915 * to itself through 2 cross-connected ports of the
2918 if (promiscuous_on) {
2919 ret = rte_eth_promiscuous_enable(portid);
2921 rte_exit(EXIT_FAILURE,
2922 "rte_eth_promiscuous_enable: err=%s, port=%u\n",
2923 rte_strerror(-ret), portid);
2925 /* initialize spinlock for each port */
2926 rte_spinlock_init(&(locks[portid]));
2929 if (!check_ptype(portid))
2930 rte_exit(EXIT_FAILURE,
2931 "PMD can not provide needed ptypes\n");
2934 check_all_ports_link_status(enabled_port_mask);
2936 if (app_mode == APP_MODE_EMPTY_POLL) {
2938 if (empty_poll_train) {
2939 policy.state = TRAINING;
2941 policy.state = MED_NORMAL;
2942 policy.med_base_edpi = ep_med_edpi;
2943 policy.hgh_base_edpi = ep_hgh_edpi;
2946 ret = rte_power_empty_poll_stat_init(&ep_params,
2950 rte_exit(EXIT_FAILURE, "empty poll init failed");
2954 /* launch per-lcore init on every lcore */
2955 if (app_mode == APP_MODE_LEGACY) {
2956 rte_eal_mp_remote_launch(main_legacy_loop, NULL, CALL_MAIN);
2957 } else if (app_mode == APP_MODE_EMPTY_POLL) {
2958 empty_poll_stop = false;
2959 rte_eal_mp_remote_launch(main_empty_poll_loop, NULL,
2961 } else if (app_mode == APP_MODE_TELEMETRY) {
2964 /* Init metrics library */
2965 rte_metrics_init(rte_socket_id());
2966 /** Register stats with metrics library */
2967 for (i = 0; i < NUM_TELSTATS; i++)
2968 ptr_strings[i] = telstats_strings[i].name;
2970 ret = rte_metrics_reg_names(ptr_strings, NUM_TELSTATS);
2972 telstats_index = ret;
2974 rte_exit(EXIT_FAILURE, "failed to register metrics names");
2976 RTE_LCORE_FOREACH_WORKER(lcore_id) {
2977 rte_spinlock_init(&stats[lcore_id].telemetry_lock);
2979 rte_timer_init(&telemetry_timer);
2980 rte_telemetry_register_cmd("/l3fwd-power/stats",
2982 "Returns global power stats. Parameters: None");
2983 rte_eal_mp_remote_launch(main_telemetry_loop, NULL,
2985 } else if (app_mode == APP_MODE_INTERRUPT) {
2986 rte_eal_mp_remote_launch(main_intr_loop, NULL, CALL_MAIN);
2987 } else if (app_mode == APP_MODE_PMD_MGMT) {
2988 /* reuse telemetry loop for PMD power management mode */
2989 rte_eal_mp_remote_launch(main_telemetry_loop, NULL, CALL_MAIN);
2992 if (app_mode == APP_MODE_EMPTY_POLL || app_mode == APP_MODE_TELEMETRY)
2993 launch_timer(rte_lcore_id());
2995 RTE_LCORE_FOREACH_WORKER(lcore_id) {
2996 if (rte_eal_wait_lcore(lcore_id) < 0)
3000 if (app_mode == APP_MODE_PMD_MGMT) {
3001 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
3002 if (rte_lcore_is_enabled(lcore_id) == 0)
3004 qconf = &lcore_conf[lcore_id];
3005 for (queue = 0; queue < qconf->n_rx_queue; ++queue) {
3006 portid = qconf->rx_queue_list[queue].port_id;
3007 queueid = qconf->rx_queue_list[queue].queue_id;
3009 rte_power_ethdev_pmgmt_queue_disable(lcore_id,
3015 RTE_ETH_FOREACH_DEV(portid)
3017 if ((enabled_port_mask & (1 << portid)) == 0)
3020 ret = rte_eth_dev_stop(portid);
3022 RTE_LOG(ERR, L3FWD_POWER, "rte_eth_dev_stop: err=%d, port=%u\n",
3025 rte_eth_dev_close(portid);
3028 if (app_mode == APP_MODE_EMPTY_POLL)
3029 rte_power_empty_poll_stat_free();
3031 if ((app_mode == APP_MODE_LEGACY || app_mode == APP_MODE_EMPTY_POLL) &&
3032 deinit_power_library())
3033 rte_exit(EXIT_FAILURE, "deinit_power_library failed\n");
3035 if (rte_eal_cleanup() < 0)
3036 RTE_LOG(ERR, L3FWD_POWER, "EAL cleanup failed\n");