2 // Copyright (c) 2010-2017 Intel Corporation
4 // Licensed under the Apache License, Version 2.0 (the "License");
5 // you may not use this file except in compliance with the License.
6 // You may obtain a copy of the License at
8 // http://www.apache.org/licenses/LICENSE-2.0
10 // Unless required by applicable law or agreed to in writing, software
11 // distributed under the License is distributed on an "AS IS" BASIS,
12 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 // See the License for the specific language governing permissions and
14 // limitations under the License.
22 #include <rte_cycles.h>
23 #include <rte_atomic.h>
24 #include <rte_table_hash.h>
25 #include <rte_memzone.h>
26 #include <rte_errno.h>
28 #include "prox_malloc.h"
36 #include "prox_args.h"
37 #include "prox_assert.h"
39 #include "prox_shared.h"
40 #include "prox_port_cfg.h"
42 #include "hash_utils.h"
43 #include "handle_lb_net.h"
44 #include "prox_cksum.h"
45 #include "thread_nop.h"
46 #include "thread_generic.h"
47 #include "thread_pipeline.h"
49 #include "handle_master.h"
51 #if RTE_VERSION < RTE_VERSION_NUM(1,8,0,0)
52 #define RTE_CACHE_LINE_SIZE CACHE_LINE_SIZE
55 uint8_t lb_nb_txrings = 0xff;
56 struct rte_ring *ctrl_rings[RTE_MAX_LCORE*MAX_TASKS_PER_CORE];
58 static void __attribute__((noreturn)) prox_usage(const char *prgname)
60 plog_info("\nUsage: %s [-f CONFIG_FILE] [-a|-e] [-m|-s|-i] [-w DEF] [-u] [-t]\n"
61 "\t-f CONFIG_FILE : configuration file to load, ./prox.cfg by default\n"
62 "\t-l LOG_FILE : log file name, ./prox.log by default\n"
63 "\t-p : include PID in log file name if default log file is used\n"
64 "\t-o DISPLAY: Set display to use, can be 'curses' (default), 'cli' or 'none'\n"
65 "\t-v verbosity : initial logging verbosity\n"
66 "\t-a : autostart all cores (by default)\n"
67 "\t-e : don't autostart\n"
68 "\t-n : Create NULL devices instead of using PCI devices, useful together with -i\n"
69 "\t-m : list supported task modes and exit\n"
70 "\t-s : check configuration file syntax and exit\n"
71 "\t-i : check initialization sequence and exit\n"
72 "\t-u : Listen on UDS /tmp/prox.sock\n"
73 "\t-t : Listen on TCP port 8474\n"
74 "\t-q : Pass argument to Lua interpreter, useful to define variables\n"
75 "\t-w : define variable using syntax varname=value\n"
76 "\t takes precedence over variables defined in CONFIG_FILE\n"
77 "\t-k : Log statistics to file \"stats_dump\" in current directory\n"
78 "\t-d : Run as daemon, the parent process will block until PROX is not initialized\n"
79 "\t-z : Ignore CPU topology, implies -i\n"
80 "\t-r : Change initial screen refresh rate. If set to a lower than 0.001 seconds,\n"
81 "\t screen refreshing will be disabled\n"
86 static void check_mixed_normal_pipeline(void)
88 struct lcore_cfg *lconf = NULL;
89 uint32_t lcore_id = -1;
91 while (prox_core_next(&lcore_id, 0) == 0) {
92 lconf = &lcore_cfg[lcore_id];
94 int all_thread_nop = 1;
98 for (uint8_t task_id = 0; task_id < lconf->n_tasks_all; ++task_id) {
99 struct task_args *targ = &lconf->targs[task_id];
100 l3 = !strcmp("l3", targ->sub_mode_str);
101 all_thread_nop = all_thread_nop && !l3 &&
102 targ->task_init->thread_x == thread_nop;
104 pipeline = pipeline || targ->task_init->thread_x == thread_pipeline;
105 generic = generic || targ->task_init->thread_x == thread_generic || l3;
107 PROX_PANIC(generic && pipeline, "Can't run both pipeline and normal thread on same core\n");
110 lconf->thread_x = thread_nop;
112 lconf->thread_x = thread_generic;
117 static void check_zero_rx(void)
119 struct lcore_cfg *lconf = NULL;
120 struct task_args *targ;
122 while (core_targ_next(&lconf, &targ, 0) == 0) {
123 if (targ->nb_rxports != 0) {
124 PROX_PANIC(task_init_flag_set(targ->task_init, TASK_FEATURE_NO_RX),
125 "\tCore %u task %u: rx_ports configured while mode %s does not use it\n", lconf->id, targ->id, targ->task_init->mode_str);
130 static void check_nb_mbuf(void)
132 struct lcore_cfg *lconf = NULL;
133 struct task_args *targ = NULL;
135 int n_txd = 0, n_rxd = 0;
137 while (core_targ_next(&lconf, &targ, 0) == 0) {
138 for (uint8_t i = 0; i < targ->nb_txports; ++i) {
139 port_id = targ->tx_port_queue[i].port;
140 n_txd = prox_port_cfg[port_id].n_txd;
142 for (uint8_t i = 0; i < targ->nb_rxports; ++i) {
143 port_id = targ->rx_port_queue[i].port;
144 n_rxd = prox_port_cfg[port_id].n_rxd;
146 if (targ->nb_mbuf <= n_rxd + n_txd + targ->nb_cache_mbuf + MAX_PKT_BURST) {
147 plog_warn("Core %d, task %d might not have enough mbufs (%d) to support %d txd, %d rxd and %d cache_mbuf\n",
148 lconf->id, targ->id, targ->nb_mbuf, n_txd, n_rxd, targ->nb_cache_mbuf);
153 static void check_missing_rx(void)
155 struct lcore_cfg *lconf = NULL, *rx_lconf = NULL, *tx_lconf = NULL;
156 struct task_args *targ, *rx_targ = NULL, *tx_targ = NULL;
157 uint8_t port_id, rx_port_id, ok;
159 while (core_targ_next(&lconf, &targ, 0) == 0) {
160 PROX_PANIC((targ->flags & TASK_ARG_RX_RING) && targ->rx_rings[0] == 0 && !targ->tx_opt_ring_task,
161 "Configuration Error - Core %u task %u Receiving from ring, but nobody xmitting to this ring\n", lconf->id, targ->id);
162 if (targ->nb_rxports == 0 && targ->nb_rxrings == 0) {
163 PROX_PANIC(!task_init_flag_set(targ->task_init, TASK_FEATURE_NO_RX),
164 "\tCore %u task %u: no rx_ports and no rx_rings configured while required by mode %s\n", lconf->id, targ->id, targ->task_init->mode_str);
169 while (core_targ_next(&lconf, &targ, 0) == 0) {
170 if (strcmp(targ->sub_mode_str, "l3") != 0)
173 PROX_PANIC((targ->nb_rxports == 0) && (targ->nb_txports == 0), "L3 task must have a RX or a TX port\n");
174 // If the L3 sub_mode receives from a port, check that there is at least one core/task
175 // transmitting to this port in L3 sub_mode
176 for (uint8_t i = 0; i < targ->nb_rxports; ++i) {
177 rx_port_id = targ->rx_port_queue[i].port;
180 while (core_targ_next(&tx_lconf, &tx_targ, 0) == 0) {
181 if ((port_id = tx_targ->tx_port_queue[0].port) == OUT_DISCARD)
183 if ((rx_port_id == port_id) && (tx_targ->flags & TASK_ARG_L3)){
188 PROX_PANIC(ok == 0, "RX L3 sub mode for port %d on core %d task %d, but no core/task transmitting on that port\n", rx_port_id, lconf->id, targ->id);
191 // If the L3 sub_mode transmits to a port, check that there is at least one core/task
192 // receiving from that port in L3 sub_mode.
193 if ((port_id = targ->tx_port_queue[0].port) == OUT_DISCARD)
197 plog_info("\tCore %d task %d transmitting to port %d in L3 mode\n", lconf->id, targ->id, port_id);
198 while (core_targ_next(&rx_lconf, &rx_targ, 0) == 0) {
199 for (uint8_t i = 0; i < rx_targ->nb_rxports; ++i) {
200 rx_port_id = rx_targ->rx_port_queue[i].port;
201 if ((rx_port_id == port_id) && (rx_targ->flags & TASK_ARG_L3)){
207 plog_info("\tCore %d task %d has found core %d task %d receiving from port %d\n", lconf->id, targ->id, rx_lconf->id, rx_targ->id, port_id);
211 PROX_PANIC(ok == 0, "L3 sub mode for port %d on core %d task %d, but no core/task receiving on that port\n", port_id, lconf->id, targ->id);
215 static void check_cfg_consistent(void)
220 check_mixed_normal_pipeline();
223 static void plog_all_rings(void)
225 struct lcore_cfg *lconf = NULL;
226 struct task_args *targ;
228 while (core_targ_next(&lconf, &targ, 0) == 0) {
229 for (uint8_t ring_idx = 0; ring_idx < targ->nb_rxrings; ++ring_idx) {
230 plog_info("\tCore %u, task %u, rx_ring[%u] %p\n", lconf->id, targ->id, ring_idx, targ->rx_rings[ring_idx]);
235 static int chain_flag_state(struct task_args *targ, uint64_t flag, int is_set)
237 if (task_init_flag_set(targ->task_init, flag) == is_set)
242 for (uint32_t i = 0; i < targ->n_prev_tasks; ++i) {
243 ret = chain_flag_state(targ->prev_tasks[i], flag, is_set);
250 static int chain_flag_always_set(struct task_args *targ, uint64_t flag)
252 return (!chain_flag_state(targ, flag, 0));
255 static int chain_flag_never_set(struct task_args *targ, uint64_t flag)
257 return (!chain_flag_state(targ, flag, 1));
260 static int chain_flag_sometimes_set(struct task_args *targ, uint64_t flag)
262 return (chain_flag_state(targ, flag, 1));
265 static void configure_if_tx_queues(struct task_args *targ, uint8_t socket)
269 for (uint8_t i = 0; i < targ->nb_txports; ++i) {
270 if_port = targ->tx_port_queue[i].port;
272 PROX_PANIC(if_port == OUT_DISCARD, "port misconfigured, exiting\n");
274 PROX_PANIC(!prox_port_cfg[if_port].active, "\tPort %u not used, skipping...\n", if_port);
276 int dsocket = prox_port_cfg[if_port].socket;
277 if (dsocket != -1 && dsocket != socket) {
278 plog_warn("TX core on socket %d while device on socket %d\n", socket, dsocket);
281 if (prox_port_cfg[if_port].tx_ring[0] == '\0') { // Rings-backed port can use single queue
282 targ->tx_port_queue[i].queue = prox_port_cfg[if_port].n_txq;
283 prox_port_cfg[if_port].n_txq++;
285 prox_port_cfg[if_port].n_txq = 1;
286 targ->tx_port_queue[i].queue = 0;
288 /* By default OFFLOAD is enabled, but if the whole
289 chain has NOOFFLOADS set all the way until the
290 first task that receives from a port, it will be
291 disabled for the destination port. */
292 #if RTE_VERSION < RTE_VERSION_NUM(18,8,0,1)
293 if (chain_flag_always_set(targ, TASK_FEATURE_TXQ_FLAGS_NOOFFLOADS)) {
294 prox_port_cfg[if_port].tx_conf.txq_flags |= ETH_TXQ_FLAGS_NOOFFLOADS;
297 if (chain_flag_always_set(targ, TASK_FEATURE_TXQ_FLAGS_NOOFFLOADS)) {
298 prox_port_cfg[if_port].requested_tx_offload &= ~(DEV_TX_OFFLOAD_IPV4_CKSUM | DEV_TX_OFFLOAD_UDP_CKSUM);
304 static void configure_if_rx_queues(struct task_args *targ, uint8_t socket)
306 struct prox_port_cfg *port;
307 for (int i = 0; i < targ->nb_rxports; i++) {
308 uint8_t if_port = targ->rx_port_queue[i].port;
310 if (if_port == OUT_DISCARD) {
314 port = &prox_port_cfg[if_port];
315 PROX_PANIC(!port->active, "Port %u not used, aborting...\n", if_port);
317 if(port->rx_ring[0] != '\0') {
321 // If the mbuf size (of the rx task) is not big enough, we might receive multiple segments
322 // This is usually the case when setting a big mtu size i.e. enabling jumbo frames.
323 // If the packets get transmitted, then multi segments will have to be enabled on the TX port
324 uint16_t max_frame_size = port->mtu + ETHER_HDR_LEN + ETHER_CRC_LEN + 2 * PROX_VLAN_TAG_SIZE;
325 if (max_frame_size + sizeof(struct rte_mbuf) + RTE_PKTMBUF_HEADROOM > targ->mbuf_size) {
326 targ->task_init->flag_features |= TASK_FEATURE_TXQ_FLAGS_MULTSEGS;
328 targ->rx_port_queue[i].queue = port->n_rxq;
329 port->pool[targ->rx_port_queue[i].queue] = targ->pool;
330 port->pool_size[targ->rx_port_queue[i].queue] = targ->nb_mbuf - 1;
333 int dsocket = port->socket;
334 if (dsocket != -1 && dsocket != socket) {
335 plog_warn("RX core on socket %d while device on socket %d\n", socket, dsocket);
340 static void configure_if_queues(void)
342 struct lcore_cfg *lconf = NULL;
343 struct task_args *targ;
346 while (core_targ_next(&lconf, &targ, 0) == 0) {
347 socket = rte_lcore_to_socket_id(lconf->id);
349 configure_if_rx_queues(targ, socket);
350 configure_if_tx_queues(targ, socket);
354 static void configure_tx_queue_flags(void)
356 struct lcore_cfg *lconf = NULL;
357 struct task_args *targ;
361 while (core_targ_next(&lconf, &targ, 0) == 0) {
362 socket = rte_lcore_to_socket_id(lconf->id);
363 for (uint8_t i = 0; i < targ->nb_txports; ++i) {
364 if_port = targ->tx_port_queue[i].port;
365 #if RTE_VERSION < RTE_VERSION_NUM(18,8,0,1)
366 /* Set the ETH_TXQ_FLAGS_NOREFCOUNT flag if none of
367 the tasks up to the task transmitting to the port
369 if (chain_flag_never_set(targ, TASK_FEATURE_TXQ_FLAGS_REFCOUNT)) {
370 prox_port_cfg[if_port].tx_conf.txq_flags |= ETH_TXQ_FLAGS_NOREFCOUNT;
373 /* Set the DEV_TX_OFFLOAD_MBUF_FAST_FREE flag if none of
374 the tasks up to the task transmitting to the port
375 use refcnt and per-queue all mbufs comes from the same mempool. */
376 if (chain_flag_never_set(targ, TASK_FEATURE_TXQ_FLAGS_REFCOUNT)) {
377 if (chain_flag_never_set(targ, TASK_FEATURE_TXQ_FLAGS_MULTIPLE_MEMPOOL))
378 prox_port_cfg[if_port].requested_tx_offload |= DEV_TX_OFFLOAD_MBUF_FAST_FREE;
385 static void configure_multi_segments(void)
387 struct lcore_cfg *lconf = NULL;
388 struct task_args *targ;
391 while (core_targ_next(&lconf, &targ, 0) == 0) {
392 for (uint8_t i = 0; i < targ->nb_txports; ++i) {
393 if_port = targ->tx_port_queue[i].port;
394 // Multi segment is disabled for most tasks. It is only enabled for tasks requiring big packets.
395 #if RTE_VERSION < RTE_VERSION_NUM(18,8,0,1)
396 // We can only enable "no multi segment" if no such task exists in the chain of tasks.
397 if (chain_flag_never_set(targ, TASK_FEATURE_TXQ_FLAGS_MULTSEGS)) {
398 prox_port_cfg[if_port].tx_conf.txq_flags |= ETH_TXQ_FLAGS_NOMULTSEGS;
401 // We enable "multi segment" if at least one task requires it in the chain of tasks.
402 if (chain_flag_sometimes_set(targ, TASK_FEATURE_TXQ_FLAGS_MULTSEGS)) {
403 prox_port_cfg[if_port].requested_tx_offload |= DEV_TX_OFFLOAD_MULTI_SEGS;
410 static const char *gen_ring_name(void)
412 static char retval[] = "XX";
413 static const char* ring_names =
414 "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
415 "abcdefghijklmnopqrstuvwxyz"
416 "[\\]^_`!\"#$%&'()*+,-./:;<="
422 retval[0] = ring_names[idx % strlen(ring_names)];
423 idx /= strlen(ring_names);
424 retval[1] = idx ? ring_names[(idx - 1) % strlen(ring_names)] : 0;
431 struct ring_init_stats {
432 uint32_t n_pkt_rings;
433 uint32_t n_ctrl_rings;
434 uint32_t n_opt_rings;
437 static uint32_t ring_init_stats_total(const struct ring_init_stats *ris)
439 return ris->n_pkt_rings + ris->n_ctrl_rings + ris->n_opt_rings;
442 static uint32_t count_incoming_tasks(uint32_t lcore_worker, uint32_t dest_task)
444 struct lcore_cfg *lconf = NULL;
445 struct task_args *targ;
449 while (core_targ_next(&lconf, &targ, 0) == 0) {
450 for (uint8_t idxx = 0; idxx < MAX_PROTOCOLS; ++idxx) {
451 for (uint8_t ridx = 0; ridx < targ->core_task_set[idxx].n_elems; ++ridx) {
452 ct = targ->core_task_set[idxx].core_task[ridx];
454 if (dest_task == ct.task && lcore_worker == ct.core)
462 static struct rte_ring *get_existing_ring(uint32_t lcore_id, uint32_t task_id)
464 if (!prox_core_active(lcore_id, 0))
467 struct lcore_cfg *lconf = &lcore_cfg[lcore_id];
469 if (task_id >= lconf->n_tasks_all)
472 if (lconf->targs[task_id].nb_rxrings == 0)
475 return lconf->targs[task_id].rx_rings[0];
478 static struct rte_ring *init_ring_between_tasks(struct lcore_cfg *lconf, struct task_args *starg,
479 const struct core_task ct, uint8_t ring_idx, int idx,
480 struct ring_init_stats *ris)
483 struct rte_ring *ring = NULL;
484 struct lcore_cfg *lworker;
485 struct task_args *dtarg;
487 PROX_ASSERT(prox_core_active(ct.core, 0));
488 lworker = &lcore_cfg[ct.core];
490 /* socket used is the one that the sending core resides on */
491 socket = rte_lcore_to_socket_id(lconf->id);
493 plog_info("\t\tCreating ring on socket %u with size %u\n"
494 "\t\t\tsource core, task and socket = %u, %u, %u\n"
495 "\t\t\tdestination core, task and socket = %u, %u, %u\n"
496 "\t\t\tdestination worker id = %u\n",
497 socket, starg->ring_size,
498 lconf->id, starg->id, socket,
499 ct.core, ct.task, rte_lcore_to_socket_id(ct.core),
503 struct rte_ring **dring = NULL;
505 if (ct.type == CTRL_TYPE_MSG)
506 dring = &lworker->ctrl_rings_m[ct.task];
507 else if (ct.type == CTRL_TYPE_PKT) {
508 dring = &lworker->ctrl_rings_p[ct.task];
509 starg->flags |= TASK_ARG_CTRL_RINGS_P;
513 ring = rte_ring_create(gen_ring_name(), starg->ring_size, socket, RING_F_SC_DEQ);
516 PROX_PANIC(ring == NULL, "Cannot create ring to connect I/O core %u with worker core %u\n", lconf->id, ct.core);
518 starg->tx_rings[starg->tot_n_txrings_inited] = ring;
519 starg->tot_n_txrings_inited++;
521 if (lconf->id == prox_cfg.master) {
522 ctrl_rings[ct.core*MAX_TASKS_PER_CORE + ct.task] = ring;
523 } else if (ct.core == prox_cfg.master) {
524 starg->ctrl_plane_ring = ring;
527 plog_info("\t\t\tCore %u task %u to -> core %u task %u ctrl_ring %s %p %s\n",
528 lconf->id, starg->id, ct.core, ct.task, ct.type == CTRL_TYPE_PKT?
529 "pkt" : "msg", ring, ring->name);
534 dtarg = &lworker->targs[ct.task];
535 lworker->targs[ct.task].worker_thread_id = ring_idx;
536 PROX_ASSERT(dtarg->flags & TASK_ARG_RX_RING);
537 PROX_ASSERT(ct.task < lworker->n_tasks_all);
539 /* If all the following conditions are met, the ring can be
541 if (!task_is_master(starg) && !task_is_master(dtarg) && starg->lconf->id == dtarg->lconf->id &&
542 starg->nb_txrings == 1 && idx == 0 && dtarg->task &&
543 dtarg->tot_rxrings == 1 && starg->task == dtarg->task - 1) {
544 plog_info("\t\tOptimizing away ring on core %u from task %u to task %u\n",
545 dtarg->lconf->id, starg->task, dtarg->task);
546 /* No need to set up ws_mbuf. */
547 starg->tx_opt_ring = 1;
548 /* During init of destination task, the buffer in the
549 source task will be initialized. */
550 dtarg->tx_opt_ring_task = starg;
556 int ring_created = 1;
557 /* Only create multi-producer rings if configured to do so AND
558 there is only one task sending to the task */
559 if ((prox_cfg.flags & DSF_MP_RINGS && count_incoming_tasks(ct.core, ct.task) > 1)
560 || (prox_cfg.flags & DSF_ENABLE_BYPASS)) {
561 ring = get_existing_ring(ct.core, ct.task);
564 plog_info("\t\tCore %u task %u creatign MP ring %p to core %u task %u\n",
565 lconf->id, starg->id, ring, ct.core, ct.task);
569 ring = rte_ring_create(gen_ring_name(), starg->ring_size, socket, RING_F_SC_DEQ);
570 plog_info("\t\tCore %u task %u using MP ring %p from core %u task %u\n",
571 lconf->id, starg->id, ring, ct.core, ct.task);
575 ring = rte_ring_create(gen_ring_name(), starg->ring_size, socket, RING_F_SP_ENQ | RING_F_SC_DEQ);
577 PROX_PANIC(ring == NULL, "Cannot create ring to connect I/O core %u with worker core %u\n", lconf->id, ct.core);
579 starg->tx_rings[starg->tot_n_txrings_inited] = ring;
580 starg->tot_n_txrings_inited++;
583 PROX_ASSERT(dtarg->nb_rxrings < MAX_RINGS_PER_TASK);
584 dtarg->rx_rings[dtarg->nb_rxrings] = ring;
586 if (dtarg->nb_rxrings > 1)
587 dtarg->task_init->flag_features |= TASK_FEATURE_TXQ_FLAGS_MULTIPLE_MEMPOOL;
589 dtarg->nb_slave_threads = starg->core_task_set[idx].n_elems;
590 dtarg->lb_friend_core = lconf->id;
591 dtarg->lb_friend_task = starg->id;
592 plog_info("\t\tWorker thread %d has core %d, task %d as a lb friend\n", ct.core, lconf->id, starg->id);
593 plog_info("\t\tCore %u task %u tx_ring[%u] -> core %u task %u rx_ring[%u] %p %s %u WT\n",
594 lconf->id, starg->id, ring_idx, ct.core, ct.task, dtarg->nb_rxrings, ring, ring->name,
595 dtarg->nb_slave_threads);
600 static void init_rings(void)
602 struct lcore_cfg *lconf = NULL;
603 struct task_args *starg;
604 struct ring_init_stats ris = {0};
606 while (core_targ_next(&lconf, &starg, 1) == 0) {
607 plog_info("\t*** Initializing rings on core %u, task %u ***\n", lconf->id, starg->id);
608 for (uint8_t idx = 0; idx < MAX_PROTOCOLS; ++idx) {
609 for (uint8_t ring_idx = 0; ring_idx < starg->core_task_set[idx].n_elems; ++ring_idx) {
610 PROX_ASSERT(ring_idx < MAX_WT_PER_LB);
611 PROX_ASSERT(starg->tot_n_txrings_inited < MAX_RINGS_PER_TASK);
613 struct core_task ct = starg->core_task_set[idx].core_task[ring_idx];
614 init_ring_between_tasks(lconf, starg, ct, ring_idx, idx, &ris);
619 plog_info("\tInitialized %d rings:\n"
620 "\t\tNumber of packet rings: %u\n"
621 "\t\tNumber of control rings: %u\n"
622 "\t\tNumber of optimized rings: %u\n",
623 ring_init_stats_total(&ris),
629 struct prox_port_cfg *port;
630 while (core_targ_next(&lconf, &starg, 1) == 0) {
631 if ((starg->task_init) && (starg->flags & TASK_ARG_L3)) {
633 ct.core = prox_cfg.master;
635 ct.type = CTRL_TYPE_PKT;
636 struct rte_ring *rx_ring = init_ring_between_tasks(lconf, starg, ct, 0, 0, &ris);
639 ct.task = starg->id;;
640 struct rte_ring *tx_ring = init_ring_between_tasks(&lcore_cfg[prox_cfg.master], lcore_cfg[prox_cfg.master].targs, ct, 0, 0, &ris);
645 static void shuffle_mempool(struct rte_mempool* mempool, uint32_t nb_mbuf)
647 struct rte_mbuf** pkts = prox_zmalloc(nb_mbuf * sizeof(*pkts), rte_socket_id());
650 while ((got < nb_mbuf) && (rte_mempool_get_bulk(mempool, (void**)(pkts + got), 1) == 0))
657 idx = rand() % nb_mbuf;
658 } while (pkts[idx] == 0);
660 rte_mempool_put_bulk(mempool, (void**)&pkts[idx], 1);
667 static void set_mbuf_size(struct task_args *targ)
669 /* mbuf size can be set
670 * - from config file (highest priority, overwriting any other config) - should only be used as workaround
671 * - defaulted to MBUF_SIZE.
672 * Except if set explicitely, ensure that size is big enough for vmxnet3 driver
677 targ->mbuf_size = MBUF_SIZE;
678 struct prox_port_cfg *port;
679 uint16_t max_frame_size = 0, min_buffer_size = 0;
681 for (int i = 0; i < targ->nb_rxports; i++) {
682 uint8_t if_port = targ->rx_port_queue[i].port;
684 if (if_port == OUT_DISCARD) {
687 port = &prox_port_cfg[if_port];
688 if (max_frame_size < port->mtu + ETHER_HDR_LEN + ETHER_CRC_LEN + 2 * PROX_VLAN_TAG_SIZE)
689 max_frame_size = port->mtu + ETHER_HDR_LEN + ETHER_CRC_LEN + 2 * PROX_VLAN_TAG_SIZE;
690 if (min_buffer_size < port->min_rx_bufsize)
691 min_buffer_size = port->min_rx_bufsize;
693 // Check whether we receive from i40e. This driver have extra mbuf size requirements
694 if (strcmp(port->short_name, "i40e") == 0)
698 // i40e supports a maximum of 5 descriptors chained
699 uint16_t required_mbuf_size = RTE_ALIGN(max_frame_size / 5, 128) + sizeof(struct rte_mbuf) + RTE_PKTMBUF_HEADROOM;
700 if (required_mbuf_size > targ->mbuf_size) {
701 targ->mbuf_size = required_mbuf_size;
702 plog_info("\t\tSetting mbuf_size to %u to support frame_size %u\n", targ->mbuf_size, max_frame_size);
705 if (min_buffer_size > targ->mbuf_size) {
706 plog_warn("Mbuf size might be too small. This might result in packet segmentation and memory leak\n");
711 static void setup_mempools_unique_per_socket(void)
715 struct lcore_cfg *lconf = NULL;
716 struct task_args *targ;
718 struct rte_mempool *pool[MAX_SOCKETS];
719 uint32_t mbuf_count[MAX_SOCKETS] = {0};
720 uint32_t nb_cache_mbuf[MAX_SOCKETS] = {0};
721 uint32_t mbuf_size[MAX_SOCKETS] = {0};
723 while (core_targ_next_early(&lconf, &targ, 0) == 0) {
724 PROX_PANIC(targ->task_init == NULL, "task_init = NULL, is mode specified for core %d, task %d ?\n", lconf->id, targ->id);
725 uint8_t socket = rte_lcore_to_socket_id(lconf->id);
726 PROX_ASSERT(socket < MAX_SOCKETS);
729 if (targ->rx_port_queue[0].port != OUT_DISCARD) {
730 struct prox_port_cfg* port_cfg = &prox_port_cfg[targ->rx_port_queue[0].port];
731 PROX_ASSERT(targ->nb_mbuf != 0);
732 mbuf_count[socket] += targ->nb_mbuf;
733 if (nb_cache_mbuf[socket] == 0)
734 nb_cache_mbuf[socket] = targ->nb_cache_mbuf;
736 PROX_PANIC(nb_cache_mbuf[socket] != targ->nb_cache_mbuf,
737 "all mbuf_cache must have the same size if using a unique mempool per socket\n");
739 if (mbuf_size[socket] == 0)
740 mbuf_size[socket] = targ->mbuf_size;
742 PROX_PANIC(mbuf_size[socket] != targ->mbuf_size,
743 "all mbuf_size must have the same size if using a unique mempool per socket\n");
747 for (int i = 0 ; i < MAX_SOCKETS; i++) {
748 if (mbuf_count[i] != 0) {
749 sprintf(name, "socket_%u_pool", i);
750 pool[i] = rte_mempool_create(name,
751 mbuf_count[i] - 1, mbuf_size[i],
753 sizeof(struct rte_pktmbuf_pool_private),
754 rte_pktmbuf_pool_init, NULL,
755 prox_pktmbuf_init, NULL,
757 PROX_PANIC(pool[i] == NULL, "\t\tError: cannot create mempool for socket %u\n", i);
758 plog_info("\t\tMempool %p size = %u * %u cache %u, socket %d\n", pool[i],
759 mbuf_count[i], mbuf_size[i], nb_cache_mbuf[i], i);
761 if (prox_cfg.flags & DSF_SHUFFLE) {
762 shuffle_mempool(pool[i], mbuf_count[i]);
768 while (core_targ_next_early(&lconf, &targ, 0) == 0) {
769 uint8_t socket = rte_lcore_to_socket_id(lconf->id);
771 if (targ->rx_port_queue[0].port != OUT_DISCARD) {
772 /* use this pool for the interface that the core is receiving from */
773 /* If one core receives from multiple ports, all the ports use the same mempool */
774 targ->pool = pool[socket];
775 /* Set the number of mbuf to the number of the unique mempool, so that the used and free work */
776 targ->nb_mbuf = mbuf_count[socket];
777 plog_info("\t\tMempool %p size = %u * %u cache %u, socket %d\n", targ->pool,
778 targ->nb_mbuf, mbuf_size[socket], targ->nb_cache_mbuf, socket);
783 static void setup_mempool_for_rx_task(struct lcore_cfg *lconf, struct task_args *targ)
785 const uint8_t socket = rte_lcore_to_socket_id(lconf->id);
786 struct prox_port_cfg *port_cfg = &prox_port_cfg[targ->rx_port_queue[0].port];
787 const struct rte_memzone *mz;
788 struct rte_mempool *mp = NULL;
790 char memzone_name[64];
795 /* allocate memory pool for packets */
796 PROX_ASSERT(targ->nb_mbuf != 0);
798 if (targ->pool_name[0] == '\0') {
799 sprintf(name, "core_%u_port_%u_pool", lconf->id, targ->id);
802 snprintf(memzone_name, sizeof(memzone_name)-1, "MP_%s", targ->pool_name);
803 mz = rte_memzone_lookup(memzone_name);
806 mp = (struct rte_mempool*)mz->addr;
808 targ->nb_mbuf = mp->size;
812 #ifdef RTE_LIBRTE_IVSHMEM_FALSE
813 if (mz != NULL && mp != NULL && mp->phys_addr != mz->ioremap_addr) {
814 /* Init mbufs with ioremap_addr for dma */
815 mp->phys_addr = mz->ioremap_addr;
816 mp->elt_pa[0] = mp->phys_addr + (mp->elt_va_start - (uintptr_t)mp);
818 struct prox_pktmbuf_reinit_args init_args;
820 init_args.lconf = lconf;
822 uint32_t elt_sz = mp->elt_size + mp->header_size + mp->trailer_size;
823 rte_mempool_obj_iter((void*)mp->elt_va_start, mp->size, elt_sz, 1,
824 mp->elt_pa, mp->pg_num, mp->pg_shift, prox_pktmbuf_reinit, &init_args);
828 /* Use this pool for the interface that the core is
829 receiving from if one core receives from multiple
830 ports, all the ports use the same mempool */
831 if (targ->pool == NULL) {
832 plog_info("\t\tCreating mempool with name '%s'\n", name);
833 targ->pool = rte_mempool_create(name,
834 targ->nb_mbuf - 1, targ->mbuf_size,
836 sizeof(struct rte_pktmbuf_pool_private),
837 rte_pktmbuf_pool_init, NULL,
838 prox_pktmbuf_init, lconf,
842 PROX_PANIC(targ->pool == NULL,
843 "\t\tError: cannot create mempool for core %u port %u: %s\n", lconf->id, targ->id, rte_strerror(rte_errno));
845 plog_info("\t\tMempool %p size = %u * %u cache %u, socket %d\n", targ->pool,
846 targ->nb_mbuf, targ->mbuf_size, targ->nb_cache_mbuf, socket);
847 if (prox_cfg.flags & DSF_SHUFFLE) {
848 shuffle_mempool(targ->pool, targ->nb_mbuf);
852 static void setup_mempools_multiple_per_socket(void)
854 struct lcore_cfg *lconf = NULL;
855 struct task_args *targ;
857 while (core_targ_next_early(&lconf, &targ, 0) == 0) {
858 PROX_PANIC(targ->task_init == NULL, "task_init = NULL, is mode specified for core %d, task %d ?\n", lconf->id, targ->id);
859 if (targ->rx_port_queue[0].port == OUT_DISCARD)
861 setup_mempool_for_rx_task(lconf, targ);
865 static void setup_mempools(void)
867 if (prox_cfg.flags & UNIQUE_MEMPOOL_PER_SOCKET)
868 setup_mempools_unique_per_socket();
870 setup_mempools_multiple_per_socket();
873 static void set_task_lconf(void)
875 struct lcore_cfg *lconf;
876 uint32_t lcore_id = -1;
878 while(prox_core_next(&lcore_id, 1) == 0) {
879 lconf = &lcore_cfg[lcore_id];
880 for (uint8_t task_id = 0; task_id < lconf->n_tasks_all; ++task_id) {
881 lconf->targs[task_id].lconf = lconf;
886 static void set_dest_threads(void)
888 struct lcore_cfg *lconf = NULL;
889 struct task_args *targ;
891 while (core_targ_next(&lconf, &targ, 0) == 0) {
892 for (uint8_t idx = 0; idx < MAX_PROTOCOLS; ++idx) {
893 for (uint8_t ring_idx = 0; ring_idx < targ->core_task_set[idx].n_elems; ++ring_idx) {
894 struct core_task ct = targ->core_task_set[idx].core_task[ring_idx];
896 struct task_args *dest_task = core_targ_get(ct.core, ct.task);
897 dest_task->prev_tasks[dest_task->n_prev_tasks++] = targ;
903 static void setup_all_task_structs_early_init(void)
905 struct lcore_cfg *lconf = NULL;
906 struct task_args *targ;
908 plog_info("\t*** Calling early init on all tasks ***\n");
909 while (core_targ_next(&lconf, &targ, 0) == 0) {
910 if (targ->task_init->early_init) {
911 targ->task_init->early_init(targ);
916 static void setup_all_task_structs(void)
918 struct lcore_cfg *lconf;
919 uint32_t lcore_id = -1;
920 struct task_base *tmaster = NULL;
922 while(prox_core_next(&lcore_id, 1) == 0) {
923 lconf = &lcore_cfg[lcore_id];
924 for (uint8_t task_id = 0; task_id < lconf->n_tasks_all; ++task_id) {
925 if (task_is_master(&lconf->targs[task_id])) {
926 plog_info("\tInitializing MASTER struct for core %d task %d\n", lcore_id, task_id);
927 lconf->tasks_all[task_id] = init_task_struct(&lconf->targs[task_id]);
928 tmaster = lconf->tasks_all[task_id];
932 PROX_PANIC(tmaster == NULL, "Can't initialize master task\n");
935 while(prox_core_next(&lcore_id, 1) == 0) {
936 lconf = &lcore_cfg[lcore_id];
937 plog_info("\tInitializing struct for core %d with %d task\n", lcore_id, lconf->n_tasks_all);
938 for (uint8_t task_id = 0; task_id < lconf->n_tasks_all; ++task_id) {
939 if (!task_is_master(&lconf->targs[task_id])) {
940 plog_info("\tInitializing struct for core %d task %d\n", lcore_id, task_id);
941 lconf->targs[task_id].tmaster = tmaster;
942 lconf->tasks_all[task_id] = init_task_struct(&lconf->targs[task_id]);
948 static void init_port_activate(void)
950 struct lcore_cfg *lconf = NULL;
951 struct task_args *targ;
954 while (core_targ_next_early(&lconf, &targ, 0) == 0) {
955 for (int i = 0; i < targ->nb_rxports; i++) {
956 port_id = targ->rx_port_queue[i].port;
957 prox_port_cfg[port_id].active = 1;
960 for (int i = 0; i < targ->nb_txports; i++) {
961 port_id = targ->tx_port_queue[i].port;
962 prox_port_cfg[port_id].active = 1;
967 /* Initialize cores and allocate mempools */
968 static void init_lcores(void)
970 struct lcore_cfg *lconf = 0;
971 uint32_t lcore_id = -1;
973 while(prox_core_next(&lcore_id, 0) == 0) {
974 uint8_t socket = rte_lcore_to_socket_id(lcore_id);
975 PROX_PANIC(socket + 1 > MAX_SOCKETS, "Can't configure core %u (on socket %u). MAX_SOCKET is set to %d\n", lcore_id, socket, MAX_SOCKETS);
978 /* need to allocate mempools as the first thing to use the lowest possible address range */
979 plog_info("=== Initializing mempools ===\n");
982 lcore_cfg_alloc_hp();
987 plog_info("=== Initializing port addresses ===\n");
990 plog_info("=== Initializing queue numbers on cores ===\n");
991 configure_if_queues();
993 plog_info("=== Initializing rings on cores ===\n");
996 configure_multi_segments();
997 configure_tx_queue_flags();
999 plog_info("=== Checking configuration consistency ===\n");
1000 check_cfg_consistent();
1005 static int setup_prox(int argc, char **argv)
1007 if (prox_read_config_file() != 0 ||
1008 prox_setup_rte(argv[0]) != 0) {
1012 if (prox_cfg.flags & DSF_CHECK_SYNTAX) {
1013 plog_info("=== Configuration file syntax has been checked ===\n\n");
1017 init_port_activate();
1018 plog_info("=== Initializing rte devices ===\n");
1019 if (!(prox_cfg.flags & DSF_USE_DUMMY_DEVICES))
1020 init_rte_ring_dev();
1021 init_rte_dev(prox_cfg.flags & DSF_USE_DUMMY_DEVICES);
1022 plog_info("=== Calibrating TSC overhead ===\n");
1024 plog_info("\tTSC running at %"PRIu64" Hz\n", rte_get_tsc_hz());
1027 plog_info("=== Initializing ports ===\n");
1030 setup_all_task_structs_early_init();
1031 plog_info("=== Initializing tasks ===\n");
1032 setup_all_task_structs();
1034 if (prox_cfg.logbuf_size) {
1035 prox_cfg.logbuf = prox_zmalloc(prox_cfg.logbuf_size, rte_socket_id());
1036 PROX_PANIC(prox_cfg.logbuf == NULL, "Failed to allocate memory for logbuf with size = %d\n", prox_cfg.logbuf_size);
1039 if (prox_cfg.flags & DSF_CHECK_INIT) {
1040 plog_info("=== Initialization sequence completed ===\n\n");
1044 /* Current way that works to disable DPDK logging */
1045 FILE *f = fopen("/dev/null", "r");
1046 rte_openlog_stream(f);
1047 plog_info("=== PROX started ===\n");
1051 static int success = 0;
1052 static void siguser_handler(int signal)
1054 if (signal == SIGUSR1)
1060 static void sigabrt_handler(__attribute__((unused)) int signum)
1062 /* restore default disposition for SIGABRT and SIGPIPE */
1063 signal(SIGABRT, SIG_DFL);
1064 signal(SIGPIPE, SIG_DFL);
1066 /* ignore further Ctrl-C */
1067 signal(SIGINT, SIG_IGN);
1069 /* more drastic exit on tedious termination signal */
1070 plog_info("Aborting...\n");
1071 if (lcore_cfg != NULL) {
1073 pthread_t thread_id, tid0, tid = pthread_self();
1074 memset(&tid0, 0, sizeof(tid0));
1076 /* cancel all threads except current one */
1078 while (prox_core_next(&lcore_id, 1) == 0) {
1079 thread_id = lcore_cfg[lcore_id].thread_id;
1080 if (pthread_equal(thread_id, tid0))
1082 if (pthread_equal(thread_id, tid))
1084 pthread_cancel(thread_id);
1087 /* wait for cancelled threads to terminate */
1089 while (prox_core_next(&lcore_id, 1) == 0) {
1090 thread_id = lcore_cfg[lcore_id].thread_id;
1091 if (pthread_equal(thread_id, tid0))
1093 if (pthread_equal(thread_id, tid))
1095 pthread_join(thread_id, NULL);
1102 /* close ports on termination signal */
1103 close_ports_atexit();
1109 static void sigterm_handler(int signum)
1111 /* abort on second Ctrl-C */
1112 if (signum == SIGINT)
1113 signal(SIGINT, sigabrt_handler);
1115 /* gracefully quit on harmless termination signal */
1116 /* ports will subsequently get closed at resulting exit */
1120 int main(int argc, char **argv)
1122 /* set en_US locale to print big numbers with ',' */
1123 setlocale(LC_NUMERIC, "en_US.utf-8");
1125 if (prox_parse_args(argc, argv) != 0){
1126 prox_usage(argv[0]);
1129 plog_init(prox_cfg.log_name, prox_cfg.log_name_pid);
1130 plog_info("=== " PROGRAM_NAME " " VERSION_STR " ===\n");
1131 plog_info("\tUsing DPDK %s\n", rte_version() + sizeof(RTE_VER_PREFIX));
1134 if (prox_cfg.flags & DSF_LIST_TASK_MODES) {
1135 /* list supported task modes and exit */
1137 return EXIT_SUCCESS;
1140 /* close ports at normal exit */
1141 atexit(close_ports_atexit);
1142 /* gracefully quit on harmless termination signals */
1143 signal(SIGHUP, sigterm_handler);
1144 signal(SIGINT, sigterm_handler);
1145 signal(SIGQUIT, sigterm_handler);
1146 signal(SIGTERM, sigterm_handler);
1147 signal(SIGUSR1, sigterm_handler);
1148 signal(SIGUSR2, sigterm_handler);
1149 /* more drastic exit on tedious termination signals */
1150 signal(SIGABRT, sigabrt_handler);
1151 signal(SIGPIPE, sigabrt_handler);
1153 if (prox_cfg.flags & DSF_DAEMON) {
1154 signal(SIGUSR1, siguser_handler);
1155 signal(SIGUSR2, siguser_handler);
1156 plog_info("=== Running in Daemon mode ===\n");
1157 plog_info("\tForking child and waiting for setup completion\n");
1159 pid_t ppid = getpid();
1162 plog_err("Failed to fork process to run in daemon mode\n");
1163 return EXIT_FAILURE;
1171 kill(ppid, SIGUSR2);
1172 return EXIT_FAILURE;
1174 if (setup_prox(argc, argv) != 0) {
1175 kill(ppid, SIGUSR2);
1176 return EXIT_FAILURE;
1179 kill(ppid, SIGUSR1);
1180 run(prox_cfg.flags);
1181 return EXIT_SUCCESS;
1185 /* Before exiting the parent, wait until the
1186 child process has finished setting up */
1188 if (prox_cfg.logbuf) {
1189 file_print(prox_cfg.logbuf);
1191 return success? EXIT_SUCCESS : EXIT_FAILURE;
1195 if (setup_prox(argc, argv) != 0)
1196 return EXIT_FAILURE;
1197 run(prox_cfg.flags);
1198 return EXIT_SUCCESS;