#include <locale.h>
#include <unistd.h>
#include <signal.h>
+#include <curses.h>
#include <rte_cycles.h>
#include <rte_atomic.h>
}
}
+static void check_nb_mbuf(void)
+{
+ struct lcore_cfg *lconf = NULL;
+ struct task_args *targ = NULL;
+ uint8_t port_id;
+ int n_txd = 0, n_rxd = 0;
+
+ while (core_targ_next(&lconf, &targ, 0) == 0) {
+ for (uint8_t i = 0; i < targ->nb_txports; ++i) {
+ port_id = targ->tx_port_queue[i].port;
+ n_txd = prox_port_cfg[port_id].n_txd;
+ }
+ for (uint8_t i = 0; i < targ->nb_rxports; ++i) {
+ port_id = targ->rx_port_queue[i].port;
+ n_rxd = prox_port_cfg[port_id].n_rxd;
+ }
+ if (targ->nb_mbuf <= n_rxd + n_txd + targ->nb_cache_mbuf + MAX_PKT_BURST) {
+ plog_warn("Core %d, task %d might not have enough mbufs (%d) to support %d txd, %d rxd and %d cache_mbuf\n",
+ lconf->id, targ->id, targ->nb_mbuf, n_txd, n_rxd, targ->nb_cache_mbuf);
+ }
+ }
+}
+
static void check_missing_rx(void)
{
struct lcore_cfg *lconf = NULL, *rx_lconf = NULL, *tx_lconf = NULL;
struct task_args *targ, *rx_targ = NULL, *tx_targ = NULL;
- struct prox_port_cfg *port;
uint8_t port_id, rx_port_id, ok;
while (core_targ_next(&lconf, &targ, 0) == 0) {
static void check_cfg_consistent(void)
{
+ check_nb_mbuf();
check_missing_rx();
check_zero_rx();
check_mixed_normal_pipeline();
return 0;
}
+static int chain_flag_always_set(struct task_args *targ, uint64_t flag)
+{
+ return (!chain_flag_state(targ, flag, 0));
+}
+
+static int chain_flag_never_set(struct task_args *targ, uint64_t flag)
+{
+ return (!chain_flag_state(targ, flag, 1));
+}
+
+static int chain_flag_sometimes_set(struct task_args *targ, uint64_t flag)
+{
+ return (chain_flag_state(targ, flag, 1));
+}
+
static void configure_if_tx_queues(struct task_args *targ, uint8_t socket)
{
uint8_t if_port;
prox_port_cfg[if_port].n_txq = 1;
targ->tx_port_queue[i].queue = 0;
}
- /* Set the ETH_TXQ_FLAGS_NOREFCOUNT flag if none of
- the tasks up to the task transmitting to the port
- does not use refcnt. */
- if (!chain_flag_state(targ, TASK_FEATURE_TXQ_FLAGS_REFCOUNT, 1)) {
- prox_port_cfg[if_port].tx_conf.txq_flags |= ETH_TXQ_FLAGS_NOREFCOUNT;
- plog_info("\t\tEnabling No refcnt on port %d\n", if_port);
- }
- else {
- plog_info("\t\tRefcnt used on port %d\n", if_port);
- }
-
/* By default OFFLOAD is enabled, but if the whole
chain has NOOFFLOADS set all the way until the
first task that receives from a port, it will be
disabled for the destination port. */
- if (chain_flag_state(targ, TASK_FEATURE_TXQ_FLAGS_NOOFFLOADS, 1)) {
+#if RTE_VERSION < RTE_VERSION_NUM(18,8,0,1)
+ if (chain_flag_always_set(targ, TASK_FEATURE_TXQ_FLAGS_NOOFFLOADS)) {
prox_port_cfg[if_port].tx_conf.txq_flags |= ETH_TXQ_FLAGS_NOOFFLOADS;
- plog_info("\t\tDisabling TX offloads on port %d\n", if_port);
- } else {
- plog_info("\t\tEnabling TX offloads on port %d\n", if_port);
}
-
- /* By default NOMULTSEGS is disabled, as drivers/NIC might split packets on RX
- It should only be enabled when we know for sure that the RX does not split packets.
- Set the ETH_TXQ_FLAGS_NOMULTSEGS flag if none of the tasks up to the task
- transmitting to the port does not use multsegs. */
- if (!chain_flag_state(targ, TASK_FEATURE_TXQ_FLAGS_NOMULTSEGS, 0)) {
- prox_port_cfg[if_port].tx_conf.txq_flags |= ETH_TXQ_FLAGS_NOMULTSEGS;
- plog_info("\t\tEnabling No MultiSegs on port %d\n", if_port);
- }
- else {
- plog_info("\t\tMultiSegs used on port %d\n", if_port);
+#else
+ if (chain_flag_always_set(targ, TASK_FEATURE_TXQ_FLAGS_NOOFFLOADS)) {
+ prox_port_cfg[if_port].requested_tx_offload &= ~(DEV_TX_OFFLOAD_IPV4_CKSUM | DEV_TX_OFFLOAD_UDP_CKSUM);
}
+#endif
}
}
static void configure_if_rx_queues(struct task_args *targ, uint8_t socket)
{
+ struct prox_port_cfg *port;
for (int i = 0; i < targ->nb_rxports; i++) {
uint8_t if_port = targ->rx_port_queue[i].port;
return;
}
- PROX_PANIC(!prox_port_cfg[if_port].active, "Port %u not used, aborting...\n", if_port);
+ port = &prox_port_cfg[if_port];
+ PROX_PANIC(!port->active, "Port %u not used, aborting...\n", if_port);
- if(prox_port_cfg[if_port].rx_ring[0] != '\0') {
- prox_port_cfg[if_port].n_rxq = 0;
+ if(port->rx_ring[0] != '\0') {
+ port->n_rxq = 0;
}
- targ->rx_port_queue[i].queue = prox_port_cfg[if_port].n_rxq;
- prox_port_cfg[if_port].pool[targ->rx_port_queue[i].queue] = targ->pool;
- prox_port_cfg[if_port].pool_size[targ->rx_port_queue[i].queue] = targ->nb_mbuf - 1;
- prox_port_cfg[if_port].n_rxq++;
+ // If the mbuf size (of the rx task) is not big enough, we might receive multiple segments
+ // This is usually the case when setting a big mtu size i.e. enabling jumbo frames.
+ // If the packets get transmitted, then multi segments will have to be enabled on the TX port
+ uint16_t max_frame_size = port->mtu + PROX_RTE_ETHER_HDR_LEN + PROX_RTE_ETHER_CRC_LEN + 2 * PROX_VLAN_TAG_SIZE;
+ if (max_frame_size + sizeof(struct rte_mbuf) + RTE_PKTMBUF_HEADROOM > targ->mbuf_size) {
+ targ->task_init->flag_features |= TASK_FEATURE_TXQ_FLAGS_MULTSEGS;
+ }
+ targ->rx_port_queue[i].queue = port->n_rxq;
+ port->pool[targ->rx_port_queue[i].queue] = targ->pool;
+ port->pool_size[targ->rx_port_queue[i].queue] = targ->nb_mbuf - 1;
+ port->n_rxq++;
- int dsocket = prox_port_cfg[if_port].socket;
+ int dsocket = port->socket;
if (dsocket != -1 && dsocket != socket) {
plog_warn("RX core on socket %d while device on socket %d\n", socket, dsocket);
}
while (core_targ_next(&lconf, &targ, 0) == 0) {
socket = rte_lcore_to_socket_id(lconf->id);
- configure_if_tx_queues(targ, socket);
configure_if_rx_queues(targ, socket);
+ configure_if_tx_queues(targ, socket);
+ }
+}
+
+static void configure_tx_queue_flags(void)
+{
+ struct lcore_cfg *lconf = NULL;
+ struct task_args *targ;
+ uint8_t socket;
+ uint8_t if_port;
+
+ while (core_targ_next(&lconf, &targ, 0) == 0) {
+ socket = rte_lcore_to_socket_id(lconf->id);
+ for (uint8_t i = 0; i < targ->nb_txports; ++i) {
+ if_port = targ->tx_port_queue[i].port;
+#if RTE_VERSION < RTE_VERSION_NUM(18,8,0,1)
+ /* Set the ETH_TXQ_FLAGS_NOREFCOUNT flag if none of
+ the tasks up to the task transmitting to the port
+ use refcnt. */
+ if (chain_flag_never_set(targ, TASK_FEATURE_TXQ_FLAGS_REFCOUNT)) {
+ prox_port_cfg[if_port].tx_conf.txq_flags |= ETH_TXQ_FLAGS_NOREFCOUNT;
+ }
+#else
+ /* Set the DEV_TX_OFFLOAD_MBUF_FAST_FREE flag if none of
+ the tasks up to the task transmitting to the port
+ use refcnt and per-queue all mbufs comes from the same mempool. */
+ if (chain_flag_never_set(targ, TASK_FEATURE_TXQ_FLAGS_REFCOUNT)) {
+ if (chain_flag_never_set(targ, TASK_FEATURE_TXQ_FLAGS_MULTIPLE_MEMPOOL))
+ prox_port_cfg[if_port].requested_tx_offload |= DEV_TX_OFFLOAD_MBUF_FAST_FREE;
+ }
+#endif
+ }
+ }
+}
+
+static void configure_multi_segments(void)
+{
+ struct lcore_cfg *lconf = NULL;
+ struct task_args *targ;
+ uint8_t if_port;
+
+ while (core_targ_next(&lconf, &targ, 0) == 0) {
+ for (uint8_t i = 0; i < targ->nb_txports; ++i) {
+ if_port = targ->tx_port_queue[i].port;
+ // Multi segment is disabled for most tasks. It is only enabled for tasks requiring big packets.
+#if RTE_VERSION < RTE_VERSION_NUM(18,8,0,1)
+ // We can only enable "no multi segment" if no such task exists in the chain of tasks.
+ if (chain_flag_never_set(targ, TASK_FEATURE_TXQ_FLAGS_MULTSEGS)) {
+ prox_port_cfg[if_port].tx_conf.txq_flags |= ETH_TXQ_FLAGS_NOMULTSEGS;
+ }
+#else
+ // We enable "multi segment" if at least one task requires it in the chain of tasks.
+ if (chain_flag_sometimes_set(targ, TASK_FEATURE_TXQ_FLAGS_MULTSEGS)) {
+ prox_port_cfg[if_port].requested_tx_offload |= DEV_TX_OFFLOAD_MULTI_SEGS;
+ }
+#endif
+ }
}
}
starg->ctrl_plane_ring = ring;
}
- plog_info("\t\tCore %u task %u to -> core %u task %u ctrl_ring %s %p %s\n",
+ plog_info("\t\t\tCore %u task %u to -> core %u task %u ctrl_ring %s %p %s\n",
lconf->id, starg->id, ct.core, ct.task, ct.type == CTRL_TYPE_PKT?
"pkt" : "msg", ring, ring->name);
ris->n_ctrl_rings++;
PROX_ASSERT(dtarg->nb_rxrings < MAX_RINGS_PER_TASK);
dtarg->rx_rings[dtarg->nb_rxrings] = ring;
++dtarg->nb_rxrings;
+ if (dtarg->nb_rxrings > 1)
+ dtarg->task_init->flag_features |= TASK_FEATURE_TXQ_FLAGS_MULTIPLE_MEMPOOL;
}
dtarg->nb_slave_threads = starg->core_task_set[idx].n_elems;
dtarg->lb_friend_core = lconf->id;
ct.core = lconf->id;
ct.task = starg->id;;
- struct rte_ring *tx_ring = init_ring_between_tasks(lcore_cfg, lcore_cfg[prox_cfg.master].targs, ct, 0, 0, &ris);
+ struct rte_ring *tx_ring = init_ring_between_tasks(&lcore_cfg[prox_cfg.master], lcore_cfg[prox_cfg.master].targs, ct, 0, 0, &ris);
}
}
}
struct rte_mbuf** pkts = prox_zmalloc(nb_mbuf * sizeof(*pkts), rte_socket_id());
uint64_t got = 0;
- while (rte_mempool_get_bulk(mempool, (void**)(pkts + got), 1) == 0)
+ while ((got < nb_mbuf) && (rte_mempool_get_bulk(mempool, (void**)(pkts + got), 1) == 0))
++got;
+ nb_mbuf = got;
while (got) {
int idx;
do {
- idx = rand() % nb_mbuf - 1;
+ idx = rand() % nb_mbuf;
} while (pkts[idx] == 0);
rte_mempool_put_bulk(mempool, (void**)&pkts[idx], 1);
prox_free(pkts);
}
+static void set_mbuf_size(struct task_args *targ)
+{
+ /* mbuf size can be set
+ * - from config file (highest priority, overwriting any other config) - should only be used as workaround
+ * - defaulted to MBUF_SIZE.
+ * Except if set explicitely, ensure that size is big enough for vmxnet3 driver
+ */
+ if (targ->mbuf_size)
+ return;
+
+ targ->mbuf_size = MBUF_SIZE;
+ struct prox_port_cfg *port;
+ uint16_t max_frame_size = 0, min_buffer_size = 0;
+ int i40e = 0;
+ for (int i = 0; i < targ->nb_rxports; i++) {
+ uint8_t if_port = targ->rx_port_queue[i].port;
+
+ if (if_port == OUT_DISCARD) {
+ continue;
+ }
+ port = &prox_port_cfg[if_port];
+ if (max_frame_size < port->mtu + PROX_RTE_ETHER_HDR_LEN + PROX_RTE_ETHER_CRC_LEN + 2 * PROX_VLAN_TAG_SIZE)
+ max_frame_size = port->mtu + PROX_RTE_ETHER_HDR_LEN + PROX_RTE_ETHER_CRC_LEN + 2 * PROX_VLAN_TAG_SIZE;
+ if (min_buffer_size < port->min_rx_bufsize)
+ min_buffer_size = port->min_rx_bufsize;
+
+ // Check whether we receive from i40e. This driver have extra mbuf size requirements
+ if (strcmp(port->short_name, "i40e") == 0)
+ i40e = 1;
+ }
+ if (i40e) {
+ // i40e supports a maximum of 5 descriptors chained
+ uint16_t required_mbuf_size = RTE_ALIGN(max_frame_size / 5, 128) + sizeof(struct rte_mbuf) + RTE_PKTMBUF_HEADROOM;
+ if (required_mbuf_size > targ->mbuf_size) {
+ targ->mbuf_size = required_mbuf_size;
+ plog_info("\t\tSetting mbuf_size to %u to support frame_size %u\n", targ->mbuf_size, max_frame_size);
+ }
+ }
+ if (min_buffer_size > targ->mbuf_size) {
+ plog_warn("Mbuf size might be too small. This might result in packet segmentation and memory leak\n");
+ }
+
+}
+
static void setup_mempools_unique_per_socket(void)
{
uint32_t flags = 0;
uint8_t socket = rte_lcore_to_socket_id(lconf->id);
PROX_ASSERT(socket < MAX_SOCKETS);
- if (targ->mbuf_size_set_explicitely)
- flags = MEMPOOL_F_NO_SPREAD;
- if ((!targ->mbuf_size_set_explicitely) && (targ->task_init->mbuf_size != 0)) {
- targ->mbuf_size = targ->task_init->mbuf_size;
- }
+ set_mbuf_size(targ);
if (targ->rx_port_queue[0].port != OUT_DISCARD) {
struct prox_port_cfg* port_cfg = &prox_port_cfg[targ->rx_port_queue[0].port];
PROX_ASSERT(targ->nb_mbuf != 0);
PROX_PANIC(mbuf_size[socket] != targ->mbuf_size,
"all mbuf_size must have the same size if using a unique mempool per socket\n");
}
- if ((!targ->mbuf_size_set_explicitely) && (strcmp(port_cfg->short_name, "vmxnet3") == 0)) {
- if (mbuf_size[socket] < MBUF_SIZE + RTE_PKTMBUF_HEADROOM)
- mbuf_size[socket] = MBUF_SIZE + RTE_PKTMBUF_HEADROOM;
- }
}
}
for (int i = 0 ; i < MAX_SOCKETS; i++) {
char memzone_name[64];
char name[64];
- /* mbuf size can be set
- * - from config file (highest priority, overwriting any other config) - should only be used as workaround
- * - through each 'mode', overwriting the default mbuf_size
- * - defaulted to MBUF_SIZE i.e. 1518 Bytes
- * Except is set expliciteky, ensure that size is big enough for vmxnet3 driver
- */
- if (targ->mbuf_size_set_explicitely) {
- flags = MEMPOOL_F_NO_SPREAD;
- /* targ->mbuf_size already set */
- }
- else if (targ->task_init->mbuf_size != 0) {
- /* mbuf_size not set through config file but set through mode */
- targ->mbuf_size = targ->task_init->mbuf_size;
- }
- else if (strcmp(port_cfg->short_name, "vmxnet3") == 0) {
- if (targ->mbuf_size < MBUF_SIZE + RTE_PKTMBUF_HEADROOM)
- targ->mbuf_size = MBUF_SIZE + RTE_PKTMBUF_HEADROOM;
- }
+ set_mbuf_size(targ);
/* allocate memory pool for packets */
PROX_ASSERT(targ->nb_mbuf != 0);
plog_info("=== Initializing rings on cores ===\n");
init_rings();
+ configure_multi_segments();
+ configure_tx_queue_flags();
+
plog_info("=== Checking configuration consistency ===\n");
check_cfg_consistent();
plog_all_rings();
-
- setup_all_task_structs_early_init();
- plog_info("=== Initializing tasks ===\n");
- setup_all_task_structs();
}
static int setup_prox(int argc, char **argv)
plog_info("=== Initializing ports ===\n");
init_port_all();
+ setup_all_task_structs_early_init();
+ plog_info("=== Initializing tasks ===\n");
+ setup_all_task_structs();
+
if (prox_cfg.logbuf_size) {
prox_cfg.logbuf = prox_zmalloc(prox_cfg.logbuf_size, rte_socket_id());
PROX_PANIC(prox_cfg.logbuf == NULL, "Failed to allocate memory for logbuf with size = %d\n", prox_cfg.logbuf_size);
quit();
}
+static void set_term_env(void)
+{
+ static const char var[] = "TERM";
+ static char str[] = "TERM=putty";
+ char *old_value, *new_value;
+ int max_ver = 0, min_ver = 0, n;
+
+ old_value = getenv(var);
+
+ const char *ncurses_version = curses_version();
+ n = sscanf(ncurses_version, "ncurses %d.%d", &max_ver, &min_ver);
+ if (n != 2) {
+ plog_info("\tUnable to extract ncurses version from %s. TERM left unchanged to %s\n", ncurses_version, old_value);
+ return;
+ } else {
+ plog_info("\tncurses version = %d.%d (%s)\n", max_ver, min_ver, ncurses_version);
+ }
+
+ if (((max_ver > 6) || ((max_ver == 6) && (min_ver >= 1))) && (strcmp(old_value, "xterm") == 0)) {
+ // On recent OSes such as RHEL 8.0, ncurses(6.1) introduced support
+ // for ECMA-48 repeat character control.
+ // Some terminal emulators use TERM=xterm but do not support this feature.
+ // In this case, printing repeating character such as "22000000 Hz" might
+ // display as 220 Hz.
+ // Other emulattors, such as tmux, use TERM=screen, and do not exhibit the issue.
+ plog_info("\tChanged TERM from %s ", old_value);
+ putenv(str);
+ new_value = getenv(var);
+ plog_info("to %s\n", new_value);
+ } else {
+ plog_info("\tTERM left unchanged to %s\n", old_value);
+ }
+}
+
int main(int argc, char **argv)
{
/* set en_US locale to print big numbers with ',' */
if (prox_parse_args(argc, argv) != 0){
prox_usage(argv[0]);
}
-
plog_init(prox_cfg.log_name, prox_cfg.log_name_pid);
- plog_info("=== " PROGRAM_NAME " " VERSION_STR " ===\n");
+ plog_info("=== " PROGRAM_NAME " %s ===\n", VERSION_STR());
plog_info("\tUsing DPDK %s\n", rte_version() + sizeof(RTE_VER_PREFIX));
+ set_term_env();
read_rdt_info();
if (prox_cfg.flags & DSF_LIST_TASK_MODES) {
Code Review / samplevnf.git / blobdiff