Modified the application to support for multiple queues
per interface. Enabled inequal worker thread counts.
Signed-off-by: Petri Savolainen <petri.savolai...@nokia.com>
---
test/performance/odp_l2fwd.c | 413 ++++++++++++++++++++++++++++++++++++-------
1 file changed, 351 insertions(+), 62 deletions(-)
diff --git a/test/performance/odp_l2fwd.c b/test/performance/odp_l2fwd.c
index d201186..67a23ed 100644
--- a/test/performance/odp_l2fwd.c
+++ b/test/performance/odp_l2fwd.c
@@ -45,6 +45,12 @@
*/
#define MAX_PKT_BURST 32
+/** Maximum number of pktio queues per interface */
+#define MAX_QUEUES 32
+
+/** Maximum number of pktio interfaces */
+#define MAX_PKTIOS 8
+
/**
* Packet input mode
*/
@@ -64,6 +70,7 @@ typedef enum pkt_in_mode_t {
typedef struct {
int cpu_count;
int if_count; /**< Number of interfaces to be used */
+ int num_workers; /**< Number of worker threads */
char **if_names; /**< Array of pointers to interface names */
pkt_in_mode_t mode; /**< Packet input mode */
int time; /**< Time in seconds to run. */
@@ -95,8 +102,20 @@ typedef union {
/**
* Thread specific arguments
*/
-typedef struct {
- int src_idx; /**< Source interface identifier */
+typedef struct thread_args_t {
+ int num_pktio;
+
+ struct {
+ odp_pktio_t rx_pktio;
+ odp_pktio_t tx_pktio;
+ odp_pktin_queue_t pktin;
+ odp_pktout_queue_t pktout;
+ int rx_idx;
+ int tx_idx;
+ int rx_queue_idx;
+ int tx_queue_idx;
+ } pktio[MAX_PKTIOS];
+
stats_t *stats; /**< Pointer to per thread stats */
} thread_args_t;
@@ -110,14 +129,24 @@ typedef struct {
appl_args_t appl;
/** Thread specific arguments */
thread_args_t thread[MAX_WORKERS];
- /** Table of pktio handles */
- odp_pktio_t pktios[ODP_CONFIG_PKTIO_ENTRIES];
/** Table of port ethernet addresses */
- odph_ethaddr_t port_eth_addr[ODP_CONFIG_PKTIO_ENTRIES];
+ odph_ethaddr_t port_eth_addr[MAX_PKTIOS];
/** Table of dst ethernet addresses */
- odph_ethaddr_t dst_eth_addr[ODP_CONFIG_PKTIO_ENTRIES];
+ odph_ethaddr_t dst_eth_addr[MAX_PKTIOS];
/** Table of dst ports */
- int dst_port[ODP_CONFIG_PKTIO_ENTRIES];
+ int dst_port[MAX_PKTIOS];
+ /** Table of pktio handles */
+ struct {
+ odp_pktio_t pktio;
+ odp_pktin_queue_t pktin[MAX_QUEUES];
+ odp_pktout_queue_t pktout[MAX_QUEUES];
+ int num_rx_thr;
+ int num_tx_thr;
+ int num_rx_queue;
+ int num_tx_queue;
+ int next_rx_queue;
+ int next_tx_queue;
+ } pktios[MAX_PKTIOS];
} args_t;
/** Global pointer to args */
@@ -190,7 +219,7 @@ static void *pktio_queue_thread(void *arg)
/* packets from the same queue are from the same interface */
dst_idx = lookup_dest_port(pkt_tbl[0]);
fill_eth_addrs(pkt_tbl, pkts, dst_idx);
- pktio_dst = gbl_args->pktios[dst_idx];
+ pktio_dst = gbl_args->pktios[dst_idx].pktio;
sent = odp_pktio_send(pktio_dst, pkt_tbl, pkts);
@@ -226,8 +255,9 @@ static inline int lookup_dest_port(odp_packet_t pkt)
pktio_src = odp_packet_input(pkt);
- for (src_idx = -1, i = 0; gbl_args->pktios[i] != ODP_PKTIO_INVALID; ++i)
- if (gbl_args->pktios[i] == pktio_src)
+ for (src_idx = -1, i = 0; gbl_args->pktios[i].pktio
+ != ODP_PKTIO_INVALID; ++i)
+ if (gbl_args->pktios[i].pktio == pktio_src)
src_idx = i;
if (src_idx == -1)
@@ -264,32 +294,29 @@ static void *pktio_direct_recv_thread(void *arg)
int thr;
int pkts;
odp_packet_t pkt_tbl[MAX_PKT_BURST];
- int src_idx, dst_idx;
- odp_pktio_t pktio_src, pktio_dst;
+ int dst_idx, num_pktio;
+ odp_pktin_queue_t pktin;
+ odp_pktout_queue_t pktout;
+ int pktio = 0;
thread_args_t *thr_args = arg;
stats_t *stats = thr_args->stats;
thr = odp_thread_id();
- src_idx = thr_args->src_idx;
- dst_idx = gbl_args->dst_port[src_idx];
- pktio_src = gbl_args->pktios[src_idx];
- pktio_dst = gbl_args->pktios[dst_idx];
+ num_pktio = thr_args->num_pktio;
+ dst_idx = thr_args->pktio[pktio].tx_idx;
+ pktin = thr_args->pktio[pktio].pktin;
+ pktout = thr_args->pktio[pktio].pktout;
- printf("[%02i] srcif:%s dstif:%s spktio:%02" PRIu64
- " dpktio:%02" PRIu64 " DIRECT RECV mode\n",
- thr,
- gbl_args->appl.if_names[src_idx],
- gbl_args->appl.if_names[dst_idx],
- odp_pktio_to_u64(pktio_src), odp_pktio_to_u64(pktio_dst));
+ printf("[%02i] num pktios %i, DIRECT RECV mode\n", thr, num_pktio);
odp_barrier_wait(&barrier);
/* Loop packets */
while (!exit_threads) {
- int sent, i;
+ int sent;
unsigned tx_drops;
- pkts = odp_pktio_recv(pktio_src, pkt_tbl, MAX_PKT_BURST);
+ pkts = odp_pktio_recv_queue(pktin, pkt_tbl, MAX_PKT_BURST);
if (odp_unlikely(pkts <= 0))
continue;
@@ -310,12 +337,14 @@ static void *pktio_direct_recv_thread(void *arg)
fill_eth_addrs(pkt_tbl, pkts, dst_idx);
- sent = odp_pktio_send(pktio_dst, pkt_tbl, pkts);
+ sent = odp_pktio_send_queue(pktout, pkt_tbl, pkts);
sent = odp_unlikely(sent < 0) ? 0 : sent;
tx_drops = pkts - sent;
if (odp_unlikely(tx_drops)) {
+ int i;
+
stats->s.tx_drops += tx_drops;
/* Drop rejected packets */
@@ -324,6 +353,16 @@ static void *pktio_direct_recv_thread(void *arg)
}
stats->s.packets += pkts;
+
+ if (num_pktio > 1) {
+ dst_idx = thr_args->pktio[pktio].tx_idx;
+ pktin = thr_args->pktio[pktio].pktin;
+ pktout = thr_args->pktio[pktio].pktout;
+ pktio++;
+ if (pktio == num_pktio)
+ pktio = 0;
+ }
+
}
/* Make sure that the last stats write is visible to readers */
@@ -335,13 +374,15 @@ static void *pktio_direct_recv_thread(void *arg)
/**
* Create a pktio handle, optionally associating a default input queue.
*
- * @param dev Name of device to open
- * @param pool Pool to associate with device for packet RX/TX
+ * @param dev Name of device to open
+ * @param index Pktio index
+ * @param pool Pool to associate with device for packet RX/TX
*
- * @return The handle of the created pktio object.
- * @retval ODP_PKTIO_INVALID if the create fails.
+ * @retval 0 on success
+ * @retval -1 on failure
*/
-static odp_pktio_t create_pktio(const char *dev, odp_pool_t pool)
+static int create_pktio(const char *dev, int index, int num_rx, int num_tx,
+ odp_pool_t pool)
{
char inq_name[ODP_QUEUE_NAME_LEN];
odp_queue_param_t qparam;
@@ -350,26 +391,94 @@ static odp_pktio_t create_pktio(const char *dev,
odp_pool_t pool)
int ret;
odp_pktio_param_t pktio_param;
odp_schedule_sync_t sync_mode;
+ odp_pktio_capability_t capa;
+ odp_pktio_input_queue_param_t in_queue_param;
+ odp_pktio_output_queue_param_t out_queue_param;
+ odp_bool_t single_rx = 1;
+ odp_bool_t single_tx = 1;
odp_pktio_param_init(&pktio_param);
- if (gbl_args->appl.mode == DIRECT_RECV)
+ if (gbl_args->appl.mode == DIRECT_RECV) {
pktio_param.in_mode = ODP_PKTIN_MODE_RECV;
- else
+ pktio_param.out_mode = ODP_PKTOUT_MODE_SEND;
+ } else {
pktio_param.in_mode = ODP_PKTIN_MODE_SCHED;
+ }
pktio = odp_pktio_open(dev, pool, &pktio_param);
if (pktio == ODP_PKTIO_INVALID) {
LOG_ERR("Error: failed to open %s\n", dev);
- return ODP_PKTIO_INVALID;
+ return -1;
}
printf("created pktio %" PRIu64 " (%s)\n",
odp_pktio_to_u64(pktio), dev);
- /* no further setup needed for direct receive mode */
- if (gbl_args->appl.mode == DIRECT_RECV)
- return pktio;
+ if (odp_pktio_capability(pktio, &capa)) {
+ LOG_ERR("Error: capability query failed %s\n", dev);
+ return -1;
+ }
+
+ if (gbl_args->appl.mode == DIRECT_RECV) {
+ if (num_rx > (int)capa.max_input_queues) {
+ printf("Sharing %i input queues between %i workers\n",
+ capa.max_input_queues, num_rx);
+ num_rx = capa.max_input_queues;
+ single_rx = 0;
+ }
+
+ if (num_tx > (int)capa.max_output_queues) {
+ printf("Sharing %i output queues between %i workers\n",
+ capa.max_output_queues, num_tx);
+ num_tx = capa.max_output_queues;
+ single_tx = 0;
+ }
+
+ odp_pktio_input_queue_param_init(&in_queue_param);
+ odp_pktio_output_queue_param_init(&out_queue_param);
+
+ in_queue_param.single_user = single_rx;
+ in_queue_param.hash_enable = 1;
+ in_queue_param.hash_proto.proto.ipv4_udp = 1;
+ in_queue_param.num_queues = num_rx;
+
+ if (odp_pktio_input_queues_config(pktio, &in_queue_param)) {
+ LOG_ERR("Error: input queue config failed %s\n", dev);
+ return -1;
+ }
+
+ out_queue_param.single_user = single_tx;
+ out_queue_param.num_queues = num_tx;
+
+ if (odp_pktio_output_queues_config(pktio, &out_queue_param)) {
+ LOG_ERR("Error: output queue config failed %s\n", dev);
+ return -1;
+ }
+
+ if (odp_pktio_pktin_queues(pktio,
+ gbl_args->pktios[index].pktin,
+ num_rx) != num_rx) {
+ LOG_ERR("Error: pktin queue query failed %s\n", dev);
+ return -1;
+ }
+
+ if (odp_pktio_pktout_queues(pktio,
+ gbl_args->pktios[index].pktout,
+ num_tx) != num_tx) {
+ LOG_ERR("Error: pktout queue query failed %s\n", dev);
+ return -1;
+ }
+
+ printf("created %i input and %i output queues on (%s)\n",
+ num_rx, num_tx, dev);
+
+ gbl_args->pktios[index].num_rx_queue = num_rx;
+ gbl_args->pktios[index].num_tx_queue = num_tx;
+ gbl_args->pktios[index].pktio = pktio;
+
+ return 0;
+ }
if (gbl_args->appl.mode == SCHED_ATOMIC)
sync_mode = ODP_SCHED_SYNC_ATOMIC;
@@ -389,16 +498,18 @@ static odp_pktio_t create_pktio(const char *dev,
odp_pool_t pool)
inq_def = odp_queue_create(inq_name, ODP_QUEUE_TYPE_PKTIN, &qparam);
if (inq_def == ODP_QUEUE_INVALID) {
LOG_ERR("Error: pktio queue creation failed\n");
- return ODP_PKTIO_INVALID;
+ return -1;
}
ret = odp_pktio_inq_setdef(pktio, inq_def);
if (ret != 0) {
LOG_ERR("Error: default input-Q setup\n");
- return ODP_PKTIO_INVALID;
+ return -1;
}
- return pktio;
+ gbl_args->pktios[index].pktio = pktio;
+
+ return 0;
}
/**
@@ -464,6 +575,167 @@ static int print_speed_stats(int num_workers, stats_t
*thr_stats,
return pkts > 100 ? 0 : -1;
}
+static void print_port_mapping(void)
+{
+ int if_count, num_workers;
+ int thr, pktio;
+
+ if_count = gbl_args->appl.if_count;
+ num_workers = gbl_args->appl.num_workers;
+
+ printf("\nWorker mapping table (port[queue])\n--------------------\n");
+
+ for (thr = 0; thr < num_workers; thr++) {
+ int rx_idx, tx_idx;
+ int rx_queue_idx, tx_queue_idx;
+ thread_args_t *thr_args = &gbl_args->thread[thr];
+ int num = thr_args->num_pktio;
+
+ printf("Worker %i\n", thr);
+
+ for (pktio = 0; pktio < num; pktio++) {
+ rx_idx = thr_args->pktio[pktio].rx_idx;
+ tx_idx = thr_args->pktio[pktio].tx_idx;
+ rx_queue_idx = thr_args->pktio[pktio].rx_queue_idx;
+ tx_queue_idx = thr_args->pktio[pktio].tx_queue_idx;
+ printf(" %i[%i] -> %i[%i]\n",
+ rx_idx, rx_queue_idx, tx_idx, tx_queue_idx);
+ }
+ }
+
+ printf("\nPort config\n--------------------\n");
+
+ for (pktio = 0; pktio < if_count; pktio++) {
+ const char *dev = gbl_args->appl.if_names[pktio];
+
+ printf("Port %i (%s)\n", pktio, dev);
+ printf(" rx workers %i\n",
+ gbl_args->pktios[pktio].num_rx_thr);
+ printf(" tx workers %i\n",
+ gbl_args->pktios[pktio].num_tx_thr);
+ printf(" rx queues %i\n",
+ gbl_args->pktios[pktio].num_rx_queue);
+ printf(" tx queues %i\n",
+ gbl_args->pktios[pktio].num_tx_queue);
+ }
+
+ printf("\n");
+}
+
+/*
+ * Bind worker threads to interfaces and calculate number of queues needed
+ *
+ * less workers (N) than interfaces (M)
+ * - assign each worker to process every Nth interface
+ * - workers process inequal number of interfaces, when M is not divisible by
N
+ * - needs only single queue per interface
+ * otherwise
+ * - assign an interface to every Mth worker
+ * - interfaces are processed by inequal number of workers, when N is not
+ * divisible by M
+ * - tries to configure a queue per worker per interface
+ * - shares queues, if interface capability does not allows a queue per worker
+ */
+static void bind_workers(void)
+{
+ int if_count, num_workers;
+ int rx_idx, tx_idx, thr, pktio;
+ thread_args_t *thr_args;
+
+ if_count = gbl_args->appl.if_count;
+ num_workers = gbl_args->appl.num_workers;
+
+ /* initialize port forwarding table */
+ for (rx_idx = 0; rx_idx < if_count; rx_idx++)
+ gbl_args->dst_port[rx_idx] = find_dest_port(rx_idx);
+
+ if (if_count > num_workers) {
+ thr = 0;
+
+ for (rx_idx = 0; rx_idx < if_count; rx_idx++) {
+ thr_args = &gbl_args->thread[thr];
+ pktio = thr_args->num_pktio;
+ tx_idx = gbl_args->dst_port[rx_idx];
+ thr_args->pktio[pktio].rx_idx = rx_idx;
+ thr_args->pktio[pktio].tx_idx = tx_idx;
+ thr_args->num_pktio++;
+
+ gbl_args->pktios[rx_idx].num_rx_thr++;
+ gbl_args->pktios[tx_idx].num_tx_thr++;
+
+ thr++;
+ if (thr >= num_workers)
+ thr = 0;
+ }
+ } else {
+ rx_idx = 0;
+
+ for (thr = 0; thr < num_workers; thr++) {
+ thr_args = &gbl_args->thread[thr];
+ pktio = thr_args->num_pktio;
+ tx_idx = gbl_args->dst_port[rx_idx];
+ thr_args->pktio[pktio].rx_idx = rx_idx;
+ thr_args->pktio[pktio].tx_idx = tx_idx;
+ thr_args->num_pktio++;
+
+ gbl_args->pktios[rx_idx].num_rx_thr++;
+ gbl_args->pktios[tx_idx].num_tx_thr++;
+
+ rx_idx++;
+ if (rx_idx >= if_count)
+ rx_idx = 0;
+ }
+ }
+}
+
+/*
+ * Bind queues to threads and fill in missing thread arguments (handles)
+ */
+static void bind_queues(void)
+{
+ int num_workers;
+ int thr, pktio;
+
+ num_workers = gbl_args->appl.num_workers;
+
+ for (thr = 0; thr < num_workers; thr++) {
+ int rx_idx, tx_idx;
+ thread_args_t *thr_args = &gbl_args->thread[thr];
+ int num = thr_args->num_pktio;
+
+ for (pktio = 0; pktio < num; pktio++) {
+ int rx_queue, tx_queue;
+
+ rx_idx = thr_args->pktio[pktio].rx_idx;
+ tx_idx = thr_args->pktio[pktio].tx_idx;
+ rx_queue = gbl_args->pktios[rx_idx].next_rx_queue;
+ tx_queue = gbl_args->pktios[tx_idx].next_tx_queue;
+
+ thr_args->pktio[pktio].rx_queue_idx = rx_queue;
+ thr_args->pktio[pktio].tx_queue_idx = tx_queue;
+ thr_args->pktio[pktio].pktin =
+ gbl_args->pktios[rx_idx].pktin[rx_queue];
+ thr_args->pktio[pktio].pktout =
+ gbl_args->pktios[tx_idx].pktout[tx_queue];
+ thr_args->pktio[pktio].rx_pktio =
+ gbl_args->pktios[rx_idx].pktio;
+ thr_args->pktio[pktio].tx_pktio =
+ gbl_args->pktios[tx_idx].pktio;
+
+ rx_queue++;
+ tx_queue++;
+
+ if (rx_queue >= gbl_args->pktios[rx_idx].num_rx_queue)
+ rx_queue = 0;
+ if (tx_queue >= gbl_args->pktios[tx_idx].num_tx_queue)
+ tx_queue = 0;
+
+ gbl_args->pktios[rx_idx].next_rx_queue = rx_queue;
+ gbl_args->pktios[tx_idx].next_tx_queue = tx_queue;
+ }
+ }
+}
+
/**
* ODP L2 forwarding main function
*/
@@ -478,10 +750,10 @@ int main(int argc, char *argv[])
odp_cpumask_t cpumask;
char cpumaskstr[ODP_CPUMASK_STR_SIZE];
odph_ethaddr_t new_addr;
- odp_pktio_t pktio;
odp_pool_param_t params;
int ret;
stats_t *stats;
+ int if_count;
/* Init ODP before calling anything else */
if (odp_init_global(NULL, NULL)) {
@@ -521,16 +793,13 @@ int main(int argc, char *argv[])
num_workers = odp_cpumask_default_worker(&cpumask, num_workers);
(void)odp_cpumask_to_str(&cpumask, cpumaskstr, sizeof(cpumaskstr));
+ gbl_args->appl.num_workers = num_workers;
+ if_count = gbl_args->appl.if_count;
+
printf("num worker threads: %i\n", num_workers);
printf("first CPU: %i\n", odp_cpumask_first(&cpumask));
printf("cpu mask: %s\n", cpumaskstr);
- if (num_workers < gbl_args->appl.if_count) {
- LOG_ERR("Error: CPU count %d less than interface count\n",
- num_workers);
- exit(EXIT_FAILURE);
- }
-
/* Create packet pool */
odp_pool_param_init(¶ms);
params.pkt.seg_len = SHM_PKT_POOL_BUF_SIZE;
@@ -546,14 +815,28 @@ int main(int argc, char *argv[])
}
odp_pool_print(pool);
- for (i = 0; i < gbl_args->appl.if_count; ++i) {
- pktio = create_pktio(gbl_args->appl.if_names[i], pool);
- if (pktio == ODP_PKTIO_INVALID)
+ bind_workers();
+
+ for (i = 0; i < if_count; ++i) {
+ const char *dev = gbl_args->appl.if_names[i];
+ int num_rx, num_tx;
+
+ /* Current default values for other than direct mode */
+ num_rx = 1;
+ num_tx = 1;
+
+ if (gbl_args->appl.mode == DIRECT_RECV) {
+ /* Request a queue per thread */
+ num_rx = gbl_args->pktios[i].num_rx_thr;
+ num_tx = gbl_args->pktios[i].num_tx_thr;
+ }
+
+ if (create_pktio(dev, i, num_rx, num_tx, pool))
exit(EXIT_FAILURE);
- gbl_args->pktios[i] = pktio;
/* Save interface ethernet address */
- if (odp_pktio_mac_addr(pktio, gbl_args->port_eth_addr[i].addr,
+ if (odp_pktio_mac_addr(gbl_args->pktios[i].pktio,
+ gbl_args->port_eth_addr[i].addr,
ODPH_ETHADDR_LEN) != ODPH_ETHADDR_LEN) {
LOG_ERR("Error: interface ethernet address unknown\n");
exit(EXIT_FAILURE);
@@ -567,12 +850,14 @@ int main(int argc, char *argv[])
new_addr.addr[5] = i;
gbl_args->dst_eth_addr[i] = new_addr;
}
-
- /* Save interface destination port */
- gbl_args->dst_port[i] = find_dest_port(i);
}
- gbl_args->pktios[i] = ODP_PKTIO_INVALID;
+ gbl_args->pktios[i].pktio = ODP_PKTIO_INVALID;
+
+ bind_queues();
+
+ if (gbl_args->appl.mode == DIRECT_RECV)
+ print_port_mapping();
memset(thread_tbl, 0, sizeof(thread_tbl));
@@ -586,12 +871,12 @@ int main(int argc, char *argv[])
odp_cpumask_t thd_mask;
void *(*thr_run_func) (void *);
- if (gbl_args->appl.mode == DIRECT_RECV)
+ if (gbl_args->appl.mode == DIRECT_RECV) {
thr_run_func = pktio_direct_recv_thread;
- else /* SCHED_NONE / SCHED_ATOMIC / SCHED_ORDERED */
+ } else { /* SCHED_NONE / SCHED_ATOMIC / SCHED_ORDERED */
thr_run_func = pktio_queue_thread;
+ }
- gbl_args->thread[i].src_idx = i % gbl_args->appl.if_count;
gbl_args->thread[i].stats = &stats[i];
odp_cpumask_zero(&thd_mask);
@@ -603,8 +888,10 @@ int main(int argc, char *argv[])
}
/* Start packet receive and transmit */
- for (i = 0; i < gbl_args->appl.if_count; ++i) {
- pktio = gbl_args->pktios[i];
+ for (i = 0; i < if_count; ++i) {
+ odp_pktio_t pktio;
+
+ pktio = gbl_args->pktios[i].pktio;
ret = odp_pktio_start(pktio);
if (ret) {
LOG_ERR("Error: unable to start %s\n",
@@ -761,7 +1048,8 @@ static void parse_args(int argc, char *argv[], appl_args_t
*appl_args)
appl_args->if_count = i;
- if (appl_args->if_count == 0) {
+ if (appl_args->if_count < 1 ||
+ appl_args->if_count > MAX_PKTIOS) {
usage(argv[0]);
exit(EXIT_FAILURE);
}
@@ -870,6 +1158,7 @@ static void usage(char *progname)
"\n"
"Mandatory OPTIONS:\n"
" -i, --interface Eth interfaces (comma-separated, no spaces)\n"
+ " Interface count min 1, max %i\n"
"\n"
"Optional OPTIONS\n"
" -m, --mode 0: Send&receive packets directly from NIC
(default)\n"
@@ -891,6 +1180,6 @@ static void usage(char *progname)
" ODP_PKTIO_DISABLE_SOCKET_MMAP\n"
" ODP_PKTIO_DISABLE_SOCKET_MMSG\n"
" can be used to advanced pkt I/O selection for linux-generic\n"
- "\n", NO_PATH(progname), NO_PATH(progname)
+ "\n", NO_PATH(progname), NO_PATH(progname), MAX_PKTIOS
);
}
--
2.6.3
_______________________________________________
lng-odp mailing list
lng-odp@lists.linaro.org
https://lists.linaro.org/mailman/listinfo/lng-odp
