odp_bench_queue.c

Microbenchmark application for queue API functions

/* SPDX-License-Identifier: BSD-3-Clause
 * Copyright (c) 2025 Nokia
 */
 
#ifndef _GNU_SOURCE
#define _GNU_SOURCE
#endif
 
#include <odp_api.h>
#include <odp/helper/odph_api.h>
 
#include <bench_common.h>
#include <export_results.h>
 
#include <getopt.h>
#include <inttypes.h>
#include <signal.h>
#include <stdlib.h>
#include <unistd.h>
 
#define TEST_MAX_BURST 32U
#define TEST_DEF_BURST 8U
#define TEST_REPEAT_COUNT 1000U
#define TEST_ROUNDS 100U
#define TEST_MAX_QUEUES (32U * 1024U)
#define TEST_MAX_BENCH 40U
#define TEST_MAX_DATA_ARR (TEST_MAX_BURST * TEST_REPEAT_COUNT)
 
#define NO_PATH(file_name) (strrchr((file_name), '/') ? \
                strrchr((file_name), '/') + 1 : (file_name))
 
#define BENCH_INFO_1(run_fn, init_fn, term_fn) \
    {.name = #run_fn, .run = run_fn, .init = init_fn, .term = term_fn, .desc = NULL}
 
#define BENCH_INFO_2(fn_name, run_fn, init_fn, term_fn) \
    {.name = fn_name, .run = run_fn, .init = init_fn, .term = term_fn, .desc = NULL}
 
#define BENCH_INFO_TM_1(run_fn, init_fn, term_fn) \
    {.name = #run_fn, .run = run_fn, .init = init_fn, .term = term_fn, .cond = NULL,\
     .max_rounds = 0}
 
#define BENCH_INFO_TM_2(fn_name, run_fn, init_fn, term_fn) \
    {.name = fn_name, .run = run_fn, .init = init_fn, .term = term_fn, .cond = NULL,\
     .max_rounds = 0}
 
typedef enum {
    M_TPUT,
    M_LATENCY
} meas_mode_t;
 
typedef struct {
    uint32_t burst_size;
    uint32_t num_queues;
    uint32_t bench_idx;
    uint32_t rounds;
    meas_mode_t mode;
    odp_bool_t is_time;
} appl_args_t;
 
typedef struct {
    appl_args_t appl;
    char cpumask_str[ODP_CPUMASK_STR_SIZE];
 
    struct {
        uint32_t num_buf;
        uint32_t buf_size;
        odp_pool_t pool;
        odp_queue_t p_q;
        odp_queue_t s_q;
    } resources;
 
    struct {
        odp_event_t evs[TEST_MAX_DATA_ARR];
        void *ptrs[TEST_MAX_DATA_ARR];
    } data;
 
    struct {
        union {
            bench_suite_t b;
            bench_tm_suite_t t;
        } s;
 
        odp_atomic_u32_t *exit;
        int *retval;
        void *args;
        int (*suite_fn)(void *args);
        int (*export_fn)(void *data);
 
        union {
            double b[TEST_MAX_BENCH];
            bench_tm_result_t t[TEST_MAX_BENCH];
        } r;
    } suite;
 
    uint8_t context[ODP_CACHE_LINE_SIZE];
} args_t;
 
static args_t *gbl_args;
 
static void init_src_events(void)
{
    uint32_t i;
    const uint32_t num = gbl_args->resources.num_buf;
    odp_buffer_t buf;
 
    for (i = 0U; i < num; i++) {
        buf = odp_buffer_alloc(gbl_args->resources.pool);
 
        if (buf == ODP_BUFFER_INVALID)
            /* There should be enough buffers to allocate so abort in case of an
             * error */
            ODPH_ABORT("Failed to allocate\n");
 
        gbl_args->data.evs[i] = odp_buffer_to_event(buf);
    }
 
    for (uint32_t j = i; j < ODPH_ARRAY_SIZE(gbl_args->data.evs); j++)
        gbl_args->data.evs[j] = ODP_EVENT_INVALID;
}
 
static void term_partial_events(uint32_t start_idx)
{
    odp_event_t ev;
 
    for (uint32_t i = start_idx; i < ODPH_ARRAY_SIZE(gbl_args->data.evs); i++) {
        ev = gbl_args->data.evs[i];
 
        if (ev != ODP_EVENT_INVALID)
            odp_event_free(ev);
    }
}
 
static void term_src_events(void)
{
    odp_event_t ev;
    uint32_t i = 0U;
 
    while (true) {
        ev = odp_queue_deq(gbl_args->resources.p_q);
 
        if (ev == ODP_EVENT_INVALID)
            break;
 
        odp_event_free(ev);
        i++;
    }
 
    term_partial_events(i);
}
 
static void init_events(void)
{
    for (uint32_t i = 0U; i < ODPH_ARRAY_SIZE(gbl_args->data.evs); i++)
        gbl_args->data.evs[i] = ODP_EVENT_INVALID;
}
 
static void fill_plain_queues(void)
{
    const uint32_t num = gbl_args->resources.num_buf;
    odp_buffer_t buf;
    odp_queue_t dst = gbl_args->resources.p_q;
 
    for (uint32_t i = 0U; i < num; i++) {
        buf = odp_buffer_alloc(gbl_args->resources.pool);
 
        if (buf == ODP_BUFFER_INVALID)
            /* There should be enough buffers to allocate so abort in case of an
             * error */
            ODPH_ABORT("Failed to allocate\n");
 
        if (odp_queue_enq(dst, odp_buffer_to_event(buf)) < 0)
            /* There should be enough room to enqueue test buffers to destination
             * queues so abort in case of an error */
            ODPH_ABORT("Failed to enqueue\n");
    }
 
    init_events();
}
 
static void term_events(void)
{
    odp_event_t ev;
 
    for (uint32_t i = 0U; i < ODPH_ARRAY_SIZE(gbl_args->data.evs); i++) {
        ev = gbl_args->data.evs[i];
 
        if (ev != ODP_EVENT_INVALID)
            odp_event_free(ev);
    }
}
 
static void term_dst_events(void)
{
    odp_event_t ev;
 
    while (true) {
        ev = odp_queue_deq(gbl_args->resources.p_q);
 
        if (ev == ODP_EVENT_INVALID)
            break;
 
        odp_event_free(ev);
    }
 
    term_events();
}
 
static void fill_scheduled_queues(void)
{
    const uint32_t num = gbl_args->resources.num_buf;
    odp_buffer_t buf;
    odp_queue_t dst = gbl_args->resources.s_q;
 
    for (uint32_t i = 0U; i < num; i++) {
        buf = odp_buffer_alloc(gbl_args->resources.pool);
 
        if (buf == ODP_BUFFER_INVALID)
            /* There should be enough buffers to allocate so abort in case of an
             * error */
            ODPH_ABORT("Failed to allocate\n");
 
        if (odp_queue_enq(dst, odp_buffer_to_event(buf)) < 0)
            /* There should be enough room to enqueue test buffers to destination
             * queues so abort in case of an error */
            ODPH_ABORT("Failed to enqueue\n");
    }
 
    init_events();
}
 
static void term_scheduled_src_events(void)
{
    odp_event_t ev;
    uint32_t i = 0U;
 
    while (true) {
        ev = odp_schedule(NULL, odp_schedule_wait_time(ODP_TIME_MSEC_IN_NS));
 
        if (ev == ODP_EVENT_INVALID)
            break;
 
        odp_event_free(ev);
        i++;
    }
 
    term_partial_events(i);
}
 
static void term_scheduled_dst_events(void)
{
    odp_event_t ev;
 
    while (true) {
        ev = odp_schedule(NULL, odp_schedule_wait_time(ODP_TIME_MSEC_IN_NS));
 
        if (ev == ODP_EVENT_INVALID)
            break;
 
        odp_event_free(ev);
    }
 
    term_events();
}
 
static int context_set_plain(void)
{
    odp_queue_t dst = gbl_args->resources.p_q;
    uint8_t *context = gbl_args->context;
    const uint32_t len = ODPH_ARRAY_SIZE(gbl_args->context);
    int ret = 0;
 
    for (uint32_t i = 0U; i < TEST_REPEAT_COUNT; i++)
        ret += odp_queue_context_set(dst, context, len);
 
    return !ret;
}
 
static int context_set_plain_tm(bench_tm_result_t *res, int repeat_count)
{
    bench_tm_stamp_t s1, s2;
    odp_queue_t dst = gbl_args->resources.p_q;
    uint8_t *context = gbl_args->context;
    const uint32_t len = ODPH_ARRAY_SIZE(gbl_args->context);
    int ret = 0;
    const uint8_t id1 = bench_tm_func_register(res, "odp_queue_context_set()");
 
    for (int i = 0U; i < repeat_count; i++) {
        bench_tm_now(res, &s1);
        ret += odp_queue_context_set(dst, context, len);
        bench_tm_now(res, &s2);
        bench_tm_func_record(&s2, &s1, res, id1);
    }
 
    return !ret;
}
 
static int context_plain(void)
{
    odp_queue_t q = gbl_args->resources.p_q;
    void **dst = gbl_args->data.ptrs;
 
    for (uint32_t i = 0U; i < TEST_REPEAT_COUNT; i++)
        dst[i] = odp_queue_context(q);
 
    return 1;
}
 
static int context_plain_tm(bench_tm_result_t *res, int repeat_count)
{
    bench_tm_stamp_t s1, s2;
    odp_queue_t q = gbl_args->resources.p_q;
    void **dst = gbl_args->data.ptrs;
    const uint8_t id1 = bench_tm_func_register(res, "odp_queue_context()");
 
    for (int i = 0U; i < repeat_count; i++) {
        bench_tm_now(res, &s1);
        dst[i] = odp_queue_context(q);
        bench_tm_now(res, &s2);
        bench_tm_func_record(&s2, &s1, res, id1);
    }
 
    return 1;
}
 
static int context_set_scheduled(void)
{
    odp_queue_t dst = gbl_args->resources.s_q;
    uint8_t *context = gbl_args->context;
    const uint32_t len = ODPH_ARRAY_SIZE(gbl_args->context);
    int ret = 0;
 
    for (uint32_t i = 0U; i < TEST_REPEAT_COUNT; i++)
        ret += odp_queue_context_set(dst, context, len);
 
    return !ret;
}
 
static int context_set_scheduled_tm(bench_tm_result_t *res, int repeat_count)
{
    bench_tm_stamp_t s1, s2;
    odp_queue_t dst = gbl_args->resources.s_q;
    uint8_t *context = gbl_args->context;
    const uint32_t len = ODPH_ARRAY_SIZE(gbl_args->context);
    int ret = 0;
    const uint8_t id1 = bench_tm_func_register(res, "odp_queue_context_set()");
 
    for (int i = 0U; i < repeat_count; i++) {
        bench_tm_now(res, &s1);
        ret += odp_queue_context_set(dst, context, len);
        bench_tm_now(res, &s2);
        bench_tm_func_record(&s2, &s1, res, id1);
    }
 
    return !ret;
}
 
static int context_scheduled(void)
{
    odp_queue_t q = gbl_args->resources.s_q;
    void **dst = gbl_args->data.ptrs;
 
    for (uint32_t i = 0U; i < TEST_REPEAT_COUNT; i++)
        dst[i] = odp_queue_context(q);
 
    return 1;
}
 
static int context_scheduled_tm(bench_tm_result_t *res, int repeat_count)
{
    bench_tm_stamp_t s1, s2;
    odp_queue_t q = gbl_args->resources.s_q;
    void **dst = gbl_args->data.ptrs;
    const uint8_t id1 = bench_tm_func_register(res, "odp_queue_context()");
 
    for (int i = 0U; i < repeat_count; i++) {
        bench_tm_now(res, &s1);
        dst[i] = odp_queue_context(q);
        bench_tm_now(res, &s2);
        bench_tm_func_record(&s2, &s1, res, id1);
    }
 
    return 1;
}
 
static int enqueue_plain(void)
{
    odp_queue_t dst = gbl_args->resources.p_q;
    odp_event_t *src = gbl_args->data.evs;
    int ret = 0;
 
    for (uint32_t i = 0U; i < TEST_REPEAT_COUNT; i++)
        ret += odp_queue_enq(dst, src[i]);
 
    return !ret;
}
 
static int enqueue_plain_tm(bench_tm_result_t *res, int repeat_count)
{
    bench_tm_stamp_t s1, s2;
    odp_queue_t dst = gbl_args->resources.p_q;
    odp_event_t *src = gbl_args->data.evs;
    int ret = 0;
    const uint8_t id1 = bench_tm_func_register(res, "odp_queue_enq()");
 
    for (int i = 0; i < repeat_count; i++) {
        bench_tm_now(res, &s1);
        ret += odp_queue_enq(dst, src[i]);
        bench_tm_now(res, &s2);
        bench_tm_func_record(&s2, &s1, res, id1);
    }
 
    return !ret;
}
 
static int enqueue_multi_plain(void)
{
    odp_queue_t dst = gbl_args->resources.p_q;
    odp_event_t *src = gbl_args->data.evs;
    int num_tot = 0, ret;
    const int num_burst = gbl_args->appl.burst_size;
 
    for (uint32_t i = 0U; i < TEST_REPEAT_COUNT; i++) {
        ret = odp_queue_enq_multi(dst, &src[num_tot], num_burst);
 
        ODPH_ASSERT(ret >= 0);
 
        num_tot += ret;
    }
 
    return 1;
}
 
static int enqueue_multi_plain_tm(bench_tm_result_t *res, int repeat_count)
{
    bench_tm_stamp_t s1, s2;
    odp_queue_t dst = gbl_args->resources.p_q;
    odp_event_t *src = gbl_args->data.evs;
    int num_tot = 0, ret;
    const int num_burst = gbl_args->appl.burst_size;
    const uint8_t id1 = bench_tm_func_register(res, "odp_queue_enq_multi()");
 
    for (int i = 0; i < repeat_count; i++) {
        bench_tm_now(res, &s1);
        ret = odp_queue_enq_multi(dst, &src[num_tot], num_burst);
        bench_tm_now(res, &s2);
 
        ODPH_ASSERT(ret >= 0);
 
        bench_tm_func_record(&s2, &s1, res, id1);
        num_tot += ret;
    }
 
    return 1;
}
 
static int dequeue_plain(void)
{
    odp_queue_t src = gbl_args->resources.p_q;
    odp_event_t *dst = gbl_args->data.evs;
 
    for (uint32_t i = 0U; i < TEST_REPEAT_COUNT; i++)
        dst[i] = odp_queue_deq(src);
 
    return 1;
}
 
static int dequeue_plain_tm(bench_tm_result_t *res, int repeat_count)
{
    bench_tm_stamp_t s1, s2;
    odp_queue_t src = gbl_args->resources.p_q;
    odp_event_t *dst = gbl_args->data.evs;
    const uint8_t id1 = bench_tm_func_register(res, "odp_queue_deq()");
 
    for (int i = 0U; i < repeat_count; i++) {
        bench_tm_now(res, &s1);
        dst[i] = odp_queue_deq(src);
        bench_tm_now(res, &s2);
        bench_tm_func_record(&s2, &s1, res, id1);
    }
 
    return 1;
}
 
static int dequeue_multi_plain(void)
{
    odp_queue_t src = gbl_args->resources.p_q;
    odp_event_t *dst = gbl_args->data.evs;
    int num_tot = 0, ret;
    const int num_burst = gbl_args->appl.burst_size;
 
    for (uint32_t i = 0U; i < TEST_REPEAT_COUNT; i++) {
        ret = odp_queue_deq_multi(src, &dst[num_tot], num_burst);
 
        ODPH_ASSERT(ret >= 0);
 
        num_tot += ret;
    }
 
    return 1;
}
 
static int dequeue_multi_plain_tm(bench_tm_result_t *res, int repeat_count)
{
    bench_tm_stamp_t s1, s2;
    odp_queue_t src = gbl_args->resources.p_q;
    odp_event_t *dst = gbl_args->data.evs;
    int num_tot = 0, ret;
    const int num_burst = gbl_args->appl.burst_size;
    const uint8_t id1 = bench_tm_func_register(res, "odp_queue_deq_multi()");
 
    for (int i = 0U; i < repeat_count; i++) {
        bench_tm_now(res, &s1);
        ret = odp_queue_deq_multi(src, &dst[num_tot], num_burst);
        bench_tm_now(res, &s2);
 
        ODPH_ASSERT(ret >= 0);
 
        bench_tm_func_record(&s2, &s1, res, id1);
        num_tot += ret;
    }
 
    return 1;
}
 
static int enqueue_scheduled(void)
{
    odp_queue_t dst = gbl_args->resources.s_q;
    odp_event_t *src = gbl_args->data.evs;
    int ret = 0;
 
    for (uint32_t i = 0U; i < TEST_REPEAT_COUNT; i++)
        ret += odp_queue_enq(dst, src[i]);
 
    return !ret;
}
 
static int enqueue_scheduled_tm(bench_tm_result_t *res, int repeat_count)
{
    bench_tm_stamp_t s1, s2;
    odp_queue_t dst = gbl_args->resources.s_q;
    odp_event_t *src = gbl_args->data.evs;
    int ret = 0;
    const uint8_t id1 = bench_tm_func_register(res, "odp_queue_enq()");
 
    for (int i = 0; i < repeat_count; i++) {
        bench_tm_now(res, &s1);
        ret += odp_queue_enq(dst, src[i]);
        bench_tm_now(res, &s2);
        bench_tm_func_record(&s2, &s1, res, id1);
    }
 
    return !ret;
}
 
static int enqueue_multi_scheduled(void)
{
    odp_queue_t dst = gbl_args->resources.s_q;
    odp_event_t *src = gbl_args->data.evs;
    int num_tot = 0, ret;
    const int num_burst = gbl_args->appl.burst_size;
 
    for (uint32_t i = 0U; i < TEST_REPEAT_COUNT; i++) {
        ret = odp_queue_enq_multi(dst, &src[num_tot], num_burst);
 
        ODPH_ASSERT(ret >= 0);
 
        num_tot += ret;
    }
 
    return 1;
}
 
static int enqueue_multi_scheduled_tm(bench_tm_result_t *res, int repeat_count)
{
    bench_tm_stamp_t s1, s2;
    odp_queue_t dst = gbl_args->resources.s_q;
    odp_event_t *src = gbl_args->data.evs;
    int num_tot = 0, ret;
    const int num_burst = gbl_args->appl.burst_size;
    const uint8_t id1 = bench_tm_func_register(res, "odp_queue_enq_multi()");
 
    for (int i = 0; i < repeat_count; i++) {
        bench_tm_now(res, &s1);
        ret = odp_queue_enq_multi(dst, &src[num_tot], num_burst);
        bench_tm_now(res, &s2);
 
        ODPH_ASSERT(ret >= 0);
 
        bench_tm_func_record(&s2, &s1, res, id1);
        num_tot += ret;
    }
 
    return 1;
}
 
static int schedule(void)
{
    odp_event_t *dst = gbl_args->data.evs;
 
    for (uint32_t i = 0U; i < TEST_REPEAT_COUNT; i++)
        dst[i] = odp_schedule(NULL, ODP_SCHED_NO_WAIT);
 
    return 1;
}
 
static int schedule_tm(bench_tm_result_t *res, int repeat_count)
{
    bench_tm_stamp_t s1, s2;
    odp_event_t *dst = gbl_args->data.evs;
    const uint8_t id1 = bench_tm_func_register(res, "odp_schedule()");
 
    for (int i = 0; i < repeat_count; i++) {
        bench_tm_now(res, &s1);
        dst[i] = odp_schedule(NULL, ODP_SCHED_NO_WAIT);
        bench_tm_now(res, &s2);
        bench_tm_func_record(&s2, &s1, res, id1);
    }
 
    return 1;
}
 
static int schedule_multi_nw(void)
{
    odp_event_t *dst = gbl_args->data.evs;
    const int num_burst = gbl_args->appl.burst_size;
    int num_recv = 0;
 
    for (uint32_t i = 0U; i < TEST_REPEAT_COUNT; i++)
        num_recv += odp_schedule_multi_no_wait(NULL, &dst[num_recv], num_burst);
 
    return 1;
}
 
static int schedule_multi_nw_tm(bench_tm_result_t *res, int repeat_count)
{
    bench_tm_stamp_t s1, s2;
    odp_event_t *dst = gbl_args->data.evs;
    const int num_burst = gbl_args->appl.burst_size;
    int num_recv = 0;
    const uint8_t id1 = bench_tm_func_register(res, "odp_schedule_multi_no_wait()");
 
    for (int i = 0; i < repeat_count; i++) {
        bench_tm_now(res, &s1);
        num_recv += odp_schedule_multi_no_wait(NULL, &dst[num_recv], num_burst);
        bench_tm_now(res, &s2);
        bench_tm_func_record(&s2, &s1, res, id1);
    }
 
    return 1;
}
 
static int schedule_multi(void)
{
    odp_event_t *dst = gbl_args->data.evs;
    const int num_burst = gbl_args->appl.burst_size;
    int num_recv = 0;
 
    for (uint32_t i = 0U; i < TEST_REPEAT_COUNT; i++)
        num_recv += odp_schedule_multi(NULL, ODP_SCHED_NO_WAIT, &dst[num_recv], num_burst);
 
    return 1;
}
 
static int schedule_multi_tm(bench_tm_result_t *res, int repeat_count)
{
    bench_tm_stamp_t s1, s2;
    odp_event_t *dst = gbl_args->data.evs;
    const int num_burst = gbl_args->appl.burst_size;
    int num_recv = 0;
    const uint8_t id1 = bench_tm_func_register(res, "odp_schedule_multi()");
 
    for (int i = 0; i < repeat_count; i++) {
        bench_tm_now(res, &s1);
        num_recv += odp_schedule_multi(NULL, ODP_SCHED_NO_WAIT, &dst[num_recv], num_burst);
        bench_tm_now(res, &s2);
        bench_tm_func_record(&s2, &s1, res, id1);
    }
 
    return 1;
}
 
bench_info_t test_suite[] = {
    BENCH_INFO_1(context_set_plain, NULL, NULL),
    BENCH_INFO_1(context_plain, NULL, NULL),
    BENCH_INFO_1(context_set_scheduled, NULL, NULL),
    BENCH_INFO_1(context_scheduled, NULL, NULL),
    BENCH_INFO_1(enqueue_plain, init_src_events, term_src_events),
    BENCH_INFO_1(enqueue_multi_plain, init_src_events, term_src_events),
    BENCH_INFO_1(dequeue_plain, fill_plain_queues, term_dst_events),
    BENCH_INFO_1(dequeue_multi_plain, fill_plain_queues, term_dst_events),
    BENCH_INFO_1(enqueue_scheduled, init_src_events, term_scheduled_src_events),
    BENCH_INFO_1(enqueue_multi_scheduled, init_src_events, term_scheduled_src_events),
    BENCH_INFO_2("schedule_empty", schedule, init_events, term_events),
    BENCH_INFO_2("schedule_multi_nw_empty", schedule_multi_nw, init_events, term_events),
    BENCH_INFO_2("schedule_multi_empty", schedule_multi, init_events, term_events),
    BENCH_INFO_2("schedule_full", schedule, fill_scheduled_queues, term_scheduled_dst_events),
    BENCH_INFO_2("schedule_multi_nw_full", schedule_multi_nw, fill_scheduled_queues,
             term_scheduled_dst_events),
    BENCH_INFO_2("schedule_multi_full", schedule_multi, fill_scheduled_queues,
             term_scheduled_dst_events),
};
 
ODP_STATIC_ASSERT(ODPH_ARRAY_SIZE(test_suite) < TEST_MAX_BENCH,
          "Result array is too small to hold all the results");
 
bench_tm_info_t test_suite_tm[] = {
    BENCH_INFO_TM_1(context_set_plain_tm, NULL, NULL),
    BENCH_INFO_TM_1(context_plain_tm, NULL, NULL),
    BENCH_INFO_TM_1(context_set_scheduled_tm, NULL, NULL),
    BENCH_INFO_TM_1(context_scheduled_tm, NULL, NULL),
    BENCH_INFO_TM_1(enqueue_plain_tm, init_src_events, term_src_events),
    BENCH_INFO_TM_1(enqueue_multi_plain_tm, init_src_events, term_src_events),
    BENCH_INFO_TM_1(dequeue_plain_tm, fill_plain_queues, term_dst_events),
    BENCH_INFO_TM_1(dequeue_multi_plain_tm, fill_plain_queues, term_dst_events),
    BENCH_INFO_TM_1(enqueue_scheduled_tm, init_src_events, term_scheduled_src_events),
    BENCH_INFO_TM_1(enqueue_multi_scheduled_tm, init_src_events, term_scheduled_src_events),
    BENCH_INFO_TM_2("schedule_empty_tm", schedule_tm, init_events, term_events),
    BENCH_INFO_TM_2("schedule_multi_nw_empty_tm", schedule_multi_nw_tm, init_events,
            term_events),
    BENCH_INFO_TM_2("schedule_multi_empty_tm", schedule_multi_tm, init_events, term_events),
    BENCH_INFO_TM_2("schedule_full_tm", schedule_tm, fill_scheduled_queues,
            term_scheduled_dst_events),
    BENCH_INFO_TM_2("schedule_multi_nw_full_tm", schedule_multi_nw_tm, fill_scheduled_queues,
            term_scheduled_dst_events),
    BENCH_INFO_TM_2("schedule_multi_full_tm", schedule_multi_tm, fill_scheduled_queues,
            term_scheduled_dst_events),
};
 
ODP_STATIC_ASSERT(ODPH_ARRAY_SIZE(test_suite_tm) < TEST_MAX_BENCH,
          "Result array is too small to hold all the results");
 
static void set_defaults(void)
{
    gbl_args->appl.burst_size = TEST_DEF_BURST;
    gbl_args->appl.num_queues = 1U; /* TODO: make CLI configurable */
    gbl_args->appl.bench_idx = 0U; /* Run all benchmarks */
    gbl_args->appl.rounds = TEST_ROUNDS;
    gbl_args->appl.mode = M_TPUT;
    gbl_args->appl.is_time = false;
    gbl_args->resources.num_buf = TEST_DEF_BURST * TEST_REPEAT_COUNT;
    gbl_args->resources.buf_size = ODP_CACHE_LINE_SIZE;
    gbl_args->resources.pool = ODP_POOL_INVALID;
    gbl_args->resources.p_q = ODP_QUEUE_INVALID;
    gbl_args->resources.s_q = ODP_QUEUE_INVALID;
}
 
static void print_usage(char *progname)
{
    printf("\n"
           "OpenDataPlane queue API microbenchmarks.\n"
           "\n"
           "Usage: %s OPTIONS\n"
           "  E.g. %s\n"
           "\n"
           "Optional OPTIONS:\n"
           "  -b, --burst <num>      Test burst size for '*_multi()' functions. Maximum %u.\n"
           "                         %u by default. Implementation capabilities may further\n"
           "                         limit the value.\n"
           "  -i, --index <idx>      Benchmark index to run indefinitely. Index should be\n"
           "                         one-based. Use 0 to run all benchmarks. 0 by default.\n"
           "                         Registered benchmarks:\n",
           NO_PATH(progname), NO_PATH(progname), TEST_MAX_BURST, TEST_DEF_BURST);
 
    /* Dump the to-be-tested functions for user-friendliness */
    for (uint32_t i = 0U; i < ODPH_ARRAY_SIZE(test_suite); ++i)
        printf("                             %u: %s\n", i + 1U, test_suite[i].name);
 
    printf("  -r, --rounds <num>     Run each test case 'num' times. %u by default.\n"
           "  -t, --time <opt>       Time measurement.\n"
           "                             0: measure CPU cycles (default)\n"
           "                             1: measure time\n"
           "  -m, --mode <mode>      Measurement mode.\n"
           "                             0: measure throughput, track average execution\n"
           "                                time (default)\n"
           "                             1: measure latency, track function minimum and\n"
           "                                maximum in addition to average execution time.\n"
           "  -h, --help             Display help and exit.\n\n"
           "\n", TEST_ROUNDS);
}
 
static void check_args(void)
{
    odp_pool_capability_t p_capa;
    odp_queue_capability_t q_capa;
    odp_schedule_capability_t s_capa;
 
    gbl_args->resources.num_buf = TEST_REPEAT_COUNT * gbl_args->appl.burst_size;
 
    if (gbl_args->appl.mode != M_TPUT && gbl_args->appl.mode != M_LATENCY) {
        printf("Invalid measurement mode: %d\n", gbl_args->appl.mode);
        exit(EXIT_FAILURE);
    }
 
    if (odp_pool_capability(&p_capa) < 0) {
        ODPH_ERR("Failed to query pool capabilities\n");
        exit(EXIT_FAILURE);
    }
 
    if (gbl_args->resources.num_buf > p_capa.buf.max_num) {
        ODPH_ERR("Invalid amount of buffers: %u (max: %u)\n",
             gbl_args->resources.num_buf, p_capa.buf.max_num);
        exit(EXIT_FAILURE);
    }
 
    if (gbl_args->resources.buf_size > p_capa.buf.max_size)
        gbl_args->resources.buf_size = p_capa.buf.max_size;
 
    if (odp_queue_capability(&q_capa) < 0) {
        ODPH_ERR("Failed to query queue capabilities\n");
        exit(EXIT_FAILURE);
    }
 
    if (gbl_args->appl.num_queues * 2U > q_capa.max_queues) {
        ODPH_ERR("Invalid number of queues: %u (max: %u, any type)\n",
             gbl_args->appl.num_queues, q_capa.max_queues);
        exit(EXIT_FAILURE);
    }
 
    if (gbl_args->appl.num_queues > q_capa.plain.max_num) {
        ODPH_ERR("Invalid number of queues: %u (max: %u, plain)\n",
             gbl_args->appl.num_queues, q_capa.max_queues);
        exit(EXIT_FAILURE);
    }
 
    if (gbl_args->resources.num_buf > q_capa.plain.max_size) {
        ODPH_ERR("Invalid queue size: %u (max: %u, plain)\n", gbl_args->resources.num_buf,
             q_capa.plain.max_size);
        exit(EXIT_FAILURE);
    }
 
    if (odp_schedule_capability(&s_capa) < 0) {
        ODPH_ERR("Failed to query schedule capabilities\n");
        exit(EXIT_FAILURE);
    }
 
    if (gbl_args->appl.num_queues > s_capa.max_queues) {
        ODPH_ERR("Invalid number of queues: %u (max: %u, scheduled)\n",
             gbl_args->appl.num_queues, s_capa.max_queues);
        exit(EXIT_FAILURE);
    }
 
    if (gbl_args->resources.num_buf > s_capa.max_queue_size) {
        ODPH_ERR("Invalid queue size: %u (max: %u, scheduled)\n",
             gbl_args->resources.num_buf, s_capa.max_queue_size);
        exit(EXIT_FAILURE);
    }
}
 
static void parse_args(int argc, char *argv[])
{
    int opt;
    static const struct option longopts[] = {
        {"burst", required_argument, NULL, 'b'},
        {"index", required_argument, NULL, 'i'},
        {"rounds", required_argument, NULL, 'r'},
        {"time", required_argument, NULL, 't'},
        {"mode", required_argument, NULL, 'm'},
        {"help", no_argument, NULL, 'h'},
        {NULL, 0, NULL, 0}
    };
    static const char *shortopts =  "b:i:r:t:m:h";
    appl_args_t *appl_args = &gbl_args->appl;
 
    while (true) {
        opt = getopt_long(argc, argv, shortopts, longopts, NULL);
 
        if (opt == -1)
            break;
 
        switch (opt) {
        case 'b':
            appl_args->burst_size = atoi(optarg);
            break;
        case 'i':
            appl_args->bench_idx = atoi(optarg);
            break;
        case 'r':
            appl_args->rounds = atoi(optarg);
            break;
        case 't':
            appl_args->is_time = atoi(optarg);
            break;
        case 'm':
            appl_args->mode = atoi(optarg);
            break;
        case 'h':
            print_usage(argv[0]);
            exit(EXIT_SUCCESS);
            break;
        default:
            break;
        }
    }
 
    check_args();
}
 
static void sig_handler(int signo ODP_UNUSED)
{
    odp_atomic_store_u32(gbl_args->suite.exit, 1);
}
 
static int setup_sig_handler(void)
{
    struct sigaction action;
 
    memset(&action, 0, sizeof(action));
    action.sa_handler = sig_handler;
 
    /* No additional signals blocked. By default, the signal which triggered
     * the handler is blocked. */
    if (sigemptyset(&action.sa_mask))
        return -1;
 
    if (sigaction(SIGINT, &action, NULL))
        return -1;
 
    return 0;
}
 
static int bench_buffer_tm_export(void *data)
{
    args_t *gbl_args = data;
    bench_tm_result_t *res;
    uint64_t num;
    int ret = 0;
    const char *unit = gbl_args->appl.is_time ? "nsec" : "cpu cycles";
 
    if (test_common_write("function name,min %s per function call,"
                  "average %s per function call,"
                  "max %s per function call\n", unit, unit, unit)) {
        ret = -1;
        goto exit;
    }
 
    for (uint32_t i = 0; i < gbl_args->suite.s.t.num_bench; i++) {
        res = &gbl_args->suite.s.t.result[i];
 
        for (int j = 0; j < res->num; j++) {
            num = res->func[j].num ? res->func[j].num : 1;
            if (test_common_write("%s,%" PRIu64 ",%" PRIu64 ",%" PRIu64 "\n",
                          res->func[j].name,
                          bench_tm_to_u64(res, &res->func[j].min),
                          bench_tm_to_u64(res, &res->func[j].tot) / num,
                          bench_tm_to_u64(res, &res->func[j].max))) {
                ret = -1;
                goto exit;
            }
        }
    }
 
exit:
    test_common_write_term();
 
    return ret;
}
 
static int bench_buffer_export(void *data)
{
    args_t *gbl_args = data;
    int ret = 0;
 
    if (test_common_write("%s", gbl_args->appl.is_time ?
                  "function name,average nsec per function call\n" :
                  "function name,average cpu cycles per function call\n")) {
        ret = -1;
        goto exit;
    }
 
    for (int i = 0; i < gbl_args->suite.s.b.num_bench; i++) {
        if (test_common_write("odp_%s,%f\n", gbl_args->suite.s.b.bench[i].desc != NULL ?
                    gbl_args->suite.s.b.bench[i].desc :
                    gbl_args->suite.s.b.bench[i].name,
                      gbl_args->suite.s.b.result[i])) {
            ret = -1;
            goto exit;
        }
    }
 
exit:
    test_common_write_term();
 
    return ret;
}
 
static void init_suite(args_t *gbl_args)
{
    if (gbl_args->appl.mode == M_LATENCY) {
        bench_tm_suite_init(&gbl_args->suite.s.t);
        gbl_args->suite.s.t.bench = test_suite_tm;
        gbl_args->suite.s.t.num_bench = ODPH_ARRAY_SIZE(test_suite_tm);
        gbl_args->suite.s.t.rounds = TEST_REPEAT_COUNT;
        gbl_args->suite.s.t.bench_idx = gbl_args->appl.bench_idx;
        gbl_args->suite.s.t.measure_time = gbl_args->appl.is_time;
        gbl_args->suite.exit = &gbl_args->suite.s.t.exit_worker;
        gbl_args->suite.retval = &gbl_args->suite.s.t.retval;
        gbl_args->suite.args = &gbl_args->suite.s.t;
        gbl_args->suite.suite_fn = bench_tm_run;
        gbl_args->suite.export_fn = bench_buffer_tm_export;
        gbl_args->suite.s.t.result = gbl_args->suite.r.t;
    } else {
        bench_suite_init(&gbl_args->suite.s.b);
        gbl_args->suite.s.b.bench = test_suite;
        gbl_args->suite.s.b.num_bench = ODPH_ARRAY_SIZE(test_suite);
        gbl_args->suite.s.b.indef_idx = gbl_args->appl.bench_idx;
        gbl_args->suite.s.b.rounds = gbl_args->appl.rounds;
        gbl_args->suite.s.b.repeat_count = TEST_REPEAT_COUNT;
        gbl_args->suite.s.b.measure_time = gbl_args->appl.is_time;
        gbl_args->suite.exit = &gbl_args->suite.s.b.exit_worker;
        gbl_args->suite.retval = &gbl_args->suite.s.b.retval;
        gbl_args->suite.args = &gbl_args->suite.s.b;
        gbl_args->suite.suite_fn = bench_run;
        gbl_args->suite.export_fn = bench_buffer_export;
        gbl_args->suite.s.b.result = gbl_args->suite.r.b;
    }
}
 
static int create_resources(void)
{
    odp_pool_param_t p_param;
    odp_queue_param_t q_param;
 
    if (odp_schedule_config(NULL)) {
        ODPH_ERR("Failed to configure scheduler\n");
        return -1;
    }
 
    odp_pool_param_init(&p_param);
    p_param.type = ODP_POOL_BUFFER;
    p_param.buf.num = gbl_args->resources.num_buf;
    p_param.buf.size = gbl_args->resources.buf_size;
    gbl_args->resources.pool = odp_pool_create("bq_pool", &p_param);
 
    if (gbl_args->resources.pool == ODP_POOL_INVALID) {
        ODPH_ERR("Failed to create test pool\n");
        return -1;
    }
 
    odp_queue_param_init(&q_param);
    q_param.enq_mode = ODP_QUEUE_OP_MT_UNSAFE;
    q_param.deq_mode = ODP_QUEUE_OP_MT_UNSAFE;
    q_param.context = gbl_args->context;
    q_param.size = gbl_args->resources.num_buf;
    gbl_args->resources.p_q = odp_queue_create("bq_plain", &q_param);
 
    if (gbl_args->resources.p_q == ODP_QUEUE_INVALID) {
        ODPH_ERR("Failed to create plain test queue\n");
        return -1;
    }
 
    q_param.type = ODP_QUEUE_TYPE_SCHED;
    gbl_args->resources.s_q = odp_queue_create("bq_sched", &q_param);
 
    if (gbl_args->resources.s_q == ODP_QUEUE_INVALID) {
        ODPH_ERR("Failed to create scheduled test queue\n");
        return -1;
    }
 
    return 0;
}
 
static void print_info(void)
{
    odp_sys_info_print();
    printf("\n"
           "odp_queue_buffer options\n"
           "------------------------\n");
 
    printf("Burst size:       %d\n", gbl_args->appl.burst_size);
    printf("CPU mask:         %s\n", gbl_args->cpumask_str);
    printf("Measurement unit: %s\n", gbl_args->appl.is_time ? "nsec" : "CPU cycles");
    printf("Test rounds:      %u\n", gbl_args->appl.rounds);
    printf("Measurement mode: %s\n", gbl_args->appl.mode == M_TPUT ? "throughput" : "latency");
    printf("\n");
}
 
static void run_worker(odp_instance_t instance, odp_cpumask_t default_mask)
{
    const int cpu = odp_cpumask_first(&default_mask);
    odp_cpumask_t cpumask;
    odph_thread_t worker_thread;
    odph_thread_common_param_t thr_common;
    odph_thread_param_t thr_param;
 
    memset(&worker_thread, 0, sizeof(odph_thread_t));
    odp_cpumask_zero(&cpumask);
    odp_cpumask_set(&cpumask, cpu);
    odph_thread_common_param_init(&thr_common);
    thr_common.instance = instance;
    thr_common.cpumask = &cpumask;
    thr_common.share_param = 1;
    odph_thread_param_init(&thr_param);
    thr_param.start = gbl_args->suite.suite_fn;
    thr_param.arg = gbl_args->suite.args;
    thr_param.thr_type = ODP_THREAD_WORKER;
 
    if (odph_thread_create(&worker_thread, &thr_common, &thr_param, 1) != 1) {
        ODPH_ERR("Creating worker thread failed\n");
        exit(EXIT_FAILURE);
    }
 
    (void)odph_thread_join(&worker_thread, 1);
}
 
static void teardown_resources(void)
{
    (void)odp_queue_destroy(gbl_args->resources.s_q);
    (void)odp_queue_destroy(gbl_args->resources.p_q);
    (void)odp_pool_destroy(gbl_args->resources.pool);
}
 
int main(int argc, char *argv[])
{
    odph_helper_options_t helper_options;
    test_common_options_t common_options;
    odp_shm_t shm;
    odp_cpumask_t default_mask;
    odp_instance_t instance;
    odp_init_t init_param;
    uint8_t ret;
 
    argc = odph_parse_options(argc, argv);
 
    if (odph_options(&helper_options)) {
        ODPH_ERR("Reading ODP helper options failed\n");
        exit(EXIT_FAILURE);
    }
 
    argc = test_common_parse_options(argc, argv);
 
    if (test_common_options(&common_options)) {
        ODPH_ERR("Reading test options failed\n");
        exit(EXIT_FAILURE);
    }
 
    odp_init_param_init(&init_param);
    init_param.mem_model = helper_options.mem_model;
 
    if (odp_init_global(&instance, &init_param, NULL)) {
        ODPH_ERR("Global init failed\n");
        exit(EXIT_FAILURE);
    }
 
    if (odp_init_local(instance, ODP_THREAD_CONTROL)) {
        ODPH_ERR("Local init failed\n");
        exit(EXIT_FAILURE);
    }
 
    shm = odp_shm_reserve("shm_args", sizeof(args_t), ODP_CACHE_LINE_SIZE, 0);
 
    if (shm == ODP_SHM_INVALID) {
        ODPH_ERR("Shared memory reserve failed\n");
        exit(EXIT_FAILURE);
    }
 
    gbl_args = odp_shm_addr(shm);
 
    if (gbl_args == NULL) {
        ODPH_ERR("Shared memory allocation failed\n");
        exit(EXIT_FAILURE);
    }
 
    memset(gbl_args, 0, sizeof(args_t));
    set_defaults();
    parse_args(argc, argv);
    init_suite(gbl_args);
 
    if (setup_sig_handler()) {
        ODPH_ERR("Signal handler setup failed\n");
        exit(EXIT_FAILURE);
    }
 
    if (odp_cpumask_default_worker(&default_mask, 1) != 1) {
        ODPH_ERR("Unable to allocate worker thread\n");
        exit(EXIT_FAILURE);
    }
 
    (void)odp_cpumask_to_str(&default_mask, gbl_args->cpumask_str,
                 sizeof(gbl_args->cpumask_str));
 
    if (create_resources())
        exit(EXIT_FAILURE);
 
    print_info();
    run_worker(instance, default_mask);
    ret = *gbl_args->suite.retval;
 
    if (ret == 0 && common_options.is_export) {
        if (gbl_args->suite.export_fn(gbl_args)) {
            ODPH_ERR("Export failed\n");
            ret = -1;
        }
    }
 
    teardown_resources();
 
    if (odp_shm_free(shm)) {
        ODPH_ERR("Shared memory free failed\n");
        exit(EXIT_FAILURE);
    }
 
    if (odp_term_local()) {
        ODPH_ERR("Local term local\n");
        exit(EXIT_FAILURE);
    }
 
    if (odp_term_global(instance)) {
        ODPH_ERR("Global term global\n");
        exit(EXIT_FAILURE);
    }
 
    return ret;
}