app/testeventdev: add perf queue worker functions
authorJerin Jacob <jerin.jacob@caviumnetworks.com>
Tue, 4 Jul 2017 04:53:20 +0000 (10:23 +0530)
committerJerin Jacob <jerin.jacob@caviumnetworks.com>
Fri, 7 Jul 2017 07:35:13 +0000 (09:35 +0200)
Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
Acked-by: Harry van Haaren <harry.van.haaren@intel.com>
app/test-eventdev/test_perf_common.h
app/test-eventdev/test_perf_queue.c

index f824695..9888e50 100644 (file)
@@ -86,6 +86,66 @@ struct perf_elt {
        uint64_t timestamp;
 } __rte_cache_aligned;
 
+#define BURST_SIZE 16
+
+#define PERF_WORKER_INIT\
+       struct worker_data *w  = arg;\
+       struct test_perf *t = w->t;\
+       struct evt_options *opt = t->opt;\
+       const uint8_t dev = w->dev_id;\
+       const uint8_t port = w->port_id;\
+       uint8_t *const sched_type_list = &t->sched_type_list[0];\
+       struct rte_mempool *const pool = t->pool;\
+       const uint8_t nb_stages = t->opt->nb_stages;\
+       const uint8_t laststage = nb_stages - 1;\
+       uint8_t cnt = 0;\
+       void *bufs[16] __rte_cache_aligned;\
+       int const sz = RTE_DIM(bufs);\
+       if (opt->verbose_level > 1)\
+               printf("%s(): lcore %d dev_id %d port=%d\n", __func__,\
+                               rte_lcore_id(), dev, port)
+
+static inline __attribute__((always_inline)) int
+perf_process_last_stage(struct rte_mempool *const pool,
+               struct rte_event *const ev, struct worker_data *const w,
+               void *bufs[], int const buf_sz, uint8_t count)
+{
+       bufs[count++] = ev->event_ptr;
+       w->processed_pkts++;
+       rte_smp_wmb();
+
+       if (unlikely(count == buf_sz)) {
+               count = 0;
+               rte_mempool_put_bulk(pool, bufs, buf_sz);
+       }
+       return count;
+}
+
+static inline __attribute__((always_inline)) uint8_t
+perf_process_last_stage_latency(struct rte_mempool *const pool,
+               struct rte_event *const ev, struct worker_data *const w,
+               void *bufs[], int const buf_sz, uint8_t count)
+{
+       uint64_t latency;
+       struct perf_elt *const m = ev->event_ptr;
+
+       bufs[count++] = ev->event_ptr;
+       w->processed_pkts++;
+
+       if (unlikely(count == buf_sz)) {
+               count = 0;
+               latency = rte_get_timer_cycles() - m->timestamp;
+               rte_mempool_put_bulk(pool, bufs, buf_sz);
+       } else {
+               latency = rte_get_timer_cycles() - m->timestamp;
+       }
+
+       w->latency += latency;
+       rte_smp_wmb();
+       return count;
+}
+
+
 static inline int
 perf_nb_event_ports(struct evt_options *opt)
 {
index 1ac8231..323d15f 100644 (file)
@@ -41,6 +41,141 @@ perf_queue_nb_event_queues(struct evt_options *opt)
        return evt_nr_active_lcores(opt->plcores) * opt->nb_stages;
 }
 
+static inline __attribute__((always_inline)) void
+mark_fwd_latency(struct rte_event *const ev,
+               const uint8_t nb_stages)
+{
+       if (unlikely((ev->queue_id % nb_stages) == 0)) {
+               struct perf_elt *const m = ev->event_ptr;
+
+               m->timestamp = rte_get_timer_cycles();
+       }
+}
+
+static inline __attribute__((always_inline)) void
+fwd_event(struct rte_event *const ev, uint8_t *const sched_type_list,
+               const uint8_t nb_stages)
+{
+       ev->queue_id++;
+       ev->sched_type = sched_type_list[ev->queue_id % nb_stages];
+       ev->op = RTE_EVENT_OP_FORWARD;
+       ev->event_type = RTE_EVENT_TYPE_CPU;
+}
+
+static int
+perf_queue_worker(void *arg, const int enable_fwd_latency)
+{
+       PERF_WORKER_INIT;
+       struct rte_event ev;
+
+       while (t->done == false) {
+               uint16_t event = rte_event_dequeue_burst(dev, port, &ev, 1, 0);
+
+               if (!event) {
+                       rte_pause();
+                       continue;
+               }
+               if (enable_fwd_latency)
+               /* first q in pipeline, mark timestamp to compute fwd latency */
+                       mark_fwd_latency(&ev, nb_stages);
+
+               /* last stage in pipeline */
+               if (unlikely((ev.queue_id % nb_stages) == laststage)) {
+                       if (enable_fwd_latency)
+                               cnt = perf_process_last_stage_latency(pool,
+                                       &ev, w, bufs, sz, cnt);
+                       else
+                               cnt = perf_process_last_stage(pool,
+                                       &ev, w, bufs, sz, cnt);
+               } else {
+                       fwd_event(&ev, sched_type_list, nb_stages);
+                       while (rte_event_enqueue_burst(dev, port, &ev, 1) != 1)
+                               rte_pause();
+               }
+       }
+       return 0;
+}
+
+static int
+perf_queue_worker_burst(void *arg, const int enable_fwd_latency)
+{
+       PERF_WORKER_INIT;
+       uint16_t i;
+       /* +1 to avoid prefetch out of array check */
+       struct rte_event ev[BURST_SIZE + 1];
+
+       while (t->done == false) {
+               uint16_t const nb_rx = rte_event_dequeue_burst(dev, port, ev,
+                               BURST_SIZE, 0);
+
+               if (!nb_rx) {
+                       rte_pause();
+                       continue;
+               }
+
+               for (i = 0; i < nb_rx; i++) {
+                       if (enable_fwd_latency) {
+                               rte_prefetch0(ev[i+1].event_ptr);
+                               /* first queue in pipeline.
+                                * mark time stamp to compute fwd latency
+                                */
+                               mark_fwd_latency(&ev[i], nb_stages);
+                       }
+                       /* last stage in pipeline */
+                       if (unlikely((ev[i].queue_id % nb_stages) ==
+                                                laststage)) {
+                               if (enable_fwd_latency)
+                                       cnt = perf_process_last_stage_latency(
+                                               pool, &ev[i], w, bufs, sz, cnt);
+                               else
+                                       cnt = perf_process_last_stage(pool,
+                                               &ev[i], w, bufs, sz, cnt);
+
+                               ev[i].op = RTE_EVENT_OP_RELEASE;
+                       } else {
+                               fwd_event(&ev[i], sched_type_list, nb_stages);
+                       }
+               }
+
+               uint16_t enq;
+
+               enq = rte_event_enqueue_burst(dev, port, ev, nb_rx);
+               while (enq < nb_rx) {
+                       enq += rte_event_enqueue_burst(dev, port,
+                                                       ev + enq, nb_rx - enq);
+               }
+       }
+       return 0;
+}
+
+static int
+worker_wrapper(void *arg)
+{
+       struct worker_data *w  = arg;
+       struct evt_options *opt = w->t->opt;
+
+       const bool burst = evt_has_burst_mode(w->dev_id);
+       const int fwd_latency = opt->fwd_latency;
+
+       /* allow compiler to optimize */
+       if (!burst && !fwd_latency)
+               return perf_queue_worker(arg, 0);
+       else if (!burst && fwd_latency)
+               return perf_queue_worker(arg, 1);
+       else if (burst && !fwd_latency)
+               return perf_queue_worker_burst(arg, 0);
+       else if (burst && fwd_latency)
+               return perf_queue_worker_burst(arg, 1);
+
+       rte_panic("invalid worker\n");
+}
+
+static int
+perf_queue_launch_lcores(struct evt_test *test, struct evt_options *opt)
+{
+       return perf_launch_lcores(test, opt, worker_wrapper);
+}
+
 static int
 perf_queue_eventdev_setup(struct evt_test *test, struct evt_options *opt)
 {
@@ -143,6 +278,7 @@ static const struct evt_test_ops perf_queue =  {
        .test_setup         = perf_test_setup,
        .mempool_setup      = perf_mempool_setup,
        .eventdev_setup     = perf_queue_eventdev_setup,
+       .launch_lcores      = perf_queue_launch_lcores,
        .eventdev_destroy   = perf_eventdev_destroy,
        .mempool_destroy    = perf_mempool_destroy,
        .test_result        = perf_test_result,