event/octeontx2: add worker enqueue functions
authorPavan Nikhilesh <pbhagavatula@marvell.com>
Fri, 28 Jun 2019 18:23:26 +0000 (23:53 +0530)
committerJerin Jacob <jerinj@marvell.com>
Wed, 3 Jul 2019 04:56:08 +0000 (06:56 +0200)
Add worker event enqueue functions.

Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
drivers/event/octeontx2/otx2_evdev.h
drivers/event/octeontx2/otx2_worker.c

index cccce1d..4f2fd33 100644 (file)
@@ -170,6 +170,14 @@ parse_kvargs_value(const char *key, const char *value, void *opaque)
        return 0;
 }
 
+uint16_t otx2_ssogws_enq(void *port, const struct rte_event *ev);
+uint16_t otx2_ssogws_enq_burst(void *port, const struct rte_event ev[],
+                              uint16_t nb_events);
+uint16_t otx2_ssogws_enq_new_burst(void *port, const struct rte_event ev[],
+                                  uint16_t nb_events);
+uint16_t otx2_ssogws_enq_fwd_burst(void *port, const struct rte_event ev[],
+                                  uint16_t nb_events);
+
 /* Init and Fini API's */
 int otx2_sso_init(struct rte_eventdev *event_dev);
 int otx2_sso_fini(struct rte_eventdev *event_dev);
index 83f535d..044c5f1 100644 (file)
@@ -3,3 +3,139 @@
  */
 
 #include "otx2_worker.h"
+
+static __rte_noinline uint8_t
+otx2_ssogws_new_event(struct otx2_ssogws *ws, const struct rte_event *ev)
+{
+       const uint32_t tag = (uint32_t)ev->event;
+       const uint8_t new_tt = ev->sched_type;
+       const uint64_t event_ptr = ev->u64;
+       const uint16_t grp = ev->queue_id;
+
+       if (ws->xaq_lmt <= *ws->fc_mem)
+               return 0;
+
+       otx2_ssogws_add_work(ws, event_ptr, tag, new_tt, grp);
+
+       return 1;
+}
+
+static __rte_always_inline void
+otx2_ssogws_fwd_swtag(struct otx2_ssogws *ws, const struct rte_event *ev)
+{
+       const uint32_t tag = (uint32_t)ev->event;
+       const uint8_t new_tt = ev->sched_type;
+       const uint8_t cur_tt = ws->cur_tt;
+
+       /* 96XX model
+        * cur_tt/new_tt     SSO_SYNC_ORDERED SSO_SYNC_ATOMIC SSO_SYNC_UNTAGGED
+        *
+        * SSO_SYNC_ORDERED        norm           norm             untag
+        * SSO_SYNC_ATOMIC         norm           norm             untag
+        * SSO_SYNC_UNTAGGED       norm           norm             NOOP
+        */
+
+       if (new_tt == SSO_SYNC_UNTAGGED) {
+               if (cur_tt != SSO_SYNC_UNTAGGED)
+                       otx2_ssogws_swtag_untag(ws);
+       } else {
+               otx2_ssogws_swtag_norm(ws, tag, new_tt);
+       }
+
+       ws->swtag_req = 1;
+}
+
+static __rte_always_inline void
+otx2_ssogws_fwd_group(struct otx2_ssogws *ws, const struct rte_event *ev,
+                     const uint16_t grp)
+{
+       const uint32_t tag = (uint32_t)ev->event;
+       const uint8_t new_tt = ev->sched_type;
+
+       otx2_write64(ev->u64, OTX2_SSOW_GET_BASE_ADDR(ws->getwrk_op) +
+                    SSOW_LF_GWS_OP_UPD_WQP_GRP1);
+       rte_smp_wmb();
+       otx2_ssogws_swtag_desched(ws, tag, new_tt, grp);
+}
+
+static __rte_always_inline void
+otx2_ssogws_forward_event(struct otx2_ssogws *ws, const struct rte_event *ev)
+{
+       const uint8_t grp = ev->queue_id;
+
+       /* Group hasn't changed, Use SWTAG to forward the event */
+       if (ws->cur_grp == grp)
+               otx2_ssogws_fwd_swtag(ws, ev);
+       else
+       /*
+        * Group has been changed for group based work pipelining,
+        * Use deschedule/add_work operation to transfer the event to
+        * new group/core
+        */
+               otx2_ssogws_fwd_group(ws, ev, grp);
+}
+
+static __rte_always_inline void
+otx2_ssogws_release_event(struct otx2_ssogws *ws)
+{
+       otx2_ssogws_swtag_flush(ws);
+}
+
+uint16_t __hot
+otx2_ssogws_enq(void *port, const struct rte_event *ev)
+{
+       struct otx2_ssogws *ws = port;
+
+       switch (ev->op) {
+       case RTE_EVENT_OP_NEW:
+               rte_smp_mb();
+               return otx2_ssogws_new_event(ws, ev);
+       case RTE_EVENT_OP_FORWARD:
+               otx2_ssogws_forward_event(ws, ev);
+               break;
+       case RTE_EVENT_OP_RELEASE:
+               otx2_ssogws_release_event(ws);
+               break;
+       default:
+               return 0;
+       }
+
+       return 1;
+}
+
+uint16_t __hot
+otx2_ssogws_enq_burst(void *port, const struct rte_event ev[],
+                     uint16_t nb_events)
+{
+       RTE_SET_USED(nb_events);
+       return otx2_ssogws_enq(port, ev);
+}
+
+uint16_t __hot
+otx2_ssogws_enq_new_burst(void *port, const struct rte_event ev[],
+                         uint16_t nb_events)
+{
+       struct otx2_ssogws *ws = port;
+       uint16_t i, rc = 1;
+
+       rte_smp_mb();
+       if (ws->xaq_lmt <= *ws->fc_mem)
+               return 0;
+
+       for (i = 0; i < nb_events && rc; i++)
+               rc = otx2_ssogws_new_event(ws,  &ev[i]);
+
+       return nb_events;
+}
+
+uint16_t __hot
+otx2_ssogws_enq_fwd_burst(void *port, const struct rte_event ev[],
+                         uint16_t nb_events)
+{
+       struct otx2_ssogws *ws = port;
+
+       RTE_SET_USED(nb_events);
+       otx2_ssogws_forward_event(ws,  ev);
+
+       return 1;
+}