[dpdk-dev] [PATCH] event/octeontx2: remove hotspots by prefetching data

[dpdk-dev] [PATCH] event/octeontx2: remove hotspots by prefetching data

[pbhagavatula]

From: Pavan Nikhilesh <pbhagavatula@marvell.com>

Recent profiling revealed few hotspots in octeontx2 event device
driver add prefetch hints to reduce stalls.

Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
---
 drivers/event/octeontx2/otx2_worker.h      | 1 +
 drivers/event/octeontx2/otx2_worker_dual.c | 1 +
 2 files changed, 2 insertions(+)

diff --git a/drivers/event/octeontx2/otx2_worker.h b/drivers/event/octeontx2/otx2_worker.h
index 6f7842199..092ba71f7 100644
--- a/drivers/event/octeontx2/otx2_worker.h
+++ b/drivers/event/octeontx2/otx2_worker.h
@@ -280,6 +280,7 @@ otx2_ssogws_event_tx(struct otx2_ssogws *ws, struct rte_event ev[],
 	struct rte_mbuf *m = ev[0].mbuf;
 	const struct otx2_eth_txq *txq = otx2_ssogws_xtract_meta(m);
 
+	rte_prefetch_non_temporal(txq);
 	/* Perform header writes before barrier for TSO */
 	otx2_nix_xmit_prepare_tso(m, flags);
 	otx2_ssogws_order(ws, !ev->sched_type);
diff --git a/drivers/event/octeontx2/otx2_worker_dual.c b/drivers/event/octeontx2/otx2_worker_dual.c
index ff68472ef..7016eee4a 100644
--- a/drivers/event/octeontx2/otx2_worker_dual.c
+++ b/drivers/event/octeontx2/otx2_worker_dual.c
@@ -148,6 +148,7 @@ otx2_ssogws_dual_deq_ ##name(void *port, struct rte_event *ev,		\
 	struct otx2_ssogws_dual *ws = port;				\
 	uint8_t gw;							\
 									\
+	rte_prefetch_non_temporal(ws);					\
 	RTE_SET_USED(timeout_ticks);					\
 	if (ws->swtag_req) {						\
 		otx2_ssogws_swtag_wait((struct otx2_ssogws *)		\
-- 
2.17.1

Re: [dpdk-dev] [PATCH] event/octeontx2: remove hotspots by prefetching data

[Jerin Jacob]

On Wed, Oct 23, 2019 at 9:37 PM <pbhagavatula@marvell.com> wrote:
>
> From: Pavan Nikhilesh <pbhagavatula@marvell.com>
>
> Recent profiling revealed few hotspots in octeontx2 event device
> driver add prefetch hints to reduce stalls.
>
> Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>


Applied to dpdk-next-eventdev/master. Thanks.


> ---
>  drivers/event/octeontx2/otx2_worker.h      | 1 +
>  drivers/event/octeontx2/otx2_worker_dual.c | 1 +
>  2 files changed, 2 insertions(+)
>
> diff --git a/drivers/event/octeontx2/otx2_worker.h b/drivers/event/octeontx2/otx2_worker.h
> index 6f7842199..092ba71f7 100644
> --- a/drivers/event/octeontx2/otx2_worker.h
> +++ b/drivers/event/octeontx2/otx2_worker.h
> @@ -280,6 +280,7 @@ otx2_ssogws_event_tx(struct otx2_ssogws *ws, struct rte_event ev[],
>         struct rte_mbuf *m = ev[0].mbuf;
>         const struct otx2_eth_txq *txq = otx2_ssogws_xtract_meta(m);
>
> +       rte_prefetch_non_temporal(txq);
>         /* Perform header writes before barrier for TSO */
>         otx2_nix_xmit_prepare_tso(m, flags);
>         otx2_ssogws_order(ws, !ev->sched_type);
> diff --git a/drivers/event/octeontx2/otx2_worker_dual.c b/drivers/event/octeontx2/otx2_worker_dual.c
> index ff68472ef..7016eee4a 100644
> --- a/drivers/event/octeontx2/otx2_worker_dual.c
> +++ b/drivers/event/octeontx2/otx2_worker_dual.c
> @@ -148,6 +148,7 @@ otx2_ssogws_dual_deq_ ##name(void *port, struct rte_event *ev,              \
>         struct otx2_ssogws_dual *ws = port;                             \
>         uint8_t gw;                                                     \
>                                                                         \
> +       rte_prefetch_non_temporal(ws);                                  \
>         RTE_SET_USED(timeout_ticks);                                    \
>         if (ws->swtag_req) {                                            \
>                 otx2_ssogws_swtag_wait((struct otx2_ssogws *)           \
> --
> 2.17.1
>