[PATCH 1/3] cnxk/event: invalidate GWC on port reset

[PATCH 1/3] cnxk/event: invalidate GWC on port reset

[pbhagavatula]

From: Pavan Nikhilesh <pbhagavatula@marvell.com>

Invalidate GWC on event port i.e., HWS reset to prevent
invalid response from SSO.

Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
---
 drivers/common/cnxk/roc_sso.c       | 31 +++++++++++++++++++++++++++++
 drivers/common/cnxk/roc_sso.h       |  2 ++
 drivers/common/cnxk/version.map     |  1 +
 drivers/event/cnxk/cn10k_eventdev.c | 19 +++++++++++++++++-
 4 files changed, 52 insertions(+), 1 deletion(-)

diff --git a/drivers/common/cnxk/roc_sso.c b/drivers/common/cnxk/roc_sso.c
index a5f48d5bbc..1ea0761531 100644
--- a/drivers/common/cnxk/roc_sso.c
+++ b/drivers/common/cnxk/roc_sso.c
@@ -357,6 +357,37 @@ roc_sso_hws_stats_get(struct roc_sso *roc_sso, uint8_t hws,
 	return rc;
 }
 
+void
+roc_sso_hws_gwc_invalidate(struct roc_sso *roc_sso, uint8_t *hws,
+			   uint8_t nb_hws)
+{
+	struct sso *sso = roc_sso_to_sso_priv(roc_sso);
+	struct ssow_lf_inv_req *req;
+	struct dev *dev = &sso->dev;
+	struct mbox *mbox;
+	int i;
+
+	if (!nb_hws)
+		return;
+
+	mbox = mbox_get(dev->mbox);
+	req = mbox_alloc_msg_sso_ws_cache_inv(mbox);
+	if (req == NULL) {
+		mbox_process(mbox);
+		req = mbox_alloc_msg_sso_ws_cache_inv(mbox);
+		if (req == NULL) {
+			mbox_put(mbox);
+			return;
+		}
+	}
+	req->hdr.ver = SSOW_INVAL_SELECTIVE_VER;
+	req->nb_hws = nb_hws;
+	for (i = 0; i < nb_hws; i++)
+		req->hws[i] = hws[i];
+	mbox_process(mbox);
+	mbox_put(mbox);
+}
+
 int
 roc_sso_hwgrp_stats_get(struct roc_sso *roc_sso, uint8_t hwgrp,
 			struct roc_sso_hwgrp_stats *stats)
diff --git a/drivers/common/cnxk/roc_sso.h b/drivers/common/cnxk/roc_sso.h
index a2bb6fcb22..8ee62afb9a 100644
--- a/drivers/common/cnxk/roc_sso.h
+++ b/drivers/common/cnxk/roc_sso.h
@@ -100,6 +100,8 @@ int __roc_api roc_sso_hwgrp_free_xaq_aura(struct roc_sso *roc_sso,
 int __roc_api roc_sso_hwgrp_stash_config(struct roc_sso *roc_sso,
 					 struct roc_sso_hwgrp_stash *stash,
 					 uint16_t nb_stash);
+void __roc_api roc_sso_hws_gwc_invalidate(struct roc_sso *roc_sso, uint8_t *hws,
+					  uint8_t nb_hws);
 
 /* Debug */
 void __roc_api roc_sso_dump(struct roc_sso *roc_sso, uint8_t nb_hws,
diff --git a/drivers/common/cnxk/version.map b/drivers/common/cnxk/version.map
index 8c71497df8..cfb7efbdc7 100644
--- a/drivers/common/cnxk/version.map
+++ b/drivers/common/cnxk/version.map
@@ -475,6 +475,7 @@ INTERNAL {
 	roc_sso_hws_base_get;
 	roc_sso_hws_link;
 	roc_sso_hws_stats_get;
+	roc_sso_hws_gwc_invalidate;
 	roc_sso_hws_unlink;
 	roc_sso_ns_to_gw;
 	roc_sso_rsrc_fini;
diff --git a/drivers/event/cnxk/cn10k_eventdev.c b/drivers/event/cnxk/cn10k_eventdev.c
index 499a3aace7..56482c20a1 100644
--- a/drivers/event/cnxk/cn10k_eventdev.c
+++ b/drivers/event/cnxk/cn10k_eventdev.c
@@ -118,6 +118,7 @@ static int
 cn10k_sso_hws_flush_events(void *hws, uint8_t queue_id, uintptr_t base,
 			   cnxk_handle_event_t fn, void *arg)
 {
+	struct cnxk_sso_evdev *dev = cnxk_sso_pmd_priv(arg);
 	uint64_t retry = CNXK_SSO_FLUSH_RETRY_MAX;
 	struct cn10k_sso_hws *ws = hws;
 	uint64_t cq_ds_cnt = 1;
@@ -128,6 +129,7 @@ cn10k_sso_hws_flush_events(void *hws, uint8_t queue_id, uintptr_t base,
 
 	plt_write64(0, base + SSO_LF_GGRP_QCTL);
 
+	roc_sso_hws_gwc_invalidate(&dev->sso, &ws->hws_id, 1);
 	plt_write64(0, ws->base + SSOW_LF_GWS_OP_GWC_INVAL);
 	req = queue_id;	    /* GGRP ID */
 	req |= BIT_ULL(18); /* Grouped */
@@ -162,6 +164,7 @@ cn10k_sso_hws_flush_events(void *hws, uint8_t queue_id, uintptr_t base,
 		return -EAGAIN;
 
 	plt_write64(0, ws->base + SSOW_LF_GWS_OP_GWC_INVAL);
+	roc_sso_hws_gwc_invalidate(&dev->sso, &ws->hws_id, 1);
 	rte_mb();
 
 	return 0;
@@ -181,6 +184,7 @@ cn10k_sso_hws_reset(void *arg, void *hws)
 	uint8_t pend_tt;
 	bool is_pend;
 
+	roc_sso_hws_gwc_invalidate(&dev->sso, &ws->hws_id, 1);
 	plt_write64(0, ws->base + SSOW_LF_GWS_OP_GWC_INVAL);
 	/* Wait till getwork/swtp/waitw/desched completes. */
 	is_pend = false;
@@ -237,6 +241,7 @@ cn10k_sso_hws_reset(void *arg, void *hws)
 	}
 
 	plt_write64(0, base + SSOW_LF_GWS_OP_GWC_INVAL);
+	roc_sso_hws_gwc_invalidate(&dev->sso, &ws->hws_id, 1);
 	rte_mb();
 }
 
@@ -670,7 +675,9 @@ cn10k_sso_configure_queue_stash(struct rte_eventdev *event_dev)
 static int
 cn10k_sso_start(struct rte_eventdev *event_dev)
 {
-	int rc;
+	struct cnxk_sso_evdev *dev = cnxk_sso_pmd_priv(event_dev);
+	uint8_t hws[RTE_EVENT_MAX_PORTS_PER_DEV];
+	int rc, i;
 
 	rc = cn10k_sso_updt_tx_adptr_data(event_dev);
 	if (rc < 0)
@@ -682,6 +689,9 @@ cn10k_sso_start(struct rte_eventdev *event_dev)
 	if (rc < 0)
 		return rc;
 	cn10k_sso_fp_fns_set(event_dev);
+	for (i = 0; i < event_dev->data->nb_ports; i++)
+		hws[i] = i;
+	roc_sso_hws_gwc_invalidate(&dev->sso, hws, event_dev->data->nb_ports);
 
 	return rc;
 }
@@ -689,6 +699,13 @@ cn10k_sso_start(struct rte_eventdev *event_dev)
 static void
 cn10k_sso_stop(struct rte_eventdev *event_dev)
 {
+	struct cnxk_sso_evdev *dev = cnxk_sso_pmd_priv(event_dev);
+	uint8_t hws[RTE_EVENT_MAX_PORTS_PER_DEV];
+	int i;
+
+	for (i = 0; i < event_dev->data->nb_ports; i++)
+		hws[i] = i;
+	roc_sso_hws_gwc_invalidate(&dev->sso, hws, event_dev->data->nb_ports);
 	cnxk_sso_stop(event_dev, cn10k_sso_hws_reset,
 		      cn10k_sso_hws_flush_events);
 }
-- 
2.41.0

[PATCH 2/3] event/cnxk: remove checks from op release

[pbhagavatula]

From: Pavan Nikhilesh <pbhagavatula@marvell.com>

Remove expensive fastpath checks from op release, remove
explicit release in CN9K tx routine.

Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
---
 drivers/event/cnxk/cn9k_worker.h | 17 -----------------
 drivers/event/cnxk/cnxk_worker.h |  6 ------
 2 files changed, 23 deletions(-)

diff --git a/drivers/event/cnxk/cn9k_worker.h b/drivers/event/cnxk/cn9k_worker.h
index 9ddab095ac..ee659e80d6 100644
--- a/drivers/event/cnxk/cn9k_worker.h
+++ b/drivers/event/cnxk/cn9k_worker.h
@@ -156,15 +156,6 @@ cn9k_sso_hws_dual_forward_event(struct cn9k_sso_hws_dual *dws, uint64_t base,
 	}
 }
 
-static __rte_always_inline void
-cn9k_sso_tx_tag_flush(uint64_t base)
-{
-	if (unlikely(CNXK_TT_FROM_TAG(plt_read64(base + SSOW_LF_GWS_TAG)) ==
-		     SSO_TT_EMPTY))
-		return;
-	plt_write64(0, base + SSOW_LF_GWS_OP_SWTAG_FLUSH);
-}
-
 static __rte_always_inline void
 cn9k_wqe_to_mbuf(uint64_t wqe, const uint64_t mbuf, uint8_t port_id,
 		 const uint32_t tag, const uint32_t flags,
@@ -727,7 +718,6 @@ cn9k_sso_hws_event_tx(uint64_t base, struct rte_event *ev, uint64_t *cmd,
 		      uint64_t *txq_data, const uint32_t flags)
 {
 	struct rte_mbuf *m = ev->mbuf;
-	uint16_t ref_cnt = m->refcnt;
 	struct cn9k_eth_txq *txq;
 
 	/* Perform header writes before barrier for TSO */
@@ -800,13 +790,6 @@ cn9k_sso_hws_event_tx(uint64_t base, struct rte_event *ev, uint64_t *cmd,
 	}
 
 done:
-	if (flags & NIX_TX_OFFLOAD_MBUF_NOFF_F) {
-		if (ref_cnt > 1)
-			return 1;
-	}
-
-	cn9k_sso_tx_tag_flush(base);
-
 	return 1;
 }
 
diff --git a/drivers/event/cnxk/cnxk_worker.h b/drivers/event/cnxk/cnxk_worker.h
index 2bd41f8a5e..0e0d728ba4 100644
--- a/drivers/event/cnxk/cnxk_worker.h
+++ b/drivers/event/cnxk/cnxk_worker.h
@@ -54,12 +54,6 @@ cnxk_sso_hws_swtag_untag(uintptr_t swtag_untag_op)
 static __rte_always_inline void
 cnxk_sso_hws_swtag_flush(uint64_t base)
 {
-	/* Ensure that there is no previous flush is pending. */
-	while (plt_read64(base + SSOW_LF_GWS_PENDSTATE) & BIT_ULL(56))
-		;
-	if (CNXK_TT_FROM_TAG(plt_read64(base + SSOW_LF_GWS_TAG)) ==
-	    SSO_TT_EMPTY)
-		return;
 	plt_write64(0, base + SSOW_LF_GWS_OP_SWTAG_FLUSH);
 }
 
-- 
2.41.0

[PATCH 3/3] common/cnxk: use local labels in asm intrinsic

[pbhagavatula]

From: Pavan Nikhilesh <pbhagavatula@marvell.com>

Using labels in asm generates them as regular function and shades
callstack in tools like gdb or perf.
Use local label instead for better visibility.

Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
---
 drivers/common/cnxk/roc_sso_dp.h | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/drivers/common/cnxk/roc_sso_dp.h b/drivers/common/cnxk/roc_sso_dp.h
index 9d30286d2f..03c5bdf7ee 100644
--- a/drivers/common/cnxk/roc_sso_dp.h
+++ b/drivers/common/cnxk/roc_sso_dp.h
@@ -13,13 +13,13 @@ roc_sso_hws_head_wait(uintptr_t base)
 
 #if defined(__aarch64__)
 	asm volatile(PLT_CPU_FEATURE_PREAMBLE
-		     "		ldr %[tag], [%[tag_op]]	\n"
-		     "		tbnz %[tag], 35, done%=		\n"
+		     "		ldr %[tag], [%[tag_op]]		\n"
+		     "		tbnz %[tag], 35, .Ldone%=	\n"
 		     "		sevl				\n"
-		     "rty%=:	wfe				\n"
-		     "		ldr %[tag], [%[tag_op]]	\n"
-		     "		tbz %[tag], 35, rty%=		\n"
-		     "done%=:					\n"
+		     ".Lrty%=:	wfe				\n"
+		     "		ldr %[tag], [%[tag_op]]		\n"
+		     "		tbz %[tag], 35, .Lrty%=		\n"
+		     ".Ldone%=:					\n"
 		     : [tag] "=&r"(tag)
 		     : [tag_op] "r"(tag_op));
 #else
-- 
2.41.0

Re: [PATCH 3/3] common/cnxk: use local labels in asm intrinsic

[Jerin Jacob]

On Tue, Sep 5, 2023 at 10:25 PM <pbhagavatula@marvell.com> wrote:
>
> From: Pavan Nikhilesh <pbhagavatula@marvell.com>
>
> Using labels in asm generates them as regular function and shades
> callstack in tools like gdb or perf.
> Use local label instead for better visibility.
>
> Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>

Series applied to dpdk-next-net-eventdev/for-main. Thanks