DPDK patches and discussions
 help / color / mirror / Atom feed
From: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
To: dev@dpdk.org
Cc: Peter Spreadborough <peter.spreadborough@broadcom.com>,
	Farah Smith <farah.smith@broadcom.com>,
	Jay Ding <jay.ding@broadcom.com>,
	Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Subject: [PATCH v1 3/4] net/bnxt/tf_ulp: performance and tuning changes for thor2 stats cache
Date: Mon, 18 Nov 2024 21:48:57 +0530	[thread overview]
Message-ID: <20241118161858.1181992-4-sriharsha.basavapatna@broadcom.com> (raw)
In-Reply-To: <20241118161858.1181992-1-sriharsha.basavapatna@broadcom.com>

From: Peter Spreadborough <peter.spreadborough@broadcom.com>

This change adds changes the act get API to expect a physical
address for the host memory rather than a pointer to a local
virtual address. The change was made because the virt->phys
API call has a very high overhead. Also included in this change
to the mutex locks and delays within the main loop.

Fixes: 0513f0af034d ("net/bnxt/tf_ulp: add stats cache for Thor2")
Reviewed-by: Farah Smith <farah.smith@broadcom.com>
Reviewed-by: Jay Ding <jay.ding@broadcom.com>
Signed-off-by: Peter Spreadborough <peter.spreadborough@broadcom.com>
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
---
 drivers/net/bnxt/tf_core/v3/tfc.h        | 12 ++++---
 drivers/net/bnxt/tf_core/v3/tfc_act.c    | 26 ++++++--------
 drivers/net/bnxt/tf_ulp/ulp_fc_mgr_tfc.c |  5 ++-
 drivers/net/bnxt/tf_ulp/ulp_sc_mgr.c     | 45 ++++++++++++++++++++----
 drivers/net/bnxt/tf_ulp/ulp_sc_mgr.h     |  3 +-
 drivers/net/bnxt/tf_ulp/ulp_sc_mgr_tfc.c |  4 +--
 6 files changed, 65 insertions(+), 30 deletions(-)

diff --git a/drivers/net/bnxt/tf_core/v3/tfc.h b/drivers/net/bnxt/tf_core/v3/tfc.h
index 98725b32c2..4176aa01bd 100644
--- a/drivers/net/bnxt/tf_core/v3/tfc.h
+++ b/drivers/net/bnxt/tf_core/v3/tfc.h
@@ -705,7 +705,9 @@ enum tfc_tbl_scope_bucket_factor {
 	TFC_TBL_SCOPE_BUCKET_FACTOR_4 = 4,
 	TFC_TBL_SCOPE_BUCKET_FACTOR_8 = 8,
 	TFC_TBL_SCOPE_BUCKET_FACTOR_16 = 16,
-	TFC_TBL_SCOPE_BUCKET_FACTOR_MAX = TFC_TBL_SCOPE_BUCKET_FACTOR_16
+	TFC_TBL_SCOPE_BUCKET_FACTOR_32 = 32,
+	TFC_TBL_SCOPE_BUCKET_FACTOR_64 = 64,
+	TFC_TBL_SCOPE_BUCKET_FACTOR_MAX = TFC_TBL_SCOPE_BUCKET_FACTOR_64
 };
 
 /**
@@ -1420,8 +1422,9 @@ int tfc_act_set(struct tfc *tfcp,
  * @param[in] cmm_clr
  *   Pointer to cmm clr
  *
- * @param[in,out] data
- *   Data read. Must be word aligned, i.e. [1:0] must be 0.
+ * @param[in,out] host_address
+ *   Data read. Must be word aligned, i.e. [1:0] must be 0. The address
+ *   must be the ret_mem_virt2iova() version of the virt address.
  *
  * @param[in,out] data_sz_words
  *   Data buffer size in words.	Size could be 8/16/24/32/64B
@@ -1437,7 +1440,8 @@ int tfc_act_get(struct tfc *tfcp,
 		struct tfc_mpc_batch_info_t *batch_info,
 		const struct tfc_cmm_info *cmm_info,
 		struct tfc_cmm_clr *clr,
-		uint8_t *data, uint16_t *data_sz_words);
+		uint64_t *host_address,
+		uint16_t *data_sz_words);
 
 /**
  * Free a CMM Resource
diff --git a/drivers/net/bnxt/tf_core/v3/tfc_act.c b/drivers/net/bnxt/tf_core/v3/tfc_act.c
index 3e215f4881..0e98bd30d7 100644
--- a/drivers/net/bnxt/tf_core/v3/tfc_act.c
+++ b/drivers/net/bnxt/tf_core/v3/tfc_act.c
@@ -367,7 +367,7 @@ int tfc_act_get_only_response(struct cfa_bld_mpcinfo *mpc_info,
 static int tfc_act_get_only(struct tfc *tfcp,
 			    struct tfc_mpc_batch_info_t *batch_info,
 			    const struct tfc_cmm_info *cmm_info,
-			    uint8_t *data,
+			    uint64_t *host_address,
 			    uint16_t *data_sz_words)
 {
 	int rc = 0;
@@ -378,7 +378,6 @@ static int tfc_act_get_only(struct tfc *tfcp,
 	uint32_t buff_len;
 	struct cfa_mpc_data_obj fields_cmd[CFA_BLD_MPC_READ_CMD_MAX_FLD] = { {0} };
 	uint32_t entry_offset;
-	uint64_t host_address;
 	struct bnxt_mpc_mbuf mpc_msg_in;
 	struct bnxt_mpc_mbuf mpc_msg_out;
 	uint32_t record_size;
@@ -405,13 +404,11 @@ static int tfc_act_get_only(struct tfc *tfcp,
 	}
 
 	/* Check that data pointer is word aligned */
-	if (unlikely(((uint64_t)data)  & 0x3ULL)) {
+	if (unlikely(*host_address  & 0x3ULL)) {
 		PMD_DRV_LOG_LINE(ERR, "data pointer not word aligned");
 		return -EINVAL;
 	}
 
-	host_address = (uint64_t)rte_mem_virt2iova(data);
-
 	/* Check that MPC APIs are bound */
 	if (unlikely(mpc_info->mpcops == NULL)) {
 		PMD_DRV_LOG_LINE(ERR, "MPC not initialized");
@@ -450,7 +447,7 @@ static int tfc_act_get_only(struct tfc *tfcp,
 
 	fields_cmd[CFA_BLD_MPC_READ_CMD_HOST_ADDRESS_FLD].field_id =
 		CFA_BLD_MPC_READ_CMD_HOST_ADDRESS_FLD;
-	fields_cmd[CFA_BLD_MPC_READ_CMD_HOST_ADDRESS_FLD].val = host_address;
+	fields_cmd[CFA_BLD_MPC_READ_CMD_HOST_ADDRESS_FLD].val = *host_address;
 
 	buff_len = TFC_MPC_MAX_TX_BYTES;
 
@@ -547,7 +544,7 @@ int tfc_act_get_clear_response(struct cfa_bld_mpcinfo *mpc_info,
 static int tfc_act_get_clear(struct tfc *tfcp,
 			     struct tfc_mpc_batch_info_t *batch_info,
 			     const struct tfc_cmm_info *cmm_info,
-			     uint8_t *data,
+			     uint64_t *host_address,
 			     uint16_t *data_sz_words,
 			     uint8_t clr_offset,
 			     uint8_t clr_size)
@@ -560,7 +557,6 @@ static int tfc_act_get_clear(struct tfc *tfcp,
 	uint32_t buff_len;
 	struct cfa_mpc_data_obj fields_cmd[CFA_BLD_MPC_READ_CLR_CMD_MAX_FLD] = { {0} };
 	uint32_t entry_offset;
-	uint64_t host_address;
 	struct bnxt_mpc_mbuf mpc_msg_in;
 	struct bnxt_mpc_mbuf mpc_msg_out;
 	uint32_t record_size;
@@ -589,13 +585,11 @@ static int tfc_act_get_clear(struct tfc *tfcp,
 	}
 
 	/* Check that data pointer is word aligned */
-	if (unlikely(((uint64_t)data)  & 0x3ULL)) {
+	if (unlikely(*host_address  & 0x3ULL)) {
 		PMD_DRV_LOG_LINE(ERR, "data pointer not word aligned");
 		return -EINVAL;
 	}
 
-	host_address = (uint64_t)rte_mem_virt2iova(data);
-
 	/* Check that MPC APIs are bound */
 	if (unlikely(mpc_info->mpcops == NULL)) {
 		PMD_DRV_LOG_LINE(ERR, "MPC not initialized");
@@ -634,7 +628,7 @@ static int tfc_act_get_clear(struct tfc *tfcp,
 
 	fields_cmd[CFA_BLD_MPC_READ_CLR_CMD_HOST_ADDRESS_FLD].field_id =
 		CFA_BLD_MPC_READ_CLR_CMD_HOST_ADDRESS_FLD;
-	fields_cmd[CFA_BLD_MPC_READ_CLR_CMD_HOST_ADDRESS_FLD].val = host_address;
+	fields_cmd[CFA_BLD_MPC_READ_CLR_CMD_HOST_ADDRESS_FLD].val = *host_address;
 
 	for (i = clr_offset; i < clr_size; i++)
 		mask |= (1 << i);
@@ -700,7 +694,8 @@ int tfc_act_get(struct tfc *tfcp,
 		struct tfc_mpc_batch_info_t *batch_info,
 		const struct tfc_cmm_info *cmm_info,
 		struct tfc_cmm_clr *clr,
-		uint8_t *data, uint16_t *data_sz_words)
+		uint64_t *host_address,
+		uint16_t *data_sz_words)
 {
 	/* It's not an error to pass clr as a Null pointer, just means that read
 	 * and clear is not being requested.  Also allow the user to manage
@@ -718,14 +713,15 @@ int tfc_act_get(struct tfc *tfcp,
 		return tfc_act_get_clear(tfcp,
 					 batch_info,
 					 cmm_info,
-					 data, data_sz_words,
+					 host_address,
+					 data_sz_words,
 					 clr->offset_in_byte / 2,
 					 clr->sz_in_byte / 2);
 	} else {
 		return tfc_act_get_only(tfcp,
 					batch_info,
 					cmm_info,
-					data,
+					host_address,
 					data_sz_words);
 	}
 }
diff --git a/drivers/net/bnxt/tf_ulp/ulp_fc_mgr_tfc.c b/drivers/net/bnxt/tf_ulp/ulp_fc_mgr_tfc.c
index e73fdcd1c7..4a1d0bf3df 100644
--- a/drivers/net/bnxt/tf_ulp/ulp_fc_mgr_tfc.c
+++ b/drivers/net/bnxt/tf_ulp/ulp_fc_mgr_tfc.c
@@ -53,6 +53,7 @@ ulp_fc_tfc_update_accum_stats(__rte_unused struct bnxt_ulp_context *ctxt,
 }
 
 static uint8_t *data;
+static uint64_t virt2iova_data;
 
 static int32_t
 ulp_fc_tfc_flow_stat_get(struct bnxt_ulp_context *ctxt,
@@ -84,6 +85,8 @@ ulp_fc_tfc_flow_stat_get(struct bnxt_ulp_context *ctxt,
 			BNXT_DRV_DBG(ERR, "Failed to allocate dma buffer\n");
 			return -EINVAL;
 		}
+
+		virt2iova_data = (uint64_t)rte_mem_virt2iova(data);
 	}
 
 	/* Ensure that data is large enough to read words */
@@ -105,7 +108,7 @@ ulp_fc_tfc_flow_stat_get(struct bnxt_ulp_context *ctxt,
 		cmm_clr.sz_in_byte = sizeof(data64[ULP_FC_TFC_PKT_CNT_OFFS]) +
 			sizeof(data64[ULP_FC_TFC_BYTE_CNT_OFFS]);
 	}
-	rc = tfc_act_get(tfcp, NULL, &cmm_info, &cmm_clr, data, &word_size);
+	rc = tfc_act_get(tfcp, NULL, &cmm_info, &cmm_clr, &virt2iova_data, &word_size);
 	if (rc) {
 		BNXT_DRV_DBG(ERR,
 			     "Failed to read stat memory hndl=0x%" PRIx64 "\n",
diff --git a/drivers/net/bnxt/tf_ulp/ulp_sc_mgr.c b/drivers/net/bnxt/tf_ulp/ulp_sc_mgr.c
index c82fdaf6dd..5fa8e240db 100644
--- a/drivers/net/bnxt/tf_ulp/ulp_sc_mgr.c
+++ b/drivers/net/bnxt/tf_ulp/ulp_sc_mgr.c
@@ -3,6 +3,8 @@
  * All rights reserved.
  */
 
+#include <sched.h>
+#include <unistd.h>
 #include <rte_common.h>
 #include <rte_cycles.h>
 #include <rte_malloc.h>
@@ -56,7 +58,9 @@ int32_t ulp_sc_mgr_init(struct bnxt_ulp_context *ctxt)
 	struct bnxt_ulp_sc_info *ulp_sc_info;
 	uint32_t stats_cache_tbl_sz;
 	uint32_t dev_id;
+	uint8_t *data;
 	int rc;
+	int i;
 
 	if (!ctxt) {
 		BNXT_DRV_DBG(DEBUG, "Invalid ULP CTXT\n");
@@ -124,6 +128,12 @@ int32_t ulp_sc_mgr_init(struct bnxt_ulp_context *ctxt)
 		goto error;
 	}
 
+	data = ulp_sc_info->read_data;
+	for (i = 0; i < ULP_SC_BATCH_SIZE; i++) {
+		ulp_sc_info->read_data_iova[i] = (uint64_t)rte_mem_virt2iova(data);
+		data += ULP_SC_PAGE_SIZE;
+	}
+
 	rc = ulp_sc_mgr_thread_start(ctxt);
 	if (rc)
 		BNXT_DRV_DBG(DEBUG, "Stats counter thread start failed\n");
@@ -241,12 +251,12 @@ static uint32_t ulp_stats_cache_main_loop(void *arg)
 					(uint64_t)sce;
 
 				rc = sc_ops->ulp_stats_cache_update(tfcp,
-								    sce->dir,
-								    data,
-								    sce->handle,
-								    &words,
-								    &batch_info,
-								    sce->reset);
+							    sce->dir,
+							    &ulp_sc_info->read_data_iova[batch],
+							    sce->handle,
+							    &words,
+							    &batch_info,
+							    sce->reset);
 				if (unlikely(rc)) {
 					/* Abort this batch */
 					PMD_DRV_LOG_LINE(ERR,
@@ -330,18 +340,39 @@ int32_t
 ulp_sc_mgr_thread_start(struct bnxt_ulp_context *ctxt)
 {
 	struct bnxt_ulp_sc_info *ulp_sc_info;
+	rte_thread_attr_t attr;
+	rte_cpuset_t mask;
+	size_t i;
 	int rc;
 
 	ulp_sc_info = bnxt_ulp_cntxt_ptr2_sc_info_get(ctxt);
 
 	if (ulp_sc_info && !(ulp_sc_info->flags & ULP_FLAG_SC_THREAD)) {
+		rte_thread_attr_init(&attr);
+
+		rte_thread_get_affinity(&mask);
+
+		for (i = 1; i < CPU_SETSIZE; i++) {
+			if (CPU_ISSET(i, &mask)) {
+				CPU_ZERO(&mask);
+				CPU_SET(i + 2, &mask);
+				break;
+			}
+		}
+
+		rc = rte_thread_attr_set_affinity(&attr, &mask);
+		if (rc)
+			return rc;
+
 		rc = rte_thread_create(&ulp_sc_info->tid,
-				       NULL,
+				       &attr,
 				       &ulp_stats_cache_main_loop,
 				       (void *)ctxt->cfg_data);
 		if (rc)
 			return rc;
 
+		rte_thread_set_prefixed_name(ulp_sc_info->tid, "ulp_sc_mgr");
+
 		ulp_sc_info->flags |= ULP_FLAG_SC_THREAD;
 	}
 
diff --git a/drivers/net/bnxt/tf_ulp/ulp_sc_mgr.h b/drivers/net/bnxt/tf_ulp/ulp_sc_mgr.h
index d5c835cd75..5501dc18fd 100644
--- a/drivers/net/bnxt/tf_ulp/ulp_sc_mgr.h
+++ b/drivers/net/bnxt/tf_ulp/ulp_sc_mgr.h
@@ -36,6 +36,7 @@ struct ulp_sc_tfc_stats_cache_entry {
 struct bnxt_ulp_sc_info {
 	struct ulp_sc_tfc_stats_cache_entry *stats_cache_tbl;
 	uint8_t		*read_data;
+	uint64_t	read_data_iova[ULP_SC_BATCH_SIZE];
 	uint32_t	flags;
 	uint32_t	num_entries;
 	uint32_t	num_counters;
@@ -48,7 +49,7 @@ struct bnxt_ulp_sc_core_ops {
 	int32_t
 	(*ulp_stats_cache_update)(struct tfc *tfcp,
 				  int dir,
-				  uint8_t *data,
+				  uint64_t *host_address,
 				  uint64_t handle,
 				  uint16_t *words,
 				  struct tfc_mpc_batch_info_t *batch_info,
diff --git a/drivers/net/bnxt/tf_ulp/ulp_sc_mgr_tfc.c b/drivers/net/bnxt/tf_ulp/ulp_sc_mgr_tfc.c
index a8141980d8..23edf9f755 100644
--- a/drivers/net/bnxt/tf_ulp/ulp_sc_mgr_tfc.c
+++ b/drivers/net/bnxt/tf_ulp/ulp_sc_mgr_tfc.c
@@ -24,7 +24,7 @@
 static int32_t
 ulp_sc_tfc_stats_cache_update(struct tfc *tfcp,
 			      int dir,
-			      uint8_t *data,
+			      uint64_t *host_address,
 			      uint64_t handle,
 			      uint16_t *words,
 			      struct tfc_mpc_batch_info_t *batch_info,
@@ -48,7 +48,7 @@ ulp_sc_tfc_stats_cache_update(struct tfc *tfcp,
 			 batch_info,
 			 &cmm_info,
 			 &cmm_clr,
-			 data,
+			 host_address,
 			 words);
 
 	return rc;
-- 
2.39.3


  parent reply	other threads:[~2024-11-18 16:07 UTC|newest]

Thread overview: 5+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2024-11-18 16:18 [PATCH v1 0/4] TruFlow fixes for Thor2 Sriharsha Basavapatna
2024-11-18 16:18 ` [PATCH v1 1/4] net/bnxt/tf_ulp: fix F1F2 vxlan counter acccumulation " Sriharsha Basavapatna
2024-11-18 16:18 ` [PATCH v1 2/4] net/bnxt/tf_ulp: fix vfr clean up and stats lockup Sriharsha Basavapatna
2024-11-18 16:18 ` Sriharsha Basavapatna [this message]
2024-11-18 16:18 ` [PATCH v1 4/4] net/bnxt/tf_ulp: update template files Sriharsha Basavapatna

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20241118161858.1181992-4-sriharsha.basavapatna@broadcom.com \
    --to=sriharsha.basavapatna@broadcom.com \
    --cc=dev@dpdk.org \
    --cc=farah.smith@broadcom.com \
    --cc=jay.ding@broadcom.com \
    --cc=peter.spreadborough@broadcom.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).