[RFC PATCH 24/27] net/intel: use separate array for desc status tracking

[Bruce Richardson <bruce.richardson@intel.com>]

Rather than writing a last_id for each individual descriptor, we can
write one only for places where the "report status" (RS) bit is set,
i.e. the descriptors which will be written back when done. The method
used for marking what descriptors are free is also changed in the
process, even if the last descriptor with the "done" bits set is past
the expected point, we only track up to the expected point, and leave
the rest to be counted as freed next time. This means that we always
have the RS/DD bits set at fixed intervals, and we always track free
slots in units of the same tx_free_thresh intervals.

Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
---
 drivers/net/intel/common/tx.h             |  4 ++
 drivers/net/intel/common/tx_scalar_fns.h  | 59 +++++++++++------------
 drivers/net/intel/i40e/i40e_rxtx.c        | 20 ++++++++
 drivers/net/intel/iavf/iavf_rxtx.c        | 19 ++++++++
 drivers/net/intel/ice/ice_rxtx.c          | 20 ++++++++
 drivers/net/intel/idpf/idpf_common_rxtx.c |  7 +++
 drivers/net/intel/idpf/idpf_rxtx.c        | 13 +++++
 7 files changed, 110 insertions(+), 32 deletions(-)

diff --git a/drivers/net/intel/common/tx.h b/drivers/net/intel/common/tx.h
index 0d11daaab3..9b3f8385e6 100644
--- a/drivers/net/intel/common/tx.h
+++ b/drivers/net/intel/common/tx.h
@@ -126,6 +126,8 @@ struct ci_tx_queue {
 		struct ci_tx_entry *sw_ring; /* virtual address of SW ring */
 		struct ci_tx_entry_vec *sw_ring_vec;
 	};
+	/* Scalar TX path: Array tracking last_id at each RS threshold boundary */
+	uint16_t *rs_last_id;
 	uint16_t nb_tx_desc;           /* number of TX descriptors */
 	uint16_t tx_tail; /* current value of tail register */
 	uint16_t nb_tx_used; /* number of TX desc used since RS bit set */
@@ -139,6 +141,8 @@ struct ci_tx_queue {
 	uint16_t tx_free_thresh;
 	/* Number of TX descriptors to use before RS bit is set. */
 	uint16_t tx_rs_thresh;
+	/* Scalar TX path: log2 of tx_rs_thresh for efficient bit operations */
+	uint8_t log2_rs_thresh;
 	uint16_t port_id;  /* Device port identifier. */
 	uint16_t queue_id; /* TX queue index. */
 	uint16_t reg_idx;
diff --git a/drivers/net/intel/common/tx_scalar_fns.h b/drivers/net/intel/common/tx_scalar_fns.h
index 55502b46ed..3d0a23eda3 100644
--- a/drivers/net/intel/common/tx_scalar_fns.h
+++ b/drivers/net/intel/common/tx_scalar_fns.h
@@ -23,37 +23,24 @@
 static __rte_always_inline int
 ci_tx_xmit_cleanup(struct ci_tx_queue *txq)
 {
-	struct ci_tx_entry *sw_ring = txq->sw_ring;
 	volatile struct ci_tx_desc *txd = txq->ci_tx_ring;
-	uint16_t last_desc_cleaned = txq->last_desc_cleaned;
-	uint16_t nb_tx_desc = txq->nb_tx_desc;
-	uint16_t desc_to_clean_to;
-	uint16_t nb_tx_to_clean;
+	const uint16_t last_desc_cleaned = txq->last_desc_cleaned;
+	const uint16_t nb_tx_desc = txq->nb_tx_desc;
 
-	/* Determine the last descriptor needing to be cleaned */
-	desc_to_clean_to = (uint16_t)(last_desc_cleaned + txq->tx_rs_thresh);
-	if (desc_to_clean_to >= nb_tx_desc)
-		desc_to_clean_to = (uint16_t)(desc_to_clean_to - nb_tx_desc);
-
-	/* Check to make sure the last descriptor to clean is done */
-	desc_to_clean_to = sw_ring[desc_to_clean_to].last_id;
+	const uint16_t rs_idx = (last_desc_cleaned == nb_tx_desc - 1) ?
+			0 :
+			(last_desc_cleaned + 1) >> txq->log2_rs_thresh;
+	uint16_t desc_to_clean_to = (rs_idx << txq->log2_rs_thresh) + (txq->tx_rs_thresh - 1);
 
 	/* Check if descriptor is done - all drivers use 0xF as done value in bits 3:0 */
-	if ((txd[desc_to_clean_to].cmd_type_offset_bsz & rte_cpu_to_le_64(CI_TXD_QW1_DTYPE_M)) !=
-			rte_cpu_to_le_64(CI_TX_DESC_DTYPE_DESC_DONE)) {
+	if ((txd[txq->rs_last_id[rs_idx]].cmd_type_offset_bsz & rte_cpu_to_le_64(CI_TXD_QW1_DTYPE_M)) !=
+			rte_cpu_to_le_64(CI_TX_DESC_DTYPE_DESC_DONE))
 		/* Descriptor not yet processed by hardware */
 		return -1;
-	}
-
-	/* Figure out how many descriptors will be cleaned */
-	if (last_desc_cleaned > desc_to_clean_to)
-		nb_tx_to_clean = (uint16_t)((nb_tx_desc - last_desc_cleaned) + desc_to_clean_to);
-	else
-		nb_tx_to_clean = (uint16_t)(desc_to_clean_to - last_desc_cleaned);
 
 	/* Update the txq to reflect the last descriptor that was cleaned */
 	txq->last_desc_cleaned = desc_to_clean_to;
-	txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + nb_tx_to_clean);
+	txq->nb_tx_free += txq->tx_rs_thresh;
 
 	return 0;
 }
@@ -232,6 +219,7 @@ ci_xmit_pkts(struct ci_tx_queue *txq,
 		uint16_t nb_ipsec = 0;
 		uint64_t ipsec_qw0 = 0, ipsec_qw1 = 0;
 		uint64_t cd_qw0, cd_qw1;
+		uint16_t pkt_rs_idx;
 		tx_pkt = *tx_pkts++;
 
 		td_cmd = CI_TX_DESC_CMD_ICRC;
@@ -272,6 +260,9 @@ ci_xmit_pkts(struct ci_tx_queue *txq,
 		if (tx_last >= txq->nb_tx_desc)
 			tx_last = (uint16_t)(tx_last - txq->nb_tx_desc);
 
+		/* Track the RS threshold bucket at packet start */
+		pkt_rs_idx = (uint16_t)(tx_id >> txq->log2_rs_thresh);
+
 		if (unlikely(nb_used > txq->nb_tx_free)) {
 			if (ci_tx_xmit_cleanup(txq) != 0) {
 				if (nb_tx == 0)
@@ -311,8 +302,7 @@ ci_xmit_pkts(struct ci_tx_queue *txq,
 
 			if (txe->mbuf)
 				rte_pktmbuf_free_seg(txe->mbuf);
-			*txe = (struct ci_tx_entry){ .mbuf = tx_pkt, .last_id = tx_last, .next_id = tx_id };
-
+			txe->mbuf = tx_pkt;
 			/* Setup TX Descriptor */
 			td_cmd |= CI_TX_DESC_CMD_EOP;
 			const uint64_t cmd_type_offset_bsz = CI_TX_DESC_DTYPE_DATA |
@@ -339,7 +329,6 @@ ci_xmit_pkts(struct ci_tx_queue *txq,
 
 			write_txd(ctx_txd, cd_qw0, cd_qw1);
 
-			txe->last_id = tx_last;
 			tx_id = txe->next_id;
 			txe = txn;
 		}
@@ -358,7 +347,6 @@ ci_xmit_pkts(struct ci_tx_queue *txq,
 			ipsec_txd[0] = ipsec_qw0;
 			ipsec_txd[1] = ipsec_qw1;
 
-			txe->last_id = tx_last;
 			tx_id = txe->next_id;
 			txe = txn;
 		}
@@ -394,7 +382,6 @@ ci_xmit_pkts(struct ci_tx_queue *txq,
 				buf_dma_addr += CI_MAX_DATA_PER_TXD;
 				slen -= CI_MAX_DATA_PER_TXD;
 
-				txe->last_id = tx_last;
 				tx_id = txe->next_id;
 				txe = txn;
 				txd = &ci_tx_ring[tx_id];
@@ -412,7 +399,6 @@ ci_xmit_pkts(struct ci_tx_queue *txq,
 				((uint64_t)td_tag << CI_TXD_QW1_L2TAG1_S);
 			write_txd(txd, buf_dma_addr, cmd_type_offset_bsz);
 
-			txe->last_id = tx_last;
 			tx_id = txe->next_id;
 			txe = txn;
 			m_seg = m_seg->next;
@@ -421,13 +407,22 @@ ci_xmit_pkts(struct ci_tx_queue *txq,
 		txq->nb_tx_used = (uint16_t)(txq->nb_tx_used + nb_used);
 		txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_used);
 
-		/* set RS bit on the last descriptor of one packet */
-		if (txq->nb_tx_used >= txq->tx_rs_thresh) {
+		/* Check if packet crosses into a new RS threshold bucket.
+		 * The RS bit is set on the last descriptor when we move from one bucket to another.
+		 * For example, with tx_rs_thresh=32 and a 5-descriptor packet using slots 30-34:
+		 *   - pkt_rs_idx = 30 >> 5 = 0 (started in bucket 0)
+		 *   - tx_last = 34, so 35 >> 5 = 1 (next packet is in bucket 1)
+		 *   - Since 0 != 1, set RS bit on descriptor 34, and record rs_last_id[0] = 34
+		 */
+		uint16_t next_rs_idx = ((tx_last + 1) >> txq->log2_rs_thresh);
+
+		if (next_rs_idx != pkt_rs_idx) {
+			/* Packet crossed into a new bucket - set RS bit on last descriptor */
 			txd->cmd_type_offset_bsz |=
 					rte_cpu_to_le_64(CI_TX_DESC_CMD_RS << CI_TXD_QW1_CMD_S);
 
-			/* Update txq RS bit counters */
-			txq->nb_tx_used = 0;
+			/* Record the last descriptor ID for the bucket we're leaving */
+			txq->rs_last_id[pkt_rs_idx] = tx_last;
 		}
 
 		if (ts_fns != NULL)
diff --git a/drivers/net/intel/i40e/i40e_rxtx.c b/drivers/net/intel/i40e/i40e_rxtx.c
index aef78c5358..1fadd0407a 100644
--- a/drivers/net/intel/i40e/i40e_rxtx.c
+++ b/drivers/net/intel/i40e/i40e_rxtx.c
@@ -24,6 +24,7 @@
 #include <rte_ip.h>
 #include <rte_net.h>
 #include <rte_vect.h>
+#include <rte_bitops.h>
 
 #include "i40e_logs.h"
 #include "base/i40e_prototype.h"
@@ -2269,6 +2270,13 @@ i40e_dev_tx_queue_setup(struct rte_eth_dev *dev,
 			     (int)queue_idx);
 		return I40E_ERR_PARAM;
 	}
+	if (!rte_is_power_of_2(tx_rs_thresh)) {
+		PMD_INIT_LOG(ERR, "tx_rs_thresh must be a power of 2. (tx_rs_thresh=%u port=%d queue=%d)",
+			     (unsigned int)tx_rs_thresh,
+			     (int)dev->data->port_id,
+			     (int)queue_idx);
+		return I40E_ERR_PARAM;
+	}
 	if ((tx_rs_thresh > 1) && (tx_conf->tx_thresh.wthresh != 0)) {
 		PMD_INIT_LOG(ERR, "TX WTHRESH must be set to 0 if "
 			     "tx_rs_thresh is greater than 1. "
@@ -2310,6 +2318,7 @@ i40e_dev_tx_queue_setup(struct rte_eth_dev *dev,
 	txq->mz = tz;
 	txq->nb_tx_desc = nb_desc;
 	txq->tx_rs_thresh = tx_rs_thresh;
+	txq->log2_rs_thresh = rte_log2_u32(tx_rs_thresh);
 	txq->tx_free_thresh = tx_free_thresh;
 	txq->queue_id = queue_idx;
 	txq->reg_idx = reg_idx;
@@ -2333,6 +2342,16 @@ i40e_dev_tx_queue_setup(struct rte_eth_dev *dev,
 		return -ENOMEM;
 	}
 
+	/* Allocate RS last_id tracking array */
+	uint16_t num_rs_buckets = nb_desc / tx_rs_thresh;
+	txq->rs_last_id = rte_zmalloc_socket(NULL, sizeof(txq->rs_last_id[0]) * num_rs_buckets,
+			RTE_CACHE_LINE_SIZE, socket_id);
+	if (txq->rs_last_id == NULL) {
+		i40e_tx_queue_release(txq);
+		PMD_DRV_LOG(ERR, "Failed to allocate memory for RS last_id array");
+		return -ENOMEM;
+	}
+
 	i40e_reset_tx_queue(txq);
 	txq->q_set = TRUE;
 
@@ -2378,6 +2397,7 @@ i40e_tx_queue_release(void *txq)
 
 	ci_txq_release_all_mbufs(q, false);
 	rte_free(q->sw_ring);
+	rte_free(q->rs_last_id);
 	rte_memzone_free(q->mz);
 	rte_free(q);
 }
diff --git a/drivers/net/intel/iavf/iavf_rxtx.c b/drivers/net/intel/iavf/iavf_rxtx.c
index f96876ca46..4517d55011 100644
--- a/drivers/net/intel/iavf/iavf_rxtx.c
+++ b/drivers/net/intel/iavf/iavf_rxtx.c
@@ -25,6 +25,7 @@
 #include <rte_ip.h>
 #include <rte_net.h>
 #include <rte_vect.h>
+#include <rte_bitops.h>
 #include <rte_vxlan.h>
 #include <rte_gtp.h>
 #include <rte_geneve.h>
@@ -204,6 +205,11 @@ check_tx_thresh(uint16_t nb_desc, uint16_t tx_rs_thresh,
 			     tx_rs_thresh, nb_desc);
 		return -EINVAL;
 	}
+	if (!rte_is_power_of_2(tx_rs_thresh)) {
+		PMD_INIT_LOG(ERR, "tx_rs_thresh must be a power of 2. (tx_rs_thresh=%u)",
+		             tx_rs_thresh);
+		return -EINVAL;
+	}
 
 	return 0;
 }
@@ -804,6 +810,7 @@ iavf_dev_tx_queue_setup(struct rte_eth_dev *dev,
 
 	txq->nb_tx_desc = nb_desc;
 	txq->tx_rs_thresh = tx_rs_thresh;
+	txq->log2_rs_thresh = rte_log2_u32(tx_rs_thresh);
 	txq->tx_free_thresh = tx_free_thresh;
 	txq->queue_id = queue_idx;
 	txq->port_id = dev->data->port_id;
@@ -827,6 +834,17 @@ iavf_dev_tx_queue_setup(struct rte_eth_dev *dev,
 		return -ENOMEM;
 	}
 
+	/* Allocate RS last_id tracking array */
+	uint16_t num_rs_buckets = nb_desc / tx_rs_thresh;
+	txq->rs_last_id = rte_zmalloc_socket(NULL, sizeof(txq->rs_last_id[0]) * num_rs_buckets,
+			RTE_CACHE_LINE_SIZE, socket_id);
+	if (txq->rs_last_id == NULL) {
+		PMD_INIT_LOG(ERR, "Failed to allocate memory for RS last_id array");
+		rte_free(txq->sw_ring);
+		rte_free(txq);
+		return -ENOMEM;
+	}
+
 	/* Allocate TX hardware ring descriptors. */
 	ring_size = sizeof(struct ci_tx_desc) * IAVF_MAX_RING_DESC;
 	ring_size = RTE_ALIGN(ring_size, IAVF_DMA_MEM_ALIGN);
@@ -1051,6 +1069,7 @@ iavf_dev_tx_queue_release(struct rte_eth_dev *dev, uint16_t qid)
 
 	ci_txq_release_all_mbufs(q, q->use_ctx);
 	rte_free(q->sw_ring);
+	rte_free(q->rs_last_id);
 	rte_memzone_free(q->mz);
 	rte_free(q);
 }
diff --git a/drivers/net/intel/ice/ice_rxtx.c b/drivers/net/intel/ice/ice_rxtx.c
index 2c73011181..a6a454ddf5 100644
--- a/drivers/net/intel/ice/ice_rxtx.c
+++ b/drivers/net/intel/ice/ice_rxtx.c
@@ -5,6 +5,7 @@
 #include <ethdev_driver.h>
 #include <rte_net.h>
 #include <rte_vect.h>
+#include <rte_bitops.h>
 
 #include "ice_rxtx.h"
 #include "ice_rxtx_vec_common.h"
@@ -1576,6 +1577,13 @@ ice_tx_queue_setup(struct rte_eth_dev *dev,
 			     (int)queue_idx);
 		return -EINVAL;
 	}
+	if (!rte_is_power_of_2(tx_rs_thresh)) {
+		PMD_INIT_LOG(ERR, "tx_rs_thresh must be a power of 2. (tx_rs_thresh=%u port=%d queue=%d)",
+			     (unsigned int)tx_rs_thresh,
+			     (int)dev->data->port_id,
+			     (int)queue_idx);
+		return -EINVAL;
+	}
 	if (tx_rs_thresh > 1 && tx_conf->tx_thresh.wthresh != 0) {
 		PMD_INIT_LOG(ERR, "TX WTHRESH must be set to 0 if "
 			     "tx_rs_thresh is greater than 1. "
@@ -1618,6 +1626,7 @@ ice_tx_queue_setup(struct rte_eth_dev *dev,
 	txq->mz = tz;
 	txq->nb_tx_desc = nb_desc;
 	txq->tx_rs_thresh = tx_rs_thresh;
+	txq->log2_rs_thresh = rte_log2_u32(tx_rs_thresh);
 	txq->tx_free_thresh = tx_free_thresh;
 	txq->queue_id = queue_idx;
 
@@ -1642,6 +1651,16 @@ ice_tx_queue_setup(struct rte_eth_dev *dev,
 		return -ENOMEM;
 	}
 
+	/* Allocate RS last_id tracking array */
+	uint16_t num_rs_buckets = nb_desc / tx_rs_thresh;
+	txq->rs_last_id = rte_zmalloc_socket(NULL, sizeof(txq->rs_last_id[0]) * num_rs_buckets,
+			RTE_CACHE_LINE_SIZE, socket_id);
+	if (txq->rs_last_id == NULL) {
+		ice_tx_queue_release(txq);
+		PMD_INIT_LOG(ERR, "Failed to allocate memory for RS last_id array");
+		return -ENOMEM;
+	}
+
 	if (vsi->type == ICE_VSI_PF && (offloads & RTE_ETH_TX_OFFLOAD_SEND_ON_TIMESTAMP)) {
 		if (hw->phy_model != ICE_PHY_E830) {
 			ice_tx_queue_release(txq);
@@ -1714,6 +1733,7 @@ ice_tx_queue_release(void *txq)
 
 	ci_txq_release_all_mbufs(q, false);
 	rte_free(q->sw_ring);
+	rte_free(q->rs_last_id);
 	if (q->tsq) {
 		rte_memzone_free(q->tsq->ts_mz);
 		rte_free(q->tsq);
diff --git a/drivers/net/intel/idpf/idpf_common_rxtx.c b/drivers/net/intel/idpf/idpf_common_rxtx.c
index 81bc45f6ef..1d123f6350 100644
--- a/drivers/net/intel/idpf/idpf_common_rxtx.c
+++ b/drivers/net/intel/idpf/idpf_common_rxtx.c
@@ -5,6 +5,7 @@
 #include <eal_export.h>
 #include <rte_mbuf_dyn.h>
 #include <rte_errno.h>
+#include <rte_bitops.h>
 
 #include "idpf_common_rxtx.h"
 #include "idpf_common_device.h"
@@ -73,6 +74,11 @@ idpf_qc_tx_thresh_check(uint16_t nb_desc, uint16_t tx_rs_thresh,
 			tx_rs_thresh, nb_desc);
 		return -EINVAL;
 	}
+	if (!rte_is_power_of_2(tx_rs_thresh)) {
+		DRV_LOG(ERR, "tx_rs_thresh must be a power of 2. (tx_rs_thresh=%u)",
+			tx_rs_thresh);
+		return -EINVAL;
+	}
 
 	return 0;
 }
@@ -333,6 +339,7 @@ idpf_qc_tx_queue_release(void *txq)
 	}
 
 	ci_txq_release_all_mbufs(q, false);
+	rte_free(q->rs_last_id);
 	rte_free(q->sw_ring);
 	rte_memzone_free(q->mz);
 	rte_free(q);
diff --git a/drivers/net/intel/idpf/idpf_rxtx.c b/drivers/net/intel/idpf/idpf_rxtx.c
index e974eb44b0..5c2516f556 100644
--- a/drivers/net/intel/idpf/idpf_rxtx.c
+++ b/drivers/net/intel/idpf/idpf_rxtx.c
@@ -437,6 +437,7 @@ idpf_tx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
 
 	txq->nb_tx_desc = nb_desc;
 	txq->tx_rs_thresh = tx_rs_thresh;
+	txq->log2_rs_thresh = rte_log2_u32(tx_rs_thresh);
 	txq->tx_free_thresh = tx_free_thresh;
 	txq->queue_id = vport->chunks_info.tx_start_qid + queue_idx;
 	txq->port_id = dev->data->port_id;
@@ -468,6 +469,15 @@ idpf_tx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
 		goto err_sw_ring_alloc;
 	}
 
+	txq->rs_last_id = rte_zmalloc_socket("idpf tx rs_last_id",
+			sizeof(txq->rs_last_id[0]) * (nb_desc >> txq->log2_rs_thresh),
+			RTE_CACHE_LINE_SIZE, socket_id);
+	if (txq->rs_last_id == NULL) {
+		PMD_INIT_LOG(ERR, "Failed to allocate memory for TX RS tracking");
+		ret = -ENOMEM;
+		goto err_rs_last_id_alloc;
+	}
+
 	if (!is_splitq) {
 		txq->ci_tx_ring = mz->addr;
 		idpf_qc_single_tx_queue_reset(txq);
@@ -490,6 +500,9 @@ idpf_tx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
 	return 0;
 
 err_complq_setup:
+	rte_free(txq->rs_last_id);
+err_rs_last_id_alloc:
+	rte_free(txq->sw_ring);
 err_sw_ring_alloc:
 	idpf_dma_zone_release(mz);
 err_mz_reserve:
-- 
2.51.0