[dpdk-dev] [PATCH v3 3/5] net/i40e: fix vector rx burst for i40e

Jeff Guo jia.guo at intel.com
Wed Sep 9 08:36:34 CEST 2020


The limitation of burst size in vector rx was removed, since it should
retrieve as much received packets as possible. And also the scattered
receive path should use a wrapper function to achieve the goal of
burst maximizing. And do some code cleaning for vector rx path.

Signed-off-by: Jeff Guo <jia.guo at intel.com>
---
 drivers/net/i40e/i40e_rxtx.h             |  1 +
 drivers/net/i40e/i40e_rxtx_vec_altivec.c | 64 ++++++++++++++++--------
 drivers/net/i40e/i40e_rxtx_vec_avx2.c    | 29 ++++++-----
 drivers/net/i40e/i40e_rxtx_vec_neon.c    | 58 +++++++++++++--------
 drivers/net/i40e/i40e_rxtx_vec_sse.c     | 58 +++++++++++++--------
 5 files changed, 133 insertions(+), 77 deletions(-)

diff --git a/drivers/net/i40e/i40e_rxtx.h b/drivers/net/i40e/i40e_rxtx.h
index 57d7b4160..01d4609f9 100644
--- a/drivers/net/i40e/i40e_rxtx.h
+++ b/drivers/net/i40e/i40e_rxtx.h
@@ -14,6 +14,7 @@
 #define RTE_I40E_MAX_RX_BURST          RTE_I40E_RXQ_REARM_THRESH
 #define RTE_I40E_TX_MAX_FREE_BUF_SZ    64
 #define RTE_I40E_DESCS_PER_LOOP    4
+#define RTE_I40E_DESCS_PER_LOOP_AVX    8
 
 #define I40E_RXBUF_SZ_1024 1024
 #define I40E_RXBUF_SZ_2048 2048
diff --git a/drivers/net/i40e/i40e_rxtx_vec_altivec.c b/drivers/net/i40e/i40e_rxtx_vec_altivec.c
index 6862a017e..345c63aa7 100644
--- a/drivers/net/i40e/i40e_rxtx_vec_altivec.c
+++ b/drivers/net/i40e/i40e_rxtx_vec_altivec.c
@@ -188,11 +188,13 @@ desc_to_ptype_v(vector unsigned long descs[4], struct rte_mbuf **rx_pkts,
 		ptype_tbl[(*(vector unsigned char *)&ptype1)[8]];
 }
 
- /* Notice:
-  * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
-  * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
-  *   numbers of DD bits
-  */
+/**
+ * vPMD raw receive routine, only accept(nb_pkts >= RTE_I40E_DESCS_PER_LOOP)
+ *
+ * Notice:
+ * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
+ * - floor align nb_pkts to a RTE_I40E_DESCS_PER_LOOP power-of-two
+ */
 static inline uint16_t
 _recv_raw_pkts_vec(struct i40e_rx_queue *rxq, struct rte_mbuf **rx_pkts,
 		   uint16_t nb_pkts, uint8_t *split_packet)
@@ -214,9 +216,6 @@ _recv_raw_pkts_vec(struct i40e_rx_queue *rxq, struct rte_mbuf **rx_pkts,
 		};
 	vector unsigned long dd_check, eop_check;
 
-	/* nb_pkts shall be less equal than RTE_I40E_MAX_RX_BURST */
-	nb_pkts = RTE_MIN(nb_pkts, RTE_I40E_MAX_RX_BURST);
-
 	/* nb_pkts has to be floor-aligned to RTE_I40E_DESCS_PER_LOOP */
 	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_I40E_DESCS_PER_LOOP);
 
@@ -447,11 +446,6 @@ _recv_raw_pkts_vec(struct i40e_rx_queue *rxq, struct rte_mbuf **rx_pkts,
 	return nb_pkts_recd;
 }
 
- /* Notice:
-  * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
-  * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
-  *   numbers of DD bits
-  */
 uint16_t
 i40e_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 		   uint16_t nb_pkts)
@@ -459,19 +453,19 @@ i40e_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 	return _recv_raw_pkts_vec(rx_queue, rx_pkts, nb_pkts, NULL);
 }
 
- /* vPMD receive routine that reassembles scattered packets
-  * Notice:
-  * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
-  * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
-  *   numbers of DD bits
-  */
-uint16_t
-i40e_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
-			     uint16_t nb_pkts)
+/**
+ * vPMD receive routine that reassembles single burst of 32 scattered packets
+ */
+static uint16_t
+i40e_recv_scattered_burst_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
+			      uint16_t nb_pkts)
 {
 	struct i40e_rx_queue *rxq = rx_queue;
 	uint8_t split_flags[RTE_I40E_VPMD_RX_BURST] = {0};
 
+	/* split_flags only can support max of RTE_I40E_VPMD_RX_BURST */
+	nb_pkts = RTE_MIN(nb_pkts, RTE_I40E_VPMD_RX_BURST);
+
 	/* get some new buffers */
 	uint16_t nb_bufs = _recv_raw_pkts_vec(rxq, rx_pkts, nb_pkts,
 			split_flags);
@@ -500,6 +494,32 @@ i40e_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 		&split_flags[i]);
 }
 
+/**
+ * vPMD receive routine that reassembles scattered packets.
+ */
+uint16_t
+i40e_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
+			     uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > RTE_I40E_VPMD_RX_BURST) {
+		uint16_t burst;
+
+		burst = i40e_recv_scattered_burst_vec(rx_queue,
+						      rx_pkts + retval,
+						      RTE_I40E_VPMD_RX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < RTE_I40E_VPMD_RX_BURST)
+			return retval;
+	}
+
+	return retval + i40e_recv_scattered_burst_vec(rx_queue,
+						      rx_pkts + retval,
+						      nb_pkts);
+}
+
 static inline void
 vtx1(volatile struct i40e_tx_desc *txdp,
 	struct rte_mbuf *pkt, uint64_t flags)
diff --git a/drivers/net/i40e/i40e_rxtx_vec_avx2.c b/drivers/net/i40e/i40e_rxtx_vec_avx2.c
index 3bcef1363..b5e6867d0 100644
--- a/drivers/net/i40e/i40e_rxtx_vec_avx2.c
+++ b/drivers/net/i40e/i40e_rxtx_vec_avx2.c
@@ -36,7 +36,7 @@ i40e_rxq_rearm(struct i40e_rx_queue *rxq)
 		    rxq->nb_rx_desc) {
 			__m128i dma_addr0;
 			dma_addr0 = _mm_setzero_si128();
-			for (i = 0; i < RTE_I40E_DESCS_PER_LOOP; i++) {
+			for (i = 0; i < RTE_I40E_DESCS_PER_LOOP_AVX; i++) {
 				rxep[i].mbuf = &rxq->fake_mbuf;
 				_mm_store_si128((__m128i *)&rxdp[i].read,
 						dma_addr0);
@@ -219,13 +219,18 @@ desc_fdir_processing_32b(volatile union i40e_rx_desc *rxdp,
 
 #define PKTLEN_SHIFT     10
 
-/* Force inline as some compilers will not inline by default. */
+/**
+ * vPMD raw receive routine, only accept(nb_pkts >= RTE_I40E_DESCS_PER_LOOP_AVX)
+ *
+ * Notice:
+ * - nb_pkts < RTE_I40E_DESCS_PER_LOOP_AVX, just return no packet
+ * - floor align nb_pkts to a RTE_I40E_DESCS_PER_LOOP_AVX power-of-two
+ * - force inline as some compilers will not inline by default
+ */
 static __rte_always_inline uint16_t
 _recv_raw_pkts_vec_avx2(struct i40e_rx_queue *rxq, struct rte_mbuf **rx_pkts,
 		uint16_t nb_pkts, uint8_t *split_packet)
 {
-#define RTE_I40E_DESCS_PER_LOOP_AVX 8
-
 	const uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
 	const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
 			0, rxq->mbuf_initializer);
@@ -729,10 +734,6 @@ _recv_raw_pkts_vec_avx2(struct i40e_rx_queue *rxq, struct rte_mbuf **rx_pkts,
 	return received;
 }
 
-/*
- * Notice:
- * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
- */
 uint16_t
 i40e_recv_pkts_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
 		   uint16_t nb_pkts)
@@ -740,10 +741,8 @@ i40e_recv_pkts_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
 	return _recv_raw_pkts_vec_avx2(rx_queue, rx_pkts, nb_pkts, NULL);
 }
 
-/*
+/**
  * vPMD receive routine that reassembles single burst of 32 scattered packets
- * Notice:
- * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
  */
 static uint16_t
 i40e_recv_scattered_burst_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
@@ -752,6 +751,9 @@ i40e_recv_scattered_burst_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
 	struct i40e_rx_queue *rxq = rx_queue;
 	uint8_t split_flags[RTE_I40E_VPMD_RX_BURST] = {0};
 
+	/* split_flags only can support max of RTE_I40E_VPMD_RX_BURST */
+	nb_pkts = RTE_MIN(nb_pkts, RTE_I40E_VPMD_RX_BURST);
+
 	/* get some new buffers */
 	uint16_t nb_bufs = _recv_raw_pkts_vec_avx2(rxq, rx_pkts, nb_pkts,
 			split_flags);
@@ -781,11 +783,8 @@ i40e_recv_scattered_burst_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
 		&split_flags[i]);
 }
 
-/*
+/**
  * vPMD receive routine that reassembles scattered packets.
- * Main receive routine that can handle arbitrary burst sizes
- * Notice:
- * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
  */
 uint16_t
 i40e_recv_scattered_pkts_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
diff --git a/drivers/net/i40e/i40e_rxtx_vec_neon.c b/drivers/net/i40e/i40e_rxtx_vec_neon.c
index 6f874e45b..143cdf4a5 100644
--- a/drivers/net/i40e/i40e_rxtx_vec_neon.c
+++ b/drivers/net/i40e/i40e_rxtx_vec_neon.c
@@ -187,11 +187,12 @@ desc_to_ptype_v(uint64x2_t descs[4], struct rte_mbuf **__rte_restrict rx_pkts,
 
 }
 
- /*
+/**
+ * vPMD raw receive routine, only accept(nb_pkts >= RTE_I40E_DESCS_PER_LOOP)
+ *
  * Notice:
  * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
- * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
- *   numbers of DD bits
+ * - floor align nb_pkts to a RTE_I40E_DESCS_PER_LOOP power-of-two
  */
 static inline uint16_t
 _recv_raw_pkts_vec(struct i40e_rx_queue *__rte_restrict rxq,
@@ -230,9 +231,6 @@ _recv_raw_pkts_vec(struct i40e_rx_queue *__rte_restrict rxq,
 		0, 0, 0       /* ignore non-length fields */
 		};
 
-	/* nb_pkts shall be less equal than RTE_I40E_MAX_RX_BURST */
-	nb_pkts = RTE_MIN(nb_pkts, RTE_I40E_MAX_RX_BURST);
-
 	/* nb_pkts has to be floor-aligned to RTE_I40E_DESCS_PER_LOOP */
 	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_I40E_DESCS_PER_LOOP);
 
@@ -426,12 +424,6 @@ _recv_raw_pkts_vec(struct i40e_rx_queue *__rte_restrict rxq,
 	return nb_pkts_recd;
 }
 
- /*
- * Notice:
- * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
- * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
- *   numbers of DD bits
- */
 uint16_t
 i40e_recv_pkts_vec(void *__rte_restrict rx_queue,
 		struct rte_mbuf **__rte_restrict rx_pkts, uint16_t nb_pkts)
@@ -439,20 +431,20 @@ i40e_recv_pkts_vec(void *__rte_restrict rx_queue,
 	return _recv_raw_pkts_vec(rx_queue, rx_pkts, nb_pkts, NULL);
 }
 
- /* vPMD receive routine that reassembles scattered packets
- * Notice:
- * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
- * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
- *   numbers of DD bits
+/**
+ * vPMD receive routine that reassembles single burst of 32 scattered packets
  */
-uint16_t
-i40e_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
-			     uint16_t nb_pkts)
+static uint16_t
+i40e_recv_scattered_burst_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
+			      uint16_t nb_pkts)
 {
 
 	struct i40e_rx_queue *rxq = rx_queue;
 	uint8_t split_flags[RTE_I40E_VPMD_RX_BURST] = {0};
 
+	/* split_flags only can support max of RTE_I40E_VPMD_RX_BURST */
+	nb_pkts = RTE_MIN(nb_pkts, RTE_I40E_VPMD_RX_BURST);
+
 	/* get some new buffers */
 	uint16_t nb_bufs = _recv_raw_pkts_vec(rxq, rx_pkts, nb_pkts,
 			split_flags);
@@ -482,6 +474,32 @@ i40e_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 		&split_flags[i]);
 }
 
+/**
+ * vPMD receive routine that reassembles scattered packets.
+ */
+uint16_t
+i40e_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
+			     uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > RTE_I40E_VPMD_RX_BURST) {
+		uint16_t burst;
+
+		burst = i40e_recv_scattered_burst_vec(rx_queue,
+						      rx_pkts + retval,
+						      RTE_I40E_VPMD_RX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < RTE_I40E_VPMD_RX_BURST)
+			return retval;
+	}
+
+	return retval + i40e_recv_scattered_burst_vec(rx_queue,
+						      rx_pkts + retval,
+						      nb_pkts);
+}
+
 static inline void
 vtx1(volatile struct i40e_tx_desc *txdp,
 		struct rte_mbuf *pkt, uint64_t flags)
diff --git a/drivers/net/i40e/i40e_rxtx_vec_sse.c b/drivers/net/i40e/i40e_rxtx_vec_sse.c
index 698518349..605912246 100644
--- a/drivers/net/i40e/i40e_rxtx_vec_sse.c
+++ b/drivers/net/i40e/i40e_rxtx_vec_sse.c
@@ -342,11 +342,12 @@ desc_to_ptype_v(__m128i descs[4], struct rte_mbuf **rx_pkts,
 	rx_pkts[3]->packet_type = ptype_tbl[_mm_extract_epi8(ptype1, 8)];
 }
 
- /*
+/**
+ * vPMD raw receive routine, only accept(nb_pkts >= RTE_I40E_DESCS_PER_LOOP)
+ *
  * Notice:
  * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
- * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
- *   numbers of DD bits
+ * - floor align nb_pkts to a RTE_I40E_DESCS_PER_LOOP power-of-two
  */
 static inline uint16_t
 _recv_raw_pkts_vec(struct i40e_rx_queue *rxq, struct rte_mbuf **rx_pkts,
@@ -378,9 +379,6 @@ _recv_raw_pkts_vec(struct i40e_rx_queue *rxq, struct rte_mbuf **rx_pkts,
 			offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
 	__m128i dd_check, eop_check;
 
-	/* nb_pkts shall be less equal than RTE_I40E_MAX_RX_BURST */
-	nb_pkts = RTE_MIN(nb_pkts, RTE_I40E_MAX_RX_BURST);
-
 	/* nb_pkts has to be floor-aligned to RTE_I40E_DESCS_PER_LOOP */
 	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_I40E_DESCS_PER_LOOP);
 
@@ -592,12 +590,6 @@ _recv_raw_pkts_vec(struct i40e_rx_queue *rxq, struct rte_mbuf **rx_pkts,
 	return nb_pkts_recd;
 }
 
- /*
- * Notice:
- * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
- * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
- *   numbers of DD bits
- */
 uint16_t
 i40e_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 		   uint16_t nb_pkts)
@@ -605,20 +597,20 @@ i40e_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 	return _recv_raw_pkts_vec(rx_queue, rx_pkts, nb_pkts, NULL);
 }
 
- /* vPMD receive routine that reassembles scattered packets
- * Notice:
- * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
- * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
- *   numbers of DD bits
+/**
+ * vPMD receive routine that reassembles single burst of 32 scattered packets
  */
-uint16_t
-i40e_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
-			     uint16_t nb_pkts)
+static uint16_t
+i40e_recv_scattered_burst_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
+			      uint16_t nb_pkts)
 {
 
 	struct i40e_rx_queue *rxq = rx_queue;
 	uint8_t split_flags[RTE_I40E_VPMD_RX_BURST] = {0};
 
+	/* split_flags only can support max of RTE_I40E_VPMD_RX_BURST */
+	nb_pkts = RTE_MIN(nb_pkts, RTE_I40E_VPMD_RX_BURST);
+
 	/* get some new buffers */
 	uint16_t nb_bufs = _recv_raw_pkts_vec(rxq, rx_pkts, nb_pkts,
 			split_flags);
@@ -648,6 +640,32 @@ i40e_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 		&split_flags[i]);
 }
 
+/**
+ * vPMD receive routine that reassembles scattered packets.
+ */
+uint16_t
+i40e_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
+			     uint16_t nb_pkts)
+{
+	uint16_t retval = 0;
+
+	while (nb_pkts > RTE_I40E_VPMD_RX_BURST) {
+		uint16_t burst;
+
+		burst = i40e_recv_scattered_burst_vec(rx_queue,
+						      rx_pkts + retval,
+						      RTE_I40E_VPMD_RX_BURST);
+		retval += burst;
+		nb_pkts -= burst;
+		if (burst < RTE_I40E_VPMD_RX_BURST)
+			return retval;
+	}
+
+	return retval + i40e_recv_scattered_burst_vec(rx_queue,
+						      rx_pkts + retval,
+						      nb_pkts);
+}
+
 static inline void
 vtx1(volatile struct i40e_tx_desc *txdp,
 		struct rte_mbuf *pkt, uint64_t flags)
-- 
2.20.1



More information about the dev mailing list