[PATCH v7 31/47] net/sxe2: fix NEON Rx ptype mapping and memory ordering
liujie5 at linkdatatechnology.com
liujie5 at linkdatatechnology.com
Mon Aug 31 04:44:03 CEST 2026
From: Jie Liu <liujie5 at linkdatatechnology.com>
Compute the descriptor lower/upper dwords (desc_lo/desc_hi) once per
loop iteration and pass them to sxe2_rx_desc_ptype_fill_neon() and
sxe2_rx_desc_offloads_para_fill_neon(). Map the packet_type lanes
from desc_lo lanes 1/3/5/7, fixing the previous lane ordering that
mapped pkt0/1/2/3 from lanes 5/1/7/3.
Drop three of the four rte_atomic_thread_fence(acquire) barriers
between the descriptor loads in sxe2_rx_pkts_common_vec_neon(); the
single acquire fence after the descriptor set is loaded is sufficient
for the Rx producer/consumer ordering.
Fixes: ac60f302cbef ("net/sxe2: add vectorized Rx and Tx")
Cc: stable at dpdk.org
Signed-off-by: Jie Liu <liujie5 at linkdatatechnology.com>
---
drivers/net/sxe2/sxe2_txrx_vec_neon.c | 132 +++++++++++++-------------
1 file changed, 68 insertions(+), 64 deletions(-)
diff --git a/drivers/net/sxe2/sxe2_txrx_vec_neon.c b/drivers/net/sxe2/sxe2_txrx_vec_neon.c
index 6b4c70c7ac..0b139b8269 100644
--- a/drivers/net/sxe2/sxe2_txrx_vec_neon.c
+++ b/drivers/net/sxe2/sxe2_txrx_vec_neon.c
@@ -213,22 +213,24 @@ uint16_t sxe2_tx_pkts_vec_neon(void *tx_queue,
}
static __rte_always_inline void
-sxe2_rx_desc_ptype_fill_neon(uint16x8_t staterr, struct rte_mbuf **__rte_restrict rx_pkts)
+sxe2_rx_desc_ptype_fill_neon(uint32x4_t desc_lo,
+ struct rte_mbuf **__rte_restrict rx_pkts,
+ const uint32_t *__rte_restrict ptype_tbl)
{
- uint16x8_t ptype_mask = {
- 0, 0x3FFULL,
- 0, 0x3FFULL,
- 0, 0x3FFULL,
- 0, 0x3FFULL,
+ const uint32x4_t ptype_mask = {
+ SXE2_RX_DESC_PTYPE_MASK_NO_SHIFT << 16,
+ SXE2_RX_DESC_PTYPE_MASK_NO_SHIFT << 16,
+ SXE2_RX_DESC_PTYPE_MASK_NO_SHIFT << 16,
+ SXE2_RX_DESC_PTYPE_MASK_NO_SHIFT << 16,
};
uint16x8_t ptype_all;
- ptype_all = vandq_u16(staterr, ptype_mask);
+ ptype_all = vreinterpretq_u16_u32(vandq_u32(desc_lo, ptype_mask));
- rx_pkts[3]->packet_type = sxe2_ptype_tbl[vgetq_lane_u16(ptype_all, 3)];
- rx_pkts[2]->packet_type = sxe2_ptype_tbl[vgetq_lane_u16(ptype_all, 7)];
- rx_pkts[1]->packet_type = sxe2_ptype_tbl[vgetq_lane_u16(ptype_all, 1)];
- rx_pkts[0]->packet_type = sxe2_ptype_tbl[vgetq_lane_u16(ptype_all, 5)];
+ rx_pkts[0]->packet_type = ptype_tbl[vgetq_lane_u16(ptype_all, 1)];
+ rx_pkts[1]->packet_type = ptype_tbl[vgetq_lane_u16(ptype_all, 3)];
+ rx_pkts[2]->packet_type = ptype_tbl[vgetq_lane_u16(ptype_all, 5)];
+ rx_pkts[3]->packet_type = ptype_tbl[vgetq_lane_u16(ptype_all, 7)];
}
static __rte_always_inline uint32x4_t
@@ -271,9 +273,10 @@ sxe2_rx_desc_fnav_flags_neon(uint64x2_t descs_arr[4])
static __rte_always_inline void
sxe2_rx_desc_offloads_para_fill_neon(struct sxe2_rx_queue *rxq,
volatile union sxe2_rx_desc *desc,
- uint64x2_t descs[4], struct rte_mbuf **rx_pkts)
+ uint64x2_t descs[4], uint32x4_t desc_lo, uint32x4_t desc_hi,
+ struct rte_mbuf **rx_pkts)
{
- uint32x4_t desc_lo, desc_hi, flags, tmp_flags;
+ uint32x4_t flags, tmp_flags;
const uint64x2_t mbuf_init = {rxq->mbuf_init_value, 0};
uint64x2_t rearm0, rearm1, rearm2, rearm3;
@@ -330,23 +333,6 @@ sxe2_rx_desc_offloads_para_fill_neon(struct sxe2_rx_queue *rxq,
0, 0, 0, 0, 0, 0, 0, 0
};
- {
- uint32x4_t d0 = vreinterpretq_u32_u64(descs[0]);
- uint32x4_t d1 = vreinterpretq_u32_u64(descs[1]);
- uint32x4_t d2 = vreinterpretq_u32_u64(descs[2]);
- uint32x4_t d3 = vreinterpretq_u32_u64(descs[3]);
- uint64x2_t f64, t64;
-
- flags = vzip2q_u32(d1, d0);
- tmp_flags = vzip2q_u32(d3, d2);
- f64 = vreinterpretq_u64_u32(flags);
- t64 = vreinterpretq_u64_u32(tmp_flags);
- desc_lo = vreinterpretq_u32_u64(vcombine_u64(vget_low_u64(f64),
- vget_low_u64(t64)));
- desc_hi = vreinterpretq_u32_u64(vcombine_u64(vget_high_u64(f64),
- vget_high_u64(t64)));
- }
-
desc_lo = vandq_u32(desc_lo, desc_msk);
desc_hi = vandq_u32(desc_hi, rss_msk);
@@ -505,25 +491,39 @@ sxe2_rx_pkts_common_vec_neon(struct sxe2_rx_queue *rxq, struct rte_mbuf **rx_pkt
uint64x2_t descs[SXE2_RX_NUM_PER_LOOP_NEON];
uint8x16_t pkt_mb1, pkt_mb2, pkt_mb3, pkt_mb4;
uint64x2_t mbp1, mbp2;
+ uint32x4_t desc_lo, desc_hi;
uint16x8_t staterr;
uint16x8_t tmp;
uint16_t bit_num;
descs[3] = vld1q_u64(RTE_CAST_PTR(uint64_t *, desc + 3));
- rte_atomic_thread_fence(rte_memory_order_acquire);
descs[2] = vld1q_u64(RTE_CAST_PTR(uint64_t *, desc + 2));
- rte_atomic_thread_fence(rte_memory_order_acquire);
descs[1] = vld1q_u64(RTE_CAST_PTR(uint64_t *, desc + 1));
- rte_atomic_thread_fence(rte_memory_order_acquire);
descs[0] = vld1q_u64(RTE_CAST_PTR(uint64_t *, desc));
rte_atomic_thread_fence(rte_memory_order_acquire);
-
descs[3] = vld1q_lane_u64(RTE_CAST_PTR(uint64_t *, desc + 3), descs[3], 0);
descs[2] = vld1q_lane_u64(RTE_CAST_PTR(uint64_t *, desc + 2), descs[2], 0);
descs[1] = vld1q_lane_u64(RTE_CAST_PTR(uint64_t *, desc + 1), descs[1], 0);
descs[0] = vld1q_lane_u64(RTE_CAST_PTR(uint64_t *, desc), descs[0], 0);
+ {
+ uint32x4_t d0 = vreinterpretq_u32_u64(descs[0]);
+ uint32x4_t d1 = vreinterpretq_u32_u64(descs[1]);
+ uint32x4_t d2 = vreinterpretq_u32_u64(descs[2]);
+ uint32x4_t d3 = vreinterpretq_u32_u64(descs[3]);
+
+ uint32x4_t q1_01 = vzip2q_u32(d0, d1);
+ uint32x4_t q1_23 = vzip2q_u32(d2, d3);
+ uint64x2_t q1_01_64 = vreinterpretq_u64_u32(q1_01);
+ uint64x2_t q1_23_64 = vreinterpretq_u64_u32(q1_23);
+
+ desc_lo = vreinterpretq_u32_u64(vcombine_u64(vget_low_u64(q1_01_64),
+ vget_low_u64(q1_23_64)));
+ desc_hi = vreinterpretq_u32_u64(vcombine_u64(vget_high_u64(q1_01_64),
+ vget_high_u64(q1_23_64)));
+ }
+
mbp1 = vld1q_u64((uint64_t *)&buffer[i]);
mbp2 = vld1q_u64((uint64_t *)&buffer[i + 2]);
@@ -543,7 +543,8 @@ sxe2_rx_pkts_common_vec_neon(struct sxe2_rx_queue *rxq, struct rte_mbuf **rx_pkt
pkt_mb1 = vqtbl1q_u8(vreinterpretq_u8_u64(descs[0]), rvp_shuf_mask);
if (do_offload) {
- sxe2_rx_desc_offloads_para_fill_neon(rxq, desc, descs, &rx_pkts[i]);
+ sxe2_rx_desc_offloads_para_fill_neon(rxq, desc, descs, desc_lo,
+ desc_hi, &rx_pkts[i]);
} else {
const uint64x2_t mbuf_init = {
rxq->mbuf_init_value,
@@ -578,55 +579,48 @@ sxe2_rx_pkts_common_vec_neon(struct sxe2_rx_queue *rxq, struct rte_mbuf **rx_pkt
rte_prefetch_non_temporal(desc + SXE2_RX_NUM_PER_LOOP_NEON);
{
- uint32x4_t d0 = vreinterpretq_u32_u64(descs[0]);
- uint32x4_t d1 = vreinterpretq_u32_u64(descs[1]);
- uint32x4_t d2 = vreinterpretq_u32_u64(descs[2]);
- uint32x4_t d3 = vreinterpretq_u32_u64(descs[3]);
- uint32x4_t sterr_tmp1 = vzip2q_u32(d1, d0);
- uint32x4_t sterr_tmp2 = vzip2q_u32(d3, d2);
- uint32x4_t sterr_u32 = vzip1q_u32(sterr_tmp1, sterr_tmp2);
-
- staterr = vreinterpretq_u16_u32(sterr_u32);
+ uint16x8_t sterr_tmp1 = vzip2q_u16(vreinterpretq_u16_u64(descs[0]),
+ vreinterpretq_u16_u64(descs[2]));
+ uint16x8_t sterr_tmp2 = vzip2q_u16(vreinterpretq_u16_u64(descs[1]),
+ vreinterpretq_u16_u64(descs[3]));
+ staterr = vzip1q_u16(sterr_tmp1, sterr_tmp2);
}
- sxe2_rx_desc_ptype_fill_neon(staterr, &rx_pkts[i]);
+ sxe2_rx_desc_ptype_fill_neon(desc_lo, &rx_pkts[i], sxe2_ptype_tbl);
if (umbcast_flags != NULL) {
- uint32x4_t umbcast_mask = {
- SXE2_RX_DESC_STATUS_UMBCAST_MASK, SXE2_RX_DESC_STATUS_UMBCAST_MASK,
- SXE2_RX_DESC_STATUS_UMBCAST_MASK, SXE2_RX_DESC_STATUS_UMBCAST_MASK,
- };
-
+ const uint32x4_t umbcast_mask =
+ vdupq_n_u32(SXE2_RX_DESC_STATUS_UMBCAST_MASK);
uint8x16_t umbcast_shuf_mask = {
- 0x0B, 0x03, 0x0F, 0x07,
+ 3, 7, 11, 15,
0xFF, 0xFF, 0xFF, 0xFF,
0xFF, 0xFF, 0xFF, 0xFF,
0xFF, 0xFF, 0xFF, 0xFF,
};
uint8x16_t umbcast_bits =
- vreinterpretq_u8_u32(vandq_u32(vreinterpretq_u32_u16(staterr),
- umbcast_mask));
+ vreinterpretq_u8_u32(vandq_u32(desc_lo, umbcast_mask));
umbcast_bits = vqtbl1q_u8(umbcast_bits, umbcast_shuf_mask);
- vst1q_lane_u32((uint32_t *)umbcast_flags,
- vreinterpretq_u32_u8(umbcast_bits), 0);
+ *(uint32_t *)umbcast_flags =
+ vgetq_lane_u32(vreinterpretq_u32_u8(umbcast_bits), 0);
umbcast_flags += SXE2_RX_NUM_PER_LOOP_NEON;
}
if (split_rxe_flags) {
uint8x16_t eop_shuf_mask = {
- 0x08, 0x00, 0x0C, 0x04,
+ 0, 2, 4, 6,
0xFF, 0xFF, 0xFF, 0xFF,
0xFF, 0xFF, 0xFF, 0xFF,
0xFF, 0xFF, 0xFF, 0xFF};
uint8x16_t eop_bits;
uint32x4_t rxe_mask = {
- 0x2080, 0x2080, 0x2080, 0x2080
+ 0x20802080, 0x20802080, 0x20802080, 0x20802080
};
uint32x4_t rxe_bits;
uint32x4_t eop_mask;
- eop_mask = vshlq_n_u32(vdupq_n_u32(1), SXE2_RX_DESC_STATUS_EOP_SHIFT);
+ eop_mask = vdupq_n_u32((1U << SXE2_RX_DESC_STATUS_EOP_SHIFT) |
+ (1U << (SXE2_RX_DESC_STATUS_EOP_SHIFT + 16)));
eop_bits = vandq_u8(vmvnq_u8(vreinterpretq_u8_u16(staterr)),
vreinterpretq_u8_u32(eop_mask));
@@ -650,12 +644,22 @@ sxe2_rx_pkts_common_vec_neon(struct sxe2_rx_queue *rxq, struct rte_mbuf **rx_pkt
}
{
- uint32x4_t dd_mask = vdupq_n_u32(1);
- uint32x4_t sterr_dd = vandq_u32(vreinterpretq_u32_u16(staterr), dd_mask);
- uint16x4_t packed_lo = vmovn_u32(sterr_dd);
- uint64_t dd64 = vget_lane_u64(vreinterpret_u64_u16(packed_lo), 0);
-
- bit_num = (uint16_t)rte_popcount64(dd64);
+ const uint16x8_t dd_check = {
+ 0x0001, 0x0001, 0x0001, 0x0001,
+ 0, 0, 0, 0
+ };
+ uint16x8_t sterr_dd;
+ uint64_t stat;
+ sterr_dd = vandq_u16(staterr, dd_check);
+ sterr_dd = vshlq_n_u16(sterr_dd, 15);
+ sterr_dd =
+ vreinterpretq_u16_s16(vshrq_n_s16(vreinterpretq_s16_u16(sterr_dd),
+ 15));
+ stat = ~vgetq_lane_u64(vreinterpretq_u64_u16(sterr_dd), 0);
+ if (likely(stat == 0))
+ bit_num = SXE2_RX_NUM_PER_LOOP_NEON;
+ else
+ bit_num = (uint16_t)(rte_ctz64(stat) / 16);
}
done_num += bit_num;
if (likely(bit_num != SXE2_RX_NUM_PER_LOOP_NEON))
--
2.52.0
More information about the dev
mailing list