BNXT changes
cover BNXT changes
This patch set includes changes proposed for BNXT PMD. Found these in code review.
Ruifeng Wang (3):
net/bnxt: defer completion index update net/bnxt: remove redundant ifdefs
net/bnxt: fix risk in Rx descriptor read in NEON path
drivers/net/bnxt/bnxt_rxr.c | 2 +- drivers/net/bnxt/bnxt_rxtx_vec_neon.c | 21 +++++++++++++++------ 2 files changed, 16 insertions(+), 7 deletions(-)
Signature · 1 line
2.25.1
PATCH 1/3 net/bnxt: defer completion index update
When no packet is received, there is no need to update completion raw cons. Moved update down to remove unnecessary store in this case.
--- drivers/net/bnxt/bnxt_rxr.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/net/bnxt/bnxt_rxr.c +1 -1
@@ -1096,7 +1096,6 @@ uint16_t bnxt_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,break ; } - cpr->cp_raw_cons = raw_cons;if (!nb_rx_pkts && !nb_rep_rx_pkts && !evt) {/* * For PMD, there is no need to keep on pushing to REARM
@@ -1105,6 +1104,7 @@ uint16_t bnxt_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, goto done; } + cpr->cp_raw_cons = raw_cons;/* Ring the completion queue doorbell. */ bnxt_db_cq (cpr);
Signature · 1 line
2.25.1
PATCH 2/3 net/bnxt: remove redundant ifdefs
In reply to Ruifeng WangNEON vector path is built only when Arm platform is 64-bit. The ifdefs in NEON path are of no use, hence remove.
--- drivers/net/bnxt/bnxt_rxtx_vec_neon.c | 4 ---- 1 file changed, 4 deletions(-)
drivers/net/bnxt/bnxt_rxtx_vec_neon.c +0 -4
@@ -220,13 +220,9 @@ recv_burst_vec_neon(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)/* Copy four mbuf pointers to output array. */ t0 =vld1q_u64 ((void *)&rxr->rx_buf_ring[mbcons]); -#ifdef RTE_ARCH_ARM64 t1 =vld1q_u64 ((void *)&rxr->rx_buf_ring[mbcons +2 ]); -#endif vst1q_u64 ((void *)&rx_pkts[i], t0); -#ifdef RTE_ARCH_ARM64 vst1q_u64 ((void *)&rx_pkts[i +2 ], t1); -#endif /* Prefetch four descriptor pairs for next iteration. */ if (i +BNXT_RX_DESCS_PER_LOOP_VEC128 < nb_pkts) {
Signature · 1 line
2.25.1
PATCH 3/3 net/bnxt: fix risk in Rx descriptor read in NEON path
In reply to Ruifeng WangRx descriptor contains a valid bit which indicates readiness of the rest of descriptor words. Hence, the word contains valid bit must be read prior to other words.
In NEON vector path, two contiguous 8B descriptor are loaded to a single NEON register. Given vector load ensures no 16B atomicity, read of the word that includes valid bit could be reordered after read of other words. In this case, data could be invalid.
Reloaded lower 64b after read barrier. This ensures what fetched is correct.
Also fixed comments that not pertains to Arm platform architecture.
--- drivers/net/bnxt/bnxt_rxtx_vec_neon.c | 17 +++++++++++++++-- 1 file changed, 15 insertions(+), 2 deletions(-)
drivers/net/bnxt/bnxt_rxtx_vec_neon.c +15 -2
@@ -231,25 +231,38 @@ recv_burst_vec_neon(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts) }/* -* Load the four current descriptors into SSE registers in -* reverse order to ensure consistent state. +* Load the four current descriptors into NEON registers. +* IO barriers are used to ensure consistent state. */ rxcmp1[3 ] =vld1q_u32 ((void *)&cpr->cp_desc_ring[cons +7 ]);rte_io_rmb (); +/* Reload lower 64b of descriptors to make it ordered after info3_v. */ + rxcmp1[3 ] =vreinterpretq_u32_u64 (vld1q_lane_u64 + ((void *)&cpr->cp_desc_ring[cons +7 ], +vreinterpretq_u64_u32 (rxcmp1[3 ]),0 )); rxcmp[3 ] =vld1q_u32 ((void *)&cpr->cp_desc_ring[cons +6 ]); rxcmp1[2 ] =vld1q_u32 ((void *)&cpr->cp_desc_ring[cons +5 ]);rte_io_rmb (); + rxcmp1[2 ] =vreinterpretq_u32_u64 (vld1q_lane_u64 + ((void *)&cpr->cp_desc_ring[cons +5 ], +vreinterpretq_u64_u32 (rxcmp1[2 ]),0 )); rxcmp[2 ] =vld1q_u32 ((void *)&cpr->cp_desc_ring[cons +4 ]); t1 =vreinterpretq_u64_u32 (vzip2q_u32 (rxcmp1[2 ], rxcmp1[3 ])); rxcmp1[1 ] =vld1q_u32 ((void *)&cpr->cp_desc_ring[cons +3 ]);rte_io_rmb (); + rxcmp1[1 ] =vreinterpretq_u32_u64 (vld1q_lane_u64 + ((void *)&cpr->cp_desc_ring[cons +3 ], +vreinterpretq_u64_u32 (rxcmp1[1 ]),0 )); rxcmp[1 ] =vld1q_u32 ((void *)&cpr->cp_desc_ring[cons +2 ]); rxcmp1[0 ] =vld1q_u32 ((void *)&cpr->cp_desc_ring[cons +1 ]);rte_io_rmb (); + rxcmp1[0 ] =vreinterpretq_u32_u64 (vld1q_lane_u64 + ((void *)&cpr->cp_desc_ring[cons +1 ], +vreinterpretq_u64_u32 (rxcmp1[0 ]),0 )); rxcmp[0 ] =vld1q_u32 ((void *)&cpr->cp_desc_ring[cons +0 ]); t0 =vreinterpretq_u64_u32 (vzip2q_u32 (rxcmp1[0 ], rxcmp1[1 ]));
Signature · 1 line
2.25.1
Quoted 8 lines · Ruifeng Wang
This patch set includes changes proposed for BNXT PMD. Found these in code review.
Ruifeng Wang (3):
net/bnxt: defer completion index update net/bnxt: remove redundant ifdefs
net/bnxt: fix risk in Rx descriptor read in NEON path
Thanks Ruifeng.
Let me review the patchset and get back.
Quoted 8 lines
drivers/net/bnxt/bnxt_rxr.c | 2 +- drivers/net/bnxt/bnxt_rxtx_vec_neon.c | 21 +++++++++++++++------ 2 files changed, 16 insertions(+), 7 deletions(-)--
2.25.1
Quoted 6 lines · Ajit Khaparde
Quoted 4 lines · Ruifeng Wang
This patch set includes changes proposed for BNXT PMD. Found these in code review.
Quoted 6 lines
Quoted 4 lines
Ruifeng Wang (3):
net/bnxt: defer completion index update net/bnxt: remove redundant ifdefs
net/bnxt: fix risk in Rx descriptor read in NEON pathThanks Ruifeng.
Let me review the patchset and get back.
Patches applied to dpdk-next-net-brcm. Thanks
Quoted 9 lines
Quoted 9 lines
drivers/net/bnxt/bnxt_rxr.c | 2 +- drivers/net/bnxt/bnxt_rxtx_vec_neon.c | 21 +++++++++++++++------ 2 files changed, 16 insertions(+), 7 deletions(-)--
2.25.1