From: Yongseok Koh <yskoh@mellanox.com>
To: adrien.mazarguil@6wind.com, nelio.laranjeiro@6wind.com,
jerin.jacob@caviumnetworks.com, jianbo.liu@arm.com
Cc: dev@dpdk.org, Yongseok Koh <yskoh@mellanox.com>, stable@dpdk.org
Subject: [dpdk-stable] [PATCH 2/2] net/mlx5: fix synchonization on polling Rx completions
Date: Tue, 26 Dec 2017 20:28:24 -0800 [thread overview]
Message-ID: <20171227042824.33373-2-yskoh@mellanox.com> (raw)
In-Reply-To: <20171227042824.33373-1-yskoh@mellanox.com>
Polling a new packet is basically sensing the generation bit in a
completion entry. For some processors not having strongly-ordered memory
model, there has to be an IO memory barrier between reading the generation
bit and other fields of the entry in order to guarantee data is not stale.
Fixes: 570acdb1da8a ("net/mlx5: add vectorized Rx/Tx burst for ARM")
Cc: stable@dpdk.org
Signed-off-by: Yongseok Koh <yskoh@mellanox.com>
Acked-by: Shahaf Shuler <shahafs@mellanox.com>
---
drivers/net/mlx5/mlx5_rxtx.c | 1 +
drivers/net/mlx5/mlx5_rxtx_vec_neon.h | 53 ++++++++++++++++++++---------------
drivers/net/mlx5/mlx5_rxtx_vec_sse.h | 2 +-
3 files changed, 32 insertions(+), 24 deletions(-)
diff --git a/drivers/net/mlx5/mlx5_rxtx.c b/drivers/net/mlx5/mlx5_rxtx.c
index 28c0ad8ab..ad7545e3c 100644
--- a/drivers/net/mlx5/mlx5_rxtx.c
+++ b/drivers/net/mlx5/mlx5_rxtx.c
@@ -1674,6 +1674,7 @@ mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
return 0;
++rxq->cq_ci;
op_own = cqe->op_own;
+ rte_io_rmb();
if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED) {
volatile struct mlx5_mini_cqe8 (*mc)[8] =
(volatile struct mlx5_mini_cqe8 (*)[8])
diff --git a/drivers/net/mlx5/mlx5_rxtx_vec_neon.h b/drivers/net/mlx5/mlx5_rxtx_vec_neon.h
index 77ce0c3e0..39b7b1953 100644
--- a/drivers/net/mlx5/mlx5_rxtx_vec_neon.h
+++ b/drivers/net/mlx5/mlx5_rxtx_vec_neon.h
@@ -798,6 +798,7 @@ rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
uint16x4_t mask;
uint16x4_t byte_cnt;
uint32x4_t ptype_info, flow_tag;
+ register uint64x2_t c0, c1, c2, c3;
uint8_t *p0, *p1, *p2, *p3;
uint8_t *e0 = (void *)&elts[pos]->pkt_len;
uint8_t *e1 = (void *)&elts[pos + 1]->pkt_len;
@@ -814,6 +815,16 @@ rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
p1 = p0 + (pkts_n - pos > 1) * sizeof(struct mlx5_cqe);
p2 = p1 + (pkts_n - pos > 2) * sizeof(struct mlx5_cqe);
p3 = p2 + (pkts_n - pos > 3) * sizeof(struct mlx5_cqe);
+ /* B.0 (CQE 3) load a block having op_own. */
+ c3 = vld1q_u64((uint64_t *)(p3 + 48));
+ /* B.0 (CQE 2) load a block having op_own. */
+ c2 = vld1q_u64((uint64_t *)(p2 + 48));
+ /* B.0 (CQE 1) load a block having op_own. */
+ c1 = vld1q_u64((uint64_t *)(p1 + 48));
+ /* B.0 (CQE 0) load a block having op_own. */
+ c0 = vld1q_u64((uint64_t *)(p0 + 48));
+ /* Synchronize for loading the rest of blocks. */
+ rte_io_rmb();
/* Prefetch next 4 CQEs. */
if (pkts_n - pos >= 2 * MLX5_VPMD_DESCS_PER_LOOP) {
unsigned int next = pos + MLX5_VPMD_DESCS_PER_LOOP;
@@ -823,50 +834,46 @@ rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
rte_prefetch_non_temporal(&cq[next + 3]);
}
__asm__ volatile (
- /* B.1 (CQE 3) load a block having op_own. */
- "ld1 {v19.16b}, [%[p3]] \n\t"
- "sub %[p3], %[p3], #48 \n\t"
- /* B.2 (CQE 3) load the rest blocks. */
+ /* B.1 (CQE 3) load the rest of blocks. */
"ld1 {v16.16b - v18.16b}, [%[p3]] \n\t"
+ /* B.2 (CQE 3) move the block having op_own. */
+ "mov v19.16b, %[c3].16b \n\t"
/* B.3 (CQE 3) extract 16B fields. */
"tbl v23.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
+ /* B.1 (CQE 2) load the rest of blocks. */
+ "ld1 {v16.16b - v18.16b}, [%[p2]] \n\t"
/* B.4 (CQE 3) adjust CRC length. */
"sub v23.8h, v23.8h, %[crc_adj].8h \n\t"
- /* B.1 (CQE 2) load a block having op_own. */
- "ld1 {v19.16b}, [%[p2]] \n\t"
- "sub %[p2], %[p2], #48 \n\t"
/* C.1 (CQE 3) generate final structure for mbuf. */
"tbl v15.16b, {v23.16b}, %[mb_shuf_m].16b \n\t"
- /* B.2 (CQE 2) load the rest blocks. */
- "ld1 {v16.16b - v18.16b}, [%[p2]] \n\t"
+ /* B.2 (CQE 2) move the block having op_own. */
+ "mov v19.16b, %[c2].16b \n\t"
/* B.3 (CQE 2) extract 16B fields. */
"tbl v22.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
+ /* B.1 (CQE 1) load the rest of blocks. */
+ "ld1 {v16.16b - v18.16b}, [%[p1]] \n\t"
/* B.4 (CQE 2) adjust CRC length. */
"sub v22.8h, v22.8h, %[crc_adj].8h \n\t"
- /* B.1 (CQE 1) load a block having op_own. */
- "ld1 {v19.16b}, [%[p1]] \n\t"
- "sub %[p1], %[p1], #48 \n\t"
/* C.1 (CQE 2) generate final structure for mbuf. */
"tbl v14.16b, {v22.16b}, %[mb_shuf_m].16b \n\t"
- /* B.2 (CQE 1) load the rest blocks. */
- "ld1 {v16.16b - v18.16b}, [%[p1]] \n\t"
+ /* B.2 (CQE 1) move the block having op_own. */
+ "mov v19.16b, %[c1].16b \n\t"
/* B.3 (CQE 1) extract 16B fields. */
"tbl v21.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
+ /* B.1 (CQE 0) load the rest of blocks. */
+ "ld1 {v16.16b - v18.16b}, [%[p0]] \n\t"
/* B.4 (CQE 1) adjust CRC length. */
"sub v21.8h, v21.8h, %[crc_adj].8h \n\t"
- /* B.1 (CQE 0) load a block having op_own. */
- "ld1 {v19.16b}, [%[p0]] \n\t"
- "sub %[p0], %[p0], #48 \n\t"
/* C.1 (CQE 1) generate final structure for mbuf. */
"tbl v13.16b, {v21.16b}, %[mb_shuf_m].16b \n\t"
- /* B.2 (CQE 0) load the rest blocks. */
- "ld1 {v16.16b - v18.16b}, [%[p0]] \n\t"
+ /* B.2 (CQE 0) move the block having op_own. */
+ "mov v19.16b, %[c0].16b \n\t"
+ /* A.1 load mbuf pointers. */
+ "ld1 {v24.2d - v25.2d}, [%[elts_p]] \n\t"
/* B.3 (CQE 0) extract 16B fields. */
"tbl v20.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
/* B.4 (CQE 0) adjust CRC length. */
"sub v20.8h, v20.8h, %[crc_adj].8h \n\t"
- /* A.1 load mbuf pointers. */
- "ld1 {v24.2d - v25.2d}, [%[elts_p]] \n\t"
/* D.1 extract op_own byte. */
"tbl %[op_own].8b, {v20.16b - v23.16b}, %[owner_shuf_m].8b \n\t"
/* C.2 (CQE 3) adjust flow mark. */
@@ -901,9 +908,9 @@ rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
[byte_cnt]"=&w"(byte_cnt),
[ptype_info]"=&w"(ptype_info),
[flow_tag]"=&w"(flow_tag)
- :[p3]"r"(p3 + 48), [p2]"r"(p2 + 48),
- [p1]"r"(p1 + 48), [p0]"r"(p0 + 48),
+ :[p3]"r"(p3), [p2]"r"(p2), [p1]"r"(p1), [p0]"r"(p0),
[e3]"r"(e3), [e2]"r"(e2), [e1]"r"(e1), [e0]"r"(e0),
+ [c3]"w"(c3), [c2]"w"(c2), [c1]"w"(c1), [c0]"w"(c0),
[elts_p]"r"(elts_p),
[pkts_p]"r"(pkts_p),
[cqe_shuf_m]"w"(cqe_shuf_m),
diff --git a/drivers/net/mlx5/mlx5_rxtx_vec_sse.h b/drivers/net/mlx5/mlx5_rxtx_vec_sse.h
index f25681184..3b90adffa 100644
--- a/drivers/net/mlx5/mlx5_rxtx_vec_sse.h
+++ b/drivers/net/mlx5/mlx5_rxtx_vec_sse.h
@@ -821,7 +821,7 @@ rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
/* B.2 copy mbuf pointers. */
_mm_storeu_si128((__m128i *)&pkts[pos], mbp1);
_mm_storeu_si128((__m128i *)&pkts[pos + 2], mbp2);
- rte_compiler_barrier();
+ rte_io_rmb();
/* C.1 load remained CQE data and extract necessary fields. */
cqe_tmp2 = _mm_load_si128((__m128i *)&cq[pos + p3]);
cqe_tmp1 = _mm_load_si128((__m128i *)&cq[pos + p2]);
--
2.11.0
next parent reply other threads:[~2017-12-27 4:28 UTC|newest]
Thread overview: 5+ messages / expand[flat|nested] mbox.gz Atom feed top
[not found] <20171227042824.33373-1-yskoh@mellanox.com>
2017-12-27 4:28 ` Yongseok Koh [this message]
[not found] ` <20180116011050.18866-1-yskoh@mellanox.com>
2018-01-16 1:10 ` [dpdk-stable] [PATCH v2 8/8] " Yongseok Koh
2018-01-16 3:53 ` Jianbo Liu
[not found] ` <20180119004430.15305-1-yskoh@mellanox.com>
2018-01-19 0:44 ` [dpdk-stable] [PATCH v3 " Yongseok Koh
[not found] ` <20180125210250.38233-1-yskoh@mellanox.com>
2018-01-25 21:02 ` [dpdk-stable] [PATCH v4 9/9] net/mlx5: fix synchronization " Yongseok Koh
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=20171227042824.33373-2-yskoh@mellanox.com \
--to=yskoh@mellanox.com \
--cc=adrien.mazarguil@6wind.com \
--cc=dev@dpdk.org \
--cc=jerin.jacob@caviumnetworks.com \
--cc=jianbo.liu@arm.com \
--cc=nelio.laranjeiro@6wind.com \
--cc=stable@dpdk.org \
/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).