summaryrefslogtreecommitdiffstats
path: root/src/plugins/rdma
diff options
context:
space:
mode:
authorBenoît Ganne <bganne@cisco.com>2020-04-24 15:52:24 +0200
committerDave Barach <openvpp@barachs.net>2020-04-27 12:19:39 +0000
commit2d25467d130485d804ad91a4df9870d684f70dd6 (patch)
tree2c8a777ef20e342eea1f8d3e36bf6ffb5bf24b86 /src/plugins/rdma
parent81284163a293759bc5c2d6a124639c6796589d15 (diff)
rdma: tx: interleave prefetches
Type: improvement Change-Id: Ic2d9b17cf5e524f3ad2a3c5343fe1230aa360e73 Signed-off-by: Benoît Ganne <bganne@cisco.com>
Diffstat (limited to 'src/plugins/rdma')
-rw-r--r--src/plugins/rdma/output.c37
1 files changed, 13 insertions, 24 deletions
diff --git a/src/plugins/rdma/output.c b/src/plugins/rdma/output.c
index 2e7fb5978a4..5b181485e49 100644
--- a/src/plugins/rdma/output.c
+++ b/src/plugins/rdma/output.c
@@ -294,7 +294,7 @@ rdma_device_output_tx_mlx5 (vlib_main_t * vm,
wrap_around:
wqe = txq->dv_sq_wqes + (tail & sq_mask);
- while (n >= 4)
+ while (n >= 8)
{
u32 flags = b[0]->flags | b[1]->flags | b[2]->flags | b[3]->flags;
if (PREDICT_FALSE (flags & VLIB_BUFFER_NEXT_PRESENT))
@@ -302,18 +302,16 @@ wrap_around:
n_left_from, n, bi, b, wqe,
tail);
- if (PREDICT_TRUE (n >= 8))
- {
- vlib_prefetch_buffer_header (b[4], LOAD);
- vlib_prefetch_buffer_header (b[5], LOAD);
- vlib_prefetch_buffer_header (b[6], LOAD);
- vlib_prefetch_buffer_header (b[7], LOAD);
- CLIB_PREFETCH (wqe + 4, 4 * sizeof (wqe[0]), STORE);
- }
-
+ vlib_prefetch_buffer_header (b[4], LOAD);
rdma_mlx5_wqe_init (wqe + 0, txq->dv_wqe_tmpl, b[0], tail + 0);
+
+ vlib_prefetch_buffer_header (b[5], LOAD);
rdma_mlx5_wqe_init (wqe + 1, txq->dv_wqe_tmpl, b[1], tail + 1);
+
+ vlib_prefetch_buffer_header (b[6], LOAD);
rdma_mlx5_wqe_init (wqe + 2, txq->dv_wqe_tmpl, b[2], tail + 2);
+
+ vlib_prefetch_buffer_header (b[7], LOAD);
rdma_mlx5_wqe_init (wqe + 3, txq->dv_wqe_tmpl, b[3], tail + 3);
b += 4;
@@ -395,33 +393,24 @@ rdma_device_output_tx_ibverb (vlib_main_t * vm,
struct ibv_sge sge[VLIB_FRAME_SIZE], *s = sge;
u32 n = n_left_from;
- while (n >= 4)
+ while (n >= 8)
{
- if (PREDICT_TRUE (n >= 8))
- {
- vlib_prefetch_buffer_header (b[4 + 0], LOAD);
- vlib_prefetch_buffer_header (b[4 + 1], LOAD);
- vlib_prefetch_buffer_header (b[4 + 2], LOAD);
- vlib_prefetch_buffer_header (b[4 + 3], LOAD);
- CLIB_PREFETCH (&s[4 + 0], 4 * sizeof (s[0]), STORE);
- clib_prefetch_store (&w[4 + 0]);
- clib_prefetch_store (&w[4 + 1]);
- clib_prefetch_store (&w[4 + 2]);
- clib_prefetch_store (&w[4 + 3]);
- }
-
+ vlib_prefetch_buffer_header (b[4], LOAD);
s[0].addr = vlib_buffer_get_current_va (b[0]);
s[0].length = b[0]->current_length;
s[0].lkey = rd->lkey;
+ vlib_prefetch_buffer_header (b[5], LOAD);
s[1].addr = vlib_buffer_get_current_va (b[1]);
s[1].length = b[1]->current_length;
s[1].lkey = rd->lkey;
+ vlib_prefetch_buffer_header (b[6], LOAD);
s[2].addr = vlib_buffer_get_current_va (b[2]);
s[2].length = b[2]->current_length;
s[2].lkey = rd->lkey;
+ vlib_prefetch_buffer_header (b[7], LOAD);
s[3].addr = vlib_buffer_get_current_va (b[3]);
s[3].length = b[3]->current_length;
s[3].lkey = rd->lkey;