aboutsummaryrefslogtreecommitdiffstats
path: root/src/vnet
diff options
context:
space:
mode:
authorZhiyong Yang <zhiyong.yang@intel.com>2020-05-15 22:32:34 +0800
committerDamjan Marion <dmarion@me.com>2020-06-27 10:26:13 +0000
commit975a3cbb130fe7f3d5415dee7c05ea5e7bc59daf (patch)
tree7e4cc9763ac70b365a4ca2665d254b024dba04e1 /src/vnet
parentd352bf8276f33f33e4af5e8a7b85dad8187f531d (diff)
l2: performance enhancement in l2input
Short Load/Stores combined with prefetching in the beginning of the loop place too much pressure on AGUs and memory accesses. The patch interleaves load/store operations with computational operations to alleviate the pain point. vlib_get_buffers is also leveraged. Redefine u8 dst_and_src[12] instead of dst[6] and src[6] in struct l2input_trace_t in order to merge two copys into one. Type: improvement Signed-off-by: Zhiyong Yang <zhiyong.yang@intel.com> Change-Id: I7d3df7732c476069235e3019c68f0f53bca9637e
Diffstat (limited to 'src/vnet')
-rw-r--r--src/vnet/l2/l2_input.c149
1 files changed, 69 insertions, 80 deletions
diff --git a/src/vnet/l2/l2_input.c b/src/vnet/l2/l2_input.c
index 8a7ab802175..542bd42c4c2 100644
--- a/src/vnet/l2/l2_input.c
+++ b/src/vnet/l2/l2_input.c
@@ -101,8 +101,7 @@ format_l2_input_features (u8 * s, va_list * args)
typedef struct
{
/* per-pkt trace data */
- u8 src[6];
- u8 dst[6];
+ u8 dst_and_src[12];
u32 next_index;
u32 sw_if_index;
} l2input_trace_t;
@@ -117,8 +116,8 @@ format_l2input_trace (u8 * s, va_list * args)
s = format (s, "l2-input: sw_if_index %d dst %U src %U",
t->sw_if_index,
- format_ethernet_address, t->dst,
- format_ethernet_address, t->src);
+ format_ethernet_address, t->dst_and_src,
+ format_ethernet_address, t->dst_and_src + 6);
return s;
}
@@ -298,10 +297,12 @@ l2input_node_inline (vlib_main_t * vm,
u32 n_left_from, *from, *to_next;
l2input_next_t next_index;
l2input_main_t *msm = &l2input_main;
+ vlib_buffer_t *bufs[VLIB_FRAME_SIZE], **b = bufs;
from = vlib_frame_vector_args (frame);
n_left_from = frame->n_vectors; /* number of packets to process */
next_index = node->cached_next_index;
+ vlib_get_buffers (vm, from, bufs, n_left_from);
while (n_left_from > 0)
{
@@ -312,30 +313,22 @@ l2input_node_inline (vlib_main_t * vm,
while (n_left_from >= 8 && n_left_to_next >= 4)
{
- u32 bi0, bi1, bi2, bi3;
- vlib_buffer_t *b0, *b1, *b2, *b3;
u32 next0, next1, next2, next3;
u32 sw_if_index0, sw_if_index1, sw_if_index2, sw_if_index3;
/* Prefetch next iteration. */
{
- vlib_buffer_t *p4, *p5, *p6, *p7;
-
- p4 = vlib_get_buffer (vm, from[4]);
- p5 = vlib_get_buffer (vm, from[5]);
- p6 = vlib_get_buffer (vm, from[6]);
- p7 = vlib_get_buffer (vm, from[7]);
/* Prefetch the buffer header and packet for the N+2 loop iteration */
- vlib_prefetch_buffer_header (p4, LOAD);
- vlib_prefetch_buffer_header (p5, LOAD);
- vlib_prefetch_buffer_header (p6, LOAD);
- vlib_prefetch_buffer_header (p7, LOAD);
+ 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 (p4->data, CLIB_CACHE_LINE_BYTES, STORE);
- CLIB_PREFETCH (p5->data, CLIB_CACHE_LINE_BYTES, STORE);
- CLIB_PREFETCH (p6->data, CLIB_CACHE_LINE_BYTES, STORE);
- CLIB_PREFETCH (p7->data, CLIB_CACHE_LINE_BYTES, STORE);
+ CLIB_PREFETCH (b[4]->data, CLIB_CACHE_LINE_BYTES, STORE);
+ CLIB_PREFETCH (b[5]->data, CLIB_CACHE_LINE_BYTES, STORE);
+ CLIB_PREFETCH (b[6]->data, CLIB_CACHE_LINE_BYTES, STORE);
+ CLIB_PREFETCH (b[7]->data, CLIB_CACHE_LINE_BYTES, STORE);
/*
* Don't bother prefetching the bridge-domain config (which
@@ -347,112 +340,108 @@ l2input_node_inline (vlib_main_t * vm,
/* speculatively enqueue b0 and b1 to the current next frame */
/* bi is "buffer index", b is pointer to the buffer */
- to_next[0] = bi0 = from[0];
- to_next[1] = bi1 = from[1];
- to_next[2] = bi2 = from[2];
- to_next[3] = bi3 = from[3];
- from += 4;
- to_next += 4;
- n_left_from -= 4;
- n_left_to_next -= 4;
-
- b0 = vlib_get_buffer (vm, bi0);
- b1 = vlib_get_buffer (vm, bi1);
- b2 = vlib_get_buffer (vm, bi2);
- b3 = vlib_get_buffer (vm, bi3);
if (do_trace)
{
/* RX interface handles */
- sw_if_index0 = vnet_buffer (b0)->sw_if_index[VLIB_RX];
- sw_if_index1 = vnet_buffer (b1)->sw_if_index[VLIB_RX];
- sw_if_index2 = vnet_buffer (b2)->sw_if_index[VLIB_RX];
- sw_if_index3 = vnet_buffer (b3)->sw_if_index[VLIB_RX];
+ sw_if_index0 = vnet_buffer (b[0])->sw_if_index[VLIB_RX];
+ sw_if_index1 = vnet_buffer (b[1])->sw_if_index[VLIB_RX];
+ sw_if_index2 = vnet_buffer (b[2])->sw_if_index[VLIB_RX];
+ sw_if_index3 = vnet_buffer (b[3])->sw_if_index[VLIB_RX];
- if (b0->flags & VLIB_BUFFER_IS_TRACED)
+ if (b[0]->flags & VLIB_BUFFER_IS_TRACED)
{
- ethernet_header_t *h0 = vlib_buffer_get_current (b0);
+ ethernet_header_t *h0 = vlib_buffer_get_current (b[0]);
l2input_trace_t *t =
- vlib_add_trace (vm, node, b0, sizeof (*t));
+ vlib_add_trace (vm, node, b[0], sizeof (*t));
t->sw_if_index = sw_if_index0;
- clib_memcpy_fast (t->src, h0->src_address, 6);
- clib_memcpy_fast (t->dst, h0->dst_address, 6);
+ clib_memcpy_fast (t->dst_and_src, h0->dst_address,
+ sizeof (h0->dst_address) +
+ sizeof (h0->src_address));
}
- if (b1->flags & VLIB_BUFFER_IS_TRACED)
+ if (b[1]->flags & VLIB_BUFFER_IS_TRACED)
{
- ethernet_header_t *h1 = vlib_buffer_get_current (b1);
+ ethernet_header_t *h1 = vlib_buffer_get_current (b[1]);
l2input_trace_t *t =
- vlib_add_trace (vm, node, b1, sizeof (*t));
+ vlib_add_trace (vm, node, b[1], sizeof (*t));
t->sw_if_index = sw_if_index1;
- clib_memcpy_fast (t->src, h1->src_address, 6);
- clib_memcpy_fast (t->dst, h1->dst_address, 6);
+ clib_memcpy_fast (t->dst_and_src, h1->dst_address,
+ sizeof (h1->dst_address) +
+ sizeof (h1->src_address));
}
- if (b2->flags & VLIB_BUFFER_IS_TRACED)
+ if (b[2]->flags & VLIB_BUFFER_IS_TRACED)
{
- ethernet_header_t *h2 = vlib_buffer_get_current (b2);
+ ethernet_header_t *h2 = vlib_buffer_get_current (b[2]);
l2input_trace_t *t =
- vlib_add_trace (vm, node, b2, sizeof (*t));
+ vlib_add_trace (vm, node, b[2], sizeof (*t));
t->sw_if_index = sw_if_index2;
- clib_memcpy_fast (t->src, h2->src_address, 6);
- clib_memcpy_fast (t->dst, h2->dst_address, 6);
+ clib_memcpy_fast (t->dst_and_src, h2->dst_address,
+ sizeof (h2->dst_address) +
+ sizeof (h2->src_address));
}
- if (b3->flags & VLIB_BUFFER_IS_TRACED)
+ if (b[3]->flags & VLIB_BUFFER_IS_TRACED)
{
- ethernet_header_t *h3 = vlib_buffer_get_current (b3);
+ ethernet_header_t *h3 = vlib_buffer_get_current (b[3]);
l2input_trace_t *t =
- vlib_add_trace (vm, node, b3, sizeof (*t));
+ vlib_add_trace (vm, node, b[3], sizeof (*t));
t->sw_if_index = sw_if_index3;
- clib_memcpy_fast (t->src, h3->src_address, 6);
- clib_memcpy_fast (t->dst, h3->dst_address, 6);
+ clib_memcpy_fast (t->dst_and_src, h3->dst_address,
+ sizeof (h3->dst_address) +
+ sizeof (h3->src_address));
}
}
- classify_and_dispatch (msm, b0, &next0);
- classify_and_dispatch (msm, b1, &next1);
- classify_and_dispatch (msm, b2, &next2);
- classify_and_dispatch (msm, b3, &next3);
+ classify_and_dispatch (msm, b[0], &next0);
+ classify_and_dispatch (msm, b[1], &next1);
+ //show the better performance when clib_memcpy_fast is put here.
+ clib_memcpy_fast (to_next, from, sizeof (from[0]) * 4);
+ to_next += 4;
+ classify_and_dispatch (msm, b[2], &next2);
+ classify_and_dispatch (msm, b[3], &next3);
+ b += 4;
+ n_left_from -= 4;
+ n_left_to_next -= 4;
/* verify speculative enqueues, maybe switch current next frame */
/* if next0==next1==next_index then nothing special needs to be done */
vlib_validate_buffer_enqueue_x4 (vm, node, next_index,
to_next, n_left_to_next,
- bi0, bi1, bi2, bi3,
+ from[0], from[1], from[2], from[3],
next0, next1, next2, next3);
+ from += 4;
}
while (n_left_from > 0 && n_left_to_next > 0)
{
- u32 bi0;
- vlib_buffer_t *b0;
u32 next0;
u32 sw_if_index0;
/* speculatively enqueue b0 to the current next frame */
- bi0 = from[0];
- to_next[0] = bi0;
- from += 1;
- to_next += 1;
- n_left_from -= 1;
- n_left_to_next -= 1;
-
- b0 = vlib_get_buffer (vm, bi0);
- if (do_trace && PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED))
+ if (do_trace && PREDICT_FALSE (b[0]->flags & VLIB_BUFFER_IS_TRACED))
{
- ethernet_header_t *h0 = vlib_buffer_get_current (b0);
- l2input_trace_t *t = vlib_add_trace (vm, node, b0, sizeof (*t));
- sw_if_index0 = vnet_buffer (b0)->sw_if_index[VLIB_RX];
+ ethernet_header_t *h0 = vlib_buffer_get_current (b[0]);
+ l2input_trace_t *t =
+ vlib_add_trace (vm, node, b[0], sizeof (*t));
+ sw_if_index0 = vnet_buffer (b[0])->sw_if_index[VLIB_RX];
t->sw_if_index = sw_if_index0;
- clib_memcpy_fast (t->src, h0->src_address, 6);
- clib_memcpy_fast (t->dst, h0->dst_address, 6);
+ clib_memcpy_fast (t->dst_and_src, h0->dst_address,
+ sizeof (h0->dst_address) +
+ sizeof (h0->src_address));
}
- classify_and_dispatch (msm, b0, &next0);
+ classify_and_dispatch (msm, b[0], &next0);
+ b += 1;
+ to_next[0] = from[0];
+ to_next += 1;
+ n_left_from -= 1;
+ n_left_to_next -= 1;
/* verify speculative enqueue, maybe switch current next frame */
vlib_validate_buffer_enqueue_x1 (vm, node, next_index,
to_next, n_left_to_next,
- bi0, next0);
+ from[0], next0);
+ from += 1;
}
vlib_put_next_frame (vm, node, next_index, n_left_to_next);