提交 9f9e9cd5 编写于 作者: T Tariq Toukan 提交者: Saeed Mahameed

net/mlx5e: Remove page_ref bulking in Striding RQ

When many packets reside on the same page, the bulking of
page_ref modifications reduces the total number of atomic
operations executed.

Besides the necessary 2 operations on page alloc/free, we
have the following extra ops per page:
- one on WQE allocation (bump refcnt to maximum possible),
- zero ops for SKBs,
- one on WQE free,
a constant of two operations in total, no matter how many
packets/SKBs actually populate the page.

Without this bulking, we have:
- no ops on WQE allocation or free,
- one op per SKB,

Comparing the two methods when PAGE_SIZE is 4K:
- As mentioned above, bulking method always executes 2 operations,
  not more, but not less.
- In the default MTU configuration (1500, stride size is 2K),
  the non-bulking method execute 2 ops as well.
- For larger MTUs with stride size of 4K, non-bulking method
  executes only a single op.
- For XDP (stride size of 4K, no SKBs), non-bulking method
  executes no ops at all!

Hence, to optimize the flows with linear SKB and XDP over Striding RQ,
we here remove the page_ref bulking method.

Performance testing:
ConnectX-5, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz.

Single core packet rate (64 bytes).

Early drop in TC: no degradation.

XDP_DROP:
before: 14,270,188 pps
after:  20,503,603 pps, 43% improvement.
Signed-off-by: NTariq Toukan <tariqt@mellanox.com>
Signed-off-by: NSaeed Mahameed <saeedm@mellanox.com>
上级 22f45398
......@@ -456,7 +456,6 @@ struct mlx5e_umr_dma_info {
struct mlx5e_mpw_info {
struct mlx5e_umr_dma_info umr;
u16 consumed_strides;
u16 skbs_frags[MLX5_MPWRQ_PAGES_PER_WQE];
DECLARE_BITMAP(xdp_xmit_bitmap, MLX5_MPWRQ_PAGES_PER_WQE);
};
......
......@@ -296,37 +296,28 @@ void mlx5e_dealloc_rx_wqe(struct mlx5e_rq *rq, u16 ix)
mlx5e_free_rx_wqe(rq, wi);
}
static inline int mlx5e_mpwqe_strides_per_page(struct mlx5e_rq *rq)
{
return rq->mpwqe.num_strides >> MLX5_MPWRQ_WQE_PAGE_ORDER;
}
static inline void mlx5e_add_skb_frag_mpwqe(struct mlx5e_rq *rq,
struct sk_buff *skb,
struct mlx5e_mpw_info *wi,
u32 page_idx, u32 frag_offset,
u32 len)
struct mlx5e_dma_info *di,
u32 frag_offset, u32 len)
{
unsigned int truesize = ALIGN(len, BIT(rq->mpwqe.log_stride_sz));
dma_sync_single_for_cpu(rq->pdev,
wi->umr.dma_info[page_idx].addr + frag_offset,
di->addr + frag_offset,
len, DMA_FROM_DEVICE);
wi->skbs_frags[page_idx]++;
page_ref_inc(di->page);
skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
wi->umr.dma_info[page_idx].page, frag_offset,
len, truesize);
di->page, frag_offset, len, truesize);
}
static inline void
mlx5e_copy_skb_header_mpwqe(struct device *pdev,
struct sk_buff *skb,
struct mlx5e_mpw_info *wi,
u32 page_idx, u32 offset,
u32 headlen)
struct mlx5e_dma_info *dma_info,
u32 offset, u32 headlen)
{
u16 headlen_pg = min_t(u32, headlen, PAGE_SIZE - offset);
struct mlx5e_dma_info *dma_info = &wi->umr.dma_info[page_idx];
unsigned int len;
/* Aligning len to sizeof(long) optimizes memcpy performance */
......@@ -351,15 +342,12 @@ void mlx5e_free_rx_mpwqe(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi)
{
const bool no_xdp_xmit =
bitmap_empty(wi->xdp_xmit_bitmap, MLX5_MPWRQ_PAGES_PER_WQE);
int pg_strides = mlx5e_mpwqe_strides_per_page(rq);
struct mlx5e_dma_info *dma_info = wi->umr.dma_info;
int i;
for (i = 0; i < MLX5_MPWRQ_PAGES_PER_WQE; i++) {
page_ref_sub(dma_info[i].page, pg_strides - wi->skbs_frags[i]);
for (i = 0; i < MLX5_MPWRQ_PAGES_PER_WQE; i++)
if (no_xdp_xmit || !test_bit(i, wi->xdp_xmit_bitmap))
mlx5e_page_release(rq, &dma_info[i], true);
}
}
static void mlx5e_post_rx_mpwqe(struct mlx5e_rq *rq)
......@@ -380,7 +368,6 @@ static void mlx5e_post_rx_mpwqe(struct mlx5e_rq *rq)
static int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
{
struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
int pg_strides = mlx5e_mpwqe_strides_per_page(rq);
struct mlx5e_dma_info *dma_info = &wi->umr.dma_info[0];
struct mlx5e_icosq *sq = &rq->channel->icosq;
struct mlx5_wq_cyc *wq = &sq->wq;
......@@ -403,10 +390,8 @@ static int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
if (unlikely(err))
goto err_unmap;
umr_wqe->inline_mtts[i].ptag = cpu_to_be64(dma_info->addr | MLX5_EN_WR);
page_ref_add(dma_info->page, pg_strides);
}
memset(wi->skbs_frags, 0, sizeof(*wi->skbs_frags) * MLX5_MPWRQ_PAGES_PER_WQE);
bitmap_zero(wi->xdp_xmit_bitmap, MLX5_MPWRQ_PAGES_PER_WQE);
wi->consumed_strides = 0;
......@@ -425,7 +410,6 @@ static int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
err_unmap:
while (--i >= 0) {
dma_info--;
page_ref_sub(dma_info->page, pg_strides);
mlx5e_page_release(rq, dma_info, true);
}
rq->stats.buff_alloc_err++;
......@@ -987,9 +971,10 @@ mlx5e_skb_from_cqe_mpwrq_nonlinear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *w
u16 cqe_bcnt, u32 head_offset, u32 page_idx)
{
u16 headlen = min_t(u16, MLX5_MPWRQ_SMALL_PACKET_THRESHOLD, cqe_bcnt);
struct mlx5e_dma_info *di = &wi->umr.dma_info[page_idx];
u32 frag_offset = head_offset + headlen;
u16 byte_cnt = cqe_bcnt - headlen;
u32 head_page_idx = page_idx;
u32 byte_cnt = cqe_bcnt - headlen;
struct mlx5e_dma_info *head_di = di;
struct sk_buff *skb;
skb = napi_alloc_skb(rq->cq.napi,
......@@ -1002,7 +987,7 @@ mlx5e_skb_from_cqe_mpwrq_nonlinear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *w
prefetchw(skb->data);
if (unlikely(frag_offset >= PAGE_SIZE)) {
page_idx++;
di++;
frag_offset -= PAGE_SIZE;
}
......@@ -1010,14 +995,14 @@ mlx5e_skb_from_cqe_mpwrq_nonlinear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *w
u32 pg_consumed_bytes =
min_t(u32, PAGE_SIZE - frag_offset, byte_cnt);
mlx5e_add_skb_frag_mpwqe(rq, skb, wi, page_idx, frag_offset,
mlx5e_add_skb_frag_mpwqe(rq, skb, di, frag_offset,
pg_consumed_bytes);
byte_cnt -= pg_consumed_bytes;
frag_offset = 0;
page_idx++;
di++;
}
/* copy header */
mlx5e_copy_skb_header_mpwqe(rq->pdev, skb, wi, head_page_idx,
mlx5e_copy_skb_header_mpwqe(rq->pdev, skb, head_di,
head_offset, headlen);
/* skb linear part was allocated with headlen and aligned to long */
skb->tail += headlen;
......@@ -1060,7 +1045,7 @@ mlx5e_skb_from_cqe_mpwrq_linear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi,
return NULL;
/* queue up for recycling/reuse */
wi->skbs_frags[page_idx]++;
page_ref_inc(di->page);
return skb;
}
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册