Re: [PATCH net-next] mlx4: Better use of order-0 pages in RX path
From: Eric Dumazet <hidden>
Date: 2017-03-14 14:21:51
Subsystem:
mellanox ethernet driver (mlx4_en), mellanox mlx4 core vpi driver, networking drivers, the rest · Maintainers:
Tariq Toukan, Andrew Lunn, "David S. Miller", Eric Dumazet, Jakub Kicinski, Paolo Abeni, Linus Torvalds
On Tue, 2017-03-14 at 06:34 -0700, Eric Dumazet wrote:
So I will leave this to Mellanox for XDP tests and upstreaming this, and will stop arguing with you, this is going nowhere.
Tariq, I will send a v2, including these changes (plus the missing include of yesterday) One is to make sure high order allocations remove __GFP_DIRECT_RECLAIM The other is changing mlx4_en_recover_from_oom() to increase by one rx_alloc_order instead of plain reset to rx_pref_alloc_order Please test XDP and tell me if you find any issues ? Thanks !
diff --git a/drivers/net/ethernet/mellanox/mlx4/en_rx.c b/drivers/net/ethernet/mellanox/mlx4/en_rx.c
index a71554649c25383bb765fa8220bc9cd490247aee..cc41f2f145541b469b52e7014659d5fdbb7dac68 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_rx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_rx.c@@ -60,8 +60,10 @@ static struct page *mlx4_alloc_page(struct mlx4_en_priv *priv, if (unlikely(!ring->pre_allocated_count)) { unsigned int order = READ_ONCE(ring->rx_alloc_order); - page = __alloc_pages_node(node, gfp | __GFP_NOMEMALLOC | - __GFP_NOWARN | __GFP_NORETRY, + page = __alloc_pages_node(node, (gfp & ~__GFP_DIRECT_RECLAIM) | + __GFP_NOMEMALLOC | + __GFP_NOWARN | + __GFP_NORETRY, order); if (page) { split_page(page, order);
@@ -412,12 +414,13 @@ int mlx4_en_activate_rx_rings(struct mlx4_en_priv *priv) } /* Under memory pressure, each ring->rx_alloc_order might be lowered - * to very small values. Periodically reset it to initial value for + * to very small values. Periodically increase t to initial value for * optimal allocations, in case stress is over. */ void mlx4_en_recover_from_oom(struct mlx4_en_priv *priv) { struct mlx4_en_rx_ring *ring; + unsigned int order; int ring_ind; if (!priv->port_up)
@@ -425,7 +428,9 @@ void mlx4_en_recover_from_oom(struct mlx4_en_priv *priv) for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) { ring = priv->rx_ring[ring_ind]; - WRITE_ONCE(ring->rx_alloc_order, ring->rx_pref_alloc_order); + order = min_t(unsigned int, ring->rx_alloc_order + 1, + ring->rx_pref_alloc_order); + WRITE_ONCE(ring->rx_alloc_order, order); } }