Merge branch 'eth-mlx4-use-the-page-pool-for-rx-buffers'

Jakub Kicinski says:

====================
eth: mlx4: use the page pool for Rx buffers

Convert mlx4 to page pool. I've been sitting on these patches for
over a year, and Jonathan Lemon had a similar series years before.
We never deployed it or sent upstream because it didn't really show
much perf win under normal load (admittedly I think the real testing
was done before Ilias's work on recycling).

During the v6.9 kernel rollout Meta's CDN team noticed that machines
with CX3 Pro (mlx4) are prone to overloads (double digit % of CPU time
spent mapping buffers in the IOMMU). The problem does not occur with
modern NICs, so I dusted off this series and reportedly it still works.
And it makes the problem go away, no overloads, perf back in line with
older kernels. Something must have changed in IOMMU code, I guess.

This series is very simple, and can very likely be optimized further.
Thing is, I don't have access to any CX3 Pro NICs. They only exist
in CDN locations which haven't had a HW refresh for a while. So I can
say this series survives a week under traffic w/ XDP enabled, but
my ability to iterate and improve is a bit limited.

v2: https://lore.kernel.org/20250211192141.619024-1-kuba@kernel.org
v1: https://lore.kernel.org/20250205031213.358973-1-kuba@kernel.org
====================

Link: https://patch.msgid.link/20250213010635.1354034-1-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

3 files changed, 53 insertions, 100 deletions

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_rx.c b/drivers/net/ethernet/mellanox/mlx4/en_rx.c
index 15c57e9517e9..b33285d755b9 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_rx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_rx.c
@@ -48,60 +48,43 @@
 #if IS_ENABLED(CONFIG_IPV6)
 #include <net/ip6_checksum.h>
 #endif
+#include <net/page_pool/helpers.h>
 
 #include "mlx4_en.h"
 
-static int mlx4_alloc_page(struct mlx4_en_priv *priv,
-			   struct mlx4_en_rx_alloc *frag,
-			   gfp_t gfp)
-{
-	struct page *page;
-	dma_addr_t dma;
-
-	page = alloc_page(gfp);
-	if (unlikely(!page))
-		return -ENOMEM;
-	dma = dma_map_page(priv->ddev, page, 0, PAGE_SIZE, priv->dma_dir);
-	if (unlikely(dma_mapping_error(priv->ddev, dma))) {
-		__free_page(page);
-		return -ENOMEM;
-	}
-	frag->page = page;
-	frag->dma = dma;
-	frag->page_offset = priv->rx_headroom;
-	return 0;
-}
-
 static int mlx4_en_alloc_frags(struct mlx4_en_priv *priv,
 			       struct mlx4_en_rx_ring *ring,
 			       struct mlx4_en_rx_desc *rx_desc,
 			       struct mlx4_en_rx_alloc *frags,
 			       gfp_t gfp)
 {
+	dma_addr_t dma;
 	int i;
 
 	for (i = 0; i < priv->num_frags; i++, frags++) {
 		if (!frags->page) {
-			if (mlx4_alloc_page(priv, frags, gfp)) {
+			frags->page = page_pool_alloc_pages(ring->pp, gfp);
+			if (!frags->page) {
 				ring->alloc_fail++;
 				return -ENOMEM;
 			}
+			page_pool_fragment_page(frags->page, 1);
+			frags->page_offset = priv->rx_headroom;
+
 			ring->rx_alloc_pages++;
 		}
-		rx_desc->data[i].addr = cpu_to_be64(frags->dma +
-						    frags->page_offset);
+		dma = page_pool_get_dma_addr(frags->page);
+		rx_desc->data[i].addr = cpu_to_be64(dma + frags->page_offset);
 	}
 	return 0;
 }
 
 static void mlx4_en_free_frag(const struct mlx4_en_priv *priv,
+			      struct mlx4_en_rx_ring *ring,
 			      struct mlx4_en_rx_alloc *frag)
 {
-	if (frag->page) {
-		dma_unmap_page(priv->ddev, frag->dma,
-			       PAGE_SIZE, priv->dma_dir);
-		__free_page(frag->page);
-	}
+	if (frag->page)
+		page_pool_put_full_page(ring->pp, frag->page, false);
 	/* We need to clear all fields, otherwise a change of priv->log_rx_info
 	 * could lead to see garbage later in frag->page.
 	 */
@@ -141,18 +124,6 @@ static int mlx4_en_prepare_rx_desc(struct mlx4_en_priv *priv,
 		(index << ring->log_stride);
 	struct mlx4_en_rx_alloc *frags = ring->rx_info +
 					(index << priv->log_rx_info);
-	if (likely(ring->page_cache.index > 0)) {
-		/* XDP uses a single page per frame */
-		if (!frags->page) {
-			ring->page_cache.index--;
-			frags->page = ring->page_cache.buf[ring->page_cache.index].page;
-			frags->dma  = ring->page_cache.buf[ring->page_cache.index].dma;
-		}
-		frags->page_offset = XDP_PACKET_HEADROOM;
-		rx_desc->data[0].addr = cpu_to_be64(frags->dma +
-						    XDP_PACKET_HEADROOM);
-		return 0;
-	}
 
 	return mlx4_en_alloc_frags(priv, ring, rx_desc, frags, gfp);
 }
@@ -178,7 +149,7 @@ static void mlx4_en_free_rx_desc(const struct mlx4_en_priv *priv,
 	frags = ring->rx_info + (index << priv->log_rx_info);
 	for (nr = 0; nr < priv->num_frags; nr++) {
 		en_dbg(DRV, priv, "Freeing fragment:%d\n", nr);
-		mlx4_en_free_frag(priv, frags + nr);
+		mlx4_en_free_frag(priv, ring, frags + nr);
 	}
 }
 
@@ -268,6 +239,7 @@ int mlx4_en_create_rx_ring(struct mlx4_en_priv *priv,
 			   u32 size, u16 stride, int node, int queue_index)
 {
 	struct mlx4_en_dev *mdev = priv->mdev;
+	struct page_pool_params pp = {};
 	struct mlx4_en_rx_ring *ring;
 	int err = -ENOMEM;
 	int tmp;
@@ -286,9 +258,26 @@ int mlx4_en_create_rx_ring(struct mlx4_en_priv *priv,
 	ring->log_stride = ffs(ring->stride) - 1;
 	ring->buf_size = ring->size * ring->stride + TXBB_SIZE;
 
-	if (xdp_rxq_info_reg(&ring->xdp_rxq, priv->dev, queue_index, 0) < 0)
+	pp.flags = PP_FLAG_DMA_MAP;
+	pp.pool_size = size * DIV_ROUND_UP(priv->rx_skb_size, PAGE_SIZE);
+	pp.nid = node;
+	pp.napi = &priv->rx_cq[queue_index]->napi;
+	pp.netdev = priv->dev;
+	pp.dev = &mdev->dev->persist->pdev->dev;
+	pp.dma_dir = priv->dma_dir;
+
+	ring->pp = page_pool_create(&pp);
+	if (!ring->pp)
 		goto err_ring;
 
+	if (xdp_rxq_info_reg(&ring->xdp_rxq, priv->dev, queue_index, 0) < 0)
+		goto err_pp;
+
+	err = xdp_rxq_info_reg_mem_model(&ring->xdp_rxq, MEM_TYPE_PAGE_POOL,
+					 ring->pp);
+	if (err)
+		goto err_xdp_info;
+
 	tmp = size * roundup_pow_of_two(MLX4_EN_MAX_RX_FRAGS *
 					sizeof(struct mlx4_en_rx_alloc));
 	ring->rx_info = kvzalloc_node(tmp, GFP_KERNEL, node);
@@ -319,6 +308,8 @@ err_info:
 	ring->rx_info = NULL;
 err_xdp_info:
 	xdp_rxq_info_unreg(&ring->xdp_rxq);
+err_pp:
+	page_pool_destroy(ring->pp);
 err_ring:
 	kfree(ring);
 	*pring = NULL;
@@ -409,26 +400,6 @@ void mlx4_en_recover_from_oom(struct mlx4_en_priv *priv)
 	}
 }
 
-/* When the rx ring is running in page-per-packet mode, a released frame can go
- * directly into a small cache, to avoid unmapping or touching the page
- * allocator. In bpf prog performance scenarios, buffers are either forwarded
- * or dropped, never converted to skbs, so every page can come directly from
- * this cache when it is sized to be a multiple of the napi budget.
- */
-bool mlx4_en_rx_recycle(struct mlx4_en_rx_ring *ring,
-			struct mlx4_en_rx_alloc *frame)
-{
-	struct mlx4_en_page_cache *cache = &ring->page_cache;
-
-	if (cache->index >= MLX4_EN_CACHE_SIZE)
-		return false;
-
-	cache->buf[cache->index].page = frame->page;
-	cache->buf[cache->index].dma = frame->dma;
-	cache->index++;
-	return true;
-}
-
 void mlx4_en_destroy_rx_ring(struct mlx4_en_priv *priv,
 			     struct mlx4_en_rx_ring **pring,
 			     u32 size, u16 stride)
@@ -445,6 +416,7 @@ void mlx4_en_destroy_rx_ring(struct mlx4_en_priv *priv,
 	xdp_rxq_info_unreg(&ring->xdp_rxq);
 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, size * stride + TXBB_SIZE);
 	kvfree(ring->rx_info);
+	page_pool_destroy(ring->pp);
 	ring->rx_info = NULL;
 	kfree(ring);
 	*pring = NULL;
@@ -453,14 +425,6 @@ void mlx4_en_destroy_rx_ring(struct mlx4_en_priv *priv,
 void mlx4_en_deactivate_rx_ring(struct mlx4_en_priv *priv,
 				struct mlx4_en_rx_ring *ring)
 {
-	int i;
-
-	for (i = 0; i < ring->page_cache.index; i++) {
-		dma_unmap_page(priv->ddev, ring->page_cache.buf[i].dma,
-			       PAGE_SIZE, priv->dma_dir);
-		put_page(ring->page_cache.buf[i].page);
-	}
-	ring->page_cache.index = 0;
 	mlx4_en_free_rx_buf(priv, ring);
 	if (ring->stride <= TXBB_SIZE)
 		ring->buf -= TXBB_SIZE;
@@ -487,7 +451,7 @@ static int mlx4_en_complete_rx_desc(struct mlx4_en_priv *priv,
 		if (unlikely(!page))
 			goto fail;
 
-		dma = frags->dma;
+		dma = page_pool_get_dma_addr(page);
 		dma_sync_single_range_for_cpu(priv->ddev, dma, frags->page_offset,
 					      frag_size, priv->dma_dir);
 
@@ -498,6 +462,7 @@ static int mlx4_en_complete_rx_desc(struct mlx4_en_priv *priv,
 		if (frag_info->frag_stride == PAGE_SIZE / 2) {
 			frags->page_offset ^= PAGE_SIZE / 2;
 			release = page_count(page) != 1 ||
+				  atomic_long_read(&page->pp_ref_count) != 1 ||
 				  page_is_pfmemalloc(page) ||
 				  page_to_nid(page) != numa_mem_id();
 		} else if (!priv->rx_headroom) {
@@ -511,10 +476,9 @@ static int mlx4_en_complete_rx_desc(struct mlx4_en_priv *priv,
 			release = frags->page_offset + frag_info->frag_size > PAGE_SIZE;
 		}
 		if (release) {
-			dma_unmap_page(priv->ddev, dma, PAGE_SIZE, priv->dma_dir);
 			frags->page = NULL;
 		} else {
-			page_ref_inc(page);
+			page_pool_ref_page(page);
 		}
 
 		nr++;
@@ -784,7 +748,8 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud
 			/* Get pointer to first fragment since we haven't
 			 * skb yet and cast it to ethhdr struct
 			 */
-			dma = frags[0].dma + frags[0].page_offset;
+			dma = page_pool_get_dma_addr(frags[0].page);
+			dma += frags[0].page_offset;
 			dma_sync_single_for_cpu(priv->ddev, dma, sizeof(*ethh),
 						DMA_FROM_DEVICE);
 
@@ -823,7 +788,8 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud
 			void *orig_data;
 			u32 act;
 
-			dma = frags[0].dma + frags[0].page_offset;
+			dma = page_pool_get_dma_addr(frags[0].page);
+			dma += frags[0].page_offset;
 			dma_sync_single_for_cpu(priv->ddev, dma,
 						priv->frag_info[0].frag_size,
 						DMA_FROM_DEVICE);
@@ -886,6 +852,7 @@ xdp_drop_no_cnt:
 		skb = napi_get_frags(&cq->napi);
 		if (unlikely(!skb))
 			goto next;
+		skb_mark_for_recycle(skb);
 
 		if (unlikely(ring->hwtstamp_rx_filter == HWTSTAMP_FILTER_ALL)) {
 			u64 timestamp = mlx4_en_get_cqe_ts(cqe);
diff --git a/drivers/net/ethernet/mellanox/mlx4/en_tx.c b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
index 1ddb11cb25f9..87f35bcbeff8 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_tx.c
@@ -44,6 +44,7 @@
 #include <linux/ipv6.h>
 #include <linux/indirect_call_wrapper.h>
 #include <net/ipv6.h>
+#include <net/page_pool/helpers.h>
 
 #include "mlx4_en.h"
 
@@ -350,16 +351,10 @@ u32 mlx4_en_recycle_tx_desc(struct mlx4_en_priv *priv,
 			    int napi_mode)
 {
 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
-	struct mlx4_en_rx_alloc frame = {
-		.page = tx_info->page,
-		.dma = tx_info->map0_dma,
-	};
-
-	if (!napi_mode || !mlx4_en_rx_recycle(ring->recycle_ring, &frame)) {
-		dma_unmap_page(priv->ddev, tx_info->map0_dma,
-			       PAGE_SIZE, priv->dma_dir);
-		put_page(tx_info->page);
-	}
+	struct page_pool *pool = ring->recycle_ring->pp;
+
+	/* Note that napi_mode = 0 means ndo_close() path, not budget = 0 */
+	page_pool_put_full_page(pool, tx_info->page, !!napi_mode);
 
 	return tx_info->nr_txbb;
 }
@@ -450,6 +445,8 @@ int mlx4_en_process_tx_cq(struct net_device *dev,
 
 	if (unlikely(!priv->port_up))
 		return 0;
+	if (unlikely(!napi_budget) && cq->type == TX_XDP)
+		return 0;
 
 	netdev_txq_bql_complete_prefetchw(ring->tx_queue);
 
@@ -1194,7 +1191,7 @@ netdev_tx_t mlx4_en_xmit_frame(struct mlx4_en_rx_ring *rx_ring,
 	tx_desc = ring->buf + (index << LOG_TXBB_SIZE);
 	data = &tx_desc->data;
 
-	dma = frame->dma;
+	dma = page_pool_get_dma_addr(frame->page);
 
 	tx_info->page = frame->page;
 	frame->page = NULL;
diff --git a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
index 28b70dcc652e..ad0d91a75184 100644
--- a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
+++ b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h
@@ -247,20 +247,11 @@ struct mlx4_en_tx_desc {
 
 struct mlx4_en_rx_alloc {
 	struct page	*page;
-	dma_addr_t	dma;
 	u32		page_offset;
 };
 
 #define MLX4_EN_CACHE_SIZE (2 * NAPI_POLL_WEIGHT)
 
-struct mlx4_en_page_cache {
-	u32 index;
-	struct {
-		struct page	*page;
-		dma_addr_t	dma;
-	} buf[MLX4_EN_CACHE_SIZE];
-};
-
 enum {
 	MLX4_EN_TX_RING_STATE_RECOVERING,
 };
@@ -335,14 +326,14 @@ struct mlx4_en_rx_ring {
 	u16 stride;
 	u16 log_stride;
 	u16 cqn;	/* index of port CQ associated with this ring */
+	u8  fcs_del;
 	u32 prod;
 	u32 cons;
 	u32 buf_size;
-	u8  fcs_del;
+	struct page_pool *pp;
 	void *buf;
 	void *rx_info;
 	struct bpf_prog __rcu *xdp_prog;
-	struct mlx4_en_page_cache page_cache;
 	unsigned long bytes;
 	unsigned long packets;
 	unsigned long csum_ok;
@@ -707,8 +698,6 @@ netdev_tx_t mlx4_en_xmit_frame(struct mlx4_en_rx_ring *rx_ring,
 			       struct mlx4_en_priv *priv, unsigned int length,
 			       int tx_ind, bool *doorbell_pending);
 void mlx4_en_xmit_doorbell(struct mlx4_en_tx_ring *ring);
-bool mlx4_en_rx_recycle(struct mlx4_en_rx_ring *ring,
-			struct mlx4_en_rx_alloc *frame);
 
 int mlx4_en_create_tx_ring(struct mlx4_en_priv *priv,
 			   struct mlx4_en_tx_ring **pring,