Merge tag 'mlx5-updates-2023-03-28' of git://git.kernel.org/pub/scm/linux/kernel/git/saeed/linux

Saeed Mahameed says:

====================
mlx5-updates-2023-03-28

Dragos Tatulea says:
====================

net/mlx5e: RX, Drop page_cache and fully use page_pool

For page allocation on the rx path, the mlx5e driver has been using an
internal page cache in tandem with the page pool. The internal page
cache uses a queue for page recycling which has the issue of head of
queue blocking.

This patch series drops the internal page_cache altogether and uses the
page_pool to implement everything that was done by the page_cache
before:
* Let the page_pool handle dma mapping and unmapping.
* Use fragmented pages with fragment counter instead of tracking via
  page ref.
* Enable skb recycling.

The patch series has the following effects on the rx path:

* Improved performance for the cases when there was low page recycling
  due to head of queue blocking in the internal page_cache. The test
  for this was running a single iperf TCP stream to a rx queue
  which is bound on the same cpu as the application.

  |-------------+--------+--------+------+---------|
  | rq type     | before | after  | unit |   diff  |
  |-------------+--------+--------+------+---------|
  | striding rq |  30.1  |  31.4  | Gbps |  4.14 % |
  | legacy rq   |  30.2  |  33.0  | Gbps |  8.48 % |
  |-------------+--------+--------+------+---------|

* Small XDP performance degradation. The test was is XDP drop
  program running on a single rx queue with small packets incoming
  it looks like this:

  |-------------+----------+----------+------+---------|
  | rq type     | before   | after    | unit |   diff  |
  |-------------+----------+----------+------+---------|
  | striding rq | 19725449 | 18544617 | pps  | -6.37 % |
  | legacy rq   | 19879931 | 18631841 | pps  | -6.70 % |
  |-------------+----------+----------+------+---------|

  This will be handled in a different patch series by adding support for
  multi-packet per page.

* For other cases the performance is roughly the same.

The above numbers were obtained on the following system:
  24 core Intel(R) Xeon(R) Platinum 8380 CPU @ 2.30GHz
  32 GB RAM
  ConnectX-7 single port

The breakdown on the patch series is the following:
* Preparations for introducing the mlx5e_frag_page struct.
* Delete the mlx5e_page_cache struct.
* Enable dma mapping from page_pool.
* Enable skb recycling and fragment counting.
* Do deferred release of pages (just before alloc) to ensure better
  page_pool cache utilization.

====================

* tag 'mlx5-updates-2023-03-28' of git://git.kernel.org/pub/scm/linux/kernel/git/saeed/linux:
  net/mlx5e: RX, Remove unnecessary recycle parameter and page_cache stats
  net/mlx5e: RX, Break the wqe bulk refill in smaller chunks
  net/mlx5e: RX, Increase WQE bulk size for legacy rq
  net/mlx5e: RX, Split off release path for xsk buffers for legacy rq
  net/mlx5e: RX, Defer page release in legacy rq for better recycling
  net/mlx5e: RX, Change wqe last_in_page field from bool to bit flags
  net/mlx5e: RX, Defer page release in striding rq for better recycling
  net/mlx5e: RX, Rename xdp_xmit_bitmap to a more generic name
  net/mlx5e: RX, Enable skb page recycling through the page_pool
  net/mlx5e: RX, Enable dma map and sync from page_pool allocator
  net/mlx5e: RX, Remove internal page_cache
  net/mlx5e: RX, Store SHAMPO header pages in array
  net/mlx5e: RX, Remove alloc unit layout constraint for striding rq
  net/mlx5e: RX, Remove alloc unit layout constraint for legacy rq
  net/mlx5e: RX, Remove mlx5e_alloc_unit argument in page allocation
====================

Link: https://lore.kernel.org/r/20230328205623.142075-1-saeed@kernel.orgSigned-off-by: Jakub Kicinski <kuba@kernel.org>

Documentation/networking/device_drivers/ethernet/mellanox/mlx5/counters.rst

@@ -346,32 +346,6 @@ the software port.
      - The number of receive packets with CQE compression on ring i [#accel]_.
      - Acceleration
 
-   * - `rx[i]_cache_reuse`
-     - The number of events of successful reuse of a page from a driver's
-       internal page cache.
-     - Acceleration
-
-   * - `rx[i]_cache_full`
-     - The number of events of full internal page cache where driver can't put a
-       page back to the cache for recycling (page will be freed).
-     - Acceleration
-
-   * - `rx[i]_cache_empty`
-     - The number of events where cache was empty - no page to give. Driver
-       shall allocate new page.
-     - Acceleration
-
-   * - `rx[i]_cache_busy`
-     - The number of events where cache head was busy and cannot be recycled.
-       Driver allocated new page.
-     - Acceleration
-
-   * - `rx[i]_cache_waive`
-     - The number of cache evacuation. This can occur due to page move to
-       another NUMA node or page was pfmemalloc-ed and should be freed as soon
-       as possible.
-     - Acceleration
-
    * - `rx[i]_arfs_err`
      - Number of flow rules that failed to be added to the flow table.
      - Error

drivers/net/ethernet/mellanox/mlx5/core/en.h

@@ -475,11 +475,6 @@ struct mlx5e_txqsq {
 	cqe_ts_to_ns               ptp_cyc2time;
 } ____cacheline_aligned_in_smp;
 
-union mlx5e_alloc_unit {
-	struct page *page;
-	struct xdp_buff *xsk;
-};
-
 /* XDP packets can be transmitted in different ways. On completion, we need to
  * distinguish between them to clean up things in a proper way.
  */
@@ -605,16 +600,35 @@ struct mlx5e_icosq {
 	struct work_struct         recover_work;
 } ____cacheline_aligned_in_smp;
 
+struct mlx5e_frag_page {
+	struct page *page;
+	u16 frags;
+};
+
+enum mlx5e_wqe_frag_flag {
+	MLX5E_WQE_FRAG_LAST_IN_PAGE,
+	MLX5E_WQE_FRAG_SKIP_RELEASE,
+};
+
 struct mlx5e_wqe_frag_info {
-	union mlx5e_alloc_unit *au;
+	union {
+		struct mlx5e_frag_page *frag_page;
+		struct xdp_buff **xskp;
+	};
 	u32 offset;
-	bool last_in_page;
+	u8 flags;
+};
+
+union mlx5e_alloc_units {
+	DECLARE_FLEX_ARRAY(struct mlx5e_frag_page, frag_pages);
+	DECLARE_FLEX_ARRAY(struct page *, pages);
+	DECLARE_FLEX_ARRAY(struct xdp_buff *, xsk_buffs);
 };
 
 struct mlx5e_mpw_info {
 	u16 consumed_strides;
-	DECLARE_BITMAP(xdp_xmit_bitmap, MLX5_MPWRQ_MAX_PAGES_PER_WQE);
-	union mlx5e_alloc_unit alloc_units[];
+	DECLARE_BITMAP(skip_release_bitmap, MLX5_MPWRQ_MAX_PAGES_PER_WQE);
+	union mlx5e_alloc_units alloc_units;
 };
 
 #define MLX5E_MAX_RX_FRAGS 4
@@ -625,11 +639,6 @@ struct mlx5e_mpw_info {
 #define MLX5E_CACHE_UNIT (MLX5_MPWRQ_MAX_PAGES_PER_WQE > NAPI_POLL_WEIGHT ? \
 			  MLX5_MPWRQ_MAX_PAGES_PER_WQE : NAPI_POLL_WEIGHT)
 #define MLX5E_CACHE_SIZE	(4 * roundup_pow_of_two(MLX5E_CACHE_UNIT))
-struct mlx5e_page_cache {
-	u32 head;
-	u32 tail;
-	struct page *page_cache[MLX5E_CACHE_SIZE];
-};
 
 struct mlx5e_rq;
 typedef void (*mlx5e_fp_handle_rx_cqe)(struct mlx5e_rq*, struct mlx5_cqe64*);
@@ -661,19 +670,24 @@ struct mlx5e_rq_frags_info {
 	struct mlx5e_rq_frag_info arr[MLX5E_MAX_RX_FRAGS];
 	u8 num_frags;
 	u8 log_num_frags;
-	u8 wqe_bulk;
+	u16 wqe_bulk;
+	u16 refill_unit;
 	u8 wqe_index_mask;
 };
 
 struct mlx5e_dma_info {
 	dma_addr_t addr;
-	struct page *page;
+	union {
+		struct mlx5e_frag_page *frag_page;
+		struct page *page;
+	};
 };
 
 struct mlx5e_shampo_hd {
 	u32 mkey;
 	struct mlx5e_dma_info *info;
-	struct page *last_page;
+	struct mlx5e_frag_page *pages;
+	u16 curr_page_index;
 	u16 hd_per_wq;
 	u16 hd_per_wqe;
 	unsigned long *bitmap;
@@ -702,7 +716,7 @@ struct mlx5e_rq {
 		struct {
 			struct mlx5_wq_cyc          wq;
 			struct mlx5e_wqe_frag_info *frags;
-			union mlx5e_alloc_unit     *alloc_units;
+			union mlx5e_alloc_units    *alloc_units;
 			struct mlx5e_rq_frags_info  info;
 			mlx5e_fp_skb_from_cqe       skb_from_cqe;
 		} wqe;
@@ -738,7 +752,6 @@ struct mlx5e_rq {
 	struct mlx5e_rq_stats *stats;
 	struct mlx5e_cq        cq;
 	struct mlx5e_cq_decomp cqd;
-	struct mlx5e_page_cache page_cache;
 	struct hwtstamp_config *tstamp;
 	struct mlx5_clock      *clock;
 	struct mlx5e_icosq    *icosq;

drivers/net/ethernet/mellanox/mlx5/core/en/params.c

@@ -667,6 +667,48 @@ static int mlx5e_max_nonlinear_mtu(int first_frag_size, int frag_size, bool xdp)
 	return first_frag_size + (MLX5E_MAX_RX_FRAGS - 2) * frag_size + PAGE_SIZE;
 }
 
+static void mlx5e_rx_compute_wqe_bulk_params(struct mlx5e_params *params,
+					     struct mlx5e_rq_frags_info *info)
+{
+	u16 bulk_bound_rq_size = (1 << params->log_rq_mtu_frames) / 4;
+	u32 bulk_bound_rq_size_in_bytes;
+	u32 sum_frag_strides = 0;
+	u32 wqe_bulk_in_bytes;
+	u16 split_factor;
+	u32 wqe_bulk;
+	int i;
+
+	for (i = 0; i < info->num_frags; i++)
+		sum_frag_strides += info->arr[i].frag_stride;
+
+	/* For MTUs larger than PAGE_SIZE, align to PAGE_SIZE to reflect
+	 * amount of consumed pages per wqe in bytes.
+	 */
+	if (sum_frag_strides > PAGE_SIZE)
+		sum_frag_strides = ALIGN(sum_frag_strides, PAGE_SIZE);
+
+	bulk_bound_rq_size_in_bytes = bulk_bound_rq_size * sum_frag_strides;
+
+#define MAX_WQE_BULK_BYTES(xdp) ((xdp ? 256 : 512) * 1024)
+
+	/* A WQE bulk should not exceed min(512KB, 1/4 of rq size). For XDP
+	 * keep bulk size smaller to avoid filling the page_pool cache on
+	 * every bulk refill.
+	 */
+	wqe_bulk_in_bytes = min_t(u32, MAX_WQE_BULK_BYTES(params->xdp_prog),
+				  bulk_bound_rq_size_in_bytes);
+	wqe_bulk = DIV_ROUND_UP(wqe_bulk_in_bytes, sum_frag_strides);
+
+	/* Make sure that allocations don't start when the page is still used
+	 * by older WQEs.
+	 */
+	info->wqe_bulk = max_t(u16, info->wqe_index_mask + 1, wqe_bulk);
+
+	split_factor = DIV_ROUND_UP(MAX_WQE_BULK_BYTES(params->xdp_prog),
+				    PP_ALLOC_CACHE_REFILL * PAGE_SIZE);
+	info->refill_unit = DIV_ROUND_UP(info->wqe_bulk, split_factor);
+}
+
 #define DEFAULT_FRAG_SIZE (2048)
 
 static int mlx5e_build_rq_frags_info(struct mlx5_core_dev *mdev,
@@ -774,11 +816,14 @@ static int mlx5e_build_rq_frags_info(struct mlx5_core_dev *mdev,
 	}
 
 out:
-	/* Bulking optimization to skip allocation until at least 8 WQEs can be
-	 * allocated in a row. At the same time, never start allocation when
-	 * the page is still used by older WQEs.
+	/* Bulking optimization to skip allocation until a large enough number
+	 * of WQEs can be allocated in a row. Bulking also influences how well
+	 * deferred page release works.
 	 */
-	info->wqe_bulk = max_t(u8, info->wqe_index_mask + 1, 8);
+	mlx5e_rx_compute_wqe_bulk_params(params, info);
+
+	mlx5_core_dbg(mdev, "%s: wqe_bulk = %u, wqe_bulk_refill_unit = %u\n",
+		      __func__, info->wqe_bulk, info->refill_unit);
 
 	info->log_num_frags = order_base_2(info->num_frags);
 

drivers/net/ethernet/mellanox/mlx5/core/en/reporter_rx.c

@@ -121,9 +121,9 @@ static int mlx5e_rx_reporter_err_icosq_cqe_recover(void *ctx)
 
 	mlx5e_reset_icosq_cc_pc(icosq);
 
-	mlx5e_free_rx_in_progress_descs(rq);
+	mlx5e_free_rx_missing_descs(rq);
 	if (xskrq)
-		mlx5e_free_rx_in_progress_descs(xskrq);
+		mlx5e_free_rx_missing_descs(xskrq);
 
 	clear_bit(MLX5E_SQ_STATE_RECOVERING, &icosq->state);
 	mlx5e_activate_icosq(icosq);

drivers/net/ethernet/mellanox/mlx5/core/en/txrx.h

@@ -65,13 +65,11 @@ int mlx5e_napi_poll(struct napi_struct *napi, int budget);
 int mlx5e_poll_ico_cq(struct mlx5e_cq *cq);
 
 /* RX */
-void mlx5e_page_dma_unmap(struct mlx5e_rq *rq, struct page *page);
-void mlx5e_page_release_dynamic(struct mlx5e_rq *rq, struct page *page, bool recycle);
 INDIRECT_CALLABLE_DECLARE(bool mlx5e_post_rx_wqes(struct mlx5e_rq *rq));
 INDIRECT_CALLABLE_DECLARE(bool mlx5e_post_rx_mpwqes(struct mlx5e_rq *rq));
 int mlx5e_poll_rx_cq(struct mlx5e_cq *cq, int budget);
 void mlx5e_free_rx_descs(struct mlx5e_rq *rq);
-void mlx5e_free_rx_in_progress_descs(struct mlx5e_rq *rq);
+void mlx5e_free_rx_missing_descs(struct mlx5e_rq *rq);
 
 static inline bool mlx5e_rx_hw_stamp(struct hwtstamp_config *config)
 {
@@ -489,7 +487,7 @@ static inline bool mlx5e_icosq_can_post_wqe(struct mlx5e_icosq *sq, u16 wqe_size
 
 static inline struct mlx5e_mpw_info *mlx5e_get_mpw_info(struct mlx5e_rq *rq, int i)
 {
-	size_t isz = struct_size(rq->mpwqe.info, alloc_units, rq->mpwqe.pages_per_wqe);
+	size_t isz = struct_size(rq->mpwqe.info, alloc_units.frag_pages, rq->mpwqe.pages_per_wqe);
 
 	return (struct mlx5e_mpw_info *)((char *)rq->mpwqe.info + array_size(i, isz));
 }

drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c

@@ -209,8 +209,6 @@ bool mlx5e_xdp_handle(struct mlx5e_rq *rq,
 			goto xdp_abort;
 		__set_bit(MLX5E_RQ_FLAG_XDP_XMIT, rq->flags);
 		__set_bit(MLX5E_RQ_FLAG_XDP_REDIRECT, rq->flags);
-		if (xdp->rxq->mem.type != MEM_TYPE_XSK_BUFF_POOL)
-			mlx5e_page_dma_unmap(rq, virt_to_page(xdp->data));
 		rq->stats->xdp_redirect++;
 		return true;
 	default:
@@ -507,7 +505,6 @@ mlx5e_xmit_xdp_frame(struct mlx5e_xdpsq *sq, struct mlx5e_xmit_data *xdptxd,
 static void mlx5e_free_xdpsq_desc(struct mlx5e_xdpsq *sq,
 				  struct mlx5e_xdp_wqe_info *wi,
 				  u32 *xsk_frames,
-				  bool recycle,
 				  struct xdp_frame_bulk *bq)
 {
 	struct mlx5e_xdp_info_fifo *xdpi_fifo = &sq->db.xdpi_fifo;
@@ -525,7 +522,8 @@ static void mlx5e_free_xdpsq_desc(struct mlx5e_xdpsq *sq,
 			break;
 		case MLX5E_XDP_XMIT_MODE_PAGE:
 			/* XDP_TX from the regular RQ */
-			mlx5e_page_release_dynamic(xdpi.page.rq, xdpi.page.page, recycle);
+			page_pool_put_defragged_page(xdpi.page.rq->page_pool,
+						     xdpi.page.page, -1, true);
 			break;
 		case MLX5E_XDP_XMIT_MODE_XSK:
 			/* AF_XDP send */
@@ -579,7 +577,7 @@ bool mlx5e_poll_xdpsq_cq(struct mlx5e_cq *cq)
 
 			sqcc += wi->num_wqebbs;
 
-			mlx5e_free_xdpsq_desc(sq, wi, &xsk_frames, true, &bq);
+			mlx5e_free_xdpsq_desc(sq, wi, &xsk_frames, &bq);
 		} while (!last_wqe);
 
 		if (unlikely(get_cqe_opcode(cqe) != MLX5_CQE_REQ)) {
@@ -626,7 +624,7 @@ void mlx5e_free_xdpsq_descs(struct mlx5e_xdpsq *sq)
 
 		sq->cc += wi->num_wqebbs;
 
-		mlx5e_free_xdpsq_desc(sq, wi, &xsk_frames, false, &bq);
+		mlx5e_free_xdpsq_desc(sq, wi, &xsk_frames, &bq);
 	}
 
 	xdp_flush_frame_bulk(&bq);

drivers/net/ethernet/mellanox/mlx5/core/en/xsk/rx.c

@@ -22,6 +22,7 @@ int mlx5e_xsk_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 	struct mlx5e_icosq *icosq = rq->icosq;
 	struct mlx5_wq_cyc *wq = &icosq->wq;
 	struct mlx5e_umr_wqe *umr_wqe;
+	struct xdp_buff **xsk_buffs;
 	int batch, i;
 	u32 offset; /* 17-bit value with MTT. */
 	u16 pi;
@@ -29,9 +30,9 @@ int mlx5e_xsk_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 	if (unlikely(!xsk_buff_can_alloc(rq->xsk_pool, rq->mpwqe.pages_per_wqe)))
 		goto err;
 
-	BUILD_BUG_ON(sizeof(wi->alloc_units[0]) != sizeof(wi->alloc_units[0].xsk));
 	XSK_CHECK_PRIV_TYPE(struct mlx5e_xdp_buff);
-	batch = xsk_buff_alloc_batch(rq->xsk_pool, (struct xdp_buff **)wi->alloc_units,
+	xsk_buffs = (struct xdp_buff **)wi->alloc_units.xsk_buffs;
+	batch = xsk_buff_alloc_batch(rq->xsk_pool, xsk_buffs,
 				     rq->mpwqe.pages_per_wqe);
 
 	/* If batch < pages_per_wqe, either:
@@ -41,8 +42,8 @@ int mlx5e_xsk_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 	 * the first error, which will mean there are no more valid descriptors.
 	 */
 	for (; batch < rq->mpwqe.pages_per_wqe; batch++) {
-		wi->alloc_units[batch].xsk = xsk_buff_alloc(rq->xsk_pool);
-		if (unlikely(!wi->alloc_units[batch].xsk))
+		xsk_buffs[batch] = xsk_buff_alloc(rq->xsk_pool);
+		if (unlikely(!xsk_buffs[batch]))
 			goto err_reuse_batch;
 	}
 
@@ -52,8 +53,8 @@ int mlx5e_xsk_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 
 	if (likely(rq->mpwqe.umr_mode == MLX5E_MPWRQ_UMR_MODE_ALIGNED)) {
 		for (i = 0; i < batch; i++) {
-			struct mlx5e_xdp_buff *mxbuf = xsk_buff_to_mxbuf(wi->alloc_units[i].xsk);
-			dma_addr_t addr = xsk_buff_xdp_get_frame_dma(wi->alloc_units[i].xsk);
+			struct mlx5e_xdp_buff *mxbuf = xsk_buff_to_mxbuf(xsk_buffs[i]);
+			dma_addr_t addr = xsk_buff_xdp_get_frame_dma(xsk_buffs[i]);
 
 			umr_wqe->inline_mtts[i] = (struct mlx5_mtt) {
 				.ptag = cpu_to_be64(addr | MLX5_EN_WR),
@@ -62,8 +63,8 @@ int mlx5e_xsk_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 		}
 	} else if (unlikely(rq->mpwqe.umr_mode == MLX5E_MPWRQ_UMR_MODE_UNALIGNED)) {
 		for (i = 0; i < batch; i++) {
-			struct mlx5e_xdp_buff *mxbuf = xsk_buff_to_mxbuf(wi->alloc_units[i].xsk);
-			dma_addr_t addr = xsk_buff_xdp_get_frame_dma(wi->alloc_units[i].xsk);
+			struct mlx5e_xdp_buff *mxbuf = xsk_buff_to_mxbuf(xsk_buffs[i]);
+			dma_addr_t addr = xsk_buff_xdp_get_frame_dma(xsk_buffs[i]);
 
 			umr_wqe->inline_ksms[i] = (struct mlx5_ksm) {
 				.key = rq->mkey_be,
@@ -75,8 +76,8 @@ int mlx5e_xsk_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 		u32 mapping_size = 1 << (rq->mpwqe.page_shift - 2);
 
 		for (i = 0; i < batch; i++) {
-			struct mlx5e_xdp_buff *mxbuf = xsk_buff_to_mxbuf(wi->alloc_units[i].xsk);
-			dma_addr_t addr = xsk_buff_xdp_get_frame_dma(wi->alloc_units[i].xsk);
+			struct mlx5e_xdp_buff *mxbuf = xsk_buff_to_mxbuf(xsk_buffs[i]);
+			dma_addr_t addr = xsk_buff_xdp_get_frame_dma(xsk_buffs[i]);
 
 			umr_wqe->inline_ksms[i << 2] = (struct mlx5_ksm) {
 				.key = rq->mkey_be,
@@ -102,8 +103,8 @@ int mlx5e_xsk_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 		__be32 frame_size = cpu_to_be32(rq->xsk_pool->chunk_size);
 
 		for (i = 0; i < batch; i++) {
-			struct mlx5e_xdp_buff *mxbuf = xsk_buff_to_mxbuf(wi->alloc_units[i].xsk);
-			dma_addr_t addr = xsk_buff_xdp_get_frame_dma(wi->alloc_units[i].xsk);
+			struct mlx5e_xdp_buff *mxbuf = xsk_buff_to_mxbuf(xsk_buffs[i]);
+			dma_addr_t addr = xsk_buff_xdp_get_frame_dma(xsk_buffs[i]);
 
 			umr_wqe->inline_klms[i << 1] = (struct mlx5_klm) {
 				.key = rq->mkey_be,
@@ -119,7 +120,7 @@ int mlx5e_xsk_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 		}
 	}
 
-	bitmap_zero(wi->xdp_xmit_bitmap, rq->mpwqe.pages_per_wqe);
+	bitmap_zero(wi->skip_release_bitmap, rq->mpwqe.pages_per_wqe);
 	wi->consumed_strides = 0;
 
 	umr_wqe->ctrl.opmod_idx_opcode =
@@ -149,7 +150,7 @@ int mlx5e_xsk_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 
 err_reuse_batch:
 	while (--batch >= 0)
-		xsk_buff_free(wi->alloc_units[batch].xsk);
+		xsk_buff_free(xsk_buffs[batch]);
 
 err:
 	rq->stats->buff_alloc_err++;
@@ -163,13 +164,10 @@ int mlx5e_xsk_alloc_rx_wqes_batched(struct mlx5e_rq *rq, u16 ix, int wqe_bulk)
 	u32 contig, alloc;
 	int i;
 
-	/* mlx5e_init_frags_partition creates a 1:1 mapping between
-	 * rq->wqe.frags and rq->wqe.alloc_units, which allows us to
-	 * allocate XDP buffers straight into alloc_units.
+	/* Each rq->wqe.frags->xskp is 1:1 mapped to an element inside the
+	 * rq->wqe.alloc_units->xsk_buffs array allocated here.
 	 */
-	BUILD_BUG_ON(sizeof(rq->wqe.alloc_units[0]) !=
-		     sizeof(rq->wqe.alloc_units[0].xsk));
-	buffs = (struct xdp_buff **)rq->wqe.alloc_units;
+	buffs = rq->wqe.alloc_units->xsk_buffs;
 	contig = mlx5_wq_cyc_get_size(wq) - ix;
 	if (wqe_bulk <= contig) {
 		alloc = xsk_buff_alloc_batch(rq->xsk_pool, buffs + ix, wqe_bulk);
@@ -189,8 +187,9 @@ int mlx5e_xsk_alloc_rx_wqes_batched(struct mlx5e_rq *rq, u16 ix, int wqe_bulk)
 		/* Assumes log_num_frags == 0. */
 		frag = &rq->wqe.frags[j];
 
-		addr = xsk_buff_xdp_get_frame_dma(frag->au->xsk);
+		addr = xsk_buff_xdp_get_frame_dma(*frag->xskp);
 		wqe->data[0].addr = cpu_to_be64(addr + rq->buff.headroom);
+		frag->flags &= ~BIT(MLX5E_WQE_FRAG_SKIP_RELEASE);
 	}
 
 	return alloc;
@@ -211,12 +210,13 @@ int mlx5e_xsk_alloc_rx_wqes(struct mlx5e_rq *rq, u16 ix, int wqe_bulk)
 		/* Assumes log_num_frags == 0. */
 		frag = &rq->wqe.frags[j];
 
-		frag->au->xsk = xsk_buff_alloc(rq->xsk_pool);
-		if (unlikely(!frag->au->xsk))
+		*frag->xskp = xsk_buff_alloc(rq->xsk_pool);
+		if (unlikely(!*frag->xskp))
 			return i;
 
-		addr = xsk_buff_xdp_get_frame_dma(frag->au->xsk);
+		addr = xsk_buff_xdp_get_frame_dma(*frag->xskp);
 		wqe->data[0].addr = cpu_to_be64(addr + rq->buff.headroom);
+		frag->flags &= ~BIT(MLX5E_WQE_FRAG_SKIP_RELEASE);
 	}
 
 	return wqe_bulk;
@@ -251,7 +251,7 @@ struct sk_buff *mlx5e_xsk_skb_from_cqe_mpwrq_linear(struct mlx5e_rq *rq,
 						    u32 head_offset,
 						    u32 page_idx)
 {
-	struct mlx5e_xdp_buff *mxbuf = xsk_buff_to_mxbuf(wi->alloc_units[page_idx].xsk);
+	struct mlx5e_xdp_buff *mxbuf = xsk_buff_to_mxbuf(wi->alloc_units.xsk_buffs[page_idx]);
 	struct bpf_prog *prog;
 
 	/* Check packet size. Note LRO doesn't use linear SKB */
@@ -291,7 +291,7 @@ struct sk_buff *mlx5e_xsk_skb_from_cqe_mpwrq_linear(struct mlx5e_rq *rq,
 	prog = rcu_dereference(rq->xdp_prog);
 	if (likely(prog && mlx5e_xdp_handle(rq, prog, mxbuf))) {
 		if (likely(__test_and_clear_bit(MLX5E_RQ_FLAG_XDP_XMIT, rq->flags)))
-			__set_bit(page_idx, wi->xdp_xmit_bitmap); /* non-atomic */
+			__set_bit(page_idx, wi->skip_release_bitmap); /* non-atomic */
 		return NULL; /* page/packet was consumed by XDP */
 	}
 
@@ -306,7 +306,7 @@ struct sk_buff *mlx5e_xsk_skb_from_cqe_linear(struct mlx5e_rq *rq,
 					      struct mlx5_cqe64 *cqe,
 					      u32 cqe_bcnt)
 {
-	struct mlx5e_xdp_buff *mxbuf = xsk_buff_to_mxbuf(wi->au->xsk);
+	struct mlx5e_xdp_buff *mxbuf = xsk_buff_to_mxbuf(*wi->xskp);
 	struct bpf_prog *prog;
 
 	/* wi->offset is not used in this function, because xdp->data and the

drivers/net/ethernet/mellanox/mlx5/core/en_main.c

@@ -262,23 +262,30 @@ static int mlx5e_rq_shampo_hd_info_alloc(struct mlx5e_rq *rq, int node)
 
 	shampo->bitmap = bitmap_zalloc_node(shampo->hd_per_wq, GFP_KERNEL,
 					    node);
-	if (!shampo->bitmap)
-		return -ENOMEM;
-
 	shampo->info = kvzalloc_node(array_size(shampo->hd_per_wq,
 						sizeof(*shampo->info)),
 				     GFP_KERNEL, node);
-	if (!shampo->info) {
-		kvfree(shampo->bitmap);
-		return -ENOMEM;
-	}
+	shampo->pages = kvzalloc_node(array_size(shampo->hd_per_wq,
+						 sizeof(*shampo->pages)),
+				     GFP_KERNEL, node);
+	if (!shampo->bitmap || !shampo->info || !shampo->pages)
+		goto err_nomem;
+
 	return 0;
+
+err_nomem:
+	kvfree(shampo->info);
+	kvfree(shampo->bitmap);
+	kvfree(shampo->pages);
+
+	return -ENOMEM;
 }
 
 static void mlx5e_rq_shampo_hd_info_free(struct mlx5e_rq *rq)
 {
 	kvfree(rq->mpwqe.shampo->bitmap);
 	kvfree(rq->mpwqe.shampo->info);
+	kvfree(rq->mpwqe.shampo->pages);
 }
 
 static int mlx5e_rq_alloc_mpwqe_info(struct mlx5e_rq *rq, int node)
@@ -286,13 +293,23 @@ static int mlx5e_rq_alloc_mpwqe_info(struct mlx5e_rq *rq, int node)
 	int wq_sz = mlx5_wq_ll_get_size(&rq->mpwqe.wq);
 	size_t alloc_size;
 
-	alloc_size = array_size(wq_sz, struct_size(rq->mpwqe.info, alloc_units,
+	alloc_size = array_size(wq_sz, struct_size(rq->mpwqe.info,
+						   alloc_units.frag_pages,
 						   rq->mpwqe.pages_per_wqe));
 
 	rq->mpwqe.info = kvzalloc_node(alloc_size, GFP_KERNEL, node);
 	if (!rq->mpwqe.info)
 		return -ENOMEM;
 
+	/* For deferred page release (release right before alloc), make sure
+	 * that on first round release is not called.
+	 */
+	for (int i = 0; i < wq_sz; i++) {
+		struct mlx5e_mpw_info *wi = mlx5e_get_mpw_info(rq, i);
+
+		bitmap_fill(wi->skip_release_bitmap, rq->mpwqe.pages_per_wqe);
+	}
+
 	mlx5e_build_umr_wqe(rq, rq->icosq, &rq->mpwqe.umr_wqe);
 
 	return 0;
@@ -499,14 +516,12 @@ static void mlx5e_init_frags_partition(struct mlx5e_rq *rq)
 	struct mlx5e_wqe_frag_info *prev = NULL;
 	int i;
 
-	if (rq->xsk_pool) {
-		/* Assumptions used by XSK batched allocator. */
-		WARN_ON(rq->wqe.info.num_frags != 1);
-		WARN_ON(rq->wqe.info.log_num_frags != 0);
-		WARN_ON(rq->wqe.info.arr[0].frag_stride != PAGE_SIZE);
-	}
+	WARN_ON(rq->xsk_pool);
+
+	next_frag.frag_page = &rq->wqe.alloc_units->frag_pages[0];
 
-	next_frag.au = &rq->wqe.alloc_units[0];
+	/* Skip first release due to deferred release. */
+	next_frag.flags = BIT(MLX5E_WQE_FRAG_SKIP_RELEASE);
 
 	for (i = 0; i < mlx5_wq_cyc_get_size(&rq->wqe.wq); i++) {
 		struct mlx5e_rq_frag_info *frag_info = &rq->wqe.info.arr[0];
@@ -516,10 +531,11 @@ static void mlx5e_init_frags_partition(struct mlx5e_rq *rq)
 
 		for (f = 0; f < rq->wqe.info.num_frags; f++, frag++) {
 			if (next_frag.offset + frag_info[f].frag_stride > PAGE_SIZE) {
-				next_frag.au++;
+				/* Pages are assigned at runtime. */
+				next_frag.frag_page++;
 				next_frag.offset = 0;
 				if (prev)
-					prev->last_in_page = true;
+					prev->flags |= BIT(MLX5E_WQE_FRAG_LAST_IN_PAGE);
 			}
 			*frag = next_frag;
 
@@ -530,25 +546,68 @@ static void mlx5e_init_frags_partition(struct mlx5e_rq *rq)
 	}
 
 	if (prev)
-		prev->last_in_page = true;
+		prev->flags |= BIT(MLX5E_WQE_FRAG_LAST_IN_PAGE);
 }
 
-static int mlx5e_init_au_list(struct mlx5e_rq *rq, int wq_sz, int node)
+static void mlx5e_init_xsk_buffs(struct mlx5e_rq *rq)
 {
+	int i;
+
+	/* Assumptions used by XSK batched allocator. */
+	WARN_ON(rq->wqe.info.num_frags != 1);
+	WARN_ON(rq->wqe.info.log_num_frags != 0);
+	WARN_ON(rq->wqe.info.arr[0].frag_stride != PAGE_SIZE);
+
+	/* Considering the above assumptions a fragment maps to a single
+	 * xsk_buff.
+	 */
+	for (i = 0; i < mlx5_wq_cyc_get_size(&rq->wqe.wq); i++) {
+		rq->wqe.frags[i].xskp = &rq->wqe.alloc_units->xsk_buffs[i];
+
+		/* Skip first release due to deferred release as WQES are
+		 * not allocated yet.
+		 */
+		rq->wqe.frags[i].flags |= BIT(MLX5E_WQE_FRAG_SKIP_RELEASE);
+	}
+}
+
+static int mlx5e_init_wqe_alloc_info(struct mlx5e_rq *rq, int node)
+{
+	int wq_sz = mlx5_wq_cyc_get_size(&rq->wqe.wq);
 	int len = wq_sz << rq->wqe.info.log_num_frags;
+	struct mlx5e_wqe_frag_info *frags;
+	union mlx5e_alloc_units *aus;
+	int aus_sz;
+
+	if (rq->xsk_pool)
+		aus_sz = sizeof(*aus->xsk_buffs);
+	else
+		aus_sz = sizeof(*aus->frag_pages);
+
+	aus = kvzalloc_node(array_size(len, aus_sz), GFP_KERNEL, node);
+	if (!aus)
+		return -ENOMEM;
 
-	rq->wqe.alloc_units = kvzalloc_node(array_size(len, sizeof(*rq->wqe.alloc_units)),
-					    GFP_KERNEL, node);
-	if (!rq->wqe.alloc_units)
+	frags = kvzalloc_node(array_size(len, sizeof(*frags)), GFP_KERNEL, node);
+	if (!frags) {
+		kvfree(aus);
 		return -ENOMEM;
+	}
+
+	rq->wqe.alloc_units = aus;
+	rq->wqe.frags = frags;
 
-	mlx5e_init_frags_partition(rq);
+	if (rq->xsk_pool)
+		mlx5e_init_xsk_buffs(rq);
+	else
+		mlx5e_init_frags_partition(rq);
 
 	return 0;
 }
 
-static void mlx5e_free_au_list(struct mlx5e_rq *rq)
+static void mlx5e_free_wqe_alloc_info(struct mlx5e_rq *rq)
 {
+	kvfree(rq->wqe.frags);
 	kvfree(rq->wqe.alloc_units);
 }
 
@@ -693,7 +752,6 @@ static int mlx5e_alloc_rq(struct mlx5e_params *params,
 			  struct mlx5e_rq_param *rqp,
 			  int node, struct mlx5e_rq *rq)
 {
-	struct page_pool_params pp_params = { 0 };
 	struct mlx5_core_dev *mdev = rq->mdev;
 	void *rqc = rqp->rqc;
 	void *rqc_wq = MLX5_ADDR_OF(rqc, rqc, wq);
@@ -778,18 +836,9 @@ static int mlx5e_alloc_rq(struct mlx5e_params *params,
 		rq->wqe.info = rqp->frags_info;
 		rq->buff.frame0_sz = rq->wqe.info.arr[0].frag_stride;
 
-		rq->wqe.frags =
-			kvzalloc_node(array_size(sizeof(*rq->wqe.frags),
-					(wq_sz << rq->wqe.info.log_num_frags)),
-				      GFP_KERNEL, node);
-		if (!rq->wqe.frags) {
-			err = -ENOMEM;
-			goto err_rq_wq_destroy;
-		}
-
-		err = mlx5e_init_au_list(rq, wq_sz, node);
+		err = mlx5e_init_wqe_alloc_info(rq, node);
 		if (err)
-			goto err_rq_frags;
+			goto err_rq_wq_destroy;
 	}
 
 	if (xsk) {
@@ -798,12 +847,15 @@ static int mlx5e_alloc_rq(struct mlx5e_params *params,
 		xsk_pool_set_rxq_info(rq->xsk_pool, &rq->xdp_rxq);
 	} else {
 		/* Create a page_pool and register it with rxq */
+		struct page_pool_params pp_params = { 0 };
+
 		pp_params.order     = 0;
-		pp_params.flags     = 0; /* No-internal DMA mapping in page_pool */
+		pp_params.flags     = PP_FLAG_DMA_MAP | PP_FLAG_DMA_SYNC_DEV | PP_FLAG_PAGE_FRAG;
 		pp_params.pool_size = pool_size;
 		pp_params.nid       = node;
 		pp_params.dev       = rq->pdev;
 		pp_params.dma_dir   = rq->buff.map_dir;
+		pp_params.max_len   = PAGE_SIZE;
 
 		/* page_pool can be used even when there is no rq->xdp_prog,
 		 * given page_pool does not handle DMA mapping there is no
@@ -869,9 +921,6 @@ static int mlx5e_alloc_rq(struct mlx5e_params *params,
 		rq->dim.mode = DIM_CQ_PERIOD_MODE_START_FROM_EQE;
 	}
 
-	rq->page_cache.head = 0;
-	rq->page_cache.tail = 0;
-
 	return 0;
 
 err_destroy_page_pool:
@@ -888,9 +937,7 @@ static int mlx5e_alloc_rq(struct mlx5e_params *params,
 		mlx5e_free_mpwqe_rq_drop_page(rq);
 		break;
 	default: /* MLX5_WQ_TYPE_CYCLIC */
-		mlx5e_free_au_list(rq);
-err_rq_frags:
-		kvfree(rq->wqe.frags);
+		mlx5e_free_wqe_alloc_info(rq);
 	}
 err_rq_wq_destroy:
 	mlx5_wq_destroy(&rq->wq_ctrl);
@@ -904,7 +951,6 @@ static int mlx5e_alloc_rq(struct mlx5e_params *params,
 static void mlx5e_free_rq(struct mlx5e_rq *rq)
 {
 	struct bpf_prog *old_prog;
-	int i;
 
 	if (xdp_rxq_info_is_reg(&rq->xdp_rxq)) {
 		old_prog = rcu_dereference_protected(rq->xdp_prog,
@@ -921,17 +967,7 @@ static void mlx5e_free_rq(struct mlx5e_rq *rq)
 		mlx5e_rq_free_shampo(rq);
 		break;
 	default: /* MLX5_WQ_TYPE_CYCLIC */
-		kvfree(rq->wqe.frags);
-		mlx5e_free_au_list(rq);
-	}
-
-	for (i = rq->page_cache.head; i != rq->page_cache.tail;
-	     i = (i + 1) & (MLX5E_CACHE_SIZE - 1)) {
-		/* With AF_XDP, page_cache is not used, so this loop is not
-		 * entered, and it's safe to call mlx5e_page_release_dynamic
-		 * directly.
-		 */
-		mlx5e_page_release_dynamic(rq, rq->page_cache.page_cache[i], false);
+		mlx5e_free_wqe_alloc_info(rq);
 	}
 
 	xdp_rxq_info_unreg(&rq->xdp_rxq);
@@ -1094,7 +1130,7 @@ int mlx5e_wait_for_min_rx_wqes(struct mlx5e_rq *rq, int wait_time)
 	return -ETIMEDOUT;
 }
 
-void mlx5e_free_rx_in_progress_descs(struct mlx5e_rq *rq)
+void mlx5e_free_rx_missing_descs(struct mlx5e_rq *rq)
 {
 	struct mlx5_wq_ll *wq;
 	u16 head;
@@ -1106,8 +1142,12 @@ void mlx5e_free_rx_in_progress_descs(struct mlx5e_rq *rq)
 	wq = &rq->mpwqe.wq;
 	head = wq->head;
 
-	/* Outstanding UMR WQEs (in progress) start at wq->head */
-	for (i = 0; i < rq->mpwqe.umr_in_progress; i++) {
+	/* Release WQEs that are in missing state: they have been
+	 * popped from the list after completion but were not freed
+	 * due to deferred release.
+	 * Also free the linked-list reserved entry, hence the "+ 1".
+	 */
+	for (i = 0; i < mlx5_wq_ll_missing(wq) + 1; i++) {
 		rq->dealloc_wqe(rq, head);
 		head = mlx5_wq_ll_get_wqe_next_ix(wq, head);
 	}
@@ -1134,7 +1174,7 @@ void mlx5e_free_rx_descs(struct mlx5e_rq *rq)
 	if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ) {
 		struct mlx5_wq_ll *wq = &rq->mpwqe.wq;
 
-		mlx5e_free_rx_in_progress_descs(rq);
+		mlx5e_free_rx_missing_descs(rq);
 
 		while (!mlx5_wq_ll_is_empty(wq)) {
 			struct mlx5e_rx_wqe_ll *wqe;
@@ -1152,12 +1192,21 @@ void mlx5e_free_rx_descs(struct mlx5e_rq *rq)
 						0, true);
 	} else {
 		struct mlx5_wq_cyc *wq = &rq->wqe.wq;
+		u16 missing = mlx5_wq_cyc_missing(wq);
+		u16 head = mlx5_wq_cyc_get_head(wq);
 
 		while (!mlx5_wq_cyc_is_empty(wq)) {
 			wqe_ix = mlx5_wq_cyc_get_tail(wq);
 			rq->dealloc_wqe(rq, wqe_ix);
 			mlx5_wq_cyc_pop(wq);
 		}
+		/* Missing slots might also contain unreleased pages due to
+		 * deferred release.
+		 */
+		while (missing--) {
+			wqe_ix = mlx5_wq_cyc_ctr2ix(wq, head++);
+			rq->dealloc_wqe(rq, wqe_ix);
+		}
 	}
 
 }

drivers/net/ethernet/mellanox/mlx5/core/en_rx.c

@@ -271,98 +271,35 @@ static inline u32 mlx5e_decompress_cqes_start(struct mlx5e_rq *rq,
 	return mlx5e_decompress_cqes_cont(rq, wq, 1, budget_rem);
 }
 
-static inline bool mlx5e_rx_cache_put(struct mlx5e_rq *rq, struct page *page)
-{
-	struct mlx5e_page_cache *cache = &rq->page_cache;
-	u32 tail_next = (cache->tail + 1) & (MLX5E_CACHE_SIZE - 1);
-	struct mlx5e_rq_stats *stats = rq->stats;
-
-	if (tail_next == cache->head) {
-		stats->cache_full++;
-		return false;
-	}
-
-	if (!dev_page_is_reusable(page)) {
-		stats->cache_waive++;
-		return false;
-	}
+#define MLX5E_PAGECNT_BIAS_MAX (PAGE_SIZE / 64)
 
-	cache->page_cache[cache->tail] = page;
-	cache->tail = tail_next;
-	return true;
-}
-
-static inline bool mlx5e_rx_cache_get(struct mlx5e_rq *rq, union mlx5e_alloc_unit *au)
+static int mlx5e_page_alloc_fragmented(struct mlx5e_rq *rq,
+				       struct mlx5e_frag_page *frag_page)
 {
-	struct mlx5e_page_cache *cache = &rq->page_cache;
-	struct mlx5e_rq_stats *stats = rq->stats;
-	dma_addr_t addr;
-
-	if (unlikely(cache->head == cache->tail)) {
-		stats->cache_empty++;
-		return false;
-	}
+	struct page *page;
 
-	if (page_ref_count(cache->page_cache[cache->head]) != 1) {
-		stats->cache_busy++;
-		return false;
-	}
-
-	au->page = cache->page_cache[cache->head];
-	cache->head = (cache->head + 1) & (MLX5E_CACHE_SIZE - 1);
-	stats->cache_reuse++;
-
-	addr = page_pool_get_dma_addr(au->page);
-	/* Non-XSK always uses PAGE_SIZE. */
-	dma_sync_single_for_device(rq->pdev, addr, PAGE_SIZE, rq->buff.map_dir);
-	return true;
-}
-
-static inline int mlx5e_page_alloc_pool(struct mlx5e_rq *rq, union mlx5e_alloc_unit *au)
-{
-	dma_addr_t addr;
-
-	if (mlx5e_rx_cache_get(rq, au))
-		return 0;
-
-	au->page = page_pool_dev_alloc_pages(rq->page_pool);
-	if (unlikely(!au->page))
+	page = page_pool_dev_alloc_pages(rq->page_pool);
+	if (unlikely(!page))
 		return -ENOMEM;
 
-	/* Non-XSK always uses PAGE_SIZE. */
-	addr = dma_map_page(rq->pdev, au->page, 0, PAGE_SIZE, rq->buff.map_dir);
-	if (unlikely(dma_mapping_error(rq->pdev, addr))) {
-		page_pool_recycle_direct(rq->page_pool, au->page);
-		au->page = NULL;
-		return -ENOMEM;
-	}
-	page_pool_set_dma_addr(au->page, addr);
-
-	return 0;
-}
+	page_pool_fragment_page(page, MLX5E_PAGECNT_BIAS_MAX);
 
-void mlx5e_page_dma_unmap(struct mlx5e_rq *rq, struct page *page)
-{
-	dma_addr_t dma_addr = page_pool_get_dma_addr(page);
+	*frag_page = (struct mlx5e_frag_page) {
+		.page	= page,
+		.frags	= 0,
+	};
 
-	dma_unmap_page_attrs(rq->pdev, dma_addr, PAGE_SIZE, rq->buff.map_dir,
-			     DMA_ATTR_SKIP_CPU_SYNC);
-	page_pool_set_dma_addr(page, 0);
+	return 0;
 }
 
-void mlx5e_page_release_dynamic(struct mlx5e_rq *rq, struct page *page, bool recycle)
+static void mlx5e_page_release_fragmented(struct mlx5e_rq *rq,
+					  struct mlx5e_frag_page *frag_page)
 {
-	if (likely(recycle)) {
-		if (mlx5e_rx_cache_put(rq, page))
-			return;
+	u16 drain_count = MLX5E_PAGECNT_BIAS_MAX - frag_page->frags;
+	struct page *page = frag_page->page;
 
-		mlx5e_page_dma_unmap(rq, page);
-		page_pool_recycle_direct(rq->page_pool, page);
-	} else {
-		mlx5e_page_dma_unmap(rq, page);
-		page_pool_release_page(rq->page_pool, page);
-		put_page(page);
-	}
+	if (page_pool_defrag_page(page, drain_count) == 0)
+		page_pool_put_defragged_page(rq->page_pool, page, -1, true);
 }
 
 static inline int mlx5e_get_rx_frag(struct mlx5e_rq *rq,
@@ -371,22 +308,31 @@ static inline int mlx5e_get_rx_frag(struct mlx5e_rq *rq,
 	int err = 0;
 
 	if (!frag->offset)
-		/* On first frag (offset == 0), replenish page (alloc_unit actually).
-		 * Other frags that point to the same alloc_unit (with a different
+		/* On first frag (offset == 0), replenish page.
+		 * Other frags that point to the same page (with a different
 		 * offset) should just use the new one without replenishing again
 		 * by themselves.
 		 */
-		err = mlx5e_page_alloc_pool(rq, frag->au);
+		err = mlx5e_page_alloc_fragmented(rq, frag->frag_page);
 
 	return err;
 }
 
+static bool mlx5e_frag_can_release(struct mlx5e_wqe_frag_info *frag)
+{
+#define CAN_RELEASE_MASK \
+	(BIT(MLX5E_WQE_FRAG_LAST_IN_PAGE) | BIT(MLX5E_WQE_FRAG_SKIP_RELEASE))
+
+#define CAN_RELEASE_VALUE BIT(MLX5E_WQE_FRAG_LAST_IN_PAGE)
+
+	return (frag->flags & CAN_RELEASE_MASK) == CAN_RELEASE_VALUE;
+}
+
 static inline void mlx5e_put_rx_frag(struct mlx5e_rq *rq,
-				     struct mlx5e_wqe_frag_info *frag,
-				     bool recycle)
+				     struct mlx5e_wqe_frag_info *frag)
 {
-	if (frag->last_in_page)
-		mlx5e_page_release_dynamic(rq, frag->au->page, recycle);
+	if (mlx5e_frag_can_release(frag))
+		mlx5e_page_release_fragmented(rq, frag->frag_page);
 }
 
 static inline struct mlx5e_wqe_frag_info *get_frag(struct mlx5e_rq *rq, u16 ix)
@@ -409,8 +355,10 @@ static int mlx5e_alloc_rx_wqe(struct mlx5e_rq *rq, struct mlx5e_rx_wqe_cyc *wqe,
 		if (unlikely(err))
 			goto free_frags;
 
+		frag->flags &= ~BIT(MLX5E_WQE_FRAG_SKIP_RELEASE);
+
 		headroom = i == 0 ? rq->buff.headroom : 0;
-		addr = page_pool_get_dma_addr(frag->au->page);
+		addr = page_pool_get_dma_addr(frag->frag_page->page);
 		wqe->data[i].addr = cpu_to_be64(addr + frag->offset + headroom);
 	}
 
@@ -418,35 +366,66 @@ static int mlx5e_alloc_rx_wqe(struct mlx5e_rq *rq, struct mlx5e_rx_wqe_cyc *wqe,
 
 free_frags:
 	while (--i >= 0)
-		mlx5e_put_rx_frag(rq, --frag, true);
+		mlx5e_put_rx_frag(rq, --frag);
 
 	return err;
 }
 
 static inline void mlx5e_free_rx_wqe(struct mlx5e_rq *rq,
-				     struct mlx5e_wqe_frag_info *wi,
-				     bool recycle)
+				     struct mlx5e_wqe_frag_info *wi)
 {
 	int i;
 
-	if (rq->xsk_pool) {
-		/* The `recycle` parameter is ignored, and the page is always
-		 * put into the Reuse Ring, because there is no way to return
-		 * the page to the userspace when the interface goes down.
-		 */
-		xsk_buff_free(wi->au->xsk);
-		return;
-	}
-
 	for (i = 0; i < rq->wqe.info.num_frags; i++, wi++)
-		mlx5e_put_rx_frag(rq, wi, recycle);
+		mlx5e_put_rx_frag(rq, wi);
+}
+
+static void mlx5e_xsk_free_rx_wqe(struct mlx5e_wqe_frag_info *wi)
+{
+	if (!(wi->flags & BIT(MLX5E_WQE_FRAG_SKIP_RELEASE)))
+		xsk_buff_free(*wi->xskp);
 }
 
 static void mlx5e_dealloc_rx_wqe(struct mlx5e_rq *rq, u16 ix)
 {
 	struct mlx5e_wqe_frag_info *wi = get_frag(rq, ix);
 
-	mlx5e_free_rx_wqe(rq, wi, false);
+	if (rq->xsk_pool)
+		mlx5e_xsk_free_rx_wqe(wi);
+	else
+		mlx5e_free_rx_wqe(rq, wi);
+}
+
+static void mlx5e_xsk_free_rx_wqes(struct mlx5e_rq *rq, u16 ix, int wqe_bulk)
+{
+	struct mlx5_wq_cyc *wq = &rq->wqe.wq;
+	int i;
+
+	for (i = 0; i < wqe_bulk; i++) {
+		int j = mlx5_wq_cyc_ctr2ix(wq, ix + i);
+		struct mlx5e_wqe_frag_info *wi;
+
+		wi = get_frag(rq, j);
+		/* The page is always put into the Reuse Ring, because there
+		 * is no way to return the page to the userspace when the
+		 * interface goes down.
+		 */
+		mlx5e_xsk_free_rx_wqe(wi);
+	}
+}
+
+static void mlx5e_free_rx_wqes(struct mlx5e_rq *rq, u16 ix, int wqe_bulk)
+{
+	struct mlx5_wq_cyc *wq = &rq->wqe.wq;
+	int i;
+
+	for (i = 0; i < wqe_bulk; i++) {
+		int j = mlx5_wq_cyc_ctr2ix(wq, ix + i);
+		struct mlx5e_wqe_frag_info *wi;
+
+		wi = get_frag(rq, j);
+		mlx5e_free_rx_wqe(rq, wi);
+	}
 }
 
 static int mlx5e_alloc_rx_wqes(struct mlx5e_rq *rq, u16 ix, int wqe_bulk)
@@ -467,18 +446,42 @@ static int mlx5e_alloc_rx_wqes(struct mlx5e_rq *rq, u16 ix, int wqe_bulk)
 	return i;
 }
 
+static int mlx5e_refill_rx_wqes(struct mlx5e_rq *rq, u16 ix, int wqe_bulk)
+{
+	int remaining = wqe_bulk;
+	int i = 0;
+
+	/* The WQE bulk is split into smaller bulks that are sized
+	 * according to the page pool cache refill size to avoid overflowing
+	 * the page pool cache due to too many page releases at once.
+	 */
+	do {
+		int refill = min_t(u16, rq->wqe.info.refill_unit, remaining);
+		int alloc_count;
+
+		mlx5e_free_rx_wqes(rq, ix + i, refill);
+		alloc_count = mlx5e_alloc_rx_wqes(rq, ix + i, refill);
+		i += alloc_count;
+		if (unlikely(alloc_count != refill))
+			break;
+
+		remaining -= refill;
+	} while (remaining);
+
+	return i;
+}
+
 static inline void
 mlx5e_add_skb_frag(struct mlx5e_rq *rq, struct sk_buff *skb,
-		   union mlx5e_alloc_unit *au, u32 frag_offset, u32 len,
+		   struct page *page, u32 frag_offset, u32 len,
 		   unsigned int truesize)
 {
-	dma_addr_t addr = page_pool_get_dma_addr(au->page);
+	dma_addr_t addr = page_pool_get_dma_addr(page);
 
 	dma_sync_single_for_cpu(rq->pdev, addr + frag_offset, len,
 				rq->buff.map_dir);
-	page_ref_inc(au->page);
 	skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
-			au->page, frag_offset, len, truesize);
+			page, frag_offset, len, truesize);
 }
 
 static inline void
@@ -496,30 +499,36 @@ mlx5e_copy_skb_header(struct mlx5e_rq *rq, struct sk_buff *skb,
 }
 
 static void
-mlx5e_free_rx_mpwqe(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi, bool recycle)
+mlx5e_free_rx_mpwqe(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi)
 {
-	union mlx5e_alloc_unit *alloc_units = wi->alloc_units;
 	bool no_xdp_xmit;
 	int i;
 
 	/* A common case for AF_XDP. */
-	if (bitmap_full(wi->xdp_xmit_bitmap, rq->mpwqe.pages_per_wqe))
+	if (bitmap_full(wi->skip_release_bitmap, rq->mpwqe.pages_per_wqe))
 		return;
 
-	no_xdp_xmit = bitmap_empty(wi->xdp_xmit_bitmap, rq->mpwqe.pages_per_wqe);
+	no_xdp_xmit = bitmap_empty(wi->skip_release_bitmap, rq->mpwqe.pages_per_wqe);
 
 	if (rq->xsk_pool) {
-		/* The `recycle` parameter is ignored, and the page is always
-		 * put into the Reuse Ring, because there is no way to return
-		 * the page to the userspace when the interface goes down.
+		struct xdp_buff **xsk_buffs = wi->alloc_units.xsk_buffs;
+
+		/* The page is always put into the Reuse Ring, because there
+		 * is no way to return the page to userspace when the interface
+		 * goes down.
 		 */
 		for (i = 0; i < rq->mpwqe.pages_per_wqe; i++)
-			if (no_xdp_xmit || !test_bit(i, wi->xdp_xmit_bitmap))
-				xsk_buff_free(alloc_units[i].xsk);
+			if (no_xdp_xmit || !test_bit(i, wi->skip_release_bitmap))
+				xsk_buff_free(xsk_buffs[i]);
 	} else {
-		for (i = 0; i < rq->mpwqe.pages_per_wqe; i++)
-			if (no_xdp_xmit || !test_bit(i, wi->xdp_xmit_bitmap))
-				mlx5e_page_release_dynamic(rq, alloc_units[i].page, recycle);
+		for (i = 0; i < rq->mpwqe.pages_per_wqe; i++) {
+			if (no_xdp_xmit || !test_bit(i, wi->skip_release_bitmap)) {
+				struct mlx5e_frag_page *frag_page;
+
+				frag_page = &wi->alloc_units.frag_pages[i];
+				mlx5e_page_release_fragmented(rq, frag_page);
+			}
+		}
 	}
 }
 
@@ -583,7 +592,8 @@ static int mlx5e_build_shampo_hd_umr(struct mlx5e_rq *rq,
 	struct mlx5e_shampo_hd *shampo = rq->mpwqe.shampo;
 	u16 entries, pi, header_offset, err, wqe_bbs, new_entries;
 	u32 lkey = rq->mdev->mlx5e_res.hw_objs.mkey;
-	struct page *page = shampo->last_page;
+	u16 page_index = shampo->curr_page_index;
+	struct mlx5e_frag_page *frag_page;
 	u64 addr = shampo->last_addr;
 	struct mlx5e_dma_info *dma_info;
 	struct mlx5e_umr_wqe *umr_wqe;
@@ -597,6 +607,8 @@ static int mlx5e_build_shampo_hd_umr(struct mlx5e_rq *rq,
 	umr_wqe = mlx5_wq_cyc_get_wqe(&sq->wq, pi);
 	build_klm_umr(sq, umr_wqe, shampo->key, index, entries, wqe_bbs);
 
+	frag_page = &shampo->pages[page_index];
+
 	for (i = 0; i < entries; i++, index++) {
 		dma_info = &shampo->info[index];
 		if (i >= klm_entries || (index < shampo->pi && shampo->pi - index <
@@ -605,16 +617,20 @@ static int mlx5e_build_shampo_hd_umr(struct mlx5e_rq *rq,
 		header_offset = (index & (MLX5E_SHAMPO_WQ_HEADER_PER_PAGE - 1)) <<
 			MLX5E_SHAMPO_LOG_MAX_HEADER_ENTRY_SIZE;
 		if (!(header_offset & (PAGE_SIZE - 1))) {
-			union mlx5e_alloc_unit au;
+			page_index = (page_index + 1) & (shampo->hd_per_wq - 1);
+			frag_page = &shampo->pages[page_index];
 
-			err = mlx5e_page_alloc_pool(rq, &au);
+			err = mlx5e_page_alloc_fragmented(rq, frag_page);
 			if (unlikely(err))
 				goto err_unmap;
-			page = dma_info->page = au.page;
-			addr = dma_info->addr = page_pool_get_dma_addr(au.page);
+
+			addr = page_pool_get_dma_addr(frag_page->page);
+
+			dma_info->addr = addr;
+			dma_info->frag_page = frag_page;
 		} else {
 			dma_info->addr = addr + header_offset;
-			dma_info->page = page;
+			dma_info->frag_page = frag_page;
 		}
 
 update_klm:
@@ -632,7 +648,7 @@ static int mlx5e_build_shampo_hd_umr(struct mlx5e_rq *rq,
 	};
 
 	shampo->pi = (shampo->pi + new_entries) & (shampo->hd_per_wq - 1);
-	shampo->last_page = page;
+	shampo->curr_page_index = page_index;
 	shampo->last_addr = addr;
 	sq->pc += wqe_bbs;
 	sq->doorbell_cseg = &umr_wqe->ctrl;
@@ -644,7 +660,7 @@ static int mlx5e_build_shampo_hd_umr(struct mlx5e_rq *rq,
 		dma_info = &shampo->info[--index];
 		if (!(i & (MLX5E_SHAMPO_WQ_HEADER_PER_PAGE - 1))) {
 			dma_info->addr = ALIGN_DOWN(dma_info->addr, PAGE_SIZE);
-			mlx5e_page_release_dynamic(rq, dma_info->page, true);
+			mlx5e_page_release_fragmented(rq, dma_info->frag_page);
 		}
 	}
 	rq->stats->buff_alloc_err++;
@@ -693,8 +709,8 @@ static int mlx5e_alloc_rx_hd_mpwqe(struct mlx5e_rq *rq)
 static int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 {
 	struct mlx5e_mpw_info *wi = mlx5e_get_mpw_info(rq, ix);
-	union mlx5e_alloc_unit *au = &wi->alloc_units[0];
 	struct mlx5e_icosq *sq = rq->icosq;
+	struct mlx5e_frag_page *frag_page;
 	struct mlx5_wq_cyc *wq = &sq->wq;
 	struct mlx5e_umr_wqe *umr_wqe;
 	u32 offset; /* 17-bit value with MTT. */
@@ -712,13 +728,15 @@ static int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 	umr_wqe = mlx5_wq_cyc_get_wqe(wq, pi);
 	memcpy(umr_wqe, &rq->mpwqe.umr_wqe, sizeof(struct mlx5e_umr_wqe));
 
-	for (i = 0; i < rq->mpwqe.pages_per_wqe; i++, au++) {
+	frag_page = &wi->alloc_units.frag_pages[0];
+
+	for (i = 0; i < rq->mpwqe.pages_per_wqe; i++, frag_page++) {
 		dma_addr_t addr;
 
-		err = mlx5e_page_alloc_pool(rq, au);
+		err = mlx5e_page_alloc_fragmented(rq, frag_page);
 		if (unlikely(err))
 			goto err_unmap;
-		addr = page_pool_get_dma_addr(au->page);
+		addr = page_pool_get_dma_addr(frag_page->page);
 		umr_wqe->inline_mtts[i] = (struct mlx5_mtt) {
 			.ptag = cpu_to_be64(addr | MLX5_EN_WR),
 		};
@@ -735,7 +753,7 @@ static int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 		       sizeof(*umr_wqe->inline_mtts) * pad);
 	}
 
-	bitmap_zero(wi->xdp_xmit_bitmap, rq->mpwqe.pages_per_wqe);
+	bitmap_zero(wi->skip_release_bitmap, rq->mpwqe.pages_per_wqe);
 	wi->consumed_strides = 0;
 
 	umr_wqe->ctrl.opmod_idx_opcode =
@@ -759,8 +777,8 @@ static int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 
 err_unmap:
 	while (--i >= 0) {
-		au--;
-		mlx5e_page_release_dynamic(rq, au->page, true);
+		frag_page--;
+		mlx5e_page_release_fragmented(rq, frag_page);
 	}
 
 err:
@@ -778,8 +796,8 @@ static int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 void mlx5e_shampo_dealloc_hd(struct mlx5e_rq *rq, u16 len, u16 start, bool close)
 {
 	struct mlx5e_shampo_hd *shampo = rq->mpwqe.shampo;
+	struct mlx5e_frag_page *deleted_page = NULL;
 	int hd_per_wq = shampo->hd_per_wq;
-	struct page *deleted_page = NULL;
 	struct mlx5e_dma_info *hd_info;
 	int i, index = start;
 
@@ -792,10 +810,12 @@ void mlx5e_shampo_dealloc_hd(struct mlx5e_rq *rq, u16 len, u16 start, bool close
 
 		hd_info = &shampo->info[index];
 		hd_info->addr = ALIGN_DOWN(hd_info->addr, PAGE_SIZE);
-		if (hd_info->page != deleted_page) {
-			deleted_page = hd_info->page;
-			mlx5e_page_release_dynamic(rq, hd_info->page, false);
+		if (hd_info->frag_page && hd_info->frag_page != deleted_page) {
+			deleted_page = hd_info->frag_page;
+			mlx5e_page_release_fragmented(rq, hd_info->frag_page);
 		}
+
+		hd_info->frag_page = NULL;
 	}
 
 	if (start + len > hd_per_wq) {
@@ -810,8 +830,8 @@ void mlx5e_shampo_dealloc_hd(struct mlx5e_rq *rq, u16 len, u16 start, bool close
 static void mlx5e_dealloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 {
 	struct mlx5e_mpw_info *wi = mlx5e_get_mpw_info(rq, ix);
-	/* Don't recycle, this function is called on rq/netdev close */
-	mlx5e_free_rx_mpwqe(rq, wi, false);
+	/* This function is called on rq/netdev close. */
+	mlx5e_free_rx_mpwqe(rq, wi);
 }
 
 INDIRECT_CALLABLE_SCOPE bool mlx5e_post_rx_wqes(struct mlx5e_rq *rq)
@@ -838,17 +858,20 @@ INDIRECT_CALLABLE_SCOPE bool mlx5e_post_rx_wqes(struct mlx5e_rq *rq)
 	 */
 	wqe_bulk -= (head + wqe_bulk) & rq->wqe.info.wqe_index_mask;
 
-	if (!rq->xsk_pool)
-		count = mlx5e_alloc_rx_wqes(rq, head, wqe_bulk);
-	else if (likely(!rq->xsk_pool->dma_need_sync))
+	if (!rq->xsk_pool) {
+		count = mlx5e_refill_rx_wqes(rq, head, wqe_bulk);
+	} else if (likely(!rq->xsk_pool->dma_need_sync)) {
+		mlx5e_xsk_free_rx_wqes(rq, head, wqe_bulk);
 		count = mlx5e_xsk_alloc_rx_wqes_batched(rq, head, wqe_bulk);
-	else
+	} else {
+		mlx5e_xsk_free_rx_wqes(rq, head, wqe_bulk);
 		/* If dma_need_sync is true, it's more efficient to call
 		 * xsk_buff_alloc in a loop, rather than xsk_buff_alloc_batch,
 		 * because the latter does the same check and returns only one
 		 * frame.
 		 */
 		count = mlx5e_xsk_alloc_rx_wqes(rq, head, wqe_bulk);
+	}
 
 	mlx5_wq_cyc_push_n(wq, count);
 	if (unlikely(count != wqe_bulk)) {
@@ -1029,6 +1052,11 @@ INDIRECT_CALLABLE_SCOPE bool mlx5e_post_rx_mpwqes(struct mlx5e_rq *rq)
 	head = rq->mpwqe.actual_wq_head;
 	i = missing;
 	do {
+		struct mlx5e_mpw_info *wi = mlx5e_get_mpw_info(rq, head);
+
+		/* Deferred free for better page pool cache usage. */
+		mlx5e_free_rx_mpwqe(rq, wi);
+
 		alloc_err = rq->xsk_pool ? mlx5e_xsk_alloc_rx_mpwqe(rq, head) :
 					   mlx5e_alloc_rx_mpwqe(rq, head);
 
@@ -1133,7 +1161,7 @@ static void *mlx5e_shampo_get_packet_hd(struct mlx5e_rq *rq, u16 header_index)
 	struct mlx5e_dma_info *last_head = &rq->mpwqe.shampo->info[header_index];
 	u16 head_offset = (last_head->addr & (PAGE_SIZE - 1)) + rq->buff.headroom;
 
-	return page_address(last_head->page) + head_offset;
+	return page_address(last_head->frag_page->page) + head_offset;
 }
 
 static void mlx5e_shampo_update_ipv4_udp_hdr(struct mlx5e_rq *rq, struct iphdr *ipv4)
@@ -1586,7 +1614,7 @@ static struct sk_buff *
 mlx5e_skb_from_cqe_linear(struct mlx5e_rq *rq, struct mlx5e_wqe_frag_info *wi,
 			  struct mlx5_cqe64 *cqe, u32 cqe_bcnt)
 {
-	union mlx5e_alloc_unit *au = wi->au;
+	struct mlx5e_frag_page *frag_page = wi->frag_page;
 	u16 rx_headroom = rq->buff.headroom;
 	struct bpf_prog *prog;
 	struct sk_buff *skb;
@@ -1595,11 +1623,11 @@ mlx5e_skb_from_cqe_linear(struct mlx5e_rq *rq, struct mlx5e_wqe_frag_info *wi,
 	dma_addr_t addr;
 	u32 frag_size;
 
-	va             = page_address(au->page) + wi->offset;
+	va             = page_address(frag_page->page) + wi->offset;
 	data           = va + rx_headroom;
 	frag_size      = MLX5_SKB_FRAG_SZ(rx_headroom + cqe_bcnt);
 
-	addr = page_pool_get_dma_addr(au->page);
+	addr = page_pool_get_dma_addr(frag_page->page);
 	dma_sync_single_range_for_cpu(rq->pdev, addr, wi->offset,
 				      frag_size, rq->buff.map_dir);
 	net_prefetch(data);
@@ -1623,7 +1651,8 @@ mlx5e_skb_from_cqe_linear(struct mlx5e_rq *rq, struct mlx5e_wqe_frag_info *wi,
 		return NULL;
 
 	/* queue up for recycling/reuse */
-	page_ref_inc(au->page);
+	skb_mark_for_recycle(skb);
+	frag_page->frags++;
 
 	return skb;
 }
@@ -1634,8 +1663,8 @@ mlx5e_skb_from_cqe_nonlinear(struct mlx5e_rq *rq, struct mlx5e_wqe_frag_info *wi
 {
 	struct mlx5e_rq_frag_info *frag_info = &rq->wqe.info.arr[0];
 	struct mlx5e_wqe_frag_info *head_wi = wi;
-	union mlx5e_alloc_unit *au = wi->au;
 	u16 rx_headroom = rq->buff.headroom;
+	struct mlx5e_frag_page *frag_page;
 	struct skb_shared_info *sinfo;
 	struct mlx5e_xdp_buff mxbuf;
 	u32 frag_consumed_bytes;
@@ -1645,10 +1674,12 @@ mlx5e_skb_from_cqe_nonlinear(struct mlx5e_rq *rq, struct mlx5e_wqe_frag_info *wi
 	u32 truesize;
 	void *va;
 
-	va = page_address(au->page) + wi->offset;
+	frag_page = wi->frag_page;
+
+	va = page_address(frag_page->page) + wi->offset;
 	frag_consumed_bytes = min_t(u32, frag_info->frag_size, cqe_bcnt);
 
-	addr = page_pool_get_dma_addr(au->page);
+	addr = page_pool_get_dma_addr(frag_page->page);
 	dma_sync_single_range_for_cpu(rq->pdev, addr, wi->offset,
 				      rq->buff.frame0_sz, rq->buff.map_dir);
 	net_prefetchw(va); /* xdp_frame data area */
@@ -1665,11 +1696,11 @@ mlx5e_skb_from_cqe_nonlinear(struct mlx5e_rq *rq, struct mlx5e_wqe_frag_info *wi
 	while (cqe_bcnt) {
 		skb_frag_t *frag;
 
-		au = wi->au;
+		frag_page = wi->frag_page;
 
 		frag_consumed_bytes = min_t(u32, frag_info->frag_size, cqe_bcnt);
 
-		addr = page_pool_get_dma_addr(au->page);
+		addr = page_pool_get_dma_addr(frag_page->page);
 		dma_sync_single_for_cpu(rq->pdev, addr + wi->offset,
 					frag_consumed_bytes, rq->buff.map_dir);
 
@@ -1683,11 +1714,12 @@ mlx5e_skb_from_cqe_nonlinear(struct mlx5e_rq *rq, struct mlx5e_wqe_frag_info *wi
 		}
 
 		frag = &sinfo->frags[sinfo->nr_frags++];
-		__skb_frag_set_page(frag, au->page);
+
+		__skb_frag_set_page(frag, frag_page->page);
 		skb_frag_off_set(frag, wi->offset);
 		skb_frag_size_set(frag, frag_consumed_bytes);
 
-		if (page_is_pfmemalloc(au->page))
+		if (page_is_pfmemalloc(frag_page->page))
 			xdp_buff_set_frag_pfmemalloc(&mxbuf.xdp);
 
 		sinfo->xdp_frags_size += frag_consumed_bytes;
@@ -1704,7 +1736,7 @@ mlx5e_skb_from_cqe_nonlinear(struct mlx5e_rq *rq, struct mlx5e_wqe_frag_info *wi
 			int i;
 
 			for (i = wi - head_wi; i < rq->wqe.info.num_frags; i++)
-				mlx5e_put_rx_frag(rq, &head_wi[i], true);
+				mlx5e_put_rx_frag(rq, &head_wi[i]);
 		}
 		return NULL; /* page/packet was consumed by XDP */
 	}
@@ -1716,21 +1748,17 @@ mlx5e_skb_from_cqe_nonlinear(struct mlx5e_rq *rq, struct mlx5e_wqe_frag_info *wi
 	if (unlikely(!skb))
 		return NULL;
 
-	page_ref_inc(head_wi->au->page);
+	skb_mark_for_recycle(skb);
+	head_wi->frag_page->frags++;
 
 	if (xdp_buff_has_frags(&mxbuf.xdp)) {
-		int i;
-
 		/* sinfo->nr_frags is reset by build_skb, calculate again. */
 		xdp_update_skb_shared_info(skb, wi - head_wi - 1,
 					   sinfo->xdp_frags_size, truesize,
 					   xdp_buff_is_frag_pfmemalloc(&mxbuf.xdp));
 
-		for (i = 0; i < sinfo->nr_frags; i++) {
-			skb_frag_t *frag = &sinfo->frags[i];
-
-			page_ref_inc(skb_frag_page(frag));
-		}
+		for (struct mlx5e_wqe_frag_info *pwi = head_wi + 1; pwi < wi; pwi++)
+			pwi->frag_page->frags++;
 	}
 
 	return skb;
@@ -1768,7 +1796,7 @@ static void mlx5e_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
 
 	if (unlikely(MLX5E_RX_ERR_CQE(cqe))) {
 		mlx5e_handle_rx_err_cqe(rq, cqe);
-		goto free_wqe;
+		goto wq_cyc_pop;
 	}
 
 	skb = INDIRECT_CALL_3(rq->wqe.skb_from_cqe,
@@ -1782,9 +1810,9 @@ static void mlx5e_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
 			/* do not return page to cache,
 			 * it will be returned on XDP_TX completion.
 			 */
-			goto wq_cyc_pop;
+			wi->flags |= BIT(MLX5E_WQE_FRAG_SKIP_RELEASE);
 		}
-		goto free_wqe;
+		goto wq_cyc_pop;
 	}
 
 	mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
@@ -1792,13 +1820,11 @@ static void mlx5e_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
 	if (mlx5e_cqe_regb_chain(cqe))
 		if (!mlx5e_tc_update_skb_nic(cqe, skb)) {
 			dev_kfree_skb_any(skb);
-			goto free_wqe;
+			goto wq_cyc_pop;
 		}
 
 	napi_gro_receive(rq->cq.napi, skb);
 
-free_wqe:
-	mlx5e_free_rx_wqe(rq, wi, true);
 wq_cyc_pop:
 	mlx5_wq_cyc_pop(wq);
 }
@@ -1822,7 +1848,7 @@ static void mlx5e_handle_rx_cqe_rep(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
 
 	if (unlikely(MLX5E_RX_ERR_CQE(cqe))) {
 		mlx5e_handle_rx_err_cqe(rq, cqe);
-		goto free_wqe;
+		goto wq_cyc_pop;
 	}
 
 	skb = INDIRECT_CALL_2(rq->wqe.skb_from_cqe,
@@ -1835,9 +1861,9 @@ static void mlx5e_handle_rx_cqe_rep(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
 			/* do not return page to cache,
 			 * it will be returned on XDP_TX completion.
 			 */
-			goto wq_cyc_pop;
+			wi->flags |= BIT(MLX5E_WQE_FRAG_SKIP_RELEASE);
 		}
-		goto free_wqe;
+		goto wq_cyc_pop;
 	}
 
 	mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
@@ -1847,8 +1873,6 @@ static void mlx5e_handle_rx_cqe_rep(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
 
 	mlx5e_rep_tc_receive(cqe, rq, skb);
 
-free_wqe:
-	mlx5e_free_rx_wqe(rq, wi, true);
 wq_cyc_pop:
 	mlx5_wq_cyc_pop(wq);
 }
@@ -1901,7 +1925,6 @@ static void mlx5e_handle_rx_cqe_mpwrq_rep(struct mlx5e_rq *rq, struct mlx5_cqe64
 
 	wq  = &rq->mpwqe.wq;
 	wqe = mlx5_wq_ll_get_wqe(wq, wqe_id);
-	mlx5e_free_rx_mpwqe(rq, wi, true);
 	mlx5_wq_ll_pop(wq, cqe->wqe_id, &wqe->next.next_wqe_index);
 }
 
@@ -1913,7 +1936,8 @@ const struct mlx5e_rx_handlers mlx5e_rx_handlers_rep = {
 
 static void
 mlx5e_fill_skb_data(struct sk_buff *skb, struct mlx5e_rq *rq,
-		    union mlx5e_alloc_unit *au, u32 data_bcnt, u32 data_offset)
+		    struct mlx5e_frag_page *frag_page,
+		    u32 data_bcnt, u32 data_offset)
 {
 	net_prefetchw(skb->data);
 
@@ -1927,12 +1951,13 @@ mlx5e_fill_skb_data(struct sk_buff *skb, struct mlx5e_rq *rq,
 		else
 			truesize = ALIGN(pg_consumed_bytes, BIT(rq->mpwqe.log_stride_sz));
 
-		mlx5e_add_skb_frag(rq, skb, au, data_offset,
+		frag_page->frags++;
+		mlx5e_add_skb_frag(rq, skb, frag_page->page, data_offset,
 				   pg_consumed_bytes, truesize);
 
 		data_bcnt -= pg_consumed_bytes;
 		data_offset = 0;
-		au++;
+		frag_page++;
 	}
 }
 
@@ -1941,11 +1966,11 @@ mlx5e_skb_from_cqe_mpwrq_nonlinear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *w
 				   struct mlx5_cqe64 *cqe, u16 cqe_bcnt, u32 head_offset,
 				   u32 page_idx)
 {
-	union mlx5e_alloc_unit *au = &wi->alloc_units[page_idx];
+	struct mlx5e_frag_page *frag_page = &wi->alloc_units.frag_pages[page_idx];
 	u16 headlen = min_t(u16, MLX5E_RX_MAX_HEAD, cqe_bcnt);
+	struct mlx5e_frag_page *head_page = frag_page;
 	u32 frag_offset    = head_offset + headlen;
 	u32 byte_cnt       = cqe_bcnt - headlen;
-	union mlx5e_alloc_unit *head_au = au;
 	struct sk_buff *skb;
 	dma_addr_t addr;
 
@@ -1960,14 +1985,15 @@ mlx5e_skb_from_cqe_mpwrq_nonlinear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *w
 
 	/* Non-linear mode, hence non-XSK, which always uses PAGE_SIZE. */
 	if (unlikely(frag_offset >= PAGE_SIZE)) {
-		au++;
+		frag_page++;
 		frag_offset -= PAGE_SIZE;
 	}
 
-	mlx5e_fill_skb_data(skb, rq, au, byte_cnt, frag_offset);
+	skb_mark_for_recycle(skb);
+	mlx5e_fill_skb_data(skb, rq, frag_page, byte_cnt, frag_offset);
 	/* copy header */
-	addr = page_pool_get_dma_addr(head_au->page);
-	mlx5e_copy_skb_header(rq, skb, head_au->page, addr,
+	addr = page_pool_get_dma_addr(head_page->page);
+	mlx5e_copy_skb_header(rq, skb, head_page->page, addr,
 			      head_offset, head_offset, headlen);
 	/* skb linear part was allocated with headlen and aligned to long */
 	skb->tail += headlen;
@@ -1981,7 +2007,7 @@ mlx5e_skb_from_cqe_mpwrq_linear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi,
 				struct mlx5_cqe64 *cqe, u16 cqe_bcnt, u32 head_offset,
 				u32 page_idx)
 {
-	union mlx5e_alloc_unit *au = &wi->alloc_units[page_idx];
+	struct mlx5e_frag_page *frag_page = &wi->alloc_units.frag_pages[page_idx];
 	u16 rx_headroom = rq->buff.headroom;
 	struct bpf_prog *prog;
 	struct sk_buff *skb;
@@ -1996,11 +2022,11 @@ mlx5e_skb_from_cqe_mpwrq_linear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi,
 		return NULL;
 	}
 
-	va             = page_address(au->page) + head_offset;
+	va             = page_address(frag_page->page) + head_offset;
 	data           = va + rx_headroom;
 	frag_size      = MLX5_SKB_FRAG_SZ(rx_headroom + cqe_bcnt);
 
-	addr = page_pool_get_dma_addr(au->page);
+	addr = page_pool_get_dma_addr(frag_page->page);
 	dma_sync_single_range_for_cpu(rq->pdev, addr, head_offset,
 				      frag_size, rq->buff.map_dir);
 	net_prefetch(data);
@@ -2013,7 +2039,7 @@ mlx5e_skb_from_cqe_mpwrq_linear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi,
 		mlx5e_fill_mxbuf(rq, cqe, va, rx_headroom, cqe_bcnt, &mxbuf);
 		if (mlx5e_xdp_handle(rq, prog, &mxbuf)) {
 			if (__test_and_clear_bit(MLX5E_RQ_FLAG_XDP_XMIT, rq->flags))
-				__set_bit(page_idx, wi->xdp_xmit_bitmap); /* non-atomic */
+				__set_bit(page_idx, wi->skip_release_bitmap); /* non-atomic */
 			return NULL; /* page/packet was consumed by XDP */
 		}
 
@@ -2027,7 +2053,8 @@ mlx5e_skb_from_cqe_mpwrq_linear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi,
 		return NULL;
 
 	/* queue up for recycling/reuse */
-	page_ref_inc(au->page);
+	skb_mark_for_recycle(skb);
+	frag_page->frags++;
 
 	return skb;
 }
@@ -2044,7 +2071,7 @@ mlx5e_skb_from_cqe_shampo(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi,
 	void *hdr, *data;
 	u32 frag_size;
 
-	hdr		= page_address(head->page) + head_offset;
+	hdr		= page_address(head->frag_page->page) + head_offset;
 	data		= hdr + rx_headroom;
 	frag_size	= MLX5_SKB_FRAG_SZ(rx_headroom + head_size);
 
@@ -2058,9 +2085,7 @@ mlx5e_skb_from_cqe_shampo(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi,
 		if (unlikely(!skb))
 			return NULL;
 
-		/* queue up for recycling/reuse */
-		page_ref_inc(head->page);
-
+		head->frag_page->frags++;
 	} else {
 		/* allocate SKB and copy header for large header */
 		rq->stats->gro_large_hds++;
@@ -2072,13 +2097,17 @@ mlx5e_skb_from_cqe_shampo(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi,
 		}
 
 		prefetchw(skb->data);
-		mlx5e_copy_skb_header(rq, skb, head->page, head->addr,
+		mlx5e_copy_skb_header(rq, skb, head->frag_page->page, head->addr,
 				      head_offset + rx_headroom,
 				      rx_headroom, head_size);
 		/* skb linear part was allocated with headlen and aligned to long */
 		skb->tail += head_size;
 		skb->len  += head_size;
 	}
+
+	/* queue up for recycling/reuse */
+	skb_mark_for_recycle(skb);
+
 	return skb;
 }
 
@@ -2123,8 +2152,10 @@ mlx5e_free_rx_shampo_hd_entry(struct mlx5e_rq *rq, u16 header_index)
 	u64 addr = shampo->info[header_index].addr;
 
 	if (((header_index + 1) & (MLX5E_SHAMPO_WQ_HEADER_PER_PAGE - 1)) == 0) {
-		shampo->info[header_index].addr = ALIGN_DOWN(addr, PAGE_SIZE);
-		mlx5e_page_release_dynamic(rq, shampo->info[header_index].page, true);
+		struct mlx5e_dma_info *dma_info = &shampo->info[header_index];
+
+		dma_info->addr = ALIGN_DOWN(addr, PAGE_SIZE);
+		mlx5e_page_release_fragmented(rq, dma_info->frag_page);
 	}
 	bitmap_clear(shampo->bitmap, header_index, 1);
 }
@@ -2145,7 +2176,6 @@ static void mlx5e_handle_rx_cqe_mpwrq_shampo(struct mlx5e_rq *rq, struct mlx5_cq
 	bool match		= cqe->shampo.match;
 	struct mlx5e_rq_stats *stats = rq->stats;
 	struct mlx5e_rx_wqe_ll *wqe;
-	union mlx5e_alloc_unit *au;
 	struct mlx5e_mpw_info *wi;
 	struct mlx5_wq_ll *wq;
 
@@ -2195,8 +2225,10 @@ static void mlx5e_handle_rx_cqe_mpwrq_shampo(struct mlx5e_rq *rq, struct mlx5_cq
 	}
 
 	if (likely(head_size)) {
-		au = &wi->alloc_units[page_idx];
-		mlx5e_fill_skb_data(*skb, rq, au, data_bcnt, data_offset);
+		struct mlx5e_frag_page *frag_page;
+
+		frag_page = &wi->alloc_units.frag_pages[page_idx];
+		mlx5e_fill_skb_data(*skb, rq, frag_page, data_bcnt, data_offset);
 	}
 
 	mlx5e_shampo_complete_rx_cqe(rq, cqe, cqe_bcnt, *skb);
@@ -2210,7 +2242,6 @@ static void mlx5e_handle_rx_cqe_mpwrq_shampo(struct mlx5e_rq *rq, struct mlx5_cq
 
 	wq  = &rq->mpwqe.wq;
 	wqe = mlx5_wq_ll_get_wqe(wq, wqe_id);
-	mlx5e_free_rx_mpwqe(rq, wi, true);
 	mlx5_wq_ll_pop(wq, cqe->wqe_id, &wqe->next.next_wqe_index);
 }
 
@@ -2270,7 +2301,6 @@ static void mlx5e_handle_rx_cqe_mpwrq(struct mlx5e_rq *rq, struct mlx5_cqe64 *cq
 
 	wq  = &rq->mpwqe.wq;
 	wqe = mlx5_wq_ll_get_wqe(wq, wqe_id);
-	mlx5e_free_rx_mpwqe(rq, wi, true);
 	mlx5_wq_ll_pop(wq, cqe->wqe_id, &wqe->next.next_wqe_index);
 }
 
@@ -2489,7 +2519,7 @@ static void mlx5i_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
 
 	if (unlikely(MLX5E_RX_ERR_CQE(cqe))) {
 		rq->stats->wqe_err++;
-		goto wq_free_wqe;
+		goto wq_cyc_pop;
 	}
 
 	skb = INDIRECT_CALL_2(rq->wqe.skb_from_cqe,
@@ -2497,17 +2527,16 @@ static void mlx5i_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
 			      mlx5e_skb_from_cqe_nonlinear,
 			      rq, wi, cqe, cqe_bcnt);
 	if (!skb)
-		goto wq_free_wqe;
+		goto wq_cyc_pop;
 
 	mlx5i_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
 	if (unlikely(!skb->dev)) {
 		dev_kfree_skb_any(skb);
-		goto wq_free_wqe;
+		goto wq_cyc_pop;
 	}
 	napi_gro_receive(rq->cq.napi, skb);
 
-wq_free_wqe:
-	mlx5e_free_rx_wqe(rq, wi, true);
+wq_cyc_pop:
 	mlx5_wq_cyc_pop(wq);
 }
 
@@ -2582,12 +2611,12 @@ static void mlx5e_trap_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe
 
 	if (unlikely(MLX5E_RX_ERR_CQE(cqe))) {
 		rq->stats->wqe_err++;
-		goto free_wqe;
+		goto wq_cyc_pop;
 	}
 
 	skb = mlx5e_skb_from_cqe_nonlinear(rq, wi, cqe, cqe_bcnt);
 	if (!skb)
-		goto free_wqe;
+		goto wq_cyc_pop;
 
 	mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
 	skb_push(skb, ETH_HLEN);
@@ -2596,8 +2625,7 @@ static void mlx5e_trap_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe
 				 rq->netdev->devlink_port);
 	dev_kfree_skb_any(skb);
 
-free_wqe:
-	mlx5e_free_rx_wqe(rq, wi, false);
+wq_cyc_pop:
 	mlx5_wq_cyc_pop(wq);
 }
 

drivers/net/ethernet/mellanox/mlx5/core/en_stats.c

@@ -179,11 +179,6 @@ static const struct counter_desc sw_stats_desc[] = {
 	{ MLX5E_DECLARE_STAT(struct mlx5e_sw_stats, rx_buff_alloc_err) },
 	{ MLX5E_DECLARE_STAT(struct mlx5e_sw_stats, rx_cqe_compress_blks) },
 	{ MLX5E_DECLARE_STAT(struct mlx5e_sw_stats, rx_cqe_compress_pkts) },
-	{ MLX5E_DECLARE_STAT(struct mlx5e_sw_stats, rx_cache_reuse) },
-	{ MLX5E_DECLARE_STAT(struct mlx5e_sw_stats, rx_cache_full) },
-	{ MLX5E_DECLARE_STAT(struct mlx5e_sw_stats, rx_cache_empty) },
-	{ MLX5E_DECLARE_STAT(struct mlx5e_sw_stats, rx_cache_busy) },
-	{ MLX5E_DECLARE_STAT(struct mlx5e_sw_stats, rx_cache_waive) },
 	{ MLX5E_DECLARE_STAT(struct mlx5e_sw_stats, rx_congst_umr) },
 	{ MLX5E_DECLARE_STAT(struct mlx5e_sw_stats, rx_arfs_err) },
 	{ MLX5E_DECLARE_STAT(struct mlx5e_sw_stats, rx_recover) },
@@ -358,11 +353,6 @@ static void mlx5e_stats_grp_sw_update_stats_rq_stats(struct mlx5e_sw_stats *s,
 	s->rx_buff_alloc_err          += rq_stats->buff_alloc_err;
 	s->rx_cqe_compress_blks       += rq_stats->cqe_compress_blks;
 	s->rx_cqe_compress_pkts       += rq_stats->cqe_compress_pkts;
-	s->rx_cache_reuse             += rq_stats->cache_reuse;
-	s->rx_cache_full              += rq_stats->cache_full;
-	s->rx_cache_empty             += rq_stats->cache_empty;
-	s->rx_cache_busy              += rq_stats->cache_busy;
-	s->rx_cache_waive             += rq_stats->cache_waive;
 	s->rx_congst_umr              += rq_stats->congst_umr;
 	s->rx_arfs_err                += rq_stats->arfs_err;
 	s->rx_recover                 += rq_stats->recover;
@@ -1978,11 +1968,6 @@ static const struct counter_desc rq_stats_desc[] = {
 	{ MLX5E_DECLARE_RX_STAT(struct mlx5e_rq_stats, buff_alloc_err) },
 	{ MLX5E_DECLARE_RX_STAT(struct mlx5e_rq_stats, cqe_compress_blks) },
 	{ MLX5E_DECLARE_RX_STAT(struct mlx5e_rq_stats, cqe_compress_pkts) },
-	{ MLX5E_DECLARE_RX_STAT(struct mlx5e_rq_stats, cache_reuse) },
-	{ MLX5E_DECLARE_RX_STAT(struct mlx5e_rq_stats, cache_full) },
-	{ MLX5E_DECLARE_RX_STAT(struct mlx5e_rq_stats, cache_empty) },
-	{ MLX5E_DECLARE_RX_STAT(struct mlx5e_rq_stats, cache_busy) },
-	{ MLX5E_DECLARE_RX_STAT(struct mlx5e_rq_stats, cache_waive) },
 	{ MLX5E_DECLARE_RX_STAT(struct mlx5e_rq_stats, congst_umr) },
 	{ MLX5E_DECLARE_RX_STAT(struct mlx5e_rq_stats, arfs_err) },
 	{ MLX5E_DECLARE_RX_STAT(struct mlx5e_rq_stats, recover) },
@@ -2163,11 +2148,6 @@ static const struct counter_desc ptp_rq_stats_desc[] = {
 	{ MLX5E_DECLARE_PTP_RQ_STAT(struct mlx5e_rq_stats, buff_alloc_err) },
 	{ MLX5E_DECLARE_PTP_RQ_STAT(struct mlx5e_rq_stats, cqe_compress_blks) },
 	{ MLX5E_DECLARE_PTP_RQ_STAT(struct mlx5e_rq_stats, cqe_compress_pkts) },
-	{ MLX5E_DECLARE_PTP_RQ_STAT(struct mlx5e_rq_stats, cache_reuse) },
-	{ MLX5E_DECLARE_PTP_RQ_STAT(struct mlx5e_rq_stats, cache_full) },
-	{ MLX5E_DECLARE_PTP_RQ_STAT(struct mlx5e_rq_stats, cache_empty) },
-	{ MLX5E_DECLARE_PTP_RQ_STAT(struct mlx5e_rq_stats, cache_busy) },
-	{ MLX5E_DECLARE_PTP_RQ_STAT(struct mlx5e_rq_stats, cache_waive) },
 	{ MLX5E_DECLARE_PTP_RQ_STAT(struct mlx5e_rq_stats, congst_umr) },
 	{ MLX5E_DECLARE_PTP_RQ_STAT(struct mlx5e_rq_stats, arfs_err) },
 	{ MLX5E_DECLARE_PTP_RQ_STAT(struct mlx5e_rq_stats, recover) },

drivers/net/ethernet/mellanox/mlx5/core/en_stats.h

@@ -193,11 +193,6 @@ struct mlx5e_sw_stats {
 	u64 rx_buff_alloc_err;
 	u64 rx_cqe_compress_blks;
 	u64 rx_cqe_compress_pkts;
-	u64 rx_cache_reuse;
-	u64 rx_cache_full;
-	u64 rx_cache_empty;
-	u64 rx_cache_busy;
-	u64 rx_cache_waive;
 	u64 rx_congst_umr;
 	u64 rx_arfs_err;
 	u64 rx_recover;
@@ -362,11 +357,6 @@ struct mlx5e_rq_stats {
 	u64 buff_alloc_err;
 	u64 cqe_compress_blks;
 	u64 cqe_compress_pkts;
-	u64 cache_reuse;
-	u64 cache_full;
-	u64 cache_empty;
-	u64 cache_busy;
-	u64 cache_waive;
 	u64 congst_umr;
 	u64 arfs_err;
 	u64 recover;