Merge branch '200GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/tnguy/next-queue

Tony Nguyen says:

====================
idpf: XDP chapter I: convert Rx to libeth

Alexander Lobakin says:

XDP for idpf is currently 5 chapters:
* convert Rx to libeth (this);
* convert Tx and stats to libeth;
* generic XDP and XSk code changes, libeth_xdp;
* actual XDP for idpf via libeth_xdp;
* XSk for idpf (^).

Part I does the following:
* splits &idpf_queue into 4 (RQ, SQ, FQ, CQ) and puts them on a diet;
* ensures optimal cacheline placement, strictly asserts CL sizes;
* moves currently unused/dead singleq mode out of line;
* reuses libeth's Rx ptype definitions and helpers;
* uses libeth's Rx buffer management for both header and payload;
* eliminates memcpy()s and coherent DMA uses on hotpath, uses
  napi_build_skb() instead of in-place short skb allocation.

Most idpf patches, except for the queue split, removes more lines
than adds.

Expect far better memory utilization and +5-8% on Rx depending on
the case (+17% on skb XDP_DROP :>).

* '200GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/tnguy/next-queue:
  idpf: use libeth Rx buffer management for payload buffer
  idpf: convert header split mode to libeth + napi_build_skb()
  libeth: support different types of buffers for Rx
  idpf: remove legacy Page Pool Ethtool stats
  idpf: reuse libeth's definitions of parsed ptype structures
  idpf: compile singleq code only under default-n CONFIG_IDPF_SINGLEQ
  idpf: merge singleq and splitq &net_device_ops
  idpf: strictly assert cachelines of queue and queue vector structures
  idpf: avoid bloating &idpf_q_vector with big %NR_CPUS
  idpf: split &idpf_queue into 4 strictly-typed queue structures
  idpf: stop using macros for accessing queue descriptors
  libeth: add cacheline / struct layout assertion helpers
  page_pool: use __cacheline_group_{begin, end}_aligned()
  cache: add __cacheline_group_{begin, end}_aligned() (+ couple more)
====================

Link: https://patch.msgid.link/20240710203031.188081-1-anthony.l.nguyen@intel.comSigned-off-by: Jakub Kicinski <kuba@kernel.org>

drivers/net/ethernet/intel/Kconfig

@@ -384,17 +384,6 @@ config IGC_LEDS
 	  Optional support for controlling the NIC LED's with the netdev
 	  LED trigger.
 
-config IDPF
-	tristate "Intel(R) Infrastructure Data Path Function Support"
-	depends on PCI_MSI
-	select DIMLIB
-	select PAGE_POOL
-	select PAGE_POOL_STATS
-	help
-	  This driver supports Intel(R) Infrastructure Data Path Function
-	  devices.
-
-	  To compile this driver as a module, choose M here. The module
-	  will be called idpf.
+source "drivers/net/ethernet/intel/idpf/Kconfig"
 
 endif # NET_VENDOR_INTEL

drivers/net/ethernet/intel/idpf/Kconfig

+# SPDX-License-Identifier: GPL-2.0-only
+# Copyright (C) 2024 Intel Corporation
+
+config IDPF
+	tristate "Intel(R) Infrastructure Data Path Function Support"
+	depends on PCI_MSI
+	select DIMLIB
+	select LIBETH
+	help
+	  This driver supports Intel(R) Infrastructure Data Path Function
+	  devices.
+
+	  To compile this driver as a module, choose M here. The module
+	  will be called idpf.
+
+if IDPF
+
+config IDPF_SINGLEQ
+	bool "idpf singleq support"
+	help
+	  This option enables support for legacy single Rx/Tx queues w/no
+	  completion and fill queues. Only enable if you have hardware which
+	  wants to work in this mode as it increases the driver size and adds
+	  runtme checks on hotpath.
+
+endif # IDPF

drivers/net/ethernet/intel/idpf/Makefile

@@ -12,7 +12,8 @@ idpf-y := \
 	idpf_ethtool.o		\
 	idpf_lib.o		\
 	idpf_main.o		\
-	idpf_singleq_txrx.o	\
 	idpf_txrx.o		\
 	idpf_virtchnl.o 	\
 	idpf_vf_dev.o
+
+idpf-$(CONFIG_IDPF_SINGLEQ)	+= idpf_singleq_txrx.o

drivers/net/ethernet/intel/idpf/idpf.h

@@ -17,10 +17,8 @@ struct idpf_vport_max_q;
 #include <linux/sctp.h>
 #include <linux/ethtool_netlink.h>
 #include <net/gro.h>
-#include <linux/dim.h>
 
 #include "virtchnl2.h"
-#include "idpf_lan_txrx.h"
 #include "idpf_txrx.h"
 #include "idpf_controlq.h"
 
@@ -266,7 +264,6 @@ struct idpf_port_stats {
  *		    the worst case.
  * @num_bufqs_per_qgrp: Buffer queues per RX queue in a given grouping
  * @bufq_desc_count: Buffer queue descriptor count
- * @bufq_size: Size of buffers in ring (e.g. 2K, 4K, etc)
  * @num_rxq_grp: Number of RX queues in a group
  * @rxq_grps: Total number of RX groups. Number of groups * number of RX per
  *	      group will yield total number of RX queues.
@@ -302,7 +299,7 @@ struct idpf_vport {
 	u16 num_txq_grp;
 	struct idpf_txq_group *txq_grps;
 	u32 txq_model;
-	struct idpf_queue **txqs;
+	struct idpf_tx_queue **txqs;
 	bool crc_enable;
 
 	u16 num_rxq;
@@ -310,11 +307,10 @@ struct idpf_vport {
 	u32 rxq_desc_count;
 	u8 num_bufqs_per_qgrp;
 	u32 bufq_desc_count[IDPF_MAX_BUFQS_PER_RXQ_GRP];
-	u32 bufq_size[IDPF_MAX_BUFQS_PER_RXQ_GRP];
 	u16 num_rxq_grp;
 	struct idpf_rxq_group *rxq_grps;
 	u32 rxq_model;
-	struct idpf_rx_ptype_decoded rx_ptype_lkup[IDPF_RX_MAX_PTYPE];
+	struct libeth_rx_pt *rx_ptype_lkup;
 
 	struct idpf_adapter *adapter;
 	struct net_device *netdev;
@@ -601,7 +597,8 @@ struct idpf_adapter {
  */
 static inline int idpf_is_queue_model_split(u16 q_model)
 {
-	return q_model == VIRTCHNL2_QUEUE_MODEL_SPLIT;
+	return !IS_ENABLED(CONFIG_IDPF_SINGLEQ) ||
+	       q_model == VIRTCHNL2_QUEUE_MODEL_SPLIT;
 }
 
 #define idpf_is_cap_ena(adapter, field, flag) \

drivers/net/ethernet/intel/idpf/idpf_lan_txrx.h

@@ -4,6 +4,8 @@
 #ifndef _IDPF_LAN_TXRX_H_
 #define _IDPF_LAN_TXRX_H_
 
+#include <linux/bits.h>
+
 enum idpf_rss_hash {
 	IDPF_HASH_INVALID			= 0,
 	/* Values 1 - 28 are reserved for future use */

drivers/net/ethernet/intel/idpf/idpf_lib.c

@@ -4,8 +4,7 @@
 #include "idpf.h"
 #include "idpf_virtchnl.h"
 
-static const struct net_device_ops idpf_netdev_ops_splitq;
-static const struct net_device_ops idpf_netdev_ops_singleq;
+static const struct net_device_ops idpf_netdev_ops;
 
 /**
  * idpf_init_vector_stack - Fill the MSIX vector stack with vector index
@@ -69,7 +68,7 @@ static void idpf_deinit_vector_stack(struct idpf_adapter *adapter)
 static void idpf_mb_intr_rel_irq(struct idpf_adapter *adapter)
 {
 	clear_bit(IDPF_MB_INTR_MODE, adapter->flags);
-	free_irq(adapter->msix_entries[0].vector, adapter);
+	kfree(free_irq(adapter->msix_entries[0].vector, adapter));
 	queue_delayed_work(adapter->mbx_wq, &adapter->mbx_task, 0);
 }
 
@@ -124,15 +123,14 @@ static void idpf_mb_irq_enable(struct idpf_adapter *adapter)
  */
 static int idpf_mb_intr_req_irq(struct idpf_adapter *adapter)
 {
-	struct idpf_q_vector *mb_vector = &adapter->mb_vector;
 	int irq_num, mb_vidx = 0, err;
+	char *name;
 
 	irq_num = adapter->msix_entries[mb_vidx].vector;
-	mb_vector->name = kasprintf(GFP_KERNEL, "%s-%s-%d",
-				    dev_driver_string(&adapter->pdev->dev),
-				    "Mailbox", mb_vidx);
-	err = request_irq(irq_num, adapter->irq_mb_handler, 0,
-			  mb_vector->name, adapter);
+	name = kasprintf(GFP_KERNEL, "%s-%s-%d",
+			 dev_driver_string(&adapter->pdev->dev),
+			 "Mailbox", mb_vidx);
+	err = request_irq(irq_num, adapter->irq_mb_handler, 0, name, adapter);
 	if (err) {
 		dev_err(&adapter->pdev->dev,
 			"IRQ request for mailbox failed, error: %d\n", err);
@@ -765,10 +763,7 @@ static int idpf_cfg_netdev(struct idpf_vport *vport)
 	}
 
 	/* assign netdev_ops */
-	if (idpf_is_queue_model_split(vport->txq_model))
-		netdev->netdev_ops = &idpf_netdev_ops_splitq;
-	else
-		netdev->netdev_ops = &idpf_netdev_ops_singleq;
+	netdev->netdev_ops = &idpf_netdev_ops;
 
 	/* setup watchdog timeout value to be 5 second */
 	netdev->watchdog_timeo = 5 * HZ;
@@ -946,6 +941,9 @@ static void idpf_decfg_netdev(struct idpf_vport *vport)
 {
 	struct idpf_adapter *adapter = vport->adapter;
 
+	kfree(vport->rx_ptype_lkup);
+	vport->rx_ptype_lkup = NULL;
+
 	unregister_netdev(vport->netdev);
 	free_netdev(vport->netdev);
 	vport->netdev = NULL;
@@ -1318,14 +1316,14 @@ static void idpf_rx_init_buf_tail(struct idpf_vport *vport)
 
 		if (idpf_is_queue_model_split(vport->rxq_model)) {
 			for (j = 0; j < vport->num_bufqs_per_qgrp; j++) {
-				struct idpf_queue *q =
+				const struct idpf_buf_queue *q =
 					&grp->splitq.bufq_sets[j].bufq;
 
 				writel(q->next_to_alloc, q->tail);
 			}
 		} else {
 			for (j = 0; j < grp->singleq.num_rxq; j++) {
-				struct idpf_queue *q =
+				const struct idpf_rx_queue *q =
 					grp->singleq.rxqs[j];
 
 				writel(q->next_to_alloc, q->tail);
@@ -1855,7 +1853,7 @@ int idpf_initiate_soft_reset(struct idpf_vport *vport,
 	enum idpf_vport_state current_state = np->state;
 	struct idpf_adapter *adapter = vport->adapter;
 	struct idpf_vport *new_vport;
-	int err, i;
+	int err;
 
 	/* If the system is low on memory, we can end up in bad state if we
 	 * free all the memory for queue resources and try to allocate them
@@ -1929,46 +1927,6 @@ int idpf_initiate_soft_reset(struct idpf_vport *vport,
 	 */
 	memcpy(vport, new_vport, offsetof(struct idpf_vport, link_speed_mbps));
 
-	/* Since idpf_vport_queues_alloc was called with new_port, the queue
-	 * back pointers are currently pointing to the local new_vport. Reset
-	 * the backpointers to the original vport here
-	 */
-	for (i = 0; i < vport->num_txq_grp; i++) {
-		struct idpf_txq_group *tx_qgrp = &vport->txq_grps[i];
-		int j;
-
-		tx_qgrp->vport = vport;
-		for (j = 0; j < tx_qgrp->num_txq; j++)
-			tx_qgrp->txqs[j]->vport = vport;
-
-		if (idpf_is_queue_model_split(vport->txq_model))
-			tx_qgrp->complq->vport = vport;
-	}
-
-	for (i = 0; i < vport->num_rxq_grp; i++) {
-		struct idpf_rxq_group *rx_qgrp = &vport->rxq_grps[i];
-		struct idpf_queue *q;
-		u16 num_rxq;
-		int j;
-
-		rx_qgrp->vport = vport;
-		for (j = 0; j < vport->num_bufqs_per_qgrp; j++)
-			rx_qgrp->splitq.bufq_sets[j].bufq.vport = vport;
-
-		if (idpf_is_queue_model_split(vport->rxq_model))
-			num_rxq = rx_qgrp->splitq.num_rxq_sets;
-		else
-			num_rxq = rx_qgrp->singleq.num_rxq;
-
-		for (j = 0; j < num_rxq; j++) {
-			if (idpf_is_queue_model_split(vport->rxq_model))
-				q = &rx_qgrp->splitq.rxq_sets[j]->rxq;
-			else
-				q = rx_qgrp->singleq.rxqs[j];
-			q->vport = vport;
-		}
-	}
-
 	if (reset_cause == IDPF_SR_Q_CHANGE)
 		idpf_vport_alloc_vec_indexes(vport);
 
@@ -2393,24 +2351,10 @@ void idpf_free_dma_mem(struct idpf_hw *hw, struct idpf_dma_mem *mem)
 	mem->pa = 0;
 }
 
-static const struct net_device_ops idpf_netdev_ops_splitq = {
-	.ndo_open = idpf_open,
-	.ndo_stop = idpf_stop,
-	.ndo_start_xmit = idpf_tx_splitq_start,
-	.ndo_features_check = idpf_features_check,
-	.ndo_set_rx_mode = idpf_set_rx_mode,
-	.ndo_validate_addr = eth_validate_addr,
-	.ndo_set_mac_address = idpf_set_mac,
-	.ndo_change_mtu = idpf_change_mtu,
-	.ndo_get_stats64 = idpf_get_stats64,
-	.ndo_set_features = idpf_set_features,
-	.ndo_tx_timeout = idpf_tx_timeout,
-};
-
-static const struct net_device_ops idpf_netdev_ops_singleq = {
+static const struct net_device_ops idpf_netdev_ops = {
 	.ndo_open = idpf_open,
 	.ndo_stop = idpf_stop,
-	.ndo_start_xmit = idpf_tx_singleq_start,
+	.ndo_start_xmit = idpf_tx_start,
 	.ndo_features_check = idpf_features_check,
 	.ndo_set_rx_mode = idpf_set_rx_mode,
 	.ndo_validate_addr = eth_validate_addr,

drivers/net/ethernet/intel/idpf/idpf_main.c

@@ -8,6 +8,7 @@
 #define DRV_SUMMARY	"Intel(R) Infrastructure Data Path Function Linux Driver"
 
 MODULE_DESCRIPTION(DRV_SUMMARY);
+MODULE_IMPORT_NS(LIBETH);
 MODULE_LICENSE("GPL");
 
 /**

drivers/net/ethernet/intel/libeth/rx.c

@@ -6,7 +6,7 @@
 /* Rx buffer management */
 
 /**
- * libeth_rx_hw_len - get the actual buffer size to be passed to HW
+ * libeth_rx_hw_len_mtu - get the actual buffer size to be passed to HW
  * @pp: &page_pool_params of the netdev to calculate the size for
  * @max_len: maximum buffer size for a single descriptor
  *
@@ -14,7 +14,7 @@
  * MTU the @dev has, HW required alignment, minimum and maximum allowed values,
  * and system's page size.
  */
-static u32 libeth_rx_hw_len(const struct page_pool_params *pp, u32 max_len)
+static u32 libeth_rx_hw_len_mtu(const struct page_pool_params *pp, u32 max_len)
 {
 	u32 len;
 
@@ -26,6 +26,118 @@ static u32 libeth_rx_hw_len(const struct page_pool_params *pp, u32 max_len)
 	return len;
 }
 
+/**
+ * libeth_rx_hw_len_truesize - get the short buffer size to be passed to HW
+ * @pp: &page_pool_params of the netdev to calculate the size for
+ * @max_len: maximum buffer size for a single descriptor
+ * @truesize: desired truesize for the buffers
+ *
+ * Return: HW-writeable length per one buffer to pass it to the HW ignoring the
+ * MTU and closest to the passed truesize. Can be used for "short" buffer
+ * queues to fragment pages more efficiently.
+ */
+static u32 libeth_rx_hw_len_truesize(const struct page_pool_params *pp,
+				     u32 max_len, u32 truesize)
+{
+	u32 min, len;
+
+	min = SKB_HEAD_ALIGN(pp->offset + LIBETH_RX_BUF_STRIDE);
+	truesize = clamp(roundup_pow_of_two(truesize), roundup_pow_of_two(min),
+			 PAGE_SIZE << LIBETH_RX_PAGE_ORDER);
+
+	len = SKB_WITH_OVERHEAD(truesize - pp->offset);
+	len = ALIGN_DOWN(len, LIBETH_RX_BUF_STRIDE) ? : LIBETH_RX_BUF_STRIDE;
+	len = min3(len, ALIGN_DOWN(max_len ? : U32_MAX, LIBETH_RX_BUF_STRIDE),
+		   pp->max_len);
+
+	return len;
+}
+
+/**
+ * libeth_rx_page_pool_params - calculate params with the stack overhead
+ * @fq: buffer queue to calculate the size for
+ * @pp: &page_pool_params of the netdev
+ *
+ * Set the PP params to will all needed stack overhead (headroom, tailroom) and
+ * both the HW buffer length and the truesize for all types of buffers. For
+ * "short" buffers, truesize never exceeds the "wanted" one; for the rest,
+ * it can be up to the page size.
+ *
+ * Return: true on success, false on invalid input params.
+ */
+static bool libeth_rx_page_pool_params(struct libeth_fq *fq,
+				       struct page_pool_params *pp)
+{
+	pp->offset = LIBETH_SKB_HEADROOM;
+	/* HW-writeable / syncable length per one page */
+	pp->max_len = LIBETH_RX_PAGE_LEN(pp->offset);
+
+	/* HW-writeable length per buffer */
+	switch (fq->type) {
+	case LIBETH_FQE_MTU:
+		fq->buf_len = libeth_rx_hw_len_mtu(pp, fq->buf_len);
+		break;
+	case LIBETH_FQE_SHORT:
+		fq->buf_len = libeth_rx_hw_len_truesize(pp, fq->buf_len,
+							fq->truesize);
+		break;
+	case LIBETH_FQE_HDR:
+		fq->buf_len = ALIGN(LIBETH_MAX_HEAD, LIBETH_RX_BUF_STRIDE);
+		break;
+	default:
+		return false;
+	}
+
+	/* Buffer size to allocate */
+	fq->truesize = roundup_pow_of_two(SKB_HEAD_ALIGN(pp->offset +
+							 fq->buf_len));
+
+	return true;
+}
+
+/**
+ * libeth_rx_page_pool_params_zc - calculate params without the stack overhead
+ * @fq: buffer queue to calculate the size for
+ * @pp: &page_pool_params of the netdev
+ *
+ * Set the PP params to exclude the stack overhead and both the buffer length
+ * and the truesize, which are equal for the data buffers. Note that this
+ * requires separate header buffers to be always active and account the
+ * overhead.
+ * With the MTU == ``PAGE_SIZE``, this allows the kernel to enable the zerocopy
+ * mode.
+ *
+ * Return: true on success, false on invalid input params.
+ */
+static bool libeth_rx_page_pool_params_zc(struct libeth_fq *fq,
+					  struct page_pool_params *pp)
+{
+	u32 mtu, max;
+
+	pp->offset = 0;
+	pp->max_len = PAGE_SIZE << LIBETH_RX_PAGE_ORDER;
+
+	switch (fq->type) {
+	case LIBETH_FQE_MTU:
+		mtu = READ_ONCE(pp->netdev->mtu);
+		break;
+	case LIBETH_FQE_SHORT:
+		mtu = fq->truesize;
+		break;
+	default:
+		return false;
+	}
+
+	mtu = roundup_pow_of_two(mtu);
+	max = min(rounddown_pow_of_two(fq->buf_len ? : U32_MAX),
+		  pp->max_len);
+
+	fq->buf_len = clamp(mtu, LIBETH_RX_BUF_STRIDE, max);
+	fq->truesize = fq->buf_len;
+
+	return true;
+}
+
 /**
  * libeth_rx_fq_create - create a PP with the default libeth settings
  * @fq: buffer queue struct to fill
@@ -44,19 +156,17 @@ int libeth_rx_fq_create(struct libeth_fq *fq, struct napi_struct *napi)
 		.netdev		= napi->dev,
 		.napi		= napi,
 		.dma_dir	= DMA_FROM_DEVICE,
-		.offset		= LIBETH_SKB_HEADROOM,
 	};
 	struct libeth_fqe *fqes;
 	struct page_pool *pool;
+	bool ret;
 
-	/* HW-writeable / syncable length per one page */
-	pp.max_len = LIBETH_RX_PAGE_LEN(pp.offset);
-
-	/* HW-writeable length per buffer */
-	fq->buf_len = libeth_rx_hw_len(&pp, fq->buf_len);
-	/* Buffer size to allocate */
-	fq->truesize = roundup_pow_of_two(SKB_HEAD_ALIGN(pp.offset +
-							 fq->buf_len));
+	if (!fq->hsplit)
+		ret = libeth_rx_page_pool_params(fq, &pp);
+	else
+		ret = libeth_rx_page_pool_params_zc(fq, &pp);
+	if (!ret)
+		return -EINVAL;
 
 	pool = page_pool_create(&pp);
 	if (IS_ERR(pool))

include/linux/cache.h

@@ -13,6 +13,32 @@
 #define SMP_CACHE_BYTES L1_CACHE_BYTES
 #endif
 
+/**
+ * SMP_CACHE_ALIGN - align a value to the L2 cacheline size
+ * @x: value to align
+ *
+ * On some architectures, L2 ("SMP") CL size is bigger than L1, and sometimes,
+ * this needs to be accounted.
+ *
+ * Return: aligned value.
+ */
+#ifndef SMP_CACHE_ALIGN
+#define SMP_CACHE_ALIGN(x)	ALIGN(x, SMP_CACHE_BYTES)
+#endif
+
+/*
+ * ``__aligned_largest`` aligns a field to the value most optimal for the
+ * target architecture to perform memory operations. Get the actual value
+ * to be able to use it anywhere else.
+ */
+#ifndef __LARGEST_ALIGN
+#define __LARGEST_ALIGN		sizeof(struct { long x; } __aligned_largest)
+#endif
+
+#ifndef LARGEST_ALIGN
+#define LARGEST_ALIGN(x)	ALIGN(x, __LARGEST_ALIGN)
+#endif
+
 /*
  * __read_mostly is used to keep rarely changing variables out of frequently
  * updated cachelines. Its use should be reserved for data that is used
@@ -95,6 +121,39 @@
 	__u8 __cacheline_group_end__##GROUP[0]
 #endif
 
+/**
+ * __cacheline_group_begin_aligned - declare an aligned group start
+ * @GROUP: name of the group
+ * @...: optional group alignment
+ *
+ * The following block inside a struct:
+ *
+ *	__cacheline_group_begin_aligned(grp);
+ *	field a;
+ *	field b;
+ *	__cacheline_group_end_aligned(grp);
+ *
+ * will always be aligned to either the specified alignment or
+ * ``SMP_CACHE_BYTES``.
+ */
+#define __cacheline_group_begin_aligned(GROUP, ...)		\
+	__cacheline_group_begin(GROUP)				\
+	__aligned((__VA_ARGS__ + 0) ? : SMP_CACHE_BYTES)
+
+/**
+ * __cacheline_group_end_aligned - declare an aligned group end
+ * @GROUP: name of the group
+ * @...: optional alignment (same as was in __cacheline_group_begin_aligned())
+ *
+ * Note that the end marker is aligned to sizeof(long) to allow more precise
+ * size assertion. It also declares a padding at the end to avoid next field
+ * falling into this cacheline.
+ */
+#define __cacheline_group_end_aligned(GROUP, ...)		\
+	__cacheline_group_end(GROUP) __aligned(sizeof(long));	\
+	struct { } __cacheline_group_pad__##GROUP		\
+	__aligned((__VA_ARGS__ + 0) ? : SMP_CACHE_BYTES)
+
 #ifndef CACHELINE_ASSERT_GROUP_MEMBER
 #define CACHELINE_ASSERT_GROUP_MEMBER(TYPE, GROUP, MEMBER) \
 	BUILD_BUG_ON(!(offsetof(TYPE, MEMBER) >= \

include/net/libeth/cache.h

+/* SPDX-License-Identifier: GPL-2.0-only */
+/* Copyright (C) 2024 Intel Corporation */
+
+#ifndef __LIBETH_CACHE_H
+#define __LIBETH_CACHE_H
+
+#include <linux/cache.h>
+
+/**
+ * libeth_cacheline_group_assert - make sure cacheline group size is expected
+ * @type: type of the structure containing the group
+ * @grp: group name inside the struct
+ * @sz: expected group size
+ */
+#if defined(CONFIG_64BIT) && SMP_CACHE_BYTES == 64
+#define libeth_cacheline_group_assert(type, grp, sz)			      \
+	static_assert(offsetof(type, __cacheline_group_end__##grp) -	      \
+		      offsetofend(type, __cacheline_group_begin__##grp) ==    \
+		      (sz))
+#define __libeth_cacheline_struct_assert(type, sz)			      \
+	static_assert(sizeof(type) == (sz))
+#else /* !CONFIG_64BIT || SMP_CACHE_BYTES != 64 */
+#define libeth_cacheline_group_assert(type, grp, sz)			      \
+	static_assert(offsetof(type, __cacheline_group_end__##grp) -	      \
+		      offsetofend(type, __cacheline_group_begin__##grp) <=    \
+		      (sz))
+#define __libeth_cacheline_struct_assert(type, sz)			      \
+	static_assert(sizeof(type) <= (sz))
+#endif /* !CONFIG_64BIT || SMP_CACHE_BYTES != 64 */
+
+#define __libeth_cls1(sz1)	SMP_CACHE_ALIGN(sz1)
+#define __libeth_cls2(sz1, sz2)	(SMP_CACHE_ALIGN(sz1) + SMP_CACHE_ALIGN(sz2))
+#define __libeth_cls3(sz1, sz2, sz3)					      \
+	(SMP_CACHE_ALIGN(sz1) + SMP_CACHE_ALIGN(sz2) + SMP_CACHE_ALIGN(sz3))
+#define __libeth_cls(...)						      \
+	CONCATENATE(__libeth_cls, COUNT_ARGS(__VA_ARGS__))(__VA_ARGS__)
+
+/**
+ * libeth_cacheline_struct_assert - make sure CL-based struct size is expected
+ * @type: type of the struct
+ * @...: from 1 to 3 CL group sizes (read-mostly, read-write, cold)
+ *
+ * When a struct contains several CL groups, it's difficult to predict its size
+ * on different architectures. The macro instead takes sizes of all of the
+ * groups the structure contains and generates the final struct size.
+ */
+#define libeth_cacheline_struct_assert(type, ...)			      \
+	__libeth_cacheline_struct_assert(type, __libeth_cls(__VA_ARGS__));    \
+	static_assert(__alignof(type) >= SMP_CACHE_BYTES)
+
+/**
+ * libeth_cacheline_set_assert - make sure CL-based struct layout is expected
+ * @type: type of the struct
+ * @ro: expected size of the read-mostly group
+ * @rw: expected size of the read-write group
+ * @c: expected size of the cold group
+ *
+ * Check that each group size is expected and then do final struct size check.
+ */
+#define libeth_cacheline_set_assert(type, ro, rw, c)			      \
+	libeth_cacheline_group_assert(type, read_mostly, ro);		      \
+	libeth_cacheline_group_assert(type, read_write, rw);		      \
+	libeth_cacheline_group_assert(type, cold, c);			      \
+	libeth_cacheline_struct_assert(type, ro, rw, c)
+
+#endif /* __LIBETH_CACHE_H */

include/net/libeth/rx.h

@@ -17,6 +17,8 @@
 #define LIBETH_MAX_HEADROOM	LIBETH_SKB_HEADROOM
 /* Link layer / L2 overhead: Ethernet, 2 VLAN tags (C + S), FCS */
 #define LIBETH_RX_LL_LEN	(ETH_HLEN + 2 * VLAN_HLEN + ETH_FCS_LEN)
+/* Maximum supported L2-L4 header length */
+#define LIBETH_MAX_HEAD		roundup_pow_of_two(max(MAX_HEADER, 256))
 
 /* Always use order-0 pages */
 #define LIBETH_RX_PAGE_ORDER	0
@@ -43,6 +45,18 @@ struct libeth_fqe {
 	u32			truesize;
 } __aligned_largest;
 
+/**
+ * enum libeth_fqe_type - enum representing types of Rx buffers
+ * @LIBETH_FQE_MTU: buffer size is determined by MTU
+ * @LIBETH_FQE_SHORT: buffer size is smaller than MTU, for short frames
+ * @LIBETH_FQE_HDR: buffer size is ```LIBETH_MAX_HEAD```-sized, for headers
+ */
+enum libeth_fqe_type {
+	LIBETH_FQE_MTU		= 0U,
+	LIBETH_FQE_SHORT,
+	LIBETH_FQE_HDR,
+};
+
 /**
  * struct libeth_fq - structure representing a buffer (fill) queue
  * @fp: hotpath part of the structure
@@ -50,6 +64,8 @@ struct libeth_fqe {
  * @fqes: array of Rx buffers
  * @truesize: size to allocate per buffer, w/overhead
  * @count: number of descriptors/buffers the queue has
+ * @type: type of the buffers this queue has
+ * @hsplit: flag whether header split is enabled
  * @buf_len: HW-writeable length per each buffer
  * @nid: ID of the closest NUMA node with memory
  */
@@ -63,6 +79,9 @@ struct libeth_fq {
 	);
 
 	/* Cold fields */
+	enum libeth_fqe_type	type:2;
+	bool			hsplit:1;
+
 	u32			buf_len;
 	int			nid;
 };

include/net/page_pool/types.h

@@ -129,6 +129,16 @@ struct page_pool_stats {
 };
 #endif
 
+/* The whole frag API block must stay within one cacheline. On 32-bit systems,
+ * sizeof(long) == sizeof(int), so that the block size is ``3 * sizeof(long)``.
+ * On 64-bit systems, the actual size is ``2 * sizeof(long) + sizeof(int)``.
+ * The closest pow-2 to both of them is ``4 * sizeof(long)``, so just use that
+ * one for simplicity.
+ * Having it aligned to a cacheline boundary may be excessive and doesn't bring
+ * any good.
+ */
+#define PAGE_POOL_FRAG_GROUP_ALIGN	(4 * sizeof(long))
+
 struct page_pool {
 	struct page_pool_params_fast p;
 
@@ -142,19 +152,11 @@ struct page_pool {
 	bool system:1;			/* This is a global percpu pool */
 #endif
 
-	/* The following block must stay within one cacheline. On 32-bit
-	 * systems, sizeof(long) == sizeof(int), so that the block size is
-	 * ``3 * sizeof(long)``. On 64-bit systems, the actual size is
-	 * ``2 * sizeof(long) + sizeof(int)``. The closest pow-2 to both of
-	 * them is ``4 * sizeof(long)``, so just use that one for simplicity.
-	 * Having it aligned to a cacheline boundary may be excessive and
-	 * doesn't bring any good.
-	 */
-	__cacheline_group_begin(frag) __aligned(4 * sizeof(long));
+	__cacheline_group_begin_aligned(frag, PAGE_POOL_FRAG_GROUP_ALIGN);
 	long frag_users;
 	netmem_ref frag_page;
 	unsigned int frag_offset;
-	__cacheline_group_end(frag);
+	__cacheline_group_end_aligned(frag, PAGE_POOL_FRAG_GROUP_ALIGN);
 
 	struct delayed_work release_dw;
 	void (*disconnect)(void *pool);

net/core/page_pool.c

@@ -178,7 +178,8 @@ static void page_pool_struct_check(void)
 	CACHELINE_ASSERT_GROUP_MEMBER(struct page_pool, frag, frag_users);
 	CACHELINE_ASSERT_GROUP_MEMBER(struct page_pool, frag, frag_page);
 	CACHELINE_ASSERT_GROUP_MEMBER(struct page_pool, frag, frag_offset);
-	CACHELINE_ASSERT_GROUP_SIZE(struct page_pool, frag, 4 * sizeof(long));
+	CACHELINE_ASSERT_GROUP_SIZE(struct page_pool, frag,
+				    PAGE_POOL_FRAG_GROUP_ALIGN);
 }
 
 static int page_pool_init(struct page_pool *pool,