Merge branch 'bpf-cpumap-enable-gro-for-xdp_pass-frames'

Alexander Lobakin says:

====================
bpf: cpumap: enable GRO for XDP_PASS frames

Several months ago, I had been looking through my old XDP hints tree[0]
to check whether some patches not directly related to hints can be sent
standalone. Roughly at the same time, Daniel appeared and asked[1] about
GRO for cpumap from that tree.

Currently, cpumap uses its own kthread which processes cpumap-redirected
frames by batches of 8, without any weighting (but with rescheduling
points). The resulting skbs get passed to the stack via
netif_receive_skb_list(), which means no GRO happens.
Even though we can't currently pass checksum status from the drivers,
in many cases GRO performs better than the listified Rx without the
aggregation, confirmed by tests.

In order to enable GRO in cpumap, we need to do the following:

* patches 1-2: decouple the GRO struct from the NAPI struct and allow
  using it out of a NAPI entity within the kernel core code;
* patch 3: switch cpumap from netif_receive_skb_list() to
  gro_receive_skb().

Additional improvements:

* patch 4: optimize XDP_PASS in cpumap by using arrays instead of linked
  lists;
* patch 5-6: introduce and use function do get skbs from the NAPI percpu
  caches by bulks, not one at a time;
* patch 7-8: use that function in veth as well and remove the one that
  was now superseded by it.

My trafficgen UDP GRO tests, small frame sizes:

                GRO off    GRO on
baseline        2.7        N/A       Mpps
patch 3         2.3        4         Mpps
patch 8         2.4        4.7       Mpps

1...3 diff      -17        +48       %
1...8 diff      -11        +74       %

Daniel reported from +14%[2] to +18%[3] of throughput in neper's TCP RR
tests. On my system however, the same test gave me up to +100%.

Note that there's a series from Lorenzo[4] which achieves the same, but
in a different way. During the discussions, the approach using a
standalone GRO instance was preferred over the threaded NAPI.

[0] https://github.com/alobakin/linux/tree/xdp_hints
[1] https://lore.kernel.org/bpf/cadda351-6e93-4568-ba26-21a760bf9a57@app.fastmail.com
[2] https://lore.kernel.org/bpf/merfatcdvwpx2lj4j2pahhwp4vihstpidws3jwljwazhh76xkd@t5vsh4gvk4mh
[3] https://lore.kernel.org/bpf/yzda66wro5twmzpmjoxvy4si5zvkehlmgtpi6brheek3sj73tj@o7kd6nurr3o6
[4] https://lore.kernel.org/bpf/20241130-cpumap-gro-v1-0-c1180b1b5758@kernel.org
====================

Link: https://patch.msgid.link/20250225171751.2268401-1-aleksander.lobakin@intel.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>

14 files changed, 315 insertions, 180 deletions

diff --git a/drivers/net/ethernet/brocade/bna/bnad.c b/drivers/net/ethernet/brocade/bna/bnad.c
index ece6f3b48327..3b9107003b00 100644
--- a/drivers/net/ethernet/brocade/bna/bnad.c
+++ b/drivers/net/ethernet/brocade/bna/bnad.c
@@ -19,6 +19,7 @@
 #include <linux/ip.h>
 #include <linux/prefetch.h>
 #include <linux/module.h>
+#include <net/gro.h>
 
 #include "bnad.h"
 #include "bna.h"
diff --git a/drivers/net/ethernet/cortina/gemini.c b/drivers/net/ethernet/cortina/gemini.c
index 991e3839858b..1f8067bdd61a 100644
--- a/drivers/net/ethernet/cortina/gemini.c
+++ b/drivers/net/ethernet/cortina/gemini.c
@@ -40,6 +40,7 @@
 #include <linux/in.h>
 #include <linux/ip.h>
 #include <linux/ipv6.h>
+#include <net/gro.h>
 
 #include "gemini.h"
 
diff --git a/drivers/net/veth.c b/drivers/net/veth.c
index ba3ae2d8092f..05f5eeef539f 100644
--- a/drivers/net/veth.c
+++ b/drivers/net/veth.c
@@ -684,8 +684,7 @@ static void veth_xdp_rcv_bulk_skb(struct veth_rq *rq, void **frames,
 	void *skbs[VETH_XDP_BATCH];
 	int i;
 
-	if (xdp_alloc_skb_bulk(skbs, n_xdpf,
-			       GFP_ATOMIC | __GFP_ZERO) < 0) {
+	if (unlikely(!napi_skb_cache_get_bulk(skbs, n_xdpf))) {
 		for (i = 0; i < n_xdpf; i++)
 			xdp_return_frame(frames[i]);
 		stats->rx_drops += n_xdpf;
diff --git a/drivers/net/wwan/t7xx/t7xx_hif_dpmaif_rx.c b/drivers/net/wwan/t7xx/t7xx_hif_dpmaif_rx.c
index 7a9c09cd4fdc..6a7a26085fc7 100644
--- a/drivers/net/wwan/t7xx/t7xx_hif_dpmaif_rx.c
+++ b/drivers/net/wwan/t7xx/t7xx_hif_dpmaif_rx.c
@@ -41,6 +41,7 @@
 #include <linux/types.h>
 #include <linux/wait.h>
 #include <linux/workqueue.h>
+#include <net/gro.h>
 
 #include "t7xx_dpmaif.h"
 #include "t7xx_hif_dpmaif.h"
diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index 2094d3edda73..26a0c4e4d963 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h
@@ -340,11 +340,27 @@ struct gro_list {
 };
 
 /*
- * size of gro hash buckets, must less than bit number of
- * napi_struct::gro_bitmask
+ * size of gro hash buckets, must be <= the number of bits in
+ * gro_node::bitmask
  */
 #define GRO_HASH_BUCKETS	8
 
+/**
+ * struct gro_node - structure to support Generic Receive Offload
+ * @bitmask: bitmask to indicate used buckets in @hash
+ * @hash: hashtable of pending aggregated skbs, separated by flows
+ * @rx_list: list of pending ``GRO_NORMAL`` skbs
+ * @rx_count: cached current length of @rx_list
+ * @cached_napi_id: napi_struct::napi_id cached for hotpath, 0 for standalone
+ */
+struct gro_node {
+	unsigned long		bitmask;
+	struct gro_list		hash[GRO_HASH_BUCKETS];
+	struct list_head	rx_list;
+	u32			rx_count;
+	u32			cached_napi_id;
+};
+
 /*
  * Structure for per-NAPI config
  */
@@ -371,7 +387,6 @@ struct napi_struct {
 	unsigned long		state;
 	int			weight;
 	u32			defer_hard_irqs_count;
-	unsigned long		gro_bitmask;
 	int			(*poll)(struct napi_struct *, int);
 #ifdef CONFIG_NETPOLL
 	/* CPU actively polling if netpoll is configured */
@@ -380,11 +395,8 @@ struct napi_struct {
 	/* CPU on which NAPI has been scheduled for processing */
 	int			list_owner;
 	struct net_device	*dev;
-	struct gro_list		gro_hash[GRO_HASH_BUCKETS];
 	struct sk_buff		*skb;
-	struct list_head	rx_list; /* Pending GRO_NORMAL skbs */
-	int			rx_count; /* length of rx_list */
-	unsigned int		napi_id; /* protected by netdev_lock */
+	struct gro_node		gro;
 	struct hrtimer		timer;
 	/* all fields past this point are write-protected by netdev_lock */
 	struct task_struct	*thread;
@@ -392,6 +404,7 @@ struct napi_struct {
 	unsigned long		irq_suspend_timeout;
 	u32			defer_hard_irqs;
 	/* control-path-only fields follow */
+	u32			napi_id;
 	struct list_head	dev_list;
 	struct hlist_node	napi_hash_node;
 	int			irq;
@@ -4131,8 +4144,14 @@ int netif_receive_skb(struct sk_buff *skb);
 int netif_receive_skb_core(struct sk_buff *skb);
 void netif_receive_skb_list_internal(struct list_head *head);
 void netif_receive_skb_list(struct list_head *head);
-gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb);
-void napi_gro_flush(struct napi_struct *napi, bool flush_old);
+gro_result_t gro_receive_skb(struct gro_node *gro, struct sk_buff *skb);
+
+static inline gro_result_t napi_gro_receive(struct napi_struct *napi,
+					    struct sk_buff *skb)
+{
+	return gro_receive_skb(&napi->gro, skb);
+}
+
 struct sk_buff *napi_get_frags(struct napi_struct *napi);
 gro_result_t napi_gro_frags(struct napi_struct *napi);
 
diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
index 171aa15f6541..14517e95a46c 100644
--- a/include/linux/skbuff.h
+++ b/include/linux/skbuff.h
@@ -1320,6 +1320,7 @@ struct sk_buff *build_skb_around(struct sk_buff *skb,
 				 void *data, unsigned int frag_size);
 void skb_attempt_defer_free(struct sk_buff *skb);
 
+u32 napi_skb_cache_get_bulk(void **skbs, u32 n);
 struct sk_buff *napi_build_skb(void *data, unsigned int frag_size);
 struct sk_buff *slab_build_skb(void *data);
 
diff --git a/include/net/busy_poll.h b/include/net/busy_poll.h
index cab6146a510a..6e172d0f6ef5 100644
--- a/include/net/busy_poll.h
+++ b/include/net/busy_poll.h
@@ -127,18 +127,24 @@ static inline void sk_busy_loop(struct sock *sk, int nonblock)
 }
 
 /* used in the NIC receive handler to mark the skb */
-static inline void skb_mark_napi_id(struct sk_buff *skb,
-				    struct napi_struct *napi)
+static inline void __skb_mark_napi_id(struct sk_buff *skb,
+				      const struct gro_node *gro)
 {
 #ifdef CONFIG_NET_RX_BUSY_POLL
 	/* If the skb was already marked with a valid NAPI ID, avoid overwriting
 	 * it.
 	 */
 	if (!napi_id_valid(skb->napi_id))
-		skb->napi_id = napi->napi_id;
+		skb->napi_id = gro->cached_napi_id;
 #endif
 }
 
+static inline void skb_mark_napi_id(struct sk_buff *skb,
+				    const struct napi_struct *napi)
+{
+	__skb_mark_napi_id(skb, &napi->gro);
+}
+
 /* used in the protocol handler to propagate the napi_id to the socket */
 static inline void sk_mark_napi_id(struct sock *sk, const struct sk_buff *skb)
 {
diff --git a/include/net/gro.h b/include/net/gro.h
index 7b548f91754b..22d3a69e4404 100644
--- a/include/net/gro.h
+++ b/include/net/gro.h
@@ -509,28 +509,46 @@ static inline int gro_receive_network_flush(const void *th, const void *th2,
 
 int skb_gro_receive(struct sk_buff *p, struct sk_buff *skb);
 int skb_gro_receive_list(struct sk_buff *p, struct sk_buff *skb);
+void __gro_flush(struct gro_node *gro, bool flush_old);
+
+static inline void gro_flush(struct gro_node *gro, bool flush_old)
+{
+	if (!gro->bitmask)
+		return;
+
+	__gro_flush(gro, flush_old);
+}
+
+static inline void napi_gro_flush(struct napi_struct *napi, bool flush_old)
+{
+	gro_flush(&napi->gro, flush_old);
+}
 
 /* Pass the currently batched GRO_NORMAL SKBs up to the stack. */
-static inline void gro_normal_list(struct napi_struct *napi)
+static inline void gro_normal_list(struct gro_node *gro)
 {
-	if (!napi->rx_count)
+	if (!gro->rx_count)
 		return;
-	netif_receive_skb_list_internal(&napi->rx_list);
-	INIT_LIST_HEAD(&napi->rx_list);
-	napi->rx_count = 0;
+	netif_receive_skb_list_internal(&gro->rx_list);
+	INIT_LIST_HEAD(&gro->rx_list);
+	gro->rx_count = 0;
 }
 
 /* Queue one GRO_NORMAL SKB up for list processing. If batch size exceeded,
  * pass the whole batch up to the stack.
  */
-static inline void gro_normal_one(struct napi_struct *napi, struct sk_buff *skb, int segs)
+static inline void gro_normal_one(struct gro_node *gro, struct sk_buff *skb,
+				  int segs)
 {
-	list_add_tail(&skb->list, &napi->rx_list);
-	napi->rx_count += segs;
-	if (napi->rx_count >= READ_ONCE(net_hotdata.gro_normal_batch))
-		gro_normal_list(napi);
+	list_add_tail(&skb->list, &gro->rx_list);
+	gro->rx_count += segs;
+	if (gro->rx_count >= READ_ONCE(net_hotdata.gro_normal_batch))
+		gro_normal_list(gro);
 }
 
+void gro_init(struct gro_node *gro);
+void gro_cleanup(struct gro_node *gro);
+
 /* This function is the alternative of 'inet_iif' and 'inet_sdif'
  * functions in case we can not rely on fields of IPCB.
  *
diff --git a/include/net/xdp.h b/include/net/xdp.h
index 4dafc5e021f1..48efacbaa35d 100644
--- a/include/net/xdp.h
+++ b/include/net/xdp.h
@@ -343,7 +343,6 @@ struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf,
 					   struct net_device *dev);
 struct sk_buff *xdp_build_skb_from_frame(struct xdp_frame *xdpf,
 					 struct net_device *dev);
-int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp);
 struct xdp_frame *xdpf_clone(struct xdp_frame *xdpf);
 
 static inline
diff --git a/kernel/bpf/cpumap.c b/kernel/bpf/cpumap.c
index 774accbd4a22..67e8a2fc1a99 100644
--- a/kernel/bpf/cpumap.c
+++ b/kernel/bpf/cpumap.c
@@ -33,8 +33,8 @@
 #include <trace/events/xdp.h>
 #include <linux/btf_ids.h>
 
-#include <linux/netdevice.h>   /* netif_receive_skb_list */
-#include <linux/etherdevice.h> /* eth_type_trans */
+#include <linux/netdevice.h>
+#include <net/gro.h>
 
 /* General idea: XDP packets getting XDP redirected to another CPU,
  * will maximum be stored/queued for one driver ->poll() call.  It is
@@ -68,6 +68,7 @@ struct bpf_cpu_map_entry {
 
 	struct bpf_cpumap_val value;
 	struct bpf_prog *prog;
+	struct gro_node gro;
 
 	struct completion kthread_running;
 	struct rcu_work free_work;
@@ -133,22 +134,23 @@ static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
 	}
 }
 
-static void cpu_map_bpf_prog_run_skb(struct bpf_cpu_map_entry *rcpu,
-				     struct list_head *listp,
-				     struct xdp_cpumap_stats *stats)
+static u32 cpu_map_bpf_prog_run_skb(struct bpf_cpu_map_entry *rcpu,
+				    void **skbs, u32 skb_n,
+				    struct xdp_cpumap_stats *stats)
 {
-	struct sk_buff *skb, *tmp;
 	struct xdp_buff xdp;
-	u32 act;
+	u32 act, pass = 0;
 	int err;
 
-	list_for_each_entry_safe(skb, tmp, listp, list) {
+	for (u32 i = 0; i < skb_n; i++) {
+		struct sk_buff *skb = skbs[i];
+
 		act = bpf_prog_run_generic_xdp(skb, &xdp, rcpu->prog);
 		switch (act) {
 		case XDP_PASS:
+			skbs[pass++] = skb;
 			break;
 		case XDP_REDIRECT:
-			skb_list_del_init(skb);
 			err = xdp_do_generic_redirect(skb->dev, skb, &xdp,
 						      rcpu->prog);
 			if (unlikely(err)) {
@@ -157,7 +159,7 @@ static void cpu_map_bpf_prog_run_skb(struct bpf_cpu_map_entry *rcpu,
 			} else {
 				stats->redirect++;
 			}
-			return;
+			break;
 		default:
 			bpf_warn_invalid_xdp_action(NULL, rcpu->prog, act);
 			fallthrough;
@@ -165,12 +167,15 @@ static void cpu_map_bpf_prog_run_skb(struct bpf_cpu_map_entry *rcpu,
 			trace_xdp_exception(skb->dev, rcpu->prog, act);
 			fallthrough;
 		case XDP_DROP:
-			skb_list_del_init(skb);
-			kfree_skb(skb);
+			napi_consume_skb(skb, true);
 			stats->drop++;
-			return;
+			break;
 		}
 	}
+
+	stats->pass += pass;
+
+	return pass;
 }
 
 static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
@@ -204,7 +209,6 @@ static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
 				stats->drop++;
 			} else {
 				frames[nframes++] = xdpf;
-				stats->pass++;
 			}
 			break;
 		case XDP_REDIRECT:
@@ -228,43 +232,65 @@ static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
 	}
 
 	xdp_clear_return_frame_no_direct();
+	stats->pass += nframes;
 
 	return nframes;
 }
 
 #define CPUMAP_BATCH 8
 
-static int cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames,
-				int xdp_n, struct xdp_cpumap_stats *stats,
-				struct list_head *list)
+struct cpu_map_ret {
+	u32 xdp_n;
+	u32 skb_n;
+};
+
+static void cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames,
+				 void **skbs, struct cpu_map_ret *ret,
+				 struct xdp_cpumap_stats *stats)
 {
 	struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx;
-	int nframes;
 
 	if (!rcpu->prog)
-		return xdp_n;
+		goto out;
 
-	rcu_read_lock_bh();
+	rcu_read_lock();
 	bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx);
 
-	nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, xdp_n, stats);
+	ret->xdp_n = cpu_map_bpf_prog_run_xdp(rcpu, frames, ret->xdp_n, stats);
+	if (unlikely(ret->skb_n))
+		ret->skb_n = cpu_map_bpf_prog_run_skb(rcpu, skbs, ret->skb_n,
+						      stats);
 
 	if (stats->redirect)
 		xdp_do_flush();
 
-	if (unlikely(!list_empty(list)))
-		cpu_map_bpf_prog_run_skb(rcpu, list, stats);
-
 	bpf_net_ctx_clear(bpf_net_ctx);
-	rcu_read_unlock_bh(); /* resched point, may call do_softirq() */
+	rcu_read_unlock();
 
-	return nframes;
+out:
+	if (unlikely(ret->skb_n) && ret->xdp_n)
+		memmove(&skbs[ret->xdp_n], skbs, ret->skb_n * sizeof(*skbs));
+}
+
+static void cpu_map_gro_flush(struct bpf_cpu_map_entry *rcpu, bool empty)
+{
+	/*
+	 * If the ring is not empty, there'll be a new iteration soon, and we
+	 * only need to do a full flush if a tick is long (> 1 ms).
+	 * If the ring is empty, to not hold GRO packets in the stack for too
+	 * long, do a full flush.
+	 * This is equivalent to how NAPI decides whether to perform a full
+	 * flush.
+	 */
+	gro_flush(&rcpu->gro, !empty && HZ >= 1000);
+	gro_normal_list(&rcpu->gro);
 }
 
 static int cpu_map_kthread_run(void *data)
 {
 	struct bpf_cpu_map_entry *rcpu = data;
 	unsigned long last_qs = jiffies;
+	u32 packets = 0;
 
 	complete(&rcpu->kthread_running);
 	set_current_state(TASK_INTERRUPTIBLE);
@@ -277,11 +303,11 @@ static int cpu_map_kthread_run(void *data)
 	while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
 		struct xdp_cpumap_stats stats = {}; /* zero stats */
 		unsigned int kmem_alloc_drops = 0, sched = 0;
-		gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
-		int i, n, m, nframes, xdp_n;
+		struct cpu_map_ret ret = { };
 		void *frames[CPUMAP_BATCH];
 		void *skbs[CPUMAP_BATCH];
-		LIST_HEAD(list);
+		u32 i, n, m;
+		bool empty;
 
 		/* Release CPU reschedule checks */
 		if (__ptr_ring_empty(rcpu->queue)) {
@@ -306,7 +332,7 @@ static int cpu_map_kthread_run(void *data)
 		 */
 		n = __ptr_ring_consume_batched(rcpu->queue, frames,
 					       CPUMAP_BATCH);
-		for (i = 0, xdp_n = 0; i < n; i++) {
+		for (i = 0; i < n; i++) {
 			void *f = frames[i];
 			struct page *page;
 
@@ -314,11 +340,11 @@ static int cpu_map_kthread_run(void *data)
 				struct sk_buff *skb = f;
 
 				__ptr_clear_bit(0, &skb);
-				list_add_tail(&skb->list, &list);
+				skbs[ret.skb_n++] = skb;
 				continue;
 			}
 
-			frames[xdp_n++] = f;
+			frames[ret.xdp_n++] = f;
 			page = virt_to_page(f);
 
 			/* Bring struct page memory area to curr CPU. Read by
@@ -328,40 +354,51 @@ static int cpu_map_kthread_run(void *data)
 			prefetchw(page);
 		}
 
+		local_bh_disable();
+
 		/* Support running another XDP prog on this CPU */
-		nframes = cpu_map_bpf_prog_run(rcpu, frames, xdp_n, &stats, &list);
-		if (nframes) {
-			m = kmem_cache_alloc_bulk(net_hotdata.skbuff_cache,
-						  gfp, nframes, skbs);
-			if (unlikely(m == 0)) {
-				for (i = 0; i < nframes; i++)
-					skbs[i] = NULL; /* effect: xdp_return_frame */
-				kmem_alloc_drops += nframes;
-			}
+		cpu_map_bpf_prog_run(rcpu, frames, skbs, &ret, &stats);
+		if (!ret.xdp_n)
+			goto stats;
+
+		m = napi_skb_cache_get_bulk(skbs, ret.xdp_n);
+		if (unlikely(m < ret.xdp_n)) {
+			for (i = m; i < ret.xdp_n; i++)
+				xdp_return_frame(frames[i]);
+
+			if (ret.skb_n)
+				memmove(&skbs[m], &skbs[ret.xdp_n],
+					ret.skb_n * sizeof(*skbs));
+
+			kmem_alloc_drops += ret.xdp_n - m;
+			ret.xdp_n = m;
 		}
 
-		local_bh_disable();
-		for (i = 0; i < nframes; i++) {
+		for (i = 0; i < ret.xdp_n; i++) {
 			struct xdp_frame *xdpf = frames[i];
-			struct sk_buff *skb = skbs[i];
 
-			skb = __xdp_build_skb_from_frame(xdpf, skb,
-							 xdpf->dev_rx);
-			if (!skb) {
-				xdp_return_frame(xdpf);
-				continue;
-			}
-
-			list_add_tail(&skb->list, &list);
+			/* Can fail only when !skb -- already handled above */
+			__xdp_build_skb_from_frame(xdpf, skbs[i], xdpf->dev_rx);
 		}
 
+stats:
 		/* Feedback loop via tracepoint.
 		 * NB: keep before recv to allow measuring enqueue/dequeue latency.
 		 */
 		trace_xdp_cpumap_kthread(rcpu->map_id, n, kmem_alloc_drops,
 					 sched, &stats);
 
-		netif_receive_skb_list(&list);
+		for (i = 0; i < ret.xdp_n + ret.skb_n; i++)
+			gro_receive_skb(&rcpu->gro, skbs[i]);
+
+		/* Flush either every 64 packets or in case of empty ring */
+		packets += n;
+		empty = __ptr_ring_empty(rcpu->queue);
+		if (packets >= NAPI_POLL_WEIGHT || empty) {
+			cpu_map_gro_flush(rcpu, empty);
+			packets = 0;
+		}
+
 		local_bh_enable(); /* resched point, may call do_softirq() */
 	}
 	__set_current_state(TASK_RUNNING);
@@ -430,6 +467,7 @@ __cpu_map_entry_alloc(struct bpf_map *map, struct bpf_cpumap_val *value,
 	rcpu->cpu    = cpu;
 	rcpu->map_id = map->id;
 	rcpu->value.qsize  = value->qsize;
+	gro_init(&rcpu->gro);
 
 	if (fd > 0 && __cpu_map_load_bpf_program(rcpu, map, fd))
 		goto free_ptr_ring;
@@ -458,6 +496,7 @@ free_prog:
 	if (rcpu->prog)
 		bpf_prog_put(rcpu->prog);
 free_ptr_ring:
+	gro_cleanup(&rcpu->gro);
 	ptr_ring_cleanup(rcpu->queue, NULL);
 free_queue:
 	kfree(rcpu->queue);
@@ -487,6 +526,7 @@ static void __cpu_map_entry_free(struct work_struct *work)
 
 	if (rcpu->prog)
 		bpf_prog_put(rcpu->prog);
+	gro_cleanup(&rcpu->gro);
 	/* The queue should be empty at this point */
 	__cpu_map_ring_cleanup(rcpu->queue);
 	ptr_ring_cleanup(rcpu->queue, NULL);
diff --git a/net/core/dev.c b/net/core/dev.c
index b0ab0169f507..2dd9e6f490e6 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -6484,7 +6484,7 @@ bool napi_complete_done(struct napi_struct *n, int work_done)
 		return false;
 
 	if (work_done) {
-		if (n->gro_bitmask)
+		if (n->gro.bitmask)
 			timeout = napi_get_gro_flush_timeout(n);
 		n->defer_hard_irqs_count = napi_get_defer_hard_irqs(n);
 	}
@@ -6494,15 +6494,14 @@ bool napi_complete_done(struct napi_struct *n, int work_done)
 		if (timeout)
 			ret = false;
 	}
-	if (n->gro_bitmask) {
-		/* When the NAPI instance uses a timeout and keeps postponing
-		 * it, we need to bound somehow the time packets are kept in
-		 * the GRO layer
-		 */
-		napi_gro_flush(n, !!timeout);
-	}
 
-	gro_normal_list(n);
+	/*
+	 * When the NAPI instance uses a timeout and keeps postponing
+	 * it, we need to bound somehow the time packets are kept in
+	 * the GRO layer.
+	 */
+	gro_flush(&n->gro, !!timeout);
+	gro_normal_list(&n->gro);
 
 	if (unlikely(!list_empty(&n->poll_list))) {
 		/* If n->poll_list is not empty, we need to mask irqs */
@@ -6566,19 +6565,15 @@ static void skb_defer_free_flush(struct softnet_data *sd)
 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule)
 {
 	if (!skip_schedule) {
-		gro_normal_list(napi);
+		gro_normal_list(&napi->gro);
 		__napi_schedule(napi);
 		return;
 	}
 
-	if (napi->gro_bitmask) {
-		/* flush too old packets
-		 * If HZ < 1000, flush all packets.
-		 */
-		napi_gro_flush(napi, HZ >= 1000);
-	}
+	/* Flush too old packets. If HZ < 1000, flush all packets */
+	gro_flush(&napi->gro, HZ >= 1000);
+	gro_normal_list(&napi->gro);
 
-	gro_normal_list(napi);
 	clear_bit(NAPI_STATE_SCHED, &napi->state);
 }
 
@@ -6685,7 +6680,7 @@ restart:
 		}
 		work = napi_poll(napi, budget);
 		trace_napi_poll(napi, work, budget);
-		gro_normal_list(napi);
+		gro_normal_list(&napi->gro);
 count:
 		if (work > 0)
 			__NET_ADD_STATS(dev_net(napi->dev),
@@ -6785,6 +6780,8 @@ void napi_resume_irqs(unsigned int napi_id)
 static void __napi_hash_add_with_id(struct napi_struct *napi,
 				    unsigned int napi_id)
 {
+	napi->gro.cached_napi_id = napi_id;
+
 	WRITE_ONCE(napi->napi_id, napi_id);
 	hlist_add_head_rcu(&napi->napi_hash_node,
 			   &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]);
@@ -6853,17 +6850,6 @@ static enum hrtimer_restart napi_watchdog(struct hrtimer *timer)
 	return HRTIMER_NORESTART;
 }
 
-static void init_gro_hash(struct napi_struct *napi)
-{
-	int i;
-
-	for (i = 0; i < GRO_HASH_BUCKETS; i++) {
-		INIT_LIST_HEAD(&napi->gro_hash[i].list);
-		napi->gro_hash[i].count = 0;
-	}
-	napi->gro_bitmask = 0;
-}
-
 int dev_set_threaded(struct net_device *dev, bool threaded)
 {
 	struct napi_struct *napi;
@@ -7191,10 +7177,8 @@ void netif_napi_add_weight_locked(struct net_device *dev,
 	INIT_HLIST_NODE(&napi->napi_hash_node);
 	hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
 	napi->timer.function = napi_watchdog;
-	init_gro_hash(napi);
+	gro_init(&napi->gro);
 	napi->skb = NULL;
-	INIT_LIST_HEAD(&napi->rx_list);
-	napi->rx_count = 0;
 	napi->poll = poll;
 	if (weight > NAPI_POLL_WEIGHT)
 		netdev_err_once(dev, "%s() called with weight %d\n", __func__,
@@ -7308,19 +7292,6 @@ void napi_enable(struct napi_struct *n)
 }
 EXPORT_SYMBOL(napi_enable);
 
-static void flush_gro_hash(struct napi_struct *napi)
-{
-	int i;
-
-	for (i = 0; i < GRO_HASH_BUCKETS; i++) {
-		struct sk_buff *skb, *n;
-
-		list_for_each_entry_safe(skb, n, &napi->gro_hash[i].list, list)
-			kfree_skb(skb);
-		napi->gro_hash[i].count = 0;
-	}
-}
-
 /* Must be called in process context */
 void __netif_napi_del_locked(struct napi_struct *napi)
 {
@@ -7343,8 +7314,7 @@ void __netif_napi_del_locked(struct napi_struct *napi)
 	list_del_rcu(&napi->dev_list);
 	napi_free_frags(napi);
 
-	flush_gro_hash(napi);
-	napi->gro_bitmask = 0;
+	gro_cleanup(&napi->gro);
 
 	if (napi->thread) {
 		kthread_stop(napi->thread);
@@ -7403,14 +7373,9 @@ static int __napi_poll(struct napi_struct *n, bool *repoll)
 		return work;
 	}
 
-	if (n->gro_bitmask) {
-		/* flush too old packets
-		 * If HZ < 1000, flush all packets.
-		 */
-		napi_gro_flush(n, HZ >= 1000);
-	}
-
-	gro_normal_list(n);
+	/* Flush too old packets. If HZ < 1000, flush all packets */
+	gro_flush(&n->gro, HZ >= 1000);
+	gro_normal_list(&n->gro);
 
 	/* Some drivers may have called napi_schedule
 	 * prior to exhausting their budget.
@@ -12439,7 +12404,7 @@ static struct hlist_head * __net_init netdev_create_hash(void)
 static int __net_init netdev_init(struct net *net)
 {
 	BUILD_BUG_ON(GRO_HASH_BUCKETS >
-		     8 * sizeof_field(struct napi_struct, gro_bitmask));
+		     BITS_PER_BYTE * sizeof_field(struct gro_node, bitmask));
 
 	INIT_LIST_HEAD(&net->dev_base_head);
 
@@ -12804,7 +12769,7 @@ static int __init net_dev_init(void)
 		INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd);
 		spin_lock_init(&sd->defer_lock);
 
-		init_gro_hash(&sd->backlog);
+		gro_init(&sd->backlog.gro);
 		sd->backlog.poll = process_backlog;
 		sd->backlog.weight = weight_p;
 		INIT_LIST_HEAD(&sd->backlog.poll_list);
diff --git a/net/core/gro.c b/net/core/gro.c
index 78b320b63174..19bd4cdaee3a 100644
--- a/net/core/gro.c
+++ b/net/core/gro.c
@@ -250,8 +250,7 @@ int skb_gro_receive_list(struct sk_buff *p, struct sk_buff *skb)
 	return 0;
 }
 
-
-static void napi_gro_complete(struct napi_struct *napi, struct sk_buff *skb)
+static void gro_complete(struct gro_node *gro, struct sk_buff *skb)
 {
 	struct list_head *head = &net_hotdata.offload_base;
 	struct packet_offload *ptype;
@@ -284,43 +283,43 @@ static void napi_gro_complete(struct napi_struct *napi, struct sk_buff *skb)
 	}
 
 out:
-	gro_normal_one(napi, skb, NAPI_GRO_CB(skb)->count);
+	gro_normal_one(gro, skb, NAPI_GRO_CB(skb)->count);
 }
 
-static void __napi_gro_flush_chain(struct napi_struct *napi, u32 index,
-				   bool flush_old)
+static void __gro_flush_chain(struct gro_node *gro, u32 index, bool flush_old)
 {
-	struct list_head *head = &napi->gro_hash[index].list;
+	struct list_head *head = &gro->hash[index].list;
 	struct sk_buff *skb, *p;
 
 	list_for_each_entry_safe_reverse(skb, p, head, list) {
 		if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
 			return;
 		skb_list_del_init(skb);
-		napi_gro_complete(napi, skb);
-		napi->gro_hash[index].count--;
+		gro_complete(gro, skb);
+		gro->hash[index].count--;
 	}
 
-	if (!napi->gro_hash[index].count)
-		__clear_bit(index, &napi->gro_bitmask);
+	if (!gro->hash[index].count)
+		__clear_bit(index, &gro->bitmask);
 }
 
-/* napi->gro_hash[].list contains packets ordered by age.
+/*
+ * gro->hash[].list contains packets ordered by age.
  * youngest packets at the head of it.
  * Complete skbs in reverse order to reduce latencies.
  */
-void napi_gro_flush(struct napi_struct *napi, bool flush_old)
+void __gro_flush(struct gro_node *gro, bool flush_old)
 {
-	unsigned long bitmask = napi->gro_bitmask;
+	unsigned long bitmask = gro->bitmask;
 	unsigned int i, base = ~0U;
 
 	while ((i = ffs(bitmask)) != 0) {
 		bitmask >>= i;
 		base += i;
-		__napi_gro_flush_chain(napi, base, flush_old);
+		__gro_flush_chain(gro, base, flush_old);
 	}
 }
-EXPORT_SYMBOL(napi_gro_flush);
+EXPORT_SYMBOL(__gro_flush);
 
 static unsigned long gro_list_prepare_tc_ext(const struct sk_buff *skb,
 					     const struct sk_buff *p,
@@ -439,7 +438,7 @@ static void gro_try_pull_from_frag0(struct sk_buff *skb)
 		gro_pull_from_frag0(skb, grow);
 }
 
-static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head)
+static void gro_flush_oldest(struct gro_node *gro, struct list_head *head)
 {
 	struct sk_buff *oldest;
 
@@ -455,14 +454,15 @@ static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head)
 	 * SKB to the chain.
 	 */
 	skb_list_del_init(oldest);
-	napi_gro_complete(napi, oldest);
+	gro_complete(gro, oldest);
 }
 
-static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
+static enum gro_result dev_gro_receive(struct gro_node *gro,
+				       struct sk_buff *skb)
 {
 	u32 bucket = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1);
-	struct gro_list *gro_list = &napi->gro_hash[bucket];
 	struct list_head *head = &net_hotdata.offload_base;
+	struct gro_list *gro_list = &gro->hash[bucket];
 	struct packet_offload *ptype;
 	__be16 type = skb->protocol;
 	struct sk_buff *pp = NULL;
@@ -526,7 +526,7 @@ found_ptype:
 
 	if (pp) {
 		skb_list_del_init(pp);
-		napi_gro_complete(napi, pp);
+		gro_complete(gro, pp);
 		gro_list->count--;
 	}
 
@@ -537,7 +537,7 @@ found_ptype:
 		goto normal;
 
 	if (unlikely(gro_list->count >= MAX_GRO_SKBS))
-		gro_flush_oldest(napi, &gro_list->list);
+		gro_flush_oldest(gro, &gro_list->list);
 	else
 		gro_list->count++;
 
@@ -551,10 +551,10 @@ found_ptype:
 	ret = GRO_HELD;
 ok:
 	if (gro_list->count) {
-		if (!test_bit(bucket, &napi->gro_bitmask))
-			__set_bit(bucket, &napi->gro_bitmask);
-	} else if (test_bit(bucket, &napi->gro_bitmask)) {
-		__clear_bit(bucket, &napi->gro_bitmask);
+		if (!test_bit(bucket, &gro->bitmask))
+			__set_bit(bucket, &gro->bitmask);
+	} else if (test_bit(bucket, &gro->bitmask)) {
+		__clear_bit(bucket, &gro->bitmask);
 	}
 
 	return ret;
@@ -593,13 +593,12 @@ struct packet_offload *gro_find_complete_by_type(__be16 type)
 }
 EXPORT_SYMBOL(gro_find_complete_by_type);
 
-static gro_result_t napi_skb_finish(struct napi_struct *napi,
-				    struct sk_buff *skb,
-				    gro_result_t ret)
+static gro_result_t gro_skb_finish(struct gro_node *gro, struct sk_buff *skb,
+				   gro_result_t ret)
 {
 	switch (ret) {
 	case GRO_NORMAL:
-		gro_normal_one(napi, skb, 1);
+		gro_normal_one(gro, skb, 1);
 		break;
 
 	case GRO_MERGED_FREE:
@@ -620,21 +619,21 @@ static gro_result_t napi_skb_finish(struct napi_struct *napi,
 	return ret;
 }
 
-gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
+gro_result_t gro_receive_skb(struct gro_node *gro, struct sk_buff *skb)
 {
 	gro_result_t ret;
 
-	skb_mark_napi_id(skb, napi);
+	__skb_mark_napi_id(skb, gro);
 	trace_napi_gro_receive_entry(skb);
 
 	skb_gro_reset_offset(skb, 0);
 
-	ret = napi_skb_finish(napi, skb, dev_gro_receive(napi, skb));
+	ret = gro_skb_finish(gro, skb, dev_gro_receive(gro, skb));
 	trace_napi_gro_receive_exit(ret);
 
 	return ret;
 }
-EXPORT_SYMBOL(napi_gro_receive);
+EXPORT_SYMBOL(gro_receive_skb);
 
 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
 {
@@ -690,7 +689,7 @@ static gro_result_t napi_frags_finish(struct napi_struct *napi,
 		__skb_push(skb, ETH_HLEN);
 		skb->protocol = eth_type_trans(skb, skb->dev);
 		if (ret == GRO_NORMAL)
-			gro_normal_one(napi, skb, 1);
+			gro_normal_one(&napi->gro, skb, 1);
 		break;
 
 	case GRO_MERGED_FREE:
@@ -759,7 +758,7 @@ gro_result_t napi_gro_frags(struct napi_struct *napi)
 
 	trace_napi_gro_frags_entry(skb);
 
-	ret = napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
+	ret = napi_frags_finish(napi, skb, dev_gro_receive(&napi->gro, skb));
 	trace_napi_gro_frags_exit(ret);
 
 	return ret;
@@ -791,3 +790,37 @@ __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
 	return sum;
 }
 EXPORT_SYMBOL(__skb_gro_checksum_complete);
+
+void gro_init(struct gro_node *gro)
+{
+	for (u32 i = 0; i < GRO_HASH_BUCKETS; i++) {
+		INIT_LIST_HEAD(&gro->hash[i].list);
+		gro->hash[i].count = 0;
+	}
+
+	gro->bitmask = 0;
+	gro->cached_napi_id = 0;
+
+	INIT_LIST_HEAD(&gro->rx_list);
+	gro->rx_count = 0;
+}
+
+void gro_cleanup(struct gro_node *gro)
+{
+	struct sk_buff *skb, *n;
+
+	for (u32 i = 0; i < GRO_HASH_BUCKETS; i++) {
+		list_for_each_entry_safe(skb, n, &gro->hash[i].list, list)
+			kfree_skb(skb);
+
+		gro->hash[i].count = 0;
+	}
+
+	gro->bitmask = 0;
+	gro->cached_napi_id = 0;
+
+	list_for_each_entry_safe(skb, n, &gro->rx_list, list)
+		kfree_skb(skb);
+
+	gro->rx_count = 0;
+}
diff --git a/net/core/skbuff.c b/net/core/skbuff.c
index 5b241c9e6f38..f12815f9c83d 100644
--- a/net/core/skbuff.c
+++ b/net/core/skbuff.c
@@ -295,6 +295,68 @@ static struct sk_buff *napi_skb_cache_get(void)
 	return skb;
 }
 
+/**
+ * napi_skb_cache_get_bulk - obtain a number of zeroed skb heads from the cache
+ * @skbs: pointer to an at least @n-sized array to fill with skb pointers
+ * @n: number of entries to provide
+ *
+ * Tries to obtain @n &sk_buff entries from the NAPI percpu cache and writes
+ * the pointers into the provided array @skbs. If there are less entries
+ * available, tries to replenish the cache and bulk-allocates the diff from
+ * the MM layer if needed.
+ * The heads are being zeroed with either memset() or %__GFP_ZERO, so they are
+ * ready for {,__}build_skb_around() and don't have any data buffers attached.
+ * Must be called *only* from the BH context.
+ *
+ * Return: number of successfully allocated skbs (@n if no actual allocation
+ *	   needed or kmem_cache_alloc_bulk() didn't fail).
+ */
+u32 napi_skb_cache_get_bulk(void **skbs, u32 n)
+{
+	struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
+	u32 bulk, total = n;
+
+	local_lock_nested_bh(&napi_alloc_cache.bh_lock);
+
+	if (nc->skb_count >= n)
+		goto get;
+
+	/* No enough cached skbs. Try refilling the cache first */
+	bulk = min(NAPI_SKB_CACHE_SIZE - nc->skb_count, NAPI_SKB_CACHE_BULK);
+	nc->skb_count += kmem_cache_alloc_bulk(net_hotdata.skbuff_cache,
+					       GFP_ATOMIC | __GFP_NOWARN, bulk,
+					       &nc->skb_cache[nc->skb_count]);
+	if (likely(nc->skb_count >= n))
+		goto get;
+
+	/* Still not enough. Bulk-allocate the missing part directly, zeroed */
+	n -= kmem_cache_alloc_bulk(net_hotdata.skbuff_cache,
+				   GFP_ATOMIC | __GFP_ZERO | __GFP_NOWARN,
+				   n - nc->skb_count, &skbs[nc->skb_count]);
+	if (likely(nc->skb_count >= n))
+		goto get;
+
+	/* kmem_cache didn't allocate the number we need, limit the output */
+	total -= n - nc->skb_count;
+	n = nc->skb_count;
+
+get:
+	for (u32 base = nc->skb_count - n, i = 0; i < n; i++) {
+		u32 cache_size = kmem_cache_size(net_hotdata.skbuff_cache);
+
+		skbs[i] = nc->skb_cache[base + i];
+
+		kasan_mempool_unpoison_object(skbs[i], cache_size);
+		memset(skbs[i], 0, offsetof(struct sk_buff, tail));
+	}
+
+	nc->skb_count -= n;
+	local_unlock_nested_bh(&napi_alloc_cache.bh_lock);
+
+	return total;
+}
+EXPORT_SYMBOL_GPL(napi_skb_cache_get_bulk);
+
 static inline void __finalize_skb_around(struct sk_buff *skb, void *data,
 					 unsigned int size)
 {
diff --git a/net/core/xdp.c b/net/core/xdp.c
index 2c6ab6fb452f..f86eedad586a 100644
--- a/net/core/xdp.c
+++ b/net/core/xdp.c
@@ -618,16 +618,6 @@ void xdp_warn(const char *msg, const char *func, const int line)
 };
 EXPORT_SYMBOL_GPL(xdp_warn);
 
-int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp)
-{
-	n_skb = kmem_cache_alloc_bulk(net_hotdata.skbuff_cache, gfp, n_skb, skbs);
-	if (unlikely(!n_skb))
-		return -ENOMEM;
-
-	return 0;
-}
-EXPORT_SYMBOL_GPL(xdp_alloc_skb_bulk);
-
 /**
  * xdp_build_skb_from_buff - create an skb from &xdp_buff
  * @xdp: &xdp_buff to convert to an skb