Merge branch 'XDP-redirect-memory-return-API'

Jesper Dangaard Brouer says:

====================
XDP redirect memory return API

Submitted against net-next, as it contains NIC driver changes.

This patchset works towards supporting different XDP RX-ring memory
allocators.  As this will be needed by the AF_XDP zero-copy mode.

The patchset uses mlx5 as the sample driver, which gets implemented
XDP_REDIRECT RX-mode, but not ndo_xdp_xmit (as this API is subject to
change thought the patchset).

A new struct xdp_frame is introduced (modeled after cpumap xdp_pkt).
And both ndo_xdp_xmit and the new xdp_return_frame end-up using this.

Support for a driver supplied allocator is implemented, and a
refurbished version of page_pool is the first return allocator type
introduced.  This will be a integration point for AF_XDP zero-copy.

The mlx5 driver evolve into using the page_pool, and see a performance
increase (with ndo_xdp_xmit out ixgbe driver) from 6Mpps to 12Mpps.

The patchset stop at 16 patches (one over limit), but more API changes
are planned.  Specifically extending ndo_xdp_xmit and xdp_return_frame
APIs to support bulking.  As this will address some known limits.

V2: Updated according to Tariq's feedback
V3: Updated based on feedback from Jason Wang and Alex Duyck
V4: Updated based on feedback from Tariq and Jason
V5: Fix SPDX license, add Tariq's reviews, improve patch desc for perf test
V6: Updated based on feedback from Eric Dumazet and Alex Duyck
V7: Adapt to i40e that got XDP_REDIRECT support in-between
V8:
 Updated based on feedback kbuild test robot, and adjust for mlx5 changes
 page_pool only compiled into kernel when drivers Kconfig 'select' feature
V9:
 Remove some inline statements, let compiler decide what to inline
 Fix return value in virtio_net driver
 Adjust for mlx5 changes in-between submissions
V10:
 Minor adjust for mlx5 requested by Tariq
 Resubmit against net-next
V11: avoid leaking info stored in frame data on page reuse
====================

Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

5 files changed, 231 insertions, 27 deletions

diff --git a/include/linux/filter.h b/include/linux/filter.h
index fc4e8f91b03d..4da8b2308174 100644
--- a/include/linux/filter.h
+++ b/include/linux/filter.h
@@ -30,6 +30,7 @@ struct sock;
 struct seccomp_data;
 struct bpf_prog_aux;
 struct xdp_rxq_info;
+struct xdp_buff;
 
 /* ArgX, context and stack frame pointer register positions. Note,
  * Arg1, Arg2, Arg3, etc are used as argument mappings of function
@@ -500,14 +501,6 @@ struct bpf_skb_data_end {
 	void *data_end;
 };
 
-struct xdp_buff {
-	void *data;
-	void *data_end;
-	void *data_meta;
-	void *data_hard_start;
-	struct xdp_rxq_info *rxq;
-};
-
 struct sk_msg_buff {
 	void *data;
 	void *data_end;
@@ -772,21 +765,6 @@ int xdp_do_redirect(struct net_device *dev,
 		    struct bpf_prog *prog);
 void xdp_do_flush_map(void);
 
-/* Drivers not supporting XDP metadata can use this helper, which
- * rejects any room expansion for metadata as a result.
- */
-static __always_inline void
-xdp_set_data_meta_invalid(struct xdp_buff *xdp)
-{
-	xdp->data_meta = xdp->data + 1;
-}
-
-static __always_inline bool
-xdp_data_meta_unsupported(const struct xdp_buff *xdp)
-{
-	return unlikely(xdp->data_meta > xdp->data);
-}
-
 void bpf_warn_invalid_xdp_action(u32 act);
 
 struct sock *do_sk_redirect_map(struct sk_buff *skb);
diff --git a/include/linux/if_tun.h b/include/linux/if_tun.h
index fd00170b494f..3d2996dc7d85 100644
--- a/include/linux/if_tun.h
+++ b/include/linux/if_tun.h
@@ -22,7 +22,7 @@
 #if defined(CONFIG_TUN) || defined(CONFIG_TUN_MODULE)
 struct socket *tun_get_socket(struct file *);
 struct ptr_ring *tun_get_tx_ring(struct file *file);
-bool tun_is_xdp_buff(void *ptr);
+bool tun_is_xdp_frame(void *ptr);
 void *tun_xdp_to_ptr(void *ptr);
 void *tun_ptr_to_xdp(void *ptr);
 void tun_ptr_free(void *ptr);
@@ -39,7 +39,7 @@ static inline struct ptr_ring *tun_get_tx_ring(struct file *f)
 {
 	return ERR_PTR(-EINVAL);
 }
-static inline bool tun_is_xdp_buff(void *ptr)
+static inline bool tun_is_xdp_frame(void *ptr)
 {
 	return false;
 }
diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index cf44503ea81a..14e0777ffcfb 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h
@@ -1165,7 +1165,7 @@ struct dev_ifalias {
  *	This function is used to set or query state related to XDP on the
  *	netdevice and manage BPF offload. See definition of
  *	enum bpf_netdev_command for details.
- * int (*ndo_xdp_xmit)(struct net_device *dev, struct xdp_buff *xdp);
+ * int (*ndo_xdp_xmit)(struct net_device *dev, struct xdp_frame *xdp);
  *	This function is used to submit a XDP packet for transmit on a
  *	netdevice.
  * void (*ndo_xdp_flush)(struct net_device *dev);
@@ -1356,7 +1356,7 @@ struct net_device_ops {
 	int			(*ndo_bpf)(struct net_device *dev,
 					   struct netdev_bpf *bpf);
 	int			(*ndo_xdp_xmit)(struct net_device *dev,
-						struct xdp_buff *xdp);
+						struct xdp_frame *xdp);
 	void			(*ndo_xdp_flush)(struct net_device *dev);
 };
 
diff --git a/include/net/page_pool.h b/include/net/page_pool.h
new file mode 100644
index 000000000000..c79087153148
--- /dev/null
+++ b/include/net/page_pool.h
@@ -0,0 +1,143 @@
+/* SPDX-License-Identifier: GPL-2.0
+ *
+ * page_pool.h
+ *	Author:	Jesper Dangaard Brouer <netoptimizer@brouer.com>
+ *	Copyright (C) 2016 Red Hat, Inc.
+ */
+
+/**
+ * DOC: page_pool allocator
+ *
+ * This page_pool allocator is optimized for the XDP mode that
+ * uses one-frame-per-page, but have fallbacks that act like the
+ * regular page allocator APIs.
+ *
+ * Basic use involve replacing alloc_pages() calls with the
+ * page_pool_alloc_pages() call.  Drivers should likely use
+ * page_pool_dev_alloc_pages() replacing dev_alloc_pages().
+ *
+ * If page_pool handles DMA mapping (use page->private), then API user
+ * is responsible for invoking page_pool_put_page() once.  In-case of
+ * elevated refcnt, the DMA state is released, assuming other users of
+ * the page will eventually call put_page().
+ *
+ * If no DMA mapping is done, then it can act as shim-layer that
+ * fall-through to alloc_page.  As no state is kept on the page, the
+ * regular put_page() call is sufficient.
+ */
+#ifndef _NET_PAGE_POOL_H
+#define _NET_PAGE_POOL_H
+
+#include <linux/mm.h> /* Needed by ptr_ring */
+#include <linux/ptr_ring.h>
+#include <linux/dma-direction.h>
+
+#define PP_FLAG_DMA_MAP 1 /* Should page_pool do the DMA map/unmap */
+#define PP_FLAG_ALL	PP_FLAG_DMA_MAP
+
+/*
+ * Fast allocation side cache array/stack
+ *
+ * The cache size and refill watermark is related to the network
+ * use-case.  The NAPI budget is 64 packets.  After a NAPI poll the RX
+ * ring is usually refilled and the max consumed elements will be 64,
+ * thus a natural max size of objects needed in the cache.
+ *
+ * Keeping room for more objects, is due to XDP_DROP use-case.  As
+ * XDP_DROP allows the opportunity to recycle objects directly into
+ * this array, as it shares the same softirq/NAPI protection.  If
+ * cache is already full (or partly full) then the XDP_DROP recycles
+ * would have to take a slower code path.
+ */
+#define PP_ALLOC_CACHE_SIZE	128
+#define PP_ALLOC_CACHE_REFILL	64
+struct pp_alloc_cache {
+	u32 count;
+	void *cache[PP_ALLOC_CACHE_SIZE];
+};
+
+struct page_pool_params {
+	unsigned int	flags;
+	unsigned int	order;
+	unsigned int	pool_size;
+	int		nid;  /* Numa node id to allocate from pages from */
+	struct device	*dev; /* device, for DMA pre-mapping purposes */
+	enum dma_data_direction dma_dir; /* DMA mapping direction */
+};
+
+struct page_pool {
+	struct rcu_head rcu;
+	struct page_pool_params p;
+
+	/*
+	 * Data structure for allocation side
+	 *
+	 * Drivers allocation side usually already perform some kind
+	 * of resource protection.  Piggyback on this protection, and
+	 * require driver to protect allocation side.
+	 *
+	 * For NIC drivers this means, allocate a page_pool per
+	 * RX-queue. As the RX-queue is already protected by
+	 * Softirq/BH scheduling and napi_schedule. NAPI schedule
+	 * guarantee that a single napi_struct will only be scheduled
+	 * on a single CPU (see napi_schedule).
+	 */
+	struct pp_alloc_cache alloc ____cacheline_aligned_in_smp;
+
+	/* Data structure for storing recycled pages.
+	 *
+	 * Returning/freeing pages is more complicated synchronization
+	 * wise, because free's can happen on remote CPUs, with no
+	 * association with allocation resource.
+	 *
+	 * Use ptr_ring, as it separates consumer and producer
+	 * effeciently, it a way that doesn't bounce cache-lines.
+	 *
+	 * TODO: Implement bulk return pages into this structure.
+	 */
+	struct ptr_ring ring;
+};
+
+struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp);
+
+static inline struct page *page_pool_dev_alloc_pages(struct page_pool *pool)
+{
+	gfp_t gfp = (GFP_ATOMIC | __GFP_NOWARN);
+
+	return page_pool_alloc_pages(pool, gfp);
+}
+
+struct page_pool *page_pool_create(const struct page_pool_params *params);
+
+void page_pool_destroy(struct page_pool *pool);
+
+/* Never call this directly, use helpers below */
+void __page_pool_put_page(struct page_pool *pool,
+			  struct page *page, bool allow_direct);
+
+static inline void page_pool_put_page(struct page_pool *pool, struct page *page)
+{
+	/* When page_pool isn't compiled-in, net/core/xdp.c doesn't
+	 * allow registering MEM_TYPE_PAGE_POOL, but shield linker.
+	 */
+#ifdef CONFIG_PAGE_POOL
+	__page_pool_put_page(pool, page, false);
+#endif
+}
+/* Very limited use-cases allow recycle direct */
+static inline void page_pool_recycle_direct(struct page_pool *pool,
+					    struct page *page)
+{
+	__page_pool_put_page(pool, page, true);
+}
+
+static inline bool is_page_pool_compiled_in(void)
+{
+#ifdef CONFIG_PAGE_POOL
+	return true;
+#else
+	return false;
+#endif
+}
+
+#endif /* _NET_PAGE_POOL_H */
diff --git a/include/net/xdp.h b/include/net/xdp.h
index b2362ddfa694..137ad5f9f40f 100644
--- a/include/net/xdp.h
+++ b/include/net/xdp.h
@@ -33,16 +33,99 @@
  * also mandatory during RX-ring setup.
  */
 
+enum xdp_mem_type {
+	MEM_TYPE_PAGE_SHARED = 0, /* Split-page refcnt based model */
+	MEM_TYPE_PAGE_ORDER0,     /* Orig XDP full page model */
+	MEM_TYPE_PAGE_POOL,
+	MEM_TYPE_MAX,
+};
+
+struct xdp_mem_info {
+	u32 type; /* enum xdp_mem_type, but known size type */
+	u32 id;
+};
+
+struct page_pool;
+
 struct xdp_rxq_info {
 	struct net_device *dev;
 	u32 queue_index;
 	u32 reg_state;
+	struct xdp_mem_info mem;
 } ____cacheline_aligned; /* perf critical, avoid false-sharing */
 
+struct xdp_buff {
+	void *data;
+	void *data_end;
+	void *data_meta;
+	void *data_hard_start;
+	struct xdp_rxq_info *rxq;
+};
+
+struct xdp_frame {
+	void *data;
+	u16 len;
+	u16 headroom;
+	u16 metasize;
+	/* Lifetime of xdp_rxq_info is limited to NAPI/enqueue time,
+	 * while mem info is valid on remote CPU.
+	 */
+	struct xdp_mem_info mem;
+	struct net_device *dev_rx; /* used by cpumap */
+};
+
+/* Convert xdp_buff to xdp_frame */
+static inline
+struct xdp_frame *convert_to_xdp_frame(struct xdp_buff *xdp)
+{
+	struct xdp_frame *xdp_frame;
+	int metasize;
+	int headroom;
+
+	/* Assure headroom is available for storing info */
+	headroom = xdp->data - xdp->data_hard_start;
+	metasize = xdp->data - xdp->data_meta;
+	metasize = metasize > 0 ? metasize : 0;
+	if (unlikely((headroom - metasize) < sizeof(*xdp_frame)))
+		return NULL;
+
+	/* Store info in top of packet */
+	xdp_frame = xdp->data_hard_start;
+
+	xdp_frame->data = xdp->data;
+	xdp_frame->len  = xdp->data_end - xdp->data;
+	xdp_frame->headroom = headroom - sizeof(*xdp_frame);
+	xdp_frame->metasize = metasize;
+
+	/* rxq only valid until napi_schedule ends, convert to xdp_mem_info */
+	xdp_frame->mem = xdp->rxq->mem;
+
+	return xdp_frame;
+}
+
+void xdp_return_frame(struct xdp_frame *xdpf);
+
 int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
 		     struct net_device *dev, u32 queue_index);
 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq);
 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq);
 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq);
+int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
+			       enum xdp_mem_type type, void *allocator);
+
+/* Drivers not supporting XDP metadata can use this helper, which
+ * rejects any room expansion for metadata as a result.
+ */
+static __always_inline void
+xdp_set_data_meta_invalid(struct xdp_buff *xdp)
+{
+	xdp->data_meta = xdp->data + 1;
+}
+
+static __always_inline bool
+xdp_data_meta_unsupported(const struct xdp_buff *xdp)
+{
+	return unlikely(xdp->data_meta > xdp->data);
+}
 
 #endif /* __LINUX_NET_XDP_H__ */