xref: /linux/include/net/xdp.h (revision ed30aef3c864f99111e16d4ea5cf29488d99a278)
1 /* SPDX-License-Identifier: GPL-2.0-only */
2 /* include/net/xdp.h
3  *
4  * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
5  */
6 #ifndef __LINUX_NET_XDP_H__
7 #define __LINUX_NET_XDP_H__
8 
9 #include <linux/skbuff.h> /* skb_shared_info */
10 
11 /**
12  * DOC: XDP RX-queue information
13  *
14  * The XDP RX-queue info (xdp_rxq_info) is associated with the driver
15  * level RX-ring queues.  It is information that is specific to how
16  * the driver have configured a given RX-ring queue.
17  *
18  * Each xdp_buff frame received in the driver carry a (pointer)
19  * reference to this xdp_rxq_info structure.  This provides the XDP
20  * data-path read-access to RX-info for both kernel and bpf-side
21  * (limited subset).
22  *
23  * For now, direct access is only safe while running in NAPI/softirq
24  * context.  Contents is read-mostly and must not be updated during
25  * driver NAPI/softirq poll.
26  *
27  * The driver usage API is a register and unregister API.
28  *
29  * The struct is not directly tied to the XDP prog.  A new XDP prog
30  * can be attached as long as it doesn't change the underlying
31  * RX-ring.  If the RX-ring does change significantly, the NIC driver
32  * naturally need to stop the RX-ring before purging and reallocating
33  * memory.  In that process the driver MUST call unregistor (which
34  * also apply for driver shutdown and unload).  The register API is
35  * also mandatory during RX-ring setup.
36  */
37 
38 enum xdp_mem_type {
39 	MEM_TYPE_PAGE_SHARED = 0, /* Split-page refcnt based model */
40 	MEM_TYPE_PAGE_ORDER0,     /* Orig XDP full page model */
41 	MEM_TYPE_PAGE_POOL,
42 	MEM_TYPE_XSK_BUFF_POOL,
43 	MEM_TYPE_MAX,
44 };
45 
46 /* XDP flags for ndo_xdp_xmit */
47 #define XDP_XMIT_FLUSH		(1U << 0)	/* doorbell signal consumer */
48 #define XDP_XMIT_FLAGS_MASK	XDP_XMIT_FLUSH
49 
50 struct xdp_mem_info {
51 	u32 type; /* enum xdp_mem_type, but known size type */
52 	u32 id;
53 };
54 
55 struct page_pool;
56 
57 struct xdp_rxq_info {
58 	struct net_device *dev;
59 	u32 queue_index;
60 	u32 reg_state;
61 	struct xdp_mem_info mem;
62 } ____cacheline_aligned; /* perf critical, avoid false-sharing */
63 
64 struct xdp_txq_info {
65 	struct net_device *dev;
66 };
67 
68 struct xdp_buff {
69 	void *data;
70 	void *data_end;
71 	void *data_meta;
72 	void *data_hard_start;
73 	struct xdp_rxq_info *rxq;
74 	struct xdp_txq_info *txq;
75 	u32 frame_sz; /* frame size to deduce data_hard_end/reserved tailroom*/
76 };
77 
78 /* Reserve memory area at end-of data area.
79  *
80  * This macro reserves tailroom in the XDP buffer by limiting the
81  * XDP/BPF data access to data_hard_end.  Notice same area (and size)
82  * is used for XDP_PASS, when constructing the SKB via build_skb().
83  */
84 #define xdp_data_hard_end(xdp)				\
85 	((xdp)->data_hard_start + (xdp)->frame_sz -	\
86 	 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)))
87 
88 static inline struct skb_shared_info *
89 xdp_get_shared_info_from_buff(struct xdp_buff *xdp)
90 {
91 	return (struct skb_shared_info *)xdp_data_hard_end(xdp);
92 }
93 
94 struct xdp_frame {
95 	void *data;
96 	u16 len;
97 	u16 headroom;
98 	u32 metasize:8;
99 	u32 frame_sz:24;
100 	/* Lifetime of xdp_rxq_info is limited to NAPI/enqueue time,
101 	 * while mem info is valid on remote CPU.
102 	 */
103 	struct xdp_mem_info mem;
104 	struct net_device *dev_rx; /* used by cpumap */
105 };
106 
107 #define XDP_BULK_QUEUE_SIZE	16
108 struct xdp_frame_bulk {
109 	int count;
110 	void *xa;
111 	void *q[XDP_BULK_QUEUE_SIZE];
112 };
113 
114 static __always_inline void xdp_frame_bulk_init(struct xdp_frame_bulk *bq)
115 {
116 	/* bq->count will be zero'ed when bq->xa gets updated */
117 	bq->xa = NULL;
118 }
119 
120 static inline struct skb_shared_info *
121 xdp_get_shared_info_from_frame(struct xdp_frame *frame)
122 {
123 	void *data_hard_start = frame->data - frame->headroom - sizeof(*frame);
124 
125 	return (struct skb_shared_info *)(data_hard_start + frame->frame_sz -
126 				SKB_DATA_ALIGN(sizeof(struct skb_shared_info)));
127 }
128 
129 struct xdp_cpumap_stats {
130 	unsigned int redirect;
131 	unsigned int pass;
132 	unsigned int drop;
133 };
134 
135 /* Clear kernel pointers in xdp_frame */
136 static inline void xdp_scrub_frame(struct xdp_frame *frame)
137 {
138 	frame->data = NULL;
139 	frame->dev_rx = NULL;
140 }
141 
142 /* Avoids inlining WARN macro in fast-path */
143 void xdp_warn(const char *msg, const char *func, const int line);
144 #define XDP_WARN(msg) xdp_warn(msg, __func__, __LINE__)
145 
146 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp);
147 
148 static inline
149 void xdp_convert_frame_to_buff(struct xdp_frame *frame, struct xdp_buff *xdp)
150 {
151 	xdp->data_hard_start = frame->data - frame->headroom - sizeof(*frame);
152 	xdp->data = frame->data;
153 	xdp->data_end = frame->data + frame->len;
154 	xdp->data_meta = frame->data - frame->metasize;
155 	xdp->frame_sz = frame->frame_sz;
156 }
157 
158 static inline
159 int xdp_update_frame_from_buff(struct xdp_buff *xdp,
160 			       struct xdp_frame *xdp_frame)
161 {
162 	int metasize, headroom;
163 
164 	/* Assure headroom is available for storing info */
165 	headroom = xdp->data - xdp->data_hard_start;
166 	metasize = xdp->data - xdp->data_meta;
167 	metasize = metasize > 0 ? metasize : 0;
168 	if (unlikely((headroom - metasize) < sizeof(*xdp_frame)))
169 		return -ENOSPC;
170 
171 	/* Catch if driver didn't reserve tailroom for skb_shared_info */
172 	if (unlikely(xdp->data_end > xdp_data_hard_end(xdp))) {
173 		XDP_WARN("Driver BUG: missing reserved tailroom");
174 		return -ENOSPC;
175 	}
176 
177 	xdp_frame->data = xdp->data;
178 	xdp_frame->len  = xdp->data_end - xdp->data;
179 	xdp_frame->headroom = headroom - sizeof(*xdp_frame);
180 	xdp_frame->metasize = metasize;
181 	xdp_frame->frame_sz = xdp->frame_sz;
182 
183 	return 0;
184 }
185 
186 /* Convert xdp_buff to xdp_frame */
187 static inline
188 struct xdp_frame *xdp_convert_buff_to_frame(struct xdp_buff *xdp)
189 {
190 	struct xdp_frame *xdp_frame;
191 
192 	if (xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL)
193 		return xdp_convert_zc_to_xdp_frame(xdp);
194 
195 	/* Store info in top of packet */
196 	xdp_frame = xdp->data_hard_start;
197 	if (unlikely(xdp_update_frame_from_buff(xdp, xdp_frame) < 0))
198 		return NULL;
199 
200 	/* rxq only valid until napi_schedule ends, convert to xdp_mem_info */
201 	xdp_frame->mem = xdp->rxq->mem;
202 
203 	return xdp_frame;
204 }
205 
206 void xdp_return_frame(struct xdp_frame *xdpf);
207 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf);
208 void xdp_return_buff(struct xdp_buff *xdp);
209 void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq);
210 void xdp_return_frame_bulk(struct xdp_frame *xdpf,
211 			   struct xdp_frame_bulk *bq);
212 
213 /* When sending xdp_frame into the network stack, then there is no
214  * return point callback, which is needed to release e.g. DMA-mapping
215  * resources with page_pool.  Thus, have explicit function to release
216  * frame resources.
217  */
218 void __xdp_release_frame(void *data, struct xdp_mem_info *mem);
219 static inline void xdp_release_frame(struct xdp_frame *xdpf)
220 {
221 	struct xdp_mem_info *mem = &xdpf->mem;
222 
223 	/* Curr only page_pool needs this */
224 	if (mem->type == MEM_TYPE_PAGE_POOL)
225 		__xdp_release_frame(xdpf->data, mem);
226 }
227 
228 int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
229 		     struct net_device *dev, u32 queue_index);
230 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq);
231 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq);
232 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq);
233 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
234 			       enum xdp_mem_type type, void *allocator);
235 void xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq);
236 
237 /* Drivers not supporting XDP metadata can use this helper, which
238  * rejects any room expansion for metadata as a result.
239  */
240 static __always_inline void
241 xdp_set_data_meta_invalid(struct xdp_buff *xdp)
242 {
243 	xdp->data_meta = xdp->data + 1;
244 }
245 
246 static __always_inline bool
247 xdp_data_meta_unsupported(const struct xdp_buff *xdp)
248 {
249 	return unlikely(xdp->data_meta > xdp->data);
250 }
251 
252 struct xdp_attachment_info {
253 	struct bpf_prog *prog;
254 	u32 flags;
255 };
256 
257 struct netdev_bpf;
258 bool xdp_attachment_flags_ok(struct xdp_attachment_info *info,
259 			     struct netdev_bpf *bpf);
260 void xdp_attachment_setup(struct xdp_attachment_info *info,
261 			  struct netdev_bpf *bpf);
262 
263 #define DEV_MAP_BULK_SIZE XDP_BULK_QUEUE_SIZE
264 
265 #endif /* __LINUX_NET_XDP_H__ */
266