xref: /openbmc/linux/include/net/xdp.h (revision dc023681)
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 	unsigned int napi_id;
63 } ____cacheline_aligned; /* perf critical, avoid false-sharing */
64 
65 struct xdp_txq_info {
66 	struct net_device *dev;
67 };
68 
69 struct xdp_buff {
70 	void *data;
71 	void *data_end;
72 	void *data_meta;
73 	void *data_hard_start;
74 	struct xdp_rxq_info *rxq;
75 	struct xdp_txq_info *txq;
76 	u32 frame_sz; /* frame size to deduce data_hard_end/reserved tailroom*/
77 };
78 
79 static __always_inline void
80 xdp_init_buff(struct xdp_buff *xdp, u32 frame_sz, struct xdp_rxq_info *rxq)
81 {
82 	xdp->frame_sz = frame_sz;
83 	xdp->rxq = rxq;
84 }
85 
86 static __always_inline void
87 xdp_prepare_buff(struct xdp_buff *xdp, unsigned char *hard_start,
88 		 int headroom, int data_len, const bool meta_valid)
89 {
90 	unsigned char *data = hard_start + headroom;
91 
92 	xdp->data_hard_start = hard_start;
93 	xdp->data = data;
94 	xdp->data_end = data + data_len;
95 	xdp->data_meta = meta_valid ? data : data + 1;
96 }
97 
98 /* Reserve memory area at end-of data area.
99  *
100  * This macro reserves tailroom in the XDP buffer by limiting the
101  * XDP/BPF data access to data_hard_end.  Notice same area (and size)
102  * is used for XDP_PASS, when constructing the SKB via build_skb().
103  */
104 #define xdp_data_hard_end(xdp)				\
105 	((xdp)->data_hard_start + (xdp)->frame_sz -	\
106 	 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)))
107 
108 static inline struct skb_shared_info *
109 xdp_get_shared_info_from_buff(struct xdp_buff *xdp)
110 {
111 	return (struct skb_shared_info *)xdp_data_hard_end(xdp);
112 }
113 
114 struct xdp_frame {
115 	void *data;
116 	u16 len;
117 	u16 headroom;
118 	u32 metasize:8;
119 	u32 frame_sz:24;
120 	/* Lifetime of xdp_rxq_info is limited to NAPI/enqueue time,
121 	 * while mem info is valid on remote CPU.
122 	 */
123 	struct xdp_mem_info mem;
124 	struct net_device *dev_rx; /* used by cpumap */
125 };
126 
127 #define XDP_BULK_QUEUE_SIZE	16
128 struct xdp_frame_bulk {
129 	int count;
130 	void *xa;
131 	void *q[XDP_BULK_QUEUE_SIZE];
132 };
133 
134 static __always_inline void xdp_frame_bulk_init(struct xdp_frame_bulk *bq)
135 {
136 	/* bq->count will be zero'ed when bq->xa gets updated */
137 	bq->xa = NULL;
138 }
139 
140 static inline struct skb_shared_info *
141 xdp_get_shared_info_from_frame(struct xdp_frame *frame)
142 {
143 	void *data_hard_start = frame->data - frame->headroom - sizeof(*frame);
144 
145 	return (struct skb_shared_info *)(data_hard_start + frame->frame_sz -
146 				SKB_DATA_ALIGN(sizeof(struct skb_shared_info)));
147 }
148 
149 struct xdp_cpumap_stats {
150 	unsigned int redirect;
151 	unsigned int pass;
152 	unsigned int drop;
153 };
154 
155 /* Clear kernel pointers in xdp_frame */
156 static inline void xdp_scrub_frame(struct xdp_frame *frame)
157 {
158 	frame->data = NULL;
159 	frame->dev_rx = NULL;
160 }
161 
162 /* Avoids inlining WARN macro in fast-path */
163 void xdp_warn(const char *msg, const char *func, const int line);
164 #define XDP_WARN(msg) xdp_warn(msg, __func__, __LINE__)
165 
166 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp);
167 struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf,
168 					   struct sk_buff *skb,
169 					   struct net_device *dev);
170 struct sk_buff *xdp_build_skb_from_frame(struct xdp_frame *xdpf,
171 					 struct net_device *dev);
172 int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp);
173 struct xdp_frame *xdpf_clone(struct xdp_frame *xdpf);
174 
175 static inline
176 void xdp_convert_frame_to_buff(struct xdp_frame *frame, struct xdp_buff *xdp)
177 {
178 	xdp->data_hard_start = frame->data - frame->headroom - sizeof(*frame);
179 	xdp->data = frame->data;
180 	xdp->data_end = frame->data + frame->len;
181 	xdp->data_meta = frame->data - frame->metasize;
182 	xdp->frame_sz = frame->frame_sz;
183 }
184 
185 static inline
186 int xdp_update_frame_from_buff(struct xdp_buff *xdp,
187 			       struct xdp_frame *xdp_frame)
188 {
189 	int metasize, headroom;
190 
191 	/* Assure headroom is available for storing info */
192 	headroom = xdp->data - xdp->data_hard_start;
193 	metasize = xdp->data - xdp->data_meta;
194 	metasize = metasize > 0 ? metasize : 0;
195 	if (unlikely((headroom - metasize) < sizeof(*xdp_frame)))
196 		return -ENOSPC;
197 
198 	/* Catch if driver didn't reserve tailroom for skb_shared_info */
199 	if (unlikely(xdp->data_end > xdp_data_hard_end(xdp))) {
200 		XDP_WARN("Driver BUG: missing reserved tailroom");
201 		return -ENOSPC;
202 	}
203 
204 	xdp_frame->data = xdp->data;
205 	xdp_frame->len  = xdp->data_end - xdp->data;
206 	xdp_frame->headroom = headroom - sizeof(*xdp_frame);
207 	xdp_frame->metasize = metasize;
208 	xdp_frame->frame_sz = xdp->frame_sz;
209 
210 	return 0;
211 }
212 
213 /* Convert xdp_buff to xdp_frame */
214 static inline
215 struct xdp_frame *xdp_convert_buff_to_frame(struct xdp_buff *xdp)
216 {
217 	struct xdp_frame *xdp_frame;
218 
219 	if (xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL)
220 		return xdp_convert_zc_to_xdp_frame(xdp);
221 
222 	/* Store info in top of packet */
223 	xdp_frame = xdp->data_hard_start;
224 	if (unlikely(xdp_update_frame_from_buff(xdp, xdp_frame) < 0))
225 		return NULL;
226 
227 	/* rxq only valid until napi_schedule ends, convert to xdp_mem_info */
228 	xdp_frame->mem = xdp->rxq->mem;
229 
230 	return xdp_frame;
231 }
232 
233 void xdp_return_frame(struct xdp_frame *xdpf);
234 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf);
235 void xdp_return_buff(struct xdp_buff *xdp);
236 void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq);
237 void xdp_return_frame_bulk(struct xdp_frame *xdpf,
238 			   struct xdp_frame_bulk *bq);
239 
240 /* When sending xdp_frame into the network stack, then there is no
241  * return point callback, which is needed to release e.g. DMA-mapping
242  * resources with page_pool.  Thus, have explicit function to release
243  * frame resources.
244  */
245 void __xdp_release_frame(void *data, struct xdp_mem_info *mem);
246 static inline void xdp_release_frame(struct xdp_frame *xdpf)
247 {
248 	struct xdp_mem_info *mem = &xdpf->mem;
249 
250 	/* Curr only page_pool needs this */
251 	if (mem->type == MEM_TYPE_PAGE_POOL)
252 		__xdp_release_frame(xdpf->data, mem);
253 }
254 
255 int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
256 		     struct net_device *dev, u32 queue_index, unsigned int napi_id);
257 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq);
258 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq);
259 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq);
260 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
261 			       enum xdp_mem_type type, void *allocator);
262 void xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq);
263 
264 /* Drivers not supporting XDP metadata can use this helper, which
265  * rejects any room expansion for metadata as a result.
266  */
267 static __always_inline void
268 xdp_set_data_meta_invalid(struct xdp_buff *xdp)
269 {
270 	xdp->data_meta = xdp->data + 1;
271 }
272 
273 static __always_inline bool
274 xdp_data_meta_unsupported(const struct xdp_buff *xdp)
275 {
276 	return unlikely(xdp->data_meta > xdp->data);
277 }
278 
279 static inline bool xdp_metalen_invalid(unsigned long metalen)
280 {
281 	return (metalen & (sizeof(__u32) - 1)) || (metalen > 32);
282 }
283 
284 struct xdp_attachment_info {
285 	struct bpf_prog *prog;
286 	u32 flags;
287 };
288 
289 struct netdev_bpf;
290 void xdp_attachment_setup(struct xdp_attachment_info *info,
291 			  struct netdev_bpf *bpf);
292 
293 #define DEV_MAP_BULK_SIZE XDP_BULK_QUEUE_SIZE
294 
295 #endif /* __LINUX_NET_XDP_H__ */
296