xref: /linux/include/net/xdp.h (revision ca853314e78b0a65c20b6a889a23c31f918d4aa2)
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 /* Reserve memory area at end-of data area.
80  *
81  * This macro reserves tailroom in the XDP buffer by limiting the
82  * XDP/BPF data access to data_hard_end.  Notice same area (and size)
83  * is used for XDP_PASS, when constructing the SKB via build_skb().
84  */
85 #define xdp_data_hard_end(xdp)				\
86 	((xdp)->data_hard_start + (xdp)->frame_sz -	\
87 	 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)))
88 
89 static inline struct skb_shared_info *
90 xdp_get_shared_info_from_buff(struct xdp_buff *xdp)
91 {
92 	return (struct skb_shared_info *)xdp_data_hard_end(xdp);
93 }
94 
95 struct xdp_frame {
96 	void *data;
97 	u16 len;
98 	u16 headroom;
99 	u32 metasize:8;
100 	u32 frame_sz:24;
101 	/* Lifetime of xdp_rxq_info is limited to NAPI/enqueue time,
102 	 * while mem info is valid on remote CPU.
103 	 */
104 	struct xdp_mem_info mem;
105 	struct net_device *dev_rx; /* used by cpumap */
106 };
107 
108 #define XDP_BULK_QUEUE_SIZE	16
109 struct xdp_frame_bulk {
110 	int count;
111 	void *xa;
112 	void *q[XDP_BULK_QUEUE_SIZE];
113 };
114 
115 static __always_inline void xdp_frame_bulk_init(struct xdp_frame_bulk *bq)
116 {
117 	/* bq->count will be zero'ed when bq->xa gets updated */
118 	bq->xa = NULL;
119 }
120 
121 static inline struct skb_shared_info *
122 xdp_get_shared_info_from_frame(struct xdp_frame *frame)
123 {
124 	void *data_hard_start = frame->data - frame->headroom - sizeof(*frame);
125 
126 	return (struct skb_shared_info *)(data_hard_start + frame->frame_sz -
127 				SKB_DATA_ALIGN(sizeof(struct skb_shared_info)));
128 }
129 
130 struct xdp_cpumap_stats {
131 	unsigned int redirect;
132 	unsigned int pass;
133 	unsigned int drop;
134 };
135 
136 /* Clear kernel pointers in xdp_frame */
137 static inline void xdp_scrub_frame(struct xdp_frame *frame)
138 {
139 	frame->data = NULL;
140 	frame->dev_rx = NULL;
141 }
142 
143 /* Avoids inlining WARN macro in fast-path */
144 void xdp_warn(const char *msg, const char *func, const int line);
145 #define XDP_WARN(msg) xdp_warn(msg, __func__, __LINE__)
146 
147 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp);
148 
149 static inline
150 void xdp_convert_frame_to_buff(struct xdp_frame *frame, struct xdp_buff *xdp)
151 {
152 	xdp->data_hard_start = frame->data - frame->headroom - sizeof(*frame);
153 	xdp->data = frame->data;
154 	xdp->data_end = frame->data + frame->len;
155 	xdp->data_meta = frame->data - frame->metasize;
156 	xdp->frame_sz = frame->frame_sz;
157 }
158 
159 static inline
160 int xdp_update_frame_from_buff(struct xdp_buff *xdp,
161 			       struct xdp_frame *xdp_frame)
162 {
163 	int metasize, headroom;
164 
165 	/* Assure headroom is available for storing info */
166 	headroom = xdp->data - xdp->data_hard_start;
167 	metasize = xdp->data - xdp->data_meta;
168 	metasize = metasize > 0 ? metasize : 0;
169 	if (unlikely((headroom - metasize) < sizeof(*xdp_frame)))
170 		return -ENOSPC;
171 
172 	/* Catch if driver didn't reserve tailroom for skb_shared_info */
173 	if (unlikely(xdp->data_end > xdp_data_hard_end(xdp))) {
174 		XDP_WARN("Driver BUG: missing reserved tailroom");
175 		return -ENOSPC;
176 	}
177 
178 	xdp_frame->data = xdp->data;
179 	xdp_frame->len  = xdp->data_end - xdp->data;
180 	xdp_frame->headroom = headroom - sizeof(*xdp_frame);
181 	xdp_frame->metasize = metasize;
182 	xdp_frame->frame_sz = xdp->frame_sz;
183 
184 	return 0;
185 }
186 
187 /* Convert xdp_buff to xdp_frame */
188 static inline
189 struct xdp_frame *xdp_convert_buff_to_frame(struct xdp_buff *xdp)
190 {
191 	struct xdp_frame *xdp_frame;
192 
193 	if (xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL)
194 		return xdp_convert_zc_to_xdp_frame(xdp);
195 
196 	/* Store info in top of packet */
197 	xdp_frame = xdp->data_hard_start;
198 	if (unlikely(xdp_update_frame_from_buff(xdp, xdp_frame) < 0))
199 		return NULL;
200 
201 	/* rxq only valid until napi_schedule ends, convert to xdp_mem_info */
202 	xdp_frame->mem = xdp->rxq->mem;
203 
204 	return xdp_frame;
205 }
206 
207 void xdp_return_frame(struct xdp_frame *xdpf);
208 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf);
209 void xdp_return_buff(struct xdp_buff *xdp);
210 void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq);
211 void xdp_return_frame_bulk(struct xdp_frame *xdpf,
212 			   struct xdp_frame_bulk *bq);
213 
214 /* When sending xdp_frame into the network stack, then there is no
215  * return point callback, which is needed to release e.g. DMA-mapping
216  * resources with page_pool.  Thus, have explicit function to release
217  * frame resources.
218  */
219 void __xdp_release_frame(void *data, struct xdp_mem_info *mem);
220 static inline void xdp_release_frame(struct xdp_frame *xdpf)
221 {
222 	struct xdp_mem_info *mem = &xdpf->mem;
223 
224 	/* Curr only page_pool needs this */
225 	if (mem->type == MEM_TYPE_PAGE_POOL)
226 		__xdp_release_frame(xdpf->data, mem);
227 }
228 
229 int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
230 		     struct net_device *dev, u32 queue_index, unsigned int napi_id);
231 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq);
232 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq);
233 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq);
234 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
235 			       enum xdp_mem_type type, void *allocator);
236 void xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq);
237 
238 /* Drivers not supporting XDP metadata can use this helper, which
239  * rejects any room expansion for metadata as a result.
240  */
241 static __always_inline void
242 xdp_set_data_meta_invalid(struct xdp_buff *xdp)
243 {
244 	xdp->data_meta = xdp->data + 1;
245 }
246 
247 static __always_inline bool
248 xdp_data_meta_unsupported(const struct xdp_buff *xdp)
249 {
250 	return unlikely(xdp->data_meta > xdp->data);
251 }
252 
253 struct xdp_attachment_info {
254 	struct bpf_prog *prog;
255 	u32 flags;
256 };
257 
258 struct netdev_bpf;
259 void xdp_attachment_setup(struct xdp_attachment_info *info,
260 			  struct netdev_bpf *bpf);
261 
262 #define DEV_MAP_BULK_SIZE XDP_BULK_QUEUE_SIZE
263 
264 #endif /* __LINUX_NET_XDP_H__ */
265