xref: /linux/include/net/xdp.h (revision fbf46565c67c626849c7ce2a326972d3008d2a91)
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 carries 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 are 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 unregister (which
34  * also applies 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 	u32 frag_size;
64 } ____cacheline_aligned; /* perf critical, avoid false-sharing */
65 
66 struct xdp_txq_info {
67 	struct net_device *dev;
68 };
69 
70 enum xdp_buff_flags {
71 	XDP_FLAGS_HAS_FRAGS		= BIT(0), /* non-linear xdp buff */
72 	XDP_FLAGS_FRAGS_PF_MEMALLOC	= BIT(1), /* xdp paged memory is under
73 						   * pressure
74 						   */
75 };
76 
77 struct xdp_buff {
78 	void *data;
79 	void *data_end;
80 	void *data_meta;
81 	void *data_hard_start;
82 	struct xdp_rxq_info *rxq;
83 	struct xdp_txq_info *txq;
84 	u32 frame_sz; /* frame size to deduce data_hard_end/reserved tailroom*/
85 	u32 flags; /* supported values defined in xdp_buff_flags */
86 };
87 
88 static __always_inline bool xdp_buff_has_frags(struct xdp_buff *xdp)
89 {
90 	return !!(xdp->flags & XDP_FLAGS_HAS_FRAGS);
91 }
92 
93 static __always_inline void xdp_buff_set_frags_flag(struct xdp_buff *xdp)
94 {
95 	xdp->flags |= XDP_FLAGS_HAS_FRAGS;
96 }
97 
98 static __always_inline void xdp_buff_clear_frags_flag(struct xdp_buff *xdp)
99 {
100 	xdp->flags &= ~XDP_FLAGS_HAS_FRAGS;
101 }
102 
103 static __always_inline bool xdp_buff_is_frag_pfmemalloc(struct xdp_buff *xdp)
104 {
105 	return !!(xdp->flags & XDP_FLAGS_FRAGS_PF_MEMALLOC);
106 }
107 
108 static __always_inline void xdp_buff_set_frag_pfmemalloc(struct xdp_buff *xdp)
109 {
110 	xdp->flags |= XDP_FLAGS_FRAGS_PF_MEMALLOC;
111 }
112 
113 static __always_inline void
114 xdp_init_buff(struct xdp_buff *xdp, u32 frame_sz, struct xdp_rxq_info *rxq)
115 {
116 	xdp->frame_sz = frame_sz;
117 	xdp->rxq = rxq;
118 	xdp->flags = 0;
119 }
120 
121 static __always_inline void
122 xdp_prepare_buff(struct xdp_buff *xdp, unsigned char *hard_start,
123 		 int headroom, int data_len, const bool meta_valid)
124 {
125 	unsigned char *data = hard_start + headroom;
126 
127 	xdp->data_hard_start = hard_start;
128 	xdp->data = data;
129 	xdp->data_end = data + data_len;
130 	xdp->data_meta = meta_valid ? data : data + 1;
131 }
132 
133 /* Reserve memory area at end-of data area.
134  *
135  * This macro reserves tailroom in the XDP buffer by limiting the
136  * XDP/BPF data access to data_hard_end.  Notice same area (and size)
137  * is used for XDP_PASS, when constructing the SKB via build_skb().
138  */
139 #define xdp_data_hard_end(xdp)				\
140 	((xdp)->data_hard_start + (xdp)->frame_sz -	\
141 	 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)))
142 
143 static inline struct skb_shared_info *
144 xdp_get_shared_info_from_buff(struct xdp_buff *xdp)
145 {
146 	return (struct skb_shared_info *)xdp_data_hard_end(xdp);
147 }
148 
149 static __always_inline unsigned int xdp_get_buff_len(struct xdp_buff *xdp)
150 {
151 	unsigned int len = xdp->data_end - xdp->data;
152 	struct skb_shared_info *sinfo;
153 
154 	if (likely(!xdp_buff_has_frags(xdp)))
155 		goto out;
156 
157 	sinfo = xdp_get_shared_info_from_buff(xdp);
158 	len += sinfo->xdp_frags_size;
159 out:
160 	return len;
161 }
162 
163 struct xdp_frame {
164 	void *data;
165 	u16 len;
166 	u16 headroom;
167 	u32 metasize; /* uses lower 8-bits */
168 	/* Lifetime of xdp_rxq_info is limited to NAPI/enqueue time,
169 	 * while mem info is valid on remote CPU.
170 	 */
171 	struct xdp_mem_info mem;
172 	struct net_device *dev_rx; /* used by cpumap */
173 	u32 frame_sz;
174 	u32 flags; /* supported values defined in xdp_buff_flags */
175 };
176 
177 static __always_inline bool xdp_frame_has_frags(struct xdp_frame *frame)
178 {
179 	return !!(frame->flags & XDP_FLAGS_HAS_FRAGS);
180 }
181 
182 static __always_inline bool xdp_frame_is_frag_pfmemalloc(struct xdp_frame *frame)
183 {
184 	return !!(frame->flags & XDP_FLAGS_FRAGS_PF_MEMALLOC);
185 }
186 
187 #define XDP_BULK_QUEUE_SIZE	16
188 struct xdp_frame_bulk {
189 	int count;
190 	void *xa;
191 	void *q[XDP_BULK_QUEUE_SIZE];
192 };
193 
194 static __always_inline void xdp_frame_bulk_init(struct xdp_frame_bulk *bq)
195 {
196 	/* bq->count will be zero'ed when bq->xa gets updated */
197 	bq->xa = NULL;
198 }
199 
200 static inline struct skb_shared_info *
201 xdp_get_shared_info_from_frame(struct xdp_frame *frame)
202 {
203 	void *data_hard_start = frame->data - frame->headroom - sizeof(*frame);
204 
205 	return (struct skb_shared_info *)(data_hard_start + frame->frame_sz -
206 				SKB_DATA_ALIGN(sizeof(struct skb_shared_info)));
207 }
208 
209 struct xdp_cpumap_stats {
210 	unsigned int redirect;
211 	unsigned int pass;
212 	unsigned int drop;
213 };
214 
215 /* Clear kernel pointers in xdp_frame */
216 static inline void xdp_scrub_frame(struct xdp_frame *frame)
217 {
218 	frame->data = NULL;
219 	frame->dev_rx = NULL;
220 }
221 
222 static inline void
223 xdp_update_skb_shared_info(struct sk_buff *skb, u8 nr_frags,
224 			   unsigned int size, unsigned int truesize,
225 			   bool pfmemalloc)
226 {
227 	skb_shinfo(skb)->nr_frags = nr_frags;
228 
229 	skb->len += size;
230 	skb->data_len += size;
231 	skb->truesize += truesize;
232 	skb->pfmemalloc |= pfmemalloc;
233 }
234 
235 /* Avoids inlining WARN macro in fast-path */
236 void xdp_warn(const char *msg, const char *func, const int line);
237 #define XDP_WARN(msg) xdp_warn(msg, __func__, __LINE__)
238 
239 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp);
240 struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf,
241 					   struct sk_buff *skb,
242 					   struct net_device *dev);
243 struct sk_buff *xdp_build_skb_from_frame(struct xdp_frame *xdpf,
244 					 struct net_device *dev);
245 int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp);
246 struct xdp_frame *xdpf_clone(struct xdp_frame *xdpf);
247 
248 static inline
249 void xdp_convert_frame_to_buff(struct xdp_frame *frame, struct xdp_buff *xdp)
250 {
251 	xdp->data_hard_start = frame->data - frame->headroom - sizeof(*frame);
252 	xdp->data = frame->data;
253 	xdp->data_end = frame->data + frame->len;
254 	xdp->data_meta = frame->data - frame->metasize;
255 	xdp->frame_sz = frame->frame_sz;
256 	xdp->flags = frame->flags;
257 }
258 
259 static inline
260 int xdp_update_frame_from_buff(struct xdp_buff *xdp,
261 			       struct xdp_frame *xdp_frame)
262 {
263 	int metasize, headroom;
264 
265 	/* Assure headroom is available for storing info */
266 	headroom = xdp->data - xdp->data_hard_start;
267 	metasize = xdp->data - xdp->data_meta;
268 	metasize = metasize > 0 ? metasize : 0;
269 	if (unlikely((headroom - metasize) < sizeof(*xdp_frame)))
270 		return -ENOSPC;
271 
272 	/* Catch if driver didn't reserve tailroom for skb_shared_info */
273 	if (unlikely(xdp->data_end > xdp_data_hard_end(xdp))) {
274 		XDP_WARN("Driver BUG: missing reserved tailroom");
275 		return -ENOSPC;
276 	}
277 
278 	xdp_frame->data = xdp->data;
279 	xdp_frame->len  = xdp->data_end - xdp->data;
280 	xdp_frame->headroom = headroom - sizeof(*xdp_frame);
281 	xdp_frame->metasize = metasize;
282 	xdp_frame->frame_sz = xdp->frame_sz;
283 	xdp_frame->flags = xdp->flags;
284 
285 	return 0;
286 }
287 
288 /* Convert xdp_buff to xdp_frame */
289 static inline
290 struct xdp_frame *xdp_convert_buff_to_frame(struct xdp_buff *xdp)
291 {
292 	struct xdp_frame *xdp_frame;
293 
294 	if (xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL)
295 		return xdp_convert_zc_to_xdp_frame(xdp);
296 
297 	/* Store info in top of packet */
298 	xdp_frame = xdp->data_hard_start;
299 	if (unlikely(xdp_update_frame_from_buff(xdp, xdp_frame) < 0))
300 		return NULL;
301 
302 	/* rxq only valid until napi_schedule ends, convert to xdp_mem_info */
303 	xdp_frame->mem = xdp->rxq->mem;
304 
305 	return xdp_frame;
306 }
307 
308 void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct,
309 		  struct xdp_buff *xdp);
310 void xdp_return_frame(struct xdp_frame *xdpf);
311 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf);
312 void xdp_return_buff(struct xdp_buff *xdp);
313 void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq);
314 void xdp_return_frame_bulk(struct xdp_frame *xdpf,
315 			   struct xdp_frame_bulk *bq);
316 
317 /* When sending xdp_frame into the network stack, then there is no
318  * return point callback, which is needed to release e.g. DMA-mapping
319  * resources with page_pool.  Thus, have explicit function to release
320  * frame resources.
321  */
322 void __xdp_release_frame(void *data, struct xdp_mem_info *mem);
323 static inline void xdp_release_frame(struct xdp_frame *xdpf)
324 {
325 	struct xdp_mem_info *mem = &xdpf->mem;
326 	struct skb_shared_info *sinfo;
327 	int i;
328 
329 	/* Curr only page_pool needs this */
330 	if (mem->type != MEM_TYPE_PAGE_POOL)
331 		return;
332 
333 	if (likely(!xdp_frame_has_frags(xdpf)))
334 		goto out;
335 
336 	sinfo = xdp_get_shared_info_from_frame(xdpf);
337 	for (i = 0; i < sinfo->nr_frags; i++) {
338 		struct page *page = skb_frag_page(&sinfo->frags[i]);
339 
340 		__xdp_release_frame(page_address(page), mem);
341 	}
342 out:
343 	__xdp_release_frame(xdpf->data, mem);
344 }
345 
346 static __always_inline unsigned int xdp_get_frame_len(struct xdp_frame *xdpf)
347 {
348 	struct skb_shared_info *sinfo;
349 	unsigned int len = xdpf->len;
350 
351 	if (likely(!xdp_frame_has_frags(xdpf)))
352 		goto out;
353 
354 	sinfo = xdp_get_shared_info_from_frame(xdpf);
355 	len += sinfo->xdp_frags_size;
356 out:
357 	return len;
358 }
359 
360 int __xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
361 		       struct net_device *dev, u32 queue_index,
362 		       unsigned int napi_id, u32 frag_size);
363 static inline int
364 xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
365 		 struct net_device *dev, u32 queue_index,
366 		 unsigned int napi_id)
367 {
368 	return __xdp_rxq_info_reg(xdp_rxq, dev, queue_index, napi_id, 0);
369 }
370 
371 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq);
372 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq);
373 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq);
374 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
375 			       enum xdp_mem_type type, void *allocator);
376 void xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq);
377 int xdp_reg_mem_model(struct xdp_mem_info *mem,
378 		      enum xdp_mem_type type, void *allocator);
379 void xdp_unreg_mem_model(struct xdp_mem_info *mem);
380 
381 /* Drivers not supporting XDP metadata can use this helper, which
382  * rejects any room expansion for metadata as a result.
383  */
384 static __always_inline void
385 xdp_set_data_meta_invalid(struct xdp_buff *xdp)
386 {
387 	xdp->data_meta = xdp->data + 1;
388 }
389 
390 static __always_inline bool
391 xdp_data_meta_unsupported(const struct xdp_buff *xdp)
392 {
393 	return unlikely(xdp->data_meta > xdp->data);
394 }
395 
396 static inline bool xdp_metalen_invalid(unsigned long metalen)
397 {
398 	return (metalen & (sizeof(__u32) - 1)) || (metalen > 32);
399 }
400 
401 struct xdp_attachment_info {
402 	struct bpf_prog *prog;
403 	u32 flags;
404 };
405 
406 struct netdev_bpf;
407 void xdp_attachment_setup(struct xdp_attachment_info *info,
408 			  struct netdev_bpf *bpf);
409 
410 #define DEV_MAP_BULK_SIZE XDP_BULK_QUEUE_SIZE
411 
412 #endif /* __LINUX_NET_XDP_H__ */
413