1 /* SPDX-License-Identifier: GPL-2.0-only */
4 * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
6 #ifndef __LINUX_NET_XDP_H__
7 #define __LINUX_NET_XDP_H__
9 #include <linux/bitfield.h>
10 #include <linux/filter.h>
11 #include <linux/netdevice.h>
12 #include <linux/skbuff.h> /* skb_shared_info */
15 * DOC: XDP RX-queue information
17 * The XDP RX-queue info (xdp_rxq_info) is associated with the driver
18 * level RX-ring queues. It is information that is specific to how
19 * the driver have configured a given RX-ring queue.
21 * Each xdp_buff frame received in the driver carries a (pointer)
22 * reference to this xdp_rxq_info structure. This provides the XDP
23 * data-path read-access to RX-info for both kernel and bpf-side
26 * For now, direct access is only safe while running in NAPI/softirq
27 * context. Contents are read-mostly and must not be updated during
28 * driver NAPI/softirq poll.
30 * The driver usage API is a register and unregister API.
32 * The struct is not directly tied to the XDP prog. A new XDP prog
33 * can be attached as long as it doesn't change the underlying
34 * RX-ring. If the RX-ring does change significantly, the NIC driver
35 * naturally need to stop the RX-ring before purging and reallocating
36 * memory. In that process the driver MUST call unregister (which
37 * also applies for driver shutdown and unload). The register API is
38 * also mandatory during RX-ring setup.
42 MEM_TYPE_PAGE_SHARED = 0, /* Split-page refcnt based model */
43 MEM_TYPE_PAGE_ORDER0, /* Orig XDP full page model */
45 MEM_TYPE_XSK_BUFF_POOL,
49 /* XDP flags for ndo_xdp_xmit */
50 #define XDP_XMIT_FLUSH (1U << 0) /* doorbell signal consumer */
51 #define XDP_XMIT_FLAGS_MASK XDP_XMIT_FLUSH
54 u32 type; /* enum xdp_mem_type, but known size type */
61 struct net_device *dev;
64 struct xdp_mem_info mem;
67 } ____cacheline_aligned; /* perf critical, avoid false-sharing */
70 struct net_device *dev;
74 XDP_FLAGS_HAS_FRAGS = BIT(0), /* non-linear xdp buff */
75 XDP_FLAGS_FRAGS_PF_MEMALLOC = BIT(1), /* xdp paged memory is under
84 void *data_hard_start;
85 struct xdp_rxq_info *rxq;
86 struct xdp_txq_info *txq;
87 u32 frame_sz; /* frame size to deduce data_hard_end/reserved tailroom*/
88 u32 flags; /* supported values defined in xdp_buff_flags */
91 static __always_inline bool xdp_buff_has_frags(struct xdp_buff *xdp)
93 return !!(xdp->flags & XDP_FLAGS_HAS_FRAGS);
96 static __always_inline void xdp_buff_set_frags_flag(struct xdp_buff *xdp)
98 xdp->flags |= XDP_FLAGS_HAS_FRAGS;
101 static __always_inline void xdp_buff_clear_frags_flag(struct xdp_buff *xdp)
103 xdp->flags &= ~XDP_FLAGS_HAS_FRAGS;
106 static __always_inline bool xdp_buff_is_frag_pfmemalloc(struct xdp_buff *xdp)
108 return !!(xdp->flags & XDP_FLAGS_FRAGS_PF_MEMALLOC);
111 static __always_inline void xdp_buff_set_frag_pfmemalloc(struct xdp_buff *xdp)
113 xdp->flags |= XDP_FLAGS_FRAGS_PF_MEMALLOC;
116 static __always_inline void
117 xdp_init_buff(struct xdp_buff *xdp, u32 frame_sz, struct xdp_rxq_info *rxq)
119 xdp->frame_sz = frame_sz;
124 static __always_inline void
125 xdp_prepare_buff(struct xdp_buff *xdp, unsigned char *hard_start,
126 int headroom, int data_len, const bool meta_valid)
128 unsigned char *data = hard_start + headroom;
130 xdp->data_hard_start = hard_start;
132 xdp->data_end = data + data_len;
133 xdp->data_meta = meta_valid ? data : data + 1;
136 /* Reserve memory area at end-of data area.
138 * This macro reserves tailroom in the XDP buffer by limiting the
139 * XDP/BPF data access to data_hard_end. Notice same area (and size)
140 * is used for XDP_PASS, when constructing the SKB via build_skb().
142 #define xdp_data_hard_end(xdp) \
143 ((xdp)->data_hard_start + (xdp)->frame_sz - \
144 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)))
146 static inline struct skb_shared_info *
147 xdp_get_shared_info_from_buff(struct xdp_buff *xdp)
149 return (struct skb_shared_info *)xdp_data_hard_end(xdp);
152 static __always_inline unsigned int xdp_get_buff_len(struct xdp_buff *xdp)
154 unsigned int len = xdp->data_end - xdp->data;
155 struct skb_shared_info *sinfo;
157 if (likely(!xdp_buff_has_frags(xdp)))
160 sinfo = xdp_get_shared_info_from_buff(xdp);
161 len += sinfo->xdp_frags_size;
170 u32 metasize; /* uses lower 8-bits */
171 /* Lifetime of xdp_rxq_info is limited to NAPI/enqueue time,
172 * while mem info is valid on remote CPU.
174 struct xdp_mem_info mem;
175 struct net_device *dev_rx; /* used by cpumap */
177 u32 flags; /* supported values defined in xdp_buff_flags */
180 static __always_inline bool xdp_frame_has_frags(struct xdp_frame *frame)
182 return !!(frame->flags & XDP_FLAGS_HAS_FRAGS);
185 static __always_inline bool xdp_frame_is_frag_pfmemalloc(struct xdp_frame *frame)
187 return !!(frame->flags & XDP_FLAGS_FRAGS_PF_MEMALLOC);
190 #define XDP_BULK_QUEUE_SIZE 16
191 struct xdp_frame_bulk {
194 void *q[XDP_BULK_QUEUE_SIZE];
197 static __always_inline void xdp_frame_bulk_init(struct xdp_frame_bulk *bq)
199 /* bq->count will be zero'ed when bq->xa gets updated */
203 static inline struct skb_shared_info *
204 xdp_get_shared_info_from_frame(struct xdp_frame *frame)
206 void *data_hard_start = frame->data - frame->headroom - sizeof(*frame);
208 return (struct skb_shared_info *)(data_hard_start + frame->frame_sz -
209 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)));
212 struct xdp_cpumap_stats {
213 unsigned int redirect;
218 /* Clear kernel pointers in xdp_frame */
219 static inline void xdp_scrub_frame(struct xdp_frame *frame)
222 frame->dev_rx = NULL;
226 xdp_update_skb_shared_info(struct sk_buff *skb, u8 nr_frags,
227 unsigned int size, unsigned int truesize,
230 skb_shinfo(skb)->nr_frags = nr_frags;
233 skb->data_len += size;
234 skb->truesize += truesize;
235 skb->pfmemalloc |= pfmemalloc;
238 /* Avoids inlining WARN macro in fast-path */
239 void xdp_warn(const char *msg, const char *func, const int line);
240 #define XDP_WARN(msg) xdp_warn(msg, __func__, __LINE__)
242 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp);
243 struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf,
245 struct net_device *dev);
246 struct sk_buff *xdp_build_skb_from_frame(struct xdp_frame *xdpf,
247 struct net_device *dev);
248 int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp);
249 struct xdp_frame *xdpf_clone(struct xdp_frame *xdpf);
252 void xdp_convert_frame_to_buff(struct xdp_frame *frame, struct xdp_buff *xdp)
254 xdp->data_hard_start = frame->data - frame->headroom - sizeof(*frame);
255 xdp->data = frame->data;
256 xdp->data_end = frame->data + frame->len;
257 xdp->data_meta = frame->data - frame->metasize;
258 xdp->frame_sz = frame->frame_sz;
259 xdp->flags = frame->flags;
263 int xdp_update_frame_from_buff(struct xdp_buff *xdp,
264 struct xdp_frame *xdp_frame)
266 int metasize, headroom;
268 /* Assure headroom is available for storing info */
269 headroom = xdp->data - xdp->data_hard_start;
270 metasize = xdp->data - xdp->data_meta;
271 metasize = metasize > 0 ? metasize : 0;
272 if (unlikely((headroom - metasize) < sizeof(*xdp_frame)))
275 /* Catch if driver didn't reserve tailroom for skb_shared_info */
276 if (unlikely(xdp->data_end > xdp_data_hard_end(xdp))) {
277 XDP_WARN("Driver BUG: missing reserved tailroom");
281 xdp_frame->data = xdp->data;
282 xdp_frame->len = xdp->data_end - xdp->data;
283 xdp_frame->headroom = headroom - sizeof(*xdp_frame);
284 xdp_frame->metasize = metasize;
285 xdp_frame->frame_sz = xdp->frame_sz;
286 xdp_frame->flags = xdp->flags;
291 /* Convert xdp_buff to xdp_frame */
293 struct xdp_frame *xdp_convert_buff_to_frame(struct xdp_buff *xdp)
295 struct xdp_frame *xdp_frame;
297 if (xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL)
298 return xdp_convert_zc_to_xdp_frame(xdp);
300 /* Store info in top of packet */
301 xdp_frame = xdp->data_hard_start;
302 if (unlikely(xdp_update_frame_from_buff(xdp, xdp_frame) < 0))
305 /* rxq only valid until napi_schedule ends, convert to xdp_mem_info */
306 xdp_frame->mem = xdp->rxq->mem;
311 void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct,
312 struct xdp_buff *xdp);
313 void xdp_return_frame(struct xdp_frame *xdpf);
314 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf);
315 void xdp_return_buff(struct xdp_buff *xdp);
316 void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq);
317 void xdp_return_frame_bulk(struct xdp_frame *xdpf,
318 struct xdp_frame_bulk *bq);
320 static __always_inline unsigned int xdp_get_frame_len(struct xdp_frame *xdpf)
322 struct skb_shared_info *sinfo;
323 unsigned int len = xdpf->len;
325 if (likely(!xdp_frame_has_frags(xdpf)))
328 sinfo = xdp_get_shared_info_from_frame(xdpf);
329 len += sinfo->xdp_frags_size;
334 int __xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
335 struct net_device *dev, u32 queue_index,
336 unsigned int napi_id, u32 frag_size);
338 xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
339 struct net_device *dev, u32 queue_index,
340 unsigned int napi_id)
342 return __xdp_rxq_info_reg(xdp_rxq, dev, queue_index, napi_id, 0);
345 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq);
346 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq);
347 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq);
348 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
349 enum xdp_mem_type type, void *allocator);
350 void xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq);
351 int xdp_reg_mem_model(struct xdp_mem_info *mem,
352 enum xdp_mem_type type, void *allocator);
353 void xdp_unreg_mem_model(struct xdp_mem_info *mem);
355 /* Drivers not supporting XDP metadata can use this helper, which
356 * rejects any room expansion for metadata as a result.
358 static __always_inline void
359 xdp_set_data_meta_invalid(struct xdp_buff *xdp)
361 xdp->data_meta = xdp->data + 1;
364 static __always_inline bool
365 xdp_data_meta_unsupported(const struct xdp_buff *xdp)
367 return unlikely(xdp->data_meta > xdp->data);
370 static inline bool xdp_metalen_invalid(unsigned long metalen)
372 return (metalen & (sizeof(__u32) - 1)) || (metalen > 32);
375 struct xdp_attachment_info {
376 struct bpf_prog *prog;
381 void xdp_attachment_setup(struct xdp_attachment_info *info,
382 struct netdev_bpf *bpf);
384 #define DEV_MAP_BULK_SIZE XDP_BULK_QUEUE_SIZE
386 #define XDP_METADATA_KFUNC_xxx \
387 XDP_METADATA_KFUNC(XDP_METADATA_KFUNC_RX_TIMESTAMP, \
388 bpf_xdp_metadata_rx_timestamp) \
389 XDP_METADATA_KFUNC(XDP_METADATA_KFUNC_RX_HASH, \
390 bpf_xdp_metadata_rx_hash) \
393 #define XDP_METADATA_KFUNC(name, _) name,
394 XDP_METADATA_KFUNC_xxx
395 #undef XDP_METADATA_KFUNC
396 MAX_XDP_METADATA_KFUNC,
399 enum xdp_rss_hash_type {
400 /* First part: Individual bits for L3/L4 types */
401 XDP_RSS_L3_IPV4 = BIT(0),
402 XDP_RSS_L3_IPV6 = BIT(1),
404 /* The fixed (L3) IPv4 and IPv6 headers can both be followed by
405 * variable/dynamic headers, IPv4 called Options and IPv6 called
406 * Extension Headers. HW RSS type can contain this info.
408 XDP_RSS_L3_DYNHDR = BIT(2),
410 /* When RSS hash covers L4 then drivers MUST set XDP_RSS_L4 bit in
411 * addition to the protocol specific bit. This ease interaction with
412 * SKBs and avoids reserving a fixed mask for future L4 protocol bits.
414 XDP_RSS_L4 = BIT(3), /* L4 based hash, proto can be unknown */
415 XDP_RSS_L4_TCP = BIT(4),
416 XDP_RSS_L4_UDP = BIT(5),
417 XDP_RSS_L4_SCTP = BIT(6),
418 XDP_RSS_L4_IPSEC = BIT(7), /* L4 based hash include IPSEC SPI */
420 /* Second part: RSS hash type combinations used for driver HW mapping */
421 XDP_RSS_TYPE_NONE = 0,
422 XDP_RSS_TYPE_L2 = XDP_RSS_TYPE_NONE,
424 XDP_RSS_TYPE_L3_IPV4 = XDP_RSS_L3_IPV4,
425 XDP_RSS_TYPE_L3_IPV6 = XDP_RSS_L3_IPV6,
426 XDP_RSS_TYPE_L3_IPV4_OPT = XDP_RSS_L3_IPV4 | XDP_RSS_L3_DYNHDR,
427 XDP_RSS_TYPE_L3_IPV6_EX = XDP_RSS_L3_IPV6 | XDP_RSS_L3_DYNHDR,
429 XDP_RSS_TYPE_L4_ANY = XDP_RSS_L4,
430 XDP_RSS_TYPE_L4_IPV4_TCP = XDP_RSS_L3_IPV4 | XDP_RSS_L4 | XDP_RSS_L4_TCP,
431 XDP_RSS_TYPE_L4_IPV4_UDP = XDP_RSS_L3_IPV4 | XDP_RSS_L4 | XDP_RSS_L4_UDP,
432 XDP_RSS_TYPE_L4_IPV4_SCTP = XDP_RSS_L3_IPV4 | XDP_RSS_L4 | XDP_RSS_L4_SCTP,
433 XDP_RSS_TYPE_L4_IPV4_IPSEC = XDP_RSS_L3_IPV4 | XDP_RSS_L4 | XDP_RSS_L4_IPSEC,
435 XDP_RSS_TYPE_L4_IPV6_TCP = XDP_RSS_L3_IPV6 | XDP_RSS_L4 | XDP_RSS_L4_TCP,
436 XDP_RSS_TYPE_L4_IPV6_UDP = XDP_RSS_L3_IPV6 | XDP_RSS_L4 | XDP_RSS_L4_UDP,
437 XDP_RSS_TYPE_L4_IPV6_SCTP = XDP_RSS_L3_IPV6 | XDP_RSS_L4 | XDP_RSS_L4_SCTP,
438 XDP_RSS_TYPE_L4_IPV6_IPSEC = XDP_RSS_L3_IPV6 | XDP_RSS_L4 | XDP_RSS_L4_IPSEC,
440 XDP_RSS_TYPE_L4_IPV6_TCP_EX = XDP_RSS_TYPE_L4_IPV6_TCP | XDP_RSS_L3_DYNHDR,
441 XDP_RSS_TYPE_L4_IPV6_UDP_EX = XDP_RSS_TYPE_L4_IPV6_UDP | XDP_RSS_L3_DYNHDR,
442 XDP_RSS_TYPE_L4_IPV6_SCTP_EX = XDP_RSS_TYPE_L4_IPV6_SCTP | XDP_RSS_L3_DYNHDR,
445 struct xdp_metadata_ops {
446 int (*xmo_rx_timestamp)(const struct xdp_md *ctx, u64 *timestamp);
447 int (*xmo_rx_hash)(const struct xdp_md *ctx, u32 *hash,
448 enum xdp_rss_hash_type *rss_type);
452 u32 bpf_xdp_metadata_kfunc_id(int id);
453 bool bpf_dev_bound_kfunc_id(u32 btf_id);
454 void xdp_set_features_flag(struct net_device *dev, xdp_features_t val);
455 void xdp_features_set_redirect_target(struct net_device *dev, bool support_sg);
456 void xdp_features_clear_redirect_target(struct net_device *dev);
458 static inline u32 bpf_xdp_metadata_kfunc_id(int id) { return 0; }
459 static inline bool bpf_dev_bound_kfunc_id(u32 btf_id) { return false; }
462 xdp_set_features_flag(struct net_device *dev, xdp_features_t val)
467 xdp_features_set_redirect_target(struct net_device *dev, bool support_sg)
472 xdp_features_clear_redirect_target(struct net_device *dev)
477 static inline void xdp_clear_features_flag(struct net_device *dev)
479 xdp_set_features_flag(dev, 0);
482 static __always_inline u32 bpf_prog_run_xdp(const struct bpf_prog *prog,
483 struct xdp_buff *xdp)
485 /* Driver XDP hooks are invoked within a single NAPI poll cycle and thus
486 * under local_bh_disable(), which provides the needed RCU protection
487 * for accessing map entries.
489 u32 act = __bpf_prog_run(prog, xdp, BPF_DISPATCHER_FUNC(xdp));
491 if (static_branch_unlikely(&bpf_master_redirect_enabled_key)) {
492 if (act == XDP_TX && netif_is_bond_slave(xdp->rxq->dev))
493 act = xdp_master_redirect(xdp);
498 #endif /* __LINUX_NET_XDP_H__ */