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 #include <uapi/linux/netdev.h> 11 #include <linux/bitfield.h> 12 13 /** 14 * DOC: XDP RX-queue information 15 * 16 * The XDP RX-queue info (xdp_rxq_info) is associated with the driver 17 * level RX-ring queues. It is information that is specific to how 18 * the driver have configured a given RX-ring queue. 19 * 20 * Each xdp_buff frame received in the driver carries a (pointer) 21 * reference to this xdp_rxq_info structure. This provides the XDP 22 * data-path read-access to RX-info for both kernel and bpf-side 23 * (limited subset). 24 * 25 * For now, direct access is only safe while running in NAPI/softirq 26 * context. Contents are read-mostly and must not be updated during 27 * driver NAPI/softirq poll. 28 * 29 * The driver usage API is a register and unregister API. 30 * 31 * The struct is not directly tied to the XDP prog. A new XDP prog 32 * can be attached as long as it doesn't change the underlying 33 * RX-ring. If the RX-ring does change significantly, the NIC driver 34 * naturally need to stop the RX-ring before purging and reallocating 35 * memory. In that process the driver MUST call unregister (which 36 * also applies for driver shutdown and unload). The register API is 37 * also mandatory during RX-ring setup. 38 */ 39 40 enum xdp_mem_type { 41 MEM_TYPE_PAGE_SHARED = 0, /* Split-page refcnt based model */ 42 MEM_TYPE_PAGE_ORDER0, /* Orig XDP full page model */ 43 MEM_TYPE_PAGE_POOL, 44 MEM_TYPE_XSK_BUFF_POOL, 45 MEM_TYPE_MAX, 46 }; 47 48 typedef u32 xdp_features_t; 49 50 /* XDP flags for ndo_xdp_xmit */ 51 #define XDP_XMIT_FLUSH (1U << 0) /* doorbell signal consumer */ 52 #define XDP_XMIT_FLAGS_MASK XDP_XMIT_FLUSH 53 54 struct xdp_mem_info { 55 u32 type; /* enum xdp_mem_type, but known size type */ 56 u32 id; 57 }; 58 59 struct page_pool; 60 61 struct xdp_rxq_info { 62 struct net_device *dev; 63 u32 queue_index; 64 u32 reg_state; 65 struct xdp_mem_info mem; 66 unsigned int napi_id; 67 u32 frag_size; 68 } ____cacheline_aligned; /* perf critical, avoid false-sharing */ 69 70 struct xdp_txq_info { 71 struct net_device *dev; 72 }; 73 74 enum xdp_buff_flags { 75 XDP_FLAGS_HAS_FRAGS = BIT(0), /* non-linear xdp buff */ 76 XDP_FLAGS_FRAGS_PF_MEMALLOC = BIT(1), /* xdp paged memory is under 77 * pressure 78 */ 79 }; 80 81 struct xdp_buff { 82 void *data; 83 void *data_end; 84 void *data_meta; 85 void *data_hard_start; 86 struct xdp_rxq_info *rxq; 87 struct xdp_txq_info *txq; 88 u32 frame_sz; /* frame size to deduce data_hard_end/reserved tailroom*/ 89 u32 flags; /* supported values defined in xdp_buff_flags */ 90 }; 91 92 static __always_inline bool xdp_buff_has_frags(struct xdp_buff *xdp) 93 { 94 return !!(xdp->flags & XDP_FLAGS_HAS_FRAGS); 95 } 96 97 static __always_inline void xdp_buff_set_frags_flag(struct xdp_buff *xdp) 98 { 99 xdp->flags |= XDP_FLAGS_HAS_FRAGS; 100 } 101 102 static __always_inline void xdp_buff_clear_frags_flag(struct xdp_buff *xdp) 103 { 104 xdp->flags &= ~XDP_FLAGS_HAS_FRAGS; 105 } 106 107 static __always_inline bool xdp_buff_is_frag_pfmemalloc(struct xdp_buff *xdp) 108 { 109 return !!(xdp->flags & XDP_FLAGS_FRAGS_PF_MEMALLOC); 110 } 111 112 static __always_inline void xdp_buff_set_frag_pfmemalloc(struct xdp_buff *xdp) 113 { 114 xdp->flags |= XDP_FLAGS_FRAGS_PF_MEMALLOC; 115 } 116 117 static __always_inline void 118 xdp_init_buff(struct xdp_buff *xdp, u32 frame_sz, struct xdp_rxq_info *rxq) 119 { 120 xdp->frame_sz = frame_sz; 121 xdp->rxq = rxq; 122 xdp->flags = 0; 123 } 124 125 static __always_inline void 126 xdp_prepare_buff(struct xdp_buff *xdp, unsigned char *hard_start, 127 int headroom, int data_len, const bool meta_valid) 128 { 129 unsigned char *data = hard_start + headroom; 130 131 xdp->data_hard_start = hard_start; 132 xdp->data = data; 133 xdp->data_end = data + data_len; 134 xdp->data_meta = meta_valid ? data : data + 1; 135 } 136 137 /* Reserve memory area at end-of data area. 138 * 139 * This macro reserves tailroom in the XDP buffer by limiting the 140 * XDP/BPF data access to data_hard_end. Notice same area (and size) 141 * is used for XDP_PASS, when constructing the SKB via build_skb(). 142 */ 143 #define xdp_data_hard_end(xdp) \ 144 ((xdp)->data_hard_start + (xdp)->frame_sz - \ 145 SKB_DATA_ALIGN(sizeof(struct skb_shared_info))) 146 147 static inline struct skb_shared_info * 148 xdp_get_shared_info_from_buff(struct xdp_buff *xdp) 149 { 150 return (struct skb_shared_info *)xdp_data_hard_end(xdp); 151 } 152 153 static __always_inline unsigned int xdp_get_buff_len(struct xdp_buff *xdp) 154 { 155 unsigned int len = xdp->data_end - xdp->data; 156 struct skb_shared_info *sinfo; 157 158 if (likely(!xdp_buff_has_frags(xdp))) 159 goto out; 160 161 sinfo = xdp_get_shared_info_from_buff(xdp); 162 len += sinfo->xdp_frags_size; 163 out: 164 return len; 165 } 166 167 struct xdp_frame { 168 void *data; 169 u16 len; 170 u16 headroom; 171 u32 metasize; /* uses lower 8-bits */ 172 /* Lifetime of xdp_rxq_info is limited to NAPI/enqueue time, 173 * while mem info is valid on remote CPU. 174 */ 175 struct xdp_mem_info mem; 176 struct net_device *dev_rx; /* used by cpumap */ 177 u32 frame_sz; 178 u32 flags; /* supported values defined in xdp_buff_flags */ 179 }; 180 181 static __always_inline bool xdp_frame_has_frags(struct xdp_frame *frame) 182 { 183 return !!(frame->flags & XDP_FLAGS_HAS_FRAGS); 184 } 185 186 static __always_inline bool xdp_frame_is_frag_pfmemalloc(struct xdp_frame *frame) 187 { 188 return !!(frame->flags & XDP_FLAGS_FRAGS_PF_MEMALLOC); 189 } 190 191 #define XDP_BULK_QUEUE_SIZE 16 192 struct xdp_frame_bulk { 193 int count; 194 void *xa; 195 void *q[XDP_BULK_QUEUE_SIZE]; 196 }; 197 198 static __always_inline void xdp_frame_bulk_init(struct xdp_frame_bulk *bq) 199 { 200 /* bq->count will be zero'ed when bq->xa gets updated */ 201 bq->xa = NULL; 202 } 203 204 static inline struct skb_shared_info * 205 xdp_get_shared_info_from_frame(struct xdp_frame *frame) 206 { 207 void *data_hard_start = frame->data - frame->headroom - sizeof(*frame); 208 209 return (struct skb_shared_info *)(data_hard_start + frame->frame_sz - 210 SKB_DATA_ALIGN(sizeof(struct skb_shared_info))); 211 } 212 213 struct xdp_cpumap_stats { 214 unsigned int redirect; 215 unsigned int pass; 216 unsigned int drop; 217 }; 218 219 /* Clear kernel pointers in xdp_frame */ 220 static inline void xdp_scrub_frame(struct xdp_frame *frame) 221 { 222 frame->data = NULL; 223 frame->dev_rx = NULL; 224 } 225 226 static inline void 227 xdp_update_skb_shared_info(struct sk_buff *skb, u8 nr_frags, 228 unsigned int size, unsigned int truesize, 229 bool pfmemalloc) 230 { 231 skb_shinfo(skb)->nr_frags = nr_frags; 232 233 skb->len += size; 234 skb->data_len += size; 235 skb->truesize += truesize; 236 skb->pfmemalloc |= pfmemalloc; 237 } 238 239 /* Avoids inlining WARN macro in fast-path */ 240 void xdp_warn(const char *msg, const char *func, const int line); 241 #define XDP_WARN(msg) xdp_warn(msg, __func__, __LINE__) 242 243 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp); 244 struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf, 245 struct sk_buff *skb, 246 struct net_device *dev); 247 struct sk_buff *xdp_build_skb_from_frame(struct xdp_frame *xdpf, 248 struct net_device *dev); 249 int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp); 250 struct xdp_frame *xdpf_clone(struct xdp_frame *xdpf); 251 252 static inline 253 void xdp_convert_frame_to_buff(struct xdp_frame *frame, struct xdp_buff *xdp) 254 { 255 xdp->data_hard_start = frame->data - frame->headroom - sizeof(*frame); 256 xdp->data = frame->data; 257 xdp->data_end = frame->data + frame->len; 258 xdp->data_meta = frame->data - frame->metasize; 259 xdp->frame_sz = frame->frame_sz; 260 xdp->flags = frame->flags; 261 } 262 263 static inline 264 int xdp_update_frame_from_buff(struct xdp_buff *xdp, 265 struct xdp_frame *xdp_frame) 266 { 267 int metasize, headroom; 268 269 /* Assure headroom is available for storing info */ 270 headroom = xdp->data - xdp->data_hard_start; 271 metasize = xdp->data - xdp->data_meta; 272 metasize = metasize > 0 ? metasize : 0; 273 if (unlikely((headroom - metasize) < sizeof(*xdp_frame))) 274 return -ENOSPC; 275 276 /* Catch if driver didn't reserve tailroom for skb_shared_info */ 277 if (unlikely(xdp->data_end > xdp_data_hard_end(xdp))) { 278 XDP_WARN("Driver BUG: missing reserved tailroom"); 279 return -ENOSPC; 280 } 281 282 xdp_frame->data = xdp->data; 283 xdp_frame->len = xdp->data_end - xdp->data; 284 xdp_frame->headroom = headroom - sizeof(*xdp_frame); 285 xdp_frame->metasize = metasize; 286 xdp_frame->frame_sz = xdp->frame_sz; 287 xdp_frame->flags = xdp->flags; 288 289 return 0; 290 } 291 292 /* Convert xdp_buff to xdp_frame */ 293 static inline 294 struct xdp_frame *xdp_convert_buff_to_frame(struct xdp_buff *xdp) 295 { 296 struct xdp_frame *xdp_frame; 297 298 if (xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL) 299 return xdp_convert_zc_to_xdp_frame(xdp); 300 301 /* Store info in top of packet */ 302 xdp_frame = xdp->data_hard_start; 303 if (unlikely(xdp_update_frame_from_buff(xdp, xdp_frame) < 0)) 304 return NULL; 305 306 /* rxq only valid until napi_schedule ends, convert to xdp_mem_info */ 307 xdp_frame->mem = xdp->rxq->mem; 308 309 return xdp_frame; 310 } 311 312 void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct, 313 struct xdp_buff *xdp); 314 void xdp_return_frame(struct xdp_frame *xdpf); 315 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf); 316 void xdp_return_buff(struct xdp_buff *xdp); 317 void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq); 318 void xdp_return_frame_bulk(struct xdp_frame *xdpf, 319 struct xdp_frame_bulk *bq); 320 321 /* When sending xdp_frame into the network stack, then there is no 322 * return point callback, which is needed to release e.g. DMA-mapping 323 * resources with page_pool. Thus, have explicit function to release 324 * frame resources. 325 */ 326 void __xdp_release_frame(void *data, struct xdp_mem_info *mem); 327 static inline void xdp_release_frame(struct xdp_frame *xdpf) 328 { 329 struct xdp_mem_info *mem = &xdpf->mem; 330 struct skb_shared_info *sinfo; 331 int i; 332 333 /* Curr only page_pool needs this */ 334 if (mem->type != MEM_TYPE_PAGE_POOL) 335 return; 336 337 if (likely(!xdp_frame_has_frags(xdpf))) 338 goto out; 339 340 sinfo = xdp_get_shared_info_from_frame(xdpf); 341 for (i = 0; i < sinfo->nr_frags; i++) { 342 struct page *page = skb_frag_page(&sinfo->frags[i]); 343 344 __xdp_release_frame(page_address(page), mem); 345 } 346 out: 347 __xdp_release_frame(xdpf->data, mem); 348 } 349 350 static __always_inline unsigned int xdp_get_frame_len(struct xdp_frame *xdpf) 351 { 352 struct skb_shared_info *sinfo; 353 unsigned int len = xdpf->len; 354 355 if (likely(!xdp_frame_has_frags(xdpf))) 356 goto out; 357 358 sinfo = xdp_get_shared_info_from_frame(xdpf); 359 len += sinfo->xdp_frags_size; 360 out: 361 return len; 362 } 363 364 int __xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq, 365 struct net_device *dev, u32 queue_index, 366 unsigned int napi_id, u32 frag_size); 367 static inline int 368 xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq, 369 struct net_device *dev, u32 queue_index, 370 unsigned int napi_id) 371 { 372 return __xdp_rxq_info_reg(xdp_rxq, dev, queue_index, napi_id, 0); 373 } 374 375 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq); 376 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq); 377 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq); 378 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq, 379 enum xdp_mem_type type, void *allocator); 380 void xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq); 381 int xdp_reg_mem_model(struct xdp_mem_info *mem, 382 enum xdp_mem_type type, void *allocator); 383 void xdp_unreg_mem_model(struct xdp_mem_info *mem); 384 385 /* Drivers not supporting XDP metadata can use this helper, which 386 * rejects any room expansion for metadata as a result. 387 */ 388 static __always_inline void 389 xdp_set_data_meta_invalid(struct xdp_buff *xdp) 390 { 391 xdp->data_meta = xdp->data + 1; 392 } 393 394 static __always_inline bool 395 xdp_data_meta_unsupported(const struct xdp_buff *xdp) 396 { 397 return unlikely(xdp->data_meta > xdp->data); 398 } 399 400 static inline bool xdp_metalen_invalid(unsigned long metalen) 401 { 402 return (metalen & (sizeof(__u32) - 1)) || (metalen > 32); 403 } 404 405 struct xdp_attachment_info { 406 struct bpf_prog *prog; 407 u32 flags; 408 }; 409 410 struct netdev_bpf; 411 void xdp_attachment_setup(struct xdp_attachment_info *info, 412 struct netdev_bpf *bpf); 413 414 #define DEV_MAP_BULK_SIZE XDP_BULK_QUEUE_SIZE 415 416 #define XDP_METADATA_KFUNC_xxx \ 417 XDP_METADATA_KFUNC(XDP_METADATA_KFUNC_RX_TIMESTAMP, \ 418 bpf_xdp_metadata_rx_timestamp) \ 419 XDP_METADATA_KFUNC(XDP_METADATA_KFUNC_RX_HASH, \ 420 bpf_xdp_metadata_rx_hash) \ 421 422 enum { 423 #define XDP_METADATA_KFUNC(name, _) name, 424 XDP_METADATA_KFUNC_xxx 425 #undef XDP_METADATA_KFUNC 426 MAX_XDP_METADATA_KFUNC, 427 }; 428 429 enum xdp_rss_hash_type { 430 /* First part: Individual bits for L3/L4 types */ 431 XDP_RSS_L3_IPV4 = BIT(0), 432 XDP_RSS_L3_IPV6 = BIT(1), 433 434 /* The fixed (L3) IPv4 and IPv6 headers can both be followed by 435 * variable/dynamic headers, IPv4 called Options and IPv6 called 436 * Extension Headers. HW RSS type can contain this info. 437 */ 438 XDP_RSS_L3_DYNHDR = BIT(2), 439 440 /* When RSS hash covers L4 then drivers MUST set XDP_RSS_L4 bit in 441 * addition to the protocol specific bit. This ease interaction with 442 * SKBs and avoids reserving a fixed mask for future L4 protocol bits. 443 */ 444 XDP_RSS_L4 = BIT(3), /* L4 based hash, proto can be unknown */ 445 XDP_RSS_L4_TCP = BIT(4), 446 XDP_RSS_L4_UDP = BIT(5), 447 XDP_RSS_L4_SCTP = BIT(6), 448 XDP_RSS_L4_IPSEC = BIT(7), /* L4 based hash include IPSEC SPI */ 449 450 /* Second part: RSS hash type combinations used for driver HW mapping */ 451 XDP_RSS_TYPE_NONE = 0, 452 XDP_RSS_TYPE_L2 = XDP_RSS_TYPE_NONE, 453 454 XDP_RSS_TYPE_L3_IPV4 = XDP_RSS_L3_IPV4, 455 XDP_RSS_TYPE_L3_IPV6 = XDP_RSS_L3_IPV6, 456 XDP_RSS_TYPE_L3_IPV4_OPT = XDP_RSS_L3_IPV4 | XDP_RSS_L3_DYNHDR, 457 XDP_RSS_TYPE_L3_IPV6_EX = XDP_RSS_L3_IPV6 | XDP_RSS_L3_DYNHDR, 458 459 XDP_RSS_TYPE_L4_ANY = XDP_RSS_L4, 460 XDP_RSS_TYPE_L4_IPV4_TCP = XDP_RSS_L3_IPV4 | XDP_RSS_L4 | XDP_RSS_L4_TCP, 461 XDP_RSS_TYPE_L4_IPV4_UDP = XDP_RSS_L3_IPV4 | XDP_RSS_L4 | XDP_RSS_L4_UDP, 462 XDP_RSS_TYPE_L4_IPV4_SCTP = XDP_RSS_L3_IPV4 | XDP_RSS_L4 | XDP_RSS_L4_SCTP, 463 XDP_RSS_TYPE_L4_IPV4_IPSEC = XDP_RSS_L3_IPV4 | XDP_RSS_L4 | XDP_RSS_L4_IPSEC, 464 465 XDP_RSS_TYPE_L4_IPV6_TCP = XDP_RSS_L3_IPV6 | XDP_RSS_L4 | XDP_RSS_L4_TCP, 466 XDP_RSS_TYPE_L4_IPV6_UDP = XDP_RSS_L3_IPV6 | XDP_RSS_L4 | XDP_RSS_L4_UDP, 467 XDP_RSS_TYPE_L4_IPV6_SCTP = XDP_RSS_L3_IPV6 | XDP_RSS_L4 | XDP_RSS_L4_SCTP, 468 XDP_RSS_TYPE_L4_IPV6_IPSEC = XDP_RSS_L3_IPV6 | XDP_RSS_L4 | XDP_RSS_L4_IPSEC, 469 470 XDP_RSS_TYPE_L4_IPV6_TCP_EX = XDP_RSS_TYPE_L4_IPV6_TCP | XDP_RSS_L3_DYNHDR, 471 XDP_RSS_TYPE_L4_IPV6_UDP_EX = XDP_RSS_TYPE_L4_IPV6_UDP | XDP_RSS_L3_DYNHDR, 472 XDP_RSS_TYPE_L4_IPV6_SCTP_EX = XDP_RSS_TYPE_L4_IPV6_SCTP | XDP_RSS_L3_DYNHDR, 473 }; 474 475 #ifdef CONFIG_NET 476 u32 bpf_xdp_metadata_kfunc_id(int id); 477 bool bpf_dev_bound_kfunc_id(u32 btf_id); 478 void xdp_set_features_flag(struct net_device *dev, xdp_features_t val); 479 void xdp_features_set_redirect_target(struct net_device *dev, bool support_sg); 480 void xdp_features_clear_redirect_target(struct net_device *dev); 481 #else 482 static inline u32 bpf_xdp_metadata_kfunc_id(int id) { return 0; } 483 static inline bool bpf_dev_bound_kfunc_id(u32 btf_id) { return false; } 484 485 static inline void 486 xdp_set_features_flag(struct net_device *dev, xdp_features_t val) 487 { 488 } 489 490 static inline void 491 xdp_features_set_redirect_target(struct net_device *dev, bool support_sg) 492 { 493 } 494 495 static inline void 496 xdp_features_clear_redirect_target(struct net_device *dev) 497 { 498 } 499 #endif 500 501 static inline void xdp_clear_features_flag(struct net_device *dev) 502 { 503 xdp_set_features_flag(dev, 0); 504 } 505 506 #endif /* __LINUX_NET_XDP_H__ */ 507