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/bitfield.h> 10 #include <linux/filter.h> 11 #include <linux/netdevice.h> 12 #include <linux/skbuff.h> /* skb_shared_info */ 13 14 /** 15 * DOC: XDP RX-queue information 16 * 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 has configured a given RX-ring queue. 20 * 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 24 * (limited subset). 25 * 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. 29 * 30 * The driver usage API is a register and unregister API. 31 * 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 needs 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. 39 */ 40 41 enum xdp_mem_type { 42 MEM_TYPE_PAGE_SHARED = 0, /* Split-page refcnt based model */ 43 MEM_TYPE_PAGE_ORDER0, /* Orig XDP full page model */ 44 MEM_TYPE_PAGE_POOL, 45 MEM_TYPE_XSK_BUFF_POOL, 46 MEM_TYPE_MAX, 47 }; 48 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 52 53 struct xdp_mem_info { 54 u32 type; /* enum xdp_mem_type, but known size type */ 55 u32 id; 56 }; 57 58 struct page_pool; 59 60 struct xdp_rxq_info { 61 struct net_device *dev; 62 u32 queue_index; 63 u32 reg_state; 64 struct xdp_mem_info mem; 65 u32 frag_size; 66 } ____cacheline_aligned; /* perf critical, avoid false-sharing */ 67 68 struct xdp_txq_info { 69 struct net_device *dev; 70 }; 71 72 enum xdp_buff_flags { 73 XDP_FLAGS_HAS_FRAGS = BIT(0), /* non-linear xdp buff */ 74 XDP_FLAGS_FRAGS_PF_MEMALLOC = BIT(1), /* xdp paged memory is under 75 * pressure 76 */ 77 }; 78 79 struct xdp_buff { 80 void *data; 81 void *data_end; 82 void *data_meta; 83 void *data_hard_start; 84 struct xdp_rxq_info *rxq; 85 struct xdp_txq_info *txq; 86 u32 frame_sz; /* frame size to deduce data_hard_end/reserved tailroom*/ 87 u32 flags; /* supported values defined in xdp_buff_flags */ 88 }; 89 90 static __always_inline bool xdp_buff_has_frags(struct xdp_buff *xdp) 91 { 92 return !!(xdp->flags & XDP_FLAGS_HAS_FRAGS); 93 } 94 95 static __always_inline void xdp_buff_set_frags_flag(struct xdp_buff *xdp) 96 { 97 xdp->flags |= XDP_FLAGS_HAS_FRAGS; 98 } 99 100 static __always_inline void xdp_buff_clear_frags_flag(struct xdp_buff *xdp) 101 { 102 xdp->flags &= ~XDP_FLAGS_HAS_FRAGS; 103 } 104 105 static __always_inline bool xdp_buff_is_frag_pfmemalloc(struct xdp_buff *xdp) 106 { 107 return !!(xdp->flags & XDP_FLAGS_FRAGS_PF_MEMALLOC); 108 } 109 110 static __always_inline void xdp_buff_set_frag_pfmemalloc(struct xdp_buff *xdp) 111 { 112 xdp->flags |= XDP_FLAGS_FRAGS_PF_MEMALLOC; 113 } 114 115 static __always_inline void 116 xdp_init_buff(struct xdp_buff *xdp, u32 frame_sz, struct xdp_rxq_info *rxq) 117 { 118 xdp->frame_sz = frame_sz; 119 xdp->rxq = rxq; 120 xdp->flags = 0; 121 } 122 123 static __always_inline void 124 xdp_prepare_buff(struct xdp_buff *xdp, unsigned char *hard_start, 125 int headroom, int data_len, const bool meta_valid) 126 { 127 unsigned char *data = hard_start + headroom; 128 129 xdp->data_hard_start = hard_start; 130 xdp->data = data; 131 xdp->data_end = data + data_len; 132 xdp->data_meta = meta_valid ? data : data + 1; 133 } 134 135 /* Reserve memory area at end-of data area. 136 * 137 * This macro reserves tailroom in the XDP buffer by limiting the 138 * XDP/BPF data access to data_hard_end. Notice same area (and size) 139 * is used for XDP_PASS, when constructing the SKB via build_skb(). 140 */ 141 #define xdp_data_hard_end(xdp) \ 142 ((xdp)->data_hard_start + (xdp)->frame_sz - \ 143 SKB_DATA_ALIGN(sizeof(struct skb_shared_info))) 144 145 static inline struct skb_shared_info * 146 xdp_get_shared_info_from_buff(struct xdp_buff *xdp) 147 { 148 return (struct skb_shared_info *)xdp_data_hard_end(xdp); 149 } 150 151 static __always_inline unsigned int xdp_get_buff_len(struct xdp_buff *xdp) 152 { 153 unsigned int len = xdp->data_end - xdp->data; 154 struct skb_shared_info *sinfo; 155 156 if (likely(!xdp_buff_has_frags(xdp))) 157 goto out; 158 159 sinfo = xdp_get_shared_info_from_buff(xdp); 160 len += sinfo->xdp_frags_size; 161 out: 162 return len; 163 } 164 165 struct xdp_frame { 166 void *data; 167 u16 len; 168 u16 headroom; 169 u32 metasize; /* uses lower 8-bits */ 170 /* Lifetime of xdp_rxq_info is limited to NAPI/enqueue time, 171 * while mem info is valid on remote CPU. 172 */ 173 struct xdp_mem_info mem; 174 struct net_device *dev_rx; /* used by cpumap */ 175 u32 frame_sz; 176 u32 flags; /* supported values defined in xdp_buff_flags */ 177 }; 178 179 static __always_inline bool xdp_frame_has_frags(struct xdp_frame *frame) 180 { 181 return !!(frame->flags & XDP_FLAGS_HAS_FRAGS); 182 } 183 184 static __always_inline bool xdp_frame_is_frag_pfmemalloc(struct xdp_frame *frame) 185 { 186 return !!(frame->flags & XDP_FLAGS_FRAGS_PF_MEMALLOC); 187 } 188 189 #define XDP_BULK_QUEUE_SIZE 16 190 struct xdp_frame_bulk { 191 int count; 192 void *xa; 193 void *q[XDP_BULK_QUEUE_SIZE]; 194 }; 195 196 static __always_inline void xdp_frame_bulk_init(struct xdp_frame_bulk *bq) 197 { 198 /* bq->count will be zero'ed when bq->xa gets updated */ 199 bq->xa = NULL; 200 } 201 202 static inline struct skb_shared_info * 203 xdp_get_shared_info_from_frame(struct xdp_frame *frame) 204 { 205 void *data_hard_start = frame->data - frame->headroom - sizeof(*frame); 206 207 return (struct skb_shared_info *)(data_hard_start + frame->frame_sz - 208 SKB_DATA_ALIGN(sizeof(struct skb_shared_info))); 209 } 210 211 struct xdp_cpumap_stats { 212 unsigned int redirect; 213 unsigned int pass; 214 unsigned int drop; 215 }; 216 217 /* Clear kernel pointers in xdp_frame */ 218 static inline void xdp_scrub_frame(struct xdp_frame *frame) 219 { 220 frame->data = NULL; 221 frame->dev_rx = NULL; 222 } 223 224 static inline void 225 xdp_update_skb_shared_info(struct sk_buff *skb, u8 nr_frags, 226 unsigned int size, unsigned int truesize, 227 bool pfmemalloc) 228 { 229 skb_shinfo(skb)->nr_frags = nr_frags; 230 231 skb->len += size; 232 skb->data_len += size; 233 skb->truesize += truesize; 234 skb->pfmemalloc |= pfmemalloc; 235 } 236 237 /* Avoids inlining WARN macro in fast-path */ 238 void xdp_warn(const char *msg, const char *func, const int line); 239 #define XDP_WARN(msg) xdp_warn(msg, __func__, __LINE__) 240 241 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp); 242 struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf, 243 struct sk_buff *skb, 244 struct net_device *dev); 245 struct sk_buff *xdp_build_skb_from_frame(struct xdp_frame *xdpf, 246 struct net_device *dev); 247 int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp); 248 struct xdp_frame *xdpf_clone(struct xdp_frame *xdpf); 249 250 static inline 251 void xdp_convert_frame_to_buff(struct xdp_frame *frame, struct xdp_buff *xdp) 252 { 253 xdp->data_hard_start = frame->data - frame->headroom - sizeof(*frame); 254 xdp->data = frame->data; 255 xdp->data_end = frame->data + frame->len; 256 xdp->data_meta = frame->data - frame->metasize; 257 xdp->frame_sz = frame->frame_sz; 258 xdp->flags = frame->flags; 259 } 260 261 static inline 262 int xdp_update_frame_from_buff(struct xdp_buff *xdp, 263 struct xdp_frame *xdp_frame) 264 { 265 int metasize, headroom; 266 267 /* Assure headroom is available for storing info */ 268 headroom = xdp->data - xdp->data_hard_start; 269 metasize = xdp->data - xdp->data_meta; 270 metasize = metasize > 0 ? metasize : 0; 271 if (unlikely((headroom - metasize) < sizeof(*xdp_frame))) 272 return -ENOSPC; 273 274 /* Catch if driver didn't reserve tailroom for skb_shared_info */ 275 if (unlikely(xdp->data_end > xdp_data_hard_end(xdp))) { 276 XDP_WARN("Driver BUG: missing reserved tailroom"); 277 return -ENOSPC; 278 } 279 280 xdp_frame->data = xdp->data; 281 xdp_frame->len = xdp->data_end - xdp->data; 282 xdp_frame->headroom = headroom - sizeof(*xdp_frame); 283 xdp_frame->metasize = metasize; 284 xdp_frame->frame_sz = xdp->frame_sz; 285 xdp_frame->flags = xdp->flags; 286 287 return 0; 288 } 289 290 /* Convert xdp_buff to xdp_frame */ 291 static inline 292 struct xdp_frame *xdp_convert_buff_to_frame(struct xdp_buff *xdp) 293 { 294 struct xdp_frame *xdp_frame; 295 296 if (xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL) 297 return xdp_convert_zc_to_xdp_frame(xdp); 298 299 /* Store info in top of packet */ 300 xdp_frame = xdp->data_hard_start; 301 if (unlikely(xdp_update_frame_from_buff(xdp, xdp_frame) < 0)) 302 return NULL; 303 304 /* rxq only valid until napi_schedule ends, convert to xdp_mem_info */ 305 xdp_frame->mem = xdp->rxq->mem; 306 307 return xdp_frame; 308 } 309 310 void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct, 311 struct xdp_buff *xdp); 312 void xdp_return_frame(struct xdp_frame *xdpf); 313 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf); 314 void xdp_return_buff(struct xdp_buff *xdp); 315 void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq); 316 void xdp_return_frame_bulk(struct xdp_frame *xdpf, 317 struct xdp_frame_bulk *bq); 318 319 static __always_inline unsigned int xdp_get_frame_len(struct xdp_frame *xdpf) 320 { 321 struct skb_shared_info *sinfo; 322 unsigned int len = xdpf->len; 323 324 if (likely(!xdp_frame_has_frags(xdpf))) 325 goto out; 326 327 sinfo = xdp_get_shared_info_from_frame(xdpf); 328 len += sinfo->xdp_frags_size; 329 out: 330 return len; 331 } 332 333 int __xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq, 334 struct net_device *dev, u32 queue_index, 335 unsigned int napi_id, u32 frag_size); 336 static inline int 337 xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq, 338 struct net_device *dev, u32 queue_index, 339 unsigned int napi_id) 340 { 341 return __xdp_rxq_info_reg(xdp_rxq, dev, queue_index, napi_id, 0); 342 } 343 344 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq); 345 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq); 346 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq); 347 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq, 348 enum xdp_mem_type type, void *allocator); 349 void xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq); 350 int xdp_reg_mem_model(struct xdp_mem_info *mem, 351 enum xdp_mem_type type, void *allocator); 352 void xdp_unreg_mem_model(struct xdp_mem_info *mem); 353 354 /* Drivers not supporting XDP metadata can use this helper, which 355 * rejects any room expansion for metadata as a result. 356 */ 357 static __always_inline void 358 xdp_set_data_meta_invalid(struct xdp_buff *xdp) 359 { 360 xdp->data_meta = xdp->data + 1; 361 } 362 363 static __always_inline bool 364 xdp_data_meta_unsupported(const struct xdp_buff *xdp) 365 { 366 return unlikely(xdp->data_meta > xdp->data); 367 } 368 369 static inline bool xdp_metalen_invalid(unsigned long metalen) 370 { 371 unsigned long meta_max; 372 373 meta_max = type_max(typeof_member(struct skb_shared_info, meta_len)); 374 BUILD_BUG_ON(!__builtin_constant_p(meta_max)); 375 376 return !IS_ALIGNED(metalen, sizeof(u32)) || metalen > meta_max; 377 } 378 379 struct xdp_attachment_info { 380 struct bpf_prog *prog; 381 u32 flags; 382 }; 383 384 struct netdev_bpf; 385 void xdp_attachment_setup(struct xdp_attachment_info *info, 386 struct netdev_bpf *bpf); 387 388 #define DEV_MAP_BULK_SIZE XDP_BULK_QUEUE_SIZE 389 390 /* Define the relationship between xdp-rx-metadata kfunc and 391 * various other entities: 392 * - xdp_rx_metadata enum 393 * - netdev netlink enum (Documentation/netlink/specs/netdev.yaml) 394 * - kfunc name 395 * - xdp_metadata_ops field 396 */ 397 #define XDP_METADATA_KFUNC_xxx \ 398 XDP_METADATA_KFUNC(XDP_METADATA_KFUNC_RX_TIMESTAMP, \ 399 NETDEV_XDP_RX_METADATA_TIMESTAMP, \ 400 bpf_xdp_metadata_rx_timestamp, \ 401 xmo_rx_timestamp) \ 402 XDP_METADATA_KFUNC(XDP_METADATA_KFUNC_RX_HASH, \ 403 NETDEV_XDP_RX_METADATA_HASH, \ 404 bpf_xdp_metadata_rx_hash, \ 405 xmo_rx_hash) \ 406 XDP_METADATA_KFUNC(XDP_METADATA_KFUNC_RX_VLAN_TAG, \ 407 NETDEV_XDP_RX_METADATA_VLAN_TAG, \ 408 bpf_xdp_metadata_rx_vlan_tag, \ 409 xmo_rx_vlan_tag) \ 410 411 enum xdp_rx_metadata { 412 #define XDP_METADATA_KFUNC(name, _, __, ___) name, 413 XDP_METADATA_KFUNC_xxx 414 #undef XDP_METADATA_KFUNC 415 MAX_XDP_METADATA_KFUNC, 416 }; 417 418 enum xdp_rss_hash_type { 419 /* First part: Individual bits for L3/L4 types */ 420 XDP_RSS_L3_IPV4 = BIT(0), 421 XDP_RSS_L3_IPV6 = BIT(1), 422 423 /* The fixed (L3) IPv4 and IPv6 headers can both be followed by 424 * variable/dynamic headers, IPv4 called Options and IPv6 called 425 * Extension Headers. HW RSS type can contain this info. 426 */ 427 XDP_RSS_L3_DYNHDR = BIT(2), 428 429 /* When RSS hash covers L4 then drivers MUST set XDP_RSS_L4 bit in 430 * addition to the protocol specific bit. This ease interaction with 431 * SKBs and avoids reserving a fixed mask for future L4 protocol bits. 432 */ 433 XDP_RSS_L4 = BIT(3), /* L4 based hash, proto can be unknown */ 434 XDP_RSS_L4_TCP = BIT(4), 435 XDP_RSS_L4_UDP = BIT(5), 436 XDP_RSS_L4_SCTP = BIT(6), 437 XDP_RSS_L4_IPSEC = BIT(7), /* L4 based hash include IPSEC SPI */ 438 XDP_RSS_L4_ICMP = BIT(8), 439 440 /* Second part: RSS hash type combinations used for driver HW mapping */ 441 XDP_RSS_TYPE_NONE = 0, 442 XDP_RSS_TYPE_L2 = XDP_RSS_TYPE_NONE, 443 444 XDP_RSS_TYPE_L3_IPV4 = XDP_RSS_L3_IPV4, 445 XDP_RSS_TYPE_L3_IPV6 = XDP_RSS_L3_IPV6, 446 XDP_RSS_TYPE_L3_IPV4_OPT = XDP_RSS_L3_IPV4 | XDP_RSS_L3_DYNHDR, 447 XDP_RSS_TYPE_L3_IPV6_EX = XDP_RSS_L3_IPV6 | XDP_RSS_L3_DYNHDR, 448 449 XDP_RSS_TYPE_L4_ANY = XDP_RSS_L4, 450 XDP_RSS_TYPE_L4_IPV4_TCP = XDP_RSS_L3_IPV4 | XDP_RSS_L4 | XDP_RSS_L4_TCP, 451 XDP_RSS_TYPE_L4_IPV4_UDP = XDP_RSS_L3_IPV4 | XDP_RSS_L4 | XDP_RSS_L4_UDP, 452 XDP_RSS_TYPE_L4_IPV4_SCTP = XDP_RSS_L3_IPV4 | XDP_RSS_L4 | XDP_RSS_L4_SCTP, 453 XDP_RSS_TYPE_L4_IPV4_IPSEC = XDP_RSS_L3_IPV4 | XDP_RSS_L4 | XDP_RSS_L4_IPSEC, 454 XDP_RSS_TYPE_L4_IPV4_ICMP = XDP_RSS_L3_IPV4 | XDP_RSS_L4 | XDP_RSS_L4_ICMP, 455 456 XDP_RSS_TYPE_L4_IPV6_TCP = XDP_RSS_L3_IPV6 | XDP_RSS_L4 | XDP_RSS_L4_TCP, 457 XDP_RSS_TYPE_L4_IPV6_UDP = XDP_RSS_L3_IPV6 | XDP_RSS_L4 | XDP_RSS_L4_UDP, 458 XDP_RSS_TYPE_L4_IPV6_SCTP = XDP_RSS_L3_IPV6 | XDP_RSS_L4 | XDP_RSS_L4_SCTP, 459 XDP_RSS_TYPE_L4_IPV6_IPSEC = XDP_RSS_L3_IPV6 | XDP_RSS_L4 | XDP_RSS_L4_IPSEC, 460 XDP_RSS_TYPE_L4_IPV6_ICMP = XDP_RSS_L3_IPV6 | XDP_RSS_L4 | XDP_RSS_L4_ICMP, 461 462 XDP_RSS_TYPE_L4_IPV6_TCP_EX = XDP_RSS_TYPE_L4_IPV6_TCP | XDP_RSS_L3_DYNHDR, 463 XDP_RSS_TYPE_L4_IPV6_UDP_EX = XDP_RSS_TYPE_L4_IPV6_UDP | XDP_RSS_L3_DYNHDR, 464 XDP_RSS_TYPE_L4_IPV6_SCTP_EX = XDP_RSS_TYPE_L4_IPV6_SCTP | XDP_RSS_L3_DYNHDR, 465 }; 466 467 struct xdp_metadata_ops { 468 int (*xmo_rx_timestamp)(const struct xdp_md *ctx, u64 *timestamp); 469 int (*xmo_rx_hash)(const struct xdp_md *ctx, u32 *hash, 470 enum xdp_rss_hash_type *rss_type); 471 int (*xmo_rx_vlan_tag)(const struct xdp_md *ctx, __be16 *vlan_proto, 472 u16 *vlan_tci); 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 static __always_inline u32 bpf_prog_run_xdp(const struct bpf_prog *prog, 507 struct xdp_buff *xdp) 508 { 509 /* Driver XDP hooks are invoked within a single NAPI poll cycle and thus 510 * under local_bh_disable(), which provides the needed RCU protection 511 * for accessing map entries. 512 */ 513 u32 act = __bpf_prog_run(prog, xdp, BPF_DISPATCHER_FUNC(xdp)); 514 515 if (static_branch_unlikely(&bpf_master_redirect_enabled_key)) { 516 if (act == XDP_TX && netif_is_bond_slave(xdp->rxq->dev)) 517 act = xdp_master_redirect(xdp); 518 } 519 520 return act; 521 } 522 #endif /* __LINUX_NET_XDP_H__ */ 523