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 41enum 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 53struct xdp_mem_info { 54 u32 type; /* enum xdp_mem_type, but known size type */ 55 u32 id; 56}; 57 58struct page_pool; 59 60struct xdp_rxq_info { 61 struct net_device *dev; 62 u32 queue_index; 63 u32 reg_state; 64 struct xdp_mem_info mem; 65 unsigned int napi_id; 66 u32 frag_size; 67} ____cacheline_aligned; /* perf critical, avoid false-sharing */ 68 69struct xdp_txq_info { 70 struct net_device *dev; 71}; 72 73enum xdp_buff_flags { 74 XDP_FLAGS_HAS_FRAGS = BIT(0), /* non-linear xdp buff */ 75 XDP_FLAGS_FRAGS_PF_MEMALLOC = BIT(1), /* xdp paged memory is under 76 * pressure 77 */ 78}; 79 80struct xdp_buff { 81 void *data; 82 void *data_end; 83 void *data_meta; 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 */ 89}; 90 91static __always_inline bool xdp_buff_has_frags(struct xdp_buff *xdp) 92{ 93 return !!(xdp->flags & XDP_FLAGS_HAS_FRAGS); 94} 95 96static __always_inline void xdp_buff_set_frags_flag(struct xdp_buff *xdp) 97{ 98 xdp->flags |= XDP_FLAGS_HAS_FRAGS; 99} 100 101static __always_inline void xdp_buff_clear_frags_flag(struct xdp_buff *xdp) 102{ 103 xdp->flags &= ~XDP_FLAGS_HAS_FRAGS; 104} 105 106static __always_inline bool xdp_buff_is_frag_pfmemalloc(struct xdp_buff *xdp) 107{ 108 return !!(xdp->flags & XDP_FLAGS_FRAGS_PF_MEMALLOC); 109} 110 111static __always_inline void xdp_buff_set_frag_pfmemalloc(struct xdp_buff *xdp) 112{ 113 xdp->flags |= XDP_FLAGS_FRAGS_PF_MEMALLOC; 114} 115 116static __always_inline void 117xdp_init_buff(struct xdp_buff *xdp, u32 frame_sz, struct xdp_rxq_info *rxq) 118{ 119 xdp->frame_sz = frame_sz; 120 xdp->rxq = rxq; 121 xdp->flags = 0; 122} 123 124static __always_inline void 125xdp_prepare_buff(struct xdp_buff *xdp, unsigned char *hard_start, 126 int headroom, int data_len, const bool meta_valid) 127{ 128 unsigned char *data = hard_start + headroom; 129 130 xdp->data_hard_start = hard_start; 131 xdp->data = data; 132 xdp->data_end = data + data_len; 133 xdp->data_meta = meta_valid ? data : data + 1; 134} 135 136/* Reserve memory area at end-of data area. 137 * 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(). 141 */ 142#define xdp_data_hard_end(xdp) \ 143 ((xdp)->data_hard_start + (xdp)->frame_sz - \ 144 SKB_DATA_ALIGN(sizeof(struct skb_shared_info))) 145 146static inline struct skb_shared_info * 147xdp_get_shared_info_from_buff(struct xdp_buff *xdp) 148{ 149 return (struct skb_shared_info *)xdp_data_hard_end(xdp); 150} 151 152static __always_inline unsigned int xdp_get_buff_len(struct xdp_buff *xdp) 153{ 154 unsigned int len = xdp->data_end - xdp->data; 155 struct skb_shared_info *sinfo; 156 157 if (likely(!xdp_buff_has_frags(xdp))) 158 goto out; 159 160 sinfo = xdp_get_shared_info_from_buff(xdp); 161 len += sinfo->xdp_frags_size; 162out: 163 return len; 164} 165 166struct xdp_frame { 167 void *data; 168 u16 len; 169 u16 headroom; 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. 173 */ 174 struct xdp_mem_info mem; 175 struct net_device *dev_rx; /* used by cpumap */ 176 u32 frame_sz; 177 u32 flags; /* supported values defined in xdp_buff_flags */ 178}; 179 180static __always_inline bool xdp_frame_has_frags(struct xdp_frame *frame) 181{ 182 return !!(frame->flags & XDP_FLAGS_HAS_FRAGS); 183} 184 185static __always_inline bool xdp_frame_is_frag_pfmemalloc(struct xdp_frame *frame) 186{ 187 return !!(frame->flags & XDP_FLAGS_FRAGS_PF_MEMALLOC); 188} 189 190#define XDP_BULK_QUEUE_SIZE 16 191struct xdp_frame_bulk { 192 int count; 193 void *xa; 194 void *q[XDP_BULK_QUEUE_SIZE]; 195}; 196 197static __always_inline void xdp_frame_bulk_init(struct xdp_frame_bulk *bq) 198{ 199 /* bq->count will be zero'ed when bq->xa gets updated */ 200 bq->xa = NULL; 201} 202 203static inline struct skb_shared_info * 204xdp_get_shared_info_from_frame(struct xdp_frame *frame) 205{ 206 void *data_hard_start = frame->data - frame->headroom - sizeof(*frame); 207 208 return (struct skb_shared_info *)(data_hard_start + frame->frame_sz - 209 SKB_DATA_ALIGN(sizeof(struct skb_shared_info))); 210} 211 212struct xdp_cpumap_stats { 213 unsigned int redirect; 214 unsigned int pass; 215 unsigned int drop; 216}; 217 218/* Clear kernel pointers in xdp_frame */ 219static inline void xdp_scrub_frame(struct xdp_frame *frame) 220{ 221 frame->data = NULL; 222 frame->dev_rx = NULL; 223} 224 225static inline void 226xdp_update_skb_shared_info(struct sk_buff *skb, u8 nr_frags, 227 unsigned int size, unsigned int truesize, 228 bool pfmemalloc) 229{ 230 skb_shinfo(skb)->nr_frags = nr_frags; 231 232 skb->len += size; 233 skb->data_len += size; 234 skb->truesize += truesize; 235 skb->pfmemalloc |= pfmemalloc; 236} 237 238/* Avoids inlining WARN macro in fast-path */ 239void xdp_warn(const char *msg, const char *func, const int line); 240#define XDP_WARN(msg) xdp_warn(msg, __func__, __LINE__) 241 242struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp); 243struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf, 244 struct sk_buff *skb, 245 struct net_device *dev); 246struct sk_buff *xdp_build_skb_from_frame(struct xdp_frame *xdpf, 247 struct net_device *dev); 248int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp); 249struct xdp_frame *xdpf_clone(struct xdp_frame *xdpf); 250 251static inline 252void xdp_convert_frame_to_buff(struct xdp_frame *frame, struct xdp_buff *xdp) 253{ 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; 260} 261 262static inline 263int xdp_update_frame_from_buff(struct xdp_buff *xdp, 264 struct xdp_frame *xdp_frame) 265{ 266 int metasize, headroom; 267 268 /* 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))) 273 return -ENOSPC; 274 275 /* 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"); 278 return -ENOSPC; 279 } 280 281 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; 287 288 return 0; 289} 290 291/* Convert xdp_buff to xdp_frame */ 292static inline 293struct xdp_frame *xdp_convert_buff_to_frame(struct xdp_buff *xdp) 294{ 295 struct xdp_frame *xdp_frame; 296 297 if (xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL) 298 return xdp_convert_zc_to_xdp_frame(xdp); 299 300 /* 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)) 303 return NULL; 304 305 /* rxq only valid until napi_schedule ends, convert to xdp_mem_info */ 306 xdp_frame->mem = xdp->rxq->mem; 307 308 return xdp_frame; 309} 310 311void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct, 312 struct xdp_buff *xdp); 313void xdp_return_frame(struct xdp_frame *xdpf); 314void xdp_return_frame_rx_napi(struct xdp_frame *xdpf); 315void xdp_return_buff(struct xdp_buff *xdp); 316void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq); 317void xdp_return_frame_bulk(struct xdp_frame *xdpf, 318 struct xdp_frame_bulk *bq); 319 320static __always_inline unsigned int xdp_get_frame_len(struct xdp_frame *xdpf) 321{ 322 struct skb_shared_info *sinfo; 323 unsigned int len = xdpf->len; 324 325 if (likely(!xdp_frame_has_frags(xdpf))) 326 goto out; 327 328 sinfo = xdp_get_shared_info_from_frame(xdpf); 329 len += sinfo->xdp_frags_size; 330out: 331 return len; 332} 333 334int __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); 337static inline int 338xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq, 339 struct net_device *dev, u32 queue_index, 340 unsigned int napi_id) 341{ 342 return __xdp_rxq_info_reg(xdp_rxq, dev, queue_index, napi_id, 0); 343} 344 345void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq); 346void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq); 347bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq); 348int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq, 349 enum xdp_mem_type type, void *allocator); 350void xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq); 351int xdp_reg_mem_model(struct xdp_mem_info *mem, 352 enum xdp_mem_type type, void *allocator); 353void xdp_unreg_mem_model(struct xdp_mem_info *mem); 354 355/* Drivers not supporting XDP metadata can use this helper, which 356 * rejects any room expansion for metadata as a result. 357 */ 358static __always_inline void 359xdp_set_data_meta_invalid(struct xdp_buff *xdp) 360{ 361 xdp->data_meta = xdp->data + 1; 362} 363 364static __always_inline bool 365xdp_data_meta_unsupported(const struct xdp_buff *xdp) 366{ 367 return unlikely(xdp->data_meta > xdp->data); 368} 369 370static inline bool xdp_metalen_invalid(unsigned long metalen) 371{ 372 unsigned long meta_max; 373 374 meta_max = type_max(typeof_member(struct skb_shared_info, meta_len)); 375 BUILD_BUG_ON(!__builtin_constant_p(meta_max)); 376 377 return !IS_ALIGNED(metalen, sizeof(u32)) || metalen > meta_max; 378} 379 380struct xdp_attachment_info { 381 struct bpf_prog *prog; 382 u32 flags; 383}; 384 385struct netdev_bpf; 386void xdp_attachment_setup(struct xdp_attachment_info *info, 387 struct netdev_bpf *bpf); 388 389#define DEV_MAP_BULK_SIZE XDP_BULK_QUEUE_SIZE 390 391/* Define the relationship between xdp-rx-metadata kfunc and 392 * various other entities: 393 * - xdp_rx_metadata enum 394 * - netdev netlink enum (Documentation/netlink/specs/netdev.yaml) 395 * - kfunc name 396 * - xdp_metadata_ops field 397 */ 398#define XDP_METADATA_KFUNC_xxx \ 399 XDP_METADATA_KFUNC(XDP_METADATA_KFUNC_RX_TIMESTAMP, \ 400 NETDEV_XDP_RX_METADATA_TIMESTAMP, \ 401 bpf_xdp_metadata_rx_timestamp, \ 402 xmo_rx_timestamp) \ 403 XDP_METADATA_KFUNC(XDP_METADATA_KFUNC_RX_HASH, \ 404 NETDEV_XDP_RX_METADATA_HASH, \ 405 bpf_xdp_metadata_rx_hash, \ 406 xmo_rx_hash) \ 407 XDP_METADATA_KFUNC(XDP_METADATA_KFUNC_RX_VLAN_TAG, \ 408 NETDEV_XDP_RX_METADATA_VLAN_TAG, \ 409 bpf_xdp_metadata_rx_vlan_tag, \ 410 xmo_rx_vlan_tag) \ 411 412enum xdp_rx_metadata { 413#define XDP_METADATA_KFUNC(name, _, __, ___) name, 414XDP_METADATA_KFUNC_xxx 415#undef XDP_METADATA_KFUNC 416MAX_XDP_METADATA_KFUNC, 417}; 418 419enum xdp_rss_hash_type { 420 /* First part: Individual bits for L3/L4 types */ 421 XDP_RSS_L3_IPV4 = BIT(0), 422 XDP_RSS_L3_IPV6 = BIT(1), 423 424 /* The fixed (L3) IPv4 and IPv6 headers can both be followed by 425 * variable/dynamic headers, IPv4 called Options and IPv6 called 426 * Extension Headers. HW RSS type can contain this info. 427 */ 428 XDP_RSS_L3_DYNHDR = BIT(2), 429 430 /* When RSS hash covers L4 then drivers MUST set XDP_RSS_L4 bit in 431 * addition to the protocol specific bit. This ease interaction with 432 * SKBs and avoids reserving a fixed mask for future L4 protocol bits. 433 */ 434 XDP_RSS_L4 = BIT(3), /* L4 based hash, proto can be unknown */ 435 XDP_RSS_L4_TCP = BIT(4), 436 XDP_RSS_L4_UDP = BIT(5), 437 XDP_RSS_L4_SCTP = BIT(6), 438 XDP_RSS_L4_IPSEC = BIT(7), /* L4 based hash include IPSEC SPI */ 439 XDP_RSS_L4_ICMP = BIT(8), 440 441 /* Second part: RSS hash type combinations used for driver HW mapping */ 442 XDP_RSS_TYPE_NONE = 0, 443 XDP_RSS_TYPE_L2 = XDP_RSS_TYPE_NONE, 444 445 XDP_RSS_TYPE_L3_IPV4 = XDP_RSS_L3_IPV4, 446 XDP_RSS_TYPE_L3_IPV6 = XDP_RSS_L3_IPV6, 447 XDP_RSS_TYPE_L3_IPV4_OPT = XDP_RSS_L3_IPV4 | XDP_RSS_L3_DYNHDR, 448 XDP_RSS_TYPE_L3_IPV6_EX = XDP_RSS_L3_IPV6 | XDP_RSS_L3_DYNHDR, 449 450 XDP_RSS_TYPE_L4_ANY = XDP_RSS_L4, 451 XDP_RSS_TYPE_L4_IPV4_TCP = XDP_RSS_L3_IPV4 | XDP_RSS_L4 | XDP_RSS_L4_TCP, 452 XDP_RSS_TYPE_L4_IPV4_UDP = XDP_RSS_L3_IPV4 | XDP_RSS_L4 | XDP_RSS_L4_UDP, 453 XDP_RSS_TYPE_L4_IPV4_SCTP = XDP_RSS_L3_IPV4 | XDP_RSS_L4 | XDP_RSS_L4_SCTP, 454 XDP_RSS_TYPE_L4_IPV4_IPSEC = XDP_RSS_L3_IPV4 | XDP_RSS_L4 | XDP_RSS_L4_IPSEC, 455 XDP_RSS_TYPE_L4_IPV4_ICMP = XDP_RSS_L3_IPV4 | XDP_RSS_L4 | XDP_RSS_L4_ICMP, 456 457 XDP_RSS_TYPE_L4_IPV6_TCP = XDP_RSS_L3_IPV6 | XDP_RSS_L4 | XDP_RSS_L4_TCP, 458 XDP_RSS_TYPE_L4_IPV6_UDP = XDP_RSS_L3_IPV6 | XDP_RSS_L4 | XDP_RSS_L4_UDP, 459 XDP_RSS_TYPE_L4_IPV6_SCTP = XDP_RSS_L3_IPV6 | XDP_RSS_L4 | XDP_RSS_L4_SCTP, 460 XDP_RSS_TYPE_L4_IPV6_IPSEC = XDP_RSS_L3_IPV6 | XDP_RSS_L4 | XDP_RSS_L4_IPSEC, 461 XDP_RSS_TYPE_L4_IPV6_ICMP = XDP_RSS_L3_IPV6 | XDP_RSS_L4 | XDP_RSS_L4_ICMP, 462 463 XDP_RSS_TYPE_L4_IPV6_TCP_EX = XDP_RSS_TYPE_L4_IPV6_TCP | XDP_RSS_L3_DYNHDR, 464 XDP_RSS_TYPE_L4_IPV6_UDP_EX = XDP_RSS_TYPE_L4_IPV6_UDP | XDP_RSS_L3_DYNHDR, 465 XDP_RSS_TYPE_L4_IPV6_SCTP_EX = XDP_RSS_TYPE_L4_IPV6_SCTP | XDP_RSS_L3_DYNHDR, 466}; 467 468struct xdp_metadata_ops { 469 int (*xmo_rx_timestamp)(const struct xdp_md *ctx, u64 *timestamp); 470 int (*xmo_rx_hash)(const struct xdp_md *ctx, u32 *hash, 471 enum xdp_rss_hash_type *rss_type); 472 int (*xmo_rx_vlan_tag)(const struct xdp_md *ctx, __be16 *vlan_proto, 473 u16 *vlan_tci); 474}; 475 476#ifdef CONFIG_NET 477u32 bpf_xdp_metadata_kfunc_id(int id); 478bool bpf_dev_bound_kfunc_id(u32 btf_id); 479void xdp_set_features_flag(struct net_device *dev, xdp_features_t val); 480void xdp_features_set_redirect_target(struct net_device *dev, bool support_sg); 481void xdp_features_clear_redirect_target(struct net_device *dev); 482#else 483static inline u32 bpf_xdp_metadata_kfunc_id(int id) { return 0; } 484static inline bool bpf_dev_bound_kfunc_id(u32 btf_id) { return false; } 485 486static inline void 487xdp_set_features_flag(struct net_device *dev, xdp_features_t val) 488{ 489} 490 491static inline void 492xdp_features_set_redirect_target(struct net_device *dev, bool support_sg) 493{ 494} 495 496static inline void 497xdp_features_clear_redirect_target(struct net_device *dev) 498{ 499} 500#endif 501 502static inline void xdp_clear_features_flag(struct net_device *dev) 503{ 504 xdp_set_features_flag(dev, 0); 505} 506 507static __always_inline u32 bpf_prog_run_xdp(const struct bpf_prog *prog, 508 struct xdp_buff *xdp) 509{ 510 /* Driver XDP hooks are invoked within a single NAPI poll cycle and thus 511 * under local_bh_disable(), which provides the needed RCU protection 512 * for accessing map entries. 513 */ 514 u32 act = __bpf_prog_run(prog, xdp, BPF_DISPATCHER_FUNC(xdp)); 515 516 if (static_branch_unlikely(&bpf_master_redirect_enabled_key)) { 517 if (act == XDP_TX && netif_is_bond_slave(xdp->rxq->dev)) 518 act = xdp_master_redirect(xdp); 519 } 520 521 return act; 522} 523#endif /* __LINUX_NET_XDP_H__ */ 524