1 // SPDX-License-Identifier: GPL-2.0-only 2 /* net/core/xdp.c 3 * 4 * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc. 5 */ 6 #include <linux/bpf.h> 7 #include <linux/filter.h> 8 #include <linux/types.h> 9 #include <linux/mm.h> 10 #include <linux/netdevice.h> 11 #include <linux/slab.h> 12 #include <linux/idr.h> 13 #include <linux/rhashtable.h> 14 #include <linux/bug.h> 15 #include <net/page_pool.h> 16 17 #include <net/xdp.h> 18 #include <net/xdp_priv.h> /* struct xdp_mem_allocator */ 19 #include <trace/events/xdp.h> 20 #include <net/xdp_sock_drv.h> 21 22 #define REG_STATE_NEW 0x0 23 #define REG_STATE_REGISTERED 0x1 24 #define REG_STATE_UNREGISTERED 0x2 25 #define REG_STATE_UNUSED 0x3 26 27 static DEFINE_IDA(mem_id_pool); 28 static DEFINE_MUTEX(mem_id_lock); 29 #define MEM_ID_MAX 0xFFFE 30 #define MEM_ID_MIN 1 31 static int mem_id_next = MEM_ID_MIN; 32 33 static bool mem_id_init; /* false */ 34 static struct rhashtable *mem_id_ht; 35 36 static u32 xdp_mem_id_hashfn(const void *data, u32 len, u32 seed) 37 { 38 const u32 *k = data; 39 const u32 key = *k; 40 41 BUILD_BUG_ON(sizeof_field(struct xdp_mem_allocator, mem.id) 42 != sizeof(u32)); 43 44 /* Use cyclic increasing ID as direct hash key */ 45 return key; 46 } 47 48 static int xdp_mem_id_cmp(struct rhashtable_compare_arg *arg, 49 const void *ptr) 50 { 51 const struct xdp_mem_allocator *xa = ptr; 52 u32 mem_id = *(u32 *)arg->key; 53 54 return xa->mem.id != mem_id; 55 } 56 57 static const struct rhashtable_params mem_id_rht_params = { 58 .nelem_hint = 64, 59 .head_offset = offsetof(struct xdp_mem_allocator, node), 60 .key_offset = offsetof(struct xdp_mem_allocator, mem.id), 61 .key_len = sizeof_field(struct xdp_mem_allocator, mem.id), 62 .max_size = MEM_ID_MAX, 63 .min_size = 8, 64 .automatic_shrinking = true, 65 .hashfn = xdp_mem_id_hashfn, 66 .obj_cmpfn = xdp_mem_id_cmp, 67 }; 68 69 static void __xdp_mem_allocator_rcu_free(struct rcu_head *rcu) 70 { 71 struct xdp_mem_allocator *xa; 72 73 xa = container_of(rcu, struct xdp_mem_allocator, rcu); 74 75 /* Allow this ID to be reused */ 76 ida_simple_remove(&mem_id_pool, xa->mem.id); 77 78 kfree(xa); 79 } 80 81 static void mem_xa_remove(struct xdp_mem_allocator *xa) 82 { 83 trace_mem_disconnect(xa); 84 85 if (!rhashtable_remove_fast(mem_id_ht, &xa->node, mem_id_rht_params)) 86 call_rcu(&xa->rcu, __xdp_mem_allocator_rcu_free); 87 } 88 89 static void mem_allocator_disconnect(void *allocator) 90 { 91 struct xdp_mem_allocator *xa; 92 struct rhashtable_iter iter; 93 94 mutex_lock(&mem_id_lock); 95 96 rhashtable_walk_enter(mem_id_ht, &iter); 97 do { 98 rhashtable_walk_start(&iter); 99 100 while ((xa = rhashtable_walk_next(&iter)) && !IS_ERR(xa)) { 101 if (xa->allocator == allocator) 102 mem_xa_remove(xa); 103 } 104 105 rhashtable_walk_stop(&iter); 106 107 } while (xa == ERR_PTR(-EAGAIN)); 108 rhashtable_walk_exit(&iter); 109 110 mutex_unlock(&mem_id_lock); 111 } 112 113 void xdp_unreg_mem_model(struct xdp_mem_info *mem) 114 { 115 struct xdp_mem_allocator *xa; 116 int type = mem->type; 117 int id = mem->id; 118 119 /* Reset mem info to defaults */ 120 mem->id = 0; 121 mem->type = 0; 122 123 if (id == 0) 124 return; 125 126 if (type == MEM_TYPE_PAGE_POOL) { 127 rcu_read_lock(); 128 xa = rhashtable_lookup(mem_id_ht, &id, mem_id_rht_params); 129 page_pool_destroy(xa->page_pool); 130 rcu_read_unlock(); 131 } 132 } 133 EXPORT_SYMBOL_GPL(xdp_unreg_mem_model); 134 135 void xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq) 136 { 137 if (xdp_rxq->reg_state != REG_STATE_REGISTERED) { 138 WARN(1, "Missing register, driver bug"); 139 return; 140 } 141 142 xdp_unreg_mem_model(&xdp_rxq->mem); 143 } 144 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg_mem_model); 145 146 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq) 147 { 148 /* Simplify driver cleanup code paths, allow unreg "unused" */ 149 if (xdp_rxq->reg_state == REG_STATE_UNUSED) 150 return; 151 152 xdp_rxq_info_unreg_mem_model(xdp_rxq); 153 154 xdp_rxq->reg_state = REG_STATE_UNREGISTERED; 155 xdp_rxq->dev = NULL; 156 } 157 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg); 158 159 static void xdp_rxq_info_init(struct xdp_rxq_info *xdp_rxq) 160 { 161 memset(xdp_rxq, 0, sizeof(*xdp_rxq)); 162 } 163 164 /* Returns 0 on success, negative on failure */ 165 int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq, 166 struct net_device *dev, u32 queue_index, unsigned int napi_id) 167 { 168 if (!dev) { 169 WARN(1, "Missing net_device from driver"); 170 return -ENODEV; 171 } 172 173 if (xdp_rxq->reg_state == REG_STATE_UNUSED) { 174 WARN(1, "Driver promised not to register this"); 175 return -EINVAL; 176 } 177 178 if (xdp_rxq->reg_state == REG_STATE_REGISTERED) { 179 WARN(1, "Missing unregister, handled but fix driver"); 180 xdp_rxq_info_unreg(xdp_rxq); 181 } 182 183 /* State either UNREGISTERED or NEW */ 184 xdp_rxq_info_init(xdp_rxq); 185 xdp_rxq->dev = dev; 186 xdp_rxq->queue_index = queue_index; 187 xdp_rxq->napi_id = napi_id; 188 189 xdp_rxq->reg_state = REG_STATE_REGISTERED; 190 return 0; 191 } 192 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg); 193 194 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq) 195 { 196 xdp_rxq->reg_state = REG_STATE_UNUSED; 197 } 198 EXPORT_SYMBOL_GPL(xdp_rxq_info_unused); 199 200 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq) 201 { 202 return (xdp_rxq->reg_state == REG_STATE_REGISTERED); 203 } 204 EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg); 205 206 static int __mem_id_init_hash_table(void) 207 { 208 struct rhashtable *rht; 209 int ret; 210 211 if (unlikely(mem_id_init)) 212 return 0; 213 214 rht = kzalloc(sizeof(*rht), GFP_KERNEL); 215 if (!rht) 216 return -ENOMEM; 217 218 ret = rhashtable_init(rht, &mem_id_rht_params); 219 if (ret < 0) { 220 kfree(rht); 221 return ret; 222 } 223 mem_id_ht = rht; 224 smp_mb(); /* mutex lock should provide enough pairing */ 225 mem_id_init = true; 226 227 return 0; 228 } 229 230 /* Allocate a cyclic ID that maps to allocator pointer. 231 * See: https://www.kernel.org/doc/html/latest/core-api/idr.html 232 * 233 * Caller must lock mem_id_lock. 234 */ 235 static int __mem_id_cyclic_get(gfp_t gfp) 236 { 237 int retries = 1; 238 int id; 239 240 again: 241 id = ida_simple_get(&mem_id_pool, mem_id_next, MEM_ID_MAX, gfp); 242 if (id < 0) { 243 if (id == -ENOSPC) { 244 /* Cyclic allocator, reset next id */ 245 if (retries--) { 246 mem_id_next = MEM_ID_MIN; 247 goto again; 248 } 249 } 250 return id; /* errno */ 251 } 252 mem_id_next = id + 1; 253 254 return id; 255 } 256 257 static bool __is_supported_mem_type(enum xdp_mem_type type) 258 { 259 if (type == MEM_TYPE_PAGE_POOL) 260 return is_page_pool_compiled_in(); 261 262 if (type >= MEM_TYPE_MAX) 263 return false; 264 265 return true; 266 } 267 268 static struct xdp_mem_allocator *__xdp_reg_mem_model(struct xdp_mem_info *mem, 269 enum xdp_mem_type type, 270 void *allocator) 271 { 272 struct xdp_mem_allocator *xdp_alloc; 273 gfp_t gfp = GFP_KERNEL; 274 int id, errno, ret; 275 void *ptr; 276 277 if (!__is_supported_mem_type(type)) 278 return ERR_PTR(-EOPNOTSUPP); 279 280 mem->type = type; 281 282 if (!allocator) { 283 if (type == MEM_TYPE_PAGE_POOL) 284 return ERR_PTR(-EINVAL); /* Setup time check page_pool req */ 285 return NULL; 286 } 287 288 /* Delay init of rhashtable to save memory if feature isn't used */ 289 if (!mem_id_init) { 290 mutex_lock(&mem_id_lock); 291 ret = __mem_id_init_hash_table(); 292 mutex_unlock(&mem_id_lock); 293 if (ret < 0) { 294 WARN_ON(1); 295 return ERR_PTR(ret); 296 } 297 } 298 299 xdp_alloc = kzalloc(sizeof(*xdp_alloc), gfp); 300 if (!xdp_alloc) 301 return ERR_PTR(-ENOMEM); 302 303 mutex_lock(&mem_id_lock); 304 id = __mem_id_cyclic_get(gfp); 305 if (id < 0) { 306 errno = id; 307 goto err; 308 } 309 mem->id = id; 310 xdp_alloc->mem = *mem; 311 xdp_alloc->allocator = allocator; 312 313 /* Insert allocator into ID lookup table */ 314 ptr = rhashtable_insert_slow(mem_id_ht, &id, &xdp_alloc->node); 315 if (IS_ERR(ptr)) { 316 ida_simple_remove(&mem_id_pool, mem->id); 317 mem->id = 0; 318 errno = PTR_ERR(ptr); 319 goto err; 320 } 321 322 if (type == MEM_TYPE_PAGE_POOL) 323 page_pool_use_xdp_mem(allocator, mem_allocator_disconnect, mem); 324 325 mutex_unlock(&mem_id_lock); 326 327 return xdp_alloc; 328 err: 329 mutex_unlock(&mem_id_lock); 330 kfree(xdp_alloc); 331 return ERR_PTR(errno); 332 } 333 334 int xdp_reg_mem_model(struct xdp_mem_info *mem, 335 enum xdp_mem_type type, void *allocator) 336 { 337 struct xdp_mem_allocator *xdp_alloc; 338 339 xdp_alloc = __xdp_reg_mem_model(mem, type, allocator); 340 if (IS_ERR(xdp_alloc)) 341 return PTR_ERR(xdp_alloc); 342 return 0; 343 } 344 EXPORT_SYMBOL_GPL(xdp_reg_mem_model); 345 346 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq, 347 enum xdp_mem_type type, void *allocator) 348 { 349 struct xdp_mem_allocator *xdp_alloc; 350 351 if (xdp_rxq->reg_state != REG_STATE_REGISTERED) { 352 WARN(1, "Missing register, driver bug"); 353 return -EFAULT; 354 } 355 356 xdp_alloc = __xdp_reg_mem_model(&xdp_rxq->mem, type, allocator); 357 if (IS_ERR(xdp_alloc)) 358 return PTR_ERR(xdp_alloc); 359 360 if (trace_mem_connect_enabled() && xdp_alloc) 361 trace_mem_connect(xdp_alloc, xdp_rxq); 362 return 0; 363 } 364 365 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model); 366 367 /* XDP RX runs under NAPI protection, and in different delivery error 368 * scenarios (e.g. queue full), it is possible to return the xdp_frame 369 * while still leveraging this protection. The @napi_direct boolean 370 * is used for those calls sites. Thus, allowing for faster recycling 371 * of xdp_frames/pages in those cases. 372 */ 373 static void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct, 374 struct xdp_buff *xdp) 375 { 376 struct xdp_mem_allocator *xa; 377 struct page *page; 378 379 switch (mem->type) { 380 case MEM_TYPE_PAGE_POOL: 381 rcu_read_lock(); 382 /* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */ 383 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params); 384 page = virt_to_head_page(data); 385 if (napi_direct && xdp_return_frame_no_direct()) 386 napi_direct = false; 387 page_pool_put_full_page(xa->page_pool, page, napi_direct); 388 rcu_read_unlock(); 389 break; 390 case MEM_TYPE_PAGE_SHARED: 391 page_frag_free(data); 392 break; 393 case MEM_TYPE_PAGE_ORDER0: 394 page = virt_to_page(data); /* Assumes order0 page*/ 395 put_page(page); 396 break; 397 case MEM_TYPE_XSK_BUFF_POOL: 398 /* NB! Only valid from an xdp_buff! */ 399 xsk_buff_free(xdp); 400 break; 401 default: 402 /* Not possible, checked in xdp_rxq_info_reg_mem_model() */ 403 WARN(1, "Incorrect XDP memory type (%d) usage", mem->type); 404 break; 405 } 406 } 407 408 void xdp_return_frame(struct xdp_frame *xdpf) 409 { 410 __xdp_return(xdpf->data, &xdpf->mem, false, NULL); 411 } 412 EXPORT_SYMBOL_GPL(xdp_return_frame); 413 414 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf) 415 { 416 __xdp_return(xdpf->data, &xdpf->mem, true, NULL); 417 } 418 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi); 419 420 /* XDP bulk APIs introduce a defer/flush mechanism to return 421 * pages belonging to the same xdp_mem_allocator object 422 * (identified via the mem.id field) in bulk to optimize 423 * I-cache and D-cache. 424 * The bulk queue size is set to 16 to be aligned to how 425 * XDP_REDIRECT bulking works. The bulk is flushed when 426 * it is full or when mem.id changes. 427 * xdp_frame_bulk is usually stored/allocated on the function 428 * call-stack to avoid locking penalties. 429 */ 430 void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq) 431 { 432 struct xdp_mem_allocator *xa = bq->xa; 433 434 if (unlikely(!xa || !bq->count)) 435 return; 436 437 page_pool_put_page_bulk(xa->page_pool, bq->q, bq->count); 438 /* bq->xa is not cleared to save lookup, if mem.id same in next bulk */ 439 bq->count = 0; 440 } 441 EXPORT_SYMBOL_GPL(xdp_flush_frame_bulk); 442 443 /* Must be called with rcu_read_lock held */ 444 void xdp_return_frame_bulk(struct xdp_frame *xdpf, 445 struct xdp_frame_bulk *bq) 446 { 447 struct xdp_mem_info *mem = &xdpf->mem; 448 struct xdp_mem_allocator *xa; 449 450 if (mem->type != MEM_TYPE_PAGE_POOL) { 451 __xdp_return(xdpf->data, &xdpf->mem, false, NULL); 452 return; 453 } 454 455 xa = bq->xa; 456 if (unlikely(!xa)) { 457 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params); 458 bq->count = 0; 459 bq->xa = xa; 460 } 461 462 if (bq->count == XDP_BULK_QUEUE_SIZE) 463 xdp_flush_frame_bulk(bq); 464 465 if (unlikely(mem->id != xa->mem.id)) { 466 xdp_flush_frame_bulk(bq); 467 bq->xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params); 468 } 469 470 bq->q[bq->count++] = xdpf->data; 471 } 472 EXPORT_SYMBOL_GPL(xdp_return_frame_bulk); 473 474 void xdp_return_buff(struct xdp_buff *xdp) 475 { 476 __xdp_return(xdp->data, &xdp->rxq->mem, true, xdp); 477 } 478 479 /* Only called for MEM_TYPE_PAGE_POOL see xdp.h */ 480 void __xdp_release_frame(void *data, struct xdp_mem_info *mem) 481 { 482 struct xdp_mem_allocator *xa; 483 struct page *page; 484 485 rcu_read_lock(); 486 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params); 487 page = virt_to_head_page(data); 488 if (xa) 489 page_pool_release_page(xa->page_pool, page); 490 rcu_read_unlock(); 491 } 492 EXPORT_SYMBOL_GPL(__xdp_release_frame); 493 494 void xdp_attachment_setup(struct xdp_attachment_info *info, 495 struct netdev_bpf *bpf) 496 { 497 if (info->prog) 498 bpf_prog_put(info->prog); 499 info->prog = bpf->prog; 500 info->flags = bpf->flags; 501 } 502 EXPORT_SYMBOL_GPL(xdp_attachment_setup); 503 504 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp) 505 { 506 unsigned int metasize, totsize; 507 void *addr, *data_to_copy; 508 struct xdp_frame *xdpf; 509 struct page *page; 510 511 /* Clone into a MEM_TYPE_PAGE_ORDER0 xdp_frame. */ 512 metasize = xdp_data_meta_unsupported(xdp) ? 0 : 513 xdp->data - xdp->data_meta; 514 totsize = xdp->data_end - xdp->data + metasize; 515 516 if (sizeof(*xdpf) + totsize > PAGE_SIZE) 517 return NULL; 518 519 page = dev_alloc_page(); 520 if (!page) 521 return NULL; 522 523 addr = page_to_virt(page); 524 xdpf = addr; 525 memset(xdpf, 0, sizeof(*xdpf)); 526 527 addr += sizeof(*xdpf); 528 data_to_copy = metasize ? xdp->data_meta : xdp->data; 529 memcpy(addr, data_to_copy, totsize); 530 531 xdpf->data = addr + metasize; 532 xdpf->len = totsize - metasize; 533 xdpf->headroom = 0; 534 xdpf->metasize = metasize; 535 xdpf->frame_sz = PAGE_SIZE; 536 xdpf->mem.type = MEM_TYPE_PAGE_ORDER0; 537 538 xsk_buff_free(xdp); 539 return xdpf; 540 } 541 EXPORT_SYMBOL_GPL(xdp_convert_zc_to_xdp_frame); 542 543 /* Used by XDP_WARN macro, to avoid inlining WARN() in fast-path */ 544 void xdp_warn(const char *msg, const char *func, const int line) 545 { 546 WARN(1, "XDP_WARN: %s(line:%d): %s\n", func, line, msg); 547 }; 548 EXPORT_SYMBOL_GPL(xdp_warn); 549 550 int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp) 551 { 552 n_skb = kmem_cache_alloc_bulk(skbuff_head_cache, gfp, 553 n_skb, skbs); 554 if (unlikely(!n_skb)) 555 return -ENOMEM; 556 557 return 0; 558 } 559 EXPORT_SYMBOL_GPL(xdp_alloc_skb_bulk); 560 561 struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf, 562 struct sk_buff *skb, 563 struct net_device *dev) 564 { 565 unsigned int headroom, frame_size; 566 void *hard_start; 567 568 /* Part of headroom was reserved to xdpf */ 569 headroom = sizeof(*xdpf) + xdpf->headroom; 570 571 /* Memory size backing xdp_frame data already have reserved 572 * room for build_skb to place skb_shared_info in tailroom. 573 */ 574 frame_size = xdpf->frame_sz; 575 576 hard_start = xdpf->data - headroom; 577 skb = build_skb_around(skb, hard_start, frame_size); 578 if (unlikely(!skb)) 579 return NULL; 580 581 skb_reserve(skb, headroom); 582 __skb_put(skb, xdpf->len); 583 if (xdpf->metasize) 584 skb_metadata_set(skb, xdpf->metasize); 585 586 /* Essential SKB info: protocol and skb->dev */ 587 skb->protocol = eth_type_trans(skb, dev); 588 589 /* Optional SKB info, currently missing: 590 * - HW checksum info (skb->ip_summed) 591 * - HW RX hash (skb_set_hash) 592 * - RX ring dev queue index (skb_record_rx_queue) 593 */ 594 595 /* Until page_pool get SKB return path, release DMA here */ 596 xdp_release_frame(xdpf); 597 598 /* Allow SKB to reuse area used by xdp_frame */ 599 xdp_scrub_frame(xdpf); 600 601 return skb; 602 } 603 EXPORT_SYMBOL_GPL(__xdp_build_skb_from_frame); 604 605 struct sk_buff *xdp_build_skb_from_frame(struct xdp_frame *xdpf, 606 struct net_device *dev) 607 { 608 struct sk_buff *skb; 609 610 skb = kmem_cache_alloc(skbuff_head_cache, GFP_ATOMIC); 611 if (unlikely(!skb)) 612 return NULL; 613 614 memset(skb, 0, offsetof(struct sk_buff, tail)); 615 616 return __xdp_build_skb_from_frame(xdpf, skb, dev); 617 } 618 EXPORT_SYMBOL_GPL(xdp_build_skb_from_frame); 619 620 struct xdp_frame *xdpf_clone(struct xdp_frame *xdpf) 621 { 622 unsigned int headroom, totalsize; 623 struct xdp_frame *nxdpf; 624 struct page *page; 625 void *addr; 626 627 headroom = xdpf->headroom + sizeof(*xdpf); 628 totalsize = headroom + xdpf->len; 629 630 if (unlikely(totalsize > PAGE_SIZE)) 631 return NULL; 632 page = dev_alloc_page(); 633 if (!page) 634 return NULL; 635 addr = page_to_virt(page); 636 637 memcpy(addr, xdpf, totalsize); 638 639 nxdpf = addr; 640 nxdpf->data = addr + headroom; 641 nxdpf->frame_sz = PAGE_SIZE; 642 nxdpf->mem.type = MEM_TYPE_PAGE_ORDER0; 643 nxdpf->mem.id = 0; 644 645 return nxdpf; 646 } 647