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 trace_mem_connect(xdp_alloc, xdp_rxq); 361 return 0; 362 } 363 364 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model); 365 366 /* XDP RX runs under NAPI protection, and in different delivery error 367 * scenarios (e.g. queue full), it is possible to return the xdp_frame 368 * while still leveraging this protection. The @napi_direct boolean 369 * is used for those calls sites. Thus, allowing for faster recycling 370 * of xdp_frames/pages in those cases. 371 */ 372 static void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct, 373 struct xdp_buff *xdp) 374 { 375 struct xdp_mem_allocator *xa; 376 struct page *page; 377 378 switch (mem->type) { 379 case MEM_TYPE_PAGE_POOL: 380 rcu_read_lock(); 381 /* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */ 382 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params); 383 page = virt_to_head_page(data); 384 if (napi_direct && xdp_return_frame_no_direct()) 385 napi_direct = false; 386 page_pool_put_full_page(xa->page_pool, page, napi_direct); 387 rcu_read_unlock(); 388 break; 389 case MEM_TYPE_PAGE_SHARED: 390 page_frag_free(data); 391 break; 392 case MEM_TYPE_PAGE_ORDER0: 393 page = virt_to_page(data); /* Assumes order0 page*/ 394 put_page(page); 395 break; 396 case MEM_TYPE_XSK_BUFF_POOL: 397 /* NB! Only valid from an xdp_buff! */ 398 xsk_buff_free(xdp); 399 break; 400 default: 401 /* Not possible, checked in xdp_rxq_info_reg_mem_model() */ 402 WARN(1, "Incorrect XDP memory type (%d) usage", mem->type); 403 break; 404 } 405 } 406 407 void xdp_return_frame(struct xdp_frame *xdpf) 408 { 409 __xdp_return(xdpf->data, &xdpf->mem, false, NULL); 410 } 411 EXPORT_SYMBOL_GPL(xdp_return_frame); 412 413 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf) 414 { 415 __xdp_return(xdpf->data, &xdpf->mem, true, NULL); 416 } 417 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi); 418 419 /* XDP bulk APIs introduce a defer/flush mechanism to return 420 * pages belonging to the same xdp_mem_allocator object 421 * (identified via the mem.id field) in bulk to optimize 422 * I-cache and D-cache. 423 * The bulk queue size is set to 16 to be aligned to how 424 * XDP_REDIRECT bulking works. The bulk is flushed when 425 * it is full or when mem.id changes. 426 * xdp_frame_bulk is usually stored/allocated on the function 427 * call-stack to avoid locking penalties. 428 */ 429 void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq) 430 { 431 struct xdp_mem_allocator *xa = bq->xa; 432 433 if (unlikely(!xa || !bq->count)) 434 return; 435 436 page_pool_put_page_bulk(xa->page_pool, bq->q, bq->count); 437 /* bq->xa is not cleared to save lookup, if mem.id same in next bulk */ 438 bq->count = 0; 439 } 440 EXPORT_SYMBOL_GPL(xdp_flush_frame_bulk); 441 442 /* Must be called with rcu_read_lock held */ 443 void xdp_return_frame_bulk(struct xdp_frame *xdpf, 444 struct xdp_frame_bulk *bq) 445 { 446 struct xdp_mem_info *mem = &xdpf->mem; 447 struct xdp_mem_allocator *xa; 448 449 if (mem->type != MEM_TYPE_PAGE_POOL) { 450 __xdp_return(xdpf->data, &xdpf->mem, false, NULL); 451 return; 452 } 453 454 xa = bq->xa; 455 if (unlikely(!xa)) { 456 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params); 457 bq->count = 0; 458 bq->xa = xa; 459 } 460 461 if (bq->count == XDP_BULK_QUEUE_SIZE) 462 xdp_flush_frame_bulk(bq); 463 464 if (unlikely(mem->id != xa->mem.id)) { 465 xdp_flush_frame_bulk(bq); 466 bq->xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params); 467 } 468 469 bq->q[bq->count++] = xdpf->data; 470 } 471 EXPORT_SYMBOL_GPL(xdp_return_frame_bulk); 472 473 void xdp_return_buff(struct xdp_buff *xdp) 474 { 475 __xdp_return(xdp->data, &xdp->rxq->mem, true, xdp); 476 } 477 478 /* Only called for MEM_TYPE_PAGE_POOL see xdp.h */ 479 void __xdp_release_frame(void *data, struct xdp_mem_info *mem) 480 { 481 struct xdp_mem_allocator *xa; 482 struct page *page; 483 484 rcu_read_lock(); 485 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params); 486 page = virt_to_head_page(data); 487 if (xa) 488 page_pool_release_page(xa->page_pool, page); 489 rcu_read_unlock(); 490 } 491 EXPORT_SYMBOL_GPL(__xdp_release_frame); 492 493 void xdp_attachment_setup(struct xdp_attachment_info *info, 494 struct netdev_bpf *bpf) 495 { 496 if (info->prog) 497 bpf_prog_put(info->prog); 498 info->prog = bpf->prog; 499 info->flags = bpf->flags; 500 } 501 EXPORT_SYMBOL_GPL(xdp_attachment_setup); 502 503 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp) 504 { 505 unsigned int metasize, totsize; 506 void *addr, *data_to_copy; 507 struct xdp_frame *xdpf; 508 struct page *page; 509 510 /* Clone into a MEM_TYPE_PAGE_ORDER0 xdp_frame. */ 511 metasize = xdp_data_meta_unsupported(xdp) ? 0 : 512 xdp->data - xdp->data_meta; 513 totsize = xdp->data_end - xdp->data + metasize; 514 515 if (sizeof(*xdpf) + totsize > PAGE_SIZE) 516 return NULL; 517 518 page = dev_alloc_page(); 519 if (!page) 520 return NULL; 521 522 addr = page_to_virt(page); 523 xdpf = addr; 524 memset(xdpf, 0, sizeof(*xdpf)); 525 526 addr += sizeof(*xdpf); 527 data_to_copy = metasize ? xdp->data_meta : xdp->data; 528 memcpy(addr, data_to_copy, totsize); 529 530 xdpf->data = addr + metasize; 531 xdpf->len = totsize - metasize; 532 xdpf->headroom = 0; 533 xdpf->metasize = metasize; 534 xdpf->frame_sz = PAGE_SIZE; 535 xdpf->mem.type = MEM_TYPE_PAGE_ORDER0; 536 537 xsk_buff_free(xdp); 538 return xdpf; 539 } 540 EXPORT_SYMBOL_GPL(xdp_convert_zc_to_xdp_frame); 541 542 /* Used by XDP_WARN macro, to avoid inlining WARN() in fast-path */ 543 void xdp_warn(const char *msg, const char *func, const int line) 544 { 545 WARN(1, "XDP_WARN: %s(line:%d): %s\n", func, line, msg); 546 }; 547 EXPORT_SYMBOL_GPL(xdp_warn); 548 549 int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp) 550 { 551 n_skb = kmem_cache_alloc_bulk(skbuff_head_cache, gfp, 552 n_skb, skbs); 553 if (unlikely(!n_skb)) 554 return -ENOMEM; 555 556 return 0; 557 } 558 EXPORT_SYMBOL_GPL(xdp_alloc_skb_bulk); 559 560 struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf, 561 struct sk_buff *skb, 562 struct net_device *dev) 563 { 564 unsigned int headroom, frame_size; 565 void *hard_start; 566 567 /* Part of headroom was reserved to xdpf */ 568 headroom = sizeof(*xdpf) + xdpf->headroom; 569 570 /* Memory size backing xdp_frame data already have reserved 571 * room for build_skb to place skb_shared_info in tailroom. 572 */ 573 frame_size = xdpf->frame_sz; 574 575 hard_start = xdpf->data - headroom; 576 skb = build_skb_around(skb, hard_start, frame_size); 577 if (unlikely(!skb)) 578 return NULL; 579 580 skb_reserve(skb, headroom); 581 __skb_put(skb, xdpf->len); 582 if (xdpf->metasize) 583 skb_metadata_set(skb, xdpf->metasize); 584 585 /* Essential SKB info: protocol and skb->dev */ 586 skb->protocol = eth_type_trans(skb, dev); 587 588 /* Optional SKB info, currently missing: 589 * - HW checksum info (skb->ip_summed) 590 * - HW RX hash (skb_set_hash) 591 * - RX ring dev queue index (skb_record_rx_queue) 592 */ 593 594 /* Until page_pool get SKB return path, release DMA here */ 595 xdp_release_frame(xdpf); 596 597 /* Allow SKB to reuse area used by xdp_frame */ 598 xdp_scrub_frame(xdpf); 599 600 return skb; 601 } 602 EXPORT_SYMBOL_GPL(__xdp_build_skb_from_frame); 603 604 struct sk_buff *xdp_build_skb_from_frame(struct xdp_frame *xdpf, 605 struct net_device *dev) 606 { 607 struct sk_buff *skb; 608 609 skb = kmem_cache_alloc(skbuff_head_cache, GFP_ATOMIC); 610 if (unlikely(!skb)) 611 return NULL; 612 613 memset(skb, 0, offsetof(struct sk_buff, tail)); 614 615 return __xdp_build_skb_from_frame(xdpf, skb, dev); 616 } 617 EXPORT_SYMBOL_GPL(xdp_build_skb_from_frame); 618 619 struct xdp_frame *xdpf_clone(struct xdp_frame *xdpf) 620 { 621 unsigned int headroom, totalsize; 622 struct xdp_frame *nxdpf; 623 struct page *page; 624 void *addr; 625 626 headroom = xdpf->headroom + sizeof(*xdpf); 627 totalsize = headroom + xdpf->len; 628 629 if (unlikely(totalsize > PAGE_SIZE)) 630 return NULL; 631 page = dev_alloc_page(); 632 if (!page) 633 return NULL; 634 addr = page_to_virt(page); 635 636 memcpy(addr, xdpf, totalsize); 637 638 nxdpf = addr; 639 nxdpf->data = addr + headroom; 640 nxdpf->frame_sz = PAGE_SIZE; 641 nxdpf->mem.type = MEM_TYPE_PAGE_ORDER0; 642 nxdpf->mem.id = 0; 643 644 return nxdpf; 645 } 646