net/core/xdp.c

   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,
 167                        unsigned int napi_id, u32 frag_size)
 168 {
 169         if (!dev) {
 170                 WARN(1, "Missing net_device from driver");
 171                 return -ENODEV;
 172         }
 173
 174         if (xdp_rxq->reg_state == REG_STATE_UNUSED) {
 175                 WARN(1, "Driver promised not to register this");
 176                 return -EINVAL;
 177         }
 178
 179         if (xdp_rxq->reg_state == REG_STATE_REGISTERED) {
 180                 WARN(1, "Missing unregister, handled but fix driver");
 181                 xdp_rxq_info_unreg(xdp_rxq);
 182         }
 183
 184         /* State either UNREGISTERED or NEW */
 185         xdp_rxq_info_init(xdp_rxq);
 186         xdp_rxq->dev = dev;
 187         xdp_rxq->queue_index = queue_index;
 188         xdp_rxq->napi_id = napi_id;
 189         xdp_rxq->frag_size = frag_size;
 190
 191         xdp_rxq->reg_state = REG_STATE_REGISTERED;
 192         return 0;
 193 }
 194 EXPORT_SYMBOL_GPL(__xdp_rxq_info_reg);
 195
 196 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq)
 197 {
 198         xdp_rxq->reg_state = REG_STATE_UNUSED;
 199 }
 200 EXPORT_SYMBOL_GPL(xdp_rxq_info_unused);
 201
 202 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq)
 203 {
 204         return (xdp_rxq->reg_state == REG_STATE_REGISTERED);
 205 }
 206 EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg);
 207
 208 static int __mem_id_init_hash_table(void)
 209 {
 210         struct rhashtable *rht;
 211         int ret;
 212
 213         if (unlikely(mem_id_init))
 214                 return 0;
 215
 216         rht = kzalloc(sizeof(*rht), GFP_KERNEL);
 217         if (!rht)
 218                 return -ENOMEM;
 219
 220         ret = rhashtable_init(rht, &mem_id_rht_params);
 221         if (ret < 0) {
 222                 kfree(rht);
 223                 return ret;
 224         }
 225         mem_id_ht = rht;
 226         smp_mb(); /* mutex lock should provide enough pairing */
 227         mem_id_init = true;
 228
 229         return 0;
 230 }
 231
 232 /* Allocate a cyclic ID that maps to allocator pointer.
 233  * See: https://www.kernel.org/doc/html/latest/core-api/idr.html
 234  *
 235  * Caller must lock mem_id_lock.
 236  */
 237 static int __mem_id_cyclic_get(gfp_t gfp)
 238 {
 239         int retries = 1;
 240         int id;
 241
 242 again:
 243         id = ida_simple_get(&mem_id_pool, mem_id_next, MEM_ID_MAX, gfp);
 244         if (id < 0) {
 245                 if (id == -ENOSPC) {
 246                         /* Cyclic allocator, reset next id */
 247                         if (retries--) {
 248                                 mem_id_next = MEM_ID_MIN;
 249                                 goto again;
 250                         }
 251                 }
 252                 return id; /* errno */
 253         }
 254         mem_id_next = id + 1;
 255
 256         return id;
 257 }
 258
 259 static bool __is_supported_mem_type(enum xdp_mem_type type)
 260 {
 261         if (type == MEM_TYPE_PAGE_POOL)
 262                 return is_page_pool_compiled_in();
 263
 264         if (type >= MEM_TYPE_MAX)
 265                 return false;
 266
 267         return true;
 268 }
 269
 270 static struct xdp_mem_allocator *__xdp_reg_mem_model(struct xdp_mem_info *mem,
 271                                                      enum xdp_mem_type type,
 272                                                      void *allocator)
 273 {
 274         struct xdp_mem_allocator *xdp_alloc;
 275         gfp_t gfp = GFP_KERNEL;
 276         int id, errno, ret;
 277         void *ptr;
 278
 279         if (!__is_supported_mem_type(type))
 280                 return ERR_PTR(-EOPNOTSUPP);
 281
 282         mem->type = type;
 283
 284         if (!allocator) {
 285                 if (type == MEM_TYPE_PAGE_POOL)
 286                         return ERR_PTR(-EINVAL); /* Setup time check page_pool req */
 287                 return NULL;
 288         }
 289
 290         /* Delay init of rhashtable to save memory if feature isn't used */
 291         if (!mem_id_init) {
 292                 mutex_lock(&mem_id_lock);
 293                 ret = __mem_id_init_hash_table();
 294                 mutex_unlock(&mem_id_lock);
 295                 if (ret < 0) {
 296                         WARN_ON(1);
 297                         return ERR_PTR(ret);
 298                 }
 299         }
 300
 301         xdp_alloc = kzalloc(sizeof(*xdp_alloc), gfp);
 302         if (!xdp_alloc)
 303                 return ERR_PTR(-ENOMEM);
 304
 305         mutex_lock(&mem_id_lock);
 306         id = __mem_id_cyclic_get(gfp);
 307         if (id < 0) {
 308                 errno = id;
 309                 goto err;
 310         }
 311         mem->id = id;
 312         xdp_alloc->mem = *mem;
 313         xdp_alloc->allocator = allocator;
 314
 315         /* Insert allocator into ID lookup table */
 316         ptr = rhashtable_insert_slow(mem_id_ht, &id, &xdp_alloc->node);
 317         if (IS_ERR(ptr)) {
 318                 ida_simple_remove(&mem_id_pool, mem->id);
 319                 mem->id = 0;
 320                 errno = PTR_ERR(ptr);
 321                 goto err;
 322         }
 323
 324         if (type == MEM_TYPE_PAGE_POOL)
 325                 page_pool_use_xdp_mem(allocator, mem_allocator_disconnect, mem);
 326
 327         mutex_unlock(&mem_id_lock);
 328
 329         return xdp_alloc;
 330 err:
 331         mutex_unlock(&mem_id_lock);
 332         kfree(xdp_alloc);
 333         return ERR_PTR(errno);
 334 }
 335
 336 int xdp_reg_mem_model(struct xdp_mem_info *mem,
 337                       enum xdp_mem_type type, void *allocator)
 338 {
 339         struct xdp_mem_allocator *xdp_alloc;
 340
 341         xdp_alloc = __xdp_reg_mem_model(mem, type, allocator);
 342         if (IS_ERR(xdp_alloc))
 343                 return PTR_ERR(xdp_alloc);
 344         return 0;
 345 }
 346 EXPORT_SYMBOL_GPL(xdp_reg_mem_model);
 347
 348 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
 349                                enum xdp_mem_type type, void *allocator)
 350 {
 351         struct xdp_mem_allocator *xdp_alloc;
 352
 353         if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
 354                 WARN(1, "Missing register, driver bug");
 355                 return -EFAULT;
 356         }
 357
 358         xdp_alloc = __xdp_reg_mem_model(&xdp_rxq->mem, type, allocator);
 359         if (IS_ERR(xdp_alloc))
 360                 return PTR_ERR(xdp_alloc);
 361
 362         if (trace_mem_connect_enabled() && xdp_alloc)
 363                 trace_mem_connect(xdp_alloc, xdp_rxq);
 364         return 0;
 365 }
 366
 367 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model);
 368
 369 /* XDP RX runs under NAPI protection, and in different delivery error
 370  * scenarios (e.g. queue full), it is possible to return the xdp_frame
 371  * while still leveraging this protection.  The @napi_direct boolean
 372  * is used for those calls sites.  Thus, allowing for faster recycling
 373  * of xdp_frames/pages in those cases.
 374  */
 375 void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct,
 376                   struct xdp_buff *xdp)
 377 {
 378         struct xdp_mem_allocator *xa;
 379         struct page *page;
 380
 381         switch (mem->type) {
 382         case MEM_TYPE_PAGE_POOL:
 383                 rcu_read_lock();
 384                 /* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */
 385                 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
 386                 page = virt_to_head_page(data);
 387                 if (napi_direct && xdp_return_frame_no_direct())
 388                         napi_direct = false;
 389                 page_pool_put_full_page(xa->page_pool, page, napi_direct);
 390                 rcu_read_unlock();
 391                 break;
 392         case MEM_TYPE_PAGE_SHARED:
 393                 page_frag_free(data);
 394                 break;
 395         case MEM_TYPE_PAGE_ORDER0:
 396                 page = virt_to_page(data); /* Assumes order0 page*/
 397                 put_page(page);
 398                 break;
 399         case MEM_TYPE_XSK_BUFF_POOL:
 400                 /* NB! Only valid from an xdp_buff! */
 401                 xsk_buff_free(xdp);
 402                 break;
 403         default:
 404                 /* Not possible, checked in xdp_rxq_info_reg_mem_model() */
 405                 WARN(1, "Incorrect XDP memory type (%d) usage", mem->type);
 406                 break;
 407         }
 408 }
 409
 410 void xdp_return_frame(struct xdp_frame *xdpf)
 411 {
 412         struct skb_shared_info *sinfo;
 413         int i;
 414
 415         if (likely(!xdp_frame_has_frags(xdpf)))
 416                 goto out;
 417
 418         sinfo = xdp_get_shared_info_from_frame(xdpf);
 419         for (i = 0; i < sinfo->nr_frags; i++) {
 420                 struct page *page = skb_frag_page(&sinfo->frags[i]);
 421
 422                 __xdp_return(page_address(page), &xdpf->mem, false, NULL);
 423         }
 424 out:
 425         __xdp_return(xdpf->data, &xdpf->mem, false, NULL);
 426 }
 427 EXPORT_SYMBOL_GPL(xdp_return_frame);
 428
 429 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf)
 430 {
 431         struct skb_shared_info *sinfo;
 432         int i;
 433
 434         if (likely(!xdp_frame_has_frags(xdpf)))
 435                 goto out;
 436
 437         sinfo = xdp_get_shared_info_from_frame(xdpf);
 438         for (i = 0; i < sinfo->nr_frags; i++) {
 439                 struct page *page = skb_frag_page(&sinfo->frags[i]);
 440
 441                 __xdp_return(page_address(page), &xdpf->mem, true, NULL);
 442         }
 443 out:
 444         __xdp_return(xdpf->data, &xdpf->mem, true, NULL);
 445 }
 446 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi);
 447
 448 /* XDP bulk APIs introduce a defer/flush mechanism to return
 449  * pages belonging to the same xdp_mem_allocator object
 450  * (identified via the mem.id field) in bulk to optimize
 451  * I-cache and D-cache.
 452  * The bulk queue size is set to 16 to be aligned to how
 453  * XDP_REDIRECT bulking works. The bulk is flushed when
 454  * it is full or when mem.id changes.
 455  * xdp_frame_bulk is usually stored/allocated on the function
 456  * call-stack to avoid locking penalties.
 457  */
 458 void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq)
 459 {
 460         struct xdp_mem_allocator *xa = bq->xa;
 461
 462         if (unlikely(!xa || !bq->count))
 463                 return;
 464
 465         page_pool_put_page_bulk(xa->page_pool, bq->q, bq->count);
 466         /* bq->xa is not cleared to save lookup, if mem.id same in next bulk */
 467         bq->count = 0;
 468 }
 469 EXPORT_SYMBOL_GPL(xdp_flush_frame_bulk);
 470
 471 /* Must be called with rcu_read_lock held */
 472 void xdp_return_frame_bulk(struct xdp_frame *xdpf,
 473                            struct xdp_frame_bulk *bq)
 474 {
 475         struct xdp_mem_info *mem = &xdpf->mem;
 476         struct xdp_mem_allocator *xa;
 477
 478         if (mem->type != MEM_TYPE_PAGE_POOL) {
 479                 xdp_return_frame(xdpf);
 480                 return;
 481         }
 482
 483         xa = bq->xa;
 484         if (unlikely(!xa)) {
 485                 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
 486                 bq->count = 0;
 487                 bq->xa = xa;
 488         }
 489
 490         if (bq->count == XDP_BULK_QUEUE_SIZE)
 491                 xdp_flush_frame_bulk(bq);
 492
 493         if (unlikely(mem->id != xa->mem.id)) {
 494                 xdp_flush_frame_bulk(bq);
 495                 bq->xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
 496         }
 497
 498         if (unlikely(xdp_frame_has_frags(xdpf))) {
 499                 struct skb_shared_info *sinfo;
 500                 int i;
 501
 502                 sinfo = xdp_get_shared_info_from_frame(xdpf);
 503                 for (i = 0; i < sinfo->nr_frags; i++) {
 504                         skb_frag_t *frag = &sinfo->frags[i];
 505
 506                         bq->q[bq->count++] = skb_frag_address(frag);
 507                         if (bq->count == XDP_BULK_QUEUE_SIZE)
 508                                 xdp_flush_frame_bulk(bq);
 509                 }
 510         }
 511         bq->q[bq->count++] = xdpf->data;
 512 }
 513 EXPORT_SYMBOL_GPL(xdp_return_frame_bulk);
 514
 515 void xdp_return_buff(struct xdp_buff *xdp)
 516 {
 517         struct skb_shared_info *sinfo;
 518         int i;
 519
 520         if (likely(!xdp_buff_has_frags(xdp)))
 521                 goto out;
 522
 523         sinfo = xdp_get_shared_info_from_buff(xdp);
 524         for (i = 0; i < sinfo->nr_frags; i++) {
 525                 struct page *page = skb_frag_page(&sinfo->frags[i]);
 526
 527                 __xdp_return(page_address(page), &xdp->rxq->mem, true, xdp);
 528         }
 529 out:
 530         __xdp_return(xdp->data, &xdp->rxq->mem, true, xdp);
 531 }
 532 EXPORT_SYMBOL_GPL(xdp_return_buff);
 533
 534 /* Only called for MEM_TYPE_PAGE_POOL see xdp.h */
 535 void __xdp_release_frame(void *data, struct xdp_mem_info *mem)
 536 {
 537         struct xdp_mem_allocator *xa;
 538         struct page *page;
 539
 540         rcu_read_lock();
 541         xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
 542         page = virt_to_head_page(data);
 543         if (xa)
 544                 page_pool_release_page(xa->page_pool, page);
 545         rcu_read_unlock();
 546 }
 547 EXPORT_SYMBOL_GPL(__xdp_release_frame);
 548
 549 void xdp_attachment_setup(struct xdp_attachment_info *info,
 550                           struct netdev_bpf *bpf)
 551 {
 552         if (info->prog)
 553                 bpf_prog_put(info->prog);
 554         info->prog = bpf->prog;
 555         info->flags = bpf->flags;
 556 }
 557 EXPORT_SYMBOL_GPL(xdp_attachment_setup);
 558
 559 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp)
 560 {
 561         unsigned int metasize, totsize;
 562         void *addr, *data_to_copy;
 563         struct xdp_frame *xdpf;
 564         struct page *page;
 565
 566         /* Clone into a MEM_TYPE_PAGE_ORDER0 xdp_frame. */
 567         metasize = xdp_data_meta_unsupported(xdp) ? 0 :
 568                    xdp->data - xdp->data_meta;
 569         totsize = xdp->data_end - xdp->data + metasize;
 570
 571         if (sizeof(*xdpf) + totsize > PAGE_SIZE)
 572                 return NULL;
 573
 574         page = dev_alloc_page();
 575         if (!page)
 576                 return NULL;
 577
 578         addr = page_to_virt(page);
 579         xdpf = addr;
 580         memset(xdpf, 0, sizeof(*xdpf));
 581
 582         addr += sizeof(*xdpf);
 583         data_to_copy = metasize ? xdp->data_meta : xdp->data;
 584         memcpy(addr, data_to_copy, totsize);
 585
 586         xdpf->data = addr + metasize;
 587         xdpf->len = totsize - metasize;
 588         xdpf->headroom = 0;
 589         xdpf->metasize = metasize;
 590         xdpf->frame_sz = PAGE_SIZE;
 591         xdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
 592
 593         xsk_buff_free(xdp);
 594         return xdpf;
 595 }
 596 EXPORT_SYMBOL_GPL(xdp_convert_zc_to_xdp_frame);
 597
 598 /* Used by XDP_WARN macro, to avoid inlining WARN() in fast-path */
 599 void xdp_warn(const char *msg, const char *func, const int line)
 600 {
 601         WARN(1, "XDP_WARN: %s(line:%d): %s\n", func, line, msg);
 602 };
 603 EXPORT_SYMBOL_GPL(xdp_warn);
 604
 605 int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp)
 606 {
 607         n_skb = kmem_cache_alloc_bulk(skbuff_head_cache, gfp,
 608                                       n_skb, skbs);
 609         if (unlikely(!n_skb))
 610                 return -ENOMEM;
 611
 612         return 0;
 613 }
 614 EXPORT_SYMBOL_GPL(xdp_alloc_skb_bulk);
 615
 616 struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf,
 617                                            struct sk_buff *skb,
 618                                            struct net_device *dev)
 619 {
 620         struct skb_shared_info *sinfo = xdp_get_shared_info_from_frame(xdpf);
 621         unsigned int headroom, frame_size;
 622         void *hard_start;
 623         u8 nr_frags;
 624
 625         /* xdp frags frame */
 626         if (unlikely(xdp_frame_has_frags(xdpf)))
 627                 nr_frags = sinfo->nr_frags;
 628
 629         /* Part of headroom was reserved to xdpf */
 630         headroom = sizeof(*xdpf) + xdpf->headroom;
 631
 632         /* Memory size backing xdp_frame data already have reserved
 633          * room for build_skb to place skb_shared_info in tailroom.
 634          */
 635         frame_size = xdpf->frame_sz;
 636
 637         hard_start = xdpf->data - headroom;
 638         skb = build_skb_around(skb, hard_start, frame_size);
 639         if (unlikely(!skb))
 640                 return NULL;
 641
 642         skb_reserve(skb, headroom);
 643         __skb_put(skb, xdpf->len);
 644         if (xdpf->metasize)
 645                 skb_metadata_set(skb, xdpf->metasize);
 646
 647         if (unlikely(xdp_frame_has_frags(xdpf)))
 648                 xdp_update_skb_shared_info(skb, nr_frags,
 649                                            sinfo->xdp_frags_size,
 650                                            nr_frags * xdpf->frame_sz,
 651                                            xdp_frame_is_frag_pfmemalloc(xdpf));
 652
 653         /* Essential SKB info: protocol and skb->dev */
 654         skb->protocol = eth_type_trans(skb, dev);
 655
 656         /* Optional SKB info, currently missing:
 657          * - HW checksum info           (skb->ip_summed)
 658          * - HW RX hash                 (skb_set_hash)
 659          * - RX ring dev queue index    (skb_record_rx_queue)
 660          */
 661
 662         /* Until page_pool get SKB return path, release DMA here */
 663         xdp_release_frame(xdpf);
 664
 665         /* Allow SKB to reuse area used by xdp_frame */
 666         xdp_scrub_frame(xdpf);
 667
 668         return skb;
 669 }
 670 EXPORT_SYMBOL_GPL(__xdp_build_skb_from_frame);
 671
 672 struct sk_buff *xdp_build_skb_from_frame(struct xdp_frame *xdpf,
 673                                          struct net_device *dev)
 674 {
 675         struct sk_buff *skb;
 676
 677         skb = kmem_cache_alloc(skbuff_head_cache, GFP_ATOMIC);
 678         if (unlikely(!skb))
 679                 return NULL;
 680
 681         memset(skb, 0, offsetof(struct sk_buff, tail));
 682
 683         return __xdp_build_skb_from_frame(xdpf, skb, dev);
 684 }
 685 EXPORT_SYMBOL_GPL(xdp_build_skb_from_frame);
 686
 687 struct xdp_frame *xdpf_clone(struct xdp_frame *xdpf)
 688 {
 689         unsigned int headroom, totalsize;
 690         struct xdp_frame *nxdpf;
 691         struct page *page;
 692         void *addr;
 693
 694         headroom = xdpf->headroom + sizeof(*xdpf);
 695         totalsize = headroom + xdpf->len;
 696
 697         if (unlikely(totalsize > PAGE_SIZE))
 698                 return NULL;
 699         page = dev_alloc_page();
 700         if (!page)
 701                 return NULL;
 702         addr = page_to_virt(page);
 703
 704         memcpy(addr, xdpf, totalsize);
 705
 706         nxdpf = addr;
 707         nxdpf->data = addr + headroom;
 708         nxdpf->frame_sz = PAGE_SIZE;
 709         nxdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
 710         nxdpf->mem.id = 0;
 711
 712         return nxdpf;
 713 }