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 page *page;
 379
 380         switch (mem->type) {
 381         case MEM_TYPE_PAGE_POOL:
 382                 page = virt_to_head_page(data);
 383                 if (napi_direct && xdp_return_frame_no_direct())
 384                         napi_direct = false;
 385                 /* No need to check ((page->pp_magic & ~0x3UL) == PP_SIGNATURE)
 386                  * as mem->type knows this a page_pool page
 387                  */
 388                 page_pool_put_full_page(page->pp, page, napi_direct);
 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         struct skb_shared_info *sinfo;
 411         int i;
 412
 413         if (likely(!xdp_frame_has_frags(xdpf)))
 414                 goto out;
 415
 416         sinfo = xdp_get_shared_info_from_frame(xdpf);
 417         for (i = 0; i < sinfo->nr_frags; i++) {
 418                 struct page *page = skb_frag_page(&sinfo->frags[i]);
 419
 420                 __xdp_return(page_address(page), &xdpf->mem, false, NULL);
 421         }
 422 out:
 423         __xdp_return(xdpf->data, &xdpf->mem, false, NULL);
 424 }
 425 EXPORT_SYMBOL_GPL(xdp_return_frame);
 426
 427 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf)
 428 {
 429         struct skb_shared_info *sinfo;
 430         int i;
 431
 432         if (likely(!xdp_frame_has_frags(xdpf)))
 433                 goto out;
 434
 435         sinfo = xdp_get_shared_info_from_frame(xdpf);
 436         for (i = 0; i < sinfo->nr_frags; i++) {
 437                 struct page *page = skb_frag_page(&sinfo->frags[i]);
 438
 439                 __xdp_return(page_address(page), &xdpf->mem, true, NULL);
 440         }
 441 out:
 442         __xdp_return(xdpf->data, &xdpf->mem, true, NULL);
 443 }
 444 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi);
 445
 446 /* XDP bulk APIs introduce a defer/flush mechanism to return
 447  * pages belonging to the same xdp_mem_allocator object
 448  * (identified via the mem.id field) in bulk to optimize
 449  * I-cache and D-cache.
 450  * The bulk queue size is set to 16 to be aligned to how
 451  * XDP_REDIRECT bulking works. The bulk is flushed when
 452  * it is full or when mem.id changes.
 453  * xdp_frame_bulk is usually stored/allocated on the function
 454  * call-stack to avoid locking penalties.
 455  */
 456 void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq)
 457 {
 458         struct xdp_mem_allocator *xa = bq->xa;
 459
 460         if (unlikely(!xa || !bq->count))
 461                 return;
 462
 463         page_pool_put_page_bulk(xa->page_pool, bq->q, bq->count);
 464         /* bq->xa is not cleared to save lookup, if mem.id same in next bulk */
 465         bq->count = 0;
 466 }
 467 EXPORT_SYMBOL_GPL(xdp_flush_frame_bulk);
 468
 469 /* Must be called with rcu_read_lock held */
 470 void xdp_return_frame_bulk(struct xdp_frame *xdpf,
 471                            struct xdp_frame_bulk *bq)
 472 {
 473         struct xdp_mem_info *mem = &xdpf->mem;
 474         struct xdp_mem_allocator *xa;
 475
 476         if (mem->type != MEM_TYPE_PAGE_POOL) {
 477                 xdp_return_frame(xdpf);
 478                 return;
 479         }
 480
 481         xa = bq->xa;
 482         if (unlikely(!xa)) {
 483                 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
 484                 bq->count = 0;
 485                 bq->xa = xa;
 486         }
 487
 488         if (bq->count == XDP_BULK_QUEUE_SIZE)
 489                 xdp_flush_frame_bulk(bq);
 490
 491         if (unlikely(mem->id != xa->mem.id)) {
 492                 xdp_flush_frame_bulk(bq);
 493                 bq->xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
 494         }
 495
 496         if (unlikely(xdp_frame_has_frags(xdpf))) {
 497                 struct skb_shared_info *sinfo;
 498                 int i;
 499
 500                 sinfo = xdp_get_shared_info_from_frame(xdpf);
 501                 for (i = 0; i < sinfo->nr_frags; i++) {
 502                         skb_frag_t *frag = &sinfo->frags[i];
 503
 504                         bq->q[bq->count++] = skb_frag_address(frag);
 505                         if (bq->count == XDP_BULK_QUEUE_SIZE)
 506                                 xdp_flush_frame_bulk(bq);
 507                 }
 508         }
 509         bq->q[bq->count++] = xdpf->data;
 510 }
 511 EXPORT_SYMBOL_GPL(xdp_return_frame_bulk);
 512
 513 void xdp_return_buff(struct xdp_buff *xdp)
 514 {
 515         struct skb_shared_info *sinfo;
 516         int i;
 517
 518         if (likely(!xdp_buff_has_frags(xdp)))
 519                 goto out;
 520
 521         sinfo = xdp_get_shared_info_from_buff(xdp);
 522         for (i = 0; i < sinfo->nr_frags; i++) {
 523                 struct page *page = skb_frag_page(&sinfo->frags[i]);
 524
 525                 __xdp_return(page_address(page), &xdp->rxq->mem, true, xdp);
 526         }
 527 out:
 528         __xdp_return(xdp->data, &xdp->rxq->mem, true, xdp);
 529 }
 530 EXPORT_SYMBOL_GPL(xdp_return_buff);
 531
 532 /* Only called for MEM_TYPE_PAGE_POOL see xdp.h */
 533 void __xdp_release_frame(void *data, struct xdp_mem_info *mem)
 534 {
 535         struct xdp_mem_allocator *xa;
 536         struct page *page;
 537
 538         rcu_read_lock();
 539         xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
 540         page = virt_to_head_page(data);
 541         if (xa)
 542                 page_pool_release_page(xa->page_pool, page);
 543         rcu_read_unlock();
 544 }
 545 EXPORT_SYMBOL_GPL(__xdp_release_frame);
 546
 547 void xdp_attachment_setup(struct xdp_attachment_info *info,
 548                           struct netdev_bpf *bpf)
 549 {
 550         if (info->prog)
 551                 bpf_prog_put(info->prog);
 552         info->prog = bpf->prog;
 553         info->flags = bpf->flags;
 554 }
 555 EXPORT_SYMBOL_GPL(xdp_attachment_setup);
 556
 557 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp)
 558 {
 559         unsigned int metasize, totsize;
 560         void *addr, *data_to_copy;
 561         struct xdp_frame *xdpf;
 562         struct page *page;
 563
 564         /* Clone into a MEM_TYPE_PAGE_ORDER0 xdp_frame. */
 565         metasize = xdp_data_meta_unsupported(xdp) ? 0 :
 566                    xdp->data - xdp->data_meta;
 567         totsize = xdp->data_end - xdp->data + metasize;
 568
 569         if (sizeof(*xdpf) + totsize > PAGE_SIZE)
 570                 return NULL;
 571
 572         page = dev_alloc_page();
 573         if (!page)
 574                 return NULL;
 575
 576         addr = page_to_virt(page);
 577         xdpf = addr;
 578         memset(xdpf, 0, sizeof(*xdpf));
 579
 580         addr += sizeof(*xdpf);
 581         data_to_copy = metasize ? xdp->data_meta : xdp->data;
 582         memcpy(addr, data_to_copy, totsize);
 583
 584         xdpf->data = addr + metasize;
 585         xdpf->len = totsize - metasize;
 586         xdpf->headroom = 0;
 587         xdpf->metasize = metasize;
 588         xdpf->frame_sz = PAGE_SIZE;
 589         xdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
 590
 591         xsk_buff_free(xdp);
 592         return xdpf;
 593 }
 594 EXPORT_SYMBOL_GPL(xdp_convert_zc_to_xdp_frame);
 595
 596 /* Used by XDP_WARN macro, to avoid inlining WARN() in fast-path */
 597 void xdp_warn(const char *msg, const char *func, const int line)
 598 {
 599         WARN(1, "XDP_WARN: %s(line:%d): %s\n", func, line, msg);
 600 };
 601 EXPORT_SYMBOL_GPL(xdp_warn);
 602
 603 int xdp_alloc_skb_bulk(void **skbs, int n_skb, gfp_t gfp)
 604 {
 605         n_skb = kmem_cache_alloc_bulk(skbuff_head_cache, gfp,
 606                                       n_skb, skbs);
 607         if (unlikely(!n_skb))
 608                 return -ENOMEM;
 609
 610         return 0;
 611 }
 612 EXPORT_SYMBOL_GPL(xdp_alloc_skb_bulk);
 613
 614 struct sk_buff *__xdp_build_skb_from_frame(struct xdp_frame *xdpf,
 615                                            struct sk_buff *skb,
 616                                            struct net_device *dev)
 617 {
 618         struct skb_shared_info *sinfo = xdp_get_shared_info_from_frame(xdpf);
 619         unsigned int headroom, frame_size;
 620         void *hard_start;
 621         u8 nr_frags;
 622
 623         /* xdp frags frame */
 624         if (unlikely(xdp_frame_has_frags(xdpf)))
 625                 nr_frags = sinfo->nr_frags;
 626
 627         /* Part of headroom was reserved to xdpf */
 628         headroom = sizeof(*xdpf) + xdpf->headroom;
 629
 630         /* Memory size backing xdp_frame data already have reserved
 631          * room for build_skb to place skb_shared_info in tailroom.
 632          */
 633         frame_size = xdpf->frame_sz;
 634
 635         hard_start = xdpf->data - headroom;
 636         skb = build_skb_around(skb, hard_start, frame_size);
 637         if (unlikely(!skb))
 638                 return NULL;
 639
 640         skb_reserve(skb, headroom);
 641         __skb_put(skb, xdpf->len);
 642         if (xdpf->metasize)
 643                 skb_metadata_set(skb, xdpf->metasize);
 644
 645         if (unlikely(xdp_frame_has_frags(xdpf)))
 646                 xdp_update_skb_shared_info(skb, nr_frags,
 647                                            sinfo->xdp_frags_size,
 648                                            nr_frags * xdpf->frame_sz,
 649                                            xdp_frame_is_frag_pfmemalloc(xdpf));
 650
 651         /* Essential SKB info: protocol and skb->dev */
 652         skb->protocol = eth_type_trans(skb, dev);
 653
 654         /* Optional SKB info, currently missing:
 655          * - HW checksum info           (skb->ip_summed)
 656          * - HW RX hash                 (skb_set_hash)
 657          * - RX ring dev queue index    (skb_record_rx_queue)
 658          */
 659
 660         /* Until page_pool get SKB return path, release DMA here */
 661         xdp_release_frame(xdpf);
 662
 663         /* Allow SKB to reuse area used by xdp_frame */
 664         xdp_scrub_frame(xdpf);
 665
 666         return skb;
 667 }
 668 EXPORT_SYMBOL_GPL(__xdp_build_skb_from_frame);
 669
 670 struct sk_buff *xdp_build_skb_from_frame(struct xdp_frame *xdpf,
 671                                          struct net_device *dev)
 672 {
 673         struct sk_buff *skb;
 674
 675         skb = kmem_cache_alloc(skbuff_head_cache, GFP_ATOMIC);
 676         if (unlikely(!skb))
 677                 return NULL;
 678
 679         memset(skb, 0, offsetof(struct sk_buff, tail));
 680
 681         return __xdp_build_skb_from_frame(xdpf, skb, dev);
 682 }
 683 EXPORT_SYMBOL_GPL(xdp_build_skb_from_frame);
 684
 685 struct xdp_frame *xdpf_clone(struct xdp_frame *xdpf)
 686 {
 687         unsigned int headroom, totalsize;
 688         struct xdp_frame *nxdpf;
 689         struct page *page;
 690         void *addr;
 691
 692         headroom = xdpf->headroom + sizeof(*xdpf);
 693         totalsize = headroom + xdpf->len;
 694
 695         if (unlikely(totalsize > PAGE_SIZE))
 696                 return NULL;
 697         page = dev_alloc_page();
 698         if (!page)
 699                 return NULL;
 700         addr = page_to_virt(page);
 701
 702         memcpy(addr, xdpf, totalsize);
 703
 704         nxdpf = addr;
 705         nxdpf->data = addr + headroom;
 706         nxdpf->frame_sz = PAGE_SIZE;
 707         nxdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
 708         nxdpf->mem.id = 0;
 709
 710         return nxdpf;
 711 }