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_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq)
 114 {
 115         struct xdp_mem_allocator *xa;
 116         int id = xdp_rxq->mem.id;
 117
 118         if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
 119                 WARN(1, "Missing register, driver bug");
 120                 return;
 121         }
 122
 123         if (id == 0)
 124                 return;
 125
 126         if (xdp_rxq->mem.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_rxq_info_unreg_mem_model);
 134
 135 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq)
 136 {
 137         /* Simplify driver cleanup code paths, allow unreg "unused" */
 138         if (xdp_rxq->reg_state == REG_STATE_UNUSED)
 139                 return;
 140
 141         WARN(!(xdp_rxq->reg_state == REG_STATE_REGISTERED), "Driver BUG");
 142
 143         xdp_rxq_info_unreg_mem_model(xdp_rxq);
 144
 145         xdp_rxq->reg_state = REG_STATE_UNREGISTERED;
 146         xdp_rxq->dev = NULL;
 147
 148         /* Reset mem info to defaults */
 149         xdp_rxq->mem.id = 0;
 150         xdp_rxq->mem.type = 0;
 151 }
 152 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg);
 153
 154 static void xdp_rxq_info_init(struct xdp_rxq_info *xdp_rxq)
 155 {
 156         memset(xdp_rxq, 0, sizeof(*xdp_rxq));
 157 }
 158
 159 /* Returns 0 on success, negative on failure */
 160 int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
 161                      struct net_device *dev, u32 queue_index, unsigned int napi_id)
 162 {
 163         if (xdp_rxq->reg_state == REG_STATE_UNUSED) {
 164                 WARN(1, "Driver promised not to register this");
 165                 return -EINVAL;
 166         }
 167
 168         if (xdp_rxq->reg_state == REG_STATE_REGISTERED) {
 169                 WARN(1, "Missing unregister, handled but fix driver");
 170                 xdp_rxq_info_unreg(xdp_rxq);
 171         }
 172
 173         if (!dev) {
 174                 WARN(1, "Missing net_device from driver");
 175                 return -ENODEV;
 176         }
 177
 178         /* State either UNREGISTERED or NEW */
 179         xdp_rxq_info_init(xdp_rxq);
 180         xdp_rxq->dev = dev;
 181         xdp_rxq->queue_index = queue_index;
 182         xdp_rxq->napi_id = napi_id;
 183
 184         xdp_rxq->reg_state = REG_STATE_REGISTERED;
 185         return 0;
 186 }
 187 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg);
 188
 189 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq)
 190 {
 191         xdp_rxq->reg_state = REG_STATE_UNUSED;
 192 }
 193 EXPORT_SYMBOL_GPL(xdp_rxq_info_unused);
 194
 195 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq)
 196 {
 197         return (xdp_rxq->reg_state == REG_STATE_REGISTERED);
 198 }
 199 EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg);
 200
 201 static int __mem_id_init_hash_table(void)
 202 {
 203         struct rhashtable *rht;
 204         int ret;
 205
 206         if (unlikely(mem_id_init))
 207                 return 0;
 208
 209         rht = kzalloc(sizeof(*rht), GFP_KERNEL);
 210         if (!rht)
 211                 return -ENOMEM;
 212
 213         ret = rhashtable_init(rht, &mem_id_rht_params);
 214         if (ret < 0) {
 215                 kfree(rht);
 216                 return ret;
 217         }
 218         mem_id_ht = rht;
 219         smp_mb(); /* mutex lock should provide enough pairing */
 220         mem_id_init = true;
 221
 222         return 0;
 223 }
 224
 225 /* Allocate a cyclic ID that maps to allocator pointer.
 226  * See: https://www.kernel.org/doc/html/latest/core-api/idr.html
 227  *
 228  * Caller must lock mem_id_lock.
 229  */
 230 static int __mem_id_cyclic_get(gfp_t gfp)
 231 {
 232         int retries = 1;
 233         int id;
 234
 235 again:
 236         id = ida_simple_get(&mem_id_pool, mem_id_next, MEM_ID_MAX, gfp);
 237         if (id < 0) {
 238                 if (id == -ENOSPC) {
 239                         /* Cyclic allocator, reset next id */
 240                         if (retries--) {
 241                                 mem_id_next = MEM_ID_MIN;
 242                                 goto again;
 243                         }
 244                 }
 245                 return id; /* errno */
 246         }
 247         mem_id_next = id + 1;
 248
 249         return id;
 250 }
 251
 252 static bool __is_supported_mem_type(enum xdp_mem_type type)
 253 {
 254         if (type == MEM_TYPE_PAGE_POOL)
 255                 return is_page_pool_compiled_in();
 256
 257         if (type >= MEM_TYPE_MAX)
 258                 return false;
 259
 260         return true;
 261 }
 262
 263 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
 264                                enum xdp_mem_type type, void *allocator)
 265 {
 266         struct xdp_mem_allocator *xdp_alloc;
 267         gfp_t gfp = GFP_KERNEL;
 268         int id, errno, ret;
 269         void *ptr;
 270
 271         if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
 272                 WARN(1, "Missing register, driver bug");
 273                 return -EFAULT;
 274         }
 275
 276         if (!__is_supported_mem_type(type))
 277                 return -EOPNOTSUPP;
 278
 279         xdp_rxq->mem.type = type;
 280
 281         if (!allocator) {
 282                 if (type == MEM_TYPE_PAGE_POOL)
 283                         return -EINVAL; /* Setup time check page_pool req */
 284                 return 0;
 285         }
 286
 287         /* Delay init of rhashtable to save memory if feature isn't used */
 288         if (!mem_id_init) {
 289                 mutex_lock(&mem_id_lock);
 290                 ret = __mem_id_init_hash_table();
 291                 mutex_unlock(&mem_id_lock);
 292                 if (ret < 0) {
 293                         WARN_ON(1);
 294                         return ret;
 295                 }
 296         }
 297
 298         xdp_alloc = kzalloc(sizeof(*xdp_alloc), gfp);
 299         if (!xdp_alloc)
 300                 return -ENOMEM;
 301
 302         mutex_lock(&mem_id_lock);
 303         id = __mem_id_cyclic_get(gfp);
 304         if (id < 0) {
 305                 errno = id;
 306                 goto err;
 307         }
 308         xdp_rxq->mem.id = id;
 309         xdp_alloc->mem  = xdp_rxq->mem;
 310         xdp_alloc->allocator = allocator;
 311
 312         /* Insert allocator into ID lookup table */
 313         ptr = rhashtable_insert_slow(mem_id_ht, &id, &xdp_alloc->node);
 314         if (IS_ERR(ptr)) {
 315                 ida_simple_remove(&mem_id_pool, xdp_rxq->mem.id);
 316                 xdp_rxq->mem.id = 0;
 317                 errno = PTR_ERR(ptr);
 318                 goto err;
 319         }
 320
 321         if (type == MEM_TYPE_PAGE_POOL)
 322                 page_pool_use_xdp_mem(allocator, mem_allocator_disconnect);
 323
 324         mutex_unlock(&mem_id_lock);
 325
 326         trace_mem_connect(xdp_alloc, xdp_rxq);
 327         return 0;
 328 err:
 329         mutex_unlock(&mem_id_lock);
 330         kfree(xdp_alloc);
 331         return errno;
 332 }
 333 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model);
 334
 335 /* XDP RX runs under NAPI protection, and in different delivery error
 336  * scenarios (e.g. queue full), it is possible to return the xdp_frame
 337  * while still leveraging this protection.  The @napi_direct boolean
 338  * is used for those calls sites.  Thus, allowing for faster recycling
 339  * of xdp_frames/pages in those cases.
 340  */
 341 static void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct,
 342                          struct xdp_buff *xdp)
 343 {
 344         struct xdp_mem_allocator *xa;
 345         struct page *page;
 346
 347         switch (mem->type) {
 348         case MEM_TYPE_PAGE_POOL:
 349                 rcu_read_lock();
 350                 /* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */
 351                 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
 352                 page = virt_to_head_page(data);
 353                 napi_direct &= !xdp_return_frame_no_direct();
 354                 page_pool_put_full_page(xa->page_pool, page, napi_direct);
 355                 rcu_read_unlock();
 356                 break;
 357         case MEM_TYPE_PAGE_SHARED:
 358                 page_frag_free(data);
 359                 break;
 360         case MEM_TYPE_PAGE_ORDER0:
 361                 page = virt_to_page(data); /* Assumes order0 page*/
 362                 put_page(page);
 363                 break;
 364         case MEM_TYPE_XSK_BUFF_POOL:
 365                 /* NB! Only valid from an xdp_buff! */
 366                 xsk_buff_free(xdp);
 367                 break;
 368         default:
 369                 /* Not possible, checked in xdp_rxq_info_reg_mem_model() */
 370                 WARN(1, "Incorrect XDP memory type (%d) usage", mem->type);
 371                 break;
 372         }
 373 }
 374
 375 void xdp_return_frame(struct xdp_frame *xdpf)
 376 {
 377         __xdp_return(xdpf->data, &xdpf->mem, false, NULL);
 378 }
 379 EXPORT_SYMBOL_GPL(xdp_return_frame);
 380
 381 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf)
 382 {
 383         __xdp_return(xdpf->data, &xdpf->mem, true, NULL);
 384 }
 385 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi);
 386
 387 /* XDP bulk APIs introduce a defer/flush mechanism to return
 388  * pages belonging to the same xdp_mem_allocator object
 389  * (identified via the mem.id field) in bulk to optimize
 390  * I-cache and D-cache.
 391  * The bulk queue size is set to 16 to be aligned to how
 392  * XDP_REDIRECT bulking works. The bulk is flushed when
 393  * it is full or when mem.id changes.
 394  * xdp_frame_bulk is usually stored/allocated on the function
 395  * call-stack to avoid locking penalties.
 396  */
 397 void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq)
 398 {
 399         struct xdp_mem_allocator *xa = bq->xa;
 400
 401         if (unlikely(!xa || !bq->count))
 402                 return;
 403
 404         page_pool_put_page_bulk(xa->page_pool, bq->q, bq->count);
 405         /* bq->xa is not cleared to save lookup, if mem.id same in next bulk */
 406         bq->count = 0;
 407 }
 408 EXPORT_SYMBOL_GPL(xdp_flush_frame_bulk);
 409
 410 /* Must be called with rcu_read_lock held */
 411 void xdp_return_frame_bulk(struct xdp_frame *xdpf,
 412                            struct xdp_frame_bulk *bq)
 413 {
 414         struct xdp_mem_info *mem = &xdpf->mem;
 415         struct xdp_mem_allocator *xa;
 416
 417         if (mem->type != MEM_TYPE_PAGE_POOL) {
 418                 __xdp_return(xdpf->data, &xdpf->mem, false, NULL);
 419                 return;
 420         }
 421
 422         xa = bq->xa;
 423         if (unlikely(!xa)) {
 424                 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
 425                 bq->count = 0;
 426                 bq->xa = xa;
 427         }
 428
 429         if (bq->count == XDP_BULK_QUEUE_SIZE)
 430                 xdp_flush_frame_bulk(bq);
 431
 432         if (unlikely(mem->id != xa->mem.id)) {
 433                 xdp_flush_frame_bulk(bq);
 434                 bq->xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
 435         }
 436
 437         bq->q[bq->count++] = xdpf->data;
 438 }
 439 EXPORT_SYMBOL_GPL(xdp_return_frame_bulk);
 440
 441 void xdp_return_buff(struct xdp_buff *xdp)
 442 {
 443         __xdp_return(xdp->data, &xdp->rxq->mem, true, xdp);
 444 }
 445
 446 /* Only called for MEM_TYPE_PAGE_POOL see xdp.h */
 447 void __xdp_release_frame(void *data, struct xdp_mem_info *mem)
 448 {
 449         struct xdp_mem_allocator *xa;
 450         struct page *page;
 451
 452         rcu_read_lock();
 453         xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
 454         page = virt_to_head_page(data);
 455         if (xa)
 456                 page_pool_release_page(xa->page_pool, page);
 457         rcu_read_unlock();
 458 }
 459 EXPORT_SYMBOL_GPL(__xdp_release_frame);
 460
 461 void xdp_attachment_setup(struct xdp_attachment_info *info,
 462                           struct netdev_bpf *bpf)
 463 {
 464         if (info->prog)
 465                 bpf_prog_put(info->prog);
 466         info->prog = bpf->prog;
 467         info->flags = bpf->flags;
 468 }
 469 EXPORT_SYMBOL_GPL(xdp_attachment_setup);
 470
 471 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp)
 472 {
 473         unsigned int metasize, totsize;
 474         void *addr, *data_to_copy;
 475         struct xdp_frame *xdpf;
 476         struct page *page;
 477
 478         /* Clone into a MEM_TYPE_PAGE_ORDER0 xdp_frame. */
 479         metasize = xdp_data_meta_unsupported(xdp) ? 0 :
 480                    xdp->data - xdp->data_meta;
 481         totsize = xdp->data_end - xdp->data + metasize;
 482
 483         if (sizeof(*xdpf) + totsize > PAGE_SIZE)
 484                 return NULL;
 485
 486         page = dev_alloc_page();
 487         if (!page)
 488                 return NULL;
 489
 490         addr = page_to_virt(page);
 491         xdpf = addr;
 492         memset(xdpf, 0, sizeof(*xdpf));
 493
 494         addr += sizeof(*xdpf);
 495         data_to_copy = metasize ? xdp->data_meta : xdp->data;
 496         memcpy(addr, data_to_copy, totsize);
 497
 498         xdpf->data = addr + metasize;
 499         xdpf->len = totsize - metasize;
 500         xdpf->headroom = 0;
 501         xdpf->metasize = metasize;
 502         xdpf->frame_sz = PAGE_SIZE;
 503         xdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
 504
 505         xsk_buff_free(xdp);
 506         return xdpf;
 507 }
 508 EXPORT_SYMBOL_GPL(xdp_convert_zc_to_xdp_frame);
 509
 510 /* Used by XDP_WARN macro, to avoid inlining WARN() in fast-path */
 511 void xdp_warn(const char *msg, const char *func, const int line)
 512 {
 513         WARN(1, "XDP_WARN: %s(line:%d): %s\n", func, line, msg);
 514 };
 515 EXPORT_SYMBOL_GPL(xdp_warn);