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) 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; 182 183 xdp_rxq->reg_state = REG_STATE_REGISTERED; 184 return 0; 185 } 186 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg); 187 188 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq) 189 { 190 xdp_rxq->reg_state = REG_STATE_UNUSED; 191 } 192 EXPORT_SYMBOL_GPL(xdp_rxq_info_unused); 193 194 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq) 195 { 196 return (xdp_rxq->reg_state == REG_STATE_REGISTERED); 197 } 198 EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg); 199 200 static int __mem_id_init_hash_table(void) 201 { 202 struct rhashtable *rht; 203 int ret; 204 205 if (unlikely(mem_id_init)) 206 return 0; 207 208 rht = kzalloc(sizeof(*rht), GFP_KERNEL); 209 if (!rht) 210 return -ENOMEM; 211 212 ret = rhashtable_init(rht, &mem_id_rht_params); 213 if (ret < 0) { 214 kfree(rht); 215 return ret; 216 } 217 mem_id_ht = rht; 218 smp_mb(); /* mutex lock should provide enough pairing */ 219 mem_id_init = true; 220 221 return 0; 222 } 223 224 /* Allocate a cyclic ID that maps to allocator pointer. 225 * See: https://www.kernel.org/doc/html/latest/core-api/idr.html 226 * 227 * Caller must lock mem_id_lock. 228 */ 229 static int __mem_id_cyclic_get(gfp_t gfp) 230 { 231 int retries = 1; 232 int id; 233 234 again: 235 id = ida_simple_get(&mem_id_pool, mem_id_next, MEM_ID_MAX, gfp); 236 if (id < 0) { 237 if (id == -ENOSPC) { 238 /* Cyclic allocator, reset next id */ 239 if (retries--) { 240 mem_id_next = MEM_ID_MIN; 241 goto again; 242 } 243 } 244 return id; /* errno */ 245 } 246 mem_id_next = id + 1; 247 248 return id; 249 } 250 251 static bool __is_supported_mem_type(enum xdp_mem_type type) 252 { 253 if (type == MEM_TYPE_PAGE_POOL) 254 return is_page_pool_compiled_in(); 255 256 if (type >= MEM_TYPE_MAX) 257 return false; 258 259 return true; 260 } 261 262 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq, 263 enum xdp_mem_type type, void *allocator) 264 { 265 struct xdp_mem_allocator *xdp_alloc; 266 gfp_t gfp = GFP_KERNEL; 267 int id, errno, ret; 268 void *ptr; 269 270 if (xdp_rxq->reg_state != REG_STATE_REGISTERED) { 271 WARN(1, "Missing register, driver bug"); 272 return -EFAULT; 273 } 274 275 if (!__is_supported_mem_type(type)) 276 return -EOPNOTSUPP; 277 278 xdp_rxq->mem.type = type; 279 280 if (!allocator) { 281 if (type == MEM_TYPE_PAGE_POOL) 282 return -EINVAL; /* Setup time check page_pool req */ 283 return 0; 284 } 285 286 /* Delay init of rhashtable to save memory if feature isn't used */ 287 if (!mem_id_init) { 288 mutex_lock(&mem_id_lock); 289 ret = __mem_id_init_hash_table(); 290 mutex_unlock(&mem_id_lock); 291 if (ret < 0) { 292 WARN_ON(1); 293 return ret; 294 } 295 } 296 297 xdp_alloc = kzalloc(sizeof(*xdp_alloc), gfp); 298 if (!xdp_alloc) 299 return -ENOMEM; 300 301 mutex_lock(&mem_id_lock); 302 id = __mem_id_cyclic_get(gfp); 303 if (id < 0) { 304 errno = id; 305 goto err; 306 } 307 xdp_rxq->mem.id = id; 308 xdp_alloc->mem = xdp_rxq->mem; 309 xdp_alloc->allocator = allocator; 310 311 /* Insert allocator into ID lookup table */ 312 ptr = rhashtable_insert_slow(mem_id_ht, &id, &xdp_alloc->node); 313 if (IS_ERR(ptr)) { 314 ida_simple_remove(&mem_id_pool, xdp_rxq->mem.id); 315 xdp_rxq->mem.id = 0; 316 errno = PTR_ERR(ptr); 317 goto err; 318 } 319 320 if (type == MEM_TYPE_PAGE_POOL) 321 page_pool_use_xdp_mem(allocator, mem_allocator_disconnect); 322 323 mutex_unlock(&mem_id_lock); 324 325 trace_mem_connect(xdp_alloc, xdp_rxq); 326 return 0; 327 err: 328 mutex_unlock(&mem_id_lock); 329 kfree(xdp_alloc); 330 return errno; 331 } 332 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model); 333 334 /* XDP RX runs under NAPI protection, and in different delivery error 335 * scenarios (e.g. queue full), it is possible to return the xdp_frame 336 * while still leveraging this protection. The @napi_direct boolean 337 * is used for those calls sites. Thus, allowing for faster recycling 338 * of xdp_frames/pages in those cases. This path is never used by the 339 * MEM_TYPE_XSK_BUFF_POOL memory type, so it's explicitly not part of 340 * the switch-statement. 341 */ 342 static void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct) 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 default: 365 /* Not possible, checked in xdp_rxq_info_reg_mem_model() */ 366 WARN(1, "Incorrect XDP memory type (%d) usage", mem->type); 367 break; 368 } 369 } 370 371 void xdp_return_frame(struct xdp_frame *xdpf) 372 { 373 __xdp_return(xdpf->data, &xdpf->mem, false); 374 } 375 EXPORT_SYMBOL_GPL(xdp_return_frame); 376 377 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf) 378 { 379 __xdp_return(xdpf->data, &xdpf->mem, true); 380 } 381 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi); 382 383 void xdp_return_buff(struct xdp_buff *xdp) 384 { 385 __xdp_return(xdp->data, &xdp->rxq->mem, true); 386 } 387 388 /* Only called for MEM_TYPE_PAGE_POOL see xdp.h */ 389 void __xdp_release_frame(void *data, struct xdp_mem_info *mem) 390 { 391 struct xdp_mem_allocator *xa; 392 struct page *page; 393 394 rcu_read_lock(); 395 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params); 396 page = virt_to_head_page(data); 397 if (xa) 398 page_pool_release_page(xa->page_pool, page); 399 rcu_read_unlock(); 400 } 401 EXPORT_SYMBOL_GPL(__xdp_release_frame); 402 403 int xdp_attachment_query(struct xdp_attachment_info *info, 404 struct netdev_bpf *bpf) 405 { 406 bpf->prog_id = info->prog ? info->prog->aux->id : 0; 407 bpf->prog_flags = info->prog ? info->flags : 0; 408 return 0; 409 } 410 EXPORT_SYMBOL_GPL(xdp_attachment_query); 411 412 bool xdp_attachment_flags_ok(struct xdp_attachment_info *info, 413 struct netdev_bpf *bpf) 414 { 415 if (info->prog && (bpf->flags ^ info->flags) & XDP_FLAGS_MODES) { 416 NL_SET_ERR_MSG(bpf->extack, 417 "program loaded with different flags"); 418 return false; 419 } 420 return true; 421 } 422 EXPORT_SYMBOL_GPL(xdp_attachment_flags_ok); 423 424 void xdp_attachment_setup(struct xdp_attachment_info *info, 425 struct netdev_bpf *bpf) 426 { 427 if (info->prog) 428 bpf_prog_put(info->prog); 429 info->prog = bpf->prog; 430 info->flags = bpf->flags; 431 } 432 EXPORT_SYMBOL_GPL(xdp_attachment_setup); 433 434 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp) 435 { 436 unsigned int metasize, totsize; 437 void *addr, *data_to_copy; 438 struct xdp_frame *xdpf; 439 struct page *page; 440 441 /* Clone into a MEM_TYPE_PAGE_ORDER0 xdp_frame. */ 442 metasize = xdp_data_meta_unsupported(xdp) ? 0 : 443 xdp->data - xdp->data_meta; 444 totsize = xdp->data_end - xdp->data + metasize; 445 446 if (sizeof(*xdpf) + totsize > PAGE_SIZE) 447 return NULL; 448 449 page = dev_alloc_page(); 450 if (!page) 451 return NULL; 452 453 addr = page_to_virt(page); 454 xdpf = addr; 455 memset(xdpf, 0, sizeof(*xdpf)); 456 457 addr += sizeof(*xdpf); 458 data_to_copy = metasize ? xdp->data_meta : xdp->data; 459 memcpy(addr, data_to_copy, totsize); 460 461 xdpf->data = addr + metasize; 462 xdpf->len = totsize - metasize; 463 xdpf->headroom = 0; 464 xdpf->metasize = metasize; 465 xdpf->mem.type = MEM_TYPE_PAGE_ORDER0; 466 467 xsk_buff_free(xdp); 468 return xdpf; 469 } 470 EXPORT_SYMBOL_GPL(xdp_convert_zc_to_xdp_frame); 471 472 /* Used by XDP_WARN macro, to avoid inlining WARN() in fast-path */ 473 void xdp_warn(const char *msg, const char *func, const int line) 474 { 475 WARN(1, "XDP_WARN: %s(line:%d): %s\n", func, line, msg); 476 }; 477 EXPORT_SYMBOL_GPL(xdp_warn); 478