1 // SPDX-License-Identifier: GPL-2.0-only 2 /* Copyright (c) 2019 Facebook */ 3 4 #include <linux/bpf.h> 5 #include <linux/bpf_verifier.h> 6 #include <linux/btf.h> 7 #include <linux/filter.h> 8 #include <linux/slab.h> 9 #include <linux/numa.h> 10 #include <linux/seq_file.h> 11 #include <linux/refcount.h> 12 #include <linux/mutex.h> 13 #include <linux/btf_ids.h> 14 #include <linux/rcupdate_wait.h> 15 #include <linux/poll.h> 16 17 struct bpf_struct_ops_value { 18 struct bpf_struct_ops_common_value common; 19 char data[] ____cacheline_aligned_in_smp; 20 }; 21 22 #define MAX_TRAMP_IMAGE_PAGES 8 23 24 struct bpf_struct_ops_map { 25 struct bpf_map map; 26 const struct bpf_struct_ops_desc *st_ops_desc; 27 /* protect map_update */ 28 struct mutex lock; 29 /* link has all the bpf_links that is populated 30 * to the func ptr of the kernel's struct 31 * (in kvalue.data). 32 */ 33 struct bpf_link **links; 34 /* ksyms for bpf trampolines */ 35 struct bpf_ksym **ksyms; 36 u32 funcs_cnt; 37 u32 image_pages_cnt; 38 /* image_pages is an array of pages that has all the trampolines 39 * that stores the func args before calling the bpf_prog. 40 */ 41 void *image_pages[MAX_TRAMP_IMAGE_PAGES]; 42 /* The owner moduler's btf. */ 43 struct btf *btf; 44 /* uvalue->data stores the kernel struct 45 * (e.g. tcp_congestion_ops) that is more useful 46 * to userspace than the kvalue. For example, 47 * the bpf_prog's id is stored instead of the kernel 48 * address of a func ptr. 49 */ 50 struct bpf_struct_ops_value *uvalue; 51 /* kvalue.data stores the actual kernel's struct 52 * (e.g. tcp_congestion_ops) that will be 53 * registered to the kernel subsystem. 54 */ 55 struct bpf_struct_ops_value kvalue; 56 }; 57 58 struct bpf_struct_ops_link { 59 struct bpf_link link; 60 struct bpf_map __rcu *map; 61 wait_queue_head_t wait_hup; 62 }; 63 64 static DEFINE_MUTEX(update_mutex); 65 66 #define VALUE_PREFIX "bpf_struct_ops_" 67 #define VALUE_PREFIX_LEN (sizeof(VALUE_PREFIX) - 1) 68 69 const struct bpf_verifier_ops bpf_struct_ops_verifier_ops = { 70 }; 71 72 const struct bpf_prog_ops bpf_struct_ops_prog_ops = { 73 #ifdef CONFIG_NET 74 .test_run = bpf_struct_ops_test_run, 75 #endif 76 }; 77 78 BTF_ID_LIST(st_ops_ids) 79 BTF_ID(struct, module) 80 BTF_ID(struct, bpf_struct_ops_common_value) 81 82 enum { 83 IDX_MODULE_ID, 84 IDX_ST_OPS_COMMON_VALUE_ID, 85 }; 86 87 extern struct btf *btf_vmlinux; 88 89 static bool is_valid_value_type(struct btf *btf, s32 value_id, 90 const struct btf_type *type, 91 const char *value_name) 92 { 93 const struct btf_type *common_value_type; 94 const struct btf_member *member; 95 const struct btf_type *vt, *mt; 96 97 vt = btf_type_by_id(btf, value_id); 98 if (btf_vlen(vt) != 2) { 99 pr_warn("The number of %s's members should be 2, but we get %d\n", 100 value_name, btf_vlen(vt)); 101 return false; 102 } 103 member = btf_type_member(vt); 104 mt = btf_type_by_id(btf, member->type); 105 common_value_type = btf_type_by_id(btf_vmlinux, 106 st_ops_ids[IDX_ST_OPS_COMMON_VALUE_ID]); 107 if (mt != common_value_type) { 108 pr_warn("The first member of %s should be bpf_struct_ops_common_value\n", 109 value_name); 110 return false; 111 } 112 member++; 113 mt = btf_type_by_id(btf, member->type); 114 if (mt != type) { 115 pr_warn("The second member of %s should be %s\n", 116 value_name, btf_name_by_offset(btf, type->name_off)); 117 return false; 118 } 119 120 return true; 121 } 122 123 static void *bpf_struct_ops_image_alloc(void) 124 { 125 void *image; 126 int err; 127 128 err = bpf_jit_charge_modmem(PAGE_SIZE); 129 if (err) 130 return ERR_PTR(err); 131 image = arch_alloc_bpf_trampoline(PAGE_SIZE); 132 if (!image) { 133 bpf_jit_uncharge_modmem(PAGE_SIZE); 134 return ERR_PTR(-ENOMEM); 135 } 136 137 return image; 138 } 139 140 void bpf_struct_ops_image_free(void *image) 141 { 142 if (image) { 143 arch_free_bpf_trampoline(image, PAGE_SIZE); 144 bpf_jit_uncharge_modmem(PAGE_SIZE); 145 } 146 } 147 148 #define MAYBE_NULL_SUFFIX "__nullable" 149 #define REFCOUNTED_SUFFIX "__ref" 150 151 /* Prepare argument info for every nullable argument of a member of a 152 * struct_ops type. 153 * 154 * Initialize a struct bpf_struct_ops_arg_info according to type info of 155 * the arguments of a stub function. (Check kCFI for more information about 156 * stub functions.) 157 * 158 * Each member in the struct_ops type has a struct bpf_struct_ops_arg_info 159 * to provide an array of struct bpf_ctx_arg_aux, which in turn provides 160 * the information that used by the verifier to check the arguments of the 161 * BPF struct_ops program assigned to the member. Here, we only care about 162 * the arguments that are marked as __nullable. 163 * 164 * The array of struct bpf_ctx_arg_aux is eventually assigned to 165 * prog->aux->ctx_arg_info of BPF struct_ops programs and passed to the 166 * verifier. (See check_struct_ops_btf_id()) 167 * 168 * arg_info->info will be the list of struct bpf_ctx_arg_aux if success. If 169 * fails, it will be kept untouched. 170 */ 171 static int prepare_arg_info(struct btf *btf, 172 const char *st_ops_name, 173 const char *member_name, 174 const struct btf_type *func_proto, void *stub_func_addr, 175 struct bpf_struct_ops_arg_info *arg_info) 176 { 177 const struct btf_type *stub_func_proto, *pointed_type; 178 bool is_nullable = false, is_refcounted = false; 179 const struct btf_param *stub_args, *args; 180 struct bpf_ctx_arg_aux *info, *info_buf; 181 u32 nargs, arg_no, info_cnt = 0; 182 char ksym[KSYM_SYMBOL_LEN]; 183 const char *stub_fname; 184 const char *suffix; 185 s32 stub_func_id; 186 u32 arg_btf_id; 187 int offset; 188 189 stub_fname = kallsyms_lookup((unsigned long)stub_func_addr, NULL, NULL, NULL, ksym); 190 if (!stub_fname) { 191 pr_warn("Cannot find the stub function name for the %s in struct %s\n", 192 member_name, st_ops_name); 193 return -ENOENT; 194 } 195 196 stub_func_id = btf_find_by_name_kind(btf, stub_fname, BTF_KIND_FUNC); 197 if (stub_func_id < 0) { 198 pr_warn("Cannot find the stub function %s in btf\n", stub_fname); 199 return -ENOENT; 200 } 201 202 stub_func_proto = btf_type_by_id(btf, stub_func_id); 203 stub_func_proto = btf_type_by_id(btf, stub_func_proto->type); 204 205 /* Check if the number of arguments of the stub function is the same 206 * as the number of arguments of the function pointer. 207 */ 208 nargs = btf_type_vlen(func_proto); 209 if (nargs != btf_type_vlen(stub_func_proto)) { 210 pr_warn("the number of arguments of the stub function %s does not match the number of arguments of the member %s of struct %s\n", 211 stub_fname, member_name, st_ops_name); 212 return -EINVAL; 213 } 214 215 if (!nargs) 216 return 0; 217 218 args = btf_params(func_proto); 219 stub_args = btf_params(stub_func_proto); 220 221 info_buf = kcalloc(nargs, sizeof(*info_buf), GFP_KERNEL); 222 if (!info_buf) 223 return -ENOMEM; 224 225 /* Prepare info for every nullable argument */ 226 info = info_buf; 227 for (arg_no = 0; arg_no < nargs; arg_no++) { 228 /* Skip arguments that is not suffixed with 229 * "__nullable or __ref". 230 */ 231 is_nullable = btf_param_match_suffix(btf, &stub_args[arg_no], 232 MAYBE_NULL_SUFFIX); 233 is_refcounted = btf_param_match_suffix(btf, &stub_args[arg_no], 234 REFCOUNTED_SUFFIX); 235 236 if (is_nullable) 237 suffix = MAYBE_NULL_SUFFIX; 238 else if (is_refcounted) 239 suffix = REFCOUNTED_SUFFIX; 240 else 241 continue; 242 243 /* Should be a pointer to struct */ 244 pointed_type = btf_type_resolve_ptr(btf, 245 args[arg_no].type, 246 &arg_btf_id); 247 if (!pointed_type || 248 !btf_type_is_struct(pointed_type)) { 249 pr_warn("stub function %s has %s tagging to an unsupported type\n", 250 stub_fname, suffix); 251 goto err_out; 252 } 253 254 offset = btf_ctx_arg_offset(btf, func_proto, arg_no); 255 if (offset < 0) { 256 pr_warn("stub function %s has an invalid trampoline ctx offset for arg#%u\n", 257 stub_fname, arg_no); 258 goto err_out; 259 } 260 261 if (args[arg_no].type != stub_args[arg_no].type) { 262 pr_warn("arg#%u type in stub function %s does not match with its original func_proto\n", 263 arg_no, stub_fname); 264 goto err_out; 265 } 266 267 /* Fill the information of the new argument */ 268 info->btf_id = arg_btf_id; 269 info->btf = btf; 270 info->offset = offset; 271 if (is_nullable) { 272 info->reg_type = PTR_TRUSTED | PTR_TO_BTF_ID | PTR_MAYBE_NULL; 273 } else if (is_refcounted) { 274 info->reg_type = PTR_TRUSTED | PTR_TO_BTF_ID; 275 info->refcounted = true; 276 } 277 278 info++; 279 info_cnt++; 280 } 281 282 if (info_cnt) { 283 arg_info->info = info_buf; 284 arg_info->cnt = info_cnt; 285 } else { 286 kfree(info_buf); 287 } 288 289 return 0; 290 291 err_out: 292 kfree(info_buf); 293 294 return -EINVAL; 295 } 296 297 /* Clean up the arg_info in a struct bpf_struct_ops_desc. */ 298 void bpf_struct_ops_desc_release(struct bpf_struct_ops_desc *st_ops_desc) 299 { 300 struct bpf_struct_ops_arg_info *arg_info; 301 int i; 302 303 arg_info = st_ops_desc->arg_info; 304 for (i = 0; i < btf_type_vlen(st_ops_desc->type); i++) 305 kfree(arg_info[i].info); 306 307 kfree(arg_info); 308 } 309 310 static bool is_module_member(const struct btf *btf, u32 id) 311 { 312 const struct btf_type *t; 313 314 t = btf_type_resolve_ptr(btf, id, NULL); 315 if (!t) 316 return false; 317 318 if (!__btf_type_is_struct(t) && !btf_type_is_fwd(t)) 319 return false; 320 321 return !strcmp(btf_name_by_offset(btf, t->name_off), "module"); 322 } 323 324 int bpf_struct_ops_supported(const struct bpf_struct_ops *st_ops, u32 moff) 325 { 326 void *func_ptr = *(void **)(st_ops->cfi_stubs + moff); 327 328 return func_ptr ? 0 : -ENOTSUPP; 329 } 330 331 int bpf_struct_ops_desc_init(struct bpf_struct_ops_desc *st_ops_desc, 332 struct btf *btf, 333 struct bpf_verifier_log *log) 334 { 335 struct bpf_struct_ops *st_ops = st_ops_desc->st_ops; 336 struct bpf_struct_ops_arg_info *arg_info; 337 const struct btf_member *member; 338 const struct btf_type *t; 339 s32 type_id, value_id; 340 char value_name[128]; 341 const char *mname; 342 int i, err; 343 344 if (strlen(st_ops->name) + VALUE_PREFIX_LEN >= 345 sizeof(value_name)) { 346 pr_warn("struct_ops name %s is too long\n", 347 st_ops->name); 348 return -EINVAL; 349 } 350 sprintf(value_name, "%s%s", VALUE_PREFIX, st_ops->name); 351 352 if (!st_ops->cfi_stubs) { 353 pr_warn("struct_ops for %s has no cfi_stubs\n", st_ops->name); 354 return -EINVAL; 355 } 356 357 type_id = btf_find_by_name_kind(btf, st_ops->name, 358 BTF_KIND_STRUCT); 359 if (type_id < 0) { 360 pr_warn("Cannot find struct %s in %s\n", 361 st_ops->name, btf_get_name(btf)); 362 return -EINVAL; 363 } 364 t = btf_type_by_id(btf, type_id); 365 if (btf_type_vlen(t) > BPF_STRUCT_OPS_MAX_NR_MEMBERS) { 366 pr_warn("Cannot support #%u members in struct %s\n", 367 btf_type_vlen(t), st_ops->name); 368 return -EINVAL; 369 } 370 371 value_id = btf_find_by_name_kind(btf, value_name, 372 BTF_KIND_STRUCT); 373 if (value_id < 0) { 374 pr_warn("Cannot find struct %s in %s\n", 375 value_name, btf_get_name(btf)); 376 return -EINVAL; 377 } 378 if (!is_valid_value_type(btf, value_id, t, value_name)) 379 return -EINVAL; 380 381 arg_info = kcalloc(btf_type_vlen(t), sizeof(*arg_info), 382 GFP_KERNEL); 383 if (!arg_info) 384 return -ENOMEM; 385 386 st_ops_desc->arg_info = arg_info; 387 st_ops_desc->type = t; 388 st_ops_desc->type_id = type_id; 389 st_ops_desc->value_id = value_id; 390 st_ops_desc->value_type = btf_type_by_id(btf, value_id); 391 392 for_each_member(i, t, member) { 393 const struct btf_type *func_proto, *ret_type; 394 void **stub_func_addr; 395 u32 moff; 396 397 moff = __btf_member_bit_offset(t, member) / 8; 398 mname = btf_name_by_offset(btf, member->name_off); 399 if (!*mname) { 400 pr_warn("anon member in struct %s is not supported\n", 401 st_ops->name); 402 err = -EOPNOTSUPP; 403 goto errout; 404 } 405 406 if (__btf_member_bitfield_size(t, member)) { 407 pr_warn("bit field member %s in struct %s is not supported\n", 408 mname, st_ops->name); 409 err = -EOPNOTSUPP; 410 goto errout; 411 } 412 413 if (!st_ops_ids[IDX_MODULE_ID] && is_module_member(btf, member->type)) { 414 pr_warn("'struct module' btf id not found. Is CONFIG_MODULES enabled? bpf_struct_ops '%s' needs module support.\n", 415 st_ops->name); 416 err = -EOPNOTSUPP; 417 goto errout; 418 } 419 420 func_proto = btf_type_resolve_func_ptr(btf, 421 member->type, 422 NULL); 423 424 /* The member is not a function pointer or 425 * the function pointer is not supported. 426 */ 427 if (!func_proto || bpf_struct_ops_supported(st_ops, moff)) 428 continue; 429 430 if (func_proto->type) { 431 ret_type = btf_type_resolve_ptr(btf, func_proto->type, NULL); 432 if (ret_type && !__btf_type_is_struct(ret_type)) { 433 pr_warn("func ptr %s in struct %s returns non-struct pointer, which is not supported\n", 434 mname, st_ops->name); 435 err = -EOPNOTSUPP; 436 goto errout; 437 } 438 } 439 440 if (btf_distill_func_proto(log, btf, 441 func_proto, mname, 442 &st_ops->func_models[i])) { 443 pr_warn("Error in parsing func ptr %s in struct %s\n", 444 mname, st_ops->name); 445 err = -EINVAL; 446 goto errout; 447 } 448 449 stub_func_addr = *(void **)(st_ops->cfi_stubs + moff); 450 err = prepare_arg_info(btf, st_ops->name, mname, 451 func_proto, stub_func_addr, 452 arg_info + i); 453 if (err) 454 goto errout; 455 } 456 457 if (st_ops->init(btf)) { 458 pr_warn("Error in init bpf_struct_ops %s\n", 459 st_ops->name); 460 err = -EINVAL; 461 goto errout; 462 } 463 464 return 0; 465 466 errout: 467 bpf_struct_ops_desc_release(st_ops_desc); 468 469 return err; 470 } 471 472 static int bpf_struct_ops_map_get_next_key(struct bpf_map *map, void *key, 473 void *next_key) 474 { 475 if (key && *(u32 *)key == 0) 476 return -ENOENT; 477 478 *(u32 *)next_key = 0; 479 return 0; 480 } 481 482 int bpf_struct_ops_map_sys_lookup_elem(struct bpf_map *map, void *key, 483 void *value) 484 { 485 struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map; 486 struct bpf_struct_ops_value *uvalue, *kvalue; 487 enum bpf_struct_ops_state state; 488 s64 refcnt; 489 490 if (unlikely(*(u32 *)key != 0)) 491 return -ENOENT; 492 493 kvalue = &st_map->kvalue; 494 /* Pair with smp_store_release() during map_update */ 495 state = smp_load_acquire(&kvalue->common.state); 496 if (state == BPF_STRUCT_OPS_STATE_INIT) { 497 memset(value, 0, map->value_size); 498 return 0; 499 } 500 501 /* No lock is needed. state and refcnt do not need 502 * to be updated together under atomic context. 503 */ 504 uvalue = value; 505 memcpy(uvalue, st_map->uvalue, map->value_size); 506 uvalue->common.state = state; 507 508 /* This value offers the user space a general estimate of how 509 * many sockets are still utilizing this struct_ops for TCP 510 * congestion control. The number might not be exact, but it 511 * should sufficiently meet our present goals. 512 */ 513 refcnt = atomic64_read(&map->refcnt) - atomic64_read(&map->usercnt); 514 refcount_set(&uvalue->common.refcnt, max_t(s64, refcnt, 0)); 515 516 return 0; 517 } 518 519 static void *bpf_struct_ops_map_lookup_elem(struct bpf_map *map, void *key) 520 { 521 return ERR_PTR(-EINVAL); 522 } 523 524 static void bpf_struct_ops_map_put_progs(struct bpf_struct_ops_map *st_map) 525 { 526 u32 i; 527 528 for (i = 0; i < st_map->funcs_cnt; i++) { 529 if (!st_map->links[i]) 530 break; 531 bpf_link_put(st_map->links[i]); 532 st_map->links[i] = NULL; 533 } 534 } 535 536 static void bpf_struct_ops_map_dissoc_progs(struct bpf_struct_ops_map *st_map) 537 { 538 u32 i; 539 540 for (i = 0; i < st_map->funcs_cnt; i++) { 541 if (!st_map->links[i]) 542 break; 543 bpf_prog_disassoc_struct_ops(st_map->links[i]->prog); 544 } 545 } 546 547 static void bpf_struct_ops_map_free_image(struct bpf_struct_ops_map *st_map) 548 { 549 int i; 550 551 for (i = 0; i < st_map->image_pages_cnt; i++) 552 bpf_struct_ops_image_free(st_map->image_pages[i]); 553 st_map->image_pages_cnt = 0; 554 } 555 556 static int check_zero_holes(const struct btf *btf, const struct btf_type *t, void *data) 557 { 558 const struct btf_member *member; 559 u32 i, moff, msize, prev_mend = 0; 560 const struct btf_type *mtype; 561 562 for_each_member(i, t, member) { 563 moff = __btf_member_bit_offset(t, member) / 8; 564 if (moff > prev_mend && 565 memchr_inv(data + prev_mend, 0, moff - prev_mend)) 566 return -EINVAL; 567 568 mtype = btf_type_by_id(btf, member->type); 569 mtype = btf_resolve_size(btf, mtype, &msize); 570 if (IS_ERR(mtype)) 571 return PTR_ERR(mtype); 572 prev_mend = moff + msize; 573 } 574 575 if (t->size > prev_mend && 576 memchr_inv(data + prev_mend, 0, t->size - prev_mend)) 577 return -EINVAL; 578 579 return 0; 580 } 581 582 static void bpf_struct_ops_link_release(struct bpf_link *link) 583 { 584 } 585 586 static void bpf_struct_ops_link_dealloc(struct bpf_link *link) 587 { 588 struct bpf_tramp_link *tlink = container_of(link, struct bpf_tramp_link, link); 589 590 kfree(tlink); 591 } 592 593 const struct bpf_link_ops bpf_struct_ops_link_lops = { 594 .release = bpf_struct_ops_link_release, 595 .dealloc = bpf_struct_ops_link_dealloc, 596 }; 597 598 int bpf_struct_ops_prepare_trampoline(struct bpf_tramp_links *tlinks, 599 struct bpf_tramp_link *link, 600 const struct btf_func_model *model, 601 void *stub_func, 602 void **_image, u32 *_image_off, 603 bool allow_alloc) 604 { 605 u32 image_off = *_image_off, flags = BPF_TRAMP_F_INDIRECT; 606 void *image = *_image; 607 int size; 608 609 tlinks[BPF_TRAMP_FENTRY].links[0] = link; 610 tlinks[BPF_TRAMP_FENTRY].nr_links = 1; 611 612 if (model->ret_size > 0) 613 flags |= BPF_TRAMP_F_RET_FENTRY_RET; 614 615 size = arch_bpf_trampoline_size(model, flags, tlinks, stub_func); 616 if (size <= 0) 617 return size ? : -EFAULT; 618 619 /* Allocate image buffer if necessary */ 620 if (!image || size > PAGE_SIZE - image_off) { 621 if (!allow_alloc) 622 return -E2BIG; 623 624 image = bpf_struct_ops_image_alloc(); 625 if (IS_ERR(image)) 626 return PTR_ERR(image); 627 image_off = 0; 628 } 629 630 size = arch_prepare_bpf_trampoline(NULL, image + image_off, 631 image + image_off + size, 632 model, flags, tlinks, stub_func); 633 if (size <= 0) { 634 if (image != *_image) 635 bpf_struct_ops_image_free(image); 636 return size ? : -EFAULT; 637 } 638 639 *_image = image; 640 *_image_off = image_off + size; 641 return 0; 642 } 643 644 static void bpf_struct_ops_ksym_init(const char *tname, const char *mname, 645 void *image, unsigned int size, 646 struct bpf_ksym *ksym) 647 { 648 snprintf(ksym->name, KSYM_NAME_LEN, "bpf__%s_%s", tname, mname); 649 INIT_LIST_HEAD_RCU(&ksym->lnode); 650 bpf_image_ksym_init(image, size, ksym); 651 } 652 653 static void bpf_struct_ops_map_add_ksyms(struct bpf_struct_ops_map *st_map) 654 { 655 u32 i; 656 657 for (i = 0; i < st_map->funcs_cnt; i++) { 658 if (!st_map->ksyms[i]) 659 break; 660 bpf_image_ksym_add(st_map->ksyms[i]); 661 } 662 } 663 664 static void bpf_struct_ops_map_del_ksyms(struct bpf_struct_ops_map *st_map) 665 { 666 u32 i; 667 668 for (i = 0; i < st_map->funcs_cnt; i++) { 669 if (!st_map->ksyms[i]) 670 break; 671 bpf_image_ksym_del(st_map->ksyms[i]); 672 } 673 } 674 675 static void bpf_struct_ops_map_free_ksyms(struct bpf_struct_ops_map *st_map) 676 { 677 u32 i; 678 679 for (i = 0; i < st_map->funcs_cnt; i++) { 680 if (!st_map->ksyms[i]) 681 break; 682 kfree(st_map->ksyms[i]); 683 st_map->ksyms[i] = NULL; 684 } 685 } 686 687 static long bpf_struct_ops_map_update_elem(struct bpf_map *map, void *key, 688 void *value, u64 flags) 689 { 690 struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map; 691 const struct bpf_struct_ops_desc *st_ops_desc = st_map->st_ops_desc; 692 const struct bpf_struct_ops *st_ops = st_ops_desc->st_ops; 693 struct bpf_struct_ops_value *uvalue, *kvalue; 694 const struct btf_type *module_type; 695 const struct btf_member *member; 696 const struct btf_type *t = st_ops_desc->type; 697 struct bpf_tramp_links *tlinks; 698 void *udata, *kdata; 699 int prog_fd, err; 700 u32 i, trampoline_start, image_off = 0; 701 void *cur_image = NULL, *image = NULL; 702 struct bpf_link **plink; 703 struct bpf_ksym **pksym; 704 const char *tname, *mname; 705 706 if (flags) 707 return -EINVAL; 708 709 if (*(u32 *)key != 0) 710 return -E2BIG; 711 712 err = check_zero_holes(st_map->btf, st_ops_desc->value_type, value); 713 if (err) 714 return err; 715 716 uvalue = value; 717 err = check_zero_holes(st_map->btf, t, uvalue->data); 718 if (err) 719 return err; 720 721 if (uvalue->common.state || refcount_read(&uvalue->common.refcnt)) 722 return -EINVAL; 723 724 tlinks = kcalloc(BPF_TRAMP_MAX, sizeof(*tlinks), GFP_KERNEL); 725 if (!tlinks) 726 return -ENOMEM; 727 728 uvalue = (struct bpf_struct_ops_value *)st_map->uvalue; 729 kvalue = (struct bpf_struct_ops_value *)&st_map->kvalue; 730 731 mutex_lock(&st_map->lock); 732 733 if (kvalue->common.state != BPF_STRUCT_OPS_STATE_INIT) { 734 err = -EBUSY; 735 goto unlock; 736 } 737 738 memcpy(uvalue, value, map->value_size); 739 740 udata = &uvalue->data; 741 kdata = &kvalue->data; 742 743 plink = st_map->links; 744 pksym = st_map->ksyms; 745 tname = btf_name_by_offset(st_map->btf, t->name_off); 746 module_type = btf_type_by_id(btf_vmlinux, st_ops_ids[IDX_MODULE_ID]); 747 for_each_member(i, t, member) { 748 const struct btf_type *mtype, *ptype; 749 struct bpf_prog *prog; 750 struct bpf_tramp_link *link; 751 struct bpf_ksym *ksym; 752 u32 moff; 753 754 moff = __btf_member_bit_offset(t, member) / 8; 755 mname = btf_name_by_offset(st_map->btf, member->name_off); 756 ptype = btf_type_resolve_ptr(st_map->btf, member->type, NULL); 757 if (ptype == module_type) { 758 if (*(void **)(udata + moff)) 759 goto reset_unlock; 760 *(void **)(kdata + moff) = BPF_MODULE_OWNER; 761 continue; 762 } 763 764 err = st_ops->init_member(t, member, kdata, udata); 765 if (err < 0) 766 goto reset_unlock; 767 768 /* The ->init_member() has handled this member */ 769 if (err > 0) 770 continue; 771 772 /* If st_ops->init_member does not handle it, 773 * we will only handle func ptrs and zero-ed members 774 * here. Reject everything else. 775 */ 776 777 /* All non func ptr member must be 0 */ 778 if (!ptype || !btf_type_is_func_proto(ptype)) { 779 u32 msize; 780 781 mtype = btf_type_by_id(st_map->btf, member->type); 782 mtype = btf_resolve_size(st_map->btf, mtype, &msize); 783 if (IS_ERR(mtype)) { 784 err = PTR_ERR(mtype); 785 goto reset_unlock; 786 } 787 788 if (memchr_inv(udata + moff, 0, msize)) { 789 err = -EINVAL; 790 goto reset_unlock; 791 } 792 793 continue; 794 } 795 796 prog_fd = (int)(*(unsigned long *)(udata + moff)); 797 /* Similar check as the attr->attach_prog_fd */ 798 if (!prog_fd) 799 continue; 800 801 prog = bpf_prog_get(prog_fd); 802 if (IS_ERR(prog)) { 803 err = PTR_ERR(prog); 804 goto reset_unlock; 805 } 806 807 if (prog->type != BPF_PROG_TYPE_STRUCT_OPS || 808 prog->aux->attach_btf_id != st_ops_desc->type_id || 809 prog->expected_attach_type != i) { 810 bpf_prog_put(prog); 811 err = -EINVAL; 812 goto reset_unlock; 813 } 814 815 /* Poison pointer on error instead of return for backward compatibility */ 816 bpf_prog_assoc_struct_ops(prog, &st_map->map); 817 818 link = kzalloc(sizeof(*link), GFP_USER); 819 if (!link) { 820 bpf_prog_put(prog); 821 err = -ENOMEM; 822 goto reset_unlock; 823 } 824 bpf_link_init(&link->link, BPF_LINK_TYPE_STRUCT_OPS, 825 &bpf_struct_ops_link_lops, prog, prog->expected_attach_type); 826 *plink++ = &link->link; 827 828 ksym = kzalloc(sizeof(*ksym), GFP_USER); 829 if (!ksym) { 830 err = -ENOMEM; 831 goto reset_unlock; 832 } 833 *pksym++ = ksym; 834 835 trampoline_start = image_off; 836 err = bpf_struct_ops_prepare_trampoline(tlinks, link, 837 &st_ops->func_models[i], 838 *(void **)(st_ops->cfi_stubs + moff), 839 &image, &image_off, 840 st_map->image_pages_cnt < MAX_TRAMP_IMAGE_PAGES); 841 if (err) 842 goto reset_unlock; 843 844 if (cur_image != image) { 845 st_map->image_pages[st_map->image_pages_cnt++] = image; 846 cur_image = image; 847 trampoline_start = 0; 848 } 849 850 *(void **)(kdata + moff) = image + trampoline_start + cfi_get_offset(); 851 852 /* put prog_id to udata */ 853 *(unsigned long *)(udata + moff) = prog->aux->id; 854 855 /* init ksym for this trampoline */ 856 bpf_struct_ops_ksym_init(tname, mname, 857 image + trampoline_start, 858 image_off - trampoline_start, 859 ksym); 860 } 861 862 if (st_ops->validate) { 863 err = st_ops->validate(kdata); 864 if (err) 865 goto reset_unlock; 866 } 867 for (i = 0; i < st_map->image_pages_cnt; i++) { 868 err = arch_protect_bpf_trampoline(st_map->image_pages[i], 869 PAGE_SIZE); 870 if (err) 871 goto reset_unlock; 872 } 873 874 if (st_map->map.map_flags & BPF_F_LINK) { 875 err = 0; 876 /* Let bpf_link handle registration & unregistration. 877 * 878 * Pair with smp_load_acquire() during lookup_elem(). 879 */ 880 smp_store_release(&kvalue->common.state, BPF_STRUCT_OPS_STATE_READY); 881 goto unlock; 882 } 883 884 err = st_ops->reg(kdata, NULL); 885 if (likely(!err)) { 886 /* This refcnt increment on the map here after 887 * 'st_ops->reg()' is secure since the state of the 888 * map must be set to INIT at this moment, and thus 889 * bpf_struct_ops_map_delete_elem() can't unregister 890 * or transition it to TOBEFREE concurrently. 891 */ 892 bpf_map_inc(map); 893 /* Pair with smp_load_acquire() during lookup_elem(). 894 * It ensures the above udata updates (e.g. prog->aux->id) 895 * can be seen once BPF_STRUCT_OPS_STATE_INUSE is set. 896 */ 897 smp_store_release(&kvalue->common.state, BPF_STRUCT_OPS_STATE_INUSE); 898 goto unlock; 899 } 900 901 /* Error during st_ops->reg(). Can happen if this struct_ops needs to be 902 * verified as a whole, after all init_member() calls. Can also happen if 903 * there was a race in registering the struct_ops (under the same name) to 904 * a sub-system through different struct_ops's maps. 905 */ 906 907 reset_unlock: 908 bpf_struct_ops_map_free_ksyms(st_map); 909 bpf_struct_ops_map_free_image(st_map); 910 bpf_struct_ops_map_put_progs(st_map); 911 memset(uvalue, 0, map->value_size); 912 memset(kvalue, 0, map->value_size); 913 unlock: 914 kfree(tlinks); 915 mutex_unlock(&st_map->lock); 916 if (!err) 917 bpf_struct_ops_map_add_ksyms(st_map); 918 return err; 919 } 920 921 static long bpf_struct_ops_map_delete_elem(struct bpf_map *map, void *key) 922 { 923 enum bpf_struct_ops_state prev_state; 924 struct bpf_struct_ops_map *st_map; 925 926 st_map = (struct bpf_struct_ops_map *)map; 927 if (st_map->map.map_flags & BPF_F_LINK) 928 return -EOPNOTSUPP; 929 930 prev_state = cmpxchg(&st_map->kvalue.common.state, 931 BPF_STRUCT_OPS_STATE_INUSE, 932 BPF_STRUCT_OPS_STATE_TOBEFREE); 933 switch (prev_state) { 934 case BPF_STRUCT_OPS_STATE_INUSE: 935 st_map->st_ops_desc->st_ops->unreg(&st_map->kvalue.data, NULL); 936 bpf_map_put(map); 937 return 0; 938 case BPF_STRUCT_OPS_STATE_TOBEFREE: 939 return -EINPROGRESS; 940 case BPF_STRUCT_OPS_STATE_INIT: 941 return -ENOENT; 942 default: 943 WARN_ON_ONCE(1); 944 /* Should never happen. Treat it as not found. */ 945 return -ENOENT; 946 } 947 } 948 949 static void bpf_struct_ops_map_seq_show_elem(struct bpf_map *map, void *key, 950 struct seq_file *m) 951 { 952 struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map; 953 void *value; 954 int err; 955 956 value = kmalloc(map->value_size, GFP_USER | __GFP_NOWARN); 957 if (!value) 958 return; 959 960 err = bpf_struct_ops_map_sys_lookup_elem(map, key, value); 961 if (!err) { 962 btf_type_seq_show(st_map->btf, 963 map->btf_vmlinux_value_type_id, 964 value, m); 965 seq_putc(m, '\n'); 966 } 967 968 kfree(value); 969 } 970 971 static void __bpf_struct_ops_map_free(struct bpf_map *map) 972 { 973 struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map; 974 975 if (st_map->links) 976 bpf_struct_ops_map_put_progs(st_map); 977 if (st_map->ksyms) 978 bpf_struct_ops_map_free_ksyms(st_map); 979 bpf_map_area_free(st_map->links); 980 bpf_map_area_free(st_map->ksyms); 981 bpf_struct_ops_map_free_image(st_map); 982 bpf_map_area_free(st_map->uvalue); 983 bpf_map_area_free(st_map); 984 } 985 986 static void bpf_struct_ops_map_free(struct bpf_map *map) 987 { 988 struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map; 989 990 /* st_ops->owner was acquired during map_alloc to implicitly holds 991 * the btf's refcnt. The acquire was only done when btf_is_module() 992 * st_map->btf cannot be NULL here. 993 */ 994 if (btf_is_module(st_map->btf)) 995 module_put(st_map->st_ops_desc->st_ops->owner); 996 997 bpf_struct_ops_map_dissoc_progs(st_map); 998 999 bpf_struct_ops_map_del_ksyms(st_map); 1000 1001 /* The struct_ops's function may switch to another struct_ops. 1002 * 1003 * For example, bpf_tcp_cc_x->init() may switch to 1004 * another tcp_cc_y by calling 1005 * setsockopt(TCP_CONGESTION, "tcp_cc_y"). 1006 * During the switch, bpf_struct_ops_put(tcp_cc_x) is called 1007 * and its refcount may reach 0 which then free its 1008 * trampoline image while tcp_cc_x is still running. 1009 * 1010 * A vanilla rcu gp is to wait for all bpf-tcp-cc prog 1011 * to finish. bpf-tcp-cc prog is non sleepable. 1012 * A rcu_tasks gp is to wait for the last few insn 1013 * in the tramopline image to finish before releasing 1014 * the trampoline image. 1015 */ 1016 synchronize_rcu_mult(call_rcu, call_rcu_tasks); 1017 1018 __bpf_struct_ops_map_free(map); 1019 } 1020 1021 static int bpf_struct_ops_map_alloc_check(union bpf_attr *attr) 1022 { 1023 if (attr->key_size != sizeof(unsigned int) || attr->max_entries != 1 || 1024 (attr->map_flags & ~(BPF_F_LINK | BPF_F_VTYPE_BTF_OBJ_FD)) || 1025 !attr->btf_vmlinux_value_type_id) 1026 return -EINVAL; 1027 return 0; 1028 } 1029 1030 static u32 count_func_ptrs(const struct btf *btf, const struct btf_type *t) 1031 { 1032 int i; 1033 u32 count; 1034 const struct btf_member *member; 1035 1036 count = 0; 1037 for_each_member(i, t, member) 1038 if (btf_type_resolve_func_ptr(btf, member->type, NULL)) 1039 count++; 1040 return count; 1041 } 1042 1043 static struct bpf_map *bpf_struct_ops_map_alloc(union bpf_attr *attr) 1044 { 1045 const struct bpf_struct_ops_desc *st_ops_desc; 1046 size_t st_map_size; 1047 struct bpf_struct_ops_map *st_map; 1048 const struct btf_type *t, *vt; 1049 struct module *mod = NULL; 1050 struct bpf_map *map; 1051 struct btf *btf; 1052 int ret; 1053 1054 if (attr->map_flags & BPF_F_VTYPE_BTF_OBJ_FD) { 1055 /* The map holds btf for its whole life time. */ 1056 btf = btf_get_by_fd(attr->value_type_btf_obj_fd); 1057 if (IS_ERR(btf)) 1058 return ERR_CAST(btf); 1059 if (!btf_is_module(btf)) { 1060 btf_put(btf); 1061 return ERR_PTR(-EINVAL); 1062 } 1063 1064 mod = btf_try_get_module(btf); 1065 /* mod holds a refcnt to btf. We don't need an extra refcnt 1066 * here. 1067 */ 1068 btf_put(btf); 1069 if (!mod) 1070 return ERR_PTR(-EINVAL); 1071 } else { 1072 btf = bpf_get_btf_vmlinux(); 1073 if (IS_ERR(btf)) 1074 return ERR_CAST(btf); 1075 if (!btf) 1076 return ERR_PTR(-ENOTSUPP); 1077 } 1078 1079 st_ops_desc = bpf_struct_ops_find_value(btf, attr->btf_vmlinux_value_type_id); 1080 if (!st_ops_desc) { 1081 ret = -ENOTSUPP; 1082 goto errout; 1083 } 1084 1085 vt = st_ops_desc->value_type; 1086 if (attr->value_size != vt->size) { 1087 ret = -EINVAL; 1088 goto errout; 1089 } 1090 1091 t = st_ops_desc->type; 1092 1093 st_map_size = sizeof(*st_map) + 1094 /* kvalue stores the 1095 * struct bpf_struct_ops_tcp_congestions_ops 1096 */ 1097 (vt->size - sizeof(struct bpf_struct_ops_value)); 1098 1099 st_map = bpf_map_area_alloc(st_map_size, NUMA_NO_NODE); 1100 if (!st_map) { 1101 ret = -ENOMEM; 1102 goto errout; 1103 } 1104 1105 st_map->st_ops_desc = st_ops_desc; 1106 map = &st_map->map; 1107 1108 st_map->uvalue = bpf_map_area_alloc(vt->size, NUMA_NO_NODE); 1109 st_map->funcs_cnt = count_func_ptrs(btf, t); 1110 st_map->links = 1111 bpf_map_area_alloc(st_map->funcs_cnt * sizeof(struct bpf_link *), 1112 NUMA_NO_NODE); 1113 1114 st_map->ksyms = 1115 bpf_map_area_alloc(st_map->funcs_cnt * sizeof(struct bpf_ksym *), 1116 NUMA_NO_NODE); 1117 if (!st_map->uvalue || !st_map->links || !st_map->ksyms) { 1118 ret = -ENOMEM; 1119 goto errout_free; 1120 } 1121 st_map->btf = btf; 1122 1123 mutex_init(&st_map->lock); 1124 bpf_map_init_from_attr(map, attr); 1125 1126 return map; 1127 1128 errout_free: 1129 __bpf_struct_ops_map_free(map); 1130 errout: 1131 module_put(mod); 1132 1133 return ERR_PTR(ret); 1134 } 1135 1136 static u64 bpf_struct_ops_map_mem_usage(const struct bpf_map *map) 1137 { 1138 struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map; 1139 const struct bpf_struct_ops_desc *st_ops_desc = st_map->st_ops_desc; 1140 const struct btf_type *vt = st_ops_desc->value_type; 1141 u64 usage; 1142 1143 usage = sizeof(*st_map) + 1144 vt->size - sizeof(struct bpf_struct_ops_value); 1145 usage += vt->size; 1146 usage += st_map->funcs_cnt * sizeof(struct bpf_link *); 1147 usage += st_map->funcs_cnt * sizeof(struct bpf_ksym *); 1148 usage += PAGE_SIZE; 1149 return usage; 1150 } 1151 1152 BTF_ID_LIST_SINGLE(bpf_struct_ops_map_btf_ids, struct, bpf_struct_ops_map) 1153 const struct bpf_map_ops bpf_struct_ops_map_ops = { 1154 .map_alloc_check = bpf_struct_ops_map_alloc_check, 1155 .map_alloc = bpf_struct_ops_map_alloc, 1156 .map_free = bpf_struct_ops_map_free, 1157 .map_get_next_key = bpf_struct_ops_map_get_next_key, 1158 .map_lookup_elem = bpf_struct_ops_map_lookup_elem, 1159 .map_delete_elem = bpf_struct_ops_map_delete_elem, 1160 .map_update_elem = bpf_struct_ops_map_update_elem, 1161 .map_seq_show_elem = bpf_struct_ops_map_seq_show_elem, 1162 .map_mem_usage = bpf_struct_ops_map_mem_usage, 1163 .map_btf_id = &bpf_struct_ops_map_btf_ids[0], 1164 }; 1165 1166 /* "const void *" because some subsystem is 1167 * passing a const (e.g. const struct tcp_congestion_ops *) 1168 */ 1169 bool bpf_struct_ops_get(const void *kdata) 1170 { 1171 struct bpf_struct_ops_value *kvalue; 1172 struct bpf_struct_ops_map *st_map; 1173 struct bpf_map *map; 1174 1175 kvalue = container_of(kdata, struct bpf_struct_ops_value, data); 1176 st_map = container_of(kvalue, struct bpf_struct_ops_map, kvalue); 1177 1178 map = __bpf_map_inc_not_zero(&st_map->map, false); 1179 return !IS_ERR(map); 1180 } 1181 EXPORT_SYMBOL_GPL(bpf_struct_ops_get); 1182 1183 void bpf_struct_ops_put(const void *kdata) 1184 { 1185 struct bpf_struct_ops_value *kvalue; 1186 struct bpf_struct_ops_map *st_map; 1187 1188 kvalue = container_of(kdata, struct bpf_struct_ops_value, data); 1189 st_map = container_of(kvalue, struct bpf_struct_ops_map, kvalue); 1190 1191 bpf_map_put(&st_map->map); 1192 } 1193 EXPORT_SYMBOL_GPL(bpf_struct_ops_put); 1194 1195 u32 bpf_struct_ops_id(const void *kdata) 1196 { 1197 struct bpf_struct_ops_value *kvalue; 1198 struct bpf_struct_ops_map *st_map; 1199 1200 kvalue = container_of(kdata, struct bpf_struct_ops_value, data); 1201 st_map = container_of(kvalue, struct bpf_struct_ops_map, kvalue); 1202 1203 return st_map->map.id; 1204 } 1205 EXPORT_SYMBOL_GPL(bpf_struct_ops_id); 1206 1207 static bool bpf_struct_ops_valid_to_reg(struct bpf_map *map) 1208 { 1209 struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map; 1210 1211 return map->map_type == BPF_MAP_TYPE_STRUCT_OPS && 1212 map->map_flags & BPF_F_LINK && 1213 /* Pair with smp_store_release() during map_update */ 1214 smp_load_acquire(&st_map->kvalue.common.state) == BPF_STRUCT_OPS_STATE_READY; 1215 } 1216 1217 static void bpf_struct_ops_map_link_dealloc(struct bpf_link *link) 1218 { 1219 struct bpf_struct_ops_link *st_link; 1220 struct bpf_struct_ops_map *st_map; 1221 1222 st_link = container_of(link, struct bpf_struct_ops_link, link); 1223 st_map = (struct bpf_struct_ops_map *) 1224 rcu_dereference_protected(st_link->map, true); 1225 if (st_map) { 1226 st_map->st_ops_desc->st_ops->unreg(&st_map->kvalue.data, link); 1227 bpf_map_put(&st_map->map); 1228 } 1229 kfree(st_link); 1230 } 1231 1232 static void bpf_struct_ops_map_link_show_fdinfo(const struct bpf_link *link, 1233 struct seq_file *seq) 1234 { 1235 struct bpf_struct_ops_link *st_link; 1236 struct bpf_map *map; 1237 1238 st_link = container_of(link, struct bpf_struct_ops_link, link); 1239 rcu_read_lock(); 1240 map = rcu_dereference(st_link->map); 1241 if (map) 1242 seq_printf(seq, "map_id:\t%d\n", map->id); 1243 rcu_read_unlock(); 1244 } 1245 1246 static int bpf_struct_ops_map_link_fill_link_info(const struct bpf_link *link, 1247 struct bpf_link_info *info) 1248 { 1249 struct bpf_struct_ops_link *st_link; 1250 struct bpf_map *map; 1251 1252 st_link = container_of(link, struct bpf_struct_ops_link, link); 1253 rcu_read_lock(); 1254 map = rcu_dereference(st_link->map); 1255 if (map) 1256 info->struct_ops.map_id = map->id; 1257 rcu_read_unlock(); 1258 return 0; 1259 } 1260 1261 static int bpf_struct_ops_map_link_update(struct bpf_link *link, struct bpf_map *new_map, 1262 struct bpf_map *expected_old_map) 1263 { 1264 struct bpf_struct_ops_map *st_map, *old_st_map; 1265 struct bpf_map *old_map; 1266 struct bpf_struct_ops_link *st_link; 1267 int err; 1268 1269 st_link = container_of(link, struct bpf_struct_ops_link, link); 1270 st_map = container_of(new_map, struct bpf_struct_ops_map, map); 1271 1272 if (!bpf_struct_ops_valid_to_reg(new_map)) 1273 return -EINVAL; 1274 1275 if (!st_map->st_ops_desc->st_ops->update) 1276 return -EOPNOTSUPP; 1277 1278 mutex_lock(&update_mutex); 1279 1280 old_map = rcu_dereference_protected(st_link->map, lockdep_is_held(&update_mutex)); 1281 if (!old_map) { 1282 err = -ENOLINK; 1283 goto err_out; 1284 } 1285 if (expected_old_map && old_map != expected_old_map) { 1286 err = -EPERM; 1287 goto err_out; 1288 } 1289 1290 old_st_map = container_of(old_map, struct bpf_struct_ops_map, map); 1291 /* The new and old struct_ops must be the same type. */ 1292 if (st_map->st_ops_desc != old_st_map->st_ops_desc) { 1293 err = -EINVAL; 1294 goto err_out; 1295 } 1296 1297 err = st_map->st_ops_desc->st_ops->update(st_map->kvalue.data, old_st_map->kvalue.data, link); 1298 if (err) 1299 goto err_out; 1300 1301 bpf_map_inc(new_map); 1302 rcu_assign_pointer(st_link->map, new_map); 1303 bpf_map_put(old_map); 1304 1305 err_out: 1306 mutex_unlock(&update_mutex); 1307 1308 return err; 1309 } 1310 1311 static int bpf_struct_ops_map_link_detach(struct bpf_link *link) 1312 { 1313 struct bpf_struct_ops_link *st_link = container_of(link, struct bpf_struct_ops_link, link); 1314 struct bpf_struct_ops_map *st_map; 1315 struct bpf_map *map; 1316 1317 mutex_lock(&update_mutex); 1318 1319 map = rcu_dereference_protected(st_link->map, lockdep_is_held(&update_mutex)); 1320 if (!map) { 1321 mutex_unlock(&update_mutex); 1322 return 0; 1323 } 1324 st_map = container_of(map, struct bpf_struct_ops_map, map); 1325 1326 st_map->st_ops_desc->st_ops->unreg(&st_map->kvalue.data, link); 1327 1328 RCU_INIT_POINTER(st_link->map, NULL); 1329 /* Pair with bpf_map_get() in bpf_struct_ops_link_create() or 1330 * bpf_map_inc() in bpf_struct_ops_map_link_update(). 1331 */ 1332 bpf_map_put(&st_map->map); 1333 1334 mutex_unlock(&update_mutex); 1335 1336 wake_up_interruptible_poll(&st_link->wait_hup, EPOLLHUP); 1337 1338 return 0; 1339 } 1340 1341 static __poll_t bpf_struct_ops_map_link_poll(struct file *file, 1342 struct poll_table_struct *pts) 1343 { 1344 struct bpf_struct_ops_link *st_link = file->private_data; 1345 1346 poll_wait(file, &st_link->wait_hup, pts); 1347 1348 return rcu_access_pointer(st_link->map) ? 0 : EPOLLHUP; 1349 } 1350 1351 static const struct bpf_link_ops bpf_struct_ops_map_lops = { 1352 .dealloc = bpf_struct_ops_map_link_dealloc, 1353 .detach = bpf_struct_ops_map_link_detach, 1354 .show_fdinfo = bpf_struct_ops_map_link_show_fdinfo, 1355 .fill_link_info = bpf_struct_ops_map_link_fill_link_info, 1356 .update_map = bpf_struct_ops_map_link_update, 1357 .poll = bpf_struct_ops_map_link_poll, 1358 }; 1359 1360 int bpf_struct_ops_link_create(union bpf_attr *attr) 1361 { 1362 struct bpf_struct_ops_link *link = NULL; 1363 struct bpf_link_primer link_primer; 1364 struct bpf_struct_ops_map *st_map; 1365 struct bpf_map *map; 1366 int err; 1367 1368 map = bpf_map_get(attr->link_create.map_fd); 1369 if (IS_ERR(map)) 1370 return PTR_ERR(map); 1371 1372 st_map = (struct bpf_struct_ops_map *)map; 1373 1374 if (!bpf_struct_ops_valid_to_reg(map)) { 1375 err = -EINVAL; 1376 goto err_out; 1377 } 1378 1379 link = kzalloc(sizeof(*link), GFP_USER); 1380 if (!link) { 1381 err = -ENOMEM; 1382 goto err_out; 1383 } 1384 bpf_link_init(&link->link, BPF_LINK_TYPE_STRUCT_OPS, &bpf_struct_ops_map_lops, NULL, 1385 attr->link_create.attach_type); 1386 1387 err = bpf_link_prime(&link->link, &link_primer); 1388 if (err) 1389 goto err_out; 1390 1391 init_waitqueue_head(&link->wait_hup); 1392 1393 /* Hold the update_mutex such that the subsystem cannot 1394 * do link->ops->detach() before the link is fully initialized. 1395 */ 1396 mutex_lock(&update_mutex); 1397 err = st_map->st_ops_desc->st_ops->reg(st_map->kvalue.data, &link->link); 1398 if (err) { 1399 mutex_unlock(&update_mutex); 1400 bpf_link_cleanup(&link_primer); 1401 link = NULL; 1402 goto err_out; 1403 } 1404 RCU_INIT_POINTER(link->map, map); 1405 mutex_unlock(&update_mutex); 1406 1407 return bpf_link_settle(&link_primer); 1408 1409 err_out: 1410 bpf_map_put(map); 1411 kfree(link); 1412 return err; 1413 } 1414 1415 int bpf_prog_assoc_struct_ops(struct bpf_prog *prog, struct bpf_map *map) 1416 { 1417 struct bpf_map *st_ops_assoc; 1418 1419 guard(mutex)(&prog->aux->st_ops_assoc_mutex); 1420 1421 st_ops_assoc = rcu_dereference_protected(prog->aux->st_ops_assoc, 1422 lockdep_is_held(&prog->aux->st_ops_assoc_mutex)); 1423 if (st_ops_assoc && st_ops_assoc == map) 1424 return 0; 1425 1426 if (st_ops_assoc) { 1427 if (prog->type != BPF_PROG_TYPE_STRUCT_OPS) 1428 return -EBUSY; 1429 1430 rcu_assign_pointer(prog->aux->st_ops_assoc, BPF_PTR_POISON); 1431 } else { 1432 /* 1433 * struct_ops map does not track associated non-struct_ops programs. 1434 * Bump the refcount to make sure st_ops_assoc is always valid. 1435 */ 1436 if (prog->type != BPF_PROG_TYPE_STRUCT_OPS) 1437 bpf_map_inc(map); 1438 1439 rcu_assign_pointer(prog->aux->st_ops_assoc, map); 1440 } 1441 1442 return 0; 1443 } 1444 1445 void bpf_prog_disassoc_struct_ops(struct bpf_prog *prog) 1446 { 1447 struct bpf_map *st_ops_assoc; 1448 1449 guard(mutex)(&prog->aux->st_ops_assoc_mutex); 1450 1451 st_ops_assoc = rcu_dereference_protected(prog->aux->st_ops_assoc, 1452 lockdep_is_held(&prog->aux->st_ops_assoc_mutex)); 1453 if (!st_ops_assoc || st_ops_assoc == BPF_PTR_POISON) 1454 return; 1455 1456 if (prog->type != BPF_PROG_TYPE_STRUCT_OPS) 1457 bpf_map_put(st_ops_assoc); 1458 1459 RCU_INIT_POINTER(prog->aux->st_ops_assoc, NULL); 1460 } 1461 1462 /* 1463 * Get a reference to the struct_ops struct (i.e., kdata) associated with a 1464 * program. Should only be called in BPF program context (e.g., in a kfunc). 1465 * 1466 * If the returned pointer is not NULL, it must points to a valid struct_ops. 1467 * The struct_ops map is not guaranteed to be initialized nor attached. 1468 * Kernel struct_ops implementers are responsible for tracking and checking 1469 * the state of the struct_ops if the use case requires an initialized or 1470 * attached struct_ops. 1471 */ 1472 void *bpf_prog_get_assoc_struct_ops(const struct bpf_prog_aux *aux) 1473 { 1474 struct bpf_struct_ops_map *st_map; 1475 struct bpf_map *st_ops_assoc; 1476 1477 st_ops_assoc = rcu_dereference_check(aux->st_ops_assoc, bpf_rcu_lock_held()); 1478 if (!st_ops_assoc || st_ops_assoc == BPF_PTR_POISON) 1479 return NULL; 1480 1481 st_map = (struct bpf_struct_ops_map *)st_ops_assoc; 1482 1483 return &st_map->kvalue.data; 1484 } 1485 EXPORT_SYMBOL_GPL(bpf_prog_get_assoc_struct_ops); 1486 1487 void bpf_map_struct_ops_info_fill(struct bpf_map_info *info, struct bpf_map *map) 1488 { 1489 struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map; 1490 1491 info->btf_vmlinux_id = btf_obj_id(st_map->btf); 1492 } 1493