1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Functions to manage eBPF programs attached to cgroups 4 * 5 * Copyright (c) 2016 Daniel Mack 6 */ 7 8 #include <linux/kernel.h> 9 #include <linux/atomic.h> 10 #include <linux/cgroup.h> 11 #include <linux/filter.h> 12 #include <linux/slab.h> 13 #include <linux/sysctl.h> 14 #include <linux/string.h> 15 #include <linux/bpf.h> 16 #include <linux/bpf-cgroup.h> 17 #include <net/sock.h> 18 #include <net/bpf_sk_storage.h> 19 20 #include "../cgroup/cgroup-internal.h" 21 22 DEFINE_STATIC_KEY_ARRAY_FALSE(cgroup_bpf_enabled_key, MAX_BPF_ATTACH_TYPE); 23 EXPORT_SYMBOL(cgroup_bpf_enabled_key); 24 25 void cgroup_bpf_offline(struct cgroup *cgrp) 26 { 27 cgroup_get(cgrp); 28 percpu_ref_kill(&cgrp->bpf.refcnt); 29 } 30 31 static void bpf_cgroup_storages_free(struct bpf_cgroup_storage *storages[]) 32 { 33 enum bpf_cgroup_storage_type stype; 34 35 for_each_cgroup_storage_type(stype) 36 bpf_cgroup_storage_free(storages[stype]); 37 } 38 39 static int bpf_cgroup_storages_alloc(struct bpf_cgroup_storage *storages[], 40 struct bpf_cgroup_storage *new_storages[], 41 enum bpf_attach_type type, 42 struct bpf_prog *prog, 43 struct cgroup *cgrp) 44 { 45 enum bpf_cgroup_storage_type stype; 46 struct bpf_cgroup_storage_key key; 47 struct bpf_map *map; 48 49 key.cgroup_inode_id = cgroup_id(cgrp); 50 key.attach_type = type; 51 52 for_each_cgroup_storage_type(stype) { 53 map = prog->aux->cgroup_storage[stype]; 54 if (!map) 55 continue; 56 57 storages[stype] = cgroup_storage_lookup((void *)map, &key, false); 58 if (storages[stype]) 59 continue; 60 61 storages[stype] = bpf_cgroup_storage_alloc(prog, stype); 62 if (IS_ERR(storages[stype])) { 63 bpf_cgroup_storages_free(new_storages); 64 return -ENOMEM; 65 } 66 67 new_storages[stype] = storages[stype]; 68 } 69 70 return 0; 71 } 72 73 static void bpf_cgroup_storages_assign(struct bpf_cgroup_storage *dst[], 74 struct bpf_cgroup_storage *src[]) 75 { 76 enum bpf_cgroup_storage_type stype; 77 78 for_each_cgroup_storage_type(stype) 79 dst[stype] = src[stype]; 80 } 81 82 static void bpf_cgroup_storages_link(struct bpf_cgroup_storage *storages[], 83 struct cgroup *cgrp, 84 enum bpf_attach_type attach_type) 85 { 86 enum bpf_cgroup_storage_type stype; 87 88 for_each_cgroup_storage_type(stype) 89 bpf_cgroup_storage_link(storages[stype], cgrp, attach_type); 90 } 91 92 /* Called when bpf_cgroup_link is auto-detached from dying cgroup. 93 * It drops cgroup and bpf_prog refcounts, and marks bpf_link as defunct. It 94 * doesn't free link memory, which will eventually be done by bpf_link's 95 * release() callback, when its last FD is closed. 96 */ 97 static void bpf_cgroup_link_auto_detach(struct bpf_cgroup_link *link) 98 { 99 cgroup_put(link->cgroup); 100 link->cgroup = NULL; 101 } 102 103 /** 104 * cgroup_bpf_release() - put references of all bpf programs and 105 * release all cgroup bpf data 106 * @work: work structure embedded into the cgroup to modify 107 */ 108 static void cgroup_bpf_release(struct work_struct *work) 109 { 110 struct cgroup *p, *cgrp = container_of(work, struct cgroup, 111 bpf.release_work); 112 struct bpf_prog_array *old_array; 113 struct list_head *storages = &cgrp->bpf.storages; 114 struct bpf_cgroup_storage *storage, *stmp; 115 116 unsigned int type; 117 118 mutex_lock(&cgroup_mutex); 119 120 for (type = 0; type < ARRAY_SIZE(cgrp->bpf.progs); type++) { 121 struct list_head *progs = &cgrp->bpf.progs[type]; 122 struct bpf_prog_list *pl, *pltmp; 123 124 list_for_each_entry_safe(pl, pltmp, progs, node) { 125 list_del(&pl->node); 126 if (pl->prog) 127 bpf_prog_put(pl->prog); 128 if (pl->link) 129 bpf_cgroup_link_auto_detach(pl->link); 130 kfree(pl); 131 static_branch_dec(&cgroup_bpf_enabled_key[type]); 132 } 133 old_array = rcu_dereference_protected( 134 cgrp->bpf.effective[type], 135 lockdep_is_held(&cgroup_mutex)); 136 bpf_prog_array_free(old_array); 137 } 138 139 list_for_each_entry_safe(storage, stmp, storages, list_cg) { 140 bpf_cgroup_storage_unlink(storage); 141 bpf_cgroup_storage_free(storage); 142 } 143 144 mutex_unlock(&cgroup_mutex); 145 146 for (p = cgroup_parent(cgrp); p; p = cgroup_parent(p)) 147 cgroup_bpf_put(p); 148 149 percpu_ref_exit(&cgrp->bpf.refcnt); 150 cgroup_put(cgrp); 151 } 152 153 /** 154 * cgroup_bpf_release_fn() - callback used to schedule releasing 155 * of bpf cgroup data 156 * @ref: percpu ref counter structure 157 */ 158 static void cgroup_bpf_release_fn(struct percpu_ref *ref) 159 { 160 struct cgroup *cgrp = container_of(ref, struct cgroup, bpf.refcnt); 161 162 INIT_WORK(&cgrp->bpf.release_work, cgroup_bpf_release); 163 queue_work(system_wq, &cgrp->bpf.release_work); 164 } 165 166 /* Get underlying bpf_prog of bpf_prog_list entry, regardless if it's through 167 * link or direct prog. 168 */ 169 static struct bpf_prog *prog_list_prog(struct bpf_prog_list *pl) 170 { 171 if (pl->prog) 172 return pl->prog; 173 if (pl->link) 174 return pl->link->link.prog; 175 return NULL; 176 } 177 178 /* count number of elements in the list. 179 * it's slow but the list cannot be long 180 */ 181 static u32 prog_list_length(struct list_head *head) 182 { 183 struct bpf_prog_list *pl; 184 u32 cnt = 0; 185 186 list_for_each_entry(pl, head, node) { 187 if (!prog_list_prog(pl)) 188 continue; 189 cnt++; 190 } 191 return cnt; 192 } 193 194 /* if parent has non-overridable prog attached, 195 * disallow attaching new programs to the descendent cgroup. 196 * if parent has overridable or multi-prog, allow attaching 197 */ 198 static bool hierarchy_allows_attach(struct cgroup *cgrp, 199 enum bpf_attach_type type) 200 { 201 struct cgroup *p; 202 203 p = cgroup_parent(cgrp); 204 if (!p) 205 return true; 206 do { 207 u32 flags = p->bpf.flags[type]; 208 u32 cnt; 209 210 if (flags & BPF_F_ALLOW_MULTI) 211 return true; 212 cnt = prog_list_length(&p->bpf.progs[type]); 213 WARN_ON_ONCE(cnt > 1); 214 if (cnt == 1) 215 return !!(flags & BPF_F_ALLOW_OVERRIDE); 216 p = cgroup_parent(p); 217 } while (p); 218 return true; 219 } 220 221 /* compute a chain of effective programs for a given cgroup: 222 * start from the list of programs in this cgroup and add 223 * all parent programs. 224 * Note that parent's F_ALLOW_OVERRIDE-type program is yielding 225 * to programs in this cgroup 226 */ 227 static int compute_effective_progs(struct cgroup *cgrp, 228 enum bpf_attach_type type, 229 struct bpf_prog_array **array) 230 { 231 struct bpf_prog_array_item *item; 232 struct bpf_prog_array *progs; 233 struct bpf_prog_list *pl; 234 struct cgroup *p = cgrp; 235 int cnt = 0; 236 237 /* count number of effective programs by walking parents */ 238 do { 239 if (cnt == 0 || (p->bpf.flags[type] & BPF_F_ALLOW_MULTI)) 240 cnt += prog_list_length(&p->bpf.progs[type]); 241 p = cgroup_parent(p); 242 } while (p); 243 244 progs = bpf_prog_array_alloc(cnt, GFP_KERNEL); 245 if (!progs) 246 return -ENOMEM; 247 248 /* populate the array with effective progs */ 249 cnt = 0; 250 p = cgrp; 251 do { 252 if (cnt > 0 && !(p->bpf.flags[type] & BPF_F_ALLOW_MULTI)) 253 continue; 254 255 list_for_each_entry(pl, &p->bpf.progs[type], node) { 256 if (!prog_list_prog(pl)) 257 continue; 258 259 item = &progs->items[cnt]; 260 item->prog = prog_list_prog(pl); 261 bpf_cgroup_storages_assign(item->cgroup_storage, 262 pl->storage); 263 cnt++; 264 } 265 } while ((p = cgroup_parent(p))); 266 267 *array = progs; 268 return 0; 269 } 270 271 static void activate_effective_progs(struct cgroup *cgrp, 272 enum bpf_attach_type type, 273 struct bpf_prog_array *old_array) 274 { 275 old_array = rcu_replace_pointer(cgrp->bpf.effective[type], old_array, 276 lockdep_is_held(&cgroup_mutex)); 277 /* free prog array after grace period, since __cgroup_bpf_run_*() 278 * might be still walking the array 279 */ 280 bpf_prog_array_free(old_array); 281 } 282 283 /** 284 * cgroup_bpf_inherit() - inherit effective programs from parent 285 * @cgrp: the cgroup to modify 286 */ 287 int cgroup_bpf_inherit(struct cgroup *cgrp) 288 { 289 /* has to use marco instead of const int, since compiler thinks 290 * that array below is variable length 291 */ 292 #define NR ARRAY_SIZE(cgrp->bpf.effective) 293 struct bpf_prog_array *arrays[NR] = {}; 294 struct cgroup *p; 295 int ret, i; 296 297 ret = percpu_ref_init(&cgrp->bpf.refcnt, cgroup_bpf_release_fn, 0, 298 GFP_KERNEL); 299 if (ret) 300 return ret; 301 302 for (p = cgroup_parent(cgrp); p; p = cgroup_parent(p)) 303 cgroup_bpf_get(p); 304 305 for (i = 0; i < NR; i++) 306 INIT_LIST_HEAD(&cgrp->bpf.progs[i]); 307 308 INIT_LIST_HEAD(&cgrp->bpf.storages); 309 310 for (i = 0; i < NR; i++) 311 if (compute_effective_progs(cgrp, i, &arrays[i])) 312 goto cleanup; 313 314 for (i = 0; i < NR; i++) 315 activate_effective_progs(cgrp, i, arrays[i]); 316 317 return 0; 318 cleanup: 319 for (i = 0; i < NR; i++) 320 bpf_prog_array_free(arrays[i]); 321 322 for (p = cgroup_parent(cgrp); p; p = cgroup_parent(p)) 323 cgroup_bpf_put(p); 324 325 percpu_ref_exit(&cgrp->bpf.refcnt); 326 327 return -ENOMEM; 328 } 329 330 static int update_effective_progs(struct cgroup *cgrp, 331 enum bpf_attach_type type) 332 { 333 struct cgroup_subsys_state *css; 334 int err; 335 336 /* allocate and recompute effective prog arrays */ 337 css_for_each_descendant_pre(css, &cgrp->self) { 338 struct cgroup *desc = container_of(css, struct cgroup, self); 339 340 if (percpu_ref_is_zero(&desc->bpf.refcnt)) 341 continue; 342 343 err = compute_effective_progs(desc, type, &desc->bpf.inactive); 344 if (err) 345 goto cleanup; 346 } 347 348 /* all allocations were successful. Activate all prog arrays */ 349 css_for_each_descendant_pre(css, &cgrp->self) { 350 struct cgroup *desc = container_of(css, struct cgroup, self); 351 352 if (percpu_ref_is_zero(&desc->bpf.refcnt)) { 353 if (unlikely(desc->bpf.inactive)) { 354 bpf_prog_array_free(desc->bpf.inactive); 355 desc->bpf.inactive = NULL; 356 } 357 continue; 358 } 359 360 activate_effective_progs(desc, type, desc->bpf.inactive); 361 desc->bpf.inactive = NULL; 362 } 363 364 return 0; 365 366 cleanup: 367 /* oom while computing effective. Free all computed effective arrays 368 * since they were not activated 369 */ 370 css_for_each_descendant_pre(css, &cgrp->self) { 371 struct cgroup *desc = container_of(css, struct cgroup, self); 372 373 bpf_prog_array_free(desc->bpf.inactive); 374 desc->bpf.inactive = NULL; 375 } 376 377 return err; 378 } 379 380 #define BPF_CGROUP_MAX_PROGS 64 381 382 static struct bpf_prog_list *find_attach_entry(struct list_head *progs, 383 struct bpf_prog *prog, 384 struct bpf_cgroup_link *link, 385 struct bpf_prog *replace_prog, 386 bool allow_multi) 387 { 388 struct bpf_prog_list *pl; 389 390 /* single-attach case */ 391 if (!allow_multi) { 392 if (list_empty(progs)) 393 return NULL; 394 return list_first_entry(progs, typeof(*pl), node); 395 } 396 397 list_for_each_entry(pl, progs, node) { 398 if (prog && pl->prog == prog && prog != replace_prog) 399 /* disallow attaching the same prog twice */ 400 return ERR_PTR(-EINVAL); 401 if (link && pl->link == link) 402 /* disallow attaching the same link twice */ 403 return ERR_PTR(-EINVAL); 404 } 405 406 /* direct prog multi-attach w/ replacement case */ 407 if (replace_prog) { 408 list_for_each_entry(pl, progs, node) { 409 if (pl->prog == replace_prog) 410 /* a match found */ 411 return pl; 412 } 413 /* prog to replace not found for cgroup */ 414 return ERR_PTR(-ENOENT); 415 } 416 417 return NULL; 418 } 419 420 /** 421 * __cgroup_bpf_attach() - Attach the program or the link to a cgroup, and 422 * propagate the change to descendants 423 * @cgrp: The cgroup which descendants to traverse 424 * @prog: A program to attach 425 * @link: A link to attach 426 * @replace_prog: Previously attached program to replace if BPF_F_REPLACE is set 427 * @type: Type of attach operation 428 * @flags: Option flags 429 * 430 * Exactly one of @prog or @link can be non-null. 431 * Must be called with cgroup_mutex held. 432 */ 433 int __cgroup_bpf_attach(struct cgroup *cgrp, 434 struct bpf_prog *prog, struct bpf_prog *replace_prog, 435 struct bpf_cgroup_link *link, 436 enum bpf_attach_type type, u32 flags) 437 { 438 u32 saved_flags = (flags & (BPF_F_ALLOW_OVERRIDE | BPF_F_ALLOW_MULTI)); 439 struct list_head *progs = &cgrp->bpf.progs[type]; 440 struct bpf_prog *old_prog = NULL; 441 struct bpf_cgroup_storage *storage[MAX_BPF_CGROUP_STORAGE_TYPE] = {}; 442 struct bpf_cgroup_storage *new_storage[MAX_BPF_CGROUP_STORAGE_TYPE] = {}; 443 struct bpf_prog_list *pl; 444 int err; 445 446 if (((flags & BPF_F_ALLOW_OVERRIDE) && (flags & BPF_F_ALLOW_MULTI)) || 447 ((flags & BPF_F_REPLACE) && !(flags & BPF_F_ALLOW_MULTI))) 448 /* invalid combination */ 449 return -EINVAL; 450 if (link && (prog || replace_prog)) 451 /* only either link or prog/replace_prog can be specified */ 452 return -EINVAL; 453 if (!!replace_prog != !!(flags & BPF_F_REPLACE)) 454 /* replace_prog implies BPF_F_REPLACE, and vice versa */ 455 return -EINVAL; 456 457 if (!hierarchy_allows_attach(cgrp, type)) 458 return -EPERM; 459 460 if (!list_empty(progs) && cgrp->bpf.flags[type] != saved_flags) 461 /* Disallow attaching non-overridable on top 462 * of existing overridable in this cgroup. 463 * Disallow attaching multi-prog if overridable or none 464 */ 465 return -EPERM; 466 467 if (prog_list_length(progs) >= BPF_CGROUP_MAX_PROGS) 468 return -E2BIG; 469 470 pl = find_attach_entry(progs, prog, link, replace_prog, 471 flags & BPF_F_ALLOW_MULTI); 472 if (IS_ERR(pl)) 473 return PTR_ERR(pl); 474 475 if (bpf_cgroup_storages_alloc(storage, new_storage, type, 476 prog ? : link->link.prog, cgrp)) 477 return -ENOMEM; 478 479 if (pl) { 480 old_prog = pl->prog; 481 } else { 482 pl = kmalloc(sizeof(*pl), GFP_KERNEL); 483 if (!pl) { 484 bpf_cgroup_storages_free(new_storage); 485 return -ENOMEM; 486 } 487 list_add_tail(&pl->node, progs); 488 } 489 490 pl->prog = prog; 491 pl->link = link; 492 bpf_cgroup_storages_assign(pl->storage, storage); 493 cgrp->bpf.flags[type] = saved_flags; 494 495 err = update_effective_progs(cgrp, type); 496 if (err) 497 goto cleanup; 498 499 if (old_prog) 500 bpf_prog_put(old_prog); 501 else 502 static_branch_inc(&cgroup_bpf_enabled_key[type]); 503 bpf_cgroup_storages_link(new_storage, cgrp, type); 504 return 0; 505 506 cleanup: 507 if (old_prog) { 508 pl->prog = old_prog; 509 pl->link = NULL; 510 } 511 bpf_cgroup_storages_free(new_storage); 512 if (!old_prog) { 513 list_del(&pl->node); 514 kfree(pl); 515 } 516 return err; 517 } 518 519 /* Swap updated BPF program for given link in effective program arrays across 520 * all descendant cgroups. This function is guaranteed to succeed. 521 */ 522 static void replace_effective_prog(struct cgroup *cgrp, 523 enum bpf_attach_type type, 524 struct bpf_cgroup_link *link) 525 { 526 struct bpf_prog_array_item *item; 527 struct cgroup_subsys_state *css; 528 struct bpf_prog_array *progs; 529 struct bpf_prog_list *pl; 530 struct list_head *head; 531 struct cgroup *cg; 532 int pos; 533 534 css_for_each_descendant_pre(css, &cgrp->self) { 535 struct cgroup *desc = container_of(css, struct cgroup, self); 536 537 if (percpu_ref_is_zero(&desc->bpf.refcnt)) 538 continue; 539 540 /* find position of link in effective progs array */ 541 for (pos = 0, cg = desc; cg; cg = cgroup_parent(cg)) { 542 if (pos && !(cg->bpf.flags[type] & BPF_F_ALLOW_MULTI)) 543 continue; 544 545 head = &cg->bpf.progs[type]; 546 list_for_each_entry(pl, head, node) { 547 if (!prog_list_prog(pl)) 548 continue; 549 if (pl->link == link) 550 goto found; 551 pos++; 552 } 553 } 554 found: 555 BUG_ON(!cg); 556 progs = rcu_dereference_protected( 557 desc->bpf.effective[type], 558 lockdep_is_held(&cgroup_mutex)); 559 item = &progs->items[pos]; 560 WRITE_ONCE(item->prog, link->link.prog); 561 } 562 } 563 564 /** 565 * __cgroup_bpf_replace() - Replace link's program and propagate the change 566 * to descendants 567 * @cgrp: The cgroup which descendants to traverse 568 * @link: A link for which to replace BPF program 569 * @type: Type of attach operation 570 * 571 * Must be called with cgroup_mutex held. 572 */ 573 static int __cgroup_bpf_replace(struct cgroup *cgrp, 574 struct bpf_cgroup_link *link, 575 struct bpf_prog *new_prog) 576 { 577 struct list_head *progs = &cgrp->bpf.progs[link->type]; 578 struct bpf_prog *old_prog; 579 struct bpf_prog_list *pl; 580 bool found = false; 581 582 if (link->link.prog->type != new_prog->type) 583 return -EINVAL; 584 585 list_for_each_entry(pl, progs, node) { 586 if (pl->link == link) { 587 found = true; 588 break; 589 } 590 } 591 if (!found) 592 return -ENOENT; 593 594 old_prog = xchg(&link->link.prog, new_prog); 595 replace_effective_prog(cgrp, link->type, link); 596 bpf_prog_put(old_prog); 597 return 0; 598 } 599 600 static int cgroup_bpf_replace(struct bpf_link *link, struct bpf_prog *new_prog, 601 struct bpf_prog *old_prog) 602 { 603 struct bpf_cgroup_link *cg_link; 604 int ret; 605 606 cg_link = container_of(link, struct bpf_cgroup_link, link); 607 608 mutex_lock(&cgroup_mutex); 609 /* link might have been auto-released by dying cgroup, so fail */ 610 if (!cg_link->cgroup) { 611 ret = -ENOLINK; 612 goto out_unlock; 613 } 614 if (old_prog && link->prog != old_prog) { 615 ret = -EPERM; 616 goto out_unlock; 617 } 618 ret = __cgroup_bpf_replace(cg_link->cgroup, cg_link, new_prog); 619 out_unlock: 620 mutex_unlock(&cgroup_mutex); 621 return ret; 622 } 623 624 static struct bpf_prog_list *find_detach_entry(struct list_head *progs, 625 struct bpf_prog *prog, 626 struct bpf_cgroup_link *link, 627 bool allow_multi) 628 { 629 struct bpf_prog_list *pl; 630 631 if (!allow_multi) { 632 if (list_empty(progs)) 633 /* report error when trying to detach and nothing is attached */ 634 return ERR_PTR(-ENOENT); 635 636 /* to maintain backward compatibility NONE and OVERRIDE cgroups 637 * allow detaching with invalid FD (prog==NULL) in legacy mode 638 */ 639 return list_first_entry(progs, typeof(*pl), node); 640 } 641 642 if (!prog && !link) 643 /* to detach MULTI prog the user has to specify valid FD 644 * of the program or link to be detached 645 */ 646 return ERR_PTR(-EINVAL); 647 648 /* find the prog or link and detach it */ 649 list_for_each_entry(pl, progs, node) { 650 if (pl->prog == prog && pl->link == link) 651 return pl; 652 } 653 return ERR_PTR(-ENOENT); 654 } 655 656 /** 657 * __cgroup_bpf_detach() - Detach the program or link from a cgroup, and 658 * propagate the change to descendants 659 * @cgrp: The cgroup which descendants to traverse 660 * @prog: A program to detach or NULL 661 * @prog: A link to detach or NULL 662 * @type: Type of detach operation 663 * 664 * At most one of @prog or @link can be non-NULL. 665 * Must be called with cgroup_mutex held. 666 */ 667 int __cgroup_bpf_detach(struct cgroup *cgrp, struct bpf_prog *prog, 668 struct bpf_cgroup_link *link, enum bpf_attach_type type) 669 { 670 struct list_head *progs = &cgrp->bpf.progs[type]; 671 u32 flags = cgrp->bpf.flags[type]; 672 struct bpf_prog_list *pl; 673 struct bpf_prog *old_prog; 674 int err; 675 676 if (prog && link) 677 /* only one of prog or link can be specified */ 678 return -EINVAL; 679 680 pl = find_detach_entry(progs, prog, link, flags & BPF_F_ALLOW_MULTI); 681 if (IS_ERR(pl)) 682 return PTR_ERR(pl); 683 684 /* mark it deleted, so it's ignored while recomputing effective */ 685 old_prog = pl->prog; 686 pl->prog = NULL; 687 pl->link = NULL; 688 689 err = update_effective_progs(cgrp, type); 690 if (err) 691 goto cleanup; 692 693 /* now can actually delete it from this cgroup list */ 694 list_del(&pl->node); 695 kfree(pl); 696 if (list_empty(progs)) 697 /* last program was detached, reset flags to zero */ 698 cgrp->bpf.flags[type] = 0; 699 if (old_prog) 700 bpf_prog_put(old_prog); 701 static_branch_dec(&cgroup_bpf_enabled_key[type]); 702 return 0; 703 704 cleanup: 705 /* restore back prog or link */ 706 pl->prog = old_prog; 707 pl->link = link; 708 return err; 709 } 710 711 /* Must be called with cgroup_mutex held to avoid races. */ 712 int __cgroup_bpf_query(struct cgroup *cgrp, const union bpf_attr *attr, 713 union bpf_attr __user *uattr) 714 { 715 __u32 __user *prog_ids = u64_to_user_ptr(attr->query.prog_ids); 716 enum bpf_attach_type type = attr->query.attach_type; 717 struct list_head *progs = &cgrp->bpf.progs[type]; 718 u32 flags = cgrp->bpf.flags[type]; 719 struct bpf_prog_array *effective; 720 struct bpf_prog *prog; 721 int cnt, ret = 0, i; 722 723 effective = rcu_dereference_protected(cgrp->bpf.effective[type], 724 lockdep_is_held(&cgroup_mutex)); 725 726 if (attr->query.query_flags & BPF_F_QUERY_EFFECTIVE) 727 cnt = bpf_prog_array_length(effective); 728 else 729 cnt = prog_list_length(progs); 730 731 if (copy_to_user(&uattr->query.attach_flags, &flags, sizeof(flags))) 732 return -EFAULT; 733 if (copy_to_user(&uattr->query.prog_cnt, &cnt, sizeof(cnt))) 734 return -EFAULT; 735 if (attr->query.prog_cnt == 0 || !prog_ids || !cnt) 736 /* return early if user requested only program count + flags */ 737 return 0; 738 if (attr->query.prog_cnt < cnt) { 739 cnt = attr->query.prog_cnt; 740 ret = -ENOSPC; 741 } 742 743 if (attr->query.query_flags & BPF_F_QUERY_EFFECTIVE) { 744 return bpf_prog_array_copy_to_user(effective, prog_ids, cnt); 745 } else { 746 struct bpf_prog_list *pl; 747 u32 id; 748 749 i = 0; 750 list_for_each_entry(pl, progs, node) { 751 prog = prog_list_prog(pl); 752 id = prog->aux->id; 753 if (copy_to_user(prog_ids + i, &id, sizeof(id))) 754 return -EFAULT; 755 if (++i == cnt) 756 break; 757 } 758 } 759 return ret; 760 } 761 762 int cgroup_bpf_prog_attach(const union bpf_attr *attr, 763 enum bpf_prog_type ptype, struct bpf_prog *prog) 764 { 765 struct bpf_prog *replace_prog = NULL; 766 struct cgroup *cgrp; 767 int ret; 768 769 cgrp = cgroup_get_from_fd(attr->target_fd); 770 if (IS_ERR(cgrp)) 771 return PTR_ERR(cgrp); 772 773 if ((attr->attach_flags & BPF_F_ALLOW_MULTI) && 774 (attr->attach_flags & BPF_F_REPLACE)) { 775 replace_prog = bpf_prog_get_type(attr->replace_bpf_fd, ptype); 776 if (IS_ERR(replace_prog)) { 777 cgroup_put(cgrp); 778 return PTR_ERR(replace_prog); 779 } 780 } 781 782 ret = cgroup_bpf_attach(cgrp, prog, replace_prog, NULL, 783 attr->attach_type, attr->attach_flags); 784 785 if (replace_prog) 786 bpf_prog_put(replace_prog); 787 cgroup_put(cgrp); 788 return ret; 789 } 790 791 int cgroup_bpf_prog_detach(const union bpf_attr *attr, enum bpf_prog_type ptype) 792 { 793 struct bpf_prog *prog; 794 struct cgroup *cgrp; 795 int ret; 796 797 cgrp = cgroup_get_from_fd(attr->target_fd); 798 if (IS_ERR(cgrp)) 799 return PTR_ERR(cgrp); 800 801 prog = bpf_prog_get_type(attr->attach_bpf_fd, ptype); 802 if (IS_ERR(prog)) 803 prog = NULL; 804 805 ret = cgroup_bpf_detach(cgrp, prog, attr->attach_type); 806 if (prog) 807 bpf_prog_put(prog); 808 809 cgroup_put(cgrp); 810 return ret; 811 } 812 813 static void bpf_cgroup_link_release(struct bpf_link *link) 814 { 815 struct bpf_cgroup_link *cg_link = 816 container_of(link, struct bpf_cgroup_link, link); 817 struct cgroup *cg; 818 819 /* link might have been auto-detached by dying cgroup already, 820 * in that case our work is done here 821 */ 822 if (!cg_link->cgroup) 823 return; 824 825 mutex_lock(&cgroup_mutex); 826 827 /* re-check cgroup under lock again */ 828 if (!cg_link->cgroup) { 829 mutex_unlock(&cgroup_mutex); 830 return; 831 } 832 833 WARN_ON(__cgroup_bpf_detach(cg_link->cgroup, NULL, cg_link, 834 cg_link->type)); 835 836 cg = cg_link->cgroup; 837 cg_link->cgroup = NULL; 838 839 mutex_unlock(&cgroup_mutex); 840 841 cgroup_put(cg); 842 } 843 844 static void bpf_cgroup_link_dealloc(struct bpf_link *link) 845 { 846 struct bpf_cgroup_link *cg_link = 847 container_of(link, struct bpf_cgroup_link, link); 848 849 kfree(cg_link); 850 } 851 852 static int bpf_cgroup_link_detach(struct bpf_link *link) 853 { 854 bpf_cgroup_link_release(link); 855 856 return 0; 857 } 858 859 static void bpf_cgroup_link_show_fdinfo(const struct bpf_link *link, 860 struct seq_file *seq) 861 { 862 struct bpf_cgroup_link *cg_link = 863 container_of(link, struct bpf_cgroup_link, link); 864 u64 cg_id = 0; 865 866 mutex_lock(&cgroup_mutex); 867 if (cg_link->cgroup) 868 cg_id = cgroup_id(cg_link->cgroup); 869 mutex_unlock(&cgroup_mutex); 870 871 seq_printf(seq, 872 "cgroup_id:\t%llu\n" 873 "attach_type:\t%d\n", 874 cg_id, 875 cg_link->type); 876 } 877 878 static int bpf_cgroup_link_fill_link_info(const struct bpf_link *link, 879 struct bpf_link_info *info) 880 { 881 struct bpf_cgroup_link *cg_link = 882 container_of(link, struct bpf_cgroup_link, link); 883 u64 cg_id = 0; 884 885 mutex_lock(&cgroup_mutex); 886 if (cg_link->cgroup) 887 cg_id = cgroup_id(cg_link->cgroup); 888 mutex_unlock(&cgroup_mutex); 889 890 info->cgroup.cgroup_id = cg_id; 891 info->cgroup.attach_type = cg_link->type; 892 return 0; 893 } 894 895 static const struct bpf_link_ops bpf_cgroup_link_lops = { 896 .release = bpf_cgroup_link_release, 897 .dealloc = bpf_cgroup_link_dealloc, 898 .detach = bpf_cgroup_link_detach, 899 .update_prog = cgroup_bpf_replace, 900 .show_fdinfo = bpf_cgroup_link_show_fdinfo, 901 .fill_link_info = bpf_cgroup_link_fill_link_info, 902 }; 903 904 int cgroup_bpf_link_attach(const union bpf_attr *attr, struct bpf_prog *prog) 905 { 906 struct bpf_link_primer link_primer; 907 struct bpf_cgroup_link *link; 908 struct cgroup *cgrp; 909 int err; 910 911 if (attr->link_create.flags) 912 return -EINVAL; 913 914 cgrp = cgroup_get_from_fd(attr->link_create.target_fd); 915 if (IS_ERR(cgrp)) 916 return PTR_ERR(cgrp); 917 918 link = kzalloc(sizeof(*link), GFP_USER); 919 if (!link) { 920 err = -ENOMEM; 921 goto out_put_cgroup; 922 } 923 bpf_link_init(&link->link, BPF_LINK_TYPE_CGROUP, &bpf_cgroup_link_lops, 924 prog); 925 link->cgroup = cgrp; 926 link->type = attr->link_create.attach_type; 927 928 err = bpf_link_prime(&link->link, &link_primer); 929 if (err) { 930 kfree(link); 931 goto out_put_cgroup; 932 } 933 934 err = cgroup_bpf_attach(cgrp, NULL, NULL, link, link->type, 935 BPF_F_ALLOW_MULTI); 936 if (err) { 937 bpf_link_cleanup(&link_primer); 938 goto out_put_cgroup; 939 } 940 941 return bpf_link_settle(&link_primer); 942 943 out_put_cgroup: 944 cgroup_put(cgrp); 945 return err; 946 } 947 948 int cgroup_bpf_prog_query(const union bpf_attr *attr, 949 union bpf_attr __user *uattr) 950 { 951 struct cgroup *cgrp; 952 int ret; 953 954 cgrp = cgroup_get_from_fd(attr->query.target_fd); 955 if (IS_ERR(cgrp)) 956 return PTR_ERR(cgrp); 957 958 ret = cgroup_bpf_query(cgrp, attr, uattr); 959 960 cgroup_put(cgrp); 961 return ret; 962 } 963 964 /** 965 * __cgroup_bpf_run_filter_skb() - Run a program for packet filtering 966 * @sk: The socket sending or receiving traffic 967 * @skb: The skb that is being sent or received 968 * @type: The type of program to be exectuted 969 * 970 * If no socket is passed, or the socket is not of type INET or INET6, 971 * this function does nothing and returns 0. 972 * 973 * The program type passed in via @type must be suitable for network 974 * filtering. No further check is performed to assert that. 975 * 976 * For egress packets, this function can return: 977 * NET_XMIT_SUCCESS (0) - continue with packet output 978 * NET_XMIT_DROP (1) - drop packet and notify TCP to call cwr 979 * NET_XMIT_CN (2) - continue with packet output and notify TCP 980 * to call cwr 981 * -EPERM - drop packet 982 * 983 * For ingress packets, this function will return -EPERM if any 984 * attached program was found and if it returned != 1 during execution. 985 * Otherwise 0 is returned. 986 */ 987 int __cgroup_bpf_run_filter_skb(struct sock *sk, 988 struct sk_buff *skb, 989 enum bpf_attach_type type) 990 { 991 unsigned int offset = skb->data - skb_network_header(skb); 992 struct sock *save_sk; 993 void *saved_data_end; 994 struct cgroup *cgrp; 995 int ret; 996 997 if (!sk || !sk_fullsock(sk)) 998 return 0; 999 1000 if (sk->sk_family != AF_INET && sk->sk_family != AF_INET6) 1001 return 0; 1002 1003 cgrp = sock_cgroup_ptr(&sk->sk_cgrp_data); 1004 save_sk = skb->sk; 1005 skb->sk = sk; 1006 __skb_push(skb, offset); 1007 1008 /* compute pointers for the bpf prog */ 1009 bpf_compute_and_save_data_end(skb, &saved_data_end); 1010 1011 if (type == BPF_CGROUP_INET_EGRESS) { 1012 ret = BPF_PROG_CGROUP_INET_EGRESS_RUN_ARRAY( 1013 cgrp->bpf.effective[type], skb, __bpf_prog_run_save_cb); 1014 } else { 1015 ret = BPF_PROG_RUN_ARRAY(cgrp->bpf.effective[type], skb, 1016 __bpf_prog_run_save_cb); 1017 ret = (ret == 1 ? 0 : -EPERM); 1018 } 1019 bpf_restore_data_end(skb, saved_data_end); 1020 __skb_pull(skb, offset); 1021 skb->sk = save_sk; 1022 1023 return ret; 1024 } 1025 EXPORT_SYMBOL(__cgroup_bpf_run_filter_skb); 1026 1027 /** 1028 * __cgroup_bpf_run_filter_sk() - Run a program on a sock 1029 * @sk: sock structure to manipulate 1030 * @type: The type of program to be exectuted 1031 * 1032 * socket is passed is expected to be of type INET or INET6. 1033 * 1034 * The program type passed in via @type must be suitable for sock 1035 * filtering. No further check is performed to assert that. 1036 * 1037 * This function will return %-EPERM if any if an attached program was found 1038 * and if it returned != 1 during execution. In all other cases, 0 is returned. 1039 */ 1040 int __cgroup_bpf_run_filter_sk(struct sock *sk, 1041 enum bpf_attach_type type) 1042 { 1043 struct cgroup *cgrp = sock_cgroup_ptr(&sk->sk_cgrp_data); 1044 int ret; 1045 1046 ret = BPF_PROG_RUN_ARRAY(cgrp->bpf.effective[type], sk, BPF_PROG_RUN); 1047 return ret == 1 ? 0 : -EPERM; 1048 } 1049 EXPORT_SYMBOL(__cgroup_bpf_run_filter_sk); 1050 1051 /** 1052 * __cgroup_bpf_run_filter_sock_addr() - Run a program on a sock and 1053 * provided by user sockaddr 1054 * @sk: sock struct that will use sockaddr 1055 * @uaddr: sockaddr struct provided by user 1056 * @type: The type of program to be exectuted 1057 * @t_ctx: Pointer to attach type specific context 1058 * 1059 * socket is expected to be of type INET or INET6. 1060 * 1061 * This function will return %-EPERM if an attached program is found and 1062 * returned value != 1 during execution. In all other cases, 0 is returned. 1063 */ 1064 int __cgroup_bpf_run_filter_sock_addr(struct sock *sk, 1065 struct sockaddr *uaddr, 1066 enum bpf_attach_type type, 1067 void *t_ctx) 1068 { 1069 struct bpf_sock_addr_kern ctx = { 1070 .sk = sk, 1071 .uaddr = uaddr, 1072 .t_ctx = t_ctx, 1073 }; 1074 struct sockaddr_storage unspec; 1075 struct cgroup *cgrp; 1076 int ret; 1077 1078 /* Check socket family since not all sockets represent network 1079 * endpoint (e.g. AF_UNIX). 1080 */ 1081 if (sk->sk_family != AF_INET && sk->sk_family != AF_INET6) 1082 return 0; 1083 1084 if (!ctx.uaddr) { 1085 memset(&unspec, 0, sizeof(unspec)); 1086 ctx.uaddr = (struct sockaddr *)&unspec; 1087 } 1088 1089 cgrp = sock_cgroup_ptr(&sk->sk_cgrp_data); 1090 ret = BPF_PROG_RUN_ARRAY(cgrp->bpf.effective[type], &ctx, BPF_PROG_RUN); 1091 1092 return ret == 1 ? 0 : -EPERM; 1093 } 1094 EXPORT_SYMBOL(__cgroup_bpf_run_filter_sock_addr); 1095 1096 /** 1097 * __cgroup_bpf_run_filter_sock_ops() - Run a program on a sock 1098 * @sk: socket to get cgroup from 1099 * @sock_ops: bpf_sock_ops_kern struct to pass to program. Contains 1100 * sk with connection information (IP addresses, etc.) May not contain 1101 * cgroup info if it is a req sock. 1102 * @type: The type of program to be exectuted 1103 * 1104 * socket passed is expected to be of type INET or INET6. 1105 * 1106 * The program type passed in via @type must be suitable for sock_ops 1107 * filtering. No further check is performed to assert that. 1108 * 1109 * This function will return %-EPERM if any if an attached program was found 1110 * and if it returned != 1 during execution. In all other cases, 0 is returned. 1111 */ 1112 int __cgroup_bpf_run_filter_sock_ops(struct sock *sk, 1113 struct bpf_sock_ops_kern *sock_ops, 1114 enum bpf_attach_type type) 1115 { 1116 struct cgroup *cgrp = sock_cgroup_ptr(&sk->sk_cgrp_data); 1117 int ret; 1118 1119 ret = BPF_PROG_RUN_ARRAY(cgrp->bpf.effective[type], sock_ops, 1120 BPF_PROG_RUN); 1121 return ret == 1 ? 0 : -EPERM; 1122 } 1123 EXPORT_SYMBOL(__cgroup_bpf_run_filter_sock_ops); 1124 1125 int __cgroup_bpf_check_dev_permission(short dev_type, u32 major, u32 minor, 1126 short access, enum bpf_attach_type type) 1127 { 1128 struct cgroup *cgrp; 1129 struct bpf_cgroup_dev_ctx ctx = { 1130 .access_type = (access << 16) | dev_type, 1131 .major = major, 1132 .minor = minor, 1133 }; 1134 int allow = 1; 1135 1136 rcu_read_lock(); 1137 cgrp = task_dfl_cgroup(current); 1138 allow = BPF_PROG_RUN_ARRAY(cgrp->bpf.effective[type], &ctx, 1139 BPF_PROG_RUN); 1140 rcu_read_unlock(); 1141 1142 return !allow; 1143 } 1144 1145 static const struct bpf_func_proto * 1146 cgroup_base_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog) 1147 { 1148 switch (func_id) { 1149 case BPF_FUNC_get_current_uid_gid: 1150 return &bpf_get_current_uid_gid_proto; 1151 case BPF_FUNC_get_local_storage: 1152 return &bpf_get_local_storage_proto; 1153 case BPF_FUNC_get_current_cgroup_id: 1154 return &bpf_get_current_cgroup_id_proto; 1155 case BPF_FUNC_perf_event_output: 1156 return &bpf_event_output_data_proto; 1157 default: 1158 return bpf_base_func_proto(func_id); 1159 } 1160 } 1161 1162 static const struct bpf_func_proto * 1163 cgroup_dev_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog) 1164 { 1165 return cgroup_base_func_proto(func_id, prog); 1166 } 1167 1168 static bool cgroup_dev_is_valid_access(int off, int size, 1169 enum bpf_access_type type, 1170 const struct bpf_prog *prog, 1171 struct bpf_insn_access_aux *info) 1172 { 1173 const int size_default = sizeof(__u32); 1174 1175 if (type == BPF_WRITE) 1176 return false; 1177 1178 if (off < 0 || off + size > sizeof(struct bpf_cgroup_dev_ctx)) 1179 return false; 1180 /* The verifier guarantees that size > 0. */ 1181 if (off % size != 0) 1182 return false; 1183 1184 switch (off) { 1185 case bpf_ctx_range(struct bpf_cgroup_dev_ctx, access_type): 1186 bpf_ctx_record_field_size(info, size_default); 1187 if (!bpf_ctx_narrow_access_ok(off, size, size_default)) 1188 return false; 1189 break; 1190 default: 1191 if (size != size_default) 1192 return false; 1193 } 1194 1195 return true; 1196 } 1197 1198 const struct bpf_prog_ops cg_dev_prog_ops = { 1199 }; 1200 1201 const struct bpf_verifier_ops cg_dev_verifier_ops = { 1202 .get_func_proto = cgroup_dev_func_proto, 1203 .is_valid_access = cgroup_dev_is_valid_access, 1204 }; 1205 1206 /** 1207 * __cgroup_bpf_run_filter_sysctl - Run a program on sysctl 1208 * 1209 * @head: sysctl table header 1210 * @table: sysctl table 1211 * @write: sysctl is being read (= 0) or written (= 1) 1212 * @buf: pointer to buffer (in and out) 1213 * @pcount: value-result argument: value is size of buffer pointed to by @buf, 1214 * result is size of @new_buf if program set new value, initial value 1215 * otherwise 1216 * @ppos: value-result argument: value is position at which read from or write 1217 * to sysctl is happening, result is new position if program overrode it, 1218 * initial value otherwise 1219 * @type: type of program to be executed 1220 * 1221 * Program is run when sysctl is being accessed, either read or written, and 1222 * can allow or deny such access. 1223 * 1224 * This function will return %-EPERM if an attached program is found and 1225 * returned value != 1 during execution. In all other cases 0 is returned. 1226 */ 1227 int __cgroup_bpf_run_filter_sysctl(struct ctl_table_header *head, 1228 struct ctl_table *table, int write, 1229 char **buf, size_t *pcount, loff_t *ppos, 1230 enum bpf_attach_type type) 1231 { 1232 struct bpf_sysctl_kern ctx = { 1233 .head = head, 1234 .table = table, 1235 .write = write, 1236 .ppos = ppos, 1237 .cur_val = NULL, 1238 .cur_len = PAGE_SIZE, 1239 .new_val = NULL, 1240 .new_len = 0, 1241 .new_updated = 0, 1242 }; 1243 struct cgroup *cgrp; 1244 loff_t pos = 0; 1245 int ret; 1246 1247 ctx.cur_val = kmalloc_track_caller(ctx.cur_len, GFP_KERNEL); 1248 if (!ctx.cur_val || 1249 table->proc_handler(table, 0, ctx.cur_val, &ctx.cur_len, &pos)) { 1250 /* Let BPF program decide how to proceed. */ 1251 ctx.cur_len = 0; 1252 } 1253 1254 if (write && *buf && *pcount) { 1255 /* BPF program should be able to override new value with a 1256 * buffer bigger than provided by user. 1257 */ 1258 ctx.new_val = kmalloc_track_caller(PAGE_SIZE, GFP_KERNEL); 1259 ctx.new_len = min_t(size_t, PAGE_SIZE, *pcount); 1260 if (ctx.new_val) { 1261 memcpy(ctx.new_val, *buf, ctx.new_len); 1262 } else { 1263 /* Let BPF program decide how to proceed. */ 1264 ctx.new_len = 0; 1265 } 1266 } 1267 1268 rcu_read_lock(); 1269 cgrp = task_dfl_cgroup(current); 1270 ret = BPF_PROG_RUN_ARRAY(cgrp->bpf.effective[type], &ctx, BPF_PROG_RUN); 1271 rcu_read_unlock(); 1272 1273 kfree(ctx.cur_val); 1274 1275 if (ret == 1 && ctx.new_updated) { 1276 kfree(*buf); 1277 *buf = ctx.new_val; 1278 *pcount = ctx.new_len; 1279 } else { 1280 kfree(ctx.new_val); 1281 } 1282 1283 return ret == 1 ? 0 : -EPERM; 1284 } 1285 1286 #ifdef CONFIG_NET 1287 static bool __cgroup_bpf_prog_array_is_empty(struct cgroup *cgrp, 1288 enum bpf_attach_type attach_type) 1289 { 1290 struct bpf_prog_array *prog_array; 1291 bool empty; 1292 1293 rcu_read_lock(); 1294 prog_array = rcu_dereference(cgrp->bpf.effective[attach_type]); 1295 empty = bpf_prog_array_is_empty(prog_array); 1296 rcu_read_unlock(); 1297 1298 return empty; 1299 } 1300 1301 static int sockopt_alloc_buf(struct bpf_sockopt_kern *ctx, int max_optlen, 1302 struct bpf_sockopt_buf *buf) 1303 { 1304 if (unlikely(max_optlen < 0)) 1305 return -EINVAL; 1306 1307 if (unlikely(max_optlen > PAGE_SIZE)) { 1308 /* We don't expose optvals that are greater than PAGE_SIZE 1309 * to the BPF program. 1310 */ 1311 max_optlen = PAGE_SIZE; 1312 } 1313 1314 if (max_optlen <= sizeof(buf->data)) { 1315 /* When the optval fits into BPF_SOCKOPT_KERN_BUF_SIZE 1316 * bytes avoid the cost of kzalloc. 1317 */ 1318 ctx->optval = buf->data; 1319 ctx->optval_end = ctx->optval + max_optlen; 1320 return max_optlen; 1321 } 1322 1323 ctx->optval = kzalloc(max_optlen, GFP_USER); 1324 if (!ctx->optval) 1325 return -ENOMEM; 1326 1327 ctx->optval_end = ctx->optval + max_optlen; 1328 1329 return max_optlen; 1330 } 1331 1332 static void sockopt_free_buf(struct bpf_sockopt_kern *ctx, 1333 struct bpf_sockopt_buf *buf) 1334 { 1335 if (ctx->optval == buf->data) 1336 return; 1337 kfree(ctx->optval); 1338 } 1339 1340 static bool sockopt_buf_allocated(struct bpf_sockopt_kern *ctx, 1341 struct bpf_sockopt_buf *buf) 1342 { 1343 return ctx->optval != buf->data; 1344 } 1345 1346 int __cgroup_bpf_run_filter_setsockopt(struct sock *sk, int *level, 1347 int *optname, char __user *optval, 1348 int *optlen, char **kernel_optval) 1349 { 1350 struct cgroup *cgrp = sock_cgroup_ptr(&sk->sk_cgrp_data); 1351 struct bpf_sockopt_buf buf = {}; 1352 struct bpf_sockopt_kern ctx = { 1353 .sk = sk, 1354 .level = *level, 1355 .optname = *optname, 1356 }; 1357 int ret, max_optlen; 1358 1359 /* Opportunistic check to see whether we have any BPF program 1360 * attached to the hook so we don't waste time allocating 1361 * memory and locking the socket. 1362 */ 1363 if (__cgroup_bpf_prog_array_is_empty(cgrp, BPF_CGROUP_SETSOCKOPT)) 1364 return 0; 1365 1366 /* Allocate a bit more than the initial user buffer for 1367 * BPF program. The canonical use case is overriding 1368 * TCP_CONGESTION(nv) to TCP_CONGESTION(cubic). 1369 */ 1370 max_optlen = max_t(int, 16, *optlen); 1371 1372 max_optlen = sockopt_alloc_buf(&ctx, max_optlen, &buf); 1373 if (max_optlen < 0) 1374 return max_optlen; 1375 1376 ctx.optlen = *optlen; 1377 1378 if (copy_from_user(ctx.optval, optval, min(*optlen, max_optlen)) != 0) { 1379 ret = -EFAULT; 1380 goto out; 1381 } 1382 1383 lock_sock(sk); 1384 ret = BPF_PROG_RUN_ARRAY(cgrp->bpf.effective[BPF_CGROUP_SETSOCKOPT], 1385 &ctx, BPF_PROG_RUN); 1386 release_sock(sk); 1387 1388 if (!ret) { 1389 ret = -EPERM; 1390 goto out; 1391 } 1392 1393 if (ctx.optlen == -1) { 1394 /* optlen set to -1, bypass kernel */ 1395 ret = 1; 1396 } else if (ctx.optlen > max_optlen || ctx.optlen < -1) { 1397 /* optlen is out of bounds */ 1398 ret = -EFAULT; 1399 } else { 1400 /* optlen within bounds, run kernel handler */ 1401 ret = 0; 1402 1403 /* export any potential modifications */ 1404 *level = ctx.level; 1405 *optname = ctx.optname; 1406 1407 /* optlen == 0 from BPF indicates that we should 1408 * use original userspace data. 1409 */ 1410 if (ctx.optlen != 0) { 1411 *optlen = ctx.optlen; 1412 /* We've used bpf_sockopt_kern->buf as an intermediary 1413 * storage, but the BPF program indicates that we need 1414 * to pass this data to the kernel setsockopt handler. 1415 * No way to export on-stack buf, have to allocate a 1416 * new buffer. 1417 */ 1418 if (!sockopt_buf_allocated(&ctx, &buf)) { 1419 void *p = kmalloc(ctx.optlen, GFP_USER); 1420 1421 if (!p) { 1422 ret = -ENOMEM; 1423 goto out; 1424 } 1425 memcpy(p, ctx.optval, ctx.optlen); 1426 *kernel_optval = p; 1427 } else { 1428 *kernel_optval = ctx.optval; 1429 } 1430 /* export and don't free sockopt buf */ 1431 return 0; 1432 } 1433 } 1434 1435 out: 1436 sockopt_free_buf(&ctx, &buf); 1437 return ret; 1438 } 1439 1440 int __cgroup_bpf_run_filter_getsockopt(struct sock *sk, int level, 1441 int optname, char __user *optval, 1442 int __user *optlen, int max_optlen, 1443 int retval) 1444 { 1445 struct cgroup *cgrp = sock_cgroup_ptr(&sk->sk_cgrp_data); 1446 struct bpf_sockopt_buf buf = {}; 1447 struct bpf_sockopt_kern ctx = { 1448 .sk = sk, 1449 .level = level, 1450 .optname = optname, 1451 .retval = retval, 1452 }; 1453 int ret; 1454 1455 /* Opportunistic check to see whether we have any BPF program 1456 * attached to the hook so we don't waste time allocating 1457 * memory and locking the socket. 1458 */ 1459 if (__cgroup_bpf_prog_array_is_empty(cgrp, BPF_CGROUP_GETSOCKOPT)) 1460 return retval; 1461 1462 ctx.optlen = max_optlen; 1463 1464 max_optlen = sockopt_alloc_buf(&ctx, max_optlen, &buf); 1465 if (max_optlen < 0) 1466 return max_optlen; 1467 1468 if (!retval) { 1469 /* If kernel getsockopt finished successfully, 1470 * copy whatever was returned to the user back 1471 * into our temporary buffer. Set optlen to the 1472 * one that kernel returned as well to let 1473 * BPF programs inspect the value. 1474 */ 1475 1476 if (get_user(ctx.optlen, optlen)) { 1477 ret = -EFAULT; 1478 goto out; 1479 } 1480 1481 if (copy_from_user(ctx.optval, optval, 1482 min(ctx.optlen, max_optlen)) != 0) { 1483 ret = -EFAULT; 1484 goto out; 1485 } 1486 } 1487 1488 lock_sock(sk); 1489 ret = BPF_PROG_RUN_ARRAY(cgrp->bpf.effective[BPF_CGROUP_GETSOCKOPT], 1490 &ctx, BPF_PROG_RUN); 1491 release_sock(sk); 1492 1493 if (!ret) { 1494 ret = -EPERM; 1495 goto out; 1496 } 1497 1498 if (ctx.optlen > max_optlen) { 1499 ret = -EFAULT; 1500 goto out; 1501 } 1502 1503 /* BPF programs only allowed to set retval to 0, not some 1504 * arbitrary value. 1505 */ 1506 if (ctx.retval != 0 && ctx.retval != retval) { 1507 ret = -EFAULT; 1508 goto out; 1509 } 1510 1511 if (ctx.optlen != 0) { 1512 if (copy_to_user(optval, ctx.optval, ctx.optlen) || 1513 put_user(ctx.optlen, optlen)) { 1514 ret = -EFAULT; 1515 goto out; 1516 } 1517 } 1518 1519 ret = ctx.retval; 1520 1521 out: 1522 sockopt_free_buf(&ctx, &buf); 1523 return ret; 1524 } 1525 1526 int __cgroup_bpf_run_filter_getsockopt_kern(struct sock *sk, int level, 1527 int optname, void *optval, 1528 int *optlen, int retval) 1529 { 1530 struct cgroup *cgrp = sock_cgroup_ptr(&sk->sk_cgrp_data); 1531 struct bpf_sockopt_kern ctx = { 1532 .sk = sk, 1533 .level = level, 1534 .optname = optname, 1535 .retval = retval, 1536 .optlen = *optlen, 1537 .optval = optval, 1538 .optval_end = optval + *optlen, 1539 }; 1540 int ret; 1541 1542 /* Note that __cgroup_bpf_run_filter_getsockopt doesn't copy 1543 * user data back into BPF buffer when reval != 0. This is 1544 * done as an optimization to avoid extra copy, assuming 1545 * kernel won't populate the data in case of an error. 1546 * Here we always pass the data and memset() should 1547 * be called if that data shouldn't be "exported". 1548 */ 1549 1550 ret = BPF_PROG_RUN_ARRAY(cgrp->bpf.effective[BPF_CGROUP_GETSOCKOPT], 1551 &ctx, BPF_PROG_RUN); 1552 if (!ret) 1553 return -EPERM; 1554 1555 if (ctx.optlen > *optlen) 1556 return -EFAULT; 1557 1558 /* BPF programs only allowed to set retval to 0, not some 1559 * arbitrary value. 1560 */ 1561 if (ctx.retval != 0 && ctx.retval != retval) 1562 return -EFAULT; 1563 1564 /* BPF programs can shrink the buffer, export the modifications. 1565 */ 1566 if (ctx.optlen != 0) 1567 *optlen = ctx.optlen; 1568 1569 return ctx.retval; 1570 } 1571 #endif 1572 1573 static ssize_t sysctl_cpy_dir(const struct ctl_dir *dir, char **bufp, 1574 size_t *lenp) 1575 { 1576 ssize_t tmp_ret = 0, ret; 1577 1578 if (dir->header.parent) { 1579 tmp_ret = sysctl_cpy_dir(dir->header.parent, bufp, lenp); 1580 if (tmp_ret < 0) 1581 return tmp_ret; 1582 } 1583 1584 ret = strscpy(*bufp, dir->header.ctl_table[0].procname, *lenp); 1585 if (ret < 0) 1586 return ret; 1587 *bufp += ret; 1588 *lenp -= ret; 1589 ret += tmp_ret; 1590 1591 /* Avoid leading slash. */ 1592 if (!ret) 1593 return ret; 1594 1595 tmp_ret = strscpy(*bufp, "/", *lenp); 1596 if (tmp_ret < 0) 1597 return tmp_ret; 1598 *bufp += tmp_ret; 1599 *lenp -= tmp_ret; 1600 1601 return ret + tmp_ret; 1602 } 1603 1604 BPF_CALL_4(bpf_sysctl_get_name, struct bpf_sysctl_kern *, ctx, char *, buf, 1605 size_t, buf_len, u64, flags) 1606 { 1607 ssize_t tmp_ret = 0, ret; 1608 1609 if (!buf) 1610 return -EINVAL; 1611 1612 if (!(flags & BPF_F_SYSCTL_BASE_NAME)) { 1613 if (!ctx->head) 1614 return -EINVAL; 1615 tmp_ret = sysctl_cpy_dir(ctx->head->parent, &buf, &buf_len); 1616 if (tmp_ret < 0) 1617 return tmp_ret; 1618 } 1619 1620 ret = strscpy(buf, ctx->table->procname, buf_len); 1621 1622 return ret < 0 ? ret : tmp_ret + ret; 1623 } 1624 1625 static const struct bpf_func_proto bpf_sysctl_get_name_proto = { 1626 .func = bpf_sysctl_get_name, 1627 .gpl_only = false, 1628 .ret_type = RET_INTEGER, 1629 .arg1_type = ARG_PTR_TO_CTX, 1630 .arg2_type = ARG_PTR_TO_MEM, 1631 .arg3_type = ARG_CONST_SIZE, 1632 .arg4_type = ARG_ANYTHING, 1633 }; 1634 1635 static int copy_sysctl_value(char *dst, size_t dst_len, char *src, 1636 size_t src_len) 1637 { 1638 if (!dst) 1639 return -EINVAL; 1640 1641 if (!dst_len) 1642 return -E2BIG; 1643 1644 if (!src || !src_len) { 1645 memset(dst, 0, dst_len); 1646 return -EINVAL; 1647 } 1648 1649 memcpy(dst, src, min(dst_len, src_len)); 1650 1651 if (dst_len > src_len) { 1652 memset(dst + src_len, '\0', dst_len - src_len); 1653 return src_len; 1654 } 1655 1656 dst[dst_len - 1] = '\0'; 1657 1658 return -E2BIG; 1659 } 1660 1661 BPF_CALL_3(bpf_sysctl_get_current_value, struct bpf_sysctl_kern *, ctx, 1662 char *, buf, size_t, buf_len) 1663 { 1664 return copy_sysctl_value(buf, buf_len, ctx->cur_val, ctx->cur_len); 1665 } 1666 1667 static const struct bpf_func_proto bpf_sysctl_get_current_value_proto = { 1668 .func = bpf_sysctl_get_current_value, 1669 .gpl_only = false, 1670 .ret_type = RET_INTEGER, 1671 .arg1_type = ARG_PTR_TO_CTX, 1672 .arg2_type = ARG_PTR_TO_UNINIT_MEM, 1673 .arg3_type = ARG_CONST_SIZE, 1674 }; 1675 1676 BPF_CALL_3(bpf_sysctl_get_new_value, struct bpf_sysctl_kern *, ctx, char *, buf, 1677 size_t, buf_len) 1678 { 1679 if (!ctx->write) { 1680 if (buf && buf_len) 1681 memset(buf, '\0', buf_len); 1682 return -EINVAL; 1683 } 1684 return copy_sysctl_value(buf, buf_len, ctx->new_val, ctx->new_len); 1685 } 1686 1687 static const struct bpf_func_proto bpf_sysctl_get_new_value_proto = { 1688 .func = bpf_sysctl_get_new_value, 1689 .gpl_only = false, 1690 .ret_type = RET_INTEGER, 1691 .arg1_type = ARG_PTR_TO_CTX, 1692 .arg2_type = ARG_PTR_TO_UNINIT_MEM, 1693 .arg3_type = ARG_CONST_SIZE, 1694 }; 1695 1696 BPF_CALL_3(bpf_sysctl_set_new_value, struct bpf_sysctl_kern *, ctx, 1697 const char *, buf, size_t, buf_len) 1698 { 1699 if (!ctx->write || !ctx->new_val || !ctx->new_len || !buf || !buf_len) 1700 return -EINVAL; 1701 1702 if (buf_len > PAGE_SIZE - 1) 1703 return -E2BIG; 1704 1705 memcpy(ctx->new_val, buf, buf_len); 1706 ctx->new_len = buf_len; 1707 ctx->new_updated = 1; 1708 1709 return 0; 1710 } 1711 1712 static const struct bpf_func_proto bpf_sysctl_set_new_value_proto = { 1713 .func = bpf_sysctl_set_new_value, 1714 .gpl_only = false, 1715 .ret_type = RET_INTEGER, 1716 .arg1_type = ARG_PTR_TO_CTX, 1717 .arg2_type = ARG_PTR_TO_MEM, 1718 .arg3_type = ARG_CONST_SIZE, 1719 }; 1720 1721 static const struct bpf_func_proto * 1722 sysctl_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog) 1723 { 1724 switch (func_id) { 1725 case BPF_FUNC_strtol: 1726 return &bpf_strtol_proto; 1727 case BPF_FUNC_strtoul: 1728 return &bpf_strtoul_proto; 1729 case BPF_FUNC_sysctl_get_name: 1730 return &bpf_sysctl_get_name_proto; 1731 case BPF_FUNC_sysctl_get_current_value: 1732 return &bpf_sysctl_get_current_value_proto; 1733 case BPF_FUNC_sysctl_get_new_value: 1734 return &bpf_sysctl_get_new_value_proto; 1735 case BPF_FUNC_sysctl_set_new_value: 1736 return &bpf_sysctl_set_new_value_proto; 1737 default: 1738 return cgroup_base_func_proto(func_id, prog); 1739 } 1740 } 1741 1742 static bool sysctl_is_valid_access(int off, int size, enum bpf_access_type type, 1743 const struct bpf_prog *prog, 1744 struct bpf_insn_access_aux *info) 1745 { 1746 const int size_default = sizeof(__u32); 1747 1748 if (off < 0 || off + size > sizeof(struct bpf_sysctl) || off % size) 1749 return false; 1750 1751 switch (off) { 1752 case bpf_ctx_range(struct bpf_sysctl, write): 1753 if (type != BPF_READ) 1754 return false; 1755 bpf_ctx_record_field_size(info, size_default); 1756 return bpf_ctx_narrow_access_ok(off, size, size_default); 1757 case bpf_ctx_range(struct bpf_sysctl, file_pos): 1758 if (type == BPF_READ) { 1759 bpf_ctx_record_field_size(info, size_default); 1760 return bpf_ctx_narrow_access_ok(off, size, size_default); 1761 } else { 1762 return size == size_default; 1763 } 1764 default: 1765 return false; 1766 } 1767 } 1768 1769 static u32 sysctl_convert_ctx_access(enum bpf_access_type type, 1770 const struct bpf_insn *si, 1771 struct bpf_insn *insn_buf, 1772 struct bpf_prog *prog, u32 *target_size) 1773 { 1774 struct bpf_insn *insn = insn_buf; 1775 u32 read_size; 1776 1777 switch (si->off) { 1778 case offsetof(struct bpf_sysctl, write): 1779 *insn++ = BPF_LDX_MEM( 1780 BPF_SIZE(si->code), si->dst_reg, si->src_reg, 1781 bpf_target_off(struct bpf_sysctl_kern, write, 1782 sizeof_field(struct bpf_sysctl_kern, 1783 write), 1784 target_size)); 1785 break; 1786 case offsetof(struct bpf_sysctl, file_pos): 1787 /* ppos is a pointer so it should be accessed via indirect 1788 * loads and stores. Also for stores additional temporary 1789 * register is used since neither src_reg nor dst_reg can be 1790 * overridden. 1791 */ 1792 if (type == BPF_WRITE) { 1793 int treg = BPF_REG_9; 1794 1795 if (si->src_reg == treg || si->dst_reg == treg) 1796 --treg; 1797 if (si->src_reg == treg || si->dst_reg == treg) 1798 --treg; 1799 *insn++ = BPF_STX_MEM( 1800 BPF_DW, si->dst_reg, treg, 1801 offsetof(struct bpf_sysctl_kern, tmp_reg)); 1802 *insn++ = BPF_LDX_MEM( 1803 BPF_FIELD_SIZEOF(struct bpf_sysctl_kern, ppos), 1804 treg, si->dst_reg, 1805 offsetof(struct bpf_sysctl_kern, ppos)); 1806 *insn++ = BPF_STX_MEM( 1807 BPF_SIZEOF(u32), treg, si->src_reg, 1808 bpf_ctx_narrow_access_offset( 1809 0, sizeof(u32), sizeof(loff_t))); 1810 *insn++ = BPF_LDX_MEM( 1811 BPF_DW, treg, si->dst_reg, 1812 offsetof(struct bpf_sysctl_kern, tmp_reg)); 1813 } else { 1814 *insn++ = BPF_LDX_MEM( 1815 BPF_FIELD_SIZEOF(struct bpf_sysctl_kern, ppos), 1816 si->dst_reg, si->src_reg, 1817 offsetof(struct bpf_sysctl_kern, ppos)); 1818 read_size = bpf_size_to_bytes(BPF_SIZE(si->code)); 1819 *insn++ = BPF_LDX_MEM( 1820 BPF_SIZE(si->code), si->dst_reg, si->dst_reg, 1821 bpf_ctx_narrow_access_offset( 1822 0, read_size, sizeof(loff_t))); 1823 } 1824 *target_size = sizeof(u32); 1825 break; 1826 } 1827 1828 return insn - insn_buf; 1829 } 1830 1831 const struct bpf_verifier_ops cg_sysctl_verifier_ops = { 1832 .get_func_proto = sysctl_func_proto, 1833 .is_valid_access = sysctl_is_valid_access, 1834 .convert_ctx_access = sysctl_convert_ctx_access, 1835 }; 1836 1837 const struct bpf_prog_ops cg_sysctl_prog_ops = { 1838 }; 1839 1840 static const struct bpf_func_proto * 1841 cg_sockopt_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog) 1842 { 1843 switch (func_id) { 1844 #ifdef CONFIG_NET 1845 case BPF_FUNC_sk_storage_get: 1846 return &bpf_sk_storage_get_proto; 1847 case BPF_FUNC_sk_storage_delete: 1848 return &bpf_sk_storage_delete_proto; 1849 #endif 1850 #ifdef CONFIG_INET 1851 case BPF_FUNC_tcp_sock: 1852 return &bpf_tcp_sock_proto; 1853 #endif 1854 default: 1855 return cgroup_base_func_proto(func_id, prog); 1856 } 1857 } 1858 1859 static bool cg_sockopt_is_valid_access(int off, int size, 1860 enum bpf_access_type type, 1861 const struct bpf_prog *prog, 1862 struct bpf_insn_access_aux *info) 1863 { 1864 const int size_default = sizeof(__u32); 1865 1866 if (off < 0 || off >= sizeof(struct bpf_sockopt)) 1867 return false; 1868 1869 if (off % size != 0) 1870 return false; 1871 1872 if (type == BPF_WRITE) { 1873 switch (off) { 1874 case offsetof(struct bpf_sockopt, retval): 1875 if (size != size_default) 1876 return false; 1877 return prog->expected_attach_type == 1878 BPF_CGROUP_GETSOCKOPT; 1879 case offsetof(struct bpf_sockopt, optname): 1880 fallthrough; 1881 case offsetof(struct bpf_sockopt, level): 1882 if (size != size_default) 1883 return false; 1884 return prog->expected_attach_type == 1885 BPF_CGROUP_SETSOCKOPT; 1886 case offsetof(struct bpf_sockopt, optlen): 1887 return size == size_default; 1888 default: 1889 return false; 1890 } 1891 } 1892 1893 switch (off) { 1894 case offsetof(struct bpf_sockopt, sk): 1895 if (size != sizeof(__u64)) 1896 return false; 1897 info->reg_type = PTR_TO_SOCKET; 1898 break; 1899 case offsetof(struct bpf_sockopt, optval): 1900 if (size != sizeof(__u64)) 1901 return false; 1902 info->reg_type = PTR_TO_PACKET; 1903 break; 1904 case offsetof(struct bpf_sockopt, optval_end): 1905 if (size != sizeof(__u64)) 1906 return false; 1907 info->reg_type = PTR_TO_PACKET_END; 1908 break; 1909 case offsetof(struct bpf_sockopt, retval): 1910 if (size != size_default) 1911 return false; 1912 return prog->expected_attach_type == BPF_CGROUP_GETSOCKOPT; 1913 default: 1914 if (size != size_default) 1915 return false; 1916 break; 1917 } 1918 return true; 1919 } 1920 1921 #define CG_SOCKOPT_ACCESS_FIELD(T, F) \ 1922 T(BPF_FIELD_SIZEOF(struct bpf_sockopt_kern, F), \ 1923 si->dst_reg, si->src_reg, \ 1924 offsetof(struct bpf_sockopt_kern, F)) 1925 1926 static u32 cg_sockopt_convert_ctx_access(enum bpf_access_type type, 1927 const struct bpf_insn *si, 1928 struct bpf_insn *insn_buf, 1929 struct bpf_prog *prog, 1930 u32 *target_size) 1931 { 1932 struct bpf_insn *insn = insn_buf; 1933 1934 switch (si->off) { 1935 case offsetof(struct bpf_sockopt, sk): 1936 *insn++ = CG_SOCKOPT_ACCESS_FIELD(BPF_LDX_MEM, sk); 1937 break; 1938 case offsetof(struct bpf_sockopt, level): 1939 if (type == BPF_WRITE) 1940 *insn++ = CG_SOCKOPT_ACCESS_FIELD(BPF_STX_MEM, level); 1941 else 1942 *insn++ = CG_SOCKOPT_ACCESS_FIELD(BPF_LDX_MEM, level); 1943 break; 1944 case offsetof(struct bpf_sockopt, optname): 1945 if (type == BPF_WRITE) 1946 *insn++ = CG_SOCKOPT_ACCESS_FIELD(BPF_STX_MEM, optname); 1947 else 1948 *insn++ = CG_SOCKOPT_ACCESS_FIELD(BPF_LDX_MEM, optname); 1949 break; 1950 case offsetof(struct bpf_sockopt, optlen): 1951 if (type == BPF_WRITE) 1952 *insn++ = CG_SOCKOPT_ACCESS_FIELD(BPF_STX_MEM, optlen); 1953 else 1954 *insn++ = CG_SOCKOPT_ACCESS_FIELD(BPF_LDX_MEM, optlen); 1955 break; 1956 case offsetof(struct bpf_sockopt, retval): 1957 if (type == BPF_WRITE) 1958 *insn++ = CG_SOCKOPT_ACCESS_FIELD(BPF_STX_MEM, retval); 1959 else 1960 *insn++ = CG_SOCKOPT_ACCESS_FIELD(BPF_LDX_MEM, retval); 1961 break; 1962 case offsetof(struct bpf_sockopt, optval): 1963 *insn++ = CG_SOCKOPT_ACCESS_FIELD(BPF_LDX_MEM, optval); 1964 break; 1965 case offsetof(struct bpf_sockopt, optval_end): 1966 *insn++ = CG_SOCKOPT_ACCESS_FIELD(BPF_LDX_MEM, optval_end); 1967 break; 1968 } 1969 1970 return insn - insn_buf; 1971 } 1972 1973 static int cg_sockopt_get_prologue(struct bpf_insn *insn_buf, 1974 bool direct_write, 1975 const struct bpf_prog *prog) 1976 { 1977 /* Nothing to do for sockopt argument. The data is kzalloc'ated. 1978 */ 1979 return 0; 1980 } 1981 1982 const struct bpf_verifier_ops cg_sockopt_verifier_ops = { 1983 .get_func_proto = cg_sockopt_func_proto, 1984 .is_valid_access = cg_sockopt_is_valid_access, 1985 .convert_ctx_access = cg_sockopt_convert_ctx_access, 1986 .gen_prologue = cg_sockopt_get_prologue, 1987 }; 1988 1989 const struct bpf_prog_ops cg_sockopt_prog_ops = { 1990 }; 1991