1 // SPDX-License-Identifier: GPL-2.0-only 2 /* Copyright (c) 2019 Facebook */ 3 #include <linux/hash.h> 4 #include <linux/bpf.h> 5 #include <linux/filter.h> 6 #include <linux/ftrace.h> 7 #include <linux/rbtree_latch.h> 8 #include <linux/perf_event.h> 9 #include <linux/btf.h> 10 #include <linux/rcupdate_trace.h> 11 #include <linux/rcupdate_wait.h> 12 #include <linux/module.h> 13 14 /* dummy _ops. The verifier will operate on target program's ops. */ 15 const struct bpf_verifier_ops bpf_extension_verifier_ops = { 16 }; 17 const struct bpf_prog_ops bpf_extension_prog_ops = { 18 }; 19 20 /* btf_vmlinux has ~22k attachable functions. 1k htab is enough. */ 21 #define TRAMPOLINE_HASH_BITS 10 22 #define TRAMPOLINE_TABLE_SIZE (1 << TRAMPOLINE_HASH_BITS) 23 24 static struct hlist_head trampoline_table[TRAMPOLINE_TABLE_SIZE]; 25 26 /* serializes access to trampoline_table */ 27 static DEFINE_MUTEX(trampoline_mutex); 28 29 void *bpf_jit_alloc_exec_page(void) 30 { 31 void *image; 32 33 image = bpf_jit_alloc_exec(PAGE_SIZE); 34 if (!image) 35 return NULL; 36 37 set_vm_flush_reset_perms(image); 38 /* Keep image as writeable. The alternative is to keep flipping ro/rw 39 * everytime new program is attached or detached. 40 */ 41 set_memory_x((long)image, 1); 42 return image; 43 } 44 45 void bpf_image_ksym_add(void *data, struct bpf_ksym *ksym) 46 { 47 ksym->start = (unsigned long) data; 48 ksym->end = ksym->start + PAGE_SIZE; 49 bpf_ksym_add(ksym); 50 perf_event_ksymbol(PERF_RECORD_KSYMBOL_TYPE_BPF, ksym->start, 51 PAGE_SIZE, false, ksym->name); 52 } 53 54 void bpf_image_ksym_del(struct bpf_ksym *ksym) 55 { 56 bpf_ksym_del(ksym); 57 perf_event_ksymbol(PERF_RECORD_KSYMBOL_TYPE_BPF, ksym->start, 58 PAGE_SIZE, true, ksym->name); 59 } 60 61 static struct bpf_trampoline *bpf_trampoline_lookup(u64 key) 62 { 63 struct bpf_trampoline *tr; 64 struct hlist_head *head; 65 int i; 66 67 mutex_lock(&trampoline_mutex); 68 head = &trampoline_table[hash_64(key, TRAMPOLINE_HASH_BITS)]; 69 hlist_for_each_entry(tr, head, hlist) { 70 if (tr->key == key) { 71 refcount_inc(&tr->refcnt); 72 goto out; 73 } 74 } 75 tr = kzalloc(sizeof(*tr), GFP_KERNEL); 76 if (!tr) 77 goto out; 78 79 tr->key = key; 80 INIT_HLIST_NODE(&tr->hlist); 81 hlist_add_head(&tr->hlist, head); 82 refcount_set(&tr->refcnt, 1); 83 mutex_init(&tr->mutex); 84 for (i = 0; i < BPF_TRAMP_MAX; i++) 85 INIT_HLIST_HEAD(&tr->progs_hlist[i]); 86 out: 87 mutex_unlock(&trampoline_mutex); 88 return tr; 89 } 90 91 static int bpf_trampoline_module_get(struct bpf_trampoline *tr) 92 { 93 struct module *mod; 94 int err = 0; 95 96 preempt_disable(); 97 mod = __module_text_address((unsigned long) tr->func.addr); 98 if (mod && !try_module_get(mod)) 99 err = -ENOENT; 100 preempt_enable(); 101 tr->mod = mod; 102 return err; 103 } 104 105 static void bpf_trampoline_module_put(struct bpf_trampoline *tr) 106 { 107 module_put(tr->mod); 108 tr->mod = NULL; 109 } 110 111 static int is_ftrace_location(void *ip) 112 { 113 long addr; 114 115 addr = ftrace_location((long)ip); 116 if (!addr) 117 return 0; 118 if (WARN_ON_ONCE(addr != (long)ip)) 119 return -EFAULT; 120 return 1; 121 } 122 123 static int unregister_fentry(struct bpf_trampoline *tr, void *old_addr) 124 { 125 void *ip = tr->func.addr; 126 int ret; 127 128 if (tr->func.ftrace_managed) 129 ret = unregister_ftrace_direct((long)ip, (long)old_addr); 130 else 131 ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, old_addr, NULL); 132 133 if (!ret) 134 bpf_trampoline_module_put(tr); 135 return ret; 136 } 137 138 static int modify_fentry(struct bpf_trampoline *tr, void *old_addr, void *new_addr) 139 { 140 void *ip = tr->func.addr; 141 int ret; 142 143 if (tr->func.ftrace_managed) 144 ret = modify_ftrace_direct((long)ip, (long)old_addr, (long)new_addr); 145 else 146 ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, old_addr, new_addr); 147 return ret; 148 } 149 150 /* first time registering */ 151 static int register_fentry(struct bpf_trampoline *tr, void *new_addr) 152 { 153 void *ip = tr->func.addr; 154 int ret; 155 156 ret = is_ftrace_location(ip); 157 if (ret < 0) 158 return ret; 159 tr->func.ftrace_managed = ret; 160 161 if (bpf_trampoline_module_get(tr)) 162 return -ENOENT; 163 164 if (tr->func.ftrace_managed) 165 ret = register_ftrace_direct((long)ip, (long)new_addr); 166 else 167 ret = bpf_arch_text_poke(ip, BPF_MOD_CALL, NULL, new_addr); 168 169 if (ret) 170 bpf_trampoline_module_put(tr); 171 return ret; 172 } 173 174 static struct bpf_tramp_progs * 175 bpf_trampoline_get_progs(const struct bpf_trampoline *tr, int *total) 176 { 177 const struct bpf_prog_aux *aux; 178 struct bpf_tramp_progs *tprogs; 179 struct bpf_prog **progs; 180 int kind; 181 182 *total = 0; 183 tprogs = kcalloc(BPF_TRAMP_MAX, sizeof(*tprogs), GFP_KERNEL); 184 if (!tprogs) 185 return ERR_PTR(-ENOMEM); 186 187 for (kind = 0; kind < BPF_TRAMP_MAX; kind++) { 188 tprogs[kind].nr_progs = tr->progs_cnt[kind]; 189 *total += tr->progs_cnt[kind]; 190 progs = tprogs[kind].progs; 191 192 hlist_for_each_entry(aux, &tr->progs_hlist[kind], tramp_hlist) 193 *progs++ = aux->prog; 194 } 195 return tprogs; 196 } 197 198 static void __bpf_tramp_image_put_deferred(struct work_struct *work) 199 { 200 struct bpf_tramp_image *im; 201 202 im = container_of(work, struct bpf_tramp_image, work); 203 bpf_image_ksym_del(&im->ksym); 204 bpf_jit_free_exec(im->image); 205 bpf_jit_uncharge_modmem(1); 206 percpu_ref_exit(&im->pcref); 207 kfree_rcu(im, rcu); 208 } 209 210 /* callback, fexit step 3 or fentry step 2 */ 211 static void __bpf_tramp_image_put_rcu(struct rcu_head *rcu) 212 { 213 struct bpf_tramp_image *im; 214 215 im = container_of(rcu, struct bpf_tramp_image, rcu); 216 INIT_WORK(&im->work, __bpf_tramp_image_put_deferred); 217 schedule_work(&im->work); 218 } 219 220 /* callback, fexit step 2. Called after percpu_ref_kill confirms. */ 221 static void __bpf_tramp_image_release(struct percpu_ref *pcref) 222 { 223 struct bpf_tramp_image *im; 224 225 im = container_of(pcref, struct bpf_tramp_image, pcref); 226 call_rcu_tasks(&im->rcu, __bpf_tramp_image_put_rcu); 227 } 228 229 /* callback, fexit or fentry step 1 */ 230 static void __bpf_tramp_image_put_rcu_tasks(struct rcu_head *rcu) 231 { 232 struct bpf_tramp_image *im; 233 234 im = container_of(rcu, struct bpf_tramp_image, rcu); 235 if (im->ip_after_call) 236 /* the case of fmod_ret/fexit trampoline and CONFIG_PREEMPTION=y */ 237 percpu_ref_kill(&im->pcref); 238 else 239 /* the case of fentry trampoline */ 240 call_rcu_tasks(&im->rcu, __bpf_tramp_image_put_rcu); 241 } 242 243 static void bpf_tramp_image_put(struct bpf_tramp_image *im) 244 { 245 /* The trampoline image that calls original function is using: 246 * rcu_read_lock_trace to protect sleepable bpf progs 247 * rcu_read_lock to protect normal bpf progs 248 * percpu_ref to protect trampoline itself 249 * rcu tasks to protect trampoline asm not covered by percpu_ref 250 * (which are few asm insns before __bpf_tramp_enter and 251 * after __bpf_tramp_exit) 252 * 253 * The trampoline is unreachable before bpf_tramp_image_put(). 254 * 255 * First, patch the trampoline to avoid calling into fexit progs. 256 * The progs will be freed even if the original function is still 257 * executing or sleeping. 258 * In case of CONFIG_PREEMPT=y use call_rcu_tasks() to wait on 259 * first few asm instructions to execute and call into 260 * __bpf_tramp_enter->percpu_ref_get. 261 * Then use percpu_ref_kill to wait for the trampoline and the original 262 * function to finish. 263 * Then use call_rcu_tasks() to make sure few asm insns in 264 * the trampoline epilogue are done as well. 265 * 266 * In !PREEMPT case the task that got interrupted in the first asm 267 * insns won't go through an RCU quiescent state which the 268 * percpu_ref_kill will be waiting for. Hence the first 269 * call_rcu_tasks() is not necessary. 270 */ 271 if (im->ip_after_call) { 272 int err = bpf_arch_text_poke(im->ip_after_call, BPF_MOD_JUMP, 273 NULL, im->ip_epilogue); 274 WARN_ON(err); 275 if (IS_ENABLED(CONFIG_PREEMPTION)) 276 call_rcu_tasks(&im->rcu, __bpf_tramp_image_put_rcu_tasks); 277 else 278 percpu_ref_kill(&im->pcref); 279 return; 280 } 281 282 /* The trampoline without fexit and fmod_ret progs doesn't call original 283 * function and doesn't use percpu_ref. 284 * Use call_rcu_tasks_trace() to wait for sleepable progs to finish. 285 * Then use call_rcu_tasks() to wait for the rest of trampoline asm 286 * and normal progs. 287 */ 288 call_rcu_tasks_trace(&im->rcu, __bpf_tramp_image_put_rcu_tasks); 289 } 290 291 static struct bpf_tramp_image *bpf_tramp_image_alloc(u64 key, u32 idx) 292 { 293 struct bpf_tramp_image *im; 294 struct bpf_ksym *ksym; 295 void *image; 296 int err = -ENOMEM; 297 298 im = kzalloc(sizeof(*im), GFP_KERNEL); 299 if (!im) 300 goto out; 301 302 err = bpf_jit_charge_modmem(1); 303 if (err) 304 goto out_free_im; 305 306 err = -ENOMEM; 307 im->image = image = bpf_jit_alloc_exec_page(); 308 if (!image) 309 goto out_uncharge; 310 311 err = percpu_ref_init(&im->pcref, __bpf_tramp_image_release, 0, GFP_KERNEL); 312 if (err) 313 goto out_free_image; 314 315 ksym = &im->ksym; 316 INIT_LIST_HEAD_RCU(&ksym->lnode); 317 snprintf(ksym->name, KSYM_NAME_LEN, "bpf_trampoline_%llu_%u", key, idx); 318 bpf_image_ksym_add(image, ksym); 319 return im; 320 321 out_free_image: 322 bpf_jit_free_exec(im->image); 323 out_uncharge: 324 bpf_jit_uncharge_modmem(1); 325 out_free_im: 326 kfree(im); 327 out: 328 return ERR_PTR(err); 329 } 330 331 static int bpf_trampoline_update(struct bpf_trampoline *tr) 332 { 333 struct bpf_tramp_image *im; 334 struct bpf_tramp_progs *tprogs; 335 u32 flags = BPF_TRAMP_F_RESTORE_REGS; 336 int err, total; 337 338 tprogs = bpf_trampoline_get_progs(tr, &total); 339 if (IS_ERR(tprogs)) 340 return PTR_ERR(tprogs); 341 342 if (total == 0) { 343 err = unregister_fentry(tr, tr->cur_image->image); 344 bpf_tramp_image_put(tr->cur_image); 345 tr->cur_image = NULL; 346 tr->selector = 0; 347 goto out; 348 } 349 350 im = bpf_tramp_image_alloc(tr->key, tr->selector); 351 if (IS_ERR(im)) { 352 err = PTR_ERR(im); 353 goto out; 354 } 355 356 if (tprogs[BPF_TRAMP_FEXIT].nr_progs || 357 tprogs[BPF_TRAMP_MODIFY_RETURN].nr_progs) 358 flags = BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_SKIP_FRAME; 359 360 err = arch_prepare_bpf_trampoline(im, im->image, im->image + PAGE_SIZE, 361 &tr->func.model, flags, tprogs, 362 tr->func.addr); 363 if (err < 0) 364 goto out; 365 366 WARN_ON(tr->cur_image && tr->selector == 0); 367 WARN_ON(!tr->cur_image && tr->selector); 368 if (tr->cur_image) 369 /* progs already running at this address */ 370 err = modify_fentry(tr, tr->cur_image->image, im->image); 371 else 372 /* first time registering */ 373 err = register_fentry(tr, im->image); 374 if (err) 375 goto out; 376 if (tr->cur_image) 377 bpf_tramp_image_put(tr->cur_image); 378 tr->cur_image = im; 379 tr->selector++; 380 out: 381 kfree(tprogs); 382 return err; 383 } 384 385 static enum bpf_tramp_prog_type bpf_attach_type_to_tramp(struct bpf_prog *prog) 386 { 387 switch (prog->expected_attach_type) { 388 case BPF_TRACE_FENTRY: 389 return BPF_TRAMP_FENTRY; 390 case BPF_MODIFY_RETURN: 391 return BPF_TRAMP_MODIFY_RETURN; 392 case BPF_TRACE_FEXIT: 393 return BPF_TRAMP_FEXIT; 394 case BPF_LSM_MAC: 395 if (!prog->aux->attach_func_proto->type) 396 /* The function returns void, we cannot modify its 397 * return value. 398 */ 399 return BPF_TRAMP_FEXIT; 400 else 401 return BPF_TRAMP_MODIFY_RETURN; 402 default: 403 return BPF_TRAMP_REPLACE; 404 } 405 } 406 407 int bpf_trampoline_link_prog(struct bpf_prog *prog, struct bpf_trampoline *tr) 408 { 409 enum bpf_tramp_prog_type kind; 410 int err = 0; 411 int cnt; 412 413 kind = bpf_attach_type_to_tramp(prog); 414 mutex_lock(&tr->mutex); 415 if (tr->extension_prog) { 416 /* cannot attach fentry/fexit if extension prog is attached. 417 * cannot overwrite extension prog either. 418 */ 419 err = -EBUSY; 420 goto out; 421 } 422 cnt = tr->progs_cnt[BPF_TRAMP_FENTRY] + tr->progs_cnt[BPF_TRAMP_FEXIT]; 423 if (kind == BPF_TRAMP_REPLACE) { 424 /* Cannot attach extension if fentry/fexit are in use. */ 425 if (cnt) { 426 err = -EBUSY; 427 goto out; 428 } 429 tr->extension_prog = prog; 430 err = bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP, NULL, 431 prog->bpf_func); 432 goto out; 433 } 434 if (cnt >= BPF_MAX_TRAMP_PROGS) { 435 err = -E2BIG; 436 goto out; 437 } 438 if (!hlist_unhashed(&prog->aux->tramp_hlist)) { 439 /* prog already linked */ 440 err = -EBUSY; 441 goto out; 442 } 443 hlist_add_head(&prog->aux->tramp_hlist, &tr->progs_hlist[kind]); 444 tr->progs_cnt[kind]++; 445 err = bpf_trampoline_update(tr); 446 if (err) { 447 hlist_del_init(&prog->aux->tramp_hlist); 448 tr->progs_cnt[kind]--; 449 } 450 out: 451 mutex_unlock(&tr->mutex); 452 return err; 453 } 454 455 /* bpf_trampoline_unlink_prog() should never fail. */ 456 int bpf_trampoline_unlink_prog(struct bpf_prog *prog, struct bpf_trampoline *tr) 457 { 458 enum bpf_tramp_prog_type kind; 459 int err; 460 461 kind = bpf_attach_type_to_tramp(prog); 462 mutex_lock(&tr->mutex); 463 if (kind == BPF_TRAMP_REPLACE) { 464 WARN_ON_ONCE(!tr->extension_prog); 465 err = bpf_arch_text_poke(tr->func.addr, BPF_MOD_JUMP, 466 tr->extension_prog->bpf_func, NULL); 467 tr->extension_prog = NULL; 468 goto out; 469 } 470 hlist_del_init(&prog->aux->tramp_hlist); 471 tr->progs_cnt[kind]--; 472 err = bpf_trampoline_update(tr); 473 out: 474 mutex_unlock(&tr->mutex); 475 return err; 476 } 477 478 struct bpf_trampoline *bpf_trampoline_get(u64 key, 479 struct bpf_attach_target_info *tgt_info) 480 { 481 struct bpf_trampoline *tr; 482 483 tr = bpf_trampoline_lookup(key); 484 if (!tr) 485 return NULL; 486 487 mutex_lock(&tr->mutex); 488 if (tr->func.addr) 489 goto out; 490 491 memcpy(&tr->func.model, &tgt_info->fmodel, sizeof(tgt_info->fmodel)); 492 tr->func.addr = (void *)tgt_info->tgt_addr; 493 out: 494 mutex_unlock(&tr->mutex); 495 return tr; 496 } 497 498 void bpf_trampoline_put(struct bpf_trampoline *tr) 499 { 500 if (!tr) 501 return; 502 mutex_lock(&trampoline_mutex); 503 if (!refcount_dec_and_test(&tr->refcnt)) 504 goto out; 505 WARN_ON_ONCE(mutex_is_locked(&tr->mutex)); 506 if (WARN_ON_ONCE(!hlist_empty(&tr->progs_hlist[BPF_TRAMP_FENTRY]))) 507 goto out; 508 if (WARN_ON_ONCE(!hlist_empty(&tr->progs_hlist[BPF_TRAMP_FEXIT]))) 509 goto out; 510 /* This code will be executed even when the last bpf_tramp_image 511 * is alive. All progs are detached from the trampoline and the 512 * trampoline image is patched with jmp into epilogue to skip 513 * fexit progs. The fentry-only trampoline will be freed via 514 * multiple rcu callbacks. 515 */ 516 hlist_del(&tr->hlist); 517 kfree(tr); 518 out: 519 mutex_unlock(&trampoline_mutex); 520 } 521 522 #define NO_START_TIME 1 523 static u64 notrace bpf_prog_start_time(void) 524 { 525 u64 start = NO_START_TIME; 526 527 if (static_branch_unlikely(&bpf_stats_enabled_key)) { 528 start = sched_clock(); 529 if (unlikely(!start)) 530 start = NO_START_TIME; 531 } 532 return start; 533 } 534 535 static void notrace inc_misses_counter(struct bpf_prog *prog) 536 { 537 struct bpf_prog_stats *stats; 538 539 stats = this_cpu_ptr(prog->stats); 540 u64_stats_update_begin(&stats->syncp); 541 stats->misses++; 542 u64_stats_update_end(&stats->syncp); 543 } 544 545 /* The logic is similar to BPF_PROG_RUN, but with an explicit 546 * rcu_read_lock() and migrate_disable() which are required 547 * for the trampoline. The macro is split into 548 * call __bpf_prog_enter 549 * call prog->bpf_func 550 * call __bpf_prog_exit 551 * 552 * __bpf_prog_enter returns: 553 * 0 - skip execution of the bpf prog 554 * 1 - execute bpf prog 555 * [2..MAX_U64] - excute bpf prog and record execution time. 556 * This is start time. 557 */ 558 u64 notrace __bpf_prog_enter(struct bpf_prog *prog) 559 __acquires(RCU) 560 { 561 rcu_read_lock(); 562 migrate_disable(); 563 if (unlikely(__this_cpu_inc_return(*(prog->active)) != 1)) { 564 inc_misses_counter(prog); 565 return 0; 566 } 567 return bpf_prog_start_time(); 568 } 569 570 static void notrace update_prog_stats(struct bpf_prog *prog, 571 u64 start) 572 { 573 struct bpf_prog_stats *stats; 574 575 if (static_branch_unlikely(&bpf_stats_enabled_key) && 576 /* static_key could be enabled in __bpf_prog_enter* 577 * and disabled in __bpf_prog_exit*. 578 * And vice versa. 579 * Hence check that 'start' is valid. 580 */ 581 start > NO_START_TIME) { 582 stats = this_cpu_ptr(prog->stats); 583 u64_stats_update_begin(&stats->syncp); 584 stats->cnt++; 585 stats->nsecs += sched_clock() - start; 586 u64_stats_update_end(&stats->syncp); 587 } 588 } 589 590 void notrace __bpf_prog_exit(struct bpf_prog *prog, u64 start) 591 __releases(RCU) 592 { 593 update_prog_stats(prog, start); 594 __this_cpu_dec(*(prog->active)); 595 migrate_enable(); 596 rcu_read_unlock(); 597 } 598 599 u64 notrace __bpf_prog_enter_sleepable(struct bpf_prog *prog) 600 { 601 rcu_read_lock_trace(); 602 migrate_disable(); 603 might_fault(); 604 if (unlikely(__this_cpu_inc_return(*(prog->active)) != 1)) { 605 inc_misses_counter(prog); 606 return 0; 607 } 608 return bpf_prog_start_time(); 609 } 610 611 void notrace __bpf_prog_exit_sleepable(struct bpf_prog *prog, u64 start) 612 { 613 update_prog_stats(prog, start); 614 __this_cpu_dec(*(prog->active)); 615 migrate_enable(); 616 rcu_read_unlock_trace(); 617 } 618 619 void notrace __bpf_tramp_enter(struct bpf_tramp_image *tr) 620 { 621 percpu_ref_get(&tr->pcref); 622 } 623 624 void notrace __bpf_tramp_exit(struct bpf_tramp_image *tr) 625 { 626 percpu_ref_put(&tr->pcref); 627 } 628 629 int __weak 630 arch_prepare_bpf_trampoline(struct bpf_tramp_image *tr, void *image, void *image_end, 631 const struct btf_func_model *m, u32 flags, 632 struct bpf_tramp_progs *tprogs, 633 void *orig_call) 634 { 635 return -ENOTSUPP; 636 } 637 638 static int __init init_trampolines(void) 639 { 640 int i; 641 642 for (i = 0; i < TRAMPOLINE_TABLE_SIZE; i++) 643 INIT_HLIST_HEAD(&trampoline_table[i]); 644 return 0; 645 } 646 late_initcall(init_trampolines); 647