1 // SPDX-License-Identifier: GPL-2.0 2 /* Copyright (c) 2011-2015 PLUMgrid, http://plumgrid.com 3 * Copyright (c) 2016 Facebook 4 */ 5 #include <linux/kernel.h> 6 #include <linux/types.h> 7 #include <linux/slab.h> 8 #include <linux/bpf.h> 9 #include <linux/bpf_perf_event.h> 10 #include <linux/filter.h> 11 #include <linux/uaccess.h> 12 #include <linux/ctype.h> 13 #include <linux/kprobes.h> 14 #include <linux/syscalls.h> 15 #include <linux/error-injection.h> 16 17 #include "trace_probe.h" 18 #include "trace.h" 19 20 u64 bpf_get_stackid(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5); 21 u64 bpf_get_stack(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5); 22 23 /** 24 * trace_call_bpf - invoke BPF program 25 * @call: tracepoint event 26 * @ctx: opaque context pointer 27 * 28 * kprobe handlers execute BPF programs via this helper. 29 * Can be used from static tracepoints in the future. 30 * 31 * Return: BPF programs always return an integer which is interpreted by 32 * kprobe handler as: 33 * 0 - return from kprobe (event is filtered out) 34 * 1 - store kprobe event into ring buffer 35 * Other values are reserved and currently alias to 1 36 */ 37 unsigned int trace_call_bpf(struct trace_event_call *call, void *ctx) 38 { 39 unsigned int ret; 40 41 if (in_nmi()) /* not supported yet */ 42 return 1; 43 44 preempt_disable(); 45 46 if (unlikely(__this_cpu_inc_return(bpf_prog_active) != 1)) { 47 /* 48 * since some bpf program is already running on this cpu, 49 * don't call into another bpf program (same or different) 50 * and don't send kprobe event into ring-buffer, 51 * so return zero here 52 */ 53 ret = 0; 54 goto out; 55 } 56 57 /* 58 * Instead of moving rcu_read_lock/rcu_dereference/rcu_read_unlock 59 * to all call sites, we did a bpf_prog_array_valid() there to check 60 * whether call->prog_array is empty or not, which is 61 * a heurisitc to speed up execution. 62 * 63 * If bpf_prog_array_valid() fetched prog_array was 64 * non-NULL, we go into trace_call_bpf() and do the actual 65 * proper rcu_dereference() under RCU lock. 66 * If it turns out that prog_array is NULL then, we bail out. 67 * For the opposite, if the bpf_prog_array_valid() fetched pointer 68 * was NULL, you'll skip the prog_array with the risk of missing 69 * out of events when it was updated in between this and the 70 * rcu_dereference() which is accepted risk. 71 */ 72 ret = BPF_PROG_RUN_ARRAY_CHECK(call->prog_array, ctx, BPF_PROG_RUN); 73 74 out: 75 __this_cpu_dec(bpf_prog_active); 76 preempt_enable(); 77 78 return ret; 79 } 80 EXPORT_SYMBOL_GPL(trace_call_bpf); 81 82 #ifdef CONFIG_BPF_KPROBE_OVERRIDE 83 BPF_CALL_2(bpf_override_return, struct pt_regs *, regs, unsigned long, rc) 84 { 85 regs_set_return_value(regs, rc); 86 override_function_with_return(regs); 87 return 0; 88 } 89 90 static const struct bpf_func_proto bpf_override_return_proto = { 91 .func = bpf_override_return, 92 .gpl_only = true, 93 .ret_type = RET_INTEGER, 94 .arg1_type = ARG_PTR_TO_CTX, 95 .arg2_type = ARG_ANYTHING, 96 }; 97 #endif 98 99 BPF_CALL_3(bpf_probe_read, void *, dst, u32, size, const void *, unsafe_ptr) 100 { 101 int ret; 102 103 ret = probe_kernel_read(dst, unsafe_ptr, size); 104 if (unlikely(ret < 0)) 105 memset(dst, 0, size); 106 107 return ret; 108 } 109 110 static const struct bpf_func_proto bpf_probe_read_proto = { 111 .func = bpf_probe_read, 112 .gpl_only = true, 113 .ret_type = RET_INTEGER, 114 .arg1_type = ARG_PTR_TO_UNINIT_MEM, 115 .arg2_type = ARG_CONST_SIZE_OR_ZERO, 116 .arg3_type = ARG_ANYTHING, 117 }; 118 119 BPF_CALL_3(bpf_probe_write_user, void *, unsafe_ptr, const void *, src, 120 u32, size) 121 { 122 /* 123 * Ensure we're in user context which is safe for the helper to 124 * run. This helper has no business in a kthread. 125 * 126 * access_ok() should prevent writing to non-user memory, but in 127 * some situations (nommu, temporary switch, etc) access_ok() does 128 * not provide enough validation, hence the check on KERNEL_DS. 129 */ 130 131 if (unlikely(in_interrupt() || 132 current->flags & (PF_KTHREAD | PF_EXITING))) 133 return -EPERM; 134 if (unlikely(uaccess_kernel())) 135 return -EPERM; 136 if (!access_ok(VERIFY_WRITE, unsafe_ptr, size)) 137 return -EPERM; 138 139 return probe_kernel_write(unsafe_ptr, src, size); 140 } 141 142 static const struct bpf_func_proto bpf_probe_write_user_proto = { 143 .func = bpf_probe_write_user, 144 .gpl_only = true, 145 .ret_type = RET_INTEGER, 146 .arg1_type = ARG_ANYTHING, 147 .arg2_type = ARG_PTR_TO_MEM, 148 .arg3_type = ARG_CONST_SIZE, 149 }; 150 151 static const struct bpf_func_proto *bpf_get_probe_write_proto(void) 152 { 153 pr_warn_ratelimited("%s[%d] is installing a program with bpf_probe_write_user helper that may corrupt user memory!", 154 current->comm, task_pid_nr(current)); 155 156 return &bpf_probe_write_user_proto; 157 } 158 159 /* 160 * Only limited trace_printk() conversion specifiers allowed: 161 * %d %i %u %x %ld %li %lu %lx %lld %lli %llu %llx %p %s 162 */ 163 BPF_CALL_5(bpf_trace_printk, char *, fmt, u32, fmt_size, u64, arg1, 164 u64, arg2, u64, arg3) 165 { 166 bool str_seen = false; 167 int mod[3] = {}; 168 int fmt_cnt = 0; 169 u64 unsafe_addr; 170 char buf[64]; 171 int i; 172 173 /* 174 * bpf_check()->check_func_arg()->check_stack_boundary() 175 * guarantees that fmt points to bpf program stack, 176 * fmt_size bytes of it were initialized and fmt_size > 0 177 */ 178 if (fmt[--fmt_size] != 0) 179 return -EINVAL; 180 181 /* check format string for allowed specifiers */ 182 for (i = 0; i < fmt_size; i++) { 183 if ((!isprint(fmt[i]) && !isspace(fmt[i])) || !isascii(fmt[i])) 184 return -EINVAL; 185 186 if (fmt[i] != '%') 187 continue; 188 189 if (fmt_cnt >= 3) 190 return -EINVAL; 191 192 /* fmt[i] != 0 && fmt[last] == 0, so we can access fmt[i + 1] */ 193 i++; 194 if (fmt[i] == 'l') { 195 mod[fmt_cnt]++; 196 i++; 197 } else if (fmt[i] == 'p' || fmt[i] == 's') { 198 mod[fmt_cnt]++; 199 /* disallow any further format extensions */ 200 if (fmt[i + 1] != 0 && 201 !isspace(fmt[i + 1]) && 202 !ispunct(fmt[i + 1])) 203 return -EINVAL; 204 fmt_cnt++; 205 if (fmt[i] == 's') { 206 if (str_seen) 207 /* allow only one '%s' per fmt string */ 208 return -EINVAL; 209 str_seen = true; 210 211 switch (fmt_cnt) { 212 case 1: 213 unsafe_addr = arg1; 214 arg1 = (long) buf; 215 break; 216 case 2: 217 unsafe_addr = arg2; 218 arg2 = (long) buf; 219 break; 220 case 3: 221 unsafe_addr = arg3; 222 arg3 = (long) buf; 223 break; 224 } 225 buf[0] = 0; 226 strncpy_from_unsafe(buf, 227 (void *) (long) unsafe_addr, 228 sizeof(buf)); 229 } 230 continue; 231 } 232 233 if (fmt[i] == 'l') { 234 mod[fmt_cnt]++; 235 i++; 236 } 237 238 if (fmt[i] != 'i' && fmt[i] != 'd' && 239 fmt[i] != 'u' && fmt[i] != 'x') 240 return -EINVAL; 241 fmt_cnt++; 242 } 243 244 /* Horrid workaround for getting va_list handling working with different 245 * argument type combinations generically for 32 and 64 bit archs. 246 */ 247 #define __BPF_TP_EMIT() __BPF_ARG3_TP() 248 #define __BPF_TP(...) \ 249 __trace_printk(0 /* Fake ip */, \ 250 fmt, ##__VA_ARGS__) 251 252 #define __BPF_ARG1_TP(...) \ 253 ((mod[0] == 2 || (mod[0] == 1 && __BITS_PER_LONG == 64)) \ 254 ? __BPF_TP(arg1, ##__VA_ARGS__) \ 255 : ((mod[0] == 1 || (mod[0] == 0 && __BITS_PER_LONG == 32)) \ 256 ? __BPF_TP((long)arg1, ##__VA_ARGS__) \ 257 : __BPF_TP((u32)arg1, ##__VA_ARGS__))) 258 259 #define __BPF_ARG2_TP(...) \ 260 ((mod[1] == 2 || (mod[1] == 1 && __BITS_PER_LONG == 64)) \ 261 ? __BPF_ARG1_TP(arg2, ##__VA_ARGS__) \ 262 : ((mod[1] == 1 || (mod[1] == 0 && __BITS_PER_LONG == 32)) \ 263 ? __BPF_ARG1_TP((long)arg2, ##__VA_ARGS__) \ 264 : __BPF_ARG1_TP((u32)arg2, ##__VA_ARGS__))) 265 266 #define __BPF_ARG3_TP(...) \ 267 ((mod[2] == 2 || (mod[2] == 1 && __BITS_PER_LONG == 64)) \ 268 ? __BPF_ARG2_TP(arg3, ##__VA_ARGS__) \ 269 : ((mod[2] == 1 || (mod[2] == 0 && __BITS_PER_LONG == 32)) \ 270 ? __BPF_ARG2_TP((long)arg3, ##__VA_ARGS__) \ 271 : __BPF_ARG2_TP((u32)arg3, ##__VA_ARGS__))) 272 273 return __BPF_TP_EMIT(); 274 } 275 276 static const struct bpf_func_proto bpf_trace_printk_proto = { 277 .func = bpf_trace_printk, 278 .gpl_only = true, 279 .ret_type = RET_INTEGER, 280 .arg1_type = ARG_PTR_TO_MEM, 281 .arg2_type = ARG_CONST_SIZE, 282 }; 283 284 const struct bpf_func_proto *bpf_get_trace_printk_proto(void) 285 { 286 /* 287 * this program might be calling bpf_trace_printk, 288 * so allocate per-cpu printk buffers 289 */ 290 trace_printk_init_buffers(); 291 292 return &bpf_trace_printk_proto; 293 } 294 295 static __always_inline int 296 get_map_perf_counter(struct bpf_map *map, u64 flags, 297 u64 *value, u64 *enabled, u64 *running) 298 { 299 struct bpf_array *array = container_of(map, struct bpf_array, map); 300 unsigned int cpu = smp_processor_id(); 301 u64 index = flags & BPF_F_INDEX_MASK; 302 struct bpf_event_entry *ee; 303 304 if (unlikely(flags & ~(BPF_F_INDEX_MASK))) 305 return -EINVAL; 306 if (index == BPF_F_CURRENT_CPU) 307 index = cpu; 308 if (unlikely(index >= array->map.max_entries)) 309 return -E2BIG; 310 311 ee = READ_ONCE(array->ptrs[index]); 312 if (!ee) 313 return -ENOENT; 314 315 return perf_event_read_local(ee->event, value, enabled, running); 316 } 317 318 BPF_CALL_2(bpf_perf_event_read, struct bpf_map *, map, u64, flags) 319 { 320 u64 value = 0; 321 int err; 322 323 err = get_map_perf_counter(map, flags, &value, NULL, NULL); 324 /* 325 * this api is ugly since we miss [-22..-2] range of valid 326 * counter values, but that's uapi 327 */ 328 if (err) 329 return err; 330 return value; 331 } 332 333 static const struct bpf_func_proto bpf_perf_event_read_proto = { 334 .func = bpf_perf_event_read, 335 .gpl_only = true, 336 .ret_type = RET_INTEGER, 337 .arg1_type = ARG_CONST_MAP_PTR, 338 .arg2_type = ARG_ANYTHING, 339 }; 340 341 BPF_CALL_4(bpf_perf_event_read_value, struct bpf_map *, map, u64, flags, 342 struct bpf_perf_event_value *, buf, u32, size) 343 { 344 int err = -EINVAL; 345 346 if (unlikely(size != sizeof(struct bpf_perf_event_value))) 347 goto clear; 348 err = get_map_perf_counter(map, flags, &buf->counter, &buf->enabled, 349 &buf->running); 350 if (unlikely(err)) 351 goto clear; 352 return 0; 353 clear: 354 memset(buf, 0, size); 355 return err; 356 } 357 358 static const struct bpf_func_proto bpf_perf_event_read_value_proto = { 359 .func = bpf_perf_event_read_value, 360 .gpl_only = true, 361 .ret_type = RET_INTEGER, 362 .arg1_type = ARG_CONST_MAP_PTR, 363 .arg2_type = ARG_ANYTHING, 364 .arg3_type = ARG_PTR_TO_UNINIT_MEM, 365 .arg4_type = ARG_CONST_SIZE, 366 }; 367 368 static DEFINE_PER_CPU(struct perf_sample_data, bpf_trace_sd); 369 370 static __always_inline u64 371 __bpf_perf_event_output(struct pt_regs *regs, struct bpf_map *map, 372 u64 flags, struct perf_sample_data *sd) 373 { 374 struct bpf_array *array = container_of(map, struct bpf_array, map); 375 unsigned int cpu = smp_processor_id(); 376 u64 index = flags & BPF_F_INDEX_MASK; 377 struct bpf_event_entry *ee; 378 struct perf_event *event; 379 380 if (index == BPF_F_CURRENT_CPU) 381 index = cpu; 382 if (unlikely(index >= array->map.max_entries)) 383 return -E2BIG; 384 385 ee = READ_ONCE(array->ptrs[index]); 386 if (!ee) 387 return -ENOENT; 388 389 event = ee->event; 390 if (unlikely(event->attr.type != PERF_TYPE_SOFTWARE || 391 event->attr.config != PERF_COUNT_SW_BPF_OUTPUT)) 392 return -EINVAL; 393 394 if (unlikely(event->oncpu != cpu)) 395 return -EOPNOTSUPP; 396 397 perf_event_output(event, sd, regs); 398 return 0; 399 } 400 401 BPF_CALL_5(bpf_perf_event_output, struct pt_regs *, regs, struct bpf_map *, map, 402 u64, flags, void *, data, u64, size) 403 { 404 struct perf_sample_data *sd = this_cpu_ptr(&bpf_trace_sd); 405 struct perf_raw_record raw = { 406 .frag = { 407 .size = size, 408 .data = data, 409 }, 410 }; 411 412 if (unlikely(flags & ~(BPF_F_INDEX_MASK))) 413 return -EINVAL; 414 415 perf_sample_data_init(sd, 0, 0); 416 sd->raw = &raw; 417 418 return __bpf_perf_event_output(regs, map, flags, sd); 419 } 420 421 static const struct bpf_func_proto bpf_perf_event_output_proto = { 422 .func = bpf_perf_event_output, 423 .gpl_only = true, 424 .ret_type = RET_INTEGER, 425 .arg1_type = ARG_PTR_TO_CTX, 426 .arg2_type = ARG_CONST_MAP_PTR, 427 .arg3_type = ARG_ANYTHING, 428 .arg4_type = ARG_PTR_TO_MEM, 429 .arg5_type = ARG_CONST_SIZE_OR_ZERO, 430 }; 431 432 static DEFINE_PER_CPU(struct pt_regs, bpf_pt_regs); 433 static DEFINE_PER_CPU(struct perf_sample_data, bpf_misc_sd); 434 435 u64 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 436 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 437 { 438 struct perf_sample_data *sd = this_cpu_ptr(&bpf_misc_sd); 439 struct pt_regs *regs = this_cpu_ptr(&bpf_pt_regs); 440 struct perf_raw_frag frag = { 441 .copy = ctx_copy, 442 .size = ctx_size, 443 .data = ctx, 444 }; 445 struct perf_raw_record raw = { 446 .frag = { 447 { 448 .next = ctx_size ? &frag : NULL, 449 }, 450 .size = meta_size, 451 .data = meta, 452 }, 453 }; 454 455 perf_fetch_caller_regs(regs); 456 perf_sample_data_init(sd, 0, 0); 457 sd->raw = &raw; 458 459 return __bpf_perf_event_output(regs, map, flags, sd); 460 } 461 462 BPF_CALL_0(bpf_get_current_task) 463 { 464 return (long) current; 465 } 466 467 static const struct bpf_func_proto bpf_get_current_task_proto = { 468 .func = bpf_get_current_task, 469 .gpl_only = true, 470 .ret_type = RET_INTEGER, 471 }; 472 473 BPF_CALL_2(bpf_current_task_under_cgroup, struct bpf_map *, map, u32, idx) 474 { 475 struct bpf_array *array = container_of(map, struct bpf_array, map); 476 struct cgroup *cgrp; 477 478 if (unlikely(idx >= array->map.max_entries)) 479 return -E2BIG; 480 481 cgrp = READ_ONCE(array->ptrs[idx]); 482 if (unlikely(!cgrp)) 483 return -EAGAIN; 484 485 return task_under_cgroup_hierarchy(current, cgrp); 486 } 487 488 static const struct bpf_func_proto bpf_current_task_under_cgroup_proto = { 489 .func = bpf_current_task_under_cgroup, 490 .gpl_only = false, 491 .ret_type = RET_INTEGER, 492 .arg1_type = ARG_CONST_MAP_PTR, 493 .arg2_type = ARG_ANYTHING, 494 }; 495 496 BPF_CALL_3(bpf_probe_read_str, void *, dst, u32, size, 497 const void *, unsafe_ptr) 498 { 499 int ret; 500 501 /* 502 * The strncpy_from_unsafe() call will likely not fill the entire 503 * buffer, but that's okay in this circumstance as we're probing 504 * arbitrary memory anyway similar to bpf_probe_read() and might 505 * as well probe the stack. Thus, memory is explicitly cleared 506 * only in error case, so that improper users ignoring return 507 * code altogether don't copy garbage; otherwise length of string 508 * is returned that can be used for bpf_perf_event_output() et al. 509 */ 510 ret = strncpy_from_unsafe(dst, unsafe_ptr, size); 511 if (unlikely(ret < 0)) 512 memset(dst, 0, size); 513 514 return ret; 515 } 516 517 static const struct bpf_func_proto bpf_probe_read_str_proto = { 518 .func = bpf_probe_read_str, 519 .gpl_only = true, 520 .ret_type = RET_INTEGER, 521 .arg1_type = ARG_PTR_TO_UNINIT_MEM, 522 .arg2_type = ARG_CONST_SIZE_OR_ZERO, 523 .arg3_type = ARG_ANYTHING, 524 }; 525 526 static const struct bpf_func_proto * 527 tracing_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog) 528 { 529 switch (func_id) { 530 case BPF_FUNC_map_lookup_elem: 531 return &bpf_map_lookup_elem_proto; 532 case BPF_FUNC_map_update_elem: 533 return &bpf_map_update_elem_proto; 534 case BPF_FUNC_map_delete_elem: 535 return &bpf_map_delete_elem_proto; 536 case BPF_FUNC_probe_read: 537 return &bpf_probe_read_proto; 538 case BPF_FUNC_ktime_get_ns: 539 return &bpf_ktime_get_ns_proto; 540 case BPF_FUNC_tail_call: 541 return &bpf_tail_call_proto; 542 case BPF_FUNC_get_current_pid_tgid: 543 return &bpf_get_current_pid_tgid_proto; 544 case BPF_FUNC_get_current_task: 545 return &bpf_get_current_task_proto; 546 case BPF_FUNC_get_current_uid_gid: 547 return &bpf_get_current_uid_gid_proto; 548 case BPF_FUNC_get_current_comm: 549 return &bpf_get_current_comm_proto; 550 case BPF_FUNC_trace_printk: 551 return bpf_get_trace_printk_proto(); 552 case BPF_FUNC_get_smp_processor_id: 553 return &bpf_get_smp_processor_id_proto; 554 case BPF_FUNC_get_numa_node_id: 555 return &bpf_get_numa_node_id_proto; 556 case BPF_FUNC_perf_event_read: 557 return &bpf_perf_event_read_proto; 558 case BPF_FUNC_probe_write_user: 559 return bpf_get_probe_write_proto(); 560 case BPF_FUNC_current_task_under_cgroup: 561 return &bpf_current_task_under_cgroup_proto; 562 case BPF_FUNC_get_prandom_u32: 563 return &bpf_get_prandom_u32_proto; 564 case BPF_FUNC_probe_read_str: 565 return &bpf_probe_read_str_proto; 566 #ifdef CONFIG_CGROUPS 567 case BPF_FUNC_get_current_cgroup_id: 568 return &bpf_get_current_cgroup_id_proto; 569 #endif 570 default: 571 return NULL; 572 } 573 } 574 575 static const struct bpf_func_proto * 576 kprobe_prog_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog) 577 { 578 switch (func_id) { 579 case BPF_FUNC_perf_event_output: 580 return &bpf_perf_event_output_proto; 581 case BPF_FUNC_get_stackid: 582 return &bpf_get_stackid_proto; 583 case BPF_FUNC_get_stack: 584 return &bpf_get_stack_proto; 585 case BPF_FUNC_perf_event_read_value: 586 return &bpf_perf_event_read_value_proto; 587 #ifdef CONFIG_BPF_KPROBE_OVERRIDE 588 case BPF_FUNC_override_return: 589 return &bpf_override_return_proto; 590 #endif 591 default: 592 return tracing_func_proto(func_id, prog); 593 } 594 } 595 596 /* bpf+kprobe programs can access fields of 'struct pt_regs' */ 597 static bool kprobe_prog_is_valid_access(int off, int size, enum bpf_access_type type, 598 const struct bpf_prog *prog, 599 struct bpf_insn_access_aux *info) 600 { 601 if (off < 0 || off >= sizeof(struct pt_regs)) 602 return false; 603 if (type != BPF_READ) 604 return false; 605 if (off % size != 0) 606 return false; 607 /* 608 * Assertion for 32 bit to make sure last 8 byte access 609 * (BPF_DW) to the last 4 byte member is disallowed. 610 */ 611 if (off + size > sizeof(struct pt_regs)) 612 return false; 613 614 return true; 615 } 616 617 const struct bpf_verifier_ops kprobe_verifier_ops = { 618 .get_func_proto = kprobe_prog_func_proto, 619 .is_valid_access = kprobe_prog_is_valid_access, 620 }; 621 622 const struct bpf_prog_ops kprobe_prog_ops = { 623 }; 624 625 BPF_CALL_5(bpf_perf_event_output_tp, void *, tp_buff, struct bpf_map *, map, 626 u64, flags, void *, data, u64, size) 627 { 628 struct pt_regs *regs = *(struct pt_regs **)tp_buff; 629 630 /* 631 * r1 points to perf tracepoint buffer where first 8 bytes are hidden 632 * from bpf program and contain a pointer to 'struct pt_regs'. Fetch it 633 * from there and call the same bpf_perf_event_output() helper inline. 634 */ 635 return ____bpf_perf_event_output(regs, map, flags, data, size); 636 } 637 638 static const struct bpf_func_proto bpf_perf_event_output_proto_tp = { 639 .func = bpf_perf_event_output_tp, 640 .gpl_only = true, 641 .ret_type = RET_INTEGER, 642 .arg1_type = ARG_PTR_TO_CTX, 643 .arg2_type = ARG_CONST_MAP_PTR, 644 .arg3_type = ARG_ANYTHING, 645 .arg4_type = ARG_PTR_TO_MEM, 646 .arg5_type = ARG_CONST_SIZE_OR_ZERO, 647 }; 648 649 BPF_CALL_3(bpf_get_stackid_tp, void *, tp_buff, struct bpf_map *, map, 650 u64, flags) 651 { 652 struct pt_regs *regs = *(struct pt_regs **)tp_buff; 653 654 /* 655 * Same comment as in bpf_perf_event_output_tp(), only that this time 656 * the other helper's function body cannot be inlined due to being 657 * external, thus we need to call raw helper function. 658 */ 659 return bpf_get_stackid((unsigned long) regs, (unsigned long) map, 660 flags, 0, 0); 661 } 662 663 static const struct bpf_func_proto bpf_get_stackid_proto_tp = { 664 .func = bpf_get_stackid_tp, 665 .gpl_only = true, 666 .ret_type = RET_INTEGER, 667 .arg1_type = ARG_PTR_TO_CTX, 668 .arg2_type = ARG_CONST_MAP_PTR, 669 .arg3_type = ARG_ANYTHING, 670 }; 671 672 BPF_CALL_4(bpf_get_stack_tp, void *, tp_buff, void *, buf, u32, size, 673 u64, flags) 674 { 675 struct pt_regs *regs = *(struct pt_regs **)tp_buff; 676 677 return bpf_get_stack((unsigned long) regs, (unsigned long) buf, 678 (unsigned long) size, flags, 0); 679 } 680 681 static const struct bpf_func_proto bpf_get_stack_proto_tp = { 682 .func = bpf_get_stack_tp, 683 .gpl_only = true, 684 .ret_type = RET_INTEGER, 685 .arg1_type = ARG_PTR_TO_CTX, 686 .arg2_type = ARG_PTR_TO_UNINIT_MEM, 687 .arg3_type = ARG_CONST_SIZE_OR_ZERO, 688 .arg4_type = ARG_ANYTHING, 689 }; 690 691 static const struct bpf_func_proto * 692 tp_prog_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog) 693 { 694 switch (func_id) { 695 case BPF_FUNC_perf_event_output: 696 return &bpf_perf_event_output_proto_tp; 697 case BPF_FUNC_get_stackid: 698 return &bpf_get_stackid_proto_tp; 699 case BPF_FUNC_get_stack: 700 return &bpf_get_stack_proto_tp; 701 default: 702 return tracing_func_proto(func_id, prog); 703 } 704 } 705 706 static bool tp_prog_is_valid_access(int off, int size, enum bpf_access_type type, 707 const struct bpf_prog *prog, 708 struct bpf_insn_access_aux *info) 709 { 710 if (off < sizeof(void *) || off >= PERF_MAX_TRACE_SIZE) 711 return false; 712 if (type != BPF_READ) 713 return false; 714 if (off % size != 0) 715 return false; 716 717 BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(__u64)); 718 return true; 719 } 720 721 const struct bpf_verifier_ops tracepoint_verifier_ops = { 722 .get_func_proto = tp_prog_func_proto, 723 .is_valid_access = tp_prog_is_valid_access, 724 }; 725 726 const struct bpf_prog_ops tracepoint_prog_ops = { 727 }; 728 729 BPF_CALL_3(bpf_perf_prog_read_value, struct bpf_perf_event_data_kern *, ctx, 730 struct bpf_perf_event_value *, buf, u32, size) 731 { 732 int err = -EINVAL; 733 734 if (unlikely(size != sizeof(struct bpf_perf_event_value))) 735 goto clear; 736 err = perf_event_read_local(ctx->event, &buf->counter, &buf->enabled, 737 &buf->running); 738 if (unlikely(err)) 739 goto clear; 740 return 0; 741 clear: 742 memset(buf, 0, size); 743 return err; 744 } 745 746 static const struct bpf_func_proto bpf_perf_prog_read_value_proto = { 747 .func = bpf_perf_prog_read_value, 748 .gpl_only = true, 749 .ret_type = RET_INTEGER, 750 .arg1_type = ARG_PTR_TO_CTX, 751 .arg2_type = ARG_PTR_TO_UNINIT_MEM, 752 .arg3_type = ARG_CONST_SIZE, 753 }; 754 755 static const struct bpf_func_proto * 756 pe_prog_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog) 757 { 758 switch (func_id) { 759 case BPF_FUNC_perf_event_output: 760 return &bpf_perf_event_output_proto_tp; 761 case BPF_FUNC_get_stackid: 762 return &bpf_get_stackid_proto_tp; 763 case BPF_FUNC_get_stack: 764 return &bpf_get_stack_proto_tp; 765 case BPF_FUNC_perf_prog_read_value: 766 return &bpf_perf_prog_read_value_proto; 767 default: 768 return tracing_func_proto(func_id, prog); 769 } 770 } 771 772 /* 773 * bpf_raw_tp_regs are separate from bpf_pt_regs used from skb/xdp 774 * to avoid potential recursive reuse issue when/if tracepoints are added 775 * inside bpf_*_event_output, bpf_get_stackid and/or bpf_get_stack 776 */ 777 static DEFINE_PER_CPU(struct pt_regs, bpf_raw_tp_regs); 778 BPF_CALL_5(bpf_perf_event_output_raw_tp, struct bpf_raw_tracepoint_args *, args, 779 struct bpf_map *, map, u64, flags, void *, data, u64, size) 780 { 781 struct pt_regs *regs = this_cpu_ptr(&bpf_raw_tp_regs); 782 783 perf_fetch_caller_regs(regs); 784 return ____bpf_perf_event_output(regs, map, flags, data, size); 785 } 786 787 static const struct bpf_func_proto bpf_perf_event_output_proto_raw_tp = { 788 .func = bpf_perf_event_output_raw_tp, 789 .gpl_only = true, 790 .ret_type = RET_INTEGER, 791 .arg1_type = ARG_PTR_TO_CTX, 792 .arg2_type = ARG_CONST_MAP_PTR, 793 .arg3_type = ARG_ANYTHING, 794 .arg4_type = ARG_PTR_TO_MEM, 795 .arg5_type = ARG_CONST_SIZE_OR_ZERO, 796 }; 797 798 BPF_CALL_3(bpf_get_stackid_raw_tp, struct bpf_raw_tracepoint_args *, args, 799 struct bpf_map *, map, u64, flags) 800 { 801 struct pt_regs *regs = this_cpu_ptr(&bpf_raw_tp_regs); 802 803 perf_fetch_caller_regs(regs); 804 /* similar to bpf_perf_event_output_tp, but pt_regs fetched differently */ 805 return bpf_get_stackid((unsigned long) regs, (unsigned long) map, 806 flags, 0, 0); 807 } 808 809 static const struct bpf_func_proto bpf_get_stackid_proto_raw_tp = { 810 .func = bpf_get_stackid_raw_tp, 811 .gpl_only = true, 812 .ret_type = RET_INTEGER, 813 .arg1_type = ARG_PTR_TO_CTX, 814 .arg2_type = ARG_CONST_MAP_PTR, 815 .arg3_type = ARG_ANYTHING, 816 }; 817 818 BPF_CALL_4(bpf_get_stack_raw_tp, struct bpf_raw_tracepoint_args *, args, 819 void *, buf, u32, size, u64, flags) 820 { 821 struct pt_regs *regs = this_cpu_ptr(&bpf_raw_tp_regs); 822 823 perf_fetch_caller_regs(regs); 824 return bpf_get_stack((unsigned long) regs, (unsigned long) buf, 825 (unsigned long) size, flags, 0); 826 } 827 828 static const struct bpf_func_proto bpf_get_stack_proto_raw_tp = { 829 .func = bpf_get_stack_raw_tp, 830 .gpl_only = true, 831 .ret_type = RET_INTEGER, 832 .arg1_type = ARG_PTR_TO_CTX, 833 .arg2_type = ARG_PTR_TO_MEM, 834 .arg3_type = ARG_CONST_SIZE_OR_ZERO, 835 .arg4_type = ARG_ANYTHING, 836 }; 837 838 static const struct bpf_func_proto * 839 raw_tp_prog_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog) 840 { 841 switch (func_id) { 842 case BPF_FUNC_perf_event_output: 843 return &bpf_perf_event_output_proto_raw_tp; 844 case BPF_FUNC_get_stackid: 845 return &bpf_get_stackid_proto_raw_tp; 846 case BPF_FUNC_get_stack: 847 return &bpf_get_stack_proto_raw_tp; 848 default: 849 return tracing_func_proto(func_id, prog); 850 } 851 } 852 853 static bool raw_tp_prog_is_valid_access(int off, int size, 854 enum bpf_access_type type, 855 const struct bpf_prog *prog, 856 struct bpf_insn_access_aux *info) 857 { 858 /* largest tracepoint in the kernel has 12 args */ 859 if (off < 0 || off >= sizeof(__u64) * 12) 860 return false; 861 if (type != BPF_READ) 862 return false; 863 if (off % size != 0) 864 return false; 865 return true; 866 } 867 868 const struct bpf_verifier_ops raw_tracepoint_verifier_ops = { 869 .get_func_proto = raw_tp_prog_func_proto, 870 .is_valid_access = raw_tp_prog_is_valid_access, 871 }; 872 873 const struct bpf_prog_ops raw_tracepoint_prog_ops = { 874 }; 875 876 static bool pe_prog_is_valid_access(int off, int size, enum bpf_access_type type, 877 const struct bpf_prog *prog, 878 struct bpf_insn_access_aux *info) 879 { 880 const int size_u64 = sizeof(u64); 881 882 if (off < 0 || off >= sizeof(struct bpf_perf_event_data)) 883 return false; 884 if (type != BPF_READ) 885 return false; 886 if (off % size != 0) { 887 if (sizeof(unsigned long) != 4) 888 return false; 889 if (size != 8) 890 return false; 891 if (off % size != 4) 892 return false; 893 } 894 895 switch (off) { 896 case bpf_ctx_range(struct bpf_perf_event_data, sample_period): 897 bpf_ctx_record_field_size(info, size_u64); 898 if (!bpf_ctx_narrow_access_ok(off, size, size_u64)) 899 return false; 900 break; 901 case bpf_ctx_range(struct bpf_perf_event_data, addr): 902 bpf_ctx_record_field_size(info, size_u64); 903 if (!bpf_ctx_narrow_access_ok(off, size, size_u64)) 904 return false; 905 break; 906 default: 907 if (size != sizeof(long)) 908 return false; 909 } 910 911 return true; 912 } 913 914 static u32 pe_prog_convert_ctx_access(enum bpf_access_type type, 915 const struct bpf_insn *si, 916 struct bpf_insn *insn_buf, 917 struct bpf_prog *prog, u32 *target_size) 918 { 919 struct bpf_insn *insn = insn_buf; 920 921 switch (si->off) { 922 case offsetof(struct bpf_perf_event_data, sample_period): 923 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct bpf_perf_event_data_kern, 924 data), si->dst_reg, si->src_reg, 925 offsetof(struct bpf_perf_event_data_kern, data)); 926 *insn++ = BPF_LDX_MEM(BPF_DW, si->dst_reg, si->dst_reg, 927 bpf_target_off(struct perf_sample_data, period, 8, 928 target_size)); 929 break; 930 case offsetof(struct bpf_perf_event_data, addr): 931 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct bpf_perf_event_data_kern, 932 data), si->dst_reg, si->src_reg, 933 offsetof(struct bpf_perf_event_data_kern, data)); 934 *insn++ = BPF_LDX_MEM(BPF_DW, si->dst_reg, si->dst_reg, 935 bpf_target_off(struct perf_sample_data, addr, 8, 936 target_size)); 937 break; 938 default: 939 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct bpf_perf_event_data_kern, 940 regs), si->dst_reg, si->src_reg, 941 offsetof(struct bpf_perf_event_data_kern, regs)); 942 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(long), si->dst_reg, si->dst_reg, 943 si->off); 944 break; 945 } 946 947 return insn - insn_buf; 948 } 949 950 const struct bpf_verifier_ops perf_event_verifier_ops = { 951 .get_func_proto = pe_prog_func_proto, 952 .is_valid_access = pe_prog_is_valid_access, 953 .convert_ctx_access = pe_prog_convert_ctx_access, 954 }; 955 956 const struct bpf_prog_ops perf_event_prog_ops = { 957 }; 958 959 static DEFINE_MUTEX(bpf_event_mutex); 960 961 #define BPF_TRACE_MAX_PROGS 64 962 963 int perf_event_attach_bpf_prog(struct perf_event *event, 964 struct bpf_prog *prog) 965 { 966 struct bpf_prog_array __rcu *old_array; 967 struct bpf_prog_array *new_array; 968 int ret = -EEXIST; 969 970 /* 971 * Kprobe override only works if they are on the function entry, 972 * and only if they are on the opt-in list. 973 */ 974 if (prog->kprobe_override && 975 (!trace_kprobe_on_func_entry(event->tp_event) || 976 !trace_kprobe_error_injectable(event->tp_event))) 977 return -EINVAL; 978 979 mutex_lock(&bpf_event_mutex); 980 981 if (event->prog) 982 goto unlock; 983 984 old_array = event->tp_event->prog_array; 985 if (old_array && 986 bpf_prog_array_length(old_array) >= BPF_TRACE_MAX_PROGS) { 987 ret = -E2BIG; 988 goto unlock; 989 } 990 991 ret = bpf_prog_array_copy(old_array, NULL, prog, &new_array); 992 if (ret < 0) 993 goto unlock; 994 995 /* set the new array to event->tp_event and set event->prog */ 996 event->prog = prog; 997 rcu_assign_pointer(event->tp_event->prog_array, new_array); 998 bpf_prog_array_free(old_array); 999 1000 unlock: 1001 mutex_unlock(&bpf_event_mutex); 1002 return ret; 1003 } 1004 1005 void perf_event_detach_bpf_prog(struct perf_event *event) 1006 { 1007 struct bpf_prog_array __rcu *old_array; 1008 struct bpf_prog_array *new_array; 1009 int ret; 1010 1011 mutex_lock(&bpf_event_mutex); 1012 1013 if (!event->prog) 1014 goto unlock; 1015 1016 old_array = event->tp_event->prog_array; 1017 ret = bpf_prog_array_copy(old_array, event->prog, NULL, &new_array); 1018 if (ret == -ENOENT) 1019 goto unlock; 1020 if (ret < 0) { 1021 bpf_prog_array_delete_safe(old_array, event->prog); 1022 } else { 1023 rcu_assign_pointer(event->tp_event->prog_array, new_array); 1024 bpf_prog_array_free(old_array); 1025 } 1026 1027 bpf_prog_put(event->prog); 1028 event->prog = NULL; 1029 1030 unlock: 1031 mutex_unlock(&bpf_event_mutex); 1032 } 1033 1034 int perf_event_query_prog_array(struct perf_event *event, void __user *info) 1035 { 1036 struct perf_event_query_bpf __user *uquery = info; 1037 struct perf_event_query_bpf query = {}; 1038 u32 *ids, prog_cnt, ids_len; 1039 int ret; 1040 1041 if (!capable(CAP_SYS_ADMIN)) 1042 return -EPERM; 1043 if (event->attr.type != PERF_TYPE_TRACEPOINT) 1044 return -EINVAL; 1045 if (copy_from_user(&query, uquery, sizeof(query))) 1046 return -EFAULT; 1047 1048 ids_len = query.ids_len; 1049 if (ids_len > BPF_TRACE_MAX_PROGS) 1050 return -E2BIG; 1051 ids = kcalloc(ids_len, sizeof(u32), GFP_USER | __GFP_NOWARN); 1052 if (!ids) 1053 return -ENOMEM; 1054 /* 1055 * The above kcalloc returns ZERO_SIZE_PTR when ids_len = 0, which 1056 * is required when user only wants to check for uquery->prog_cnt. 1057 * There is no need to check for it since the case is handled 1058 * gracefully in bpf_prog_array_copy_info. 1059 */ 1060 1061 mutex_lock(&bpf_event_mutex); 1062 ret = bpf_prog_array_copy_info(event->tp_event->prog_array, 1063 ids, 1064 ids_len, 1065 &prog_cnt); 1066 mutex_unlock(&bpf_event_mutex); 1067 1068 if (copy_to_user(&uquery->prog_cnt, &prog_cnt, sizeof(prog_cnt)) || 1069 copy_to_user(uquery->ids, ids, ids_len * sizeof(u32))) 1070 ret = -EFAULT; 1071 1072 kfree(ids); 1073 return ret; 1074 } 1075 1076 extern struct bpf_raw_event_map __start__bpf_raw_tp[]; 1077 extern struct bpf_raw_event_map __stop__bpf_raw_tp[]; 1078 1079 struct bpf_raw_event_map *bpf_find_raw_tracepoint(const char *name) 1080 { 1081 struct bpf_raw_event_map *btp = __start__bpf_raw_tp; 1082 1083 for (; btp < __stop__bpf_raw_tp; btp++) { 1084 if (!strcmp(btp->tp->name, name)) 1085 return btp; 1086 } 1087 return NULL; 1088 } 1089 1090 static __always_inline 1091 void __bpf_trace_run(struct bpf_prog *prog, u64 *args) 1092 { 1093 rcu_read_lock(); 1094 preempt_disable(); 1095 (void) BPF_PROG_RUN(prog, args); 1096 preempt_enable(); 1097 rcu_read_unlock(); 1098 } 1099 1100 #define UNPACK(...) __VA_ARGS__ 1101 #define REPEAT_1(FN, DL, X, ...) FN(X) 1102 #define REPEAT_2(FN, DL, X, ...) FN(X) UNPACK DL REPEAT_1(FN, DL, __VA_ARGS__) 1103 #define REPEAT_3(FN, DL, X, ...) FN(X) UNPACK DL REPEAT_2(FN, DL, __VA_ARGS__) 1104 #define REPEAT_4(FN, DL, X, ...) FN(X) UNPACK DL REPEAT_3(FN, DL, __VA_ARGS__) 1105 #define REPEAT_5(FN, DL, X, ...) FN(X) UNPACK DL REPEAT_4(FN, DL, __VA_ARGS__) 1106 #define REPEAT_6(FN, DL, X, ...) FN(X) UNPACK DL REPEAT_5(FN, DL, __VA_ARGS__) 1107 #define REPEAT_7(FN, DL, X, ...) FN(X) UNPACK DL REPEAT_6(FN, DL, __VA_ARGS__) 1108 #define REPEAT_8(FN, DL, X, ...) FN(X) UNPACK DL REPEAT_7(FN, DL, __VA_ARGS__) 1109 #define REPEAT_9(FN, DL, X, ...) FN(X) UNPACK DL REPEAT_8(FN, DL, __VA_ARGS__) 1110 #define REPEAT_10(FN, DL, X, ...) FN(X) UNPACK DL REPEAT_9(FN, DL, __VA_ARGS__) 1111 #define REPEAT_11(FN, DL, X, ...) FN(X) UNPACK DL REPEAT_10(FN, DL, __VA_ARGS__) 1112 #define REPEAT_12(FN, DL, X, ...) FN(X) UNPACK DL REPEAT_11(FN, DL, __VA_ARGS__) 1113 #define REPEAT(X, FN, DL, ...) REPEAT_##X(FN, DL, __VA_ARGS__) 1114 1115 #define SARG(X) u64 arg##X 1116 #define COPY(X) args[X] = arg##X 1117 1118 #define __DL_COM (,) 1119 #define __DL_SEM (;) 1120 1121 #define __SEQ_0_11 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11 1122 1123 #define BPF_TRACE_DEFN_x(x) \ 1124 void bpf_trace_run##x(struct bpf_prog *prog, \ 1125 REPEAT(x, SARG, __DL_COM, __SEQ_0_11)) \ 1126 { \ 1127 u64 args[x]; \ 1128 REPEAT(x, COPY, __DL_SEM, __SEQ_0_11); \ 1129 __bpf_trace_run(prog, args); \ 1130 } \ 1131 EXPORT_SYMBOL_GPL(bpf_trace_run##x) 1132 BPF_TRACE_DEFN_x(1); 1133 BPF_TRACE_DEFN_x(2); 1134 BPF_TRACE_DEFN_x(3); 1135 BPF_TRACE_DEFN_x(4); 1136 BPF_TRACE_DEFN_x(5); 1137 BPF_TRACE_DEFN_x(6); 1138 BPF_TRACE_DEFN_x(7); 1139 BPF_TRACE_DEFN_x(8); 1140 BPF_TRACE_DEFN_x(9); 1141 BPF_TRACE_DEFN_x(10); 1142 BPF_TRACE_DEFN_x(11); 1143 BPF_TRACE_DEFN_x(12); 1144 1145 static int __bpf_probe_register(struct bpf_raw_event_map *btp, struct bpf_prog *prog) 1146 { 1147 struct tracepoint *tp = btp->tp; 1148 1149 /* 1150 * check that program doesn't access arguments beyond what's 1151 * available in this tracepoint 1152 */ 1153 if (prog->aux->max_ctx_offset > btp->num_args * sizeof(u64)) 1154 return -EINVAL; 1155 1156 return tracepoint_probe_register(tp, (void *)btp->bpf_func, prog); 1157 } 1158 1159 int bpf_probe_register(struct bpf_raw_event_map *btp, struct bpf_prog *prog) 1160 { 1161 int err; 1162 1163 mutex_lock(&bpf_event_mutex); 1164 err = __bpf_probe_register(btp, prog); 1165 mutex_unlock(&bpf_event_mutex); 1166 return err; 1167 } 1168 1169 int bpf_probe_unregister(struct bpf_raw_event_map *btp, struct bpf_prog *prog) 1170 { 1171 int err; 1172 1173 mutex_lock(&bpf_event_mutex); 1174 err = tracepoint_probe_unregister(btp->tp, (void *)btp->bpf_func, prog); 1175 mutex_unlock(&bpf_event_mutex); 1176 return err; 1177 } 1178 1179 int bpf_get_perf_event_info(const struct perf_event *event, u32 *prog_id, 1180 u32 *fd_type, const char **buf, 1181 u64 *probe_offset, u64 *probe_addr) 1182 { 1183 bool is_tracepoint, is_syscall_tp; 1184 struct bpf_prog *prog; 1185 int flags, err = 0; 1186 1187 prog = event->prog; 1188 if (!prog) 1189 return -ENOENT; 1190 1191 /* not supporting BPF_PROG_TYPE_PERF_EVENT yet */ 1192 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) 1193 return -EOPNOTSUPP; 1194 1195 *prog_id = prog->aux->id; 1196 flags = event->tp_event->flags; 1197 is_tracepoint = flags & TRACE_EVENT_FL_TRACEPOINT; 1198 is_syscall_tp = is_syscall_trace_event(event->tp_event); 1199 1200 if (is_tracepoint || is_syscall_tp) { 1201 *buf = is_tracepoint ? event->tp_event->tp->name 1202 : event->tp_event->name; 1203 *fd_type = BPF_FD_TYPE_TRACEPOINT; 1204 *probe_offset = 0x0; 1205 *probe_addr = 0x0; 1206 } else { 1207 /* kprobe/uprobe */ 1208 err = -EOPNOTSUPP; 1209 #ifdef CONFIG_KPROBE_EVENTS 1210 if (flags & TRACE_EVENT_FL_KPROBE) 1211 err = bpf_get_kprobe_info(event, fd_type, buf, 1212 probe_offset, probe_addr, 1213 event->attr.type == PERF_TYPE_TRACEPOINT); 1214 #endif 1215 #ifdef CONFIG_UPROBE_EVENTS 1216 if (flags & TRACE_EVENT_FL_UPROBE) 1217 err = bpf_get_uprobe_info(event, fd_type, buf, 1218 probe_offset, 1219 event->attr.type == PERF_TYPE_TRACEPOINT); 1220 #endif 1221 } 1222 1223 return err; 1224 } 1225