1 /* Copyright (c) 2011-2015 PLUMgrid, http://plumgrid.com
2 * Copyright (c) 2016 Facebook
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of version 2 of the GNU General Public
6 * License as published by the Free Software Foundation.
8 #include <linux/kernel.h>
9 #include <linux/types.h>
10 #include <linux/slab.h>
11 #include <linux/bpf.h>
12 #include <linux/bpf_perf_event.h>
13 #include <linux/filter.h>
14 #include <linux/uaccess.h>
15 #include <linux/ctype.h>
18 u64 bpf_get_stackid(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5);
21 * trace_call_bpf - invoke BPF program
22 * @call: tracepoint event
23 * @ctx: opaque context pointer
25 * kprobe handlers execute BPF programs via this helper.
26 * Can be used from static tracepoints in the future.
28 * Return: BPF programs always return an integer which is interpreted by
30 * 0 - return from kprobe (event is filtered out)
31 * 1 - store kprobe event into ring buffer
32 * Other values are reserved and currently alias to 1
34 unsigned int trace_call_bpf(struct trace_event_call *call, void *ctx)
38 if (in_nmi()) /* not supported yet */
43 if (unlikely(__this_cpu_inc_return(bpf_prog_active) != 1)) {
45 * since some bpf program is already running on this cpu,
46 * don't call into another bpf program (same or different)
47 * and don't send kprobe event into ring-buffer,
55 * Instead of moving rcu_read_lock/rcu_dereference/rcu_read_unlock
56 * to all call sites, we did a bpf_prog_array_valid() there to check
57 * whether call->prog_array is empty or not, which is
58 * a heurisitc to speed up execution.
60 * If bpf_prog_array_valid() fetched prog_array was
61 * non-NULL, we go into trace_call_bpf() and do the actual
62 * proper rcu_dereference() under RCU lock.
63 * If it turns out that prog_array is NULL then, we bail out.
64 * For the opposite, if the bpf_prog_array_valid() fetched pointer
65 * was NULL, you'll skip the prog_array with the risk of missing
66 * out of events when it was updated in between this and the
67 * rcu_dereference() which is accepted risk.
69 ret = BPF_PROG_RUN_ARRAY_CHECK(call->prog_array, ctx, BPF_PROG_RUN);
72 __this_cpu_dec(bpf_prog_active);
77 EXPORT_SYMBOL_GPL(trace_call_bpf);
79 BPF_CALL_3(bpf_probe_read, void *, dst, u32, size, const void *, unsafe_ptr)
83 ret = probe_kernel_read(dst, unsafe_ptr, size);
84 if (unlikely(ret < 0))
90 static const struct bpf_func_proto bpf_probe_read_proto = {
91 .func = bpf_probe_read,
93 .ret_type = RET_INTEGER,
94 .arg1_type = ARG_PTR_TO_UNINIT_MEM,
95 .arg2_type = ARG_CONST_SIZE_OR_ZERO,
96 .arg3_type = ARG_ANYTHING,
99 BPF_CALL_3(bpf_probe_write_user, void *, unsafe_ptr, const void *, src,
103 * Ensure we're in user context which is safe for the helper to
104 * run. This helper has no business in a kthread.
106 * access_ok() should prevent writing to non-user memory, but in
107 * some situations (nommu, temporary switch, etc) access_ok() does
108 * not provide enough validation, hence the check on KERNEL_DS.
111 if (unlikely(in_interrupt() ||
112 current->flags & (PF_KTHREAD | PF_EXITING)))
114 if (unlikely(uaccess_kernel()))
116 if (!access_ok(VERIFY_WRITE, unsafe_ptr, size))
119 return probe_kernel_write(unsafe_ptr, src, size);
122 static const struct bpf_func_proto bpf_probe_write_user_proto = {
123 .func = bpf_probe_write_user,
125 .ret_type = RET_INTEGER,
126 .arg1_type = ARG_ANYTHING,
127 .arg2_type = ARG_PTR_TO_MEM,
128 .arg3_type = ARG_CONST_SIZE,
131 static const struct bpf_func_proto *bpf_get_probe_write_proto(void)
133 pr_warn_ratelimited("%s[%d] is installing a program with bpf_probe_write_user helper that may corrupt user memory!",
134 current->comm, task_pid_nr(current));
136 return &bpf_probe_write_user_proto;
140 * Only limited trace_printk() conversion specifiers allowed:
141 * %d %i %u %x %ld %li %lu %lx %lld %lli %llu %llx %p %s
143 BPF_CALL_5(bpf_trace_printk, char *, fmt, u32, fmt_size, u64, arg1,
144 u64, arg2, u64, arg3)
146 bool str_seen = false;
154 * bpf_check()->check_func_arg()->check_stack_boundary()
155 * guarantees that fmt points to bpf program stack,
156 * fmt_size bytes of it were initialized and fmt_size > 0
158 if (fmt[--fmt_size] != 0)
161 /* check format string for allowed specifiers */
162 for (i = 0; i < fmt_size; i++) {
163 if ((!isprint(fmt[i]) && !isspace(fmt[i])) || !isascii(fmt[i]))
172 /* fmt[i] != 0 && fmt[last] == 0, so we can access fmt[i + 1] */
177 } else if (fmt[i] == 'p' || fmt[i] == 's') {
180 if (!isspace(fmt[i]) && !ispunct(fmt[i]) && fmt[i] != 0)
183 if (fmt[i - 1] == 's') {
185 /* allow only one '%s' per fmt string */
204 strncpy_from_unsafe(buf,
205 (void *) (long) unsafe_addr,
216 if (fmt[i] != 'i' && fmt[i] != 'd' &&
217 fmt[i] != 'u' && fmt[i] != 'x')
222 /* Horrid workaround for getting va_list handling working with different
223 * argument type combinations generically for 32 and 64 bit archs.
225 #define __BPF_TP_EMIT() __BPF_ARG3_TP()
226 #define __BPF_TP(...) \
227 __trace_printk(1 /* Fake ip will not be printed. */, \
230 #define __BPF_ARG1_TP(...) \
231 ((mod[0] == 2 || (mod[0] == 1 && __BITS_PER_LONG == 64)) \
232 ? __BPF_TP(arg1, ##__VA_ARGS__) \
233 : ((mod[0] == 1 || (mod[0] == 0 && __BITS_PER_LONG == 32)) \
234 ? __BPF_TP((long)arg1, ##__VA_ARGS__) \
235 : __BPF_TP((u32)arg1, ##__VA_ARGS__)))
237 #define __BPF_ARG2_TP(...) \
238 ((mod[1] == 2 || (mod[1] == 1 && __BITS_PER_LONG == 64)) \
239 ? __BPF_ARG1_TP(arg2, ##__VA_ARGS__) \
240 : ((mod[1] == 1 || (mod[1] == 0 && __BITS_PER_LONG == 32)) \
241 ? __BPF_ARG1_TP((long)arg2, ##__VA_ARGS__) \
242 : __BPF_ARG1_TP((u32)arg2, ##__VA_ARGS__)))
244 #define __BPF_ARG3_TP(...) \
245 ((mod[2] == 2 || (mod[2] == 1 && __BITS_PER_LONG == 64)) \
246 ? __BPF_ARG2_TP(arg3, ##__VA_ARGS__) \
247 : ((mod[2] == 1 || (mod[2] == 0 && __BITS_PER_LONG == 32)) \
248 ? __BPF_ARG2_TP((long)arg3, ##__VA_ARGS__) \
249 : __BPF_ARG2_TP((u32)arg3, ##__VA_ARGS__)))
251 return __BPF_TP_EMIT();
254 static const struct bpf_func_proto bpf_trace_printk_proto = {
255 .func = bpf_trace_printk,
257 .ret_type = RET_INTEGER,
258 .arg1_type = ARG_PTR_TO_MEM,
259 .arg2_type = ARG_CONST_SIZE,
262 const struct bpf_func_proto *bpf_get_trace_printk_proto(void)
265 * this program might be calling bpf_trace_printk,
266 * so allocate per-cpu printk buffers
268 trace_printk_init_buffers();
270 return &bpf_trace_printk_proto;
273 static __always_inline int
274 get_map_perf_counter(struct bpf_map *map, u64 flags,
275 u64 *value, u64 *enabled, u64 *running)
277 struct bpf_array *array = container_of(map, struct bpf_array, map);
278 unsigned int cpu = smp_processor_id();
279 u64 index = flags & BPF_F_INDEX_MASK;
280 struct bpf_event_entry *ee;
282 if (unlikely(flags & ~(BPF_F_INDEX_MASK)))
284 if (index == BPF_F_CURRENT_CPU)
286 if (unlikely(index >= array->map.max_entries))
289 ee = READ_ONCE(array->ptrs[index]);
293 return perf_event_read_local(ee->event, value, enabled, running);
296 BPF_CALL_2(bpf_perf_event_read, struct bpf_map *, map, u64, flags)
301 err = get_map_perf_counter(map, flags, &value, NULL, NULL);
303 * this api is ugly since we miss [-22..-2] range of valid
304 * counter values, but that's uapi
311 static const struct bpf_func_proto bpf_perf_event_read_proto = {
312 .func = bpf_perf_event_read,
314 .ret_type = RET_INTEGER,
315 .arg1_type = ARG_CONST_MAP_PTR,
316 .arg2_type = ARG_ANYTHING,
319 BPF_CALL_4(bpf_perf_event_read_value, struct bpf_map *, map, u64, flags,
320 struct bpf_perf_event_value *, buf, u32, size)
324 if (unlikely(size != sizeof(struct bpf_perf_event_value)))
326 err = get_map_perf_counter(map, flags, &buf->counter, &buf->enabled,
332 memset(buf, 0, size);
336 static const struct bpf_func_proto bpf_perf_event_read_value_proto = {
337 .func = bpf_perf_event_read_value,
339 .ret_type = RET_INTEGER,
340 .arg1_type = ARG_CONST_MAP_PTR,
341 .arg2_type = ARG_ANYTHING,
342 .arg3_type = ARG_PTR_TO_UNINIT_MEM,
343 .arg4_type = ARG_CONST_SIZE,
346 static DEFINE_PER_CPU(struct perf_sample_data, bpf_sd);
348 static __always_inline u64
349 __bpf_perf_event_output(struct pt_regs *regs, struct bpf_map *map,
350 u64 flags, struct perf_raw_record *raw)
352 struct bpf_array *array = container_of(map, struct bpf_array, map);
353 struct perf_sample_data *sd = this_cpu_ptr(&bpf_sd);
354 unsigned int cpu = smp_processor_id();
355 u64 index = flags & BPF_F_INDEX_MASK;
356 struct bpf_event_entry *ee;
357 struct perf_event *event;
359 if (index == BPF_F_CURRENT_CPU)
361 if (unlikely(index >= array->map.max_entries))
364 ee = READ_ONCE(array->ptrs[index]);
369 if (unlikely(event->attr.type != PERF_TYPE_SOFTWARE ||
370 event->attr.config != PERF_COUNT_SW_BPF_OUTPUT))
373 if (unlikely(event->oncpu != cpu))
376 perf_sample_data_init(sd, 0, 0);
378 perf_event_output(event, sd, regs);
382 BPF_CALL_5(bpf_perf_event_output, struct pt_regs *, regs, struct bpf_map *, map,
383 u64, flags, void *, data, u64, size)
385 struct perf_raw_record raw = {
392 if (unlikely(flags & ~(BPF_F_INDEX_MASK)))
395 return __bpf_perf_event_output(regs, map, flags, &raw);
398 static const struct bpf_func_proto bpf_perf_event_output_proto = {
399 .func = bpf_perf_event_output,
401 .ret_type = RET_INTEGER,
402 .arg1_type = ARG_PTR_TO_CTX,
403 .arg2_type = ARG_CONST_MAP_PTR,
404 .arg3_type = ARG_ANYTHING,
405 .arg4_type = ARG_PTR_TO_MEM,
406 .arg5_type = ARG_CONST_SIZE_OR_ZERO,
409 static DEFINE_PER_CPU(struct pt_regs, bpf_pt_regs);
411 u64 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
412 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
414 struct pt_regs *regs = this_cpu_ptr(&bpf_pt_regs);
415 struct perf_raw_frag frag = {
420 struct perf_raw_record raw = {
423 .next = ctx_size ? &frag : NULL,
430 perf_fetch_caller_regs(regs);
432 return __bpf_perf_event_output(regs, map, flags, &raw);
435 BPF_CALL_0(bpf_get_current_task)
437 return (long) current;
440 static const struct bpf_func_proto bpf_get_current_task_proto = {
441 .func = bpf_get_current_task,
443 .ret_type = RET_INTEGER,
446 BPF_CALL_2(bpf_current_task_under_cgroup, struct bpf_map *, map, u32, idx)
448 struct bpf_array *array = container_of(map, struct bpf_array, map);
451 if (unlikely(in_interrupt()))
453 if (unlikely(idx >= array->map.max_entries))
456 cgrp = READ_ONCE(array->ptrs[idx]);
460 return task_under_cgroup_hierarchy(current, cgrp);
463 static const struct bpf_func_proto bpf_current_task_under_cgroup_proto = {
464 .func = bpf_current_task_under_cgroup,
466 .ret_type = RET_INTEGER,
467 .arg1_type = ARG_CONST_MAP_PTR,
468 .arg2_type = ARG_ANYTHING,
471 BPF_CALL_3(bpf_probe_read_str, void *, dst, u32, size,
472 const void *, unsafe_ptr)
477 * The strncpy_from_unsafe() call will likely not fill the entire
478 * buffer, but that's okay in this circumstance as we're probing
479 * arbitrary memory anyway similar to bpf_probe_read() and might
480 * as well probe the stack. Thus, memory is explicitly cleared
481 * only in error case, so that improper users ignoring return
482 * code altogether don't copy garbage; otherwise length of string
483 * is returned that can be used for bpf_perf_event_output() et al.
485 ret = strncpy_from_unsafe(dst, unsafe_ptr, size);
486 if (unlikely(ret < 0))
487 memset(dst, 0, size);
492 static const struct bpf_func_proto bpf_probe_read_str_proto = {
493 .func = bpf_probe_read_str,
495 .ret_type = RET_INTEGER,
496 .arg1_type = ARG_PTR_TO_UNINIT_MEM,
497 .arg2_type = ARG_CONST_SIZE_OR_ZERO,
498 .arg3_type = ARG_ANYTHING,
501 static const struct bpf_func_proto *tracing_func_proto(enum bpf_func_id func_id)
504 case BPF_FUNC_map_lookup_elem:
505 return &bpf_map_lookup_elem_proto;
506 case BPF_FUNC_map_update_elem:
507 return &bpf_map_update_elem_proto;
508 case BPF_FUNC_map_delete_elem:
509 return &bpf_map_delete_elem_proto;
510 case BPF_FUNC_probe_read:
511 return &bpf_probe_read_proto;
512 case BPF_FUNC_ktime_get_ns:
513 return &bpf_ktime_get_ns_proto;
514 case BPF_FUNC_tail_call:
515 return &bpf_tail_call_proto;
516 case BPF_FUNC_get_current_pid_tgid:
517 return &bpf_get_current_pid_tgid_proto;
518 case BPF_FUNC_get_current_task:
519 return &bpf_get_current_task_proto;
520 case BPF_FUNC_get_current_uid_gid:
521 return &bpf_get_current_uid_gid_proto;
522 case BPF_FUNC_get_current_comm:
523 return &bpf_get_current_comm_proto;
524 case BPF_FUNC_trace_printk:
525 return bpf_get_trace_printk_proto();
526 case BPF_FUNC_get_smp_processor_id:
527 return &bpf_get_smp_processor_id_proto;
528 case BPF_FUNC_get_numa_node_id:
529 return &bpf_get_numa_node_id_proto;
530 case BPF_FUNC_perf_event_read:
531 return &bpf_perf_event_read_proto;
532 case BPF_FUNC_probe_write_user:
533 return bpf_get_probe_write_proto();
534 case BPF_FUNC_current_task_under_cgroup:
535 return &bpf_current_task_under_cgroup_proto;
536 case BPF_FUNC_get_prandom_u32:
537 return &bpf_get_prandom_u32_proto;
538 case BPF_FUNC_probe_read_str:
539 return &bpf_probe_read_str_proto;
545 static const struct bpf_func_proto *kprobe_prog_func_proto(enum bpf_func_id func_id)
548 case BPF_FUNC_perf_event_output:
549 return &bpf_perf_event_output_proto;
550 case BPF_FUNC_get_stackid:
551 return &bpf_get_stackid_proto;
552 case BPF_FUNC_perf_event_read_value:
553 return &bpf_perf_event_read_value_proto;
555 return tracing_func_proto(func_id);
559 /* bpf+kprobe programs can access fields of 'struct pt_regs' */
560 static bool kprobe_prog_is_valid_access(int off, int size, enum bpf_access_type type,
561 struct bpf_insn_access_aux *info)
563 if (off < 0 || off >= sizeof(struct pt_regs))
565 if (type != BPF_READ)
570 * Assertion for 32 bit to make sure last 8 byte access
571 * (BPF_DW) to the last 4 byte member is disallowed.
573 if (off + size > sizeof(struct pt_regs))
579 const struct bpf_verifier_ops kprobe_verifier_ops = {
580 .get_func_proto = kprobe_prog_func_proto,
581 .is_valid_access = kprobe_prog_is_valid_access,
584 const struct bpf_prog_ops kprobe_prog_ops = {
587 BPF_CALL_5(bpf_perf_event_output_tp, void *, tp_buff, struct bpf_map *, map,
588 u64, flags, void *, data, u64, size)
590 struct pt_regs *regs = *(struct pt_regs **)tp_buff;
593 * r1 points to perf tracepoint buffer where first 8 bytes are hidden
594 * from bpf program and contain a pointer to 'struct pt_regs'. Fetch it
595 * from there and call the same bpf_perf_event_output() helper inline.
597 return ____bpf_perf_event_output(regs, map, flags, data, size);
600 static const struct bpf_func_proto bpf_perf_event_output_proto_tp = {
601 .func = bpf_perf_event_output_tp,
603 .ret_type = RET_INTEGER,
604 .arg1_type = ARG_PTR_TO_CTX,
605 .arg2_type = ARG_CONST_MAP_PTR,
606 .arg3_type = ARG_ANYTHING,
607 .arg4_type = ARG_PTR_TO_MEM,
608 .arg5_type = ARG_CONST_SIZE_OR_ZERO,
611 BPF_CALL_3(bpf_get_stackid_tp, void *, tp_buff, struct bpf_map *, map,
614 struct pt_regs *regs = *(struct pt_regs **)tp_buff;
617 * Same comment as in bpf_perf_event_output_tp(), only that this time
618 * the other helper's function body cannot be inlined due to being
619 * external, thus we need to call raw helper function.
621 return bpf_get_stackid((unsigned long) regs, (unsigned long) map,
625 static const struct bpf_func_proto bpf_get_stackid_proto_tp = {
626 .func = bpf_get_stackid_tp,
628 .ret_type = RET_INTEGER,
629 .arg1_type = ARG_PTR_TO_CTX,
630 .arg2_type = ARG_CONST_MAP_PTR,
631 .arg3_type = ARG_ANYTHING,
634 BPF_CALL_3(bpf_perf_prog_read_value_tp, struct bpf_perf_event_data_kern *, ctx,
635 struct bpf_perf_event_value *, buf, u32, size)
639 if (unlikely(size != sizeof(struct bpf_perf_event_value)))
641 err = perf_event_read_local(ctx->event, &buf->counter, &buf->enabled,
647 memset(buf, 0, size);
651 static const struct bpf_func_proto bpf_perf_prog_read_value_proto_tp = {
652 .func = bpf_perf_prog_read_value_tp,
654 .ret_type = RET_INTEGER,
655 .arg1_type = ARG_PTR_TO_CTX,
656 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
657 .arg3_type = ARG_CONST_SIZE,
660 static const struct bpf_func_proto *tp_prog_func_proto(enum bpf_func_id func_id)
663 case BPF_FUNC_perf_event_output:
664 return &bpf_perf_event_output_proto_tp;
665 case BPF_FUNC_get_stackid:
666 return &bpf_get_stackid_proto_tp;
667 case BPF_FUNC_perf_prog_read_value:
668 return &bpf_perf_prog_read_value_proto_tp;
670 return tracing_func_proto(func_id);
674 static bool tp_prog_is_valid_access(int off, int size, enum bpf_access_type type,
675 struct bpf_insn_access_aux *info)
677 if (off < sizeof(void *) || off >= PERF_MAX_TRACE_SIZE)
679 if (type != BPF_READ)
684 BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(__u64));
688 const struct bpf_verifier_ops tracepoint_verifier_ops = {
689 .get_func_proto = tp_prog_func_proto,
690 .is_valid_access = tp_prog_is_valid_access,
693 const struct bpf_prog_ops tracepoint_prog_ops = {
696 static bool pe_prog_is_valid_access(int off, int size, enum bpf_access_type type,
697 struct bpf_insn_access_aux *info)
699 const int size_sp = FIELD_SIZEOF(struct bpf_perf_event_data,
702 if (off < 0 || off >= sizeof(struct bpf_perf_event_data))
704 if (type != BPF_READ)
710 case bpf_ctx_range(struct bpf_perf_event_data, sample_period):
711 bpf_ctx_record_field_size(info, size_sp);
712 if (!bpf_ctx_narrow_access_ok(off, size, size_sp))
716 if (size != sizeof(long))
723 static u32 pe_prog_convert_ctx_access(enum bpf_access_type type,
724 const struct bpf_insn *si,
725 struct bpf_insn *insn_buf,
726 struct bpf_prog *prog, u32 *target_size)
728 struct bpf_insn *insn = insn_buf;
731 case offsetof(struct bpf_perf_event_data, sample_period):
732 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct bpf_perf_event_data_kern,
733 data), si->dst_reg, si->src_reg,
734 offsetof(struct bpf_perf_event_data_kern, data));
735 *insn++ = BPF_LDX_MEM(BPF_DW, si->dst_reg, si->dst_reg,
736 bpf_target_off(struct perf_sample_data, period, 8,
740 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct bpf_perf_event_data_kern,
741 regs), si->dst_reg, si->src_reg,
742 offsetof(struct bpf_perf_event_data_kern, regs));
743 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(long), si->dst_reg, si->dst_reg,
748 return insn - insn_buf;
751 const struct bpf_verifier_ops perf_event_verifier_ops = {
752 .get_func_proto = tp_prog_func_proto,
753 .is_valid_access = pe_prog_is_valid_access,
754 .convert_ctx_access = pe_prog_convert_ctx_access,
757 const struct bpf_prog_ops perf_event_prog_ops = {
760 static DEFINE_MUTEX(bpf_event_mutex);
762 int perf_event_attach_bpf_prog(struct perf_event *event,
763 struct bpf_prog *prog)
765 struct bpf_prog_array __rcu *old_array;
766 struct bpf_prog_array *new_array;
769 mutex_lock(&bpf_event_mutex);
774 old_array = event->tp_event->prog_array;
775 ret = bpf_prog_array_copy(old_array, NULL, prog, &new_array);
779 /* set the new array to event->tp_event and set event->prog */
781 rcu_assign_pointer(event->tp_event->prog_array, new_array);
782 bpf_prog_array_free(old_array);
785 mutex_unlock(&bpf_event_mutex);
789 void perf_event_detach_bpf_prog(struct perf_event *event)
791 struct bpf_prog_array __rcu *old_array;
792 struct bpf_prog_array *new_array;
795 mutex_lock(&bpf_event_mutex);
800 old_array = event->tp_event->prog_array;
801 ret = bpf_prog_array_copy(old_array, event->prog, NULL, &new_array);
803 bpf_prog_array_delete_safe(old_array, event->prog);
805 rcu_assign_pointer(event->tp_event->prog_array, new_array);
806 bpf_prog_array_free(old_array);
809 bpf_prog_put(event->prog);
813 mutex_unlock(&bpf_event_mutex);