kernel/trace/bpf_trace.c

   1 /* Copyright (c) 2011-2015 PLUMgrid, http://plumgrid.com
   2  * Copyright (c) 2016 Facebook
   3  *
   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.
   7  */
   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>
  16 #include <linux/kprobes.h>
  17 #include <linux/error-injection.h>
  18
  19 #include "trace_probe.h"
  20 #include "trace.h"
  21
  22 u64 bpf_get_stackid(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5);
  23
  24 /**
  25  * trace_call_bpf - invoke BPF program
  26  * @call: tracepoint event
  27  * @ctx: opaque context pointer
  28  *
  29  * kprobe handlers execute BPF programs via this helper.
  30  * Can be used from static tracepoints in the future.
  31  *
  32  * Return: BPF programs always return an integer which is interpreted by
  33  * kprobe handler as:
  34  * 0 - return from kprobe (event is filtered out)
  35  * 1 - store kprobe event into ring buffer
  36  * Other values are reserved and currently alias to 1
  37  */
  38 unsigned int trace_call_bpf(struct trace_event_call *call, void *ctx)
  39 {
  40         unsigned int ret;
  41
  42         if (in_nmi()) /* not supported yet */
  43                 return 1;
  44
  45         preempt_disable();
  46
  47         if (unlikely(__this_cpu_inc_return(bpf_prog_active) != 1)) {
  48                 /*
  49                  * since some bpf program is already running on this cpu,
  50                  * don't call into another bpf program (same or different)
  51                  * and don't send kprobe event into ring-buffer,
  52                  * so return zero here
  53                  */
  54                 ret = 0;
  55                 goto out;
  56         }
  57
  58         /*
  59          * Instead of moving rcu_read_lock/rcu_dereference/rcu_read_unlock
  60          * to all call sites, we did a bpf_prog_array_valid() there to check
  61          * whether call->prog_array is empty or not, which is
  62          * a heurisitc to speed up execution.
  63          *
  64          * If bpf_prog_array_valid() fetched prog_array was
  65          * non-NULL, we go into trace_call_bpf() and do the actual
  66          * proper rcu_dereference() under RCU lock.
  67          * If it turns out that prog_array is NULL then, we bail out.
  68          * For the opposite, if the bpf_prog_array_valid() fetched pointer
  69          * was NULL, you'll skip the prog_array with the risk of missing
  70          * out of events when it was updated in between this and the
  71          * rcu_dereference() which is accepted risk.
  72          */
  73         ret = BPF_PROG_RUN_ARRAY_CHECK(call->prog_array, ctx, BPF_PROG_RUN);
  74
  75  out:
  76         __this_cpu_dec(bpf_prog_active);
  77         preempt_enable();
  78
  79         return ret;
  80 }
  81 EXPORT_SYMBOL_GPL(trace_call_bpf);
  82
  83 #ifdef CONFIG_BPF_KPROBE_OVERRIDE
  84 BPF_CALL_2(bpf_override_return, struct pt_regs *, regs, unsigned long, rc)
  85 {
  86         regs_set_return_value(regs, rc);
  87         override_function_with_return(regs);
  88         return 0;
  89 }
  90
  91 static const struct bpf_func_proto bpf_override_return_proto = {
  92         .func           = bpf_override_return,
  93         .gpl_only       = true,
  94         .ret_type       = RET_INTEGER,
  95         .arg1_type      = ARG_PTR_TO_CTX,
  96         .arg2_type      = ARG_ANYTHING,
  97 };
  98 #endif
  99
 100 BPF_CALL_3(bpf_probe_read, void *, dst, u32, size, const void *, unsafe_ptr)
 101 {
 102         int ret;
 103
 104         ret = probe_kernel_read(dst, unsafe_ptr, size);
 105         if (unlikely(ret < 0))
 106                 memset(dst, 0, size);
 107
 108         return ret;
 109 }
 110
 111 static const struct bpf_func_proto bpf_probe_read_proto = {
 112         .func           = bpf_probe_read,
 113         .gpl_only       = true,
 114         .ret_type       = RET_INTEGER,
 115         .arg1_type      = ARG_PTR_TO_UNINIT_MEM,
 116         .arg2_type      = ARG_CONST_SIZE_OR_ZERO,
 117         .arg3_type      = ARG_ANYTHING,
 118 };
 119
 120 BPF_CALL_3(bpf_probe_write_user, void *, unsafe_ptr, const void *, src,
 121            u32, size)
 122 {
 123         /*
 124          * Ensure we're in user context which is safe for the helper to
 125          * run. This helper has no business in a kthread.
 126          *
 127          * access_ok() should prevent writing to non-user memory, but in
 128          * some situations (nommu, temporary switch, etc) access_ok() does
 129          * not provide enough validation, hence the check on KERNEL_DS.
 130          */
 131
 132         if (unlikely(in_interrupt() ||
 133                      current->flags & (PF_KTHREAD | PF_EXITING)))
 134                 return -EPERM;
 135         if (unlikely(uaccess_kernel()))
 136                 return -EPERM;
 137         if (!access_ok(VERIFY_WRITE, unsafe_ptr, size))
 138                 return -EPERM;
 139
 140         return probe_kernel_write(unsafe_ptr, src, size);
 141 }
 142
 143 static const struct bpf_func_proto bpf_probe_write_user_proto = {
 144         .func           = bpf_probe_write_user,
 145         .gpl_only       = true,
 146         .ret_type       = RET_INTEGER,
 147         .arg1_type      = ARG_ANYTHING,
 148         .arg2_type      = ARG_PTR_TO_MEM,
 149         .arg3_type      = ARG_CONST_SIZE,
 150 };
 151
 152 static const struct bpf_func_proto *bpf_get_probe_write_proto(void)
 153 {
 154         pr_warn_ratelimited("%s[%d] is installing a program with bpf_probe_write_user helper that may corrupt user memory!",
 155                             current->comm, task_pid_nr(current));
 156
 157         return &bpf_probe_write_user_proto;
 158 }
 159
 160 /*
 161  * Only limited trace_printk() conversion specifiers allowed:
 162  * %d %i %u %x %ld %li %lu %lx %lld %lli %llu %llx %p %s
 163  */
 164 BPF_CALL_5(bpf_trace_printk, char *, fmt, u32, fmt_size, u64, arg1,
 165            u64, arg2, u64, arg3)
 166 {
 167         bool str_seen = false;
 168         int mod[3] = {};
 169         int fmt_cnt = 0;
 170         u64 unsafe_addr;
 171         char buf[64];
 172         int i;
 173
 174         /*
 175          * bpf_check()->check_func_arg()->check_stack_boundary()
 176          * guarantees that fmt points to bpf program stack,
 177          * fmt_size bytes of it were initialized and fmt_size > 0
 178          */
 179         if (fmt[--fmt_size] != 0)
 180                 return -EINVAL;
 181
 182         /* check format string for allowed specifiers */
 183         for (i = 0; i < fmt_size; i++) {
 184                 if ((!isprint(fmt[i]) && !isspace(fmt[i])) || !isascii(fmt[i]))
 185                         return -EINVAL;
 186
 187                 if (fmt[i] != '%')
 188                         continue;
 189
 190                 if (fmt_cnt >= 3)
 191                         return -EINVAL;
 192
 193                 /* fmt[i] != 0 && fmt[last] == 0, so we can access fmt[i + 1] */
 194                 i++;
 195                 if (fmt[i] == 'l') {
 196                         mod[fmt_cnt]++;
 197                         i++;
 198                 } else if (fmt[i] == 'p' || fmt[i] == 's') {
 199                         mod[fmt_cnt]++;
 200                         i++;
 201                         if (!isspace(fmt[i]) && !ispunct(fmt[i]) && fmt[i] != 0)
 202                                 return -EINVAL;
 203                         fmt_cnt++;
 204                         if (fmt[i - 1] == 's') {
 205                                 if (str_seen)
 206                                         /* allow only one '%s' per fmt string */
 207                                         return -EINVAL;
 208                                 str_seen = true;
 209
 210                                 switch (fmt_cnt) {
 211                                 case 1:
 212                                         unsafe_addr = arg1;
 213                                         arg1 = (long) buf;
 214                                         break;
 215                                 case 2:
 216                                         unsafe_addr = arg2;
 217                                         arg2 = (long) buf;
 218                                         break;
 219                                 case 3:
 220                                         unsafe_addr = arg3;
 221                                         arg3 = (long) buf;
 222                                         break;
 223                                 }
 224                                 buf[0] = 0;
 225                                 strncpy_from_unsafe(buf,
 226                                                     (void *) (long) unsafe_addr,
 227                                                     sizeof(buf));
 228                         }
 229                         continue;
 230                 }
 231
 232                 if (fmt[i] == 'l') {
 233                         mod[fmt_cnt]++;
 234                         i++;
 235                 }
 236
 237                 if (fmt[i] != 'i' && fmt[i] != 'd' &&
 238                     fmt[i] != 'u' && fmt[i] != 'x')
 239                         return -EINVAL;
 240                 fmt_cnt++;
 241         }
 242
 243 /* Horrid workaround for getting va_list handling working with different
 244  * argument type combinations generically for 32 and 64 bit archs.
 245  */
 246 #define __BPF_TP_EMIT() __BPF_ARG3_TP()
 247 #define __BPF_TP(...)                                                   \
 248         __trace_printk(0 /* Fake ip */,                                 \
 249                        fmt, ##__VA_ARGS__)
 250
 251 #define __BPF_ARG1_TP(...)                                              \
 252         ((mod[0] == 2 || (mod[0] == 1 && __BITS_PER_LONG == 64))        \
 253           ? __BPF_TP(arg1, ##__VA_ARGS__)                               \
 254           : ((mod[0] == 1 || (mod[0] == 0 && __BITS_PER_LONG == 32))    \
 255               ? __BPF_TP((long)arg1, ##__VA_ARGS__)                     \
 256               : __BPF_TP((u32)arg1, ##__VA_ARGS__)))
 257
 258 #define __BPF_ARG2_TP(...)                                              \
 259         ((mod[1] == 2 || (mod[1] == 1 && __BITS_PER_LONG == 64))        \
 260           ? __BPF_ARG1_TP(arg2, ##__VA_ARGS__)                          \
 261           : ((mod[1] == 1 || (mod[1] == 0 && __BITS_PER_LONG == 32))    \
 262               ? __BPF_ARG1_TP((long)arg2, ##__VA_ARGS__)                \
 263               : __BPF_ARG1_TP((u32)arg2, ##__VA_ARGS__)))
 264
 265 #define __BPF_ARG3_TP(...)                                              \
 266         ((mod[2] == 2 || (mod[2] == 1 && __BITS_PER_LONG == 64))        \
 267           ? __BPF_ARG2_TP(arg3, ##__VA_ARGS__)                          \
 268           : ((mod[2] == 1 || (mod[2] == 0 && __BITS_PER_LONG == 32))    \
 269               ? __BPF_ARG2_TP((long)arg3, ##__VA_ARGS__)                \
 270               : __BPF_ARG2_TP((u32)arg3, ##__VA_ARGS__)))
 271
 272         return __BPF_TP_EMIT();
 273 }
 274
 275 static const struct bpf_func_proto bpf_trace_printk_proto = {
 276         .func           = bpf_trace_printk,
 277         .gpl_only       = true,
 278         .ret_type       = RET_INTEGER,
 279         .arg1_type      = ARG_PTR_TO_MEM,
 280         .arg2_type      = ARG_CONST_SIZE,
 281 };
 282
 283 const struct bpf_func_proto *bpf_get_trace_printk_proto(void)
 284 {
 285         /*
 286          * this program might be calling bpf_trace_printk,
 287          * so allocate per-cpu printk buffers
 288          */
 289         trace_printk_init_buffers();
 290
 291         return &bpf_trace_printk_proto;
 292 }
 293
 294 static __always_inline int
 295 get_map_perf_counter(struct bpf_map *map, u64 flags,
 296                      u64 *value, u64 *enabled, u64 *running)
 297 {
 298         struct bpf_array *array = container_of(map, struct bpf_array, map);
 299         unsigned int cpu = smp_processor_id();
 300         u64 index = flags & BPF_F_INDEX_MASK;
 301         struct bpf_event_entry *ee;
 302
 303         if (unlikely(flags & ~(BPF_F_INDEX_MASK)))
 304                 return -EINVAL;
 305         if (index == BPF_F_CURRENT_CPU)
 306                 index = cpu;
 307         if (unlikely(index >= array->map.max_entries))
 308                 return -E2BIG;
 309
 310         ee = READ_ONCE(array->ptrs[index]);
 311         if (!ee)
 312                 return -ENOENT;
 313
 314         return perf_event_read_local(ee->event, value, enabled, running);
 315 }
 316
 317 BPF_CALL_2(bpf_perf_event_read, struct bpf_map *, map, u64, flags)
 318 {
 319         u64 value = 0;
 320         int err;
 321
 322         err = get_map_perf_counter(map, flags, &value, NULL, NULL);
 323         /*
 324          * this api is ugly since we miss [-22..-2] range of valid
 325          * counter values, but that's uapi
 326          */
 327         if (err)
 328                 return err;
 329         return value;
 330 }
 331
 332 static const struct bpf_func_proto bpf_perf_event_read_proto = {
 333         .func           = bpf_perf_event_read,
 334         .gpl_only       = true,
 335         .ret_type       = RET_INTEGER,
 336         .arg1_type      = ARG_CONST_MAP_PTR,
 337         .arg2_type      = ARG_ANYTHING,
 338 };
 339
 340 BPF_CALL_4(bpf_perf_event_read_value, struct bpf_map *, map, u64, flags,
 341            struct bpf_perf_event_value *, buf, u32, size)
 342 {
 343         int err = -EINVAL;
 344
 345         if (unlikely(size != sizeof(struct bpf_perf_event_value)))
 346                 goto clear;
 347         err = get_map_perf_counter(map, flags, &buf->counter, &buf->enabled,
 348                                    &buf->running);
 349         if (unlikely(err))
 350                 goto clear;
 351         return 0;
 352 clear:
 353         memset(buf, 0, size);
 354         return err;
 355 }
 356
 357 static const struct bpf_func_proto bpf_perf_event_read_value_proto = {
 358         .func           = bpf_perf_event_read_value,
 359         .gpl_only       = true,
 360         .ret_type       = RET_INTEGER,
 361         .arg1_type      = ARG_CONST_MAP_PTR,
 362         .arg2_type      = ARG_ANYTHING,
 363         .arg3_type      = ARG_PTR_TO_UNINIT_MEM,
 364         .arg4_type      = ARG_CONST_SIZE,
 365 };
 366
 367 static DEFINE_PER_CPU(struct perf_sample_data, bpf_trace_sd);
 368
 369 static __always_inline u64
 370 __bpf_perf_event_output(struct pt_regs *regs, struct bpf_map *map,
 371                         u64 flags, struct perf_sample_data *sd)
 372 {
 373         struct bpf_array *array = container_of(map, struct bpf_array, map);
 374         unsigned int cpu = smp_processor_id();
 375         u64 index = flags & BPF_F_INDEX_MASK;
 376         struct bpf_event_entry *ee;
 377         struct perf_event *event;
 378
 379         if (index == BPF_F_CURRENT_CPU)
 380                 index = cpu;
 381         if (unlikely(index >= array->map.max_entries))
 382                 return -E2BIG;
 383
 384         ee = READ_ONCE(array->ptrs[index]);
 385         if (!ee)
 386                 return -ENOENT;
 387
 388         event = ee->event;
 389         if (unlikely(event->attr.type != PERF_TYPE_SOFTWARE ||
 390                      event->attr.config != PERF_COUNT_SW_BPF_OUTPUT))
 391                 return -EINVAL;
 392
 393         if (unlikely(event->oncpu != cpu))
 394                 return -EOPNOTSUPP;
 395
 396         perf_event_output(event, sd, regs);
 397         return 0;
 398 }
 399
 400 BPF_CALL_5(bpf_perf_event_output, struct pt_regs *, regs, struct bpf_map *, map,
 401            u64, flags, void *, data, u64, size)
 402 {
 403         struct perf_sample_data *sd = this_cpu_ptr(&bpf_trace_sd);
 404         struct perf_raw_record raw = {
 405                 .frag = {
 406                         .size = size,
 407                         .data = data,
 408                 },
 409         };
 410
 411         if (unlikely(flags & ~(BPF_F_INDEX_MASK)))
 412                 return -EINVAL;
 413
 414         perf_sample_data_init(sd, 0, 0);
 415         sd->raw = &raw;
 416
 417         return __bpf_perf_event_output(regs, map, flags, sd);
 418 }
 419
 420 static const struct bpf_func_proto bpf_perf_event_output_proto = {
 421         .func           = bpf_perf_event_output,
 422         .gpl_only       = true,
 423         .ret_type       = RET_INTEGER,
 424         .arg1_type      = ARG_PTR_TO_CTX,
 425         .arg2_type      = ARG_CONST_MAP_PTR,
 426         .arg3_type      = ARG_ANYTHING,
 427         .arg4_type      = ARG_PTR_TO_MEM,
 428         .arg5_type      = ARG_CONST_SIZE_OR_ZERO,
 429 };
 430
 431 static DEFINE_PER_CPU(struct pt_regs, bpf_pt_regs);
 432 static DEFINE_PER_CPU(struct perf_sample_data, bpf_misc_sd);
 433
 434 u64 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
 435                      void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
 436 {
 437         struct perf_sample_data *sd = this_cpu_ptr(&bpf_misc_sd);
 438         struct pt_regs *regs = this_cpu_ptr(&bpf_pt_regs);
 439         struct perf_raw_frag frag = {
 440                 .copy           = ctx_copy,
 441                 .size           = ctx_size,
 442                 .data           = ctx,
 443         };
 444         struct perf_raw_record raw = {
 445                 .frag = {
 446                         {
 447                                 .next   = ctx_size ? &frag : NULL,
 448                         },
 449                         .size   = meta_size,
 450                         .data   = meta,
 451                 },
 452         };
 453
 454         perf_fetch_caller_regs(regs);
 455         perf_sample_data_init(sd, 0, 0);
 456         sd->raw = &raw;
 457
 458         return __bpf_perf_event_output(regs, map, flags, sd);
 459 }
 460
 461 BPF_CALL_0(bpf_get_current_task)
 462 {
 463         return (long) current;
 464 }
 465
 466 static const struct bpf_func_proto bpf_get_current_task_proto = {
 467         .func           = bpf_get_current_task,
 468         .gpl_only       = true,
 469         .ret_type       = RET_INTEGER,
 470 };
 471
 472 BPF_CALL_2(bpf_current_task_under_cgroup, struct bpf_map *, map, u32, idx)
 473 {
 474         struct bpf_array *array = container_of(map, struct bpf_array, map);
 475         struct cgroup *cgrp;
 476
 477         if (unlikely(in_interrupt()))
 478                 return -EINVAL;
 479         if (unlikely(idx >= array->map.max_entries))
 480                 return -E2BIG;
 481
 482         cgrp = READ_ONCE(array->ptrs[idx]);
 483         if (unlikely(!cgrp))
 484                 return -EAGAIN;
 485
 486         return task_under_cgroup_hierarchy(current, cgrp);
 487 }
 488
 489 static const struct bpf_func_proto bpf_current_task_under_cgroup_proto = {
 490         .func           = bpf_current_task_under_cgroup,
 491         .gpl_only       = false,
 492         .ret_type       = RET_INTEGER,
 493         .arg1_type      = ARG_CONST_MAP_PTR,
 494         .arg2_type      = ARG_ANYTHING,
 495 };
 496
 497 BPF_CALL_3(bpf_probe_read_str, void *, dst, u32, size,
 498            const void *, unsafe_ptr)
 499 {
 500         int ret;
 501
 502         /*
 503          * The strncpy_from_unsafe() call will likely not fill the entire
 504          * buffer, but that's okay in this circumstance as we're probing
 505          * arbitrary memory anyway similar to bpf_probe_read() and might
 506          * as well probe the stack. Thus, memory is explicitly cleared
 507          * only in error case, so that improper users ignoring return
 508          * code altogether don't copy garbage; otherwise length of string
 509          * is returned that can be used for bpf_perf_event_output() et al.
 510          */
 511         ret = strncpy_from_unsafe(dst, unsafe_ptr, size);
 512         if (unlikely(ret < 0))
 513                 memset(dst, 0, size);
 514
 515         return ret;
 516 }
 517
 518 static const struct bpf_func_proto bpf_probe_read_str_proto = {
 519         .func           = bpf_probe_read_str,
 520         .gpl_only       = true,
 521         .ret_type       = RET_INTEGER,
 522         .arg1_type      = ARG_PTR_TO_UNINIT_MEM,
 523         .arg2_type      = ARG_CONST_SIZE_OR_ZERO,
 524         .arg3_type      = ARG_ANYTHING,
 525 };
 526
 527 static const struct bpf_func_proto *tracing_func_proto(enum bpf_func_id func_id)
 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         default:
 567                 return NULL;
 568         }
 569 }
 570
 571 static const struct bpf_func_proto *kprobe_prog_func_proto(enum bpf_func_id func_id)
 572 {
 573         switch (func_id) {
 574         case BPF_FUNC_perf_event_output:
 575                 return &bpf_perf_event_output_proto;
 576         case BPF_FUNC_get_stackid:
 577                 return &bpf_get_stackid_proto;
 578         case BPF_FUNC_perf_event_read_value:
 579                 return &bpf_perf_event_read_value_proto;
 580 #ifdef CONFIG_BPF_KPROBE_OVERRIDE
 581         case BPF_FUNC_override_return:
 582                 return &bpf_override_return_proto;
 583 #endif
 584         default:
 585                 return tracing_func_proto(func_id);
 586         }
 587 }
 588
 589 /* bpf+kprobe programs can access fields of 'struct pt_regs' */
 590 static bool kprobe_prog_is_valid_access(int off, int size, enum bpf_access_type type,
 591                                         struct bpf_insn_access_aux *info)
 592 {
 593         if (off < 0 || off >= sizeof(struct pt_regs))
 594                 return false;
 595         if (type != BPF_READ)
 596                 return false;
 597         if (off % size != 0)
 598                 return false;
 599         /*
 600          * Assertion for 32 bit to make sure last 8 byte access
 601          * (BPF_DW) to the last 4 byte member is disallowed.
 602          */
 603         if (off + size > sizeof(struct pt_regs))
 604                 return false;
 605
 606         return true;
 607 }
 608
 609 const struct bpf_verifier_ops kprobe_verifier_ops = {
 610         .get_func_proto  = kprobe_prog_func_proto,
 611         .is_valid_access = kprobe_prog_is_valid_access,
 612 };
 613
 614 const struct bpf_prog_ops kprobe_prog_ops = {
 615 };
 616
 617 BPF_CALL_5(bpf_perf_event_output_tp, void *, tp_buff, struct bpf_map *, map,
 618            u64, flags, void *, data, u64, size)
 619 {
 620         struct pt_regs *regs = *(struct pt_regs **)tp_buff;
 621
 622         /*
 623          * r1 points to perf tracepoint buffer where first 8 bytes are hidden
 624          * from bpf program and contain a pointer to 'struct pt_regs'. Fetch it
 625          * from there and call the same bpf_perf_event_output() helper inline.
 626          */
 627         return ____bpf_perf_event_output(regs, map, flags, data, size);
 628 }
 629
 630 static const struct bpf_func_proto bpf_perf_event_output_proto_tp = {
 631         .func           = bpf_perf_event_output_tp,
 632         .gpl_only       = true,
 633         .ret_type       = RET_INTEGER,
 634         .arg1_type      = ARG_PTR_TO_CTX,
 635         .arg2_type      = ARG_CONST_MAP_PTR,
 636         .arg3_type      = ARG_ANYTHING,
 637         .arg4_type      = ARG_PTR_TO_MEM,
 638         .arg5_type      = ARG_CONST_SIZE_OR_ZERO,
 639 };
 640
 641 BPF_CALL_3(bpf_get_stackid_tp, void *, tp_buff, struct bpf_map *, map,
 642            u64, flags)
 643 {
 644         struct pt_regs *regs = *(struct pt_regs **)tp_buff;
 645
 646         /*
 647          * Same comment as in bpf_perf_event_output_tp(), only that this time
 648          * the other helper's function body cannot be inlined due to being
 649          * external, thus we need to call raw helper function.
 650          */
 651         return bpf_get_stackid((unsigned long) regs, (unsigned long) map,
 652                                flags, 0, 0);
 653 }
 654
 655 static const struct bpf_func_proto bpf_get_stackid_proto_tp = {
 656         .func           = bpf_get_stackid_tp,
 657         .gpl_only       = true,
 658         .ret_type       = RET_INTEGER,
 659         .arg1_type      = ARG_PTR_TO_CTX,
 660         .arg2_type      = ARG_CONST_MAP_PTR,
 661         .arg3_type      = ARG_ANYTHING,
 662 };
 663
 664 static const struct bpf_func_proto *tp_prog_func_proto(enum bpf_func_id func_id)
 665 {
 666         switch (func_id) {
 667         case BPF_FUNC_perf_event_output:
 668                 return &bpf_perf_event_output_proto_tp;
 669         case BPF_FUNC_get_stackid:
 670                 return &bpf_get_stackid_proto_tp;
 671         default:
 672                 return tracing_func_proto(func_id);
 673         }
 674 }
 675
 676 static bool tp_prog_is_valid_access(int off, int size, enum bpf_access_type type,
 677                                     struct bpf_insn_access_aux *info)
 678 {
 679         if (off < sizeof(void *) || off >= PERF_MAX_TRACE_SIZE)
 680                 return false;
 681         if (type != BPF_READ)
 682                 return false;
 683         if (off % size != 0)
 684                 return false;
 685
 686         BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(__u64));
 687         return true;
 688 }
 689
 690 const struct bpf_verifier_ops tracepoint_verifier_ops = {
 691         .get_func_proto  = tp_prog_func_proto,
 692         .is_valid_access = tp_prog_is_valid_access,
 693 };
 694
 695 const struct bpf_prog_ops tracepoint_prog_ops = {
 696 };
 697
 698 BPF_CALL_3(bpf_perf_prog_read_value, struct bpf_perf_event_data_kern *, ctx,
 699            struct bpf_perf_event_value *, buf, u32, size)
 700 {
 701         int err = -EINVAL;
 702
 703         if (unlikely(size != sizeof(struct bpf_perf_event_value)))
 704                 goto clear;
 705         err = perf_event_read_local(ctx->event, &buf->counter, &buf->enabled,
 706                                     &buf->running);
 707         if (unlikely(err))
 708                 goto clear;
 709         return 0;
 710 clear:
 711         memset(buf, 0, size);
 712         return err;
 713 }
 714
 715 static const struct bpf_func_proto bpf_perf_prog_read_value_proto = {
 716          .func           = bpf_perf_prog_read_value,
 717          .gpl_only       = true,
 718          .ret_type       = RET_INTEGER,
 719          .arg1_type      = ARG_PTR_TO_CTX,
 720          .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
 721          .arg3_type      = ARG_CONST_SIZE,
 722 };
 723
 724 static const struct bpf_func_proto *pe_prog_func_proto(enum bpf_func_id func_id)
 725 {
 726         switch (func_id) {
 727         case BPF_FUNC_perf_event_output:
 728                 return &bpf_perf_event_output_proto_tp;
 729         case BPF_FUNC_get_stackid:
 730                 return &bpf_get_stackid_proto_tp;
 731         case BPF_FUNC_perf_prog_read_value:
 732                 return &bpf_perf_prog_read_value_proto;
 733         default:
 734                 return tracing_func_proto(func_id);
 735         }
 736 }
 737
 738 static bool pe_prog_is_valid_access(int off, int size, enum bpf_access_type type,
 739                                     struct bpf_insn_access_aux *info)
 740 {
 741         const int size_sp = FIELD_SIZEOF(struct bpf_perf_event_data,
 742                                          sample_period);
 743
 744         if (off < 0 || off >= sizeof(struct bpf_perf_event_data))
 745                 return false;
 746         if (type != BPF_READ)
 747                 return false;
 748         if (off % size != 0)
 749                 return false;
 750
 751         switch (off) {
 752         case bpf_ctx_range(struct bpf_perf_event_data, sample_period):
 753                 bpf_ctx_record_field_size(info, size_sp);
 754                 if (!bpf_ctx_narrow_access_ok(off, size, size_sp))
 755                         return false;
 756                 break;
 757         default:
 758                 if (size != sizeof(long))
 759                         return false;
 760         }
 761
 762         return true;
 763 }
 764
 765 static u32 pe_prog_convert_ctx_access(enum bpf_access_type type,
 766                                       const struct bpf_insn *si,
 767                                       struct bpf_insn *insn_buf,
 768                                       struct bpf_prog *prog, u32 *target_size)
 769 {
 770         struct bpf_insn *insn = insn_buf;
 771
 772         switch (si->off) {
 773         case offsetof(struct bpf_perf_event_data, sample_period):
 774                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct bpf_perf_event_data_kern,
 775                                                        data), si->dst_reg, si->src_reg,
 776                                       offsetof(struct bpf_perf_event_data_kern, data));
 777                 *insn++ = BPF_LDX_MEM(BPF_DW, si->dst_reg, si->dst_reg,
 778                                       bpf_target_off(struct perf_sample_data, period, 8,
 779                                                      target_size));
 780                 break;
 781         default:
 782                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct bpf_perf_event_data_kern,
 783                                                        regs), si->dst_reg, si->src_reg,
 784                                       offsetof(struct bpf_perf_event_data_kern, regs));
 785                 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(long), si->dst_reg, si->dst_reg,
 786                                       si->off);
 787                 break;
 788         }
 789
 790         return insn - insn_buf;
 791 }
 792
 793 const struct bpf_verifier_ops perf_event_verifier_ops = {
 794         .get_func_proto         = pe_prog_func_proto,
 795         .is_valid_access        = pe_prog_is_valid_access,
 796         .convert_ctx_access     = pe_prog_convert_ctx_access,
 797 };
 798
 799 const struct bpf_prog_ops perf_event_prog_ops = {
 800 };
 801
 802 static DEFINE_MUTEX(bpf_event_mutex);
 803
 804 #define BPF_TRACE_MAX_PROGS 64
 805
 806 int perf_event_attach_bpf_prog(struct perf_event *event,
 807                                struct bpf_prog *prog)
 808 {
 809         struct bpf_prog_array __rcu *old_array;
 810         struct bpf_prog_array *new_array;
 811         int ret = -EEXIST;
 812
 813         /*
 814          * Kprobe override only works if they are on the function entry,
 815          * and only if they are on the opt-in list.
 816          */
 817         if (prog->kprobe_override &&
 818             (!trace_kprobe_on_func_entry(event->tp_event) ||
 819              !trace_kprobe_error_injectable(event->tp_event)))
 820                 return -EINVAL;
 821
 822         mutex_lock(&bpf_event_mutex);
 823
 824         if (event->prog)
 825                 goto unlock;
 826
 827         old_array = event->tp_event->prog_array;
 828         if (old_array &&
 829             bpf_prog_array_length(old_array) >= BPF_TRACE_MAX_PROGS) {
 830                 ret = -E2BIG;
 831                 goto unlock;
 832         }
 833
 834         ret = bpf_prog_array_copy(old_array, NULL, prog, &new_array);
 835         if (ret < 0)
 836                 goto unlock;
 837
 838         /* set the new array to event->tp_event and set event->prog */
 839         event->prog = prog;
 840         rcu_assign_pointer(event->tp_event->prog_array, new_array);
 841         bpf_prog_array_free(old_array);
 842
 843 unlock:
 844         mutex_unlock(&bpf_event_mutex);
 845         return ret;
 846 }
 847
 848 void perf_event_detach_bpf_prog(struct perf_event *event)
 849 {
 850         struct bpf_prog_array __rcu *old_array;
 851         struct bpf_prog_array *new_array;
 852         int ret;
 853
 854         mutex_lock(&bpf_event_mutex);
 855
 856         if (!event->prog)
 857                 goto unlock;
 858
 859         old_array = event->tp_event->prog_array;
 860         ret = bpf_prog_array_copy(old_array, event->prog, NULL, &new_array);
 861         if (ret < 0) {
 862                 bpf_prog_array_delete_safe(old_array, event->prog);
 863         } else {
 864                 rcu_assign_pointer(event->tp_event->prog_array, new_array);
 865                 bpf_prog_array_free(old_array);
 866         }
 867
 868         bpf_prog_put(event->prog);
 869         event->prog = NULL;
 870
 871 unlock:
 872         mutex_unlock(&bpf_event_mutex);
 873 }
 874
 875 int perf_event_query_prog_array(struct perf_event *event, void __user *info)
 876 {
 877         struct perf_event_query_bpf __user *uquery = info;
 878         struct perf_event_query_bpf query = {};
 879         int ret;
 880
 881         if (!capable(CAP_SYS_ADMIN))
 882                 return -EPERM;
 883         if (event->attr.type != PERF_TYPE_TRACEPOINT)
 884                 return -EINVAL;
 885         if (copy_from_user(&query, uquery, sizeof(query)))
 886                 return -EFAULT;
 887         if (query.ids_len > BPF_TRACE_MAX_PROGS)
 888                 return -E2BIG;
 889
 890         mutex_lock(&bpf_event_mutex);
 891         ret = bpf_prog_array_copy_info(event->tp_event->prog_array,
 892                                        uquery->ids,
 893                                        query.ids_len,
 894                                        &uquery->prog_cnt);
 895         mutex_unlock(&bpf_event_mutex);
 896
 897         return ret;
 898 }