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 *
 528 tracing_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
 529 {
 530         switch (func_id) {
 531         case BPF_FUNC_map_lookup_elem:
 532                 return &bpf_map_lookup_elem_proto;
 533         case BPF_FUNC_map_update_elem:
 534                 return &bpf_map_update_elem_proto;
 535         case BPF_FUNC_map_delete_elem:
 536                 return &bpf_map_delete_elem_proto;
 537         case BPF_FUNC_probe_read:
 538                 return &bpf_probe_read_proto;
 539         case BPF_FUNC_ktime_get_ns:
 540                 return &bpf_ktime_get_ns_proto;
 541         case BPF_FUNC_tail_call:
 542                 return &bpf_tail_call_proto;
 543         case BPF_FUNC_get_current_pid_tgid:
 544                 return &bpf_get_current_pid_tgid_proto;
 545         case BPF_FUNC_get_current_task:
 546                 return &bpf_get_current_task_proto;
 547         case BPF_FUNC_get_current_uid_gid:
 548                 return &bpf_get_current_uid_gid_proto;
 549         case BPF_FUNC_get_current_comm:
 550                 return &bpf_get_current_comm_proto;
 551         case BPF_FUNC_trace_printk:
 552                 return bpf_get_trace_printk_proto();
 553         case BPF_FUNC_get_smp_processor_id:
 554                 return &bpf_get_smp_processor_id_proto;
 555         case BPF_FUNC_get_numa_node_id:
 556                 return &bpf_get_numa_node_id_proto;
 557         case BPF_FUNC_perf_event_read:
 558                 return &bpf_perf_event_read_proto;
 559         case BPF_FUNC_probe_write_user:
 560                 return bpf_get_probe_write_proto();
 561         case BPF_FUNC_current_task_under_cgroup:
 562                 return &bpf_current_task_under_cgroup_proto;
 563         case BPF_FUNC_get_prandom_u32:
 564                 return &bpf_get_prandom_u32_proto;
 565         case BPF_FUNC_probe_read_str:
 566                 return &bpf_probe_read_str_proto;
 567         default:
 568                 return NULL;
 569         }
 570 }
 571
 572 static const struct bpf_func_proto *
 573 kprobe_prog_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
 574 {
 575         switch (func_id) {
 576         case BPF_FUNC_perf_event_output:
 577                 return &bpf_perf_event_output_proto;
 578         case BPF_FUNC_get_stackid:
 579                 return &bpf_get_stackid_proto;
 580         case BPF_FUNC_perf_event_read_value:
 581                 return &bpf_perf_event_read_value_proto;
 582 #ifdef CONFIG_BPF_KPROBE_OVERRIDE
 583         case BPF_FUNC_override_return:
 584                 return &bpf_override_return_proto;
 585 #endif
 586         default:
 587                 return tracing_func_proto(func_id, prog);
 588         }
 589 }
 590
 591 /* bpf+kprobe programs can access fields of 'struct pt_regs' */
 592 static bool kprobe_prog_is_valid_access(int off, int size, enum bpf_access_type type,
 593                                         const struct bpf_prog *prog,
 594                                         struct bpf_insn_access_aux *info)
 595 {
 596         if (off < 0 || off >= sizeof(struct pt_regs))
 597                 return false;
 598         if (type != BPF_READ)
 599                 return false;
 600         if (off % size != 0)
 601                 return false;
 602         /*
 603          * Assertion for 32 bit to make sure last 8 byte access
 604          * (BPF_DW) to the last 4 byte member is disallowed.
 605          */
 606         if (off + size > sizeof(struct pt_regs))
 607                 return false;
 608
 609         return true;
 610 }
 611
 612 const struct bpf_verifier_ops kprobe_verifier_ops = {
 613         .get_func_proto  = kprobe_prog_func_proto,
 614         .is_valid_access = kprobe_prog_is_valid_access,
 615 };
 616
 617 const struct bpf_prog_ops kprobe_prog_ops = {
 618 };
 619
 620 BPF_CALL_5(bpf_perf_event_output_tp, void *, tp_buff, struct bpf_map *, map,
 621            u64, flags, void *, data, u64, size)
 622 {
 623         struct pt_regs *regs = *(struct pt_regs **)tp_buff;
 624
 625         /*
 626          * r1 points to perf tracepoint buffer where first 8 bytes are hidden
 627          * from bpf program and contain a pointer to 'struct pt_regs'. Fetch it
 628          * from there and call the same bpf_perf_event_output() helper inline.
 629          */
 630         return ____bpf_perf_event_output(regs, map, flags, data, size);
 631 }
 632
 633 static const struct bpf_func_proto bpf_perf_event_output_proto_tp = {
 634         .func           = bpf_perf_event_output_tp,
 635         .gpl_only       = true,
 636         .ret_type       = RET_INTEGER,
 637         .arg1_type      = ARG_PTR_TO_CTX,
 638         .arg2_type      = ARG_CONST_MAP_PTR,
 639         .arg3_type      = ARG_ANYTHING,
 640         .arg4_type      = ARG_PTR_TO_MEM,
 641         .arg5_type      = ARG_CONST_SIZE_OR_ZERO,
 642 };
 643
 644 BPF_CALL_3(bpf_get_stackid_tp, void *, tp_buff, struct bpf_map *, map,
 645            u64, flags)
 646 {
 647         struct pt_regs *regs = *(struct pt_regs **)tp_buff;
 648
 649         /*
 650          * Same comment as in bpf_perf_event_output_tp(), only that this time
 651          * the other helper's function body cannot be inlined due to being
 652          * external, thus we need to call raw helper function.
 653          */
 654         return bpf_get_stackid((unsigned long) regs, (unsigned long) map,
 655                                flags, 0, 0);
 656 }
 657
 658 static const struct bpf_func_proto bpf_get_stackid_proto_tp = {
 659         .func           = bpf_get_stackid_tp,
 660         .gpl_only       = true,
 661         .ret_type       = RET_INTEGER,
 662         .arg1_type      = ARG_PTR_TO_CTX,
 663         .arg2_type      = ARG_CONST_MAP_PTR,
 664         .arg3_type      = ARG_ANYTHING,
 665 };
 666
 667 static const struct bpf_func_proto *
 668 tp_prog_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
 669 {
 670         switch (func_id) {
 671         case BPF_FUNC_perf_event_output:
 672                 return &bpf_perf_event_output_proto_tp;
 673         case BPF_FUNC_get_stackid:
 674                 return &bpf_get_stackid_proto_tp;
 675         default:
 676                 return tracing_func_proto(func_id, prog);
 677         }
 678 }
 679
 680 static bool tp_prog_is_valid_access(int off, int size, enum bpf_access_type type,
 681                                     const struct bpf_prog *prog,
 682                                     struct bpf_insn_access_aux *info)
 683 {
 684         if (off < sizeof(void *) || off >= PERF_MAX_TRACE_SIZE)
 685                 return false;
 686         if (type != BPF_READ)
 687                 return false;
 688         if (off % size != 0)
 689                 return false;
 690
 691         BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(__u64));
 692         return true;
 693 }
 694
 695 const struct bpf_verifier_ops tracepoint_verifier_ops = {
 696         .get_func_proto  = tp_prog_func_proto,
 697         .is_valid_access = tp_prog_is_valid_access,
 698 };
 699
 700 const struct bpf_prog_ops tracepoint_prog_ops = {
 701 };
 702
 703 BPF_CALL_3(bpf_perf_prog_read_value, struct bpf_perf_event_data_kern *, ctx,
 704            struct bpf_perf_event_value *, buf, u32, size)
 705 {
 706         int err = -EINVAL;
 707
 708         if (unlikely(size != sizeof(struct bpf_perf_event_value)))
 709                 goto clear;
 710         err = perf_event_read_local(ctx->event, &buf->counter, &buf->enabled,
 711                                     &buf->running);
 712         if (unlikely(err))
 713                 goto clear;
 714         return 0;
 715 clear:
 716         memset(buf, 0, size);
 717         return err;
 718 }
 719
 720 static const struct bpf_func_proto bpf_perf_prog_read_value_proto = {
 721          .func           = bpf_perf_prog_read_value,
 722          .gpl_only       = true,
 723          .ret_type       = RET_INTEGER,
 724          .arg1_type      = ARG_PTR_TO_CTX,
 725          .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
 726          .arg3_type      = ARG_CONST_SIZE,
 727 };
 728
 729 static const struct bpf_func_proto *
 730 pe_prog_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
 731 {
 732         switch (func_id) {
 733         case BPF_FUNC_perf_event_output:
 734                 return &bpf_perf_event_output_proto_tp;
 735         case BPF_FUNC_get_stackid:
 736                 return &bpf_get_stackid_proto_tp;
 737         case BPF_FUNC_perf_prog_read_value:
 738                 return &bpf_perf_prog_read_value_proto;
 739         default:
 740                 return tracing_func_proto(func_id, prog);
 741         }
 742 }
 743
 744 /*
 745  * bpf_raw_tp_regs are separate from bpf_pt_regs used from skb/xdp
 746  * to avoid potential recursive reuse issue when/if tracepoints are added
 747  * inside bpf_*_event_output and/or bpf_get_stack_id
 748  */
 749 static DEFINE_PER_CPU(struct pt_regs, bpf_raw_tp_regs);
 750 BPF_CALL_5(bpf_perf_event_output_raw_tp, struct bpf_raw_tracepoint_args *, args,
 751            struct bpf_map *, map, u64, flags, void *, data, u64, size)
 752 {
 753         struct pt_regs *regs = this_cpu_ptr(&bpf_raw_tp_regs);
 754
 755         perf_fetch_caller_regs(regs);
 756         return ____bpf_perf_event_output(regs, map, flags, data, size);
 757 }
 758
 759 static const struct bpf_func_proto bpf_perf_event_output_proto_raw_tp = {
 760         .func           = bpf_perf_event_output_raw_tp,
 761         .gpl_only       = true,
 762         .ret_type       = RET_INTEGER,
 763         .arg1_type      = ARG_PTR_TO_CTX,
 764         .arg2_type      = ARG_CONST_MAP_PTR,
 765         .arg3_type      = ARG_ANYTHING,
 766         .arg4_type      = ARG_PTR_TO_MEM,
 767         .arg5_type      = ARG_CONST_SIZE_OR_ZERO,
 768 };
 769
 770 BPF_CALL_3(bpf_get_stackid_raw_tp, struct bpf_raw_tracepoint_args *, args,
 771            struct bpf_map *, map, u64, flags)
 772 {
 773         struct pt_regs *regs = this_cpu_ptr(&bpf_raw_tp_regs);
 774
 775         perf_fetch_caller_regs(regs);
 776         /* similar to bpf_perf_event_output_tp, but pt_regs fetched differently */
 777         return bpf_get_stackid((unsigned long) regs, (unsigned long) map,
 778                                flags, 0, 0);
 779 }
 780
 781 static const struct bpf_func_proto bpf_get_stackid_proto_raw_tp = {
 782         .func           = bpf_get_stackid_raw_tp,
 783         .gpl_only       = true,
 784         .ret_type       = RET_INTEGER,
 785         .arg1_type      = ARG_PTR_TO_CTX,
 786         .arg2_type      = ARG_CONST_MAP_PTR,
 787         .arg3_type      = ARG_ANYTHING,
 788 };
 789
 790 static const struct bpf_func_proto *
 791 raw_tp_prog_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
 792 {
 793         switch (func_id) {
 794         case BPF_FUNC_perf_event_output:
 795                 return &bpf_perf_event_output_proto_raw_tp;
 796         case BPF_FUNC_get_stackid:
 797                 return &bpf_get_stackid_proto_raw_tp;
 798         default:
 799                 return tracing_func_proto(func_id, prog);
 800         }
 801 }
 802
 803 static bool raw_tp_prog_is_valid_access(int off, int size,
 804                                         enum bpf_access_type type,
 805                                         const struct bpf_prog *prog,
 806                                         struct bpf_insn_access_aux *info)
 807 {
 808         /* largest tracepoint in the kernel has 12 args */
 809         if (off < 0 || off >= sizeof(__u64) * 12)
 810                 return false;
 811         if (type != BPF_READ)
 812                 return false;
 813         if (off % size != 0)
 814                 return false;
 815         return true;
 816 }
 817
 818 const struct bpf_verifier_ops raw_tracepoint_verifier_ops = {
 819         .get_func_proto  = raw_tp_prog_func_proto,
 820         .is_valid_access = raw_tp_prog_is_valid_access,
 821 };
 822
 823 const struct bpf_prog_ops raw_tracepoint_prog_ops = {
 824 };
 825
 826 static bool pe_prog_is_valid_access(int off, int size, enum bpf_access_type type,
 827                                     const struct bpf_prog *prog,
 828                                     struct bpf_insn_access_aux *info)
 829 {
 830         const int size_u64 = sizeof(u64);
 831
 832         if (off < 0 || off >= sizeof(struct bpf_perf_event_data))
 833                 return false;
 834         if (type != BPF_READ)
 835                 return false;
 836         if (off % size != 0)
 837                 return false;
 838
 839         switch (off) {
 840         case bpf_ctx_range(struct bpf_perf_event_data, sample_period):
 841                 bpf_ctx_record_field_size(info, size_u64);
 842                 if (!bpf_ctx_narrow_access_ok(off, size, size_u64))
 843                         return false;
 844                 break;
 845         case bpf_ctx_range(struct bpf_perf_event_data, addr):
 846                 bpf_ctx_record_field_size(info, size_u64);
 847                 if (!bpf_ctx_narrow_access_ok(off, size, size_u64))
 848                         return false;
 849                 break;
 850         default:
 851                 if (size != sizeof(long))
 852                         return false;
 853         }
 854
 855         return true;
 856 }
 857
 858 static u32 pe_prog_convert_ctx_access(enum bpf_access_type type,
 859                                       const struct bpf_insn *si,
 860                                       struct bpf_insn *insn_buf,
 861                                       struct bpf_prog *prog, u32 *target_size)
 862 {
 863         struct bpf_insn *insn = insn_buf;
 864
 865         switch (si->off) {
 866         case offsetof(struct bpf_perf_event_data, sample_period):
 867                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct bpf_perf_event_data_kern,
 868                                                        data), si->dst_reg, si->src_reg,
 869                                       offsetof(struct bpf_perf_event_data_kern, data));
 870                 *insn++ = BPF_LDX_MEM(BPF_DW, si->dst_reg, si->dst_reg,
 871                                       bpf_target_off(struct perf_sample_data, period, 8,
 872                                                      target_size));
 873                 break;
 874         case offsetof(struct bpf_perf_event_data, addr):
 875                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct bpf_perf_event_data_kern,
 876                                                        data), si->dst_reg, si->src_reg,
 877                                       offsetof(struct bpf_perf_event_data_kern, data));
 878                 *insn++ = BPF_LDX_MEM(BPF_DW, si->dst_reg, si->dst_reg,
 879                                       bpf_target_off(struct perf_sample_data, addr, 8,
 880                                                      target_size));
 881                 break;
 882         default:
 883                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct bpf_perf_event_data_kern,
 884                                                        regs), si->dst_reg, si->src_reg,
 885                                       offsetof(struct bpf_perf_event_data_kern, regs));
 886                 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(long), si->dst_reg, si->dst_reg,
 887                                       si->off);
 888                 break;
 889         }
 890
 891         return insn - insn_buf;
 892 }
 893
 894 const struct bpf_verifier_ops perf_event_verifier_ops = {
 895         .get_func_proto         = pe_prog_func_proto,
 896         .is_valid_access        = pe_prog_is_valid_access,
 897         .convert_ctx_access     = pe_prog_convert_ctx_access,
 898 };
 899
 900 const struct bpf_prog_ops perf_event_prog_ops = {
 901 };
 902
 903 static DEFINE_MUTEX(bpf_event_mutex);
 904
 905 #define BPF_TRACE_MAX_PROGS 64
 906
 907 int perf_event_attach_bpf_prog(struct perf_event *event,
 908                                struct bpf_prog *prog)
 909 {
 910         struct bpf_prog_array __rcu *old_array;
 911         struct bpf_prog_array *new_array;
 912         int ret = -EEXIST;
 913
 914         /*
 915          * Kprobe override only works if they are on the function entry,
 916          * and only if they are on the opt-in list.
 917          */
 918         if (prog->kprobe_override &&
 919             (!trace_kprobe_on_func_entry(event->tp_event) ||
 920              !trace_kprobe_error_injectable(event->tp_event)))
 921                 return -EINVAL;
 922
 923         mutex_lock(&bpf_event_mutex);
 924
 925         if (event->prog)
 926                 goto unlock;
 927
 928         old_array = event->tp_event->prog_array;
 929         if (old_array &&
 930             bpf_prog_array_length(old_array) >= BPF_TRACE_MAX_PROGS) {
 931                 ret = -E2BIG;
 932                 goto unlock;
 933         }
 934
 935         ret = bpf_prog_array_copy(old_array, NULL, prog, &new_array);
 936         if (ret < 0)
 937                 goto unlock;
 938
 939         /* set the new array to event->tp_event and set event->prog */
 940         event->prog = prog;
 941         rcu_assign_pointer(event->tp_event->prog_array, new_array);
 942         bpf_prog_array_free(old_array);
 943
 944 unlock:
 945         mutex_unlock(&bpf_event_mutex);
 946         return ret;
 947 }
 948
 949 void perf_event_detach_bpf_prog(struct perf_event *event)
 950 {
 951         struct bpf_prog_array __rcu *old_array;
 952         struct bpf_prog_array *new_array;
 953         int ret;
 954
 955         mutex_lock(&bpf_event_mutex);
 956
 957         if (!event->prog)
 958                 goto unlock;
 959
 960         old_array = event->tp_event->prog_array;
 961         ret = bpf_prog_array_copy(old_array, event->prog, NULL, &new_array);
 962         if (ret < 0) {
 963                 bpf_prog_array_delete_safe(old_array, event->prog);
 964         } else {
 965                 rcu_assign_pointer(event->tp_event->prog_array, new_array);
 966                 bpf_prog_array_free(old_array);
 967         }
 968
 969         bpf_prog_put(event->prog);
 970         event->prog = NULL;
 971
 972 unlock:
 973         mutex_unlock(&bpf_event_mutex);
 974 }
 975
 976 int perf_event_query_prog_array(struct perf_event *event, void __user *info)
 977 {
 978         struct perf_event_query_bpf __user *uquery = info;
 979         struct perf_event_query_bpf query = {};
 980         u32 *ids, prog_cnt, ids_len;
 981         int ret;
 982
 983         if (!capable(CAP_SYS_ADMIN))
 984                 return -EPERM;
 985         if (event->attr.type != PERF_TYPE_TRACEPOINT)
 986                 return -EINVAL;
 987         if (copy_from_user(&query, uquery, sizeof(query)))
 988                 return -EFAULT;
 989
 990         ids_len = query.ids_len;
 991         if (ids_len > BPF_TRACE_MAX_PROGS)
 992                 return -E2BIG;
 993         ids = kcalloc(ids_len, sizeof(u32), GFP_USER | __GFP_NOWARN);
 994         if (!ids)
 995                 return -ENOMEM;
 996         /*
 997          * The above kcalloc returns ZERO_SIZE_PTR when ids_len = 0, which
 998          * is required when user only wants to check for uquery->prog_cnt.
 999          * There is no need to check for it since the case is handled
1000          * gracefully in bpf_prog_array_copy_info.
1001          */
1002
1003         mutex_lock(&bpf_event_mutex);
1004         ret = bpf_prog_array_copy_info(event->tp_event->prog_array,
1005                                        ids,
1006                                        ids_len,
1007                                        &prog_cnt);
1008         mutex_unlock(&bpf_event_mutex);
1009
1010         if (copy_to_user(&uquery->prog_cnt, &prog_cnt, sizeof(prog_cnt)) ||
1011             copy_to_user(uquery->ids, ids, ids_len * sizeof(u32)))
1012                 ret = -EFAULT;
1013
1014         kfree(ids);
1015         return ret;
1016 }
1017
1018 extern struct bpf_raw_event_map __start__bpf_raw_tp[];
1019 extern struct bpf_raw_event_map __stop__bpf_raw_tp[];
1020
1021 struct bpf_raw_event_map *bpf_find_raw_tracepoint(const char *name)
1022 {
1023         struct bpf_raw_event_map *btp = __start__bpf_raw_tp;
1024
1025         for (; btp < __stop__bpf_raw_tp; btp++) {
1026                 if (!strcmp(btp->tp->name, name))
1027                         return btp;
1028         }
1029         return NULL;
1030 }
1031
1032 static __always_inline
1033 void __bpf_trace_run(struct bpf_prog *prog, u64 *args)
1034 {
1035         rcu_read_lock();
1036         preempt_disable();
1037         (void) BPF_PROG_RUN(prog, args);
1038         preempt_enable();
1039         rcu_read_unlock();
1040 }
1041
1042 #define UNPACK(...)                     __VA_ARGS__
1043 #define REPEAT_1(FN, DL, X, ...)        FN(X)
1044 #define REPEAT_2(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_1(FN, DL, __VA_ARGS__)
1045 #define REPEAT_3(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_2(FN, DL, __VA_ARGS__)
1046 #define REPEAT_4(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_3(FN, DL, __VA_ARGS__)
1047 #define REPEAT_5(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_4(FN, DL, __VA_ARGS__)
1048 #define REPEAT_6(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_5(FN, DL, __VA_ARGS__)
1049 #define REPEAT_7(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_6(FN, DL, __VA_ARGS__)
1050 #define REPEAT_8(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_7(FN, DL, __VA_ARGS__)
1051 #define REPEAT_9(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_8(FN, DL, __VA_ARGS__)
1052 #define REPEAT_10(FN, DL, X, ...)       FN(X) UNPACK DL REPEAT_9(FN, DL, __VA_ARGS__)
1053 #define REPEAT_11(FN, DL, X, ...)       FN(X) UNPACK DL REPEAT_10(FN, DL, __VA_ARGS__)
1054 #define REPEAT_12(FN, DL, X, ...)       FN(X) UNPACK DL REPEAT_11(FN, DL, __VA_ARGS__)
1055 #define REPEAT(X, FN, DL, ...)          REPEAT_##X(FN, DL, __VA_ARGS__)
1056
1057 #define SARG(X)         u64 arg##X
1058 #define COPY(X)         args[X] = arg##X
1059
1060 #define __DL_COM        (,)
1061 #define __DL_SEM        (;)
1062
1063 #define __SEQ_0_11      0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11
1064
1065 #define BPF_TRACE_DEFN_x(x)                                             \
1066         void bpf_trace_run##x(struct bpf_prog *prog,                    \
1067                               REPEAT(x, SARG, __DL_COM, __SEQ_0_11))    \
1068         {                                                               \
1069                 u64 args[x];                                            \
1070                 REPEAT(x, COPY, __DL_SEM, __SEQ_0_11);                  \
1071                 __bpf_trace_run(prog, args);                            \
1072         }                                                               \
1073         EXPORT_SYMBOL_GPL(bpf_trace_run##x)
1074 BPF_TRACE_DEFN_x(1);
1075 BPF_TRACE_DEFN_x(2);
1076 BPF_TRACE_DEFN_x(3);
1077 BPF_TRACE_DEFN_x(4);
1078 BPF_TRACE_DEFN_x(5);
1079 BPF_TRACE_DEFN_x(6);
1080 BPF_TRACE_DEFN_x(7);
1081 BPF_TRACE_DEFN_x(8);
1082 BPF_TRACE_DEFN_x(9);
1083 BPF_TRACE_DEFN_x(10);
1084 BPF_TRACE_DEFN_x(11);
1085 BPF_TRACE_DEFN_x(12);
1086
1087 static int __bpf_probe_register(struct bpf_raw_event_map *btp, struct bpf_prog *prog)
1088 {
1089         struct tracepoint *tp = btp->tp;
1090
1091         /*
1092          * check that program doesn't access arguments beyond what's
1093          * available in this tracepoint
1094          */
1095         if (prog->aux->max_ctx_offset > btp->num_args * sizeof(u64))
1096                 return -EINVAL;
1097
1098         return tracepoint_probe_register(tp, (void *)btp->bpf_func, prog);
1099 }
1100
1101 int bpf_probe_register(struct bpf_raw_event_map *btp, struct bpf_prog *prog)
1102 {
1103         int err;
1104
1105         mutex_lock(&bpf_event_mutex);
1106         err = __bpf_probe_register(btp, prog);
1107         mutex_unlock(&bpf_event_mutex);
1108         return err;
1109 }
1110
1111 int bpf_probe_unregister(struct bpf_raw_event_map *btp, struct bpf_prog *prog)
1112 {
1113         int err;
1114
1115         mutex_lock(&bpf_event_mutex);
1116         err = tracepoint_probe_unregister(btp->tp, (void *)btp->bpf_func, prog);
1117         mutex_unlock(&bpf_event_mutex);
1118         return err;
1119 }