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 "trace.h"
  17
  18 u64 bpf_get_stackid(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5);
  19
  20 /**
  21  * trace_call_bpf - invoke BPF program
  22  * @call: tracepoint event
  23  * @ctx: opaque context pointer
  24  *
  25  * kprobe handlers execute BPF programs via this helper.
  26  * Can be used from static tracepoints in the future.
  27  *
  28  * Return: BPF programs always return an integer which is interpreted by
  29  * kprobe handler as:
  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
  33  */
  34 unsigned int trace_call_bpf(struct trace_event_call *call, void *ctx)
  35 {
  36         unsigned int ret;
  37
  38         if (in_nmi()) /* not supported yet */
  39                 return 1;
  40
  41         preempt_disable();
  42
  43         if (unlikely(__this_cpu_inc_return(bpf_prog_active) != 1)) {
  44                 /*
  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,
  48                  * so return zero here
  49                  */
  50                 ret = 0;
  51                 goto out;
  52         }
  53
  54         /*
  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.
  59          *
  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.
  68          */
  69         ret = BPF_PROG_RUN_ARRAY_CHECK(call->prog_array, ctx, BPF_PROG_RUN);
  70
  71  out:
  72         __this_cpu_dec(bpf_prog_active);
  73         preempt_enable();
  74
  75         return ret;
  76 }
  77 EXPORT_SYMBOL_GPL(trace_call_bpf);
  78
  79 BPF_CALL_3(bpf_probe_read, void *, dst, u32, size, const void *, unsafe_ptr)
  80 {
  81         int ret;
  82
  83         ret = probe_kernel_read(dst, unsafe_ptr, size);
  84         if (unlikely(ret < 0))
  85                 memset(dst, 0, size);
  86
  87         return ret;
  88 }
  89
  90 static const struct bpf_func_proto bpf_probe_read_proto = {
  91         .func           = bpf_probe_read,
  92         .gpl_only       = true,
  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,
  97 };
  98
  99 BPF_CALL_3(bpf_probe_write_user, void *, unsafe_ptr, const void *, src,
 100            u32, size)
 101 {
 102         /*
 103          * Ensure we're in user context which is safe for the helper to
 104          * run. This helper has no business in a kthread.
 105          *
 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.
 109          */
 110
 111         if (unlikely(in_interrupt() ||
 112                      current->flags & (PF_KTHREAD | PF_EXITING)))
 113                 return -EPERM;
 114         if (unlikely(uaccess_kernel()))
 115                 return -EPERM;
 116         if (!access_ok(VERIFY_WRITE, unsafe_ptr, size))
 117                 return -EPERM;
 118
 119         return probe_kernel_write(unsafe_ptr, src, size);
 120 }
 121
 122 static const struct bpf_func_proto bpf_probe_write_user_proto = {
 123         .func           = bpf_probe_write_user,
 124         .gpl_only       = true,
 125         .ret_type       = RET_INTEGER,
 126         .arg1_type      = ARG_ANYTHING,
 127         .arg2_type      = ARG_PTR_TO_MEM,
 128         .arg3_type      = ARG_CONST_SIZE,
 129 };
 130
 131 static const struct bpf_func_proto *bpf_get_probe_write_proto(void)
 132 {
 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));
 135
 136         return &bpf_probe_write_user_proto;
 137 }
 138
 139 /*
 140  * Only limited trace_printk() conversion specifiers allowed:
 141  * %d %i %u %x %ld %li %lu %lx %lld %lli %llu %llx %p %s
 142  */
 143 BPF_CALL_5(bpf_trace_printk, char *, fmt, u32, fmt_size, u64, arg1,
 144            u64, arg2, u64, arg3)
 145 {
 146         bool str_seen = false;
 147         int mod[3] = {};
 148         int fmt_cnt = 0;
 149         u64 unsafe_addr;
 150         char buf[64];
 151         int i;
 152
 153         /*
 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
 157          */
 158         if (fmt[--fmt_size] != 0)
 159                 return -EINVAL;
 160
 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]))
 164                         return -EINVAL;
 165
 166                 if (fmt[i] != '%')
 167                         continue;
 168
 169                 if (fmt_cnt >= 3)
 170                         return -EINVAL;
 171
 172                 /* fmt[i] != 0 && fmt[last] == 0, so we can access fmt[i + 1] */
 173                 i++;
 174                 if (fmt[i] == 'l') {
 175                         mod[fmt_cnt]++;
 176                         i++;
 177                 } else if (fmt[i] == 'p' || fmt[i] == 's') {
 178                         mod[fmt_cnt]++;
 179                         i++;
 180                         if (!isspace(fmt[i]) && !ispunct(fmt[i]) && fmt[i] != 0)
 181                                 return -EINVAL;
 182                         fmt_cnt++;
 183                         if (fmt[i - 1] == 's') {
 184                                 if (str_seen)
 185                                         /* allow only one '%s' per fmt string */
 186                                         return -EINVAL;
 187                                 str_seen = true;
 188
 189                                 switch (fmt_cnt) {
 190                                 case 1:
 191                                         unsafe_addr = arg1;
 192                                         arg1 = (long) buf;
 193                                         break;
 194                                 case 2:
 195                                         unsafe_addr = arg2;
 196                                         arg2 = (long) buf;
 197                                         break;
 198                                 case 3:
 199                                         unsafe_addr = arg3;
 200                                         arg3 = (long) buf;
 201                                         break;
 202                                 }
 203                                 buf[0] = 0;
 204                                 strncpy_from_unsafe(buf,
 205                                                     (void *) (long) unsafe_addr,
 206                                                     sizeof(buf));
 207                         }
 208                         continue;
 209                 }
 210
 211                 if (fmt[i] == 'l') {
 212                         mod[fmt_cnt]++;
 213                         i++;
 214                 }
 215
 216                 if (fmt[i] != 'i' && fmt[i] != 'd' &&
 217                     fmt[i] != 'u' && fmt[i] != 'x')
 218                         return -EINVAL;
 219                 fmt_cnt++;
 220         }
 221
 222 /* Horrid workaround for getting va_list handling working with different
 223  * argument type combinations generically for 32 and 64 bit archs.
 224  */
 225 #define __BPF_TP_EMIT() __BPF_ARG3_TP()
 226 #define __BPF_TP(...)                                                   \
 227         __trace_printk(1 /* Fake ip will not be printed. */,            \
 228                        fmt, ##__VA_ARGS__)
 229
 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__)))
 236
 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__)))
 243
 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__)))
 250
 251         return __BPF_TP_EMIT();
 252 }
 253
 254 static const struct bpf_func_proto bpf_trace_printk_proto = {
 255         .func           = bpf_trace_printk,
 256         .gpl_only       = true,
 257         .ret_type       = RET_INTEGER,
 258         .arg1_type      = ARG_PTR_TO_MEM,
 259         .arg2_type      = ARG_CONST_SIZE,
 260 };
 261
 262 const struct bpf_func_proto *bpf_get_trace_printk_proto(void)
 263 {
 264         /*
 265          * this program might be calling bpf_trace_printk,
 266          * so allocate per-cpu printk buffers
 267          */
 268         trace_printk_init_buffers();
 269
 270         return &bpf_trace_printk_proto;
 271 }
 272
 273 static __always_inline int
 274 get_map_perf_counter(struct bpf_map *map, u64 flags,
 275                      u64 *value, u64 *enabled, u64 *running)
 276 {
 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;
 281
 282         if (unlikely(flags & ~(BPF_F_INDEX_MASK)))
 283                 return -EINVAL;
 284         if (index == BPF_F_CURRENT_CPU)
 285                 index = cpu;
 286         if (unlikely(index >= array->map.max_entries))
 287                 return -E2BIG;
 288
 289         ee = READ_ONCE(array->ptrs[index]);
 290         if (!ee)
 291                 return -ENOENT;
 292
 293         return perf_event_read_local(ee->event, value, enabled, running);
 294 }
 295
 296 BPF_CALL_2(bpf_perf_event_read, struct bpf_map *, map, u64, flags)
 297 {
 298         u64 value = 0;
 299         int err;
 300
 301         err = get_map_perf_counter(map, flags, &value, NULL, NULL);
 302         /*
 303          * this api is ugly since we miss [-22..-2] range of valid
 304          * counter values, but that's uapi
 305          */
 306         if (err)
 307                 return err;
 308         return value;
 309 }
 310
 311 static const struct bpf_func_proto bpf_perf_event_read_proto = {
 312         .func           = bpf_perf_event_read,
 313         .gpl_only       = true,
 314         .ret_type       = RET_INTEGER,
 315         .arg1_type      = ARG_CONST_MAP_PTR,
 316         .arg2_type      = ARG_ANYTHING,
 317 };
 318
 319 BPF_CALL_4(bpf_perf_event_read_value, struct bpf_map *, map, u64, flags,
 320            struct bpf_perf_event_value *, buf, u32, size)
 321 {
 322         int err = -EINVAL;
 323
 324         if (unlikely(size != sizeof(struct bpf_perf_event_value)))
 325                 goto clear;
 326         err = get_map_perf_counter(map, flags, &buf->counter, &buf->enabled,
 327                                    &buf->running);
 328         if (unlikely(err))
 329                 goto clear;
 330         return 0;
 331 clear:
 332         memset(buf, 0, size);
 333         return err;
 334 }
 335
 336 static const struct bpf_func_proto bpf_perf_event_read_value_proto = {
 337         .func           = bpf_perf_event_read_value,
 338         .gpl_only       = true,
 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,
 344 };
 345
 346 static DEFINE_PER_CPU(struct perf_sample_data, bpf_trace_sd);
 347
 348 static __always_inline u64
 349 __bpf_perf_event_output(struct pt_regs *regs, struct bpf_map *map,
 350                         u64 flags, struct perf_sample_data *sd)
 351 {
 352         struct bpf_array *array = container_of(map, struct bpf_array, map);
 353         unsigned int cpu = smp_processor_id();
 354         u64 index = flags & BPF_F_INDEX_MASK;
 355         struct bpf_event_entry *ee;
 356         struct perf_event *event;
 357
 358         if (index == BPF_F_CURRENT_CPU)
 359                 index = cpu;
 360         if (unlikely(index >= array->map.max_entries))
 361                 return -E2BIG;
 362
 363         ee = READ_ONCE(array->ptrs[index]);
 364         if (!ee)
 365                 return -ENOENT;
 366
 367         event = ee->event;
 368         if (unlikely(event->attr.type != PERF_TYPE_SOFTWARE ||
 369                      event->attr.config != PERF_COUNT_SW_BPF_OUTPUT))
 370                 return -EINVAL;
 371
 372         if (unlikely(event->oncpu != cpu))
 373                 return -EOPNOTSUPP;
 374
 375         perf_event_output(event, sd, regs);
 376         return 0;
 377 }
 378
 379 BPF_CALL_5(bpf_perf_event_output, struct pt_regs *, regs, struct bpf_map *, map,
 380            u64, flags, void *, data, u64, size)
 381 {
 382         struct perf_sample_data *sd = this_cpu_ptr(&bpf_trace_sd);
 383         struct perf_raw_record raw = {
 384                 .frag = {
 385                         .size = size,
 386                         .data = data,
 387                 },
 388         };
 389
 390         if (unlikely(flags & ~(BPF_F_INDEX_MASK)))
 391                 return -EINVAL;
 392
 393         perf_sample_data_init(sd, 0, 0);
 394         sd->raw = &raw;
 395
 396         return __bpf_perf_event_output(regs, map, flags, sd);
 397 }
 398
 399 static const struct bpf_func_proto bpf_perf_event_output_proto = {
 400         .func           = bpf_perf_event_output,
 401         .gpl_only       = true,
 402         .ret_type       = RET_INTEGER,
 403         .arg1_type      = ARG_PTR_TO_CTX,
 404         .arg2_type      = ARG_CONST_MAP_PTR,
 405         .arg3_type      = ARG_ANYTHING,
 406         .arg4_type      = ARG_PTR_TO_MEM,
 407         .arg5_type      = ARG_CONST_SIZE_OR_ZERO,
 408 };
 409
 410 static DEFINE_PER_CPU(struct pt_regs, bpf_pt_regs);
 411 static DEFINE_PER_CPU(struct perf_sample_data, bpf_misc_sd);
 412
 413 u64 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
 414                      void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
 415 {
 416         struct perf_sample_data *sd = this_cpu_ptr(&bpf_misc_sd);
 417         struct pt_regs *regs = this_cpu_ptr(&bpf_pt_regs);
 418         struct perf_raw_frag frag = {
 419                 .copy           = ctx_copy,
 420                 .size           = ctx_size,
 421                 .data           = ctx,
 422         };
 423         struct perf_raw_record raw = {
 424                 .frag = {
 425                         {
 426                                 .next   = ctx_size ? &frag : NULL,
 427                         },
 428                         .size   = meta_size,
 429                         .data   = meta,
 430                 },
 431         };
 432
 433         perf_fetch_caller_regs(regs);
 434         perf_sample_data_init(sd, 0, 0);
 435         sd->raw = &raw;
 436
 437         return __bpf_perf_event_output(regs, map, flags, sd);
 438 }
 439
 440 BPF_CALL_0(bpf_get_current_task)
 441 {
 442         return (long) current;
 443 }
 444
 445 static const struct bpf_func_proto bpf_get_current_task_proto = {
 446         .func           = bpf_get_current_task,
 447         .gpl_only       = true,
 448         .ret_type       = RET_INTEGER,
 449 };
 450
 451 BPF_CALL_2(bpf_current_task_under_cgroup, struct bpf_map *, map, u32, idx)
 452 {
 453         struct bpf_array *array = container_of(map, struct bpf_array, map);
 454         struct cgroup *cgrp;
 455
 456         if (unlikely(in_interrupt()))
 457                 return -EINVAL;
 458         if (unlikely(idx >= array->map.max_entries))
 459                 return -E2BIG;
 460
 461         cgrp = READ_ONCE(array->ptrs[idx]);
 462         if (unlikely(!cgrp))
 463                 return -EAGAIN;
 464
 465         return task_under_cgroup_hierarchy(current, cgrp);
 466 }
 467
 468 static const struct bpf_func_proto bpf_current_task_under_cgroup_proto = {
 469         .func           = bpf_current_task_under_cgroup,
 470         .gpl_only       = false,
 471         .ret_type       = RET_INTEGER,
 472         .arg1_type      = ARG_CONST_MAP_PTR,
 473         .arg2_type      = ARG_ANYTHING,
 474 };
 475
 476 BPF_CALL_3(bpf_probe_read_str, void *, dst, u32, size,
 477            const void *, unsafe_ptr)
 478 {
 479         int ret;
 480
 481         /*
 482          * The strncpy_from_unsafe() call will likely not fill the entire
 483          * buffer, but that's okay in this circumstance as we're probing
 484          * arbitrary memory anyway similar to bpf_probe_read() and might
 485          * as well probe the stack. Thus, memory is explicitly cleared
 486          * only in error case, so that improper users ignoring return
 487          * code altogether don't copy garbage; otherwise length of string
 488          * is returned that can be used for bpf_perf_event_output() et al.
 489          */
 490         ret = strncpy_from_unsafe(dst, unsafe_ptr, size);
 491         if (unlikely(ret < 0))
 492                 memset(dst, 0, size);
 493
 494         return ret;
 495 }
 496
 497 static const struct bpf_func_proto bpf_probe_read_str_proto = {
 498         .func           = bpf_probe_read_str,
 499         .gpl_only       = true,
 500         .ret_type       = RET_INTEGER,
 501         .arg1_type      = ARG_PTR_TO_UNINIT_MEM,
 502         .arg2_type      = ARG_CONST_SIZE_OR_ZERO,
 503         .arg3_type      = ARG_ANYTHING,
 504 };
 505
 506 static const struct bpf_func_proto *tracing_func_proto(enum bpf_func_id func_id)
 507 {
 508         switch (func_id) {
 509         case BPF_FUNC_map_lookup_elem:
 510                 return &bpf_map_lookup_elem_proto;
 511         case BPF_FUNC_map_update_elem:
 512                 return &bpf_map_update_elem_proto;
 513         case BPF_FUNC_map_delete_elem:
 514                 return &bpf_map_delete_elem_proto;
 515         case BPF_FUNC_probe_read:
 516                 return &bpf_probe_read_proto;
 517         case BPF_FUNC_ktime_get_ns:
 518                 return &bpf_ktime_get_ns_proto;
 519         case BPF_FUNC_tail_call:
 520                 return &bpf_tail_call_proto;
 521         case BPF_FUNC_get_current_pid_tgid:
 522                 return &bpf_get_current_pid_tgid_proto;
 523         case BPF_FUNC_get_current_task:
 524                 return &bpf_get_current_task_proto;
 525         case BPF_FUNC_get_current_uid_gid:
 526                 return &bpf_get_current_uid_gid_proto;
 527         case BPF_FUNC_get_current_comm:
 528                 return &bpf_get_current_comm_proto;
 529         case BPF_FUNC_trace_printk:
 530                 return bpf_get_trace_printk_proto();
 531         case BPF_FUNC_get_smp_processor_id:
 532                 return &bpf_get_smp_processor_id_proto;
 533         case BPF_FUNC_get_numa_node_id:
 534                 return &bpf_get_numa_node_id_proto;
 535         case BPF_FUNC_perf_event_read:
 536                 return &bpf_perf_event_read_proto;
 537         case BPF_FUNC_probe_write_user:
 538                 return bpf_get_probe_write_proto();
 539         case BPF_FUNC_current_task_under_cgroup:
 540                 return &bpf_current_task_under_cgroup_proto;
 541         case BPF_FUNC_get_prandom_u32:
 542                 return &bpf_get_prandom_u32_proto;
 543         case BPF_FUNC_probe_read_str:
 544                 return &bpf_probe_read_str_proto;
 545         default:
 546                 return NULL;
 547         }
 548 }
 549
 550 static const struct bpf_func_proto *kprobe_prog_func_proto(enum bpf_func_id func_id)
 551 {
 552         switch (func_id) {
 553         case BPF_FUNC_perf_event_output:
 554                 return &bpf_perf_event_output_proto;
 555         case BPF_FUNC_get_stackid:
 556                 return &bpf_get_stackid_proto;
 557         case BPF_FUNC_perf_event_read_value:
 558                 return &bpf_perf_event_read_value_proto;
 559         default:
 560                 return tracing_func_proto(func_id);
 561         }
 562 }
 563
 564 /* bpf+kprobe programs can access fields of 'struct pt_regs' */
 565 static bool kprobe_prog_is_valid_access(int off, int size, enum bpf_access_type type,
 566                                         struct bpf_insn_access_aux *info)
 567 {
 568         if (off < 0 || off >= sizeof(struct pt_regs))
 569                 return false;
 570         if (type != BPF_READ)
 571                 return false;
 572         if (off % size != 0)
 573                 return false;
 574         /*
 575          * Assertion for 32 bit to make sure last 8 byte access
 576          * (BPF_DW) to the last 4 byte member is disallowed.
 577          */
 578         if (off + size > sizeof(struct pt_regs))
 579                 return false;
 580
 581         return true;
 582 }
 583
 584 const struct bpf_verifier_ops kprobe_verifier_ops = {
 585         .get_func_proto  = kprobe_prog_func_proto,
 586         .is_valid_access = kprobe_prog_is_valid_access,
 587 };
 588
 589 const struct bpf_prog_ops kprobe_prog_ops = {
 590 };
 591
 592 BPF_CALL_5(bpf_perf_event_output_tp, void *, tp_buff, struct bpf_map *, map,
 593            u64, flags, void *, data, u64, size)
 594 {
 595         struct pt_regs *regs = *(struct pt_regs **)tp_buff;
 596
 597         /*
 598          * r1 points to perf tracepoint buffer where first 8 bytes are hidden
 599          * from bpf program and contain a pointer to 'struct pt_regs'. Fetch it
 600          * from there and call the same bpf_perf_event_output() helper inline.
 601          */
 602         return ____bpf_perf_event_output(regs, map, flags, data, size);
 603 }
 604
 605 static const struct bpf_func_proto bpf_perf_event_output_proto_tp = {
 606         .func           = bpf_perf_event_output_tp,
 607         .gpl_only       = true,
 608         .ret_type       = RET_INTEGER,
 609         .arg1_type      = ARG_PTR_TO_CTX,
 610         .arg2_type      = ARG_CONST_MAP_PTR,
 611         .arg3_type      = ARG_ANYTHING,
 612         .arg4_type      = ARG_PTR_TO_MEM,
 613         .arg5_type      = ARG_CONST_SIZE_OR_ZERO,
 614 };
 615
 616 BPF_CALL_3(bpf_get_stackid_tp, void *, tp_buff, struct bpf_map *, map,
 617            u64, flags)
 618 {
 619         struct pt_regs *regs = *(struct pt_regs **)tp_buff;
 620
 621         /*
 622          * Same comment as in bpf_perf_event_output_tp(), only that this time
 623          * the other helper's function body cannot be inlined due to being
 624          * external, thus we need to call raw helper function.
 625          */
 626         return bpf_get_stackid((unsigned long) regs, (unsigned long) map,
 627                                flags, 0, 0);
 628 }
 629
 630 static const struct bpf_func_proto bpf_get_stackid_proto_tp = {
 631         .func           = bpf_get_stackid_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 };
 638
 639 BPF_CALL_3(bpf_perf_prog_read_value_tp, struct bpf_perf_event_data_kern *, ctx,
 640            struct bpf_perf_event_value *, buf, u32, size)
 641 {
 642         int err = -EINVAL;
 643
 644         if (unlikely(size != sizeof(struct bpf_perf_event_value)))
 645                 goto clear;
 646         err = perf_event_read_local(ctx->event, &buf->counter, &buf->enabled,
 647                                     &buf->running);
 648         if (unlikely(err))
 649                 goto clear;
 650         return 0;
 651 clear:
 652         memset(buf, 0, size);
 653         return err;
 654 }
 655
 656 static const struct bpf_func_proto bpf_perf_prog_read_value_proto_tp = {
 657          .func           = bpf_perf_prog_read_value_tp,
 658          .gpl_only       = true,
 659          .ret_type       = RET_INTEGER,
 660          .arg1_type      = ARG_PTR_TO_CTX,
 661          .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
 662          .arg3_type      = ARG_CONST_SIZE,
 663 };
 664
 665 static const struct bpf_func_proto *tp_prog_func_proto(enum bpf_func_id func_id)
 666 {
 667         switch (func_id) {
 668         case BPF_FUNC_perf_event_output:
 669                 return &bpf_perf_event_output_proto_tp;
 670         case BPF_FUNC_get_stackid:
 671                 return &bpf_get_stackid_proto_tp;
 672         case BPF_FUNC_perf_prog_read_value:
 673                 return &bpf_perf_prog_read_value_proto_tp;
 674         default:
 675                 return tracing_func_proto(func_id);
 676         }
 677 }
 678
 679 static bool tp_prog_is_valid_access(int off, int size, enum bpf_access_type type,
 680                                     struct bpf_insn_access_aux *info)
 681 {
 682         if (off < sizeof(void *) || off >= PERF_MAX_TRACE_SIZE)
 683                 return false;
 684         if (type != BPF_READ)
 685                 return false;
 686         if (off % size != 0)
 687                 return false;
 688
 689         BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(__u64));
 690         return true;
 691 }
 692
 693 const struct bpf_verifier_ops tracepoint_verifier_ops = {
 694         .get_func_proto  = tp_prog_func_proto,
 695         .is_valid_access = tp_prog_is_valid_access,
 696 };
 697
 698 const struct bpf_prog_ops tracepoint_prog_ops = {
 699 };
 700
 701 static bool pe_prog_is_valid_access(int off, int size, enum bpf_access_type type,
 702                                     struct bpf_insn_access_aux *info)
 703 {
 704         const int size_sp = FIELD_SIZEOF(struct bpf_perf_event_data,
 705                                          sample_period);
 706
 707         if (off < 0 || off >= sizeof(struct bpf_perf_event_data))
 708                 return false;
 709         if (type != BPF_READ)
 710                 return false;
 711         if (off % size != 0)
 712                 return false;
 713
 714         switch (off) {
 715         case bpf_ctx_range(struct bpf_perf_event_data, sample_period):
 716                 bpf_ctx_record_field_size(info, size_sp);
 717                 if (!bpf_ctx_narrow_access_ok(off, size, size_sp))
 718                         return false;
 719                 break;
 720         default:
 721                 if (size != sizeof(long))
 722                         return false;
 723         }
 724
 725         return true;
 726 }
 727
 728 static u32 pe_prog_convert_ctx_access(enum bpf_access_type type,
 729                                       const struct bpf_insn *si,
 730                                       struct bpf_insn *insn_buf,
 731                                       struct bpf_prog *prog, u32 *target_size)
 732 {
 733         struct bpf_insn *insn = insn_buf;
 734
 735         switch (si->off) {
 736         case offsetof(struct bpf_perf_event_data, sample_period):
 737                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct bpf_perf_event_data_kern,
 738                                                        data), si->dst_reg, si->src_reg,
 739                                       offsetof(struct bpf_perf_event_data_kern, data));
 740                 *insn++ = BPF_LDX_MEM(BPF_DW, si->dst_reg, si->dst_reg,
 741                                       bpf_target_off(struct perf_sample_data, period, 8,
 742                                                      target_size));
 743                 break;
 744         default:
 745                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct bpf_perf_event_data_kern,
 746                                                        regs), si->dst_reg, si->src_reg,
 747                                       offsetof(struct bpf_perf_event_data_kern, regs));
 748                 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(long), si->dst_reg, si->dst_reg,
 749                                       si->off);
 750                 break;
 751         }
 752
 753         return insn - insn_buf;
 754 }
 755
 756 const struct bpf_verifier_ops perf_event_verifier_ops = {
 757         .get_func_proto         = tp_prog_func_proto,
 758         .is_valid_access        = pe_prog_is_valid_access,
 759         .convert_ctx_access     = pe_prog_convert_ctx_access,
 760 };
 761
 762 const struct bpf_prog_ops perf_event_prog_ops = {
 763 };
 764
 765 static DEFINE_MUTEX(bpf_event_mutex);
 766
 767 #define BPF_TRACE_MAX_PROGS 64
 768
 769 int perf_event_attach_bpf_prog(struct perf_event *event,
 770                                struct bpf_prog *prog)
 771 {
 772         struct bpf_prog_array __rcu *old_array;
 773         struct bpf_prog_array *new_array;
 774         int ret = -EEXIST;
 775
 776         mutex_lock(&bpf_event_mutex);
 777
 778         if (event->prog)
 779                 goto unlock;
 780
 781         old_array = event->tp_event->prog_array;
 782         if (old_array &&
 783             bpf_prog_array_length(old_array) >= BPF_TRACE_MAX_PROGS) {
 784                 ret = -E2BIG;
 785                 goto unlock;
 786         }
 787
 788         ret = bpf_prog_array_copy(old_array, NULL, prog, &new_array);
 789         if (ret < 0)
 790                 goto unlock;
 791
 792         /* set the new array to event->tp_event and set event->prog */
 793         event->prog = prog;
 794         rcu_assign_pointer(event->tp_event->prog_array, new_array);
 795         bpf_prog_array_free(old_array);
 796
 797 unlock:
 798         mutex_unlock(&bpf_event_mutex);
 799         return ret;
 800 }
 801
 802 void perf_event_detach_bpf_prog(struct perf_event *event)
 803 {
 804         struct bpf_prog_array __rcu *old_array;
 805         struct bpf_prog_array *new_array;
 806         int ret;
 807
 808         mutex_lock(&bpf_event_mutex);
 809
 810         if (!event->prog)
 811                 goto unlock;
 812
 813         old_array = event->tp_event->prog_array;
 814         ret = bpf_prog_array_copy(old_array, event->prog, NULL, &new_array);
 815         if (ret < 0) {
 816                 bpf_prog_array_delete_safe(old_array, event->prog);
 817         } else {
 818                 rcu_assign_pointer(event->tp_event->prog_array, new_array);
 819                 bpf_prog_array_free(old_array);
 820         }
 821
 822         bpf_prog_put(event->prog);
 823         event->prog = NULL;
 824
 825 unlock:
 826         mutex_unlock(&bpf_event_mutex);
 827 }