kernel/trace/bpf_trace.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Copyright (c) 2011-2015 PLUMgrid, http://plumgrid.com
   3  * Copyright (c) 2016 Facebook
   4  */
   5 #include <linux/kernel.h>
   6 #include <linux/types.h>
   7 #include <linux/slab.h>
   8 #include <linux/bpf.h>
   9 #include <linux/bpf_perf_event.h>
  10 #include <linux/filter.h>
  11 #include <linux/uaccess.h>
  12 #include <linux/ctype.h>
  13 #include <linux/kprobes.h>
  14 #include <linux/syscalls.h>
  15 #include <linux/error-injection.h>
  16
  17 #include <asm/tlb.h>
  18
  19 #include "trace_probe.h"
  20 #include "trace.h"
  21
  22 #define bpf_event_rcu_dereference(p)                                    \
  23         rcu_dereference_protected(p, lockdep_is_held(&bpf_event_mutex))
  24
  25 #ifdef CONFIG_MODULES
  26 struct bpf_trace_module {
  27         struct module *module;
  28         struct list_head list;
  29 };
  30
  31 static LIST_HEAD(bpf_trace_modules);
  32 static DEFINE_MUTEX(bpf_module_mutex);
  33
  34 static struct bpf_raw_event_map *bpf_get_raw_tracepoint_module(const char *name)
  35 {
  36         struct bpf_raw_event_map *btp, *ret = NULL;
  37         struct bpf_trace_module *btm;
  38         unsigned int i;
  39
  40         mutex_lock(&bpf_module_mutex);
  41         list_for_each_entry(btm, &bpf_trace_modules, list) {
  42                 for (i = 0; i < btm->module->num_bpf_raw_events; ++i) {
  43                         btp = &btm->module->bpf_raw_events[i];
  44                         if (!strcmp(btp->tp->name, name)) {
  45                                 if (try_module_get(btm->module))
  46                                         ret = btp;
  47                                 goto out;
  48                         }
  49                 }
  50         }
  51 out:
  52         mutex_unlock(&bpf_module_mutex);
  53         return ret;
  54 }
  55 #else
  56 static struct bpf_raw_event_map *bpf_get_raw_tracepoint_module(const char *name)
  57 {
  58         return NULL;
  59 }
  60 #endif /* CONFIG_MODULES */
  61
  62 u64 bpf_get_stackid(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5);
  63 u64 bpf_get_stack(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5);
  64
  65 /**
  66  * trace_call_bpf - invoke BPF program
  67  * @call: tracepoint event
  68  * @ctx: opaque context pointer
  69  *
  70  * kprobe handlers execute BPF programs via this helper.
  71  * Can be used from static tracepoints in the future.
  72  *
  73  * Return: BPF programs always return an integer which is interpreted by
  74  * kprobe handler as:
  75  * 0 - return from kprobe (event is filtered out)
  76  * 1 - store kprobe event into ring buffer
  77  * Other values are reserved and currently alias to 1
  78  */
  79 unsigned int trace_call_bpf(struct trace_event_call *call, void *ctx)
  80 {
  81         unsigned int ret;
  82
  83         if (in_nmi()) /* not supported yet */
  84                 return 1;
  85
  86         cant_sleep();
  87
  88         if (unlikely(__this_cpu_inc_return(bpf_prog_active) != 1)) {
  89                 /*
  90                  * since some bpf program is already running on this cpu,
  91                  * don't call into another bpf program (same or different)
  92                  * and don't send kprobe event into ring-buffer,
  93                  * so return zero here
  94                  */
  95                 ret = 0;
  96                 goto out;
  97         }
  98
  99         /*
 100          * Instead of moving rcu_read_lock/rcu_dereference/rcu_read_unlock
 101          * to all call sites, we did a bpf_prog_array_valid() there to check
 102          * whether call->prog_array is empty or not, which is
 103          * a heurisitc to speed up execution.
 104          *
 105          * If bpf_prog_array_valid() fetched prog_array was
 106          * non-NULL, we go into trace_call_bpf() and do the actual
 107          * proper rcu_dereference() under RCU lock.
 108          * If it turns out that prog_array is NULL then, we bail out.
 109          * For the opposite, if the bpf_prog_array_valid() fetched pointer
 110          * was NULL, you'll skip the prog_array with the risk of missing
 111          * out of events when it was updated in between this and the
 112          * rcu_dereference() which is accepted risk.
 113          */
 114         ret = BPF_PROG_RUN_ARRAY_CHECK(call->prog_array, ctx, BPF_PROG_RUN);
 115
 116  out:
 117         __this_cpu_dec(bpf_prog_active);
 118
 119         return ret;
 120 }
 121
 122 #ifdef CONFIG_BPF_KPROBE_OVERRIDE
 123 BPF_CALL_2(bpf_override_return, struct pt_regs *, regs, unsigned long, rc)
 124 {
 125         regs_set_return_value(regs, rc);
 126         override_function_with_return(regs);
 127         return 0;
 128 }
 129
 130 static const struct bpf_func_proto bpf_override_return_proto = {
 131         .func           = bpf_override_return,
 132         .gpl_only       = true,
 133         .ret_type       = RET_INTEGER,
 134         .arg1_type      = ARG_PTR_TO_CTX,
 135         .arg2_type      = ARG_ANYTHING,
 136 };
 137 #endif
 138
 139 BPF_CALL_3(bpf_probe_read_user, void *, dst, u32, size,
 140            const void __user *, unsafe_ptr)
 141 {
 142         int ret = probe_user_read(dst, unsafe_ptr, size);
 143
 144         if (unlikely(ret < 0))
 145                 memset(dst, 0, size);
 146
 147         return ret;
 148 }
 149
 150 static const struct bpf_func_proto bpf_probe_read_user_proto = {
 151         .func           = bpf_probe_read_user,
 152         .gpl_only       = true,
 153         .ret_type       = RET_INTEGER,
 154         .arg1_type      = ARG_PTR_TO_UNINIT_MEM,
 155         .arg2_type      = ARG_CONST_SIZE_OR_ZERO,
 156         .arg3_type      = ARG_ANYTHING,
 157 };
 158
 159 BPF_CALL_3(bpf_probe_read_user_str, void *, dst, u32, size,
 160            const void __user *, unsafe_ptr)
 161 {
 162         int ret = strncpy_from_unsafe_user(dst, unsafe_ptr, size);
 163
 164         if (unlikely(ret < 0))
 165                 memset(dst, 0, size);
 166
 167         return ret;
 168 }
 169
 170 static const struct bpf_func_proto bpf_probe_read_user_str_proto = {
 171         .func           = bpf_probe_read_user_str,
 172         .gpl_only       = true,
 173         .ret_type       = RET_INTEGER,
 174         .arg1_type      = ARG_PTR_TO_UNINIT_MEM,
 175         .arg2_type      = ARG_CONST_SIZE_OR_ZERO,
 176         .arg3_type      = ARG_ANYTHING,
 177 };
 178
 179 static __always_inline int
 180 bpf_probe_read_kernel_common(void *dst, u32 size, const void *unsafe_ptr,
 181                              const bool compat)
 182 {
 183         int ret = security_locked_down(LOCKDOWN_BPF_READ);
 184
 185         if (unlikely(ret < 0))
 186                 goto out;
 187         ret = compat ? probe_kernel_read(dst, unsafe_ptr, size) :
 188               probe_kernel_read_strict(dst, unsafe_ptr, size);
 189         if (unlikely(ret < 0))
 190 out:
 191                 memset(dst, 0, size);
 192         return ret;
 193 }
 194
 195 BPF_CALL_3(bpf_probe_read_kernel, void *, dst, u32, size,
 196            const void *, unsafe_ptr)
 197 {
 198         return bpf_probe_read_kernel_common(dst, size, unsafe_ptr, false);
 199 }
 200
 201 static const struct bpf_func_proto bpf_probe_read_kernel_proto = {
 202         .func           = bpf_probe_read_kernel,
 203         .gpl_only       = true,
 204         .ret_type       = RET_INTEGER,
 205         .arg1_type      = ARG_PTR_TO_UNINIT_MEM,
 206         .arg2_type      = ARG_CONST_SIZE_OR_ZERO,
 207         .arg3_type      = ARG_ANYTHING,
 208 };
 209
 210 BPF_CALL_3(bpf_probe_read_compat, void *, dst, u32, size,
 211            const void *, unsafe_ptr)
 212 {
 213         return bpf_probe_read_kernel_common(dst, size, unsafe_ptr, true);
 214 }
 215
 216 static const struct bpf_func_proto bpf_probe_read_compat_proto = {
 217         .func           = bpf_probe_read_compat,
 218         .gpl_only       = true,
 219         .ret_type       = RET_INTEGER,
 220         .arg1_type      = ARG_PTR_TO_UNINIT_MEM,
 221         .arg2_type      = ARG_CONST_SIZE_OR_ZERO,
 222         .arg3_type      = ARG_ANYTHING,
 223 };
 224
 225 static __always_inline int
 226 bpf_probe_read_kernel_str_common(void *dst, u32 size, const void *unsafe_ptr,
 227                                  const bool compat)
 228 {
 229         int ret = security_locked_down(LOCKDOWN_BPF_READ);
 230
 231         if (unlikely(ret < 0))
 232                 goto out;
 233         /*
 234          * The strncpy_from_unsafe_*() call will likely not fill the entire
 235          * buffer, but that's okay in this circumstance as we're probing
 236          * arbitrary memory anyway similar to bpf_probe_read_*() and might
 237          * as well probe the stack. Thus, memory is explicitly cleared
 238          * only in error case, so that improper users ignoring return
 239          * code altogether don't copy garbage; otherwise length of string
 240          * is returned that can be used for bpf_perf_event_output() et al.
 241          */
 242         ret = compat ? strncpy_from_unsafe(dst, unsafe_ptr, size) :
 243               strncpy_from_unsafe_strict(dst, unsafe_ptr, size);
 244         if (unlikely(ret < 0))
 245 out:
 246                 memset(dst, 0, size);
 247         return ret;
 248 }
 249
 250 BPF_CALL_3(bpf_probe_read_kernel_str, void *, dst, u32, size,
 251            const void *, unsafe_ptr)
 252 {
 253         return bpf_probe_read_kernel_str_common(dst, size, unsafe_ptr, false);
 254 }
 255
 256 static const struct bpf_func_proto bpf_probe_read_kernel_str_proto = {
 257         .func           = bpf_probe_read_kernel_str,
 258         .gpl_only       = true,
 259         .ret_type       = RET_INTEGER,
 260         .arg1_type      = ARG_PTR_TO_UNINIT_MEM,
 261         .arg2_type      = ARG_CONST_SIZE_OR_ZERO,
 262         .arg3_type      = ARG_ANYTHING,
 263 };
 264
 265 BPF_CALL_3(bpf_probe_read_compat_str, void *, dst, u32, size,
 266            const void *, unsafe_ptr)
 267 {
 268         return bpf_probe_read_kernel_str_common(dst, size, unsafe_ptr, true);
 269 }
 270
 271 static const struct bpf_func_proto bpf_probe_read_compat_str_proto = {
 272         .func           = bpf_probe_read_compat_str,
 273         .gpl_only       = true,
 274         .ret_type       = RET_INTEGER,
 275         .arg1_type      = ARG_PTR_TO_UNINIT_MEM,
 276         .arg2_type      = ARG_CONST_SIZE_OR_ZERO,
 277         .arg3_type      = ARG_ANYTHING,
 278 };
 279
 280 BPF_CALL_3(bpf_probe_write_user, void __user *, unsafe_ptr, const void *, src,
 281            u32, size)
 282 {
 283         /*
 284          * Ensure we're in user context which is safe for the helper to
 285          * run. This helper has no business in a kthread.
 286          *
 287          * access_ok() should prevent writing to non-user memory, but in
 288          * some situations (nommu, temporary switch, etc) access_ok() does
 289          * not provide enough validation, hence the check on KERNEL_DS.
 290          *
 291          * nmi_uaccess_okay() ensures the probe is not run in an interim
 292          * state, when the task or mm are switched. This is specifically
 293          * required to prevent the use of temporary mm.
 294          */
 295
 296         if (unlikely(in_interrupt() ||
 297                      current->flags & (PF_KTHREAD | PF_EXITING)))
 298                 return -EPERM;
 299         if (unlikely(uaccess_kernel()))
 300                 return -EPERM;
 301         if (unlikely(!nmi_uaccess_okay()))
 302                 return -EPERM;
 303
 304         return probe_user_write(unsafe_ptr, src, size);
 305 }
 306
 307 static const struct bpf_func_proto bpf_probe_write_user_proto = {
 308         .func           = bpf_probe_write_user,
 309         .gpl_only       = true,
 310         .ret_type       = RET_INTEGER,
 311         .arg1_type      = ARG_ANYTHING,
 312         .arg2_type      = ARG_PTR_TO_MEM,
 313         .arg3_type      = ARG_CONST_SIZE,
 314 };
 315
 316 static const struct bpf_func_proto *bpf_get_probe_write_proto(void)
 317 {
 318         pr_warn_ratelimited("%s[%d] is installing a program with bpf_probe_write_user helper that may corrupt user memory!",
 319                             current->comm, task_pid_nr(current));
 320
 321         return &bpf_probe_write_user_proto;
 322 }
 323
 324 /*
 325  * Only limited trace_printk() conversion specifiers allowed:
 326  * %d %i %u %x %ld %li %lu %lx %lld %lli %llu %llx %p %pks %pus %s
 327  */
 328 BPF_CALL_5(bpf_trace_printk, char *, fmt, u32, fmt_size, u64, arg1,
 329            u64, arg2, u64, arg3)
 330 {
 331         int i, mod[3] = {}, fmt_cnt = 0;
 332         char buf[64], fmt_ptype;
 333         void *unsafe_ptr = NULL;
 334         bool str_seen = false;
 335
 336         /*
 337          * bpf_check()->check_func_arg()->check_stack_boundary()
 338          * guarantees that fmt points to bpf program stack,
 339          * fmt_size bytes of it were initialized and fmt_size > 0
 340          */
 341         if (fmt[--fmt_size] != 0)
 342                 return -EINVAL;
 343
 344         /* check format string for allowed specifiers */
 345         for (i = 0; i < fmt_size; i++) {
 346                 if ((!isprint(fmt[i]) && !isspace(fmt[i])) || !isascii(fmt[i]))
 347                         return -EINVAL;
 348
 349                 if (fmt[i] != '%')
 350                         continue;
 351
 352                 if (fmt_cnt >= 3)
 353                         return -EINVAL;
 354
 355                 /* fmt[i] != 0 && fmt[last] == 0, so we can access fmt[i + 1] */
 356                 i++;
 357                 if (fmt[i] == 'l') {
 358                         mod[fmt_cnt]++;
 359                         i++;
 360                 } else if (fmt[i] == 'p') {
 361                         mod[fmt_cnt]++;
 362                         if ((fmt[i + 1] == 'k' ||
 363                              fmt[i + 1] == 'u') &&
 364                             fmt[i + 2] == 's') {
 365                                 fmt_ptype = fmt[i + 1];
 366                                 i += 2;
 367                                 goto fmt_str;
 368                         }
 369
 370                         /* disallow any further format extensions */
 371                         if (fmt[i + 1] != 0 &&
 372                             !isspace(fmt[i + 1]) &&
 373                             !ispunct(fmt[i + 1]))
 374                                 return -EINVAL;
 375
 376                         goto fmt_next;
 377                 } else if (fmt[i] == 's') {
 378                         mod[fmt_cnt]++;
 379                         fmt_ptype = fmt[i];
 380 fmt_str:
 381                         if (str_seen)
 382                                 /* allow only one '%s' per fmt string */
 383                                 return -EINVAL;
 384                         str_seen = true;
 385
 386                         if (fmt[i + 1] != 0 &&
 387                             !isspace(fmt[i + 1]) &&
 388                             !ispunct(fmt[i + 1]))
 389                                 return -EINVAL;
 390
 391                         switch (fmt_cnt) {
 392                         case 0:
 393                                 unsafe_ptr = (void *)(long)arg1;
 394                                 arg1 = (long)buf;
 395                                 break;
 396                         case 1:
 397                                 unsafe_ptr = (void *)(long)arg2;
 398                                 arg2 = (long)buf;
 399                                 break;
 400                         case 2:
 401                                 unsafe_ptr = (void *)(long)arg3;
 402                                 arg3 = (long)buf;
 403                                 break;
 404                         }
 405
 406                         buf[0] = 0;
 407                         switch (fmt_ptype) {
 408                         case 's':
 409 #ifdef CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE
 410                                 strncpy_from_unsafe(buf, unsafe_ptr,
 411                                                     sizeof(buf));
 412                                 break;
 413 #endif
 414                         case 'k':
 415                                 strncpy_from_unsafe_strict(buf, unsafe_ptr,
 416                                                            sizeof(buf));
 417                                 break;
 418                         case 'u':
 419                                 strncpy_from_unsafe_user(buf,
 420                                         (__force void __user *)unsafe_ptr,
 421                                                          sizeof(buf));
 422                                 break;
 423                         }
 424                         goto fmt_next;
 425                 }
 426
 427                 if (fmt[i] == 'l') {
 428                         mod[fmt_cnt]++;
 429                         i++;
 430                 }
 431
 432                 if (fmt[i] != 'i' && fmt[i] != 'd' &&
 433                     fmt[i] != 'u' && fmt[i] != 'x')
 434                         return -EINVAL;
 435 fmt_next:
 436                 fmt_cnt++;
 437         }
 438
 439 /* Horrid workaround for getting va_list handling working with different
 440  * argument type combinations generically for 32 and 64 bit archs.
 441  */
 442 #define __BPF_TP_EMIT() __BPF_ARG3_TP()
 443 #define __BPF_TP(...)                                                   \
 444         __trace_printk(0 /* Fake ip */,                                 \
 445                        fmt, ##__VA_ARGS__)
 446
 447 #define __BPF_ARG1_TP(...)                                              \
 448         ((mod[0] == 2 || (mod[0] == 1 && __BITS_PER_LONG == 64))        \
 449           ? __BPF_TP(arg1, ##__VA_ARGS__)                               \
 450           : ((mod[0] == 1 || (mod[0] == 0 && __BITS_PER_LONG == 32))    \
 451               ? __BPF_TP((long)arg1, ##__VA_ARGS__)                     \
 452               : __BPF_TP((u32)arg1, ##__VA_ARGS__)))
 453
 454 #define __BPF_ARG2_TP(...)                                              \
 455         ((mod[1] == 2 || (mod[1] == 1 && __BITS_PER_LONG == 64))        \
 456           ? __BPF_ARG1_TP(arg2, ##__VA_ARGS__)                          \
 457           : ((mod[1] == 1 || (mod[1] == 0 && __BITS_PER_LONG == 32))    \
 458               ? __BPF_ARG1_TP((long)arg2, ##__VA_ARGS__)                \
 459               : __BPF_ARG1_TP((u32)arg2, ##__VA_ARGS__)))
 460
 461 #define __BPF_ARG3_TP(...)                                              \
 462         ((mod[2] == 2 || (mod[2] == 1 && __BITS_PER_LONG == 64))        \
 463           ? __BPF_ARG2_TP(arg3, ##__VA_ARGS__)                          \
 464           : ((mod[2] == 1 || (mod[2] == 0 && __BITS_PER_LONG == 32))    \
 465               ? __BPF_ARG2_TP((long)arg3, ##__VA_ARGS__)                \
 466               : __BPF_ARG2_TP((u32)arg3, ##__VA_ARGS__)))
 467
 468         return __BPF_TP_EMIT();
 469 }
 470
 471 static const struct bpf_func_proto bpf_trace_printk_proto = {
 472         .func           = bpf_trace_printk,
 473         .gpl_only       = true,
 474         .ret_type       = RET_INTEGER,
 475         .arg1_type      = ARG_PTR_TO_MEM,
 476         .arg2_type      = ARG_CONST_SIZE,
 477 };
 478
 479 const struct bpf_func_proto *bpf_get_trace_printk_proto(void)
 480 {
 481         /*
 482          * this program might be calling bpf_trace_printk,
 483          * so allocate per-cpu printk buffers
 484          */
 485         trace_printk_init_buffers();
 486
 487         return &bpf_trace_printk_proto;
 488 }
 489
 490 static __always_inline int
 491 get_map_perf_counter(struct bpf_map *map, u64 flags,
 492                      u64 *value, u64 *enabled, u64 *running)
 493 {
 494         struct bpf_array *array = container_of(map, struct bpf_array, map);
 495         unsigned int cpu = smp_processor_id();
 496         u64 index = flags & BPF_F_INDEX_MASK;
 497         struct bpf_event_entry *ee;
 498
 499         if (unlikely(flags & ~(BPF_F_INDEX_MASK)))
 500                 return -EINVAL;
 501         if (index == BPF_F_CURRENT_CPU)
 502                 index = cpu;
 503         if (unlikely(index >= array->map.max_entries))
 504                 return -E2BIG;
 505
 506         ee = READ_ONCE(array->ptrs[index]);
 507         if (!ee)
 508                 return -ENOENT;
 509
 510         return perf_event_read_local(ee->event, value, enabled, running);
 511 }
 512
 513 BPF_CALL_2(bpf_perf_event_read, struct bpf_map *, map, u64, flags)
 514 {
 515         u64 value = 0;
 516         int err;
 517
 518         err = get_map_perf_counter(map, flags, &value, NULL, NULL);
 519         /*
 520          * this api is ugly since we miss [-22..-2] range of valid
 521          * counter values, but that's uapi
 522          */
 523         if (err)
 524                 return err;
 525         return value;
 526 }
 527
 528 static const struct bpf_func_proto bpf_perf_event_read_proto = {
 529         .func           = bpf_perf_event_read,
 530         .gpl_only       = true,
 531         .ret_type       = RET_INTEGER,
 532         .arg1_type      = ARG_CONST_MAP_PTR,
 533         .arg2_type      = ARG_ANYTHING,
 534 };
 535
 536 BPF_CALL_4(bpf_perf_event_read_value, struct bpf_map *, map, u64, flags,
 537            struct bpf_perf_event_value *, buf, u32, size)
 538 {
 539         int err = -EINVAL;
 540
 541         if (unlikely(size != sizeof(struct bpf_perf_event_value)))
 542                 goto clear;
 543         err = get_map_perf_counter(map, flags, &buf->counter, &buf->enabled,
 544                                    &buf->running);
 545         if (unlikely(err))
 546                 goto clear;
 547         return 0;
 548 clear:
 549         memset(buf, 0, size);
 550         return err;
 551 }
 552
 553 static const struct bpf_func_proto bpf_perf_event_read_value_proto = {
 554         .func           = bpf_perf_event_read_value,
 555         .gpl_only       = true,
 556         .ret_type       = RET_INTEGER,
 557         .arg1_type      = ARG_CONST_MAP_PTR,
 558         .arg2_type      = ARG_ANYTHING,
 559         .arg3_type      = ARG_PTR_TO_UNINIT_MEM,
 560         .arg4_type      = ARG_CONST_SIZE,
 561 };
 562
 563 static __always_inline u64
 564 __bpf_perf_event_output(struct pt_regs *regs, struct bpf_map *map,
 565                         u64 flags, struct perf_sample_data *sd)
 566 {
 567         struct bpf_array *array = container_of(map, struct bpf_array, map);
 568         unsigned int cpu = smp_processor_id();
 569         u64 index = flags & BPF_F_INDEX_MASK;
 570         struct bpf_event_entry *ee;
 571         struct perf_event *event;
 572
 573         if (index == BPF_F_CURRENT_CPU)
 574                 index = cpu;
 575         if (unlikely(index >= array->map.max_entries))
 576                 return -E2BIG;
 577
 578         ee = READ_ONCE(array->ptrs[index]);
 579         if (!ee)
 580                 return -ENOENT;
 581
 582         event = ee->event;
 583         if (unlikely(event->attr.type != PERF_TYPE_SOFTWARE ||
 584                      event->attr.config != PERF_COUNT_SW_BPF_OUTPUT))
 585                 return -EINVAL;
 586
 587         if (unlikely(event->oncpu != cpu))
 588                 return -EOPNOTSUPP;
 589
 590         return perf_event_output(event, sd, regs);
 591 }
 592
 593 /*
 594  * Support executing tracepoints in normal, irq, and nmi context that each call
 595  * bpf_perf_event_output
 596  */
 597 struct bpf_trace_sample_data {
 598         struct perf_sample_data sds[3];
 599 };
 600
 601 static DEFINE_PER_CPU(struct bpf_trace_sample_data, bpf_trace_sds);
 602 static DEFINE_PER_CPU(int, bpf_trace_nest_level);
 603 BPF_CALL_5(bpf_perf_event_output, struct pt_regs *, regs, struct bpf_map *, map,
 604            u64, flags, void *, data, u64, size)
 605 {
 606         struct bpf_trace_sample_data *sds = this_cpu_ptr(&bpf_trace_sds);
 607         int nest_level = this_cpu_inc_return(bpf_trace_nest_level);
 608         struct perf_raw_record raw = {
 609                 .frag = {
 610                         .size = size,
 611                         .data = data,
 612                 },
 613         };
 614         struct perf_sample_data *sd;
 615         int err;
 616
 617         if (WARN_ON_ONCE(nest_level > ARRAY_SIZE(sds->sds))) {
 618                 err = -EBUSY;
 619                 goto out;
 620         }
 621
 622         sd = &sds->sds[nest_level - 1];
 623
 624         if (unlikely(flags & ~(BPF_F_INDEX_MASK))) {
 625                 err = -EINVAL;
 626                 goto out;
 627         }
 628
 629         perf_sample_data_init(sd, 0, 0);
 630         sd->raw = &raw;
 631
 632         err = __bpf_perf_event_output(regs, map, flags, sd);
 633
 634 out:
 635         this_cpu_dec(bpf_trace_nest_level);
 636         return err;
 637 }
 638
 639 static const struct bpf_func_proto bpf_perf_event_output_proto = {
 640         .func           = bpf_perf_event_output,
 641         .gpl_only       = true,
 642         .ret_type       = RET_INTEGER,
 643         .arg1_type      = ARG_PTR_TO_CTX,
 644         .arg2_type      = ARG_CONST_MAP_PTR,
 645         .arg3_type      = ARG_ANYTHING,
 646         .arg4_type      = ARG_PTR_TO_MEM,
 647         .arg5_type      = ARG_CONST_SIZE_OR_ZERO,
 648 };
 649
 650 static DEFINE_PER_CPU(int, bpf_event_output_nest_level);
 651 struct bpf_nested_pt_regs {
 652         struct pt_regs regs[3];
 653 };
 654 static DEFINE_PER_CPU(struct bpf_nested_pt_regs, bpf_pt_regs);
 655 static DEFINE_PER_CPU(struct bpf_trace_sample_data, bpf_misc_sds);
 656
 657 u64 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
 658                      void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
 659 {
 660         int nest_level = this_cpu_inc_return(bpf_event_output_nest_level);
 661         struct perf_raw_frag frag = {
 662                 .copy           = ctx_copy,
 663                 .size           = ctx_size,
 664                 .data           = ctx,
 665         };
 666         struct perf_raw_record raw = {
 667                 .frag = {
 668                         {
 669                                 .next   = ctx_size ? &frag : NULL,
 670                         },
 671                         .size   = meta_size,
 672                         .data   = meta,
 673                 },
 674         };
 675         struct perf_sample_data *sd;
 676         struct pt_regs *regs;
 677         u64 ret;
 678
 679         if (WARN_ON_ONCE(nest_level > ARRAY_SIZE(bpf_misc_sds.sds))) {
 680                 ret = -EBUSY;
 681                 goto out;
 682         }
 683         sd = this_cpu_ptr(&bpf_misc_sds.sds[nest_level - 1]);
 684         regs = this_cpu_ptr(&bpf_pt_regs.regs[nest_level - 1]);
 685
 686         perf_fetch_caller_regs(regs);
 687         perf_sample_data_init(sd, 0, 0);
 688         sd->raw = &raw;
 689
 690         ret = __bpf_perf_event_output(regs, map, flags, sd);
 691 out:
 692         this_cpu_dec(bpf_event_output_nest_level);
 693         return ret;
 694 }
 695
 696 BPF_CALL_0(bpf_get_current_task)
 697 {
 698         return (long) current;
 699 }
 700
 701 static const struct bpf_func_proto bpf_get_current_task_proto = {
 702         .func           = bpf_get_current_task,
 703         .gpl_only       = true,
 704         .ret_type       = RET_INTEGER,
 705 };
 706
 707 BPF_CALL_2(bpf_current_task_under_cgroup, struct bpf_map *, map, u32, idx)
 708 {
 709         struct bpf_array *array = container_of(map, struct bpf_array, map);
 710         struct cgroup *cgrp;
 711
 712         if (unlikely(idx >= array->map.max_entries))
 713                 return -E2BIG;
 714
 715         cgrp = READ_ONCE(array->ptrs[idx]);
 716         if (unlikely(!cgrp))
 717                 return -EAGAIN;
 718
 719         return task_under_cgroup_hierarchy(current, cgrp);
 720 }
 721
 722 static const struct bpf_func_proto bpf_current_task_under_cgroup_proto = {
 723         .func           = bpf_current_task_under_cgroup,
 724         .gpl_only       = false,
 725         .ret_type       = RET_INTEGER,
 726         .arg1_type      = ARG_CONST_MAP_PTR,
 727         .arg2_type      = ARG_ANYTHING,
 728 };
 729
 730 struct send_signal_irq_work {
 731         struct irq_work irq_work;
 732         struct task_struct *task;
 733         u32 sig;
 734         enum pid_type type;
 735 };
 736
 737 static DEFINE_PER_CPU(struct send_signal_irq_work, send_signal_work);
 738
 739 static void do_bpf_send_signal(struct irq_work *entry)
 740 {
 741         struct send_signal_irq_work *work;
 742
 743         work = container_of(entry, struct send_signal_irq_work, irq_work);
 744         group_send_sig_info(work->sig, SEND_SIG_PRIV, work->task, work->type);
 745 }
 746
 747 static int bpf_send_signal_common(u32 sig, enum pid_type type)
 748 {
 749         struct send_signal_irq_work *work = NULL;
 750
 751         /* Similar to bpf_probe_write_user, task needs to be
 752          * in a sound condition and kernel memory access be
 753          * permitted in order to send signal to the current
 754          * task.
 755          */
 756         if (unlikely(current->flags & (PF_KTHREAD | PF_EXITING)))
 757                 return -EPERM;
 758         if (unlikely(uaccess_kernel()))
 759                 return -EPERM;
 760         if (unlikely(!nmi_uaccess_okay()))
 761                 return -EPERM;
 762
 763         if (irqs_disabled()) {
 764                 /* Do an early check on signal validity. Otherwise,
 765                  * the error is lost in deferred irq_work.
 766                  */
 767                 if (unlikely(!valid_signal(sig)))
 768                         return -EINVAL;
 769
 770                 work = this_cpu_ptr(&send_signal_work);
 771                 if (atomic_read(&work->irq_work.flags) & IRQ_WORK_BUSY)
 772                         return -EBUSY;
 773
 774                 /* Add the current task, which is the target of sending signal,
 775                  * to the irq_work. The current task may change when queued
 776                  * irq works get executed.
 777                  */
 778                 work->task = current;
 779                 work->sig = sig;
 780                 work->type = type;
 781                 irq_work_queue(&work->irq_work);
 782                 return 0;
 783         }
 784
 785         return group_send_sig_info(sig, SEND_SIG_PRIV, current, type);
 786 }
 787
 788 BPF_CALL_1(bpf_send_signal, u32, sig)
 789 {
 790         return bpf_send_signal_common(sig, PIDTYPE_TGID);
 791 }
 792
 793 static const struct bpf_func_proto bpf_send_signal_proto = {
 794         .func           = bpf_send_signal,
 795         .gpl_only       = false,
 796         .ret_type       = RET_INTEGER,
 797         .arg1_type      = ARG_ANYTHING,
 798 };
 799
 800 BPF_CALL_1(bpf_send_signal_thread, u32, sig)
 801 {
 802         return bpf_send_signal_common(sig, PIDTYPE_PID);
 803 }
 804
 805 static const struct bpf_func_proto bpf_send_signal_thread_proto = {
 806         .func           = bpf_send_signal_thread,
 807         .gpl_only       = false,
 808         .ret_type       = RET_INTEGER,
 809         .arg1_type      = ARG_ANYTHING,
 810 };
 811
 812 const struct bpf_func_proto *
 813 bpf_tracing_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
 814 {
 815         switch (func_id) {
 816         case BPF_FUNC_map_lookup_elem:
 817                 return &bpf_map_lookup_elem_proto;
 818         case BPF_FUNC_map_update_elem:
 819                 return &bpf_map_update_elem_proto;
 820         case BPF_FUNC_map_delete_elem:
 821                 return &bpf_map_delete_elem_proto;
 822         case BPF_FUNC_map_push_elem:
 823                 return &bpf_map_push_elem_proto;
 824         case BPF_FUNC_map_pop_elem:
 825                 return &bpf_map_pop_elem_proto;
 826         case BPF_FUNC_map_peek_elem:
 827                 return &bpf_map_peek_elem_proto;
 828         case BPF_FUNC_ktime_get_ns:
 829                 return &bpf_ktime_get_ns_proto;
 830         case BPF_FUNC_tail_call:
 831                 return &bpf_tail_call_proto;
 832         case BPF_FUNC_get_current_pid_tgid:
 833                 return &bpf_get_current_pid_tgid_proto;
 834         case BPF_FUNC_get_current_task:
 835                 return &bpf_get_current_task_proto;
 836         case BPF_FUNC_get_current_uid_gid:
 837                 return &bpf_get_current_uid_gid_proto;
 838         case BPF_FUNC_get_current_comm:
 839                 return &bpf_get_current_comm_proto;
 840         case BPF_FUNC_trace_printk:
 841                 return bpf_get_trace_printk_proto();
 842         case BPF_FUNC_get_smp_processor_id:
 843                 return &bpf_get_smp_processor_id_proto;
 844         case BPF_FUNC_get_numa_node_id:
 845                 return &bpf_get_numa_node_id_proto;
 846         case BPF_FUNC_perf_event_read:
 847                 return &bpf_perf_event_read_proto;
 848         case BPF_FUNC_probe_write_user:
 849                 return bpf_get_probe_write_proto();
 850         case BPF_FUNC_current_task_under_cgroup:
 851                 return &bpf_current_task_under_cgroup_proto;
 852         case BPF_FUNC_get_prandom_u32:
 853                 return &bpf_get_prandom_u32_proto;
 854         case BPF_FUNC_probe_read_user:
 855                 return &bpf_probe_read_user_proto;
 856         case BPF_FUNC_probe_read_kernel:
 857                 return &bpf_probe_read_kernel_proto;
 858         case BPF_FUNC_probe_read_user_str:
 859                 return &bpf_probe_read_user_str_proto;
 860         case BPF_FUNC_probe_read_kernel_str:
 861                 return &bpf_probe_read_kernel_str_proto;
 862 #ifdef CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE
 863         case BPF_FUNC_probe_read:
 864                 return &bpf_probe_read_compat_proto;
 865         case BPF_FUNC_probe_read_str:
 866                 return &bpf_probe_read_compat_str_proto;
 867 #endif
 868 #ifdef CONFIG_CGROUPS
 869         case BPF_FUNC_get_current_cgroup_id:
 870                 return &bpf_get_current_cgroup_id_proto;
 871 #endif
 872         case BPF_FUNC_send_signal:
 873                 return &bpf_send_signal_proto;
 874         case BPF_FUNC_send_signal_thread:
 875                 return &bpf_send_signal_thread_proto;
 876         case BPF_FUNC_perf_event_read_value:
 877                 return &bpf_perf_event_read_value_proto;
 878         case BPF_FUNC_get_ns_current_pid_tgid:
 879                 return &bpf_get_ns_current_pid_tgid_proto;
 880         default:
 881                 return NULL;
 882         }
 883 }
 884
 885 static const struct bpf_func_proto *
 886 kprobe_prog_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
 887 {
 888         switch (func_id) {
 889         case BPF_FUNC_perf_event_output:
 890                 return &bpf_perf_event_output_proto;
 891         case BPF_FUNC_get_stackid:
 892                 return &bpf_get_stackid_proto;
 893         case BPF_FUNC_get_stack:
 894                 return &bpf_get_stack_proto;
 895 #ifdef CONFIG_BPF_KPROBE_OVERRIDE
 896         case BPF_FUNC_override_return:
 897                 return &bpf_override_return_proto;
 898 #endif
 899         default:
 900                 return bpf_tracing_func_proto(func_id, prog);
 901         }
 902 }
 903
 904 /* bpf+kprobe programs can access fields of 'struct pt_regs' */
 905 static bool kprobe_prog_is_valid_access(int off, int size, enum bpf_access_type type,
 906                                         const struct bpf_prog *prog,
 907                                         struct bpf_insn_access_aux *info)
 908 {
 909         if (off < 0 || off >= sizeof(struct pt_regs))
 910                 return false;
 911         if (type != BPF_READ)
 912                 return false;
 913         if (off % size != 0)
 914                 return false;
 915         /*
 916          * Assertion for 32 bit to make sure last 8 byte access
 917          * (BPF_DW) to the last 4 byte member is disallowed.
 918          */
 919         if (off + size > sizeof(struct pt_regs))
 920                 return false;
 921
 922         return true;
 923 }
 924
 925 const struct bpf_verifier_ops kprobe_verifier_ops = {
 926         .get_func_proto  = kprobe_prog_func_proto,
 927         .is_valid_access = kprobe_prog_is_valid_access,
 928 };
 929
 930 const struct bpf_prog_ops kprobe_prog_ops = {
 931 };
 932
 933 BPF_CALL_5(bpf_perf_event_output_tp, void *, tp_buff, struct bpf_map *, map,
 934            u64, flags, void *, data, u64, size)
 935 {
 936         struct pt_regs *regs = *(struct pt_regs **)tp_buff;
 937
 938         /*
 939          * r1 points to perf tracepoint buffer where first 8 bytes are hidden
 940          * from bpf program and contain a pointer to 'struct pt_regs'. Fetch it
 941          * from there and call the same bpf_perf_event_output() helper inline.
 942          */
 943         return ____bpf_perf_event_output(regs, map, flags, data, size);
 944 }
 945
 946 static const struct bpf_func_proto bpf_perf_event_output_proto_tp = {
 947         .func           = bpf_perf_event_output_tp,
 948         .gpl_only       = true,
 949         .ret_type       = RET_INTEGER,
 950         .arg1_type      = ARG_PTR_TO_CTX,
 951         .arg2_type      = ARG_CONST_MAP_PTR,
 952         .arg3_type      = ARG_ANYTHING,
 953         .arg4_type      = ARG_PTR_TO_MEM,
 954         .arg5_type      = ARG_CONST_SIZE_OR_ZERO,
 955 };
 956
 957 BPF_CALL_3(bpf_get_stackid_tp, void *, tp_buff, struct bpf_map *, map,
 958            u64, flags)
 959 {
 960         struct pt_regs *regs = *(struct pt_regs **)tp_buff;
 961
 962         /*
 963          * Same comment as in bpf_perf_event_output_tp(), only that this time
 964          * the other helper's function body cannot be inlined due to being
 965          * external, thus we need to call raw helper function.
 966          */
 967         return bpf_get_stackid((unsigned long) regs, (unsigned long) map,
 968                                flags, 0, 0);
 969 }
 970
 971 static const struct bpf_func_proto bpf_get_stackid_proto_tp = {
 972         .func           = bpf_get_stackid_tp,
 973         .gpl_only       = true,
 974         .ret_type       = RET_INTEGER,
 975         .arg1_type      = ARG_PTR_TO_CTX,
 976         .arg2_type      = ARG_CONST_MAP_PTR,
 977         .arg3_type      = ARG_ANYTHING,
 978 };
 979
 980 BPF_CALL_4(bpf_get_stack_tp, void *, tp_buff, void *, buf, u32, size,
 981            u64, flags)
 982 {
 983         struct pt_regs *regs = *(struct pt_regs **)tp_buff;
 984
 985         return bpf_get_stack((unsigned long) regs, (unsigned long) buf,
 986                              (unsigned long) size, flags, 0);
 987 }
 988
 989 static const struct bpf_func_proto bpf_get_stack_proto_tp = {
 990         .func           = bpf_get_stack_tp,
 991         .gpl_only       = true,
 992         .ret_type       = RET_INTEGER,
 993         .arg1_type      = ARG_PTR_TO_CTX,
 994         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
 995         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 996         .arg4_type      = ARG_ANYTHING,
 997 };
 998
 999 static const struct bpf_func_proto *
1000 tp_prog_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
1001 {
1002         switch (func_id) {
1003         case BPF_FUNC_perf_event_output:
1004                 return &bpf_perf_event_output_proto_tp;
1005         case BPF_FUNC_get_stackid:
1006                 return &bpf_get_stackid_proto_tp;
1007         case BPF_FUNC_get_stack:
1008                 return &bpf_get_stack_proto_tp;
1009         default:
1010                 return bpf_tracing_func_proto(func_id, prog);
1011         }
1012 }
1013
1014 static bool tp_prog_is_valid_access(int off, int size, enum bpf_access_type type,
1015                                     const struct bpf_prog *prog,
1016                                     struct bpf_insn_access_aux *info)
1017 {
1018         if (off < sizeof(void *) || off >= PERF_MAX_TRACE_SIZE)
1019                 return false;
1020         if (type != BPF_READ)
1021                 return false;
1022         if (off % size != 0)
1023                 return false;
1024
1025         BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(__u64));
1026         return true;
1027 }
1028
1029 const struct bpf_verifier_ops tracepoint_verifier_ops = {
1030         .get_func_proto  = tp_prog_func_proto,
1031         .is_valid_access = tp_prog_is_valid_access,
1032 };
1033
1034 const struct bpf_prog_ops tracepoint_prog_ops = {
1035 };
1036
1037 BPF_CALL_3(bpf_perf_prog_read_value, struct bpf_perf_event_data_kern *, ctx,
1038            struct bpf_perf_event_value *, buf, u32, size)
1039 {
1040         int err = -EINVAL;
1041
1042         if (unlikely(size != sizeof(struct bpf_perf_event_value)))
1043                 goto clear;
1044         err = perf_event_read_local(ctx->event, &buf->counter, &buf->enabled,
1045                                     &buf->running);
1046         if (unlikely(err))
1047                 goto clear;
1048         return 0;
1049 clear:
1050         memset(buf, 0, size);
1051         return err;
1052 }
1053
1054 static const struct bpf_func_proto bpf_perf_prog_read_value_proto = {
1055          .func           = bpf_perf_prog_read_value,
1056          .gpl_only       = true,
1057          .ret_type       = RET_INTEGER,
1058          .arg1_type      = ARG_PTR_TO_CTX,
1059          .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
1060          .arg3_type      = ARG_CONST_SIZE,
1061 };
1062
1063 BPF_CALL_4(bpf_read_branch_records, struct bpf_perf_event_data_kern *, ctx,
1064            void *, buf, u32, size, u64, flags)
1065 {
1066 #ifndef CONFIG_X86
1067         return -ENOENT;
1068 #else
1069         static const u32 br_entry_size = sizeof(struct perf_branch_entry);
1070         struct perf_branch_stack *br_stack = ctx->data->br_stack;
1071         u32 to_copy;
1072
1073         if (unlikely(flags & ~BPF_F_GET_BRANCH_RECORDS_SIZE))
1074                 return -EINVAL;
1075
1076         if (unlikely(!br_stack))
1077                 return -EINVAL;
1078
1079         if (flags & BPF_F_GET_BRANCH_RECORDS_SIZE)
1080                 return br_stack->nr * br_entry_size;
1081
1082         if (!buf || (size % br_entry_size != 0))
1083                 return -EINVAL;
1084
1085         to_copy = min_t(u32, br_stack->nr * br_entry_size, size);
1086         memcpy(buf, br_stack->entries, to_copy);
1087
1088         return to_copy;
1089 #endif
1090 }
1091
1092 static const struct bpf_func_proto bpf_read_branch_records_proto = {
1093         .func           = bpf_read_branch_records,
1094         .gpl_only       = true,
1095         .ret_type       = RET_INTEGER,
1096         .arg1_type      = ARG_PTR_TO_CTX,
1097         .arg2_type      = ARG_PTR_TO_MEM_OR_NULL,
1098         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
1099         .arg4_type      = ARG_ANYTHING,
1100 };
1101
1102 static const struct bpf_func_proto *
1103 pe_prog_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
1104 {
1105         switch (func_id) {
1106         case BPF_FUNC_perf_event_output:
1107                 return &bpf_perf_event_output_proto_tp;
1108         case BPF_FUNC_get_stackid:
1109                 return &bpf_get_stackid_proto_tp;
1110         case BPF_FUNC_get_stack:
1111                 return &bpf_get_stack_proto_tp;
1112         case BPF_FUNC_perf_prog_read_value:
1113                 return &bpf_perf_prog_read_value_proto;
1114         case BPF_FUNC_read_branch_records:
1115                 return &bpf_read_branch_records_proto;
1116         default:
1117                 return bpf_tracing_func_proto(func_id, prog);
1118         }
1119 }
1120
1121 /*
1122  * bpf_raw_tp_regs are separate from bpf_pt_regs used from skb/xdp
1123  * to avoid potential recursive reuse issue when/if tracepoints are added
1124  * inside bpf_*_event_output, bpf_get_stackid and/or bpf_get_stack.
1125  *
1126  * Since raw tracepoints run despite bpf_prog_active, support concurrent usage
1127  * in normal, irq, and nmi context.
1128  */
1129 struct bpf_raw_tp_regs {
1130         struct pt_regs regs[3];
1131 };
1132 static DEFINE_PER_CPU(struct bpf_raw_tp_regs, bpf_raw_tp_regs);
1133 static DEFINE_PER_CPU(int, bpf_raw_tp_nest_level);
1134 static struct pt_regs *get_bpf_raw_tp_regs(void)
1135 {
1136         struct bpf_raw_tp_regs *tp_regs = this_cpu_ptr(&bpf_raw_tp_regs);
1137         int nest_level = this_cpu_inc_return(bpf_raw_tp_nest_level);
1138
1139         if (WARN_ON_ONCE(nest_level > ARRAY_SIZE(tp_regs->regs))) {
1140                 this_cpu_dec(bpf_raw_tp_nest_level);
1141                 return ERR_PTR(-EBUSY);
1142         }
1143
1144         return &tp_regs->regs[nest_level - 1];
1145 }
1146
1147 static void put_bpf_raw_tp_regs(void)
1148 {
1149         this_cpu_dec(bpf_raw_tp_nest_level);
1150 }
1151
1152 BPF_CALL_5(bpf_perf_event_output_raw_tp, struct bpf_raw_tracepoint_args *, args,
1153            struct bpf_map *, map, u64, flags, void *, data, u64, size)
1154 {
1155         struct pt_regs *regs = get_bpf_raw_tp_regs();
1156         int ret;
1157
1158         if (IS_ERR(regs))
1159                 return PTR_ERR(regs);
1160
1161         perf_fetch_caller_regs(regs);
1162         ret = ____bpf_perf_event_output(regs, map, flags, data, size);
1163
1164         put_bpf_raw_tp_regs();
1165         return ret;
1166 }
1167
1168 static const struct bpf_func_proto bpf_perf_event_output_proto_raw_tp = {
1169         .func           = bpf_perf_event_output_raw_tp,
1170         .gpl_only       = true,
1171         .ret_type       = RET_INTEGER,
1172         .arg1_type      = ARG_PTR_TO_CTX,
1173         .arg2_type      = ARG_CONST_MAP_PTR,
1174         .arg3_type      = ARG_ANYTHING,
1175         .arg4_type      = ARG_PTR_TO_MEM,
1176         .arg5_type      = ARG_CONST_SIZE_OR_ZERO,
1177 };
1178
1179 extern const struct bpf_func_proto bpf_skb_output_proto;
1180 extern const struct bpf_func_proto bpf_xdp_output_proto;
1181
1182 BPF_CALL_3(bpf_get_stackid_raw_tp, struct bpf_raw_tracepoint_args *, args,
1183            struct bpf_map *, map, u64, flags)
1184 {
1185         struct pt_regs *regs = get_bpf_raw_tp_regs();
1186         int ret;
1187
1188         if (IS_ERR(regs))
1189                 return PTR_ERR(regs);
1190
1191         perf_fetch_caller_regs(regs);
1192         /* similar to bpf_perf_event_output_tp, but pt_regs fetched differently */
1193         ret = bpf_get_stackid((unsigned long) regs, (unsigned long) map,
1194                               flags, 0, 0);
1195         put_bpf_raw_tp_regs();
1196         return ret;
1197 }
1198
1199 static const struct bpf_func_proto bpf_get_stackid_proto_raw_tp = {
1200         .func           = bpf_get_stackid_raw_tp,
1201         .gpl_only       = true,
1202         .ret_type       = RET_INTEGER,
1203         .arg1_type      = ARG_PTR_TO_CTX,
1204         .arg2_type      = ARG_CONST_MAP_PTR,
1205         .arg3_type      = ARG_ANYTHING,
1206 };
1207
1208 BPF_CALL_4(bpf_get_stack_raw_tp, struct bpf_raw_tracepoint_args *, args,
1209            void *, buf, u32, size, u64, flags)
1210 {
1211         struct pt_regs *regs = get_bpf_raw_tp_regs();
1212         int ret;
1213
1214         if (IS_ERR(regs))
1215                 return PTR_ERR(regs);
1216
1217         perf_fetch_caller_regs(regs);
1218         ret = bpf_get_stack((unsigned long) regs, (unsigned long) buf,
1219                             (unsigned long) size, flags, 0);
1220         put_bpf_raw_tp_regs();
1221         return ret;
1222 }
1223
1224 static const struct bpf_func_proto bpf_get_stack_proto_raw_tp = {
1225         .func           = bpf_get_stack_raw_tp,
1226         .gpl_only       = true,
1227         .ret_type       = RET_INTEGER,
1228         .arg1_type      = ARG_PTR_TO_CTX,
1229         .arg2_type      = ARG_PTR_TO_MEM,
1230         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
1231         .arg4_type      = ARG_ANYTHING,
1232 };
1233
1234 static const struct bpf_func_proto *
1235 raw_tp_prog_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
1236 {
1237         switch (func_id) {
1238         case BPF_FUNC_perf_event_output:
1239                 return &bpf_perf_event_output_proto_raw_tp;
1240         case BPF_FUNC_get_stackid:
1241                 return &bpf_get_stackid_proto_raw_tp;
1242         case BPF_FUNC_get_stack:
1243                 return &bpf_get_stack_proto_raw_tp;
1244         default:
1245                 return bpf_tracing_func_proto(func_id, prog);
1246         }
1247 }
1248
1249 static const struct bpf_func_proto *
1250 tracing_prog_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
1251 {
1252         switch (func_id) {
1253 #ifdef CONFIG_NET
1254         case BPF_FUNC_skb_output:
1255                 return &bpf_skb_output_proto;
1256         case BPF_FUNC_xdp_output:
1257                 return &bpf_xdp_output_proto;
1258 #endif
1259         default:
1260                 return raw_tp_prog_func_proto(func_id, prog);
1261         }
1262 }
1263
1264 static bool raw_tp_prog_is_valid_access(int off, int size,
1265                                         enum bpf_access_type type,
1266                                         const struct bpf_prog *prog,
1267                                         struct bpf_insn_access_aux *info)
1268 {
1269         if (off < 0 || off >= sizeof(__u64) * MAX_BPF_FUNC_ARGS)
1270                 return false;
1271         if (type != BPF_READ)
1272                 return false;
1273         if (off % size != 0)
1274                 return false;
1275         return true;
1276 }
1277
1278 static bool tracing_prog_is_valid_access(int off, int size,
1279                                          enum bpf_access_type type,
1280                                          const struct bpf_prog *prog,
1281                                          struct bpf_insn_access_aux *info)
1282 {
1283         if (off < 0 || off >= sizeof(__u64) * MAX_BPF_FUNC_ARGS)
1284                 return false;
1285         if (type != BPF_READ)
1286                 return false;
1287         if (off % size != 0)
1288                 return false;
1289         return btf_ctx_access(off, size, type, prog, info);
1290 }
1291
1292 int __weak bpf_prog_test_run_tracing(struct bpf_prog *prog,
1293                                      const union bpf_attr *kattr,
1294                                      union bpf_attr __user *uattr)
1295 {
1296         return -ENOTSUPP;
1297 }
1298
1299 const struct bpf_verifier_ops raw_tracepoint_verifier_ops = {
1300         .get_func_proto  = raw_tp_prog_func_proto,
1301         .is_valid_access = raw_tp_prog_is_valid_access,
1302 };
1303
1304 const struct bpf_prog_ops raw_tracepoint_prog_ops = {
1305 };
1306
1307 const struct bpf_verifier_ops tracing_verifier_ops = {
1308         .get_func_proto  = tracing_prog_func_proto,
1309         .is_valid_access = tracing_prog_is_valid_access,
1310 };
1311
1312 const struct bpf_prog_ops tracing_prog_ops = {
1313         .test_run = bpf_prog_test_run_tracing,
1314 };
1315
1316 static bool raw_tp_writable_prog_is_valid_access(int off, int size,
1317                                                  enum bpf_access_type type,
1318                                                  const struct bpf_prog *prog,
1319                                                  struct bpf_insn_access_aux *info)
1320 {
1321         if (off == 0) {
1322                 if (size != sizeof(u64) || type != BPF_READ)
1323                         return false;
1324                 info->reg_type = PTR_TO_TP_BUFFER;
1325         }
1326         return raw_tp_prog_is_valid_access(off, size, type, prog, info);
1327 }
1328
1329 const struct bpf_verifier_ops raw_tracepoint_writable_verifier_ops = {
1330         .get_func_proto  = raw_tp_prog_func_proto,
1331         .is_valid_access = raw_tp_writable_prog_is_valid_access,
1332 };
1333
1334 const struct bpf_prog_ops raw_tracepoint_writable_prog_ops = {
1335 };
1336
1337 static bool pe_prog_is_valid_access(int off, int size, enum bpf_access_type type,
1338                                     const struct bpf_prog *prog,
1339                                     struct bpf_insn_access_aux *info)
1340 {
1341         const int size_u64 = sizeof(u64);
1342
1343         if (off < 0 || off >= sizeof(struct bpf_perf_event_data))
1344                 return false;
1345         if (type != BPF_READ)
1346                 return false;
1347         if (off % size != 0) {
1348                 if (sizeof(unsigned long) != 4)
1349                         return false;
1350                 if (size != 8)
1351                         return false;
1352                 if (off % size != 4)
1353                         return false;
1354         }
1355
1356         switch (off) {
1357         case bpf_ctx_range(struct bpf_perf_event_data, sample_period):
1358                 bpf_ctx_record_field_size(info, size_u64);
1359                 if (!bpf_ctx_narrow_access_ok(off, size, size_u64))
1360                         return false;
1361                 break;
1362         case bpf_ctx_range(struct bpf_perf_event_data, addr):
1363                 bpf_ctx_record_field_size(info, size_u64);
1364                 if (!bpf_ctx_narrow_access_ok(off, size, size_u64))
1365                         return false;
1366                 break;
1367         default:
1368                 if (size != sizeof(long))
1369                         return false;
1370         }
1371
1372         return true;
1373 }
1374
1375 static u32 pe_prog_convert_ctx_access(enum bpf_access_type type,
1376                                       const struct bpf_insn *si,
1377                                       struct bpf_insn *insn_buf,
1378                                       struct bpf_prog *prog, u32 *target_size)
1379 {
1380         struct bpf_insn *insn = insn_buf;
1381
1382         switch (si->off) {
1383         case offsetof(struct bpf_perf_event_data, sample_period):
1384                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct bpf_perf_event_data_kern,
1385                                                        data), si->dst_reg, si->src_reg,
1386                                       offsetof(struct bpf_perf_event_data_kern, data));
1387                 *insn++ = BPF_LDX_MEM(BPF_DW, si->dst_reg, si->dst_reg,
1388                                       bpf_target_off(struct perf_sample_data, period, 8,
1389                                                      target_size));
1390                 break;
1391         case offsetof(struct bpf_perf_event_data, addr):
1392                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct bpf_perf_event_data_kern,
1393                                                        data), si->dst_reg, si->src_reg,
1394                                       offsetof(struct bpf_perf_event_data_kern, data));
1395                 *insn++ = BPF_LDX_MEM(BPF_DW, si->dst_reg, si->dst_reg,
1396                                       bpf_target_off(struct perf_sample_data, addr, 8,
1397                                                      target_size));
1398                 break;
1399         default:
1400                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct bpf_perf_event_data_kern,
1401                                                        regs), si->dst_reg, si->src_reg,
1402                                       offsetof(struct bpf_perf_event_data_kern, regs));
1403                 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(long), si->dst_reg, si->dst_reg,
1404                                       si->off);
1405                 break;
1406         }
1407
1408         return insn - insn_buf;
1409 }
1410
1411 const struct bpf_verifier_ops perf_event_verifier_ops = {
1412         .get_func_proto         = pe_prog_func_proto,
1413         .is_valid_access        = pe_prog_is_valid_access,
1414         .convert_ctx_access     = pe_prog_convert_ctx_access,
1415 };
1416
1417 const struct bpf_prog_ops perf_event_prog_ops = {
1418 };
1419
1420 static DEFINE_MUTEX(bpf_event_mutex);
1421
1422 #define BPF_TRACE_MAX_PROGS 64
1423
1424 int perf_event_attach_bpf_prog(struct perf_event *event,
1425                                struct bpf_prog *prog)
1426 {
1427         struct bpf_prog_array *old_array;
1428         struct bpf_prog_array *new_array;
1429         int ret = -EEXIST;
1430
1431         /*
1432          * Kprobe override only works if they are on the function entry,
1433          * and only if they are on the opt-in list.
1434          */
1435         if (prog->kprobe_override &&
1436             (!trace_kprobe_on_func_entry(event->tp_event) ||
1437              !trace_kprobe_error_injectable(event->tp_event)))
1438                 return -EINVAL;
1439
1440         mutex_lock(&bpf_event_mutex);
1441
1442         if (event->prog)
1443                 goto unlock;
1444
1445         old_array = bpf_event_rcu_dereference(event->tp_event->prog_array);
1446         if (old_array &&
1447             bpf_prog_array_length(old_array) >= BPF_TRACE_MAX_PROGS) {
1448                 ret = -E2BIG;
1449                 goto unlock;
1450         }
1451
1452         ret = bpf_prog_array_copy(old_array, NULL, prog, &new_array);
1453         if (ret < 0)
1454                 goto unlock;
1455
1456         /* set the new array to event->tp_event and set event->prog */
1457         event->prog = prog;
1458         rcu_assign_pointer(event->tp_event->prog_array, new_array);
1459         bpf_prog_array_free(old_array);
1460
1461 unlock:
1462         mutex_unlock(&bpf_event_mutex);
1463         return ret;
1464 }
1465
1466 void perf_event_detach_bpf_prog(struct perf_event *event)
1467 {
1468         struct bpf_prog_array *old_array;
1469         struct bpf_prog_array *new_array;
1470         int ret;
1471
1472         mutex_lock(&bpf_event_mutex);
1473
1474         if (!event->prog)
1475                 goto unlock;
1476
1477         old_array = bpf_event_rcu_dereference(event->tp_event->prog_array);
1478         ret = bpf_prog_array_copy(old_array, event->prog, NULL, &new_array);
1479         if (ret == -ENOENT)
1480                 goto unlock;
1481         if (ret < 0) {
1482                 bpf_prog_array_delete_safe(old_array, event->prog);
1483         } else {
1484                 rcu_assign_pointer(event->tp_event->prog_array, new_array);
1485                 bpf_prog_array_free(old_array);
1486         }
1487
1488         bpf_prog_put(event->prog);
1489         event->prog = NULL;
1490
1491 unlock:
1492         mutex_unlock(&bpf_event_mutex);
1493 }
1494
1495 int perf_event_query_prog_array(struct perf_event *event, void __user *info)
1496 {
1497         struct perf_event_query_bpf __user *uquery = info;
1498         struct perf_event_query_bpf query = {};
1499         struct bpf_prog_array *progs;
1500         u32 *ids, prog_cnt, ids_len;
1501         int ret;
1502
1503         if (!capable(CAP_SYS_ADMIN))
1504                 return -EPERM;
1505         if (event->attr.type != PERF_TYPE_TRACEPOINT)
1506                 return -EINVAL;
1507         if (copy_from_user(&query, uquery, sizeof(query)))
1508                 return -EFAULT;
1509
1510         ids_len = query.ids_len;
1511         if (ids_len > BPF_TRACE_MAX_PROGS)
1512                 return -E2BIG;
1513         ids = kcalloc(ids_len, sizeof(u32), GFP_USER | __GFP_NOWARN);
1514         if (!ids)
1515                 return -ENOMEM;
1516         /*
1517          * The above kcalloc returns ZERO_SIZE_PTR when ids_len = 0, which
1518          * is required when user only wants to check for uquery->prog_cnt.
1519          * There is no need to check for it since the case is handled
1520          * gracefully in bpf_prog_array_copy_info.
1521          */
1522
1523         mutex_lock(&bpf_event_mutex);
1524         progs = bpf_event_rcu_dereference(event->tp_event->prog_array);
1525         ret = bpf_prog_array_copy_info(progs, ids, ids_len, &prog_cnt);
1526         mutex_unlock(&bpf_event_mutex);
1527
1528         if (copy_to_user(&uquery->prog_cnt, &prog_cnt, sizeof(prog_cnt)) ||
1529             copy_to_user(uquery->ids, ids, ids_len * sizeof(u32)))
1530                 ret = -EFAULT;
1531
1532         kfree(ids);
1533         return ret;
1534 }
1535
1536 extern struct bpf_raw_event_map __start__bpf_raw_tp[];
1537 extern struct bpf_raw_event_map __stop__bpf_raw_tp[];
1538
1539 struct bpf_raw_event_map *bpf_get_raw_tracepoint(const char *name)
1540 {
1541         struct bpf_raw_event_map *btp = __start__bpf_raw_tp;
1542
1543         for (; btp < __stop__bpf_raw_tp; btp++) {
1544                 if (!strcmp(btp->tp->name, name))
1545                         return btp;
1546         }
1547
1548         return bpf_get_raw_tracepoint_module(name);
1549 }
1550
1551 void bpf_put_raw_tracepoint(struct bpf_raw_event_map *btp)
1552 {
1553         struct module *mod = __module_address((unsigned long)btp);
1554
1555         if (mod)
1556                 module_put(mod);
1557 }
1558
1559 static __always_inline
1560 void __bpf_trace_run(struct bpf_prog *prog, u64 *args)
1561 {
1562         cant_sleep();
1563         rcu_read_lock();
1564         (void) BPF_PROG_RUN(prog, args);
1565         rcu_read_unlock();
1566 }
1567
1568 #define UNPACK(...)                     __VA_ARGS__
1569 #define REPEAT_1(FN, DL, X, ...)        FN(X)
1570 #define REPEAT_2(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_1(FN, DL, __VA_ARGS__)
1571 #define REPEAT_3(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_2(FN, DL, __VA_ARGS__)
1572 #define REPEAT_4(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_3(FN, DL, __VA_ARGS__)
1573 #define REPEAT_5(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_4(FN, DL, __VA_ARGS__)
1574 #define REPEAT_6(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_5(FN, DL, __VA_ARGS__)
1575 #define REPEAT_7(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_6(FN, DL, __VA_ARGS__)
1576 #define REPEAT_8(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_7(FN, DL, __VA_ARGS__)
1577 #define REPEAT_9(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_8(FN, DL, __VA_ARGS__)
1578 #define REPEAT_10(FN, DL, X, ...)       FN(X) UNPACK DL REPEAT_9(FN, DL, __VA_ARGS__)
1579 #define REPEAT_11(FN, DL, X, ...)       FN(X) UNPACK DL REPEAT_10(FN, DL, __VA_ARGS__)
1580 #define REPEAT_12(FN, DL, X, ...)       FN(X) UNPACK DL REPEAT_11(FN, DL, __VA_ARGS__)
1581 #define REPEAT(X, FN, DL, ...)          REPEAT_##X(FN, DL, __VA_ARGS__)
1582
1583 #define SARG(X)         u64 arg##X
1584 #define COPY(X)         args[X] = arg##X
1585
1586 #define __DL_COM        (,)
1587 #define __DL_SEM        (;)
1588
1589 #define __SEQ_0_11      0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11
1590
1591 #define BPF_TRACE_DEFN_x(x)                                             \
1592         void bpf_trace_run##x(struct bpf_prog *prog,                    \
1593                               REPEAT(x, SARG, __DL_COM, __SEQ_0_11))    \
1594         {                                                               \
1595                 u64 args[x];                                            \
1596                 REPEAT(x, COPY, __DL_SEM, __SEQ_0_11);                  \
1597                 __bpf_trace_run(prog, args);                            \
1598         }                                                               \
1599         EXPORT_SYMBOL_GPL(bpf_trace_run##x)
1600 BPF_TRACE_DEFN_x(1);
1601 BPF_TRACE_DEFN_x(2);
1602 BPF_TRACE_DEFN_x(3);
1603 BPF_TRACE_DEFN_x(4);
1604 BPF_TRACE_DEFN_x(5);
1605 BPF_TRACE_DEFN_x(6);
1606 BPF_TRACE_DEFN_x(7);
1607 BPF_TRACE_DEFN_x(8);
1608 BPF_TRACE_DEFN_x(9);
1609 BPF_TRACE_DEFN_x(10);
1610 BPF_TRACE_DEFN_x(11);
1611 BPF_TRACE_DEFN_x(12);
1612
1613 static int __bpf_probe_register(struct bpf_raw_event_map *btp, struct bpf_prog *prog)
1614 {
1615         struct tracepoint *tp = btp->tp;
1616
1617         /*
1618          * check that program doesn't access arguments beyond what's
1619          * available in this tracepoint
1620          */
1621         if (prog->aux->max_ctx_offset > btp->num_args * sizeof(u64))
1622                 return -EINVAL;
1623
1624         if (prog->aux->max_tp_access > btp->writable_size)
1625                 return -EINVAL;
1626
1627         return tracepoint_probe_register(tp, (void *)btp->bpf_func, prog);
1628 }
1629
1630 int bpf_probe_register(struct bpf_raw_event_map *btp, struct bpf_prog *prog)
1631 {
1632         return __bpf_probe_register(btp, prog);
1633 }
1634
1635 int bpf_probe_unregister(struct bpf_raw_event_map *btp, struct bpf_prog *prog)
1636 {
1637         return tracepoint_probe_unregister(btp->tp, (void *)btp->bpf_func, prog);
1638 }
1639
1640 int bpf_get_perf_event_info(const struct perf_event *event, u32 *prog_id,
1641                             u32 *fd_type, const char **buf,
1642                             u64 *probe_offset, u64 *probe_addr)
1643 {
1644         bool is_tracepoint, is_syscall_tp;
1645         struct bpf_prog *prog;
1646         int flags, err = 0;
1647
1648         prog = event->prog;
1649         if (!prog)
1650                 return -ENOENT;
1651
1652         /* not supporting BPF_PROG_TYPE_PERF_EVENT yet */
1653         if (prog->type == BPF_PROG_TYPE_PERF_EVENT)
1654                 return -EOPNOTSUPP;
1655
1656         *prog_id = prog->aux->id;
1657         flags = event->tp_event->flags;
1658         is_tracepoint = flags & TRACE_EVENT_FL_TRACEPOINT;
1659         is_syscall_tp = is_syscall_trace_event(event->tp_event);
1660
1661         if (is_tracepoint || is_syscall_tp) {
1662                 *buf = is_tracepoint ? event->tp_event->tp->name
1663                                      : event->tp_event->name;
1664                 *fd_type = BPF_FD_TYPE_TRACEPOINT;
1665                 *probe_offset = 0x0;
1666                 *probe_addr = 0x0;
1667         } else {
1668                 /* kprobe/uprobe */
1669                 err = -EOPNOTSUPP;
1670 #ifdef CONFIG_KPROBE_EVENTS
1671                 if (flags & TRACE_EVENT_FL_KPROBE)
1672                         err = bpf_get_kprobe_info(event, fd_type, buf,
1673                                                   probe_offset, probe_addr,
1674                                                   event->attr.type == PERF_TYPE_TRACEPOINT);
1675 #endif
1676 #ifdef CONFIG_UPROBE_EVENTS
1677                 if (flags & TRACE_EVENT_FL_UPROBE)
1678                         err = bpf_get_uprobe_info(event, fd_type, buf,
1679                                                   probe_offset,
1680                                                   event->attr.type == PERF_TYPE_TRACEPOINT);
1681 #endif
1682         }
1683
1684         return err;
1685 }
1686
1687 static int __init send_signal_irq_work_init(void)
1688 {
1689         int cpu;
1690         struct send_signal_irq_work *work;
1691
1692         for_each_possible_cpu(cpu) {
1693                 work = per_cpu_ptr(&send_signal_work, cpu);
1694                 init_irq_work(&work->irq_work, do_bpf_send_signal);
1695         }
1696         return 0;
1697 }
1698
1699 subsys_initcall(send_signal_irq_work_init);
1700
1701 #ifdef CONFIG_MODULES
1702 static int bpf_event_notify(struct notifier_block *nb, unsigned long op,
1703                             void *module)
1704 {
1705         struct bpf_trace_module *btm, *tmp;
1706         struct module *mod = module;
1707
1708         if (mod->num_bpf_raw_events == 0 ||
1709             (op != MODULE_STATE_COMING && op != MODULE_STATE_GOING))
1710                 return 0;
1711
1712         mutex_lock(&bpf_module_mutex);
1713
1714         switch (op) {
1715         case MODULE_STATE_COMING:
1716                 btm = kzalloc(sizeof(*btm), GFP_KERNEL);
1717                 if (btm) {
1718                         btm->module = module;
1719                         list_add(&btm->list, &bpf_trace_modules);
1720                 }
1721                 break;
1722         case MODULE_STATE_GOING:
1723                 list_for_each_entry_safe(btm, tmp, &bpf_trace_modules, list) {
1724                         if (btm->module == module) {
1725                                 list_del(&btm->list);
1726                                 kfree(btm);
1727                                 break;
1728                         }
1729                 }
1730                 break;
1731         }
1732
1733         mutex_unlock(&bpf_module_mutex);
1734
1735         return 0;
1736 }
1737
1738 static struct notifier_block bpf_module_nb = {
1739         .notifier_call = bpf_event_notify,
1740 };
1741
1742 static int __init bpf_event_init(void)
1743 {
1744         register_module_notifier(&bpf_module_nb);
1745         return 0;
1746 }
1747
1748 fs_initcall(bpf_event_init);
1749 #endif /* CONFIG_MODULES */