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 %s
 327  */
 328 BPF_CALL_5(bpf_trace_printk, char *, fmt, u32, fmt_size, u64, arg1,
 329            u64, arg2, u64, arg3)
 330 {
 331         bool str_seen = false;
 332         int mod[3] = {};
 333         int fmt_cnt = 0;
 334         u64 unsafe_addr;
 335         char buf[64];
 336         int i;
 337
 338         /*
 339          * bpf_check()->check_func_arg()->check_stack_boundary()
 340          * guarantees that fmt points to bpf program stack,
 341          * fmt_size bytes of it were initialized and fmt_size > 0
 342          */
 343         if (fmt[--fmt_size] != 0)
 344                 return -EINVAL;
 345
 346         /* check format string for allowed specifiers */
 347         for (i = 0; i < fmt_size; i++) {
 348                 if ((!isprint(fmt[i]) && !isspace(fmt[i])) || !isascii(fmt[i]))
 349                         return -EINVAL;
 350
 351                 if (fmt[i] != '%')
 352                         continue;
 353
 354                 if (fmt_cnt >= 3)
 355                         return -EINVAL;
 356
 357                 /* fmt[i] != 0 && fmt[last] == 0, so we can access fmt[i + 1] */
 358                 i++;
 359                 if (fmt[i] == 'l') {
 360                         mod[fmt_cnt]++;
 361                         i++;
 362                 } else if (fmt[i] == 'p' || fmt[i] == 's') {
 363                         mod[fmt_cnt]++;
 364                         /* disallow any further format extensions */
 365                         if (fmt[i + 1] != 0 &&
 366                             !isspace(fmt[i + 1]) &&
 367                             !ispunct(fmt[i + 1]))
 368                                 return -EINVAL;
 369                         fmt_cnt++;
 370                         if (fmt[i] == 's') {
 371                                 if (str_seen)
 372                                         /* allow only one '%s' per fmt string */
 373                                         return -EINVAL;
 374                                 str_seen = true;
 375
 376                                 switch (fmt_cnt) {
 377                                 case 1:
 378                                         unsafe_addr = arg1;
 379                                         arg1 = (long) buf;
 380                                         break;
 381                                 case 2:
 382                                         unsafe_addr = arg2;
 383                                         arg2 = (long) buf;
 384                                         break;
 385                                 case 3:
 386                                         unsafe_addr = arg3;
 387                                         arg3 = (long) buf;
 388                                         break;
 389                                 }
 390                                 buf[0] = 0;
 391                                 strncpy_from_unsafe(buf,
 392                                                     (void *) (long) unsafe_addr,
 393                                                     sizeof(buf));
 394                         }
 395                         continue;
 396                 }
 397
 398                 if (fmt[i] == 'l') {
 399                         mod[fmt_cnt]++;
 400                         i++;
 401                 }
 402
 403                 if (fmt[i] != 'i' && fmt[i] != 'd' &&
 404                     fmt[i] != 'u' && fmt[i] != 'x')
 405                         return -EINVAL;
 406                 fmt_cnt++;
 407         }
 408
 409 /* Horrid workaround for getting va_list handling working with different
 410  * argument type combinations generically for 32 and 64 bit archs.
 411  */
 412 #define __BPF_TP_EMIT() __BPF_ARG3_TP()
 413 #define __BPF_TP(...)                                                   \
 414         __trace_printk(0 /* Fake ip */,                                 \
 415                        fmt, ##__VA_ARGS__)
 416
 417 #define __BPF_ARG1_TP(...)                                              \
 418         ((mod[0] == 2 || (mod[0] == 1 && __BITS_PER_LONG == 64))        \
 419           ? __BPF_TP(arg1, ##__VA_ARGS__)                               \
 420           : ((mod[0] == 1 || (mod[0] == 0 && __BITS_PER_LONG == 32))    \
 421               ? __BPF_TP((long)arg1, ##__VA_ARGS__)                     \
 422               : __BPF_TP((u32)arg1, ##__VA_ARGS__)))
 423
 424 #define __BPF_ARG2_TP(...)                                              \
 425         ((mod[1] == 2 || (mod[1] == 1 && __BITS_PER_LONG == 64))        \
 426           ? __BPF_ARG1_TP(arg2, ##__VA_ARGS__)                          \
 427           : ((mod[1] == 1 || (mod[1] == 0 && __BITS_PER_LONG == 32))    \
 428               ? __BPF_ARG1_TP((long)arg2, ##__VA_ARGS__)                \
 429               : __BPF_ARG1_TP((u32)arg2, ##__VA_ARGS__)))
 430
 431 #define __BPF_ARG3_TP(...)                                              \
 432         ((mod[2] == 2 || (mod[2] == 1 && __BITS_PER_LONG == 64))        \
 433           ? __BPF_ARG2_TP(arg3, ##__VA_ARGS__)                          \
 434           : ((mod[2] == 1 || (mod[2] == 0 && __BITS_PER_LONG == 32))    \
 435               ? __BPF_ARG2_TP((long)arg3, ##__VA_ARGS__)                \
 436               : __BPF_ARG2_TP((u32)arg3, ##__VA_ARGS__)))
 437
 438         return __BPF_TP_EMIT();
 439 }
 440
 441 static const struct bpf_func_proto bpf_trace_printk_proto = {
 442         .func           = bpf_trace_printk,
 443         .gpl_only       = true,
 444         .ret_type       = RET_INTEGER,
 445         .arg1_type      = ARG_PTR_TO_MEM,
 446         .arg2_type      = ARG_CONST_SIZE,
 447 };
 448
 449 const struct bpf_func_proto *bpf_get_trace_printk_proto(void)
 450 {
 451         /*
 452          * this program might be calling bpf_trace_printk,
 453          * so allocate per-cpu printk buffers
 454          */
 455         trace_printk_init_buffers();
 456
 457         return &bpf_trace_printk_proto;
 458 }
 459
 460 static __always_inline int
 461 get_map_perf_counter(struct bpf_map *map, u64 flags,
 462                      u64 *value, u64 *enabled, u64 *running)
 463 {
 464         struct bpf_array *array = container_of(map, struct bpf_array, map);
 465         unsigned int cpu = smp_processor_id();
 466         u64 index = flags & BPF_F_INDEX_MASK;
 467         struct bpf_event_entry *ee;
 468
 469         if (unlikely(flags & ~(BPF_F_INDEX_MASK)))
 470                 return -EINVAL;
 471         if (index == BPF_F_CURRENT_CPU)
 472                 index = cpu;
 473         if (unlikely(index >= array->map.max_entries))
 474                 return -E2BIG;
 475
 476         ee = READ_ONCE(array->ptrs[index]);
 477         if (!ee)
 478                 return -ENOENT;
 479
 480         return perf_event_read_local(ee->event, value, enabled, running);
 481 }
 482
 483 BPF_CALL_2(bpf_perf_event_read, struct bpf_map *, map, u64, flags)
 484 {
 485         u64 value = 0;
 486         int err;
 487
 488         err = get_map_perf_counter(map, flags, &value, NULL, NULL);
 489         /*
 490          * this api is ugly since we miss [-22..-2] range of valid
 491          * counter values, but that's uapi
 492          */
 493         if (err)
 494                 return err;
 495         return value;
 496 }
 497
 498 static const struct bpf_func_proto bpf_perf_event_read_proto = {
 499         .func           = bpf_perf_event_read,
 500         .gpl_only       = true,
 501         .ret_type       = RET_INTEGER,
 502         .arg1_type      = ARG_CONST_MAP_PTR,
 503         .arg2_type      = ARG_ANYTHING,
 504 };
 505
 506 BPF_CALL_4(bpf_perf_event_read_value, struct bpf_map *, map, u64, flags,
 507            struct bpf_perf_event_value *, buf, u32, size)
 508 {
 509         int err = -EINVAL;
 510
 511         if (unlikely(size != sizeof(struct bpf_perf_event_value)))
 512                 goto clear;
 513         err = get_map_perf_counter(map, flags, &buf->counter, &buf->enabled,
 514                                    &buf->running);
 515         if (unlikely(err))
 516                 goto clear;
 517         return 0;
 518 clear:
 519         memset(buf, 0, size);
 520         return err;
 521 }
 522
 523 static const struct bpf_func_proto bpf_perf_event_read_value_proto = {
 524         .func           = bpf_perf_event_read_value,
 525         .gpl_only       = true,
 526         .ret_type       = RET_INTEGER,
 527         .arg1_type      = ARG_CONST_MAP_PTR,
 528         .arg2_type      = ARG_ANYTHING,
 529         .arg3_type      = ARG_PTR_TO_UNINIT_MEM,
 530         .arg4_type      = ARG_CONST_SIZE,
 531 };
 532
 533 static __always_inline u64
 534 __bpf_perf_event_output(struct pt_regs *regs, struct bpf_map *map,
 535                         u64 flags, struct perf_sample_data *sd)
 536 {
 537         struct bpf_array *array = container_of(map, struct bpf_array, map);
 538         unsigned int cpu = smp_processor_id();
 539         u64 index = flags & BPF_F_INDEX_MASK;
 540         struct bpf_event_entry *ee;
 541         struct perf_event *event;
 542
 543         if (index == BPF_F_CURRENT_CPU)
 544                 index = cpu;
 545         if (unlikely(index >= array->map.max_entries))
 546                 return -E2BIG;
 547
 548         ee = READ_ONCE(array->ptrs[index]);
 549         if (!ee)
 550                 return -ENOENT;
 551
 552         event = ee->event;
 553         if (unlikely(event->attr.type != PERF_TYPE_SOFTWARE ||
 554                      event->attr.config != PERF_COUNT_SW_BPF_OUTPUT))
 555                 return -EINVAL;
 556
 557         if (unlikely(event->oncpu != cpu))
 558                 return -EOPNOTSUPP;
 559
 560         return perf_event_output(event, sd, regs);
 561 }
 562
 563 /*
 564  * Support executing tracepoints in normal, irq, and nmi context that each call
 565  * bpf_perf_event_output
 566  */
 567 struct bpf_trace_sample_data {
 568         struct perf_sample_data sds[3];
 569 };
 570
 571 static DEFINE_PER_CPU(struct bpf_trace_sample_data, bpf_trace_sds);
 572 static DEFINE_PER_CPU(int, bpf_trace_nest_level);
 573 BPF_CALL_5(bpf_perf_event_output, struct pt_regs *, regs, struct bpf_map *, map,
 574            u64, flags, void *, data, u64, size)
 575 {
 576         struct bpf_trace_sample_data *sds = this_cpu_ptr(&bpf_trace_sds);
 577         int nest_level = this_cpu_inc_return(bpf_trace_nest_level);
 578         struct perf_raw_record raw = {
 579                 .frag = {
 580                         .size = size,
 581                         .data = data,
 582                 },
 583         };
 584         struct perf_sample_data *sd;
 585         int err;
 586
 587         if (WARN_ON_ONCE(nest_level > ARRAY_SIZE(sds->sds))) {
 588                 err = -EBUSY;
 589                 goto out;
 590         }
 591
 592         sd = &sds->sds[nest_level - 1];
 593
 594         if (unlikely(flags & ~(BPF_F_INDEX_MASK))) {
 595                 err = -EINVAL;
 596                 goto out;
 597         }
 598
 599         perf_sample_data_init(sd, 0, 0);
 600         sd->raw = &raw;
 601
 602         err = __bpf_perf_event_output(regs, map, flags, sd);
 603
 604 out:
 605         this_cpu_dec(bpf_trace_nest_level);
 606         return err;
 607 }
 608
 609 static const struct bpf_func_proto bpf_perf_event_output_proto = {
 610         .func           = bpf_perf_event_output,
 611         .gpl_only       = true,
 612         .ret_type       = RET_INTEGER,
 613         .arg1_type      = ARG_PTR_TO_CTX,
 614         .arg2_type      = ARG_CONST_MAP_PTR,
 615         .arg3_type      = ARG_ANYTHING,
 616         .arg4_type      = ARG_PTR_TO_MEM,
 617         .arg5_type      = ARG_CONST_SIZE_OR_ZERO,
 618 };
 619
 620 static DEFINE_PER_CPU(int, bpf_event_output_nest_level);
 621 struct bpf_nested_pt_regs {
 622         struct pt_regs regs[3];
 623 };
 624 static DEFINE_PER_CPU(struct bpf_nested_pt_regs, bpf_pt_regs);
 625 static DEFINE_PER_CPU(struct bpf_trace_sample_data, bpf_misc_sds);
 626
 627 u64 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
 628                      void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
 629 {
 630         int nest_level = this_cpu_inc_return(bpf_event_output_nest_level);
 631         struct perf_raw_frag frag = {
 632                 .copy           = ctx_copy,
 633                 .size           = ctx_size,
 634                 .data           = ctx,
 635         };
 636         struct perf_raw_record raw = {
 637                 .frag = {
 638                         {
 639                                 .next   = ctx_size ? &frag : NULL,
 640                         },
 641                         .size   = meta_size,
 642                         .data   = meta,
 643                 },
 644         };
 645         struct perf_sample_data *sd;
 646         struct pt_regs *regs;
 647         u64 ret;
 648
 649         if (WARN_ON_ONCE(nest_level > ARRAY_SIZE(bpf_misc_sds.sds))) {
 650                 ret = -EBUSY;
 651                 goto out;
 652         }
 653         sd = this_cpu_ptr(&bpf_misc_sds.sds[nest_level - 1]);
 654         regs = this_cpu_ptr(&bpf_pt_regs.regs[nest_level - 1]);
 655
 656         perf_fetch_caller_regs(regs);
 657         perf_sample_data_init(sd, 0, 0);
 658         sd->raw = &raw;
 659
 660         ret = __bpf_perf_event_output(regs, map, flags, sd);
 661 out:
 662         this_cpu_dec(bpf_event_output_nest_level);
 663         return ret;
 664 }
 665
 666 BPF_CALL_0(bpf_get_current_task)
 667 {
 668         return (long) current;
 669 }
 670
 671 static const struct bpf_func_proto bpf_get_current_task_proto = {
 672         .func           = bpf_get_current_task,
 673         .gpl_only       = true,
 674         .ret_type       = RET_INTEGER,
 675 };
 676
 677 BPF_CALL_2(bpf_current_task_under_cgroup, struct bpf_map *, map, u32, idx)
 678 {
 679         struct bpf_array *array = container_of(map, struct bpf_array, map);
 680         struct cgroup *cgrp;
 681
 682         if (unlikely(idx >= array->map.max_entries))
 683                 return -E2BIG;
 684
 685         cgrp = READ_ONCE(array->ptrs[idx]);
 686         if (unlikely(!cgrp))
 687                 return -EAGAIN;
 688
 689         return task_under_cgroup_hierarchy(current, cgrp);
 690 }
 691
 692 static const struct bpf_func_proto bpf_current_task_under_cgroup_proto = {
 693         .func           = bpf_current_task_under_cgroup,
 694         .gpl_only       = false,
 695         .ret_type       = RET_INTEGER,
 696         .arg1_type      = ARG_CONST_MAP_PTR,
 697         .arg2_type      = ARG_ANYTHING,
 698 };
 699
 700 struct send_signal_irq_work {
 701         struct irq_work irq_work;
 702         struct task_struct *task;
 703         u32 sig;
 704         enum pid_type type;
 705 };
 706
 707 static DEFINE_PER_CPU(struct send_signal_irq_work, send_signal_work);
 708
 709 static void do_bpf_send_signal(struct irq_work *entry)
 710 {
 711         struct send_signal_irq_work *work;
 712
 713         work = container_of(entry, struct send_signal_irq_work, irq_work);
 714         group_send_sig_info(work->sig, SEND_SIG_PRIV, work->task, work->type);
 715 }
 716
 717 static int bpf_send_signal_common(u32 sig, enum pid_type type)
 718 {
 719         struct send_signal_irq_work *work = NULL;
 720
 721         /* Similar to bpf_probe_write_user, task needs to be
 722          * in a sound condition and kernel memory access be
 723          * permitted in order to send signal to the current
 724          * task.
 725          */
 726         if (unlikely(current->flags & (PF_KTHREAD | PF_EXITING)))
 727                 return -EPERM;
 728         if (unlikely(uaccess_kernel()))
 729                 return -EPERM;
 730         if (unlikely(!nmi_uaccess_okay()))
 731                 return -EPERM;
 732
 733         if (irqs_disabled()) {
 734                 /* Do an early check on signal validity. Otherwise,
 735                  * the error is lost in deferred irq_work.
 736                  */
 737                 if (unlikely(!valid_signal(sig)))
 738                         return -EINVAL;
 739
 740                 work = this_cpu_ptr(&send_signal_work);
 741                 if (atomic_read(&work->irq_work.flags) & IRQ_WORK_BUSY)
 742                         return -EBUSY;
 743
 744                 /* Add the current task, which is the target of sending signal,
 745                  * to the irq_work. The current task may change when queued
 746                  * irq works get executed.
 747                  */
 748                 work->task = current;
 749                 work->sig = sig;
 750                 work->type = type;
 751                 irq_work_queue(&work->irq_work);
 752                 return 0;
 753         }
 754
 755         return group_send_sig_info(sig, SEND_SIG_PRIV, current, type);
 756 }
 757
 758 BPF_CALL_1(bpf_send_signal, u32, sig)
 759 {
 760         return bpf_send_signal_common(sig, PIDTYPE_TGID);
 761 }
 762
 763 static const struct bpf_func_proto bpf_send_signal_proto = {
 764         .func           = bpf_send_signal,
 765         .gpl_only       = false,
 766         .ret_type       = RET_INTEGER,
 767         .arg1_type      = ARG_ANYTHING,
 768 };
 769
 770 BPF_CALL_1(bpf_send_signal_thread, u32, sig)
 771 {
 772         return bpf_send_signal_common(sig, PIDTYPE_PID);
 773 }
 774
 775 static const struct bpf_func_proto bpf_send_signal_thread_proto = {
 776         .func           = bpf_send_signal_thread,
 777         .gpl_only       = false,
 778         .ret_type       = RET_INTEGER,
 779         .arg1_type      = ARG_ANYTHING,
 780 };
 781
 782 const struct bpf_func_proto *
 783 bpf_tracing_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
 784 {
 785         switch (func_id) {
 786         case BPF_FUNC_map_lookup_elem:
 787                 return &bpf_map_lookup_elem_proto;
 788         case BPF_FUNC_map_update_elem:
 789                 return &bpf_map_update_elem_proto;
 790         case BPF_FUNC_map_delete_elem:
 791                 return &bpf_map_delete_elem_proto;
 792         case BPF_FUNC_map_push_elem:
 793                 return &bpf_map_push_elem_proto;
 794         case BPF_FUNC_map_pop_elem:
 795                 return &bpf_map_pop_elem_proto;
 796         case BPF_FUNC_map_peek_elem:
 797                 return &bpf_map_peek_elem_proto;
 798         case BPF_FUNC_ktime_get_ns:
 799                 return &bpf_ktime_get_ns_proto;
 800         case BPF_FUNC_tail_call:
 801                 return &bpf_tail_call_proto;
 802         case BPF_FUNC_get_current_pid_tgid:
 803                 return &bpf_get_current_pid_tgid_proto;
 804         case BPF_FUNC_get_current_task:
 805                 return &bpf_get_current_task_proto;
 806         case BPF_FUNC_get_current_uid_gid:
 807                 return &bpf_get_current_uid_gid_proto;
 808         case BPF_FUNC_get_current_comm:
 809                 return &bpf_get_current_comm_proto;
 810         case BPF_FUNC_trace_printk:
 811                 return bpf_get_trace_printk_proto();
 812         case BPF_FUNC_get_smp_processor_id:
 813                 return &bpf_get_smp_processor_id_proto;
 814         case BPF_FUNC_get_numa_node_id:
 815                 return &bpf_get_numa_node_id_proto;
 816         case BPF_FUNC_perf_event_read:
 817                 return &bpf_perf_event_read_proto;
 818         case BPF_FUNC_probe_write_user:
 819                 return bpf_get_probe_write_proto();
 820         case BPF_FUNC_current_task_under_cgroup:
 821                 return &bpf_current_task_under_cgroup_proto;
 822         case BPF_FUNC_get_prandom_u32:
 823                 return &bpf_get_prandom_u32_proto;
 824         case BPF_FUNC_probe_read_user:
 825                 return &bpf_probe_read_user_proto;
 826         case BPF_FUNC_probe_read_kernel:
 827                 return &bpf_probe_read_kernel_proto;
 828         case BPF_FUNC_probe_read:
 829                 return &bpf_probe_read_compat_proto;
 830         case BPF_FUNC_probe_read_user_str:
 831                 return &bpf_probe_read_user_str_proto;
 832         case BPF_FUNC_probe_read_kernel_str:
 833                 return &bpf_probe_read_kernel_str_proto;
 834         case BPF_FUNC_probe_read_str:
 835                 return &bpf_probe_read_compat_str_proto;
 836 #ifdef CONFIG_CGROUPS
 837         case BPF_FUNC_get_current_cgroup_id:
 838                 return &bpf_get_current_cgroup_id_proto;
 839 #endif
 840         case BPF_FUNC_send_signal:
 841                 return &bpf_send_signal_proto;
 842         case BPF_FUNC_send_signal_thread:
 843                 return &bpf_send_signal_thread_proto;
 844         case BPF_FUNC_perf_event_read_value:
 845                 return &bpf_perf_event_read_value_proto;
 846         case BPF_FUNC_get_ns_current_pid_tgid:
 847                 return &bpf_get_ns_current_pid_tgid_proto;
 848         default:
 849                 return NULL;
 850         }
 851 }
 852
 853 static const struct bpf_func_proto *
 854 kprobe_prog_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
 855 {
 856         switch (func_id) {
 857         case BPF_FUNC_perf_event_output:
 858                 return &bpf_perf_event_output_proto;
 859         case BPF_FUNC_get_stackid:
 860                 return &bpf_get_stackid_proto;
 861         case BPF_FUNC_get_stack:
 862                 return &bpf_get_stack_proto;
 863 #ifdef CONFIG_BPF_KPROBE_OVERRIDE
 864         case BPF_FUNC_override_return:
 865                 return &bpf_override_return_proto;
 866 #endif
 867         default:
 868                 return bpf_tracing_func_proto(func_id, prog);
 869         }
 870 }
 871
 872 /* bpf+kprobe programs can access fields of 'struct pt_regs' */
 873 static bool kprobe_prog_is_valid_access(int off, int size, enum bpf_access_type type,
 874                                         const struct bpf_prog *prog,
 875                                         struct bpf_insn_access_aux *info)
 876 {
 877         if (off < 0 || off >= sizeof(struct pt_regs))
 878                 return false;
 879         if (type != BPF_READ)
 880                 return false;
 881         if (off % size != 0)
 882                 return false;
 883         /*
 884          * Assertion for 32 bit to make sure last 8 byte access
 885          * (BPF_DW) to the last 4 byte member is disallowed.
 886          */
 887         if (off + size > sizeof(struct pt_regs))
 888                 return false;
 889
 890         return true;
 891 }
 892
 893 const struct bpf_verifier_ops kprobe_verifier_ops = {
 894         .get_func_proto  = kprobe_prog_func_proto,
 895         .is_valid_access = kprobe_prog_is_valid_access,
 896 };
 897
 898 const struct bpf_prog_ops kprobe_prog_ops = {
 899 };
 900
 901 BPF_CALL_5(bpf_perf_event_output_tp, void *, tp_buff, struct bpf_map *, map,
 902            u64, flags, void *, data, u64, size)
 903 {
 904         struct pt_regs *regs = *(struct pt_regs **)tp_buff;
 905
 906         /*
 907          * r1 points to perf tracepoint buffer where first 8 bytes are hidden
 908          * from bpf program and contain a pointer to 'struct pt_regs'. Fetch it
 909          * from there and call the same bpf_perf_event_output() helper inline.
 910          */
 911         return ____bpf_perf_event_output(regs, map, flags, data, size);
 912 }
 913
 914 static const struct bpf_func_proto bpf_perf_event_output_proto_tp = {
 915         .func           = bpf_perf_event_output_tp,
 916         .gpl_only       = true,
 917         .ret_type       = RET_INTEGER,
 918         .arg1_type      = ARG_PTR_TO_CTX,
 919         .arg2_type      = ARG_CONST_MAP_PTR,
 920         .arg3_type      = ARG_ANYTHING,
 921         .arg4_type      = ARG_PTR_TO_MEM,
 922         .arg5_type      = ARG_CONST_SIZE_OR_ZERO,
 923 };
 924
 925 BPF_CALL_3(bpf_get_stackid_tp, void *, tp_buff, struct bpf_map *, map,
 926            u64, flags)
 927 {
 928         struct pt_regs *regs = *(struct pt_regs **)tp_buff;
 929
 930         /*
 931          * Same comment as in bpf_perf_event_output_tp(), only that this time
 932          * the other helper's function body cannot be inlined due to being
 933          * external, thus we need to call raw helper function.
 934          */
 935         return bpf_get_stackid((unsigned long) regs, (unsigned long) map,
 936                                flags, 0, 0);
 937 }
 938
 939 static const struct bpf_func_proto bpf_get_stackid_proto_tp = {
 940         .func           = bpf_get_stackid_tp,
 941         .gpl_only       = true,
 942         .ret_type       = RET_INTEGER,
 943         .arg1_type      = ARG_PTR_TO_CTX,
 944         .arg2_type      = ARG_CONST_MAP_PTR,
 945         .arg3_type      = ARG_ANYTHING,
 946 };
 947
 948 BPF_CALL_4(bpf_get_stack_tp, void *, tp_buff, void *, buf, u32, size,
 949            u64, flags)
 950 {
 951         struct pt_regs *regs = *(struct pt_regs **)tp_buff;
 952
 953         return bpf_get_stack((unsigned long) regs, (unsigned long) buf,
 954                              (unsigned long) size, flags, 0);
 955 }
 956
 957 static const struct bpf_func_proto bpf_get_stack_proto_tp = {
 958         .func           = bpf_get_stack_tp,
 959         .gpl_only       = true,
 960         .ret_type       = RET_INTEGER,
 961         .arg1_type      = ARG_PTR_TO_CTX,
 962         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
 963         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 964         .arg4_type      = ARG_ANYTHING,
 965 };
 966
 967 static const struct bpf_func_proto *
 968 tp_prog_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
 969 {
 970         switch (func_id) {
 971         case BPF_FUNC_perf_event_output:
 972                 return &bpf_perf_event_output_proto_tp;
 973         case BPF_FUNC_get_stackid:
 974                 return &bpf_get_stackid_proto_tp;
 975         case BPF_FUNC_get_stack:
 976                 return &bpf_get_stack_proto_tp;
 977         default:
 978                 return bpf_tracing_func_proto(func_id, prog);
 979         }
 980 }
 981
 982 static bool tp_prog_is_valid_access(int off, int size, enum bpf_access_type type,
 983                                     const struct bpf_prog *prog,
 984                                     struct bpf_insn_access_aux *info)
 985 {
 986         if (off < sizeof(void *) || off >= PERF_MAX_TRACE_SIZE)
 987                 return false;
 988         if (type != BPF_READ)
 989                 return false;
 990         if (off % size != 0)
 991                 return false;
 992
 993         BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(__u64));
 994         return true;
 995 }
 996
 997 const struct bpf_verifier_ops tracepoint_verifier_ops = {
 998         .get_func_proto  = tp_prog_func_proto,
 999         .is_valid_access = tp_prog_is_valid_access,
1000 };
1001
1002 const struct bpf_prog_ops tracepoint_prog_ops = {
1003 };
1004
1005 BPF_CALL_3(bpf_perf_prog_read_value, struct bpf_perf_event_data_kern *, ctx,
1006            struct bpf_perf_event_value *, buf, u32, size)
1007 {
1008         int err = -EINVAL;
1009
1010         if (unlikely(size != sizeof(struct bpf_perf_event_value)))
1011                 goto clear;
1012         err = perf_event_read_local(ctx->event, &buf->counter, &buf->enabled,
1013                                     &buf->running);
1014         if (unlikely(err))
1015                 goto clear;
1016         return 0;
1017 clear:
1018         memset(buf, 0, size);
1019         return err;
1020 }
1021
1022 static const struct bpf_func_proto bpf_perf_prog_read_value_proto = {
1023          .func           = bpf_perf_prog_read_value,
1024          .gpl_only       = true,
1025          .ret_type       = RET_INTEGER,
1026          .arg1_type      = ARG_PTR_TO_CTX,
1027          .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
1028          .arg3_type      = ARG_CONST_SIZE,
1029 };
1030
1031 BPF_CALL_4(bpf_read_branch_records, struct bpf_perf_event_data_kern *, ctx,
1032            void *, buf, u32, size, u64, flags)
1033 {
1034 #ifndef CONFIG_X86
1035         return -ENOENT;
1036 #else
1037         static const u32 br_entry_size = sizeof(struct perf_branch_entry);
1038         struct perf_branch_stack *br_stack = ctx->data->br_stack;
1039         u32 to_copy;
1040
1041         if (unlikely(flags & ~BPF_F_GET_BRANCH_RECORDS_SIZE))
1042                 return -EINVAL;
1043
1044         if (unlikely(!br_stack))
1045                 return -EINVAL;
1046
1047         if (flags & BPF_F_GET_BRANCH_RECORDS_SIZE)
1048                 return br_stack->nr * br_entry_size;
1049
1050         if (!buf || (size % br_entry_size != 0))
1051                 return -EINVAL;
1052
1053         to_copy = min_t(u32, br_stack->nr * br_entry_size, size);
1054         memcpy(buf, br_stack->entries, to_copy);
1055
1056         return to_copy;
1057 #endif
1058 }
1059
1060 static const struct bpf_func_proto bpf_read_branch_records_proto = {
1061         .func           = bpf_read_branch_records,
1062         .gpl_only       = true,
1063         .ret_type       = RET_INTEGER,
1064         .arg1_type      = ARG_PTR_TO_CTX,
1065         .arg2_type      = ARG_PTR_TO_MEM_OR_NULL,
1066         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
1067         .arg4_type      = ARG_ANYTHING,
1068 };
1069
1070 static const struct bpf_func_proto *
1071 pe_prog_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
1072 {
1073         switch (func_id) {
1074         case BPF_FUNC_perf_event_output:
1075                 return &bpf_perf_event_output_proto_tp;
1076         case BPF_FUNC_get_stackid:
1077                 return &bpf_get_stackid_proto_tp;
1078         case BPF_FUNC_get_stack:
1079                 return &bpf_get_stack_proto_tp;
1080         case BPF_FUNC_perf_prog_read_value:
1081                 return &bpf_perf_prog_read_value_proto;
1082         case BPF_FUNC_read_branch_records:
1083                 return &bpf_read_branch_records_proto;
1084         default:
1085                 return bpf_tracing_func_proto(func_id, prog);
1086         }
1087 }
1088
1089 /*
1090  * bpf_raw_tp_regs are separate from bpf_pt_regs used from skb/xdp
1091  * to avoid potential recursive reuse issue when/if tracepoints are added
1092  * inside bpf_*_event_output, bpf_get_stackid and/or bpf_get_stack.
1093  *
1094  * Since raw tracepoints run despite bpf_prog_active, support concurrent usage
1095  * in normal, irq, and nmi context.
1096  */
1097 struct bpf_raw_tp_regs {
1098         struct pt_regs regs[3];
1099 };
1100 static DEFINE_PER_CPU(struct bpf_raw_tp_regs, bpf_raw_tp_regs);
1101 static DEFINE_PER_CPU(int, bpf_raw_tp_nest_level);
1102 static struct pt_regs *get_bpf_raw_tp_regs(void)
1103 {
1104         struct bpf_raw_tp_regs *tp_regs = this_cpu_ptr(&bpf_raw_tp_regs);
1105         int nest_level = this_cpu_inc_return(bpf_raw_tp_nest_level);
1106
1107         if (WARN_ON_ONCE(nest_level > ARRAY_SIZE(tp_regs->regs))) {
1108                 this_cpu_dec(bpf_raw_tp_nest_level);
1109                 return ERR_PTR(-EBUSY);
1110         }
1111
1112         return &tp_regs->regs[nest_level - 1];
1113 }
1114
1115 static void put_bpf_raw_tp_regs(void)
1116 {
1117         this_cpu_dec(bpf_raw_tp_nest_level);
1118 }
1119
1120 BPF_CALL_5(bpf_perf_event_output_raw_tp, struct bpf_raw_tracepoint_args *, args,
1121            struct bpf_map *, map, u64, flags, void *, data, u64, size)
1122 {
1123         struct pt_regs *regs = get_bpf_raw_tp_regs();
1124         int ret;
1125
1126         if (IS_ERR(regs))
1127                 return PTR_ERR(regs);
1128
1129         perf_fetch_caller_regs(regs);
1130         ret = ____bpf_perf_event_output(regs, map, flags, data, size);
1131
1132         put_bpf_raw_tp_regs();
1133         return ret;
1134 }
1135
1136 static const struct bpf_func_proto bpf_perf_event_output_proto_raw_tp = {
1137         .func           = bpf_perf_event_output_raw_tp,
1138         .gpl_only       = true,
1139         .ret_type       = RET_INTEGER,
1140         .arg1_type      = ARG_PTR_TO_CTX,
1141         .arg2_type      = ARG_CONST_MAP_PTR,
1142         .arg3_type      = ARG_ANYTHING,
1143         .arg4_type      = ARG_PTR_TO_MEM,
1144         .arg5_type      = ARG_CONST_SIZE_OR_ZERO,
1145 };
1146
1147 extern const struct bpf_func_proto bpf_skb_output_proto;
1148 extern const struct bpf_func_proto bpf_xdp_output_proto;
1149
1150 BPF_CALL_3(bpf_get_stackid_raw_tp, struct bpf_raw_tracepoint_args *, args,
1151            struct bpf_map *, map, u64, flags)
1152 {
1153         struct pt_regs *regs = get_bpf_raw_tp_regs();
1154         int ret;
1155
1156         if (IS_ERR(regs))
1157                 return PTR_ERR(regs);
1158
1159         perf_fetch_caller_regs(regs);
1160         /* similar to bpf_perf_event_output_tp, but pt_regs fetched differently */
1161         ret = bpf_get_stackid((unsigned long) regs, (unsigned long) map,
1162                               flags, 0, 0);
1163         put_bpf_raw_tp_regs();
1164         return ret;
1165 }
1166
1167 static const struct bpf_func_proto bpf_get_stackid_proto_raw_tp = {
1168         .func           = bpf_get_stackid_raw_tp,
1169         .gpl_only       = true,
1170         .ret_type       = RET_INTEGER,
1171         .arg1_type      = ARG_PTR_TO_CTX,
1172         .arg2_type      = ARG_CONST_MAP_PTR,
1173         .arg3_type      = ARG_ANYTHING,
1174 };
1175
1176 BPF_CALL_4(bpf_get_stack_raw_tp, struct bpf_raw_tracepoint_args *, args,
1177            void *, buf, u32, size, u64, flags)
1178 {
1179         struct pt_regs *regs = get_bpf_raw_tp_regs();
1180         int ret;
1181
1182         if (IS_ERR(regs))
1183                 return PTR_ERR(regs);
1184
1185         perf_fetch_caller_regs(regs);
1186         ret = bpf_get_stack((unsigned long) regs, (unsigned long) buf,
1187                             (unsigned long) size, flags, 0);
1188         put_bpf_raw_tp_regs();
1189         return ret;
1190 }
1191
1192 static const struct bpf_func_proto bpf_get_stack_proto_raw_tp = {
1193         .func           = bpf_get_stack_raw_tp,
1194         .gpl_only       = true,
1195         .ret_type       = RET_INTEGER,
1196         .arg1_type      = ARG_PTR_TO_CTX,
1197         .arg2_type      = ARG_PTR_TO_MEM,
1198         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
1199         .arg4_type      = ARG_ANYTHING,
1200 };
1201
1202 static const struct bpf_func_proto *
1203 raw_tp_prog_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
1204 {
1205         switch (func_id) {
1206         case BPF_FUNC_perf_event_output:
1207                 return &bpf_perf_event_output_proto_raw_tp;
1208         case BPF_FUNC_get_stackid:
1209                 return &bpf_get_stackid_proto_raw_tp;
1210         case BPF_FUNC_get_stack:
1211                 return &bpf_get_stack_proto_raw_tp;
1212         default:
1213                 return bpf_tracing_func_proto(func_id, prog);
1214         }
1215 }
1216
1217 static const struct bpf_func_proto *
1218 tracing_prog_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
1219 {
1220         switch (func_id) {
1221 #ifdef CONFIG_NET
1222         case BPF_FUNC_skb_output:
1223                 return &bpf_skb_output_proto;
1224         case BPF_FUNC_xdp_output:
1225                 return &bpf_xdp_output_proto;
1226 #endif
1227         default:
1228                 return raw_tp_prog_func_proto(func_id, prog);
1229         }
1230 }
1231
1232 static bool raw_tp_prog_is_valid_access(int off, int size,
1233                                         enum bpf_access_type type,
1234                                         const struct bpf_prog *prog,
1235                                         struct bpf_insn_access_aux *info)
1236 {
1237         if (off < 0 || off >= sizeof(__u64) * MAX_BPF_FUNC_ARGS)
1238                 return false;
1239         if (type != BPF_READ)
1240                 return false;
1241         if (off % size != 0)
1242                 return false;
1243         return true;
1244 }
1245
1246 static bool tracing_prog_is_valid_access(int off, int size,
1247                                          enum bpf_access_type type,
1248                                          const struct bpf_prog *prog,
1249                                          struct bpf_insn_access_aux *info)
1250 {
1251         if (off < 0 || off >= sizeof(__u64) * MAX_BPF_FUNC_ARGS)
1252                 return false;
1253         if (type != BPF_READ)
1254                 return false;
1255         if (off % size != 0)
1256                 return false;
1257         return btf_ctx_access(off, size, type, prog, info);
1258 }
1259
1260 int __weak bpf_prog_test_run_tracing(struct bpf_prog *prog,
1261                                      const union bpf_attr *kattr,
1262                                      union bpf_attr __user *uattr)
1263 {
1264         return -ENOTSUPP;
1265 }
1266
1267 const struct bpf_verifier_ops raw_tracepoint_verifier_ops = {
1268         .get_func_proto  = raw_tp_prog_func_proto,
1269         .is_valid_access = raw_tp_prog_is_valid_access,
1270 };
1271
1272 const struct bpf_prog_ops raw_tracepoint_prog_ops = {
1273 };
1274
1275 const struct bpf_verifier_ops tracing_verifier_ops = {
1276         .get_func_proto  = tracing_prog_func_proto,
1277         .is_valid_access = tracing_prog_is_valid_access,
1278 };
1279
1280 const struct bpf_prog_ops tracing_prog_ops = {
1281         .test_run = bpf_prog_test_run_tracing,
1282 };
1283
1284 static bool raw_tp_writable_prog_is_valid_access(int off, int size,
1285                                                  enum bpf_access_type type,
1286                                                  const struct bpf_prog *prog,
1287                                                  struct bpf_insn_access_aux *info)
1288 {
1289         if (off == 0) {
1290                 if (size != sizeof(u64) || type != BPF_READ)
1291                         return false;
1292                 info->reg_type = PTR_TO_TP_BUFFER;
1293         }
1294         return raw_tp_prog_is_valid_access(off, size, type, prog, info);
1295 }
1296
1297 const struct bpf_verifier_ops raw_tracepoint_writable_verifier_ops = {
1298         .get_func_proto  = raw_tp_prog_func_proto,
1299         .is_valid_access = raw_tp_writable_prog_is_valid_access,
1300 };
1301
1302 const struct bpf_prog_ops raw_tracepoint_writable_prog_ops = {
1303 };
1304
1305 static bool pe_prog_is_valid_access(int off, int size, enum bpf_access_type type,
1306                                     const struct bpf_prog *prog,
1307                                     struct bpf_insn_access_aux *info)
1308 {
1309         const int size_u64 = sizeof(u64);
1310
1311         if (off < 0 || off >= sizeof(struct bpf_perf_event_data))
1312                 return false;
1313         if (type != BPF_READ)
1314                 return false;
1315         if (off % size != 0) {
1316                 if (sizeof(unsigned long) != 4)
1317                         return false;
1318                 if (size != 8)
1319                         return false;
1320                 if (off % size != 4)
1321                         return false;
1322         }
1323
1324         switch (off) {
1325         case bpf_ctx_range(struct bpf_perf_event_data, sample_period):
1326                 bpf_ctx_record_field_size(info, size_u64);
1327                 if (!bpf_ctx_narrow_access_ok(off, size, size_u64))
1328                         return false;
1329                 break;
1330         case bpf_ctx_range(struct bpf_perf_event_data, addr):
1331                 bpf_ctx_record_field_size(info, size_u64);
1332                 if (!bpf_ctx_narrow_access_ok(off, size, size_u64))
1333                         return false;
1334                 break;
1335         default:
1336                 if (size != sizeof(long))
1337                         return false;
1338         }
1339
1340         return true;
1341 }
1342
1343 static u32 pe_prog_convert_ctx_access(enum bpf_access_type type,
1344                                       const struct bpf_insn *si,
1345                                       struct bpf_insn *insn_buf,
1346                                       struct bpf_prog *prog, u32 *target_size)
1347 {
1348         struct bpf_insn *insn = insn_buf;
1349
1350         switch (si->off) {
1351         case offsetof(struct bpf_perf_event_data, sample_period):
1352                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct bpf_perf_event_data_kern,
1353                                                        data), si->dst_reg, si->src_reg,
1354                                       offsetof(struct bpf_perf_event_data_kern, data));
1355                 *insn++ = BPF_LDX_MEM(BPF_DW, si->dst_reg, si->dst_reg,
1356                                       bpf_target_off(struct perf_sample_data, period, 8,
1357                                                      target_size));
1358                 break;
1359         case offsetof(struct bpf_perf_event_data, addr):
1360                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct bpf_perf_event_data_kern,
1361                                                        data), si->dst_reg, si->src_reg,
1362                                       offsetof(struct bpf_perf_event_data_kern, data));
1363                 *insn++ = BPF_LDX_MEM(BPF_DW, si->dst_reg, si->dst_reg,
1364                                       bpf_target_off(struct perf_sample_data, addr, 8,
1365                                                      target_size));
1366                 break;
1367         default:
1368                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct bpf_perf_event_data_kern,
1369                                                        regs), si->dst_reg, si->src_reg,
1370                                       offsetof(struct bpf_perf_event_data_kern, regs));
1371                 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(long), si->dst_reg, si->dst_reg,
1372                                       si->off);
1373                 break;
1374         }
1375
1376         return insn - insn_buf;
1377 }
1378
1379 const struct bpf_verifier_ops perf_event_verifier_ops = {
1380         .get_func_proto         = pe_prog_func_proto,
1381         .is_valid_access        = pe_prog_is_valid_access,
1382         .convert_ctx_access     = pe_prog_convert_ctx_access,
1383 };
1384
1385 const struct bpf_prog_ops perf_event_prog_ops = {
1386 };
1387
1388 static DEFINE_MUTEX(bpf_event_mutex);
1389
1390 #define BPF_TRACE_MAX_PROGS 64
1391
1392 int perf_event_attach_bpf_prog(struct perf_event *event,
1393                                struct bpf_prog *prog)
1394 {
1395         struct bpf_prog_array *old_array;
1396         struct bpf_prog_array *new_array;
1397         int ret = -EEXIST;
1398
1399         /*
1400          * Kprobe override only works if they are on the function entry,
1401          * and only if they are on the opt-in list.
1402          */
1403         if (prog->kprobe_override &&
1404             (!trace_kprobe_on_func_entry(event->tp_event) ||
1405              !trace_kprobe_error_injectable(event->tp_event)))
1406                 return -EINVAL;
1407
1408         mutex_lock(&bpf_event_mutex);
1409
1410         if (event->prog)
1411                 goto unlock;
1412
1413         old_array = bpf_event_rcu_dereference(event->tp_event->prog_array);
1414         if (old_array &&
1415             bpf_prog_array_length(old_array) >= BPF_TRACE_MAX_PROGS) {
1416                 ret = -E2BIG;
1417                 goto unlock;
1418         }
1419
1420         ret = bpf_prog_array_copy(old_array, NULL, prog, &new_array);
1421         if (ret < 0)
1422                 goto unlock;
1423
1424         /* set the new array to event->tp_event and set event->prog */
1425         event->prog = prog;
1426         rcu_assign_pointer(event->tp_event->prog_array, new_array);
1427         bpf_prog_array_free(old_array);
1428
1429 unlock:
1430         mutex_unlock(&bpf_event_mutex);
1431         return ret;
1432 }
1433
1434 void perf_event_detach_bpf_prog(struct perf_event *event)
1435 {
1436         struct bpf_prog_array *old_array;
1437         struct bpf_prog_array *new_array;
1438         int ret;
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         ret = bpf_prog_array_copy(old_array, event->prog, NULL, &new_array);
1447         if (ret == -ENOENT)
1448                 goto unlock;
1449         if (ret < 0) {
1450                 bpf_prog_array_delete_safe(old_array, event->prog);
1451         } else {
1452                 rcu_assign_pointer(event->tp_event->prog_array, new_array);
1453                 bpf_prog_array_free(old_array);
1454         }
1455
1456         bpf_prog_put(event->prog);
1457         event->prog = NULL;
1458
1459 unlock:
1460         mutex_unlock(&bpf_event_mutex);
1461 }
1462
1463 int perf_event_query_prog_array(struct perf_event *event, void __user *info)
1464 {
1465         struct perf_event_query_bpf __user *uquery = info;
1466         struct perf_event_query_bpf query = {};
1467         struct bpf_prog_array *progs;
1468         u32 *ids, prog_cnt, ids_len;
1469         int ret;
1470
1471         if (!capable(CAP_SYS_ADMIN))
1472                 return -EPERM;
1473         if (event->attr.type != PERF_TYPE_TRACEPOINT)
1474                 return -EINVAL;
1475         if (copy_from_user(&query, uquery, sizeof(query)))
1476                 return -EFAULT;
1477
1478         ids_len = query.ids_len;
1479         if (ids_len > BPF_TRACE_MAX_PROGS)
1480                 return -E2BIG;
1481         ids = kcalloc(ids_len, sizeof(u32), GFP_USER | __GFP_NOWARN);
1482         if (!ids)
1483                 return -ENOMEM;
1484         /*
1485          * The above kcalloc returns ZERO_SIZE_PTR when ids_len = 0, which
1486          * is required when user only wants to check for uquery->prog_cnt.
1487          * There is no need to check for it since the case is handled
1488          * gracefully in bpf_prog_array_copy_info.
1489          */
1490
1491         mutex_lock(&bpf_event_mutex);
1492         progs = bpf_event_rcu_dereference(event->tp_event->prog_array);
1493         ret = bpf_prog_array_copy_info(progs, ids, ids_len, &prog_cnt);
1494         mutex_unlock(&bpf_event_mutex);
1495
1496         if (copy_to_user(&uquery->prog_cnt, &prog_cnt, sizeof(prog_cnt)) ||
1497             copy_to_user(uquery->ids, ids, ids_len * sizeof(u32)))
1498                 ret = -EFAULT;
1499
1500         kfree(ids);
1501         return ret;
1502 }
1503
1504 extern struct bpf_raw_event_map __start__bpf_raw_tp[];
1505 extern struct bpf_raw_event_map __stop__bpf_raw_tp[];
1506
1507 struct bpf_raw_event_map *bpf_get_raw_tracepoint(const char *name)
1508 {
1509         struct bpf_raw_event_map *btp = __start__bpf_raw_tp;
1510
1511         for (; btp < __stop__bpf_raw_tp; btp++) {
1512                 if (!strcmp(btp->tp->name, name))
1513                         return btp;
1514         }
1515
1516         return bpf_get_raw_tracepoint_module(name);
1517 }
1518
1519 void bpf_put_raw_tracepoint(struct bpf_raw_event_map *btp)
1520 {
1521         struct module *mod = __module_address((unsigned long)btp);
1522
1523         if (mod)
1524                 module_put(mod);
1525 }
1526
1527 static __always_inline
1528 void __bpf_trace_run(struct bpf_prog *prog, u64 *args)
1529 {
1530         cant_sleep();
1531         rcu_read_lock();
1532         (void) BPF_PROG_RUN(prog, args);
1533         rcu_read_unlock();
1534 }
1535
1536 #define UNPACK(...)                     __VA_ARGS__
1537 #define REPEAT_1(FN, DL, X, ...)        FN(X)
1538 #define REPEAT_2(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_1(FN, DL, __VA_ARGS__)
1539 #define REPEAT_3(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_2(FN, DL, __VA_ARGS__)
1540 #define REPEAT_4(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_3(FN, DL, __VA_ARGS__)
1541 #define REPEAT_5(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_4(FN, DL, __VA_ARGS__)
1542 #define REPEAT_6(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_5(FN, DL, __VA_ARGS__)
1543 #define REPEAT_7(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_6(FN, DL, __VA_ARGS__)
1544 #define REPEAT_8(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_7(FN, DL, __VA_ARGS__)
1545 #define REPEAT_9(FN, DL, X, ...)        FN(X) UNPACK DL REPEAT_8(FN, DL, __VA_ARGS__)
1546 #define REPEAT_10(FN, DL, X, ...)       FN(X) UNPACK DL REPEAT_9(FN, DL, __VA_ARGS__)
1547 #define REPEAT_11(FN, DL, X, ...)       FN(X) UNPACK DL REPEAT_10(FN, DL, __VA_ARGS__)
1548 #define REPEAT_12(FN, DL, X, ...)       FN(X) UNPACK DL REPEAT_11(FN, DL, __VA_ARGS__)
1549 #define REPEAT(X, FN, DL, ...)          REPEAT_##X(FN, DL, __VA_ARGS__)
1550
1551 #define SARG(X)         u64 arg##X
1552 #define COPY(X)         args[X] = arg##X
1553
1554 #define __DL_COM        (,)
1555 #define __DL_SEM        (;)
1556
1557 #define __SEQ_0_11      0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11
1558
1559 #define BPF_TRACE_DEFN_x(x)                                             \
1560         void bpf_trace_run##x(struct bpf_prog *prog,                    \
1561                               REPEAT(x, SARG, __DL_COM, __SEQ_0_11))    \
1562         {                                                               \
1563                 u64 args[x];                                            \
1564                 REPEAT(x, COPY, __DL_SEM, __SEQ_0_11);                  \
1565                 __bpf_trace_run(prog, args);                            \
1566         }                                                               \
1567         EXPORT_SYMBOL_GPL(bpf_trace_run##x)
1568 BPF_TRACE_DEFN_x(1);
1569 BPF_TRACE_DEFN_x(2);
1570 BPF_TRACE_DEFN_x(3);
1571 BPF_TRACE_DEFN_x(4);
1572 BPF_TRACE_DEFN_x(5);
1573 BPF_TRACE_DEFN_x(6);
1574 BPF_TRACE_DEFN_x(7);
1575 BPF_TRACE_DEFN_x(8);
1576 BPF_TRACE_DEFN_x(9);
1577 BPF_TRACE_DEFN_x(10);
1578 BPF_TRACE_DEFN_x(11);
1579 BPF_TRACE_DEFN_x(12);
1580
1581 static int __bpf_probe_register(struct bpf_raw_event_map *btp, struct bpf_prog *prog)
1582 {
1583         struct tracepoint *tp = btp->tp;
1584
1585         /*
1586          * check that program doesn't access arguments beyond what's
1587          * available in this tracepoint
1588          */
1589         if (prog->aux->max_ctx_offset > btp->num_args * sizeof(u64))
1590                 return -EINVAL;
1591
1592         if (prog->aux->max_tp_access > btp->writable_size)
1593                 return -EINVAL;
1594
1595         return tracepoint_probe_register(tp, (void *)btp->bpf_func, prog);
1596 }
1597
1598 int bpf_probe_register(struct bpf_raw_event_map *btp, struct bpf_prog *prog)
1599 {
1600         return __bpf_probe_register(btp, prog);
1601 }
1602
1603 int bpf_probe_unregister(struct bpf_raw_event_map *btp, struct bpf_prog *prog)
1604 {
1605         return tracepoint_probe_unregister(btp->tp, (void *)btp->bpf_func, prog);
1606 }
1607
1608 int bpf_get_perf_event_info(const struct perf_event *event, u32 *prog_id,
1609                             u32 *fd_type, const char **buf,
1610                             u64 *probe_offset, u64 *probe_addr)
1611 {
1612         bool is_tracepoint, is_syscall_tp;
1613         struct bpf_prog *prog;
1614         int flags, err = 0;
1615
1616         prog = event->prog;
1617         if (!prog)
1618                 return -ENOENT;
1619
1620         /* not supporting BPF_PROG_TYPE_PERF_EVENT yet */
1621         if (prog->type == BPF_PROG_TYPE_PERF_EVENT)
1622                 return -EOPNOTSUPP;
1623
1624         *prog_id = prog->aux->id;
1625         flags = event->tp_event->flags;
1626         is_tracepoint = flags & TRACE_EVENT_FL_TRACEPOINT;
1627         is_syscall_tp = is_syscall_trace_event(event->tp_event);
1628
1629         if (is_tracepoint || is_syscall_tp) {
1630                 *buf = is_tracepoint ? event->tp_event->tp->name
1631                                      : event->tp_event->name;
1632                 *fd_type = BPF_FD_TYPE_TRACEPOINT;
1633                 *probe_offset = 0x0;
1634                 *probe_addr = 0x0;
1635         } else {
1636                 /* kprobe/uprobe */
1637                 err = -EOPNOTSUPP;
1638 #ifdef CONFIG_KPROBE_EVENTS
1639                 if (flags & TRACE_EVENT_FL_KPROBE)
1640                         err = bpf_get_kprobe_info(event, fd_type, buf,
1641                                                   probe_offset, probe_addr,
1642                                                   event->attr.type == PERF_TYPE_TRACEPOINT);
1643 #endif
1644 #ifdef CONFIG_UPROBE_EVENTS
1645                 if (flags & TRACE_EVENT_FL_UPROBE)
1646                         err = bpf_get_uprobe_info(event, fd_type, buf,
1647                                                   probe_offset,
1648                                                   event->attr.type == PERF_TYPE_TRACEPOINT);
1649 #endif
1650         }
1651
1652         return err;
1653 }
1654
1655 static int __init send_signal_irq_work_init(void)
1656 {
1657         int cpu;
1658         struct send_signal_irq_work *work;
1659
1660         for_each_possible_cpu(cpu) {
1661                 work = per_cpu_ptr(&send_signal_work, cpu);
1662                 init_irq_work(&work->irq_work, do_bpf_send_signal);
1663         }
1664         return 0;
1665 }
1666
1667 subsys_initcall(send_signal_irq_work_init);
1668
1669 #ifdef CONFIG_MODULES
1670 static int bpf_event_notify(struct notifier_block *nb, unsigned long op,
1671                             void *module)
1672 {
1673         struct bpf_trace_module *btm, *tmp;
1674         struct module *mod = module;
1675
1676         if (mod->num_bpf_raw_events == 0 ||
1677             (op != MODULE_STATE_COMING && op != MODULE_STATE_GOING))
1678                 return 0;
1679
1680         mutex_lock(&bpf_module_mutex);
1681
1682         switch (op) {
1683         case MODULE_STATE_COMING:
1684                 btm = kzalloc(sizeof(*btm), GFP_KERNEL);
1685                 if (btm) {
1686                         btm->module = module;
1687                         list_add(&btm->list, &bpf_trace_modules);
1688                 }
1689                 break;
1690         case MODULE_STATE_GOING:
1691                 list_for_each_entry_safe(btm, tmp, &bpf_trace_modules, list) {
1692                         if (btm->module == module) {
1693                                 list_del(&btm->list);
1694                                 kfree(btm);
1695                                 break;
1696                         }
1697                 }
1698                 break;
1699         }
1700
1701         mutex_unlock(&bpf_module_mutex);
1702
1703         return 0;
1704 }
1705
1706 static struct notifier_block bpf_module_nb = {
1707         .notifier_call = bpf_event_notify,
1708 };
1709
1710 static int __init bpf_event_init(void)
1711 {
1712         register_module_notifier(&bpf_module_nb);
1713         return 0;
1714 }
1715
1716 fs_initcall(bpf_event_init);
1717 #endif /* CONFIG_MODULES */