kernel/trace/trace_syscalls.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <trace/syscall.h>
   3 #include <trace/events/syscalls.h>
   4 #include <linux/syscalls.h>
   5 #include <linux/slab.h>
   6 #include <linux/kernel.h>
   7 #include <linux/module.h>       /* for MODULE_NAME_LEN via KSYM_SYMBOL_LEN */
   8 #include <linux/ftrace.h>
   9 #include <linux/perf_event.h>
  10 #include <linux/xarray.h>
  11 #include <asm/syscall.h>
  12
  13 #include "trace_output.h"
  14 #include "trace.h"
  15
  16 static DEFINE_MUTEX(syscall_trace_lock);
  17
  18 static int syscall_enter_register(struct trace_event_call *event,
  19                                  enum trace_reg type, void *data);
  20 static int syscall_exit_register(struct trace_event_call *event,
  21                                  enum trace_reg type, void *data);
  22
  23 static struct list_head *
  24 syscall_get_enter_fields(struct trace_event_call *call)
  25 {
  26         struct syscall_metadata *entry = call->data;
  27
  28         return &entry->enter_fields;
  29 }
  30
  31 extern struct syscall_metadata *__start_syscalls_metadata[];
  32 extern struct syscall_metadata *__stop_syscalls_metadata[];
  33
  34 static DEFINE_XARRAY(syscalls_metadata_sparse);
  35 static struct syscall_metadata **syscalls_metadata;
  36
  37 #ifndef ARCH_HAS_SYSCALL_MATCH_SYM_NAME
  38 static inline bool arch_syscall_match_sym_name(const char *sym, const char *name)
  39 {
  40         /*
  41          * Only compare after the "sys" prefix. Archs that use
  42          * syscall wrappers may have syscalls symbols aliases prefixed
  43          * with ".SyS" or ".sys" instead of "sys", leading to an unwanted
  44          * mismatch.
  45          */
  46         return !strcmp(sym + 3, name + 3);
  47 }
  48 #endif
  49
  50 #ifdef ARCH_TRACE_IGNORE_COMPAT_SYSCALLS
  51 /*
  52  * Some architectures that allow for 32bit applications
  53  * to run on a 64bit kernel, do not map the syscalls for
  54  * the 32bit tasks the same as they do for 64bit tasks.
  55  *
  56  *     *cough*x86*cough*
  57  *
  58  * In such a case, instead of reporting the wrong syscalls,
  59  * simply ignore them.
  60  *
  61  * For an arch to ignore the compat syscalls it needs to
  62  * define ARCH_TRACE_IGNORE_COMPAT_SYSCALLS as well as
  63  * define the function arch_trace_is_compat_syscall() to let
  64  * the tracing system know that it should ignore it.
  65  */
  66 static int
  67 trace_get_syscall_nr(struct task_struct *task, struct pt_regs *regs)
  68 {
  69         if (unlikely(arch_trace_is_compat_syscall(regs)))
  70                 return -1;
  71
  72         return syscall_get_nr(task, regs);
  73 }
  74 #else
  75 static inline int
  76 trace_get_syscall_nr(struct task_struct *task, struct pt_regs *regs)
  77 {
  78         return syscall_get_nr(task, regs);
  79 }
  80 #endif /* ARCH_TRACE_IGNORE_COMPAT_SYSCALLS */
  81
  82 static __init struct syscall_metadata *
  83 find_syscall_meta(unsigned long syscall)
  84 {
  85         struct syscall_metadata **start;
  86         struct syscall_metadata **stop;
  87         char str[KSYM_SYMBOL_LEN];
  88
  89
  90         start = __start_syscalls_metadata;
  91         stop = __stop_syscalls_metadata;
  92         kallsyms_lookup(syscall, NULL, NULL, NULL, str);
  93
  94         if (arch_syscall_match_sym_name(str, "sys_ni_syscall"))
  95                 return NULL;
  96
  97         for ( ; start < stop; start++) {
  98                 if ((*start)->name && arch_syscall_match_sym_name(str, (*start)->name))
  99                         return *start;
 100         }
 101         return NULL;
 102 }
 103
 104 static struct syscall_metadata *syscall_nr_to_meta(int nr)
 105 {
 106         if (IS_ENABLED(CONFIG_HAVE_SPARSE_SYSCALL_NR))
 107                 return xa_load(&syscalls_metadata_sparse, (unsigned long)nr);
 108
 109         if (!syscalls_metadata || nr >= NR_syscalls || nr < 0)
 110                 return NULL;
 111
 112         return syscalls_metadata[nr];
 113 }
 114
 115 const char *get_syscall_name(int syscall)
 116 {
 117         struct syscall_metadata *entry;
 118
 119         entry = syscall_nr_to_meta(syscall);
 120         if (!entry)
 121                 return NULL;
 122
 123         return entry->name;
 124 }
 125
 126 static enum print_line_t
 127 print_syscall_enter(struct trace_iterator *iter, int flags,
 128                     struct trace_event *event)
 129 {
 130         struct trace_array *tr = iter->tr;
 131         struct trace_seq *s = &iter->seq;
 132         struct trace_entry *ent = iter->ent;
 133         struct syscall_trace_enter *trace;
 134         struct syscall_metadata *entry;
 135         int i, syscall;
 136
 137         trace = (typeof(trace))ent;
 138         syscall = trace->nr;
 139         entry = syscall_nr_to_meta(syscall);
 140
 141         if (!entry)
 142                 goto end;
 143
 144         if (entry->enter_event->event.type != ent->type) {
 145                 WARN_ON_ONCE(1);
 146                 goto end;
 147         }
 148
 149         trace_seq_printf(s, "%s(", entry->name);
 150
 151         for (i = 0; i < entry->nb_args; i++) {
 152
 153                 if (trace_seq_has_overflowed(s))
 154                         goto end;
 155
 156                 /* parameter types */
 157                 if (tr && tr->trace_flags & TRACE_ITER_VERBOSE)
 158                         trace_seq_printf(s, "%s ", entry->types[i]);
 159
 160                 /* parameter values */
 161                 trace_seq_printf(s, "%s: %lx%s", entry->args[i],
 162                                  trace->args[i],
 163                                  i == entry->nb_args - 1 ? "" : ", ");
 164         }
 165
 166         trace_seq_putc(s, ')');
 167 end:
 168         trace_seq_putc(s, '\n');
 169
 170         return trace_handle_return(s);
 171 }
 172
 173 static enum print_line_t
 174 print_syscall_exit(struct trace_iterator *iter, int flags,
 175                    struct trace_event *event)
 176 {
 177         struct trace_seq *s = &iter->seq;
 178         struct trace_entry *ent = iter->ent;
 179         struct syscall_trace_exit *trace;
 180         int syscall;
 181         struct syscall_metadata *entry;
 182
 183         trace = (typeof(trace))ent;
 184         syscall = trace->nr;
 185         entry = syscall_nr_to_meta(syscall);
 186
 187         if (!entry) {
 188                 trace_seq_putc(s, '\n');
 189                 goto out;
 190         }
 191
 192         if (entry->exit_event->event.type != ent->type) {
 193                 WARN_ON_ONCE(1);
 194                 return TRACE_TYPE_UNHANDLED;
 195         }
 196
 197         trace_seq_printf(s, "%s -> 0x%lx\n", entry->name,
 198                                 trace->ret);
 199
 200  out:
 201         return trace_handle_return(s);
 202 }
 203
 204 #define SYSCALL_FIELD(_type, _name) {                                   \
 205         .type = #_type, .name = #_name,                                 \
 206         .size = sizeof(_type), .align = __alignof__(_type),             \
 207         .is_signed = is_signed_type(_type), .filter_type = FILTER_OTHER }
 208
 209 static int __init
 210 __set_enter_print_fmt(struct syscall_metadata *entry, char *buf, int len)
 211 {
 212         int i;
 213         int pos = 0;
 214
 215         /* When len=0, we just calculate the needed length */
 216 #define LEN_OR_ZERO (len ? len - pos : 0)
 217
 218         pos += snprintf(buf + pos, LEN_OR_ZERO, "\"");
 219         for (i = 0; i < entry->nb_args; i++) {
 220                 pos += snprintf(buf + pos, LEN_OR_ZERO, "%s: 0x%%0%zulx%s",
 221                                 entry->args[i], sizeof(unsigned long),
 222                                 i == entry->nb_args - 1 ? "" : ", ");
 223         }
 224         pos += snprintf(buf + pos, LEN_OR_ZERO, "\"");
 225
 226         for (i = 0; i < entry->nb_args; i++) {
 227                 pos += snprintf(buf + pos, LEN_OR_ZERO,
 228                                 ", ((unsigned long)(REC->%s))", entry->args[i]);
 229         }
 230
 231 #undef LEN_OR_ZERO
 232
 233         /* return the length of print_fmt */
 234         return pos;
 235 }
 236
 237 static int __init set_syscall_print_fmt(struct trace_event_call *call)
 238 {
 239         char *print_fmt;
 240         int len;
 241         struct syscall_metadata *entry = call->data;
 242
 243         if (entry->enter_event != call) {
 244                 call->print_fmt = "\"0x%lx\", REC->ret";
 245                 return 0;
 246         }
 247
 248         /* First: called with 0 length to calculate the needed length */
 249         len = __set_enter_print_fmt(entry, NULL, 0);
 250
 251         print_fmt = kmalloc(len + 1, GFP_KERNEL);
 252         if (!print_fmt)
 253                 return -ENOMEM;
 254
 255         /* Second: actually write the @print_fmt */
 256         __set_enter_print_fmt(entry, print_fmt, len + 1);
 257         call->print_fmt = print_fmt;
 258
 259         return 0;
 260 }
 261
 262 static void __init free_syscall_print_fmt(struct trace_event_call *call)
 263 {
 264         struct syscall_metadata *entry = call->data;
 265
 266         if (entry->enter_event == call)
 267                 kfree(call->print_fmt);
 268 }
 269
 270 static int __init syscall_enter_define_fields(struct trace_event_call *call)
 271 {
 272         struct syscall_trace_enter trace;
 273         struct syscall_metadata *meta = call->data;
 274         int offset = offsetof(typeof(trace), args);
 275         int ret = 0;
 276         int i;
 277
 278         for (i = 0; i < meta->nb_args; i++) {
 279                 ret = trace_define_field(call, meta->types[i],
 280                                          meta->args[i], offset,
 281                                          sizeof(unsigned long), 0,
 282                                          FILTER_OTHER);
 283                 if (ret)
 284                         break;
 285                 offset += sizeof(unsigned long);
 286         }
 287
 288         return ret;
 289 }
 290
 291 static void ftrace_syscall_enter(void *data, struct pt_regs *regs, long id)
 292 {
 293         struct trace_array *tr = data;
 294         struct trace_event_file *trace_file;
 295         struct syscall_trace_enter *entry;
 296         struct syscall_metadata *sys_data;
 297         struct trace_event_buffer fbuffer;
 298         unsigned long args[6];
 299         int syscall_nr;
 300         int size;
 301
 302         /*
 303          * Syscall probe called with preemption enabled, but the ring
 304          * buffer and per-cpu data require preemption to be disabled.
 305          */
 306         might_fault();
 307         guard(preempt_notrace)();
 308
 309         syscall_nr = trace_get_syscall_nr(current, regs);
 310         if (syscall_nr < 0 || syscall_nr >= NR_syscalls)
 311                 return;
 312
 313         /* Here we're inside tp handler's rcu_read_lock_sched (__DO_TRACE) */
 314         trace_file = rcu_dereference_sched(tr->enter_syscall_files[syscall_nr]);
 315         if (!trace_file)
 316                 return;
 317
 318         if (trace_trigger_soft_disabled(trace_file))
 319                 return;
 320
 321         sys_data = syscall_nr_to_meta(syscall_nr);
 322         if (!sys_data)
 323                 return;
 324
 325         size = sizeof(*entry) + sizeof(unsigned long) * sys_data->nb_args;
 326
 327         entry = trace_event_buffer_reserve(&fbuffer, trace_file, size);
 328         if (!entry)
 329                 return;
 330
 331         entry = ring_buffer_event_data(fbuffer.event);
 332         entry->nr = syscall_nr;
 333         syscall_get_arguments(current, regs, args);
 334         memcpy(entry->args, args, sizeof(unsigned long) * sys_data->nb_args);
 335
 336         trace_event_buffer_commit(&fbuffer);
 337 }
 338
 339 static void ftrace_syscall_exit(void *data, struct pt_regs *regs, long ret)
 340 {
 341         struct trace_array *tr = data;
 342         struct trace_event_file *trace_file;
 343         struct syscall_trace_exit *entry;
 344         struct syscall_metadata *sys_data;
 345         struct trace_event_buffer fbuffer;
 346         int syscall_nr;
 347
 348         /*
 349          * Syscall probe called with preemption enabled, but the ring
 350          * buffer and per-cpu data require preemption to be disabled.
 351          */
 352         might_fault();
 353         guard(preempt_notrace)();
 354
 355         syscall_nr = trace_get_syscall_nr(current, regs);
 356         if (syscall_nr < 0 || syscall_nr >= NR_syscalls)
 357                 return;
 358
 359         /* Here we're inside tp handler's rcu_read_lock_sched (__DO_TRACE()) */
 360         trace_file = rcu_dereference_sched(tr->exit_syscall_files[syscall_nr]);
 361         if (!trace_file)
 362                 return;
 363
 364         if (trace_trigger_soft_disabled(trace_file))
 365                 return;
 366
 367         sys_data = syscall_nr_to_meta(syscall_nr);
 368         if (!sys_data)
 369                 return;
 370
 371         entry = trace_event_buffer_reserve(&fbuffer, trace_file, sizeof(*entry));
 372         if (!entry)
 373                 return;
 374
 375         entry = ring_buffer_event_data(fbuffer.event);
 376         entry->nr = syscall_nr;
 377         entry->ret = syscall_get_return_value(current, regs);
 378
 379         trace_event_buffer_commit(&fbuffer);
 380 }
 381
 382 static int reg_event_syscall_enter(struct trace_event_file *file,
 383                                    struct trace_event_call *call)
 384 {
 385         struct trace_array *tr = file->tr;
 386         int ret = 0;
 387         int num;
 388
 389         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 390         if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
 391                 return -ENOSYS;
 392         mutex_lock(&syscall_trace_lock);
 393         if (!tr->sys_refcount_enter)
 394                 ret = register_trace_sys_enter(ftrace_syscall_enter, tr);
 395         if (!ret) {
 396                 rcu_assign_pointer(tr->enter_syscall_files[num], file);
 397                 tr->sys_refcount_enter++;
 398         }
 399         mutex_unlock(&syscall_trace_lock);
 400         return ret;
 401 }
 402
 403 static void unreg_event_syscall_enter(struct trace_event_file *file,
 404                                       struct trace_event_call *call)
 405 {
 406         struct trace_array *tr = file->tr;
 407         int num;
 408
 409         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 410         if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
 411                 return;
 412         mutex_lock(&syscall_trace_lock);
 413         tr->sys_refcount_enter--;
 414         RCU_INIT_POINTER(tr->enter_syscall_files[num], NULL);
 415         if (!tr->sys_refcount_enter)
 416                 unregister_trace_sys_enter(ftrace_syscall_enter, tr);
 417         mutex_unlock(&syscall_trace_lock);
 418 }
 419
 420 static int reg_event_syscall_exit(struct trace_event_file *file,
 421                                   struct trace_event_call *call)
 422 {
 423         struct trace_array *tr = file->tr;
 424         int ret = 0;
 425         int num;
 426
 427         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 428         if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
 429                 return -ENOSYS;
 430         mutex_lock(&syscall_trace_lock);
 431         if (!tr->sys_refcount_exit)
 432                 ret = register_trace_sys_exit(ftrace_syscall_exit, tr);
 433         if (!ret) {
 434                 rcu_assign_pointer(tr->exit_syscall_files[num], file);
 435                 tr->sys_refcount_exit++;
 436         }
 437         mutex_unlock(&syscall_trace_lock);
 438         return ret;
 439 }
 440
 441 static void unreg_event_syscall_exit(struct trace_event_file *file,
 442                                      struct trace_event_call *call)
 443 {
 444         struct trace_array *tr = file->tr;
 445         int num;
 446
 447         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 448         if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
 449                 return;
 450         mutex_lock(&syscall_trace_lock);
 451         tr->sys_refcount_exit--;
 452         RCU_INIT_POINTER(tr->exit_syscall_files[num], NULL);
 453         if (!tr->sys_refcount_exit)
 454                 unregister_trace_sys_exit(ftrace_syscall_exit, tr);
 455         mutex_unlock(&syscall_trace_lock);
 456 }
 457
 458 static int __init init_syscall_trace(struct trace_event_call *call)
 459 {
 460         int id;
 461         int num;
 462
 463         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 464         if (num < 0 || num >= NR_syscalls) {
 465                 pr_debug("syscall %s metadata not mapped, disabling ftrace event\n",
 466                                 ((struct syscall_metadata *)call->data)->name);
 467                 return -ENOSYS;
 468         }
 469
 470         if (set_syscall_print_fmt(call) < 0)
 471                 return -ENOMEM;
 472
 473         id = trace_event_raw_init(call);
 474
 475         if (id < 0) {
 476                 free_syscall_print_fmt(call);
 477                 return id;
 478         }
 479
 480         return id;
 481 }
 482
 483 static struct trace_event_fields __refdata syscall_enter_fields_array[] = {
 484         SYSCALL_FIELD(int, __syscall_nr),
 485         { .type = TRACE_FUNCTION_TYPE,
 486           .define_fields = syscall_enter_define_fields },
 487         {}
 488 };
 489
 490 struct trace_event_functions enter_syscall_print_funcs = {
 491         .trace          = print_syscall_enter,
 492 };
 493
 494 struct trace_event_functions exit_syscall_print_funcs = {
 495         .trace          = print_syscall_exit,
 496 };
 497
 498 struct trace_event_class __refdata event_class_syscall_enter = {
 499         .system         = "syscalls",
 500         .reg            = syscall_enter_register,
 501         .fields_array   = syscall_enter_fields_array,
 502         .get_fields     = syscall_get_enter_fields,
 503         .raw_init       = init_syscall_trace,
 504 };
 505
 506 struct trace_event_class __refdata event_class_syscall_exit = {
 507         .system         = "syscalls",
 508         .reg            = syscall_exit_register,
 509         .fields_array   = (struct trace_event_fields[]){
 510                 SYSCALL_FIELD(int, __syscall_nr),
 511                 SYSCALL_FIELD(long, ret),
 512                 {}
 513         },
 514         .fields         = LIST_HEAD_INIT(event_class_syscall_exit.fields),
 515         .raw_init       = init_syscall_trace,
 516 };
 517
 518 unsigned long __init __weak arch_syscall_addr(int nr)
 519 {
 520         return (unsigned long)sys_call_table[nr];
 521 }
 522
 523 void __init init_ftrace_syscalls(void)
 524 {
 525         struct syscall_metadata *meta;
 526         unsigned long addr;
 527         int i;
 528         void *ret;
 529
 530         if (!IS_ENABLED(CONFIG_HAVE_SPARSE_SYSCALL_NR)) {
 531                 syscalls_metadata = kcalloc(NR_syscalls,
 532                                         sizeof(*syscalls_metadata),
 533                                         GFP_KERNEL);
 534                 if (!syscalls_metadata) {
 535                         WARN_ON(1);
 536                         return;
 537                 }
 538         }
 539
 540         for (i = 0; i < NR_syscalls; i++) {
 541                 addr = arch_syscall_addr(i);
 542                 meta = find_syscall_meta(addr);
 543                 if (!meta)
 544                         continue;
 545
 546                 meta->syscall_nr = i;
 547
 548                 if (!IS_ENABLED(CONFIG_HAVE_SPARSE_SYSCALL_NR)) {
 549                         syscalls_metadata[i] = meta;
 550                 } else {
 551                         ret = xa_store(&syscalls_metadata_sparse, i, meta,
 552                                         GFP_KERNEL);
 553                         WARN(xa_is_err(ret),
 554                                 "Syscall memory allocation failed\n");
 555                 }
 556
 557         }
 558 }
 559
 560 #ifdef CONFIG_PERF_EVENTS
 561
 562 static DECLARE_BITMAP(enabled_perf_enter_syscalls, NR_syscalls);
 563 static DECLARE_BITMAP(enabled_perf_exit_syscalls, NR_syscalls);
 564 static int sys_perf_refcount_enter;
 565 static int sys_perf_refcount_exit;
 566
 567 static int perf_call_bpf_enter(struct trace_event_call *call, struct pt_regs *regs,
 568                                struct syscall_metadata *sys_data,
 569                                struct syscall_trace_enter *rec)
 570 {
 571         struct syscall_tp_t {
 572                 struct trace_entry ent;
 573                 int syscall_nr;
 574                 unsigned long args[SYSCALL_DEFINE_MAXARGS];
 575         } __aligned(8) param;
 576         int i;
 577
 578         BUILD_BUG_ON(sizeof(param.ent) < sizeof(void *));
 579
 580         /* bpf prog requires 'regs' to be the first member in the ctx (a.k.a. &param) */
 581         perf_fetch_caller_regs(regs);
 582         *(struct pt_regs **)&param = regs;
 583         param.syscall_nr = rec->nr;
 584         for (i = 0; i < sys_data->nb_args; i++)
 585                 param.args[i] = rec->args[i];
 586         return trace_call_bpf(call, &param);
 587 }
 588
 589 static void perf_syscall_enter(void *ignore, struct pt_regs *regs, long id)
 590 {
 591         struct syscall_metadata *sys_data;
 592         struct syscall_trace_enter *rec;
 593         struct pt_regs *fake_regs;
 594         struct hlist_head *head;
 595         unsigned long args[6];
 596         bool valid_prog_array;
 597         int syscall_nr;
 598         int rctx;
 599         int size;
 600
 601         /*
 602          * Syscall probe called with preemption enabled, but the ring
 603          * buffer and per-cpu data require preemption to be disabled.
 604          */
 605         might_fault();
 606         guard(preempt_notrace)();
 607
 608         syscall_nr = trace_get_syscall_nr(current, regs);
 609         if (syscall_nr < 0 || syscall_nr >= NR_syscalls)
 610                 return;
 611         if (!test_bit(syscall_nr, enabled_perf_enter_syscalls))
 612                 return;
 613
 614         sys_data = syscall_nr_to_meta(syscall_nr);
 615         if (!sys_data)
 616                 return;
 617
 618         head = this_cpu_ptr(sys_data->enter_event->perf_events);
 619         valid_prog_array = bpf_prog_array_valid(sys_data->enter_event);
 620         if (!valid_prog_array && hlist_empty(head))
 621                 return;
 622
 623         /* get the size after alignment with the u32 buffer size field */
 624         size = sizeof(unsigned long) * sys_data->nb_args + sizeof(*rec);
 625         size = ALIGN(size + sizeof(u32), sizeof(u64));
 626         size -= sizeof(u32);
 627
 628         rec = perf_trace_buf_alloc(size, &fake_regs, &rctx);
 629         if (!rec)
 630                 return;
 631
 632         rec->nr = syscall_nr;
 633         syscall_get_arguments(current, regs, args);
 634         memcpy(&rec->args, args, sizeof(unsigned long) * sys_data->nb_args);
 635
 636         if ((valid_prog_array &&
 637              !perf_call_bpf_enter(sys_data->enter_event, fake_regs, sys_data, rec)) ||
 638             hlist_empty(head)) {
 639                 perf_swevent_put_recursion_context(rctx);
 640                 return;
 641         }
 642
 643         perf_trace_buf_submit(rec, size, rctx,
 644                               sys_data->enter_event->event.type, 1, regs,
 645                               head, NULL);
 646 }
 647
 648 static int perf_sysenter_enable(struct trace_event_call *call)
 649 {
 650         int ret = 0;
 651         int num;
 652
 653         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 654
 655         mutex_lock(&syscall_trace_lock);
 656         if (!sys_perf_refcount_enter)
 657                 ret = register_trace_sys_enter(perf_syscall_enter, NULL);
 658         if (ret) {
 659                 pr_info("event trace: Could not activate syscall entry trace point");
 660         } else {
 661                 set_bit(num, enabled_perf_enter_syscalls);
 662                 sys_perf_refcount_enter++;
 663         }
 664         mutex_unlock(&syscall_trace_lock);
 665         return ret;
 666 }
 667
 668 static void perf_sysenter_disable(struct trace_event_call *call)
 669 {
 670         int num;
 671
 672         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 673
 674         mutex_lock(&syscall_trace_lock);
 675         sys_perf_refcount_enter--;
 676         clear_bit(num, enabled_perf_enter_syscalls);
 677         if (!sys_perf_refcount_enter)
 678                 unregister_trace_sys_enter(perf_syscall_enter, NULL);
 679         mutex_unlock(&syscall_trace_lock);
 680 }
 681
 682 static int perf_call_bpf_exit(struct trace_event_call *call, struct pt_regs *regs,
 683                               struct syscall_trace_exit *rec)
 684 {
 685         struct syscall_tp_t {
 686                 struct trace_entry ent;
 687                 int syscall_nr;
 688                 unsigned long ret;
 689         } __aligned(8) param;
 690
 691         /* bpf prog requires 'regs' to be the first member in the ctx (a.k.a. &param) */
 692         perf_fetch_caller_regs(regs);
 693         *(struct pt_regs **)&param = regs;
 694         param.syscall_nr = rec->nr;
 695         param.ret = rec->ret;
 696         return trace_call_bpf(call, &param);
 697 }
 698
 699 static void perf_syscall_exit(void *ignore, struct pt_regs *regs, long ret)
 700 {
 701         struct syscall_metadata *sys_data;
 702         struct syscall_trace_exit *rec;
 703         struct pt_regs *fake_regs;
 704         struct hlist_head *head;
 705         bool valid_prog_array;
 706         int syscall_nr;
 707         int rctx;
 708         int size;
 709
 710         /*
 711          * Syscall probe called with preemption enabled, but the ring
 712          * buffer and per-cpu data require preemption to be disabled.
 713          */
 714         might_fault();
 715         guard(preempt_notrace)();
 716
 717         syscall_nr = trace_get_syscall_nr(current, regs);
 718         if (syscall_nr < 0 || syscall_nr >= NR_syscalls)
 719                 return;
 720         if (!test_bit(syscall_nr, enabled_perf_exit_syscalls))
 721                 return;
 722
 723         sys_data = syscall_nr_to_meta(syscall_nr);
 724         if (!sys_data)
 725                 return;
 726
 727         head = this_cpu_ptr(sys_data->exit_event->perf_events);
 728         valid_prog_array = bpf_prog_array_valid(sys_data->exit_event);
 729         if (!valid_prog_array && hlist_empty(head))
 730                 return;
 731
 732         /* We can probably do that at build time */
 733         size = ALIGN(sizeof(*rec) + sizeof(u32), sizeof(u64));
 734         size -= sizeof(u32);
 735
 736         rec = perf_trace_buf_alloc(size, &fake_regs, &rctx);
 737         if (!rec)
 738                 return;
 739
 740         rec->nr = syscall_nr;
 741         rec->ret = syscall_get_return_value(current, regs);
 742
 743         if ((valid_prog_array &&
 744              !perf_call_bpf_exit(sys_data->exit_event, fake_regs, rec)) ||
 745             hlist_empty(head)) {
 746                 perf_swevent_put_recursion_context(rctx);
 747                 return;
 748         }
 749
 750         perf_trace_buf_submit(rec, size, rctx, sys_data->exit_event->event.type,
 751                               1, regs, head, NULL);
 752 }
 753
 754 static int perf_sysexit_enable(struct trace_event_call *call)
 755 {
 756         int ret = 0;
 757         int num;
 758
 759         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 760
 761         mutex_lock(&syscall_trace_lock);
 762         if (!sys_perf_refcount_exit)
 763                 ret = register_trace_sys_exit(perf_syscall_exit, NULL);
 764         if (ret) {
 765                 pr_info("event trace: Could not activate syscall exit trace point");
 766         } else {
 767                 set_bit(num, enabled_perf_exit_syscalls);
 768                 sys_perf_refcount_exit++;
 769         }
 770         mutex_unlock(&syscall_trace_lock);
 771         return ret;
 772 }
 773
 774 static void perf_sysexit_disable(struct trace_event_call *call)
 775 {
 776         int num;
 777
 778         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 779
 780         mutex_lock(&syscall_trace_lock);
 781         sys_perf_refcount_exit--;
 782         clear_bit(num, enabled_perf_exit_syscalls);
 783         if (!sys_perf_refcount_exit)
 784                 unregister_trace_sys_exit(perf_syscall_exit, NULL);
 785         mutex_unlock(&syscall_trace_lock);
 786 }
 787
 788 #endif /* CONFIG_PERF_EVENTS */
 789
 790 static int syscall_enter_register(struct trace_event_call *event,
 791                                  enum trace_reg type, void *data)
 792 {
 793         struct trace_event_file *file = data;
 794
 795         switch (type) {
 796         case TRACE_REG_REGISTER:
 797                 return reg_event_syscall_enter(file, event);
 798         case TRACE_REG_UNREGISTER:
 799                 unreg_event_syscall_enter(file, event);
 800                 return 0;
 801
 802 #ifdef CONFIG_PERF_EVENTS
 803         case TRACE_REG_PERF_REGISTER:
 804                 return perf_sysenter_enable(event);
 805         case TRACE_REG_PERF_UNREGISTER:
 806                 perf_sysenter_disable(event);
 807                 return 0;
 808         case TRACE_REG_PERF_OPEN:
 809         case TRACE_REG_PERF_CLOSE:
 810         case TRACE_REG_PERF_ADD:
 811         case TRACE_REG_PERF_DEL:
 812                 return 0;
 813 #endif
 814         }
 815         return 0;
 816 }
 817
 818 static int syscall_exit_register(struct trace_event_call *event,
 819                                  enum trace_reg type, void *data)
 820 {
 821         struct trace_event_file *file = data;
 822
 823         switch (type) {
 824         case TRACE_REG_REGISTER:
 825                 return reg_event_syscall_exit(file, event);
 826         case TRACE_REG_UNREGISTER:
 827                 unreg_event_syscall_exit(file, event);
 828                 return 0;
 829
 830 #ifdef CONFIG_PERF_EVENTS
 831         case TRACE_REG_PERF_REGISTER:
 832                 return perf_sysexit_enable(event);
 833         case TRACE_REG_PERF_UNREGISTER:
 834                 perf_sysexit_disable(event);
 835                 return 0;
 836         case TRACE_REG_PERF_OPEN:
 837         case TRACE_REG_PERF_CLOSE:
 838         case TRACE_REG_PERF_ADD:
 839         case TRACE_REG_PERF_DEL:
 840                 return 0;
 841 #endif
 842         }
 843         return 0;
 844 }