arch/x86/entry/common.c

   1 /*
   2  * common.c - C code for kernel entry and exit
   3  * Copyright (c) 2015 Andrew Lutomirski
   4  * GPL v2
   5  *
   6  * Based on asm and ptrace code by many authors.  The code here originated
   7  * in ptrace.c and signal.c.
   8  */
   9
  10 #include <linux/kernel.h>
  11 #include <linux/sched.h>
  12 #include <linux/sched/task_stack.h>
  13 #include <linux/mm.h>
  14 #include <linux/smp.h>
  15 #include <linux/errno.h>
  16 #include <linux/ptrace.h>
  17 #include <linux/tracehook.h>
  18 #include <linux/audit.h>
  19 #include <linux/seccomp.h>
  20 #include <linux/signal.h>
  21 #include <linux/export.h>
  22 #include <linux/context_tracking.h>
  23 #include <linux/user-return-notifier.h>
  24 #include <linux/nospec.h>
  25 #include <linux/uprobes.h>
  26 #include <linux/livepatch.h>
  27 #include <linux/syscalls.h>
  28
  29 #include <asm/desc.h>
  30 #include <asm/traps.h>
  31 #include <asm/vdso.h>
  32 #include <linux/uaccess.h>
  33 #include <asm/cpufeature.h>
  34 #include <asm/nospec-branch.h>
  35
  36 #define CREATE_TRACE_POINTS
  37 #include <trace/events/syscalls.h>
  38
  39 #ifdef CONFIG_CONTEXT_TRACKING
  40 /* Called on entry from user mode with IRQs off. */
  41 __visible inline void enter_from_user_mode(void)
  42 {
  43         CT_WARN_ON(ct_state() != CONTEXT_USER);
  44         user_exit_irqoff();
  45 }
  46 #else
  47 static inline void enter_from_user_mode(void) {}
  48 #endif
  49
  50 static void do_audit_syscall_entry(struct pt_regs *regs, u32 arch)
  51 {
  52 #ifdef CONFIG_X86_64
  53         if (arch == AUDIT_ARCH_X86_64) {
  54                 audit_syscall_entry(regs->orig_ax, regs->di,
  55                                     regs->si, regs->dx, regs->r10);
  56         } else
  57 #endif
  58         {
  59                 audit_syscall_entry(regs->orig_ax, regs->bx,
  60                                     regs->cx, regs->dx, regs->si);
  61         }
  62 }
  63
  64 /*
  65  * Returns the syscall nr to run (which should match regs->orig_ax) or -1
  66  * to skip the syscall.
  67  */
  68 static long syscall_trace_enter(struct pt_regs *regs)
  69 {
  70         u32 arch = in_ia32_syscall() ? AUDIT_ARCH_I386 : AUDIT_ARCH_X86_64;
  71
  72         struct thread_info *ti = current_thread_info();
  73         unsigned long ret = 0;
  74         bool emulated = false;
  75         u32 work;
  76
  77         if (IS_ENABLED(CONFIG_DEBUG_ENTRY))
  78                 BUG_ON(regs != task_pt_regs(current));
  79
  80         work = READ_ONCE(ti->flags) & _TIF_WORK_SYSCALL_ENTRY;
  81
  82         if (unlikely(work & _TIF_SYSCALL_EMU))
  83                 emulated = true;
  84
  85         if ((emulated || (work & _TIF_SYSCALL_TRACE)) &&
  86             tracehook_report_syscall_entry(regs))
  87                 return -1L;
  88
  89         if (emulated)
  90                 return -1L;
  91
  92 #ifdef CONFIG_SECCOMP
  93         /*
  94          * Do seccomp after ptrace, to catch any tracer changes.
  95          */
  96         if (work & _TIF_SECCOMP) {
  97                 struct seccomp_data sd;
  98
  99                 sd.arch = arch;
 100                 sd.nr = regs->orig_ax;
 101                 sd.instruction_pointer = regs->ip;
 102 #ifdef CONFIG_X86_64
 103                 if (arch == AUDIT_ARCH_X86_64) {
 104                         sd.args[0] = regs->di;
 105                         sd.args[1] = regs->si;
 106                         sd.args[2] = regs->dx;
 107                         sd.args[3] = regs->r10;
 108                         sd.args[4] = regs->r8;
 109                         sd.args[5] = regs->r9;
 110                 } else
 111 #endif
 112                 {
 113                         sd.args[0] = regs->bx;
 114                         sd.args[1] = regs->cx;
 115                         sd.args[2] = regs->dx;
 116                         sd.args[3] = regs->si;
 117                         sd.args[4] = regs->di;
 118                         sd.args[5] = regs->bp;
 119                 }
 120
 121                 ret = __secure_computing(&sd);
 122                 if (ret == -1)
 123                         return ret;
 124         }
 125 #endif
 126
 127         if (unlikely(test_thread_flag(TIF_SYSCALL_TRACEPOINT)))
 128                 trace_sys_enter(regs, regs->orig_ax);
 129
 130         do_audit_syscall_entry(regs, arch);
 131
 132         return ret ?: regs->orig_ax;
 133 }
 134
 135 #define EXIT_TO_USERMODE_LOOP_FLAGS                             \
 136         (_TIF_SIGPENDING | _TIF_NOTIFY_RESUME | _TIF_UPROBE |   \
 137          _TIF_NEED_RESCHED | _TIF_USER_RETURN_NOTIFY | _TIF_PATCH_PENDING)
 138
 139 static void exit_to_usermode_loop(struct pt_regs *regs, u32 cached_flags)
 140 {
 141         /*
 142          * In order to return to user mode, we need to have IRQs off with
 143          * none of EXIT_TO_USERMODE_LOOP_FLAGS set.  Several of these flags
 144          * can be set at any time on preemptable kernels if we have IRQs on,
 145          * so we need to loop.  Disabling preemption wouldn't help: doing the
 146          * work to clear some of the flags can sleep.
 147          */
 148         while (true) {
 149                 /* We have work to do. */
 150                 local_irq_enable();
 151
 152                 if (cached_flags & _TIF_NEED_RESCHED)
 153                         schedule();
 154
 155                 if (cached_flags & _TIF_UPROBE)
 156                         uprobe_notify_resume(regs);
 157
 158                 if (cached_flags & _TIF_PATCH_PENDING)
 159                         klp_update_patch_state(current);
 160
 161                 /* deal with pending signal delivery */
 162                 if (cached_flags & _TIF_SIGPENDING)
 163                         do_signal(regs);
 164
 165                 if (cached_flags & _TIF_NOTIFY_RESUME) {
 166                         clear_thread_flag(TIF_NOTIFY_RESUME);
 167                         tracehook_notify_resume(regs);
 168                         rseq_handle_notify_resume(NULL, regs);
 169                 }
 170
 171                 if (cached_flags & _TIF_USER_RETURN_NOTIFY)
 172                         fire_user_return_notifiers();
 173
 174                 /* Disable IRQs and retry */
 175                 local_irq_disable();
 176
 177                 cached_flags = READ_ONCE(current_thread_info()->flags);
 178
 179                 if (!(cached_flags & EXIT_TO_USERMODE_LOOP_FLAGS))
 180                         break;
 181         }
 182 }
 183
 184 /* Called with IRQs disabled. */
 185 __visible inline void prepare_exit_to_usermode(struct pt_regs *regs)
 186 {
 187         struct thread_info *ti = current_thread_info();
 188         u32 cached_flags;
 189
 190         addr_limit_user_check();
 191
 192         lockdep_assert_irqs_disabled();
 193         lockdep_sys_exit();
 194
 195         cached_flags = READ_ONCE(ti->flags);
 196
 197         if (unlikely(cached_flags & EXIT_TO_USERMODE_LOOP_FLAGS))
 198                 exit_to_usermode_loop(regs, cached_flags);
 199
 200 #ifdef CONFIG_COMPAT
 201         /*
 202          * Compat syscalls set TS_COMPAT.  Make sure we clear it before
 203          * returning to user mode.  We need to clear it *after* signal
 204          * handling, because syscall restart has a fixup for compat
 205          * syscalls.  The fixup is exercised by the ptrace_syscall_32
 206          * selftest.
 207          *
 208          * We also need to clear TS_REGS_POKED_I386: the 32-bit tracer
 209          * special case only applies after poking regs and before the
 210          * very next return to user mode.
 211          */
 212         ti->status &= ~(TS_COMPAT|TS_I386_REGS_POKED);
 213 #endif
 214
 215         user_enter_irqoff();
 216
 217         mds_user_clear_cpu_buffers();
 218 }
 219
 220 #define SYSCALL_EXIT_WORK_FLAGS                         \
 221         (_TIF_SYSCALL_TRACE | _TIF_SYSCALL_AUDIT |      \
 222          _TIF_SINGLESTEP | _TIF_SYSCALL_TRACEPOINT)
 223
 224 static void syscall_slow_exit_work(struct pt_regs *regs, u32 cached_flags)
 225 {
 226         bool step;
 227
 228         audit_syscall_exit(regs);
 229
 230         if (cached_flags & _TIF_SYSCALL_TRACEPOINT)
 231                 trace_sys_exit(regs, regs->ax);
 232
 233         /*
 234          * If TIF_SYSCALL_EMU is set, we only get here because of
 235          * TIF_SINGLESTEP (i.e. this is PTRACE_SYSEMU_SINGLESTEP).
 236          * We already reported this syscall instruction in
 237          * syscall_trace_enter().
 238          */
 239         step = unlikely(
 240                 (cached_flags & (_TIF_SINGLESTEP | _TIF_SYSCALL_EMU))
 241                 == _TIF_SINGLESTEP);
 242         if (step || cached_flags & _TIF_SYSCALL_TRACE)
 243                 tracehook_report_syscall_exit(regs, step);
 244 }
 245
 246 /*
 247  * Called with IRQs on and fully valid regs.  Returns with IRQs off in a
 248  * state such that we can immediately switch to user mode.
 249  */
 250 __visible inline void syscall_return_slowpath(struct pt_regs *regs)
 251 {
 252         struct thread_info *ti = current_thread_info();
 253         u32 cached_flags = READ_ONCE(ti->flags);
 254
 255         CT_WARN_ON(ct_state() != CONTEXT_KERNEL);
 256
 257         if (IS_ENABLED(CONFIG_PROVE_LOCKING) &&
 258             WARN(irqs_disabled(), "syscall %ld left IRQs disabled", regs->orig_ax))
 259                 local_irq_enable();
 260
 261         rseq_syscall(regs);
 262
 263         /*
 264          * First do one-time work.  If these work items are enabled, we
 265          * want to run them exactly once per syscall exit with IRQs on.
 266          */
 267         if (unlikely(cached_flags & SYSCALL_EXIT_WORK_FLAGS))
 268                 syscall_slow_exit_work(regs, cached_flags);
 269
 270         local_irq_disable();
 271         prepare_exit_to_usermode(regs);
 272 }
 273
 274 #ifdef CONFIG_X86_64
 275 __visible void do_syscall_64(unsigned long nr, struct pt_regs *regs)
 276 {
 277         struct thread_info *ti;
 278
 279         enter_from_user_mode();
 280         local_irq_enable();
 281         ti = current_thread_info();
 282         if (READ_ONCE(ti->flags) & _TIF_WORK_SYSCALL_ENTRY)
 283                 nr = syscall_trace_enter(regs);
 284
 285         /*
 286          * NB: Native and x32 syscalls are dispatched from the same
 287          * table.  The only functional difference is the x32 bit in
 288          * regs->orig_ax, which changes the behavior of some syscalls.
 289          */
 290         nr &= __SYSCALL_MASK;
 291         if (likely(nr < NR_syscalls)) {
 292                 nr = array_index_nospec(nr, NR_syscalls);
 293                 regs->ax = sys_call_table[nr](regs);
 294         }
 295
 296         syscall_return_slowpath(regs);
 297 }
 298 #endif
 299
 300 #if defined(CONFIG_X86_32) || defined(CONFIG_IA32_EMULATION)
 301 /*
 302  * Does a 32-bit syscall.  Called with IRQs on in CONTEXT_KERNEL.  Does
 303  * all entry and exit work and returns with IRQs off.  This function is
 304  * extremely hot in workloads that use it, and it's usually called from
 305  * do_fast_syscall_32, so forcibly inline it to improve performance.
 306  */
 307 static __always_inline void do_syscall_32_irqs_on(struct pt_regs *regs)
 308 {
 309         struct thread_info *ti = current_thread_info();
 310         unsigned int nr = (unsigned int)regs->orig_ax;
 311
 312 #ifdef CONFIG_IA32_EMULATION
 313         ti->status |= TS_COMPAT;
 314 #endif
 315
 316         if (READ_ONCE(ti->flags) & _TIF_WORK_SYSCALL_ENTRY) {
 317                 /*
 318                  * Subtlety here: if ptrace pokes something larger than
 319                  * 2^32-1 into orig_ax, this truncates it.  This may or
 320                  * may not be necessary, but it matches the old asm
 321                  * behavior.
 322                  */
 323                 nr = syscall_trace_enter(regs);
 324         }
 325
 326         if (likely(nr < IA32_NR_syscalls)) {
 327                 nr = array_index_nospec(nr, IA32_NR_syscalls);
 328 #ifdef CONFIG_IA32_EMULATION
 329                 regs->ax = ia32_sys_call_table[nr](regs);
 330 #else
 331                 /*
 332                  * It's possible that a 32-bit syscall implementation
 333                  * takes a 64-bit parameter but nonetheless assumes that
 334                  * the high bits are zero.  Make sure we zero-extend all
 335                  * of the args.
 336                  */
 337                 regs->ax = ia32_sys_call_table[nr](
 338                         (unsigned int)regs->bx, (unsigned int)regs->cx,
 339                         (unsigned int)regs->dx, (unsigned int)regs->si,
 340                         (unsigned int)regs->di, (unsigned int)regs->bp);
 341 #endif /* CONFIG_IA32_EMULATION */
 342         }
 343
 344         syscall_return_slowpath(regs);
 345 }
 346
 347 /* Handles int $0x80 */
 348 __visible void do_int80_syscall_32(struct pt_regs *regs)
 349 {
 350         enter_from_user_mode();
 351         local_irq_enable();
 352         do_syscall_32_irqs_on(regs);
 353 }
 354
 355 /* Returns 0 to return using IRET or 1 to return using SYSEXIT/SYSRETL. */
 356 __visible long do_fast_syscall_32(struct pt_regs *regs)
 357 {
 358         /*
 359          * Called using the internal vDSO SYSENTER/SYSCALL32 calling
 360          * convention.  Adjust regs so it looks like we entered using int80.
 361          */
 362
 363         unsigned long landing_pad = (unsigned long)current->mm->context.vdso +
 364                 vdso_image_32.sym_int80_landing_pad;
 365
 366         /*
 367          * SYSENTER loses EIP, and even SYSCALL32 needs us to skip forward
 368          * so that 'regs->ip -= 2' lands back on an int $0x80 instruction.
 369          * Fix it up.
 370          */
 371         regs->ip = landing_pad;
 372
 373         enter_from_user_mode();
 374
 375         local_irq_enable();
 376
 377         /* Fetch EBP from where the vDSO stashed it. */
 378         if (
 379 #ifdef CONFIG_X86_64
 380                 /*
 381                  * Micro-optimization: the pointer we're following is explicitly
 382                  * 32 bits, so it can't be out of range.
 383                  */
 384                 __get_user(*(u32 *)&regs->bp,
 385                             (u32 __user __force *)(unsigned long)(u32)regs->sp)
 386 #else
 387                 get_user(*(u32 *)&regs->bp,
 388                          (u32 __user __force *)(unsigned long)(u32)regs->sp)
 389 #endif
 390                 ) {
 391
 392                 /* User code screwed up. */
 393                 local_irq_disable();
 394                 regs->ax = -EFAULT;
 395                 prepare_exit_to_usermode(regs);
 396                 return 0;       /* Keep it simple: use IRET. */
 397         }
 398
 399         /* Now this is just like a normal syscall. */
 400         do_syscall_32_irqs_on(regs);
 401
 402 #ifdef CONFIG_X86_64
 403         /*
 404          * Opportunistic SYSRETL: if possible, try to return using SYSRETL.
 405          * SYSRETL is available on all 64-bit CPUs, so we don't need to
 406          * bother with SYSEXIT.
 407          *
 408          * Unlike 64-bit opportunistic SYSRET, we can't check that CX == IP,
 409          * because the ECX fixup above will ensure that this is essentially
 410          * never the case.
 411          */
 412         return regs->cs == __USER32_CS && regs->ss == __USER_DS &&
 413                 regs->ip == landing_pad &&
 414                 (regs->flags & (X86_EFLAGS_RF | X86_EFLAGS_TF)) == 0;
 415 #else
 416         /*
 417          * Opportunistic SYSEXIT: if possible, try to return using SYSEXIT.
 418          *
 419          * Unlike 64-bit opportunistic SYSRET, we can't check that CX == IP,
 420          * because the ECX fixup above will ensure that this is essentially
 421          * never the case.
 422          *
 423          * We don't allow syscalls at all from VM86 mode, but we still
 424          * need to check VM, because we might be returning from sys_vm86.
 425          */
 426         return static_cpu_has(X86_FEATURE_SEP) &&
 427                 regs->cs == __USER_CS && regs->ss == __USER_DS &&
 428                 regs->ip == landing_pad &&
 429                 (regs->flags & (X86_EFLAGS_RF | X86_EFLAGS_TF | X86_EFLAGS_VM)) == 0;
 430 #endif
 431 }
 432 #endif