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