arch/x86/entry/common.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * common.c - C code for kernel entry and exit
   4  * Copyright (c) 2015 Andrew Lutomirski
   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 #include <linux/uaccess.h>
  29
  30 #include <asm/desc.h>
  31 #include <asm/traps.h>
  32 #include <asm/vdso.h>
  33 #include <asm/cpufeature.h>
  34 #include <asm/fpu/api.h>
  35 #include <asm/nospec-branch.h>
  36
  37 #define CREATE_TRACE_POINTS
  38 #include <trace/events/syscalls.h>
  39
  40 #ifdef CONFIG_CONTEXT_TRACKING
  41 /* Called on entry from user mode with IRQs off. */
  42 __visible inline void enter_from_user_mode(void)
  43 {
  44         CT_WARN_ON(ct_state() != CONTEXT_USER);
  45         user_exit_irqoff();
  46 }
  47 #else
  48 static inline void enter_from_user_mode(void) {}
  49 #endif
  50
  51 static void do_audit_syscall_entry(struct pt_regs *regs, u32 arch)
  52 {
  53 #ifdef CONFIG_X86_64
  54         if (arch == AUDIT_ARCH_X86_64) {
  55                 audit_syscall_entry(regs->orig_ax, regs->di,
  56                                     regs->si, regs->dx, regs->r10);
  57         } else
  58 #endif
  59         {
  60                 audit_syscall_entry(regs->orig_ax, regs->bx,
  61                                     regs->cx, regs->dx, regs->si);
  62         }
  63 }
  64
  65 /*
  66  * Returns the syscall nr to run (which should match regs->orig_ax) or -1
  67  * to skip the syscall.
  68  */
  69 static long syscall_trace_enter(struct pt_regs *regs)
  70 {
  71         u32 arch = in_ia32_syscall() ? AUDIT_ARCH_I386 : AUDIT_ARCH_X86_64;
  72
  73         struct thread_info *ti = current_thread_info();
  74         unsigned long ret = 0;
  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);
  81
  82         if (work & (_TIF_SYSCALL_TRACE | _TIF_SYSCALL_EMU)) {
  83                 ret = tracehook_report_syscall_entry(regs);
  84                 if (ret || (work & _TIF_SYSCALL_EMU))
  85                         return -1L;
  86         }
  87
  88 #ifdef CONFIG_SECCOMP
  89         /*
  90          * Do seccomp after ptrace, to catch any tracer changes.
  91          */
  92         if (work & _TIF_SECCOMP) {
  93                 struct seccomp_data sd;
  94
  95                 sd.arch = arch;
  96                 sd.nr = regs->orig_ax;
  97                 sd.instruction_pointer = regs->ip;
  98 #ifdef CONFIG_X86_64
  99                 if (arch == AUDIT_ARCH_X86_64) {
 100                         sd.args[0] = regs->di;
 101                         sd.args[1] = regs->si;
 102                         sd.args[2] = regs->dx;
 103                         sd.args[3] = regs->r10;
 104                         sd.args[4] = regs->r8;
 105                         sd.args[5] = regs->r9;
 106                 } else
 107 #endif
 108                 {
 109                         sd.args[0] = regs->bx;
 110                         sd.args[1] = regs->cx;
 111                         sd.args[2] = regs->dx;
 112                         sd.args[3] = regs->si;
 113                         sd.args[4] = regs->di;
 114                         sd.args[5] = regs->bp;
 115                 }
 116
 117                 ret = __secure_computing(&sd);
 118                 if (ret == -1)
 119                         return ret;
 120         }
 121 #endif
 122
 123         if (unlikely(test_thread_flag(TIF_SYSCALL_TRACEPOINT)))
 124                 trace_sys_enter(regs, regs->orig_ax);
 125
 126         do_audit_syscall_entry(regs, arch);
 127
 128         return ret ?: regs->orig_ax;
 129 }
 130
 131 #define EXIT_TO_USERMODE_LOOP_FLAGS                             \
 132         (_TIF_SIGPENDING | _TIF_NOTIFY_RESUME | _TIF_UPROBE |   \
 133          _TIF_NEED_RESCHED | _TIF_USER_RETURN_NOTIFY | _TIF_PATCH_PENDING)
 134
 135 static void exit_to_usermode_loop(struct pt_regs *regs, u32 cached_flags)
 136 {
 137         /*
 138          * In order to return to user mode, we need to have IRQs off with
 139          * none of EXIT_TO_USERMODE_LOOP_FLAGS set.  Several of these flags
 140          * can be set at any time on preemptible kernels if we have IRQs on,
 141          * so we need to loop.  Disabling preemption wouldn't help: doing the
 142          * work to clear some of the flags can sleep.
 143          */
 144         while (true) {
 145                 /* We have work to do. */
 146                 local_irq_enable();
 147
 148                 if (cached_flags & _TIF_NEED_RESCHED)
 149                         schedule();
 150
 151                 if (cached_flags & _TIF_UPROBE)
 152                         uprobe_notify_resume(regs);
 153
 154                 if (cached_flags & _TIF_PATCH_PENDING)
 155                         klp_update_patch_state(current);
 156
 157                 /* deal with pending signal delivery */
 158                 if (cached_flags & _TIF_SIGPENDING)
 159                         do_signal(regs);
 160
 161                 if (cached_flags & _TIF_NOTIFY_RESUME) {
 162                         clear_thread_flag(TIF_NOTIFY_RESUME);
 163                         tracehook_notify_resume(regs);
 164                         rseq_handle_notify_resume(NULL, regs);
 165                 }
 166
 167                 if (cached_flags & _TIF_USER_RETURN_NOTIFY)
 168                         fire_user_return_notifiers();
 169
 170                 /* Disable IRQs and retry */
 171                 local_irq_disable();
 172
 173                 cached_flags = READ_ONCE(current_thread_info()->flags);
 174
 175                 if (!(cached_flags & EXIT_TO_USERMODE_LOOP_FLAGS))
 176                         break;
 177         }
 178 }
 179
 180 /* Called with IRQs disabled. */
 181 __visible inline void prepare_exit_to_usermode(struct pt_regs *regs)
 182 {
 183         struct thread_info *ti = current_thread_info();
 184         u32 cached_flags;
 185
 186         addr_limit_user_check();
 187
 188         lockdep_assert_irqs_disabled();
 189         lockdep_sys_exit();
 190
 191         cached_flags = READ_ONCE(ti->flags);
 192
 193         if (unlikely(cached_flags & EXIT_TO_USERMODE_LOOP_FLAGS))
 194                 exit_to_usermode_loop(regs, cached_flags);
 195
 196         /* Reload ti->flags; we may have rescheduled above. */
 197         cached_flags = READ_ONCE(ti->flags);
 198
 199         fpregs_assert_state_consistent();
 200         if (unlikely(cached_flags & _TIF_NEED_FPU_LOAD))
 201                 switch_fpu_return();
 202
 203 #ifdef CONFIG_COMPAT
 204         /*
 205          * Compat syscalls set TS_COMPAT.  Make sure we clear it before
 206          * returning to user mode.  We need to clear it *after* signal
 207          * handling, because syscall restart has a fixup for compat
 208          * syscalls.  The fixup is exercised by the ptrace_syscall_32
 209          * selftest.
 210          *
 211          * We also need to clear TS_REGS_POKED_I386: the 32-bit tracer
 212          * special case only applies after poking regs and before the
 213          * very next return to user mode.
 214          */
 215         ti->status &= ~(TS_COMPAT|TS_I386_REGS_POKED);
 216 #endif
 217
 218         user_enter_irqoff();
 219
 220         mds_user_clear_cpu_buffers();
 221 }
 222
 223 #define SYSCALL_EXIT_WORK_FLAGS                         \
 224         (_TIF_SYSCALL_TRACE | _TIF_SYSCALL_AUDIT |      \
 225          _TIF_SINGLESTEP | _TIF_SYSCALL_TRACEPOINT)
 226
 227 static void syscall_slow_exit_work(struct pt_regs *regs, u32 cached_flags)
 228 {
 229         bool step;
 230
 231         audit_syscall_exit(regs);
 232
 233         if (cached_flags & _TIF_SYSCALL_TRACEPOINT)
 234                 trace_sys_exit(regs, regs->ax);
 235
 236         /*
 237          * If TIF_SYSCALL_EMU is set, we only get here because of
 238          * TIF_SINGLESTEP (i.e. this is PTRACE_SYSEMU_SINGLESTEP).
 239          * We already reported this syscall instruction in
 240          * syscall_trace_enter().
 241          */
 242         step = unlikely(
 243                 (cached_flags & (_TIF_SINGLESTEP | _TIF_SYSCALL_EMU))
 244                 == _TIF_SINGLESTEP);
 245         if (step || cached_flags & _TIF_SYSCALL_TRACE)
 246                 tracehook_report_syscall_exit(regs, step);
 247 }
 248
 249 /*
 250  * Called with IRQs on and fully valid regs.  Returns with IRQs off in a
 251  * state such that we can immediately switch to user mode.
 252  */
 253 __visible inline void syscall_return_slowpath(struct pt_regs *regs)
 254 {
 255         struct thread_info *ti = current_thread_info();
 256         u32 cached_flags = READ_ONCE(ti->flags);
 257
 258         CT_WARN_ON(ct_state() != CONTEXT_KERNEL);
 259
 260         if (IS_ENABLED(CONFIG_PROVE_LOCKING) &&
 261             WARN(irqs_disabled(), "syscall %ld left IRQs disabled", regs->orig_ax))
 262                 local_irq_enable();
 263
 264         rseq_syscall(regs);
 265
 266         /*
 267          * First do one-time work.  If these work items are enabled, we
 268          * want to run them exactly once per syscall exit with IRQs on.
 269          */
 270         if (unlikely(cached_flags & SYSCALL_EXIT_WORK_FLAGS))
 271                 syscall_slow_exit_work(regs, cached_flags);
 272
 273         local_irq_disable();
 274         prepare_exit_to_usermode(regs);
 275 }
 276
 277 #ifdef CONFIG_X86_64
 278 __visible void do_syscall_64(unsigned long nr, struct pt_regs *regs)
 279 {
 280         struct thread_info *ti;
 281
 282         enter_from_user_mode();
 283         local_irq_enable();
 284         ti = current_thread_info();
 285         if (READ_ONCE(ti->flags) & _TIF_WORK_SYSCALL_ENTRY)
 286                 nr = syscall_trace_enter(regs);
 287
 288         if (likely(nr < NR_syscalls)) {
 289                 nr = array_index_nospec(nr, NR_syscalls);
 290                 regs->ax = sys_call_table[nr](regs);
 291 #ifdef CONFIG_X86_X32_ABI
 292         } else if (likely((nr & __X32_SYSCALL_BIT) &&
 293                           (nr & ~__X32_SYSCALL_BIT) < X32_NR_syscalls)) {
 294                 nr = array_index_nospec(nr & ~__X32_SYSCALL_BIT,
 295                                         X32_NR_syscalls);
 296                 regs->ax = x32_sys_call_table[nr](regs);
 297 #endif
 298         }
 299
 300         syscall_return_slowpath(regs);
 301 }
 302 #endif
 303
 304 #if defined(CONFIG_X86_32) || defined(CONFIG_IA32_EMULATION)
 305 /*
 306  * Does a 32-bit syscall.  Called with IRQs on in CONTEXT_KERNEL.  Does
 307  * all entry and exit work and returns with IRQs off.  This function is
 308  * extremely hot in workloads that use it, and it's usually called from
 309  * do_fast_syscall_32, so forcibly inline it to improve performance.
 310  */
 311 static __always_inline void do_syscall_32_irqs_on(struct pt_regs *regs)
 312 {
 313         struct thread_info *ti = current_thread_info();
 314         unsigned int nr = (unsigned int)regs->orig_ax;
 315
 316 #ifdef CONFIG_IA32_EMULATION
 317         ti->status |= TS_COMPAT;
 318 #endif
 319
 320         if (READ_ONCE(ti->flags) & _TIF_WORK_SYSCALL_ENTRY) {
 321                 /*
 322                  * Subtlety here: if ptrace pokes something larger than
 323                  * 2^32-1 into orig_ax, this truncates it.  This may or
 324                  * may not be necessary, but it matches the old asm
 325                  * behavior.
 326                  */
 327                 nr = syscall_trace_enter(regs);
 328         }
 329
 330         if (likely(nr < IA32_NR_syscalls)) {
 331                 nr = array_index_nospec(nr, IA32_NR_syscalls);
 332 #ifdef CONFIG_IA32_EMULATION
 333                 regs->ax = ia32_sys_call_table[nr](regs);
 334 #else
 335                 /*
 336                  * It's possible that a 32-bit syscall implementation
 337                  * takes a 64-bit parameter but nonetheless assumes that
 338                  * the high bits are zero.  Make sure we zero-extend all
 339                  * of the args.
 340                  */
 341                 regs->ax = ia32_sys_call_table[nr](
 342                         (unsigned int)regs->bx, (unsigned int)regs->cx,
 343                         (unsigned int)regs->dx, (unsigned int)regs->si,
 344                         (unsigned int)regs->di, (unsigned int)regs->bp);
 345 #endif /* CONFIG_IA32_EMULATION */
 346         }
 347
 348         syscall_return_slowpath(regs);
 349 }
 350
 351 /* Handles int $0x80 */
 352 __visible void do_int80_syscall_32(struct pt_regs *regs)
 353 {
 354         enter_from_user_mode();
 355         local_irq_enable();
 356         do_syscall_32_irqs_on(regs);
 357 }
 358
 359 /* Returns 0 to return using IRET or 1 to return using SYSEXIT/SYSRETL. */
 360 __visible long do_fast_syscall_32(struct pt_regs *regs)
 361 {
 362         /*
 363          * Called using the internal vDSO SYSENTER/SYSCALL32 calling
 364          * convention.  Adjust regs so it looks like we entered using int80.
 365          */
 366
 367         unsigned long landing_pad = (unsigned long)current->mm->context.vdso +
 368                 vdso_image_32.sym_int80_landing_pad;
 369
 370         /*
 371          * SYSENTER loses EIP, and even SYSCALL32 needs us to skip forward
 372          * so that 'regs->ip -= 2' lands back on an int $0x80 instruction.
 373          * Fix it up.
 374          */
 375         regs->ip = landing_pad;
 376
 377         enter_from_user_mode();
 378
 379         local_irq_enable();
 380
 381         /* Fetch EBP from where the vDSO stashed it. */
 382         if (
 383 #ifdef CONFIG_X86_64
 384                 /*
 385                  * Micro-optimization: the pointer we're following is explicitly
 386                  * 32 bits, so it can't be out of range.
 387                  */
 388                 __get_user(*(u32 *)&regs->bp,
 389                             (u32 __user __force *)(unsigned long)(u32)regs->sp)
 390 #else
 391                 get_user(*(u32 *)&regs->bp,
 392                          (u32 __user __force *)(unsigned long)(u32)regs->sp)
 393 #endif
 394                 ) {
 395
 396                 /* User code screwed up. */
 397                 local_irq_disable();
 398                 regs->ax = -EFAULT;
 399                 prepare_exit_to_usermode(regs);
 400                 return 0;       /* Keep it simple: use IRET. */
 401         }
 402
 403         /* Now this is just like a normal syscall. */
 404         do_syscall_32_irqs_on(regs);
 405
 406 #ifdef CONFIG_X86_64
 407         /*
 408          * Opportunistic SYSRETL: if possible, try to return using SYSRETL.
 409          * SYSRETL is available on all 64-bit CPUs, so we don't need to
 410          * bother with SYSEXIT.
 411          *
 412          * Unlike 64-bit opportunistic SYSRET, we can't check that CX == IP,
 413          * because the ECX fixup above will ensure that this is essentially
 414          * never the case.
 415          */
 416         return regs->cs == __USER32_CS && regs->ss == __USER_DS &&
 417                 regs->ip == landing_pad &&
 418                 (regs->flags & (X86_EFLAGS_RF | X86_EFLAGS_TF)) == 0;
 419 #else
 420         /*
 421          * Opportunistic SYSEXIT: if possible, try to return using SYSEXIT.
 422          *
 423          * Unlike 64-bit opportunistic SYSRET, we can't check that CX == IP,
 424          * because the ECX fixup above will ensure that this is essentially
 425          * never the case.
 426          *
 427          * We don't allow syscalls at all from VM86 mode, but we still
 428          * need to check VM, because we might be returning from sys_vm86.
 429          */
 430         return static_cpu_has(X86_FEATURE_SEP) &&
 431                 regs->cs == __USER_CS && regs->ss == __USER_DS &&
 432                 regs->ip == landing_pad &&
 433                 (regs->flags & (X86_EFLAGS_RF | X86_EFLAGS_TF | X86_EFLAGS_VM)) == 0;
 434 #endif
 435 }
 436 #endif