kernel/entry/common.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include <linux/context_tracking.h>
   4 #include <linux/entry-common.h>
   5 #include <linux/resume_user_mode.h>
   6 #include <linux/highmem.h>
   7 #include <linux/jump_label.h>
   8 #include <linux/livepatch.h>
   9 #include <linux/audit.h>
  10 #include <linux/tick.h>
  11
  12 #include "common.h"
  13
  14 #define CREATE_TRACE_POINTS
  15 #include <trace/events/syscalls.h>
  16
  17 /* See comment for enter_from_user_mode() in entry-common.h */
  18 static __always_inline void __enter_from_user_mode(struct pt_regs *regs)
  19 {
  20         arch_check_user_regs(regs);
  21         lockdep_hardirqs_off(CALLER_ADDR0);
  22
  23         CT_WARN_ON(ct_state() != CONTEXT_USER);
  24         user_exit_irqoff();
  25
  26         instrumentation_begin();
  27         trace_hardirqs_off_finish();
  28         instrumentation_end();
  29 }
  30
  31 void noinstr enter_from_user_mode(struct pt_regs *regs)
  32 {
  33         __enter_from_user_mode(regs);
  34 }
  35
  36 static inline void syscall_enter_audit(struct pt_regs *regs, long syscall)
  37 {
  38         if (unlikely(audit_context())) {
  39                 unsigned long args[6];
  40
  41                 syscall_get_arguments(current, regs, args);
  42                 audit_syscall_entry(syscall, args[0], args[1], args[2], args[3]);
  43         }
  44 }
  45
  46 static long syscall_trace_enter(struct pt_regs *regs, long syscall,
  47                                 unsigned long work)
  48 {
  49         long ret = 0;
  50
  51         /*
  52          * Handle Syscall User Dispatch.  This must comes first, since
  53          * the ABI here can be something that doesn't make sense for
  54          * other syscall_work features.
  55          */
  56         if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
  57                 if (syscall_user_dispatch(regs))
  58                         return -1L;
  59         }
  60
  61         /* Handle ptrace */
  62         if (work & (SYSCALL_WORK_SYSCALL_TRACE | SYSCALL_WORK_SYSCALL_EMU)) {
  63                 ret = ptrace_report_syscall_entry(regs);
  64                 if (ret || (work & SYSCALL_WORK_SYSCALL_EMU))
  65                         return -1L;
  66         }
  67
  68         /* Do seccomp after ptrace, to catch any tracer changes. */
  69         if (work & SYSCALL_WORK_SECCOMP) {
  70                 ret = __secure_computing(NULL);
  71                 if (ret == -1L)
  72                         return ret;
  73         }
  74
  75         /* Either of the above might have changed the syscall number */
  76         syscall = syscall_get_nr(current, regs);
  77
  78         if (unlikely(work & SYSCALL_WORK_SYSCALL_TRACEPOINT))
  79                 trace_sys_enter(regs, syscall);
  80
  81         syscall_enter_audit(regs, syscall);
  82
  83         return ret ? : syscall;
  84 }
  85
  86 static __always_inline long
  87 __syscall_enter_from_user_work(struct pt_regs *regs, long syscall)
  88 {
  89         unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
  90
  91         if (work & SYSCALL_WORK_ENTER)
  92                 syscall = syscall_trace_enter(regs, syscall, work);
  93
  94         return syscall;
  95 }
  96
  97 long syscall_enter_from_user_mode_work(struct pt_regs *regs, long syscall)
  98 {
  99         return __syscall_enter_from_user_work(regs, syscall);
 100 }
 101
 102 noinstr long syscall_enter_from_user_mode(struct pt_regs *regs, long syscall)
 103 {
 104         long ret;
 105
 106         __enter_from_user_mode(regs);
 107
 108         instrumentation_begin();
 109         local_irq_enable();
 110         ret = __syscall_enter_from_user_work(regs, syscall);
 111         instrumentation_end();
 112
 113         return ret;
 114 }
 115
 116 noinstr void syscall_enter_from_user_mode_prepare(struct pt_regs *regs)
 117 {
 118         __enter_from_user_mode(regs);
 119         instrumentation_begin();
 120         local_irq_enable();
 121         instrumentation_end();
 122 }
 123
 124 /* See comment for exit_to_user_mode() in entry-common.h */
 125 static __always_inline void __exit_to_user_mode(void)
 126 {
 127         instrumentation_begin();
 128         trace_hardirqs_on_prepare();
 129         lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 130         instrumentation_end();
 131
 132         user_enter_irqoff();
 133         arch_exit_to_user_mode();
 134         lockdep_hardirqs_on(CALLER_ADDR0);
 135 }
 136
 137 void noinstr exit_to_user_mode(void)
 138 {
 139         __exit_to_user_mode();
 140 }
 141
 142 /* Workaround to allow gradual conversion of architecture code */
 143 void __weak arch_do_signal_or_restart(struct pt_regs *regs) { }
 144
 145 #ifdef CONFIG_RT_DELAYED_SIGNALS
 146 static inline void raise_delayed_signal(void)
 147 {
 148         if (unlikely(current->forced_info.si_signo)) {
 149                 force_sig_info(&current->forced_info);
 150                 current->forced_info.si_signo = 0;
 151         }
 152 }
 153 #else
 154 static inline void raise_delayed_signal(void) { }
 155 #endif
 156
 157 static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
 158                                             unsigned long ti_work)
 159 {
 160         /*
 161          * Before returning to user space ensure that all pending work
 162          * items have been completed.
 163          */
 164         while (ti_work & EXIT_TO_USER_MODE_WORK) {
 165
 166                 local_irq_enable_exit_to_user(ti_work);
 167
 168                 if (ti_work & _TIF_NEED_RESCHED)
 169                         schedule();
 170
 171                 raise_delayed_signal();
 172
 173                 if (ti_work & _TIF_UPROBE)
 174                         uprobe_notify_resume(regs);
 175
 176                 if (ti_work & _TIF_PATCH_PENDING)
 177                         klp_update_patch_state(current);
 178
 179                 if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
 180                         arch_do_signal_or_restart(regs);
 181
 182                 if (ti_work & _TIF_NOTIFY_RESUME)
 183                         resume_user_mode_work(regs);
 184
 185                 /* Architecture specific TIF work */
 186                 arch_exit_to_user_mode_work(regs, ti_work);
 187
 188                 /*
 189                  * Disable interrupts and reevaluate the work flags as they
 190                  * might have changed while interrupts and preemption was
 191                  * enabled above.
 192                  */
 193                 local_irq_disable_exit_to_user();
 194
 195                 /* Check if any of the above work has queued a deferred wakeup */
 196                 tick_nohz_user_enter_prepare();
 197
 198                 ti_work = read_thread_flags();
 199         }
 200
 201         /* Return the latest work state for arch_exit_to_user_mode() */
 202         return ti_work;
 203 }
 204
 205 static void exit_to_user_mode_prepare(struct pt_regs *regs)
 206 {
 207         unsigned long ti_work = read_thread_flags();
 208
 209         lockdep_assert_irqs_disabled();
 210
 211         /* Flush pending rcuog wakeup before the last need_resched() check */
 212         tick_nohz_user_enter_prepare();
 213
 214         if (unlikely(ti_work & EXIT_TO_USER_MODE_WORK))
 215                 ti_work = exit_to_user_mode_loop(regs, ti_work);
 216
 217         arch_exit_to_user_mode_prepare(regs, ti_work);
 218
 219         /* Ensure that the address limit is intact and no locks are held */
 220         addr_limit_user_check();
 221         kmap_assert_nomap();
 222         lockdep_assert_irqs_disabled();
 223         lockdep_sys_exit();
 224 }
 225
 226 /*
 227  * If SYSCALL_EMU is set, then the only reason to report is when
 228  * SINGLESTEP is set (i.e. PTRACE_SYSEMU_SINGLESTEP).  This syscall
 229  * instruction has been already reported in syscall_enter_from_user_mode().
 230  */
 231 static inline bool report_single_step(unsigned long work)
 232 {
 233         if (work & SYSCALL_WORK_SYSCALL_EMU)
 234                 return false;
 235
 236         return work & SYSCALL_WORK_SYSCALL_EXIT_TRAP;
 237 }
 238
 239 static void syscall_exit_work(struct pt_regs *regs, unsigned long work)
 240 {
 241         bool step;
 242
 243         /*
 244          * If the syscall was rolled back due to syscall user dispatching,
 245          * then the tracers below are not invoked for the same reason as
 246          * the entry side was not invoked in syscall_trace_enter(): The ABI
 247          * of these syscalls is unknown.
 248          */
 249         if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
 250                 if (unlikely(current->syscall_dispatch.on_dispatch)) {
 251                         current->syscall_dispatch.on_dispatch = false;
 252                         return;
 253                 }
 254         }
 255
 256         audit_syscall_exit(regs);
 257
 258         if (work & SYSCALL_WORK_SYSCALL_TRACEPOINT)
 259                 trace_sys_exit(regs, syscall_get_return_value(current, regs));
 260
 261         step = report_single_step(work);
 262         if (step || work & SYSCALL_WORK_SYSCALL_TRACE)
 263                 ptrace_report_syscall_exit(regs, step);
 264 }
 265
 266 /*
 267  * Syscall specific exit to user mode preparation. Runs with interrupts
 268  * enabled.
 269  */
 270 static void syscall_exit_to_user_mode_prepare(struct pt_regs *regs)
 271 {
 272         unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
 273         unsigned long nr = syscall_get_nr(current, regs);
 274
 275         CT_WARN_ON(ct_state() != CONTEXT_KERNEL);
 276
 277         if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
 278                 if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr))
 279                         local_irq_enable();
 280         }
 281
 282         rseq_syscall(regs);
 283
 284         /*
 285          * Do one-time syscall specific work. If these work items are
 286          * enabled, we want to run them exactly once per syscall exit with
 287          * interrupts enabled.
 288          */
 289         if (unlikely(work & SYSCALL_WORK_EXIT))
 290                 syscall_exit_work(regs, work);
 291 }
 292
 293 static __always_inline void __syscall_exit_to_user_mode_work(struct pt_regs *regs)
 294 {
 295         syscall_exit_to_user_mode_prepare(regs);
 296         local_irq_disable_exit_to_user();
 297         exit_to_user_mode_prepare(regs);
 298 }
 299
 300 void syscall_exit_to_user_mode_work(struct pt_regs *regs)
 301 {
 302         __syscall_exit_to_user_mode_work(regs);
 303 }
 304
 305 __visible noinstr void syscall_exit_to_user_mode(struct pt_regs *regs)
 306 {
 307         instrumentation_begin();
 308         __syscall_exit_to_user_mode_work(regs);
 309         instrumentation_end();
 310         __exit_to_user_mode();
 311 }
 312
 313 noinstr void irqentry_enter_from_user_mode(struct pt_regs *regs)
 314 {
 315         __enter_from_user_mode(regs);
 316 }
 317
 318 noinstr void irqentry_exit_to_user_mode(struct pt_regs *regs)
 319 {
 320         instrumentation_begin();
 321         exit_to_user_mode_prepare(regs);
 322         instrumentation_end();
 323         __exit_to_user_mode();
 324 }
 325
 326 noinstr irqentry_state_t irqentry_enter(struct pt_regs *regs)
 327 {
 328         irqentry_state_t ret = {
 329                 .exit_rcu = false,
 330         };
 331
 332         if (user_mode(regs)) {
 333                 irqentry_enter_from_user_mode(regs);
 334                 return ret;
 335         }
 336
 337         /*
 338          * If this entry hit the idle task invoke rcu_irq_enter() whether
 339          * RCU is watching or not.
 340          *
 341          * Interrupts can nest when the first interrupt invokes softirq
 342          * processing on return which enables interrupts.
 343          *
 344          * Scheduler ticks in the idle task can mark quiescent state and
 345          * terminate a grace period, if and only if the timer interrupt is
 346          * not nested into another interrupt.
 347          *
 348          * Checking for rcu_is_watching() here would prevent the nesting
 349          * interrupt to invoke rcu_irq_enter(). If that nested interrupt is
 350          * the tick then rcu_flavor_sched_clock_irq() would wrongfully
 351          * assume that it is the first interrupt and eventually claim
 352          * quiescent state and end grace periods prematurely.
 353          *
 354          * Unconditionally invoke rcu_irq_enter() so RCU state stays
 355          * consistent.
 356          *
 357          * TINY_RCU does not support EQS, so let the compiler eliminate
 358          * this part when enabled.
 359          */
 360         if (!IS_ENABLED(CONFIG_TINY_RCU) && is_idle_task(current)) {
 361                 /*
 362                  * If RCU is not watching then the same careful
 363                  * sequence vs. lockdep and tracing is required
 364                  * as in irqentry_enter_from_user_mode().
 365                  */
 366                 lockdep_hardirqs_off(CALLER_ADDR0);
 367                 rcu_irq_enter();
 368                 instrumentation_begin();
 369                 trace_hardirqs_off_finish();
 370                 instrumentation_end();
 371
 372                 ret.exit_rcu = true;
 373                 return ret;
 374         }
 375
 376         /*
 377          * If RCU is watching then RCU only wants to check whether it needs
 378          * to restart the tick in NOHZ mode. rcu_irq_enter_check_tick()
 379          * already contains a warning when RCU is not watching, so no point
 380          * in having another one here.
 381          */
 382         lockdep_hardirqs_off(CALLER_ADDR0);
 383         instrumentation_begin();
 384         rcu_irq_enter_check_tick();
 385         trace_hardirqs_off_finish();
 386         instrumentation_end();
 387
 388         return ret;
 389 }
 390
 391 void raw_irqentry_exit_cond_resched(void)
 392 {
 393         if (!preempt_count()) {
 394                 /* Sanity check RCU and thread stack */
 395                 rcu_irq_exit_check_preempt();
 396                 if (IS_ENABLED(CONFIG_DEBUG_ENTRY))
 397                         WARN_ON_ONCE(!on_thread_stack());
 398                 if (need_resched())
 399                         preempt_schedule_irq();
 400         }
 401 }
 402 #ifdef CONFIG_PREEMPT_DYNAMIC
 403 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
 404 DEFINE_STATIC_CALL(irqentry_exit_cond_resched, raw_irqentry_exit_cond_resched);
 405 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
 406 DEFINE_STATIC_KEY_TRUE(sk_dynamic_irqentry_exit_cond_resched);
 407 void dynamic_irqentry_exit_cond_resched(void)
 408 {
 409         if (!static_key_unlikely(&sk_dynamic_irqentry_exit_cond_resched))
 410                 return;
 411         raw_irqentry_exit_cond_resched();
 412 }
 413 #endif
 414 #endif
 415
 416 noinstr void irqentry_exit(struct pt_regs *regs, irqentry_state_t state)
 417 {
 418         lockdep_assert_irqs_disabled();
 419
 420         /* Check whether this returns to user mode */
 421         if (user_mode(regs)) {
 422                 irqentry_exit_to_user_mode(regs);
 423         } else if (!regs_irqs_disabled(regs)) {
 424                 /*
 425                  * If RCU was not watching on entry this needs to be done
 426                  * carefully and needs the same ordering of lockdep/tracing
 427                  * and RCU as the return to user mode path.
 428                  */
 429                 if (state.exit_rcu) {
 430                         instrumentation_begin();
 431                         /* Tell the tracer that IRET will enable interrupts */
 432                         trace_hardirqs_on_prepare();
 433                         lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 434                         instrumentation_end();
 435                         rcu_irq_exit();
 436                         lockdep_hardirqs_on(CALLER_ADDR0);
 437                         return;
 438                 }
 439
 440                 instrumentation_begin();
 441                 if (IS_ENABLED(CONFIG_PREEMPTION))
 442                         irqentry_exit_cond_resched();
 443
 444                 /* Covers both tracing and lockdep */
 445                 trace_hardirqs_on();
 446                 instrumentation_end();
 447         } else {
 448                 /*
 449                  * IRQ flags state is correct already. Just tell RCU if it
 450                  * was not watching on entry.
 451                  */
 452                 if (state.exit_rcu)
 453                         rcu_irq_exit();
 454         }
 455 }
 456
 457 irqentry_state_t noinstr irqentry_nmi_enter(struct pt_regs *regs)
 458 {
 459         irqentry_state_t irq_state;
 460
 461         irq_state.lockdep = lockdep_hardirqs_enabled();
 462
 463         __nmi_enter();
 464         lockdep_hardirqs_off(CALLER_ADDR0);
 465         lockdep_hardirq_enter();
 466         rcu_nmi_enter();
 467
 468         instrumentation_begin();
 469         trace_hardirqs_off_finish();
 470         ftrace_nmi_enter();
 471         instrumentation_end();
 472
 473         return irq_state;
 474 }
 475
 476 void noinstr irqentry_nmi_exit(struct pt_regs *regs, irqentry_state_t irq_state)
 477 {
 478         instrumentation_begin();
 479         ftrace_nmi_exit();
 480         if (irq_state.lockdep) {
 481                 trace_hardirqs_on_prepare();
 482                 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 483         }
 484         instrumentation_end();
 485
 486         rcu_nmi_exit();
 487         lockdep_hardirq_exit();
 488         if (irq_state.lockdep)
 489                 lockdep_hardirqs_on(CALLER_ADDR0);
 490         __nmi_exit();
 491 }