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