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
 177                 /* Architecture specific TIF work */
 178                 arch_exit_to_user_mode_work(regs, ti_work);
 179
 180                 /*
 181                  * Disable interrupts and reevaluate the work flags as they
 182                  * might have changed while interrupts and preemption was
 183                  * enabled above.
 184                  */
 185                 local_irq_disable_exit_to_user();
 186
 187                 /* Check if any of the above work has queued a deferred wakeup */
 188                 tick_nohz_user_enter_prepare();
 189
 190                 ti_work = READ_ONCE(current_thread_info()->flags);
 191         }
 192
 193         /* Return the latest work state for arch_exit_to_user_mode() */
 194         return ti_work;
 195 }
 196
 197 static void exit_to_user_mode_prepare(struct pt_regs *regs)
 198 {
 199         unsigned long ti_work = READ_ONCE(current_thread_info()->flags);
 200
 201         lockdep_assert_irqs_disabled();
 202
 203         /* Flush pending rcuog wakeup before the last need_resched() check */
 204         tick_nohz_user_enter_prepare();
 205
 206         if (unlikely(ti_work & EXIT_TO_USER_MODE_WORK))
 207                 ti_work = exit_to_user_mode_loop(regs, ti_work);
 208
 209         arch_exit_to_user_mode_prepare(regs, ti_work);
 210
 211         /* Ensure that the address limit is intact and no locks are held */
 212         addr_limit_user_check();
 213         kmap_assert_nomap();
 214         lockdep_assert_irqs_disabled();
 215         lockdep_sys_exit();
 216 }
 217
 218 /*
 219  * If SYSCALL_EMU is set, then the only reason to report is when
 220  * SINGLESTEP is set (i.e. PTRACE_SYSEMU_SINGLESTEP).  This syscall
 221  * instruction has been already reported in syscall_enter_from_user_mode().
 222  */
 223 static inline bool report_single_step(unsigned long work)
 224 {
 225         if (work & SYSCALL_WORK_SYSCALL_EMU)
 226                 return false;
 227
 228         return work & SYSCALL_WORK_SYSCALL_EXIT_TRAP;
 229 }
 230
 231 static void syscall_exit_work(struct pt_regs *regs, unsigned long work)
 232 {
 233         bool step;
 234
 235         /*
 236          * If the syscall was rolled back due to syscall user dispatching,
 237          * then the tracers below are not invoked for the same reason as
 238          * the entry side was not invoked in syscall_trace_enter(): The ABI
 239          * of these syscalls is unknown.
 240          */
 241         if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
 242                 if (unlikely(current->syscall_dispatch.on_dispatch)) {
 243                         current->syscall_dispatch.on_dispatch = false;
 244                         return;
 245                 }
 246         }
 247
 248         audit_syscall_exit(regs);
 249
 250         if (work & SYSCALL_WORK_SYSCALL_TRACEPOINT)
 251                 trace_sys_exit(regs, syscall_get_return_value(current, regs));
 252
 253         step = report_single_step(work);
 254         if (step || work & SYSCALL_WORK_SYSCALL_TRACE)
 255                 arch_syscall_exit_tracehook(regs, step);
 256 }
 257
 258 /*
 259  * Syscall specific exit to user mode preparation. Runs with interrupts
 260  * enabled.
 261  */
 262 static void syscall_exit_to_user_mode_prepare(struct pt_regs *regs)
 263 {
 264         unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
 265         unsigned long nr = syscall_get_nr(current, regs);
 266
 267         CT_WARN_ON(ct_state() != CONTEXT_KERNEL);
 268
 269         if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
 270                 if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr))
 271                         local_irq_enable();
 272         }
 273
 274         rseq_syscall(regs);
 275
 276         /*
 277          * Do one-time syscall specific work. If these work items are
 278          * enabled, we want to run them exactly once per syscall exit with
 279          * interrupts enabled.
 280          */
 281         if (unlikely(work & SYSCALL_WORK_EXIT))
 282                 syscall_exit_work(regs, work);
 283 }
 284
 285 static __always_inline void __syscall_exit_to_user_mode_work(struct pt_regs *regs)
 286 {
 287         syscall_exit_to_user_mode_prepare(regs);
 288         local_irq_disable_exit_to_user();
 289         exit_to_user_mode_prepare(regs);
 290 }
 291
 292 void syscall_exit_to_user_mode_work(struct pt_regs *regs)
 293 {
 294         __syscall_exit_to_user_mode_work(regs);
 295 }
 296
 297 __visible noinstr void syscall_exit_to_user_mode(struct pt_regs *regs)
 298 {
 299         instrumentation_begin();
 300         __syscall_exit_to_user_mode_work(regs);
 301         instrumentation_end();
 302         __exit_to_user_mode();
 303 }
 304
 305 noinstr void irqentry_enter_from_user_mode(struct pt_regs *regs)
 306 {
 307         __enter_from_user_mode(regs);
 308 }
 309
 310 noinstr void irqentry_exit_to_user_mode(struct pt_regs *regs)
 311 {
 312         instrumentation_begin();
 313         exit_to_user_mode_prepare(regs);
 314         instrumentation_end();
 315         __exit_to_user_mode();
 316 }
 317
 318 noinstr irqentry_state_t irqentry_enter(struct pt_regs *regs)
 319 {
 320         irqentry_state_t ret = {
 321                 .exit_rcu = false,
 322         };
 323
 324         if (user_mode(regs)) {
 325                 irqentry_enter_from_user_mode(regs);
 326                 return ret;
 327         }
 328
 329         /*
 330          * If this entry hit the idle task invoke rcu_irq_enter() whether
 331          * RCU is watching or not.
 332          *
 333          * Interrupts can nest when the first interrupt invokes softirq
 334          * processing on return which enables interrupts.
 335          *
 336          * Scheduler ticks in the idle task can mark quiescent state and
 337          * terminate a grace period, if and only if the timer interrupt is
 338          * not nested into another interrupt.
 339          *
 340          * Checking for rcu_is_watching() here would prevent the nesting
 341          * interrupt to invoke rcu_irq_enter(). If that nested interrupt is
 342          * the tick then rcu_flavor_sched_clock_irq() would wrongfully
 343          * assume that it is the first interrupt and eventually claim
 344          * quiescent state and end grace periods prematurely.
 345          *
 346          * Unconditionally invoke rcu_irq_enter() so RCU state stays
 347          * consistent.
 348          *
 349          * TINY_RCU does not support EQS, so let the compiler eliminate
 350          * this part when enabled.
 351          */
 352         if (!IS_ENABLED(CONFIG_TINY_RCU) && is_idle_task(current)) {
 353                 /*
 354                  * If RCU is not watching then the same careful
 355                  * sequence vs. lockdep and tracing is required
 356                  * as in irqentry_enter_from_user_mode().
 357                  */
 358                 lockdep_hardirqs_off(CALLER_ADDR0);
 359                 rcu_irq_enter();
 360                 instrumentation_begin();
 361                 trace_hardirqs_off_finish();
 362                 instrumentation_end();
 363
 364                 ret.exit_rcu = true;
 365                 return ret;
 366         }
 367
 368         /*
 369          * If RCU is watching then RCU only wants to check whether it needs
 370          * to restart the tick in NOHZ mode. rcu_irq_enter_check_tick()
 371          * already contains a warning when RCU is not watching, so no point
 372          * in having another one here.
 373          */
 374         lockdep_hardirqs_off(CALLER_ADDR0);
 375         instrumentation_begin();
 376         rcu_irq_enter_check_tick();
 377         trace_hardirqs_off_finish();
 378         instrumentation_end();
 379
 380         return ret;
 381 }
 382
 383 void irqentry_exit_cond_resched(void)
 384 {
 385         if (!preempt_count()) {
 386                 /* Sanity check RCU and thread stack */
 387                 rcu_irq_exit_check_preempt();
 388                 if (IS_ENABLED(CONFIG_DEBUG_ENTRY))
 389                         WARN_ON_ONCE(!on_thread_stack());
 390                 if (need_resched())
 391                         preempt_schedule_irq();
 392         }
 393 }
 394 #ifdef CONFIG_PREEMPT_DYNAMIC
 395 DEFINE_STATIC_CALL(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
 396 #endif
 397
 398 noinstr void irqentry_exit(struct pt_regs *regs, irqentry_state_t state)
 399 {
 400         lockdep_assert_irqs_disabled();
 401
 402         /* Check whether this returns to user mode */
 403         if (user_mode(regs)) {
 404                 irqentry_exit_to_user_mode(regs);
 405         } else if (!regs_irqs_disabled(regs)) {
 406                 /*
 407                  * If RCU was not watching on entry this needs to be done
 408                  * carefully and needs the same ordering of lockdep/tracing
 409                  * and RCU as the return to user mode path.
 410                  */
 411                 if (state.exit_rcu) {
 412                         instrumentation_begin();
 413                         /* Tell the tracer that IRET will enable interrupts */
 414                         trace_hardirqs_on_prepare();
 415                         lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 416                         instrumentation_end();
 417                         rcu_irq_exit();
 418                         lockdep_hardirqs_on(CALLER_ADDR0);
 419                         return;
 420                 }
 421
 422                 instrumentation_begin();
 423                 if (IS_ENABLED(CONFIG_PREEMPTION)) {
 424 #ifdef CONFIG_PREEMPT_DYNAMIC
 425                         static_call(irqentry_exit_cond_resched)();
 426 #else
 427                         irqentry_exit_cond_resched();
 428 #endif
 429                 }
 430                 /* Covers both tracing and lockdep */
 431                 trace_hardirqs_on();
 432                 instrumentation_end();
 433         } else {
 434                 /*
 435                  * IRQ flags state is correct already. Just tell RCU if it
 436                  * was not watching on entry.
 437                  */
 438                 if (state.exit_rcu)
 439                         rcu_irq_exit();
 440         }
 441 }
 442
 443 irqentry_state_t noinstr irqentry_nmi_enter(struct pt_regs *regs)
 444 {
 445         irqentry_state_t irq_state;
 446
 447         irq_state.lockdep = lockdep_hardirqs_enabled();
 448
 449         __nmi_enter();
 450         lockdep_hardirqs_off(CALLER_ADDR0);
 451         lockdep_hardirq_enter();
 452         rcu_nmi_enter();
 453
 454         instrumentation_begin();
 455         trace_hardirqs_off_finish();
 456         ftrace_nmi_enter();
 457         instrumentation_end();
 458
 459         return irq_state;
 460 }
 461
 462 void noinstr irqentry_nmi_exit(struct pt_regs *regs, irqentry_state_t irq_state)
 463 {
 464         instrumentation_begin();
 465         ftrace_nmi_exit();
 466         if (irq_state.lockdep) {
 467                 trace_hardirqs_on_prepare();
 468                 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 469         }
 470         instrumentation_end();
 471
 472         rcu_nmi_exit();
 473         lockdep_hardirq_exit();
 474         if (irq_state.lockdep)
 475                 lockdep_hardirqs_on(CALLER_ADDR0);
 476         __nmi_exit();
 477 }