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/livepatch.h>
   6 #include <linux/audit.h>
   7
   8 #include "common.h"
   9
  10 #define CREATE_TRACE_POINTS
  11 #include <trace/events/syscalls.h>
  12
  13 /**
  14  * enter_from_user_mode - Establish state when coming from user mode
  15  *
  16  * Syscall/interrupt entry disables interrupts, but user mode is traced as
  17  * interrupts enabled. Also with NO_HZ_FULL RCU might be idle.
  18  *
  19  * 1) Tell lockdep that interrupts are disabled
  20  * 2) Invoke context tracking if enabled to reactivate RCU
  21  * 3) Trace interrupts off state
  22  */
  23 static __always_inline void enter_from_user_mode(struct pt_regs *regs)
  24 {
  25         arch_check_user_regs(regs);
  26         lockdep_hardirqs_off(CALLER_ADDR0);
  27
  28         CT_WARN_ON(ct_state() != CONTEXT_USER);
  29         user_exit_irqoff();
  30
  31         instrumentation_begin();
  32         trace_hardirqs_off_finish();
  33         instrumentation_end();
  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 = arch_syscall_enter_tracehook(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 /**
 125  * exit_to_user_mode - Fixup state when exiting to user mode
 126  *
 127  * Syscall/interupt exit enables interrupts, but the kernel state is
 128  * interrupts disabled when this is invoked. Also tell RCU about it.
 129  *
 130  * 1) Trace interrupts on state
 131  * 2) Invoke context tracking if enabled to adjust RCU state
 132  * 3) Invoke architecture specific last minute exit code, e.g. speculation
 133  *    mitigations, etc.
 134  * 4) Tell lockdep that interrupts are enabled
 135  */
 136 static __always_inline void exit_to_user_mode(void)
 137 {
 138         instrumentation_begin();
 139         trace_hardirqs_on_prepare();
 140         lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 141         instrumentation_end();
 142
 143         user_enter_irqoff();
 144         arch_exit_to_user_mode();
 145         lockdep_hardirqs_on(CALLER_ADDR0);
 146 }
 147
 148 /* Workaround to allow gradual conversion of architecture code */
 149 void __weak arch_do_signal_or_restart(struct pt_regs *regs, bool has_signal) { }
 150
 151 static void handle_signal_work(struct pt_regs *regs, unsigned long ti_work)
 152 {
 153         if (ti_work & _TIF_NOTIFY_SIGNAL)
 154                 tracehook_notify_signal();
 155
 156         arch_do_signal_or_restart(regs, ti_work & _TIF_SIGPENDING);
 157 }
 158
 159 static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
 160                                             unsigned long ti_work)
 161 {
 162         /*
 163          * Before returning to user space ensure that all pending work
 164          * items have been completed.
 165          */
 166         while (ti_work & EXIT_TO_USER_MODE_WORK) {
 167
 168                 local_irq_enable_exit_to_user(ti_work);
 169
 170                 if (ti_work & _TIF_NEED_RESCHED)
 171                         schedule();
 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                         handle_signal_work(regs, ti_work);
 181
 182                 if (ti_work & _TIF_NOTIFY_RESUME) {
 183                         tracehook_notify_resume(regs);
 184                         rseq_handle_notify_resume(NULL, regs);
 185                 }
 186
 187                 /* Architecture specific TIF work */
 188                 arch_exit_to_user_mode_work(regs, ti_work);
 189
 190                 /*
 191                  * Disable interrupts and reevaluate the work flags as they
 192                  * might have changed while interrupts and preemption was
 193                  * enabled above.
 194                  */
 195                 local_irq_disable_exit_to_user();
 196                 ti_work = READ_ONCE(current_thread_info()->flags);
 197         }
 198
 199         /* Return the latest work state for arch_exit_to_user_mode() */
 200         return ti_work;
 201 }
 202
 203 static void exit_to_user_mode_prepare(struct pt_regs *regs)
 204 {
 205         unsigned long ti_work = READ_ONCE(current_thread_info()->flags);
 206
 207         lockdep_assert_irqs_disabled();
 208
 209         if (unlikely(ti_work & EXIT_TO_USER_MODE_WORK))
 210                 ti_work = exit_to_user_mode_loop(regs, ti_work);
 211
 212         arch_exit_to_user_mode_prepare(regs, ti_work);
 213
 214         /* Ensure that the address limit is intact and no locks are held */
 215         addr_limit_user_check();
 216         lockdep_assert_irqs_disabled();
 217         lockdep_sys_exit();
 218 }
 219
 220 #ifndef _TIF_SINGLESTEP
 221 static inline bool report_single_step(unsigned long work)
 222 {
 223         return false;
 224 }
 225 #else
 226 /*
 227  * If SYSCALL_EMU is set, then the only reason to report is when
 228  * TIF_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 !!(current_thread_info()->flags & _TIF_SINGLESTEP);
 237 }
 238 #endif
 239
 240
 241 static void syscall_exit_work(struct pt_regs *regs, unsigned long work)
 242 {
 243         bool step;
 244
 245         /*
 246          * If the syscall was rolled back due to syscall user dispatching,
 247          * then the tracers below are not invoked for the same reason as
 248          * the entry side was not invoked in syscall_trace_enter(): The ABI
 249          * of these syscalls is unknown.
 250          */
 251         if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
 252                 if (unlikely(current->syscall_dispatch.on_dispatch)) {
 253                         current->syscall_dispatch.on_dispatch = false;
 254                         return;
 255                 }
 256         }
 257
 258         audit_syscall_exit(regs);
 259
 260         if (work & SYSCALL_WORK_SYSCALL_TRACEPOINT)
 261                 trace_sys_exit(regs, syscall_get_return_value(current, regs));
 262
 263         step = report_single_step(work);
 264         if (step || work & SYSCALL_WORK_SYSCALL_TRACE)
 265                 arch_syscall_exit_tracehook(regs, step);
 266 }
 267
 268 /*
 269  * Syscall specific exit to user mode preparation. Runs with interrupts
 270  * enabled.
 271  */
 272 static void syscall_exit_to_user_mode_prepare(struct pt_regs *regs)
 273 {
 274         unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
 275         unsigned long nr = syscall_get_nr(current, regs);
 276
 277         CT_WARN_ON(ct_state() != CONTEXT_KERNEL);
 278
 279         if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
 280                 if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr))
 281                         local_irq_enable();
 282         }
 283
 284         rseq_syscall(regs);
 285
 286         /*
 287          * Do one-time syscall specific work. If these work items are
 288          * enabled, we want to run them exactly once per syscall exit with
 289          * interrupts enabled.
 290          */
 291         if (unlikely(work & SYSCALL_WORK_EXIT))
 292                 syscall_exit_work(regs, work);
 293 }
 294
 295 __visible noinstr void syscall_exit_to_user_mode(struct pt_regs *regs)
 296 {
 297         instrumentation_begin();
 298         syscall_exit_to_user_mode_prepare(regs);
 299         local_irq_disable_exit_to_user();
 300         exit_to_user_mode_prepare(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 interupt 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
 395 noinstr void irqentry_exit(struct pt_regs *regs, irqentry_state_t state)
 396 {
 397         lockdep_assert_irqs_disabled();
 398
 399         /* Check whether this returns to user mode */
 400         if (user_mode(regs)) {
 401                 irqentry_exit_to_user_mode(regs);
 402         } else if (!regs_irqs_disabled(regs)) {
 403                 /*
 404                  * If RCU was not watching on entry this needs to be done
 405                  * carefully and needs the same ordering of lockdep/tracing
 406                  * and RCU as the return to user mode path.
 407                  */
 408                 if (state.exit_rcu) {
 409                         instrumentation_begin();
 410                         /* Tell the tracer that IRET will enable interrupts */
 411                         trace_hardirqs_on_prepare();
 412                         lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 413                         instrumentation_end();
 414                         rcu_irq_exit();
 415                         lockdep_hardirqs_on(CALLER_ADDR0);
 416                         return;
 417                 }
 418
 419                 instrumentation_begin();
 420                 if (IS_ENABLED(CONFIG_PREEMPTION))
 421                         irqentry_exit_cond_resched();
 422                 /* Covers both tracing and lockdep */
 423                 trace_hardirqs_on();
 424                 instrumentation_end();
 425         } else {
 426                 /*
 427                  * IRQ flags state is correct already. Just tell RCU if it
 428                  * was not watching on entry.
 429                  */
 430                 if (state.exit_rcu)
 431                         rcu_irq_exit();
 432         }
 433 }
 434
 435 irqentry_state_t noinstr irqentry_nmi_enter(struct pt_regs *regs)
 436 {
 437         irqentry_state_t irq_state;
 438
 439         irq_state.lockdep = lockdep_hardirqs_enabled();
 440
 441         __nmi_enter();
 442         lockdep_hardirqs_off(CALLER_ADDR0);
 443         lockdep_hardirq_enter();
 444         rcu_nmi_enter();
 445
 446         instrumentation_begin();
 447         trace_hardirqs_off_finish();
 448         ftrace_nmi_enter();
 449         instrumentation_end();
 450
 451         return irq_state;
 452 }
 453
 454 void noinstr irqentry_nmi_exit(struct pt_regs *regs, irqentry_state_t irq_state)
 455 {
 456         instrumentation_begin();
 457         ftrace_nmi_exit();
 458         if (irq_state.lockdep) {
 459                 trace_hardirqs_on_prepare();
 460                 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 461         }
 462         instrumentation_end();
 463
 464         rcu_nmi_exit();
 465         lockdep_hardirq_exit();
 466         if (irq_state.lockdep)
 467                 lockdep_hardirqs_on(CALLER_ADDR0);
 468         __nmi_exit();
 469 }