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
   9 #include "common.h"
  10
  11 #define CREATE_TRACE_POINTS
  12 #include <trace/events/syscalls.h>
  13
  14 /* See comment for enter_from_user_mode() in entry-common.h */
  15 static __always_inline void __enter_from_user_mode(struct pt_regs *regs)
  16 {
  17         arch_check_user_regs(regs);
  18         lockdep_hardirqs_off(CALLER_ADDR0);
  19
  20         CT_WARN_ON(ct_state() != CONTEXT_USER);
  21         user_exit_irqoff();
  22
  23         instrumentation_begin();
  24         trace_hardirqs_off_finish();
  25         instrumentation_end();
  26 }
  27
  28 void noinstr enter_from_user_mode(struct pt_regs *regs)
  29 {
  30         __enter_from_user_mode(regs);
  31 }
  32
  33 static inline void syscall_enter_audit(struct pt_regs *regs, long syscall)
  34 {
  35         if (unlikely(audit_context())) {
  36                 unsigned long args[6];
  37
  38                 syscall_get_arguments(current, regs, args);
  39                 audit_syscall_entry(syscall, args[0], args[1], args[2], args[3]);
  40         }
  41 }
  42
  43 static long syscall_trace_enter(struct pt_regs *regs, long syscall,
  44                                 unsigned long work)
  45 {
  46         long ret = 0;
  47
  48         /*
  49          * Handle Syscall User Dispatch.  This must comes first, since
  50          * the ABI here can be something that doesn't make sense for
  51          * other syscall_work features.
  52          */
  53         if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
  54                 if (syscall_user_dispatch(regs))
  55                         return -1L;
  56         }
  57
  58         /* Handle ptrace */
  59         if (work & (SYSCALL_WORK_SYSCALL_TRACE | SYSCALL_WORK_SYSCALL_EMU)) {
  60                 ret = arch_syscall_enter_tracehook(regs);
  61                 if (ret || (work & SYSCALL_WORK_SYSCALL_EMU))
  62                         return -1L;
  63         }
  64
  65         /* Do seccomp after ptrace, to catch any tracer changes. */
  66         if (work & SYSCALL_WORK_SECCOMP) {
  67                 ret = __secure_computing(NULL);
  68                 if (ret == -1L)
  69                         return ret;
  70         }
  71
  72         /* Either of the above might have changed the syscall number */
  73         syscall = syscall_get_nr(current, regs);
  74
  75         if (unlikely(work & SYSCALL_WORK_SYSCALL_TRACEPOINT))
  76                 trace_sys_enter(regs, syscall);
  77
  78         syscall_enter_audit(regs, syscall);
  79
  80         return ret ? : syscall;
  81 }
  82
  83 static __always_inline long
  84 __syscall_enter_from_user_work(struct pt_regs *regs, long syscall)
  85 {
  86         unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
  87
  88         if (work & SYSCALL_WORK_ENTER)
  89                 syscall = syscall_trace_enter(regs, syscall, work);
  90
  91         return syscall;
  92 }
  93
  94 long syscall_enter_from_user_mode_work(struct pt_regs *regs, long syscall)
  95 {
  96         return __syscall_enter_from_user_work(regs, syscall);
  97 }
  98
  99 noinstr long syscall_enter_from_user_mode(struct pt_regs *regs, long syscall)
 100 {
 101         long ret;
 102
 103         __enter_from_user_mode(regs);
 104
 105         instrumentation_begin();
 106         local_irq_enable();
 107         ret = __syscall_enter_from_user_work(regs, syscall);
 108         instrumentation_end();
 109
 110         return ret;
 111 }
 112
 113 noinstr void syscall_enter_from_user_mode_prepare(struct pt_regs *regs)
 114 {
 115         __enter_from_user_mode(regs);
 116         instrumentation_begin();
 117         local_irq_enable();
 118         instrumentation_end();
 119 }
 120
 121 /* See comment for exit_to_user_mode() in entry-common.h */
 122 static __always_inline void __exit_to_user_mode(void)
 123 {
 124         instrumentation_begin();
 125         trace_hardirqs_on_prepare();
 126         lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 127         instrumentation_end();
 128
 129         user_enter_irqoff();
 130         arch_exit_to_user_mode();
 131         lockdep_hardirqs_on(CALLER_ADDR0);
 132 }
 133
 134 void noinstr exit_to_user_mode(void)
 135 {
 136         __exit_to_user_mode();
 137 }
 138
 139 /* Workaround to allow gradual conversion of architecture code */
 140 void __weak arch_do_signal_or_restart(struct pt_regs *regs, bool has_signal) { }
 141
 142 static void handle_signal_work(struct pt_regs *regs, unsigned long ti_work)
 143 {
 144         if (ti_work & _TIF_NOTIFY_SIGNAL)
 145                 tracehook_notify_signal();
 146
 147         arch_do_signal_or_restart(regs, ti_work & _TIF_SIGPENDING);
 148 }
 149
 150 static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
 151                                             unsigned long ti_work)
 152 {
 153         /*
 154          * Before returning to user space ensure that all pending work
 155          * items have been completed.
 156          */
 157         while (ti_work & EXIT_TO_USER_MODE_WORK) {
 158
 159                 local_irq_enable_exit_to_user(ti_work);
 160
 161                 if (ti_work & _TIF_NEED_RESCHED)
 162                         schedule();
 163
 164                 if (ti_work & _TIF_UPROBE)
 165                         uprobe_notify_resume(regs);
 166
 167                 if (ti_work & _TIF_PATCH_PENDING)
 168                         klp_update_patch_state(current);
 169
 170                 if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
 171                         handle_signal_work(regs, ti_work);
 172
 173                 if (ti_work & _TIF_NOTIFY_RESUME) {
 174                         tracehook_notify_resume(regs);
 175                         rseq_handle_notify_resume(NULL, regs);
 176                 }
 177
 178                 /* Architecture specific TIF work */
 179                 arch_exit_to_user_mode_work(regs, ti_work);
 180
 181                 /*
 182                  * Disable interrupts and reevaluate the work flags as they
 183                  * might have changed while interrupts and preemption was
 184                  * enabled above.
 185                  */
 186                 local_irq_disable_exit_to_user();
 187                 ti_work = READ_ONCE(current_thread_info()->flags);
 188         }
 189
 190         /* Return the latest work state for arch_exit_to_user_mode() */
 191         return ti_work;
 192 }
 193
 194 static void exit_to_user_mode_prepare(struct pt_regs *regs)
 195 {
 196         unsigned long ti_work = READ_ONCE(current_thread_info()->flags);
 197
 198         lockdep_assert_irqs_disabled();
 199
 200         if (unlikely(ti_work & EXIT_TO_USER_MODE_WORK))
 201                 ti_work = exit_to_user_mode_loop(regs, ti_work);
 202
 203         arch_exit_to_user_mode_prepare(regs, ti_work);
 204
 205         /* Ensure that the address limit is intact and no locks are held */
 206         addr_limit_user_check();
 207         kmap_assert_nomap();
 208         lockdep_assert_irqs_disabled();
 209         lockdep_sys_exit();
 210 }
 211
 212 #ifndef _TIF_SINGLESTEP
 213 static inline bool report_single_step(unsigned long work)
 214 {
 215         return false;
 216 }
 217 #else
 218 /*
 219  * If SYSCALL_EMU is set, then the only reason to report is when
 220  * TIF_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 !!(current_thread_info()->flags & _TIF_SINGLESTEP);
 229 }
 230 #endif
 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 interupt 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
 397 noinstr void irqentry_exit(struct pt_regs *regs, irqentry_state_t state)
 398 {
 399         lockdep_assert_irqs_disabled();
 400
 401         /* Check whether this returns to user mode */
 402         if (user_mode(regs)) {
 403                 irqentry_exit_to_user_mode(regs);
 404         } else if (!regs_irqs_disabled(regs)) {
 405                 /*
 406                  * If RCU was not watching on entry this needs to be done
 407                  * carefully and needs the same ordering of lockdep/tracing
 408                  * and RCU as the return to user mode path.
 409                  */
 410                 if (state.exit_rcu) {
 411                         instrumentation_begin();
 412                         /* Tell the tracer that IRET will enable interrupts */
 413                         trace_hardirqs_on_prepare();
 414                         lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 415                         instrumentation_end();
 416                         rcu_irq_exit();
 417                         lockdep_hardirqs_on(CALLER_ADDR0);
 418                         return;
 419                 }
 420
 421                 instrumentation_begin();
 422                 if (IS_ENABLED(CONFIG_PREEMPTION))
 423                         irqentry_exit_cond_resched();
 424                 /* Covers both tracing and lockdep */
 425                 trace_hardirqs_on();
 426                 instrumentation_end();
 427         } else {
 428                 /*
 429                  * IRQ flags state is correct already. Just tell RCU if it
 430                  * was not watching on entry.
 431                  */
 432                 if (state.exit_rcu)
 433                         rcu_irq_exit();
 434         }
 435 }
 436
 437 irqentry_state_t noinstr irqentry_nmi_enter(struct pt_regs *regs)
 438 {
 439         irqentry_state_t irq_state;
 440
 441         irq_state.lockdep = lockdep_hardirqs_enabled();
 442
 443         __nmi_enter();
 444         lockdep_hardirqs_off(CALLER_ADDR0);
 445         lockdep_hardirq_enter();
 446         rcu_nmi_enter();
 447
 448         instrumentation_begin();
 449         trace_hardirqs_off_finish();
 450         ftrace_nmi_enter();
 451         instrumentation_end();
 452
 453         return irq_state;
 454 }
 455
 456 void noinstr irqentry_nmi_exit(struct pt_regs *regs, irqentry_state_t irq_state)
 457 {
 458         instrumentation_begin();
 459         ftrace_nmi_exit();
 460         if (irq_state.lockdep) {
 461                 trace_hardirqs_on_prepare();
 462                 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 463         }
 464         instrumentation_end();
 465
 466         rcu_nmi_exit();
 467         lockdep_hardirq_exit();
 468         if (irq_state.lockdep)
 469                 lockdep_hardirqs_on(CALLER_ADDR0);
 470         __nmi_exit();
 471 }