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/kmsan.h>
   9 #include <linux/livepatch.h>
  10 #include <linux/audit.h>
  11 #include <linux/tick.h>
  12
  13 #include "common.h"
  14
  15 #define CREATE_TRACE_POINTS
  16 #include <trace/events/syscalls.h>
  17
  18 /* See comment for enter_from_user_mode() in entry-common.h */
  19 static __always_inline void __enter_from_user_mode(struct pt_regs *regs)
  20 {
  21         arch_enter_from_user_mode(regs);
  22         lockdep_hardirqs_off(CALLER_ADDR0);
  23
  24         CT_WARN_ON(__ct_state() != CONTEXT_USER);
  25         user_exit_irqoff();
  26
  27         instrumentation_begin();
  28         kmsan_unpoison_entry_regs(regs);
  29         trace_hardirqs_off_finish();
  30         instrumentation_end();
  31 }
  32
  33 void noinstr enter_from_user_mode(struct pt_regs *regs)
  34 {
  35         __enter_from_user_mode(regs);
  36 }
  37
  38 static inline void syscall_enter_audit(struct pt_regs *regs, long syscall)
  39 {
  40         if (unlikely(audit_context())) {
  41                 unsigned long args[6];
  42
  43                 syscall_get_arguments(current, regs, args);
  44                 audit_syscall_entry(syscall, args[0], args[1], args[2], args[3]);
  45         }
  46 }
  47
  48 static long syscall_trace_enter(struct pt_regs *regs, long syscall,
  49                                 unsigned long work)
  50 {
  51         long ret = 0;
  52
  53         /*
  54          * Handle Syscall User Dispatch.  This must comes first, since
  55          * the ABI here can be something that doesn't make sense for
  56          * other syscall_work features.
  57          */
  58         if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
  59                 if (syscall_user_dispatch(regs))
  60                         return -1L;
  61         }
  62
  63         /* Handle ptrace */
  64         if (work & (SYSCALL_WORK_SYSCALL_TRACE | SYSCALL_WORK_SYSCALL_EMU)) {
  65                 ret = ptrace_report_syscall_entry(regs);
  66                 if (ret || (work & SYSCALL_WORK_SYSCALL_EMU))
  67                         return -1L;
  68         }
  69
  70         /* Do seccomp after ptrace, to catch any tracer changes. */
  71         if (work & SYSCALL_WORK_SECCOMP) {
  72                 ret = __secure_computing(NULL);
  73                 if (ret == -1L)
  74                         return ret;
  75         }
  76
  77         /* Either of the above might have changed the syscall number */
  78         syscall = syscall_get_nr(current, regs);
  79
  80         if (unlikely(work & SYSCALL_WORK_SYSCALL_TRACEPOINT)) {
  81                 trace_sys_enter(regs, syscall);
  82                 /*
  83                  * Probes or BPF hooks in the tracepoint may have changed the
  84                  * system call number as well.
  85                  */
  86                 syscall = syscall_get_nr(current, regs);
  87         }
  88
  89         syscall_enter_audit(regs, syscall);
  90
  91         return ret ? : syscall;
  92 }
  93
  94 static __always_inline long
  95 __syscall_enter_from_user_work(struct pt_regs *regs, long syscall)
  96 {
  97         unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
  98
  99         if (work & SYSCALL_WORK_ENTER)
 100                 syscall = syscall_trace_enter(regs, syscall, work);
 101
 102         return syscall;
 103 }
 104
 105 long syscall_enter_from_user_mode_work(struct pt_regs *regs, long syscall)
 106 {
 107         return __syscall_enter_from_user_work(regs, syscall);
 108 }
 109
 110 noinstr long syscall_enter_from_user_mode(struct pt_regs *regs, long syscall)
 111 {
 112         long ret;
 113
 114         __enter_from_user_mode(regs);
 115
 116         instrumentation_begin();
 117         local_irq_enable();
 118         ret = __syscall_enter_from_user_work(regs, syscall);
 119         instrumentation_end();
 120
 121         return ret;
 122 }
 123
 124 noinstr void syscall_enter_from_user_mode_prepare(struct pt_regs *regs)
 125 {
 126         __enter_from_user_mode(regs);
 127         instrumentation_begin();
 128         local_irq_enable();
 129         instrumentation_end();
 130 }
 131
 132 /* See comment for exit_to_user_mode() in entry-common.h */
 133 static __always_inline void __exit_to_user_mode(void)
 134 {
 135         instrumentation_begin();
 136         trace_hardirqs_on_prepare();
 137         lockdep_hardirqs_on_prepare();
 138         instrumentation_end();
 139
 140         user_enter_irqoff();
 141         arch_exit_to_user_mode();
 142         lockdep_hardirqs_on(CALLER_ADDR0);
 143 }
 144
 145 void noinstr exit_to_user_mode(void)
 146 {
 147         __exit_to_user_mode();
 148 }
 149
 150 /* Workaround to allow gradual conversion of architecture code */
 151 void __weak arch_do_signal_or_restart(struct pt_regs *regs) { }
 152
 153 static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
 154                                             unsigned long ti_work)
 155 {
 156         /*
 157          * Before returning to user space ensure that all pending work
 158          * items have been completed.
 159          */
 160         while (ti_work & EXIT_TO_USER_MODE_WORK) {
 161
 162                 local_irq_enable_exit_to_user(ti_work);
 163
 164                 if (ti_work & _TIF_NEED_RESCHED)
 165                         schedule();
 166
 167                 if (ti_work & _TIF_UPROBE)
 168                         uprobe_notify_resume(regs);
 169
 170                 if (ti_work & _TIF_PATCH_PENDING)
 171                         klp_update_patch_state(current);
 172
 173                 if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
 174                         arch_do_signal_or_restart(regs);
 175
 176                 if (ti_work & _TIF_NOTIFY_RESUME)
 177                         resume_user_mode_work(regs);
 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_thread_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;
 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         ti_work = read_thread_flags();
 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 kernel state is sane for a return to userspace */
 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                 ptrace_report_syscall_exit(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 ct_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 ct_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 ct_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                 ct_irq_enter();
 362                 instrumentation_begin();
 363                 kmsan_unpoison_entry_regs(regs);
 364                 trace_hardirqs_off_finish();
 365                 instrumentation_end();
 366
 367                 ret.exit_rcu = true;
 368                 return ret;
 369         }
 370
 371         /*
 372          * If RCU is watching then RCU only wants to check whether it needs
 373          * to restart the tick in NOHZ mode. rcu_irq_enter_check_tick()
 374          * already contains a warning when RCU is not watching, so no point
 375          * in having another one here.
 376          */
 377         lockdep_hardirqs_off(CALLER_ADDR0);
 378         instrumentation_begin();
 379         kmsan_unpoison_entry_regs(regs);
 380         rcu_irq_enter_check_tick();
 381         trace_hardirqs_off_finish();
 382         instrumentation_end();
 383
 384         return ret;
 385 }
 386
 387 void raw_irqentry_exit_cond_resched(void)
 388 {
 389         if (!preempt_count()) {
 390                 /* Sanity check RCU and thread stack */
 391                 rcu_irq_exit_check_preempt();
 392                 if (IS_ENABLED(CONFIG_DEBUG_ENTRY))
 393                         WARN_ON_ONCE(!on_thread_stack());
 394                 if (need_resched())
 395                         preempt_schedule_irq();
 396         }
 397 }
 398 #ifdef CONFIG_PREEMPT_DYNAMIC
 399 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
 400 DEFINE_STATIC_CALL(irqentry_exit_cond_resched, raw_irqentry_exit_cond_resched);
 401 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
 402 DEFINE_STATIC_KEY_TRUE(sk_dynamic_irqentry_exit_cond_resched);
 403 void dynamic_irqentry_exit_cond_resched(void)
 404 {
 405         if (!static_branch_unlikely(&sk_dynamic_irqentry_exit_cond_resched))
 406                 return;
 407         raw_irqentry_exit_cond_resched();
 408 }
 409 #endif
 410 #endif
 411
 412 noinstr void irqentry_exit(struct pt_regs *regs, irqentry_state_t state)
 413 {
 414         lockdep_assert_irqs_disabled();
 415
 416         /* Check whether this returns to user mode */
 417         if (user_mode(regs)) {
 418                 irqentry_exit_to_user_mode(regs);
 419         } else if (!regs_irqs_disabled(regs)) {
 420                 /*
 421                  * If RCU was not watching on entry this needs to be done
 422                  * carefully and needs the same ordering of lockdep/tracing
 423                  * and RCU as the return to user mode path.
 424                  */
 425                 if (state.exit_rcu) {
 426                         instrumentation_begin();
 427                         /* Tell the tracer that IRET will enable interrupts */
 428                         trace_hardirqs_on_prepare();
 429                         lockdep_hardirqs_on_prepare();
 430                         instrumentation_end();
 431                         ct_irq_exit();
 432                         lockdep_hardirqs_on(CALLER_ADDR0);
 433                         return;
 434                 }
 435
 436                 instrumentation_begin();
 437                 if (IS_ENABLED(CONFIG_PREEMPTION))
 438                         irqentry_exit_cond_resched();
 439
 440                 /* Covers both tracing and lockdep */
 441                 trace_hardirqs_on();
 442                 instrumentation_end();
 443         } else {
 444                 /*
 445                  * IRQ flags state is correct already. Just tell RCU if it
 446                  * was not watching on entry.
 447                  */
 448                 if (state.exit_rcu)
 449                         ct_irq_exit();
 450         }
 451 }
 452
 453 irqentry_state_t noinstr irqentry_nmi_enter(struct pt_regs *regs)
 454 {
 455         irqentry_state_t irq_state;
 456
 457         irq_state.lockdep = lockdep_hardirqs_enabled();
 458
 459         __nmi_enter();
 460         lockdep_hardirqs_off(CALLER_ADDR0);
 461         lockdep_hardirq_enter();
 462         ct_nmi_enter();
 463
 464         instrumentation_begin();
 465         kmsan_unpoison_entry_regs(regs);
 466         trace_hardirqs_off_finish();
 467         ftrace_nmi_enter();
 468         instrumentation_end();
 469
 470         return irq_state;
 471 }
 472
 473 void noinstr irqentry_nmi_exit(struct pt_regs *regs, irqentry_state_t irq_state)
 474 {
 475         instrumentation_begin();
 476         ftrace_nmi_exit();
 477         if (irq_state.lockdep) {
 478                 trace_hardirqs_on_prepare();
 479                 lockdep_hardirqs_on_prepare();
 480         }
 481         instrumentation_end();
 482
 483         ct_nmi_exit();
 484         lockdep_hardirq_exit();
 485         if (irq_state.lockdep)
 486                 lockdep_hardirqs_on(CALLER_ADDR0);
 487         __nmi_exit();
 488 }