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         syscall_enter_audit(regs, syscall);
  84
  85         return ret ? : syscall;
  86 }
  87
  88 static __always_inline long
  89 __syscall_enter_from_user_work(struct pt_regs *regs, long syscall)
  90 {
  91         unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
  92
  93         if (work & SYSCALL_WORK_ENTER)
  94                 syscall = syscall_trace_enter(regs, syscall, work);
  95
  96         return syscall;
  97 }
  98
  99 long syscall_enter_from_user_mode_work(struct pt_regs *regs, long syscall)
 100 {
 101         return __syscall_enter_from_user_work(regs, syscall);
 102 }
 103
 104 noinstr long syscall_enter_from_user_mode(struct pt_regs *regs, long syscall)
 105 {
 106         long ret;
 107
 108         __enter_from_user_mode(regs);
 109
 110         instrumentation_begin();
 111         local_irq_enable();
 112         ret = __syscall_enter_from_user_work(regs, syscall);
 113         instrumentation_end();
 114
 115         return ret;
 116 }
 117
 118 noinstr void syscall_enter_from_user_mode_prepare(struct pt_regs *regs)
 119 {
 120         __enter_from_user_mode(regs);
 121         instrumentation_begin();
 122         local_irq_enable();
 123         instrumentation_end();
 124 }
 125
 126 /* See comment for exit_to_user_mode() in entry-common.h */
 127 static __always_inline void __exit_to_user_mode(void)
 128 {
 129         instrumentation_begin();
 130         trace_hardirqs_on_prepare();
 131         lockdep_hardirqs_on_prepare();
 132         instrumentation_end();
 133
 134         user_enter_irqoff();
 135         arch_exit_to_user_mode();
 136         lockdep_hardirqs_on(CALLER_ADDR0);
 137 }
 138
 139 void noinstr exit_to_user_mode(void)
 140 {
 141         __exit_to_user_mode();
 142 }
 143
 144 /* Workaround to allow gradual conversion of architecture code */
 145 void __weak arch_do_signal_or_restart(struct pt_regs *regs) { }
 146
 147 static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
 148                                             unsigned long ti_work)
 149 {
 150         /*
 151          * Before returning to user space ensure that all pending work
 152          * items have been completed.
 153          */
 154         while (ti_work & EXIT_TO_USER_MODE_WORK) {
 155
 156                 local_irq_enable_exit_to_user(ti_work);
 157
 158                 if (ti_work & _TIF_NEED_RESCHED)
 159                         schedule();
 160
 161                 if (ti_work & _TIF_UPROBE)
 162                         uprobe_notify_resume(regs);
 163
 164                 if (ti_work & _TIF_PATCH_PENDING)
 165                         klp_update_patch_state(current);
 166
 167                 if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
 168                         arch_do_signal_or_restart(regs);
 169
 170                 if (ti_work & _TIF_NOTIFY_RESUME)
 171                         resume_user_mode_work(regs);
 172
 173                 /* Architecture specific TIF work */
 174                 arch_exit_to_user_mode_work(regs, ti_work);
 175
 176                 /*
 177                  * Disable interrupts and reevaluate the work flags as they
 178                  * might have changed while interrupts and preemption was
 179                  * enabled above.
 180                  */
 181                 local_irq_disable_exit_to_user();
 182
 183                 /* Check if any of the above work has queued a deferred wakeup */
 184                 tick_nohz_user_enter_prepare();
 185
 186                 ti_work = read_thread_flags();
 187         }
 188
 189         /* Return the latest work state for arch_exit_to_user_mode() */
 190         return ti_work;
 191 }
 192
 193 static void exit_to_user_mode_prepare(struct pt_regs *regs)
 194 {
 195         unsigned long ti_work;
 196
 197         lockdep_assert_irqs_disabled();
 198
 199         /* Flush pending rcuog wakeup before the last need_resched() check */
 200         tick_nohz_user_enter_prepare();
 201
 202         ti_work = read_thread_flags();
 203         if (unlikely(ti_work & EXIT_TO_USER_MODE_WORK))
 204                 ti_work = exit_to_user_mode_loop(regs, ti_work);
 205
 206         arch_exit_to_user_mode_prepare(regs, ti_work);
 207
 208         /* Ensure that the address limit is intact and no locks are held */
 209         addr_limit_user_check();
 210         kmap_assert_nomap();
 211         lockdep_assert_irqs_disabled();
 212         lockdep_sys_exit();
 213 }
 214
 215 /*
 216  * If SYSCALL_EMU is set, then the only reason to report is when
 217  * SINGLESTEP is set (i.e. PTRACE_SYSEMU_SINGLESTEP).  This syscall
 218  * instruction has been already reported in syscall_enter_from_user_mode().
 219  */
 220 static inline bool report_single_step(unsigned long work)
 221 {
 222         if (work & SYSCALL_WORK_SYSCALL_EMU)
 223                 return false;
 224
 225         return work & SYSCALL_WORK_SYSCALL_EXIT_TRAP;
 226 }
 227
 228 static void syscall_exit_work(struct pt_regs *regs, unsigned long work)
 229 {
 230         bool step;
 231
 232         /*
 233          * If the syscall was rolled back due to syscall user dispatching,
 234          * then the tracers below are not invoked for the same reason as
 235          * the entry side was not invoked in syscall_trace_enter(): The ABI
 236          * of these syscalls is unknown.
 237          */
 238         if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
 239                 if (unlikely(current->syscall_dispatch.on_dispatch)) {
 240                         current->syscall_dispatch.on_dispatch = false;
 241                         return;
 242                 }
 243         }
 244
 245         audit_syscall_exit(regs);
 246
 247         if (work & SYSCALL_WORK_SYSCALL_TRACEPOINT)
 248                 trace_sys_exit(regs, syscall_get_return_value(current, regs));
 249
 250         step = report_single_step(work);
 251         if (step || work & SYSCALL_WORK_SYSCALL_TRACE)
 252                 ptrace_report_syscall_exit(regs, step);
 253 }
 254
 255 /*
 256  * Syscall specific exit to user mode preparation. Runs with interrupts
 257  * enabled.
 258  */
 259 static void syscall_exit_to_user_mode_prepare(struct pt_regs *regs)
 260 {
 261         unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
 262         unsigned long nr = syscall_get_nr(current, regs);
 263
 264         CT_WARN_ON(ct_state() != CONTEXT_KERNEL);
 265
 266         if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
 267                 if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr))
 268                         local_irq_enable();
 269         }
 270
 271         rseq_syscall(regs);
 272
 273         /*
 274          * Do one-time syscall specific work. If these work items are
 275          * enabled, we want to run them exactly once per syscall exit with
 276          * interrupts enabled.
 277          */
 278         if (unlikely(work & SYSCALL_WORK_EXIT))
 279                 syscall_exit_work(regs, work);
 280 }
 281
 282 static __always_inline void __syscall_exit_to_user_mode_work(struct pt_regs *regs)
 283 {
 284         syscall_exit_to_user_mode_prepare(regs);
 285         local_irq_disable_exit_to_user();
 286         exit_to_user_mode_prepare(regs);
 287 }
 288
 289 void syscall_exit_to_user_mode_work(struct pt_regs *regs)
 290 {
 291         __syscall_exit_to_user_mode_work(regs);
 292 }
 293
 294 __visible noinstr void syscall_exit_to_user_mode(struct pt_regs *regs)
 295 {
 296         instrumentation_begin();
 297         __syscall_exit_to_user_mode_work(regs);
 298         instrumentation_end();
 299         __exit_to_user_mode();
 300 }
 301
 302 noinstr void irqentry_enter_from_user_mode(struct pt_regs *regs)
 303 {
 304         __enter_from_user_mode(regs);
 305 }
 306
 307 noinstr void irqentry_exit_to_user_mode(struct pt_regs *regs)
 308 {
 309         instrumentation_begin();
 310         exit_to_user_mode_prepare(regs);
 311         instrumentation_end();
 312         __exit_to_user_mode();
 313 }
 314
 315 noinstr irqentry_state_t irqentry_enter(struct pt_regs *regs)
 316 {
 317         irqentry_state_t ret = {
 318                 .exit_rcu = false,
 319         };
 320
 321         if (user_mode(regs)) {
 322                 irqentry_enter_from_user_mode(regs);
 323                 return ret;
 324         }
 325
 326         /*
 327          * If this entry hit the idle task invoke ct_irq_enter() whether
 328          * RCU is watching or not.
 329          *
 330          * Interrupts can nest when the first interrupt invokes softirq
 331          * processing on return which enables interrupts.
 332          *
 333          * Scheduler ticks in the idle task can mark quiescent state and
 334          * terminate a grace period, if and only if the timer interrupt is
 335          * not nested into another interrupt.
 336          *
 337          * Checking for rcu_is_watching() here would prevent the nesting
 338          * interrupt to invoke ct_irq_enter(). If that nested interrupt is
 339          * the tick then rcu_flavor_sched_clock_irq() would wrongfully
 340          * assume that it is the first interrupt and eventually claim
 341          * quiescent state and end grace periods prematurely.
 342          *
 343          * Unconditionally invoke ct_irq_enter() so RCU state stays
 344          * consistent.
 345          *
 346          * TINY_RCU does not support EQS, so let the compiler eliminate
 347          * this part when enabled.
 348          */
 349         if (!IS_ENABLED(CONFIG_TINY_RCU) && is_idle_task(current)) {
 350                 /*
 351                  * If RCU is not watching then the same careful
 352                  * sequence vs. lockdep and tracing is required
 353                  * as in irqentry_enter_from_user_mode().
 354                  */
 355                 lockdep_hardirqs_off(CALLER_ADDR0);
 356                 ct_irq_enter();
 357                 instrumentation_begin();
 358                 kmsan_unpoison_entry_regs(regs);
 359                 trace_hardirqs_off_finish();
 360                 instrumentation_end();
 361
 362                 ret.exit_rcu = true;
 363                 return ret;
 364         }
 365
 366         /*
 367          * If RCU is watching then RCU only wants to check whether it needs
 368          * to restart the tick in NOHZ mode. rcu_irq_enter_check_tick()
 369          * already contains a warning when RCU is not watching, so no point
 370          * in having another one here.
 371          */
 372         lockdep_hardirqs_off(CALLER_ADDR0);
 373         instrumentation_begin();
 374         kmsan_unpoison_entry_regs(regs);
 375         rcu_irq_enter_check_tick();
 376         trace_hardirqs_off_finish();
 377         instrumentation_end();
 378
 379         return ret;
 380 }
 381
 382 void raw_irqentry_exit_cond_resched(void)
 383 {
 384         if (!preempt_count()) {
 385                 /* Sanity check RCU and thread stack */
 386                 rcu_irq_exit_check_preempt();
 387                 if (IS_ENABLED(CONFIG_DEBUG_ENTRY))
 388                         WARN_ON_ONCE(!on_thread_stack());
 389                 if (need_resched())
 390                         preempt_schedule_irq();
 391         }
 392 }
 393 #ifdef CONFIG_PREEMPT_DYNAMIC
 394 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
 395 DEFINE_STATIC_CALL(irqentry_exit_cond_resched, raw_irqentry_exit_cond_resched);
 396 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
 397 DEFINE_STATIC_KEY_TRUE(sk_dynamic_irqentry_exit_cond_resched);
 398 void dynamic_irqentry_exit_cond_resched(void)
 399 {
 400         if (!static_branch_unlikely(&sk_dynamic_irqentry_exit_cond_resched))
 401                 return;
 402         raw_irqentry_exit_cond_resched();
 403 }
 404 #endif
 405 #endif
 406
 407 noinstr void irqentry_exit(struct pt_regs *regs, irqentry_state_t state)
 408 {
 409         lockdep_assert_irqs_disabled();
 410
 411         /* Check whether this returns to user mode */
 412         if (user_mode(regs)) {
 413                 irqentry_exit_to_user_mode(regs);
 414         } else if (!regs_irqs_disabled(regs)) {
 415                 /*
 416                  * If RCU was not watching on entry this needs to be done
 417                  * carefully and needs the same ordering of lockdep/tracing
 418                  * and RCU as the return to user mode path.
 419                  */
 420                 if (state.exit_rcu) {
 421                         instrumentation_begin();
 422                         /* Tell the tracer that IRET will enable interrupts */
 423                         trace_hardirqs_on_prepare();
 424                         lockdep_hardirqs_on_prepare();
 425                         instrumentation_end();
 426                         ct_irq_exit();
 427                         lockdep_hardirqs_on(CALLER_ADDR0);
 428                         return;
 429                 }
 430
 431                 instrumentation_begin();
 432                 if (IS_ENABLED(CONFIG_PREEMPTION))
 433                         irqentry_exit_cond_resched();
 434
 435                 /* Covers both tracing and lockdep */
 436                 trace_hardirqs_on();
 437                 instrumentation_end();
 438         } else {
 439                 /*
 440                  * IRQ flags state is correct already. Just tell RCU if it
 441                  * was not watching on entry.
 442                  */
 443                 if (state.exit_rcu)
 444                         ct_irq_exit();
 445         }
 446 }
 447
 448 irqentry_state_t noinstr irqentry_nmi_enter(struct pt_regs *regs)
 449 {
 450         irqentry_state_t irq_state;
 451
 452         irq_state.lockdep = lockdep_hardirqs_enabled();
 453
 454         __nmi_enter();
 455         lockdep_hardirqs_off(CALLER_ADDR0);
 456         lockdep_hardirq_enter();
 457         ct_nmi_enter();
 458
 459         instrumentation_begin();
 460         kmsan_unpoison_entry_regs(regs);
 461         trace_hardirqs_off_finish();
 462         ftrace_nmi_enter();
 463         instrumentation_end();
 464
 465         return irq_state;
 466 }
 467
 468 void noinstr irqentry_nmi_exit(struct pt_regs *regs, irqentry_state_t irq_state)
 469 {
 470         instrumentation_begin();
 471         ftrace_nmi_exit();
 472         if (irq_state.lockdep) {
 473                 trace_hardirqs_on_prepare();
 474                 lockdep_hardirqs_on_prepare();
 475         }
 476         instrumentation_end();
 477
 478         ct_nmi_exit();
 479         lockdep_hardirq_exit();
 480         if (irq_state.lockdep)
 481                 lockdep_hardirqs_on(CALLER_ADDR0);
 482         __nmi_exit();
 483 }