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
 188                 /* Check if any of the above work has queued a deferred wakeup */
 189                 rcu_nocb_flush_deferred_wakeup();
 190
 191                 ti_work = READ_ONCE(current_thread_info()->flags);
 192         }
 193
 194         /* Return the latest work state for arch_exit_to_user_mode() */
 195         return ti_work;
 196 }
 197
 198 static void exit_to_user_mode_prepare(struct pt_regs *regs)
 199 {
 200         unsigned long ti_work = READ_ONCE(current_thread_info()->flags);
 201
 202         lockdep_assert_irqs_disabled();
 203
 204         /* Flush pending rcuog wakeup before the last need_resched() check */
 205         rcu_nocb_flush_deferred_wakeup();
 206
 207         if (unlikely(ti_work & EXIT_TO_USER_MODE_WORK))
 208                 ti_work = exit_to_user_mode_loop(regs, ti_work);
 209
 210         arch_exit_to_user_mode_prepare(regs, ti_work);
 211
 212         /* Ensure that the address limit is intact and no locks are held */
 213         addr_limit_user_check();
 214         kmap_assert_nomap();
 215         lockdep_assert_irqs_disabled();
 216         lockdep_sys_exit();
 217 }
 218
 219 /*
 220  * If SYSCALL_EMU is set, then the only reason to report is when
 221  * SINGLESTEP is set (i.e. PTRACE_SYSEMU_SINGLESTEP).  This syscall
 222  * instruction has been already reported in syscall_enter_from_user_mode().
 223  */
 224 static inline bool report_single_step(unsigned long work)
 225 {
 226         if (work & SYSCALL_WORK_SYSCALL_EMU)
 227                 return false;
 228
 229         return work & SYSCALL_WORK_SYSCALL_EXIT_TRAP;
 230 }
 231
 232 static void syscall_exit_work(struct pt_regs *regs, unsigned long work)
 233 {
 234         bool step;
 235
 236         /*
 237          * If the syscall was rolled back due to syscall user dispatching,
 238          * then the tracers below are not invoked for the same reason as
 239          * the entry side was not invoked in syscall_trace_enter(): The ABI
 240          * of these syscalls is unknown.
 241          */
 242         if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
 243                 if (unlikely(current->syscall_dispatch.on_dispatch)) {
 244                         current->syscall_dispatch.on_dispatch = false;
 245                         return;
 246                 }
 247         }
 248
 249         audit_syscall_exit(regs);
 250
 251         if (work & SYSCALL_WORK_SYSCALL_TRACEPOINT)
 252                 trace_sys_exit(regs, syscall_get_return_value(current, regs));
 253
 254         step = report_single_step(work);
 255         if (step || work & SYSCALL_WORK_SYSCALL_TRACE)
 256                 arch_syscall_exit_tracehook(regs, step);
 257 }
 258
 259 /*
 260  * Syscall specific exit to user mode preparation. Runs with interrupts
 261  * enabled.
 262  */
 263 static void syscall_exit_to_user_mode_prepare(struct pt_regs *regs)
 264 {
 265         unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
 266         unsigned long nr = syscall_get_nr(current, regs);
 267
 268         CT_WARN_ON(ct_state() != CONTEXT_KERNEL);
 269
 270         if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
 271                 if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr))
 272                         local_irq_enable();
 273         }
 274
 275         rseq_syscall(regs);
 276
 277         /*
 278          * Do one-time syscall specific work. If these work items are
 279          * enabled, we want to run them exactly once per syscall exit with
 280          * interrupts enabled.
 281          */
 282         if (unlikely(work & SYSCALL_WORK_EXIT))
 283                 syscall_exit_work(regs, work);
 284 }
 285
 286 static __always_inline void __syscall_exit_to_user_mode_work(struct pt_regs *regs)
 287 {
 288         syscall_exit_to_user_mode_prepare(regs);
 289         local_irq_disable_exit_to_user();
 290         exit_to_user_mode_prepare(regs);
 291 }
 292
 293 void syscall_exit_to_user_mode_work(struct pt_regs *regs)
 294 {
 295         __syscall_exit_to_user_mode_work(regs);
 296 }
 297
 298 __visible noinstr void syscall_exit_to_user_mode(struct pt_regs *regs)
 299 {
 300         instrumentation_begin();
 301         __syscall_exit_to_user_mode_work(regs);
 302         instrumentation_end();
 303         __exit_to_user_mode();
 304 }
 305
 306 noinstr void irqentry_enter_from_user_mode(struct pt_regs *regs)
 307 {
 308         __enter_from_user_mode(regs);
 309 }
 310
 311 noinstr void irqentry_exit_to_user_mode(struct pt_regs *regs)
 312 {
 313         instrumentation_begin();
 314         exit_to_user_mode_prepare(regs);
 315         instrumentation_end();
 316         __exit_to_user_mode();
 317 }
 318
 319 noinstr irqentry_state_t irqentry_enter(struct pt_regs *regs)
 320 {
 321         irqentry_state_t ret = {
 322                 .exit_rcu = false,
 323         };
 324
 325         if (user_mode(regs)) {
 326                 irqentry_enter_from_user_mode(regs);
 327                 return ret;
 328         }
 329
 330         /*
 331          * If this entry hit the idle task invoke rcu_irq_enter() whether
 332          * RCU is watching or not.
 333          *
 334          * Interrupts can nest when the first interrupt invokes softirq
 335          * processing on return which enables interrupts.
 336          *
 337          * Scheduler ticks in the idle task can mark quiescent state and
 338          * terminate a grace period, if and only if the timer interrupt is
 339          * not nested into another interrupt.
 340          *
 341          * Checking for rcu_is_watching() here would prevent the nesting
 342          * interrupt to invoke rcu_irq_enter(). If that nested interrupt is
 343          * the tick then rcu_flavor_sched_clock_irq() would wrongfully
 344          * assume that it is the first interrupt and eventually claim
 345          * quiescent state and end grace periods prematurely.
 346          *
 347          * Unconditionally invoke rcu_irq_enter() so RCU state stays
 348          * consistent.
 349          *
 350          * TINY_RCU does not support EQS, so let the compiler eliminate
 351          * this part when enabled.
 352          */
 353         if (!IS_ENABLED(CONFIG_TINY_RCU) && is_idle_task(current)) {
 354                 /*
 355                  * If RCU is not watching then the same careful
 356                  * sequence vs. lockdep and tracing is required
 357                  * as in irqentry_enter_from_user_mode().
 358                  */
 359                 lockdep_hardirqs_off(CALLER_ADDR0);
 360                 rcu_irq_enter();
 361                 instrumentation_begin();
 362                 trace_hardirqs_off_finish();
 363                 instrumentation_end();
 364
 365                 ret.exit_rcu = true;
 366                 return ret;
 367         }
 368
 369         /*
 370          * If RCU is watching then RCU only wants to check whether it needs
 371          * to restart the tick in NOHZ mode. rcu_irq_enter_check_tick()
 372          * already contains a warning when RCU is not watching, so no point
 373          * in having another one here.
 374          */
 375         lockdep_hardirqs_off(CALLER_ADDR0);
 376         instrumentation_begin();
 377         rcu_irq_enter_check_tick();
 378         trace_hardirqs_off_finish();
 379         instrumentation_end();
 380
 381         return ret;
 382 }
 383
 384 void irqentry_exit_cond_resched(void)
 385 {
 386         if (!preempt_count()) {
 387                 /* Sanity check RCU and thread stack */
 388                 rcu_irq_exit_check_preempt();
 389                 if (IS_ENABLED(CONFIG_DEBUG_ENTRY))
 390                         WARN_ON_ONCE(!on_thread_stack());
 391                 if (need_resched())
 392                         preempt_schedule_irq();
 393         }
 394 }
 395 #ifdef CONFIG_PREEMPT_DYNAMIC
 396 DEFINE_STATIC_CALL(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
 397 #endif
 398
 399 noinstr void irqentry_exit(struct pt_regs *regs, irqentry_state_t state)
 400 {
 401         lockdep_assert_irqs_disabled();
 402
 403         /* Check whether this returns to user mode */
 404         if (user_mode(regs)) {
 405                 irqentry_exit_to_user_mode(regs);
 406         } else if (!regs_irqs_disabled(regs)) {
 407                 /*
 408                  * If RCU was not watching on entry this needs to be done
 409                  * carefully and needs the same ordering of lockdep/tracing
 410                  * and RCU as the return to user mode path.
 411                  */
 412                 if (state.exit_rcu) {
 413                         instrumentation_begin();
 414                         /* Tell the tracer that IRET will enable interrupts */
 415                         trace_hardirqs_on_prepare();
 416                         lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 417                         instrumentation_end();
 418                         rcu_irq_exit();
 419                         lockdep_hardirqs_on(CALLER_ADDR0);
 420                         return;
 421                 }
 422
 423                 instrumentation_begin();
 424                 if (IS_ENABLED(CONFIG_PREEMPTION)) {
 425 #ifdef CONFIG_PREEMPT_DYNAMIC
 426                         static_call(irqentry_exit_cond_resched)();
 427 #else
 428                         irqentry_exit_cond_resched();
 429 #endif
 430                 }
 431                 /* Covers both tracing and lockdep */
 432                 trace_hardirqs_on();
 433                 instrumentation_end();
 434         } else {
 435                 /*
 436                  * IRQ flags state is correct already. Just tell RCU if it
 437                  * was not watching on entry.
 438                  */
 439                 if (state.exit_rcu)
 440                         rcu_irq_exit();
 441         }
 442 }
 443
 444 irqentry_state_t noinstr irqentry_nmi_enter(struct pt_regs *regs)
 445 {
 446         irqentry_state_t irq_state;
 447
 448         irq_state.lockdep = lockdep_hardirqs_enabled();
 449
 450         __nmi_enter();
 451         lockdep_hardirqs_off(CALLER_ADDR0);
 452         lockdep_hardirq_enter();
 453         rcu_nmi_enter();
 454
 455         instrumentation_begin();
 456         trace_hardirqs_off_finish();
 457         ftrace_nmi_enter();
 458         instrumentation_end();
 459
 460         return irq_state;
 461 }
 462
 463 void noinstr irqentry_nmi_exit(struct pt_regs *regs, irqentry_state_t irq_state)
 464 {
 465         instrumentation_begin();
 466         ftrace_nmi_exit();
 467         if (irq_state.lockdep) {
 468                 trace_hardirqs_on_prepare();
 469                 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 470         }
 471         instrumentation_end();
 472
 473         rcu_nmi_exit();
 474         lockdep_hardirq_exit();
 475         if (irq_state.lockdep)
 476                 lockdep_hardirqs_on(CALLER_ADDR0);
 477         __nmi_exit();
 478 }