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 #define CREATE_TRACE_POINTS
   9 #include <trace/events/syscalls.h>
  10
  11 /**
  12  * enter_from_user_mode - Establish state when coming from user mode
  13  *
  14  * Syscall/interrupt entry disables interrupts, but user mode is traced as
  15  * interrupts enabled. Also with NO_HZ_FULL RCU might be idle.
  16  *
  17  * 1) Tell lockdep that interrupts are disabled
  18  * 2) Invoke context tracking if enabled to reactivate RCU
  19  * 3) Trace interrupts off state
  20  */
  21 static __always_inline void enter_from_user_mode(struct pt_regs *regs)
  22 {
  23         arch_check_user_regs(regs);
  24         lockdep_hardirqs_off(CALLER_ADDR0);
  25
  26         CT_WARN_ON(ct_state() != CONTEXT_USER);
  27         user_exit_irqoff();
  28
  29         instrumentation_begin();
  30         trace_hardirqs_off_finish();
  31         instrumentation_end();
  32 }
  33
  34 static inline void syscall_enter_audit(struct pt_regs *regs, long syscall)
  35 {
  36         if (unlikely(audit_context())) {
  37                 unsigned long args[6];
  38
  39                 syscall_get_arguments(current, regs, args);
  40                 audit_syscall_entry(syscall, args[0], args[1], args[2], args[3]);
  41         }
  42 }
  43
  44 static long syscall_trace_enter(struct pt_regs *regs, long syscall,
  45                                 unsigned long ti_work, unsigned long work)
  46 {
  47         long ret = 0;
  48
  49         /* Handle ptrace */
  50         if (work & (SYSCALL_WORK_SYSCALL_TRACE | SYSCALL_WORK_SYSCALL_EMU)) {
  51                 ret = arch_syscall_enter_tracehook(regs);
  52                 if (ret || (work & SYSCALL_WORK_SYSCALL_EMU))
  53                         return -1L;
  54         }
  55
  56         /* Do seccomp after ptrace, to catch any tracer changes. */
  57         if (work & SYSCALL_WORK_SECCOMP) {
  58                 ret = __secure_computing(NULL);
  59                 if (ret == -1L)
  60                         return ret;
  61         }
  62
  63         /* Either of the above might have changed the syscall number */
  64         syscall = syscall_get_nr(current, regs);
  65
  66         if (unlikely(work & SYSCALL_WORK_SYSCALL_TRACEPOINT))
  67                 trace_sys_enter(regs, syscall);
  68
  69         syscall_enter_audit(regs, syscall);
  70
  71         return ret ? : syscall;
  72 }
  73
  74 static __always_inline long
  75 __syscall_enter_from_user_work(struct pt_regs *regs, long syscall)
  76 {
  77         unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
  78         unsigned long ti_work;
  79
  80         ti_work = READ_ONCE(current_thread_info()->flags);
  81         if (work & SYSCALL_WORK_ENTER || ti_work & SYSCALL_ENTER_WORK)
  82                 syscall = syscall_trace_enter(regs, syscall, ti_work, work);
  83
  84         return syscall;
  85 }
  86
  87 long syscall_enter_from_user_mode_work(struct pt_regs *regs, long syscall)
  88 {
  89         return __syscall_enter_from_user_work(regs, syscall);
  90 }
  91
  92 noinstr long syscall_enter_from_user_mode(struct pt_regs *regs, long syscall)
  93 {
  94         long ret;
  95
  96         enter_from_user_mode(regs);
  97
  98         instrumentation_begin();
  99         local_irq_enable();
 100         ret = __syscall_enter_from_user_work(regs, syscall);
 101         instrumentation_end();
 102
 103         return ret;
 104 }
 105
 106 noinstr void syscall_enter_from_user_mode_prepare(struct pt_regs *regs)
 107 {
 108         enter_from_user_mode(regs);
 109         instrumentation_begin();
 110         local_irq_enable();
 111         instrumentation_end();
 112 }
 113
 114 /**
 115  * exit_to_user_mode - Fixup state when exiting to user mode
 116  *
 117  * Syscall/interupt exit enables interrupts, but the kernel state is
 118  * interrupts disabled when this is invoked. Also tell RCU about it.
 119  *
 120  * 1) Trace interrupts on state
 121  * 2) Invoke context tracking if enabled to adjust RCU state
 122  * 3) Invoke architecture specific last minute exit code, e.g. speculation
 123  *    mitigations, etc.
 124  * 4) Tell lockdep that interrupts are enabled
 125  */
 126 static __always_inline void exit_to_user_mode(void)
 127 {
 128         instrumentation_begin();
 129         trace_hardirqs_on_prepare();
 130         lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 131         instrumentation_end();
 132
 133         user_enter_irqoff();
 134         arch_exit_to_user_mode();
 135         lockdep_hardirqs_on(CALLER_ADDR0);
 136 }
 137
 138 /* Workaround to allow gradual conversion of architecture code */
 139 void __weak arch_do_signal_or_restart(struct pt_regs *regs, bool has_signal) { }
 140
 141 static void handle_signal_work(struct pt_regs *regs, unsigned long ti_work)
 142 {
 143         if (ti_work & _TIF_NOTIFY_SIGNAL)
 144                 tracehook_notify_signal();
 145
 146         arch_do_signal_or_restart(regs, ti_work & _TIF_SIGPENDING);
 147 }
 148
 149 static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
 150                                             unsigned long ti_work)
 151 {
 152         /*
 153          * Before returning to user space ensure that all pending work
 154          * items have been completed.
 155          */
 156         while (ti_work & EXIT_TO_USER_MODE_WORK) {
 157
 158                 local_irq_enable_exit_to_user(ti_work);
 159
 160                 if (ti_work & _TIF_NEED_RESCHED)
 161                         schedule();
 162
 163                 if (ti_work & _TIF_UPROBE)
 164                         uprobe_notify_resume(regs);
 165
 166                 if (ti_work & _TIF_PATCH_PENDING)
 167                         klp_update_patch_state(current);
 168
 169                 if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
 170                         handle_signal_work(regs, ti_work);
 171
 172                 if (ti_work & _TIF_NOTIFY_RESUME) {
 173                         tracehook_notify_resume(regs);
 174                         rseq_handle_notify_resume(NULL, regs);
 175                 }
 176
 177                 /* Architecture specific TIF work */
 178                 arch_exit_to_user_mode_work(regs, ti_work);
 179
 180                 /*
 181                  * Disable interrupts and reevaluate the work flags as they
 182                  * might have changed while interrupts and preemption was
 183                  * enabled above.
 184                  */
 185                 local_irq_disable_exit_to_user();
 186                 ti_work = READ_ONCE(current_thread_info()->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 = READ_ONCE(current_thread_info()->flags);
 196
 197         lockdep_assert_irqs_disabled();
 198
 199         if (unlikely(ti_work & EXIT_TO_USER_MODE_WORK))
 200                 ti_work = exit_to_user_mode_loop(regs, ti_work);
 201
 202         arch_exit_to_user_mode_prepare(regs, ti_work);
 203
 204         /* Ensure that the address limit is intact and no locks are held */
 205         addr_limit_user_check();
 206         lockdep_assert_irqs_disabled();
 207         lockdep_sys_exit();
 208 }
 209
 210 #ifndef _TIF_SINGLESTEP
 211 static inline bool report_single_step(unsigned long work)
 212 {
 213         return false;
 214 }
 215 #else
 216 /*
 217  * If SYSCALL_EMU is set, then the only reason to report is when
 218  * TIF_SINGLESTEP is set (i.e. PTRACE_SYSEMU_SINGLESTEP).  This syscall
 219  * instruction has been already reported in syscall_enter_from_user_mode().
 220  */
 221 static inline bool report_single_step(unsigned long work)
 222 {
 223         if (!(work & SYSCALL_WORK_SYSCALL_EMU))
 224                 return false;
 225
 226         return !!(current_thread_info()->flags & _TIF_SINGLESTEP);
 227 }
 228 #endif
 229
 230 static void syscall_exit_work(struct pt_regs *regs, unsigned long ti_work,
 231                               unsigned long work)
 232 {
 233         bool step;
 234
 235         audit_syscall_exit(regs);
 236
 237         if (work & SYSCALL_WORK_SYSCALL_TRACEPOINT)
 238                 trace_sys_exit(regs, syscall_get_return_value(current, regs));
 239
 240         step = report_single_step(work);
 241         if (step || work & SYSCALL_WORK_SYSCALL_TRACE)
 242                 arch_syscall_exit_tracehook(regs, step);
 243 }
 244
 245 /*
 246  * Syscall specific exit to user mode preparation. Runs with interrupts
 247  * enabled.
 248  */
 249 static void syscall_exit_to_user_mode_prepare(struct pt_regs *regs)
 250 {
 251         unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
 252         u32 cached_flags = READ_ONCE(current_thread_info()->flags);
 253         unsigned long nr = syscall_get_nr(current, regs);
 254
 255         CT_WARN_ON(ct_state() != CONTEXT_KERNEL);
 256
 257         if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
 258                 if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr))
 259                         local_irq_enable();
 260         }
 261
 262         rseq_syscall(regs);
 263
 264         /*
 265          * Do one-time syscall specific work. If these work items are
 266          * enabled, we want to run them exactly once per syscall exit with
 267          * interrupts enabled.
 268          */
 269         if (unlikely(work & SYSCALL_WORK_EXIT || cached_flags & SYSCALL_EXIT_WORK))
 270                 syscall_exit_work(regs, cached_flags, work);
 271 }
 272
 273 __visible noinstr void syscall_exit_to_user_mode(struct pt_regs *regs)
 274 {
 275         instrumentation_begin();
 276         syscall_exit_to_user_mode_prepare(regs);
 277         local_irq_disable_exit_to_user();
 278         exit_to_user_mode_prepare(regs);
 279         instrumentation_end();
 280         exit_to_user_mode();
 281 }
 282
 283 noinstr void irqentry_enter_from_user_mode(struct pt_regs *regs)
 284 {
 285         enter_from_user_mode(regs);
 286 }
 287
 288 noinstr void irqentry_exit_to_user_mode(struct pt_regs *regs)
 289 {
 290         instrumentation_begin();
 291         exit_to_user_mode_prepare(regs);
 292         instrumentation_end();
 293         exit_to_user_mode();
 294 }
 295
 296 noinstr irqentry_state_t irqentry_enter(struct pt_regs *regs)
 297 {
 298         irqentry_state_t ret = {
 299                 .exit_rcu = false,
 300         };
 301
 302         if (user_mode(regs)) {
 303                 irqentry_enter_from_user_mode(regs);
 304                 return ret;
 305         }
 306
 307         /*
 308          * If this entry hit the idle task invoke rcu_irq_enter() whether
 309          * RCU is watching or not.
 310          *
 311          * Interrupts can nest when the first interrupt invokes softirq
 312          * processing on return which enables interrupts.
 313          *
 314          * Scheduler ticks in the idle task can mark quiescent state and
 315          * terminate a grace period, if and only if the timer interrupt is
 316          * not nested into another interrupt.
 317          *
 318          * Checking for rcu_is_watching() here would prevent the nesting
 319          * interrupt to invoke rcu_irq_enter(). If that nested interrupt is
 320          * the tick then rcu_flavor_sched_clock_irq() would wrongfully
 321          * assume that it is the first interupt and eventually claim
 322          * quiescent state and end grace periods prematurely.
 323          *
 324          * Unconditionally invoke rcu_irq_enter() so RCU state stays
 325          * consistent.
 326          *
 327          * TINY_RCU does not support EQS, so let the compiler eliminate
 328          * this part when enabled.
 329          */
 330         if (!IS_ENABLED(CONFIG_TINY_RCU) && is_idle_task(current)) {
 331                 /*
 332                  * If RCU is not watching then the same careful
 333                  * sequence vs. lockdep and tracing is required
 334                  * as in irqentry_enter_from_user_mode().
 335                  */
 336                 lockdep_hardirqs_off(CALLER_ADDR0);
 337                 rcu_irq_enter();
 338                 instrumentation_begin();
 339                 trace_hardirqs_off_finish();
 340                 instrumentation_end();
 341
 342                 ret.exit_rcu = true;
 343                 return ret;
 344         }
 345
 346         /*
 347          * If RCU is watching then RCU only wants to check whether it needs
 348          * to restart the tick in NOHZ mode. rcu_irq_enter_check_tick()
 349          * already contains a warning when RCU is not watching, so no point
 350          * in having another one here.
 351          */
 352         lockdep_hardirqs_off(CALLER_ADDR0);
 353         instrumentation_begin();
 354         rcu_irq_enter_check_tick();
 355         trace_hardirqs_off_finish();
 356         instrumentation_end();
 357
 358         return ret;
 359 }
 360
 361 void irqentry_exit_cond_resched(void)
 362 {
 363         if (!preempt_count()) {
 364                 /* Sanity check RCU and thread stack */
 365                 rcu_irq_exit_check_preempt();
 366                 if (IS_ENABLED(CONFIG_DEBUG_ENTRY))
 367                         WARN_ON_ONCE(!on_thread_stack());
 368                 if (need_resched())
 369                         preempt_schedule_irq();
 370         }
 371 }
 372
 373 noinstr void irqentry_exit(struct pt_regs *regs, irqentry_state_t state)
 374 {
 375         lockdep_assert_irqs_disabled();
 376
 377         /* Check whether this returns to user mode */
 378         if (user_mode(regs)) {
 379                 irqentry_exit_to_user_mode(regs);
 380         } else if (!regs_irqs_disabled(regs)) {
 381                 /*
 382                  * If RCU was not watching on entry this needs to be done
 383                  * carefully and needs the same ordering of lockdep/tracing
 384                  * and RCU as the return to user mode path.
 385                  */
 386                 if (state.exit_rcu) {
 387                         instrumentation_begin();
 388                         /* Tell the tracer that IRET will enable interrupts */
 389                         trace_hardirqs_on_prepare();
 390                         lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 391                         instrumentation_end();
 392                         rcu_irq_exit();
 393                         lockdep_hardirqs_on(CALLER_ADDR0);
 394                         return;
 395                 }
 396
 397                 instrumentation_begin();
 398                 if (IS_ENABLED(CONFIG_PREEMPTION))
 399                         irqentry_exit_cond_resched();
 400                 /* Covers both tracing and lockdep */
 401                 trace_hardirqs_on();
 402                 instrumentation_end();
 403         } else {
 404                 /*
 405                  * IRQ flags state is correct already. Just tell RCU if it
 406                  * was not watching on entry.
 407                  */
 408                 if (state.exit_rcu)
 409                         rcu_irq_exit();
 410         }
 411 }
 412
 413 irqentry_state_t noinstr irqentry_nmi_enter(struct pt_regs *regs)
 414 {
 415         irqentry_state_t irq_state;
 416
 417         irq_state.lockdep = lockdep_hardirqs_enabled();
 418
 419         __nmi_enter();
 420         lockdep_hardirqs_off(CALLER_ADDR0);
 421         lockdep_hardirq_enter();
 422         rcu_nmi_enter();
 423
 424         instrumentation_begin();
 425         trace_hardirqs_off_finish();
 426         ftrace_nmi_enter();
 427         instrumentation_end();
 428
 429         return irq_state;
 430 }
 431
 432 void noinstr irqentry_nmi_exit(struct pt_regs *regs, irqentry_state_t irq_state)
 433 {
 434         instrumentation_begin();
 435         ftrace_nmi_exit();
 436         if (irq_state.lockdep) {
 437                 trace_hardirqs_on_prepare();
 438                 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 439         }
 440         instrumentation_end();
 441
 442         rcu_nmi_exit();
 443         lockdep_hardirq_exit();
 444         if (irq_state.lockdep)
 445                 lockdep_hardirqs_on(CALLER_ADDR0);
 446         __nmi_exit();
 447 }