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)
  46 {
  47         long ret = 0;
  48
  49         /* Handle ptrace */
  50         if (ti_work & (_TIF_SYSCALL_TRACE | _TIF_SYSCALL_EMU)) {
  51                 ret = arch_syscall_enter_tracehook(regs);
  52                 if (ret || (ti_work & _TIF_SYSCALL_EMU))
  53                         return -1L;
  54         }
  55
  56         /* Do seccomp after ptrace, to catch any tracer changes. */
  57         if (ti_work & _TIF_SECCOMP) {
  58                 ret = __secure_computing(NULL);
  59                 if (ret == -1L)
  60                         return ret;
  61         }
  62
  63         if (unlikely(ti_work & _TIF_SYSCALL_TRACEPOINT))
  64                 trace_sys_enter(regs, syscall);
  65
  66         syscall_enter_audit(regs, syscall);
  67
  68         return ret ? : syscall;
  69 }
  70
  71 noinstr long syscall_enter_from_user_mode(struct pt_regs *regs, long syscall)
  72 {
  73         unsigned long ti_work;
  74
  75         enter_from_user_mode(regs);
  76         instrumentation_begin();
  77
  78         local_irq_enable();
  79         ti_work = READ_ONCE(current_thread_info()->flags);
  80         if (ti_work & SYSCALL_ENTER_WORK)
  81                 syscall = syscall_trace_enter(regs, syscall, ti_work);
  82         instrumentation_end();
  83
  84         return syscall;
  85 }
  86
  87 /**
  88  * exit_to_user_mode - Fixup state when exiting to user mode
  89  *
  90  * Syscall/interupt exit enables interrupts, but the kernel state is
  91  * interrupts disabled when this is invoked. Also tell RCU about it.
  92  *
  93  * 1) Trace interrupts on state
  94  * 2) Invoke context tracking if enabled to adjust RCU state
  95  * 3) Invoke architecture specific last minute exit code, e.g. speculation
  96  *    mitigations, etc.
  97  * 4) Tell lockdep that interrupts are enabled
  98  */
  99 static __always_inline void exit_to_user_mode(void)
 100 {
 101         instrumentation_begin();
 102         trace_hardirqs_on_prepare();
 103         lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 104         instrumentation_end();
 105
 106         user_enter_irqoff();
 107         arch_exit_to_user_mode();
 108         lockdep_hardirqs_on(CALLER_ADDR0);
 109 }
 110
 111 /* Workaround to allow gradual conversion of architecture code */
 112 void __weak arch_do_signal_or_restart(struct pt_regs *regs, bool has_signal) { }
 113
 114 static void handle_signal_work(struct pt_regs *regs, unsigned long ti_work)
 115 {
 116         if (ti_work & _TIF_NOTIFY_SIGNAL)
 117                 tracehook_notify_signal();
 118
 119         arch_do_signal_or_restart(regs, ti_work & _TIF_SIGPENDING);
 120 }
 121
 122 static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
 123                                             unsigned long ti_work)
 124 {
 125         /*
 126          * Before returning to user space ensure that all pending work
 127          * items have been completed.
 128          */
 129         while (ti_work & EXIT_TO_USER_MODE_WORK) {
 130
 131                 local_irq_enable_exit_to_user(ti_work);
 132
 133                 if (ti_work & _TIF_NEED_RESCHED)
 134                         schedule();
 135
 136                 if (ti_work & _TIF_UPROBE)
 137                         uprobe_notify_resume(regs);
 138
 139                 if (ti_work & _TIF_PATCH_PENDING)
 140                         klp_update_patch_state(current);
 141
 142                 if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
 143                         handle_signal_work(regs, ti_work);
 144
 145                 if (ti_work & _TIF_NOTIFY_RESUME) {
 146                         clear_thread_flag(TIF_NOTIFY_RESUME);
 147                         tracehook_notify_resume(regs);
 148                         rseq_handle_notify_resume(NULL, regs);
 149                 }
 150
 151                 /* Architecture specific TIF work */
 152                 arch_exit_to_user_mode_work(regs, ti_work);
 153
 154                 /*
 155                  * Disable interrupts and reevaluate the work flags as they
 156                  * might have changed while interrupts and preemption was
 157                  * enabled above.
 158                  */
 159                 local_irq_disable_exit_to_user();
 160                 ti_work = READ_ONCE(current_thread_info()->flags);
 161         }
 162
 163         /* Return the latest work state for arch_exit_to_user_mode() */
 164         return ti_work;
 165 }
 166
 167 static void exit_to_user_mode_prepare(struct pt_regs *regs)
 168 {
 169         unsigned long ti_work = READ_ONCE(current_thread_info()->flags);
 170
 171         lockdep_assert_irqs_disabled();
 172
 173         if (unlikely(ti_work & EXIT_TO_USER_MODE_WORK))
 174                 ti_work = exit_to_user_mode_loop(regs, ti_work);
 175
 176         arch_exit_to_user_mode_prepare(regs, ti_work);
 177
 178         /* Ensure that the address limit is intact and no locks are held */
 179         addr_limit_user_check();
 180         lockdep_assert_irqs_disabled();
 181         lockdep_sys_exit();
 182 }
 183
 184 #ifndef _TIF_SINGLESTEP
 185 static inline bool report_single_step(unsigned long ti_work)
 186 {
 187         return false;
 188 }
 189 #else
 190 /*
 191  * If TIF_SYSCALL_EMU is set, then the only reason to report is when
 192  * TIF_SINGLESTEP is set (i.e. PTRACE_SYSEMU_SINGLESTEP).  This syscall
 193  * instruction has been already reported in syscall_enter_from_usermode().
 194  */
 195 #define SYSEMU_STEP     (_TIF_SINGLESTEP | _TIF_SYSCALL_EMU)
 196
 197 static inline bool report_single_step(unsigned long ti_work)
 198 {
 199         return (ti_work & SYSEMU_STEP) == _TIF_SINGLESTEP;
 200 }
 201 #endif
 202
 203 static void syscall_exit_work(struct pt_regs *regs, unsigned long ti_work)
 204 {
 205         bool step;
 206
 207         audit_syscall_exit(regs);
 208
 209         if (ti_work & _TIF_SYSCALL_TRACEPOINT)
 210                 trace_sys_exit(regs, syscall_get_return_value(current, regs));
 211
 212         step = report_single_step(ti_work);
 213         if (step || ti_work & _TIF_SYSCALL_TRACE)
 214                 arch_syscall_exit_tracehook(regs, step);
 215 }
 216
 217 /*
 218  * Syscall specific exit to user mode preparation. Runs with interrupts
 219  * enabled.
 220  */
 221 static void syscall_exit_to_user_mode_prepare(struct pt_regs *regs)
 222 {
 223         u32 cached_flags = READ_ONCE(current_thread_info()->flags);
 224         unsigned long nr = syscall_get_nr(current, regs);
 225
 226         CT_WARN_ON(ct_state() != CONTEXT_KERNEL);
 227
 228         if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
 229                 if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr))
 230                         local_irq_enable();
 231         }
 232
 233         rseq_syscall(regs);
 234
 235         /*
 236          * Do one-time syscall specific work. If these work items are
 237          * enabled, we want to run them exactly once per syscall exit with
 238          * interrupts enabled.
 239          */
 240         if (unlikely(cached_flags & SYSCALL_EXIT_WORK))
 241                 syscall_exit_work(regs, cached_flags);
 242 }
 243
 244 __visible noinstr void syscall_exit_to_user_mode(struct pt_regs *regs)
 245 {
 246         instrumentation_begin();
 247         syscall_exit_to_user_mode_prepare(regs);
 248         local_irq_disable_exit_to_user();
 249         exit_to_user_mode_prepare(regs);
 250         instrumentation_end();
 251         exit_to_user_mode();
 252 }
 253
 254 noinstr void irqentry_enter_from_user_mode(struct pt_regs *regs)
 255 {
 256         enter_from_user_mode(regs);
 257 }
 258
 259 noinstr void irqentry_exit_to_user_mode(struct pt_regs *regs)
 260 {
 261         instrumentation_begin();
 262         exit_to_user_mode_prepare(regs);
 263         instrumentation_end();
 264         exit_to_user_mode();
 265 }
 266
 267 noinstr irqentry_state_t irqentry_enter(struct pt_regs *regs)
 268 {
 269         irqentry_state_t ret = {
 270                 .exit_rcu = false,
 271         };
 272
 273         if (user_mode(regs)) {
 274                 irqentry_enter_from_user_mode(regs);
 275                 return ret;
 276         }
 277
 278         /*
 279          * If this entry hit the idle task invoke rcu_irq_enter() whether
 280          * RCU is watching or not.
 281          *
 282          * Interupts can nest when the first interrupt invokes softirq
 283          * processing on return which enables interrupts.
 284          *
 285          * Scheduler ticks in the idle task can mark quiescent state and
 286          * terminate a grace period, if and only if the timer interrupt is
 287          * not nested into another interrupt.
 288          *
 289          * Checking for __rcu_is_watching() here would prevent the nesting
 290          * interrupt to invoke rcu_irq_enter(). If that nested interrupt is
 291          * the tick then rcu_flavor_sched_clock_irq() would wrongfully
 292          * assume that it is the first interupt and eventually claim
 293          * quiescient state and end grace periods prematurely.
 294          *
 295          * Unconditionally invoke rcu_irq_enter() so RCU state stays
 296          * consistent.
 297          *
 298          * TINY_RCU does not support EQS, so let the compiler eliminate
 299          * this part when enabled.
 300          */
 301         if (!IS_ENABLED(CONFIG_TINY_RCU) && is_idle_task(current)) {
 302                 /*
 303                  * If RCU is not watching then the same careful
 304                  * sequence vs. lockdep and tracing is required
 305                  * as in irq_enter_from_user_mode().
 306                  */
 307                 lockdep_hardirqs_off(CALLER_ADDR0);
 308                 rcu_irq_enter();
 309                 instrumentation_begin();
 310                 trace_hardirqs_off_finish();
 311                 instrumentation_end();
 312
 313                 ret.exit_rcu = true;
 314                 return ret;
 315         }
 316
 317         /*
 318          * If RCU is watching then RCU only wants to check whether it needs
 319          * to restart the tick in NOHZ mode. rcu_irq_enter_check_tick()
 320          * already contains a warning when RCU is not watching, so no point
 321          * in having another one here.
 322          */
 323         instrumentation_begin();
 324         rcu_irq_enter_check_tick();
 325         /* Use the combo lockdep/tracing function */
 326         trace_hardirqs_off();
 327         instrumentation_end();
 328
 329         return ret;
 330 }
 331
 332 void irqentry_exit_cond_resched(void)
 333 {
 334         if (!preempt_count()) {
 335                 /* Sanity check RCU and thread stack */
 336                 rcu_irq_exit_check_preempt();
 337                 if (IS_ENABLED(CONFIG_DEBUG_ENTRY))
 338                         WARN_ON_ONCE(!on_thread_stack());
 339                 if (need_resched())
 340                         preempt_schedule_irq();
 341         }
 342 }
 343
 344 noinstr void irqentry_exit(struct pt_regs *regs, irqentry_state_t state)
 345 {
 346         lockdep_assert_irqs_disabled();
 347
 348         /* Check whether this returns to user mode */
 349         if (user_mode(regs)) {
 350                 irqentry_exit_to_user_mode(regs);
 351         } else if (!regs_irqs_disabled(regs)) {
 352                 /*
 353                  * If RCU was not watching on entry this needs to be done
 354                  * carefully and needs the same ordering of lockdep/tracing
 355                  * and RCU as the return to user mode path.
 356                  */
 357                 if (state.exit_rcu) {
 358                         instrumentation_begin();
 359                         /* Tell the tracer that IRET will enable interrupts */
 360                         trace_hardirqs_on_prepare();
 361                         lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 362                         instrumentation_end();
 363                         rcu_irq_exit();
 364                         lockdep_hardirqs_on(CALLER_ADDR0);
 365                         return;
 366                 }
 367
 368                 instrumentation_begin();
 369                 if (IS_ENABLED(CONFIG_PREEMPTION))
 370                         irqentry_exit_cond_resched();
 371                 /* Covers both tracing and lockdep */
 372                 trace_hardirqs_on();
 373                 instrumentation_end();
 374         } else {
 375                 /*
 376                  * IRQ flags state is correct already. Just tell RCU if it
 377                  * was not watching on entry.
 378                  */
 379                 if (state.exit_rcu)
 380                         rcu_irq_exit();
 381         }
 382 }