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