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