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         /* Either of the above might have changed the syscall number */
  64         syscall = syscall_get_nr(current, regs);
  65
  66         if (unlikely(ti_work & _TIF_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 ti_work;
  78
  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
  83         return syscall;
  84 }
  85
  86 long syscall_enter_from_user_mode_work(struct pt_regs *regs, long syscall)
  87 {
  88         return __syscall_enter_from_user_work(regs, syscall);
  89 }
  90
  91 noinstr long syscall_enter_from_user_mode(struct pt_regs *regs, long syscall)
  92 {
  93         long ret;
  94
  95         enter_from_user_mode(regs);
  96
  97         instrumentation_begin();
  98         local_irq_enable();
  99         ret = __syscall_enter_from_user_work(regs, syscall);
 100         instrumentation_end();
 101
 102         return ret;
 103 }
 104
 105 noinstr void syscall_enter_from_user_mode_prepare(struct pt_regs *regs)
 106 {
 107         enter_from_user_mode(regs);
 108         instrumentation_begin();
 109         local_irq_enable();
 110         instrumentation_end();
 111 }
 112
 113 /**
 114  * exit_to_user_mode - Fixup state when exiting to user mode
 115  *
 116  * Syscall/interupt exit enables interrupts, but the kernel state is
 117  * interrupts disabled when this is invoked. Also tell RCU about it.
 118  *
 119  * 1) Trace interrupts on state
 120  * 2) Invoke context tracking if enabled to adjust RCU state
 121  * 3) Invoke architecture specific last minute exit code, e.g. speculation
 122  *    mitigations, etc.
 123  * 4) Tell lockdep that interrupts are enabled
 124  */
 125 static __always_inline void exit_to_user_mode(void)
 126 {
 127         instrumentation_begin();
 128         trace_hardirqs_on_prepare();
 129         lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 130         instrumentation_end();
 131
 132         user_enter_irqoff();
 133         arch_exit_to_user_mode();
 134         lockdep_hardirqs_on(CALLER_ADDR0);
 135 }
 136
 137 /* Workaround to allow gradual conversion of architecture code */
 138 void __weak arch_do_signal(struct pt_regs *regs) { }
 139
 140 static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
 141                                             unsigned long ti_work)
 142 {
 143         /*
 144          * Before returning to user space ensure that all pending work
 145          * items have been completed.
 146          */
 147         while (ti_work & EXIT_TO_USER_MODE_WORK) {
 148
 149                 local_irq_enable_exit_to_user(ti_work);
 150
 151                 if (ti_work & _TIF_NEED_RESCHED)
 152                         schedule();
 153
 154                 if (ti_work & _TIF_UPROBE)
 155                         uprobe_notify_resume(regs);
 156
 157                 if (ti_work & _TIF_PATCH_PENDING)
 158                         klp_update_patch_state(current);
 159
 160                 if (ti_work & _TIF_SIGPENDING)
 161                         arch_do_signal(regs);
 162
 163                 if (ti_work & _TIF_NOTIFY_RESUME) {
 164                         clear_thread_flag(TIF_NOTIFY_RESUME);
 165                         tracehook_notify_resume(regs);
 166                         rseq_handle_notify_resume(NULL, regs);
 167                 }
 168
 169                 /* Architecture specific TIF work */
 170                 arch_exit_to_user_mode_work(regs, ti_work);
 171
 172                 /*
 173                  * Disable interrupts and reevaluate the work flags as they
 174                  * might have changed while interrupts and preemption was
 175                  * enabled above.
 176                  */
 177                 local_irq_disable_exit_to_user();
 178                 ti_work = READ_ONCE(current_thread_info()->flags);
 179         }
 180
 181         /* Return the latest work state for arch_exit_to_user_mode() */
 182         return ti_work;
 183 }
 184
 185 static void exit_to_user_mode_prepare(struct pt_regs *regs)
 186 {
 187         unsigned long ti_work = READ_ONCE(current_thread_info()->flags);
 188
 189         lockdep_assert_irqs_disabled();
 190
 191         if (unlikely(ti_work & EXIT_TO_USER_MODE_WORK))
 192                 ti_work = exit_to_user_mode_loop(regs, ti_work);
 193
 194         arch_exit_to_user_mode_prepare(regs, ti_work);
 195
 196         /* Ensure that the address limit is intact and no locks are held */
 197         addr_limit_user_check();
 198         lockdep_assert_irqs_disabled();
 199         lockdep_sys_exit();
 200 }
 201
 202 #ifndef _TIF_SINGLESTEP
 203 static inline bool report_single_step(unsigned long ti_work)
 204 {
 205         return false;
 206 }
 207 #else
 208 /*
 209  * If TIF_SYSCALL_EMU is set, then the only reason to report is when
 210  * TIF_SINGLESTEP is set (i.e. PTRACE_SYSEMU_SINGLESTEP).  This syscall
 211  * instruction has been already reported in syscall_enter_from_user_mode().
 212  */
 213 #define SYSEMU_STEP     (_TIF_SINGLESTEP | _TIF_SYSCALL_EMU)
 214
 215 static inline bool report_single_step(unsigned long ti_work)
 216 {
 217         return (ti_work & SYSEMU_STEP) == _TIF_SINGLESTEP;
 218 }
 219 #endif
 220
 221 static void syscall_exit_work(struct pt_regs *regs, unsigned long ti_work)
 222 {
 223         bool step;
 224
 225         audit_syscall_exit(regs);
 226
 227         if (ti_work & _TIF_SYSCALL_TRACEPOINT)
 228                 trace_sys_exit(regs, syscall_get_return_value(current, regs));
 229
 230         step = report_single_step(ti_work);
 231         if (step || ti_work & _TIF_SYSCALL_TRACE)
 232                 arch_syscall_exit_tracehook(regs, step);
 233 }
 234
 235 /*
 236  * Syscall specific exit to user mode preparation. Runs with interrupts
 237  * enabled.
 238  */
 239 static void syscall_exit_to_user_mode_prepare(struct pt_regs *regs)
 240 {
 241         u32 cached_flags = READ_ONCE(current_thread_info()->flags);
 242         unsigned long nr = syscall_get_nr(current, regs);
 243
 244         CT_WARN_ON(ct_state() != CONTEXT_KERNEL);
 245
 246         if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
 247                 if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr))
 248                         local_irq_enable();
 249         }
 250
 251         rseq_syscall(regs);
 252
 253         /*
 254          * Do one-time syscall specific work. If these work items are
 255          * enabled, we want to run them exactly once per syscall exit with
 256          * interrupts enabled.
 257          */
 258         if (unlikely(cached_flags & SYSCALL_EXIT_WORK))
 259                 syscall_exit_work(regs, cached_flags);
 260 }
 261
 262 __visible noinstr void syscall_exit_to_user_mode(struct pt_regs *regs)
 263 {
 264         instrumentation_begin();
 265         syscall_exit_to_user_mode_prepare(regs);
 266         local_irq_disable_exit_to_user();
 267         exit_to_user_mode_prepare(regs);
 268         instrumentation_end();
 269         exit_to_user_mode();
 270 }
 271
 272 noinstr void irqentry_enter_from_user_mode(struct pt_regs *regs)
 273 {
 274         enter_from_user_mode(regs);
 275 }
 276
 277 noinstr void irqentry_exit_to_user_mode(struct pt_regs *regs)
 278 {
 279         instrumentation_begin();
 280         exit_to_user_mode_prepare(regs);
 281         instrumentation_end();
 282         exit_to_user_mode();
 283 }
 284
 285 noinstr irqentry_state_t irqentry_enter(struct pt_regs *regs)
 286 {
 287         irqentry_state_t ret = {
 288                 .exit_rcu = false,
 289         };
 290
 291         if (user_mode(regs)) {
 292                 irqentry_enter_from_user_mode(regs);
 293                 return ret;
 294         }
 295
 296         /*
 297          * If this entry hit the idle task invoke rcu_irq_enter() whether
 298          * RCU is watching or not.
 299          *
 300          * Interupts can nest when the first interrupt invokes softirq
 301          * processing on return which enables interrupts.
 302          *
 303          * Scheduler ticks in the idle task can mark quiescent state and
 304          * terminate a grace period, if and only if the timer interrupt is
 305          * not nested into another interrupt.
 306          *
 307          * Checking for rcu_is_watching() here would prevent the nesting
 308          * interrupt to invoke rcu_irq_enter(). If that nested interrupt is
 309          * the tick then rcu_flavor_sched_clock_irq() would wrongfully
 310          * assume that it is the first interupt and eventually claim
 311          * quiescient state and end grace periods prematurely.
 312          *
 313          * Unconditionally invoke rcu_irq_enter() so RCU state stays
 314          * consistent.
 315          *
 316          * TINY_RCU does not support EQS, so let the compiler eliminate
 317          * this part when enabled.
 318          */
 319         if (!IS_ENABLED(CONFIG_TINY_RCU) && is_idle_task(current)) {
 320                 /*
 321                  * If RCU is not watching then the same careful
 322                  * sequence vs. lockdep and tracing is required
 323                  * as in irq_enter_from_user_mode().
 324                  */
 325                 lockdep_hardirqs_off(CALLER_ADDR0);
 326                 rcu_irq_enter();
 327                 instrumentation_begin();
 328                 trace_hardirqs_off_finish();
 329                 instrumentation_end();
 330
 331                 ret.exit_rcu = true;
 332                 return ret;
 333         }
 334
 335         /*
 336          * If RCU is watching then RCU only wants to check whether it needs
 337          * to restart the tick in NOHZ mode. rcu_irq_enter_check_tick()
 338          * already contains a warning when RCU is not watching, so no point
 339          * in having another one here.
 340          */
 341         instrumentation_begin();
 342         rcu_irq_enter_check_tick();
 343         /* Use the combo lockdep/tracing function */
 344         trace_hardirqs_off();
 345         instrumentation_end();
 346
 347         return ret;
 348 }
 349
 350 void irqentry_exit_cond_resched(void)
 351 {
 352         if (!preempt_count()) {
 353                 /* Sanity check RCU and thread stack */
 354                 rcu_irq_exit_check_preempt();
 355                 if (IS_ENABLED(CONFIG_DEBUG_ENTRY))
 356                         WARN_ON_ONCE(!on_thread_stack());
 357                 if (need_resched())
 358                         preempt_schedule_irq();
 359         }
 360 }
 361
 362 noinstr void irqentry_exit(struct pt_regs *regs, irqentry_state_t state)
 363 {
 364         lockdep_assert_irqs_disabled();
 365
 366         /* Check whether this returns to user mode */
 367         if (user_mode(regs)) {
 368                 irqentry_exit_to_user_mode(regs);
 369         } else if (!regs_irqs_disabled(regs)) {
 370                 /*
 371                  * If RCU was not watching on entry this needs to be done
 372                  * carefully and needs the same ordering of lockdep/tracing
 373                  * and RCU as the return to user mode path.
 374                  */
 375                 if (state.exit_rcu) {
 376                         instrumentation_begin();
 377                         /* Tell the tracer that IRET will enable interrupts */
 378                         trace_hardirqs_on_prepare();
 379                         lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 380                         instrumentation_end();
 381                         rcu_irq_exit();
 382                         lockdep_hardirqs_on(CALLER_ADDR0);
 383                         return;
 384                 }
 385
 386                 instrumentation_begin();
 387                 if (IS_ENABLED(CONFIG_PREEMPTION))
 388                         irqentry_exit_cond_resched();
 389                 /* Covers both tracing and lockdep */
 390                 trace_hardirqs_on();
 391                 instrumentation_end();
 392         } else {
 393                 /*
 394                  * IRQ flags state is correct already. Just tell RCU if it
 395                  * was not watching on entry.
 396                  */
 397                 if (state.exit_rcu)
 398                         rcu_irq_exit();
 399         }
 400 }