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                         tracehook_notify_resume(regs);
 165                         rseq_handle_notify_resume(NULL, regs);
 166                 }
 167
 168                 /* Architecture specific TIF work */
 169                 arch_exit_to_user_mode_work(regs, ti_work);
 170
 171                 /*
 172                  * Disable interrupts and reevaluate the work flags as they
 173                  * might have changed while interrupts and preemption was
 174                  * enabled above.
 175                  */
 176                 local_irq_disable_exit_to_user();
 177                 ti_work = READ_ONCE(current_thread_info()->flags);
 178         }
 179
 180         /* Return the latest work state for arch_exit_to_user_mode() */
 181         return ti_work;
 182 }
 183
 184 static void exit_to_user_mode_prepare(struct pt_regs *regs)
 185 {
 186         unsigned long ti_work = READ_ONCE(current_thread_info()->flags);
 187
 188         lockdep_assert_irqs_disabled();
 189
 190         if (unlikely(ti_work & EXIT_TO_USER_MODE_WORK))
 191                 ti_work = exit_to_user_mode_loop(regs, ti_work);
 192
 193         arch_exit_to_user_mode_prepare(regs, ti_work);
 194
 195         /* Ensure that the address limit is intact and no locks are held */
 196         addr_limit_user_check();
 197         lockdep_assert_irqs_disabled();
 198         lockdep_sys_exit();
 199 }
 200
 201 #ifndef _TIF_SINGLESTEP
 202 static inline bool report_single_step(unsigned long ti_work)
 203 {
 204         return false;
 205 }
 206 #else
 207 /*
 208  * If TIF_SYSCALL_EMU is set, then the only reason to report is when
 209  * TIF_SINGLESTEP is set (i.e. PTRACE_SYSEMU_SINGLESTEP).  This syscall
 210  * instruction has been already reported in syscall_enter_from_user_mode().
 211  */
 212 #define SYSEMU_STEP     (_TIF_SINGLESTEP | _TIF_SYSCALL_EMU)
 213
 214 static inline bool report_single_step(unsigned long ti_work)
 215 {
 216         return (ti_work & SYSEMU_STEP) == _TIF_SINGLESTEP;
 217 }
 218 #endif
 219
 220 static void syscall_exit_work(struct pt_regs *regs, unsigned long ti_work)
 221 {
 222         bool step;
 223
 224         audit_syscall_exit(regs);
 225
 226         if (ti_work & _TIF_SYSCALL_TRACEPOINT)
 227                 trace_sys_exit(regs, syscall_get_return_value(current, regs));
 228
 229         step = report_single_step(ti_work);
 230         if (step || ti_work & _TIF_SYSCALL_TRACE)
 231                 arch_syscall_exit_tracehook(regs, step);
 232 }
 233
 234 /*
 235  * Syscall specific exit to user mode preparation. Runs with interrupts
 236  * enabled.
 237  */
 238 static void syscall_exit_to_user_mode_prepare(struct pt_regs *regs)
 239 {
 240         u32 cached_flags = READ_ONCE(current_thread_info()->flags);
 241         unsigned long nr = syscall_get_nr(current, regs);
 242
 243         CT_WARN_ON(ct_state() != CONTEXT_KERNEL);
 244
 245         if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
 246                 if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr))
 247                         local_irq_enable();
 248         }
 249
 250         rseq_syscall(regs);
 251
 252         /*
 253          * Do one-time syscall specific work. If these work items are
 254          * enabled, we want to run them exactly once per syscall exit with
 255          * interrupts enabled.
 256          */
 257         if (unlikely(cached_flags & SYSCALL_EXIT_WORK))
 258                 syscall_exit_work(regs, cached_flags);
 259 }
 260
 261 __visible noinstr void syscall_exit_to_user_mode(struct pt_regs *regs)
 262 {
 263         instrumentation_begin();
 264         syscall_exit_to_user_mode_prepare(regs);
 265         local_irq_disable_exit_to_user();
 266         exit_to_user_mode_prepare(regs);
 267         instrumentation_end();
 268         exit_to_user_mode();
 269 }
 270
 271 noinstr void irqentry_enter_from_user_mode(struct pt_regs *regs)
 272 {
 273         enter_from_user_mode(regs);
 274 }
 275
 276 noinstr void irqentry_exit_to_user_mode(struct pt_regs *regs)
 277 {
 278         instrumentation_begin();
 279         exit_to_user_mode_prepare(regs);
 280         instrumentation_end();
 281         exit_to_user_mode();
 282 }
 283
 284 noinstr irqentry_state_t irqentry_enter(struct pt_regs *regs)
 285 {
 286         irqentry_state_t ret = {
 287                 .exit_rcu = false,
 288         };
 289
 290         if (user_mode(regs)) {
 291                 irqentry_enter_from_user_mode(regs);
 292                 return ret;
 293         }
 294
 295         /*
 296          * If this entry hit the idle task invoke rcu_irq_enter() whether
 297          * RCU is watching or not.
 298          *
 299          * Interupts can nest when the first interrupt invokes softirq
 300          * processing on return which enables interrupts.
 301          *
 302          * Scheduler ticks in the idle task can mark quiescent state and
 303          * terminate a grace period, if and only if the timer interrupt is
 304          * not nested into another interrupt.
 305          *
 306          * Checking for rcu_is_watching() here would prevent the nesting
 307          * interrupt to invoke rcu_irq_enter(). If that nested interrupt is
 308          * the tick then rcu_flavor_sched_clock_irq() would wrongfully
 309          * assume that it is the first interupt and eventually claim
 310          * quiescient state and end grace periods prematurely.
 311          *
 312          * Unconditionally invoke rcu_irq_enter() so RCU state stays
 313          * consistent.
 314          *
 315          * TINY_RCU does not support EQS, so let the compiler eliminate
 316          * this part when enabled.
 317          */
 318         if (!IS_ENABLED(CONFIG_TINY_RCU) && is_idle_task(current)) {
 319                 /*
 320                  * If RCU is not watching then the same careful
 321                  * sequence vs. lockdep and tracing is required
 322                  * as in irq_enter_from_user_mode().
 323                  */
 324                 lockdep_hardirqs_off(CALLER_ADDR0);
 325                 rcu_irq_enter();
 326                 instrumentation_begin();
 327                 trace_hardirqs_off_finish();
 328                 instrumentation_end();
 329
 330                 ret.exit_rcu = true;
 331                 return ret;
 332         }
 333
 334         /*
 335          * If RCU is watching then RCU only wants to check whether it needs
 336          * to restart the tick in NOHZ mode. rcu_irq_enter_check_tick()
 337          * already contains a warning when RCU is not watching, so no point
 338          * in having another one here.
 339          */
 340         lockdep_hardirqs_off(CALLER_ADDR0);
 341         instrumentation_begin();
 342         rcu_irq_enter_check_tick();
 343         trace_hardirqs_off_finish();
 344         instrumentation_end();
 345
 346         return ret;
 347 }
 348
 349 void irqentry_exit_cond_resched(void)
 350 {
 351         if (!preempt_count()) {
 352                 /* Sanity check RCU and thread stack */
 353                 rcu_irq_exit_check_preempt();
 354                 if (IS_ENABLED(CONFIG_DEBUG_ENTRY))
 355                         WARN_ON_ONCE(!on_thread_stack());
 356                 if (need_resched())
 357                         preempt_schedule_irq();
 358         }
 359 }
 360
 361 noinstr void irqentry_exit(struct pt_regs *regs, irqentry_state_t state)
 362 {
 363         lockdep_assert_irqs_disabled();
 364
 365         /* Check whether this returns to user mode */
 366         if (user_mode(regs)) {
 367                 irqentry_exit_to_user_mode(regs);
 368         } else if (!regs_irqs_disabled(regs)) {
 369                 /*
 370                  * If RCU was not watching on entry this needs to be done
 371                  * carefully and needs the same ordering of lockdep/tracing
 372                  * and RCU as the return to user mode path.
 373                  */
 374                 if (state.exit_rcu) {
 375                         instrumentation_begin();
 376                         /* Tell the tracer that IRET will enable interrupts */
 377                         trace_hardirqs_on_prepare();
 378                         lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 379                         instrumentation_end();
 380                         rcu_irq_exit();
 381                         lockdep_hardirqs_on(CALLER_ADDR0);
 382                         return;
 383                 }
 384
 385                 instrumentation_begin();
 386                 if (IS_ENABLED(CONFIG_PREEMPTION))
 387                         irqentry_exit_cond_resched();
 388                 /* Covers both tracing and lockdep */
 389                 trace_hardirqs_on();
 390                 instrumentation_end();
 391         } else {
 392                 /*
 393                  * IRQ flags state is correct already. Just tell RCU if it
 394                  * was not watching on entry.
 395                  */
 396                 if (state.exit_rcu)
 397                         rcu_irq_exit();
 398         }
 399 }