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/resume_user_mode.h>
   6 #include <linux/highmem.h>
   7 #include <linux/jump_label.h>
   8 #include <linux/kmsan.h>
   9 #include <linux/livepatch.h>
  10 #include <linux/audit.h>
  11 #include <linux/tick.h>
  12
  13 #include "common.h"
  14
  15 #define CREATE_TRACE_POINTS
  16 #include <trace/events/syscalls.h>
  17
  18 static inline void syscall_enter_audit(struct pt_regs *regs, long syscall)
  19 {
  20         if (unlikely(audit_context())) {
  21                 unsigned long args[6];
  22
  23                 syscall_get_arguments(current, regs, args);
  24                 audit_syscall_entry(syscall, args[0], args[1], args[2], args[3]);
  25         }
  26 }
  27
  28 long syscall_trace_enter(struct pt_regs *regs, long syscall,
  29                                 unsigned long work)
  30 {
  31         long ret = 0;
  32
  33         /*
  34          * Handle Syscall User Dispatch.  This must comes first, since
  35          * the ABI here can be something that doesn't make sense for
  36          * other syscall_work features.
  37          */
  38         if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
  39                 if (syscall_user_dispatch(regs))
  40                         return -1L;
  41         }
  42
  43         /* Handle ptrace */
  44         if (work & (SYSCALL_WORK_SYSCALL_TRACE | SYSCALL_WORK_SYSCALL_EMU)) {
  45                 ret = ptrace_report_syscall_entry(regs);
  46                 if (ret || (work & SYSCALL_WORK_SYSCALL_EMU))
  47                         return -1L;
  48         }
  49
  50         /* Do seccomp after ptrace, to catch any tracer changes. */
  51         if (work & SYSCALL_WORK_SECCOMP) {
  52                 ret = __secure_computing(NULL);
  53                 if (ret == -1L)
  54                         return ret;
  55         }
  56
  57         /* Either of the above might have changed the syscall number */
  58         syscall = syscall_get_nr(current, regs);
  59
  60         if (unlikely(work & SYSCALL_WORK_SYSCALL_TRACEPOINT)) {
  61                 trace_sys_enter(regs, syscall);
  62                 /*
  63                  * Probes or BPF hooks in the tracepoint may have changed the
  64                  * system call number as well.
  65                  */
  66                 syscall = syscall_get_nr(current, regs);
  67         }
  68
  69         syscall_enter_audit(regs, syscall);
  70
  71         return ret ? : syscall;
  72 }
  73
  74 noinstr void syscall_enter_from_user_mode_prepare(struct pt_regs *regs)
  75 {
  76         enter_from_user_mode(regs);
  77         instrumentation_begin();
  78         local_irq_enable();
  79         instrumentation_end();
  80 }
  81
  82 /* Workaround to allow gradual conversion of architecture code */
  83 void __weak arch_do_signal_or_restart(struct pt_regs *regs) { }
  84
  85 /**
  86  * exit_to_user_mode_loop - do any pending work before leaving to user space
  87  * @regs:       Pointer to pt_regs on entry stack
  88  * @ti_work:    TIF work flags as read by the caller
  89  */
  90 __always_inline unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
  91                                                      unsigned long ti_work)
  92 {
  93         /*
  94          * Before returning to user space ensure that all pending work
  95          * items have been completed.
  96          */
  97         while (ti_work & EXIT_TO_USER_MODE_WORK) {
  98
  99                 local_irq_enable_exit_to_user(ti_work);
 100
 101                 if (ti_work & (_TIF_NEED_RESCHED | _TIF_NEED_RESCHED_LAZY))
 102                         schedule();
 103
 104                 if (ti_work & _TIF_UPROBE)
 105                         uprobe_notify_resume(regs);
 106
 107                 if (ti_work & _TIF_PATCH_PENDING)
 108                         klp_update_patch_state(current);
 109
 110                 if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
 111                         arch_do_signal_or_restart(regs);
 112
 113                 if (ti_work & _TIF_NOTIFY_RESUME)
 114                         resume_user_mode_work(regs);
 115
 116                 /* Architecture specific TIF work */
 117                 arch_exit_to_user_mode_work(regs, ti_work);
 118
 119                 /*
 120                  * Disable interrupts and reevaluate the work flags as they
 121                  * might have changed while interrupts and preemption was
 122                  * enabled above.
 123                  */
 124                 local_irq_disable_exit_to_user();
 125
 126                 /* Check if any of the above work has queued a deferred wakeup */
 127                 tick_nohz_user_enter_prepare();
 128
 129                 ti_work = read_thread_flags();
 130         }
 131
 132         /* Return the latest work state for arch_exit_to_user_mode() */
 133         return ti_work;
 134 }
 135
 136 /*
 137  * If SYSCALL_EMU is set, then the only reason to report is when
 138  * SINGLESTEP is set (i.e. PTRACE_SYSEMU_SINGLESTEP).  This syscall
 139  * instruction has been already reported in syscall_enter_from_user_mode().
 140  */
 141 static inline bool report_single_step(unsigned long work)
 142 {
 143         if (work & SYSCALL_WORK_SYSCALL_EMU)
 144                 return false;
 145
 146         return work & SYSCALL_WORK_SYSCALL_EXIT_TRAP;
 147 }
 148
 149 static void syscall_exit_work(struct pt_regs *regs, unsigned long work)
 150 {
 151         bool step;
 152
 153         /*
 154          * If the syscall was rolled back due to syscall user dispatching,
 155          * then the tracers below are not invoked for the same reason as
 156          * the entry side was not invoked in syscall_trace_enter(): The ABI
 157          * of these syscalls is unknown.
 158          */
 159         if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
 160                 if (unlikely(current->syscall_dispatch.on_dispatch)) {
 161                         current->syscall_dispatch.on_dispatch = false;
 162                         return;
 163                 }
 164         }
 165
 166         audit_syscall_exit(regs);
 167
 168         if (work & SYSCALL_WORK_SYSCALL_TRACEPOINT)
 169                 trace_sys_exit(regs, syscall_get_return_value(current, regs));
 170
 171         step = report_single_step(work);
 172         if (step || work & SYSCALL_WORK_SYSCALL_TRACE)
 173                 ptrace_report_syscall_exit(regs, step);
 174 }
 175
 176 /*
 177  * Syscall specific exit to user mode preparation. Runs with interrupts
 178  * enabled.
 179  */
 180 static void syscall_exit_to_user_mode_prepare(struct pt_regs *regs)
 181 {
 182         unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
 183         unsigned long nr = syscall_get_nr(current, regs);
 184
 185         CT_WARN_ON(ct_state() != CT_STATE_KERNEL);
 186
 187         if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
 188                 if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr))
 189                         local_irq_enable();
 190         }
 191
 192         rseq_syscall(regs);
 193
 194         /*
 195          * Do one-time syscall specific work. If these work items are
 196          * enabled, we want to run them exactly once per syscall exit with
 197          * interrupts enabled.
 198          */
 199         if (unlikely(work & SYSCALL_WORK_EXIT))
 200                 syscall_exit_work(regs, work);
 201 }
 202
 203 static __always_inline void __syscall_exit_to_user_mode_work(struct pt_regs *regs)
 204 {
 205         syscall_exit_to_user_mode_prepare(regs);
 206         local_irq_disable_exit_to_user();
 207         exit_to_user_mode_prepare(regs);
 208 }
 209
 210 void syscall_exit_to_user_mode_work(struct pt_regs *regs)
 211 {
 212         __syscall_exit_to_user_mode_work(regs);
 213 }
 214
 215 __visible noinstr void syscall_exit_to_user_mode(struct pt_regs *regs)
 216 {
 217         instrumentation_begin();
 218         __syscall_exit_to_user_mode_work(regs);
 219         instrumentation_end();
 220         exit_to_user_mode();
 221 }
 222
 223 noinstr void irqentry_enter_from_user_mode(struct pt_regs *regs)
 224 {
 225         enter_from_user_mode(regs);
 226 }
 227
 228 noinstr void irqentry_exit_to_user_mode(struct pt_regs *regs)
 229 {
 230         instrumentation_begin();
 231         exit_to_user_mode_prepare(regs);
 232         instrumentation_end();
 233         exit_to_user_mode();
 234 }
 235
 236 noinstr irqentry_state_t irqentry_enter(struct pt_regs *regs)
 237 {
 238         irqentry_state_t ret = {
 239                 .exit_rcu = false,
 240         };
 241
 242         if (user_mode(regs)) {
 243                 irqentry_enter_from_user_mode(regs);
 244                 return ret;
 245         }
 246
 247         /*
 248          * If this entry hit the idle task invoke ct_irq_enter() whether
 249          * RCU is watching or not.
 250          *
 251          * Interrupts can nest when the first interrupt invokes softirq
 252          * processing on return which enables interrupts.
 253          *
 254          * Scheduler ticks in the idle task can mark quiescent state and
 255          * terminate a grace period, if and only if the timer interrupt is
 256          * not nested into another interrupt.
 257          *
 258          * Checking for rcu_is_watching() here would prevent the nesting
 259          * interrupt to invoke ct_irq_enter(). If that nested interrupt is
 260          * the tick then rcu_flavor_sched_clock_irq() would wrongfully
 261          * assume that it is the first interrupt and eventually claim
 262          * quiescent state and end grace periods prematurely.
 263          *
 264          * Unconditionally invoke ct_irq_enter() so RCU state stays
 265          * consistent.
 266          *
 267          * TINY_RCU does not support EQS, so let the compiler eliminate
 268          * this part when enabled.
 269          */
 270         if (!IS_ENABLED(CONFIG_TINY_RCU) && is_idle_task(current)) {
 271                 /*
 272                  * If RCU is not watching then the same careful
 273                  * sequence vs. lockdep and tracing is required
 274                  * as in irqentry_enter_from_user_mode().
 275                  */
 276                 lockdep_hardirqs_off(CALLER_ADDR0);
 277                 ct_irq_enter();
 278                 instrumentation_begin();
 279                 kmsan_unpoison_entry_regs(regs);
 280                 trace_hardirqs_off_finish();
 281                 instrumentation_end();
 282
 283                 ret.exit_rcu = true;
 284                 return ret;
 285         }
 286
 287         /*
 288          * If RCU is watching then RCU only wants to check whether it needs
 289          * to restart the tick in NOHZ mode. rcu_irq_enter_check_tick()
 290          * already contains a warning when RCU is not watching, so no point
 291          * in having another one here.
 292          */
 293         lockdep_hardirqs_off(CALLER_ADDR0);
 294         instrumentation_begin();
 295         kmsan_unpoison_entry_regs(regs);
 296         rcu_irq_enter_check_tick();
 297         trace_hardirqs_off_finish();
 298         instrumentation_end();
 299
 300         return ret;
 301 }
 302
 303 void raw_irqentry_exit_cond_resched(void)
 304 {
 305         if (!preempt_count()) {
 306                 /* Sanity check RCU and thread stack */
 307                 rcu_irq_exit_check_preempt();
 308                 if (IS_ENABLED(CONFIG_DEBUG_ENTRY))
 309                         WARN_ON_ONCE(!on_thread_stack());
 310                 if (need_resched())
 311                         preempt_schedule_irq();
 312         }
 313 }
 314 #ifdef CONFIG_PREEMPT_DYNAMIC
 315 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
 316 DEFINE_STATIC_CALL(irqentry_exit_cond_resched, raw_irqentry_exit_cond_resched);
 317 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
 318 DEFINE_STATIC_KEY_TRUE(sk_dynamic_irqentry_exit_cond_resched);
 319 void dynamic_irqentry_exit_cond_resched(void)
 320 {
 321         if (!static_branch_unlikely(&sk_dynamic_irqentry_exit_cond_resched))
 322                 return;
 323         raw_irqentry_exit_cond_resched();
 324 }
 325 #endif
 326 #endif
 327
 328 noinstr void irqentry_exit(struct pt_regs *regs, irqentry_state_t state)
 329 {
 330         lockdep_assert_irqs_disabled();
 331
 332         /* Check whether this returns to user mode */
 333         if (user_mode(regs)) {
 334                 irqentry_exit_to_user_mode(regs);
 335         } else if (!regs_irqs_disabled(regs)) {
 336                 /*
 337                  * If RCU was not watching on entry this needs to be done
 338                  * carefully and needs the same ordering of lockdep/tracing
 339                  * and RCU as the return to user mode path.
 340                  */
 341                 if (state.exit_rcu) {
 342                         instrumentation_begin();
 343                         /* Tell the tracer that IRET will enable interrupts */
 344                         trace_hardirqs_on_prepare();
 345                         lockdep_hardirqs_on_prepare();
 346                         instrumentation_end();
 347                         ct_irq_exit();
 348                         lockdep_hardirqs_on(CALLER_ADDR0);
 349                         return;
 350                 }
 351
 352                 instrumentation_begin();
 353                 if (IS_ENABLED(CONFIG_PREEMPTION))
 354                         irqentry_exit_cond_resched();
 355
 356                 /* Covers both tracing and lockdep */
 357                 trace_hardirqs_on();
 358                 instrumentation_end();
 359         } else {
 360                 /*
 361                  * IRQ flags state is correct already. Just tell RCU if it
 362                  * was not watching on entry.
 363                  */
 364                 if (state.exit_rcu)
 365                         ct_irq_exit();
 366         }
 367 }
 368
 369 irqentry_state_t noinstr irqentry_nmi_enter(struct pt_regs *regs)
 370 {
 371         irqentry_state_t irq_state;
 372
 373         irq_state.lockdep = lockdep_hardirqs_enabled();
 374
 375         __nmi_enter();
 376         lockdep_hardirqs_off(CALLER_ADDR0);
 377         lockdep_hardirq_enter();
 378         ct_nmi_enter();
 379
 380         instrumentation_begin();
 381         kmsan_unpoison_entry_regs(regs);
 382         trace_hardirqs_off_finish();
 383         ftrace_nmi_enter();
 384         instrumentation_end();
 385
 386         return irq_state;
 387 }
 388
 389 void noinstr irqentry_nmi_exit(struct pt_regs *regs, irqentry_state_t irq_state)
 390 {
 391         instrumentation_begin();
 392         ftrace_nmi_exit();
 393         if (irq_state.lockdep) {
 394                 trace_hardirqs_on_prepare();
 395                 lockdep_hardirqs_on_prepare();
 396         }
 397         instrumentation_end();
 398
 399         ct_nmi_exit();
 400         lockdep_hardirq_exit();
 401         if (irq_state.lockdep)
 402                 lockdep_hardirqs_on(CALLER_ADDR0);
 403         __nmi_exit();
 404 }