1 // SPDX-License-Identifier: GPL-2.0
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/livepatch.h>
9 #include <linux/audit.h>
10 #include <linux/tick.h>
14 #define CREATE_TRACE_POINTS
15 #include <trace/events/syscalls.h>
17 /* See comment for enter_from_user_mode() in entry-common.h */
18 static __always_inline void __enter_from_user_mode(struct pt_regs *regs)
20 arch_check_user_regs(regs);
21 lockdep_hardirqs_off(CALLER_ADDR0);
23 CT_WARN_ON(ct_state() != CONTEXT_USER);
26 instrumentation_begin();
27 trace_hardirqs_off_finish();
28 instrumentation_end();
31 void noinstr enter_from_user_mode(struct pt_regs *regs)
33 __enter_from_user_mode(regs);
36 static inline void syscall_enter_audit(struct pt_regs *regs, long syscall)
38 if (unlikely(audit_context())) {
39 unsigned long args[6];
41 syscall_get_arguments(current, regs, args);
42 audit_syscall_entry(syscall, args[0], args[1], args[2], args[3]);
46 static long syscall_trace_enter(struct pt_regs *regs, long syscall,
52 * Handle Syscall User Dispatch. This must comes first, since
53 * the ABI here can be something that doesn't make sense for
54 * other syscall_work features.
56 if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
57 if (syscall_user_dispatch(regs))
62 if (work & (SYSCALL_WORK_SYSCALL_TRACE | SYSCALL_WORK_SYSCALL_EMU)) {
63 ret = ptrace_report_syscall_entry(regs);
64 if (ret || (work & SYSCALL_WORK_SYSCALL_EMU))
68 /* Do seccomp after ptrace, to catch any tracer changes. */
69 if (work & SYSCALL_WORK_SECCOMP) {
70 ret = __secure_computing(NULL);
75 /* Either of the above might have changed the syscall number */
76 syscall = syscall_get_nr(current, regs);
78 if (unlikely(work & SYSCALL_WORK_SYSCALL_TRACEPOINT))
79 trace_sys_enter(regs, syscall);
81 syscall_enter_audit(regs, syscall);
83 return ret ? : syscall;
86 static __always_inline long
87 __syscall_enter_from_user_work(struct pt_regs *regs, long syscall)
89 unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
91 if (work & SYSCALL_WORK_ENTER)
92 syscall = syscall_trace_enter(regs, syscall, work);
97 long syscall_enter_from_user_mode_work(struct pt_regs *regs, long syscall)
99 return __syscall_enter_from_user_work(regs, syscall);
102 noinstr long syscall_enter_from_user_mode(struct pt_regs *regs, long syscall)
106 __enter_from_user_mode(regs);
108 instrumentation_begin();
110 ret = __syscall_enter_from_user_work(regs, syscall);
111 instrumentation_end();
116 noinstr void syscall_enter_from_user_mode_prepare(struct pt_regs *regs)
118 __enter_from_user_mode(regs);
119 instrumentation_begin();
121 instrumentation_end();
124 /* See comment for exit_to_user_mode() in entry-common.h */
125 static __always_inline void __exit_to_user_mode(void)
127 instrumentation_begin();
128 trace_hardirqs_on_prepare();
129 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
130 instrumentation_end();
133 arch_exit_to_user_mode();
134 lockdep_hardirqs_on(CALLER_ADDR0);
137 void noinstr exit_to_user_mode(void)
139 __exit_to_user_mode();
142 /* Workaround to allow gradual conversion of architecture code */
143 void __weak arch_do_signal_or_restart(struct pt_regs *regs) { }
145 #ifdef CONFIG_RT_DELAYED_SIGNALS
146 static inline void raise_delayed_signal(void)
148 if (unlikely(current->forced_info.si_signo)) {
149 force_sig_info(¤t->forced_info);
150 current->forced_info.si_signo = 0;
154 static inline void raise_delayed_signal(void) { }
157 static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
158 unsigned long ti_work)
161 * Before returning to user space ensure that all pending work
162 * items have been completed.
164 while (ti_work & EXIT_TO_USER_MODE_WORK) {
166 local_irq_enable_exit_to_user(ti_work);
168 if (ti_work & _TIF_NEED_RESCHED)
171 raise_delayed_signal();
173 if (ti_work & _TIF_UPROBE)
174 uprobe_notify_resume(regs);
176 if (ti_work & _TIF_PATCH_PENDING)
177 klp_update_patch_state(current);
179 if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
180 arch_do_signal_or_restart(regs);
182 if (ti_work & _TIF_NOTIFY_RESUME)
183 resume_user_mode_work(regs);
185 /* Architecture specific TIF work */
186 arch_exit_to_user_mode_work(regs, ti_work);
189 * Disable interrupts and reevaluate the work flags as they
190 * might have changed while interrupts and preemption was
193 local_irq_disable_exit_to_user();
195 /* Check if any of the above work has queued a deferred wakeup */
196 tick_nohz_user_enter_prepare();
198 ti_work = read_thread_flags();
201 /* Return the latest work state for arch_exit_to_user_mode() */
205 static void exit_to_user_mode_prepare(struct pt_regs *regs)
207 unsigned long ti_work = read_thread_flags();
209 lockdep_assert_irqs_disabled();
211 /* Flush pending rcuog wakeup before the last need_resched() check */
212 tick_nohz_user_enter_prepare();
214 if (unlikely(ti_work & EXIT_TO_USER_MODE_WORK))
215 ti_work = exit_to_user_mode_loop(regs, ti_work);
217 arch_exit_to_user_mode_prepare(regs, ti_work);
219 /* Ensure that the address limit is intact and no locks are held */
220 addr_limit_user_check();
222 lockdep_assert_irqs_disabled();
227 * If SYSCALL_EMU is set, then the only reason to report is when
228 * SINGLESTEP is set (i.e. PTRACE_SYSEMU_SINGLESTEP). This syscall
229 * instruction has been already reported in syscall_enter_from_user_mode().
231 static inline bool report_single_step(unsigned long work)
233 if (work & SYSCALL_WORK_SYSCALL_EMU)
236 return work & SYSCALL_WORK_SYSCALL_EXIT_TRAP;
239 static void syscall_exit_work(struct pt_regs *regs, unsigned long work)
244 * If the syscall was rolled back due to syscall user dispatching,
245 * then the tracers below are not invoked for the same reason as
246 * the entry side was not invoked in syscall_trace_enter(): The ABI
247 * of these syscalls is unknown.
249 if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
250 if (unlikely(current->syscall_dispatch.on_dispatch)) {
251 current->syscall_dispatch.on_dispatch = false;
256 audit_syscall_exit(regs);
258 if (work & SYSCALL_WORK_SYSCALL_TRACEPOINT)
259 trace_sys_exit(regs, syscall_get_return_value(current, regs));
261 step = report_single_step(work);
262 if (step || work & SYSCALL_WORK_SYSCALL_TRACE)
263 ptrace_report_syscall_exit(regs, step);
267 * Syscall specific exit to user mode preparation. Runs with interrupts
270 static void syscall_exit_to_user_mode_prepare(struct pt_regs *regs)
272 unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
273 unsigned long nr = syscall_get_nr(current, regs);
275 CT_WARN_ON(ct_state() != CONTEXT_KERNEL);
277 if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
278 if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr))
285 * Do one-time syscall specific work. If these work items are
286 * enabled, we want to run them exactly once per syscall exit with
287 * interrupts enabled.
289 if (unlikely(work & SYSCALL_WORK_EXIT))
290 syscall_exit_work(regs, work);
293 static __always_inline void __syscall_exit_to_user_mode_work(struct pt_regs *regs)
295 syscall_exit_to_user_mode_prepare(regs);
296 local_irq_disable_exit_to_user();
297 exit_to_user_mode_prepare(regs);
300 void syscall_exit_to_user_mode_work(struct pt_regs *regs)
302 __syscall_exit_to_user_mode_work(regs);
305 __visible noinstr void syscall_exit_to_user_mode(struct pt_regs *regs)
307 instrumentation_begin();
308 __syscall_exit_to_user_mode_work(regs);
309 instrumentation_end();
310 __exit_to_user_mode();
313 noinstr void irqentry_enter_from_user_mode(struct pt_regs *regs)
315 __enter_from_user_mode(regs);
318 noinstr void irqentry_exit_to_user_mode(struct pt_regs *regs)
320 instrumentation_begin();
321 exit_to_user_mode_prepare(regs);
322 instrumentation_end();
323 __exit_to_user_mode();
326 noinstr irqentry_state_t irqentry_enter(struct pt_regs *regs)
328 irqentry_state_t ret = {
332 if (user_mode(regs)) {
333 irqentry_enter_from_user_mode(regs);
338 * If this entry hit the idle task invoke rcu_irq_enter() whether
339 * RCU is watching or not.
341 * Interrupts can nest when the first interrupt invokes softirq
342 * processing on return which enables interrupts.
344 * Scheduler ticks in the idle task can mark quiescent state and
345 * terminate a grace period, if and only if the timer interrupt is
346 * not nested into another interrupt.
348 * Checking for rcu_is_watching() here would prevent the nesting
349 * interrupt to invoke rcu_irq_enter(). If that nested interrupt is
350 * the tick then rcu_flavor_sched_clock_irq() would wrongfully
351 * assume that it is the first interrupt and eventually claim
352 * quiescent state and end grace periods prematurely.
354 * Unconditionally invoke rcu_irq_enter() so RCU state stays
357 * TINY_RCU does not support EQS, so let the compiler eliminate
358 * this part when enabled.
360 if (!IS_ENABLED(CONFIG_TINY_RCU) && is_idle_task(current)) {
362 * If RCU is not watching then the same careful
363 * sequence vs. lockdep and tracing is required
364 * as in irqentry_enter_from_user_mode().
366 lockdep_hardirqs_off(CALLER_ADDR0);
368 instrumentation_begin();
369 trace_hardirqs_off_finish();
370 instrumentation_end();
377 * If RCU is watching then RCU only wants to check whether it needs
378 * to restart the tick in NOHZ mode. rcu_irq_enter_check_tick()
379 * already contains a warning when RCU is not watching, so no point
380 * in having another one here.
382 lockdep_hardirqs_off(CALLER_ADDR0);
383 instrumentation_begin();
384 rcu_irq_enter_check_tick();
385 trace_hardirqs_off_finish();
386 instrumentation_end();
391 void raw_irqentry_exit_cond_resched(void)
393 if (!preempt_count()) {
394 /* Sanity check RCU and thread stack */
395 rcu_irq_exit_check_preempt();
396 if (IS_ENABLED(CONFIG_DEBUG_ENTRY))
397 WARN_ON_ONCE(!on_thread_stack());
399 preempt_schedule_irq();
402 #ifdef CONFIG_PREEMPT_DYNAMIC
403 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
404 DEFINE_STATIC_CALL(irqentry_exit_cond_resched, raw_irqentry_exit_cond_resched);
405 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
406 DEFINE_STATIC_KEY_TRUE(sk_dynamic_irqentry_exit_cond_resched);
407 void dynamic_irqentry_exit_cond_resched(void)
409 if (!static_key_unlikely(&sk_dynamic_irqentry_exit_cond_resched))
411 raw_irqentry_exit_cond_resched();
416 noinstr void irqentry_exit(struct pt_regs *regs, irqentry_state_t state)
418 lockdep_assert_irqs_disabled();
420 /* Check whether this returns to user mode */
421 if (user_mode(regs)) {
422 irqentry_exit_to_user_mode(regs);
423 } else if (!regs_irqs_disabled(regs)) {
425 * If RCU was not watching on entry this needs to be done
426 * carefully and needs the same ordering of lockdep/tracing
427 * and RCU as the return to user mode path.
429 if (state.exit_rcu) {
430 instrumentation_begin();
431 /* Tell the tracer that IRET will enable interrupts */
432 trace_hardirqs_on_prepare();
433 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
434 instrumentation_end();
436 lockdep_hardirqs_on(CALLER_ADDR0);
440 instrumentation_begin();
441 if (IS_ENABLED(CONFIG_PREEMPTION))
442 irqentry_exit_cond_resched();
444 /* Covers both tracing and lockdep */
446 instrumentation_end();
449 * IRQ flags state is correct already. Just tell RCU if it
450 * was not watching on entry.
457 irqentry_state_t noinstr irqentry_nmi_enter(struct pt_regs *regs)
459 irqentry_state_t irq_state;
461 irq_state.lockdep = lockdep_hardirqs_enabled();
464 lockdep_hardirqs_off(CALLER_ADDR0);
465 lockdep_hardirq_enter();
468 instrumentation_begin();
469 trace_hardirqs_off_finish();
471 instrumentation_end();
476 void noinstr irqentry_nmi_exit(struct pt_regs *regs, irqentry_state_t irq_state)
478 instrumentation_begin();
480 if (irq_state.lockdep) {
481 trace_hardirqs_on_prepare();
482 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
484 instrumentation_end();
487 lockdep_hardirq_exit();
488 if (irq_state.lockdep)
489 lockdep_hardirqs_on(CALLER_ADDR0);