1 // SPDX-License-Identifier: GPL-2.0
3 #include <linux/context_tracking.h>
4 #include <linux/entry-common.h>
5 #include <linux/livepatch.h>
6 #include <linux/audit.h>
10 #define CREATE_TRACE_POINTS
11 #include <trace/events/syscalls.h>
14 * enter_from_user_mode - Establish state when coming from user mode
16 * Syscall/interrupt entry disables interrupts, but user mode is traced as
17 * interrupts enabled. Also with NO_HZ_FULL RCU might be idle.
19 * 1) Tell lockdep that interrupts are disabled
20 * 2) Invoke context tracking if enabled to reactivate RCU
21 * 3) Trace interrupts off state
23 static __always_inline void enter_from_user_mode(struct pt_regs *regs)
25 arch_check_user_regs(regs);
26 lockdep_hardirqs_off(CALLER_ADDR0);
28 CT_WARN_ON(ct_state() != CONTEXT_USER);
31 instrumentation_begin();
32 trace_hardirqs_off_finish();
33 instrumentation_end();
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 = arch_syscall_enter_tracehook(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();
125 * exit_to_user_mode - Fixup state when exiting to user mode
127 * Syscall/interupt exit enables interrupts, but the kernel state is
128 * interrupts disabled when this is invoked. Also tell RCU about it.
130 * 1) Trace interrupts on state
131 * 2) Invoke context tracking if enabled to adjust RCU state
132 * 3) Invoke architecture specific last minute exit code, e.g. speculation
134 * 4) Tell lockdep that interrupts are enabled
136 static __always_inline void exit_to_user_mode(void)
138 instrumentation_begin();
139 trace_hardirqs_on_prepare();
140 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
141 instrumentation_end();
144 arch_exit_to_user_mode();
145 lockdep_hardirqs_on(CALLER_ADDR0);
148 /* Workaround to allow gradual conversion of architecture code */
149 void __weak arch_do_signal_or_restart(struct pt_regs *regs, bool has_signal) { }
151 static void handle_signal_work(struct pt_regs *regs, unsigned long ti_work)
153 if (ti_work & _TIF_NOTIFY_SIGNAL)
154 tracehook_notify_signal();
156 arch_do_signal_or_restart(regs, ti_work & _TIF_SIGPENDING);
159 static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
160 unsigned long ti_work)
163 * Before returning to user space ensure that all pending work
164 * items have been completed.
166 while (ti_work & EXIT_TO_USER_MODE_WORK) {
168 local_irq_enable_exit_to_user(ti_work);
170 if (ti_work & _TIF_NEED_RESCHED)
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 handle_signal_work(regs, ti_work);
182 if (ti_work & _TIF_NOTIFY_RESUME) {
183 tracehook_notify_resume(regs);
184 rseq_handle_notify_resume(NULL, regs);
187 /* Architecture specific TIF work */
188 arch_exit_to_user_mode_work(regs, ti_work);
191 * Disable interrupts and reevaluate the work flags as they
192 * might have changed while interrupts and preemption was
195 local_irq_disable_exit_to_user();
196 ti_work = READ_ONCE(current_thread_info()->flags);
199 /* Return the latest work state for arch_exit_to_user_mode() */
203 static void exit_to_user_mode_prepare(struct pt_regs *regs)
205 unsigned long ti_work = READ_ONCE(current_thread_info()->flags);
207 lockdep_assert_irqs_disabled();
209 if (unlikely(ti_work & EXIT_TO_USER_MODE_WORK))
210 ti_work = exit_to_user_mode_loop(regs, ti_work);
212 arch_exit_to_user_mode_prepare(regs, ti_work);
214 /* Ensure that the address limit is intact and no locks are held */
215 addr_limit_user_check();
216 lockdep_assert_irqs_disabled();
220 #ifndef _TIF_SINGLESTEP
221 static inline bool report_single_step(unsigned long work)
227 * If SYSCALL_EMU is set, then the only reason to report is when
228 * TIF_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 !!(current_thread_info()->flags & _TIF_SINGLESTEP);
241 static void syscall_exit_work(struct pt_regs *regs, unsigned long work)
246 * If the syscall was rolled back due to syscall user dispatching,
247 * then the tracers below are not invoked for the same reason as
248 * the entry side was not invoked in syscall_trace_enter(): The ABI
249 * of these syscalls is unknown.
251 if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
252 if (unlikely(current->syscall_dispatch.on_dispatch)) {
253 current->syscall_dispatch.on_dispatch = false;
258 audit_syscall_exit(regs);
260 if (work & SYSCALL_WORK_SYSCALL_TRACEPOINT)
261 trace_sys_exit(regs, syscall_get_return_value(current, regs));
263 step = report_single_step(work);
264 if (step || work & SYSCALL_WORK_SYSCALL_TRACE)
265 arch_syscall_exit_tracehook(regs, step);
269 * Syscall specific exit to user mode preparation. Runs with interrupts
272 static void syscall_exit_to_user_mode_prepare(struct pt_regs *regs)
274 unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
275 unsigned long nr = syscall_get_nr(current, regs);
277 CT_WARN_ON(ct_state() != CONTEXT_KERNEL);
279 if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
280 if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr))
287 * Do one-time syscall specific work. If these work items are
288 * enabled, we want to run them exactly once per syscall exit with
289 * interrupts enabled.
291 if (unlikely(work & SYSCALL_WORK_EXIT))
292 syscall_exit_work(regs, work);
295 __visible noinstr void syscall_exit_to_user_mode(struct pt_regs *regs)
297 instrumentation_begin();
298 syscall_exit_to_user_mode_prepare(regs);
299 local_irq_disable_exit_to_user();
300 exit_to_user_mode_prepare(regs);
301 instrumentation_end();
305 noinstr void irqentry_enter_from_user_mode(struct pt_regs *regs)
307 enter_from_user_mode(regs);
310 noinstr void irqentry_exit_to_user_mode(struct pt_regs *regs)
312 instrumentation_begin();
313 exit_to_user_mode_prepare(regs);
314 instrumentation_end();
318 noinstr irqentry_state_t irqentry_enter(struct pt_regs *regs)
320 irqentry_state_t ret = {
324 if (user_mode(regs)) {
325 irqentry_enter_from_user_mode(regs);
330 * If this entry hit the idle task invoke rcu_irq_enter() whether
331 * RCU is watching or not.
333 * Interrupts can nest when the first interrupt invokes softirq
334 * processing on return which enables interrupts.
336 * Scheduler ticks in the idle task can mark quiescent state and
337 * terminate a grace period, if and only if the timer interrupt is
338 * not nested into another interrupt.
340 * Checking for rcu_is_watching() here would prevent the nesting
341 * interrupt to invoke rcu_irq_enter(). If that nested interrupt is
342 * the tick then rcu_flavor_sched_clock_irq() would wrongfully
343 * assume that it is the first interupt and eventually claim
344 * quiescent state and end grace periods prematurely.
346 * Unconditionally invoke rcu_irq_enter() so RCU state stays
349 * TINY_RCU does not support EQS, so let the compiler eliminate
350 * this part when enabled.
352 if (!IS_ENABLED(CONFIG_TINY_RCU) && is_idle_task(current)) {
354 * If RCU is not watching then the same careful
355 * sequence vs. lockdep and tracing is required
356 * as in irqentry_enter_from_user_mode().
358 lockdep_hardirqs_off(CALLER_ADDR0);
360 instrumentation_begin();
361 trace_hardirqs_off_finish();
362 instrumentation_end();
369 * If RCU is watching then RCU only wants to check whether it needs
370 * to restart the tick in NOHZ mode. rcu_irq_enter_check_tick()
371 * already contains a warning when RCU is not watching, so no point
372 * in having another one here.
374 lockdep_hardirqs_off(CALLER_ADDR0);
375 instrumentation_begin();
376 rcu_irq_enter_check_tick();
377 trace_hardirqs_off_finish();
378 instrumentation_end();
383 void irqentry_exit_cond_resched(void)
385 if (!preempt_count()) {
386 /* Sanity check RCU and thread stack */
387 rcu_irq_exit_check_preempt();
388 if (IS_ENABLED(CONFIG_DEBUG_ENTRY))
389 WARN_ON_ONCE(!on_thread_stack());
391 preempt_schedule_irq();
395 noinstr void irqentry_exit(struct pt_regs *regs, irqentry_state_t state)
397 lockdep_assert_irqs_disabled();
399 /* Check whether this returns to user mode */
400 if (user_mode(regs)) {
401 irqentry_exit_to_user_mode(regs);
402 } else if (!regs_irqs_disabled(regs)) {
404 * If RCU was not watching on entry this needs to be done
405 * carefully and needs the same ordering of lockdep/tracing
406 * and RCU as the return to user mode path.
408 if (state.exit_rcu) {
409 instrumentation_begin();
410 /* Tell the tracer that IRET will enable interrupts */
411 trace_hardirqs_on_prepare();
412 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
413 instrumentation_end();
415 lockdep_hardirqs_on(CALLER_ADDR0);
419 instrumentation_begin();
420 if (IS_ENABLED(CONFIG_PREEMPTION))
421 irqentry_exit_cond_resched();
422 /* Covers both tracing and lockdep */
424 instrumentation_end();
427 * IRQ flags state is correct already. Just tell RCU if it
428 * was not watching on entry.
435 irqentry_state_t noinstr irqentry_nmi_enter(struct pt_regs *regs)
437 irqentry_state_t irq_state;
439 irq_state.lockdep = lockdep_hardirqs_enabled();
442 lockdep_hardirqs_off(CALLER_ADDR0);
443 lockdep_hardirq_enter();
446 instrumentation_begin();
447 trace_hardirqs_off_finish();
449 instrumentation_end();
454 void noinstr irqentry_nmi_exit(struct pt_regs *regs, irqentry_state_t irq_state)
456 instrumentation_begin();
458 if (irq_state.lockdep) {
459 trace_hardirqs_on_prepare();
460 lockdep_hardirqs_on_prepare(CALLER_ADDR0);
462 instrumentation_end();
465 lockdep_hardirq_exit();
466 if (irq_state.lockdep)
467 lockdep_hardirqs_on(CALLER_ADDR0);