Linux Audio

Check our new training course

Loading...
v5.9
  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 */
 21static __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
 34static 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
 44static 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
 74static __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
 86long syscall_enter_from_user_mode_work(struct pt_regs *regs, long syscall)
 87{
 88	return __syscall_enter_from_user_work(regs, syscall);
 89}
 90
 91noinstr 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
105noinstr 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 */
125static __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 */
138void __weak arch_do_signal(struct pt_regs *regs) { }
139
140static 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
185static 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
203static 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_usermode().
212 */
213#define SYSEMU_STEP	(_TIF_SINGLESTEP | _TIF_SYSCALL_EMU)
214
215static inline bool report_single_step(unsigned long ti_work)
216{
217	return (ti_work & SYSEMU_STEP) == _TIF_SINGLESTEP;
 
 
 
218}
219#endif
220
221static 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 */
239static 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
272noinstr void irqentry_enter_from_user_mode(struct pt_regs *regs)
273{
274	enter_from_user_mode(regs);
275}
276
277noinstr 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
285noinstr 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
350void 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
362noinstr 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}
v6.8
  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
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 18static 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
 28long 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	syscall_enter_audit(regs, syscall);
 64
 65	return ret ? : syscall;
 66}
 67
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 68noinstr void syscall_enter_from_user_mode_prepare(struct pt_regs *regs)
 69{
 70	enter_from_user_mode(regs);
 71	instrumentation_begin();
 72	local_irq_enable();
 73	instrumentation_end();
 74}
 75
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 76/* Workaround to allow gradual conversion of architecture code */
 77void __weak arch_do_signal_or_restart(struct pt_regs *regs) { }
 78
 79/**
 80 * exit_to_user_mode_loop - do any pending work before leaving to user space
 81 * @regs:	Pointer to pt_regs on entry stack
 82 * @ti_work:	TIF work flags as read by the caller
 83 */
 84__always_inline unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
 85						     unsigned long ti_work)
 86{
 87	/*
 88	 * Before returning to user space ensure that all pending work
 89	 * items have been completed.
 90	 */
 91	while (ti_work & EXIT_TO_USER_MODE_WORK) {
 92
 93		local_irq_enable_exit_to_user(ti_work);
 94
 95		if (ti_work & _TIF_NEED_RESCHED)
 96			schedule();
 97
 98		if (ti_work & _TIF_UPROBE)
 99			uprobe_notify_resume(regs);
100
101		if (ti_work & _TIF_PATCH_PENDING)
102			klp_update_patch_state(current);
103
104		if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
105			arch_do_signal_or_restart(regs);
106
107		if (ti_work & _TIF_NOTIFY_RESUME)
108			resume_user_mode_work(regs);
 
 
 
109
110		/* Architecture specific TIF work */
111		arch_exit_to_user_mode_work(regs, ti_work);
112
113		/*
114		 * Disable interrupts and reevaluate the work flags as they
115		 * might have changed while interrupts and preemption was
116		 * enabled above.
117		 */
118		local_irq_disable_exit_to_user();
119
120		/* Check if any of the above work has queued a deferred wakeup */
121		tick_nohz_user_enter_prepare();
122
123		ti_work = read_thread_flags();
124	}
125
126	/* Return the latest work state for arch_exit_to_user_mode() */
127	return ti_work;
128}
129
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
130/*
131 * If SYSCALL_EMU is set, then the only reason to report is when
132 * SINGLESTEP is set (i.e. PTRACE_SYSEMU_SINGLESTEP).  This syscall
133 * instruction has been already reported in syscall_enter_from_user_mode().
134 */
135static inline bool report_single_step(unsigned long work)
 
 
136{
137	if (work & SYSCALL_WORK_SYSCALL_EMU)
138		return false;
139
140	return work & SYSCALL_WORK_SYSCALL_EXIT_TRAP;
141}
 
142
143static void syscall_exit_work(struct pt_regs *regs, unsigned long work)
144{
145	bool step;
146
147	/*
148	 * If the syscall was rolled back due to syscall user dispatching,
149	 * then the tracers below are not invoked for the same reason as
150	 * the entry side was not invoked in syscall_trace_enter(): The ABI
151	 * of these syscalls is unknown.
152	 */
153	if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
154		if (unlikely(current->syscall_dispatch.on_dispatch)) {
155			current->syscall_dispatch.on_dispatch = false;
156			return;
157		}
158	}
159
160	audit_syscall_exit(regs);
161
162	if (work & SYSCALL_WORK_SYSCALL_TRACEPOINT)
163		trace_sys_exit(regs, syscall_get_return_value(current, regs));
164
165	step = report_single_step(work);
166	if (step || work & SYSCALL_WORK_SYSCALL_TRACE)
167		ptrace_report_syscall_exit(regs, step);
168}
169
170/*
171 * Syscall specific exit to user mode preparation. Runs with interrupts
172 * enabled.
173 */
174static void syscall_exit_to_user_mode_prepare(struct pt_regs *regs)
175{
176	unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
177	unsigned long nr = syscall_get_nr(current, regs);
178
179	CT_WARN_ON(ct_state() != CONTEXT_KERNEL);
180
181	if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
182		if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr))
183			local_irq_enable();
184	}
185
186	rseq_syscall(regs);
187
188	/*
189	 * Do one-time syscall specific work. If these work items are
190	 * enabled, we want to run them exactly once per syscall exit with
191	 * interrupts enabled.
192	 */
193	if (unlikely(work & SYSCALL_WORK_EXIT))
194		syscall_exit_work(regs, work);
195}
196
197static __always_inline void __syscall_exit_to_user_mode_work(struct pt_regs *regs)
198{
 
199	syscall_exit_to_user_mode_prepare(regs);
200	local_irq_disable_exit_to_user();
201	exit_to_user_mode_prepare(regs);
202}
203
204void syscall_exit_to_user_mode_work(struct pt_regs *regs)
205{
206	__syscall_exit_to_user_mode_work(regs);
207}
208
209__visible noinstr void syscall_exit_to_user_mode(struct pt_regs *regs)
210{
211	instrumentation_begin();
212	__syscall_exit_to_user_mode_work(regs);
213	instrumentation_end();
214	exit_to_user_mode();
215}
216
217noinstr void irqentry_enter_from_user_mode(struct pt_regs *regs)
218{
219	enter_from_user_mode(regs);
220}
221
222noinstr void irqentry_exit_to_user_mode(struct pt_regs *regs)
223{
224	instrumentation_begin();
225	exit_to_user_mode_prepare(regs);
226	instrumentation_end();
227	exit_to_user_mode();
228}
229
230noinstr irqentry_state_t irqentry_enter(struct pt_regs *regs)
231{
232	irqentry_state_t ret = {
233		.exit_rcu = false,
234	};
235
236	if (user_mode(regs)) {
237		irqentry_enter_from_user_mode(regs);
238		return ret;
239	}
240
241	/*
242	 * If this entry hit the idle task invoke ct_irq_enter() whether
243	 * RCU is watching or not.
244	 *
245	 * Interrupts can nest when the first interrupt invokes softirq
246	 * processing on return which enables interrupts.
247	 *
248	 * Scheduler ticks in the idle task can mark quiescent state and
249	 * terminate a grace period, if and only if the timer interrupt is
250	 * not nested into another interrupt.
251	 *
252	 * Checking for rcu_is_watching() here would prevent the nesting
253	 * interrupt to invoke ct_irq_enter(). If that nested interrupt is
254	 * the tick then rcu_flavor_sched_clock_irq() would wrongfully
255	 * assume that it is the first interrupt and eventually claim
256	 * quiescent state and end grace periods prematurely.
257	 *
258	 * Unconditionally invoke ct_irq_enter() so RCU state stays
259	 * consistent.
260	 *
261	 * TINY_RCU does not support EQS, so let the compiler eliminate
262	 * this part when enabled.
263	 */
264	if (!IS_ENABLED(CONFIG_TINY_RCU) && is_idle_task(current)) {
265		/*
266		 * If RCU is not watching then the same careful
267		 * sequence vs. lockdep and tracing is required
268		 * as in irqentry_enter_from_user_mode().
269		 */
270		lockdep_hardirqs_off(CALLER_ADDR0);
271		ct_irq_enter();
272		instrumentation_begin();
273		kmsan_unpoison_entry_regs(regs);
274		trace_hardirqs_off_finish();
275		instrumentation_end();
276
277		ret.exit_rcu = true;
278		return ret;
279	}
280
281	/*
282	 * If RCU is watching then RCU only wants to check whether it needs
283	 * to restart the tick in NOHZ mode. rcu_irq_enter_check_tick()
284	 * already contains a warning when RCU is not watching, so no point
285	 * in having another one here.
286	 */
287	lockdep_hardirqs_off(CALLER_ADDR0);
288	instrumentation_begin();
289	kmsan_unpoison_entry_regs(regs);
290	rcu_irq_enter_check_tick();
291	trace_hardirqs_off_finish();
 
292	instrumentation_end();
293
294	return ret;
295}
296
297void raw_irqentry_exit_cond_resched(void)
298{
299	if (!preempt_count()) {
300		/* Sanity check RCU and thread stack */
301		rcu_irq_exit_check_preempt();
302		if (IS_ENABLED(CONFIG_DEBUG_ENTRY))
303			WARN_ON_ONCE(!on_thread_stack());
304		if (need_resched())
305			preempt_schedule_irq();
306	}
307}
308#ifdef CONFIG_PREEMPT_DYNAMIC
309#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
310DEFINE_STATIC_CALL(irqentry_exit_cond_resched, raw_irqentry_exit_cond_resched);
311#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
312DEFINE_STATIC_KEY_TRUE(sk_dynamic_irqentry_exit_cond_resched);
313void dynamic_irqentry_exit_cond_resched(void)
314{
315	if (!static_branch_unlikely(&sk_dynamic_irqentry_exit_cond_resched))
316		return;
317	raw_irqentry_exit_cond_resched();
318}
319#endif
320#endif
321
322noinstr void irqentry_exit(struct pt_regs *regs, irqentry_state_t state)
323{
324	lockdep_assert_irqs_disabled();
325
326	/* Check whether this returns to user mode */
327	if (user_mode(regs)) {
328		irqentry_exit_to_user_mode(regs);
329	} else if (!regs_irqs_disabled(regs)) {
330		/*
331		 * If RCU was not watching on entry this needs to be done
332		 * carefully and needs the same ordering of lockdep/tracing
333		 * and RCU as the return to user mode path.
334		 */
335		if (state.exit_rcu) {
336			instrumentation_begin();
337			/* Tell the tracer that IRET will enable interrupts */
338			trace_hardirqs_on_prepare();
339			lockdep_hardirqs_on_prepare();
340			instrumentation_end();
341			ct_irq_exit();
342			lockdep_hardirqs_on(CALLER_ADDR0);
343			return;
344		}
345
346		instrumentation_begin();
347		if (IS_ENABLED(CONFIG_PREEMPTION))
348			irqentry_exit_cond_resched();
349
350		/* Covers both tracing and lockdep */
351		trace_hardirqs_on();
352		instrumentation_end();
353	} else {
354		/*
355		 * IRQ flags state is correct already. Just tell RCU if it
356		 * was not watching on entry.
357		 */
358		if (state.exit_rcu)
359			ct_irq_exit();
360	}
361}
362
363irqentry_state_t noinstr irqentry_nmi_enter(struct pt_regs *regs)
364{
365	irqentry_state_t irq_state;
366
367	irq_state.lockdep = lockdep_hardirqs_enabled();
368
369	__nmi_enter();
370	lockdep_hardirqs_off(CALLER_ADDR0);
371	lockdep_hardirq_enter();
372	ct_nmi_enter();
373
374	instrumentation_begin();
375	kmsan_unpoison_entry_regs(regs);
376	trace_hardirqs_off_finish();
377	ftrace_nmi_enter();
378	instrumentation_end();
379
380	return irq_state;
381}
382
383void noinstr irqentry_nmi_exit(struct pt_regs *regs, irqentry_state_t irq_state)
384{
385	instrumentation_begin();
386	ftrace_nmi_exit();
387	if (irq_state.lockdep) {
388		trace_hardirqs_on_prepare();
389		lockdep_hardirqs_on_prepare();
390	}
391	instrumentation_end();
392
393	ct_nmi_exit();
394	lockdep_hardirq_exit();
395	if (irq_state.lockdep)
396		lockdep_hardirqs_on(CALLER_ADDR0);
397	__nmi_exit();
398}