xref: /linux/kernel/entry/common.c (revision c43267e6794a36013fd495a4d81bf7f748fe4615)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 #include <linux/irq-entry-common.h>
4 #include <linux/resume_user_mode.h>
5 #include <linux/highmem.h>
6 #include <linux/jump_label.h>
7 #include <linux/kmsan.h>
8 #include <linux/livepatch.h>
9 #include <linux/tick.h>
10 
11 /* Workaround to allow gradual conversion of architecture code */
arch_do_signal_or_restart(struct pt_regs * regs)12 void __weak arch_do_signal_or_restart(struct pt_regs *regs) { }
13 
14 #ifdef CONFIG_HAVE_GENERIC_TIF_BITS
15 #define EXIT_TO_USER_MODE_WORK_LOOP	(EXIT_TO_USER_MODE_WORK & ~_TIF_RSEQ)
16 #else
17 #define EXIT_TO_USER_MODE_WORK_LOOP	(EXIT_TO_USER_MODE_WORK)
18 #endif
19 
20 /* TIF bits, which prevent a time slice extension. */
21 #ifdef CONFIG_PREEMPT_RT
22 /*
23  * Since rseq slice ext has a direct correlation to the worst case
24  * scheduling latency (schedule is delayed after all), only have it affect
25  * LAZY reschedules on PREEMPT_RT for now.
26  *
27  * However, since this delay is only applicable to userspace, a value
28  * for rseq_slice_extension_nsec that is strictly less than the worst case
29  * kernel space preempt_disable() region, should mean the scheduling latency
30  * is not affected, even for !LAZY.
31  *
32  * However, since this value depends on the hardware at hand, it cannot be
33  * pre-determined in any sensible way. Hence punt on this problem for now.
34  */
35 # define TIF_SLICE_EXT_SCHED	(_TIF_NEED_RESCHED_LAZY)
36 #else
37 # define TIF_SLICE_EXT_SCHED	(_TIF_NEED_RESCHED | _TIF_NEED_RESCHED_LAZY)
38 #endif
39 #define TIF_SLICE_EXT_DENY	(EXIT_TO_USER_MODE_WORK & ~TIF_SLICE_EXT_SCHED)
40 
__exit_to_user_mode_loop(struct pt_regs * regs,unsigned long ti_work)41 static __always_inline unsigned long __exit_to_user_mode_loop(struct pt_regs *regs,
42 							      unsigned long ti_work)
43 {
44 	/*
45 	 * Before returning to user space ensure that all pending work
46 	 * items have been completed.
47 	 */
48 	while (ti_work & EXIT_TO_USER_MODE_WORK_LOOP) {
49 
50 		local_irq_enable();
51 
52 		if (ti_work & (_TIF_NEED_RESCHED | _TIF_NEED_RESCHED_LAZY)) {
53 			if (!rseq_grant_slice_extension(ti_work, TIF_SLICE_EXT_DENY))
54 				schedule();
55 		}
56 
57 		if (ti_work & _TIF_UPROBE)
58 			uprobe_notify_resume(regs);
59 
60 		if (ti_work & _TIF_PATCH_PENDING)
61 			klp_update_patch_state(current);
62 
63 		if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
64 			arch_do_signal_or_restart(regs);
65 
66 		if (ti_work & _TIF_NOTIFY_RESUME)
67 			resume_user_mode_work(regs);
68 
69 		/* Architecture specific TIF work */
70 		arch_exit_to_user_mode_work(regs, ti_work);
71 
72 		/*
73 		 * Disable interrupts and reevaluate the work flags as they
74 		 * might have changed while interrupts and preemption was
75 		 * enabled above.
76 		 */
77 		local_irq_disable();
78 
79 		/* Check if any of the above work has queued a deferred wakeup */
80 		tick_nohz_user_enter_prepare();
81 
82 		ti_work = read_thread_flags();
83 	}
84 
85 	/* Return the latest work state for arch_exit_to_user_mode() */
86 	return ti_work;
87 }
88 
89 /**
90  * exit_to_user_mode_loop - do any pending work before leaving to user space
91  * @regs:	Pointer to pt_regs on entry stack
92  * @ti_work:	TIF work flags as read by the caller
93  */
exit_to_user_mode_loop(struct pt_regs * regs,unsigned long ti_work)94 __always_inline unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
95 						     unsigned long ti_work)
96 {
97 	for (;;) {
98 		ti_work = __exit_to_user_mode_loop(regs, ti_work);
99 
100 		if (likely(!rseq_exit_to_user_mode_restart(regs, ti_work)))
101 			return ti_work;
102 		ti_work = read_thread_flags();
103 	}
104 }
105 
irqentry_enter(struct pt_regs * regs)106 noinstr irqentry_state_t irqentry_enter(struct pt_regs *regs)
107 {
108 	if (user_mode(regs)) {
109 		irqentry_state_t ret = {
110 			.exit_rcu = false,
111 		};
112 
113 		irqentry_enter_from_user_mode(regs);
114 		return ret;
115 	}
116 
117 	return irqentry_enter_from_kernel_mode(regs);
118 }
119 
120 /**
121  * arch_irqentry_exit_need_resched - Architecture specific need resched function
122  *
123  * Invoked from raw_irqentry_exit_cond_resched() to check if resched is needed.
124  * Defaults return true.
125  *
126  * The main purpose is to permit arch to avoid preemption of a task from an IRQ.
127  */
128 static inline bool arch_irqentry_exit_need_resched(void);
129 
130 #ifndef arch_irqentry_exit_need_resched
arch_irqentry_exit_need_resched(void)131 static inline bool arch_irqentry_exit_need_resched(void) { return true; }
132 #endif
133 
raw_irqentry_exit_cond_resched(void)134 void raw_irqentry_exit_cond_resched(void)
135 {
136 	if (!preempt_count()) {
137 		/* Sanity check RCU and thread stack */
138 		rcu_irq_exit_check_preempt();
139 		if (IS_ENABLED(CONFIG_DEBUG_ENTRY))
140 			WARN_ON_ONCE(!on_thread_stack());
141 		if (need_resched() && arch_irqentry_exit_need_resched())
142 			preempt_schedule_irq();
143 	}
144 }
145 #ifdef CONFIG_PREEMPT_DYNAMIC
146 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
147 DEFINE_STATIC_CALL(irqentry_exit_cond_resched, raw_irqentry_exit_cond_resched);
148 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
149 DEFINE_STATIC_KEY_TRUE(sk_dynamic_irqentry_exit_cond_resched);
dynamic_irqentry_exit_cond_resched(void)150 void dynamic_irqentry_exit_cond_resched(void)
151 {
152 	if (!static_branch_unlikely(&sk_dynamic_irqentry_exit_cond_resched))
153 		return;
154 	raw_irqentry_exit_cond_resched();
155 }
156 #endif
157 #endif
158 
irqentry_exit(struct pt_regs * regs,irqentry_state_t state)159 noinstr void irqentry_exit(struct pt_regs *regs, irqentry_state_t state)
160 {
161 	if (user_mode(regs))
162 		irqentry_exit_to_user_mode(regs);
163 	else
164 		irqentry_exit_to_kernel_mode(regs, state);
165 }
166 
irqentry_nmi_enter(struct pt_regs * regs)167 irqentry_state_t noinstr irqentry_nmi_enter(struct pt_regs *regs)
168 {
169 	irqentry_state_t irq_state;
170 
171 	irq_state.lockdep = lockdep_hardirqs_enabled();
172 
173 	__nmi_enter();
174 	lockdep_hardirqs_off(CALLER_ADDR0);
175 	lockdep_hardirq_enter();
176 	ct_nmi_enter();
177 
178 	instrumentation_begin();
179 	kmsan_unpoison_entry_regs(regs);
180 	trace_hardirqs_off_finish();
181 	ftrace_nmi_enter();
182 	instrumentation_end();
183 
184 	return irq_state;
185 }
186 
irqentry_nmi_exit(struct pt_regs * regs,irqentry_state_t irq_state)187 void noinstr irqentry_nmi_exit(struct pt_regs *regs, irqentry_state_t irq_state)
188 {
189 	instrumentation_begin();
190 	ftrace_nmi_exit();
191 	if (irq_state.lockdep) {
192 		trace_hardirqs_on_prepare();
193 		lockdep_hardirqs_on_prepare();
194 	}
195 	instrumentation_end();
196 
197 	ct_nmi_exit();
198 	lockdep_hardirq_exit();
199 	if (irq_state.lockdep)
200 		lockdep_hardirqs_on(CALLER_ADDR0);
201 	__nmi_exit();
202 }
203