xref: /linux/kernel/entry/common.c (revision 257ca10c7317d4a424e48bb95d14ca53a1f1dd6f)
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 	syscall_enter_audit(regs, syscall);
64 
65 	return ret ? : syscall;
66 }
67 
68 noinstr 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 */
77 void __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  */
135 static 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 
143 static 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  */
174 static 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 
197 static __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 
204 void 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 
217 noinstr void irqentry_enter_from_user_mode(struct pt_regs *regs)
218 {
219 	enter_from_user_mode(regs);
220 }
221 
222 noinstr 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 
230 noinstr 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 
297 void 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)
310 DEFINE_STATIC_CALL(irqentry_exit_cond_resched, raw_irqentry_exit_cond_resched);
311 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
312 DEFINE_STATIC_KEY_TRUE(sk_dynamic_irqentry_exit_cond_resched);
313 void 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 
322 noinstr 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 
363 irqentry_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 
383 void 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 }
399