projects/linux·blob v6.0 · refs

477 lines · 11.9 KiB

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/livepatch.h>
9#include <linux/audit.h>
10#include <linux/tick.h>
11
12#include "common.h"
13
14#define CREATE_TRACE_POINTS
15#include <trace/events/syscalls.h>
16
17/* See comment for enter_from_user_mode() in entry-common.h */
18static __always_inline void __enter_from_user_mode(struct pt_regs *regs)
19{
20 arch_enter_from_user_mode(regs);
21 lockdep_hardirqs_off(CALLER_ADDR0);
22
23 CT_WARN_ON(ct_state() != CONTEXT_USER);
24 user_exit_irqoff();
25
26 instrumentation_begin();
27 trace_hardirqs_off_finish();
28 instrumentation_end();
29}
30
31void noinstr enter_from_user_mode(struct pt_regs *regs)
32{
33 __enter_from_user_mode(regs);
34}
35
36static inline void syscall_enter_audit(struct pt_regs *regs, long syscall)
37{
38 if (unlikely(audit_context())) {
39 unsigned long args[6];
40
41 syscall_get_arguments(current, regs, args);
42 audit_syscall_entry(syscall, args[0], args[1], args[2], args[3]);
43 }
44}
45
46static long syscall_trace_enter(struct pt_regs *regs, long syscall,
47 unsigned long work)
48{
49 long ret = 0;
50
51 /*
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.
55 */
56 if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
57 if (syscall_user_dispatch(regs))
58 return -1L;
59 }
60
61 /* Handle ptrace */
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))
65 return -1L;
66 }
67
68 /* Do seccomp after ptrace, to catch any tracer changes. */
69 if (work & SYSCALL_WORK_SECCOMP) {
70 ret = __secure_computing(NULL);
71 if (ret == -1L)
72 return ret;
73 }
74
75 /* Either of the above might have changed the syscall number */
76 syscall = syscall_get_nr(current, regs);
77
78 if (unlikely(work & SYSCALL_WORK_SYSCALL_TRACEPOINT))
79 trace_sys_enter(regs, syscall);
80
81 syscall_enter_audit(regs, syscall);
82
83 return ret ? : syscall;
84}
85
86static __always_inline long
87__syscall_enter_from_user_work(struct pt_regs *regs, long syscall)
88{
89 unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
90
91 if (work & SYSCALL_WORK_ENTER)
92 syscall = syscall_trace_enter(regs, syscall, work);
93
94 return syscall;
95}
96
97long syscall_enter_from_user_mode_work(struct pt_regs *regs, long syscall)
98{
99 return __syscall_enter_from_user_work(regs, syscall);
100}
101
102noinstr long syscall_enter_from_user_mode(struct pt_regs *regs, long syscall)
103{
104 long ret;
105
106 __enter_from_user_mode(regs);
107
108 instrumentation_begin();
109 local_irq_enable();
110 ret = __syscall_enter_from_user_work(regs, syscall);
111 instrumentation_end();
112
113 return ret;
114}
115
116noinstr void syscall_enter_from_user_mode_prepare(struct pt_regs *regs)
117{
118 __enter_from_user_mode(regs);
119 instrumentation_begin();
120 local_irq_enable();
121 instrumentation_end();
122}
123
124/* See comment for exit_to_user_mode() in entry-common.h */
125static __always_inline void __exit_to_user_mode(void)
126{
127 instrumentation_begin();
128 trace_hardirqs_on_prepare();
129 lockdep_hardirqs_on_prepare();
130 instrumentation_end();
131
132 user_enter_irqoff();
133 arch_exit_to_user_mode();
134 lockdep_hardirqs_on(CALLER_ADDR0);
135}
136
137void noinstr exit_to_user_mode(void)
138{
139 __exit_to_user_mode();
140}
141
142/* Workaround to allow gradual conversion of architecture code */
143void __weak arch_do_signal_or_restart(struct pt_regs *regs) { }
144
145static unsigned long exit_to_user_mode_loop(struct pt_regs *regs,
146 unsigned long ti_work)
147{
148 /*
149 * Before returning to user space ensure that all pending work
150 * items have been completed.
151 */
152 while (ti_work & EXIT_TO_USER_MODE_WORK) {
153
154 local_irq_enable_exit_to_user(ti_work);
155
156 if (ti_work & _TIF_NEED_RESCHED)
157 schedule();
158
159 if (ti_work & _TIF_UPROBE)
160 uprobe_notify_resume(regs);
161
162 if (ti_work & _TIF_PATCH_PENDING)
163 klp_update_patch_state(current);
164
165 if (ti_work & (_TIF_SIGPENDING | _TIF_NOTIFY_SIGNAL))
166 arch_do_signal_or_restart(regs);
167
168 if (ti_work & _TIF_NOTIFY_RESUME)
169 resume_user_mode_work(regs);
170
171 /* Architecture specific TIF work */
172 arch_exit_to_user_mode_work(regs, ti_work);
173
174 /*
175 * Disable interrupts and reevaluate the work flags as they
176 * might have changed while interrupts and preemption was
177 * enabled above.
178 */
179 local_irq_disable_exit_to_user();
180
181 /* Check if any of the above work has queued a deferred wakeup */
182 tick_nohz_user_enter_prepare();
183
184 ti_work = read_thread_flags();
185 }
186
187 /* Return the latest work state for arch_exit_to_user_mode() */
188 return ti_work;
189}
190
191static void exit_to_user_mode_prepare(struct pt_regs *regs)
192{
193 unsigned long ti_work = read_thread_flags();
194
195 lockdep_assert_irqs_disabled();
196
197 /* Flush pending rcuog wakeup before the last need_resched() check */
198 tick_nohz_user_enter_prepare();
199
200 if (unlikely(ti_work & EXIT_TO_USER_MODE_WORK))
201 ti_work = exit_to_user_mode_loop(regs, ti_work);
202
203 arch_exit_to_user_mode_prepare(regs, ti_work);
204
205 /* Ensure that the address limit is intact and no locks are held */
206 addr_limit_user_check();
207 kmap_assert_nomap();
208 lockdep_assert_irqs_disabled();
209 lockdep_sys_exit();
210}
211
212/*
213 * If SYSCALL_EMU is set, then the only reason to report is when
214 * SINGLESTEP is set (i.e. PTRACE_SYSEMU_SINGLESTEP). This syscall
215 * instruction has been already reported in syscall_enter_from_user_mode().
216 */
217static inline bool report_single_step(unsigned long work)
218{
219 if (work & SYSCALL_WORK_SYSCALL_EMU)
220 return false;
221
222 return work & SYSCALL_WORK_SYSCALL_EXIT_TRAP;
223}
224
225static void syscall_exit_work(struct pt_regs *regs, unsigned long work)
226{
227 bool step;
228
229 /*
230 * If the syscall was rolled back due to syscall user dispatching,
231 * then the tracers below are not invoked for the same reason as
232 * the entry side was not invoked in syscall_trace_enter(): The ABI
233 * of these syscalls is unknown.
234 */
235 if (work & SYSCALL_WORK_SYSCALL_USER_DISPATCH) {
236 if (unlikely(current->syscall_dispatch.on_dispatch)) {
237 current->syscall_dispatch.on_dispatch = false;
238 return;
239 }
240 }
241
242 audit_syscall_exit(regs);
243
244 if (work & SYSCALL_WORK_SYSCALL_TRACEPOINT)
245 trace_sys_exit(regs, syscall_get_return_value(current, regs));
246
247 step = report_single_step(work);
248 if (step || work & SYSCALL_WORK_SYSCALL_TRACE)
249 ptrace_report_syscall_exit(regs, step);
250}
251
252/*
253 * Syscall specific exit to user mode preparation. Runs with interrupts
254 * enabled.
255 */
256static void syscall_exit_to_user_mode_prepare(struct pt_regs *regs)
257{
258 unsigned long work = READ_ONCE(current_thread_info()->syscall_work);
259 unsigned long nr = syscall_get_nr(current, regs);
260
261 CT_WARN_ON(ct_state() != CONTEXT_KERNEL);
262
263 if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
264 if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr))
265 local_irq_enable();
266 }
267
268 rseq_syscall(regs);
269
270 /*
271 * Do one-time syscall specific work. If these work items are
272 * enabled, we want to run them exactly once per syscall exit with
273 * interrupts enabled.
274 */
275 if (unlikely(work & SYSCALL_WORK_EXIT))
276 syscall_exit_work(regs, work);
277}
278
279static __always_inline void __syscall_exit_to_user_mode_work(struct pt_regs *regs)
280{
281 syscall_exit_to_user_mode_prepare(regs);
282 local_irq_disable_exit_to_user();
283 exit_to_user_mode_prepare(regs);
284}
285
286void syscall_exit_to_user_mode_work(struct pt_regs *regs)
287{
288 __syscall_exit_to_user_mode_work(regs);
289}
290
291__visible noinstr void syscall_exit_to_user_mode(struct pt_regs *regs)
292{
293 instrumentation_begin();
294 __syscall_exit_to_user_mode_work(regs);
295 instrumentation_end();
296 __exit_to_user_mode();
297}
298
299noinstr void irqentry_enter_from_user_mode(struct pt_regs *regs)
300{
301 __enter_from_user_mode(regs);
302}
303
304noinstr void irqentry_exit_to_user_mode(struct pt_regs *regs)
305{
306 instrumentation_begin();
307 exit_to_user_mode_prepare(regs);
308 instrumentation_end();
309 __exit_to_user_mode();
310}
311
312noinstr irqentry_state_t irqentry_enter(struct pt_regs *regs)
313{
314 irqentry_state_t ret = {
315 .exit_rcu = false,
316 };
317
318 if (user_mode(regs)) {
319 irqentry_enter_from_user_mode(regs);
320 return ret;
321 }
322
323 /*
324 * If this entry hit the idle task invoke ct_irq_enter() whether
325 * RCU is watching or not.
326 *
327 * Interrupts can nest when the first interrupt invokes softirq
328 * processing on return which enables interrupts.
329 *
330 * Scheduler ticks in the idle task can mark quiescent state and
331 * terminate a grace period, if and only if the timer interrupt is
332 * not nested into another interrupt.
333 *
334 * Checking for rcu_is_watching() here would prevent the nesting
335 * interrupt to invoke ct_irq_enter(). If that nested interrupt is
336 * the tick then rcu_flavor_sched_clock_irq() would wrongfully
337 * assume that it is the first interrupt and eventually claim
338 * quiescent state and end grace periods prematurely.
339 *
340 * Unconditionally invoke ct_irq_enter() so RCU state stays
341 * consistent.
342 *
343 * TINY_RCU does not support EQS, so let the compiler eliminate
344 * this part when enabled.
345 */
346 if (!IS_ENABLED(CONFIG_TINY_RCU) && is_idle_task(current)) {
347 /*
348 * If RCU is not watching then the same careful
349 * sequence vs. lockdep and tracing is required
350 * as in irqentry_enter_from_user_mode().
351 */
352 lockdep_hardirqs_off(CALLER_ADDR0);
353 ct_irq_enter();
354 instrumentation_begin();
355 trace_hardirqs_off_finish();
356 instrumentation_end();
357
358 ret.exit_rcu = true;
359 return ret;
360 }
361
362 /*
363 * If RCU is watching then RCU only wants to check whether it needs
364 * to restart the tick in NOHZ mode. rcu_irq_enter_check_tick()
365 * already contains a warning when RCU is not watching, so no point
366 * in having another one here.
367 */
368 lockdep_hardirqs_off(CALLER_ADDR0);
369 instrumentation_begin();
370 rcu_irq_enter_check_tick();
371 trace_hardirqs_off_finish();
372 instrumentation_end();
373
374 return ret;
375}
376
377void raw_irqentry_exit_cond_resched(void)
378{
379 if (!preempt_count()) {
380 /* Sanity check RCU and thread stack */
381 rcu_irq_exit_check_preempt();
382 if (IS_ENABLED(CONFIG_DEBUG_ENTRY))
383 WARN_ON_ONCE(!on_thread_stack());
384 if (need_resched())
385 preempt_schedule_irq();
386 }
387}
388#ifdef CONFIG_PREEMPT_DYNAMIC
389#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
390DEFINE_STATIC_CALL(irqentry_exit_cond_resched, raw_irqentry_exit_cond_resched);
391#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
392DEFINE_STATIC_KEY_TRUE(sk_dynamic_irqentry_exit_cond_resched);
393void dynamic_irqentry_exit_cond_resched(void)
394{
395 if (!static_branch_unlikely(&sk_dynamic_irqentry_exit_cond_resched))
396 return;
397 raw_irqentry_exit_cond_resched();
398}
399#endif
400#endif
401
402noinstr void irqentry_exit(struct pt_regs *regs, irqentry_state_t state)
403{
404 lockdep_assert_irqs_disabled();
405
406 /* Check whether this returns to user mode */
407 if (user_mode(regs)) {
408 irqentry_exit_to_user_mode(regs);
409 } else if (!regs_irqs_disabled(regs)) {
410 /*
411 * If RCU was not watching on entry this needs to be done
412 * carefully and needs the same ordering of lockdep/tracing
413 * and RCU as the return to user mode path.
414 */
415 if (state.exit_rcu) {
416 instrumentation_begin();
417 /* Tell the tracer that IRET will enable interrupts */
418 trace_hardirqs_on_prepare();
419 lockdep_hardirqs_on_prepare();
420 instrumentation_end();
421 ct_irq_exit();
422 lockdep_hardirqs_on(CALLER_ADDR0);
423 return;
424 }
425
426 instrumentation_begin();
427 if (IS_ENABLED(CONFIG_PREEMPTION))
428 irqentry_exit_cond_resched();
429
430 /* Covers both tracing and lockdep */
431 trace_hardirqs_on();
432 instrumentation_end();
433 } else {
434 /*
435 * IRQ flags state is correct already. Just tell RCU if it
436 * was not watching on entry.
437 */
438 if (state.exit_rcu)
439 ct_irq_exit();
440 }
441}
442
443irqentry_state_t noinstr irqentry_nmi_enter(struct pt_regs *regs)
444{
445 irqentry_state_t irq_state;
446
447 irq_state.lockdep = lockdep_hardirqs_enabled();
448
449 __nmi_enter();
450 lockdep_hardirqs_off(CALLER_ADDR0);
451 lockdep_hardirq_enter();
452 ct_nmi_enter();
453
454 instrumentation_begin();
455 trace_hardirqs_off_finish();
456 ftrace_nmi_enter();
457 instrumentation_end();
458
459 return irq_state;
460}
461
462void noinstr irqentry_nmi_exit(struct pt_regs *regs, irqentry_state_t irq_state)
463{
464 instrumentation_begin();
465 ftrace_nmi_exit();
466 if (irq_state.lockdep) {
467 trace_hardirqs_on_prepare();
468 lockdep_hardirqs_on_prepare();
469 }
470 instrumentation_end();
471
472 ct_nmi_exit();
473 lockdep_hardirq_exit();
474 if (irq_state.lockdep)
475 lockdep_hardirqs_on(CALLER_ADDR0);
476 __nmi_exit();
477}