projects/linux·blob master · refs

731 lines · 20.4 KiB

1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * transition.c - Kernel Live Patching transition functions
4 *
5 * Copyright (C) 2015-2016 Josh Poimboeuf <[email protected]>
6 */
7
8#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
10#include <linux/cpu.h>
11#include <linux/stacktrace.h>
12#include <linux/static_call.h>
13#include "core.h"
14#include "patch.h"
15#include "transition.h"
16
17#define MAX_STACK_ENTRIES 100
18static DEFINE_PER_CPU(unsigned long[MAX_STACK_ENTRIES], klp_stack_entries);
19
20#define STACK_ERR_BUF_SIZE 128
21
22#define SIGNALS_TIMEOUT 15
23
24struct klp_patch *klp_transition_patch;
25
26static int klp_target_state = KLP_TRANSITION_IDLE;
27
28static unsigned int klp_signals_cnt;
29
30/*
31 * When a livepatch is in progress, enable klp stack checking in
32 * schedule(). This helps CPU-bound kthreads get patched.
33 */
34
35DEFINE_STATIC_KEY_FALSE(klp_sched_try_switch_key);
36
37#define klp_resched_enable() static_branch_enable(&klp_sched_try_switch_key)
38#define klp_resched_disable() static_branch_disable(&klp_sched_try_switch_key)
39
40/*
41 * This work can be performed periodically to finish patching or unpatching any
42 * "straggler" tasks which failed to transition in the first attempt.
43 */
44static void klp_transition_work_fn(struct work_struct *work)
45{
46 mutex_lock(&klp_mutex);
47
48 if (klp_transition_patch)
49 klp_try_complete_transition();
50
51 mutex_unlock(&klp_mutex);
52}
53static DECLARE_DELAYED_WORK(klp_transition_work, klp_transition_work_fn);
54
55/*
56 * This function is just a stub to implement a hard force
57 * of synchronize_rcu(). This requires synchronizing
58 * tasks even in userspace and idle.
59 */
60static void klp_sync(struct work_struct *work)
61{
62}
63
64/*
65 * We allow to patch also functions where RCU is not watching,
66 * e.g. before user_exit(). We can not rely on the RCU infrastructure
67 * to do the synchronization. Instead hard force the sched synchronization.
68 *
69 * This approach allows to use RCU functions for manipulating func_stack
70 * safely.
71 */
72static void klp_synchronize_transition(void)
73{
74 schedule_on_each_cpu(klp_sync);
75}
76
77/*
78 * The transition to the target patch state is complete. Clean up the data
79 * structures.
80 */
81static void klp_complete_transition(void)
82{
83 struct klp_object *obj;
84 struct klp_func *func;
85 struct task_struct *g, *task;
86 unsigned int cpu;
87
88 pr_debug("'%s': completing %s transition\n",
89 klp_transition_patch->mod->name,
90 klp_target_state == KLP_TRANSITION_PATCHED ? "patching" : "unpatching");
91
92 if (klp_transition_patch->replace && klp_target_state == KLP_TRANSITION_PATCHED) {
93 klp_unpatch_replaced_patches(klp_transition_patch);
94 klp_discard_nops(klp_transition_patch);
95 }
96
97 if (klp_target_state == KLP_TRANSITION_UNPATCHED) {
98 /*
99 * All tasks have transitioned to KLP_TRANSITION_UNPATCHED so we can now
100 * remove the new functions from the func_stack.
101 */
102 klp_unpatch_objects(klp_transition_patch);
103
104 /*
105 * Make sure klp_ftrace_handler() can no longer see functions
106 * from this patch on the ops->func_stack. Otherwise, after
107 * func->transition gets cleared, the handler may choose a
108 * removed function.
109 */
110 klp_synchronize_transition();
111 }
112
113 klp_for_each_object(klp_transition_patch, obj)
114 klp_for_each_func(obj, func)
115 func->transition = false;
116
117 /* Prevent klp_ftrace_handler() from seeing KLP_TRANSITION_IDLE state */
118 if (klp_target_state == KLP_TRANSITION_PATCHED)
119 klp_synchronize_transition();
120
121 read_lock(&tasklist_lock);
122 for_each_process_thread(g, task) {
123 WARN_ON_ONCE(test_tsk_thread_flag(task, TIF_PATCH_PENDING));
124 task->patch_state = KLP_TRANSITION_IDLE;
125 }
126 read_unlock(&tasklist_lock);
127
128 for_each_possible_cpu(cpu) {
129 task = idle_task(cpu);
130 WARN_ON_ONCE(test_tsk_thread_flag(task, TIF_PATCH_PENDING));
131 task->patch_state = KLP_TRANSITION_IDLE;
132 }
133
134 klp_for_each_object(klp_transition_patch, obj) {
135 if (!klp_is_object_loaded(obj))
136 continue;
137 if (klp_target_state == KLP_TRANSITION_PATCHED)
138 klp_post_patch_callback(obj);
139 else if (klp_target_state == KLP_TRANSITION_UNPATCHED)
140 klp_post_unpatch_callback(obj);
141 }
142
143 pr_notice("'%s': %s complete\n", klp_transition_patch->mod->name,
144 klp_target_state == KLP_TRANSITION_PATCHED ? "patching" : "unpatching");
145
146 klp_target_state = KLP_TRANSITION_IDLE;
147 klp_transition_patch = NULL;
148}
149
150/*
151 * This is called in the error path, to cancel a transition before it has
152 * started, i.e. klp_init_transition() has been called but
153 * klp_start_transition() hasn't. If the transition *has* been started,
154 * klp_reverse_transition() should be used instead.
155 */
156void klp_cancel_transition(void)
157{
158 if (WARN_ON_ONCE(klp_target_state != KLP_TRANSITION_PATCHED))
159 return;
160
161 pr_debug("'%s': canceling patching transition, going to unpatch\n",
162 klp_transition_patch->mod->name);
163
164 klp_target_state = KLP_TRANSITION_UNPATCHED;
165 klp_complete_transition();
166}
167
168/*
169 * Switch the patched state of the task to the set of functions in the target
170 * patch state.
171 *
172 * NOTE: If task is not 'current', the caller must ensure the task is inactive.
173 * Otherwise klp_ftrace_handler() might read the wrong 'patch_state' value.
174 */
175void klp_update_patch_state(struct task_struct *task)
176{
177 /*
178 * A variant of synchronize_rcu() is used to allow patching functions
179 * where RCU is not watching, see klp_synchronize_transition().
180 */
181 preempt_disable_notrace();
182
183 /*
184 * This test_and_clear_tsk_thread_flag() call also serves as a read
185 * barrier (smp_rmb) for two cases:
186 *
187 * 1) Enforce the order of the TIF_PATCH_PENDING read and the
188 * klp_target_state read. The corresponding write barriers are in
189 * klp_init_transition() and klp_reverse_transition().
190 *
191 * 2) Enforce the order of the TIF_PATCH_PENDING read and a future read
192 * of func->transition, if klp_ftrace_handler() is called later on
193 * the same CPU. See __klp_disable_patch().
194 */
195 if (test_and_clear_tsk_thread_flag(task, TIF_PATCH_PENDING))
196 task->patch_state = READ_ONCE(klp_target_state);
197
198 preempt_enable_notrace();
199}
200
201/*
202 * Determine whether the given stack trace includes any references to a
203 * to-be-patched or to-be-unpatched function.
204 */
205static int klp_check_stack_func(struct klp_func *func, unsigned long *entries,
206 unsigned int nr_entries)
207{
208 unsigned long func_addr, func_size, address;
209 struct klp_ops *ops;
210 int i;
211
212 if (klp_target_state == KLP_TRANSITION_UNPATCHED) {
213 /*
214 * Check for the to-be-unpatched function
215 * (the func itself).
216 */
217 func_addr = (unsigned long)func->new_func;
218 func_size = func->new_size;
219 } else {
220 /*
221 * Check for the to-be-patched function
222 * (the previous func).
223 */
224 ops = klp_find_ops(func->old_func);
225
226 if (list_is_singular(&ops->func_stack)) {
227 /* original function */
228 func_addr = (unsigned long)func->old_func;
229 func_size = func->old_size;
230 } else {
231 /* previously patched function */
232 struct klp_func *prev;
233
234 prev = list_next_entry(func, stack_node);
235 func_addr = (unsigned long)prev->new_func;
236 func_size = prev->new_size;
237 }
238 }
239
240 for (i = 0; i < nr_entries; i++) {
241 address = entries[i];
242
243 if (address >= func_addr && address < func_addr + func_size)
244 return -EAGAIN;
245 }
246
247 return 0;
248}
249
250/*
251 * Determine whether it's safe to transition the task to the target patch state
252 * by looking for any to-be-patched or to-be-unpatched functions on its stack.
253 */
254static int klp_check_stack(struct task_struct *task, const char **oldname)
255{
256 unsigned long *entries = this_cpu_ptr(klp_stack_entries);
257 struct klp_object *obj;
258 struct klp_func *func;
259 int ret, nr_entries;
260
261 /* Protect 'klp_stack_entries' */
262 lockdep_assert_preemption_disabled();
263
264 ret = stack_trace_save_tsk_reliable(task, entries, MAX_STACK_ENTRIES);
265 if (ret < 0)
266 return -EINVAL;
267 nr_entries = ret;
268
269 klp_for_each_object(klp_transition_patch, obj) {
270 if (!obj->patched)
271 continue;
272 klp_for_each_func(obj, func) {
273 ret = klp_check_stack_func(func, entries, nr_entries);
274 if (ret) {
275 *oldname = func->old_name;
276 return -EADDRINUSE;
277 }
278 }
279 }
280
281 return 0;
282}
283
284static int klp_check_and_switch_task(struct task_struct *task, void *arg)
285{
286 int ret;
287
288 if (task_curr(task) && task != current)
289 return -EBUSY;
290
291 ret = klp_check_stack(task, arg);
292 if (ret)
293 return ret;
294
295 clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
296 task->patch_state = klp_target_state;
297 return 0;
298}
299
300/*
301 * Try to safely switch a task to the target patch state. If it's currently
302 * running, or it's sleeping on a to-be-patched or to-be-unpatched function, or
303 * if the stack is unreliable, return false.
304 */
305static bool klp_try_switch_task(struct task_struct *task)
306{
307 const char *old_name;
308 int ret;
309
310 /* check if this task has already switched over */
311 if (task->patch_state == klp_target_state)
312 return true;
313
314 /*
315 * For arches which don't have reliable stack traces, we have to rely
316 * on other methods (e.g., switching tasks at kernel exit).
317 */
318 if (!klp_have_reliable_stack())
319 return false;
320
321 /*
322 * Now try to check the stack for any to-be-patched or to-be-unpatched
323 * functions. If all goes well, switch the task to the target patch
324 * state.
325 */
326 if (task == current)
327 ret = klp_check_and_switch_task(current, &old_name);
328 else
329 ret = task_call_func(task, klp_check_and_switch_task, &old_name);
330
331 switch (ret) {
332 case 0: /* success */
333 break;
334
335 case -EBUSY: /* klp_check_and_switch_task() */
336 pr_debug("%s: %s:%d is running\n",
337 __func__, task->comm, task->pid);
338 break;
339 case -EINVAL: /* klp_check_and_switch_task() */
340 pr_debug("%s: %s:%d has an unreliable stack\n",
341 __func__, task->comm, task->pid);
342 break;
343 case -EADDRINUSE: /* klp_check_and_switch_task() */
344 pr_debug("%s: %s:%d is sleeping on function %s\n",
345 __func__, task->comm, task->pid, old_name);
346 break;
347
348 default:
349 pr_debug("%s: Unknown error code (%d) when trying to switch %s:%d\n",
350 __func__, ret, task->comm, task->pid);
351 break;
352 }
353
354 return !ret;
355}
356
357void __klp_sched_try_switch(void)
358{
359 /*
360 * This function is called from __schedule() while a context switch is
361 * about to happen. Preemption is already disabled and klp_mutex
362 * can't be acquired.
363 * Disabled preemption is used to prevent racing with other callers of
364 * klp_try_switch_task(). Thanks to task_call_func() they won't be
365 * able to switch to this task while it's running.
366 */
367 lockdep_assert_preemption_disabled();
368
369 if (likely(!klp_patch_pending(current)))
370 return;
371
372 /*
373 * Enforce the order of the TIF_PATCH_PENDING read above and the
374 * klp_target_state read in klp_try_switch_task(). The corresponding
375 * write barriers are in klp_init_transition() and
376 * klp_reverse_transition().
377 */
378 smp_rmb();
379
380 klp_try_switch_task(current);
381}
382
383/*
384 * Sends a fake signal to all non-kthread tasks with TIF_PATCH_PENDING set.
385 * Kthreads with TIF_PATCH_PENDING set are woken up.
386 */
387static void klp_send_signals(void)
388{
389 struct task_struct *g, *task;
390
391 if (klp_signals_cnt == SIGNALS_TIMEOUT)
392 pr_notice("signaling remaining tasks\n");
393
394 read_lock(&tasklist_lock);
395 for_each_process_thread(g, task) {
396 if (!klp_patch_pending(task))
397 continue;
398
399 /*
400 * There is a small race here. We could see TIF_PATCH_PENDING
401 * set and decide to wake up a kthread or send a fake signal.
402 * Meanwhile the task could migrate itself and the action
403 * would be meaningless. It is not serious though.
404 */
405 if (task->flags & PF_KTHREAD) {
406 /*
407 * Wake up a kthread which sleeps interruptedly and
408 * still has not been migrated.
409 */
410 wake_up_state(task, TASK_INTERRUPTIBLE);
411 } else {
412 /*
413 * Send fake signal to all non-kthread tasks which are
414 * still not migrated.
415 */
416 set_notify_signal(task);
417 }
418 }
419 read_unlock(&tasklist_lock);
420}
421
422/*
423 * Try to switch all remaining tasks to the target patch state by walking the
424 * stacks of sleeping tasks and looking for any to-be-patched or
425 * to-be-unpatched functions. If such functions are found, the task can't be
426 * switched yet.
427 *
428 * If any tasks are still stuck in the initial patch state, schedule a retry.
429 */
430void klp_try_complete_transition(void)
431{
432 unsigned int cpu;
433 struct task_struct *g, *task;
434 struct klp_patch *patch;
435 bool complete = true;
436
437 WARN_ON_ONCE(klp_target_state == KLP_TRANSITION_IDLE);
438
439 /*
440 * Try to switch the tasks to the target patch state by walking their
441 * stacks and looking for any to-be-patched or to-be-unpatched
442 * functions. If such functions are found on a stack, or if the stack
443 * is deemed unreliable, the task can't be switched yet.
444 *
445 * Usually this will transition most (or all) of the tasks on a system
446 * unless the patch includes changes to a very common function.
447 */
448 read_lock(&tasklist_lock);
449 for_each_process_thread(g, task)
450 if (!klp_try_switch_task(task))
451 complete = false;
452 read_unlock(&tasklist_lock);
453
454 /*
455 * Ditto for the idle "swapper" tasks.
456 */
457 cpus_read_lock();
458 for_each_possible_cpu(cpu) {
459 task = idle_task(cpu);
460 if (cpu_online(cpu)) {
461 if (!klp_try_switch_task(task)) {
462 complete = false;
463 /* Make idle task go through the main loop. */
464 wake_up_if_idle(cpu);
465 }
466 } else if (task->patch_state != klp_target_state) {
467 /* offline idle tasks can be switched immediately */
468 clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
469 task->patch_state = klp_target_state;
470 }
471 }
472 cpus_read_unlock();
473
474 if (!complete) {
475 if (klp_signals_cnt && !(klp_signals_cnt % SIGNALS_TIMEOUT))
476 klp_send_signals();
477 klp_signals_cnt++;
478
479 /*
480 * Some tasks weren't able to be switched over. Try again
481 * later and/or wait for other methods like kernel exit
482 * switching.
483 */
484 schedule_delayed_work(&klp_transition_work,
485 round_jiffies_relative(HZ));
486 return;
487 }
488
489 /* Done! Now cleanup the data structures. */
490 klp_resched_disable();
491 patch = klp_transition_patch;
492 klp_complete_transition();
493
494 /*
495 * It would make more sense to free the unused patches in
496 * klp_complete_transition() but it is called also
497 * from klp_cancel_transition().
498 */
499 if (!patch->enabled)
500 klp_free_patch_async(patch);
501 else if (patch->replace)
502 klp_free_replaced_patches_async(patch);
503}
504
505/*
506 * Start the transition to the specified target patch state so tasks can begin
507 * switching to it.
508 */
509void klp_start_transition(void)
510{
511 struct task_struct *g, *task;
512 unsigned int cpu;
513
514 WARN_ON_ONCE(klp_target_state == KLP_TRANSITION_IDLE);
515
516 pr_notice("'%s': starting %s transition\n",
517 klp_transition_patch->mod->name,
518 klp_target_state == KLP_TRANSITION_PATCHED ? "patching" : "unpatching");
519
520 /*
521 * Mark all normal tasks as needing a patch state update. They'll
522 * switch either in klp_try_complete_transition() or as they exit the
523 * kernel.
524 */
525 read_lock(&tasklist_lock);
526 for_each_process_thread(g, task)
527 if (task->patch_state != klp_target_state)
528 set_tsk_thread_flag(task, TIF_PATCH_PENDING);
529 read_unlock(&tasklist_lock);
530
531 /*
532 * Mark all idle tasks as needing a patch state update. They'll switch
533 * either in klp_try_complete_transition() or at the idle loop switch
534 * point.
535 */
536 for_each_possible_cpu(cpu) {
537 task = idle_task(cpu);
538 if (task->patch_state != klp_target_state)
539 set_tsk_thread_flag(task, TIF_PATCH_PENDING);
540 }
541
542 klp_resched_enable();
543
544 klp_signals_cnt = 0;
545}
546
547/*
548 * Initialize the global target patch state and all tasks to the initial patch
549 * state, and initialize all function transition states to true in preparation
550 * for patching or unpatching.
551 */
552void klp_init_transition(struct klp_patch *patch, int state)
553{
554 struct task_struct *g, *task;
555 unsigned int cpu;
556 struct klp_object *obj;
557 struct klp_func *func;
558 int initial_state = !state;
559
560 WARN_ON_ONCE(klp_target_state != KLP_TRANSITION_IDLE);
561
562 klp_transition_patch = patch;
563
564 /*
565 * Set the global target patch state which tasks will switch to. This
566 * has no effect until the TIF_PATCH_PENDING flags get set later.
567 */
568 klp_target_state = state;
569
570 pr_debug("'%s': initializing %s transition\n", patch->mod->name,
571 klp_target_state == KLP_TRANSITION_PATCHED ? "patching" : "unpatching");
572
573 /*
574 * Initialize all tasks to the initial patch state to prepare them for
575 * switching to the target state.
576 */
577 read_lock(&tasklist_lock);
578 for_each_process_thread(g, task) {
579 WARN_ON_ONCE(task->patch_state != KLP_TRANSITION_IDLE);
580 task->patch_state = initial_state;
581 }
582 read_unlock(&tasklist_lock);
583
584 /*
585 * Ditto for the idle "swapper" tasks.
586 */
587 for_each_possible_cpu(cpu) {
588 task = idle_task(cpu);
589 WARN_ON_ONCE(task->patch_state != KLP_TRANSITION_IDLE);
590 task->patch_state = initial_state;
591 }
592
593 /*
594 * Enforce the order of the task->patch_state initializations and the
595 * func->transition updates to ensure that klp_ftrace_handler() doesn't
596 * see a func in transition with a task->patch_state of KLP_TRANSITION_IDLE.
597 *
598 * Also enforce the order of the klp_target_state write and future
599 * TIF_PATCH_PENDING writes to ensure klp_update_patch_state() and
600 * __klp_sched_try_switch() don't set a task->patch_state to
601 * KLP_TRANSITION_IDLE.
602 */
603 smp_wmb();
604
605 /*
606 * Set the func transition states so klp_ftrace_handler() will know to
607 * switch to the transition logic.
608 *
609 * When patching, the funcs aren't yet in the func_stack and will be
610 * made visible to the ftrace handler shortly by the calls to
611 * klp_patch_object().
612 *
613 * When unpatching, the funcs are already in the func_stack and so are
614 * already visible to the ftrace handler.
615 */
616 klp_for_each_object(patch, obj)
617 klp_for_each_func(obj, func)
618 func->transition = true;
619}
620
621/*
622 * This function can be called in the middle of an existing transition to
623 * reverse the direction of the target patch state. This can be done to
624 * effectively cancel an existing enable or disable operation if there are any
625 * tasks which are stuck in the initial patch state.
626 */
627void klp_reverse_transition(void)
628{
629 unsigned int cpu;
630 struct task_struct *g, *task;
631
632 pr_debug("'%s': reversing transition from %s\n",
633 klp_transition_patch->mod->name,
634 klp_target_state == KLP_TRANSITION_PATCHED ? "patching to unpatching" :
635 "unpatching to patching");
636
637 /*
638 * Clear all TIF_PATCH_PENDING flags to prevent races caused by
639 * klp_update_patch_state() or __klp_sched_try_switch() running in
640 * parallel with the reverse transition.
641 */
642 read_lock(&tasklist_lock);
643 for_each_process_thread(g, task)
644 clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
645 read_unlock(&tasklist_lock);
646
647 for_each_possible_cpu(cpu)
648 clear_tsk_thread_flag(idle_task(cpu), TIF_PATCH_PENDING);
649
650 /*
651 * Make sure all existing invocations of klp_update_patch_state() and
652 * __klp_sched_try_switch() see the cleared TIF_PATCH_PENDING before
653 * starting the reverse transition.
654 */
655 klp_synchronize_transition();
656
657 /*
658 * All patching has stopped, now re-initialize the global variables to
659 * prepare for the reverse transition.
660 */
661 klp_transition_patch->enabled = !klp_transition_patch->enabled;
662 klp_target_state = !klp_target_state;
663
664 /*
665 * Enforce the order of the klp_target_state write and the
666 * TIF_PATCH_PENDING writes in klp_start_transition() to ensure
667 * klp_update_patch_state() and __klp_sched_try_switch() don't set
668 * task->patch_state to the wrong value.
669 */
670 smp_wmb();
671
672 klp_start_transition();
673}
674
675/* Called from copy_process() during fork */
676void klp_copy_process(struct task_struct *child)
677{
678
679 /*
680 * The parent process may have gone through a KLP transition since
681 * the thread flag was copied in setup_thread_stack earlier. Bring
682 * the task flag up to date with the parent here.
683 *
684 * The operation is serialized against all klp_*_transition()
685 * operations by the tasklist_lock. The only exceptions are
686 * klp_update_patch_state(current) and __klp_sched_try_switch(), but we
687 * cannot race with them because we are current.
688 */
689 if (test_tsk_thread_flag(current, TIF_PATCH_PENDING))
690 set_tsk_thread_flag(child, TIF_PATCH_PENDING);
691 else
692 clear_tsk_thread_flag(child, TIF_PATCH_PENDING);
693
694 child->patch_state = current->patch_state;
695}
696
697/*
698 * Drop TIF_PATCH_PENDING of all tasks on admin's request. This forces an
699 * existing transition to finish.
700 *
701 * NOTE: klp_update_patch_state(task) requires the task to be inactive or
702 * 'current'. This is not the case here and the consistency model could be
703 * broken. Administrator, who is the only one to execute the
704 * klp_force_transitions(), has to be aware of this.
705 */
706void klp_force_transition(void)
707{
708 struct klp_patch *patch;
709 struct task_struct *g, *task;
710 unsigned int cpu;
711
712 pr_warn("forcing remaining tasks to the patched state\n");
713
714 read_lock(&tasklist_lock);
715 for_each_process_thread(g, task)
716 klp_update_patch_state(task);
717 read_unlock(&tasklist_lock);
718
719 for_each_possible_cpu(cpu)
720 klp_update_patch_state(idle_task(cpu));
721
722 /* Set forced flag for patches being removed. */
723 if (klp_target_state == KLP_TRANSITION_UNPATCHED)
724 klp_transition_patch->forced = true;
725 else if (klp_transition_patch->replace) {
726 klp_for_each_patch(patch) {
727 if (patch != klp_transition_patch)
728 patch->forced = true;
729 }
730 }
731}