Loading...
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 <jpoimboe@redhat.com>
6 */
7
8#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
10#include <linux/cpu.h>
11#include <linux/stacktrace.h>
12#include "core.h"
13#include "patch.h"
14#include "transition.h"
15
16#define MAX_STACK_ENTRIES 100
17#define STACK_ERR_BUF_SIZE 128
18
19#define SIGNALS_TIMEOUT 15
20
21struct klp_patch *klp_transition_patch;
22
23static int klp_target_state = KLP_UNDEFINED;
24
25static unsigned int klp_signals_cnt;
26
27/*
28 * This work can be performed periodically to finish patching or unpatching any
29 * "straggler" tasks which failed to transition in the first attempt.
30 */
31static void klp_transition_work_fn(struct work_struct *work)
32{
33 mutex_lock(&klp_mutex);
34
35 if (klp_transition_patch)
36 klp_try_complete_transition();
37
38 mutex_unlock(&klp_mutex);
39}
40static DECLARE_DELAYED_WORK(klp_transition_work, klp_transition_work_fn);
41
42/*
43 * This function is just a stub to implement a hard force
44 * of synchronize_rcu(). This requires synchronizing
45 * tasks even in userspace and idle.
46 */
47static void klp_sync(struct work_struct *work)
48{
49}
50
51/*
52 * We allow to patch also functions where RCU is not watching,
53 * e.g. before user_exit(). We can not rely on the RCU infrastructure
54 * to do the synchronization. Instead hard force the sched synchronization.
55 *
56 * This approach allows to use RCU functions for manipulating func_stack
57 * safely.
58 */
59static void klp_synchronize_transition(void)
60{
61 schedule_on_each_cpu(klp_sync);
62}
63
64/*
65 * The transition to the target patch state is complete. Clean up the data
66 * structures.
67 */
68static void klp_complete_transition(void)
69{
70 struct klp_object *obj;
71 struct klp_func *func;
72 struct task_struct *g, *task;
73 unsigned int cpu;
74
75 pr_debug("'%s': completing %s transition\n",
76 klp_transition_patch->mod->name,
77 klp_target_state == KLP_PATCHED ? "patching" : "unpatching");
78
79 if (klp_transition_patch->replace && klp_target_state == KLP_PATCHED) {
80 klp_unpatch_replaced_patches(klp_transition_patch);
81 klp_discard_nops(klp_transition_patch);
82 }
83
84 if (klp_target_state == KLP_UNPATCHED) {
85 /*
86 * All tasks have transitioned to KLP_UNPATCHED so we can now
87 * remove the new functions from the func_stack.
88 */
89 klp_unpatch_objects(klp_transition_patch);
90
91 /*
92 * Make sure klp_ftrace_handler() can no longer see functions
93 * from this patch on the ops->func_stack. Otherwise, after
94 * func->transition gets cleared, the handler may choose a
95 * removed function.
96 */
97 klp_synchronize_transition();
98 }
99
100 klp_for_each_object(klp_transition_patch, obj)
101 klp_for_each_func(obj, func)
102 func->transition = false;
103
104 /* Prevent klp_ftrace_handler() from seeing KLP_UNDEFINED state */
105 if (klp_target_state == KLP_PATCHED)
106 klp_synchronize_transition();
107
108 read_lock(&tasklist_lock);
109 for_each_process_thread(g, task) {
110 WARN_ON_ONCE(test_tsk_thread_flag(task, TIF_PATCH_PENDING));
111 task->patch_state = KLP_UNDEFINED;
112 }
113 read_unlock(&tasklist_lock);
114
115 for_each_possible_cpu(cpu) {
116 task = idle_task(cpu);
117 WARN_ON_ONCE(test_tsk_thread_flag(task, TIF_PATCH_PENDING));
118 task->patch_state = KLP_UNDEFINED;
119 }
120
121 klp_for_each_object(klp_transition_patch, obj) {
122 if (!klp_is_object_loaded(obj))
123 continue;
124 if (klp_target_state == KLP_PATCHED)
125 klp_post_patch_callback(obj);
126 else if (klp_target_state == KLP_UNPATCHED)
127 klp_post_unpatch_callback(obj);
128 }
129
130 pr_notice("'%s': %s complete\n", klp_transition_patch->mod->name,
131 klp_target_state == KLP_PATCHED ? "patching" : "unpatching");
132
133 klp_target_state = KLP_UNDEFINED;
134 klp_transition_patch = NULL;
135}
136
137/*
138 * This is called in the error path, to cancel a transition before it has
139 * started, i.e. klp_init_transition() has been called but
140 * klp_start_transition() hasn't. If the transition *has* been started,
141 * klp_reverse_transition() should be used instead.
142 */
143void klp_cancel_transition(void)
144{
145 if (WARN_ON_ONCE(klp_target_state != KLP_PATCHED))
146 return;
147
148 pr_debug("'%s': canceling patching transition, going to unpatch\n",
149 klp_transition_patch->mod->name);
150
151 klp_target_state = KLP_UNPATCHED;
152 klp_complete_transition();
153}
154
155/*
156 * Switch the patched state of the task to the set of functions in the target
157 * patch state.
158 *
159 * NOTE: If task is not 'current', the caller must ensure the task is inactive.
160 * Otherwise klp_ftrace_handler() might read the wrong 'patch_state' value.
161 */
162void klp_update_patch_state(struct task_struct *task)
163{
164 /*
165 * A variant of synchronize_rcu() is used to allow patching functions
166 * where RCU is not watching, see klp_synchronize_transition().
167 */
168 preempt_disable_notrace();
169
170 /*
171 * This test_and_clear_tsk_thread_flag() call also serves as a read
172 * barrier (smp_rmb) for two cases:
173 *
174 * 1) Enforce the order of the TIF_PATCH_PENDING read and the
175 * klp_target_state read. The corresponding write barrier is in
176 * klp_init_transition().
177 *
178 * 2) Enforce the order of the TIF_PATCH_PENDING read and a future read
179 * of func->transition, if klp_ftrace_handler() is called later on
180 * the same CPU. See __klp_disable_patch().
181 */
182 if (test_and_clear_tsk_thread_flag(task, TIF_PATCH_PENDING))
183 task->patch_state = READ_ONCE(klp_target_state);
184
185 preempt_enable_notrace();
186}
187
188/*
189 * Determine whether the given stack trace includes any references to a
190 * to-be-patched or to-be-unpatched function.
191 */
192static int klp_check_stack_func(struct klp_func *func, unsigned long *entries,
193 unsigned int nr_entries)
194{
195 unsigned long func_addr, func_size, address;
196 struct klp_ops *ops;
197 int i;
198
199 if (klp_target_state == KLP_UNPATCHED) {
200 /*
201 * Check for the to-be-unpatched function
202 * (the func itself).
203 */
204 func_addr = (unsigned long)func->new_func;
205 func_size = func->new_size;
206 } else {
207 /*
208 * Check for the to-be-patched function
209 * (the previous func).
210 */
211 ops = klp_find_ops(func->old_func);
212
213 if (list_is_singular(&ops->func_stack)) {
214 /* original function */
215 func_addr = (unsigned long)func->old_func;
216 func_size = func->old_size;
217 } else {
218 /* previously patched function */
219 struct klp_func *prev;
220
221 prev = list_next_entry(func, stack_node);
222 func_addr = (unsigned long)prev->new_func;
223 func_size = prev->new_size;
224 }
225 }
226
227 for (i = 0; i < nr_entries; i++) {
228 address = entries[i];
229
230 if (address >= func_addr && address < func_addr + func_size)
231 return -EAGAIN;
232 }
233
234 return 0;
235}
236
237/*
238 * Determine whether it's safe to transition the task to the target patch state
239 * by looking for any to-be-patched or to-be-unpatched functions on its stack.
240 */
241static int klp_check_stack(struct task_struct *task, const char **oldname)
242{
243 static unsigned long entries[MAX_STACK_ENTRIES];
244 struct klp_object *obj;
245 struct klp_func *func;
246 int ret, nr_entries;
247
248 ret = stack_trace_save_tsk_reliable(task, entries, ARRAY_SIZE(entries));
249 if (ret < 0)
250 return -EINVAL;
251 nr_entries = ret;
252
253 klp_for_each_object(klp_transition_patch, obj) {
254 if (!obj->patched)
255 continue;
256 klp_for_each_func(obj, func) {
257 ret = klp_check_stack_func(func, entries, nr_entries);
258 if (ret) {
259 *oldname = func->old_name;
260 return -EADDRINUSE;
261 }
262 }
263 }
264
265 return 0;
266}
267
268static int klp_check_and_switch_task(struct task_struct *task, void *arg)
269{
270 int ret;
271
272 if (task_curr(task) && task != current)
273 return -EBUSY;
274
275 ret = klp_check_stack(task, arg);
276 if (ret)
277 return ret;
278
279 clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
280 task->patch_state = klp_target_state;
281 return 0;
282}
283
284/*
285 * Try to safely switch a task to the target patch state. If it's currently
286 * running, or it's sleeping on a to-be-patched or to-be-unpatched function, or
287 * if the stack is unreliable, return false.
288 */
289static bool klp_try_switch_task(struct task_struct *task)
290{
291 const char *old_name;
292 int ret;
293
294 /* check if this task has already switched over */
295 if (task->patch_state == klp_target_state)
296 return true;
297
298 /*
299 * For arches which don't have reliable stack traces, we have to rely
300 * on other methods (e.g., switching tasks at kernel exit).
301 */
302 if (!klp_have_reliable_stack())
303 return false;
304
305 /*
306 * Now try to check the stack for any to-be-patched or to-be-unpatched
307 * functions. If all goes well, switch the task to the target patch
308 * state.
309 */
310 ret = task_call_func(task, klp_check_and_switch_task, &old_name);
311 switch (ret) {
312 case 0: /* success */
313 break;
314
315 case -EBUSY: /* klp_check_and_switch_task() */
316 pr_debug("%s: %s:%d is running\n",
317 __func__, task->comm, task->pid);
318 break;
319 case -EINVAL: /* klp_check_and_switch_task() */
320 pr_debug("%s: %s:%d has an unreliable stack\n",
321 __func__, task->comm, task->pid);
322 break;
323 case -EADDRINUSE: /* klp_check_and_switch_task() */
324 pr_debug("%s: %s:%d is sleeping on function %s\n",
325 __func__, task->comm, task->pid, old_name);
326 break;
327
328 default:
329 pr_debug("%s: Unknown error code (%d) when trying to switch %s:%d\n",
330 __func__, ret, task->comm, task->pid);
331 break;
332 }
333
334 return !ret;
335}
336
337/*
338 * Sends a fake signal to all non-kthread tasks with TIF_PATCH_PENDING set.
339 * Kthreads with TIF_PATCH_PENDING set are woken up.
340 */
341static void klp_send_signals(void)
342{
343 struct task_struct *g, *task;
344
345 if (klp_signals_cnt == SIGNALS_TIMEOUT)
346 pr_notice("signaling remaining tasks\n");
347
348 read_lock(&tasklist_lock);
349 for_each_process_thread(g, task) {
350 if (!klp_patch_pending(task))
351 continue;
352
353 /*
354 * There is a small race here. We could see TIF_PATCH_PENDING
355 * set and decide to wake up a kthread or send a fake signal.
356 * Meanwhile the task could migrate itself and the action
357 * would be meaningless. It is not serious though.
358 */
359 if (task->flags & PF_KTHREAD) {
360 /*
361 * Wake up a kthread which sleeps interruptedly and
362 * still has not been migrated.
363 */
364 wake_up_state(task, TASK_INTERRUPTIBLE);
365 } else {
366 /*
367 * Send fake signal to all non-kthread tasks which are
368 * still not migrated.
369 */
370 set_notify_signal(task);
371 }
372 }
373 read_unlock(&tasklist_lock);
374}
375
376/*
377 * Try to switch all remaining tasks to the target patch state by walking the
378 * stacks of sleeping tasks and looking for any to-be-patched or
379 * to-be-unpatched functions. If such functions are found, the task can't be
380 * switched yet.
381 *
382 * If any tasks are still stuck in the initial patch state, schedule a retry.
383 */
384void klp_try_complete_transition(void)
385{
386 unsigned int cpu;
387 struct task_struct *g, *task;
388 struct klp_patch *patch;
389 bool complete = true;
390
391 WARN_ON_ONCE(klp_target_state == KLP_UNDEFINED);
392
393 /*
394 * Try to switch the tasks to the target patch state by walking their
395 * stacks and looking for any to-be-patched or to-be-unpatched
396 * functions. If such functions are found on a stack, or if the stack
397 * is deemed unreliable, the task can't be switched yet.
398 *
399 * Usually this will transition most (or all) of the tasks on a system
400 * unless the patch includes changes to a very common function.
401 */
402 read_lock(&tasklist_lock);
403 for_each_process_thread(g, task)
404 if (!klp_try_switch_task(task))
405 complete = false;
406 read_unlock(&tasklist_lock);
407
408 /*
409 * Ditto for the idle "swapper" tasks.
410 */
411 cpus_read_lock();
412 for_each_possible_cpu(cpu) {
413 task = idle_task(cpu);
414 if (cpu_online(cpu)) {
415 if (!klp_try_switch_task(task)) {
416 complete = false;
417 /* Make idle task go through the main loop. */
418 wake_up_if_idle(cpu);
419 }
420 } else if (task->patch_state != klp_target_state) {
421 /* offline idle tasks can be switched immediately */
422 clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
423 task->patch_state = klp_target_state;
424 }
425 }
426 cpus_read_unlock();
427
428 if (!complete) {
429 if (klp_signals_cnt && !(klp_signals_cnt % SIGNALS_TIMEOUT))
430 klp_send_signals();
431 klp_signals_cnt++;
432
433 /*
434 * Some tasks weren't able to be switched over. Try again
435 * later and/or wait for other methods like kernel exit
436 * switching.
437 */
438 schedule_delayed_work(&klp_transition_work,
439 round_jiffies_relative(HZ));
440 return;
441 }
442
443 /* we're done, now cleanup the data structures */
444 patch = klp_transition_patch;
445 klp_complete_transition();
446
447 /*
448 * It would make more sense to free the unused patches in
449 * klp_complete_transition() but it is called also
450 * from klp_cancel_transition().
451 */
452 if (!patch->enabled)
453 klp_free_patch_async(patch);
454 else if (patch->replace)
455 klp_free_replaced_patches_async(patch);
456}
457
458/*
459 * Start the transition to the specified target patch state so tasks can begin
460 * switching to it.
461 */
462void klp_start_transition(void)
463{
464 struct task_struct *g, *task;
465 unsigned int cpu;
466
467 WARN_ON_ONCE(klp_target_state == KLP_UNDEFINED);
468
469 pr_notice("'%s': starting %s transition\n",
470 klp_transition_patch->mod->name,
471 klp_target_state == KLP_PATCHED ? "patching" : "unpatching");
472
473 /*
474 * Mark all normal tasks as needing a patch state update. They'll
475 * switch either in klp_try_complete_transition() or as they exit the
476 * kernel.
477 */
478 read_lock(&tasklist_lock);
479 for_each_process_thread(g, task)
480 if (task->patch_state != klp_target_state)
481 set_tsk_thread_flag(task, TIF_PATCH_PENDING);
482 read_unlock(&tasklist_lock);
483
484 /*
485 * Mark all idle tasks as needing a patch state update. They'll switch
486 * either in klp_try_complete_transition() or at the idle loop switch
487 * point.
488 */
489 for_each_possible_cpu(cpu) {
490 task = idle_task(cpu);
491 if (task->patch_state != klp_target_state)
492 set_tsk_thread_flag(task, TIF_PATCH_PENDING);
493 }
494
495 klp_signals_cnt = 0;
496}
497
498/*
499 * Initialize the global target patch state and all tasks to the initial patch
500 * state, and initialize all function transition states to true in preparation
501 * for patching or unpatching.
502 */
503void klp_init_transition(struct klp_patch *patch, int state)
504{
505 struct task_struct *g, *task;
506 unsigned int cpu;
507 struct klp_object *obj;
508 struct klp_func *func;
509 int initial_state = !state;
510
511 WARN_ON_ONCE(klp_target_state != KLP_UNDEFINED);
512
513 klp_transition_patch = patch;
514
515 /*
516 * Set the global target patch state which tasks will switch to. This
517 * has no effect until the TIF_PATCH_PENDING flags get set later.
518 */
519 klp_target_state = state;
520
521 pr_debug("'%s': initializing %s transition\n", patch->mod->name,
522 klp_target_state == KLP_PATCHED ? "patching" : "unpatching");
523
524 /*
525 * Initialize all tasks to the initial patch state to prepare them for
526 * switching to the target state.
527 */
528 read_lock(&tasklist_lock);
529 for_each_process_thread(g, task) {
530 WARN_ON_ONCE(task->patch_state != KLP_UNDEFINED);
531 task->patch_state = initial_state;
532 }
533 read_unlock(&tasklist_lock);
534
535 /*
536 * Ditto for the idle "swapper" tasks.
537 */
538 for_each_possible_cpu(cpu) {
539 task = idle_task(cpu);
540 WARN_ON_ONCE(task->patch_state != KLP_UNDEFINED);
541 task->patch_state = initial_state;
542 }
543
544 /*
545 * Enforce the order of the task->patch_state initializations and the
546 * func->transition updates to ensure that klp_ftrace_handler() doesn't
547 * see a func in transition with a task->patch_state of KLP_UNDEFINED.
548 *
549 * Also enforce the order of the klp_target_state write and future
550 * TIF_PATCH_PENDING writes to ensure klp_update_patch_state() doesn't
551 * set a task->patch_state to KLP_UNDEFINED.
552 */
553 smp_wmb();
554
555 /*
556 * Set the func transition states so klp_ftrace_handler() will know to
557 * switch to the transition logic.
558 *
559 * When patching, the funcs aren't yet in the func_stack and will be
560 * made visible to the ftrace handler shortly by the calls to
561 * klp_patch_object().
562 *
563 * When unpatching, the funcs are already in the func_stack and so are
564 * already visible to the ftrace handler.
565 */
566 klp_for_each_object(patch, obj)
567 klp_for_each_func(obj, func)
568 func->transition = true;
569}
570
571/*
572 * This function can be called in the middle of an existing transition to
573 * reverse the direction of the target patch state. This can be done to
574 * effectively cancel an existing enable or disable operation if there are any
575 * tasks which are stuck in the initial patch state.
576 */
577void klp_reverse_transition(void)
578{
579 unsigned int cpu;
580 struct task_struct *g, *task;
581
582 pr_debug("'%s': reversing transition from %s\n",
583 klp_transition_patch->mod->name,
584 klp_target_state == KLP_PATCHED ? "patching to unpatching" :
585 "unpatching to patching");
586
587 klp_transition_patch->enabled = !klp_transition_patch->enabled;
588
589 klp_target_state = !klp_target_state;
590
591 /*
592 * Clear all TIF_PATCH_PENDING flags to prevent races caused by
593 * klp_update_patch_state() running in parallel with
594 * klp_start_transition().
595 */
596 read_lock(&tasklist_lock);
597 for_each_process_thread(g, task)
598 clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
599 read_unlock(&tasklist_lock);
600
601 for_each_possible_cpu(cpu)
602 clear_tsk_thread_flag(idle_task(cpu), TIF_PATCH_PENDING);
603
604 /* Let any remaining calls to klp_update_patch_state() complete */
605 klp_synchronize_transition();
606
607 klp_start_transition();
608}
609
610/* Called from copy_process() during fork */
611void klp_copy_process(struct task_struct *child)
612{
613
614 /*
615 * The parent process may have gone through a KLP transition since
616 * the thread flag was copied in setup_thread_stack earlier. Bring
617 * the task flag up to date with the parent here.
618 *
619 * The operation is serialized against all klp_*_transition()
620 * operations by the tasklist_lock. The only exception is
621 * klp_update_patch_state(current), but we cannot race with
622 * that because we are current.
623 */
624 if (test_tsk_thread_flag(current, TIF_PATCH_PENDING))
625 set_tsk_thread_flag(child, TIF_PATCH_PENDING);
626 else
627 clear_tsk_thread_flag(child, TIF_PATCH_PENDING);
628
629 child->patch_state = current->patch_state;
630}
631
632/*
633 * Drop TIF_PATCH_PENDING of all tasks on admin's request. This forces an
634 * existing transition to finish.
635 *
636 * NOTE: klp_update_patch_state(task) requires the task to be inactive or
637 * 'current'. This is not the case here and the consistency model could be
638 * broken. Administrator, who is the only one to execute the
639 * klp_force_transitions(), has to be aware of this.
640 */
641void klp_force_transition(void)
642{
643 struct klp_patch *patch;
644 struct task_struct *g, *task;
645 unsigned int cpu;
646
647 pr_warn("forcing remaining tasks to the patched state\n");
648
649 read_lock(&tasklist_lock);
650 for_each_process_thread(g, task)
651 klp_update_patch_state(task);
652 read_unlock(&tasklist_lock);
653
654 for_each_possible_cpu(cpu)
655 klp_update_patch_state(idle_task(cpu));
656
657 /* Set forced flag for patches being removed. */
658 if (klp_target_state == KLP_UNPATCHED)
659 klp_transition_patch->forced = true;
660 else if (klp_transition_patch->replace) {
661 klp_for_each_patch(patch) {
662 if (patch != klp_transition_patch)
663 patch->forced = true;
664 }
665 }
666}
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 <jpoimboe@redhat.com>
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_UNDEFINED;
27
28static unsigned int klp_signals_cnt;
29
30/*
31 * When a livepatch is in progress, enable klp stack checking in
32 * cond_resched(). This helps CPU-bound kthreads get patched.
33 */
34#if defined(CONFIG_PREEMPT_DYNAMIC) && defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
35
36#define klp_cond_resched_enable() sched_dynamic_klp_enable()
37#define klp_cond_resched_disable() sched_dynamic_klp_disable()
38
39#else /* !CONFIG_PREEMPT_DYNAMIC || !CONFIG_HAVE_PREEMPT_DYNAMIC_CALL */
40
41DEFINE_STATIC_KEY_FALSE(klp_sched_try_switch_key);
42EXPORT_SYMBOL(klp_sched_try_switch_key);
43
44#define klp_cond_resched_enable() static_branch_enable(&klp_sched_try_switch_key)
45#define klp_cond_resched_disable() static_branch_disable(&klp_sched_try_switch_key)
46
47#endif /* CONFIG_PREEMPT_DYNAMIC && CONFIG_HAVE_PREEMPT_DYNAMIC_CALL */
48
49/*
50 * This work can be performed periodically to finish patching or unpatching any
51 * "straggler" tasks which failed to transition in the first attempt.
52 */
53static void klp_transition_work_fn(struct work_struct *work)
54{
55 mutex_lock(&klp_mutex);
56
57 if (klp_transition_patch)
58 klp_try_complete_transition();
59
60 mutex_unlock(&klp_mutex);
61}
62static DECLARE_DELAYED_WORK(klp_transition_work, klp_transition_work_fn);
63
64/*
65 * This function is just a stub to implement a hard force
66 * of synchronize_rcu(). This requires synchronizing
67 * tasks even in userspace and idle.
68 */
69static void klp_sync(struct work_struct *work)
70{
71}
72
73/*
74 * We allow to patch also functions where RCU is not watching,
75 * e.g. before user_exit(). We can not rely on the RCU infrastructure
76 * to do the synchronization. Instead hard force the sched synchronization.
77 *
78 * This approach allows to use RCU functions for manipulating func_stack
79 * safely.
80 */
81static void klp_synchronize_transition(void)
82{
83 schedule_on_each_cpu(klp_sync);
84}
85
86/*
87 * The transition to the target patch state is complete. Clean up the data
88 * structures.
89 */
90static void klp_complete_transition(void)
91{
92 struct klp_object *obj;
93 struct klp_func *func;
94 struct task_struct *g, *task;
95 unsigned int cpu;
96
97 pr_debug("'%s': completing %s transition\n",
98 klp_transition_patch->mod->name,
99 klp_target_state == KLP_PATCHED ? "patching" : "unpatching");
100
101 if (klp_transition_patch->replace && klp_target_state == KLP_PATCHED) {
102 klp_unpatch_replaced_patches(klp_transition_patch);
103 klp_discard_nops(klp_transition_patch);
104 }
105
106 if (klp_target_state == KLP_UNPATCHED) {
107 /*
108 * All tasks have transitioned to KLP_UNPATCHED so we can now
109 * remove the new functions from the func_stack.
110 */
111 klp_unpatch_objects(klp_transition_patch);
112
113 /*
114 * Make sure klp_ftrace_handler() can no longer see functions
115 * from this patch on the ops->func_stack. Otherwise, after
116 * func->transition gets cleared, the handler may choose a
117 * removed function.
118 */
119 klp_synchronize_transition();
120 }
121
122 klp_for_each_object(klp_transition_patch, obj)
123 klp_for_each_func(obj, func)
124 func->transition = false;
125
126 /* Prevent klp_ftrace_handler() from seeing KLP_UNDEFINED state */
127 if (klp_target_state == KLP_PATCHED)
128 klp_synchronize_transition();
129
130 read_lock(&tasklist_lock);
131 for_each_process_thread(g, task) {
132 WARN_ON_ONCE(test_tsk_thread_flag(task, TIF_PATCH_PENDING));
133 task->patch_state = KLP_UNDEFINED;
134 }
135 read_unlock(&tasklist_lock);
136
137 for_each_possible_cpu(cpu) {
138 task = idle_task(cpu);
139 WARN_ON_ONCE(test_tsk_thread_flag(task, TIF_PATCH_PENDING));
140 task->patch_state = KLP_UNDEFINED;
141 }
142
143 klp_for_each_object(klp_transition_patch, obj) {
144 if (!klp_is_object_loaded(obj))
145 continue;
146 if (klp_target_state == KLP_PATCHED)
147 klp_post_patch_callback(obj);
148 else if (klp_target_state == KLP_UNPATCHED)
149 klp_post_unpatch_callback(obj);
150 }
151
152 pr_notice("'%s': %s complete\n", klp_transition_patch->mod->name,
153 klp_target_state == KLP_PATCHED ? "patching" : "unpatching");
154
155 klp_target_state = KLP_UNDEFINED;
156 klp_transition_patch = NULL;
157}
158
159/*
160 * This is called in the error path, to cancel a transition before it has
161 * started, i.e. klp_init_transition() has been called but
162 * klp_start_transition() hasn't. If the transition *has* been started,
163 * klp_reverse_transition() should be used instead.
164 */
165void klp_cancel_transition(void)
166{
167 if (WARN_ON_ONCE(klp_target_state != KLP_PATCHED))
168 return;
169
170 pr_debug("'%s': canceling patching transition, going to unpatch\n",
171 klp_transition_patch->mod->name);
172
173 klp_target_state = KLP_UNPATCHED;
174 klp_complete_transition();
175}
176
177/*
178 * Switch the patched state of the task to the set of functions in the target
179 * patch state.
180 *
181 * NOTE: If task is not 'current', the caller must ensure the task is inactive.
182 * Otherwise klp_ftrace_handler() might read the wrong 'patch_state' value.
183 */
184void klp_update_patch_state(struct task_struct *task)
185{
186 /*
187 * A variant of synchronize_rcu() is used to allow patching functions
188 * where RCU is not watching, see klp_synchronize_transition().
189 */
190 preempt_disable_notrace();
191
192 /*
193 * This test_and_clear_tsk_thread_flag() call also serves as a read
194 * barrier (smp_rmb) for two cases:
195 *
196 * 1) Enforce the order of the TIF_PATCH_PENDING read and the
197 * klp_target_state read. The corresponding write barriers are in
198 * klp_init_transition() and klp_reverse_transition().
199 *
200 * 2) Enforce the order of the TIF_PATCH_PENDING read and a future read
201 * of func->transition, if klp_ftrace_handler() is called later on
202 * the same CPU. See __klp_disable_patch().
203 */
204 if (test_and_clear_tsk_thread_flag(task, TIF_PATCH_PENDING))
205 task->patch_state = READ_ONCE(klp_target_state);
206
207 preempt_enable_notrace();
208}
209
210/*
211 * Determine whether the given stack trace includes any references to a
212 * to-be-patched or to-be-unpatched function.
213 */
214static int klp_check_stack_func(struct klp_func *func, unsigned long *entries,
215 unsigned int nr_entries)
216{
217 unsigned long func_addr, func_size, address;
218 struct klp_ops *ops;
219 int i;
220
221 if (klp_target_state == KLP_UNPATCHED) {
222 /*
223 * Check for the to-be-unpatched function
224 * (the func itself).
225 */
226 func_addr = (unsigned long)func->new_func;
227 func_size = func->new_size;
228 } else {
229 /*
230 * Check for the to-be-patched function
231 * (the previous func).
232 */
233 ops = klp_find_ops(func->old_func);
234
235 if (list_is_singular(&ops->func_stack)) {
236 /* original function */
237 func_addr = (unsigned long)func->old_func;
238 func_size = func->old_size;
239 } else {
240 /* previously patched function */
241 struct klp_func *prev;
242
243 prev = list_next_entry(func, stack_node);
244 func_addr = (unsigned long)prev->new_func;
245 func_size = prev->new_size;
246 }
247 }
248
249 for (i = 0; i < nr_entries; i++) {
250 address = entries[i];
251
252 if (address >= func_addr && address < func_addr + func_size)
253 return -EAGAIN;
254 }
255
256 return 0;
257}
258
259/*
260 * Determine whether it's safe to transition the task to the target patch state
261 * by looking for any to-be-patched or to-be-unpatched functions on its stack.
262 */
263static int klp_check_stack(struct task_struct *task, const char **oldname)
264{
265 unsigned long *entries = this_cpu_ptr(klp_stack_entries);
266 struct klp_object *obj;
267 struct klp_func *func;
268 int ret, nr_entries;
269
270 /* Protect 'klp_stack_entries' */
271 lockdep_assert_preemption_disabled();
272
273 ret = stack_trace_save_tsk_reliable(task, entries, MAX_STACK_ENTRIES);
274 if (ret < 0)
275 return -EINVAL;
276 nr_entries = ret;
277
278 klp_for_each_object(klp_transition_patch, obj) {
279 if (!obj->patched)
280 continue;
281 klp_for_each_func(obj, func) {
282 ret = klp_check_stack_func(func, entries, nr_entries);
283 if (ret) {
284 *oldname = func->old_name;
285 return -EADDRINUSE;
286 }
287 }
288 }
289
290 return 0;
291}
292
293static int klp_check_and_switch_task(struct task_struct *task, void *arg)
294{
295 int ret;
296
297 if (task_curr(task) && task != current)
298 return -EBUSY;
299
300 ret = klp_check_stack(task, arg);
301 if (ret)
302 return ret;
303
304 clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
305 task->patch_state = klp_target_state;
306 return 0;
307}
308
309/*
310 * Try to safely switch a task to the target patch state. If it's currently
311 * running, or it's sleeping on a to-be-patched or to-be-unpatched function, or
312 * if the stack is unreliable, return false.
313 */
314static bool klp_try_switch_task(struct task_struct *task)
315{
316 const char *old_name;
317 int ret;
318
319 /* check if this task has already switched over */
320 if (task->patch_state == klp_target_state)
321 return true;
322
323 /*
324 * For arches which don't have reliable stack traces, we have to rely
325 * on other methods (e.g., switching tasks at kernel exit).
326 */
327 if (!klp_have_reliable_stack())
328 return false;
329
330 /*
331 * Now try to check the stack for any to-be-patched or to-be-unpatched
332 * functions. If all goes well, switch the task to the target patch
333 * state.
334 */
335 if (task == current)
336 ret = klp_check_and_switch_task(current, &old_name);
337 else
338 ret = task_call_func(task, klp_check_and_switch_task, &old_name);
339
340 switch (ret) {
341 case 0: /* success */
342 break;
343
344 case -EBUSY: /* klp_check_and_switch_task() */
345 pr_debug("%s: %s:%d is running\n",
346 __func__, task->comm, task->pid);
347 break;
348 case -EINVAL: /* klp_check_and_switch_task() */
349 pr_debug("%s: %s:%d has an unreliable stack\n",
350 __func__, task->comm, task->pid);
351 break;
352 case -EADDRINUSE: /* klp_check_and_switch_task() */
353 pr_debug("%s: %s:%d is sleeping on function %s\n",
354 __func__, task->comm, task->pid, old_name);
355 break;
356
357 default:
358 pr_debug("%s: Unknown error code (%d) when trying to switch %s:%d\n",
359 __func__, ret, task->comm, task->pid);
360 break;
361 }
362
363 return !ret;
364}
365
366void __klp_sched_try_switch(void)
367{
368 if (likely(!klp_patch_pending(current)))
369 return;
370
371 /*
372 * This function is called from cond_resched() which is called in many
373 * places throughout the kernel. Using the klp_mutex here might
374 * deadlock.
375 *
376 * Instead, disable preemption to prevent racing with other callers of
377 * klp_try_switch_task(). Thanks to task_call_func() they won't be
378 * able to switch this task while it's running.
379 */
380 preempt_disable();
381
382 /*
383 * Make sure current didn't get patched between the above check and
384 * preempt_disable().
385 */
386 if (unlikely(!klp_patch_pending(current)))
387 goto out;
388
389 /*
390 * Enforce the order of the TIF_PATCH_PENDING read above and the
391 * klp_target_state read in klp_try_switch_task(). The corresponding
392 * write barriers are in klp_init_transition() and
393 * klp_reverse_transition().
394 */
395 smp_rmb();
396
397 klp_try_switch_task(current);
398
399out:
400 preempt_enable();
401}
402EXPORT_SYMBOL(__klp_sched_try_switch);
403
404/*
405 * Sends a fake signal to all non-kthread tasks with TIF_PATCH_PENDING set.
406 * Kthreads with TIF_PATCH_PENDING set are woken up.
407 */
408static void klp_send_signals(void)
409{
410 struct task_struct *g, *task;
411
412 if (klp_signals_cnt == SIGNALS_TIMEOUT)
413 pr_notice("signaling remaining tasks\n");
414
415 read_lock(&tasklist_lock);
416 for_each_process_thread(g, task) {
417 if (!klp_patch_pending(task))
418 continue;
419
420 /*
421 * There is a small race here. We could see TIF_PATCH_PENDING
422 * set and decide to wake up a kthread or send a fake signal.
423 * Meanwhile the task could migrate itself and the action
424 * would be meaningless. It is not serious though.
425 */
426 if (task->flags & PF_KTHREAD) {
427 /*
428 * Wake up a kthread which sleeps interruptedly and
429 * still has not been migrated.
430 */
431 wake_up_state(task, TASK_INTERRUPTIBLE);
432 } else {
433 /*
434 * Send fake signal to all non-kthread tasks which are
435 * still not migrated.
436 */
437 set_notify_signal(task);
438 }
439 }
440 read_unlock(&tasklist_lock);
441}
442
443/*
444 * Try to switch all remaining tasks to the target patch state by walking the
445 * stacks of sleeping tasks and looking for any to-be-patched or
446 * to-be-unpatched functions. If such functions are found, the task can't be
447 * switched yet.
448 *
449 * If any tasks are still stuck in the initial patch state, schedule a retry.
450 */
451void klp_try_complete_transition(void)
452{
453 unsigned int cpu;
454 struct task_struct *g, *task;
455 struct klp_patch *patch;
456 bool complete = true;
457
458 WARN_ON_ONCE(klp_target_state == KLP_UNDEFINED);
459
460 /*
461 * Try to switch the tasks to the target patch state by walking their
462 * stacks and looking for any to-be-patched or to-be-unpatched
463 * functions. If such functions are found on a stack, or if the stack
464 * is deemed unreliable, the task can't be switched yet.
465 *
466 * Usually this will transition most (or all) of the tasks on a system
467 * unless the patch includes changes to a very common function.
468 */
469 read_lock(&tasklist_lock);
470 for_each_process_thread(g, task)
471 if (!klp_try_switch_task(task))
472 complete = false;
473 read_unlock(&tasklist_lock);
474
475 /*
476 * Ditto for the idle "swapper" tasks.
477 */
478 cpus_read_lock();
479 for_each_possible_cpu(cpu) {
480 task = idle_task(cpu);
481 if (cpu_online(cpu)) {
482 if (!klp_try_switch_task(task)) {
483 complete = false;
484 /* Make idle task go through the main loop. */
485 wake_up_if_idle(cpu);
486 }
487 } else if (task->patch_state != klp_target_state) {
488 /* offline idle tasks can be switched immediately */
489 clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
490 task->patch_state = klp_target_state;
491 }
492 }
493 cpus_read_unlock();
494
495 if (!complete) {
496 if (klp_signals_cnt && !(klp_signals_cnt % SIGNALS_TIMEOUT))
497 klp_send_signals();
498 klp_signals_cnt++;
499
500 /*
501 * Some tasks weren't able to be switched over. Try again
502 * later and/or wait for other methods like kernel exit
503 * switching.
504 */
505 schedule_delayed_work(&klp_transition_work,
506 round_jiffies_relative(HZ));
507 return;
508 }
509
510 /* Done! Now cleanup the data structures. */
511 klp_cond_resched_disable();
512 patch = klp_transition_patch;
513 klp_complete_transition();
514
515 /*
516 * It would make more sense to free the unused patches in
517 * klp_complete_transition() but it is called also
518 * from klp_cancel_transition().
519 */
520 if (!patch->enabled)
521 klp_free_patch_async(patch);
522 else if (patch->replace)
523 klp_free_replaced_patches_async(patch);
524}
525
526/*
527 * Start the transition to the specified target patch state so tasks can begin
528 * switching to it.
529 */
530void klp_start_transition(void)
531{
532 struct task_struct *g, *task;
533 unsigned int cpu;
534
535 WARN_ON_ONCE(klp_target_state == KLP_UNDEFINED);
536
537 pr_notice("'%s': starting %s transition\n",
538 klp_transition_patch->mod->name,
539 klp_target_state == KLP_PATCHED ? "patching" : "unpatching");
540
541 /*
542 * Mark all normal tasks as needing a patch state update. They'll
543 * switch either in klp_try_complete_transition() or as they exit the
544 * kernel.
545 */
546 read_lock(&tasklist_lock);
547 for_each_process_thread(g, task)
548 if (task->patch_state != klp_target_state)
549 set_tsk_thread_flag(task, TIF_PATCH_PENDING);
550 read_unlock(&tasklist_lock);
551
552 /*
553 * Mark all idle tasks as needing a patch state update. They'll switch
554 * either in klp_try_complete_transition() or at the idle loop switch
555 * point.
556 */
557 for_each_possible_cpu(cpu) {
558 task = idle_task(cpu);
559 if (task->patch_state != klp_target_state)
560 set_tsk_thread_flag(task, TIF_PATCH_PENDING);
561 }
562
563 klp_cond_resched_enable();
564
565 klp_signals_cnt = 0;
566}
567
568/*
569 * Initialize the global target patch state and all tasks to the initial patch
570 * state, and initialize all function transition states to true in preparation
571 * for patching or unpatching.
572 */
573void klp_init_transition(struct klp_patch *patch, int state)
574{
575 struct task_struct *g, *task;
576 unsigned int cpu;
577 struct klp_object *obj;
578 struct klp_func *func;
579 int initial_state = !state;
580
581 WARN_ON_ONCE(klp_target_state != KLP_UNDEFINED);
582
583 klp_transition_patch = patch;
584
585 /*
586 * Set the global target patch state which tasks will switch to. This
587 * has no effect until the TIF_PATCH_PENDING flags get set later.
588 */
589 klp_target_state = state;
590
591 pr_debug("'%s': initializing %s transition\n", patch->mod->name,
592 klp_target_state == KLP_PATCHED ? "patching" : "unpatching");
593
594 /*
595 * Initialize all tasks to the initial patch state to prepare them for
596 * switching to the target state.
597 */
598 read_lock(&tasklist_lock);
599 for_each_process_thread(g, task) {
600 WARN_ON_ONCE(task->patch_state != KLP_UNDEFINED);
601 task->patch_state = initial_state;
602 }
603 read_unlock(&tasklist_lock);
604
605 /*
606 * Ditto for the idle "swapper" tasks.
607 */
608 for_each_possible_cpu(cpu) {
609 task = idle_task(cpu);
610 WARN_ON_ONCE(task->patch_state != KLP_UNDEFINED);
611 task->patch_state = initial_state;
612 }
613
614 /*
615 * Enforce the order of the task->patch_state initializations and the
616 * func->transition updates to ensure that klp_ftrace_handler() doesn't
617 * see a func in transition with a task->patch_state of KLP_UNDEFINED.
618 *
619 * Also enforce the order of the klp_target_state write and future
620 * TIF_PATCH_PENDING writes to ensure klp_update_patch_state() and
621 * __klp_sched_try_switch() don't set a task->patch_state to
622 * KLP_UNDEFINED.
623 */
624 smp_wmb();
625
626 /*
627 * Set the func transition states so klp_ftrace_handler() will know to
628 * switch to the transition logic.
629 *
630 * When patching, the funcs aren't yet in the func_stack and will be
631 * made visible to the ftrace handler shortly by the calls to
632 * klp_patch_object().
633 *
634 * When unpatching, the funcs are already in the func_stack and so are
635 * already visible to the ftrace handler.
636 */
637 klp_for_each_object(patch, obj)
638 klp_for_each_func(obj, func)
639 func->transition = true;
640}
641
642/*
643 * This function can be called in the middle of an existing transition to
644 * reverse the direction of the target patch state. This can be done to
645 * effectively cancel an existing enable or disable operation if there are any
646 * tasks which are stuck in the initial patch state.
647 */
648void klp_reverse_transition(void)
649{
650 unsigned int cpu;
651 struct task_struct *g, *task;
652
653 pr_debug("'%s': reversing transition from %s\n",
654 klp_transition_patch->mod->name,
655 klp_target_state == KLP_PATCHED ? "patching to unpatching" :
656 "unpatching to patching");
657
658 /*
659 * Clear all TIF_PATCH_PENDING flags to prevent races caused by
660 * klp_update_patch_state() or __klp_sched_try_switch() running in
661 * parallel with the reverse transition.
662 */
663 read_lock(&tasklist_lock);
664 for_each_process_thread(g, task)
665 clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
666 read_unlock(&tasklist_lock);
667
668 for_each_possible_cpu(cpu)
669 clear_tsk_thread_flag(idle_task(cpu), TIF_PATCH_PENDING);
670
671 /*
672 * Make sure all existing invocations of klp_update_patch_state() and
673 * __klp_sched_try_switch() see the cleared TIF_PATCH_PENDING before
674 * starting the reverse transition.
675 */
676 klp_synchronize_transition();
677
678 /*
679 * All patching has stopped, now re-initialize the global variables to
680 * prepare for the reverse transition.
681 */
682 klp_transition_patch->enabled = !klp_transition_patch->enabled;
683 klp_target_state = !klp_target_state;
684
685 /*
686 * Enforce the order of the klp_target_state write and the
687 * TIF_PATCH_PENDING writes in klp_start_transition() to ensure
688 * klp_update_patch_state() and __klp_sched_try_switch() don't set
689 * task->patch_state to the wrong value.
690 */
691 smp_wmb();
692
693 klp_start_transition();
694}
695
696/* Called from copy_process() during fork */
697void klp_copy_process(struct task_struct *child)
698{
699
700 /*
701 * The parent process may have gone through a KLP transition since
702 * the thread flag was copied in setup_thread_stack earlier. Bring
703 * the task flag up to date with the parent here.
704 *
705 * The operation is serialized against all klp_*_transition()
706 * operations by the tasklist_lock. The only exceptions are
707 * klp_update_patch_state(current) and __klp_sched_try_switch(), but we
708 * cannot race with them because we are current.
709 */
710 if (test_tsk_thread_flag(current, TIF_PATCH_PENDING))
711 set_tsk_thread_flag(child, TIF_PATCH_PENDING);
712 else
713 clear_tsk_thread_flag(child, TIF_PATCH_PENDING);
714
715 child->patch_state = current->patch_state;
716}
717
718/*
719 * Drop TIF_PATCH_PENDING of all tasks on admin's request. This forces an
720 * existing transition to finish.
721 *
722 * NOTE: klp_update_patch_state(task) requires the task to be inactive or
723 * 'current'. This is not the case here and the consistency model could be
724 * broken. Administrator, who is the only one to execute the
725 * klp_force_transitions(), has to be aware of this.
726 */
727void klp_force_transition(void)
728{
729 struct klp_patch *patch;
730 struct task_struct *g, *task;
731 unsigned int cpu;
732
733 pr_warn("forcing remaining tasks to the patched state\n");
734
735 read_lock(&tasklist_lock);
736 for_each_process_thread(g, task)
737 klp_update_patch_state(task);
738 read_unlock(&tasklist_lock);
739
740 for_each_possible_cpu(cpu)
741 klp_update_patch_state(idle_task(cpu));
742
743 /* Set forced flag for patches being removed. */
744 if (klp_target_state == KLP_UNPATCHED)
745 klp_transition_patch->forced = true;
746 else if (klp_transition_patch->replace) {
747 klp_for_each_patch(patch) {
748 if (patch != klp_transition_patch)
749 patch->forced = true;
750 }
751 }
752}