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