b004a1fb603236f31cd5780962c098bfcf8fe743
[muen/linux.git] / kernel / livepatch / transition.c
1 /*
2  * transition.c - Kernel Live Patching transition functions
3  *
4  * Copyright (C) 2015-2016 Josh Poimboeuf <jpoimboe@redhat.com>
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License
8  * as published by the Free Software Foundation; either version 2
9  * of the License, or (at your option) any later version.
10  *
11  * This program is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
14  * GNU General Public License for more details.
15  *
16  * You should have received a copy of the GNU General Public License
17  * along with this program; if not, see <http://www.gnu.org/licenses/>.
18  */
19
20 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
21
22 #include <linux/cpu.h>
23 #include <linux/stacktrace.h>
24 #include "core.h"
25 #include "patch.h"
26 #include "transition.h"
27 #include "../sched/sched.h"
28
29 #define MAX_STACK_ENTRIES  100
30 #define STACK_ERR_BUF_SIZE 128
31
32 struct klp_patch *klp_transition_patch;
33
34 static int klp_target_state = KLP_UNDEFINED;
35
36 /*
37  * This work can be performed periodically to finish patching or unpatching any
38  * "straggler" tasks which failed to transition in the first attempt.
39  */
40 static void klp_transition_work_fn(struct work_struct *work)
41 {
42         mutex_lock(&klp_mutex);
43
44         if (klp_transition_patch)
45                 klp_try_complete_transition();
46
47         mutex_unlock(&klp_mutex);
48 }
49 static DECLARE_DELAYED_WORK(klp_transition_work, klp_transition_work_fn);
50
51 /*
52  * This function is just a stub to implement a hard force
53  * of synchronize_sched(). This requires synchronizing
54  * tasks even in userspace and idle.
55  */
56 static void klp_sync(struct work_struct *work)
57 {
58 }
59
60 /*
61  * We allow to patch also functions where RCU is not watching,
62  * e.g. before user_exit(). We can not rely on the RCU infrastructure
63  * to do the synchronization. Instead hard force the sched synchronization.
64  *
65  * This approach allows to use RCU functions for manipulating func_stack
66  * safely.
67  */
68 static void klp_synchronize_transition(void)
69 {
70         schedule_on_each_cpu(klp_sync);
71 }
72
73 /*
74  * The transition to the target patch state is complete.  Clean up the data
75  * structures.
76  */
77 static void klp_complete_transition(void)
78 {
79         struct klp_object *obj;
80         struct klp_func *func;
81         struct task_struct *g, *task;
82         unsigned int cpu;
83         bool immediate_func = false;
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         if (klp_transition_patch->immediate)
102                 goto done;
103
104         klp_for_each_object(klp_transition_patch, obj) {
105                 klp_for_each_func(obj, func) {
106                         func->transition = false;
107                         if (func->immediate)
108                                 immediate_func = true;
109                 }
110         }
111
112         if (klp_target_state == KLP_UNPATCHED && !immediate_func)
113                 module_put(klp_transition_patch->mod);
114
115         /* Prevent klp_ftrace_handler() from seeing KLP_UNDEFINED state */
116         if (klp_target_state == KLP_PATCHED)
117                 klp_synchronize_transition();
118
119         read_lock(&tasklist_lock);
120         for_each_process_thread(g, task) {
121                 WARN_ON_ONCE(test_tsk_thread_flag(task, TIF_PATCH_PENDING));
122                 task->patch_state = KLP_UNDEFINED;
123         }
124         read_unlock(&tasklist_lock);
125
126         for_each_possible_cpu(cpu) {
127                 task = idle_task(cpu);
128                 WARN_ON_ONCE(test_tsk_thread_flag(task, TIF_PATCH_PENDING));
129                 task->patch_state = KLP_UNDEFINED;
130         }
131
132 done:
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  */
143 void klp_cancel_transition(void)
144 {
145         if (WARN_ON_ONCE(klp_target_state != KLP_PATCHED))
146                 return;
147
148         klp_target_state = KLP_UNPATCHED;
149         klp_complete_transition();
150 }
151
152 /*
153  * Switch the patched state of the task to the set of functions in the target
154  * patch state.
155  *
156  * NOTE: If task is not 'current', the caller must ensure the task is inactive.
157  * Otherwise klp_ftrace_handler() might read the wrong 'patch_state' value.
158  */
159 void klp_update_patch_state(struct task_struct *task)
160 {
161         /*
162          * A variant of synchronize_sched() is used to allow patching functions
163          * where RCU is not watching, see klp_synchronize_transition().
164          */
165         preempt_disable_notrace();
166
167         /*
168          * This test_and_clear_tsk_thread_flag() call also serves as a read
169          * barrier (smp_rmb) for two cases:
170          *
171          * 1) Enforce the order of the TIF_PATCH_PENDING read and the
172          *    klp_target_state read.  The corresponding write barrier is in
173          *    klp_init_transition().
174          *
175          * 2) Enforce the order of the TIF_PATCH_PENDING read and a future read
176          *    of func->transition, if klp_ftrace_handler() is called later on
177          *    the same CPU.  See __klp_disable_patch().
178          */
179         if (test_and_clear_tsk_thread_flag(task, TIF_PATCH_PENDING))
180                 task->patch_state = READ_ONCE(klp_target_state);
181
182         preempt_enable_notrace();
183 }
184
185 /*
186  * Determine whether the given stack trace includes any references to a
187  * to-be-patched or to-be-unpatched function.
188  */
189 static int klp_check_stack_func(struct klp_func *func,
190                                 struct stack_trace *trace)
191 {
192         unsigned long func_addr, func_size, address;
193         struct klp_ops *ops;
194         int i;
195
196         if (func->immediate)
197                 return 0;
198
199         for (i = 0; i < trace->nr_entries; i++) {
200                 address = trace->entries[i];
201
202                 if (klp_target_state == KLP_UNPATCHED) {
203                          /*
204                           * Check for the to-be-unpatched function
205                           * (the func itself).
206                           */
207                         func_addr = (unsigned long)func->new_func;
208                         func_size = func->new_size;
209                 } else {
210                         /*
211                          * Check for the to-be-patched function
212                          * (the previous func).
213                          */
214                         ops = klp_find_ops(func->old_addr);
215
216                         if (list_is_singular(&ops->func_stack)) {
217                                 /* original function */
218                                 func_addr = func->old_addr;
219                                 func_size = func->old_size;
220                         } else {
221                                 /* previously patched function */
222                                 struct klp_func *prev;
223
224                                 prev = list_next_entry(func, stack_node);
225                                 func_addr = (unsigned long)prev->new_func;
226                                 func_size = prev->new_size;
227                         }
228                 }
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  */
241 static int klp_check_stack(struct task_struct *task, char *err_buf)
242 {
243         static unsigned long entries[MAX_STACK_ENTRIES];
244         struct stack_trace trace;
245         struct klp_object *obj;
246         struct klp_func *func;
247         int ret;
248
249         trace.skip = 0;
250         trace.nr_entries = 0;
251         trace.max_entries = MAX_STACK_ENTRIES;
252         trace.entries = entries;
253         ret = save_stack_trace_tsk_reliable(task, &trace);
254         WARN_ON_ONCE(ret == -ENOSYS);
255         if (ret) {
256                 snprintf(err_buf, STACK_ERR_BUF_SIZE,
257                          "%s: %s:%d has an unreliable stack\n",
258                          __func__, task->comm, task->pid);
259                 return ret;
260         }
261
262         klp_for_each_object(klp_transition_patch, obj) {
263                 if (!obj->patched)
264                         continue;
265                 klp_for_each_func(obj, func) {
266                         ret = klp_check_stack_func(func, &trace);
267                         if (ret) {
268                                 snprintf(err_buf, STACK_ERR_BUF_SIZE,
269                                          "%s: %s:%d is sleeping on function %s\n",
270                                          __func__, task->comm, task->pid,
271                                          func->old_name);
272                                 return ret;
273                         }
274                 }
275         }
276
277         return 0;
278 }
279
280 /*
281  * Try to safely switch a task to the target patch state.  If it's currently
282  * running, or it's sleeping on a to-be-patched or to-be-unpatched function, or
283  * if the stack is unreliable, return false.
284  */
285 static bool klp_try_switch_task(struct task_struct *task)
286 {
287         struct rq *rq;
288         struct rq_flags flags;
289         int ret;
290         bool success = false;
291         char err_buf[STACK_ERR_BUF_SIZE];
292
293         err_buf[0] = '\0';
294
295         /* check if this task has already switched over */
296         if (task->patch_state == klp_target_state)
297                 return true;
298
299         /*
300          * For arches which don't have reliable stack traces, we have to rely
301          * on other methods (e.g., switching tasks at kernel exit).
302          */
303         if (!klp_have_reliable_stack())
304                 return false;
305
306         /*
307          * Now try to check the stack for any to-be-patched or to-be-unpatched
308          * functions.  If all goes well, switch the task to the target patch
309          * state.
310          */
311         rq = task_rq_lock(task, &flags);
312
313         if (task_running(rq, task) && task != current) {
314                 snprintf(err_buf, STACK_ERR_BUF_SIZE,
315                          "%s: %s:%d is running\n", __func__, task->comm,
316                          task->pid);
317                 goto done;
318         }
319
320         ret = klp_check_stack(task, err_buf);
321         if (ret)
322                 goto done;
323
324         success = true;
325
326         clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
327         task->patch_state = klp_target_state;
328
329 done:
330         task_rq_unlock(rq, task, &flags);
331
332         /*
333          * Due to console deadlock issues, pr_debug() can't be used while
334          * holding the task rq lock.  Instead we have to use a temporary buffer
335          * and print the debug message after releasing the lock.
336          */
337         if (err_buf[0] != '\0')
338                 pr_debug("%s", err_buf);
339
340         return success;
341
342 }
343
344 /*
345  * Try to switch all remaining tasks to the target patch state by walking the
346  * stacks of sleeping tasks and looking for any to-be-patched or
347  * to-be-unpatched functions.  If such functions are found, the task can't be
348  * switched yet.
349  *
350  * If any tasks are still stuck in the initial patch state, schedule a retry.
351  */
352 void klp_try_complete_transition(void)
353 {
354         unsigned int cpu;
355         struct task_struct *g, *task;
356         bool complete = true;
357
358         WARN_ON_ONCE(klp_target_state == KLP_UNDEFINED);
359
360         /*
361          * If the patch can be applied or reverted immediately, skip the
362          * per-task transitions.
363          */
364         if (klp_transition_patch->immediate)
365                 goto success;
366
367         /*
368          * Try to switch the tasks to the target patch state by walking their
369          * stacks and looking for any to-be-patched or to-be-unpatched
370          * functions.  If such functions are found on a stack, or if the stack
371          * is deemed unreliable, the task can't be switched yet.
372          *
373          * Usually this will transition most (or all) of the tasks on a system
374          * unless the patch includes changes to a very common function.
375          */
376         read_lock(&tasklist_lock);
377         for_each_process_thread(g, task)
378                 if (!klp_try_switch_task(task))
379                         complete = false;
380         read_unlock(&tasklist_lock);
381
382         /*
383          * Ditto for the idle "swapper" tasks.
384          */
385         get_online_cpus();
386         for_each_possible_cpu(cpu) {
387                 task = idle_task(cpu);
388                 if (cpu_online(cpu)) {
389                         if (!klp_try_switch_task(task))
390                                 complete = false;
391                 } else if (task->patch_state != klp_target_state) {
392                         /* offline idle tasks can be switched immediately */
393                         clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
394                         task->patch_state = klp_target_state;
395                 }
396         }
397         put_online_cpus();
398
399         if (!complete) {
400                 /*
401                  * Some tasks weren't able to be switched over.  Try again
402                  * later and/or wait for other methods like kernel exit
403                  * switching.
404                  */
405                 schedule_delayed_work(&klp_transition_work,
406                                       round_jiffies_relative(HZ));
407                 return;
408         }
409
410 success:
411         pr_notice("'%s': %s complete\n", klp_transition_patch->mod->name,
412                   klp_target_state == KLP_PATCHED ? "patching" : "unpatching");
413
414         /* we're done, now cleanup the data structures */
415         klp_complete_transition();
416 }
417
418 /*
419  * Start the transition to the specified target patch state so tasks can begin
420  * switching to it.
421  */
422 void klp_start_transition(void)
423 {
424         struct task_struct *g, *task;
425         unsigned int cpu;
426
427         WARN_ON_ONCE(klp_target_state == KLP_UNDEFINED);
428
429         pr_notice("'%s': %s...\n", klp_transition_patch->mod->name,
430                   klp_target_state == KLP_PATCHED ? "patching" : "unpatching");
431
432         /*
433          * If the patch can be applied or reverted immediately, skip the
434          * per-task transitions.
435          */
436         if (klp_transition_patch->immediate)
437                 return;
438
439         /*
440          * Mark all normal tasks as needing a patch state update.  They'll
441          * switch either in klp_try_complete_transition() or as they exit the
442          * kernel.
443          */
444         read_lock(&tasklist_lock);
445         for_each_process_thread(g, task)
446                 if (task->patch_state != klp_target_state)
447                         set_tsk_thread_flag(task, TIF_PATCH_PENDING);
448         read_unlock(&tasklist_lock);
449
450         /*
451          * Mark all idle tasks as needing a patch state update.  They'll switch
452          * either in klp_try_complete_transition() or at the idle loop switch
453          * point.
454          */
455         for_each_possible_cpu(cpu) {
456                 task = idle_task(cpu);
457                 if (task->patch_state != klp_target_state)
458                         set_tsk_thread_flag(task, TIF_PATCH_PENDING);
459         }
460 }
461
462 /*
463  * Initialize the global target patch state and all tasks to the initial patch
464  * state, and initialize all function transition states to true in preparation
465  * for patching or unpatching.
466  */
467 void klp_init_transition(struct klp_patch *patch, int state)
468 {
469         struct task_struct *g, *task;
470         unsigned int cpu;
471         struct klp_object *obj;
472         struct klp_func *func;
473         int initial_state = !state;
474
475         WARN_ON_ONCE(klp_target_state != KLP_UNDEFINED);
476
477         klp_transition_patch = patch;
478
479         /*
480          * Set the global target patch state which tasks will switch to.  This
481          * has no effect until the TIF_PATCH_PENDING flags get set later.
482          */
483         klp_target_state = state;
484
485         /*
486          * If the patch can be applied or reverted immediately, skip the
487          * per-task transitions.
488          */
489         if (patch->immediate)
490                 return;
491
492         /*
493          * Initialize all tasks to the initial patch state to prepare them for
494          * switching to the target state.
495          */
496         read_lock(&tasklist_lock);
497         for_each_process_thread(g, task) {
498                 WARN_ON_ONCE(task->patch_state != KLP_UNDEFINED);
499                 task->patch_state = initial_state;
500         }
501         read_unlock(&tasklist_lock);
502
503         /*
504          * Ditto for the idle "swapper" tasks.
505          */
506         for_each_possible_cpu(cpu) {
507                 task = idle_task(cpu);
508                 WARN_ON_ONCE(task->patch_state != KLP_UNDEFINED);
509                 task->patch_state = initial_state;
510         }
511
512         /*
513          * Enforce the order of the task->patch_state initializations and the
514          * func->transition updates to ensure that klp_ftrace_handler() doesn't
515          * see a func in transition with a task->patch_state of KLP_UNDEFINED.
516          *
517          * Also enforce the order of the klp_target_state write and future
518          * TIF_PATCH_PENDING writes to ensure klp_update_patch_state() doesn't
519          * set a task->patch_state to KLP_UNDEFINED.
520          */
521         smp_wmb();
522
523         /*
524          * Set the func transition states so klp_ftrace_handler() will know to
525          * switch to the transition logic.
526          *
527          * When patching, the funcs aren't yet in the func_stack and will be
528          * made visible to the ftrace handler shortly by the calls to
529          * klp_patch_object().
530          *
531          * When unpatching, the funcs are already in the func_stack and so are
532          * already visible to the ftrace handler.
533          */
534         klp_for_each_object(patch, obj)
535                 klp_for_each_func(obj, func)
536                         func->transition = true;
537 }
538
539 /*
540  * This function can be called in the middle of an existing transition to
541  * reverse the direction of the target patch state.  This can be done to
542  * effectively cancel an existing enable or disable operation if there are any
543  * tasks which are stuck in the initial patch state.
544  */
545 void klp_reverse_transition(void)
546 {
547         unsigned int cpu;
548         struct task_struct *g, *task;
549
550         klp_transition_patch->enabled = !klp_transition_patch->enabled;
551
552         klp_target_state = !klp_target_state;
553
554         /*
555          * Clear all TIF_PATCH_PENDING flags to prevent races caused by
556          * klp_update_patch_state() running in parallel with
557          * klp_start_transition().
558          */
559         read_lock(&tasklist_lock);
560         for_each_process_thread(g, task)
561                 clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
562         read_unlock(&tasklist_lock);
563
564         for_each_possible_cpu(cpu)
565                 clear_tsk_thread_flag(idle_task(cpu), TIF_PATCH_PENDING);
566
567         /* Let any remaining calls to klp_update_patch_state() complete */
568         klp_synchronize_transition();
569
570         klp_start_transition();
571 }
572
573 /* Called from copy_process() during fork */
574 void klp_copy_process(struct task_struct *child)
575 {
576         child->patch_state = current->patch_state;
577
578         /* TIF_PATCH_PENDING gets copied in setup_thread_stack() */
579 }