kernel/livepatch/transition.c

   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         WARN_ON_ONCE(ret == -ENOSYS);
 251         if (ret < 0) {
 252                 snprintf(err_buf, STACK_ERR_BUF_SIZE,
 253                          "%s: %s:%d has an unreliable stack\n",
 254                          __func__, task->comm, task->pid);
 255                 return ret;
 256         }
 257         nr_entries = ret;
 258
 259         klp_for_each_object(klp_transition_patch, obj) {
 260                 if (!obj->patched)
 261                         continue;
 262                 klp_for_each_func(obj, func) {
 263                         ret = klp_check_stack_func(func, entries, nr_entries);
 264                         if (ret) {
 265                                 snprintf(err_buf, STACK_ERR_BUF_SIZE,
 266                                          "%s: %s:%d is sleeping on function %s\n",
 267                                          __func__, task->comm, task->pid,
 268                                          func->old_name);
 269                                 return ret;
 270                         }
 271                 }
 272         }
 273
 274         return 0;
 275 }
 276
 277 /*
 278  * Try to safely switch a task to the target patch state.  If it's currently
 279  * running, or it's sleeping on a to-be-patched or to-be-unpatched function, or
 280  * if the stack is unreliable, return false.
 281  */
 282 static bool klp_try_switch_task(struct task_struct *task)
 283 {
 284         struct rq *rq;
 285         struct rq_flags flags;
 286         int ret;
 287         bool success = false;
 288         char err_buf[STACK_ERR_BUF_SIZE];
 289
 290         err_buf[0] = '\0';
 291
 292         /* check if this task has already switched over */
 293         if (task->patch_state == klp_target_state)
 294                 return true;
 295
 296         /*
 297          * Now try to check the stack for any to-be-patched or to-be-unpatched
 298          * functions.  If all goes well, switch the task to the target patch
 299          * state.
 300          */
 301         rq = task_rq_lock(task, &flags);
 302
 303         if (task_running(rq, task) && task != current) {
 304                 snprintf(err_buf, STACK_ERR_BUF_SIZE,
 305                          "%s: %s:%d is running\n", __func__, task->comm,
 306                          task->pid);
 307                 goto done;
 308         }
 309
 310         ret = klp_check_stack(task, err_buf);
 311         if (ret)
 312                 goto done;
 313
 314         success = true;
 315
 316         clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
 317         task->patch_state = klp_target_state;
 318
 319 done:
 320         task_rq_unlock(rq, task, &flags);
 321
 322         /*
 323          * Due to console deadlock issues, pr_debug() can't be used while
 324          * holding the task rq lock.  Instead we have to use a temporary buffer
 325          * and print the debug message after releasing the lock.
 326          */
 327         if (err_buf[0] != '\0')
 328                 pr_debug("%s", err_buf);
 329
 330         return success;
 331
 332 }
 333
 334 /*
 335  * Sends a fake signal to all non-kthread tasks with TIF_PATCH_PENDING set.
 336  * Kthreads with TIF_PATCH_PENDING set are woken up.
 337  */
 338 static void klp_send_signals(void)
 339 {
 340         struct task_struct *g, *task;
 341
 342         if (klp_signals_cnt == SIGNALS_TIMEOUT)
 343                 pr_notice("signaling remaining tasks\n");
 344
 345         read_lock(&tasklist_lock);
 346         for_each_process_thread(g, task) {
 347                 if (!klp_patch_pending(task))
 348                         continue;
 349
 350                 /*
 351                  * There is a small race here. We could see TIF_PATCH_PENDING
 352                  * set and decide to wake up a kthread or send a fake signal.
 353                  * Meanwhile the task could migrate itself and the action
 354                  * would be meaningless. It is not serious though.
 355                  */
 356                 if (task->flags & PF_KTHREAD) {
 357                         /*
 358                          * Wake up a kthread which sleeps interruptedly and
 359                          * still has not been migrated.
 360                          */
 361                         wake_up_state(task, TASK_INTERRUPTIBLE);
 362                 } else {
 363                         /*
 364                          * Send fake signal to all non-kthread tasks which are
 365                          * still not migrated.
 366                          */
 367                         spin_lock_irq(&task->sighand->siglock);
 368                         signal_wake_up(task, 0);
 369                         spin_unlock_irq(&task->sighand->siglock);
 370                 }
 371         }
 372         read_unlock(&tasklist_lock);
 373 }
 374
 375 /*
 376  * Try to switch all remaining tasks to the target patch state by walking the
 377  * stacks of sleeping tasks and looking for any to-be-patched or
 378  * to-be-unpatched functions.  If such functions are found, the task can't be
 379  * switched yet.
 380  *
 381  * If any tasks are still stuck in the initial patch state, schedule a retry.
 382  */
 383 void klp_try_complete_transition(void)
 384 {
 385         unsigned int cpu;
 386         struct task_struct *g, *task;
 387         struct klp_patch *patch;
 388         bool complete = true;
 389
 390         WARN_ON_ONCE(klp_target_state == KLP_UNDEFINED);
 391
 392         /*
 393          * Try to switch the tasks to the target patch state by walking their
 394          * stacks and looking for any to-be-patched or to-be-unpatched
 395          * functions.  If such functions are found on a stack, or if the stack
 396          * is deemed unreliable, the task can't be switched yet.
 397          *
 398          * Usually this will transition most (or all) of the tasks on a system
 399          * unless the patch includes changes to a very common function.
 400          */
 401         read_lock(&tasklist_lock);
 402         for_each_process_thread(g, task)
 403                 if (!klp_try_switch_task(task))
 404                         complete = false;
 405         read_unlock(&tasklist_lock);
 406
 407         /*
 408          * Ditto for the idle "swapper" tasks.
 409          */
 410         get_online_cpus();
 411         for_each_possible_cpu(cpu) {
 412                 task = idle_task(cpu);
 413                 if (cpu_online(cpu)) {
 414                         if (!klp_try_switch_task(task))
 415                                 complete = false;
 416                 } else if (task->patch_state != klp_target_state) {
 417                         /* offline idle tasks can be switched immediately */
 418                         clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
 419                         task->patch_state = klp_target_state;
 420                 }
 421         }
 422         put_online_cpus();
 423
 424         if (!complete) {
 425                 if (klp_signals_cnt && !(klp_signals_cnt % SIGNALS_TIMEOUT))
 426                         klp_send_signals();
 427                 klp_signals_cnt++;
 428
 429                 /*
 430                  * Some tasks weren't able to be switched over.  Try again
 431                  * later and/or wait for other methods like kernel exit
 432                  * switching.
 433                  */
 434                 schedule_delayed_work(&klp_transition_work,
 435                                       round_jiffies_relative(HZ));
 436                 return;
 437         }
 438
 439         /* we're done, now cleanup the data structures */
 440         patch = klp_transition_patch;
 441         klp_complete_transition();
 442
 443         /*
 444          * It would make more sense to free the patch in
 445          * klp_complete_transition() but it is called also
 446          * from klp_cancel_transition().
 447          */
 448         if (!patch->enabled) {
 449                 klp_free_patch_start(patch);
 450                 schedule_work(&patch->free_work);
 451         }
 452 }
 453
 454 /*
 455  * Start the transition to the specified target patch state so tasks can begin
 456  * switching to it.
 457  */
 458 void klp_start_transition(void)
 459 {
 460         struct task_struct *g, *task;
 461         unsigned int cpu;
 462
 463         WARN_ON_ONCE(klp_target_state == KLP_UNDEFINED);
 464
 465         pr_notice("'%s': starting %s transition\n",
 466                   klp_transition_patch->mod->name,
 467                   klp_target_state == KLP_PATCHED ? "patching" : "unpatching");
 468
 469         /*
 470          * Mark all normal tasks as needing a patch state update.  They'll
 471          * switch either in klp_try_complete_transition() or as they exit the
 472          * kernel.
 473          */
 474         read_lock(&tasklist_lock);
 475         for_each_process_thread(g, task)
 476                 if (task->patch_state != klp_target_state)
 477                         set_tsk_thread_flag(task, TIF_PATCH_PENDING);
 478         read_unlock(&tasklist_lock);
 479
 480         /*
 481          * Mark all idle tasks as needing a patch state update.  They'll switch
 482          * either in klp_try_complete_transition() or at the idle loop switch
 483          * point.
 484          */
 485         for_each_possible_cpu(cpu) {
 486                 task = idle_task(cpu);
 487                 if (task->patch_state != klp_target_state)
 488                         set_tsk_thread_flag(task, TIF_PATCH_PENDING);
 489         }
 490
 491         klp_signals_cnt = 0;
 492 }
 493
 494 /*
 495  * Initialize the global target patch state and all tasks to the initial patch
 496  * state, and initialize all function transition states to true in preparation
 497  * for patching or unpatching.
 498  */
 499 void klp_init_transition(struct klp_patch *patch, int state)
 500 {
 501         struct task_struct *g, *task;
 502         unsigned int cpu;
 503         struct klp_object *obj;
 504         struct klp_func *func;
 505         int initial_state = !state;
 506
 507         WARN_ON_ONCE(klp_target_state != KLP_UNDEFINED);
 508
 509         klp_transition_patch = patch;
 510
 511         /*
 512          * Set the global target patch state which tasks will switch to.  This
 513          * has no effect until the TIF_PATCH_PENDING flags get set later.
 514          */
 515         klp_target_state = state;
 516
 517         pr_debug("'%s': initializing %s transition\n", patch->mod->name,
 518                  klp_target_state == KLP_PATCHED ? "patching" : "unpatching");
 519
 520         /*
 521          * Initialize all tasks to the initial patch state to prepare them for
 522          * switching to the target state.
 523          */
 524         read_lock(&tasklist_lock);
 525         for_each_process_thread(g, task) {
 526                 WARN_ON_ONCE(task->patch_state != KLP_UNDEFINED);
 527                 task->patch_state = initial_state;
 528         }
 529         read_unlock(&tasklist_lock);
 530
 531         /*
 532          * Ditto for the idle "swapper" tasks.
 533          */
 534         for_each_possible_cpu(cpu) {
 535                 task = idle_task(cpu);
 536                 WARN_ON_ONCE(task->patch_state != KLP_UNDEFINED);
 537                 task->patch_state = initial_state;
 538         }
 539
 540         /*
 541          * Enforce the order of the task->patch_state initializations and the
 542          * func->transition updates to ensure that klp_ftrace_handler() doesn't
 543          * see a func in transition with a task->patch_state of KLP_UNDEFINED.
 544          *
 545          * Also enforce the order of the klp_target_state write and future
 546          * TIF_PATCH_PENDING writes to ensure klp_update_patch_state() doesn't
 547          * set a task->patch_state to KLP_UNDEFINED.
 548          */
 549         smp_wmb();
 550
 551         /*
 552          * Set the func transition states so klp_ftrace_handler() will know to
 553          * switch to the transition logic.
 554          *
 555          * When patching, the funcs aren't yet in the func_stack and will be
 556          * made visible to the ftrace handler shortly by the calls to
 557          * klp_patch_object().
 558          *
 559          * When unpatching, the funcs are already in the func_stack and so are
 560          * already visible to the ftrace handler.
 561          */
 562         klp_for_each_object(patch, obj)
 563                 klp_for_each_func(obj, func)
 564                         func->transition = true;
 565 }
 566
 567 /*
 568  * This function can be called in the middle of an existing transition to
 569  * reverse the direction of the target patch state.  This can be done to
 570  * effectively cancel an existing enable or disable operation if there are any
 571  * tasks which are stuck in the initial patch state.
 572  */
 573 void klp_reverse_transition(void)
 574 {
 575         unsigned int cpu;
 576         struct task_struct *g, *task;
 577
 578         pr_debug("'%s': reversing transition from %s\n",
 579                  klp_transition_patch->mod->name,
 580                  klp_target_state == KLP_PATCHED ? "patching to unpatching" :
 581                                                    "unpatching to patching");
 582
 583         klp_transition_patch->enabled = !klp_transition_patch->enabled;
 584
 585         klp_target_state = !klp_target_state;
 586
 587         /*
 588          * Clear all TIF_PATCH_PENDING flags to prevent races caused by
 589          * klp_update_patch_state() running in parallel with
 590          * klp_start_transition().
 591          */
 592         read_lock(&tasklist_lock);
 593         for_each_process_thread(g, task)
 594                 clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
 595         read_unlock(&tasklist_lock);
 596
 597         for_each_possible_cpu(cpu)
 598                 clear_tsk_thread_flag(idle_task(cpu), TIF_PATCH_PENDING);
 599
 600         /* Let any remaining calls to klp_update_patch_state() complete */
 601         klp_synchronize_transition();
 602
 603         klp_start_transition();
 604 }
 605
 606 /* Called from copy_process() during fork */
 607 void klp_copy_process(struct task_struct *child)
 608 {
 609         child->patch_state = current->patch_state;
 610
 611         /* TIF_PATCH_PENDING gets copied in setup_thread_stack() */
 612 }
 613
 614 /*
 615  * Drop TIF_PATCH_PENDING of all tasks on admin's request. This forces an
 616  * existing transition to finish.
 617  *
 618  * NOTE: klp_update_patch_state(task) requires the task to be inactive or
 619  * 'current'. This is not the case here and the consistency model could be
 620  * broken. Administrator, who is the only one to execute the
 621  * klp_force_transitions(), has to be aware of this.
 622  */
 623 void klp_force_transition(void)
 624 {
 625         struct klp_patch *patch;
 626         struct task_struct *g, *task;
 627         unsigned int cpu;
 628
 629         pr_warn("forcing remaining tasks to the patched state\n");
 630
 631         read_lock(&tasklist_lock);
 632         for_each_process_thread(g, task)
 633                 klp_update_patch_state(task);
 634         read_unlock(&tasklist_lock);
 635
 636         for_each_possible_cpu(cpu)
 637                 klp_update_patch_state(idle_task(cpu));
 638
 639         klp_for_each_patch(patch)
 640                 patch->forced = true;
 641 }