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 <linux/tracehook.h>
  13 #include "core.h"
  14 #include "patch.h"
  15 #include "transition.h"
  16 #include "../sched/sched.h"
  17
  18 #define MAX_STACK_ENTRIES  100
  19 #define STACK_ERR_BUF_SIZE 128
  20
  21 #define SIGNALS_TIMEOUT 15
  22
  23 struct klp_patch *klp_transition_patch;
  24
  25 static int klp_target_state = KLP_UNDEFINED;
  26
  27 static unsigned int klp_signals_cnt;
  28
  29 /*
  30  * This work can be performed periodically to finish patching or unpatching any
  31  * "straggler" tasks which failed to transition in the first attempt.
  32  */
  33 static void klp_transition_work_fn(struct work_struct *work)
  34 {
  35         mutex_lock(&klp_mutex);
  36
  37         if (klp_transition_patch)
  38                 klp_try_complete_transition();
  39
  40         mutex_unlock(&klp_mutex);
  41 }
  42 static DECLARE_DELAYED_WORK(klp_transition_work, klp_transition_work_fn);
  43
  44 /*
  45  * This function is just a stub to implement a hard force
  46  * of synchronize_rcu(). This requires synchronizing
  47  * tasks even in userspace and idle.
  48  */
  49 static void klp_sync(struct work_struct *work)
  50 {
  51 }
  52
  53 /*
  54  * We allow to patch also functions where RCU is not watching,
  55  * e.g. before user_exit(). We can not rely on the RCU infrastructure
  56  * to do the synchronization. Instead hard force the sched synchronization.
  57  *
  58  * This approach allows to use RCU functions for manipulating func_stack
  59  * safely.
  60  */
  61 static void klp_synchronize_transition(void)
  62 {
  63         schedule_on_each_cpu(klp_sync);
  64 }
  65
  66 /*
  67  * The transition to the target patch state is complete.  Clean up the data
  68  * structures.
  69  */
  70 static void klp_complete_transition(void)
  71 {
  72         struct klp_object *obj;
  73         struct klp_func *func;
  74         struct task_struct *g, *task;
  75         unsigned int cpu;
  76
  77         pr_debug("'%s': completing %s transition\n",
  78                  klp_transition_patch->mod->name,
  79                  klp_target_state == KLP_PATCHED ? "patching" : "unpatching");
  80
  81         if (klp_transition_patch->replace && klp_target_state == KLP_PATCHED) {
  82                 klp_unpatch_replaced_patches(klp_transition_patch);
  83                 klp_discard_nops(klp_transition_patch);
  84         }
  85
  86         if (klp_target_state == KLP_UNPATCHED) {
  87                 /*
  88                  * All tasks have transitioned to KLP_UNPATCHED so we can now
  89                  * remove the new functions from the func_stack.
  90                  */
  91                 klp_unpatch_objects(klp_transition_patch);
  92
  93                 /*
  94                  * Make sure klp_ftrace_handler() can no longer see functions
  95                  * from this patch on the ops->func_stack.  Otherwise, after
  96                  * func->transition gets cleared, the handler may choose a
  97                  * removed function.
  98                  */
  99                 klp_synchronize_transition();
 100         }
 101
 102         klp_for_each_object(klp_transition_patch, obj)
 103                 klp_for_each_func(obj, func)
 104                         func->transition = false;
 105
 106         /* Prevent klp_ftrace_handler() from seeing KLP_UNDEFINED state */
 107         if (klp_target_state == KLP_PATCHED)
 108                 klp_synchronize_transition();
 109
 110         read_lock(&tasklist_lock);
 111         for_each_process_thread(g, task) {
 112                 WARN_ON_ONCE(test_tsk_thread_flag(task, TIF_PATCH_PENDING));
 113                 task->patch_state = KLP_UNDEFINED;
 114         }
 115         read_unlock(&tasklist_lock);
 116
 117         for_each_possible_cpu(cpu) {
 118                 task = idle_task(cpu);
 119                 WARN_ON_ONCE(test_tsk_thread_flag(task, TIF_PATCH_PENDING));
 120                 task->patch_state = KLP_UNDEFINED;
 121         }
 122
 123         klp_for_each_object(klp_transition_patch, obj) {
 124                 if (!klp_is_object_loaded(obj))
 125                         continue;
 126                 if (klp_target_state == KLP_PATCHED)
 127                         klp_post_patch_callback(obj);
 128                 else if (klp_target_state == KLP_UNPATCHED)
 129                         klp_post_unpatch_callback(obj);
 130         }
 131
 132         pr_notice("'%s': %s complete\n", klp_transition_patch->mod->name,
 133                   klp_target_state == KLP_PATCHED ? "patching" : "unpatching");
 134
 135         klp_target_state = KLP_UNDEFINED;
 136         klp_transition_patch = NULL;
 137 }
 138
 139 /*
 140  * This is called in the error path, to cancel a transition before it has
 141  * started, i.e. klp_init_transition() has been called but
 142  * klp_start_transition() hasn't.  If the transition *has* been started,
 143  * klp_reverse_transition() should be used instead.
 144  */
 145 void klp_cancel_transition(void)
 146 {
 147         if (WARN_ON_ONCE(klp_target_state != KLP_PATCHED))
 148                 return;
 149
 150         pr_debug("'%s': canceling patching transition, going to unpatch\n",
 151                  klp_transition_patch->mod->name);
 152
 153         klp_target_state = KLP_UNPATCHED;
 154         klp_complete_transition();
 155 }
 156
 157 /*
 158  * Switch the patched state of the task to the set of functions in the target
 159  * patch state.
 160  *
 161  * NOTE: If task is not 'current', the caller must ensure the task is inactive.
 162  * Otherwise klp_ftrace_handler() might read the wrong 'patch_state' value.
 163  */
 164 void klp_update_patch_state(struct task_struct *task)
 165 {
 166         /*
 167          * A variant of synchronize_rcu() is used to allow patching functions
 168          * where RCU is not watching, see klp_synchronize_transition().
 169          */
 170         preempt_disable_notrace();
 171
 172         /*
 173          * This test_and_clear_tsk_thread_flag() call also serves as a read
 174          * barrier (smp_rmb) for two cases:
 175          *
 176          * 1) Enforce the order of the TIF_PATCH_PENDING read and the
 177          *    klp_target_state read.  The corresponding write barrier is in
 178          *    klp_init_transition().
 179          *
 180          * 2) Enforce the order of the TIF_PATCH_PENDING read and a future read
 181          *    of func->transition, if klp_ftrace_handler() is called later on
 182          *    the same CPU.  See __klp_disable_patch().
 183          */
 184         if (test_and_clear_tsk_thread_flag(task, TIF_PATCH_PENDING))
 185                 task->patch_state = READ_ONCE(klp_target_state);
 186
 187         preempt_enable_notrace();
 188 }
 189
 190 /*
 191  * Determine whether the given stack trace includes any references to a
 192  * to-be-patched or to-be-unpatched function.
 193  */
 194 static int klp_check_stack_func(struct klp_func *func, unsigned long *entries,
 195                                 unsigned int nr_entries)
 196 {
 197         unsigned long func_addr, func_size, address;
 198         struct klp_ops *ops;
 199         int i;
 200
 201         for (i = 0; i < nr_entries; i++) {
 202                 address = entries[i];
 203
 204                 if (klp_target_state == KLP_UNPATCHED) {
 205                          /*
 206                           * Check for the to-be-unpatched function
 207                           * (the func itself).
 208                           */
 209                         func_addr = (unsigned long)func->new_func;
 210                         func_size = func->new_size;
 211                 } else {
 212                         /*
 213                          * Check for the to-be-patched function
 214                          * (the previous func).
 215                          */
 216                         ops = klp_find_ops(func->old_func);
 217
 218                         if (list_is_singular(&ops->func_stack)) {
 219                                 /* original function */
 220                                 func_addr = (unsigned long)func->old_func;
 221                                 func_size = func->old_size;
 222                         } else {
 223                                 /* previously patched function */
 224                                 struct klp_func *prev;
 225
 226                                 prev = list_next_entry(func, stack_node);
 227                                 func_addr = (unsigned long)prev->new_func;
 228                                 func_size = prev->new_size;
 229                         }
 230                 }
 231
 232                 if (address >= func_addr && address < func_addr + func_size)
 233                         return -EAGAIN;
 234         }
 235
 236         return 0;
 237 }
 238
 239 /*
 240  * Determine whether it's safe to transition the task to the target patch state
 241  * by looking for any to-be-patched or to-be-unpatched functions on its stack.
 242  */
 243 static int klp_check_stack(struct task_struct *task, char *err_buf)
 244 {
 245         static unsigned long entries[MAX_STACK_ENTRIES];
 246         struct klp_object *obj;
 247         struct klp_func *func;
 248         int ret, nr_entries;
 249
 250         ret = stack_trace_save_tsk_reliable(task, entries, ARRAY_SIZE(entries));
 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         static char err_buf[STACK_ERR_BUF_SIZE];
 285         struct rq *rq;
 286         struct rq_flags flags;
 287         int ret;
 288         bool success = false;
 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          * For arches which don't have reliable stack traces, we have to rely
 298          * on other methods (e.g., switching tasks at kernel exit).
 299          */
 300         if (!klp_have_reliable_stack())
 301                 return false;
 302
 303         /*
 304          * Now try to check the stack for any to-be-patched or to-be-unpatched
 305          * functions.  If all goes well, switch the task to the target patch
 306          * state.
 307          */
 308         rq = task_rq_lock(task, &flags);
 309
 310         if (task_running(rq, task) && task != current) {
 311                 snprintf(err_buf, STACK_ERR_BUF_SIZE,
 312                          "%s: %s:%d is running\n", __func__, task->comm,
 313                          task->pid);
 314                 goto done;
 315         }
 316
 317         ret = klp_check_stack(task, err_buf);
 318         if (ret)
 319                 goto done;
 320
 321         success = true;
 322
 323         clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
 324         task->patch_state = klp_target_state;
 325
 326 done:
 327         task_rq_unlock(rq, task, &flags);
 328
 329         /*
 330          * Due to console deadlock issues, pr_debug() can't be used while
 331          * holding the task rq lock.  Instead we have to use a temporary buffer
 332          * and print the debug message after releasing the lock.
 333          */
 334         if (err_buf[0] != '\0')
 335                 pr_debug("%s", err_buf);
 336
 337         return success;
 338 }
 339
 340 /*
 341  * Sends a fake signal to all non-kthread tasks with TIF_PATCH_PENDING set.
 342  * Kthreads with TIF_PATCH_PENDING set are woken up.
 343  */
 344 static void klp_send_signals(void)
 345 {
 346         struct task_struct *g, *task;
 347
 348         if (klp_signals_cnt == SIGNALS_TIMEOUT)
 349                 pr_notice("signaling remaining tasks\n");
 350
 351         read_lock(&tasklist_lock);
 352         for_each_process_thread(g, task) {
 353                 if (!klp_patch_pending(task))
 354                         continue;
 355
 356                 /*
 357                  * There is a small race here. We could see TIF_PATCH_PENDING
 358                  * set and decide to wake up a kthread or send a fake signal.
 359                  * Meanwhile the task could migrate itself and the action
 360                  * would be meaningless. It is not serious though.
 361                  */
 362                 if (task->flags & PF_KTHREAD) {
 363                         /*
 364                          * Wake up a kthread which sleeps interruptedly and
 365                          * still has not been migrated.
 366                          */
 367                         wake_up_state(task, TASK_INTERRUPTIBLE);
 368                 } else {
 369                         /*
 370                          * Send fake signal to all non-kthread tasks which are
 371                          * still not migrated.
 372                          */
 373                         set_notify_signal(task);
 374                 }
 375         }
 376         read_unlock(&tasklist_lock);
 377 }
 378
 379 /*
 380  * Try to switch all remaining tasks to the target patch state by walking the
 381  * stacks of sleeping tasks and looking for any to-be-patched or
 382  * to-be-unpatched functions.  If such functions are found, the task can't be
 383  * switched yet.
 384  *
 385  * If any tasks are still stuck in the initial patch state, schedule a retry.
 386  */
 387 void klp_try_complete_transition(void)
 388 {
 389         unsigned int cpu;
 390         struct task_struct *g, *task;
 391         struct klp_patch *patch;
 392         bool complete = true;
 393
 394         WARN_ON_ONCE(klp_target_state == KLP_UNDEFINED);
 395
 396         /*
 397          * Try to switch the tasks to the target patch state by walking their
 398          * stacks and looking for any to-be-patched or to-be-unpatched
 399          * functions.  If such functions are found on a stack, or if the stack
 400          * is deemed unreliable, the task can't be switched yet.
 401          *
 402          * Usually this will transition most (or all) of the tasks on a system
 403          * unless the patch includes changes to a very common function.
 404          */
 405         read_lock(&tasklist_lock);
 406         for_each_process_thread(g, task)
 407                 if (!klp_try_switch_task(task))
 408                         complete = false;
 409         read_unlock(&tasklist_lock);
 410
 411         /*
 412          * Ditto for the idle "swapper" tasks.
 413          */
 414         get_online_cpus();
 415         for_each_possible_cpu(cpu) {
 416                 task = idle_task(cpu);
 417                 if (cpu_online(cpu)) {
 418                         if (!klp_try_switch_task(task))
 419                                 complete = false;
 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         put_online_cpus();
 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  */
 462 void 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  */
 503 void 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  */
 577 void 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 */
 611 void klp_copy_process(struct task_struct *child)
 612 {
 613         child->patch_state = current->patch_state;
 614
 615         /* TIF_PATCH_PENDING gets copied in setup_thread_stack() */
 616 }
 617
 618 /*
 619  * Drop TIF_PATCH_PENDING of all tasks on admin's request. This forces an
 620  * existing transition to finish.
 621  *
 622  * NOTE: klp_update_patch_state(task) requires the task to be inactive or
 623  * 'current'. This is not the case here and the consistency model could be
 624  * broken. Administrator, who is the only one to execute the
 625  * klp_force_transitions(), has to be aware of this.
 626  */
 627 void klp_force_transition(void)
 628 {
 629         struct klp_patch *patch;
 630         struct task_struct *g, *task;
 631         unsigned int cpu;
 632
 633         pr_warn("forcing remaining tasks to the patched state\n");
 634
 635         read_lock(&tasklist_lock);
 636         for_each_process_thread(g, task)
 637                 klp_update_patch_state(task);
 638         read_unlock(&tasklist_lock);
 639
 640         for_each_possible_cpu(cpu)
 641                 klp_update_patch_state(idle_task(cpu));
 642
 643         klp_for_each_patch(patch)
 644                 patch->forced = true;
 645 }