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