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