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