kernel/kcsan/core.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * KCSAN core runtime.
   4  *
   5  * Copyright (C) 2019, Google LLC.
   6  */
   7
   8 #define pr_fmt(fmt) "kcsan: " fmt
   9
  10 #include <linux/atomic.h>
  11 #include <linux/bug.h>
  12 #include <linux/delay.h>
  13 #include <linux/export.h>
  14 #include <linux/init.h>
  15 #include <linux/kernel.h>
  16 #include <linux/list.h>
  17 #include <linux/moduleparam.h>
  18 #include <linux/percpu.h>
  19 #include <linux/preempt.h>
  20 #include <linux/sched.h>
  21 #include <linux/uaccess.h>
  22
  23 #include "encoding.h"
  24 #include "kcsan.h"
  25 #include "permissive.h"
  26
  27 static bool kcsan_early_enable = IS_ENABLED(CONFIG_KCSAN_EARLY_ENABLE);
  28 unsigned int kcsan_udelay_task = CONFIG_KCSAN_UDELAY_TASK;
  29 unsigned int kcsan_udelay_interrupt = CONFIG_KCSAN_UDELAY_INTERRUPT;
  30 static long kcsan_skip_watch = CONFIG_KCSAN_SKIP_WATCH;
  31 static bool kcsan_interrupt_watcher = IS_ENABLED(CONFIG_KCSAN_INTERRUPT_WATCHER);
  32
  33 #ifdef MODULE_PARAM_PREFIX
  34 #undef MODULE_PARAM_PREFIX
  35 #endif
  36 #define MODULE_PARAM_PREFIX "kcsan."
  37 module_param_named(early_enable, kcsan_early_enable, bool, 0);
  38 module_param_named(udelay_task, kcsan_udelay_task, uint, 0644);
  39 module_param_named(udelay_interrupt, kcsan_udelay_interrupt, uint, 0644);
  40 module_param_named(skip_watch, kcsan_skip_watch, long, 0644);
  41 module_param_named(interrupt_watcher, kcsan_interrupt_watcher, bool, 0444);
  42
  43 bool kcsan_enabled;
  44
  45 /* Per-CPU kcsan_ctx for interrupts */
  46 static DEFINE_PER_CPU(struct kcsan_ctx, kcsan_cpu_ctx) = {
  47         .disable_count          = 0,
  48         .atomic_next            = 0,
  49         .atomic_nest_count      = 0,
  50         .in_flat_atomic         = false,
  51         .access_mask            = 0,
  52         .scoped_accesses        = {LIST_POISON1, NULL},
  53 };
  54
  55 /*
  56  * Helper macros to index into adjacent slots, starting from address slot
  57  * itself, followed by the right and left slots.
  58  *
  59  * The purpose is 2-fold:
  60  *
  61  *      1. if during insertion the address slot is already occupied, check if
  62  *         any adjacent slots are free;
  63  *      2. accesses that straddle a slot boundary due to size that exceeds a
  64  *         slot's range may check adjacent slots if any watchpoint matches.
  65  *
  66  * Note that accesses with very large size may still miss a watchpoint; however,
  67  * given this should be rare, this is a reasonable trade-off to make, since this
  68  * will avoid:
  69  *
  70  *      1. excessive contention between watchpoint checks and setup;
  71  *      2. larger number of simultaneous watchpoints without sacrificing
  72  *         performance.
  73  *
  74  * Example: SLOT_IDX values for KCSAN_CHECK_ADJACENT=1, where i is [0, 1, 2]:
  75  *
  76  *   slot=0:  [ 1,  2,  0]
  77  *   slot=9:  [10, 11,  9]
  78  *   slot=63: [64, 65, 63]
  79  */
  80 #define SLOT_IDX(slot, i) (slot + ((i + KCSAN_CHECK_ADJACENT) % NUM_SLOTS))
  81
  82 /*
  83  * SLOT_IDX_FAST is used in the fast-path. Not first checking the address's primary
  84  * slot (middle) is fine if we assume that races occur rarely. The set of
  85  * indices {SLOT_IDX(slot, i) | i in [0, NUM_SLOTS)} is equivalent to
  86  * {SLOT_IDX_FAST(slot, i) | i in [0, NUM_SLOTS)}.
  87  */
  88 #define SLOT_IDX_FAST(slot, i) (slot + i)
  89
  90 /*
  91  * Watchpoints, with each entry encoded as defined in encoding.h: in order to be
  92  * able to safely update and access a watchpoint without introducing locking
  93  * overhead, we encode each watchpoint as a single atomic long. The initial
  94  * zero-initialized state matches INVALID_WATCHPOINT.
  95  *
  96  * Add NUM_SLOTS-1 entries to account for overflow; this helps avoid having to
  97  * use more complicated SLOT_IDX_FAST calculation with modulo in the fast-path.
  98  */
  99 static atomic_long_t watchpoints[CONFIG_KCSAN_NUM_WATCHPOINTS + NUM_SLOTS-1];
 100
 101 /*
 102  * Instructions to skip watching counter, used in should_watch(). We use a
 103  * per-CPU counter to avoid excessive contention.
 104  */
 105 static DEFINE_PER_CPU(long, kcsan_skip);
 106
 107 /* For kcsan_prandom_u32_max(). */
 108 static DEFINE_PER_CPU(u32, kcsan_rand_state);
 109
 110 static __always_inline atomic_long_t *find_watchpoint(unsigned long addr,
 111                                                       size_t size,
 112                                                       bool expect_write,
 113                                                       long *encoded_watchpoint)
 114 {
 115         const int slot = watchpoint_slot(addr);
 116         const unsigned long addr_masked = addr & WATCHPOINT_ADDR_MASK;
 117         atomic_long_t *watchpoint;
 118         unsigned long wp_addr_masked;
 119         size_t wp_size;
 120         bool is_write;
 121         int i;
 122
 123         BUILD_BUG_ON(CONFIG_KCSAN_NUM_WATCHPOINTS < NUM_SLOTS);
 124
 125         for (i = 0; i < NUM_SLOTS; ++i) {
 126                 watchpoint = &watchpoints[SLOT_IDX_FAST(slot, i)];
 127                 *encoded_watchpoint = atomic_long_read(watchpoint);
 128                 if (!decode_watchpoint(*encoded_watchpoint, &wp_addr_masked,
 129                                        &wp_size, &is_write))
 130                         continue;
 131
 132                 if (expect_write && !is_write)
 133                         continue;
 134
 135                 /* Check if the watchpoint matches the access. */
 136                 if (matching_access(wp_addr_masked, wp_size, addr_masked, size))
 137                         return watchpoint;
 138         }
 139
 140         return NULL;
 141 }
 142
 143 static inline atomic_long_t *
 144 insert_watchpoint(unsigned long addr, size_t size, bool is_write)
 145 {
 146         const int slot = watchpoint_slot(addr);
 147         const long encoded_watchpoint = encode_watchpoint(addr, size, is_write);
 148         atomic_long_t *watchpoint;
 149         int i;
 150
 151         /* Check slot index logic, ensuring we stay within array bounds. */
 152         BUILD_BUG_ON(SLOT_IDX(0, 0) != KCSAN_CHECK_ADJACENT);
 153         BUILD_BUG_ON(SLOT_IDX(0, KCSAN_CHECK_ADJACENT+1) != 0);
 154         BUILD_BUG_ON(SLOT_IDX(CONFIG_KCSAN_NUM_WATCHPOINTS-1, KCSAN_CHECK_ADJACENT) != ARRAY_SIZE(watchpoints)-1);
 155         BUILD_BUG_ON(SLOT_IDX(CONFIG_KCSAN_NUM_WATCHPOINTS-1, KCSAN_CHECK_ADJACENT+1) != ARRAY_SIZE(watchpoints) - NUM_SLOTS);
 156
 157         for (i = 0; i < NUM_SLOTS; ++i) {
 158                 long expect_val = INVALID_WATCHPOINT;
 159
 160                 /* Try to acquire this slot. */
 161                 watchpoint = &watchpoints[SLOT_IDX(slot, i)];
 162                 if (atomic_long_try_cmpxchg_relaxed(watchpoint, &expect_val, encoded_watchpoint))
 163                         return watchpoint;
 164         }
 165
 166         return NULL;
 167 }
 168
 169 /*
 170  * Return true if watchpoint was successfully consumed, false otherwise.
 171  *
 172  * This may return false if:
 173  *
 174  *      1. another thread already consumed the watchpoint;
 175  *      2. the thread that set up the watchpoint already removed it;
 176  *      3. the watchpoint was removed and then re-used.
 177  */
 178 static __always_inline bool
 179 try_consume_watchpoint(atomic_long_t *watchpoint, long encoded_watchpoint)
 180 {
 181         return atomic_long_try_cmpxchg_relaxed(watchpoint, &encoded_watchpoint, CONSUMED_WATCHPOINT);
 182 }
 183
 184 /* Return true if watchpoint was not touched, false if already consumed. */
 185 static inline bool consume_watchpoint(atomic_long_t *watchpoint)
 186 {
 187         return atomic_long_xchg_relaxed(watchpoint, CONSUMED_WATCHPOINT) != CONSUMED_WATCHPOINT;
 188 }
 189
 190 /* Remove the watchpoint -- its slot may be reused after. */
 191 static inline void remove_watchpoint(atomic_long_t *watchpoint)
 192 {
 193         atomic_long_set(watchpoint, INVALID_WATCHPOINT);
 194 }
 195
 196 static __always_inline struct kcsan_ctx *get_ctx(void)
 197 {
 198         /*
 199          * In interrupts, use raw_cpu_ptr to avoid unnecessary checks, that would
 200          * also result in calls that generate warnings in uaccess regions.
 201          */
 202         return in_task() ? &current->kcsan_ctx : raw_cpu_ptr(&kcsan_cpu_ctx);
 203 }
 204
 205 /* Check scoped accesses; never inline because this is a slow-path! */
 206 static noinline void kcsan_check_scoped_accesses(void)
 207 {
 208         struct kcsan_ctx *ctx = get_ctx();
 209         struct list_head *prev_save = ctx->scoped_accesses.prev;
 210         struct kcsan_scoped_access *scoped_access;
 211
 212         ctx->scoped_accesses.prev = NULL;  /* Avoid recursion. */
 213         list_for_each_entry(scoped_access, &ctx->scoped_accesses, list)
 214                 __kcsan_check_access(scoped_access->ptr, scoped_access->size, scoped_access->type);
 215         ctx->scoped_accesses.prev = prev_save;
 216 }
 217
 218 /* Rules for generic atomic accesses. Called from fast-path. */
 219 static __always_inline bool
 220 is_atomic(const volatile void *ptr, size_t size, int type, struct kcsan_ctx *ctx)
 221 {
 222         if (type & KCSAN_ACCESS_ATOMIC)
 223                 return true;
 224
 225         /*
 226          * Unless explicitly declared atomic, never consider an assertion access
 227          * as atomic. This allows using them also in atomic regions, such as
 228          * seqlocks, without implicitly changing their semantics.
 229          */
 230         if (type & KCSAN_ACCESS_ASSERT)
 231                 return false;
 232
 233         if (IS_ENABLED(CONFIG_KCSAN_ASSUME_PLAIN_WRITES_ATOMIC) &&
 234             (type & KCSAN_ACCESS_WRITE) && size <= sizeof(long) &&
 235             !(type & KCSAN_ACCESS_COMPOUND) && IS_ALIGNED((unsigned long)ptr, size))
 236                 return true; /* Assume aligned writes up to word size are atomic. */
 237
 238         if (ctx->atomic_next > 0) {
 239                 /*
 240                  * Because we do not have separate contexts for nested
 241                  * interrupts, in case atomic_next is set, we simply assume that
 242                  * the outer interrupt set atomic_next. In the worst case, we
 243                  * will conservatively consider operations as atomic. This is a
 244                  * reasonable trade-off to make, since this case should be
 245                  * extremely rare; however, even if extremely rare, it could
 246                  * lead to false positives otherwise.
 247                  */
 248                 if ((hardirq_count() >> HARDIRQ_SHIFT) < 2)
 249                         --ctx->atomic_next; /* in task, or outer interrupt */
 250                 return true;
 251         }
 252
 253         return ctx->atomic_nest_count > 0 || ctx->in_flat_atomic;
 254 }
 255
 256 static __always_inline bool
 257 should_watch(const volatile void *ptr, size_t size, int type, struct kcsan_ctx *ctx)
 258 {
 259         /*
 260          * Never set up watchpoints when memory operations are atomic.
 261          *
 262          * Need to check this first, before kcsan_skip check below: (1) atomics
 263          * should not count towards skipped instructions, and (2) to actually
 264          * decrement kcsan_atomic_next for consecutive instruction stream.
 265          */
 266         if (is_atomic(ptr, size, type, ctx))
 267                 return false;
 268
 269         if (this_cpu_dec_return(kcsan_skip) >= 0)
 270                 return false;
 271
 272         /*
 273          * NOTE: If we get here, kcsan_skip must always be reset in slow path
 274          * via reset_kcsan_skip() to avoid underflow.
 275          */
 276
 277         /* this operation should be watched */
 278         return true;
 279 }
 280
 281 /*
 282  * Returns a pseudo-random number in interval [0, ep_ro). Simple linear
 283  * congruential generator, using constants from "Numerical Recipes".
 284  */
 285 static u32 kcsan_prandom_u32_max(u32 ep_ro)
 286 {
 287         u32 state = this_cpu_read(kcsan_rand_state);
 288
 289         state = 1664525 * state + 1013904223;
 290         this_cpu_write(kcsan_rand_state, state);
 291
 292         return state % ep_ro;
 293 }
 294
 295 static inline void reset_kcsan_skip(void)
 296 {
 297         long skip_count = kcsan_skip_watch -
 298                           (IS_ENABLED(CONFIG_KCSAN_SKIP_WATCH_RANDOMIZE) ?
 299                                    kcsan_prandom_u32_max(kcsan_skip_watch) :
 300                                    0);
 301         this_cpu_write(kcsan_skip, skip_count);
 302 }
 303
 304 static __always_inline bool kcsan_is_enabled(struct kcsan_ctx *ctx)
 305 {
 306         return READ_ONCE(kcsan_enabled) && !ctx->disable_count;
 307 }
 308
 309 /* Introduce delay depending on context and configuration. */
 310 static void delay_access(int type)
 311 {
 312         unsigned int delay = in_task() ? kcsan_udelay_task : kcsan_udelay_interrupt;
 313         /* For certain access types, skew the random delay to be longer. */
 314         unsigned int skew_delay_order =
 315                 (type & (KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_ASSERT)) ? 1 : 0;
 316
 317         delay -= IS_ENABLED(CONFIG_KCSAN_DELAY_RANDOMIZE) ?
 318                                kcsan_prandom_u32_max(delay >> skew_delay_order) :
 319                                0;
 320         udelay(delay);
 321 }
 322
 323 void kcsan_save_irqtrace(struct task_struct *task)
 324 {
 325 #ifdef CONFIG_TRACE_IRQFLAGS
 326         task->kcsan_save_irqtrace = task->irqtrace;
 327 #endif
 328 }
 329
 330 void kcsan_restore_irqtrace(struct task_struct *task)
 331 {
 332 #ifdef CONFIG_TRACE_IRQFLAGS
 333         task->irqtrace = task->kcsan_save_irqtrace;
 334 #endif
 335 }
 336
 337 /*
 338  * Pull everything together: check_access() below contains the performance
 339  * critical operations; the fast-path (including check_access) functions should
 340  * all be inlinable by the instrumentation functions.
 341  *
 342  * The slow-path (kcsan_found_watchpoint, kcsan_setup_watchpoint) are
 343  * non-inlinable -- note that, we prefix these with "kcsan_" to ensure they can
 344  * be filtered from the stacktrace, as well as give them unique names for the
 345  * UACCESS whitelist of objtool. Each function uses user_access_save/restore(),
 346  * since they do not access any user memory, but instrumentation is still
 347  * emitted in UACCESS regions.
 348  */
 349
 350 static noinline void kcsan_found_watchpoint(const volatile void *ptr,
 351                                             size_t size,
 352                                             int type,
 353                                             atomic_long_t *watchpoint,
 354                                             long encoded_watchpoint)
 355 {
 356         const bool is_assert = (type & KCSAN_ACCESS_ASSERT) != 0;
 357         struct kcsan_ctx *ctx = get_ctx();
 358         unsigned long flags;
 359         bool consumed;
 360
 361         /*
 362          * We know a watchpoint exists. Let's try to keep the race-window
 363          * between here and finally consuming the watchpoint below as small as
 364          * possible -- avoid unneccessarily complex code until consumed.
 365          */
 366
 367         if (!kcsan_is_enabled(ctx))
 368                 return;
 369
 370         /*
 371          * The access_mask check relies on value-change comparison. To avoid
 372          * reporting a race where e.g. the writer set up the watchpoint, but the
 373          * reader has access_mask!=0, we have to ignore the found watchpoint.
 374          */
 375         if (ctx->access_mask)
 376                 return;
 377
 378         /*
 379          * If the other thread does not want to ignore the access, and there was
 380          * a value change as a result of this thread's operation, we will still
 381          * generate a report of unknown origin.
 382          *
 383          * Use CONFIG_KCSAN_REPORT_RACE_UNKNOWN_ORIGIN=n to filter.
 384          */
 385         if (!is_assert && kcsan_ignore_address(ptr))
 386                 return;
 387
 388         /*
 389          * Consuming the watchpoint must be guarded by kcsan_is_enabled() to
 390          * avoid erroneously triggering reports if the context is disabled.
 391          */
 392         consumed = try_consume_watchpoint(watchpoint, encoded_watchpoint);
 393
 394         /* keep this after try_consume_watchpoint */
 395         flags = user_access_save();
 396
 397         if (consumed) {
 398                 kcsan_save_irqtrace(current);
 399                 kcsan_report_set_info(ptr, size, type, watchpoint - watchpoints);
 400                 kcsan_restore_irqtrace(current);
 401         } else {
 402                 /*
 403                  * The other thread may not print any diagnostics, as it has
 404                  * already removed the watchpoint, or another thread consumed
 405                  * the watchpoint before this thread.
 406                  */
 407                 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_REPORT_RACES]);
 408         }
 409
 410         if (is_assert)
 411                 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_ASSERT_FAILURES]);
 412         else
 413                 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_DATA_RACES]);
 414
 415         user_access_restore(flags);
 416 }
 417
 418 static noinline void
 419 kcsan_setup_watchpoint(const volatile void *ptr, size_t size, int type)
 420 {
 421         const bool is_write = (type & KCSAN_ACCESS_WRITE) != 0;
 422         const bool is_assert = (type & KCSAN_ACCESS_ASSERT) != 0;
 423         atomic_long_t *watchpoint;
 424         u64 old, new, diff;
 425         unsigned long access_mask;
 426         enum kcsan_value_change value_change = KCSAN_VALUE_CHANGE_MAYBE;
 427         unsigned long ua_flags = user_access_save();
 428         struct kcsan_ctx *ctx = get_ctx();
 429         unsigned long irq_flags = 0;
 430
 431         /*
 432          * Always reset kcsan_skip counter in slow-path to avoid underflow; see
 433          * should_watch().
 434          */
 435         reset_kcsan_skip();
 436
 437         if (!kcsan_is_enabled(ctx))
 438                 goto out;
 439
 440         /*
 441          * Check to-ignore addresses after kcsan_is_enabled(), as we may access
 442          * memory that is not yet initialized during early boot.
 443          */
 444         if (!is_assert && kcsan_ignore_address(ptr))
 445                 goto out;
 446
 447         if (!check_encodable((unsigned long)ptr, size)) {
 448                 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_UNENCODABLE_ACCESSES]);
 449                 goto out;
 450         }
 451
 452         /*
 453          * Save and restore the IRQ state trace touched by KCSAN, since KCSAN's
 454          * runtime is entered for every memory access, and potentially useful
 455          * information is lost if dirtied by KCSAN.
 456          */
 457         kcsan_save_irqtrace(current);
 458         if (!kcsan_interrupt_watcher)
 459                 local_irq_save(irq_flags);
 460
 461         watchpoint = insert_watchpoint((unsigned long)ptr, size, is_write);
 462         if (watchpoint == NULL) {
 463                 /*
 464                  * Out of capacity: the size of 'watchpoints', and the frequency
 465                  * with which should_watch() returns true should be tweaked so
 466                  * that this case happens very rarely.
 467                  */
 468                 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_NO_CAPACITY]);
 469                 goto out_unlock;
 470         }
 471
 472         atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_SETUP_WATCHPOINTS]);
 473         atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_USED_WATCHPOINTS]);
 474
 475         /*
 476          * Read the current value, to later check and infer a race if the data
 477          * was modified via a non-instrumented access, e.g. from a device.
 478          */
 479         old = 0;
 480         switch (size) {
 481         case 1:
 482                 old = READ_ONCE(*(const u8 *)ptr);
 483                 break;
 484         case 2:
 485                 old = READ_ONCE(*(const u16 *)ptr);
 486                 break;
 487         case 4:
 488                 old = READ_ONCE(*(const u32 *)ptr);
 489                 break;
 490         case 8:
 491                 old = READ_ONCE(*(const u64 *)ptr);
 492                 break;
 493         default:
 494                 break; /* ignore; we do not diff the values */
 495         }
 496
 497         /*
 498          * Delay this thread, to increase probability of observing a racy
 499          * conflicting access.
 500          */
 501         delay_access(type);
 502
 503         /*
 504          * Re-read value, and check if it is as expected; if not, we infer a
 505          * racy access.
 506          */
 507         access_mask = ctx->access_mask;
 508         new = 0;
 509         switch (size) {
 510         case 1:
 511                 new = READ_ONCE(*(const u8 *)ptr);
 512                 break;
 513         case 2:
 514                 new = READ_ONCE(*(const u16 *)ptr);
 515                 break;
 516         case 4:
 517                 new = READ_ONCE(*(const u32 *)ptr);
 518                 break;
 519         case 8:
 520                 new = READ_ONCE(*(const u64 *)ptr);
 521                 break;
 522         default:
 523                 break; /* ignore; we do not diff the values */
 524         }
 525
 526         diff = old ^ new;
 527         if (access_mask)
 528                 diff &= access_mask;
 529
 530         /*
 531          * Check if we observed a value change.
 532          *
 533          * Also check if the data race should be ignored (the rules depend on
 534          * non-zero diff); if it is to be ignored, the below rules for
 535          * KCSAN_VALUE_CHANGE_MAYBE apply.
 536          */
 537         if (diff && !kcsan_ignore_data_race(size, type, old, new, diff))
 538                 value_change = KCSAN_VALUE_CHANGE_TRUE;
 539
 540         /* Check if this access raced with another. */
 541         if (!consume_watchpoint(watchpoint)) {
 542                 /*
 543                  * Depending on the access type, map a value_change of MAYBE to
 544                  * TRUE (always report) or FALSE (never report).
 545                  */
 546                 if (value_change == KCSAN_VALUE_CHANGE_MAYBE) {
 547                         if (access_mask != 0) {
 548                                 /*
 549                                  * For access with access_mask, we require a
 550                                  * value-change, as it is likely that races on
 551                                  * ~access_mask bits are expected.
 552                                  */
 553                                 value_change = KCSAN_VALUE_CHANGE_FALSE;
 554                         } else if (size > 8 || is_assert) {
 555                                 /* Always assume a value-change. */
 556                                 value_change = KCSAN_VALUE_CHANGE_TRUE;
 557                         }
 558                 }
 559
 560                 /*
 561                  * No need to increment 'data_races' counter, as the racing
 562                  * thread already did.
 563                  *
 564                  * Count 'assert_failures' for each failed ASSERT access,
 565                  * therefore both this thread and the racing thread may
 566                  * increment this counter.
 567                  */
 568                 if (is_assert && value_change == KCSAN_VALUE_CHANGE_TRUE)
 569                         atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_ASSERT_FAILURES]);
 570
 571                 kcsan_report_known_origin(ptr, size, type, value_change,
 572                                           watchpoint - watchpoints,
 573                                           old, new, access_mask);
 574         } else if (value_change == KCSAN_VALUE_CHANGE_TRUE) {
 575                 /* Inferring a race, since the value should not have changed. */
 576
 577                 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_RACES_UNKNOWN_ORIGIN]);
 578                 if (is_assert)
 579                         atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_ASSERT_FAILURES]);
 580
 581                 if (IS_ENABLED(CONFIG_KCSAN_REPORT_RACE_UNKNOWN_ORIGIN) || is_assert)
 582                         kcsan_report_unknown_origin(ptr, size, type, old, new, access_mask);
 583         }
 584
 585         /*
 586          * Remove watchpoint; must be after reporting, since the slot may be
 587          * reused after this point.
 588          */
 589         remove_watchpoint(watchpoint);
 590         atomic_long_dec(&kcsan_counters[KCSAN_COUNTER_USED_WATCHPOINTS]);
 591 out_unlock:
 592         if (!kcsan_interrupt_watcher)
 593                 local_irq_restore(irq_flags);
 594         kcsan_restore_irqtrace(current);
 595 out:
 596         user_access_restore(ua_flags);
 597 }
 598
 599 static __always_inline void check_access(const volatile void *ptr, size_t size,
 600                                          int type)
 601 {
 602         const bool is_write = (type & KCSAN_ACCESS_WRITE) != 0;
 603         atomic_long_t *watchpoint;
 604         long encoded_watchpoint;
 605
 606         /*
 607          * Do nothing for 0 sized check; this comparison will be optimized out
 608          * for constant sized instrumentation (__tsan_{read,write}N).
 609          */
 610         if (unlikely(size == 0))
 611                 return;
 612
 613         /*
 614          * Avoid user_access_save in fast-path: find_watchpoint is safe without
 615          * user_access_save, as the address that ptr points to is only used to
 616          * check if a watchpoint exists; ptr is never dereferenced.
 617          */
 618         watchpoint = find_watchpoint((unsigned long)ptr, size, !is_write,
 619                                      &encoded_watchpoint);
 620         /*
 621          * It is safe to check kcsan_is_enabled() after find_watchpoint in the
 622          * slow-path, as long as no state changes that cause a race to be
 623          * detected and reported have occurred until kcsan_is_enabled() is
 624          * checked.
 625          */
 626
 627         if (unlikely(watchpoint != NULL))
 628                 kcsan_found_watchpoint(ptr, size, type, watchpoint,
 629                                        encoded_watchpoint);
 630         else {
 631                 struct kcsan_ctx *ctx = get_ctx(); /* Call only once in fast-path. */
 632
 633                 if (unlikely(should_watch(ptr, size, type, ctx)))
 634                         kcsan_setup_watchpoint(ptr, size, type);
 635                 else if (unlikely(ctx->scoped_accesses.prev))
 636                         kcsan_check_scoped_accesses();
 637         }
 638 }
 639
 640 /* === Public interface ===================================================== */
 641
 642 void __init kcsan_init(void)
 643 {
 644         int cpu;
 645
 646         BUG_ON(!in_task());
 647
 648         for_each_possible_cpu(cpu)
 649                 per_cpu(kcsan_rand_state, cpu) = (u32)get_cycles();
 650
 651         /*
 652          * We are in the init task, and no other tasks should be running;
 653          * WRITE_ONCE without memory barrier is sufficient.
 654          */
 655         if (kcsan_early_enable) {
 656                 pr_info("enabled early\n");
 657                 WRITE_ONCE(kcsan_enabled, true);
 658         }
 659
 660         if (IS_ENABLED(CONFIG_KCSAN_REPORT_VALUE_CHANGE_ONLY) ||
 661             IS_ENABLED(CONFIG_KCSAN_ASSUME_PLAIN_WRITES_ATOMIC) ||
 662             IS_ENABLED(CONFIG_KCSAN_PERMISSIVE) ||
 663             IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {
 664                 pr_warn("non-strict mode configured - use CONFIG_KCSAN_STRICT=y to see all data races\n");
 665         } else {
 666                 pr_info("strict mode configured\n");
 667         }
 668 }
 669
 670 /* === Exported interface =================================================== */
 671
 672 void kcsan_disable_current(void)
 673 {
 674         ++get_ctx()->disable_count;
 675 }
 676 EXPORT_SYMBOL(kcsan_disable_current);
 677
 678 void kcsan_enable_current(void)
 679 {
 680         if (get_ctx()->disable_count-- == 0) {
 681                 /*
 682                  * Warn if kcsan_enable_current() calls are unbalanced with
 683                  * kcsan_disable_current() calls, which causes disable_count to
 684                  * become negative and should not happen.
 685                  */
 686                 kcsan_disable_current(); /* restore to 0, KCSAN still enabled */
 687                 kcsan_disable_current(); /* disable to generate warning */
 688                 WARN(1, "Unbalanced %s()", __func__);
 689                 kcsan_enable_current();
 690         }
 691 }
 692 EXPORT_SYMBOL(kcsan_enable_current);
 693
 694 void kcsan_enable_current_nowarn(void)
 695 {
 696         if (get_ctx()->disable_count-- == 0)
 697                 kcsan_disable_current();
 698 }
 699 EXPORT_SYMBOL(kcsan_enable_current_nowarn);
 700
 701 void kcsan_nestable_atomic_begin(void)
 702 {
 703         /*
 704          * Do *not* check and warn if we are in a flat atomic region: nestable
 705          * and flat atomic regions are independent from each other.
 706          * See include/linux/kcsan.h: struct kcsan_ctx comments for more
 707          * comments.
 708          */
 709
 710         ++get_ctx()->atomic_nest_count;
 711 }
 712 EXPORT_SYMBOL(kcsan_nestable_atomic_begin);
 713
 714 void kcsan_nestable_atomic_end(void)
 715 {
 716         if (get_ctx()->atomic_nest_count-- == 0) {
 717                 /*
 718                  * Warn if kcsan_nestable_atomic_end() calls are unbalanced with
 719                  * kcsan_nestable_atomic_begin() calls, which causes
 720                  * atomic_nest_count to become negative and should not happen.
 721                  */
 722                 kcsan_nestable_atomic_begin(); /* restore to 0 */
 723                 kcsan_disable_current(); /* disable to generate warning */
 724                 WARN(1, "Unbalanced %s()", __func__);
 725                 kcsan_enable_current();
 726         }
 727 }
 728 EXPORT_SYMBOL(kcsan_nestable_atomic_end);
 729
 730 void kcsan_flat_atomic_begin(void)
 731 {
 732         get_ctx()->in_flat_atomic = true;
 733 }
 734 EXPORT_SYMBOL(kcsan_flat_atomic_begin);
 735
 736 void kcsan_flat_atomic_end(void)
 737 {
 738         get_ctx()->in_flat_atomic = false;
 739 }
 740 EXPORT_SYMBOL(kcsan_flat_atomic_end);
 741
 742 void kcsan_atomic_next(int n)
 743 {
 744         get_ctx()->atomic_next = n;
 745 }
 746 EXPORT_SYMBOL(kcsan_atomic_next);
 747
 748 void kcsan_set_access_mask(unsigned long mask)
 749 {
 750         get_ctx()->access_mask = mask;
 751 }
 752 EXPORT_SYMBOL(kcsan_set_access_mask);
 753
 754 struct kcsan_scoped_access *
 755 kcsan_begin_scoped_access(const volatile void *ptr, size_t size, int type,
 756                           struct kcsan_scoped_access *sa)
 757 {
 758         struct kcsan_ctx *ctx = get_ctx();
 759
 760         __kcsan_check_access(ptr, size, type);
 761
 762         ctx->disable_count++; /* Disable KCSAN, in case list debugging is on. */
 763
 764         INIT_LIST_HEAD(&sa->list);
 765         sa->ptr = ptr;
 766         sa->size = size;
 767         sa->type = type;
 768
 769         if (!ctx->scoped_accesses.prev) /* Lazy initialize list head. */
 770                 INIT_LIST_HEAD(&ctx->scoped_accesses);
 771         list_add(&sa->list, &ctx->scoped_accesses);
 772
 773         ctx->disable_count--;
 774         return sa;
 775 }
 776 EXPORT_SYMBOL(kcsan_begin_scoped_access);
 777
 778 void kcsan_end_scoped_access(struct kcsan_scoped_access *sa)
 779 {
 780         struct kcsan_ctx *ctx = get_ctx();
 781
 782         if (WARN(!ctx->scoped_accesses.prev, "Unbalanced %s()?", __func__))
 783                 return;
 784
 785         ctx->disable_count++; /* Disable KCSAN, in case list debugging is on. */
 786
 787         list_del(&sa->list);
 788         if (list_empty(&ctx->scoped_accesses))
 789                 /*
 790                  * Ensure we do not enter kcsan_check_scoped_accesses()
 791                  * slow-path if unnecessary, and avoids requiring list_empty()
 792                  * in the fast-path (to avoid a READ_ONCE() and potential
 793                  * uaccess warning).
 794                  */
 795                 ctx->scoped_accesses.prev = NULL;
 796
 797         ctx->disable_count--;
 798
 799         __kcsan_check_access(sa->ptr, sa->size, sa->type);
 800 }
 801 EXPORT_SYMBOL(kcsan_end_scoped_access);
 802
 803 void __kcsan_check_access(const volatile void *ptr, size_t size, int type)
 804 {
 805         check_access(ptr, size, type);
 806 }
 807 EXPORT_SYMBOL(__kcsan_check_access);
 808
 809 /*
 810  * KCSAN uses the same instrumentation that is emitted by supported compilers
 811  * for ThreadSanitizer (TSAN).
 812  *
 813  * When enabled, the compiler emits instrumentation calls (the functions
 814  * prefixed with "__tsan" below) for all loads and stores that it generated;
 815  * inline asm is not instrumented.
 816  *
 817  * Note that, not all supported compiler versions distinguish aligned/unaligned
 818  * accesses, but e.g. recent versions of Clang do. We simply alias the unaligned
 819  * version to the generic version, which can handle both.
 820  */
 821
 822 #define DEFINE_TSAN_READ_WRITE(size)                                           \
 823         void __tsan_read##size(void *ptr);                                     \
 824         void __tsan_read##size(void *ptr)                                      \
 825         {                                                                      \
 826                 check_access(ptr, size, 0);                                    \
 827         }                                                                      \
 828         EXPORT_SYMBOL(__tsan_read##size);                                      \
 829         void __tsan_unaligned_read##size(void *ptr)                            \
 830                 __alias(__tsan_read##size);                                    \
 831         EXPORT_SYMBOL(__tsan_unaligned_read##size);                            \
 832         void __tsan_write##size(void *ptr);                                    \
 833         void __tsan_write##size(void *ptr)                                     \
 834         {                                                                      \
 835                 check_access(ptr, size, KCSAN_ACCESS_WRITE);                   \
 836         }                                                                      \
 837         EXPORT_SYMBOL(__tsan_write##size);                                     \
 838         void __tsan_unaligned_write##size(void *ptr)                           \
 839                 __alias(__tsan_write##size);                                   \
 840         EXPORT_SYMBOL(__tsan_unaligned_write##size);                           \
 841         void __tsan_read_write##size(void *ptr);                               \
 842         void __tsan_read_write##size(void *ptr)                                \
 843         {                                                                      \
 844                 check_access(ptr, size,                                        \
 845                              KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE);      \
 846         }                                                                      \
 847         EXPORT_SYMBOL(__tsan_read_write##size);                                \
 848         void __tsan_unaligned_read_write##size(void *ptr)                      \
 849                 __alias(__tsan_read_write##size);                              \
 850         EXPORT_SYMBOL(__tsan_unaligned_read_write##size)
 851
 852 DEFINE_TSAN_READ_WRITE(1);
 853 DEFINE_TSAN_READ_WRITE(2);
 854 DEFINE_TSAN_READ_WRITE(4);
 855 DEFINE_TSAN_READ_WRITE(8);
 856 DEFINE_TSAN_READ_WRITE(16);
 857
 858 void __tsan_read_range(void *ptr, size_t size);
 859 void __tsan_read_range(void *ptr, size_t size)
 860 {
 861         check_access(ptr, size, 0);
 862 }
 863 EXPORT_SYMBOL(__tsan_read_range);
 864
 865 void __tsan_write_range(void *ptr, size_t size);
 866 void __tsan_write_range(void *ptr, size_t size)
 867 {
 868         check_access(ptr, size, KCSAN_ACCESS_WRITE);
 869 }
 870 EXPORT_SYMBOL(__tsan_write_range);
 871
 872 /*
 873  * Use of explicit volatile is generally disallowed [1], however, volatile is
 874  * still used in various concurrent context, whether in low-level
 875  * synchronization primitives or for legacy reasons.
 876  * [1] https://lwn.net/Articles/233479/
 877  *
 878  * We only consider volatile accesses atomic if they are aligned and would pass
 879  * the size-check of compiletime_assert_rwonce_type().
 880  */
 881 #define DEFINE_TSAN_VOLATILE_READ_WRITE(size)                                  \
 882         void __tsan_volatile_read##size(void *ptr);                            \
 883         void __tsan_volatile_read##size(void *ptr)                             \
 884         {                                                                      \
 885                 const bool is_atomic = size <= sizeof(long long) &&            \
 886                                        IS_ALIGNED((unsigned long)ptr, size);   \
 887                 if (IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS) && is_atomic)      \
 888                         return;                                                \
 889                 check_access(ptr, size, is_atomic ? KCSAN_ACCESS_ATOMIC : 0);  \
 890         }                                                                      \
 891         EXPORT_SYMBOL(__tsan_volatile_read##size);                             \
 892         void __tsan_unaligned_volatile_read##size(void *ptr)                   \
 893                 __alias(__tsan_volatile_read##size);                           \
 894         EXPORT_SYMBOL(__tsan_unaligned_volatile_read##size);                   \
 895         void __tsan_volatile_write##size(void *ptr);                           \
 896         void __tsan_volatile_write##size(void *ptr)                            \
 897         {                                                                      \
 898                 const bool is_atomic = size <= sizeof(long long) &&            \
 899                                        IS_ALIGNED((unsigned long)ptr, size);   \
 900                 if (IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS) && is_atomic)      \
 901                         return;                                                \
 902                 check_access(ptr, size,                                        \
 903                              KCSAN_ACCESS_WRITE |                              \
 904                                      (is_atomic ? KCSAN_ACCESS_ATOMIC : 0));   \
 905         }                                                                      \
 906         EXPORT_SYMBOL(__tsan_volatile_write##size);                            \
 907         void __tsan_unaligned_volatile_write##size(void *ptr)                  \
 908                 __alias(__tsan_volatile_write##size);                          \
 909         EXPORT_SYMBOL(__tsan_unaligned_volatile_write##size)
 910
 911 DEFINE_TSAN_VOLATILE_READ_WRITE(1);
 912 DEFINE_TSAN_VOLATILE_READ_WRITE(2);
 913 DEFINE_TSAN_VOLATILE_READ_WRITE(4);
 914 DEFINE_TSAN_VOLATILE_READ_WRITE(8);
 915 DEFINE_TSAN_VOLATILE_READ_WRITE(16);
 916
 917 /*
 918  * The below are not required by KCSAN, but can still be emitted by the
 919  * compiler.
 920  */
 921 void __tsan_func_entry(void *call_pc);
 922 void __tsan_func_entry(void *call_pc)
 923 {
 924 }
 925 EXPORT_SYMBOL(__tsan_func_entry);
 926 void __tsan_func_exit(void);
 927 void __tsan_func_exit(void)
 928 {
 929 }
 930 EXPORT_SYMBOL(__tsan_func_exit);
 931 void __tsan_init(void);
 932 void __tsan_init(void)
 933 {
 934 }
 935 EXPORT_SYMBOL(__tsan_init);
 936
 937 /*
 938  * Instrumentation for atomic builtins (__atomic_*, __sync_*).
 939  *
 940  * Normal kernel code _should not_ be using them directly, but some
 941  * architectures may implement some or all atomics using the compilers'
 942  * builtins.
 943  *
 944  * Note: If an architecture decides to fully implement atomics using the
 945  * builtins, because they are implicitly instrumented by KCSAN (and KASAN,
 946  * etc.), implementing the ARCH_ATOMIC interface (to get instrumentation via
 947  * atomic-instrumented) is no longer necessary.
 948  *
 949  * TSAN instrumentation replaces atomic accesses with calls to any of the below
 950  * functions, whose job is to also execute the operation itself.
 951  */
 952
 953 #define DEFINE_TSAN_ATOMIC_LOAD_STORE(bits)                                                        \
 954         u##bits __tsan_atomic##bits##_load(const u##bits *ptr, int memorder);                      \
 955         u##bits __tsan_atomic##bits##_load(const u##bits *ptr, int memorder)                       \
 956         {                                                                                          \
 957                 if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
 958                         check_access(ptr, bits / BITS_PER_BYTE, KCSAN_ACCESS_ATOMIC);              \
 959                 }                                                                                  \
 960                 return __atomic_load_n(ptr, memorder);                                             \
 961         }                                                                                          \
 962         EXPORT_SYMBOL(__tsan_atomic##bits##_load);                                                 \
 963         void __tsan_atomic##bits##_store(u##bits *ptr, u##bits v, int memorder);                   \
 964         void __tsan_atomic##bits##_store(u##bits *ptr, u##bits v, int memorder)                    \
 965         {                                                                                          \
 966                 if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
 967                         check_access(ptr, bits / BITS_PER_BYTE,                                    \
 968                                      KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ATOMIC);                    \
 969                 }                                                                                  \
 970                 __atomic_store_n(ptr, v, memorder);                                                \
 971         }                                                                                          \
 972         EXPORT_SYMBOL(__tsan_atomic##bits##_store)
 973
 974 #define DEFINE_TSAN_ATOMIC_RMW(op, bits, suffix)                                                   \
 975         u##bits __tsan_atomic##bits##_##op(u##bits *ptr, u##bits v, int memorder);                 \
 976         u##bits __tsan_atomic##bits##_##op(u##bits *ptr, u##bits v, int memorder)                  \
 977         {                                                                                          \
 978                 if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
 979                         check_access(ptr, bits / BITS_PER_BYTE,                                    \
 980                                      KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE |                  \
 981                                              KCSAN_ACCESS_ATOMIC);                                 \
 982                 }                                                                                  \
 983                 return __atomic_##op##suffix(ptr, v, memorder);                                    \
 984         }                                                                                          \
 985         EXPORT_SYMBOL(__tsan_atomic##bits##_##op)
 986
 987 /*
 988  * Note: CAS operations are always classified as write, even in case they
 989  * fail. We cannot perform check_access() after a write, as it might lead to
 990  * false positives, in cases such as:
 991  *
 992  *      T0: __atomic_compare_exchange_n(&p->flag, &old, 1, ...)
 993  *
 994  *      T1: if (__atomic_load_n(&p->flag, ...)) {
 995  *              modify *p;
 996  *              p->flag = 0;
 997  *          }
 998  *
 999  * The only downside is that, if there are 3 threads, with one CAS that
1000  * succeeds, another CAS that fails, and an unmarked racing operation, we may
1001  * point at the wrong CAS as the source of the race. However, if we assume that
1002  * all CAS can succeed in some other execution, the data race is still valid.
1003  */
1004 #define DEFINE_TSAN_ATOMIC_CMPXCHG(bits, strength, weak)                                           \
1005         int __tsan_atomic##bits##_compare_exchange_##strength(u##bits *ptr, u##bits *exp,          \
1006                                                               u##bits val, int mo, int fail_mo);   \
1007         int __tsan_atomic##bits##_compare_exchange_##strength(u##bits *ptr, u##bits *exp,          \
1008                                                               u##bits val, int mo, int fail_mo)    \
1009         {                                                                                          \
1010                 if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
1011                         check_access(ptr, bits / BITS_PER_BYTE,                                    \
1012                                      KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE |                  \
1013                                              KCSAN_ACCESS_ATOMIC);                                 \
1014                 }                                                                                  \
1015                 return __atomic_compare_exchange_n(ptr, exp, val, weak, mo, fail_mo);              \
1016         }                                                                                          \
1017         EXPORT_SYMBOL(__tsan_atomic##bits##_compare_exchange_##strength)
1018
1019 #define DEFINE_TSAN_ATOMIC_CMPXCHG_VAL(bits)                                                       \
1020         u##bits __tsan_atomic##bits##_compare_exchange_val(u##bits *ptr, u##bits exp, u##bits val, \
1021                                                            int mo, int fail_mo);                   \
1022         u##bits __tsan_atomic##bits##_compare_exchange_val(u##bits *ptr, u##bits exp, u##bits val, \
1023                                                            int mo, int fail_mo)                    \
1024         {                                                                                          \
1025                 if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
1026                         check_access(ptr, bits / BITS_PER_BYTE,                                    \
1027                                      KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE |                  \
1028                                              KCSAN_ACCESS_ATOMIC);                                 \
1029                 }                                                                                  \
1030                 __atomic_compare_exchange_n(ptr, &exp, val, 0, mo, fail_mo);                       \
1031                 return exp;                                                                        \
1032         }                                                                                          \
1033         EXPORT_SYMBOL(__tsan_atomic##bits##_compare_exchange_val)
1034
1035 #define DEFINE_TSAN_ATOMIC_OPS(bits)                                                               \
1036         DEFINE_TSAN_ATOMIC_LOAD_STORE(bits);                                                       \
1037         DEFINE_TSAN_ATOMIC_RMW(exchange, bits, _n);                                                \
1038         DEFINE_TSAN_ATOMIC_RMW(fetch_add, bits, );                                                 \
1039         DEFINE_TSAN_ATOMIC_RMW(fetch_sub, bits, );                                                 \
1040         DEFINE_TSAN_ATOMIC_RMW(fetch_and, bits, );                                                 \
1041         DEFINE_TSAN_ATOMIC_RMW(fetch_or, bits, );                                                  \
1042         DEFINE_TSAN_ATOMIC_RMW(fetch_xor, bits, );                                                 \
1043         DEFINE_TSAN_ATOMIC_RMW(fetch_nand, bits, );                                                \
1044         DEFINE_TSAN_ATOMIC_CMPXCHG(bits, strong, 0);                                               \
1045         DEFINE_TSAN_ATOMIC_CMPXCHG(bits, weak, 1);                                                 \
1046         DEFINE_TSAN_ATOMIC_CMPXCHG_VAL(bits)
1047
1048 DEFINE_TSAN_ATOMIC_OPS(8);
1049 DEFINE_TSAN_ATOMIC_OPS(16);
1050 DEFINE_TSAN_ATOMIC_OPS(32);
1051 DEFINE_TSAN_ATOMIC_OPS(64);
1052
1053 void __tsan_atomic_thread_fence(int memorder);
1054 void __tsan_atomic_thread_fence(int memorder)
1055 {
1056         __atomic_thread_fence(memorder);
1057 }
1058 EXPORT_SYMBOL(__tsan_atomic_thread_fence);
1059
1060 void __tsan_atomic_signal_fence(int memorder);
1061 void __tsan_atomic_signal_fence(int memorder) { }
1062 EXPORT_SYMBOL(__tsan_atomic_signal_fence);