kernel/kcsan/core.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #define pr_fmt(fmt) "kcsan: " fmt
   4
   5 #include <linux/atomic.h>
   6 #include <linux/bug.h>
   7 #include <linux/delay.h>
   8 #include <linux/export.h>
   9 #include <linux/init.h>
  10 #include <linux/kernel.h>
  11 #include <linux/list.h>
  12 #include <linux/moduleparam.h>
  13 #include <linux/percpu.h>
  14 #include <linux/preempt.h>
  15 #include <linux/random.h>
  16 #include <linux/sched.h>
  17 #include <linux/uaccess.h>
  18
  19 #include "atomic.h"
  20 #include "encoding.h"
  21 #include "kcsan.h"
  22
  23 static bool kcsan_early_enable = IS_ENABLED(CONFIG_KCSAN_EARLY_ENABLE);
  24 unsigned int kcsan_udelay_task = CONFIG_KCSAN_UDELAY_TASK;
  25 unsigned int kcsan_udelay_interrupt = CONFIG_KCSAN_UDELAY_INTERRUPT;
  26 static long kcsan_skip_watch = CONFIG_KCSAN_SKIP_WATCH;
  27 static bool kcsan_interrupt_watcher = IS_ENABLED(CONFIG_KCSAN_INTERRUPT_WATCHER);
  28
  29 #ifdef MODULE_PARAM_PREFIX
  30 #undef MODULE_PARAM_PREFIX
  31 #endif
  32 #define MODULE_PARAM_PREFIX "kcsan."
  33 module_param_named(early_enable, kcsan_early_enable, bool, 0);
  34 module_param_named(udelay_task, kcsan_udelay_task, uint, 0644);
  35 module_param_named(udelay_interrupt, kcsan_udelay_interrupt, uint, 0644);
  36 module_param_named(skip_watch, kcsan_skip_watch, long, 0644);
  37 module_param_named(interrupt_watcher, kcsan_interrupt_watcher, bool, 0444);
  38
  39 bool kcsan_enabled;
  40
  41 /* Per-CPU kcsan_ctx for interrupts */
  42 static DEFINE_PER_CPU(struct kcsan_ctx, kcsan_cpu_ctx) = {
  43         .disable_count          = 0,
  44         .atomic_next            = 0,
  45         .atomic_nest_count      = 0,
  46         .in_flat_atomic         = false,
  47         .access_mask            = 0,
  48         .scoped_accesses        = {LIST_POISON1, NULL},
  49 };
  50
  51 /*
  52  * Helper macros to index into adjacent slots, starting from address slot
  53  * itself, followed by the right and left slots.
  54  *
  55  * The purpose is 2-fold:
  56  *
  57  *      1. if during insertion the address slot is already occupied, check if
  58  *         any adjacent slots are free;
  59  *      2. accesses that straddle a slot boundary due to size that exceeds a
  60  *         slot's range may check adjacent slots if any watchpoint matches.
  61  *
  62  * Note that accesses with very large size may still miss a watchpoint; however,
  63  * given this should be rare, this is a reasonable trade-off to make, since this
  64  * will avoid:
  65  *
  66  *      1. excessive contention between watchpoint checks and setup;
  67  *      2. larger number of simultaneous watchpoints without sacrificing
  68  *         performance.
  69  *
  70  * Example: SLOT_IDX values for KCSAN_CHECK_ADJACENT=1, where i is [0, 1, 2]:
  71  *
  72  *   slot=0:  [ 1,  2,  0]
  73  *   slot=9:  [10, 11,  9]
  74  *   slot=63: [64, 65, 63]
  75  */
  76 #define SLOT_IDX(slot, i) (slot + ((i + KCSAN_CHECK_ADJACENT) % NUM_SLOTS))
  77
  78 /*
  79  * SLOT_IDX_FAST is used in the fast-path. Not first checking the address's primary
  80  * slot (middle) is fine if we assume that races occur rarely. The set of
  81  * indices {SLOT_IDX(slot, i) | i in [0, NUM_SLOTS)} is equivalent to
  82  * {SLOT_IDX_FAST(slot, i) | i in [0, NUM_SLOTS)}.
  83  */
  84 #define SLOT_IDX_FAST(slot, i) (slot + i)
  85
  86 /*
  87  * Watchpoints, with each entry encoded as defined in encoding.h: in order to be
  88  * able to safely update and access a watchpoint without introducing locking
  89  * overhead, we encode each watchpoint as a single atomic long. The initial
  90  * zero-initialized state matches INVALID_WATCHPOINT.
  91  *
  92  * Add NUM_SLOTS-1 entries to account for overflow; this helps avoid having to
  93  * use more complicated SLOT_IDX_FAST calculation with modulo in the fast-path.
  94  */
  95 static atomic_long_t watchpoints[CONFIG_KCSAN_NUM_WATCHPOINTS + NUM_SLOTS-1];
  96
  97 /*
  98  * Instructions to skip watching counter, used in should_watch(). We use a
  99  * per-CPU counter to avoid excessive contention.
 100  */
 101 static DEFINE_PER_CPU(long, kcsan_skip);
 102
 103 static __always_inline atomic_long_t *find_watchpoint(unsigned long addr,
 104                                                       size_t size,
 105                                                       bool expect_write,
 106                                                       long *encoded_watchpoint)
 107 {
 108         const int slot = watchpoint_slot(addr);
 109         const unsigned long addr_masked = addr & WATCHPOINT_ADDR_MASK;
 110         atomic_long_t *watchpoint;
 111         unsigned long wp_addr_masked;
 112         size_t wp_size;
 113         bool is_write;
 114         int i;
 115
 116         BUILD_BUG_ON(CONFIG_KCSAN_NUM_WATCHPOINTS < NUM_SLOTS);
 117
 118         for (i = 0; i < NUM_SLOTS; ++i) {
 119                 watchpoint = &watchpoints[SLOT_IDX_FAST(slot, i)];
 120                 *encoded_watchpoint = atomic_long_read(watchpoint);
 121                 if (!decode_watchpoint(*encoded_watchpoint, &wp_addr_masked,
 122                                        &wp_size, &is_write))
 123                         continue;
 124
 125                 if (expect_write && !is_write)
 126                         continue;
 127
 128                 /* Check if the watchpoint matches the access. */
 129                 if (matching_access(wp_addr_masked, wp_size, addr_masked, size))
 130                         return watchpoint;
 131         }
 132
 133         return NULL;
 134 }
 135
 136 static inline atomic_long_t *
 137 insert_watchpoint(unsigned long addr, size_t size, bool is_write)
 138 {
 139         const int slot = watchpoint_slot(addr);
 140         const long encoded_watchpoint = encode_watchpoint(addr, size, is_write);
 141         atomic_long_t *watchpoint;
 142         int i;
 143
 144         /* Check slot index logic, ensuring we stay within array bounds. */
 145         BUILD_BUG_ON(SLOT_IDX(0, 0) != KCSAN_CHECK_ADJACENT);
 146         BUILD_BUG_ON(SLOT_IDX(0, KCSAN_CHECK_ADJACENT+1) != 0);
 147         BUILD_BUG_ON(SLOT_IDX(CONFIG_KCSAN_NUM_WATCHPOINTS-1, KCSAN_CHECK_ADJACENT) != ARRAY_SIZE(watchpoints)-1);
 148         BUILD_BUG_ON(SLOT_IDX(CONFIG_KCSAN_NUM_WATCHPOINTS-1, KCSAN_CHECK_ADJACENT+1) != ARRAY_SIZE(watchpoints) - NUM_SLOTS);
 149
 150         for (i = 0; i < NUM_SLOTS; ++i) {
 151                 long expect_val = INVALID_WATCHPOINT;
 152
 153                 /* Try to acquire this slot. */
 154                 watchpoint = &watchpoints[SLOT_IDX(slot, i)];
 155                 if (atomic_long_try_cmpxchg_relaxed(watchpoint, &expect_val, encoded_watchpoint))
 156                         return watchpoint;
 157         }
 158
 159         return NULL;
 160 }
 161
 162 /*
 163  * Return true if watchpoint was successfully consumed, false otherwise.
 164  *
 165  * This may return false if:
 166  *
 167  *      1. another thread already consumed the watchpoint;
 168  *      2. the thread that set up the watchpoint already removed it;
 169  *      3. the watchpoint was removed and then re-used.
 170  */
 171 static __always_inline bool
 172 try_consume_watchpoint(atomic_long_t *watchpoint, long encoded_watchpoint)
 173 {
 174         return atomic_long_try_cmpxchg_relaxed(watchpoint, &encoded_watchpoint, CONSUMED_WATCHPOINT);
 175 }
 176
 177 /* Return true if watchpoint was not touched, false if already consumed. */
 178 static inline bool consume_watchpoint(atomic_long_t *watchpoint)
 179 {
 180         return atomic_long_xchg_relaxed(watchpoint, CONSUMED_WATCHPOINT) != CONSUMED_WATCHPOINT;
 181 }
 182
 183 /* Remove the watchpoint -- its slot may be reused after. */
 184 static inline void remove_watchpoint(atomic_long_t *watchpoint)
 185 {
 186         atomic_long_set(watchpoint, INVALID_WATCHPOINT);
 187 }
 188
 189 static __always_inline struct kcsan_ctx *get_ctx(void)
 190 {
 191         /*
 192          * In interrupts, use raw_cpu_ptr to avoid unnecessary checks, that would
 193          * also result in calls that generate warnings in uaccess regions.
 194          */
 195         return in_task() ? &current->kcsan_ctx : raw_cpu_ptr(&kcsan_cpu_ctx);
 196 }
 197
 198 /* Check scoped accesses; never inline because this is a slow-path! */
 199 static noinline void kcsan_check_scoped_accesses(void)
 200 {
 201         struct kcsan_ctx *ctx = get_ctx();
 202         struct list_head *prev_save = ctx->scoped_accesses.prev;
 203         struct kcsan_scoped_access *scoped_access;
 204
 205         ctx->scoped_accesses.prev = NULL;  /* Avoid recursion. */
 206         list_for_each_entry(scoped_access, &ctx->scoped_accesses, list)
 207                 __kcsan_check_access(scoped_access->ptr, scoped_access->size, scoped_access->type);
 208         ctx->scoped_accesses.prev = prev_save;
 209 }
 210
 211 /* Rules for generic atomic accesses. Called from fast-path. */
 212 static __always_inline bool
 213 is_atomic(const volatile void *ptr, size_t size, int type, struct kcsan_ctx *ctx)
 214 {
 215         if (type & KCSAN_ACCESS_ATOMIC)
 216                 return true;
 217
 218         /*
 219          * Unless explicitly declared atomic, never consider an assertion access
 220          * as atomic. This allows using them also in atomic regions, such as
 221          * seqlocks, without implicitly changing their semantics.
 222          */
 223         if (type & KCSAN_ACCESS_ASSERT)
 224                 return false;
 225
 226         if (IS_ENABLED(CONFIG_KCSAN_ASSUME_PLAIN_WRITES_ATOMIC) &&
 227             (type & KCSAN_ACCESS_WRITE) && size <= sizeof(long) &&
 228             !(type & KCSAN_ACCESS_COMPOUND) && IS_ALIGNED((unsigned long)ptr, size))
 229                 return true; /* Assume aligned writes up to word size are atomic. */
 230
 231         if (ctx->atomic_next > 0) {
 232                 /*
 233                  * Because we do not have separate contexts for nested
 234                  * interrupts, in case atomic_next is set, we simply assume that
 235                  * the outer interrupt set atomic_next. In the worst case, we
 236                  * will conservatively consider operations as atomic. This is a
 237                  * reasonable trade-off to make, since this case should be
 238                  * extremely rare; however, even if extremely rare, it could
 239                  * lead to false positives otherwise.
 240                  */
 241                 if ((hardirq_count() >> HARDIRQ_SHIFT) < 2)
 242                         --ctx->atomic_next; /* in task, or outer interrupt */
 243                 return true;
 244         }
 245
 246         return ctx->atomic_nest_count > 0 || ctx->in_flat_atomic;
 247 }
 248
 249 static __always_inline bool
 250 should_watch(const volatile void *ptr, size_t size, int type, struct kcsan_ctx *ctx)
 251 {
 252         /*
 253          * Never set up watchpoints when memory operations are atomic.
 254          *
 255          * Need to check this first, before kcsan_skip check below: (1) atomics
 256          * should not count towards skipped instructions, and (2) to actually
 257          * decrement kcsan_atomic_next for consecutive instruction stream.
 258          */
 259         if (is_atomic(ptr, size, type, ctx))
 260                 return false;
 261
 262         if (this_cpu_dec_return(kcsan_skip) >= 0)
 263                 return false;
 264
 265         /*
 266          * NOTE: If we get here, kcsan_skip must always be reset in slow path
 267          * via reset_kcsan_skip() to avoid underflow.
 268          */
 269
 270         /* this operation should be watched */
 271         return true;
 272 }
 273
 274 static inline void reset_kcsan_skip(void)
 275 {
 276         long skip_count = kcsan_skip_watch -
 277                           (IS_ENABLED(CONFIG_KCSAN_SKIP_WATCH_RANDOMIZE) ?
 278                                    prandom_u32_max(kcsan_skip_watch) :
 279                                    0);
 280         this_cpu_write(kcsan_skip, skip_count);
 281 }
 282
 283 static __always_inline bool kcsan_is_enabled(void)
 284 {
 285         return READ_ONCE(kcsan_enabled) && get_ctx()->disable_count == 0;
 286 }
 287
 288 static inline unsigned int get_delay(int type)
 289 {
 290         unsigned int delay = in_task() ? kcsan_udelay_task : kcsan_udelay_interrupt;
 291         /* For certain access types, skew the random delay to be longer. */
 292         unsigned int skew_delay_order =
 293                 (type & (KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_ASSERT)) ? 1 : 0;
 294
 295         return delay - (IS_ENABLED(CONFIG_KCSAN_DELAY_RANDOMIZE) ?
 296                                 prandom_u32_max(delay >> skew_delay_order) :
 297                                 0);
 298 }
 299
 300 void kcsan_save_irqtrace(struct task_struct *task)
 301 {
 302 #ifdef CONFIG_TRACE_IRQFLAGS
 303         task->kcsan_save_irqtrace = task->irqtrace;
 304 #endif
 305 }
 306
 307 void kcsan_restore_irqtrace(struct task_struct *task)
 308 {
 309 #ifdef CONFIG_TRACE_IRQFLAGS
 310         task->irqtrace = task->kcsan_save_irqtrace;
 311 #endif
 312 }
 313
 314 /*
 315  * Pull everything together: check_access() below contains the performance
 316  * critical operations; the fast-path (including check_access) functions should
 317  * all be inlinable by the instrumentation functions.
 318  *
 319  * The slow-path (kcsan_found_watchpoint, kcsan_setup_watchpoint) are
 320  * non-inlinable -- note that, we prefix these with "kcsan_" to ensure they can
 321  * be filtered from the stacktrace, as well as give them unique names for the
 322  * UACCESS whitelist of objtool. Each function uses user_access_save/restore(),
 323  * since they do not access any user memory, but instrumentation is still
 324  * emitted in UACCESS regions.
 325  */
 326
 327 static noinline void kcsan_found_watchpoint(const volatile void *ptr,
 328                                             size_t size,
 329                                             int type,
 330                                             atomic_long_t *watchpoint,
 331                                             long encoded_watchpoint)
 332 {
 333         unsigned long flags;
 334         bool consumed;
 335
 336         if (!kcsan_is_enabled())
 337                 return;
 338
 339         /*
 340          * The access_mask check relies on value-change comparison. To avoid
 341          * reporting a race where e.g. the writer set up the watchpoint, but the
 342          * reader has access_mask!=0, we have to ignore the found watchpoint.
 343          */
 344         if (get_ctx()->access_mask != 0)
 345                 return;
 346
 347         /*
 348          * Consume the watchpoint as soon as possible, to minimize the chances
 349          * of !consumed. Consuming the watchpoint must always be guarded by
 350          * kcsan_is_enabled() check, as otherwise we might erroneously
 351          * triggering reports when disabled.
 352          */
 353         consumed = try_consume_watchpoint(watchpoint, encoded_watchpoint);
 354
 355         /* keep this after try_consume_watchpoint */
 356         flags = user_access_save();
 357
 358         if (consumed) {
 359                 kcsan_save_irqtrace(current);
 360                 kcsan_report(ptr, size, type, KCSAN_VALUE_CHANGE_MAYBE,
 361                              KCSAN_REPORT_CONSUMED_WATCHPOINT,
 362                              watchpoint - watchpoints);
 363                 kcsan_restore_irqtrace(current);
 364         } else {
 365                 /*
 366                  * The other thread may not print any diagnostics, as it has
 367                  * already removed the watchpoint, or another thread consumed
 368                  * the watchpoint before this thread.
 369                  */
 370                 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_REPORT_RACES]);
 371         }
 372
 373         if ((type & KCSAN_ACCESS_ASSERT) != 0)
 374                 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_ASSERT_FAILURES]);
 375         else
 376                 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_DATA_RACES]);
 377
 378         user_access_restore(flags);
 379 }
 380
 381 static noinline void
 382 kcsan_setup_watchpoint(const volatile void *ptr, size_t size, int type)
 383 {
 384         const bool is_write = (type & KCSAN_ACCESS_WRITE) != 0;
 385         const bool is_assert = (type & KCSAN_ACCESS_ASSERT) != 0;
 386         atomic_long_t *watchpoint;
 387         union {
 388                 u8 _1;
 389                 u16 _2;
 390                 u32 _4;
 391                 u64 _8;
 392         } expect_value;
 393         unsigned long access_mask;
 394         enum kcsan_value_change value_change = KCSAN_VALUE_CHANGE_MAYBE;
 395         unsigned long ua_flags = user_access_save();
 396         unsigned long irq_flags = 0;
 397
 398         /*
 399          * Always reset kcsan_skip counter in slow-path to avoid underflow; see
 400          * should_watch().
 401          */
 402         reset_kcsan_skip();
 403
 404         if (!kcsan_is_enabled())
 405                 goto out;
 406
 407         /*
 408          * Special atomic rules: unlikely to be true, so we check them here in
 409          * the slow-path, and not in the fast-path in is_atomic(). Call after
 410          * kcsan_is_enabled(), as we may access memory that is not yet
 411          * initialized during early boot.
 412          */
 413         if (!is_assert && kcsan_is_atomic_special(ptr))
 414                 goto out;
 415
 416         if (!check_encodable((unsigned long)ptr, size)) {
 417                 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_UNENCODABLE_ACCESSES]);
 418                 goto out;
 419         }
 420
 421         /*
 422          * Save and restore the IRQ state trace touched by KCSAN, since KCSAN's
 423          * runtime is entered for every memory access, and potentially useful
 424          * information is lost if dirtied by KCSAN.
 425          */
 426         kcsan_save_irqtrace(current);
 427         if (!kcsan_interrupt_watcher)
 428                 local_irq_save(irq_flags);
 429
 430         watchpoint = insert_watchpoint((unsigned long)ptr, size, is_write);
 431         if (watchpoint == NULL) {
 432                 /*
 433                  * Out of capacity: the size of 'watchpoints', and the frequency
 434                  * with which should_watch() returns true should be tweaked so
 435                  * that this case happens very rarely.
 436                  */
 437                 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_NO_CAPACITY]);
 438                 goto out_unlock;
 439         }
 440
 441         atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_SETUP_WATCHPOINTS]);
 442         atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_USED_WATCHPOINTS]);
 443
 444         /*
 445          * Read the current value, to later check and infer a race if the data
 446          * was modified via a non-instrumented access, e.g. from a device.
 447          */
 448         expect_value._8 = 0;
 449         switch (size) {
 450         case 1:
 451                 expect_value._1 = READ_ONCE(*(const u8 *)ptr);
 452                 break;
 453         case 2:
 454                 expect_value._2 = READ_ONCE(*(const u16 *)ptr);
 455                 break;
 456         case 4:
 457                 expect_value._4 = READ_ONCE(*(const u32 *)ptr);
 458                 break;
 459         case 8:
 460                 expect_value._8 = READ_ONCE(*(const u64 *)ptr);
 461                 break;
 462         default:
 463                 break; /* ignore; we do not diff the values */
 464         }
 465
 466         if (IS_ENABLED(CONFIG_KCSAN_DEBUG)) {
 467                 kcsan_disable_current();
 468                 pr_err("watching %s, size: %zu, addr: %px [slot: %d, encoded: %lx]\n",
 469                        is_write ? "write" : "read", size, ptr,
 470                        watchpoint_slot((unsigned long)ptr),
 471                        encode_watchpoint((unsigned long)ptr, size, is_write));
 472                 kcsan_enable_current();
 473         }
 474
 475         /*
 476          * Delay this thread, to increase probability of observing a racy
 477          * conflicting access.
 478          */
 479         udelay(get_delay(type));
 480
 481         /*
 482          * Re-read value, and check if it is as expected; if not, we infer a
 483          * racy access.
 484          */
 485         access_mask = get_ctx()->access_mask;
 486         switch (size) {
 487         case 1:
 488                 expect_value._1 ^= READ_ONCE(*(const u8 *)ptr);
 489                 if (access_mask)
 490                         expect_value._1 &= (u8)access_mask;
 491                 break;
 492         case 2:
 493                 expect_value._2 ^= READ_ONCE(*(const u16 *)ptr);
 494                 if (access_mask)
 495                         expect_value._2 &= (u16)access_mask;
 496                 break;
 497         case 4:
 498                 expect_value._4 ^= READ_ONCE(*(const u32 *)ptr);
 499                 if (access_mask)
 500                         expect_value._4 &= (u32)access_mask;
 501                 break;
 502         case 8:
 503                 expect_value._8 ^= READ_ONCE(*(const u64 *)ptr);
 504                 if (access_mask)
 505                         expect_value._8 &= (u64)access_mask;
 506                 break;
 507         default:
 508                 break; /* ignore; we do not diff the values */
 509         }
 510
 511         /* Were we able to observe a value-change? */
 512         if (expect_value._8 != 0)
 513                 value_change = KCSAN_VALUE_CHANGE_TRUE;
 514
 515         /* Check if this access raced with another. */
 516         if (!consume_watchpoint(watchpoint)) {
 517                 /*
 518                  * Depending on the access type, map a value_change of MAYBE to
 519                  * TRUE (always report) or FALSE (never report).
 520                  */
 521                 if (value_change == KCSAN_VALUE_CHANGE_MAYBE) {
 522                         if (access_mask != 0) {
 523                                 /*
 524                                  * For access with access_mask, we require a
 525                                  * value-change, as it is likely that races on
 526                                  * ~access_mask bits are expected.
 527                                  */
 528                                 value_change = KCSAN_VALUE_CHANGE_FALSE;
 529                         } else if (size > 8 || is_assert) {
 530                                 /* Always assume a value-change. */
 531                                 value_change = KCSAN_VALUE_CHANGE_TRUE;
 532                         }
 533                 }
 534
 535                 /*
 536                  * No need to increment 'data_races' counter, as the racing
 537                  * thread already did.
 538                  *
 539                  * Count 'assert_failures' for each failed ASSERT access,
 540                  * therefore both this thread and the racing thread may
 541                  * increment this counter.
 542                  */
 543                 if (is_assert && value_change == KCSAN_VALUE_CHANGE_TRUE)
 544                         atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_ASSERT_FAILURES]);
 545
 546                 kcsan_report(ptr, size, type, value_change, KCSAN_REPORT_RACE_SIGNAL,
 547                              watchpoint - watchpoints);
 548         } else if (value_change == KCSAN_VALUE_CHANGE_TRUE) {
 549                 /* Inferring a race, since the value should not have changed. */
 550
 551                 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_RACES_UNKNOWN_ORIGIN]);
 552                 if (is_assert)
 553                         atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_ASSERT_FAILURES]);
 554
 555                 if (IS_ENABLED(CONFIG_KCSAN_REPORT_RACE_UNKNOWN_ORIGIN) || is_assert)
 556                         kcsan_report(ptr, size, type, KCSAN_VALUE_CHANGE_TRUE,
 557                                      KCSAN_REPORT_RACE_UNKNOWN_ORIGIN,
 558                                      watchpoint - watchpoints);
 559         }
 560
 561         /*
 562          * Remove watchpoint; must be after reporting, since the slot may be
 563          * reused after this point.
 564          */
 565         remove_watchpoint(watchpoint);
 566         atomic_long_dec(&kcsan_counters[KCSAN_COUNTER_USED_WATCHPOINTS]);
 567 out_unlock:
 568         if (!kcsan_interrupt_watcher)
 569                 local_irq_restore(irq_flags);
 570         kcsan_restore_irqtrace(current);
 571 out:
 572         user_access_restore(ua_flags);
 573 }
 574
 575 static __always_inline void check_access(const volatile void *ptr, size_t size,
 576                                          int type)
 577 {
 578         const bool is_write = (type & KCSAN_ACCESS_WRITE) != 0;
 579         atomic_long_t *watchpoint;
 580         long encoded_watchpoint;
 581
 582         /*
 583          * Do nothing for 0 sized check; this comparison will be optimized out
 584          * for constant sized instrumentation (__tsan_{read,write}N).
 585          */
 586         if (unlikely(size == 0))
 587                 return;
 588
 589         /*
 590          * Avoid user_access_save in fast-path: find_watchpoint is safe without
 591          * user_access_save, as the address that ptr points to is only used to
 592          * check if a watchpoint exists; ptr is never dereferenced.
 593          */
 594         watchpoint = find_watchpoint((unsigned long)ptr, size, !is_write,
 595                                      &encoded_watchpoint);
 596         /*
 597          * It is safe to check kcsan_is_enabled() after find_watchpoint in the
 598          * slow-path, as long as no state changes that cause a race to be
 599          * detected and reported have occurred until kcsan_is_enabled() is
 600          * checked.
 601          */
 602
 603         if (unlikely(watchpoint != NULL))
 604                 kcsan_found_watchpoint(ptr, size, type, watchpoint,
 605                                        encoded_watchpoint);
 606         else {
 607                 struct kcsan_ctx *ctx = get_ctx(); /* Call only once in fast-path. */
 608
 609                 if (unlikely(should_watch(ptr, size, type, ctx)))
 610                         kcsan_setup_watchpoint(ptr, size, type);
 611                 else if (unlikely(ctx->scoped_accesses.prev))
 612                         kcsan_check_scoped_accesses();
 613         }
 614 }
 615
 616 /* === Public interface ===================================================== */
 617
 618 void __init kcsan_init(void)
 619 {
 620         BUG_ON(!in_task());
 621
 622         kcsan_debugfs_init();
 623
 624         /*
 625          * We are in the init task, and no other tasks should be running;
 626          * WRITE_ONCE without memory barrier is sufficient.
 627          */
 628         if (kcsan_early_enable) {
 629                 pr_info("enabled early\n");
 630                 WRITE_ONCE(kcsan_enabled, true);
 631         }
 632 }
 633
 634 /* === Exported interface =================================================== */
 635
 636 void kcsan_disable_current(void)
 637 {
 638         ++get_ctx()->disable_count;
 639 }
 640 EXPORT_SYMBOL(kcsan_disable_current);
 641
 642 void kcsan_enable_current(void)
 643 {
 644         if (get_ctx()->disable_count-- == 0) {
 645                 /*
 646                  * Warn if kcsan_enable_current() calls are unbalanced with
 647                  * kcsan_disable_current() calls, which causes disable_count to
 648                  * become negative and should not happen.
 649                  */
 650                 kcsan_disable_current(); /* restore to 0, KCSAN still enabled */
 651                 kcsan_disable_current(); /* disable to generate warning */
 652                 WARN(1, "Unbalanced %s()", __func__);
 653                 kcsan_enable_current();
 654         }
 655 }
 656 EXPORT_SYMBOL(kcsan_enable_current);
 657
 658 void kcsan_enable_current_nowarn(void)
 659 {
 660         if (get_ctx()->disable_count-- == 0)
 661                 kcsan_disable_current();
 662 }
 663 EXPORT_SYMBOL(kcsan_enable_current_nowarn);
 664
 665 void kcsan_nestable_atomic_begin(void)
 666 {
 667         /*
 668          * Do *not* check and warn if we are in a flat atomic region: nestable
 669          * and flat atomic regions are independent from each other.
 670          * See include/linux/kcsan.h: struct kcsan_ctx comments for more
 671          * comments.
 672          */
 673
 674         ++get_ctx()->atomic_nest_count;
 675 }
 676 EXPORT_SYMBOL(kcsan_nestable_atomic_begin);
 677
 678 void kcsan_nestable_atomic_end(void)
 679 {
 680         if (get_ctx()->atomic_nest_count-- == 0) {
 681                 /*
 682                  * Warn if kcsan_nestable_atomic_end() calls are unbalanced with
 683                  * kcsan_nestable_atomic_begin() calls, which causes
 684                  * atomic_nest_count to become negative and should not happen.
 685                  */
 686                 kcsan_nestable_atomic_begin(); /* restore to 0 */
 687                 kcsan_disable_current(); /* disable to generate warning */
 688                 WARN(1, "Unbalanced %s()", __func__);
 689                 kcsan_enable_current();
 690         }
 691 }
 692 EXPORT_SYMBOL(kcsan_nestable_atomic_end);
 693
 694 void kcsan_flat_atomic_begin(void)
 695 {
 696         get_ctx()->in_flat_atomic = true;
 697 }
 698 EXPORT_SYMBOL(kcsan_flat_atomic_begin);
 699
 700 void kcsan_flat_atomic_end(void)
 701 {
 702         get_ctx()->in_flat_atomic = false;
 703 }
 704 EXPORT_SYMBOL(kcsan_flat_atomic_end);
 705
 706 void kcsan_atomic_next(int n)
 707 {
 708         get_ctx()->atomic_next = n;
 709 }
 710 EXPORT_SYMBOL(kcsan_atomic_next);
 711
 712 void kcsan_set_access_mask(unsigned long mask)
 713 {
 714         get_ctx()->access_mask = mask;
 715 }
 716 EXPORT_SYMBOL(kcsan_set_access_mask);
 717
 718 struct kcsan_scoped_access *
 719 kcsan_begin_scoped_access(const volatile void *ptr, size_t size, int type,
 720                           struct kcsan_scoped_access *sa)
 721 {
 722         struct kcsan_ctx *ctx = get_ctx();
 723
 724         __kcsan_check_access(ptr, size, type);
 725
 726         ctx->disable_count++; /* Disable KCSAN, in case list debugging is on. */
 727
 728         INIT_LIST_HEAD(&sa->list);
 729         sa->ptr = ptr;
 730         sa->size = size;
 731         sa->type = type;
 732
 733         if (!ctx->scoped_accesses.prev) /* Lazy initialize list head. */
 734                 INIT_LIST_HEAD(&ctx->scoped_accesses);
 735         list_add(&sa->list, &ctx->scoped_accesses);
 736
 737         ctx->disable_count--;
 738         return sa;
 739 }
 740 EXPORT_SYMBOL(kcsan_begin_scoped_access);
 741
 742 void kcsan_end_scoped_access(struct kcsan_scoped_access *sa)
 743 {
 744         struct kcsan_ctx *ctx = get_ctx();
 745
 746         if (WARN(!ctx->scoped_accesses.prev, "Unbalanced %s()?", __func__))
 747                 return;
 748
 749         ctx->disable_count++; /* Disable KCSAN, in case list debugging is on. */
 750
 751         list_del(&sa->list);
 752         if (list_empty(&ctx->scoped_accesses))
 753                 /*
 754                  * Ensure we do not enter kcsan_check_scoped_accesses()
 755                  * slow-path if unnecessary, and avoids requiring list_empty()
 756                  * in the fast-path (to avoid a READ_ONCE() and potential
 757                  * uaccess warning).
 758                  */
 759                 ctx->scoped_accesses.prev = NULL;
 760
 761         ctx->disable_count--;
 762
 763         __kcsan_check_access(sa->ptr, sa->size, sa->type);
 764 }
 765 EXPORT_SYMBOL(kcsan_end_scoped_access);
 766
 767 void __kcsan_check_access(const volatile void *ptr, size_t size, int type)
 768 {
 769         check_access(ptr, size, type);
 770 }
 771 EXPORT_SYMBOL(__kcsan_check_access);
 772
 773 /*
 774  * KCSAN uses the same instrumentation that is emitted by supported compilers
 775  * for ThreadSanitizer (TSAN).
 776  *
 777  * When enabled, the compiler emits instrumentation calls (the functions
 778  * prefixed with "__tsan" below) for all loads and stores that it generated;
 779  * inline asm is not instrumented.
 780  *
 781  * Note that, not all supported compiler versions distinguish aligned/unaligned
 782  * accesses, but e.g. recent versions of Clang do. We simply alias the unaligned
 783  * version to the generic version, which can handle both.
 784  */
 785
 786 #define DEFINE_TSAN_READ_WRITE(size)                                           \
 787         void __tsan_read##size(void *ptr);                                     \
 788         void __tsan_read##size(void *ptr)                                      \
 789         {                                                                      \
 790                 check_access(ptr, size, 0);                                    \
 791         }                                                                      \
 792         EXPORT_SYMBOL(__tsan_read##size);                                      \
 793         void __tsan_unaligned_read##size(void *ptr)                            \
 794                 __alias(__tsan_read##size);                                    \
 795         EXPORT_SYMBOL(__tsan_unaligned_read##size);                            \
 796         void __tsan_write##size(void *ptr);                                    \
 797         void __tsan_write##size(void *ptr)                                     \
 798         {                                                                      \
 799                 check_access(ptr, size, KCSAN_ACCESS_WRITE);                   \
 800         }                                                                      \
 801         EXPORT_SYMBOL(__tsan_write##size);                                     \
 802         void __tsan_unaligned_write##size(void *ptr)                           \
 803                 __alias(__tsan_write##size);                                   \
 804         EXPORT_SYMBOL(__tsan_unaligned_write##size);                           \
 805         void __tsan_read_write##size(void *ptr);                               \
 806         void __tsan_read_write##size(void *ptr)                                \
 807         {                                                                      \
 808                 check_access(ptr, size,                                        \
 809                              KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE);      \
 810         }                                                                      \
 811         EXPORT_SYMBOL(__tsan_read_write##size);                                \
 812         void __tsan_unaligned_read_write##size(void *ptr)                      \
 813                 __alias(__tsan_read_write##size);                              \
 814         EXPORT_SYMBOL(__tsan_unaligned_read_write##size)
 815
 816 DEFINE_TSAN_READ_WRITE(1);
 817 DEFINE_TSAN_READ_WRITE(2);
 818 DEFINE_TSAN_READ_WRITE(4);
 819 DEFINE_TSAN_READ_WRITE(8);
 820 DEFINE_TSAN_READ_WRITE(16);
 821
 822 void __tsan_read_range(void *ptr, size_t size);
 823 void __tsan_read_range(void *ptr, size_t size)
 824 {
 825         check_access(ptr, size, 0);
 826 }
 827 EXPORT_SYMBOL(__tsan_read_range);
 828
 829 void __tsan_write_range(void *ptr, size_t size);
 830 void __tsan_write_range(void *ptr, size_t size)
 831 {
 832         check_access(ptr, size, KCSAN_ACCESS_WRITE);
 833 }
 834 EXPORT_SYMBOL(__tsan_write_range);
 835
 836 /*
 837  * Use of explicit volatile is generally disallowed [1], however, volatile is
 838  * still used in various concurrent context, whether in low-level
 839  * synchronization primitives or for legacy reasons.
 840  * [1] https://lwn.net/Articles/233479/
 841  *
 842  * We only consider volatile accesses atomic if they are aligned and would pass
 843  * the size-check of compiletime_assert_rwonce_type().
 844  */
 845 #define DEFINE_TSAN_VOLATILE_READ_WRITE(size)                                  \
 846         void __tsan_volatile_read##size(void *ptr);                            \
 847         void __tsan_volatile_read##size(void *ptr)                             \
 848         {                                                                      \
 849                 const bool is_atomic = size <= sizeof(long long) &&            \
 850                                        IS_ALIGNED((unsigned long)ptr, size);   \
 851                 if (IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS) && is_atomic)      \
 852                         return;                                                \
 853                 check_access(ptr, size, is_atomic ? KCSAN_ACCESS_ATOMIC : 0);  \
 854         }                                                                      \
 855         EXPORT_SYMBOL(__tsan_volatile_read##size);                             \
 856         void __tsan_unaligned_volatile_read##size(void *ptr)                   \
 857                 __alias(__tsan_volatile_read##size);                           \
 858         EXPORT_SYMBOL(__tsan_unaligned_volatile_read##size);                   \
 859         void __tsan_volatile_write##size(void *ptr);                           \
 860         void __tsan_volatile_write##size(void *ptr)                            \
 861         {                                                                      \
 862                 const bool is_atomic = size <= sizeof(long long) &&            \
 863                                        IS_ALIGNED((unsigned long)ptr, size);   \
 864                 if (IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS) && is_atomic)      \
 865                         return;                                                \
 866                 check_access(ptr, size,                                        \
 867                              KCSAN_ACCESS_WRITE |                              \
 868                                      (is_atomic ? KCSAN_ACCESS_ATOMIC : 0));   \
 869         }                                                                      \
 870         EXPORT_SYMBOL(__tsan_volatile_write##size);                            \
 871         void __tsan_unaligned_volatile_write##size(void *ptr)                  \
 872                 __alias(__tsan_volatile_write##size);                          \
 873         EXPORT_SYMBOL(__tsan_unaligned_volatile_write##size)
 874
 875 DEFINE_TSAN_VOLATILE_READ_WRITE(1);
 876 DEFINE_TSAN_VOLATILE_READ_WRITE(2);
 877 DEFINE_TSAN_VOLATILE_READ_WRITE(4);
 878 DEFINE_TSAN_VOLATILE_READ_WRITE(8);
 879 DEFINE_TSAN_VOLATILE_READ_WRITE(16);
 880
 881 /*
 882  * The below are not required by KCSAN, but can still be emitted by the
 883  * compiler.
 884  */
 885 void __tsan_func_entry(void *call_pc);
 886 void __tsan_func_entry(void *call_pc)
 887 {
 888 }
 889 EXPORT_SYMBOL(__tsan_func_entry);
 890 void __tsan_func_exit(void);
 891 void __tsan_func_exit(void)
 892 {
 893 }
 894 EXPORT_SYMBOL(__tsan_func_exit);
 895 void __tsan_init(void);
 896 void __tsan_init(void)
 897 {
 898 }
 899 EXPORT_SYMBOL(__tsan_init);
 900
 901 /*
 902  * Instrumentation for atomic builtins (__atomic_*, __sync_*).
 903  *
 904  * Normal kernel code _should not_ be using them directly, but some
 905  * architectures may implement some or all atomics using the compilers'
 906  * builtins.
 907  *
 908  * Note: If an architecture decides to fully implement atomics using the
 909  * builtins, because they are implicitly instrumented by KCSAN (and KASAN,
 910  * etc.), implementing the ARCH_ATOMIC interface (to get instrumentation via
 911  * atomic-instrumented) is no longer necessary.
 912  *
 913  * TSAN instrumentation replaces atomic accesses with calls to any of the below
 914  * functions, whose job is to also execute the operation itself.
 915  */
 916
 917 #define DEFINE_TSAN_ATOMIC_LOAD_STORE(bits)                                                        \
 918         u##bits __tsan_atomic##bits##_load(const u##bits *ptr, int memorder);                      \
 919         u##bits __tsan_atomic##bits##_load(const u##bits *ptr, int memorder)                       \
 920         {                                                                                          \
 921                 if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
 922                         check_access(ptr, bits / BITS_PER_BYTE, KCSAN_ACCESS_ATOMIC);              \
 923                 }                                                                                  \
 924                 return __atomic_load_n(ptr, memorder);                                             \
 925         }                                                                                          \
 926         EXPORT_SYMBOL(__tsan_atomic##bits##_load);                                                 \
 927         void __tsan_atomic##bits##_store(u##bits *ptr, u##bits v, int memorder);                   \
 928         void __tsan_atomic##bits##_store(u##bits *ptr, u##bits v, int memorder)                    \
 929         {                                                                                          \
 930                 if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
 931                         check_access(ptr, bits / BITS_PER_BYTE,                                    \
 932                                      KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ATOMIC);                    \
 933                 }                                                                                  \
 934                 __atomic_store_n(ptr, v, memorder);                                                \
 935         }                                                                                          \
 936         EXPORT_SYMBOL(__tsan_atomic##bits##_store)
 937
 938 #define DEFINE_TSAN_ATOMIC_RMW(op, bits, suffix)                                                   \
 939         u##bits __tsan_atomic##bits##_##op(u##bits *ptr, u##bits v, int memorder);                 \
 940         u##bits __tsan_atomic##bits##_##op(u##bits *ptr, u##bits v, int memorder)                  \
 941         {                                                                                          \
 942                 if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
 943                         check_access(ptr, bits / BITS_PER_BYTE,                                    \
 944                                      KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE |                  \
 945                                              KCSAN_ACCESS_ATOMIC);                                 \
 946                 }                                                                                  \
 947                 return __atomic_##op##suffix(ptr, v, memorder);                                    \
 948         }                                                                                          \
 949         EXPORT_SYMBOL(__tsan_atomic##bits##_##op)
 950
 951 /*
 952  * Note: CAS operations are always classified as write, even in case they
 953  * fail. We cannot perform check_access() after a write, as it might lead to
 954  * false positives, in cases such as:
 955  *
 956  *      T0: __atomic_compare_exchange_n(&p->flag, &old, 1, ...)
 957  *
 958  *      T1: if (__atomic_load_n(&p->flag, ...)) {
 959  *              modify *p;
 960  *              p->flag = 0;
 961  *          }
 962  *
 963  * The only downside is that, if there are 3 threads, with one CAS that
 964  * succeeds, another CAS that fails, and an unmarked racing operation, we may
 965  * point at the wrong CAS as the source of the race. However, if we assume that
 966  * all CAS can succeed in some other execution, the data race is still valid.
 967  */
 968 #define DEFINE_TSAN_ATOMIC_CMPXCHG(bits, strength, weak)                                           \
 969         int __tsan_atomic##bits##_compare_exchange_##strength(u##bits *ptr, u##bits *exp,          \
 970                                                               u##bits val, int mo, int fail_mo);   \
 971         int __tsan_atomic##bits##_compare_exchange_##strength(u##bits *ptr, u##bits *exp,          \
 972                                                               u##bits val, int mo, int fail_mo)    \
 973         {                                                                                          \
 974                 if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
 975                         check_access(ptr, bits / BITS_PER_BYTE,                                    \
 976                                      KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE |                  \
 977                                              KCSAN_ACCESS_ATOMIC);                                 \
 978                 }                                                                                  \
 979                 return __atomic_compare_exchange_n(ptr, exp, val, weak, mo, fail_mo);              \
 980         }                                                                                          \
 981         EXPORT_SYMBOL(__tsan_atomic##bits##_compare_exchange_##strength)
 982
 983 #define DEFINE_TSAN_ATOMIC_CMPXCHG_VAL(bits)                                                       \
 984         u##bits __tsan_atomic##bits##_compare_exchange_val(u##bits *ptr, u##bits exp, u##bits val, \
 985                                                            int mo, int fail_mo);                   \
 986         u##bits __tsan_atomic##bits##_compare_exchange_val(u##bits *ptr, u##bits exp, u##bits val, \
 987                                                            int mo, int fail_mo)                    \
 988         {                                                                                          \
 989                 if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
 990                         check_access(ptr, bits / BITS_PER_BYTE,                                    \
 991                                      KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE |                  \
 992                                              KCSAN_ACCESS_ATOMIC);                                 \
 993                 }                                                                                  \
 994                 __atomic_compare_exchange_n(ptr, &exp, val, 0, mo, fail_mo);                       \
 995                 return exp;                                                                        \
 996         }                                                                                          \
 997         EXPORT_SYMBOL(__tsan_atomic##bits##_compare_exchange_val)
 998
 999 #define DEFINE_TSAN_ATOMIC_OPS(bits)                                                               \
1000         DEFINE_TSAN_ATOMIC_LOAD_STORE(bits);                                                       \
1001         DEFINE_TSAN_ATOMIC_RMW(exchange, bits, _n);                                                \
1002         DEFINE_TSAN_ATOMIC_RMW(fetch_add, bits, );                                                 \
1003         DEFINE_TSAN_ATOMIC_RMW(fetch_sub, bits, );                                                 \
1004         DEFINE_TSAN_ATOMIC_RMW(fetch_and, bits, );                                                 \
1005         DEFINE_TSAN_ATOMIC_RMW(fetch_or, bits, );                                                  \
1006         DEFINE_TSAN_ATOMIC_RMW(fetch_xor, bits, );                                                 \
1007         DEFINE_TSAN_ATOMIC_RMW(fetch_nand, bits, );                                                \
1008         DEFINE_TSAN_ATOMIC_CMPXCHG(bits, strong, 0);                                               \
1009         DEFINE_TSAN_ATOMIC_CMPXCHG(bits, weak, 1);                                                 \
1010         DEFINE_TSAN_ATOMIC_CMPXCHG_VAL(bits)
1011
1012 DEFINE_TSAN_ATOMIC_OPS(8);
1013 DEFINE_TSAN_ATOMIC_OPS(16);
1014 DEFINE_TSAN_ATOMIC_OPS(32);
1015 DEFINE_TSAN_ATOMIC_OPS(64);
1016
1017 void __tsan_atomic_thread_fence(int memorder);
1018 void __tsan_atomic_thread_fence(int memorder)
1019 {
1020         __atomic_thread_fence(memorder);
1021 }
1022 EXPORT_SYMBOL(__tsan_atomic_thread_fence);
1023
1024 void __tsan_atomic_signal_fence(int memorder);
1025 void __tsan_atomic_signal_fence(int memorder) { }
1026 EXPORT_SYMBOL(__tsan_atomic_signal_fence);