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/minmax.h>
  18 #include <linux/moduleparam.h>
  19 #include <linux/percpu.h>
  20 #include <linux/preempt.h>
  21 #include <linux/sched.h>
  22 #include <linux/string.h>
  23 #include <linux/uaccess.h>
  24
  25 #include "encoding.h"
  26 #include "kcsan.h"
  27 #include "permissive.h"
  28
  29 static bool kcsan_early_enable = IS_ENABLED(CONFIG_KCSAN_EARLY_ENABLE);
  30 unsigned int kcsan_udelay_task = CONFIG_KCSAN_UDELAY_TASK;
  31 unsigned int kcsan_udelay_interrupt = CONFIG_KCSAN_UDELAY_INTERRUPT;
  32 static long kcsan_skip_watch = CONFIG_KCSAN_SKIP_WATCH;
  33 static bool kcsan_interrupt_watcher = IS_ENABLED(CONFIG_KCSAN_INTERRUPT_WATCHER);
  34
  35 #ifdef MODULE_PARAM_PREFIX
  36 #undef MODULE_PARAM_PREFIX
  37 #endif
  38 #define MODULE_PARAM_PREFIX "kcsan."
  39 module_param_named(early_enable, kcsan_early_enable, bool, 0);
  40 module_param_named(udelay_task, kcsan_udelay_task, uint, 0644);
  41 module_param_named(udelay_interrupt, kcsan_udelay_interrupt, uint, 0644);
  42 module_param_named(skip_watch, kcsan_skip_watch, long, 0644);
  43 module_param_named(interrupt_watcher, kcsan_interrupt_watcher, bool, 0444);
  44
  45 #ifdef CONFIG_KCSAN_WEAK_MEMORY
  46 static bool kcsan_weak_memory = true;
  47 module_param_named(weak_memory, kcsan_weak_memory, bool, 0644);
  48 #else
  49 #define kcsan_weak_memory false
  50 #endif
  51
  52 bool kcsan_enabled;
  53
  54 /* Per-CPU kcsan_ctx for interrupts */
  55 static DEFINE_PER_CPU(struct kcsan_ctx, kcsan_cpu_ctx) = {
  56         .scoped_accesses        = {LIST_POISON1, NULL},
  57 };
  58
  59 /*
  60  * Helper macros to index into adjacent slots, starting from address slot
  61  * itself, followed by the right and left slots.
  62  *
  63  * The purpose is 2-fold:
  64  *
  65  *      1. if during insertion the address slot is already occupied, check if
  66  *         any adjacent slots are free;
  67  *      2. accesses that straddle a slot boundary due to size that exceeds a
  68  *         slot's range may check adjacent slots if any watchpoint matches.
  69  *
  70  * Note that accesses with very large size may still miss a watchpoint; however,
  71  * given this should be rare, this is a reasonable trade-off to make, since this
  72  * will avoid:
  73  *
  74  *      1. excessive contention between watchpoint checks and setup;
  75  *      2. larger number of simultaneous watchpoints without sacrificing
  76  *         performance.
  77  *
  78  * Example: SLOT_IDX values for KCSAN_CHECK_ADJACENT=1, where i is [0, 1, 2]:
  79  *
  80  *   slot=0:  [ 1,  2,  0]
  81  *   slot=9:  [10, 11,  9]
  82  *   slot=63: [64, 65, 63]
  83  */
  84 #define SLOT_IDX(slot, i) (slot + ((i + KCSAN_CHECK_ADJACENT) % NUM_SLOTS))
  85
  86 /*
  87  * SLOT_IDX_FAST is used in the fast-path. Not first checking the address's primary
  88  * slot (middle) is fine if we assume that races occur rarely. The set of
  89  * indices {SLOT_IDX(slot, i) | i in [0, NUM_SLOTS)} is equivalent to
  90  * {SLOT_IDX_FAST(slot, i) | i in [0, NUM_SLOTS)}.
  91  */
  92 #define SLOT_IDX_FAST(slot, i) (slot + i)
  93
  94 /*
  95  * Watchpoints, with each entry encoded as defined in encoding.h: in order to be
  96  * able to safely update and access a watchpoint without introducing locking
  97  * overhead, we encode each watchpoint as a single atomic long. The initial
  98  * zero-initialized state matches INVALID_WATCHPOINT.
  99  *
 100  * Add NUM_SLOTS-1 entries to account for overflow; this helps avoid having to
 101  * use more complicated SLOT_IDX_FAST calculation with modulo in the fast-path.
 102  */
 103 static atomic_long_t watchpoints[CONFIG_KCSAN_NUM_WATCHPOINTS + NUM_SLOTS-1];
 104
 105 /*
 106  * Instructions to skip watching counter, used in should_watch(). We use a
 107  * per-CPU counter to avoid excessive contention.
 108  */
 109 static DEFINE_PER_CPU(long, kcsan_skip);
 110
 111 /* For kcsan_prandom_u32_max(). */
 112 static DEFINE_PER_CPU(u32, kcsan_rand_state);
 113
 114 static __always_inline atomic_long_t *find_watchpoint(unsigned long addr,
 115                                                       size_t size,
 116                                                       bool expect_write,
 117                                                       long *encoded_watchpoint)
 118 {
 119         const int slot = watchpoint_slot(addr);
 120         const unsigned long addr_masked = addr & WATCHPOINT_ADDR_MASK;
 121         atomic_long_t *watchpoint;
 122         unsigned long wp_addr_masked;
 123         size_t wp_size;
 124         bool is_write;
 125         int i;
 126
 127         BUILD_BUG_ON(CONFIG_KCSAN_NUM_WATCHPOINTS < NUM_SLOTS);
 128
 129         for (i = 0; i < NUM_SLOTS; ++i) {
 130                 watchpoint = &watchpoints[SLOT_IDX_FAST(slot, i)];
 131                 *encoded_watchpoint = atomic_long_read(watchpoint);
 132                 if (!decode_watchpoint(*encoded_watchpoint, &wp_addr_masked,
 133                                        &wp_size, &is_write))
 134                         continue;
 135
 136                 if (expect_write && !is_write)
 137                         continue;
 138
 139                 /* Check if the watchpoint matches the access. */
 140                 if (matching_access(wp_addr_masked, wp_size, addr_masked, size))
 141                         return watchpoint;
 142         }
 143
 144         return NULL;
 145 }
 146
 147 static inline atomic_long_t *
 148 insert_watchpoint(unsigned long addr, size_t size, bool is_write)
 149 {
 150         const int slot = watchpoint_slot(addr);
 151         const long encoded_watchpoint = encode_watchpoint(addr, size, is_write);
 152         atomic_long_t *watchpoint;
 153         int i;
 154
 155         /* Check slot index logic, ensuring we stay within array bounds. */
 156         BUILD_BUG_ON(SLOT_IDX(0, 0) != KCSAN_CHECK_ADJACENT);
 157         BUILD_BUG_ON(SLOT_IDX(0, KCSAN_CHECK_ADJACENT+1) != 0);
 158         BUILD_BUG_ON(SLOT_IDX(CONFIG_KCSAN_NUM_WATCHPOINTS-1, KCSAN_CHECK_ADJACENT) != ARRAY_SIZE(watchpoints)-1);
 159         BUILD_BUG_ON(SLOT_IDX(CONFIG_KCSAN_NUM_WATCHPOINTS-1, KCSAN_CHECK_ADJACENT+1) != ARRAY_SIZE(watchpoints) - NUM_SLOTS);
 160
 161         for (i = 0; i < NUM_SLOTS; ++i) {
 162                 long expect_val = INVALID_WATCHPOINT;
 163
 164                 /* Try to acquire this slot. */
 165                 watchpoint = &watchpoints[SLOT_IDX(slot, i)];
 166                 if (atomic_long_try_cmpxchg_relaxed(watchpoint, &expect_val, encoded_watchpoint))
 167                         return watchpoint;
 168         }
 169
 170         return NULL;
 171 }
 172
 173 /*
 174  * Return true if watchpoint was successfully consumed, false otherwise.
 175  *
 176  * This may return false if:
 177  *
 178  *      1. another thread already consumed the watchpoint;
 179  *      2. the thread that set up the watchpoint already removed it;
 180  *      3. the watchpoint was removed and then re-used.
 181  */
 182 static __always_inline bool
 183 try_consume_watchpoint(atomic_long_t *watchpoint, long encoded_watchpoint)
 184 {
 185         return atomic_long_try_cmpxchg_relaxed(watchpoint, &encoded_watchpoint, CONSUMED_WATCHPOINT);
 186 }
 187
 188 /* Return true if watchpoint was not touched, false if already consumed. */
 189 static inline bool consume_watchpoint(atomic_long_t *watchpoint)
 190 {
 191         return atomic_long_xchg_relaxed(watchpoint, CONSUMED_WATCHPOINT) != CONSUMED_WATCHPOINT;
 192 }
 193
 194 /* Remove the watchpoint -- its slot may be reused after. */
 195 static inline void remove_watchpoint(atomic_long_t *watchpoint)
 196 {
 197         atomic_long_set(watchpoint, INVALID_WATCHPOINT);
 198 }
 199
 200 static __always_inline struct kcsan_ctx *get_ctx(void)
 201 {
 202         /*
 203          * In interrupts, use raw_cpu_ptr to avoid unnecessary checks, that would
 204          * also result in calls that generate warnings in uaccess regions.
 205          */
 206         return in_task() ? &current->kcsan_ctx : raw_cpu_ptr(&kcsan_cpu_ctx);
 207 }
 208
 209 static __always_inline void
 210 check_access(const volatile void *ptr, size_t size, int type, unsigned long ip);
 211
 212 /* Check scoped accesses; never inline because this is a slow-path! */
 213 static noinline void kcsan_check_scoped_accesses(void)
 214 {
 215         struct kcsan_ctx *ctx = get_ctx();
 216         struct kcsan_scoped_access *scoped_access;
 217
 218         if (ctx->disable_scoped)
 219                 return;
 220
 221         ctx->disable_scoped++;
 222         list_for_each_entry(scoped_access, &ctx->scoped_accesses, list) {
 223                 check_access(scoped_access->ptr, scoped_access->size,
 224                              scoped_access->type, scoped_access->ip);
 225         }
 226         ctx->disable_scoped--;
 227 }
 228
 229 /* Rules for generic atomic accesses. Called from fast-path. */
 230 static __always_inline bool
 231 is_atomic(struct kcsan_ctx *ctx, const volatile void *ptr, size_t size, int type)
 232 {
 233         if (type & KCSAN_ACCESS_ATOMIC)
 234                 return true;
 235
 236         /*
 237          * Unless explicitly declared atomic, never consider an assertion access
 238          * as atomic. This allows using them also in atomic regions, such as
 239          * seqlocks, without implicitly changing their semantics.
 240          */
 241         if (type & KCSAN_ACCESS_ASSERT)
 242                 return false;
 243
 244         if (IS_ENABLED(CONFIG_KCSAN_ASSUME_PLAIN_WRITES_ATOMIC) &&
 245             (type & KCSAN_ACCESS_WRITE) && size <= sizeof(long) &&
 246             !(type & KCSAN_ACCESS_COMPOUND) && IS_ALIGNED((unsigned long)ptr, size))
 247                 return true; /* Assume aligned writes up to word size are atomic. */
 248
 249         if (ctx->atomic_next > 0) {
 250                 /*
 251                  * Because we do not have separate contexts for nested
 252                  * interrupts, in case atomic_next is set, we simply assume that
 253                  * the outer interrupt set atomic_next. In the worst case, we
 254                  * will conservatively consider operations as atomic. This is a
 255                  * reasonable trade-off to make, since this case should be
 256                  * extremely rare; however, even if extremely rare, it could
 257                  * lead to false positives otherwise.
 258                  */
 259                 if ((hardirq_count() >> HARDIRQ_SHIFT) < 2)
 260                         --ctx->atomic_next; /* in task, or outer interrupt */
 261                 return true;
 262         }
 263
 264         return ctx->atomic_nest_count > 0 || ctx->in_flat_atomic;
 265 }
 266
 267 static __always_inline bool
 268 should_watch(struct kcsan_ctx *ctx, const volatile void *ptr, size_t size, int type)
 269 {
 270         /*
 271          * Never set up watchpoints when memory operations are atomic.
 272          *
 273          * Need to check this first, before kcsan_skip check below: (1) atomics
 274          * should not count towards skipped instructions, and (2) to actually
 275          * decrement kcsan_atomic_next for consecutive instruction stream.
 276          */
 277         if (is_atomic(ctx, ptr, size, type))
 278                 return false;
 279
 280         if (this_cpu_dec_return(kcsan_skip) >= 0)
 281                 return false;
 282
 283         /*
 284          * NOTE: If we get here, kcsan_skip must always be reset in slow path
 285          * via reset_kcsan_skip() to avoid underflow.
 286          */
 287
 288         /* this operation should be watched */
 289         return true;
 290 }
 291
 292 /*
 293  * Returns a pseudo-random number in interval [0, ep_ro). Simple linear
 294  * congruential generator, using constants from "Numerical Recipes".
 295  */
 296 static u32 kcsan_prandom_u32_max(u32 ep_ro)
 297 {
 298         u32 state = this_cpu_read(kcsan_rand_state);
 299
 300         state = 1664525 * state + 1013904223;
 301         this_cpu_write(kcsan_rand_state, state);
 302
 303         return state % ep_ro;
 304 }
 305
 306 static inline void reset_kcsan_skip(void)
 307 {
 308         long skip_count = kcsan_skip_watch -
 309                           (IS_ENABLED(CONFIG_KCSAN_SKIP_WATCH_RANDOMIZE) ?
 310                                    kcsan_prandom_u32_max(kcsan_skip_watch) :
 311                                    0);
 312         this_cpu_write(kcsan_skip, skip_count);
 313 }
 314
 315 static __always_inline bool kcsan_is_enabled(struct kcsan_ctx *ctx)
 316 {
 317         return READ_ONCE(kcsan_enabled) && !ctx->disable_count;
 318 }
 319
 320 /* Introduce delay depending on context and configuration. */
 321 static void delay_access(int type)
 322 {
 323         unsigned int delay = in_task() ? kcsan_udelay_task : kcsan_udelay_interrupt;
 324         /* For certain access types, skew the random delay to be longer. */
 325         unsigned int skew_delay_order =
 326                 (type & (KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_ASSERT)) ? 1 : 0;
 327
 328         delay -= IS_ENABLED(CONFIG_KCSAN_DELAY_RANDOMIZE) ?
 329                                kcsan_prandom_u32_max(delay >> skew_delay_order) :
 330                                0;
 331         udelay(delay);
 332 }
 333
 334 /*
 335  * Reads the instrumented memory for value change detection; value change
 336  * detection is currently done for accesses up to a size of 8 bytes.
 337  */
 338 static __always_inline u64 read_instrumented_memory(const volatile void *ptr, size_t size)
 339 {
 340         /*
 341          * In the below we don't necessarily need the read of the location to
 342          * be atomic, and we don't use READ_ONCE(), since all we need for race
 343          * detection is to observe 2 different values.
 344          *
 345          * Furthermore, on certain architectures (such as arm64), READ_ONCE()
 346          * may turn into more complex instructions than a plain load that cannot
 347          * do unaligned accesses.
 348          */
 349         switch (size) {
 350         case 1:  return *(const volatile u8 *)ptr;
 351         case 2:  return *(const volatile u16 *)ptr;
 352         case 4:  return *(const volatile u32 *)ptr;
 353         case 8:  return *(const volatile u64 *)ptr;
 354         default: return 0; /* Ignore; we do not diff the values. */
 355         }
 356 }
 357
 358 void kcsan_save_irqtrace(struct task_struct *task)
 359 {
 360 #ifdef CONFIG_TRACE_IRQFLAGS
 361         task->kcsan_save_irqtrace = task->irqtrace;
 362 #endif
 363 }
 364
 365 void kcsan_restore_irqtrace(struct task_struct *task)
 366 {
 367 #ifdef CONFIG_TRACE_IRQFLAGS
 368         task->irqtrace = task->kcsan_save_irqtrace;
 369 #endif
 370 }
 371
 372 static __always_inline int get_kcsan_stack_depth(void)
 373 {
 374 #ifdef CONFIG_KCSAN_WEAK_MEMORY
 375         return current->kcsan_stack_depth;
 376 #else
 377         BUILD_BUG();
 378         return 0;
 379 #endif
 380 }
 381
 382 static __always_inline void add_kcsan_stack_depth(int val)
 383 {
 384 #ifdef CONFIG_KCSAN_WEAK_MEMORY
 385         current->kcsan_stack_depth += val;
 386 #else
 387         BUILD_BUG();
 388 #endif
 389 }
 390
 391 static __always_inline struct kcsan_scoped_access *get_reorder_access(struct kcsan_ctx *ctx)
 392 {
 393 #ifdef CONFIG_KCSAN_WEAK_MEMORY
 394         return ctx->disable_scoped ? NULL : &ctx->reorder_access;
 395 #else
 396         return NULL;
 397 #endif
 398 }
 399
 400 static __always_inline bool
 401 find_reorder_access(struct kcsan_ctx *ctx, const volatile void *ptr, size_t size,
 402                     int type, unsigned long ip)
 403 {
 404         struct kcsan_scoped_access *reorder_access = get_reorder_access(ctx);
 405
 406         if (!reorder_access)
 407                 return false;
 408
 409         /*
 410          * Note: If accesses are repeated while reorder_access is identical,
 411          * never matches the new access, because !(type & KCSAN_ACCESS_SCOPED).
 412          */
 413         return reorder_access->ptr == ptr && reorder_access->size == size &&
 414                reorder_access->type == type && reorder_access->ip == ip;
 415 }
 416
 417 static inline void
 418 set_reorder_access(struct kcsan_ctx *ctx, const volatile void *ptr, size_t size,
 419                    int type, unsigned long ip)
 420 {
 421         struct kcsan_scoped_access *reorder_access = get_reorder_access(ctx);
 422
 423         if (!reorder_access || !kcsan_weak_memory)
 424                 return;
 425
 426         /*
 427          * To avoid nested interrupts or scheduler (which share kcsan_ctx)
 428          * reading an inconsistent reorder_access, ensure that the below has
 429          * exclusive access to reorder_access by disallowing concurrent use.
 430          */
 431         ctx->disable_scoped++;
 432         barrier();
 433         reorder_access->ptr             = ptr;
 434         reorder_access->size            = size;
 435         reorder_access->type            = type | KCSAN_ACCESS_SCOPED;
 436         reorder_access->ip              = ip;
 437         reorder_access->stack_depth     = get_kcsan_stack_depth();
 438         barrier();
 439         ctx->disable_scoped--;
 440 }
 441
 442 /*
 443  * Pull everything together: check_access() below contains the performance
 444  * critical operations; the fast-path (including check_access) functions should
 445  * all be inlinable by the instrumentation functions.
 446  *
 447  * The slow-path (kcsan_found_watchpoint, kcsan_setup_watchpoint) are
 448  * non-inlinable -- note that, we prefix these with "kcsan_" to ensure they can
 449  * be filtered from the stacktrace, as well as give them unique names for the
 450  * UACCESS whitelist of objtool. Each function uses user_access_save/restore(),
 451  * since they do not access any user memory, but instrumentation is still
 452  * emitted in UACCESS regions.
 453  */
 454
 455 static noinline void kcsan_found_watchpoint(const volatile void *ptr,
 456                                             size_t size,
 457                                             int type,
 458                                             unsigned long ip,
 459                                             atomic_long_t *watchpoint,
 460                                             long encoded_watchpoint)
 461 {
 462         const bool is_assert = (type & KCSAN_ACCESS_ASSERT) != 0;
 463         struct kcsan_ctx *ctx = get_ctx();
 464         unsigned long flags;
 465         bool consumed;
 466
 467         /*
 468          * We know a watchpoint exists. Let's try to keep the race-window
 469          * between here and finally consuming the watchpoint below as small as
 470          * possible -- avoid unneccessarily complex code until consumed.
 471          */
 472
 473         if (!kcsan_is_enabled(ctx))
 474                 return;
 475
 476         /*
 477          * The access_mask check relies on value-change comparison. To avoid
 478          * reporting a race where e.g. the writer set up the watchpoint, but the
 479          * reader has access_mask!=0, we have to ignore the found watchpoint.
 480          *
 481          * reorder_access is never created from an access with access_mask set.
 482          */
 483         if (ctx->access_mask && !find_reorder_access(ctx, ptr, size, type, ip))
 484                 return;
 485
 486         /*
 487          * If the other thread does not want to ignore the access, and there was
 488          * a value change as a result of this thread's operation, we will still
 489          * generate a report of unknown origin.
 490          *
 491          * Use CONFIG_KCSAN_REPORT_RACE_UNKNOWN_ORIGIN=n to filter.
 492          */
 493         if (!is_assert && kcsan_ignore_address(ptr))
 494                 return;
 495
 496         /*
 497          * Consuming the watchpoint must be guarded by kcsan_is_enabled() to
 498          * avoid erroneously triggering reports if the context is disabled.
 499          */
 500         consumed = try_consume_watchpoint(watchpoint, encoded_watchpoint);
 501
 502         /* keep this after try_consume_watchpoint */
 503         flags = user_access_save();
 504
 505         if (consumed) {
 506                 kcsan_save_irqtrace(current);
 507                 kcsan_report_set_info(ptr, size, type, ip, watchpoint - watchpoints);
 508                 kcsan_restore_irqtrace(current);
 509         } else {
 510                 /*
 511                  * The other thread may not print any diagnostics, as it has
 512                  * already removed the watchpoint, or another thread consumed
 513                  * the watchpoint before this thread.
 514                  */
 515                 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_REPORT_RACES]);
 516         }
 517
 518         if (is_assert)
 519                 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_ASSERT_FAILURES]);
 520         else
 521                 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_DATA_RACES]);
 522
 523         user_access_restore(flags);
 524 }
 525
 526 static noinline void
 527 kcsan_setup_watchpoint(const volatile void *ptr, size_t size, int type, unsigned long ip)
 528 {
 529         const bool is_write = (type & KCSAN_ACCESS_WRITE) != 0;
 530         const bool is_assert = (type & KCSAN_ACCESS_ASSERT) != 0;
 531         atomic_long_t *watchpoint;
 532         u64 old, new, diff;
 533         enum kcsan_value_change value_change = KCSAN_VALUE_CHANGE_MAYBE;
 534         bool interrupt_watcher = kcsan_interrupt_watcher;
 535         unsigned long ua_flags = user_access_save();
 536         struct kcsan_ctx *ctx = get_ctx();
 537         unsigned long access_mask = ctx->access_mask;
 538         unsigned long irq_flags = 0;
 539         bool is_reorder_access;
 540
 541         /*
 542          * Always reset kcsan_skip counter in slow-path to avoid underflow; see
 543          * should_watch().
 544          */
 545         reset_kcsan_skip();
 546
 547         if (!kcsan_is_enabled(ctx))
 548                 goto out;
 549
 550         /*
 551          * Check to-ignore addresses after kcsan_is_enabled(), as we may access
 552          * memory that is not yet initialized during early boot.
 553          */
 554         if (!is_assert && kcsan_ignore_address(ptr))
 555                 goto out;
 556
 557         if (!check_encodable((unsigned long)ptr, size)) {
 558                 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_UNENCODABLE_ACCESSES]);
 559                 goto out;
 560         }
 561
 562         /*
 563          * The local CPU cannot observe reordering of its own accesses, and
 564          * therefore we need to take care of 2 cases to avoid false positives:
 565          *
 566          *      1. Races of the reordered access with interrupts. To avoid, if
 567          *         the current access is reorder_access, disable interrupts.
 568          *      2. Avoid races of scoped accesses from nested interrupts (below).
 569          */
 570         is_reorder_access = find_reorder_access(ctx, ptr, size, type, ip);
 571         if (is_reorder_access)
 572                 interrupt_watcher = false;
 573         /*
 574          * Avoid races of scoped accesses from nested interrupts (or scheduler).
 575          * Assume setting up a watchpoint for a non-scoped (normal) access that
 576          * also conflicts with a current scoped access. In a nested interrupt,
 577          * which shares the context, it would check a conflicting scoped access.
 578          * To avoid, disable scoped access checking.
 579          */
 580         ctx->disable_scoped++;
 581
 582         /*
 583          * Save and restore the IRQ state trace touched by KCSAN, since KCSAN's
 584          * runtime is entered for every memory access, and potentially useful
 585          * information is lost if dirtied by KCSAN.
 586          */
 587         kcsan_save_irqtrace(current);
 588         if (!interrupt_watcher)
 589                 local_irq_save(irq_flags);
 590
 591         watchpoint = insert_watchpoint((unsigned long)ptr, size, is_write);
 592         if (watchpoint == NULL) {
 593                 /*
 594                  * Out of capacity: the size of 'watchpoints', and the frequency
 595                  * with which should_watch() returns true should be tweaked so
 596                  * that this case happens very rarely.
 597                  */
 598                 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_NO_CAPACITY]);
 599                 goto out_unlock;
 600         }
 601
 602         atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_SETUP_WATCHPOINTS]);
 603         atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_USED_WATCHPOINTS]);
 604
 605         /*
 606          * Read the current value, to later check and infer a race if the data
 607          * was modified via a non-instrumented access, e.g. from a device.
 608          */
 609         old = is_reorder_access ? 0 : read_instrumented_memory(ptr, size);
 610
 611         /*
 612          * Delay this thread, to increase probability of observing a racy
 613          * conflicting access.
 614          */
 615         delay_access(type);
 616
 617         /*
 618          * Re-read value, and check if it is as expected; if not, we infer a
 619          * racy access.
 620          */
 621         if (!is_reorder_access) {
 622                 new = read_instrumented_memory(ptr, size);
 623         } else {
 624                 /*
 625                  * Reordered accesses cannot be used for value change detection,
 626                  * because the memory location may no longer be accessible and
 627                  * could result in a fault.
 628                  */
 629                 new = 0;
 630                 access_mask = 0;
 631         }
 632
 633         diff = old ^ new;
 634         if (access_mask)
 635                 diff &= access_mask;
 636
 637         /*
 638          * Check if we observed a value change.
 639          *
 640          * Also check if the data race should be ignored (the rules depend on
 641          * non-zero diff); if it is to be ignored, the below rules for
 642          * KCSAN_VALUE_CHANGE_MAYBE apply.
 643          */
 644         if (diff && !kcsan_ignore_data_race(size, type, old, new, diff))
 645                 value_change = KCSAN_VALUE_CHANGE_TRUE;
 646
 647         /* Check if this access raced with another. */
 648         if (!consume_watchpoint(watchpoint)) {
 649                 /*
 650                  * Depending on the access type, map a value_change of MAYBE to
 651                  * TRUE (always report) or FALSE (never report).
 652                  */
 653                 if (value_change == KCSAN_VALUE_CHANGE_MAYBE) {
 654                         if (access_mask != 0) {
 655                                 /*
 656                                  * For access with access_mask, we require a
 657                                  * value-change, as it is likely that races on
 658                                  * ~access_mask bits are expected.
 659                                  */
 660                                 value_change = KCSAN_VALUE_CHANGE_FALSE;
 661                         } else if (size > 8 || is_assert) {
 662                                 /* Always assume a value-change. */
 663                                 value_change = KCSAN_VALUE_CHANGE_TRUE;
 664                         }
 665                 }
 666
 667                 /*
 668                  * No need to increment 'data_races' counter, as the racing
 669                  * thread already did.
 670                  *
 671                  * Count 'assert_failures' for each failed ASSERT access,
 672                  * therefore both this thread and the racing thread may
 673                  * increment this counter.
 674                  */
 675                 if (is_assert && value_change == KCSAN_VALUE_CHANGE_TRUE)
 676                         atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_ASSERT_FAILURES]);
 677
 678                 kcsan_report_known_origin(ptr, size, type, ip,
 679                                           value_change, watchpoint - watchpoints,
 680                                           old, new, access_mask);
 681         } else if (value_change == KCSAN_VALUE_CHANGE_TRUE) {
 682                 /* Inferring a race, since the value should not have changed. */
 683
 684                 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_RACES_UNKNOWN_ORIGIN]);
 685                 if (is_assert)
 686                         atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_ASSERT_FAILURES]);
 687
 688                 if (IS_ENABLED(CONFIG_KCSAN_REPORT_RACE_UNKNOWN_ORIGIN) || is_assert) {
 689                         kcsan_report_unknown_origin(ptr, size, type, ip,
 690                                                     old, new, access_mask);
 691                 }
 692         }
 693
 694         /*
 695          * Remove watchpoint; must be after reporting, since the slot may be
 696          * reused after this point.
 697          */
 698         remove_watchpoint(watchpoint);
 699         atomic_long_dec(&kcsan_counters[KCSAN_COUNTER_USED_WATCHPOINTS]);
 700
 701 out_unlock:
 702         if (!interrupt_watcher)
 703                 local_irq_restore(irq_flags);
 704         kcsan_restore_irqtrace(current);
 705         ctx->disable_scoped--;
 706
 707         /*
 708          * Reordered accesses cannot be used for value change detection,
 709          * therefore never consider for reordering if access_mask is set.
 710          * ASSERT_EXCLUSIVE are not real accesses, ignore them as well.
 711          */
 712         if (!access_mask && !is_assert)
 713                 set_reorder_access(ctx, ptr, size, type, ip);
 714 out:
 715         user_access_restore(ua_flags);
 716 }
 717
 718 static __always_inline void
 719 check_access(const volatile void *ptr, size_t size, int type, unsigned long ip)
 720 {
 721         atomic_long_t *watchpoint;
 722         long encoded_watchpoint;
 723
 724         /*
 725          * Do nothing for 0 sized check; this comparison will be optimized out
 726          * for constant sized instrumentation (__tsan_{read,write}N).
 727          */
 728         if (unlikely(size == 0))
 729                 return;
 730
 731 again:
 732         /*
 733          * Avoid user_access_save in fast-path: find_watchpoint is safe without
 734          * user_access_save, as the address that ptr points to is only used to
 735          * check if a watchpoint exists; ptr is never dereferenced.
 736          */
 737         watchpoint = find_watchpoint((unsigned long)ptr, size,
 738                                      !(type & KCSAN_ACCESS_WRITE),
 739                                      &encoded_watchpoint);
 740         /*
 741          * It is safe to check kcsan_is_enabled() after find_watchpoint in the
 742          * slow-path, as long as no state changes that cause a race to be
 743          * detected and reported have occurred until kcsan_is_enabled() is
 744          * checked.
 745          */
 746
 747         if (unlikely(watchpoint != NULL))
 748                 kcsan_found_watchpoint(ptr, size, type, ip, watchpoint, encoded_watchpoint);
 749         else {
 750                 struct kcsan_ctx *ctx = get_ctx(); /* Call only once in fast-path. */
 751
 752                 if (unlikely(should_watch(ctx, ptr, size, type))) {
 753                         kcsan_setup_watchpoint(ptr, size, type, ip);
 754                         return;
 755                 }
 756
 757                 if (!(type & KCSAN_ACCESS_SCOPED)) {
 758                         struct kcsan_scoped_access *reorder_access = get_reorder_access(ctx);
 759
 760                         if (reorder_access) {
 761                                 /*
 762                                  * reorder_access check: simulates reordering of
 763                                  * the access after subsequent operations.
 764                                  */
 765                                 ptr = reorder_access->ptr;
 766                                 type = reorder_access->type;
 767                                 ip = reorder_access->ip;
 768                                 /*
 769                                  * Upon a nested interrupt, this context's
 770                                  * reorder_access can be modified (shared ctx).
 771                                  * We know that upon return, reorder_access is
 772                                  * always invalidated by setting size to 0 via
 773                                  * __tsan_func_exit(). Therefore we must read
 774                                  * and check size after the other fields.
 775                                  */
 776                                 barrier();
 777                                 size = READ_ONCE(reorder_access->size);
 778                                 if (size)
 779                                         goto again;
 780                         }
 781                 }
 782
 783                 /*
 784                  * Always checked last, right before returning from runtime;
 785                  * if reorder_access is valid, checked after it was checked.
 786                  */
 787                 if (unlikely(ctx->scoped_accesses.prev))
 788                         kcsan_check_scoped_accesses();
 789         }
 790 }
 791
 792 /* === Public interface ===================================================== */
 793
 794 void __init kcsan_init(void)
 795 {
 796         int cpu;
 797
 798         BUG_ON(!in_task());
 799
 800         for_each_possible_cpu(cpu)
 801                 per_cpu(kcsan_rand_state, cpu) = (u32)get_cycles();
 802
 803         /*
 804          * We are in the init task, and no other tasks should be running;
 805          * WRITE_ONCE without memory barrier is sufficient.
 806          */
 807         if (kcsan_early_enable) {
 808                 pr_info("enabled early\n");
 809                 WRITE_ONCE(kcsan_enabled, true);
 810         }
 811
 812         if (IS_ENABLED(CONFIG_KCSAN_REPORT_VALUE_CHANGE_ONLY) ||
 813             IS_ENABLED(CONFIG_KCSAN_ASSUME_PLAIN_WRITES_ATOMIC) ||
 814             IS_ENABLED(CONFIG_KCSAN_PERMISSIVE) ||
 815             IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {
 816                 pr_warn("non-strict mode configured - use CONFIG_KCSAN_STRICT=y to see all data races\n");
 817         } else {
 818                 pr_info("strict mode configured\n");
 819         }
 820 }
 821
 822 /* === Exported interface =================================================== */
 823
 824 void kcsan_disable_current(void)
 825 {
 826         ++get_ctx()->disable_count;
 827 }
 828 EXPORT_SYMBOL(kcsan_disable_current);
 829
 830 void kcsan_enable_current(void)
 831 {
 832         if (get_ctx()->disable_count-- == 0) {
 833                 /*
 834                  * Warn if kcsan_enable_current() calls are unbalanced with
 835                  * kcsan_disable_current() calls, which causes disable_count to
 836                  * become negative and should not happen.
 837                  */
 838                 kcsan_disable_current(); /* restore to 0, KCSAN still enabled */
 839                 kcsan_disable_current(); /* disable to generate warning */
 840                 WARN(1, "Unbalanced %s()", __func__);
 841                 kcsan_enable_current();
 842         }
 843 }
 844 EXPORT_SYMBOL(kcsan_enable_current);
 845
 846 void kcsan_enable_current_nowarn(void)
 847 {
 848         if (get_ctx()->disable_count-- == 0)
 849                 kcsan_disable_current();
 850 }
 851 EXPORT_SYMBOL(kcsan_enable_current_nowarn);
 852
 853 void kcsan_nestable_atomic_begin(void)
 854 {
 855         /*
 856          * Do *not* check and warn if we are in a flat atomic region: nestable
 857          * and flat atomic regions are independent from each other.
 858          * See include/linux/kcsan.h: struct kcsan_ctx comments for more
 859          * comments.
 860          */
 861
 862         ++get_ctx()->atomic_nest_count;
 863 }
 864 EXPORT_SYMBOL(kcsan_nestable_atomic_begin);
 865
 866 void kcsan_nestable_atomic_end(void)
 867 {
 868         if (get_ctx()->atomic_nest_count-- == 0) {
 869                 /*
 870                  * Warn if kcsan_nestable_atomic_end() calls are unbalanced with
 871                  * kcsan_nestable_atomic_begin() calls, which causes
 872                  * atomic_nest_count to become negative and should not happen.
 873                  */
 874                 kcsan_nestable_atomic_begin(); /* restore to 0 */
 875                 kcsan_disable_current(); /* disable to generate warning */
 876                 WARN(1, "Unbalanced %s()", __func__);
 877                 kcsan_enable_current();
 878         }
 879 }
 880 EXPORT_SYMBOL(kcsan_nestable_atomic_end);
 881
 882 void kcsan_flat_atomic_begin(void)
 883 {
 884         get_ctx()->in_flat_atomic = true;
 885 }
 886 EXPORT_SYMBOL(kcsan_flat_atomic_begin);
 887
 888 void kcsan_flat_atomic_end(void)
 889 {
 890         get_ctx()->in_flat_atomic = false;
 891 }
 892 EXPORT_SYMBOL(kcsan_flat_atomic_end);
 893
 894 void kcsan_atomic_next(int n)
 895 {
 896         get_ctx()->atomic_next = n;
 897 }
 898 EXPORT_SYMBOL(kcsan_atomic_next);
 899
 900 void kcsan_set_access_mask(unsigned long mask)
 901 {
 902         get_ctx()->access_mask = mask;
 903 }
 904 EXPORT_SYMBOL(kcsan_set_access_mask);
 905
 906 struct kcsan_scoped_access *
 907 kcsan_begin_scoped_access(const volatile void *ptr, size_t size, int type,
 908                           struct kcsan_scoped_access *sa)
 909 {
 910         struct kcsan_ctx *ctx = get_ctx();
 911
 912         check_access(ptr, size, type, _RET_IP_);
 913
 914         ctx->disable_count++; /* Disable KCSAN, in case list debugging is on. */
 915
 916         INIT_LIST_HEAD(&sa->list);
 917         sa->ptr = ptr;
 918         sa->size = size;
 919         sa->type = type;
 920         sa->ip = _RET_IP_;
 921
 922         if (!ctx->scoped_accesses.prev) /* Lazy initialize list head. */
 923                 INIT_LIST_HEAD(&ctx->scoped_accesses);
 924         list_add(&sa->list, &ctx->scoped_accesses);
 925
 926         ctx->disable_count--;
 927         return sa;
 928 }
 929 EXPORT_SYMBOL(kcsan_begin_scoped_access);
 930
 931 void kcsan_end_scoped_access(struct kcsan_scoped_access *sa)
 932 {
 933         struct kcsan_ctx *ctx = get_ctx();
 934
 935         if (WARN(!ctx->scoped_accesses.prev, "Unbalanced %s()?", __func__))
 936                 return;
 937
 938         ctx->disable_count++; /* Disable KCSAN, in case list debugging is on. */
 939
 940         list_del(&sa->list);
 941         if (list_empty(&ctx->scoped_accesses))
 942                 /*
 943                  * Ensure we do not enter kcsan_check_scoped_accesses()
 944                  * slow-path if unnecessary, and avoids requiring list_empty()
 945                  * in the fast-path (to avoid a READ_ONCE() and potential
 946                  * uaccess warning).
 947                  */
 948                 ctx->scoped_accesses.prev = NULL;
 949
 950         ctx->disable_count--;
 951
 952         check_access(sa->ptr, sa->size, sa->type, sa->ip);
 953 }
 954 EXPORT_SYMBOL(kcsan_end_scoped_access);
 955
 956 void __kcsan_check_access(const volatile void *ptr, size_t size, int type)
 957 {
 958         check_access(ptr, size, type, _RET_IP_);
 959 }
 960 EXPORT_SYMBOL(__kcsan_check_access);
 961
 962 #define DEFINE_MEMORY_BARRIER(name, order_before_cond)                          \
 963         void __kcsan_##name(void)                                               \
 964         {                                                                       \
 965                 struct kcsan_scoped_access *sa = get_reorder_access(get_ctx()); \
 966                 if (!sa)                                                        \
 967                         return;                                                 \
 968                 if (order_before_cond)                                          \
 969                         sa->size = 0;                                           \
 970         }                                                                       \
 971         EXPORT_SYMBOL(__kcsan_##name)
 972
 973 DEFINE_MEMORY_BARRIER(mb, true);
 974 DEFINE_MEMORY_BARRIER(wmb, sa->type & (KCSAN_ACCESS_WRITE | KCSAN_ACCESS_COMPOUND));
 975 DEFINE_MEMORY_BARRIER(rmb, !(sa->type & KCSAN_ACCESS_WRITE) || (sa->type & KCSAN_ACCESS_COMPOUND));
 976 DEFINE_MEMORY_BARRIER(release, true);
 977
 978 /*
 979  * KCSAN uses the same instrumentation that is emitted by supported compilers
 980  * for ThreadSanitizer (TSAN).
 981  *
 982  * When enabled, the compiler emits instrumentation calls (the functions
 983  * prefixed with "__tsan" below) for all loads and stores that it generated;
 984  * inline asm is not instrumented.
 985  *
 986  * Note that, not all supported compiler versions distinguish aligned/unaligned
 987  * accesses, but e.g. recent versions of Clang do. We simply alias the unaligned
 988  * version to the generic version, which can handle both.
 989  */
 990
 991 #define DEFINE_TSAN_READ_WRITE(size)                                           \
 992         void __tsan_read##size(void *ptr);                                     \
 993         void __tsan_read##size(void *ptr)                                      \
 994         {                                                                      \
 995                 check_access(ptr, size, 0, _RET_IP_);                          \
 996         }                                                                      \
 997         EXPORT_SYMBOL(__tsan_read##size);                                      \
 998         void __tsan_unaligned_read##size(void *ptr)                            \
 999                 __alias(__tsan_read##size);                                    \
1000         EXPORT_SYMBOL(__tsan_unaligned_read##size);                            \
1001         void __tsan_write##size(void *ptr);                                    \
1002         void __tsan_write##size(void *ptr)                                     \
1003         {                                                                      \
1004                 check_access(ptr, size, KCSAN_ACCESS_WRITE, _RET_IP_);         \
1005         }                                                                      \
1006         EXPORT_SYMBOL(__tsan_write##size);                                     \
1007         void __tsan_unaligned_write##size(void *ptr)                           \
1008                 __alias(__tsan_write##size);                                   \
1009         EXPORT_SYMBOL(__tsan_unaligned_write##size);                           \
1010         void __tsan_read_write##size(void *ptr);                               \
1011         void __tsan_read_write##size(void *ptr)                                \
1012         {                                                                      \
1013                 check_access(ptr, size,                                        \
1014                              KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE,       \
1015                              _RET_IP_);                                        \
1016         }                                                                      \
1017         EXPORT_SYMBOL(__tsan_read_write##size);                                \
1018         void __tsan_unaligned_read_write##size(void *ptr)                      \
1019                 __alias(__tsan_read_write##size);                              \
1020         EXPORT_SYMBOL(__tsan_unaligned_read_write##size)
1021
1022 DEFINE_TSAN_READ_WRITE(1);
1023 DEFINE_TSAN_READ_WRITE(2);
1024 DEFINE_TSAN_READ_WRITE(4);
1025 DEFINE_TSAN_READ_WRITE(8);
1026 DEFINE_TSAN_READ_WRITE(16);
1027
1028 void __tsan_read_range(void *ptr, size_t size);
1029 void __tsan_read_range(void *ptr, size_t size)
1030 {
1031         check_access(ptr, size, 0, _RET_IP_);
1032 }
1033 EXPORT_SYMBOL(__tsan_read_range);
1034
1035 void __tsan_write_range(void *ptr, size_t size);
1036 void __tsan_write_range(void *ptr, size_t size)
1037 {
1038         check_access(ptr, size, KCSAN_ACCESS_WRITE, _RET_IP_);
1039 }
1040 EXPORT_SYMBOL(__tsan_write_range);
1041
1042 /*
1043  * Use of explicit volatile is generally disallowed [1], however, volatile is
1044  * still used in various concurrent context, whether in low-level
1045  * synchronization primitives or for legacy reasons.
1046  * [1] https://lwn.net/Articles/233479/
1047  *
1048  * We only consider volatile accesses atomic if they are aligned and would pass
1049  * the size-check of compiletime_assert_rwonce_type().
1050  */
1051 #define DEFINE_TSAN_VOLATILE_READ_WRITE(size)                                  \
1052         void __tsan_volatile_read##size(void *ptr);                            \
1053         void __tsan_volatile_read##size(void *ptr)                             \
1054         {                                                                      \
1055                 const bool is_atomic = size <= sizeof(long long) &&            \
1056                                        IS_ALIGNED((unsigned long)ptr, size);   \
1057                 if (IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS) && is_atomic)      \
1058                         return;                                                \
1059                 check_access(ptr, size, is_atomic ? KCSAN_ACCESS_ATOMIC : 0,   \
1060                              _RET_IP_);                                        \
1061         }                                                                      \
1062         EXPORT_SYMBOL(__tsan_volatile_read##size);                             \
1063         void __tsan_unaligned_volatile_read##size(void *ptr)                   \
1064                 __alias(__tsan_volatile_read##size);                           \
1065         EXPORT_SYMBOL(__tsan_unaligned_volatile_read##size);                   \
1066         void __tsan_volatile_write##size(void *ptr);                           \
1067         void __tsan_volatile_write##size(void *ptr)                            \
1068         {                                                                      \
1069                 const bool is_atomic = size <= sizeof(long long) &&            \
1070                                        IS_ALIGNED((unsigned long)ptr, size);   \
1071                 if (IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS) && is_atomic)      \
1072                         return;                                                \
1073                 check_access(ptr, size,                                        \
1074                              KCSAN_ACCESS_WRITE |                              \
1075                                      (is_atomic ? KCSAN_ACCESS_ATOMIC : 0),    \
1076                              _RET_IP_);                                        \
1077         }                                                                      \
1078         EXPORT_SYMBOL(__tsan_volatile_write##size);                            \
1079         void __tsan_unaligned_volatile_write##size(void *ptr)                  \
1080                 __alias(__tsan_volatile_write##size);                          \
1081         EXPORT_SYMBOL(__tsan_unaligned_volatile_write##size)
1082
1083 DEFINE_TSAN_VOLATILE_READ_WRITE(1);
1084 DEFINE_TSAN_VOLATILE_READ_WRITE(2);
1085 DEFINE_TSAN_VOLATILE_READ_WRITE(4);
1086 DEFINE_TSAN_VOLATILE_READ_WRITE(8);
1087 DEFINE_TSAN_VOLATILE_READ_WRITE(16);
1088
1089 /*
1090  * Function entry and exit are used to determine the validty of reorder_access.
1091  * Reordering of the access ends at the end of the function scope where the
1092  * access happened. This is done for two reasons:
1093  *
1094  *      1. Artificially limits the scope where missing barriers are detected.
1095  *         This minimizes false positives due to uninstrumented functions that
1096  *         contain the required barriers but were missed.
1097  *
1098  *      2. Simplifies generating the stack trace of the access.
1099  */
1100 void __tsan_func_entry(void *call_pc);
1101 noinline void __tsan_func_entry(void *call_pc)
1102 {
1103         if (!IS_ENABLED(CONFIG_KCSAN_WEAK_MEMORY))
1104                 return;
1105
1106         add_kcsan_stack_depth(1);
1107 }
1108 EXPORT_SYMBOL(__tsan_func_entry);
1109
1110 void __tsan_func_exit(void);
1111 noinline void __tsan_func_exit(void)
1112 {
1113         struct kcsan_scoped_access *reorder_access;
1114
1115         if (!IS_ENABLED(CONFIG_KCSAN_WEAK_MEMORY))
1116                 return;
1117
1118         reorder_access = get_reorder_access(get_ctx());
1119         if (!reorder_access)
1120                 goto out;
1121
1122         if (get_kcsan_stack_depth() <= reorder_access->stack_depth) {
1123                 /*
1124                  * Access check to catch cases where write without a barrier
1125                  * (supposed release) was last access in function: because
1126                  * instrumentation is inserted before the real access, a data
1127                  * race due to the write giving up a c-s would only be caught if
1128                  * we do the conflicting access after.
1129                  */
1130                 check_access(reorder_access->ptr, reorder_access->size,
1131                              reorder_access->type, reorder_access->ip);
1132                 reorder_access->size = 0;
1133                 reorder_access->stack_depth = INT_MIN;
1134         }
1135 out:
1136         add_kcsan_stack_depth(-1);
1137 }
1138 EXPORT_SYMBOL(__tsan_func_exit);
1139
1140 void __tsan_init(void);
1141 void __tsan_init(void)
1142 {
1143 }
1144 EXPORT_SYMBOL(__tsan_init);
1145
1146 /*
1147  * Instrumentation for atomic builtins (__atomic_*, __sync_*).
1148  *
1149  * Normal kernel code _should not_ be using them directly, but some
1150  * architectures may implement some or all atomics using the compilers'
1151  * builtins.
1152  *
1153  * Note: If an architecture decides to fully implement atomics using the
1154  * builtins, because they are implicitly instrumented by KCSAN (and KASAN,
1155  * etc.), implementing the ARCH_ATOMIC interface (to get instrumentation via
1156  * atomic-instrumented) is no longer necessary.
1157  *
1158  * TSAN instrumentation replaces atomic accesses with calls to any of the below
1159  * functions, whose job is to also execute the operation itself.
1160  */
1161
1162 static __always_inline void kcsan_atomic_builtin_memorder(int memorder)
1163 {
1164         if (memorder == __ATOMIC_RELEASE ||
1165             memorder == __ATOMIC_SEQ_CST ||
1166             memorder == __ATOMIC_ACQ_REL)
1167                 __kcsan_release();
1168 }
1169
1170 #define DEFINE_TSAN_ATOMIC_LOAD_STORE(bits)                                                        \
1171         u##bits __tsan_atomic##bits##_load(const u##bits *ptr, int memorder);                      \
1172         u##bits __tsan_atomic##bits##_load(const u##bits *ptr, int memorder)                       \
1173         {                                                                                          \
1174                 kcsan_atomic_builtin_memorder(memorder);                                           \
1175                 if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
1176                         check_access(ptr, bits / BITS_PER_BYTE, KCSAN_ACCESS_ATOMIC, _RET_IP_);    \
1177                 }                                                                                  \
1178                 return __atomic_load_n(ptr, memorder);                                             \
1179         }                                                                                          \
1180         EXPORT_SYMBOL(__tsan_atomic##bits##_load);                                                 \
1181         void __tsan_atomic##bits##_store(u##bits *ptr, u##bits v, int memorder);                   \
1182         void __tsan_atomic##bits##_store(u##bits *ptr, u##bits v, int memorder)                    \
1183         {                                                                                          \
1184                 kcsan_atomic_builtin_memorder(memorder);                                           \
1185                 if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
1186                         check_access(ptr, bits / BITS_PER_BYTE,                                    \
1187                                      KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ATOMIC, _RET_IP_);          \
1188                 }                                                                                  \
1189                 __atomic_store_n(ptr, v, memorder);                                                \
1190         }                                                                                          \
1191         EXPORT_SYMBOL(__tsan_atomic##bits##_store)
1192
1193 #define DEFINE_TSAN_ATOMIC_RMW(op, bits, suffix)                                                   \
1194         u##bits __tsan_atomic##bits##_##op(u##bits *ptr, u##bits v, int memorder);                 \
1195         u##bits __tsan_atomic##bits##_##op(u##bits *ptr, u##bits v, int memorder)                  \
1196         {                                                                                          \
1197                 kcsan_atomic_builtin_memorder(memorder);                                           \
1198                 if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
1199                         check_access(ptr, bits / BITS_PER_BYTE,                                    \
1200                                      KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE |                  \
1201                                              KCSAN_ACCESS_ATOMIC, _RET_IP_);                       \
1202                 }                                                                                  \
1203                 return __atomic_##op##suffix(ptr, v, memorder);                                    \
1204         }                                                                                          \
1205         EXPORT_SYMBOL(__tsan_atomic##bits##_##op)
1206
1207 /*
1208  * Note: CAS operations are always classified as write, even in case they
1209  * fail. We cannot perform check_access() after a write, as it might lead to
1210  * false positives, in cases such as:
1211  *
1212  *      T0: __atomic_compare_exchange_n(&p->flag, &old, 1, ...)
1213  *
1214  *      T1: if (__atomic_load_n(&p->flag, ...)) {
1215  *              modify *p;
1216  *              p->flag = 0;
1217  *          }
1218  *
1219  * The only downside is that, if there are 3 threads, with one CAS that
1220  * succeeds, another CAS that fails, and an unmarked racing operation, we may
1221  * point at the wrong CAS as the source of the race. However, if we assume that
1222  * all CAS can succeed in some other execution, the data race is still valid.
1223  */
1224 #define DEFINE_TSAN_ATOMIC_CMPXCHG(bits, strength, weak)                                           \
1225         int __tsan_atomic##bits##_compare_exchange_##strength(u##bits *ptr, u##bits *exp,          \
1226                                                               u##bits val, int mo, int fail_mo);   \
1227         int __tsan_atomic##bits##_compare_exchange_##strength(u##bits *ptr, u##bits *exp,          \
1228                                                               u##bits val, int mo, int fail_mo)    \
1229         {                                                                                          \
1230                 kcsan_atomic_builtin_memorder(mo);                                                 \
1231                 if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
1232                         check_access(ptr, bits / BITS_PER_BYTE,                                    \
1233                                      KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE |                  \
1234                                              KCSAN_ACCESS_ATOMIC, _RET_IP_);                       \
1235                 }                                                                                  \
1236                 return __atomic_compare_exchange_n(ptr, exp, val, weak, mo, fail_mo);              \
1237         }                                                                                          \
1238         EXPORT_SYMBOL(__tsan_atomic##bits##_compare_exchange_##strength)
1239
1240 #define DEFINE_TSAN_ATOMIC_CMPXCHG_VAL(bits)                                                       \
1241         u##bits __tsan_atomic##bits##_compare_exchange_val(u##bits *ptr, u##bits exp, u##bits val, \
1242                                                            int mo, int fail_mo);                   \
1243         u##bits __tsan_atomic##bits##_compare_exchange_val(u##bits *ptr, u##bits exp, u##bits val, \
1244                                                            int mo, int fail_mo)                    \
1245         {                                                                                          \
1246                 kcsan_atomic_builtin_memorder(mo);                                                 \
1247                 if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
1248                         check_access(ptr, bits / BITS_PER_BYTE,                                    \
1249                                      KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE |                  \
1250                                              KCSAN_ACCESS_ATOMIC, _RET_IP_);                       \
1251                 }                                                                                  \
1252                 __atomic_compare_exchange_n(ptr, &exp, val, 0, mo, fail_mo);                       \
1253                 return exp;                                                                        \
1254         }                                                                                          \
1255         EXPORT_SYMBOL(__tsan_atomic##bits##_compare_exchange_val)
1256
1257 #define DEFINE_TSAN_ATOMIC_OPS(bits)                                                               \
1258         DEFINE_TSAN_ATOMIC_LOAD_STORE(bits);                                                       \
1259         DEFINE_TSAN_ATOMIC_RMW(exchange, bits, _n);                                                \
1260         DEFINE_TSAN_ATOMIC_RMW(fetch_add, bits, );                                                 \
1261         DEFINE_TSAN_ATOMIC_RMW(fetch_sub, bits, );                                                 \
1262         DEFINE_TSAN_ATOMIC_RMW(fetch_and, bits, );                                                 \
1263         DEFINE_TSAN_ATOMIC_RMW(fetch_or, bits, );                                                  \
1264         DEFINE_TSAN_ATOMIC_RMW(fetch_xor, bits, );                                                 \
1265         DEFINE_TSAN_ATOMIC_RMW(fetch_nand, bits, );                                                \
1266         DEFINE_TSAN_ATOMIC_CMPXCHG(bits, strong, 0);                                               \
1267         DEFINE_TSAN_ATOMIC_CMPXCHG(bits, weak, 1);                                                 \
1268         DEFINE_TSAN_ATOMIC_CMPXCHG_VAL(bits)
1269
1270 DEFINE_TSAN_ATOMIC_OPS(8);
1271 DEFINE_TSAN_ATOMIC_OPS(16);
1272 DEFINE_TSAN_ATOMIC_OPS(32);
1273 #ifdef CONFIG_64BIT
1274 DEFINE_TSAN_ATOMIC_OPS(64);
1275 #endif
1276
1277 void __tsan_atomic_thread_fence(int memorder);
1278 void __tsan_atomic_thread_fence(int memorder)
1279 {
1280         kcsan_atomic_builtin_memorder(memorder);
1281         __atomic_thread_fence(memorder);
1282 }
1283 EXPORT_SYMBOL(__tsan_atomic_thread_fence);
1284
1285 /*
1286  * In instrumented files, we emit instrumentation for barriers by mapping the
1287  * kernel barriers to an __atomic_signal_fence(), which is interpreted specially
1288  * and otherwise has no relation to a real __atomic_signal_fence(). No known
1289  * kernel code uses __atomic_signal_fence().
1290  *
1291  * Since fsanitize=thread instrumentation handles __atomic_signal_fence(), which
1292  * are turned into calls to __tsan_atomic_signal_fence(), such instrumentation
1293  * can be disabled via the __no_kcsan function attribute (vs. an explicit call
1294  * which could not). When __no_kcsan is requested, __atomic_signal_fence()
1295  * generates no code.
1296  *
1297  * Note: The result of using __atomic_signal_fence() with KCSAN enabled is
1298  * potentially limiting the compiler's ability to reorder operations; however,
1299  * if barriers were instrumented with explicit calls (without LTO), the compiler
1300  * couldn't optimize much anyway. The result of a hypothetical architecture
1301  * using __atomic_signal_fence() in normal code would be KCSAN false negatives.
1302  */
1303 void __tsan_atomic_signal_fence(int memorder);
1304 noinline void __tsan_atomic_signal_fence(int memorder)
1305 {
1306         switch (memorder) {
1307         case __KCSAN_BARRIER_TO_SIGNAL_FENCE_mb:
1308                 __kcsan_mb();
1309                 break;
1310         case __KCSAN_BARRIER_TO_SIGNAL_FENCE_wmb:
1311                 __kcsan_wmb();
1312                 break;
1313         case __KCSAN_BARRIER_TO_SIGNAL_FENCE_rmb:
1314                 __kcsan_rmb();
1315                 break;
1316         case __KCSAN_BARRIER_TO_SIGNAL_FENCE_release:
1317                 __kcsan_release();
1318                 break;
1319         default:
1320                 break;
1321         }
1322 }
1323 EXPORT_SYMBOL(__tsan_atomic_signal_fence);
1324
1325 #ifdef __HAVE_ARCH_MEMSET
1326 void *__tsan_memset(void *s, int c, size_t count);
1327 noinline void *__tsan_memset(void *s, int c, size_t count)
1328 {
1329         /*
1330          * Instead of not setting up watchpoints where accessed size is greater
1331          * than MAX_ENCODABLE_SIZE, truncate checked size to MAX_ENCODABLE_SIZE.
1332          */
1333         size_t check_len = min_t(size_t, count, MAX_ENCODABLE_SIZE);
1334
1335         check_access(s, check_len, KCSAN_ACCESS_WRITE, _RET_IP_);
1336         return memset(s, c, count);
1337 }
1338 #else
1339 void *__tsan_memset(void *s, int c, size_t count) __alias(memset);
1340 #endif
1341 EXPORT_SYMBOL(__tsan_memset);
1342
1343 #ifdef __HAVE_ARCH_MEMMOVE
1344 void *__tsan_memmove(void *dst, const void *src, size_t len);
1345 noinline void *__tsan_memmove(void *dst, const void *src, size_t len)
1346 {
1347         size_t check_len = min_t(size_t, len, MAX_ENCODABLE_SIZE);
1348
1349         check_access(dst, check_len, KCSAN_ACCESS_WRITE, _RET_IP_);
1350         check_access(src, check_len, 0, _RET_IP_);
1351         return memmove(dst, src, len);
1352 }
1353 #else
1354 void *__tsan_memmove(void *dst, const void *src, size_t len) __alias(memmove);
1355 #endif
1356 EXPORT_SYMBOL(__tsan_memmove);
1357
1358 #ifdef __HAVE_ARCH_MEMCPY
1359 void *__tsan_memcpy(void *dst, const void *src, size_t len);
1360 noinline void *__tsan_memcpy(void *dst, const void *src, size_t len)
1361 {
1362         size_t check_len = min_t(size_t, len, MAX_ENCODABLE_SIZE);
1363
1364         check_access(dst, check_len, KCSAN_ACCESS_WRITE, _RET_IP_);
1365         check_access(src, check_len, 0, _RET_IP_);
1366         return memcpy(dst, src, len);
1367 }
1368 #else
1369 void *__tsan_memcpy(void *dst, const void *src, size_t len) __alias(memcpy);
1370 #endif
1371 EXPORT_SYMBOL(__tsan_memcpy);