// SPDX-License-Identifier: GPL-2.0
/*
 * KCSAN core runtime.
 *
 * Copyright (C) 2019, Google LLC.
 */

#define pr_fmt(fmt) "kcsan: " fmt

#include <linux/atomic.h>
#include <linux/bug.h>
#include <linux/delay.h>
#include <linux/export.h>
#include <linux/init.h>
#include <linux/kernel.h>
#include <linux/list.h>
#include <linux/minmax.h>
#include <linux/moduleparam.h>
#include <linux/percpu.h>
#include <linux/preempt.h>
#include <linux/sched.h>
#include <linux/string.h>
#include <linux/uaccess.h>

#include "encoding.h"
#include "kcsan.h"
#include "permissive.h"

static bool kcsan_early_enable = IS_ENABLED(CONFIG_KCSAN_EARLY_ENABLE);
unsigned int kcsan_udelay_task = CONFIG_KCSAN_UDELAY_TASK;
unsigned int kcsan_udelay_interrupt = CONFIG_KCSAN_UDELAY_INTERRUPT;
static long kcsan_skip_watch = CONFIG_KCSAN_SKIP_WATCH;
static bool kcsan_interrupt_watcher = IS_ENABLED(CONFIG_KCSAN_INTERRUPT_WATCHER);

#ifdef MODULE_PARAM_PREFIX
#undef MODULE_PARAM_PREFIX
#endif
#define MODULE_PARAM_PREFIX "kcsan."
module_param_named(early_enable, kcsan_early_enable, bool, 0);
module_param_named(udelay_task, kcsan_udelay_task, uint, 0644);
module_param_named(udelay_interrupt, kcsan_udelay_interrupt, uint, 0644);
module_param_named(skip_watch, kcsan_skip_watch, long, 0644);
module_param_named(interrupt_watcher, kcsan_interrupt_watcher, bool, 0444);

#ifdef CONFIG_KCSAN_WEAK_MEMORY
static bool kcsan_weak_memory = true;
module_param_named(weak_memory, kcsan_weak_memory, bool, 0644);
#else
#define kcsan_weak_memory false
#endif

bool kcsan_enabled;

/* Per-CPU kcsan_ctx for interrupts */
static DEFINE_PER_CPU(struct kcsan_ctx, kcsan_cpu_ctx) = {
 .scoped_accesses = {LIST_POISON1, NULL},
};

/*
 * Helper macros to index into adjacent slots, starting from address slot
 * itself, followed by the right and left slots.
 *
 * The purpose is 2-fold:
 *
 * 1. if during insertion the address slot is already occupied, check if
 *    any adjacent slots are free;
 * 2. accesses that straddle a slot boundary due to size that exceeds a
 *    slot's range may check adjacent slots if any watchpoint matches.
 *
 * Note that accesses with very large size may still miss a watchpoint; however,
 * given this should be rare, this is a reasonable trade-off to make, since this
 * will avoid:
 *
 * 1. excessive contention between watchpoint checks and setup;
 * 2. larger number of simultaneous watchpoints without sacrificing
 *    performance.
 *
 * Example: SLOT_IDX values for KCSAN_CHECK_ADJACENT=1, where i is [0, 1, 2]:
 *
 *   slot=0:  [ 1,  2,  0]
 *   slot=9:  [10, 11,  9]
 *   slot=63: [64, 65, 63]
 */
#define SLOT_IDX(slot, i) (slot + ((i + KCSAN_CHECK_ADJACENT) % NUM_SLOTS))

/*
 * SLOT_IDX_FAST is used in the fast-path. Not first checking the address's primary
 * slot (middle) is fine if we assume that races occur rarely. The set of
 * indices {SLOT_IDX(slot, i) | i in [0, NUM_SLOTS)} is equivalent to
 * {SLOT_IDX_FAST(slot, i) | i in [0, NUM_SLOTS)}.
 */
#define SLOT_IDX_FAST(slot, i) (slot + i)

/*
 * Watchpoints, with each entry encoded as defined in encoding.h: in order to be
 * able to safely update and access a watchpoint without introducing locking
 * overhead, we encode each watchpoint as a single atomic long. The initial
 * zero-initialized state matches INVALID_WATCHPOINT.
 *
 * Add NUM_SLOTS-1 entries to account for overflow; this helps avoid having to
 * use more complicated SLOT_IDX_FAST calculation with modulo in the fast-path.
 */
static atomic_long_t watchpoints[CONFIG_KCSAN_NUM_WATCHPOINTS + NUM_SLOTS-1];

/*
 * Instructions to skip watching counter, used in should_watch(). We use a
 * per-CPU counter to avoid excessive contention.
 */
static DEFINE_PER_CPU(long, kcsan_skip);

/* For kcsan_prandom_u32_max(). */
static DEFINE_PER_CPU(u32, kcsan_rand_state);

static __always_inline atomic_long_t *find_watchpoint(unsigned long addr,
            size_t size,
            bool expect_write,
            long *encoded_watchpoint)
{
 const int slot = watchpoint_slot(addr);
 const unsigned long addr_masked = addr & WATCHPOINT_ADDR_MASK;
 atomic_long_t *watchpoint;
 unsigned long wp_addr_masked;
 size_t wp_size;
 bool is_write;
 int i;

 BUILD_BUG_ON(CONFIG_KCSAN_NUM_WATCHPOINTS < NUM_SLOTS);

 for (i = 0; i < NUM_SLOTS; ++i) {
  watchpoint = &watchpoints[SLOT_IDX_FAST(slot, i)];
  *encoded_watchpoint = atomic_long_read(watchpoint);
  if (!decode_watchpoint(*encoded_watchpoint, &wp_addr_masked,
           &wp_size, &is_write))
   continue;

  if (expect_write && !is_write)
   continue;

  /* Check if the watchpoint matches the access. */
  if (matching_access(wp_addr_masked, wp_size, addr_masked, size))
   return watchpoint;
 }

 return NULL;
}

static inline atomic_long_t *
insert_watchpoint(unsigned long addr, size_t size, bool is_write)
{
 const int slot = watchpoint_slot(addr);
 const long encoded_watchpoint = encode_watchpoint(addr, size, is_write);
 atomic_long_t *watchpoint;
 int i;

 /* Check slot index logic, ensuring we stay within array bounds. */
 BUILD_BUG_ON(SLOT_IDX(0, 0) != KCSAN_CHECK_ADJACENT);
 BUILD_BUG_ON(SLOT_IDX(0, KCSAN_CHECK_ADJACENT+1) != 0);
 BUILD_BUG_ON(SLOT_IDX(CONFIG_KCSAN_NUM_WATCHPOINTS-1, KCSAN_CHECK_ADJACENT) != ARRAY_SIZE(watchpoints)-1);
 BUILD_BUG_ON(SLOT_IDX(CONFIG_KCSAN_NUM_WATCHPOINTS-1, KCSAN_CHECK_ADJACENT+1) != ARRAY_SIZE(watchpoints) - NUM_SLOTS);

 for (i = 0; i < NUM_SLOTS; ++i) {
  long expect_val = INVALID_WATCHPOINT;

  /* Try to acquire this slot. */
  watchpoint = &watchpoints[SLOT_IDX(slot, i)];
  if (atomic_long_try_cmpxchg_relaxed(watchpoint, &expect_val, encoded_watchpoint))
   return watchpoint;
 }

 return NULL;
}

/*
 * Return true if watchpoint was successfully consumed, false otherwise.
 *
 * This may return false if:
 *
 * 1. another thread already consumed the watchpoint;
 * 2. the thread that set up the watchpoint already removed it;
 * 3. the watchpoint was removed and then re-used.
 */
static __always_inline bool
try_consume_watchpoint(atomic_long_t *watchpoint, long encoded_watchpoint)
{
 return atomic_long_try_cmpxchg_relaxed(watchpoint, &encoded_watchpoint, CONSUMED_WATCHPOINT);
}

/* Return true if watchpoint was not touched, false if already consumed. */
static inline bool consume_watchpoint(atomic_long_t *watchpoint)
{
 return atomic_long_xchg_relaxed(watchpoint, CONSUMED_WATCHPOINT) != CONSUMED_WATCHPOINT;
}

/* Remove the watchpoint -- its slot may be reused after. */
static inline void remove_watchpoint(atomic_long_t *watchpoint)
{
 atomic_long_set(watchpoint, INVALID_WATCHPOINT);
}

static __always_inline struct kcsan_ctx *get_ctx(void)
{
 /*
  * In interrupts, use raw_cpu_ptr to avoid unnecessary checks, that would
  * also result in calls that generate warnings in uaccess regions.
 */
 return in_task() ? ¤t->kcsan_ctx : raw_cpu_ptr(&kcsan_cpu_ctx);
}

static __always_inline void
check_access(const volatile void *ptr, size_t size, int type, unsigned long ip);

/* Check scoped accesses; never inline because this is a slow-path! */
static noinline void kcsan_check_scoped_accesses(void)
{
 struct kcsan_ctx *ctx = get_ctx();
 struct kcsan_scoped_access *scoped_access;

 if (ctx->disable_scoped)
  return;

 ctx->disable_scoped++;
 list_for_each_entry(scoped_access, &ctx->scoped_accesses, list) {
  check_access(scoped_access->ptr, scoped_access->size,
        scoped_access->type, scoped_access->ip);
 }
 ctx->disable_scoped--;
}

/* Rules for generic atomic accesses. Called from fast-path. */
static __always_inline bool
is_atomic(struct kcsan_ctx *ctx, const volatile void *ptr, size_t size, int type)
{
 if (type & KCSAN_ACCESS_ATOMIC)
  return true;

 /*
  * Unless explicitly declared atomic, never consider an assertion access
  * as atomic. This allows using them also in atomic regions, such as
  * seqlocks, without implicitly changing their semantics.
 */
 if (type & KCSAN_ACCESS_ASSERT)
  return false;

 if (IS_ENABLED(CONFIG_KCSAN_ASSUME_PLAIN_WRITES_ATOMIC) &&
     (type & KCSAN_ACCESS_WRITE) && size <= sizeof(long) &&
     !(type & KCSAN_ACCESS_COMPOUND) && IS_ALIGNED((unsigned long)ptr, size))
  return true; /* Assume aligned writes up to word size are atomic. */

 if (ctx->atomic_next > 0) {
  /*
   * Because we do not have separate contexts for nested
   * interrupts, in case atomic_next is set, we simply assume that
   * the outer interrupt set atomic_next. In the worst case, we
   * will conservatively consider operations as atomic. This is a
   * reasonable trade-off to make, since this case should be
   * extremely rare; however, even if extremely rare, it could
   * lead to false positives otherwise.
 */
  if ((hardirq_count() >> HARDIRQ_SHIFT) < 2)
   --ctx->atomic_next; /* in task, or outer interrupt */
  return true;
 }

 return ctx->atomic_nest_count > 0 || ctx->in_flat_atomic;
}

static __always_inline bool
should_watch(struct kcsan_ctx *ctx, const volatile void *ptr, size_t size, int type)
{
 /*
  * Never set up watchpoints when memory operations are atomic.
  *
  * Need to check this first, before kcsan_skip check below: (1) atomics
  * should not count towards skipped instructions, and (2) to actually
  * decrement kcsan_atomic_next for consecutive instruction stream.
 */
 if (is_atomic(ctx, ptr, size, type))
  return false;

 if (this_cpu_dec_return(kcsan_skip) >= 0)
  return false;

 /*
  * NOTE: If we get here, kcsan_skip must always be reset in slow path
  * via reset_kcsan_skip() to avoid underflow.
 */

 /* this operation should be watched */
 return true;
}

/*
 * Returns a pseudo-random number in interval [0, ep_ro). Simple linear
 * congruential generator, using constants from "Numerical Recipes".
 */
static u32 kcsan_prandom_u32_max(u32 ep_ro)
{
 u32 state = this_cpu_read(kcsan_rand_state);

 state = 1664525 * state + 1013904223;
 this_cpu_write(kcsan_rand_state, state);

 return state % ep_ro;
}

static inline void reset_kcsan_skip(void)
{
 long skip_count = kcsan_skip_watch -
     (IS_ENABLED(CONFIG_KCSAN_SKIP_WATCH_RANDOMIZE) ?
       kcsan_prandom_u32_max(kcsan_skip_watch) :
       0);
 this_cpu_write(kcsan_skip, skip_count);
}

static __always_inline bool kcsan_is_enabled(struct kcsan_ctx *ctx)
{
 return READ_ONCE(kcsan_enabled) && !ctx->disable_count;
}

/* Introduce delay depending on context and configuration. */
static void delay_access(int type)
{
 unsigned int delay = in_task() ? kcsan_udelay_task : kcsan_udelay_interrupt;
 /* For certain access types, skew the random delay to be longer. */
 unsigned int skew_delay_order =
  (type & (KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_ASSERT)) ? 1 : 0;

 delay -= IS_ENABLED(CONFIG_KCSAN_DELAY_RANDOMIZE) ?
          kcsan_prandom_u32_max(delay >> skew_delay_order) :
          0;
 udelay(delay);
}

/*
 * Reads the instrumented memory for value change detection; value change
 * detection is currently done for accesses up to a size of 8 bytes.
 */
static __always_inline u64 read_instrumented_memory(const volatile void *ptr, size_t size)
{
 /*
  * In the below we don't necessarily need the read of the location to
  * be atomic, and we don't use READ_ONCE(), since all we need for race
  * detection is to observe 2 different values.
  *
  * Furthermore, on certain architectures (such as arm64), READ_ONCE()
  * may turn into more complex instructions than a plain load that cannot
  * do unaligned accesses.
 */
 switch (size) {
 case 1:  return *(const volatile u8 *)ptr;
 case 2:  return *(const volatile u16 *)ptr;
 case 4:  return *(const volatile u32 *)ptr;
 case 8:  return *(const volatile u64 *)ptr;
 default: return 0; /* Ignore; we do not diff the values. */
 }
}

void kcsan_save_irqtrace(struct task_struct *task)
{
#ifdef CONFIG_TRACE_IRQFLAGS
 task->kcsan_save_irqtrace = task->irqtrace;
#endif
}

void kcsan_restore_irqtrace(struct task_struct *task)
{
#ifdef CONFIG_TRACE_IRQFLAGS
 task->irqtrace = task->kcsan_save_irqtrace;
#endif
}

static __always_inline int get_kcsan_stack_depth(void)
{
#ifdef CONFIG_KCSAN_WEAK_MEMORY
 return current->kcsan_stack_depth;
#else
 BUILD_BUG();
 return 0;
#endif
}

static __always_inline void add_kcsan_stack_depth(int val)
{
#ifdef CONFIG_KCSAN_WEAK_MEMORY
 current->kcsan_stack_depth += val;
#else
 BUILD_BUG();
#endif
}

static __always_inline struct kcsan_scoped_access *get_reorder_access(struct kcsan_ctx *ctx)
{
#ifdef CONFIG_KCSAN_WEAK_MEMORY
 return ctx->disable_scoped ? NULL : &ctx->reorder_access;
#else
 return NULL;
#endif
}

static __always_inline bool
find_reorder_access(struct kcsan_ctx *ctx, const volatile void *ptr, size_t size,
      int type, unsigned long ip)
{
 struct kcsan_scoped_access *reorder_access = get_reorder_access(ctx);

 if (!reorder_access)
  return false;

 /*
  * Note: If accesses are repeated while reorder_access is identical,
  * never matches the new access, because !(type & KCSAN_ACCESS_SCOPED).
 */
 return reorder_access->ptr == ptr && reorder_access->size == size &&
        reorder_access->type == type && reorder_access->ip == ip;
}

static inline void
set_reorder_access(struct kcsan_ctx *ctx, const volatile void *ptr, size_t size,
     int type, unsigned long ip)
{
 struct kcsan_scoped_access *reorder_access = get_reorder_access(ctx);

 if (!reorder_access || !kcsan_weak_memory)
  return;

 /*
  * To avoid nested interrupts or scheduler (which share kcsan_ctx)
  * reading an inconsistent reorder_access, ensure that the below has
  * exclusive access to reorder_access by disallowing concurrent use.
 */
 ctx->disable_scoped++;
 barrier();
 reorder_access->ptr  = ptr;
 reorder_access->size  = size;
 reorder_access->type  = type | KCSAN_ACCESS_SCOPED;
 reorder_access->ip  = ip;
 reorder_access->stack_depth = get_kcsan_stack_depth();
 barrier();
 ctx->disable_scoped--;
}

/*
 * Pull everything together: check_access() below contains the performance
 * critical operations; the fast-path (including check_access) functions should
 * all be inlinable by the instrumentation functions.
 *
 * The slow-path (kcsan_found_watchpoint, kcsan_setup_watchpoint) are
 * non-inlinable -- note that, we prefix these with "kcsan_" to ensure they can
 * be filtered from the stacktrace, as well as give them unique names for the
 * UACCESS whitelist of objtool. Each function uses user_access_save/restore(),
 * since they do not access any user memory, but instrumentation is still
 * emitted in UACCESS regions.
 */

static noinline void kcsan_found_watchpoint(const volatile void *ptr,
         size_t size,
         int type,
         unsigned long ip,
         atomic_long_t *watchpoint,
         long encoded_watchpoint)
{
 const bool is_assert = (type & KCSAN_ACCESS_ASSERT) != 0;
 struct kcsan_ctx *ctx = get_ctx();
 unsigned long flags;
 bool consumed;

 /*
  * We know a watchpoint exists. Let's try to keep the race-window
  * between here and finally consuming the watchpoint below as small as
  * possible -- avoid unneccessarily complex code until consumed.
 */

 if (!kcsan_is_enabled(ctx))
  return;

 /*
  * The access_mask check relies on value-change comparison. To avoid
  * reporting a race where e.g. the writer set up the watchpoint, but the
  * reader has access_mask!=0, we have to ignore the found watchpoint.
  *
  * reorder_access is never created from an access with access_mask set.
 */
 if (ctx->access_mask && !find_reorder_access(ctx, ptr, size, type, ip))
  return;

 /*
  * If the other thread does not want to ignore the access, and there was
  * a value change as a result of this thread's operation, we will still
  * generate a report of unknown origin.
  *
  * Use CONFIG_KCSAN_REPORT_RACE_UNKNOWN_ORIGIN=n to filter.
 */
 if (!is_assert && kcsan_ignore_address(ptr))
  return;

 /*
  * Consuming the watchpoint must be guarded by kcsan_is_enabled() to
  * avoid erroneously triggering reports if the context is disabled.
 */
 consumed = try_consume_watchpoint(watchpoint, encoded_watchpoint);

 /* keep this after try_consume_watchpoint */
 flags = user_access_save();

 if (consumed) {
  kcsan_save_irqtrace(current);
  kcsan_report_set_info(ptr, size, type, ip, watchpoint - watchpoints);
  kcsan_restore_irqtrace(current);
 } else {
  /*
   * The other thread may not print any diagnostics, as it has
   * already removed the watchpoint, or another thread consumed
   * the watchpoint before this thread.
 */
  atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_REPORT_RACES]);
 }

 if (is_assert)
  atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_ASSERT_FAILURES]);
 else
  atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_DATA_RACES]);

 user_access_restore(flags);
}

static noinline void
kcsan_setup_watchpoint(const volatile void *ptr, size_t size, int type, unsigned long ip)
{
 const bool is_write = (type & KCSAN_ACCESS_WRITE) != 0;
 const bool is_assert = (type & KCSAN_ACCESS_ASSERT) != 0;
 atomic_long_t *watchpoint;
 u64 old, new, diff;
 enum kcsan_value_change value_change = KCSAN_VALUE_CHANGE_MAYBE;
 bool interrupt_watcher = kcsan_interrupt_watcher;
 unsigned long ua_flags = user_access_save();
 struct kcsan_ctx *ctx = get_ctx();
 unsigned long access_mask = ctx->access_mask;
 unsigned long irq_flags = 0;
 bool is_reorder_access;

 /*
  * Always reset kcsan_skip counter in slow-path to avoid underflow; see
  * should_watch().
 */
 reset_kcsan_skip();

 if (!kcsan_is_enabled(ctx))
  goto out;

 /*
  * Check to-ignore addresses after kcsan_is_enabled(), as we may access
  * memory that is not yet initialized during early boot.
 */
 if (!is_assert && kcsan_ignore_address(ptr))
  goto out;

 if (!check_encodable((unsigned long)ptr, size)) {
  atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_UNENCODABLE_ACCESSES]);
  goto out;
 }

 /*
  * The local CPU cannot observe reordering of its own accesses, and
  * therefore we need to take care of 2 cases to avoid false positives:
  *
  * 1. Races of the reordered access with interrupts. To avoid, if
  *    the current access is reorder_access, disable interrupts.
  * 2. Avoid races of scoped accesses from nested interrupts (below).
 */
 is_reorder_access = find_reorder_access(ctx, ptr, size, type, ip);
 if (is_reorder_access)
  interrupt_watcher = false;
 /*
  * Avoid races of scoped accesses from nested interrupts (or scheduler).
  * Assume setting up a watchpoint for a non-scoped (normal) access that
  * also conflicts with a current scoped access. In a nested interrupt,
  * which shares the context, it would check a conflicting scoped access.
  * To avoid, disable scoped access checking.
 */
 ctx->disable_scoped++;

 /*
  * Save and restore the IRQ state trace touched by KCSAN, since KCSAN's
  * runtime is entered for every memory access, and potentially useful
  * information is lost if dirtied by KCSAN.
 */
 kcsan_save_irqtrace(current);
 if (!interrupt_watcher)
  local_irq_save(irq_flags);

 watchpoint = insert_watchpoint((unsigned long)ptr, size, is_write);
 if (watchpoint == NULL) {
  /*
   * Out of capacity: the size of 'watchpoints', and the frequency
   * with which should_watch() returns true should be tweaked so
   * that this case happens very rarely.
 */
  atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_NO_CAPACITY]);
  goto out_unlock;
 }

 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_SETUP_WATCHPOINTS]);
 atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_USED_WATCHPOINTS]);

 /*
  * Read the current value, to later check and infer a race if the data
  * was modified via a non-instrumented access, e.g. from a device.
 */
 old = is_reorder_access ? 0 : read_instrumented_memory(ptr, size);

 /*
  * Delay this thread, to increase probability of observing a racy
  * conflicting access.
 */
 delay_access(type);

 /*
  * Re-read value, and check if it is as expected; if not, we infer a
  * racy access.
 */
 if (!is_reorder_access) {
  new = read_instrumented_memory(ptr, size);
 } else {
  /*
   * Reordered accesses cannot be used for value change detection,
   * because the memory location may no longer be accessible and
   * could result in a fault.
 */
  new = 0;
  access_mask = 0;
 }

 diff = old ^ new;
 if (access_mask)
  diff &= access_mask;

 /*
  * Check if we observed a value change.
  *
  * Also check if the data race should be ignored (the rules depend on
  * non-zero diff); if it is to be ignored, the below rules for
  * KCSAN_VALUE_CHANGE_MAYBE apply.
 */
 if (diff && !kcsan_ignore_data_race(size, type, old, new, diff))
  value_change = KCSAN_VALUE_CHANGE_TRUE;

 /* Check if this access raced with another. */
 if (!consume_watchpoint(watchpoint)) {
  /*
   * Depending on the access type, map a value_change of MAYBE to
   * TRUE (always report) or FALSE (never report).
 */
  if (value_change == KCSAN_VALUE_CHANGE_MAYBE) {
   if (access_mask != 0) {
    /*
     * For access with access_mask, we require a
     * value-change, as it is likely that races on
     * ~access_mask bits are expected.
 */
    value_change = KCSAN_VALUE_CHANGE_FALSE;
   } else if (size > 8 || is_assert) {
    /* Always assume a value-change. */
    value_change = KCSAN_VALUE_CHANGE_TRUE;
   }
  }

  /*
   * No need to increment 'data_races' counter, as the racing
   * thread already did.
   *
   * Count 'assert_failures' for each failed ASSERT access,
   * therefore both this thread and the racing thread may
   * increment this counter.
 */
  if (is_assert && value_change == KCSAN_VALUE_CHANGE_TRUE)
   atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_ASSERT_FAILURES]);

  kcsan_report_known_origin(ptr, size, type, ip,
       value_change, watchpoint - watchpoints,
       old, new, access_mask);
 } else if (value_change == KCSAN_VALUE_CHANGE_TRUE) {
  /* Inferring a race, since the value should not have changed. */

  atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_RACES_UNKNOWN_ORIGIN]);
  if (is_assert)
   atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_ASSERT_FAILURES]);

  if (IS_ENABLED(CONFIG_KCSAN_REPORT_RACE_UNKNOWN_ORIGIN) || is_assert) {
   kcsan_report_unknown_origin(ptr, size, type, ip,
          old, new, access_mask);
  }
 }

 /*
  * Remove watchpoint; must be after reporting, since the slot may be
  * reused after this point.
 */
 remove_watchpoint(watchpoint);
 atomic_long_dec(&kcsan_counters[KCSAN_COUNTER_USED_WATCHPOINTS]);

out_unlock:
 if (!interrupt_watcher)
  local_irq_restore(irq_flags);
 kcsan_restore_irqtrace(current);
 ctx->disable_scoped--;

 /*
  * Reordered accesses cannot be used for value change detection,
  * therefore never consider for reordering if access_mask is set.
  * ASSERT_EXCLUSIVE are not real accesses, ignore them as well.
 */
 if (!access_mask && !is_assert)
  set_reorder_access(ctx, ptr, size, type, ip);
out:
 user_access_restore(ua_flags);
}

static __always_inline void
check_access(const volatile void *ptr, size_t size, int type, unsigned long ip)
{
 atomic_long_t *watchpoint;
 long encoded_watchpoint;

 /*
  * Do nothing for 0 sized check; this comparison will be optimized out
  * for constant sized instrumentation (__tsan_{read,write}N).
 */
 if (unlikely(size == 0))
  return;

again:
 /*
  * Avoid user_access_save in fast-path: find_watchpoint is safe without
  * user_access_save, as the address that ptr points to is only used to
  * check if a watchpoint exists; ptr is never dereferenced.
 */
 watchpoint = find_watchpoint((unsigned long)ptr, size,
         !(type & KCSAN_ACCESS_WRITE),
         &encoded_watchpoint);
 /*
  * It is safe to check kcsan_is_enabled() after find_watchpoint in the
  * slow-path, as long as no state changes that cause a race to be
  * detected and reported have occurred until kcsan_is_enabled() is
  * checked.
 */

 if (unlikely(watchpoint != NULL))
  kcsan_found_watchpoint(ptr, size, type, ip, watchpoint, encoded_watchpoint);
 else {
  struct kcsan_ctx *ctx = get_ctx(); /* Call only once in fast-path. */

  if (unlikely(should_watch(ctx, ptr, size, type))) {
   kcsan_setup_watchpoint(ptr, size, type, ip);
   return;
  }

  if (!(type & KCSAN_ACCESS_SCOPED)) {
   struct kcsan_scoped_access *reorder_access = get_reorder_access(ctx);

   if (reorder_access) {
    /*
     * reorder_access check: simulates reordering of
     * the access after subsequent operations.
 */
    ptr = reorder_access->ptr;
    type = reorder_access->type;
    ip = reorder_access->ip;
    /*
     * Upon a nested interrupt, this context's
     * reorder_access can be modified (shared ctx).
     * We know that upon return, reorder_access is
     * always invalidated by setting size to 0 via
     * __tsan_func_exit(). Therefore we must read
     * and check size after the other fields.
 */
    barrier();
    size = READ_ONCE(reorder_access->size);
    if (size)
     goto again;
   }
  }

  /*
   * Always checked last, right before returning from runtime;
   * if reorder_access is valid, checked after it was checked.
 */
  if (unlikely(ctx->scoped_accesses.prev))
   kcsan_check_scoped_accesses();
 }
}

/* === Public interface ===================================================== */

void __init kcsan_init(void)
{
 int cpu;

 BUG_ON(!in_task());

 for_each_possible_cpu(cpu)
  per_cpu(kcsan_rand_state, cpu) = (u32)get_cycles();

 /*
  * We are in the init task, and no other tasks should be running;
  * WRITE_ONCE without memory barrier is sufficient.
 */
 if (kcsan_early_enable) {
  pr_info("enabled early\n");
  WRITE_ONCE(kcsan_enabled, true);
 }

 if (IS_ENABLED(CONFIG_KCSAN_REPORT_VALUE_CHANGE_ONLY) ||
     IS_ENABLED(CONFIG_KCSAN_ASSUME_PLAIN_WRITES_ATOMIC) ||
     IS_ENABLED(CONFIG_KCSAN_PERMISSIVE) ||
     IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {
  pr_warn("non-strict mode configured - use CONFIG_KCSAN_STRICT=y to see all data races\n");
 } else {
  pr_info("strict mode configured\n");
 }
}

/* === Exported interface =================================================== */

void kcsan_disable_current(void)
{
 ++get_ctx()->disable_count;
}
EXPORT_SYMBOL(kcsan_disable_current);

void kcsan_enable_current(void)
{
 if (get_ctx()->disable_count-- == 0) {
  /*
   * Warn if kcsan_enable_current() calls are unbalanced with
   * kcsan_disable_current() calls, which causes disable_count to
   * become negative and should not happen.
 */
  kcsan_disable_current(); /* restore to 0, KCSAN still enabled */
  kcsan_disable_current(); /* disable to generate warning */
  WARN(1, "Unbalanced %s()", __func__);
  kcsan_enable_current();
 }
}
EXPORT_SYMBOL(kcsan_enable_current);

void kcsan_enable_current_nowarn(void)
{
 if (get_ctx()->disable_count-- == 0)
  kcsan_disable_current();
}
EXPORT_SYMBOL(kcsan_enable_current_nowarn);

void kcsan_nestable_atomic_begin(void)
{
 /*
  * Do *not* check and warn if we are in a flat atomic region: nestable
  * and flat atomic regions are independent from each other.
  * See include/linux/kcsan.h: struct kcsan_ctx comments for more
  * comments.
 */

 ++get_ctx()->atomic_nest_count;
}
EXPORT_SYMBOL(kcsan_nestable_atomic_begin);

void kcsan_nestable_atomic_end(void)
{
 if (get_ctx()->atomic_nest_count-- == 0) {
  /*
   * Warn if kcsan_nestable_atomic_end() calls are unbalanced with
   * kcsan_nestable_atomic_begin() calls, which causes
   * atomic_nest_count to become negative and should not happen.
 */
  kcsan_nestable_atomic_begin(); /* restore to 0 */
  kcsan_disable_current(); /* disable to generate warning */
  WARN(1, "Unbalanced %s()", __func__);
  kcsan_enable_current();
 }
}
EXPORT_SYMBOL(kcsan_nestable_atomic_end);

void kcsan_flat_atomic_begin(void)
{
 get_ctx()->in_flat_atomic = true;
}
EXPORT_SYMBOL(kcsan_flat_atomic_begin);

void kcsan_flat_atomic_end(void)
{
 get_ctx()->in_flat_atomic = false;
}
EXPORT_SYMBOL(kcsan_flat_atomic_end);

void kcsan_atomic_next(int n)
{
 get_ctx()->atomic_next = n;
}
EXPORT_SYMBOL(kcsan_atomic_next);

void kcsan_set_access_mask(unsigned long mask)
{
 get_ctx()->access_mask = mask;
}
EXPORT_SYMBOL(kcsan_set_access_mask);

struct kcsan_scoped_access *
kcsan_begin_scoped_access(const volatile void *ptr, size_t size, int type,
     struct kcsan_scoped_access *sa)
{
 struct kcsan_ctx *ctx = get_ctx();

 check_access(ptr, size, type, _RET_IP_);

 ctx->disable_count++; /* Disable KCSAN, in case list debugging is on. */

 INIT_LIST_HEAD(&sa->list);
 sa->ptr = ptr;
 sa->size = size;
 sa->type = type;
 sa->ip = _RET_IP_;

 if (!ctx->scoped_accesses.prev) /* Lazy initialize list head. */
  INIT_LIST_HEAD(&ctx->scoped_accesses);
 list_add(&sa->list, &ctx->scoped_accesses);

 ctx->disable_count--;
 return sa;
}
EXPORT_SYMBOL(kcsan_begin_scoped_access);

void kcsan_end_scoped_access(struct kcsan_scoped_access *sa)
{
 struct kcsan_ctx *ctx = get_ctx();

 if (WARN(!ctx->scoped_accesses.prev, "Unbalanced %s()?", __func__))
  return;

 ctx->disable_count++; /* Disable KCSAN, in case list debugging is on. */

 list_del(&sa->list);
 if (list_empty(&ctx->scoped_accesses))
  /*
   * Ensure we do not enter kcsan_check_scoped_accesses()
   * slow-path if unnecessary, and avoids requiring list_empty()
   * in the fast-path (to avoid a READ_ONCE() and potential
   * uaccess warning).
 */
  ctx->scoped_accesses.prev = NULL;

 ctx->disable_count--;

 check_access(sa->ptr, sa->size, sa->type, sa->ip);
}
EXPORT_SYMBOL(kcsan_end_scoped_access);

void __kcsan_check_access(const volatile void *ptr, size_t size, int type)
{
 check_access(ptr, size, type, _RET_IP_);
}
EXPORT_SYMBOL(__kcsan_check_access);

#define DEFINE_MEMORY_BARRIER(name, order_before_cond)    \
 void __kcsan_##name(void)      \
 {         \
  struct kcsan_scoped_access *sa = get_reorder_access(get_ctx()); \
  if (!sa)       \
   return;       \
  if (order_before_cond)      \
   sa->size = 0;      \
 }         \
 EXPORT_SYMBOL(__kcsan_##name)

DEFINE_MEMORY_BARRIER(mb, true);
DEFINE_MEMORY_BARRIER(wmb, sa->type & (KCSAN_ACCESS_WRITE | KCSAN_ACCESS_COMPOUND));
DEFINE_MEMORY_BARRIER(rmb, !(sa->type & KCSAN_ACCESS_WRITE) || (sa->type & KCSAN_ACCESS_COMPOUND));
DEFINE_MEMORY_BARRIER(release, true);

/*
 * KCSAN uses the same instrumentation that is emitted by supported compilers
 * for ThreadSanitizer (TSAN).
 *
 * When enabled, the compiler emits instrumentation calls (the functions
 * prefixed with "__tsan" below) for all loads and stores that it generated;
 * inline asm is not instrumented.
 *
 * Note that, not all supported compiler versions distinguish aligned/unaligned
 * accesses, but e.g. recent versions of Clang do. We simply alias the unaligned
 * version to the generic version, which can handle both.
 */

#define DEFINE_TSAN_READ_WRITE(size)                                           \
 void __tsan_read##size(void *ptr);                                     \
 void __tsan_read##size(void *ptr)                                      \
 {                                                                      \
  check_access(ptr, size, 0, _RET_IP_);                          \
 }                                                                      \
 EXPORT_SYMBOL(__tsan_read##size);                                      \
 void __tsan_unaligned_read##size(void *ptr)                            \
  __alias(__tsan_read##size);                                    \
 EXPORT_SYMBOL(__tsan_unaligned_read##size);                            \
 void __tsan_write##size(void *ptr);                                    \
 void __tsan_write##size(void *ptr)                                     \
 {                                                                      \
  check_access(ptr, size, KCSAN_ACCESS_WRITE, _RET_IP_);         \
 }                                                                      \
 EXPORT_SYMBOL(__tsan_write##size);                                     \
 void __tsan_unaligned_write##size(void *ptr)                           \
  __alias(__tsan_write##size);                                   \
 EXPORT_SYMBOL(__tsan_unaligned_write##size);                           \
 void __tsan_read_write##size(void *ptr);                               \
 void __tsan_read_write##size(void *ptr)                                \
 {                                                                      \
  check_access(ptr, size,                                        \
        KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE,       \
        _RET_IP_);                                        \
 }                                                                      \
 EXPORT_SYMBOL(__tsan_read_write##size);                                \
 void __tsan_unaligned_read_write##size(void *ptr)                      \
  __alias(__tsan_read_write##size);                              \
 EXPORT_SYMBOL(__tsan_unaligned_read_write##size)

DEFINE_TSAN_READ_WRITE(1);
DEFINE_TSAN_READ_WRITE(2);
DEFINE_TSAN_READ_WRITE(4);
DEFINE_TSAN_READ_WRITE(8);
DEFINE_TSAN_READ_WRITE(16);

void __tsan_read_range(void *ptr, size_t size);
void __tsan_read_range(void *ptr, size_t size)
{
 check_access(ptr, size, 0, _RET_IP_);
}
EXPORT_SYMBOL(__tsan_read_range);

void __tsan_write_range(void *ptr, size_t size);
void __tsan_write_range(void *ptr, size_t size)
{
 check_access(ptr, size, KCSAN_ACCESS_WRITE, _RET_IP_);
}
EXPORT_SYMBOL(__tsan_write_range);

/*
 * Use of explicit volatile is generally disallowed [1], however, volatile is
 * still used in various concurrent context, whether in low-level
 * synchronization primitives or for legacy reasons.
 * [1] https://lwn.net/Articles/233479/
 *
 * We only consider volatile accesses atomic if they are aligned and would pass
 * the size-check of compiletime_assert_rwonce_type().
 */
#define DEFINE_TSAN_VOLATILE_READ_WRITE(size)                                  \
 void __tsan_volatile_read##size(void *ptr);                            \
 void __tsan_volatile_read##size(void *ptr)                             \
 {                                                                      \
  const bool is_atomic = size <= sizeof(long long) &&            \
           IS_ALIGNED((unsigned long)ptr, size);   \
  if (IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS) && is_atomic)      \
   return;                                                \
  check_access(ptr, size, is_atomic ? KCSAN_ACCESS_ATOMIC : 0,   \
        _RET_IP_);                                        \
 }                                                                      \
 EXPORT_SYMBOL(__tsan_volatile_read##size);                             \
 void __tsan_unaligned_volatile_read##size(void *ptr)                   \
  __alias(__tsan_volatile_read##size);                           \
 EXPORT_SYMBOL(__tsan_unaligned_volatile_read##size);                   \
 void __tsan_volatile_write##size(void *ptr);                           \
 void __tsan_volatile_write##size(void *ptr)                            \
 {                                                                      \
  const bool is_atomic = size <= sizeof(long long) &&            \
           IS_ALIGNED((unsigned long)ptr, size);   \
  if (IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS) && is_atomic)      \
   return;                                                \
  check_access(ptr, size,                                        \
        KCSAN_ACCESS_WRITE |                              \
         (is_atomic ? KCSAN_ACCESS_ATOMIC : 0),    \
        _RET_IP_);                                        \
 }                                                                      \
 EXPORT_SYMBOL(__tsan_volatile_write##size);                            \
 void __tsan_unaligned_volatile_write##size(void *ptr)                  \
  __alias(__tsan_volatile_write##size);                          \
 EXPORT_SYMBOL(__tsan_unaligned_volatile_write##size)

DEFINE_TSAN_VOLATILE_READ_WRITE(1);
DEFINE_TSAN_VOLATILE_READ_WRITE(2);
DEFINE_TSAN_VOLATILE_READ_WRITE(4);
DEFINE_TSAN_VOLATILE_READ_WRITE(8);
DEFINE_TSAN_VOLATILE_READ_WRITE(16);

/*
 * Function entry and exit are used to determine the validty of reorder_access.
 * Reordering of the access ends at the end of the function scope where the
 * access happened. This is done for two reasons:
 *
 * 1. Artificially limits the scope where missing barriers are detected.
 *    This minimizes false positives due to uninstrumented functions that
 *    contain the required barriers but were missed.
 *
 * 2. Simplifies generating the stack trace of the access.
 */
void __tsan_func_entry(void *call_pc);
noinline void __tsan_func_entry(void *call_pc)
{
 if (!IS_ENABLED(CONFIG_KCSAN_WEAK_MEMORY))
  return;

 add_kcsan_stack_depth(1);
}
EXPORT_SYMBOL(__tsan_func_entry);

void __tsan_func_exit(void);
noinline void __tsan_func_exit(void)
{
 struct kcsan_scoped_access *reorder_access;

 if (!IS_ENABLED(CONFIG_KCSAN_WEAK_MEMORY))
  return;

 reorder_access = get_reorder_access(get_ctx());
 if (!reorder_access)
  goto out;

 if (get_kcsan_stack_depth() <= reorder_access->stack_depth) {
  /*
   * Access check to catch cases where write without a barrier
   * (supposed release) was last access in function: because
   * instrumentation is inserted before the real access, a data
   * race due to the write giving up a c-s would only be caught if
   * we do the conflicting access after.
 */
  check_access(reorder_access->ptr, reorder_access->size,
        reorder_access->type, reorder_access->ip);
  reorder_access->size = 0;
  reorder_access->stack_depth = INT_MIN;
 }
out:
 add_kcsan_stack_depth(-1);
}
EXPORT_SYMBOL(__tsan_func_exit);

void __tsan_init(void);
void __tsan_init(void)
{
}
EXPORT_SYMBOL(__tsan_init);

/*
 * Instrumentation for atomic builtins (__atomic_*, __sync_*).
 *
 * Normal kernel code _should not_ be using them directly, but some
 * architectures may implement some or all atomics using the compilers'
 * builtins.
 *
 * Note: If an architecture decides to fully implement atomics using the
 * builtins, because they are implicitly instrumented by KCSAN (and KASAN,
 * etc.), implementing the ARCH_ATOMIC interface (to get instrumentation via
 * atomic-instrumented) is no longer necessary.
 *
 * TSAN instrumentation replaces atomic accesses with calls to any of the below
 * functions, whose job is to also execute the operation itself.
 */

static __always_inline void kcsan_atomic_builtin_memorder(int memorder)
{
 if (memorder == __ATOMIC_RELEASE ||
     memorder == __ATOMIC_SEQ_CST ||
     memorder == __ATOMIC_ACQ_REL)
  __kcsan_release();
}

#define DEFINE_TSAN_ATOMIC_LOAD_STORE(bits)                                                        \
 u##bits __tsan_atomic##bits##_load(const u##bits *ptr, int memorder);                      \
 u##bits __tsan_atomic##bits##_load(const u##bits *ptr, int memorder)                       \
 {                                                                                          \
  kcsan_atomic_builtin_memorder(memorder);                                           \
  if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
   check_access(ptr, bits / BITS_PER_BYTE, KCSAN_ACCESS_ATOMIC, _RET_IP_);    \
  }                                                                                  \
  return __atomic_load_n(ptr, memorder);                                             \
 }                                                                                          \
 EXPORT_SYMBOL(__tsan_atomic##bits##_load);                                                 \
 void __tsan_atomic##bits##_store(u##bits *ptr, u##bits v, int memorder);                   \
 void __tsan_atomic##bits##_store(u##bits *ptr, u##bits v, int memorder)                    \
 {                                                                                          \
  kcsan_atomic_builtin_memorder(memorder);                                           \
  if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
   check_access(ptr, bits / BITS_PER_BYTE,                                    \
         KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ATOMIC, _RET_IP_);          \
  }                                                                                  \
  __atomic_store_n(ptr, v, memorder);                                                \
 }                                                                                          \
 EXPORT_SYMBOL(__tsan_atomic##bits##_store)

#define DEFINE_TSAN_ATOMIC_RMW(op, bits, suffix)                                                   \
 u##bits __tsan_atomic##bits##_##op(u##bits *ptr, u##bits v, int memorder);                 \
 u##bits __tsan_atomic##bits##_##op(u##bits *ptr, u##bits v, int memorder)                  \
 {                                                                                          \
  kcsan_atomic_builtin_memorder(memorder);                                           \
  if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
   check_access(ptr, bits / BITS_PER_BYTE,                                    \
         KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE |                  \
          KCSAN_ACCESS_ATOMIC, _RET_IP_);                       \
  }                                                                                  \
  return __atomic_##op##suffix(ptr, v, memorder);                                    \
 }                                                                                          \
 EXPORT_SYMBOL(__tsan_atomic##bits##_##op)

/*
 * Note: CAS operations are always classified as write, even in case they
 * fail. We cannot perform check_access() after a write, as it might lead to
 * false positives, in cases such as:
 *
 * T0: __atomic_compare_exchange_n(&p->flag, &old, 1, ...)
 *
 * T1: if (__atomic_load_n(&p->flag, ...)) {
 *  modify *p;
 *  p->flag = 0;
 *     }
 *
 * The only downside is that, if there are 3 threads, with one CAS that
 * succeeds, another CAS that fails, and an unmarked racing operation, we may
 * point at the wrong CAS as the source of the race. However, if we assume that
 * all CAS can succeed in some other execution, the data race is still valid.
 */
#define DEFINE_TSAN_ATOMIC_CMPXCHG(bits, strength, weak)                                           \
 int __tsan_atomic##bits##_compare_exchange_##strength(u##bits *ptr, u##bits *exp,          \
             u##bits val, int mo, int fail_mo);   \
 int __tsan_atomic##bits##_compare_exchange_##strength(u##bits *ptr, u##bits *exp,          \
             u##bits val, int mo, int fail_mo)    \
 {                                                                                          \
  kcsan_atomic_builtin_memorder(mo);                                                 \
  if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
   check_access(ptr, bits / BITS_PER_BYTE,                                    \
         KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE |                  \
          KCSAN_ACCESS_ATOMIC, _RET_IP_);                       \
  }                                                                                  \
  return __atomic_compare_exchange_n(ptr, exp, val, weak, mo, fail_mo);              \
 }                                                                                          \
 EXPORT_SYMBOL(__tsan_atomic##bits##_compare_exchange_##strength)

#define DEFINE_TSAN_ATOMIC_CMPXCHG_VAL(bits)                                                       \
 u##bits __tsan_atomic##bits##_compare_exchange_val(u##bits *ptr, u##bits exp, u##bits val, \
          int mo, int fail_mo);                   \
 u##bits __tsan_atomic##bits##_compare_exchange_val(u##bits *ptr, u##bits exp, u##bits val, \
          int mo, int fail_mo)                    \
 {                                                                                          \
  kcsan_atomic_builtin_memorder(mo);                                                 \
  if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
   check_access(ptr, bits / BITS_PER_BYTE,                                    \
         KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE |                  \
          KCSAN_ACCESS_ATOMIC, _RET_IP_);                       \
  }                                                                                  \
  __atomic_compare_exchange_n(ptr, &exp, val, 0, mo, fail_mo);                       \
  return exp;                                                                        \
 }                                                                                          \
 EXPORT_SYMBOL(__tsan_atomic##bits##_compare_exchange_val)

#define DEFINE_TSAN_ATOMIC_OPS(bits)                                                               \
 DEFINE_TSAN_ATOMIC_LOAD_STORE(bits);                                                       \
 DEFINE_TSAN_ATOMIC_RMW(exchange, bits, _n);                                                \
 DEFINE_TSAN_ATOMIC_RMW(fetch_add, bits, );                                                 \
 DEFINE_TSAN_ATOMIC_RMW(fetch_sub, bits, );                                                 \
 DEFINE_TSAN_ATOMIC_RMW(fetch_and, bits, );                                                 \
 DEFINE_TSAN_ATOMIC_RMW(fetch_or, bits, );                                                  \
 DEFINE_TSAN_ATOMIC_RMW(fetch_xor, bits, );                                                 \
 DEFINE_TSAN_ATOMIC_RMW(fetch_nand, bits, );                                                \
 DEFINE_TSAN_ATOMIC_CMPXCHG(bits, strong, 0);                                               \
 DEFINE_TSAN_ATOMIC_CMPXCHG(bits, weak, 1);                                                 \
 DEFINE_TSAN_ATOMIC_CMPXCHG_VAL(bits)

DEFINE_TSAN_ATOMIC_OPS(8);
DEFINE_TSAN_ATOMIC_OPS(16);
DEFINE_TSAN_ATOMIC_OPS(32);
#ifdef CONFIG_64BIT
DEFINE_TSAN_ATOMIC_OPS(64);
#endif

void __tsan_atomic_thread_fence(int memorder);
void __tsan_atomic_thread_fence(int memorder)
{
 kcsan_atomic_builtin_memorder(memorder);
 __atomic_thread_fence(memorder);
}
EXPORT_SYMBOL(__tsan_atomic_thread_fence);

/*
 * In instrumented files, we emit instrumentation for barriers by mapping the
 * kernel barriers to an __atomic_signal_fence(), which is interpreted specially
 * and otherwise has no relation to a real __atomic_signal_fence(). No known
 * kernel code uses __atomic_signal_fence().
 *
 * Since fsanitize=thread instrumentation handles __atomic_signal_fence(), which
 * are turned into calls to __tsan_atomic_signal_fence(), such instrumentation
 * can be disabled via the __no_kcsan function attribute (vs. an explicit call
 * which could not). When __no_kcsan is requested, __atomic_signal_fence()
 * generates no code.
 *
 * Note: The result of using __atomic_signal_fence() with KCSAN enabled is
 * potentially limiting the compiler's ability to reorder operations; however,
 * if barriers were instrumented with explicit calls (without LTO), the compiler
 * couldn't optimize much anyway. The result of a hypothetical architecture
 * using __atomic_signal_fence() in normal code would be KCSAN false negatives.
 */
void __tsan_atomic_signal_fence(int memorder);
noinline void __tsan_atomic_signal_fence(int memorder)
{
 switch (memorder) {
 case __KCSAN_BARRIER_TO_SIGNAL_FENCE_mb:
  __kcsan_mb();
  break;
 case __KCSAN_BARRIER_TO_SIGNAL_FENCE_wmb:
  __kcsan_wmb();
  break;
 case __KCSAN_BARRIER_TO_SIGNAL_FENCE_rmb:
  __kcsan_rmb();
  break;
 case __KCSAN_BARRIER_TO_SIGNAL_FENCE_release:
  __kcsan_release();
  break;
 default:
  break;
 }
}
EXPORT_SYMBOL(__tsan_atomic_signal_fence);

#ifdef __HAVE_ARCH_MEMSET
void *__tsan_memset(void *s, int c, size_t count);
noinline void *__tsan_memset(void *s, int c, size_t count)
{
 /*
  * Instead of not setting up watchpoints where accessed size is greater
  * than MAX_ENCODABLE_SIZE, truncate checked size to MAX_ENCODABLE_SIZE.
 */
 size_t check_len = min_t(size_t, count, MAX_ENCODABLE_SIZE);

 check_access(s, check_len, KCSAN_ACCESS_WRITE, _RET_IP_);
 return memset(s, c, count);
}
#else
void *__tsan_memset(void *s, int c, size_t count) __alias(memset);
#endif
EXPORT_SYMBOL(__tsan_memset);

#ifdef __HAVE_ARCH_MEMMOVE
void *__tsan_memmove(void *dst, const void *src, size_t len);
noinline void *__tsan_memmove(void *dst, const void *src, size_t len)
{
 size_t check_len = min_t(size_t, len, MAX_ENCODABLE_SIZE);

 check_access(dst, check_len, KCSAN_ACCESS_WRITE, _RET_IP_);
 check_access(src, check_len, 0, _RET_IP_);
 return memmove(dst, src, len);
}
#else
void *__tsan_memmove(void *dst, const void *src, size_t len) __alias(memmove);
#endif
EXPORT_SYMBOL(__tsan_memmove);

#ifdef __HAVE_ARCH_MEMCPY
void *__tsan_memcpy(void *dst, const void *src, size_t len);
noinline void *__tsan_memcpy(void *dst, const void *src, size_t len)
{
 size_t check_len = min_t(size_t, len, MAX_ENCODABLE_SIZE);

 check_access(dst, check_len, KCSAN_ACCESS_WRITE, _RET_IP_);
 check_access(src, check_len, 0, _RET_IP_);
 return memcpy(dst, src, len);
}
#else
void *__tsan_memcpy(void *dst, const void *src, size_t len) __alias(memcpy);
#endif
EXPORT_SYMBOL(__tsan_memcpy);