Re: [PATCH 5/5] kcsan: Introduce ASSERT_EXCLUSIVE_BITS(var, mask)

[John Hubbard]

On 2/10/20 10:43 AM, Marco Elver wrote:
> This introduces ASSERT_EXCLUSIVE_BITS(var, mask).
> ASSERT_EXCLUSIVE_BITS(var, mask) will cause KCSAN to assume that the
> following access is safe w.r.t. data races (however, please see the
> docbook comment for disclaimer here).
> 
> For more context on why this was considered necessary, please see:
>   http://lkml.kernel.org/r/1580995070-25139-1-git-send-email-cai@xxxxxx
> 
> In particular, data races between reads (that use @mask bits of an
> access that should not be modified concurrently) and writes (that change
> ~@mask bits not used by the read) should ordinarily be marked. After
> marking these, we would no longer be able to detect harmful races
> between reads to @mask bits and writes to @mask bits.

I know this is "just" the commit log, but as long as I'm reviewing the
whole thing...to make the above a little clearer, see if you like this 
revised wording:

In particular, before this patch, data races between reads (that use
@mask bits of an access that should not be modified concurrently) and
writes (that change ~@mask bits not used by the readers) would have
been annotated with "data_race()". However, doing so would then hide
real problems: we would no longer be able to detect harmful races
between reads to @mask bits and writes to @mask bits.


> 
> Therefore, by using ASSERT_EXCLUSIVE_BITS(var, mask), we accomplish:
> 
>   1. No new macros introduced elsewhere; since there are numerous ways in
>      which we can extract the same bits, a one-size-fits-all macro is
>      less preferred.

This somehow confuses me a lot. Maybe say it like this:

1. Avoid a proliferation of specific macros at the call sites: by including a
   mask in the argument list, we can use the same macro in a wide variety of 
   call sites, regardless of which bits in a field each call site uses.

?

> 
>   2. The existing code does not need to be modified (although READ_ONCE()
>      may still be advisable if we cannot prove that the data race is
>      always safe).
> 
>   3. We catch bugs where the exclusive bits are modified concurrently.
> 
>   4. We document properties of the current code.
> 
> Signed-off-by: Marco Elver <elver@xxxxxxxxxx>
> Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> Cc: David Hildenbrand <david@xxxxxxxxxx>
> Cc: Jan Kara <jack@xxxxxxx>
> Cc: John Hubbard <jhubbard@xxxxxxxxxx>
> Cc: Paul E. McKenney <paulmck@xxxxxxxxxx>
> Cc: Qian Cai <cai@xxxxxx>
> ---
>  include/linux/kcsan-checks.h | 57 ++++++++++++++++++++++++++++++++----
>  kernel/kcsan/debugfs.c       | 15 +++++++++-
>  2 files changed, 65 insertions(+), 7 deletions(-)
> 
> diff --git a/include/linux/kcsan-checks.h b/include/linux/kcsan-checks.h
> index 4ef5233ff3f04..eae6030cd4348 100644
> --- a/include/linux/kcsan-checks.h
> +++ b/include/linux/kcsan-checks.h
> @@ -152,9 +152,9 @@ static inline void kcsan_check_access(const volatile void *ptr, size_t size,
>  #endif
>  
>  /**
> - * ASSERT_EXCLUSIVE_WRITER - assert no other threads are writing @var
> + * ASSERT_EXCLUSIVE_WRITER - assert no concurrent writes to @var
>   *
> - * Assert that there are no other threads writing @var; other readers are
> + * Assert that there are no concurrent writes to @var; other readers are
>   * allowed. This assertion can be used to specify properties of concurrent code,
>   * where violation cannot be detected as a normal data race.
>   *
> @@ -171,11 +171,11 @@ static inline void kcsan_check_access(const volatile void *ptr, size_t size,
>  	__kcsan_check_access(&(var), sizeof(var), KCSAN_ACCESS_ASSERT)
>  
>  /**
> - * ASSERT_EXCLUSIVE_ACCESS - assert no other threads are accessing @var
> + * ASSERT_EXCLUSIVE_ACCESS - assert no concurrent accesses to @var
>   *
> - * Assert that no other thread is accessing @var (no readers nor writers). This
> - * assertion can be used to specify properties of concurrent code, where
> - * violation cannot be detected as a normal data race.
> + * Assert that there are no concurrent accesses to @var (no readers nor
> + * writers). This assertion can be used to specify properties of concurrent
> + * code, where violation cannot be detected as a normal data race.
>   *
>   * For example, in a reference-counting algorithm where exclusive access is
>   * expected after the refcount reaches 0. We can check that this property
> @@ -191,4 +191,49 @@ static inline void kcsan_check_access(const volatile void *ptr, size_t size,
>  #define ASSERT_EXCLUSIVE_ACCESS(var)                                           \
>  	__kcsan_check_access(&(var), sizeof(var), KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT)
>  
> +/**
> + * ASSERT_EXCLUSIVE_BITS - assert no concurrent writes to subset of bits in @var
> + *
> + * [Bit-granular variant of ASSERT_EXCLUSIVE_WRITER(var)]


No need for the square brackets, unless that's some emerging convention in the
documentation world. 


> + *
> + * Assert that there are no concurrent writes to a subset of bits in @var;
> + * concurrent readers are permitted. Concurrent writes (or reads) to ~@mask bits
> + * are ignored. This assertion can be used to specify properties of concurrent
> + * code, where marked accesses imply violations cannot be detected as a normal
> + * data race.


How about this wording:

/*
 * Assert that there are no concurrent writes to a subset of bits in @var;
 * concurrent readers are permitted. Concurrent writes (or reads) to ~@mask bits
 * are ignored. This assertion provides more detailed, bit-level information to
 * the KCSAN system than most of the other (word granularity) annotations. As
 * such, it allows KCSAN to safely overlook some bits while still continuing to
 * check the remaining bits for unsafe access patterns.
 *
 * Use this if you have some bits that are read-only, and other bits that are
 * not, within a variable.
 */

?


> + *
> + * For example, this may be used when certain bits of @var may only be modified
> + * when holding the appropriate lock, but other bits may still be modified
> + * concurrently. Writers, where other bits may change concurrently, could use
> + * the assertion as follows:
> + *
> + *	spin_lock(&foo_lock);
> + *	ASSERT_EXCLUSIVE_BITS(flags, FOO_MASK);
> + *	old_flags = READ_ONCE(flags);
> + *	new_flags = (old_flags & ~FOO_MASK) | (new_foo << FOO_SHIFT);
> + *	if (cmpxchg(&flags, old_flags, new_flags) != old_flags) { ... }
> + *	spin_unlock(&foo_lock);
> + *
> + * Readers, could use it as follows:
> + *
> + *	ASSERT_EXCLUSIVE_BITS(flags, FOO_MASK);
> + *	foo = (READ_ONCE(flags) & FOO_MASK) >> FOO_SHIFT;


In the general case (which is what this documentation covers), the
READ_ONCE() is not required. So this should either leave it out, or
explain that it's not necessarily required.


> + *
> + * NOTE: The access that immediately follows is assumed to access the masked
> + * bits only, and safe w.r.t. data races. While marking this access is optional
> + * from KCSAN's point-of-view, it may still be advisable to do so, since we
> + * cannot reason about all possible compiler optimizations when it comes to bit
> + * manipulations (on the reader and writer side).
> + *
> + * @var variable to assert on
> + * @mask only check for modifications to bits set in @mask
> + */
> +#define ASSERT_EXCLUSIVE_BITS(var, mask)                                       \


This API looks good to me.


> +	do {                                                                   \
> +		kcsan_set_access_mask(mask);                                   \
> +		__kcsan_check_access(&(var), sizeof(var), KCSAN_ACCESS_ASSERT);\
> +		kcsan_set_access_mask(0);                                      \
> +		kcsan_atomic_next(1);                                          \
> +	} while (0)
> +
>  #endif /* _LINUX_KCSAN_CHECKS_H */
> diff --git a/kernel/kcsan/debugfs.c b/kernel/kcsan/debugfs.c
> index 9bbba0e57c9b3..2ff1961239778 100644
> --- a/kernel/kcsan/debugfs.c
> +++ b/kernel/kcsan/debugfs.c
> @@ -100,8 +100,10 @@ static noinline void microbenchmark(unsigned long iters)
>   * debugfs file from multiple tasks to generate real conflicts and show reports.
>   */
>  static long test_dummy;
> +static long test_flags;
>  static noinline void test_thread(unsigned long iters)
>  {
> +	const long CHANGE_BITS = 0xff00ff00ff00ff00L;
>  	const struct kcsan_ctx ctx_save = current->kcsan_ctx;
>  	cycles_t cycles;
>  
> @@ -109,16 +111,27 @@ static noinline void test_thread(unsigned long iters)
>  	memset(&current->kcsan_ctx, 0, sizeof(current->kcsan_ctx));
>  
>  	pr_info("KCSAN: %s begin | iters: %lu\n", __func__, iters);
> +	pr_info("test_dummy@%px, test_flags@%px\n", &test_dummy, &test_flags);
>  
>  	cycles = get_cycles();
>  	while (iters--) {
> +		/* These all should generate reports. */
>  		__kcsan_check_read(&test_dummy, sizeof(test_dummy));
> -		__kcsan_check_write(&test_dummy, sizeof(test_dummy));
>  		ASSERT_EXCLUSIVE_WRITER(test_dummy);
>  		ASSERT_EXCLUSIVE_ACCESS(test_dummy);
>  
> +		ASSERT_EXCLUSIVE_BITS(test_flags, ~CHANGE_BITS); /* no report */
> +		__kcsan_check_read(&test_flags, sizeof(test_flags)); /* no report */
> +
> +		ASSERT_EXCLUSIVE_BITS(test_flags, CHANGE_BITS); /* report */
> +		__kcsan_check_read(&test_flags, sizeof(test_flags)); /* no report */
> +
>  		/* not actually instrumented */
>  		WRITE_ONCE(test_dummy, iters);  /* to observe value-change */
> +		__kcsan_check_write(&test_dummy, sizeof(test_dummy));
> +
> +		test_flags ^= CHANGE_BITS; /* generate value-change */
> +		__kcsan_check_write(&test_flags, sizeof(test_flags));
>  	}
>  	cycles = get_cycles() - cycles;
>  
> 



thanks,
-- 
John Hubbard
NVIDIA