[PATCH v2 00/35] bitops: add atomic find_bit() operations

Andy Shevchenko andriy.shevchenko at linux.intel.com
Mon Dec 4 05:07:04 PST 2023

On Sun, Dec 03, 2023 at 11:23:47AM -0800, Yury Norov wrote:
> Add helpers around test_and_{set,clear}_bit() that allow to search for
> clear or set bits and flip them atomically.
> The target patterns may look like this:
> 	for (idx = 0; idx < nbits; idx++)
> 		if (test_and_clear_bit(idx, bitmap))
> 			do_something(idx);
> Or like this:
> 	do {
> 		bit = find_first_bit(bitmap, nbits);
> 		if (bit >= nbits)
> 			return nbits;
> 	} while (!test_and_clear_bit(bit, bitmap));
> 	return bit;
> In both cases, the opencoded loop may be converted to a single function
> or iterator call. Correspondingly:
> 	for_each_test_and_clear_bit(idx, bitmap, nbits)
> 		do_something(idx);
> Or:
> 	return find_and_clear_bit(bitmap, nbits);
> Obviously, the less routine code people have to write themself, the
> less probability to make a mistake.
> Those are not only handy helpers but also resolve a non-trivial
> issue of using non-atomic find_bit() together with atomic
> test_and_{set,clear)_bit().
> The trick is that find_bit() implies that the bitmap is a regular
> non-volatile piece of memory, and compiler is allowed to use such
> optimization techniques like re-fetching memory instead of caching it.
> For example, find_first_bit() is implemented like this:
>       for (idx = 0; idx * BITS_PER_LONG < sz; idx++) {
>               val = addr[idx];
>               if (val) {
>                       sz = min(idx * BITS_PER_LONG + __ffs(val), sz);
>                       break;
>               }
>       }
> On register-memory architectures, like x86, compiler may decide to
> access memory twice - first time to compare against 0, and second time
> to fetch its value to pass it to __ffs().
> When running find_first_bit() on volatile memory, the memory may get
> changed in-between, and for instance, it may lead to passing 0 to
> __ffs(), which is undefined. This is a potentially dangerous call.
> find_and_clear_bit() as a wrapper around test_and_clear_bit()
> naturally treats underlying bitmap as a volatile memory and prevents
> compiler from such optimizations.
> Now that KCSAN is catching exactly this type of situations and warns on
> undercover memory modifications. We can use it to reveal improper usage
> of find_bit(), and convert it to atomic find_and_*_bit() as appropriate.
> The 1st patch of the series adds the following atomic primitives:
> 	find_and_set_bit(addr, nbits);
> 	find_and_set_next_bit(addr, nbits, start);
> 	...
> Here find_and_{set,clear} part refers to the corresponding
> test_and_{set,clear}_bit function. Suffixes like _wrap or _lock
> derive their semantics from corresponding find() or test() functions.
> For brevity, the naming omits the fact that we search for zero bit in
> find_and_set, and correspondingly search for set bit in find_and_clear
> functions.
> The patch also adds iterators with atomic semantics, like
> for_each_test_and_set_bit(). Here, the naming rule is to simply prefix
> corresponding atomic operation with 'for_each'.
> This series is a result of discussion [1]. All find_bit() functions imply
> exclusive access to the bitmaps. However, KCSAN reports quite a number
> of warnings related to find_bit() API. Some of them are not pointing
> to real bugs because in many situations people intentionally allow
> concurrent bitmap operations.
> If so, find_bit() can be annotated such that KCSAN will ignore it:
>         bit = data_race(find_first_bit(bitmap, nbits));
> This series addresses the other important case where people really need
> atomic find ops. As the following patches show, the resulting code
> looks safer and more verbose comparing to opencoded loops followed by
> atomic bit flips.
> In [1] Mirsad reported 2% slowdown in a single-thread search test when
> switching find_bit() function to treat bitmaps as volatile arrays. On
> the other hand, kernel robot in the same thread reported +3.7% to the
> performance of will-it-scale.per_thread_ops test.
> Assuming that our compilers are sane and generate better code against
> properly annotated data, the above discrepancy doesn't look weird. When
> running on non-volatile bitmaps, plain find_bit() outperforms atomic
> find_and_bit(), and vice-versa.


In some cases the better improvements can be achieved by switching
the (very) old code to utilise IDA framework.

With Best Regards,
Andy Shevchenko

More information about the ath10k mailing list