On Fri, May 15, 2020 at 01:26:49PM -0400, Brian Foster wrote:
> Ok, that might prohibit using a bitop in the commit path. I'd still like
> to see actual numbers on that, though, just to see where on the spectrum
> it lands. I'm also wondering if the fast path logic mentioned above
> could be implemented like the following (using bitops instead of the
> spinlock):
>
> if (test_bit(XLOG_CIL_EMPTY, ...) &&
> test_and_clear_bit(XLOG_CIL_EMPTY, ...)) {
> <steal CIL res>
> }
>
> That type of pattern seems to be used in at least a few other places in
> the kernel (e.g. filemap_check_errors(), wb_start_writeback(),
> __blk_mq_tag_busy()), presumably for similar reasons.
Ok, that seems reasonable given that there is other code using the
same pattern to avoid atomic ops. Overhead will be no different to
the test/lock/retest pattern I've been using...
Cheers,
Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
