[$] The balance between features and performance in the block layer

Back in September, LWN reported on a series
of block-layer optimizations that enabled a suitably equipped system to
sustain 3.5 million I/O operations per second (IOPS). That
optimization work has continued since then, and those 3.5 million IOPS
would be a deeply disappointing result now. A recent disagreement over the
addition of a new feature has highlighted the potential cost of a heavily
optimized block layer, though; when is a feature deemed important enough to
outweigh the drive for maximum performance?