Hello,
one of our customers reported a noticeable performance regression with his
raid0 setup after updating to newer kernels. After investigation of the block
traces it was obvious that raid0 code generates very suboptimal IO pattern for
large requests that are not properly aligned to the raid chunk boundary (see
patch 2 for detailed description). The regression has been introduced already 6
years ago by commit f00d7c85be9e ("md/raid0: fix up bio splitting.") but likely
because it requires multimegabyte requests (only possible after multipage bvec
work) which are not aligned to the chunk boundary (which is usually possible
only if the filesystems are created with suboptimal parameters) it went
unnoticed for a long time. These patches fix the regression.
Honza
