On Thu, 2012-01-05 at 20:38 -0800, Tejun Heo wrote:

Hello,

On Thu, Jan 5, 2012 at 8:19 PM, Shaohua Li [off-list ref] wrote:

quoted

quoted

For detailed discussion of the bug:

 http://thread.gmane.org/gmane.linux.kernel.next/20064/focus=20159

this is overkill. when plug is added, we found huge performance
regression, that's why we add INSERT_SORT_MERGE.

With what workload? I suspect most of the improvements is from merging
across different cfqqs, no? The whole recursive thing can't be very
useful if cross-cfqq merging isn't allowed. Maybe there are specific
cases where last_merge hint merging can be specially effective. I
don't know. Regardless, this is an apparent bug and the block tree
will be pushed mainline pretty soon. If necessary, fix it better
later. Doing complex things inside merge window usually isn't a good
idea.

it's not related to the recursive merge. I forgot the detail of the
workload why we add INSERT_SORT_MERGE, it's added several months ago
anyway.
Obviously we shouldn't do complex things inside the merge window. I just
didn't agree to disable INSERT_SORT_MERGE, which will bring performance
regression for sure. Can we just change CFQ like your first path?

Thanks,
Shaohua