On 11/24/2014 09:20 PM, Mike Galbraith wrote:
On Tue, 2014-11-25 at 00:35 +0100, Thomas Gleixner wrote:
Aside of the general issues I have with this (see the inline replies
to your changelog) the overall impression of this patch is that it is
a half baken and carelessly cobbled together extract of some data base
specific kernel hackery, which I prefer not to see at all.
It culminates in a lumbering pseudo RT class of task disguised as a fair
class task. I'd expect more gain by twiddling knobs to let last buddy
do its job than the 3% mentioned.
You could perhaps create a SUPER_BATCH class that is not wakeup
preempted by any fair class task of <= priority, not only BATCH and
IDLE, but that's as nasty as this patch, though loads prettier. The
tick time thing doesn't feel right at all... if you're hurt badly by the
tick, you're likely holding the lock too long methinks.
-Mike
It is definitely not an attempt to solve any kind of RT problem. It
would be a poor attempt if it indeed attempted to solve an RT problem.
RT is all about guarantees. This patch does not help there at all and
hence I have no intention of ever applying anything like this to
SCHED_FIFO or SCHED_RR.
This problem is not caused by task holding the lock too long. It is
caused by the task happening to acquire the lock just before its current
timeslice is up. In that case, it does not matter how long the task
holds the lock for.
--
Khalid
--
To unsubscribe from this list: send the line "unsubscribe linux-api" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
