Merge branch 'vtimer'

Heiko Carstens says:

====================

s390's virtual timer infrastructure has only a single user and the
semantics of the interface are not straightforward: a virtual timer
fires if the sum of elapsed CPU time over all CPUs exceeds a specified
value.

This is suboptimal since the number of active CPUs defines how
frequently the timer fires. With many CPUs and short intervals this can
be very frequent. Furthermore the required accounting in the vtime code
uses a single 64-bit value to accumulate elapsed CPU time. In busy
systems this can lead to cache line bouncing when the value is updated
concurrently on several CPUs; which is the reason why other accounting
mechanisms use per-CPU data structures for this purpose.

In practice this is not a problem with the appldata default interval
of 10 seconds, however the interval is configurable via sysctl with a
minimum value of 1ms.

Given that there is only a single user, emulate the virtual timer in
appldata using delayed work: the work is scheduled for the minimum
possible wall-clock time until the configured CPU-time interval elapses
(remaining CPU time / number of online CPUs), with a minimum delay of
100ms. The work then reads per-CPU statistics of all online CPUs to
check whether enough CPU time has elapsed, and if so runs the
registered callbacks.

The calculation is more expensive, but avoids potential cache line
bouncing problem. In addition it allows to remove the entire vtimer
infrastructure code with a subsequent patch.

====================

Signed-off-by: Heiko Carstens <hca@linux.ibm.com>

0 files changed, 0 insertions, 0 deletions