On Wed, Feb 27, 2008 at 11:42 AM, Heiko Carstens <[EMAIL PROTECTED]> wrote:
> Oh yes, you're absolutely right. Linux issues the ESSA instruction once to > figure out if it will work or not (program check or not), just to tell > later if the kernel paramater cmma=... makes any sense. > So there's no way z/VM could know if Linux is actually using cmma or if it > was just a test. I have no idea what's possible with an undocumented instruction ;-) If nothing else, we can get a CP PTF to only enable it on the 2nd ESSA instruction... IMHO the Linux kernel option to enable / disable the feature is pure evil. Assuming that the ESSA will give a program check when system is not able or when CP says the virtual machine should not, then that should be enough to enable or disable the kernel doing ESSA instructions. > This can be done better by e.g. doing the test later after the kernel > parameter has been parsed. On the other side I'm quite sure there must > be a feature flag somewhere which indicates if the ESSA instruction > works or not. Testing feature flags should be the method to go anyway. Over lunch, Pieter and myself concluded how fortunate we are with the picked defaults for page state bits. With a less defensive design, the mismatch between the two settings would have resulted in data integrity problems. We concluded that there can be no benefit in having the two settings different, and I recommend people don't do that. Current design defaults in a mismatch in relevant configurations, so folks should be aware action is required when they start play with SLES10. >From our measurements with the feature properly enabled, I noticed that CP does not react as I expected it would (should) have done. In real life customers will be running a mix of Linux servers where some are cmma capable and some are not. I expect that might get a well tuned system out-of-tune. Until I have a better understanding of CMMA, I would not enable it on my system. Considering that it takes a virtual machine reset to disable the feature, it might be wise for people to disable it at system level until we know more. Rob PS It would also be interesting to hear which configurations were compared in the measurements that IBM published. If CMMA was disabled with the evil kernel option, then we may be comparing different things. -- Rob van der Heij Velocity Software, Inc http://velocitysoftware.com/ ---------------------------------------------------------------------- For LINUX-390 subscribe / signoff / archive access instructions, send email to [EMAIL PROTECTED] with the message: INFO LINUX-390 or visit http://www.marist.edu/htbin/wlvindex?LINUX-390
