review request for 6673124

[Xiaobin Lu]

Webrev: http://webrev.invokedynamic.info/xiaobin.lu/6673124/webrev/

Details:

Hotspot needs to compute the number of available CPUs available to the 
java process in order to determine which flavor of VM to be used or how 
many GC threads needs to created for parallel collector. Also, the user 
could call Runtime.availableProcessors() and use that result to 
determine how many threads needs to be created for their applications, etc.

Here is the current way to compute that number on Solaris where the 
problem might be seen. First, we check whether the process is bound to a 
existing processor set, if yes, we then check how many processors inside 
that processor set and use that number as the available CPUs. If not, we 
just return the number of total online CPUs as the number of available 
CPUs. This ignores the fact that even the process is NOT running inside 
a processor set, however, if there is any processor set around, the CPUs 
in that processor set won't be available to the java process. So the 
correct way of doing this is to ignore whether the process is bound to a 
processor set or not, as long as "pset_bind" returns successfully, we 
then use "pset_info" to check the available CPUs to us.

We decided not to invent another flag to fall back to the old behavior 
which is absolutely wrong. Plus it is hard to come up with a sensible 
name for that flag, not to mention the user has to dig through 
globals.hpp to find out that flag to use.

Reviewed by:

Verified by:
Experiments made on multi-core Solaris machines and trace put in the 
code to verify the return value is correct.
JPRT.

Thanks,
-Xiaobin