Re: FreeBSD GSOC proposal in 2014

On Friday, March 14, 2014 3:02:18 am Wojciech A. Koszek wrote:
> On Thu, Mar 13, 2014 at 09:56:35PM -0400, yan cui wrote:
> > Hi all,
> > 
> >     I write this mail to make my question clear. I know witness can be 
used
> > to detect wrong lock order in the kernel. However, can it be used to do
> > lock profiling (what I mean is to report the information such as which
> > locks are most contended and print some related statistics such as calling
> > graph, etc)?
> > In other words, is it enough to finish the task by porting witness to the
> > pthread library?
> >
> 
> Yan,
> 
> To my knowledge WITNESS is the only tool for lock order verification.
> 
> For lock profiling in the FreeBSD kernel there's a KTR subsystem. KTR
> mechanism is basically like syslog() in the user-space, but for the kernel.
> KTR subsystem will receive messages from KTR API that is placed in the
> FreeBSD kernel. Messages get stored on the list of some sort. List can be
> exported to a file. File you can later analyze. 
> 
> Jeff wrote a Python app which can be used for pre-processing the KTR logs
> from scheduler and protting them visually. Link:
> 
> http://svnweb.freebsd.org/base/head/tools/sched/schedgraph.py
> 
> Instead of porting witness to pthreads, maybe we could evaluate expanding
> WITNESS to cover kern_umtx? This could prove to be more universal.
> 
> Wojciech

There is a dedicated lock profiler (LOCK_PROFILING) in the kernel.  A
previous GSoC student from an earlier year has already re-implemented both
LOCK_PROFILING and WITNESS for pthreads.

-- 
John Baldwin