On Thu, Sep 18, 2003 at 05:05:42PM +0300, Jarkko Tolvanen wrote:
> How stable/usable is LWP implementation on top of pthreads?

Whatever is in CVS is actually stable enough to run both venus and
codasrv successfully. It really was the only way that I could use
valgrind to find some memory problems. Valgrind doesn't like LWP's
under the cover stack switching tricks.

There are three reasons why we're not really using the pthread
implementation.

- I've discovered that some versions of the LinuxThreads pthread
  implementation have serious flaws.

  f.i. I have reason to believe that on a RedHat-5.1 machine, a critical
  region protected by a mutex was entered by at least 2 threads at the
  same time. Now RedHat-5 is clearly a very old release, but these kinds
  of race conditions are very subtle and hard to diagnose. There are
  enough problems that still need to be found and resolved in Coda
  already.

- The pthread version has a significantly higher context switch overhead
  making it somewhat slower, and only the RVM truncate thread is really
  taking advantage of being able to run concurrently.

- At the time I started it, there was no pthread implementation for
  Windows (cygwin), which meant that we still had to maintain/use the
  old LWP implementation either way.

> I found a test program in lwp-1.10/src (testlwp.c), but that didn't work
> when I tried to run it on my Linux PC with the pthreaded LWP library
> compiled from lwp-1.10/src-pt (the main thread is blocked on a condition
> variable and the child thread fails to release it but instead runs a
> simple loop forever). The lwp-pt directory didn't seem to have any test
> programs. Is there any other working test programs?

I used the test programs from the lwp/src directory myself (and ofcourse
venus/codasrv).

> (I updated LWP 1.10 with some pthreaded LWP changes I could found from the
> CVS repository with viewcvs)

Yeah, some of those changes in CVS are definitely required to run with
venus.

Jan