Re: critical section

[Ross Johnson <rpj@ise.canberra.edu.au>]

Eli Ofenstein wrote:
> 
> > -----Original Message-----
> > From: Scott McCaskill [ mailto:scott@magruder.org ]
> > Sent: Tuesday, July 31, 2001 5:52 PM
> > To: Ye Liu
> > Cc: pthreads-win32@sourceware.cygnus.com
> > Subject: Re: critical section
> >
> >
> > ----- Original Message -----
> > From: "Ye Liu" <yliu@tibco.com>
> > To: "Scott McCaskill" <scott@magruder.org>
> > Cc: <pthreads-win32@sourceware.cygnus.com>
> > Sent: Tuesday, July 31, 2001 5:36 PM
> > Subject: Re: critical section
> >
> >
> > > In the book of "Programming with POSIX Threads",  the
> > author metioned
> > >
> > > "You cannot lock a mutex when the calling thread already
> > has that mutex
> > locked."
> > >
> > > My previous understanding is "a mutex cannot be locked twice", which
> > obviously
> > > is wrong.

The quote from the book is describing a special case and your previous
understanding is correct generally for non-recursive mutexes.

> > >
> > > If I use a non-recursive mutex, when a thread try to lock
> > the mutex which
> > is
> > > already locked by another one, what happens to the calling
> > thread? Spin or
> > > yield?
> > >
> >
> > I don't know for sure, but I would expect it to yield.  It
> > seems like the
> > spinning that your code is doing would be purely wasteful unless the
> > spinning thread and the mutex-holding thread are on different
> > processors.
> > Can you give us a better idea of what you're trying to accomplish that
> > pthread_mutex won't do for you?
> 
> That's a big "unless" :)
> 
> In many implementations, a hybrid approach is used in which the thread spins
> for a period of time before yielding, in order to allow fast critical
> sections to complete.  Additionally, implementations are often aware of the
> CPU count of the system, and will adopt an immediate-yield policy on single
> CPU systems.
> 

The pthread_mutex_* routines follow the Open Group specification
manual page at:

(the following URL is all one line)

http://www.rtlinux.org/documents/documentation/man_pages/susv2/xsh/pthread_mutex_lock.html

Those in pthreads-win32 don't spin. However, the latest
pthreads-win32 snapshot includes the following specialised
spinlock functions:

pthread_spin_init
pthread_spin_destroy
pthread_spin_lock
pthread_spin_unlock
pthread_spin_trylock
and
PTHREAD_SPINLOCK_INITIALIZER

POSIX_SPIN_LOCKS is defined.

These are CPU affinity mask-aware and do do an immediate yield
when used in single CPU processes (even on multi-CPU systems),
spinning otherwise.

Note that, in accordance with the POSIX 1003.1j standard, these
routines don't yield after a predefined spin count. They are
intended to be used for very short duration critical sections
[on multi-CPU systems].

These routines have been tested on a single CPU system but
not yet on a multi-CPU system. If someone is able to do that
I would like to know of any problems that arise (there is
a group of programs already written for this purpose in the
"tests" directory - called spin*.c).

Hope this helps.
Ross

-- 
+-------------------------+---+
| Ross Johnson            |   | "Come down off the cross
| Management & Technology |___|  We can use the wood" - Tom Waits
| Building 11                 |
| University of Canberra      | eMail: rpj@ise.canberra.edu.au
| ACT    2601                 | WWW:  
http://public.ise.canberra.edu.au/~rpj/
| AUSTRALIA                   |
+-----------------------------+