cvs commit: src/share/man/man9 Makefile condvar.9 lock.9 mi_switch.9 mtx_pool.9 mutex.9 rwlock.9 sleep.9 sleepqueue.9 sx.9 thread_exit.9 src/sys/kern kern_synch.c src/sys/sys mutex.h rwlock.h sleepqueue.h sx.h systm.h

[John Baldwin]

On Saturday 10 March 2007 19:11, Attilio Rao wrote:
> 2007/3/10, Pawel Jakub Dawidek <pjd at freebsd.org>:
> > On Sat, Mar 10, 2007 at 12:44:26PM +0100, Attilio Rao wrote:
> > > 2007/3/9, John Baldwin <jhb at freebsd.org>:
> > > >I don't have a date set for removing msleep(), esp. given it's wide 
use.
> > > >I would like to remove it and all the spl*() functions in 8.0 if we can
> > > >swing it.
> > > >
> > > >I also have patches to let condition variables work with rwlocks and sx
> > > >locks, but the current implementation results in an API "explosion"
> > > >since each of the cv_*wait*() functions grows a cv_*wait*_rw() version 
for
> > > >rwlocks and a cv_*waut*_sx() version for use with sx locks.  One 
possibility
> > > >would be to just cast the lock argument to (struct lock_object *) since 
all
> > > >of our locks have a lock_object as the first member, but then you use 
having
> > > >the compiler do type checking, and I'm really not willing to give up on
> > > >that.  Too easy to have evil bugs that way.  I suppose we could use 
some
> > > >evil macro that used typeof() but that would be very gcc specific?
> > > >
> > > >I guess one other possibility is to standardize on the field name for
> > > >the lock_object, calling it lo_object instead of mtx_object, rw_object,
> > > >sx_object, etc.  Anyone else have any ideas?
> > >
> > > What about adding a new function like:
> > >
> > > static __inline struct lock_object *
> > > mtx_export_lc(struct mtx *m)
> > > {
> > >
> > >        return (&m->mtx_object);
> > > }
> > >
> > > to be per-interface (so having sx_export_lc() and rw_export_lc() too)
> > > and than using in this way:
> > >
> > > static struct mtx foo_lock;
> > > static struct cv foo_cv;
> > > ...
> > >
> > > mtx_lock(&foo_lock);
> > > ...
> > > cv_wait(&foo_cv, mtx_export_lc(&foo_lock));
> > >
> > > (obviously using new struct lock_object methods you added for 
locking/unlocking)
> > >
> > > It sounds reasonable to you?
> >
> > This is ugly. If we really need to provide information about which type
> > of lock we are using, I'd probably prefer cv_wait_<locktype>().
> >
> > What about something like this:
> >
> > #define cv_wait(cv, lock)       do {
> >         switch (LO_CLASSINDEX((struct lock_object *)(lock))) {
> >         case 1:
> >                 cv_wait_mtx(cv, lock);
> >                 break;
> >         case 2:
> >                 cv_wait_sx(cv, lock);
> >                 break;
> >         case 3:
> >                 cv_wait_rw(cv, lock);
> >                 break;
> >         default:
> >                 panic("Invalid lock.");
> >         }
> > } while (0)
> 
> This is exactly what John is trying to avoid.
> You have however to export cv_wait_*() & friends in the public
> namespace and at this point you don't need such wrapper.
> 
> I know it is not so elegant, but the other solutions are uglier.
> Having a function returning the lock object per-primitive is the most
> suitable, IMHO.

No, that's more typing than _rw and _sx.  Here is what I want to happen if 
possible:

cv_wait(cv, mtx);

cv_wait(cv, rw);

cv_wait(cv, sx);

and have the the compiler figure it out.  Basically, trying to shoehorn some 
C++ into C since mtx, rw, and sx are sub-classes of 'lock_object'. :)  That 
is, I'd like it to do something like this:

#define	cv_wait(cv, lock) do { 				\
	if (typeof(lock) == (struct mtx *))		\
		_cv_wait(cv, &lock->mtx_object);	\
	else if (typeof(lock) == (struct rwlock *))	\
		_cv_wait(cv, &lock->rw_object);		\
	else if (typeof(lock) == (struct sx *))		\
		_cv_wait(cv, &lock->sx_object);		\
	else						\
		compile_error;				\
} while (0)

So you still get type checking, etc.  I'm thinking maybe the simplest thing to 
do is to rename 'mtx_object', 'rw_object', and 'sx_object' fields to all 
be 'lock_object' and then do this:

#define	cv_wait(cv, lock)	_cv_wait((cv), &(lock)->lock_object)

-- 
John Baldwin