Re: [RFC PATCH 2.6.17-mm1 4/3] ieee1394: convert ieee1394_transactions from semaphores to waitqueue

[Stefan Richter <stefanr@s5r6.in-berlin.de>]

Arjan van de Ven wrote:
> On Sat, 2006-06-24 at 19:45 +0200, Stefan Richter wrote:
>>following semaphores remain in the ieee1394 subsystem:
>>
>>highlevel.c:  hl_drivers_sem    (RW semaphore)
>>nodemgr.c:    subsys.rwsem      (driver core's RW semaphores)
>>raw1394.c:    fi->complete_sem  (completion semaphore)

> can this last one move to an actual completion? That would get rid of it
> nicely ;)

Hmm. There are dozens of points in raw1394 which call 
__queue_complete_req() which up()s the semaphore. ("fi" is the 
private_data of a struct file. Multiple outstanding requests may be 
associated with a file.) Then there are two places which wait on the 
semaphore:

raw1394_read(), called when somebody reads /dev/raw1394:
	if (file->f_flags & O_NONBLOCK) {
		if (down_trylock(&fi->complete_sem))
			return -EAGAIN;
	} else {
		if (down_interruptible(&fi->complete_sem))
			return -ERESTARTSYS;
	}

raw1394_release(), called when somebody closes (releases) /dev/raw1394:
	done = 0;
	while (!done) {
		/* free all complete requests */
		/* set "done" if there are no more pending requests */
		if (!done)
			down_interruptible(&fi->complete_sem);
	}
	/* cleanup, free fi */

It looks like fi->complete_sem is a actually a counting semaphore. It 
could perhaps be replaced by a wait queue plus an atomic counter. There 
is even already a wait queue in "fi" for use with poll_wait() via 
raw1394_poll().
-- 
Stefan Richter
-=====-=-==- -==- ==---
http://arcgraph.de/sr/