Re: [HACK] convert i_alloc_sem for direct_io.c craziness!

Hi Zach!

Thanks for the responses.

--
On Tue, 2 Oct 2007, Zach Brown wrote:

On Oct 1, 2007, at 1:39 PM, Peter Zijlstra wrote:

quoted

On Mon, 2007-10-01 at 12:52 -0700, Zach Brown wrote:

quoted

Do you have any suggestions for locking constructs that RT would
prefer?

Basically, anything that maps to a simple mutex. Anything more complex
gets real messy real quick.

I'm worried that the aio+dio implementation of concurrent pending IOs
just doesn't map well to PI.

Would a hack with a mutex and counts help at all?  It seems like it
would still have the same problem.  The count increments don't
transfer ownership to the count decrements and the wake up.

io submission from tasks:
	down(&inode->i_mutex);
	atomic_inc(&inode->in_flight);
	up(&inode->i_mutex);

io completion from interrupts:
	if(atomic_dec_and_test(&inode->in_flight))
		wake_up(&inode->waiting);

file allocation in tasks:
	down(&inode->i_mutex);
	wait_event(inode->waiting, atomic_read(&inode->in_flight) == 0);
	up(&inode->i_mutex);

(yeah, yeah, starvation -- it's just a demonstration)

This looks more like a completion. Actually, completions don't have PI
either, but they are usually OK.

In any case, this seems like it's not a very high priority now that
RT has Steven's work-around.  If it does become a priority can you
guys let linux-fsdevel know?

Actually, it may still be a high priority. We have one BUG currently that
seems to trigger starvation. But we don't know yet if it's a bug in the
test or in the FS code yet.  I don't know the full details yet, but we do
have someone currently looking at it. These tests are what found this
issue in the first place.

When more info becomes available, I will definitely CC the linux-fsdevel
list. Thanks for letting me know about it. (I usually get confused by all
the different lists that are out there).

-- Steve