Re: [patch] convert aio event reap to use atomic-op instead of spin_lock

["Ken Chen" <kenchen@google.com>]

On 4/12/07, Ken Chen <kenchen@google.com> wrote:
> On 4/12/07, Jeff Moyer <jmoyer@redhat.com> wrote:
> > I didn't see any response to Zach's request for code that actually
> > tests out the shared ring buffer.  Do you have such code?
>
> Yes, I do.  I was stress testing the code since last night.  After 20+
> hours of stress run with fio and aio-stress, now I'm posting it with
> confidence.
>
> I modified libaio's io_getevents to take advantage of new user level
> reap function. The feature is exported out via ring->compat_features.
> btw, is compat_feature suppose to be a version number or a bit mask?
> I think bitmask make more sense and more flexible.

Additional patch on the kernel side to export the new features.  On
top of patch posted at:
http://marc.info/?l=linux-kernel&m=117636401818057&w=2

--- a/include/linux/aio.h
+++ b/include/linux/aio.h
@@ -138,8 +138,11 @@ #define init_sync_kiocb(x, filp)			\
 		init_wait((&(x)->ki_wait));             \
 	} while (0)

+#define AIO_RING_BASE		1
+#define AIO_RING_USER_REAP	2
+
 #define AIO_RING_MAGIC			0xa10a10a1
-#define AIO_RING_COMPAT_FEATURES	1
+#define AIO_RING_COMPAT_FEATURES	(AIO_RING_BASE | AIO_RING_USER_REAP)
 #define AIO_RING_INCOMPAT_FEATURES	0
 struct aio_ring {
 	unsigned	id;	/* kernel internal index number */