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

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

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.

(warning: some lines are extremely long in the patch and my email
client will probably mangle it badly).


diff -Nurp libaio-0.3.104/src/io_getevents.c
libaio-0.3.104-new/src/io_getevents.c
--- libaio-0.3.104/src/io_getevents.c	2003-06-18 12:58:21.000000000 -0700
+++ libaio-0.3.104-new/src/io_getevents.c	2007-04-12 17:35:06.000000000 -0700
@@ -21,10 +21,13 @@
#include <stdlib.h>
#include <time.h>
#include "syscall.h"
+#include <asm/system.h>

io_syscall5(int, __io_getevents_0_4, io_getevents, io_context_t, ctx,
long, min_nr, long, nr, struct io_event *, events, struct timespec *,
timeout)

#define AIO_RING_MAGIC                  0xa10a10a1
+#define AIO_RING_BASE			1
+#define AIO_RING_USER_REAP		2

/* Ben will hate me for this */
struct aio_ring {
@@ -41,7 +44,11 @@ struct aio_ring {

int io_getevents_0_4(io_context_t ctx, long min_nr, long nr, struct
io_event * events, struct timespec * timeout)
{
+	long i = 0, ret;
+	unsigned head;
+	struct io_event *evt_base;
	struct aio_ring *ring;
+
	ring = (struct aio_ring*)ctx;
	if (ring==NULL || ring->magic != AIO_RING_MAGIC)
		goto do_syscall;
@@ -49,9 +56,35 @@ int io_getevents_0_4(io_context_t ctx, l
		if (ring->head == ring->tail)
			return 0;
	}
-	
+
+	if (!(ring->compat_features & AIO_RING_USER_REAP))
+		goto do_syscall;
+
+	if (min_nr > nr || min_nr < 0 || nr < 0)
+		return -EINVAL;
+
+	evt_base = (struct io_event *) (ring + 1);
+	while (i < nr) {
+		head = ring->head;
+		if (head == ring->tail)
+			break;
+
+		*events = evt_base[head & (ring->nr - 1)];
+		if (head == cmpxchg(&ring->head, head, head + 1)) {
+			events++;
+			i++;
+		}
+	}
+
+	if (i >= min_nr)
+		return i;
+
do_syscall:	
-	return __io_getevents_0_4(ctx, min_nr, nr, events, timeout);
+	ret = __io_getevents_0_4(ctx, min_nr - i, nr - i, events, timeout);
+	if (ret >= 0)
+		return i + ret;
+	else
+		return i ? i : ret;
}

DEFSYMVER(io_getevents_0_4, io_getevents, 0.4)
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/