Samuel Thibault, le Wed 13 Feb 2008 13:59:59 +0000, a écrit :
> stubdom: optimize block io completion polling
> by not polling all the time but just when some requests have completed.
Oops, sorry, although it never happened during my tests, I'm pretty sure
there is a race, please use this patch instead:
stubdom: optimize block io completion polling
by not polling all the time but just when some requests have completed.
Signed-off-by: Samuel Thibault <samuel.thibault@xxxxxxxxxxxxx>
diff -r e85399173769 extras/mini-os/blkfront.c
--- a/extras/mini-os/blkfront.c Tue Feb 12 16:59:08 2008 +0000
+++ b/extras/mini-os/blkfront.c Wed Feb 13 15:55:47 2008 +0000
@@ -327,6 +327,11 @@
struct blkif_response *rsp;
moretodo:
+#ifdef HAVE_LIBC
+ files[dev->fd].read = 0;
+ mb(); /* Make sure to let the handler set read to 1 before we start
looking at the ring */
+#endif
+
rp = dev->ring.sring->rsp_prod;
rmb(); /* Ensure we see queued responses up to 'rp'. */
cons = dev->ring.rsp_cons;
--- a/tools/ioemu/block-vbd.c Tue Feb 12 16:59:08 2008 +0000
+++ b/tools/ioemu/block-vbd.c Wed Feb 13 15:55:47 2008 +0000
@@ -67,6 +67,12 @@
return 100;
}
+static void vbd_io_completed(void *opaque)
+{
+ BDRVVbdState *s = opaque;
+ blkfront_aio_poll(s->dev);
+}
+
static int vbd_open(BlockDriverState *bs, const char *filename, int flags)
{
BDRVVbdState *s = bs->opaque;
@@ -85,6 +91,7 @@
}
s->fd = blkfront_open(s->dev);
+ qemu_set_fd_handler(s->fd, vbd_io_completed, NULL, s);
QEMU_LIST_INSERT_HEAD(&vbds, s, list);
@@ -102,9 +109,6 @@
void qemu_aio_poll(void)
{
- BDRVVbdState *s;
- for (s = vbds.lh_first; s; s = s->list.le_next)
- blkfront_aio_poll(s->dev);
}
/* Wait for all IO requests to complete. */
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel
|
Home