From: Rusty Russell on
On Sun, 9 May 2010 06:27:33 pm Michael S. Tsirkin wrote:
> On Fri, May 07, 2010 at 12:35:39PM +0930, Rusty Russell wrote:
> > Then there's padding to page boundary. That puts us on a cacheline again
> > for the used ring; also 2 bytes per entry.
> >
> Hmm, is used ring really 2 bytes per entry?

Err, no, I am an idiot.

> /* u32 is used here for ids for padding reasons. */
> struct vring_used_elem {
> /* Index of start of used descriptor chain. */
> __u32 id;
> /* Total length of the descriptor chain which was used (written to) */
> __u32 len;
> };
> struct vring_used {
> __u16 flags;
> __u16 idx;
> struct vring_used_elem ring[];
> };

OK, now I get it. Sorry, I was focussed on the avail ring.

> I thought that used ring has 8 bytes per entry, and that struct
> vring_used is aligned at page boundary, this
> would mean that ring element is at offset 4 bytes from page boundary.
> Thus with cacheline size 128 bytes, each 4th element crosses
> a cacheline boundary. If we had a 4 byte padding after idx, each
> used element would always be completely within a single cacheline.

I think the numbers are: every 16th entry hits two cachelines. So currently
the first 15 entries are "free" (assuming we hit the idx cacheline anyway),
then 1 in 16 cost 2 cachelines. That makes the aligned version win when
N > 240.

But, we access the array linearly. So the extra cacheline cost is in fact
amortized. I doubt it could be measured, but maybe vring_get_buf() should
prefetch? While you're there, we could use an & rather than a mod on the
calculation, which may actually be measurable :)

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)
More majordomo info at
Please read the FAQ at