Re: [RFC PATCH v2 03/14] xsk: add umem fill queue support and mmap

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Mar 27, 2018 at 06:59:08PM +0200, Björn Töpel wrote:
> @@ -30,4 +31,18 @@ struct xdp_umem_reg {
>  	__u32 frame_headroom; /* Frame head room */
>  };
>  
> +/* Pgoff for mmaping the rings */
> +#define XDP_UMEM_PGOFF_FILL_QUEUE	0x100000000
> +
> +struct xdp_queue {
> +	__u32 head_idx __attribute__((aligned(64)));
> +	__u32 tail_idx __attribute__((aligned(64)));
> +};
> +
> +/* Used for the fill and completion queues for buffers */
> +struct xdp_umem_queue {
> +	struct xdp_queue ptrs;
> +	__u32 desc[0] __attribute__((aligned(64)));
> +};
> +
>  #endif /* _LINUX_IF_XDP_H */

So IIUC it's a head/tail ring of 32 bit descriptors.

In my experience (from implementing ptr_ring) this
implies that head/tail cache lines bounce a lot between
CPUs. Caching will help some. You are also forced to
use barriers to check validity which is slow on
some architectures.

If instead you can use a special descriptor value (e.g. 0) as
a valid signal, things work much better:

- you read descriptor atomically, if it's not 0 it's fine
- same with write - write 0 to pass it to the other side
- there is a data dependency so no need for barriers (except on dec alpha)
- no need for power of 2 limitations, you can make it any size you like
- easy to resize too

architecture (if not implementation) would be shared with ptr_ring
so some of the optimization ideas like batched updates could
be lifted from there.

When I was building ptr_ring, any head/tail design underperformed
storing valid flag with data itself. YMMV.

-- 
MST