Re: [PATCH net-next-2.6] sfc: Reduce size of efx_rx_buffer by unionising skb and page

[Ben Hutchings <bhutchings@solarflare.com>]

On Tue, 2011-03-01 at 12:24 -0800, David Miller wrote:
> From: Ben Hutchings <bhutchings@solarflare.com>
> Date: Tue, 01 Mar 2011 01:21:30 +0000
> 
> > From: Steve Hodgson <shodgson@solarflare.com>
> > 
> > [bwh: Forward-ported to net-next-2.6.]
> > Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
> 
> I'm going to apply this, but you can take this further by making
> the buffer an "unsigned long" and encode the pointer plus a one-bit
> binary state there to indicate what it is.

We were primarily concerned with reducing the size of the software ring
for RX queues on 64-bit (DMA address & pointer) systems, where the
structure layout is:

Original:

struct efx_rx_buffer {
       dma_addr_t dma_addr;		// 0
       struct sk_buff *skb;		// 8
       struct page *page;		// 16
       char *data;			// 24
       unsigned int len;		// 32
};					// size = 40

Current:

struct efx_rx_buffer {
       dma_addr_t dma_addr;		// 0
       union {
               struct sk_buff *skb;	// 8
               struct page *page;	// 8
       } u;
       unsigned int len;		// 16
       bool is_page;			// 20
};					// size = 24

Combining the flag with the pointers wouldn't improve on this:

struct efx_rx_buffer {
       dma_addr_t dma_addr;		// 0
       unsigned long magic_pointer;	// 8
       unsigned int len;		// 16
};					// size = 24

Ben.

-- 
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.