Fwd: [RFC][PATCH 4/5] I/OAT DMA support and TCP acceleration

Sorry, bounced from vger the first time due to formatting

---------- Forwarded message ----------
From: Chris Leech <redacted>
Date: Dec 20, 2005 11:05 PM
Subject: Re: [RFC][PATCH 4/5] I/OAT DMA support and TCP acceleration
To: Eric Dumazet <redacted>
Cc: lkml <redacted>, netdev
[off-list ref], "Grover, Andrew" [off-list ref],
"Ronciak, John" [off-list ref]

On 12/20/05, Eric Dumazet [off-list ref] wrote:

 Please consider not enlarging cb[] if not CONFIG_NET_DMA ?

I mean, most machines wont have a compatable NIC, so why should they pay the
price (memory, cpu) in a critical structure named sk_buff ?

#ifdef CONFIG_NET_DMA
typedef dma_cookie_t net_dma_cookie_t;
#else
typedef struct {} net_dma_cookie_t;
#endif

...

        char   cb[40+sizeof(net_dma_cookie_t)];

 That could be a good way to deal with it.  Actually, I should double
check the length of tcp_skb_cb.  I took a quick look and thought that
there might be some room left there anyway, even though the comment in
tcp.h says otherwise.

 -Chris

From: Chris Leech <redacted>
Date: Tue, 20 Dec 2005 23:10:07 -0800

 That could be a good way to deal with it.  Actually, I should double
check the length of tcp_skb_cb.  I took a quick look and thought that
there might be some room left there anyway, even though the comment in
tcp.h says otherwise.

There isn't, it's basically full on 64-bit systems with ipv6 enabled.

Just put the DMA cookie object explicitly into struct sk_buff,
protected by CONFIG_NET_DMA.