Re: [PATCH 1/2] af_packet: use vmalloc_to_page() instead for the addresss returned by vmalloc()

[Eric Dumazet <eric.dumazet@gmail.com>]

Le mardi 30 novembre 2010 à 21:56 +0800, Changli Gao a écrit :
> The following commit causes the pgv->buffer may point to the memory
> returned by vmalloc(). And we can't use virt_to_page() for the vmalloc
> address.
> 
> This patch introduces a new inline function pgv_to_page(), which calls
> vmalloc_to_page() for the vmalloc address, and virt_to_page() for the
> __get_free_pages address.
> 
>     commit 0e3125c755445664f00ad036e4fc2cd32fd52877
>     Author: Neil Horman <nhorman@tuxdriver.com>
>     Date:   Tue Nov 16 10:26:47 2010 -0800
> 
>     packet: Enhance AF_PACKET implementation to not require high order contiguous memory allocation (v4)
> 

nice catch.

> Signed-off-by: Changli Gao <xiaosuo@gmail.com>
> ---
>  net/packet/af_packet.c |   21 ++++++++++++++-------
>  1 file changed, 14 insertions(+), 7 deletions(-)
> diff --git a/net/packet/af_packet.c b/net/packet/af_packet.c
> index 422705d..0171b20 100644
> --- a/net/packet/af_packet.c
> +++ b/net/packet/af_packet.c
> @@ -224,6 +224,13 @@ struct packet_skb_cb {
>  
>  #define PACKET_SKB_CB(__skb)	((struct packet_skb_cb *)((__skb)->cb))
>  
> +static inline struct page *pgv_to_page(void *addr)
> +{
> +	if (is_vmalloc_addr(addr))
> +		return vmalloc_to_page(addr);

Hmm, I am wondering if calling vmalloc_to_page(addr) several times for
each packet is not too expensive ? I believe it is.

What about caching "struct page *" pointer somewhere ?

Then later we have :

> -		p_start = virt_to_page(h.raw);
> -		p_end = virt_to_page(h_end);
> +		p_start = pgv_to_page(h.raw);
> +		p_end = pgv_to_page(h_end);
>  		while (p_start <= p_end) {
>  			flush_dcache_page(p_start);
>  			p_start++;

This was OK before Neil patch... after vmalloc(), assumption that
p_start can be incremented is completely wrong.

To fix this, we need something else than your patch.