kernel 3.2.27 on arm: WARNING: at mm/page_alloc.c:2109 __alloc_pages_nodemask+0x1d4/0x68c()

On Fri, 2012-10-05 at 14:51 +0200, Maxime Bizon wrote:

> > New convention would be : pass number of needed bytes after current
> > tail, not after current end.
> 
> Fully agree on this
> 

Here is the proposal :

Change all occurrences of :

pskb_expand_head(skb, 0, 0, gfp)

to

pskb_realloc_head(skb, gfp)   
[ an inline calling a common function ]



And few other pskb_expand_head() calls a new function

pskb_may_expand_head(skb, nhead, ntail, gfp) 
[ an inline calling same common function ]


nhead : number of needed bytes before [data,tail] portion
         (might be already < skb_headroom(skb))

ntail : number of needed bytes after [data,tail] portion
        (might be already > skb_tailroom(skb))

For example, net/xfrm/xfrm_output.c would be like the following :

(Note how all the various ugly tests will be centralized in the common
function)

And the common function would now know the caller intent.




--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

On Fri, 2012-10-05 at 15:02 +0200, Eric Dumazet wrote:

> On Fri, 2012-10-05 at 14:51 +0200, Maxime Bizon wrote:
> 
> > > New convention would be : pass number of needed bytes after current
> > > tail, not after current end.
> > 
> > Fully agree on this
> > 
> 
> Here is the proposal :

Looks fine

What is your plan for the actual pskb_expand_head() code now that you
will have absolute values for headroom & tailroom ?

Because there will still be callers that have no clue of required
tailroom (nor further headroom requirement), like skb_cow() in
vlan_reorder_header().

On Fri, 2012-10-05 at 16:50 +0200, Maxime Bizon wrote:
> On Fri, 2012-10-05 at 15:02 +0200, Eric Dumazet wrote:
> 
> > On Fri, 2012-10-05 at 14:51 +0200, Maxime Bizon wrote:
> > 
> > > > New convention would be : pass number of needed bytes after current
> > > > tail, not after current end.
> > > 
> > > Fully agree on this
> > > 
> > 
> > Here is the proposal :
> 
> Looks fine
> 
> What is your plan for the actual pskb_expand_head() code now that you
> will have absolute values for headroom & tailroom ?
> 

They stay relative values.

For example, netlink_trim() really wants to shrink the skb head.


> Because there will still be callers that have no clue of required
> tailroom (nor further headroom requirement), like skb_cow() in
> vlan_reorder_header().
> 

What I plan is to not shrink size, unless specifically asked.

Its 3.8 material anyway, so a stable fix is needed on skb_recycle() and
NET_SKB_PAD minimal value.



--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

On Fri, 2012-10-05 at 17:04 +0200, Eric Dumazet wrote:
> 
> 
> What I plan is to not shrink size, unless specifically asked.
> 
> Its 3.8 material anyway, so a stable fix is needed on skb_recycle()
> and NET_SKB_PAD minimal value.

You think removing skb_recycle() is too big a change for stable ?

Driver change is simple, as recycling is not guaranteed today you have
this:

 if (!try_recycle(skb))
    skb = alloc_skb()
> 
we just remove the try_recycle part, we are not adding any new code
path.


I'm not the right person to give you a correct NET_SKB_PAD value, I have
a lot of out of tree patches in my kernels, some custom drivers as well
that needs quite a lot of headroom, and uncommon network setup.

On Fri, 2012-10-05 at 17:15 +0200, Maxime Bizon wrote:

> You think removing skb_recycle() is too big a change for stable ?
> 
> Driver change is simple, as recycling is not guaranteed today you have
> this:
> 
>  if (!try_recycle(skb))
>     skb = alloc_skb()
> > 
> we just remove the try_recycle part, we are not adding any new code
> path.
> 
> 
> I'm not the right person to give you a correct NET_SKB_PAD value, I have
> a lot of out of tree patches in my kernels, some custom drivers as well
> that needs quite a lot of headroom, and uncommon network setup.
> 

OK, I'll send the patch to remove skb_recycle()


--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html