Message-ID: <1349434194.16710.44.camel@sakura.staff.proxad.net>
Subject: Re: kernel 3.2.27 on arm: WARNING: at mm/page_alloc.c:2109
 __alloc_pages_nodemask+0x1d4/0x68c()
From: Maxime Bizon <mbizon@freebox.fr>
Reply-To: mbizon@freebox.fr
To: Eric Dumazet <eric.dumazet@gmail.com>
Cc: David Madore <david+ml@madore.org>, Francois Romieu <romieu@fr.zoreil.com>,
        netdev@vger.kernel.org, linux-kernel@vger.kernel.org,
        Hugh Dickins <hughd@google.com>
Date: Fri, 05 Oct 2012 12:49:54 +0200
In-Reply-To: <1349422868.21172.38.camel@edumazet-glaptop>
References: <20120829002548.GA7063@aldebaran.gro-tsen.net>
	 <alpine.LSU.2.00.1208311900330.1936@eggly.anvils>
	 <1349366521.2532.12.camel@sakura.staff.proxad.net>
	 <1349368171.16011.79.camel@edumazet-glaptop>
	 <1349422868.21172.38.camel@edumazet-glaptop>
Organization: Freebox
Content-Type: text/plain; charset="ANSI_X3.4-1968"
Content-Transfer-Encoding: 7bit
Mime-Version: 1.0
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1997
Lines: 59


On Fri, 2012-10-05 at 09:41 +0200, Eric Dumazet wrote:

> By the way, the commit you pointed has no effect on the reallocation
> performed by pskb_expand_head() :

The commit has a side effect, because the problem appeared after it was
merged (and goes away if I revert it)

> int size = nhead + skb_end_offset(skb) + ntail;
> 
> So pskb_expand_head() always assumed the current head is fully used, and
> because we have some kmalloc-power-of-two contraints, each time
> pskb_expand_head() is called with a non zero (nhead + ntail) we double
> the skb->head ksize.

That is true, but only after the commit I mentioned.

Before that commit, we indeed reallocate skb->head to twice the size,
but skb->end is *not* positioned at the end of newly allocated data. So
on the next pskb_expand_head(), if head and tail are not big values, the
kmalloc() will be of the same size.


The commit adds this after allocation:

size = SKB_WITH_OVERHEAD(ksize(data))
[...]
skb->end      = skb->head + size;

so on the next pskb_expand_head, we are going to allocate twice the size
for sure.

> So why are we using skb_end_offset(skb) here is the question.
> 
> I guess it could be (skb_tail_pointer(skb) - skb->head) on some uses.

I think your patch is wrong, ntail is not the new tailroom size, it's
what missing to the current tailroom size, by adding ntail + nhead +
tail_offset we are removing previous tailroom.

We cannot shrink the skb that way here I guess, a caller may check
needed headroom & tailroom, calls with nhead=1/ntail=0 because only
headroom is missing, but after the call tailroom would be less than
before the call.

Why don't we juste reallocate to this size:

MAX(current_alloc_size, nhead + ntail + current_end - current_head)

-- 
Maxime


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/