Re: PROBLEM: System call 'sendmsg' of process ospfd (quagga) causes kernel oops

Le mardi 18 octobre 2011 à 12:45 +0200, Herbert Xu a écrit :

On Tue, Oct 18, 2011 at 12:23:43PM +0200, Eric Dumazet wrote:

quoted

You're right, if reallocations are OK in all paths.

If it wasn't OK then making needed_headroom constant won't work
anyway.

quoted

We'll need to change LL_RESERVED_SPACE() / LL_RESERVED_SPACE_EXTRA() /
LL_ALLOCATED_SPACE() macros and provide the [read once] values, instead
of a [read once] pointer to values.

I'm not sure what you mean here.  I don't see any need to change
these macros.  All we need is to save the value in a local variable:

	hh_len = LL_RESERVED_SPACE(dev);

	skb = alloc_skb(hh_len + len);
	skb_reserve(skb, hh_len);

Not really Herbert. Please read again my patch changelog.

In the bug we try to fix, we have :

skb = sock_alloc_send_skb(sk, ... + LL_ALLOCATED_SPACE(rt->dst.dev) 

... < increase of dev->needed_headroom by another cpu/task >

skb_reserve(skb, LL_RESERVED_SPACE(rt->dst.dev));

skb_put() -> crash because we reserved too much space

So we really want LL_ALLOCATED_SPACE() and LL_RESERVED_SPACE() use the
same needed_headroom, or else you can have LL_RESERVED_SPACE() >
LL_ALLOCATED_SPACE().

There are several way to fix this, but this kind of code assumed the
dev->needed... values were consistent for the whole block.