Re: netpoll + xmit_lock == deadlock

[Matt Mackall <mpm@selenic.com>]

On Mon, 2009-08-03 at 12:15 -0700, David Miller wrote:
> From: Ben Hutchings <bhutchings@solarflare.com>
> Date: Mon, 03 Aug 2009 13:05:57 +0100
> 
> > On Sun, 2009-08-02 at 13:07 -0700, David Miller wrote:
> >> My position has always been that such printk's are simply
> >> not allowed.  (check archives if you don't believe me :-)
> >>
> >> The locking is going to get rediculious if we start having
> >> to account for this.
> > 
> > I agree with that, but this does seem quite restrictive.  How can we be
> > sure that none of the kernel functions used by a driver's TX path (e.g.
> > kmalloc or DMA-mapping) will print debug or warning messages?  If such
> > guarantees exist, they do not seem to be documented.
> 
> Indeed, that's a good counter-argument.
> 
> Ho hum... so someone show me how ugly the locking is going
> to get :-)

First, I don't think we can solve the problem of 'reliably deliver
printks inside the TX path'. If the driver needs to printk here, odds
are good that sending isn't possible.

So I think we should be clear up front that the solution we're looking
for is 'don't crash the box when trying to printk from the tx path'.

I think the most straightforward, driver-agnostic way to do this
is to have netpoll wrap a driver's TX entrypoint such that recursion is
disabled while in that path. This might actually simplify netpoll's
internal locking ugliness. Something like:

int netpoll_xmit_wrapper(struct sk_buff *skb, struct net_device *dev)
{
	struct netpoll_info *npinfo = dev->npinfo;
	int ret;

	npinfo->recurse += 1; /* add appropriate locking */
	ret = npinfo->orig_tx(skb, dev);
	npinfo->recurse -= 1;
	
	return ret;
}

-- 
http://selenic.com : development and support for Mercurial and Linux