Re: Problem with dev_kfree_skb_any() in 2.6.0

[Jeff Garzik <jgarzik@pobox.com>]

On Mon, Dec 29, 2003 at 08:51:57PM -0800, David S. Miller wrote:
> On Mon, 29 Dec 2003 23:09:14 -0500
> Jeff Garzik <jgarzik@pobox.com> wrote:
> 
> > Not really...  pretty much _all_ TX queue packet freeing occurs inside 
> > an irq handler and inside the driver spinlock.  Further, we don't want 
> > to reinvent some sort of "queue skb for freeing" code in every driver.
> 
> There is one important detail not mentioned.
> 
> If we let the TX free occur in cpu IRQ disabled context, the
> BH to actually do the work will occur as some indeterminate
> time in the future after the top level IRQ spinlock release
> occurs.
> 
> Unlike local_bh_enable(), local_irq_enable() does not run
> softirq work.  Similarly when comparing IRQ handler return
> (which also runs softirq work if pending).
> 
> This is the most important reason why the suggested change is wrong.

OK, agreed.  But fixing it in the driver is still incorrect, also.

We need a single solution in the net stack, not a per-driver solution.

Look at the purpose behind his patch...

	Jeff