Re: [PATCH 1/2] list.h: add debug version of list_empty

[Phil Carmody <ext-phil.2.carmody@nokia.com>]

On 21/03/11 16:52 -0700, ext Andrew Morton wrote:
> On Tue, 15 Mar 2011 15:08:42 +0200
> Phil Carmody <ext-phil.2.carmody@nokia.com> wrote:
> > +int list_empty(const struct list_head *head)
> > +{
> > +	if ((head->prev == LIST_POISON2) || (head->prev == LIST_POISON1))
> > +		WARN(1, "list_empty performed on a node "
> > +		     "at %p removed from a list.\n", head);
> > +	else
> > +		WARN((head->prev == head) != (head->next == head),
> > +		     "list_empty corruption. %p<-%p->%p is half-empty.\n",
> > +		     head->prev, head, head->next);
 
> The second warning here is triggering maybe a hundred times from all
> over the place just when booting the kernel.
> 
> Here's the first two:
> 
> [   64.295941] WARNING: at lib/list_debug.c:89 list_empty+0x79/0x85()
> [   64.296129] list_empty corruption. ffff880255bcb788<-ffff880255bcb788->ffff88024c3a3c20 is half-empty.


OK, so the patch is working as expected. Perhaps my expectations were wrong.

Looking at list.h I was sure that lists should always be either circular or
poisoned both ends. The above is a rho-shape, this == prev.

Traditional list_empty() returns false on such a node, so it should be
possible to list_del() it. But then next->prev will be set to this->prev 
which is this. So this will never be deleted from the list. That situation
rings warning bells in my head. Which I guess is what the patch was trying
to concretise.

I presume the above are x86_64, I'll see if I can get access to such a 
machine in the next few days, or reproduce it on one of the architectures
I do have here.

Phil