Re: 2.5.70-bk16: nfs crash

[Dipankar Sarma <dipankar@in.ibm.com>]

On Thu, Jun 12, 2003 at 09:18:11AM -0700, Linus Torvalds wrote:
> 
> On Thu, 12 Jun 2003, Linus Torvalds wrote:
> > 
> > If you depend on not re-initializing the pointers, you should not use the 
> > "xxx_del()" function, and you should document it.
> 
> Besides, the code doesn't actually depend on not re-initializing the 
> pointers, it depends on the _forward_ pointers still being walkable in 
> case some other CPU is traversing the list just as we remove the entry.
> 
> Which means that I think the proper patch is to (a) document this and also
> (b) poison the back pointer.

That should work. However, I do have once concern. At the generic
list macro level, we don't know if the lockfree traversal is being
done in forward or backward direction. So, I am not sure if
list_del_rcu() should poison the backward pointer or atleast
document the fact that RCU-based traversal would not work on
backward pointers. hlist_del_rcu() can indeed poison pprev. 


> 
> A patch like the attached, in short.

Since we are at it, I can submit a patch documenting the rest of
hlist functions, if you like.

Thanks
Dipankar