Re: [RFC] NUMA futex hashing

[Eric Dumazet <dada1@cosmosbay.com>]

On Tuesday 08 August 2006 14:47, Andi Kleen wrote:
> On Tuesday 08 August 2006 14:29, Eric Dumazet wrote:
> > On Tuesday 08 August 2006 12:36, Andi Kleen wrote:
> > > > We may have special case for PRIVATE futexes (they dont need to be
> > > > chained in a global table, but a process private table)
> > >
> > > What do you mean with PRIVATE futex?
> > >
> > > Even if the futex mapping is only visible by a single MM mmap_sem is
> > > still needed to protect against other threads doing mmap.
> >
> > Hum... I would call that a user error.
> >
> > If a thread is munmap()ing the vma that contains active futexes, result
> > is undefined.
>
> We can't allow anything that could crash the kernel, corrupt a kernel,
> data structure, allow writing to freed memory etc.  No matter how
> defined it is or not. Working with a vma that doesn't have
> an existence guarantee would be just that.

As I said, we do not walk the vmas anymore, no crashes are ever possible.

Just keep a process private list of 'private futexes' , indexed by their 
virtual address. This list can be of course stored in a efficient data 
structure, an AVL or RB tree, or hash table.

The validity of the virtual address is still tested by normal get_user() 
call.. If the memory was freed by a thread, then a normal EFAULT error will 
be reported... eventually.

Eric