59b
Subj : Re: A question about atomic_ptr
To   : comp.programming.threads
From : Joe Seigh
Date : Sat Apr 16 2005 08:31 am

On Fri, 15 Apr 2005 22:03:14 -0700, Chris Thomasson <_no_damn_spam_cristom@_no_damn_comcast.net_spam> wrote:

>>> Ok, Joe said that "It needs the virtual storage to always be valid so
>>> the load locked won't segfault."
>>>
>>> Not sexy.
>>>
>
> [...]
>
>> BTW, it's the same restriction lock-free LIFO stack using DWCAS w/ version
>> counts has.
>
> So, atomic_ptr for PPC-64 needs to have some sort of static/tracked storage
> wrt atomic_ptr_ref's? You can get around all of that by allocating static or
> "collected" arrays of atomic_ptr_ref's. Then you would use the index to
> atomic_ptr_ref as the "pointer" to it.

PPC-64 has ldarx/stdcx.  It just needs the storage to always be valid for
loading from event if its no longer a valid object.  Same as the lock-free
LIFO stack needs when it does a load of the next pointer field in what it
thinks is the item on the top of the stack.  Not until it does the dwcas
on the stack head/version number does it validate that the value it
fetched was good.

This is different than the lock-free FIFO queue which needs to pool the
nodes since the storage value fetched does matter.

DWCAS based refcounting doesn't have this problem since the refcount is
incremented atomically with the fetch of the pointer.

-- 
Joe Seigh

.

0