Xen devel - Mar 2009 - atomic counters larger than 24 bits?

In tmem, I have some counters that I need to keep coherent,
preferably without a lock, so I''m using atomic_inc/dec etc.
However the comments in include/asm-x86/atomic.h say repeatedly
that the "useful range of an atomic_t is only 24 bits".

First, are the comments true in the xen implementation?
(I don''t see any code that restricts them to anything less
than "int".)  If they ARE limited to 24 bits, what''s the
recommended approach for larger atomic operations... or
do I just need to always use a lock?

Thanks,
Dan

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Dan Magenheimer, le Thu 26 Mar 2009 19:38:13 +0000, a écrit
:
> In tmem, I have some counters that I need to keep coherent,
> preferably without a lock, so I''m using atomic_inc/dec etc.
> However the comments in include/asm-x86/atomic.h say repeatedly
> that the "useful range of an atomic_t is only 24 bits".
IIRC the historical reason for this is that some Linux architectures
weren''t providing more.  The x86 implementation does support signed 32
bits. I believe the ia64 implementation does too.

Samuel

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

On 26/03/2009 23:26, "Samuel Thibault"
<samuel.thibault@ens-lyon.org> wrote:
> Dan Magenheimer, le Thu 26 Mar 2009 19:38:13 +0000, a écrit :
>> In tmem, I have some counters that I need to keep coherent,
>> preferably without a lock, so I''m using atomic_inc/dec etc.
>> However the comments in include/asm-x86/atomic.h say repeatedly
>> that the "useful range of an atomic_t is only 24 bits".
> 
> IIRC the historical reason for this is that some Linux architectures
> weren''t providing more.  The x86 implementation does support
signed 32
> bits. I believe the ia64 implementation does too.
Yes, I''ll fix the x86/atomic.h comments. Our atomic_t type is signed
and at
least 32 bits. I don''t see we''ll ever port to an architecture
without
cmpxchg or similar to easily support ''full width'' atomic_t.

 -- Keir



_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Dan Magenheimer wrote:
> In tmem, I have some counters that I need to keep coherent,
> preferably without a lock, so I''m using atomic_inc/dec etc.
> However the comments in include/asm-x86/atomic.h say repeatedly
> that the "useful range of an atomic_t is only 24 bits".
>   
Old sparcs had no way to implement proper atomics, so they used to 
reserve a byte in the atomic_t as a spinlock.  But I don''t think
that''s
true any more; they have a separate lock hash, and atomics in the kernel 
are at least 32-bits.
> First, are the comments true in the xen implementation?
> (I don''t see any code that restricts them to anything less
> than "int".)  If they ARE limited to 24 bits, what''s the
> recommended approach for larger atomic operations... or
> do I just need to always use a lock?
>   
I don''t think the limitation ever applied to Xen.

    J

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel