Re: [PATCH] MTRR: Fix race causing set_mtrr to go into infinite loop

[Loic Prylli <loic@myri.com>]

On 6/25/2007 6:34 PM, Andi Kleen wrote:
> On Tuesday 26 June 2007 00:05:17 Chuck Ebbert wrote:
>   
>> On 06/25/2007 05:38 PM, Loic Prylli wrote:
>>
>> [cc: Andi]
>>
>>     
>>> Processors synchronization in set_mtrr requires the .gate field
>>> to be set after .count field is properly initialized. Without an explicit
>>> barrier, the compiler was reordering those memory stores. That was sometimes
>>> causing a processor (in ipi_handler) to see the .gate change and
>>> decrement .count before the latter is set by set_mtrr() (which
>>> then hangs in a infinite loop with irqs disabled).
>>>       
>
> Hmm, perhaps we should just put the smp_wmb into atomic_set().
> Near all other atomic operations have memory barriers too. I think
> that would be the better fix.
>
> -Andi
>   


In Documentation/atomic_ops.txt atomic_set/atomic_read are described as
nothing more than a type-safe assignement or reading, without any extra
semantics. For other atomic operations, the rule is that any atomic
operation that doesn't return a value doesn't come with a barrier (and
any operation that returns the atomic value must have memory barriers).

So I guess you are suggesting to change the doc and the implementation
for all arches.

I should admit I did not knew a number of atomic operations did not
imply memory-barriers before. But maybe the extra cost might not be
completely negligible, especially if, for consistency with other
"barrier-implied" atomic operations, a new memory barrier is put before
and after,

Are you suggested changing just atomic_set(), or also other barrier-free
atomic operations :"atomic_dec", "atomic_inc", "atomic_add", "atomic_sub" ?

Independently of what is done to atomic, what about not making the .gate
field an atomic_t, but a simple "int" in the mttr code, since the only
operations done on it are atomic_read and atomic_set?


Loic