Re: [PATCH] splice : two smp_mb() can be omitted

[Nick Piggin <nickpiggin@yahoo.com.au>]

Eric Dumazet wrote:

>On Tuesday 31 October 2006 10:40, Nick Piggin wrote:
>
>
>>Uh, there is nothing that says mutex_unlock or any unlock
>>functions contain an implicit smp_mb(). What is given is that the
>>lock and unlock obey aquire and release memory ordering,
>>respectively.
>>
>>a = x;
>>xxx_unlock
>>b = y;
>>
>>In this situation, the load of y can be executed before that of x.
>>And some architectures will even do so (i386 can, because the
>>unlock is an unprefixed store; ia64 can, because it uses a release
>>barrier in the unlock).
>>
>
>Hum... it seems your mutex_unlock() i386/x86_64 copy is not same as mine :)
>

OK, replace xxx with mutex, and what I've said still holds true for ia64.

>Maybe we could document the fact that mutex_{lock|unlock}() has or has not an 
>implicit smp_mb().
>

It does not, none of the unlock functions ever have.

>If not, delete smp_mb() calls from include/asm-generic/mutex-dec.h 
>

They should be deleted (and from mutex-xchg). NOT because there is no 
need for
a memory barrier, but because the atomic_alter_value_and_return_something
functions always provide a barrier before and after the operation, as per
Documentation/atomic_ops.txt

Again, lock / unlock operations require acquire / release consistency. 
This is a
memory ordering operation. It is not equivalent to smp_mb, though.

--

Send instant messages to your online friends http://au.messenger.yahoo.com