Re: [PATCH 1/5] pid: Implement task references.

[Eric Dumazet <dada1@cosmosbay.com>]

Kyle Moffett a écrit :
> On Jan 30, 2006, at 00:19, Eric Dumazet wrote:
>> -    if (atomic_dec_and_test(&kref->refcount)) {
>> +    /*
>> +     * if current count is one, we are the last user and can release 
>> object
>> +     * right now, avoiding an atomic operation on 'refcount'
>> +     */
>> +    if ((atomic_read(&kref->refcount) == 1) ||
> 
> Uhh, I think you got this test reversed.  Didn't you mean != 1?  
> Otherwise you only do the dec_and_test when the refcount is one, which 
> means that you leak everything kref-ed.
> 

Not at all :)

Your mail is just another proof why kref is a good abstraction :)

If you are the last user of a kref, (refcount = 1), then
you are sure that nobody else but you is using the object, and as we are 
kref_put() this object, the atomic_dec_and-test *will* set the count the 
object and you are going to release() object.

The release() function is not going to look at kref_count again, just free the 
resources and the object.

Maybe a change in the documentation is necessary to explain this point 
(release() can e called while the apparent krefcount is 1)

Or in kref_put doing this :

if (atomic_read(&kref->refcount) == 1) {
	atomic_set(&kref->refcount, 0);
	release(kref);
}
if (atomic_dec_and_test(&kref->refcount)) {
         release(kref);
         return 1;
}


Eric