Lockless patterns: more read-modify-write operations

pbonzini — Sun, 21 Mar 2021 11:12:56 +0000

Yes, it is.

I agree that one should use preexisting APIs; rather than
atomic_dec_and_lock the article points to kref_put_lock as the higher-level API that you want to use, I guess to some extent it's a matter of taste.

Lockless patterns: more read-modify-write operations

dgc — Sat, 20 Mar 2021 23:48:13 +0000

Isn't your put_gadget() example just demonstrating why the atomic_dec_and_lock() primitive exists?

Wouldn't it be better to teach people this pattern for atomically grabbing a lock when dropping the last reference to an object as it is much easier to understand and hard to get wrong?

put_gadget()
{
       if (!atomic_dec_and_lock(&g->refcnt, &lock))
             return;
       list_del(&g->node);
       spin_unlock(&lock);
}

This is assuming, of course, that an unlocked get side is using atomic_inc_not_zero() to avoid get/put races...

-Dave.

LWN: Comments on "Lockless patterns: more read-modify-write operations"

Lockless patterns: more read-modify-write operations

Lockless patterns: more read-modify-write operations