Re: [PATCH 3/8] rhashtable: use cmpxchg() to protect ->future_tbl.

From: Herbert Xu
Date: Sat May 05 2018 - 05:27:41 EST


On Fri, May 04, 2018 at 01:54:14PM +1000, NeilBrown wrote:
> Rather than borrowing one of the bucket locks to
> protect ->future_tbl updates, use cmpxchg().
> This gives more freedom to change how bucket locking
> is implemented.
>
> Signed-off-by: NeilBrown <neilb@xxxxxxxx>

This looks nice.

> - spin_unlock_bh(old_tbl->locks);
> + rcu_assign_pointer(tmp, new_tbl);

Do we need this barrier since cmpxchg is supposed to provide memory
barrier semantics?

> + if (cmpxchg(&old_tbl->future_tbl, NULL, tmp) != NULL)
> + return -EEXIST;

Thanks,
--
Email: Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt