Re: Q. hlist_bl_add_head_rcu() in d_alloc_parallel()

On Sat, Jun 25, 2016 at 06:25:58PM -0700, Linus Torvalds wrote:

On Sat, Jun 25, 2016 at 3:54 PM, Al Viro [off-list ref] wrote:

quoted

BTW, speaking of that area - is there any reason why dentry_cmp()
isn't simply
        return dentry_string_cmp(lockless_dereference(dentry->d_name.name),
                                 ct, tcount);
other than "it predates lockless_dereference()"?

No, the only important thing is that it's a read-once, and has the
smp_read_barrier_depends().

So lockless_derefence() looks like the right thing to do, it is just
that - as you suspected - the code predates that concept.

... and as the matter of fact, patch to that effect (only cosmetically different
from what I'd cooked) had been posted by He Kuang back in March.  At least
I'd spotted that before pushing mine out...

The patch posted back then applied, with slightly edited commit message.  Apologies
for missing it...