On Tue, Sep 10, 2013 at 4:37 PM, Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> From a quick look, this looks pretty broken:
>
> if (list_lru_add(&dentry->d_sb->s_dentry_lru, &dentry->d_lru))
> this_cpu_inc(nr_dentry_unused);
> dentry->d_flags |= DCACHE_LRU_LIST;
>
> because if that list_lru_add() can fail, then we shouldn't set the
> DCACHE_LRU_LIST bit either.
>
> That said, I don't see how it can fail. We only do this with the
> dentry locked, and when it's not already on the LRU list. So I think
> the "if()" is just misleading and unnecessary - but the code works.
So I thought you'd clean this up. Looking again, it still seems really
confused, and I'm finding actual bugs.
You don't clear the DCACHE_LRU_LIST when you remove dentries from the
d_lru list. In other cases (like shrink_dentry_list), you clear just
the DCACHE_SHRINK_LIST.
As a result, the "if ()" isn't necessarily unnecessary, but there are
actual bugs. It looks like the dentry can be removed from the d_lru
lists without the bit ever getting cleared, and if that happens, it
will never be moved back.
The rule for DCACHE_LRU_LIST was - and should be - that the bit is set
IFF the d_lru list is not empty. So it gets set when a dentry is moved
to the LRU lists, but it _stays_ set if the dentry is moved to the
shrink_list. It then gets cleared when the dentry is removed from any
d_lru list (ie "list_del_init()").
I'll walk through the code, it looked suspicious. Maybe there's
something subtle that makes it work, but I don't see it.
Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-next" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
