Re: linux-next: manual merge of the akpm tree with Linus' tree

[Al Viro <viro@ZenIV.linux.org.uk>]

On Fri, Sep 13, 2013 at 12:12:20PM -0700, Linus Torvalds wrote:

> It tries to consolidate the dentry LRU stuff into a few helper
> functions that right now have anal checking of the flags. Maybe I
> overdid it, but the code was really confusing, and I think we got the
> free dentry counts wrong, and the bits wrong too, so I tried to be
> extra careful.
> 
> There are several cases:
>  - d_lru_add/del: fairly obvious
>  - d_lru_isolate: this is when the LRU callbacks ask us to remove the
> entry from the list. This is different from d_lru_del() only in that
> it uses the raw list removal, not the lru list helper function. I'm
> not sure that's right, but that's what the code used to do.

It is right - for one thing, we are holding the lock on that LRU list,
so list_lru_del() would deadlock right there.  For another, the same
list_lru_walk (OK, list_lru_walk_node()) will do ->nr_items decrement
when we return LRU_REMOVED to it, so we don't want to do it twice.
Plain list_del_init() is correct here.

>  - d_lru_shrink_move: move from the "global" lru list to a private shrinker list
>  - d_shrink_add/del: fairly obvious.
> 
> And then "denty_lru_add/del" that actually take the current state into
> account and do the right thing. Those we had before, I'm just
> explaining the difference from the low-level operations that have
> fixed "from this state to that" semantics

Looks sane; FWIW, the variant I'm playing with uses two independent
flags for "shrinker" and "per-sb", but AFAICS that doesn't yield better
code.  Feel free to slap my acked-by on it.