Re: [PATCH] vfs: keep inodes with page cache off the inode shrinker LRU

[Johannes Weiner]

On Wed, Feb 12, 2020 at 10:26:45AM -0800, Andrew Morton wrote:
> On Wed, 12 Feb 2020 11:35:40 -0500 Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
> 
> > Since the cache purging code was written for highmem scenarios, how
> > about making it specific to CONFIG_HIGHMEM at least?
> 
> Why do I have memories of suggesting this a couple of weeks ago ;)

Sorry, you did. I went back and found your email now. It completely
slipped my mind after that thread went off into another direction.

> > That way we improve the situation for the more common setups, without
> > regressing highmem configurations. And if somebody wanted to improve
> > the CONFIG_HIGHMEM behavior as well, they could still do so.
> > 
> > Somethig like the below delta on top of my patch?
> 
> Does it need to be that complicated?  What's wrong with
> 
> --- a/fs/inode.c~a
> +++ a/fs/inode.c
> @@ -761,6 +761,10 @@ static enum lru_status inode_lru_isolate
>  		return LRU_ROTATE;
>  	}
>  
> +#ifdef CONFIG_HIGHMEM
> +	/*
> +	 * lengthy blah
> +	 */
>  	if (inode_has_buffers(inode) || inode->i_data.nrpages) {
>  		__iget(inode);
>  		spin_unlock(&inode->i_lock);
> @@ -779,6 +783,7 @@ static enum lru_status inode_lru_isolate
>  		spin_lock(lru_lock);
>  		return LRU_RETRY;
>  	}
> +#endif

Pages can show up here even under !CONFIG_HIGHMEM. Because of the lock
order to maintain LRU state (i_lock -> xa_lock), when the page cache
inserts new pages it doesn't unlink the inode from the LRU atomically,
and the shrinker might get here before inode_pages_set(). In that case
we need the shrinker to punt the inode off the LRU (the #else branch).

>  	WARN_ON(inode->i_state & I_NEW);
>  	inode->i_state |= I_FREEING;
> _
> 
> Whatever we do will need plenty of testing.  It wouldn't surprise me
> if there are people who unknowingly benefit from this code on
> 64-bit machines.

If we agree this is the way to go, I can put the patch into our tree
and gather data from the Facebook fleet before we merge it.